Skip to contentSkip to conversation

rib_trie / Fix inflate_threshold_root. Now=15 size=11 bits

PATCH Applied Paweł Staszewski 10 participants 99 messages Jun 25, 2009 – Jul 20, 2009 1 message not in the archive

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

Hello ALL

Some time ago i report this:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

and now with 2.6.29 / 2.6.29.1 / 2.6.29.3 and 2.6.30 it back dmesg output:
oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

cat /proc/net/fib_triestat
Basic info: size of leaf: 40 bytes, size of tnode: 56 bytes. Main:

       Aver depth:     2.28
       Max depth:      6
       Leaves:         276539
       Prefixes:       289922
       Internal nodes: 66762
         1: 35046  2: 13824  3: 9508  4: 4897  5: 2331  6: 1149  7: 5  
9: 1  18: 1
       Pointers: 691228

Null ptrs: 347928
Total size: 35709 kB

Counters:
---------
gets = 26276593
backtracks = 547306
semantic match passed = 26188746
semantic match miss = 1117
null node hit= 27285055
skipped node resize = 0

Local:

       Aver depth:     3.33
       Max depth:      4
       Leaves:         9
       Prefixes:       10

Internal nodes: 8
1: 8
Pointers: 16
Null ptrs: 0
Total size: 2 kB

Counters:
---------
gets = 26642350
backtracks = 1282818
semantic match passed = 18166
semantic match miss = 0
null node hit= 0
skipped node resize = 0

This machine is running bgpd with two bgp peers / full route table

cat /proc/meminfo

MemTotal:       12279032 kB
MemFree:        11521920 kB
Buffers:           80288 kB
Cached:            34416 kB
SwapCached:            0 kB
Active:           286816 kB
Inactive:          82024 kB
Active(anon):     254296 kB
Inactive(anon):        0 kB
Active(file):      32520 kB
Inactive(file):    82024 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:        987988 kB
SwapFree:         987988 kB
Dirty:              1140 kB
Writeback:             0 kB
AnonPages:        254164 kB
Mapped:             5440 kB
Slab:             365084 kB
SReclaimable:      28784 kB
SUnreclaim:       336300 kB
PageTables:         2104 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     7127504 kB
Committed_AS:     267704 kB
VmallocTotal:   34359738367 kB
VmallocUsed:       11824 kB
VmallocChunk:   34359707815 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       2048 kB
DirectMap4k:        3392 kB
DirectMap2M:    12578816 kB

Interfaces mtu is1500

#1 lore raw
Quoted 15 lines · Paweł Staszewski

Hello ALL

Some time ago i report this:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

and now with 2.6.29 / 2.6.29.1 / 2.6.29.3 and 2.6.30 it back dmesg output:
oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

Quoted 8 lines

Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

Curious, you seem to hit an old alloc_pages limit()... (MAX_ORDER allocation)

Your root node has 2^18 = 262144 pointers of 8 bytes -> 2097152 bytes (+ header -> 4194304 bytes)

But since following commit, we should use vmalloc() so this PAGE_SIZE<<10) limit should not anymore be applied.

Could you do a "cat /proc/vmallocinfo" just to check your big tnodes are vmalloced() ?

commit 15be75cdb5db442d0e33d37b20832b88f3ccd383

Author: Stephen Hemminger <[email protected]>
Date:   Thu Apr 10 02:56:38 2008 -0700

IPV4: fib_trie use vmalloc for large tnodes

Use vmalloc rather than alloc_pages to avoid wasting memory. The problem is that tnode structure has a power of 2 sized array, plus a header. So the current code wastes almost half the memory allocated because it always needs the next bigger size to hold that small header.

This is similar to an earlier patch by Eric, but instead of a list and lock, I used a workqueue to handle the fact that vfree can't be done in interrupt context.

Quoted 90 lines

cat /proc/net/fib_triestat
Basic info: size of leaf: 40 bytes, size of tnode: 56 bytes. Main:

       Aver depth:     2.28
       Max depth:      6
       Leaves:         276539
       Prefixes:       289922
       Internal nodes: 66762
         1: 35046  2: 13824  3: 9508  4: 4897  5: 2331  6: 1149  7: 5 
9: 1  18: 1
       Pointers: 691228

Null ptrs: 347928
Total size: 35709 kB

Counters:
---------
gets = 26276593
backtracks = 547306
semantic match passed = 26188746
semantic match miss = 1117
null node hit= 27285055
skipped node resize = 0

Local:

       Aver depth:     3.33
       Max depth:      4
       Leaves:         9
       Prefixes:       10

Internal nodes: 8
1: 8
Pointers: 16
Null ptrs: 0
Total size: 2 kB

Counters:
---------
gets = 26642350
backtracks = 1282818
semantic match passed = 18166
semantic match miss = 0
null node hit= 0
skipped node resize = 0

This machine is running bgpd with two bgp peers / full route table

cat /proc/meminfo

MemTotal:       12279032 kB
MemFree:        11521920 kB
Buffers:           80288 kB
Cached:            34416 kB
SwapCached:            0 kB
Active:           286816 kB
Inactive:          82024 kB
Active(anon):     254296 kB
Inactive(anon):        0 kB
Active(file):      32520 kB
Inactive(file):    82024 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:        987988 kB
SwapFree:         987988 kB
Dirty:              1140 kB
Writeback:             0 kB
AnonPages:        254164 kB
Mapped:             5440 kB
Slab:             365084 kB
SReclaimable:      28784 kB
SUnreclaim:       336300 kB
PageTables:         2104 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     7127504 kB
Committed_AS:     267704 kB
VmallocTotal:   34359738367 kB
VmallocUsed:       11824 kB
VmallocChunk:   34359707815 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       2048 kB
DirectMap4k:        3392 kB
DirectMap2M:    12578816 kB

Interfaces mtu is1500

#2 lore raw

cat /proc/vmallocinfo
0xf7ffe000-0xf8000000 8192 acpi_tb_verify_table+0x1d/0x46 phys=dfe6a000 ioremap
0xf8000000-0xf8007000 28672 acpi_tb_verify_table+0x1d/0x46 phys=dfef5000 ioremap

0xf8008000-0xf800a000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfef2000 ioremap
0xf800c000-0xf800e000    8192 acpi_ex_system_memory_space_handler+0xd6/0x208 phys=fed1f000 ioremap
0xf8010000-0xf8012000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfefb000 ioremap
0xf8014000-0xf8016000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfef4000 ioremap
0xf8018000-0xf801a000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfef3000 ioremap
0xf801c000-0xf801e000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfef1000 ioremap
0xf8020000-0xf8022000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfef0000 ioremap
0xf8024000-0xf8026000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfeef000 ioremap
0xf8028000-0xf802a000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfeee000 ioremap
0xf802c000-0xf802e000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfeed000 ioremap
0xf8030000-0xf8032000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfeec000 ioremap
0xf8038000-0xf803d000   20480 ich_force_enable_hpet+0x69/0x15a phys=fed1c000 ioremap
0xf803e000-0xf8040000    8192 hpet_enable+0x2a/0x21b phys=fed00000 ioremap
0xf8040000-0xf8046000   24576 alloc_iommu+0x18d/0x1d4 phys=feb00000 ioremap
0xf8048000-0xf804a000    8192 pcim_iomap+0x2f/0x3a phys=e1b21000 ioremap
0xf804c000-0xf804e000    8192 e1000_probe+0x229/0xa73 phys=e1b20000 ioremap
0xf804f000-0xf8051000    8192 reiserfs_init_bitmap_cache+0x32/0x65 pages=1 vmalloc
0xf8052000-0xf8064000   73728 journal_init+0x30/0x82a pages=17 vmalloc
0xf8065000-0xf8067000    8192 reiserfs_allocate_list_bitmaps+0x27/0x7e pages=1 vmalloc
0xf8068000-0xf806a000    8192 reiserfs_allocate_list_bitmaps+0x27/0x7e pages=1 vmalloc
0xf806b000-0xf806d000    8192 reiserfs_allocate_list_bitmaps+0x27/0x7e pages=1 vmalloc
0xf806e000-0xf8070000    8192 reiserfs_allocate_list_bitmaps+0x27/0x7e pages=1 vmalloc
0xf8071000-0xf8073000    8192 reiserfs_allocate_list_bitmaps+0x27/0x7e pages=1 vmalloc
0xf8080000-0xf80a1000  135168 e1000_probe+0x1ca/0xa73 phys=e1b00000 ioremap
0xf80a2000-0xf80a6000   16384 e1000e_setup_rx_resources+0x20/0xf7 pages=3 vmalloc
0xf80a7000-0xf80ab000   16384 e1000e_setup_tx_resources+0x17/0x96 pages=3 vmalloc
0xf80ac000-0xf80b0000   16384 e1000e_setup_rx_resources+0x20/0xf7 pages=3 vmalloc
0xf80b1000-0xf80b5000   16384 e1000e_setup_tx_resources+0x17/0x96 pages=3 vmalloc
0xf80c0000-0xf80e1000  135168 e1000_probe+0x1ca/0xa73 phys=e1a60000 ioremap
0xf8100000-0xf8121000  135168 e1000_probe+0x1ca/0xa73 phys=e1a20000 ioremap
0xf8122000-0xf81b3000  593920 journal_init+0x65b/0x82a pages=144 vmalloc
0xf81b4000-0xf822f000  503808 sys_swapon+0x392/0x8f3 pages=122 vmalloc

0xf846a000-0xf856c000 1056768 tnode_new+0x35/0x65 pages=257 vmalloc

Eric Dumazet pisze:

Quoted 152 lines
Quoted 15 lines · Paweł Staszewski

Hello ALL

Some time ago i report this:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

and now with 2.6.29 / 2.6.29.1 / 2.6.29.3 and 2.6.30 it back dmesg output:
oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

Quoted 9 lines

Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

Curious, you seem to hit an old alloc_pages limit()... (MAX_ORDER allocation)

Your root node has 2^18 = 262144 pointers of 8 bytes -> 2097152 bytes (+ header -> 4194304 bytes)

But since following commit, we should use vmalloc() so this PAGE_SIZE<<10) limit should not anymore be applied.

Could you do a "cat /proc/vmallocinfo" just to check your big tnodes are vmalloced() ?

commit 15be75cdb5db442d0e33d37b20832b88f3ccd383

Author: Stephen Hemminger <[email protected]>
Date:   Thu Apr 10 02:56:38 2008 -0700

IPV4: fib_trie use vmalloc for large tnodes

Use vmalloc rather than alloc_pages to avoid wasting memory. The problem is that tnode structure has a power of 2 sized array, plus a header. So the current code wastes almost half the memory allocated because it always needs the next bigger size to hold that small header.

This is similar to an earlier patch by Eric, but instead of a list and lock, I used a workqueue to handle the fact that vfree can't be done in interrupt context.

Quoted 90 lines

cat /proc/net/fib_triestat
Basic info: size of leaf: 40 bytes, size of tnode: 56 bytes. Main:

       Aver depth:     2.28
       Max depth:      6
       Leaves:         276539
       Prefixes:       289922
       Internal nodes: 66762
         1: 35046  2: 13824  3: 9508  4: 4897  5: 2331  6: 1149  7: 5 
9: 1  18: 1
       Pointers: 691228

Null ptrs: 347928
Total size: 35709 kB

Counters:
---------
gets = 26276593
backtracks = 547306
semantic match passed = 26188746
semantic match miss = 1117
null node hit= 27285055
skipped node resize = 0

Local:

       Aver depth:     3.33
       Max depth:      4
       Leaves:         9
       Prefixes:       10

Internal nodes: 8
1: 8
Pointers: 16
Null ptrs: 0
Total size: 2 kB

Counters:
---------
gets = 26642350
backtracks = 1282818
semantic match passed = 18166
semantic match miss = 0
null node hit= 0
skipped node resize = 0

This machine is running bgpd with two bgp peers / full route table

cat /proc/meminfo

MemTotal:       12279032 kB
MemFree:        11521920 kB
Buffers:           80288 kB
Cached:            34416 kB
SwapCached:            0 kB
Active:           286816 kB
Inactive:          82024 kB
Active(anon):     254296 kB
Inactive(anon):        0 kB
Active(file):      32520 kB
Inactive(file):    82024 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:        987988 kB
SwapFree:         987988 kB
Dirty:              1140 kB
Writeback:             0 kB
AnonPages:        254164 kB
Mapped:             5440 kB
Slab:             365084 kB
SReclaimable:      28784 kB
SUnreclaim:       336300 kB
PageTables:         2104 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     7127504 kB
Committed_AS:     267704 kB
VmallocTotal:   34359738367 kB
VmallocUsed:       11824 kB
VmallocChunk:   34359707815 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       2048 kB
DirectMap4k:        3392 kB
DirectMap2M:    12578816 kB

Interfaces mtu is1500

#3 lore raw
Quoted 56 lines · Paweł Staszewski

cat /proc/vmallocinfo
0xf7ffe000-0xf8000000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfe6a000 ioremap
0xf8000000-0xf8007000 28672 acpi_tb_verify_table+0x1d/0x46
phys=dfef5000 ioremap
0xf8008000-0xf800a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef2000 ioremap
0xf800c000-0xf800e000 8192

acpi_ex_system_memory_space_handler+0xd6/0x208 phys=fed1f000 ioremap
0xf8010000-0xf8012000    8192 acpi_tb_verify_table+0x1d/0x46

phys=dfefb000 ioremap
0xf8014000-0xf8016000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef4000 ioremap
0xf8018000-0xf801a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef3000 ioremap
0xf801c000-0xf801e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef1000 ioremap
0xf8020000-0xf8022000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef0000 ioremap
0xf8024000-0xf8026000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeef000 ioremap
0xf8028000-0xf802a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeee000 ioremap
0xf802c000-0xf802e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeed000 ioremap
0xf8030000-0xf8032000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeec000 ioremap
0xf8038000-0xf803d000 20480 ich_force_enable_hpet+0x69/0x15a
phys=fed1c000 ioremap
0xf803e000-0xf8040000 8192 hpet_enable+0x2a/0x21b phys=fed00000 ioremap
0xf8040000-0xf8046000 24576 alloc_iommu+0x18d/0x1d4 phys=feb00000 ioremap

0xf8048000-0xf804a000    8192 pcim_iomap+0x2f/0x3a phys=e1b21000 ioremap
0xf804c000-0xf804e000    8192 e1000_probe+0x229/0xa73 phys=e1b20000 ioremap
0xf804f000-0xf8051000    8192 reiserfs_init_bitmap_cache+0x32/0x65

pages=1 vmalloc
0xf8052000-0xf8064000 73728 journal_init+0x30/0x82a pages=17 vmalloc
0xf8065000-0xf8067000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8068000-0xf806a000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806b000-0xf806d000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806e000-0xf8070000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8071000-0xf8073000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8080000-0xf80a1000 135168 e1000_probe+0x1ca/0xa73 phys=e1b00000 ioremap
0xf80a2000-0xf80a6000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80a7000-0xf80ab000 16384 e1000e_setup_tx_resources+0x17/0x96
pages=3 vmalloc
0xf80ac000-0xf80b0000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80b1000-0xf80b5000 16384 e1000e_setup_tx_resources+0x17/0x96
pages=3 vmalloc

Quoted 5 lines
0xf80c0000-0xf80e1000  135168 e1000_probe+0x1ca/0xa73 phys=e1a60000 ioremap
0xf8100000-0xf8121000  135168 e1000_probe+0x1ca/0xa73 phys=e1a20000 ioremap
0xf8122000-0xf81b3000  593920 journal_init+0x65b/0x82a pages=144 vmalloc
0xf81b4000-0xf822f000  503808 sys_swapon+0x392/0x8f3 pages=122 vmalloc

0xf846a000-0xf856c000 1056768 tnode_new+0x35/0x65 pages=257 vmalloc

This is from a 32 bit kernel.

This doesnt match your previous /proc/meminfo (from a 64bit kernel on a 12 GB machine)

Of course, I would like /proc/vmallocinfo on your loaded router, not from a dev machine :)

Quoted 166 lines

Eric Dumazet pisze:

Quoted 156 lines
Quoted 15 lines · Paweł Staszewski

Hello ALL

Some time ago i report this:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

and now with 2.6.29 / 2.6.29.1 / 2.6.29.3 and 2.6.30 it back dmesg output:
oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

Quoted 9 lines

Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

Curious, you seem to hit an old alloc_pages limit()... (MAX_ORDER allocation)

Your root node has 2^18 = 262144 pointers of 8 bytes -> 2097152 bytes (+ header -> 4194304 bytes)

But since following commit, we should use vmalloc() so this PAGE_SIZE<<10) limit
should not anymore be applied.

Could you do a "cat /proc/vmallocinfo" just to check your big tnodes are vmalloced() ?

commit 15be75cdb5db442d0e33d37b20832b88f3ccd383

Author: Stephen Hemminger <[email protected]>
Date:   Thu Apr 10 02:56:38 2008 -0700

IPV4: fib_trie use vmalloc for large tnodes

Use vmalloc rather than alloc_pages to avoid wasting memory. The problem is that tnode structure has a power of 2 sized array, plus a header. So the current code wastes almost half the memory allocated because it always needs the next bigger size to hold that small header.

This is similar to an earlier patch by Eric, but instead of a list and lock, I used a workqueue to handle the fact that vfree can't be done in interrupt context.

Quoted 90 lines

cat /proc/net/fib_triestat
Basic info: size of leaf: 40 bytes, size of tnode: 56 bytes. Main:

       Aver depth:     2.28
       Max depth:      6
       Leaves:         276539
       Prefixes:       289922
       Internal nodes: 66762
         1: 35046  2: 13824  3: 9508  4: 4897  5: 2331  6: 1149  7: 5
9: 1  18: 1
       Pointers: 691228

Null ptrs: 347928
Total size: 35709 kB

Counters:
---------
gets = 26276593
backtracks = 547306
semantic match passed = 26188746
semantic match miss = 1117
null node hit= 27285055
skipped node resize = 0

Local:

       Aver depth:     3.33
       Max depth:      4
       Leaves:         9
       Prefixes:       10

Internal nodes: 8
1: 8
Pointers: 16
Null ptrs: 0
Total size: 2 kB

Counters:
---------
gets = 26642350
backtracks = 1282818
semantic match passed = 18166
semantic match miss = 0
null node hit= 0
skipped node resize = 0

This machine is running bgpd with two bgp peers / full route table

cat /proc/meminfo

MemTotal:       12279032 kB
MemFree:        11521920 kB
Buffers:           80288 kB
Cached:            34416 kB
SwapCached:            0 kB
Active:           286816 kB
Inactive:          82024 kB
Active(anon):     254296 kB
Inactive(anon):        0 kB
Active(file):      32520 kB
Inactive(file):    82024 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:        987988 kB
SwapFree:         987988 kB
Dirty:              1140 kB
Writeback:             0 kB
AnonPages:        254164 kB
Mapped:             5440 kB
Slab:             365084 kB
SReclaimable:      28784 kB
SUnreclaim:       336300 kB
PageTables:         2104 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     7127504 kB
Committed_AS:     267704 kB
VmallocTotal:   34359738367 kB
VmallocUsed:       11824 kB
VmallocChunk:   34359707815 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       2048 kB
DirectMap4k:        3392 kB
DirectMap2M:    12578816 kB

Interfaces mtu is1500

Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#4 lore raw

Eric Dumazet pisze:

Quoted 64 lines
Quoted 55 lines · Paweł Staszewski

cat /proc/vmallocinfo
0xf7ffe000-0xf8000000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfe6a000 ioremap
0xf8000000-0xf8007000 28672 acpi_tb_verify_table+0x1d/0x46
phys=dfef5000 ioremap
0xf8008000-0xf800a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef2000 ioremap
0xf800c000-0xf800e000 8192

acpi_ex_system_memory_space_handler+0xd6/0x208 phys=fed1f000 ioremap
0xf8010000-0xf8012000    8192 acpi_tb_verify_table+0x1d/0x46

phys=dfefb000 ioremap
0xf8014000-0xf8016000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef4000 ioremap
0xf8018000-0xf801a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef3000 ioremap
0xf801c000-0xf801e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef1000 ioremap
0xf8020000-0xf8022000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef0000 ioremap
0xf8024000-0xf8026000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeef000 ioremap
0xf8028000-0xf802a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeee000 ioremap
0xf802c000-0xf802e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeed000 ioremap
0xf8030000-0xf8032000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeec000 ioremap
0xf8038000-0xf803d000 20480 ich_force_enable_hpet+0x69/0x15a
phys=fed1c000 ioremap
0xf803e000-0xf8040000 8192 hpet_enable+0x2a/0x21b phys=fed00000 ioremap
0xf8040000-0xf8046000 24576 alloc_iommu+0x18d/0x1d4 phys=feb00000 ioremap

0xf8048000-0xf804a000    8192 pcim_iomap+0x2f/0x3a phys=e1b21000 ioremap
0xf804c000-0xf804e000    8192 e1000_probe+0x229/0xa73 phys=e1b20000 ioremap
0xf804f000-0xf8051000    8192 reiserfs_init_bitmap_cache+0x32/0x65

pages=1 vmalloc
0xf8052000-0xf8064000 73728 journal_init+0x30/0x82a pages=17 vmalloc
0xf8065000-0xf8067000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8068000-0xf806a000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806b000-0xf806d000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806e000-0xf8070000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8071000-0xf8073000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8080000-0xf80a1000 135168 e1000_probe+0x1ca/0xa73 phys=e1b00000 ioremap
0xf80a2000-0xf80a6000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80a7000-0xf80ab000 16384 e1000e_setup_tx_resources+0x17/0x96
pages=3 vmalloc
0xf80ac000-0xf80b0000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80b1000-0xf80b5000 16384 e1000e_setup_tx_resources+0x17/0x96
pages=3 vmalloc

Quoted 6 lines
0xf80c0000-0xf80e1000  135168 e1000_probe+0x1ca/0xa73 phys=e1a60000 ioremap
0xf8100000-0xf8121000  135168 e1000_probe+0x1ca/0xa73 phys=e1a20000 ioremap
0xf8122000-0xf81b3000  593920 journal_init+0x65b/0x82a pages=144 vmalloc
0xf81b4000-0xf822f000  503808 sys_swapon+0x392/0x8f3 pages=122 vmalloc

0xf846a000-0xf856c000 1056768 tnode_new+0x35/0x65 pages=257 vmalloc

Quoted 8 lines

This is from a 32 bit kernel.

This doesnt match your previous /proc/meminfo (from a 64bit kernel on a 12 GB machine)

Of course, I would like /proc/vmallocinfo on your loaded router, not from a dev machine :)

Yes sorry for no info about it.
I test the same kernel configurations on one 32bit machine and second 64bit

here is meminfo from this 32bit machine working on kernel 2.6.30 cat /proc/meminfo

MemTotal:        3625444 kB
MemFree:         3043648 kB
Buffers:          133968 kB
Cached:            36316 kB
SwapCached:            0 kB
Active:           256868 kB
Inactive:          76252 kB
Active(anon):     163064 kB
Inactive(anon):        0 kB
Active(file):      93804 kB
Inactive(file):    76252 kB
Unevictable:           0 kB
Mlocked:               0 kB
HighTotal:       2758160 kB
HighFree:        2556136 kB
LowTotal:         867284 kB
LowFree:          487512 kB
SwapTotal:        995896 kB
SwapFree:         995896 kB
Dirty:              3624 kB
Writeback:             0 kB
AnonPages:        162912 kB
Mapped:             3612 kB
Slab:             235888 kB
SReclaimable:      46408 kB
SUnreclaim:       189480 kB
PageTables:          384 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     2808616 kB
Committed_AS:     170648 kB
VmallocTotal:     122880 kB
VmallocUsed:        2876 kB
VmallocChunk:     109824 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       4096 kB
DirectMap4k:        8184 kB
DirectMap4M:      901120 kB

and vmallocinfo

cat /proc/vmallocinfo
0xf7ffe000-0xf8000000 8192 acpi_tb_verify_table+0x1d/0x46 phys=dfe6a000 ioremap
0xf8000000-0xf8007000 28672 acpi_tb_verify_table+0x1d/0x46 phys=dfef5000 ioremap

0xf8008000-0xf800a000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfef2000 ioremap
0xf800c000-0xf800e000    8192 acpi_ex_system_memory_space_handler+0xd6/0x208 phys=fed1f000 ioremap
0xf8010000-0xf8012000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfefb000 ioremap
0xf8014000-0xf8016000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfef4000 ioremap
0xf8018000-0xf801a000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfef3000 ioremap
0xf801c000-0xf801e000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfef1000 ioremap
0xf8020000-0xf8022000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfef0000 ioremap
0xf8024000-0xf8026000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfeef000 ioremap
0xf8028000-0xf802a000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfeee000 ioremap
0xf802c000-0xf802e000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfeed000 ioremap
0xf8030000-0xf8032000    8192 acpi_tb_verify_table+0x1d/0x46 phys=dfeec000 ioremap
0xf8038000-0xf803d000   20480 ich_force_enable_hpet+0x69/0x15a phys=fed1c000 ioremap
0xf803e000-0xf8040000    8192 hpet_enable+0x2a/0x21b phys=fed00000 ioremap
0xf8040000-0xf8046000   24576 alloc_iommu+0x18d/0x1d4 phys=feb00000 ioremap
0xf8048000-0xf804a000    8192 pcim_iomap+0x2f/0x3a phys=e1b21000 ioremap
0xf804c000-0xf804e000    8192 e1000_probe+0x229/0xa73 phys=e1b20000 ioremap
0xf804f000-0xf8051000    8192 reiserfs_init_bitmap_cache+0x32/0x65 pages=1 vmalloc
0xf8052000-0xf8064000   73728 journal_init+0x30/0x82a pages=17 vmalloc
0xf8065000-0xf8067000    8192 reiserfs_allocate_list_bitmaps+0x27/0x7e pages=1 vmalloc
0xf8068000-0xf806a000    8192 reiserfs_allocate_list_bitmaps+0x27/0x7e pages=1 vmalloc
0xf806b000-0xf806d000    8192 reiserfs_allocate_list_bitmaps+0x27/0x7e pages=1 vmalloc
0xf806e000-0xf8070000    8192 reiserfs_allocate_list_bitmaps+0x27/0x7e pages=1 vmalloc
0xf8071000-0xf8073000    8192 reiserfs_allocate_list_bitmaps+0x27/0x7e pages=1 vmalloc
0xf8080000-0xf80a1000  135168 e1000_probe+0x1ca/0xa73 phys=e1b00000 ioremap
0xf80a2000-0xf80a6000   16384 e1000e_setup_rx_resources+0x20/0xf7 pages=3 vmalloc
0xf80a7000-0xf80ab000   16384 e1000e_setup_tx_resources+0x17/0x96 pages=3 vmalloc
0xf80ac000-0xf80b0000   16384 e1000e_setup_rx_resources+0x20/0xf7 pages=3 vmalloc
0xf80b1000-0xf80b5000   16384 e1000e_setup_tx_resources+0x17/0x96 pages=3 vmalloc
0xf80c0000-0xf80e1000  135168 e1000_probe+0x1ca/0xa73 phys=e1a60000 ioremap
0xf8100000-0xf8121000  135168 e1000_probe+0x1ca/0xa73 phys=e1a20000 ioremap
0xf8122000-0xf81b3000  593920 journal_init+0x65b/0x82a pages=144 vmalloc
0xf81b4000-0xf822f000  503808 sys_swapon+0x392/0x8f3 pages=122 vmalloc

0xf8bbc000-0xf8cbe000 1056768 tnode_new+0x35/0x65 pages=257 vmalloc

And next machine with kernel 2.6.29.3
dmesg:
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
cat /proc/meminfo

MemTotal:        2072652 kB
MemFree:          496960 kB
Buffers:          267620 kB
Cached:           895212 kB
SwapCached:            0 kB
Active:           675744 kB
Inactive:         703312 kB
Active(anon):     215848 kB
Inactive(anon):        0 kB
Active(file):     459896 kB
Inactive(file):   703312 kB
Unevictable:           0 kB
Mlocked:               0 kB
HighTotal:       1186696 kB
HighFree:         151156 kB
LowTotal:         885956 kB
LowFree:          345804 kB
SwapTotal:       1975984 kB
SwapFree:        1975984 kB
Dirty:                20 kB
Writeback:             0 kB
AnonPages:        215724 kB
Mapped:             6120 kB
Slab:             186652 kB
SReclaimable:     125832 kB
SUnreclaim:        60820 kB
PageTables:          416 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     3012308 kB
Committed_AS:     223692 kB
VmallocTotal:     122880 kB
VmallocUsed:        3192 kB
VmallocChunk:     108436 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       4096 kB
DirectMap4k:        8184 kB
DirectMap4M:      901120 kB
cat /proc/vmallocinfo
0xf7ffe000-0xf8000000    8192 acpi_tb_verify_table+0x1d/0x46 phys=7fee0000 ioremap
0xf8000000-0xf8005000   20480 acpi_tb_verify_table+0x1d/0x46 phys=7fee3000 ioremap
0xf8006000-0xf8008000    8192 acpi_tb_verify_table+0x1d/0x46 phys=7fee3000 ioremap
0xf800a000-0xf800c000    8192 acpi_tb_verify_table+0x1d/0x46 phys=7fee6000 ioremap
0xf800d000-0xf800f000    8192 reiserfs_init_bitmap_cache+0x3b/0x80 pages=1 vmalloc
0xf8010000-0xf8022000   73728 journal_init+0x30/0x8f0 pages=17 vmalloc
0xf8023000-0xf8025000    8192 reiserfs_allocate_list_bitmaps+0x2d/0x90 pages=1 vmalloc
0xf8026000-0xf8028000    8192 reiserfs_allocate_list_bitmaps+0x2d/0x90 pages=1 vmalloc
0xf8029000-0xf802b000    8192 reiserfs_allocate_list_bitmaps+0x2d/0x90 pages=1 vmalloc
0xf802c000-0xf802e000    8192 reiserfs_allocate_list_bitmaps+0x2d/0x90 pages=1 vmalloc
0xf802f000-0xf8031000    8192 reiserfs_allocate_list_bitmaps+0x2d/0x90 pages=1 vmalloc
0xf803e000-0xf8040000    8192 e1000_setup_all_tx_resources+0x57/0x660 pages=1 vmalloc
0xf8040000-0xf8061000  135168 e1000_probe+0x207/0xeb0 phys=f5000000 ioremap
0xf8062000-0xf8064000    8192 e1000_setup_all_rx_resources+0x57/0x6d0 pages=1 vmalloc
0xf8065000-0xf8067000    8192 e1000_setup_all_tx_resources+0x57/0x660 pages=1 vmalloc
0xf8068000-0xf806a000    8192 e1000_setup_all_rx_resources+0x57/0x6d0 pages=1 vmalloc
0xf806b000-0xf806d000    8192 e1000_setup_all_tx_resources+0x57/0x660 pages=1 vmalloc
0xf806e000-0xf8070000    8192 e1000_setup_all_rx_resources+0x57/0x6d0 pages=1 vmalloc
0xf8080000-0xf80a1000  135168 e1000_probe+0x207/0xeb0 phys=f1040000 ioremap
0xf80c0000-0xf80e1000  135168 e1000_probe+0x207/0xeb0 phys=f4000000 ioremap
0xf80e2000-0xf8173000  593920 journal_init+0x56e/0x8f0 pages=144 vmalloc
0xf8174000-0xf8267000  995328 sys_swapon+0x548/0xa30 pages=242 vmalloc

0xf8d17000-0xf8e19000 1056768 tnode_new+0x7f/0x90 pages=257 vmalloc

because i have this info on 5 machines that working in ibgp mesh
And only one 64bit dev machine that is one of failover member - but i kill this machine after upgrade to kernel 2.6.31-rc1

Quoted 175 lines
Quoted 154 lines

Eric Dumazet pisze:

Quoted 152 lines
Quoted 15 lines · Paweł Staszewski

Hello ALL

Some time ago i report this:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

and now with 2.6.29 / 2.6.29.1 / 2.6.29.3 and 2.6.30 it back dmesg output:
oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

Quoted 10 lines

Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

Curious, you seem to hit an old alloc_pages limit()... (MAX_ORDER allocation)

Your root node has 2^18 = 262144 pointers of 8 bytes -> 2097152 bytes (+ header -> 4194304 bytes)

But since following commit, we should use vmalloc() so this PAGE_SIZE<<10) limit
should not anymore be applied.

Could you do a "cat /proc/vmallocinfo" just to check your big tnodes are vmalloced() ?

commit 15be75cdb5db442d0e33d37b20832b88f3ccd383

Author: Stephen Hemminger <[email protected]>
Date:   Thu Apr 10 02:56:38 2008 -0700

IPV4: fib_trie use vmalloc for large tnodes

Use vmalloc rather than alloc_pages to avoid wasting memory. The problem is that tnode structure has a power of 2 sized array, plus a header. So the current code wastes almost half the memory allocated because it always needs the next bigger size to hold that small header.

This is similar to an earlier patch by Eric, but instead of a list and lock, I used a workqueue to handle the fact that vfree can't be done in interrupt context.

Quoted 88 lines

cat /proc/net/fib_triestat
Basic info: size of leaf: 40 bytes, size of tnode: 56 bytes. Main:

       Aver depth:     2.28
       Max depth:      6
       Leaves:         276539
       Prefixes:       289922
       Internal nodes: 66762
         1: 35046  2: 13824  3: 9508  4: 4897  5: 2331  6: 1149  7: 5
9: 1  18: 1
       Pointers: 691228

Null ptrs: 347928
Total size: 35709 kB

Counters:
---------
gets = 26276593
backtracks = 547306
semantic match passed = 26188746
semantic match miss = 1117
null node hit= 27285055
skipped node resize = 0

Local:

       Aver depth:     3.33
       Max depth:      4
       Leaves:         9
       Prefixes:       10

Internal nodes: 8
1: 8
Pointers: 16
Null ptrs: 0
Total size: 2 kB

Counters:
---------
gets = 26642350
backtracks = 1282818
semantic match passed = 18166
semantic match miss = 0
null node hit= 0
skipped node resize = 0

This machine is running bgpd with two bgp peers / full route table

cat /proc/meminfo

MemTotal:       12279032 kB
MemFree:        11521920 kB
Buffers:           80288 kB
Cached:            34416 kB
SwapCached:            0 kB
Active:           286816 kB
Inactive:          82024 kB
Active(anon):     254296 kB
Inactive(anon):        0 kB
Active(file):      32520 kB
Inactive(file):    82024 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:        987988 kB
SwapFree:         987988 kB
Dirty:              1140 kB
Writeback:             0 kB
AnonPages:        254164 kB
Mapped:             5440 kB
Slab:             365084 kB
SReclaimable:      28784 kB
SUnreclaim:       336300 kB
PageTables:         2104 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     7127504 kB
Committed_AS:     267704 kB
VmallocTotal:   34359738367 kB
VmallocUsed:       11824 kB
VmallocChunk:   34359707815 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       2048 kB
DirectMap4k:        3392 kB
DirectMap2M:    12578816 kB
Quoted 13 lines
Quoted 6 lines
Quoted 3 lines

Interfaces mtu is1500

Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#5 lore raw
Quoted 286 lines · Paweł Staszewski

Eric Dumazet pisze:

Quoted 70 lines
Quoted 58 lines · Paweł Staszewski

cat /proc/vmallocinfo
0xf7ffe000-0xf8000000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfe6a000 ioremap
0xf8000000-0xf8007000 28672 acpi_tb_verify_table+0x1d/0x46
phys=dfef5000 ioremap
0xf8008000-0xf800a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef2000 ioremap
0xf800c000-0xf800e000 8192

acpi_ex_system_memory_space_handler+0xd6/0x208 phys=fed1f000 ioremap
0xf8010000-0xf8012000    8192 acpi_tb_verify_table+0x1d/0x46

phys=dfefb000 ioremap
0xf8014000-0xf8016000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef4000 ioremap
0xf8018000-0xf801a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef3000 ioremap
0xf801c000-0xf801e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef1000 ioremap
0xf8020000-0xf8022000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef0000 ioremap
0xf8024000-0xf8026000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeef000 ioremap
0xf8028000-0xf802a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeee000 ioremap
0xf802c000-0xf802e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeed000 ioremap
0xf8030000-0xf8032000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeec000 ioremap
0xf8038000-0xf803d000 20480 ich_force_enable_hpet+0x69/0x15a
phys=fed1c000 ioremap
0xf803e000-0xf8040000 8192 hpet_enable+0x2a/0x21b phys=fed00000
ioremap
0xf8040000-0xf8046000 24576 alloc_iommu+0x18d/0x1d4 phys=feb00000
ioremap
0xf8048000-0xf804a000 8192 pcim_iomap+0x2f/0x3a phys=e1b21000 ioremap
0xf804c000-0xf804e000 8192 e1000_probe+0x229/0xa73 phys=e1b20000
ioremap
0xf804f000-0xf8051000 8192 reiserfs_init_bitmap_cache+0x32/0x65
pages=1 vmalloc
0xf8052000-0xf8064000 73728 journal_init+0x30/0x82a pages=17 vmalloc
0xf8065000-0xf8067000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8068000-0xf806a000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806b000-0xf806d000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806e000-0xf8070000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8071000-0xf8073000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8080000-0xf80a1000 135168 e1000_probe+0x1ca/0xa73 phys=e1b00000 ioremap
0xf80a2000-0xf80a6000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80a7000-0xf80ab000 16384 e1000e_setup_tx_resources+0x17/0x96
pages=3 vmalloc
0xf80ac000-0xf80b0000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80b1000-0xf80b5000 16384 e1000e_setup_tx_resources+0x17/0x96

Quoted 9 lines

pages=3 vmalloc
0xf80c0000-0xf80e1000 135168 e1000_probe+0x1ca/0xa73 phys=e1a60000 ioremap
0xf8100000-0xf8121000 135168 e1000_probe+0x1ca/0xa73 phys=e1a20000 ioremap
0xf8122000-0xf81b3000 593920 journal_init+0x65b/0x82a pages=144 vmalloc 0xf81b4000-0xf822f000 503808 sys_swapon+0x392/0x8f3 pages=122 vmalloc 0xf846a000-0xf856c000 1056768 tnode_new+0x35/0x65 pages=257 vmalloc

Quoted 9 lines

This is from a 32 bit kernel.

This doesnt match your previous /proc/meminfo (from a 64bit kernel on a 12 GB machine)

Of course, I would like /proc/vmallocinfo on your loaded router, not from a dev machine :)

Yes sorry for no info about it.
I test the same kernel configurations on one 32bit machine and second 64bit

here is meminfo from this 32bit machine working on kernel 2.6.30 cat /proc/meminfo

MemTotal:        3625444 kB
MemFree:         3043648 kB
Buffers:          133968 kB
Cached:            36316 kB
SwapCached:            0 kB
Active:           256868 kB
Inactive:          76252 kB
Active(anon):     163064 kB
Inactive(anon):        0 kB
Active(file):      93804 kB
Inactive(file):    76252 kB
Unevictable:           0 kB
Mlocked:               0 kB
HighTotal:       2758160 kB
HighFree:        2556136 kB
LowTotal:         867284 kB
LowFree:          487512 kB
SwapTotal:        995896 kB
SwapFree:         995896 kB
Dirty:              3624 kB
Writeback:             0 kB
AnonPages:        162912 kB
Mapped:             3612 kB
Slab:             235888 kB
SReclaimable:      46408 kB
SUnreclaim:       189480 kB
PageTables:          384 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     2808616 kB
Committed_AS:     170648 kB
VmallocTotal:     122880 kB
VmallocUsed:        2876 kB
VmallocChunk:     109824 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       4096 kB
DirectMap4k:        8184 kB
DirectMap4M:      901120 kB

and vmallocinfo

cat /proc/vmallocinfo
0xf7ffe000-0xf8000000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfe6a000 ioremap
0xf8000000-0xf8007000 28672 acpi_tb_verify_table+0x1d/0x46
phys=dfef5000 ioremap
0xf8008000-0xf800a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef2000 ioremap
0xf800c000-0xf800e000 8192

acpi_ex_system_memory_space_handler+0xd6/0x208 phys=fed1f000 ioremap
0xf8010000-0xf8012000    8192 acpi_tb_verify_table+0x1d/0x46

phys=dfefb000 ioremap
0xf8014000-0xf8016000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef4000 ioremap
0xf8018000-0xf801a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef3000 ioremap
0xf801c000-0xf801e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef1000 ioremap
0xf8020000-0xf8022000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef0000 ioremap
0xf8024000-0xf8026000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeef000 ioremap
0xf8028000-0xf802a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeee000 ioremap
0xf802c000-0xf802e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeed000 ioremap
0xf8030000-0xf8032000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeec000 ioremap
0xf8038000-0xf803d000 20480 ich_force_enable_hpet+0x69/0x15a
phys=fed1c000 ioremap
0xf803e000-0xf8040000 8192 hpet_enable+0x2a/0x21b phys=fed00000 ioremap
0xf8040000-0xf8046000 24576 alloc_iommu+0x18d/0x1d4 phys=feb00000 ioremap

0xf8048000-0xf804a000    8192 pcim_iomap+0x2f/0x3a phys=e1b21000 ioremap
0xf804c000-0xf804e000    8192 e1000_probe+0x229/0xa73 phys=e1b20000 ioremap
0xf804f000-0xf8051000    8192 reiserfs_init_bitmap_cache+0x32/0x65

pages=1 vmalloc
0xf8052000-0xf8064000 73728 journal_init+0x30/0x82a pages=17 vmalloc
0xf8065000-0xf8067000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8068000-0xf806a000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806b000-0xf806d000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806e000-0xf8070000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8071000-0xf8073000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8080000-0xf80a1000 135168 e1000_probe+0x1ca/0xa73 phys=e1b00000 ioremap
0xf80a2000-0xf80a6000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80a7000-0xf80ab000 16384 e1000e_setup_tx_resources+0x17/0x96
pages=3 vmalloc
0xf80ac000-0xf80b0000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80b1000-0xf80b5000 16384 e1000e_setup_tx_resources+0x17/0x96
pages=3 vmalloc

0xf80c0000-0xf80e1000  135168 e1000_probe+0x1ca/0xa73 phys=e1a60000 ioremap
0xf8100000-0xf8121000  135168 e1000_probe+0x1ca/0xa73 phys=e1a20000 ioremap
0xf8122000-0xf81b3000  593920 journal_init+0x65b/0x82a pages=144 vmalloc
0xf81b4000-0xf822f000  503808 sys_swapon+0x392/0x8f3 pages=122 vmalloc

0xf8bbc000-0xf8cbe000 1056768 tnode_new+0x35/0x65 pages=257 vmalloc

And next machine with kernel 2.6.29.3
dmesg:
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
cat /proc/meminfo

MemTotal:        2072652 kB
MemFree:          496960 kB
Buffers:          267620 kB
Cached:           895212 kB
SwapCached:            0 kB
Active:           675744 kB
Inactive:         703312 kB
Active(anon):     215848 kB
Inactive(anon):        0 kB
Active(file):     459896 kB
Inactive(file):   703312 kB
Unevictable:           0 kB
Mlocked:               0 kB
HighTotal:       1186696 kB
HighFree:         151156 kB
LowTotal:         885956 kB
LowFree:          345804 kB
SwapTotal:       1975984 kB
SwapFree:        1975984 kB
Dirty:                20 kB
Writeback:             0 kB
AnonPages:        215724 kB
Mapped:             6120 kB
Slab:             186652 kB
SReclaimable:     125832 kB
SUnreclaim:        60820 kB
PageTables:          416 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     3012308 kB
Committed_AS:     223692 kB
VmallocTotal:     122880 kB
VmallocUsed:        3192 kB
VmallocChunk:     108436 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       4096 kB
DirectMap4k:        8184 kB
DirectMap4M:      901120 kB

cat /proc/vmallocinfo
0xf7ffe000-0xf8000000 8192 acpi_tb_verify_table+0x1d/0x46
phys=7fee0000 ioremap
0xf8000000-0xf8005000 20480 acpi_tb_verify_table+0x1d/0x46
phys=7fee3000 ioremap
0xf8006000-0xf8008000 8192 acpi_tb_verify_table+0x1d/0x46
phys=7fee3000 ioremap
0xf800a000-0xf800c000 8192 acpi_tb_verify_table+0x1d/0x46
phys=7fee6000 ioremap
0xf800d000-0xf800f000 8192 reiserfs_init_bitmap_cache+0x3b/0x80
pages=1 vmalloc
0xf8010000-0xf8022000 73728 journal_init+0x30/0x8f0 pages=17 vmalloc
0xf8023000-0xf8025000 8192 reiserfs_allocate_list_bitmaps+0x2d/0x90
pages=1 vmalloc
0xf8026000-0xf8028000 8192 reiserfs_allocate_list_bitmaps+0x2d/0x90
pages=1 vmalloc
0xf8029000-0xf802b000 8192 reiserfs_allocate_list_bitmaps+0x2d/0x90
pages=1 vmalloc
0xf802c000-0xf802e000 8192 reiserfs_allocate_list_bitmaps+0x2d/0x90
pages=1 vmalloc
0xf802f000-0xf8031000 8192 reiserfs_allocate_list_bitmaps+0x2d/0x90
pages=1 vmalloc
0xf803e000-0xf8040000 8192 e1000_setup_all_tx_resources+0x57/0x660
pages=1 vmalloc
0xf8040000-0xf8061000 135168 e1000_probe+0x207/0xeb0 phys=f5000000 ioremap
0xf8062000-0xf8064000 8192 e1000_setup_all_rx_resources+0x57/0x6d0
pages=1 vmalloc
0xf8065000-0xf8067000 8192 e1000_setup_all_tx_resources+0x57/0x660
pages=1 vmalloc
0xf8068000-0xf806a000 8192 e1000_setup_all_rx_resources+0x57/0x6d0
pages=1 vmalloc
0xf806b000-0xf806d000 8192 e1000_setup_all_tx_resources+0x57/0x660
pages=1 vmalloc
0xf806e000-0xf8070000 8192 e1000_setup_all_rx_resources+0x57/0x6d0
pages=1 vmalloc

0xf8080000-0xf80a1000  135168 e1000_probe+0x207/0xeb0 phys=f1040000 ioremap
0xf80c0000-0xf80e1000  135168 e1000_probe+0x207/0xeb0 phys=f4000000 ioremap
0xf80e2000-0xf8173000  593920 journal_init+0x56e/0x8f0 pages=144 vmalloc
0xf8174000-0xf8267000  995328 sys_swapon+0x548/0xa30 pages=242 vmalloc
Quoted 6 lines

0xf8d17000-0xf8e19000 1056768 tnode_new+0x7f/0x90 pages=257 vmalloc

because i have this info on 5 machines that working in ibgp mesh And only one 64bit dev machine that is one of failover member - but i kill this machine after upgrade to kernel 2.6.31-rc1

Yes, I was a fool to ask you to try 2.6.31-rc1, sorry.

Even 2.6.30 is too young for a production machine.

2.6.29.5 contains the fixes, Pawel, did you tried this version ?

#6 lore raw

Eric Dumazet pisze:

Quoted 293 lines
Quoted 291 lines · Paweł Staszewski

Eric Dumazet pisze:

Quoted 71 lines
Quoted 58 lines · Paweł Staszewski

cat /proc/vmallocinfo
0xf7ffe000-0xf8000000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfe6a000 ioremap
0xf8000000-0xf8007000 28672 acpi_tb_verify_table+0x1d/0x46
phys=dfef5000 ioremap
0xf8008000-0xf800a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef2000 ioremap
0xf800c000-0xf800e000 8192

acpi_ex_system_memory_space_handler+0xd6/0x208 phys=fed1f000 ioremap
0xf8010000-0xf8012000    8192 acpi_tb_verify_table+0x1d/0x46

phys=dfefb000 ioremap
0xf8014000-0xf8016000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef4000 ioremap
0xf8018000-0xf801a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef3000 ioremap
0xf801c000-0xf801e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef1000 ioremap
0xf8020000-0xf8022000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef0000 ioremap
0xf8024000-0xf8026000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeef000 ioremap
0xf8028000-0xf802a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeee000 ioremap
0xf802c000-0xf802e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeed000 ioremap
0xf8030000-0xf8032000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeec000 ioremap
0xf8038000-0xf803d000 20480 ich_force_enable_hpet+0x69/0x15a
phys=fed1c000 ioremap
0xf803e000-0xf8040000 8192 hpet_enable+0x2a/0x21b phys=fed00000
ioremap
0xf8040000-0xf8046000 24576 alloc_iommu+0x18d/0x1d4 phys=feb00000
ioremap
0xf8048000-0xf804a000 8192 pcim_iomap+0x2f/0x3a phys=e1b21000 ioremap
0xf804c000-0xf804e000 8192 e1000_probe+0x229/0xa73 phys=e1b20000
ioremap
0xf804f000-0xf8051000 8192 reiserfs_init_bitmap_cache+0x32/0x65
pages=1 vmalloc
0xf8052000-0xf8064000 73728 journal_init+0x30/0x82a pages=17 vmalloc
0xf8065000-0xf8067000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8068000-0xf806a000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806b000-0xf806d000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806e000-0xf8070000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8071000-0xf8073000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8080000-0xf80a1000 135168 e1000_probe+0x1ca/0xa73 phys=e1b00000 ioremap
0xf80a2000-0xf80a6000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80a7000-0xf80ab000 16384 e1000e_setup_tx_resources+0x17/0x96
pages=3 vmalloc
0xf80ac000-0xf80b0000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80b1000-0xf80b5000 16384 e1000e_setup_tx_resources+0x17/0x96

Quoted 10 lines

pages=3 vmalloc
0xf80c0000-0xf80e1000 135168 e1000_probe+0x1ca/0xa73 phys=e1a60000 ioremap
0xf8100000-0xf8121000 135168 e1000_probe+0x1ca/0xa73 phys=e1a20000 ioremap
0xf8122000-0xf81b3000 593920 journal_init+0x65b/0x82a pages=144 vmalloc 0xf81b4000-0xf822f000 503808 sys_swapon+0x392/0x8f3 pages=122 vmalloc 0xf846a000-0xf856c000 1056768 tnode_new+0x35/0x65 pages=257 vmalloc

Quoted 10 lines

This is from a 32 bit kernel.

This doesnt match your previous /proc/meminfo (from a 64bit kernel on a 12 GB machine)

Of course, I would like /proc/vmallocinfo on your loaded router, not from a dev machine :)

Yes sorry for no info about it.
I test the same kernel configurations on one 32bit machine and second 64bit

here is meminfo from this 32bit machine working on kernel 2.6.30 cat /proc/meminfo

MemTotal:        3625444 kB
MemFree:         3043648 kB
Buffers:          133968 kB
Cached:            36316 kB
SwapCached:            0 kB
Active:           256868 kB
Inactive:          76252 kB
Active(anon):     163064 kB
Inactive(anon):        0 kB
Active(file):      93804 kB
Inactive(file):    76252 kB
Unevictable:           0 kB
Mlocked:               0 kB
HighTotal:       2758160 kB
HighFree:        2556136 kB
LowTotal:         867284 kB
LowFree:          487512 kB
SwapTotal:        995896 kB
SwapFree:         995896 kB
Dirty:              3624 kB
Writeback:             0 kB
AnonPages:        162912 kB
Mapped:             3612 kB
Slab:             235888 kB
SReclaimable:      46408 kB
SUnreclaim:       189480 kB
PageTables:          384 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     2808616 kB
Committed_AS:     170648 kB
VmallocTotal:     122880 kB
VmallocUsed:        2876 kB
VmallocChunk:     109824 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       4096 kB
DirectMap4k:        8184 kB
DirectMap4M:      901120 kB

and vmallocinfo

cat /proc/vmallocinfo
0xf7ffe000-0xf8000000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfe6a000 ioremap
0xf8000000-0xf8007000 28672 acpi_tb_verify_table+0x1d/0x46
phys=dfef5000 ioremap
0xf8008000-0xf800a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef2000 ioremap
0xf800c000-0xf800e000 8192

acpi_ex_system_memory_space_handler+0xd6/0x208 phys=fed1f000 ioremap
0xf8010000-0xf8012000    8192 acpi_tb_verify_table+0x1d/0x46

phys=dfefb000 ioremap
0xf8014000-0xf8016000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef4000 ioremap
0xf8018000-0xf801a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef3000 ioremap
0xf801c000-0xf801e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef1000 ioremap
0xf8020000-0xf8022000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfef0000 ioremap
0xf8024000-0xf8026000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeef000 ioremap
0xf8028000-0xf802a000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeee000 ioremap
0xf802c000-0xf802e000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeed000 ioremap
0xf8030000-0xf8032000 8192 acpi_tb_verify_table+0x1d/0x46
phys=dfeec000 ioremap
0xf8038000-0xf803d000 20480 ich_force_enable_hpet+0x69/0x15a
phys=fed1c000 ioremap
0xf803e000-0xf8040000 8192 hpet_enable+0x2a/0x21b phys=fed00000 ioremap
0xf8040000-0xf8046000 24576 alloc_iommu+0x18d/0x1d4 phys=feb00000 ioremap

0xf8048000-0xf804a000    8192 pcim_iomap+0x2f/0x3a phys=e1b21000 ioremap
0xf804c000-0xf804e000    8192 e1000_probe+0x229/0xa73 phys=e1b20000 ioremap
0xf804f000-0xf8051000    8192 reiserfs_init_bitmap_cache+0x32/0x65

pages=1 vmalloc
0xf8052000-0xf8064000 73728 journal_init+0x30/0x82a pages=17 vmalloc
0xf8065000-0xf8067000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8068000-0xf806a000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806b000-0xf806d000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf806e000-0xf8070000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8071000-0xf8073000 8192 reiserfs_allocate_list_bitmaps+0x27/0x7e
pages=1 vmalloc
0xf8080000-0xf80a1000 135168 e1000_probe+0x1ca/0xa73 phys=e1b00000 ioremap
0xf80a2000-0xf80a6000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80a7000-0xf80ab000 16384 e1000e_setup_tx_resources+0x17/0x96
pages=3 vmalloc
0xf80ac000-0xf80b0000 16384 e1000e_setup_rx_resources+0x20/0xf7
pages=3 vmalloc
0xf80b1000-0xf80b5000 16384 e1000e_setup_tx_resources+0x17/0x96
pages=3 vmalloc

0xf80c0000-0xf80e1000  135168 e1000_probe+0x1ca/0xa73 phys=e1a60000 ioremap
0xf8100000-0xf8121000  135168 e1000_probe+0x1ca/0xa73 phys=e1a20000 ioremap
0xf8122000-0xf81b3000  593920 journal_init+0x65b/0x82a pages=144 vmalloc
0xf81b4000-0xf822f000  503808 sys_swapon+0x392/0x8f3 pages=122 vmalloc

0xf8bbc000-0xf8cbe000 1056768 tnode_new+0x35/0x65 pages=257 vmalloc

And next machine with kernel 2.6.29.3
dmesg:
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
Fix inflate_threshold_root. Now=15 size=11 bits
cat /proc/meminfo

MemTotal:        2072652 kB
MemFree:          496960 kB
Buffers:          267620 kB
Cached:           895212 kB
SwapCached:            0 kB
Active:           675744 kB
Inactive:         703312 kB
Active(anon):     215848 kB
Inactive(anon):        0 kB
Active(file):     459896 kB
Inactive(file):   703312 kB
Unevictable:           0 kB
Mlocked:               0 kB
HighTotal:       1186696 kB
HighFree:         151156 kB
LowTotal:         885956 kB
LowFree:          345804 kB
SwapTotal:       1975984 kB
SwapFree:        1975984 kB
Dirty:                20 kB
Writeback:             0 kB
AnonPages:        215724 kB
Mapped:             6120 kB
Slab:             186652 kB
SReclaimable:     125832 kB
SUnreclaim:        60820 kB
PageTables:          416 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:     3012308 kB
Committed_AS:     223692 kB
VmallocTotal:     122880 kB
VmallocUsed:        3192 kB
VmallocChunk:     108436 kB
HugePages_Total:       0
HugePages_Free:        0
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:       4096 kB
DirectMap4k:        8184 kB
DirectMap4M:      901120 kB

cat /proc/vmallocinfo
0xf7ffe000-0xf8000000 8192 acpi_tb_verify_table+0x1d/0x46
phys=7fee0000 ioremap
0xf8000000-0xf8005000 20480 acpi_tb_verify_table+0x1d/0x46
phys=7fee3000 ioremap
0xf8006000-0xf8008000 8192 acpi_tb_verify_table+0x1d/0x46
phys=7fee3000 ioremap
0xf800a000-0xf800c000 8192 acpi_tb_verify_table+0x1d/0x46
phys=7fee6000 ioremap
0xf800d000-0xf800f000 8192 reiserfs_init_bitmap_cache+0x3b/0x80
pages=1 vmalloc
0xf8010000-0xf8022000 73728 journal_init+0x30/0x8f0 pages=17 vmalloc
0xf8023000-0xf8025000 8192 reiserfs_allocate_list_bitmaps+0x2d/0x90
pages=1 vmalloc
0xf8026000-0xf8028000 8192 reiserfs_allocate_list_bitmaps+0x2d/0x90
pages=1 vmalloc
0xf8029000-0xf802b000 8192 reiserfs_allocate_list_bitmaps+0x2d/0x90
pages=1 vmalloc
0xf802c000-0xf802e000 8192 reiserfs_allocate_list_bitmaps+0x2d/0x90
pages=1 vmalloc
0xf802f000-0xf8031000 8192 reiserfs_allocate_list_bitmaps+0x2d/0x90
pages=1 vmalloc
0xf803e000-0xf8040000 8192 e1000_setup_all_tx_resources+0x57/0x660
pages=1 vmalloc
0xf8040000-0xf8061000 135168 e1000_probe+0x207/0xeb0 phys=f5000000 ioremap
0xf8062000-0xf8064000 8192 e1000_setup_all_rx_resources+0x57/0x6d0
pages=1 vmalloc
0xf8065000-0xf8067000 8192 e1000_setup_all_tx_resources+0x57/0x660
pages=1 vmalloc
0xf8068000-0xf806a000 8192 e1000_setup_all_rx_resources+0x57/0x6d0
pages=1 vmalloc
0xf806b000-0xf806d000 8192 e1000_setup_all_tx_resources+0x57/0x660
pages=1 vmalloc
0xf806e000-0xf8070000 8192 e1000_setup_all_rx_resources+0x57/0x6d0
pages=1 vmalloc

0xf8080000-0xf80a1000  135168 e1000_probe+0x207/0xeb0 phys=f1040000 ioremap
0xf80c0000-0xf80e1000  135168 e1000_probe+0x207/0xeb0 phys=f4000000 ioremap
0xf80e2000-0xf8173000  593920 journal_init+0x56e/0x8f0 pages=144 vmalloc
0xf8174000-0xf8267000  995328 sys_swapon+0x548/0xa30 pages=242 vmalloc

0xf8d17000-0xf8e19000 1056768 tnode_new+0x7f/0x90 pages=257 vmalloc

Quoted 8 lines
Quoted 4 lines

because i have this info on 5 machines that working in ibgp mesh And only one 64bit dev machine that is one of failover member - but i kill this machine after upgrade to kernel 2.6.31-rc1

Yes, I was a fool to ask you to try 2.6.31-rc1, sorry.

No problem with this test i lost only one test failover and no traffic lost when system switch to primary routers. :)

Quoted 2 lines

Even 2.6.30 is too young for a production machine.

I alvays make like this - i have iBGP mesh with main access path of machines on stable 2.6.28.9 kernels and second failover path on machines that use newest kernel for testing in this case 2.6.29 but after some problems i try also 2.6.30 yestarday.

Quoted 4 lines

2.6.29.5 contains the fixes, Pawel, did you tried this version ?

I will try 2.6.29.5 today

Thanks
Paweł Staszewski

#7 lore raw
Quoted 14 lines · Paweł Staszewski

Eric Dumazet pisze:

Quoted 4 lines

Yes, I was a fool to ask you to try 2.6.31-rc1, sorry.

No problem with this test i lost only one test failover and no traffic lost when system switch to primary routers. :)

Quoted 2 lines

Even 2.6.30 is too young for a production machine.

I alvays make like this - i have iBGP mesh with main access path of machines on stable 2.6.28.9 kernels and second failover path on machines that use newest kernel for testing in this case 2.6.29 but after some problems i try also 2.6.30 yestarday.

Quoted 6 lines
Quoted 4 lines

2.6.29.5 contains the fixes, Pawel, did you tried this version ?

I will try 2.6.29.5 today

OK thanks

Please report (while machine has enough load) output of

rtstat -c20 -i1

(rtstat is a symbolic link to lnstat, if not provided by your distro)

#8 lore raw
Quoted 12 lines

OK thanks

Please report (while machine has enough load) output of

rtstat -c20 -i1

(rtstat is a symbolic link to lnstat, if not provided by your distro)

here you have
rtstat -i 1 -c 20 rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|rt_cache|

entries|  in_hit|in_slow_|in_slow_|in_no_ro|  in_brd|in_marti|in_marti| out_hit|out_slow|out_slow|gc_total|gc_ignor|gc_goal_|gc_dst_o|in_hlist|out_hlis|
       |        |     tot|      mc|     ute|        |  an_dst|  
an_src|        |    _tot|     _mc|        |      ed|    miss| verflow| _search|t_search|
  93362|22930850| 1671866|       0|    1369|       2|       0|       
0|   53432|    1324|       0|       0|       0|       0|       0| 4783896|   11985|
  92067|  101426|    5315|       0|       2|       0|       0|       
0|     258|       2|       0|       0|       0|       0|       0|   21893|       6|
  90561|  100094|    4666|       0|       6|       0|       0|       
0|     267|       1|       0|       0|       0|       0|       0|   23433|      30|
  90101|   98672|    5630|       0|       2|       0|       0|       
0|     253|       0|       0|       0|       0|       0|       0|   24386|      34|
  89994|   99962|    5654|       0|       6|       0|       0|       
0|     266|       2|       0|       0|       0|       0|       0|   26251|      38|
  95209|   91974|   14860|       0|       9|       0|       0|       
0|     236|      31|       0|       0|       0|       0|       0|   14238|      35|
  95323|  101714|   10126|       0|      14|       0|       0|       
0|     255|       9|       0|       0|       0|       0|       0|    8532|      21|
  94814|   99918|    8539|       0|       5|       0|       0|       
0|     258|       4|       0|       0|       0|       0|       0|   11069|      24|
  98510|   93929|   12672|       0|      13|       0|       0|       
0|     238|      31|       0|       0|       0|       0|       0|   12704|      34|
  98983|   96131|   11128|       0|      12|       0|       0|       
0|     252|      10|       0|       0|       0|       0|       0|    7142|      18|
  98824|   99036|    8995|       0|       5|       0|       0|       
0|     256|       3|       0|       0|       0|       0|       0|    9343|      16|
  97868|  100032|    7544|       0|       5|       0|       0|       
0|     254|       1|       0|       0|       0|       0|       0|   11902|      17|
  96929|  101942|    6722|       0|       7|       0|       0|       
0|     263|       3|       0|       0|       0|       0|       0|   13778|      46|
  95932|  100725|    6217|       0|       4|       0|       0|       
0|     259|       3|       0|       0|       0|       0|       0|   15175|      47|
  94432|  102074|    5549|       0|       7|       0|       0|       
0|     268|       2|       0|       0|       0|       0|       0|   16996|      54|
  92986|  103602|    5187|       0|       3|       0|       0|       
0|     260|       0|       0|       0|       0|       0|       0|   18333|      43|
  91387|  103934|    4666|       0|       6|       0|       0|       
0|     261|       3|       0|       0|       0|       0|       0|   19316|      46|
  90615|  104916|    5333|       0|       5|       0|       0|       
0|     260|       4|       0|       0|       0|       0|       0|   21376|      48|
  89941|  101375|    5189|       0|       8|       0|       0|       
0|     270|       0|       0|       0|       0|       0|       0|   22249|      47|
  89744|  101425|    5529|       0|       6|       0|       0|       
0|     263|       4|       0|       0|       0|       0|       0|   24089|      57|

ath the same time cpu load

19:24:34     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
19:24:35     all    0.00    0.00    0.00    0.00    1.39   20.36    0.00    0.00   78.25
19:24:35       0    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:35       1    0.00    0.00    0.00    0.00    5.00   74.00    0.00    0.00   21.00
19:24:35       2    0.00    0.00    0.00    0.00    5.00   73.00    0.00    0.00   22.00
19:24:35       3    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:35       4    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:35       5    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:35       6    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:35       7    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:35     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
19:24:36     all    0.00    0.00    0.00    0.00    1.21   16.03    0.00    0.00   82.77
19:24:36       0    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:36       1    0.00    0.00    0.00    0.00    5.05   75.76    0.00    0.00   19.19
19:24:36       2    0.00    0.00    0.99    0.00    5.94   69.31    0.00    0.00   23.76
19:24:36       3    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:36       4    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:36       5    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:36       6    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:36       7    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:36     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
19:24:37     all    0.00    0.00    0.14    0.00    1.64   20.19    0.00    0.00   78.04
19:24:37       0    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:37       1    0.00    0.00    0.00    0.00    5.94   73.27    0.00    0.00   20.79
19:24:37       2    0.00    0.00    0.00    0.00    7.00   75.00    0.00    0.00   18.00
19:24:37       3    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:37       4    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:37       5    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:37       6    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:37       7    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:37     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
19:24:38     all    0.00    0.00    0.00    0.00    0.90   14.24    0.00    0.00   84.86
19:24:38       0    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:38       1    0.00    0.00    0.00    0.00    4.00   73.00    0.00    0.00   23.00
19:24:38       2    0.00    0.00    0.00    0.00    5.05   69.70    0.00    0.00   25.25
19:24:38       3    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:38       4    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:38       5    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:38       6    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:38       7    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:38     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
19:24:39     all    0.00    0.00    0.00    0.00    2.43   29.80    0.00    0.00   67.77
19:24:39       0    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:39       1    0.00    0.00    0.00    0.00    5.00   67.00    0.00    0.00   28.00
19:24:39       2    0.00    0.00    0.00    0.00    5.94   67.33    0.00    0.00   26.73
19:24:39       3    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:39       4    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:39       5    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:39       6    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:39       7    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:39     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
19:24:40     all    0.00    0.00    0.00    0.00    1.43   14.15    0.00    0.00   84.42
19:24:40       0    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:40       1    0.00    0.00    0.00    0.00    5.94   68.32    0.00    0.00   25.74
19:24:40       2    0.00    0.00    0.00    0.00    7.07   71.72    0.00    0.00   21.21
19:24:40       3    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:40       4    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:40       5    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:40       6    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:40       7    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:40     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
19:24:41     all    0.00    0.00    0.00    0.00    1.40   17.68    0.00    0.00   80.92
19:24:41       0    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:41       1    0.00    0.00    0.00    0.00    6.06   70.71    0.00    0.00   23.23
19:24:41       2    0.00    0.00    0.00    0.00    5.88   67.65    0.00    0.00   26.47
19:24:41       3    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:41       4    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:41       5    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:41       6    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00
19:24:41       7    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  100.00

cat /proc/interrupts
CPU0 CPU1 CPU2 CPU3 CPU4
CPU5 CPU6 CPU7
0: 42 0 0 1 0
1 0 0 IO-APIC-edge timer
1: 0 0 0 1 0
0 0 1 IO-APIC-edge i8042
9: 0 0 0 0 0
0 0 0 IO-APIC-fasteoi acpi

14:          0          0          0          0          0          0          0          0   IO-APIC-edge      ide0
15:          0          0          0          0          0          0          0          0   IO-APIC-edge      ide1
29:    1139988    4692793      89662          3          0          1          0          3   PCI-MSI-edge      eth0
30:          0          2    6207546          1          0          3          0          0   PCI-MSI-edge      eth1
31:          0          1          1          0          0          0          0          0   PCI-MSI-edge
32:          0          0          0          0          0          0          2          0   PCI-MSI-edge
33:          1          1          0          0          0          0          0          0   PCI-MSI-edge
34:          0          0          0          1          0          1          0          0   PCI-MSI-edge
35:          0          0          0          1          0          0          0          1   PCI-MSI-edge
36:          0          0          0          0          1          0          0          1   PCI-MSI-edge
37:          1          0          0          0          0          1          0          0   PCI-MSI-edge
38:          0          0          1          0          1          0          0          0   PCI-MSI-edge
39:          0          0          2          0          0          0          0          0   PCI-MSI-edge
40:          0          0          0          0          0          0          2          0   PCI-MSI-edge
41:          0          2          0          0          0          0          0          0   PCI-MSI-edge
42:          0          0          0          0          0          2          0          0   PCI-MSI-edge
43:          0          0          0          2          0          0          0          0   PCI-MSI-edge
44:          0          0          0          0          0          0          0          2   PCI-MSI-edge
45:          2          0          0          0          0          0          0          0   PCI-MSI-edge
46:          0          0          0          0          2          0          0          0   PCI-MSI-edge
48:        191        200        185        213        219        219        227        214   PCI-MSI-edge      ahci
49:          0          1          1          0          0          2          1          0   PCI-MSI-edge      ioat-msi
NMI:          0          0          0          0          0          0          0          0   Non-maskable interrupts
LOC:    1083019    6233788    7735401      12394      15178      10718      21192       8515   Local timer interrupts
RES:        921         44         33         20         13          8         10         12   Rescheduling interrupts
CAL:         20         85         88         87         90         90         91         86   Function call interrupts
TLB:        103        114        918        929         95        113        973        990   TLB shootdowns
SPU:          0          0          0          0          0          0          0          0   Spurious interrupts
ERR:          0
MIS:          0

i use smp_affinity
eth0 is on cpu1
eth1 is on cpu2

all test on kernel 2.6.29.5

and only one info in dmesg about Fix inflate_threshold Fix inflate_threshold_root. Now=15 size=11 bits

This info appear when bgpd process start to learn routes from peers

#9 lore raw
In reply to Eric Dumazet
Quoted 25 lines · Eric Dumazet

Pawe? Staszewski a ?crit :

Quoted 15 lines

Hello ALL

Some time ago i report this:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

and now with 2.6.29 / 2.6.29.1 / 2.6.29.3 and 2.6.30 it back dmesg output:
oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

Quoted 8 lines

Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

Quoted 7 lines

Curious, you seem to hit an old alloc_pages limit()... (MAX_ORDER allocation)

Your root node has 2^18 = 262144 pointers of 8 bytes -> 2097152 bytes (+ header -> 4194304 bytes)

But since following commit, we should use vmalloc() so this PAGE_SIZE<<10) limit should not anymore be applied.

On the other hand, even if there is no problem with memory, it seems because of hitting max_resize the threshold should be changed, e.g. by reverting the patch below.

Jarek P.

commit 965ffea43d4ebe8cd7b9fee78d651268dd7d23c5

Author: Robert Olsson <[email protected]>
Date:   Mon Mar 19 16:29:58 2007 -0700

[IPV4]: fib_trie root node settings

The threshold for root node can be more aggressive set to get better tree compression. The new setting mekes the root grow from 16 to 19 bits and substansial improvemnt in Aver depth this with the current table of 214393 prefixes

But really the dynamic resize should need more investigation both in terms convergence and performance and maybe it should be possible to change...

Maybe just for the brave to start with or we may have to back this out.

net/ipv4/fib_trie.c +2 -2
index 5d2b43d..9be7da7 100644
@@ -292,8 +292,8 @@ static inline void check_tnode(const struct tnode *tn)
 
 static int halve_threshold = 25;
 static int inflate_threshold = 50;
-static int halve_threshold_root = 15;
-static int inflate_threshold_root = 25;
+static int halve_threshold_root = 8;
+static int inflate_threshold_root = 15;
 
 
 static void __alias_free_mem(struct rcu_head *head)
#10 lore raw
Quoted 11 lines · Jarek Poplawski
Quoted 7 lines
Quoted 7 lines

oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

On the other hand, even if there is no problem with memory, it seems because of hitting max_resize the threshold should be changed, e.g. by reverting the patch below.

You seem to have some temporary memory problem. So the printout might be a bit misleading in this case. We really like to keep the root node as big as we can to keep the tree as flat as possible for performance reasons. (We're even more motivated now when we can disable the route cache)

So I'll guess the next insert/delete inflates the root node to be within the interval. So I'll assume this just a temporary failure?

I would be nice to have *threshholds* settable by /proc or /sys. I would use this in the other direction to trade memory for even faster lookups.

But maybe experts memory allocation has some good suggestions.

Cheers.

					--ro
#11 lore raw
Quoted 20 lines · Robert Olsson
Quoted 11 lines · Jarek Poplawski
Quoted 7 lines
Quoted 7 lines

oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

On the other hand, even if there is no problem with memory, it seems because of hitting max_resize the threshold should be changed, e.g. by reverting the patch below.

You seem to have some temporary memory problem. So the printout might be a bit misleading in this case. We really like to keep the root node as big as we can to keep the tree as flat as possible for performance reasons. (We're even more motivated now when we can disable the route cache)

Quoted 8 lines

So I'll guess the next insert/delete inflates the root node to be within the interval. So I'll assume this just a temporary failure?

I would be nice to have *threshholds* settable by /proc or /sys. I would use this in the other direction to trade memory for even faster lookups.

But maybe experts memory allocation has some good suggestions.

Pawel has reported these problems for a long time:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

So, until it's fully investigated, it seems some 'fast' fix is needed here.

Cheers,
Jarek P.

#12 lore raw
Quoted 5 lines · Jarek Poplawski

Pawel has reported these problems for a long time:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

So, until it's fully investigated, it seems some 'fast' fix is needed here.

I have never reported these problems but I am definitely seeing the same
message on kernel 2.6.29.5, usually, when one of my BGP peers goes down. So, just a "me too".

Regards,

Jorge

-----------------

[ 1198.333854] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1198.437028] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1198.460848] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1199.240223] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1199.279723] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1199.383081] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1200.154893] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1200.191711] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1200.223242] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1200.270299] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1200.355795] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1206.239254] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1206.271995] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1206.349351] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1206.384676] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1206.428801] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1206.457315] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1206.485710] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1206.513691] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1209.039681] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1209.069224] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1209.108840] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1209.141450] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1209.172317] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1209.197824] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1209.224711] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1209.251566] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1209.289603] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1211.561178] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1211.598062] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1211.633238] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1211.684420] Fix inflate_threshold_root. Now=15 size=11 bits
[ 1216.507853] Fix inflate_threshold_root. Now=15 size=11 bits
-----------------
cat /proc/meminfo
MemTotal:         515732 kB
MemFree:          139544 kB
Buffers:            4992 kB
Cached:             8488 kB
SwapCached:            0 kB
Active:           295904 kB
Inactive:           8132 kB
Active(anon):     291716 kB
Inactive(anon):        0 kB
Active(file):       4188 kB
Inactive(file):     8132 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:             0 kB
SwapFree:              0 kB
Dirty:                 0 kB
Writeback:             0 kB
AnonPages:        290556 kB
Mapped:             2320 kB
Slab:              42392 kB
SReclaimable:       1096 kB
SUnreclaim:        41296 kB
PageTables:          512 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:      257864 kB
Committed_AS:     294496 kB
VmallocTotal:     515448 kB
VmallocUsed:        3140 kB
VmallocChunk:     501096 kB
DirectMap4k:        8128 kB
DirectMap4M:      516096 kB
-----------------
cat /proc/vmallocinfo
0xe07f0000-0xe07f5000   20480 acpi_tb_verify_table+0x20/0x4a phys=1fff0000 ioremap
0xe07f6000-0xe07f8000    8192 acpi_tb_verify_table+0x20/0x4a phys=1ffff000 ioremap
0xe07fa000-0xe07fc000    8192 acpi_tb_verify_table+0x20/0x4a phys=1fff0000 ioremap
0xe07fe000-0xe0800000    8192 acpi_tb_verify_table+0x20/0x4a phys=1fff0000 ioremap
0xe0801000-0xe080d000   49152 cramfs_uncompress_init+0x18/0x57 pages=11 vmalloc
0xe080e000-0xe0810000    8192 e100_probe+0x1db/0x471 phys=fdde0000 ioremap
0xe0812000-0xe0814000    8192 e100_probe+0x1db/0x471 phys=fdd80000 ioremap
0xe0816000-0xe0818000    8192 e100_probe+0x1db/0x471 phys=fbbf0000 ioremap
0xe081a000-0xe081c000    8192 e100_probe+0x1db/0x471 phys=fbbe0000 ioremap
0xe081e000-0xe0820000    8192 ahc_linux_pci_reserve_mem_region+0x49/0x72

[aic7xxx] phys=fe9f0000 ioremap
0xe0820000-0xe0822000 8192 module_alloc_update_bounds+0x8/0x2c pages=1 vmalloc
0xe0822000-0xe0825000 12288 module_alloc_update_bounds+0x8/0x2c pages=2 vmalloc
0xe0826000-0xe0828000 8192 module_alloc_update_bounds+0x8/0x2c pages=1 vmalloc

0xe082d000-0xe0832000   20480 module_alloc_update_bounds+0x8/0x2c pages=4 vmalloc
0xe0833000-0xe0836000   12288 module_alloc_update_bounds+0x8/0x2c pages=2 vmalloc
0xe0838000-0xe083c000   16384 module_alloc_update_bounds+0x8/0x2c pages=3 vmalloc
0xe083c000-0xe0840000   16384 module_alloc_update_bounds+0x8/0x2c pages=3 vmalloc
0xe0840000-0xe0861000  135168 e1000_probe+0x18a/0x83b phys=febc0000 ioremap
0xe0862000-0xe0871000   61440 module_alloc_update_bounds+0x8/0x2c pages=14 vmalloc
0xe0875000-0xe087f000   40960 module_alloc_update_bounds+0x8/0x2c pages=9 vmalloc
0xe0880000-0xe0889000   36864 module_alloc_update_bounds+0x8/0x2c pages=8 vmalloc
0xe088d000-0xe0897000   40960 module_alloc_update_bounds+0x8/0x2c pages=9 vmalloc
0xe0897000-0xe08b5000  122880 module_alloc_update_bounds+0x8/0x2c pages=29 vmalloc
0xe08bd000-0xe08c1000   16384 module_alloc_update_bounds+0x8/0x2c pages=3 vmalloc
0xe08c7000-0xe08c9000    8192 module_alloc_update_bounds+0x8/0x2c pages=1 vmalloc
0xe08c9000-0xe08cc000   12288 module_alloc_update_bounds+0x8/0x2c pages=2 vmalloc
0xe08d3000-0xe08d7000   16384 module_alloc_update_bounds+0x8/0x2c pages=3 vmalloc
0xe08d8000-0xe08dd000   20480 module_alloc_update_bounds+0x8/0x2c pages=4 vmalloc
0xe08de000-0xe08e5000   28672 module_alloc_update_bounds+0x8/0x2c pages=6 vmalloc
0xe08e6000-0xe08e8000    8192 module_alloc_update_bounds+0x8/0x2c pages=1 vmalloc
0xe08e8000-0xe08ed000   20480 module_alloc_update_bounds+0x8/0x2c pages=4 vmalloc
0xe08f3000-0xe08f6000   12288 module_alloc_update_bounds+0x8/0x2c pages=2 vmalloc
0xe08ff000-0xe0902000   12288 module_alloc_update_bounds+0x8/0x2c pages=2 vmalloc
0xe090e000-0xe0914000   24576 module_alloc_update_bounds+0x8/0x2c pages=5 vmalloc
0xe091e000-0xe0922000   16384 module_alloc_update_bounds+0x8/0x2c pages=3 vmalloc
0xe092e000-0xe0934000   24576 module_alloc_update_bounds+0x8/0x2c pages=5 vmalloc
0xe0935000-0xe0942000   53248 module_alloc_update_bounds+0x8/0x2c pages=12 vmalloc

0xe095d000-0xe0979000 114688 module_alloc_update_bounds+0x8/0x2c pages=27 vmalloc
0xe097e000-0xe0980000 8192 ahc_linux_pci_reserve_mem_region+0x49/0x72
[aic7xxx] phys=fe9e0000 ioremap
0xe0990000-0xe0992000 8192 module_alloc_update_bounds+0x8/0x2c pages=1 vmalloc
0xe099a000-0xe099c000 8192 module_alloc_update_bounds+0x8/0x2c pages=1 vmalloc
0xe0a00000-0xe0b01000 1052672 he_start+0x204/0x1126 [he] phys=fe800000 ioremap 0xe14bf000-0xe15c1000 1056768 tnode_new+0x18/0x48 pages=257 vmalloc
0xe15ed000-0xe15f0000 12288 xt_alloc_table_info+0x68/0x97 [x_tables] pages=2
vmalloc
0xe15f1000-0xe15f4000 12288 xt_alloc_table_info+0x68/0x97 [x_tables] pages=2
vmalloc
-----------------

#13 lore raw
In reply to Jarek Poplawski
Quoted 7 lines · Jarek Poplawski
Quoted 1 line

But maybe memory allocation experts has some good suggestions.

Pawel has reported these problems for a long time:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

So, until it's fully investigated, it seems some 'fast' fix is needed here.

We talked about having a fixed pre-allocated root-node long ago but it's only optimisation for routers w. full BGP. Best if memory problems got solved.

Cheers

						--ro
#14 lore raw
Quoted 5 lines · Robert Olsson
Quoted 2 lines · Jarek Poplawski
Quoted 1 line

But maybe memory allocation experts has some good suggestions.

Quoted 9 lines
Quoted 5 lines

Pawel has reported these problems for a long time:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

So, until it's fully investigated, it seems some 'fast' fix is needed here.

We talked about having a fixed pre-allocated root-node long ago but it's only optimisation for routers w. full BGP. Best if memory problems got solved.

I think the current process of rebalancing can allocate and hold unnecessarily long a lot of 'temp' memory, so probably something like the patch below could be useful. It should be applied to the 2.6.30 after two patches below (from 2.6.31-rc). (Alas I can't even compile-test it now).

Cheers,
Jarek P.

--- (for testing)

 net/ipv4/fib_trie.c |   24 ++++++++++++++++++------
 1 files changed, 18 insertions(+), 6 deletions(-)
net/ipv4/fib_trie.c +18 -6
index 012cf5a..c2fc862 100644
@@ -366,6 +366,14 @@ static void __tnode_vfree(struct work_struct *arg)
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +410,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1020,19 +1028,23 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 		tnode_put_child_reorg((struct tnode *)tp, cindex,
 				      (struct node *)tn, wasfull);
 
-		tp = node_parent((struct node *) tn);
+		synchronize_rcu();
 		tnode_free_flush();
+		tp = node_parent((struct node *) tn);
 		if (!tp)
 			break;
 		tn = tp;
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }

---
commit e0f7cb8c8cc6cccce28d2ce39ad8c60d23c3799f

Author: Jarek Poplawski <[email protected]>
Date:   Mon Jun 15 02:31:29 2009 -0700

ipv4: Fix fib_trie rebalancing

While doing trie_rebalance(): resize(), inflate(), halve() RCU free tnodes before updating their parents. It depends on RCU delaying the real destruction, but if RCU readers start after call_rcu() and before parent update they could access freed memory.

It is currently prevented with preempt_disable() on the update side, but it's not safe, except maybe classic RCU, plus it conflicts with memory allocations with GFP_KERNEL flag used from these functions.

This patch explicitly delays freeing of tnodes by adding them to the list, which is flushed after the update is finished.

net/ipv4/fib_trie.c +37 -10

122 lines not on this page · raw message

---
commit 7b85576d15bf2574b0a451108f59f9ad4170dd3f

Author: Jarek Poplawski <[email protected]>
Date:   Thu Jun 18 00:28:51 2009 -0700

ipv4: Fix fib_trie rebalancing, part 2

My previous patch, which explicitly delays freeing of tnodes by adding them to the list to flush them after the update is finished, isn't strict enough. It treats exceptionally tnodes without parent, assuming they are newly created, so "invisible" for the read side yet.

But the top tnode doesn't have parent as well, so we have to exclude all exceptions (at least until a better way is found). Additionally we need to move rcu assignment of this node before flushing, so the return type of the trie_rebalance() function is changed.

net/ipv4/fib_trie.c +10 -14
index d1a39b1..012cf5a 100644
@@ -391,13 +391,8 @@ static inline void tnode_free(struct tnode *tn)
 static void tnode_free_safe(struct tnode *tn)
 {
 	BUG_ON(IS_LEAF(tn));
-
-	if (node_parent((struct node *) tn)) {
-		tn->tnode_free = tnode_free_head;
-		tnode_free_head = tn;
-	} else {
-		tnode_free(tn);
-	}
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
 }
 
 static void tnode_free_flush(void)
@@ -1009,7 +1004,7 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
 	t_key cindex, key;
@@ -1033,12 +1028,13 @@ static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn)) {
+	if (IS_TNODE(tn))
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-		tnode_free_flush();
-	}
 
-	return (struct node *)tn;
+	rcu_assign_pointer(t->trie, (struct node *)tn);
+	tnode_free_flush();
+
+	return;
 }
 
 /* only used from updater-side */
@@ -1186,7 +1182,7 @@ static struct list_head *fib_insert_node(struct trie *t, u32 key, int plen)
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp);
 done:
 	return fa_head;
 }
@@ -1605,7 +1601,7 @@ static void trie_leaf_remove(struct trie *t, struct leaf *l)
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 
#15 lore raw
Quoted 16 lines · Jarek Poplawski
Quoted 5 lines · Robert Olsson
Quoted 2 lines · Jarek Poplawski
Quoted 1 line

But maybe memory allocation experts has some good suggestions.

Quoted 9 lines
Quoted 5 lines

Pawel has reported these problems for a long time:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

So, until it's fully investigated, it seems some 'fast' fix is needed here.

We talked about having a fixed pre-allocated root-node long ago but it's only optimisation for routers w. full BGP. Best if memory problems got solved.

Quoted 6 lines

I think the current process of rebalancing can allocate and hold unnecessarily long a lot of 'temp' memory, so probably something like the patch below could be useful. It should be applied to the 2.6.30 after two patches below (from 2.6.31-rc). (Alas I can't even compile-test it now).

Alternatively here is a faster version with less synchronize_rcu().

Jarek P.

--- (take 2 - for testing)

 net/ipv4/fib_trie.c |   27 +++++++++++++++++++++------
 1 files changed, 21 insertions(+), 6 deletions(-)
net/ipv4/fib_trie.c +21 -6
index 012cf5a..2936b2e 100644
@@ -366,6 +366,14 @@ static void __tnode_vfree(struct work_struct *arg)
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +410,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1021,18 +1029,25 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
-		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }

---
commit e0f7cb8c8cc6cccce28d2ce39ad8c60d23c3799f

Author: Jarek Poplawski <[email protected]>
Date:   Mon Jun 15 02:31:29 2009 -0700

ipv4: Fix fib_trie rebalancing

While doing trie_rebalance(): resize(), inflate(), halve() RCU free tnodes before updating their parents. It depends on RCU delaying the real destruction, but if RCU readers start after call_rcu() and before parent update they could access freed memory.

It is currently prevented with preempt_disable() on the update side, but it's not safe, except maybe classic RCU, plus it conflicts with memory allocations with GFP_KERNEL flag used from these functions.

This patch explicitly delays freeing of tnodes by adding them to the list, which is flushed after the update is finished.

net/ipv4/fib_trie.c +37 -10

122 lines not on this page · raw message

---
commit 7b85576d15bf2574b0a451108f59f9ad4170dd3f

Author: Jarek Poplawski <[email protected]>
Date:   Thu Jun 18 00:28:51 2009 -0700

ipv4: Fix fib_trie rebalancing, part 2

My previous patch, which explicitly delays freeing of tnodes by adding them to the list to flush them after the update is finished, isn't strict enough. It treats exceptionally tnodes without parent, assuming they are newly created, so "invisible" for the read side yet.

But the top tnode doesn't have parent as well, so we have to exclude all exceptions (at least until a better way is found). Additionally we need to move rcu assignment of this node before flushing, so the return type of the trie_rebalance() function is changed.

net/ipv4/fib_trie.c +10 -14
index d1a39b1..012cf5a 100644
@@ -391,13 +391,8 @@ static inline void tnode_free(struct tnode *tn)
 static void tnode_free_safe(struct tnode *tn)
 {
 	BUG_ON(IS_LEAF(tn));
-
-	if (node_parent((struct node *) tn)) {
-		tn->tnode_free = tnode_free_head;
-		tnode_free_head = tn;
-	} else {
-		tnode_free(tn);
-	}
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
 }
 
 static void tnode_free_flush(void)
@@ -1009,7 +1004,7 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
 	t_key cindex, key;
@@ -1033,12 +1028,13 @@ static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn)) {
+	if (IS_TNODE(tn))
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-		tnode_free_flush();
-	}
 
-	return (struct node *)tn;
+	rcu_assign_pointer(t->trie, (struct node *)tn);
+	tnode_free_flush();
+
+	return;
 }
 
 /* only used from updater-side */
@@ -1186,7 +1182,7 @@ static struct list_head *fib_insert_node(struct trie *t, u32 key, int plen)
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp);
 done:
 	return fa_head;
 }
@@ -1605,7 +1601,7 @@ static void trie_leaf_remove(struct trie *t, struct leaf *l)
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 
#16 lore raw

Jarek Poplawski writes:

Thanks,

Should be worth testing so we synchronize_rcu instead of doing call_rcu's

Cheers

					--ro
Quoted 321 lines

Alternatively here is a faster version with less synchronize_rcu().

Jarek P.

--- (take 2 - for testing)

 net/ipv4/fib_trie.c |   27 +++++++++++++++++++++------
 1 files changed, 21 insertions(+), 6 deletions(-)
net/ipv4/fib_trie.c +21 -7
index 012cf5a..2936b2e 100644
@@ -366,6 +366,14 @@ static void __tnode_vfree(struct work_struct *arg)
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +410,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1021,18 +1029,25 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
-		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }
 
 
 
---

commit e0f7cb8c8cc6cccce28d2ce39ad8c60d23c3799f

Author: Jarek Poplawski <[email protected]>
Date:   Mon Jun 15 02:31:29 2009 -0700

ipv4: Fix fib_trie rebalancing

While doing trie_rebalance(): resize(), inflate(), halve() RCU free tnodes before updating their parents. It depends on RCU delaying the real destruction, but if RCU readers start after call_rcu() and before parent update they could access freed memory.

It is currently prevented with preempt_disable() on the update side, but it's not safe, except maybe classic RCU, plus it conflicts with memory allocations with GFP_KERNEL flag used from these functions.

This patch explicitly delays freeing of tnodes by adding them to the list, which is flushed after the update is finished.

net/ipv4/fib_trie.c +37 -11
index 538d2a9..d1a39b1 100644
@@ -123,6 +123,7 @@ struct tnode {
 	union {
 		struct rcu_head rcu;
 		struct work_struct work;
+		struct tnode *tnode_free;
 	};
 	struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct tnode *tn, int i, struct node *n,
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -385,6 +388,29 @@ static inline void tnode_free(struct tnode *tn)
 		call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+	BUG_ON(IS_LEAF(tn));
+
+	if (node_parent((struct node *) tn)) {
+		tn->tnode_free = tnode_free_head;
+		tnode_free_head = tn;
+	} else {
+		tnode_free(tn);
+	}
+}
+
+static void tnode_free_flush(void)
+{
+	struct tnode *tn;
+
+	while ((tn = tnode_free_head)) {
+		tnode_free_head = tn->tnode_free;
+		tn->tnode_free = NULL;
+		tnode_free(tn);
+	}
+}
+
 static struct leaf *leaf_new(void)
 {
 	struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +521,7 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 
 	/* No children */
 	if (tn->empty_children == tnode_child_length(tn)) {
-		tnode_free(tn);
+		tnode_free_safe(tn);
 		return NULL;
 	}
 	/* One child */
@@ -509,7 +535,7 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 
 			/* compress one level */
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 	/*
@@ -670,7 +696,7 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 			/* compress one level */
 
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 
@@ -756,7 +782,7 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn)
 			put_child(t, tn, 2*i, inode->child[0]);
 			put_child(t, tn, 2*i+1, inode->child[1]);
 
-			tnode_free(inode);
+			tnode_free_safe(inode);
 			continue;
 		}
 
@@ -801,9 +827,9 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn)
 		put_child(t, tn, 2*i, resize(t, left));
 		put_child(t, tn, 2*i+1, resize(t, right));
 
-		tnode_free(inode);
+		tnode_free_safe(inode);
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -885,7 +911,7 @@ static struct tnode *halve(struct trie *t, struct tnode *tn)
 		put_child(t, newBinNode, 1, right);
 		put_child(t, tn, i/2, resize(t, newBinNode));
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -989,7 +1015,6 @@ static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
 	t_key cindex, key;
 	struct tnode *tp;
 
-	preempt_disable();
 	key = tn->key;
 
 	while (tn != NULL && (tp = node_parent((struct node *)tn)) != NULL) {
@@ -1001,16 +1026,18 @@ static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
+		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
+		tnode_free_flush();
+	}
 
-	preempt_enable();
 	return (struct node *)tn;
 }
 
---

commit 7b85576d15bf2574b0a451108f59f9ad4170dd3f

Author: Jarek Poplawski <[email protected]>
Date:   Thu Jun 18 00:28:51 2009 -0700

ipv4: Fix fib_trie rebalancing, part 2

My previous patch, which explicitly delays freeing of tnodes by adding them to the list to flush them after the update is finished, isn't strict enough. It treats exceptionally tnodes without parent, assuming they are newly created, so "invisible" for the read side yet.

But the top tnode doesn't have parent as well, so we have to exclude all exceptions (at least until a better way is found). Additionally we need to move rcu assignment of this node before flushing, so the return type of the trie_rebalance() function is changed.

net/ipv4/fib_trie.c +10 -15
index d1a39b1..012cf5a 100644
@@ -391,13 +391,8 @@ static inline void tnode_free(struct tnode *tn)
 static void tnode_free_safe(struct tnode *tn)
 {
 	BUG_ON(IS_LEAF(tn));
-
-	if (node_parent((struct node *) tn)) {
-		tn->tnode_free = tnode_free_head;
-		tnode_free_head = tn;
-	} else {
-		tnode_free(tn);
-	}
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
 }
 
 static void tnode_free_flush(void)
@@ -1009,7 +1004,7 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
 	t_key cindex, key;
@@ -1033,12 +1028,13 @@ static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn)) {
+	if (IS_TNODE(tn))
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-		tnode_free_flush();
-	}
 
-	return (struct node *)tn;
+	rcu_assign_pointer(t->trie, (struct node *)tn);
+	tnode_free_flush();
+
+	return;
 }
 
 /* only used from updater-side */
@@ -1186,7 +1182,7 @@ static struct list_head *fib_insert_node(struct trie *t, u32 key, int plen)
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp);
 done:
 	return fa_head;
 }
@@ -1605,7 +1601,7 @@ static void trie_leaf_remove(struct trie *t, struct leaf *l)
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#17 lore raw
Quoted 7 lines · Robert Olsson

Jarek Poplawski writes:

Thanks,

Should be worth testing so we synchronize_rcu instead of doing call_rcu's

Alas take 2 (nor 1) doesn't compile, so here it is again.

Thanks,
Jarek P.
--- (take 3 - for testing)

 net/ipv4/fib_trie.c |   30 ++++++++++++++++++++++++------
 1 files changed, 24 insertions(+), 6 deletions(-)
net/ipv4/fib_trie.c +24 -6
index 012cf5a..1a4c4b7 100644
@@ -366,6 +366,17 @@ static void __tnode_vfree(struct work_struct *arg)
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	size_t size = sizeof(struct tnode) +
+		      (sizeof(struct node *) << tn->bits);
+
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +413,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1021,18 +1032,25 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
-		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }
#18 lore raw
Quoted 10 lines · Jarek Poplawski
Quoted 7 lines · Robert Olsson

Jarek Poplawski writes:

Thanks,

Should be worth testing so we synchronize_rcu instead of doing call_rcu's

Alas take 2 (nor 1) doesn't compile, so here it is again.

So the idea is to balance memory and latency, so that large changes (those affecting the root node) get at least one synchronize_rcu(), while smaller changes just use call_rcu(), correct? This means that the amount of memory awaiting an RCU grace period is limited, but the algorithm avoids per-node synchronize_rcu() overhead.

If I understand the goal correctly, looks good! (Give or take my limited understanding of fib_trie and is usage, of course.)

Thanx, Paul

Quoted 73 lines

Thanks,
Jarek P.
--- (take 3 - for testing)

 net/ipv4/fib_trie.c |   30 ++++++++++++++++++++++++------
 1 files changed, 24 insertions(+), 6 deletions(-)
net/ipv4/fib_trie.c +24 -7
index 012cf5a..1a4c4b7 100644
@@ -366,6 +366,17 @@ static void __tnode_vfree(struct work_struct *arg)
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	size_t size = sizeof(struct tnode) +
+		      (sizeof(struct node *) << tn->bits);
+
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +413,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1021,18 +1032,25 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
-		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#19 lore raw
Quoted 12 lines · Paul E. McKenney
Quoted 10 lines · Jarek Poplawski
Quoted 7 lines · Robert Olsson

Jarek Poplawski writes:

Thanks,

Should be worth testing so we synchronize_rcu instead of doing call_rcu's

Alas take 2 (nor 1) doesn't compile, so here it is again.

Quoted 8 lines

So the idea is to balance memory and latency, so that large changes (those affecting the root node) get at least one synchronize_rcu(), while smaller changes just use call_rcu(), correct? This means that the amount of memory awaiting an RCU grace period is limited, but the algorithm avoids per-node synchronize_rcu() overhead.

If I understand the goal correctly, looks good! (Give or take my limited understanding of fib_trie and is usage, of course.)

The goal is practically to replace all call_rcu() during trie_rebalance() with synchronize_rcu() (except some freeing after ENOMEM). I guess currently (<= 2.6.30) call_rcu() can free this memory after trie_rebalance() has finished, that's why there were problems with enabled preemption. So this patch tries to do/force this a bit earlier - at least before the top/largest node is rebalanced.

Thanks,
Jarek P.

Quoted 76 lines

Thanx, Paul

Quoted 73 lines

Thanks,
Jarek P.
--- (take 3 - for testing)

 net/ipv4/fib_trie.c |   30 ++++++++++++++++++++++++------
 1 files changed, 24 insertions(+), 6 deletions(-)
net/ipv4/fib_trie.c +24 -7
index 012cf5a..1a4c4b7 100644
@@ -366,6 +366,17 @@ static void __tnode_vfree(struct work_struct *arg)
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	size_t size = sizeof(struct tnode) +
+		      (sizeof(struct node *) << tn->bits);
+
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +413,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1021,18 +1032,25 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
-		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#20 lore raw
Quoted 22 lines · Jarek Poplawski
Quoted 12 lines · Paul E. McKenney
Quoted 10 lines · Jarek Poplawski
Quoted 7 lines · Robert Olsson

Jarek Poplawski writes:

Thanks,

Should be worth testing so we synchronize_rcu instead of doing call_rcu's

Alas take 2 (nor 1) doesn't compile, so here it is again.

Quoted 8 lines

So the idea is to balance memory and latency, so that large changes (those affecting the root node) get at least one synchronize_rcu(), while smaller changes just use call_rcu(), correct? This means that the amount of memory awaiting an RCU grace period is limited, but the algorithm avoids per-node synchronize_rcu() overhead.

If I understand the goal correctly, looks good! (Give or take my limited understanding of fib_trie and is usage, of course.)

Quoted 7 lines

The goal is practically to replace all call_rcu() during trie_rebalance() with synchronize_rcu() (except some freeing after ENOMEM). I guess currently (<= 2.6.30) call_rcu() can free this memory after trie_rebalance() has finished, that's why there were problems with enabled preemption. So this patch tries to do/force this a bit earlier - at least before the top/largest node is rebalanced.

On the other hand, we could probably stay with call_rcu() plus only one synchronize_rcu() before the top node's resize() if you think it's enough here?

Thanks,
Jarek P.

Quoted 77 lines
Quoted 77 lines

Thanx, Paul

Quoted 73 lines

Thanks,
Jarek P.
--- (take 3 - for testing)

 net/ipv4/fib_trie.c |   30 ++++++++++++++++++++++++------
 1 files changed, 24 insertions(+), 6 deletions(-)
net/ipv4/fib_trie.c +24 -7
index 012cf5a..1a4c4b7 100644
@@ -366,6 +366,17 @@ static void __tnode_vfree(struct work_struct *arg)
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	size_t size = sizeof(struct tnode) +
+		      (sizeof(struct node *) << tn->bits);
+
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +413,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1021,18 +1032,25 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
-		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#21 lore raw
Quoted 31 lines · Jarek Poplawski
Quoted 22 lines · Jarek Poplawski
Quoted 12 lines · Paul E. McKenney
Quoted 10 lines · Jarek Poplawski
Quoted 7 lines · Robert Olsson

Jarek Poplawski writes:

Thanks,

Should be worth testing so we synchronize_rcu instead of doing call_rcu's

Alas take 2 (nor 1) doesn't compile, so here it is again.

Quoted 8 lines

So the idea is to balance memory and latency, so that large changes (those affecting the root node) get at least one synchronize_rcu(), while smaller changes just use call_rcu(), correct? This means that the amount of memory awaiting an RCU grace period is limited, but the algorithm avoids per-node synchronize_rcu() overhead.

If I understand the goal correctly, looks good! (Give or take my limited understanding of fib_trie and is usage, of course.)

Quoted 7 lines

The goal is practically to replace all call_rcu() during trie_rebalance() with synchronize_rcu() (except some freeing after ENOMEM). I guess currently (<= 2.6.30) call_rcu() can free this memory after trie_rebalance() has finished, that's why there were problems with enabled preemption. So this patch tries to do/force this a bit earlier - at least before the top/largest node is rebalanced.

Quoted 3 lines

On the other hand, we could probably stay with call_rcu() plus only one synchronize_rcu() before the top node's resize() if you think it's enough here?

Well, my first task is to understand the problem/goal. ;-)

My guess from what you said above is that use of call_rcu(), when combined with changes to the trie in rapid succession, is resulting in excessive memory awaiting a grace period. Is this the case, or am I confused?

Thanx, Paul

Quoted 80 lines

Thanks,
Jarek P.

Quoted 77 lines
Quoted 77 lines

Thanx, Paul

Quoted 73 lines

Thanks,
Jarek P.
--- (take 3 - for testing)

 net/ipv4/fib_trie.c |   30 ++++++++++++++++++++++++------
 1 files changed, 24 insertions(+), 6 deletions(-)
net/ipv4/fib_trie.c +24 -7
index 012cf5a..1a4c4b7 100644
@@ -366,6 +366,17 @@ static void __tnode_vfree(struct work_struct *arg)
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	size_t size = sizeof(struct tnode) +
+		      (sizeof(struct node *) << tn->bits);
+
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +413,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1021,18 +1032,25 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
-		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#22 lore raw
Quoted 36 lines · Paul E. McKenney
Quoted 31 lines · Jarek Poplawski
Quoted 22 lines · Jarek Poplawski
Quoted 12 lines · Paul E. McKenney
Quoted 10 lines · Jarek Poplawski
Quoted 7 lines · Robert Olsson

Jarek Poplawski writes:

Thanks,

Should be worth testing so we synchronize_rcu instead of doing call_rcu's

Alas take 2 (nor 1) doesn't compile, so here it is again.

Quoted 8 lines

So the idea is to balance memory and latency, so that large changes (those affecting the root node) get at least one synchronize_rcu(), while smaller changes just use call_rcu(), correct? This means that the amount of memory awaiting an RCU grace period is limited, but the algorithm avoids per-node synchronize_rcu() overhead.

If I understand the goal correctly, looks good! (Give or take my limited understanding of fib_trie and is usage, of course.)

Quoted 7 lines

The goal is practically to replace all call_rcu() during trie_rebalance() with synchronize_rcu() (except some freeing after ENOMEM). I guess currently (<= 2.6.30) call_rcu() can free this memory after trie_rebalance() has finished, that's why there were problems with enabled preemption. So this patch tries to do/force this a bit earlier - at least before the top/largest node is rebalanced.

Quoted 3 lines

On the other hand, we could probably stay with call_rcu() plus only one synchronize_rcu() before the top node's resize() if you think it's enough here?

Quoted 6 lines

Well, my first task is to understand the problem/goal. ;-)

My guess from what you said above is that use of call_rcu(), when combined with changes to the trie in rapid succession, is resulting in excessive memory awaiting a grace period. Is this the case, or am I confused?

Exactly! (I guess... ;-)

Thanks,
Jarek P.

Quoted 78 lines
Quoted 78 lines
Quoted 78 lines
Quoted 78 lines

Thanx, Paul

Quoted 73 lines

Thanks,
Jarek P.
--- (take 3 - for testing)

 net/ipv4/fib_trie.c |   30 ++++++++++++++++++++++++------
 1 files changed, 24 insertions(+), 6 deletions(-)
net/ipv4/fib_trie.c +24 -7
index 012cf5a..1a4c4b7 100644
@@ -366,6 +366,17 @@ static void __tnode_vfree(struct work_struct *arg)
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	size_t size = sizeof(struct tnode) +
+		      (sizeof(struct node *) << tn->bits);
+
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +413,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1021,18 +1032,25 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
-		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#23 lore raw
Quoted 37 lines · Jarek Poplawski
Quoted 36 lines · Paul E. McKenney
Quoted 34 lines · Jarek Poplawski
Quoted 22 lines · Jarek Poplawski
Quoted 12 lines · Paul E. McKenney
Quoted 10 lines · Jarek Poplawski
Quoted 7 lines · Robert Olsson

Jarek Poplawski writes:

Thanks,

Should be worth testing so we synchronize_rcu instead of doing call_rcu's

Alas take 2 (nor 1) doesn't compile, so here it is again.

Quoted 8 lines

So the idea is to balance memory and latency, so that large changes (those affecting the root node) get at least one synchronize_rcu(), while smaller changes just use call_rcu(), correct? This means that the amount of memory awaiting an RCU grace period is limited, but the algorithm avoids per-node synchronize_rcu() overhead.

If I understand the goal correctly, looks good! (Give or take my limited understanding of fib_trie and is usage, of course.)

Quoted 7 lines

The goal is practically to replace all call_rcu() during trie_rebalance() with synchronize_rcu() (except some freeing after ENOMEM). I guess currently (<= 2.6.30) call_rcu() can free this memory after trie_rebalance() has finished, that's why there were problems with enabled preemption. So this patch tries to do/force this a bit earlier - at least before the top/largest node is rebalanced.

On the other hand, we could probably stay with call_rcu() plus only one synchronize_rcu() before the top node's resize() if you think it's enough here?

Quoted 8 lines
Quoted 6 lines

Well, my first task is to understand the problem/goal. ;-)

My guess from what you said above is that use of call_rcu(), when combined with changes to the trie in rapid succession, is resulting in excessive memory awaiting a grace period. Is this the case, or am I confused?

Exactly! (I guess... ;-)

;-)

In that case, simply invoking synchronize_rcu() every once and awhile should take care of things. This could be at the end of every large trie operation, or you could even count the call_rcu() invocations and do a synchronize_rcu() every 100th, 1,000th, or whatever, based on the amount of memory available.

Thanx, Paul

#24 lore raw

On Fri, Jun 26, 2009 at 10:05:38AM -0700, Paul E. McKenney wrote:
...

Quoted 5 lines

In that case, simply invoking synchronize_rcu() every once and awhile should take care of things. This could be at the end of every large trie operation, or you could even count the call_rcu() invocations and do a synchronize_rcu() every 100th, 1,000th, or whatever, based on the amount of memory available.

OK, for now the minimal change for testing (2.6.30 needs previously mentioned two commits from 2.6.31-rc). (I guess I'll send it with a changelog after net-next is opened.)

Thanks,
Jarek P.
--- (take 4 - for testing)

 net/ipv4/fib_trie.c |    8 ++++++--
 1 files changed, 6 insertions(+), 2 deletions(-)
net/ipv4/fib_trie.c +6 -2
index 012cf5a..98b31a1 100644
@@ -1008,7 +1008,7 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
 	t_key cindex, key;
-	struct tnode *tp;
+	struct tnode *tp, *oldtnode = tn;
 
 	key = tn->key;
 
@@ -1028,8 +1028,12 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
+		/* force memory freeing after last changes */
+		if (oldtnode != tn)
+			synchronize_rcu();
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
+	}
 
 	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();
#25 lore raw
Quoted 11 lines · Jarek Poplawski

On Fri, Jun 26, 2009 at 10:05:38AM -0700, Paul E. McKenney wrote:
...

Quoted 5 lines

In that case, simply invoking synchronize_rcu() every once and awhile should take care of things. This could be at the end of every large trie operation, or you could even count the call_rcu() invocations and do a synchronize_rcu() every 100th, 1,000th, or whatever, based on the amount of memory available.

OK, for now the minimal change for testing (2.6.30 needs previously mentioned two commits from 2.6.31-rc). (I guess I'll send it with a changelog after net-next is opened.)

Looks promising to me!!!

Thanx, Paul

Quoted 34 lines

Thanks,
Jarek P.
--- (take 4 - for testing)

 net/ipv4/fib_trie.c |    8 ++++++--
 1 files changed, 6 insertions(+), 2 deletions(-)
net/ipv4/fib_trie.c +6 -2
index 012cf5a..98b31a1 100644
@@ -1008,7 +1008,7 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
 	t_key cindex, key;
-	struct tnode *tp;
+	struct tnode *tp, *oldtnode = tn;
 
 	key = tn->key;
 
@@ -1028,8 +1028,12 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
+		/* force memory freeing after last changes */
+		if (oldtnode != tn)
+			synchronize_rcu();
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
+	}
 
 	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();
#26 lore raw
Quoted 9 lines · Paul E. McKenney
Quoted 8 lines · Jarek Poplawski

On Fri, Jun 26, 2009 at 10:05:38AM -0700, Paul E. McKenney wrote:
...

Quoted 5 lines

In that case, simply invoking synchronize_rcu() every once and awhile should take care of things. This could be at the end of every large trie operation, or you could even count the call_rcu() invocations and do a synchronize_rcu() every 100th, 1,000th, or whatever, based on the amount of memory available.

Quoted 6 lines
Quoted 3 lines

OK, for now the minimal change for testing (2.6.30 needs previously mentioned two commits from 2.6.31-rc). (I guess I'll send it with a changelog after net-next is opened.)

Looks promising to me!!!

Alas, after rethinking, there is one detail which bothers me. Those largest allocs here are done with vmalloc and freed with RCU by schedule_work(). So, I wonder if just because of this, the previous version doing it directly isn't more reliable anyway. Of course, it's my bad I didn't point it while describing the problem earlier. (I knew I missed something...;-)

Thanks,
Jarek P.

#27 lore raw
In reply to Jarek Poplawski

Yes looks like a good solution but maybe it safest to synchronize unconditionally?

net/ipv4/fib_trie.c +4 -1
index 012cf5a..9cb8623 100644
@@ -1028,8 +1028,11 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
+		/* force memory freeing after last changes */
+		synchronize_rcu();
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
+	}
 
 	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();

Cheers

						--ro
Quoted 34 lines · Jarek Poplawski
 net/ipv4/fib_trie.c |    8 ++++++--
 1 files changed, 6 insertions(+), 2 deletions(-)
net/ipv4/fib_trie.c +6 -3
index 012cf5a..98b31a1 100644
@@ -1008,7 +1008,7 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
 	t_key cindex, key;
-	struct tnode *tp;
+	struct tnode *tp, *oldtnode = tn;
 
 	key = tn->key;
 
@@ -1028,8 +1028,12 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
+		/* force memory freeing after last changes */
+		if (oldtnode != tn)
+			synchronize_rcu();
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
+	}
 
 	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#28 lore raw
Quoted 3 lines · Robert Olsson

Yes looks like a good solution but maybe it safest to synchronize unconditionally?

Hmm... I lost around half an hour for this doubt... Sure! (Unless there are some strange cases which very often create and destroy very small tables?)

Thanks,
Jarek P.

Quoted 59 lines
net/ipv4/fib_trie.c +4 -1
index 012cf5a..9cb8623 100644
@@ -1028,8 +1028,11 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
+		/* force memory freeing after last changes */
+		synchronize_rcu();
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
+	}
 
 	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();

Cheers

						--ro
Quoted 34 lines · Jarek Poplawski
 net/ipv4/fib_trie.c |    8 ++++++--
 1 files changed, 6 insertions(+), 2 deletions(-)
net/ipv4/fib_trie.c +6 -3
index 012cf5a..98b31a1 100644
@@ -1008,7 +1008,7 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
 	t_key cindex, key;
-	struct tnode *tp;
+	struct tnode *tp, *oldtnode = tn;
 
 	key = tn->key;
 
@@ -1028,8 +1028,12 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 	}
 
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
+		/* force memory freeing after last changes */
+		if (oldtnode != tn)
+			synchronize_rcu();
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
+	}
 
 	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#29 lore raw
Quoted 8 lines · Jarek Poplawski
Quoted 3 lines · Robert Olsson

Yes looks like a good solution but maybe it safest to synchronize unconditionally?

Hmm... I lost around half an hour for this doubt... Sure! (Unless there are some strange cases which very often create and destroy very small tables?)

...or maybe even only updating such small tables very often?

Btw., Robert, I wondered about some design details lately, especially about pointer to a parent. I didn't see it in the basic docs, so it seems it could be avoided. It seems to be a problem with RCU, unless I miss something: if there were no going back from children to parents it seems we could free those "temporary" (created by inflate() and halve() and destroyed before resize() has finished) earlier.

Another problem with this, it seems, are possibly false lookups (if we go back to the new parent which doesn't have it's parent or other nodes updated). So, was there so much performance gain to introduce this?

Thanks,
Jarek P.

#30 lore raw
In reply to Robert Olsson

Robert Olsson wrote, On 06/26/2009 11:19 AM:

Quoted 19 lines
Quoted 11 lines · Jarek Poplawski
Quoted 7 lines
Quoted 7 lines

oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits Fix inflate_threshold_root. Now=15 size=11 bits

On the other hand, even if there is no problem with memory, it seems because of hitting max_resize the threshold should be changed, e.g. by reverting the patch below.

You seem to have some temporary memory problem. So the printout might be a bit misleading in this case. We really like to keep the root node as big as we can to keep the tree as flat as possible for performance reasons. (We're even more motivated now when we can disable the route cache)

Quoted 7 lines

So I'll guess the next insert/delete inflates the root node to be within the interval. So I'll assume this just a temporary failure?

I would be nice to have *threshholds* settable by /proc or /sys. I would use this in the other direction to trade memory for even faster lookups.

But maybe experts memory allocation has some good suggestions.

Robert, you and Eric pointed at memory problems, so I thought I missed something. But after the second look I see "skipped node resize" should show this, but it's always zero on these reports. So, isn't it possible the current inflate_threshold_root is simply unreachable with some conditions, at least within 10 loops?

Then these settable thresholds might be more useful here than memory fixes, but here is some idea to try handle this automatically within some limits. The patch below increases inflate_threshold_root (only) up to ~50% of its initial value if needed, and should be able to go back sometimes.

Pawel and Jorge, could you try this? (It applies to 2.6.29 too, with some offsets.)

Thanks,
Jarek P.
---

 net/ipv4/fib_trie.c |   23 ++++++++++++++++-------
 1 files changed, 16 insertions(+), 7 deletions(-)
net/ipv4/fib_trie.c +16 -7
index 012cf5a..1dc1bb4 100644
@@ -318,6 +318,7 @@ static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
 static const int halve_threshold_root = 8;
 static const int inflate_threshold_root = 15;
+static int inflate_threshold_root_fix;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -602,7 +603,8 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	/* Keep root node larger  */
 
 	if (!tn->parent)
-		inflate_threshold_use = inflate_threshold_root;
+		inflate_threshold_use = inflate_threshold_root +
+					inflate_threshold_root_fix;
 	else
 		inflate_threshold_use = inflate_threshold;
 
@@ -626,15 +628,22 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	}
 
 	if (max_resize < 0) {
-		if (!tn->parent)
-			pr_warning("Fix inflate_threshold_root."
-				   " Now=%d size=%d bits\n",
-				   inflate_threshold_root, tn->bits);
-		else
+		if (!tn->parent) {
+			if (inflate_threshold_root_fix * 2 <
+			    inflate_threshold_root)
+				inflate_threshold_root_fix++;
+			else
+				pr_warning("Fix inflate_threshold_root."
+					   " Now=%d size=%d bits fix=%d\n",
+					   inflate_threshold_root, tn->bits,
+					   inflate_threshold_root_fix);
+		} else {
 			pr_warning("Fix inflate_threshold."
 				   " Now=%d size=%d bits\n",
 				   inflate_threshold, tn->bits);
-	}
+		}
+	} else if (max_resize < 5 && !tn->parent && inflate_threshold_root_fix)
+		inflate_threshold_root_fix--;
 
 	check_tnode(tn);
 
#31 lore raw

On Sat, Jun 27, 2009 at 09:20:57PM +0200, Jarek Poplawski wrote:
...

Quoted 8 lines

Then these settable thresholds might be more useful here than memory fixes, but here is some idea to try handle this automatically within some limits. The patch below increases inflate_threshold_root (only) up to ~50% of its initial value if needed, and should be able to go back sometimes.

Pawel and Jorge, could you try this? (It applies to 2.6.29 too, with some offsets.)

A tiny adjustment in the last if...

Jarek P.
--- (take 2)

 net/ipv4/fib_trie.c |   23 ++++++++++++++++-------
 1 files changed, 16 insertions(+), 7 deletions(-)
net/ipv4/fib_trie.c +16 -7
index 012cf5a..1dc1bb4 100644
@@ -318,6 +318,7 @@ static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
 static const int halve_threshold_root = 8;
 static const int inflate_threshold_root = 15;
+static int inflate_threshold_root_fix;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -602,7 +603,8 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	/* Keep root node larger  */
 
 	if (!tn->parent)
-		inflate_threshold_use = inflate_threshold_root;
+		inflate_threshold_use = inflate_threshold_root +
+					inflate_threshold_root_fix;
 	else
 		inflate_threshold_use = inflate_threshold;
 
@@ -626,15 +628,22 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	}
 
 	if (max_resize < 0) {
-		if (!tn->parent)
-			pr_warning("Fix inflate_threshold_root."
-				   " Now=%d size=%d bits\n",
-				   inflate_threshold_root, tn->bits);
-		else
+		if (!tn->parent) {
+			if (inflate_threshold_root_fix * 2 <
+			    inflate_threshold_root)
+				inflate_threshold_root_fix++;
+			else
+				pr_warning("Fix inflate_threshold_root."
+					   " Now=%d size=%d bits fix=%d\n",
+					   inflate_threshold_root, tn->bits,
+					   inflate_threshold_root_fix);
+		} else {
 			pr_warning("Fix inflate_threshold."
 				   " Now=%d size=%d bits\n",
 				   inflate_threshold, tn->bits);
-	}
+		}
+	} else if (max_resize > 4 && !tn->parent && inflate_threshold_root_fix)
+		inflate_threshold_root_fix--;
 
 	check_tnode(tn);
 
#32 lore raw

Thanks Jarek

I apply this patch to 2.6.29.5
For some results we must wait to "rush hours" when there will be more traffic / routes :)

Regards
Paweł Staszewski

Jarek Poplawski pisze:

Quoted 79 lines · Jarek Poplawski

...

Quoted 9 lines

Then these settable thresholds might be more useful here than memory fixes, but here is some idea to try handle this automatically within some limits. The patch below increases inflate_threshold_root (only) up to ~50% of its initial value if needed, and should be able to go back sometimes.

Pawel and Jorge, could you try this? (It applies to 2.6.29 too, with some offsets.)

A tiny adjustment in the last if...

Jarek P.
--- (take 2)

 net/ipv4/fib_trie.c |   23 ++++++++++++++++-------
 1 files changed, 16 insertions(+), 7 deletions(-)
net/ipv4/fib_trie.c +16 -8
index 012cf5a..1dc1bb4 100644
@@ -318,6 +318,7 @@ static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
 static const int halve_threshold_root = 8;
 static const int inflate_threshold_root = 15;
+static int inflate_threshold_root_fix;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -602,7 +603,8 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	/* Keep root node larger  */
 
 	if (!tn->parent)
-		inflate_threshold_use = inflate_threshold_root;
+		inflate_threshold_use = inflate_threshold_root +
+					inflate_threshold_root_fix;
 	else
 		inflate_threshold_use = inflate_threshold;
 
@@ -626,15 +628,22 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	}
 
 	if (max_resize < 0) {
-		if (!tn->parent)
-			pr_warning("Fix inflate_threshold_root."
-				   " Now=%d size=%d bits\n",
-				   inflate_threshold_root, tn->bits);
-		else
+		if (!tn->parent) {
+			if (inflate_threshold_root_fix * 2 <
+			    inflate_threshold_root)
+				inflate_threshold_root_fix++;
+			else
+				pr_warning("Fix inflate_threshold_root."
+					   " Now=%d size=%d bits fix=%d\n",
+					   inflate_threshold_root, tn->bits,
+					   inflate_threshold_root_fix);
+		} else {
 			pr_warning("Fix inflate_threshold."
 				   " Now=%d size=%d bits\n",
 				   inflate_threshold, tn->bits);
-	}
+		}
+	} else if (max_resize > 4 && !tn->parent && inflate_threshold_root_fix)
+		inflate_threshold_root_fix--;
 
 	check_tnode(tn);
 
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#33 lore raw
In reply to Jarek Poplawski

When testing please monitor size of root node and and aver depth

Cheers

				--ro
Quoted 74 lines · Jarek Poplawski

On Sat, Jun 27, 2009 at 09:20:57PM +0200, Jarek Poplawski wrote:
...

Quoted 8 lines

Then these settable thresholds might be more useful here than memory fixes, but here is some idea to try handle this automatically within some limits. The patch below increases inflate_threshold_root (only) up to ~50% of its initial value if needed, and should be able to go back sometimes.

Pawel and Jorge, could you try this? (It applies to 2.6.29 too, with some offsets.)

A tiny adjustment in the last if...

Jarek P.
--- (take 2)

 net/ipv4/fib_trie.c |   23 ++++++++++++++++-------
 1 files changed, 16 insertions(+), 7 deletions(-)
net/ipv4/fib_trie.c +16 -8
index 012cf5a..1dc1bb4 100644
@@ -318,6 +318,7 @@ static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
 static const int halve_threshold_root = 8;
 static const int inflate_threshold_root = 15;
+static int inflate_threshold_root_fix;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -602,7 +603,8 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	/* Keep root node larger  */
 
 	if (!tn->parent)
-		inflate_threshold_use = inflate_threshold_root;
+		inflate_threshold_use = inflate_threshold_root +
+					inflate_threshold_root_fix;
 	else
 		inflate_threshold_use = inflate_threshold;
 
@@ -626,15 +628,22 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	}
 
 	if (max_resize < 0) {
-		if (!tn->parent)
-			pr_warning("Fix inflate_threshold_root."
-				   " Now=%d size=%d bits\n",
-				   inflate_threshold_root, tn->bits);
-		else
+		if (!tn->parent) {
+			if (inflate_threshold_root_fix * 2 <
+			    inflate_threshold_root)
+				inflate_threshold_root_fix++;
+			else
+				pr_warning("Fix inflate_threshold_root."
+					   " Now=%d size=%d bits fix=%d\n",
+					   inflate_threshold_root, tn->bits,
+					   inflate_threshold_root_fix);
+		} else {
 			pr_warning("Fix inflate_threshold."
 				   " Now=%d size=%d bits\n",
 				   inflate_threshold, tn->bits);
-	}
+		}
+	} else if (max_resize > 4 && !tn->parent && inflate_threshold_root_fix)
+		inflate_threshold_root_fix--;
 
 	check_tnode(tn);
 
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#34 lore raw

Robert Olsson pisze:

Quoted 5 lines

When testing please monitor size of root node and and aver depth

Cheers

				--ro

Some fib_triestats - kernel.2.6.29.5 with first Jarek patch.

Dump every 10sec:

Mon Jun 29 11:54:31 2009
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.28
       Max depth:      7
       Leaves:         276978
       Prefixes:       290448
       Internal nodes: 66813
         1: 34703  2: 13944  3: 9921  4: 4807  5: 2273  6: 1158  7: 5  
9: 1  18: 1
       Pointers: 691606

Null ptrs: 347816
Total size: 18403 kB

Counters:
---------
gets = 390981859
backtracks = 5332465 semantic match passed = 390452936 semantic match miss = 30198 null node hit= 375522207 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 391017445
backtracks = 121012874 semantic match passed = 37565 semantic match miss = 0
null node hit= 261583 skipped node resize = 0

Mon Jun 29 11:54:41 2009
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.28
       Max depth:      7
       Leaves:         276976
       Prefixes:       290446
       Internal nodes: 66813
         1: 34703  2: 13944  3: 9921  4: 4807  5: 2273  6: 1158  7: 5  
9: 1  18: 1
       Pointers: 691606

Null ptrs: 347818
Total size: 18403 kB

Counters:
---------
gets = 391061852
backtracks = 5334173 semantic match passed = 390532664 semantic match miss = 30199 null node hit= 375595706 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 391097445
backtracks = 121039213 semantic match passed = 37570 semantic match miss = 0
null node hit= 261589 skipped node resize = 0

Mon Jun 29 11:54:51 2009
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.28
       Max depth:      7
       Leaves:         276978
       Prefixes:       290448
       Internal nodes: 66813
         1: 34703  2: 13944  3: 9921  4: 4807  5: 2273  6: 1158  7: 5  
9: 1  18: 1
       Pointers: 691606

Null ptrs: 347816
Total size: 18403 kB

Counters:
---------
gets = 391177325
backtracks = 5336127 semantic match passed = 390647917 semantic match miss = 30208 null node hit= 375699713 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 391212932
backtracks = 121075919 semantic match passed = 37586 semantic match miss = 0
null node hit= 261701 skipped node resize = 0

Mon Jun 29 11:55:01 2009
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.28
       Max depth:      7
       Leaves:         276978
       Prefixes:       290448
       Internal nodes: 66813
         1: 34703  2: 13944  3: 9921  4: 4807  5: 2273  6: 1158  7: 5  
9: 1  18: 1
       Pointers: 691606

Null ptrs: 347816
Total size: 18403 kB

Counters:
---------
gets = 391254016
backtracks = 5337816 semantic match passed = 390724361 semantic match miss = 30214 null node hit= 375768712 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 391289631
backtracks = 121101285 semantic match passed = 37598 semantic match miss = 0
null node hit= 261707 skipped node resize = 0

Mon Jun 29 11:55:11 2009
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.28
       Max depth:      7
       Leaves:         276976
       Prefixes:       290445
       Internal nodes: 66812
         1: 34702  2: 13944  3: 9921  4: 4807  5: 2273  6: 1158  7: 5  
9: 1  18: 1
       Pointers: 691604

Null ptrs: 347817
Total size: 18402 kB

Counters:
---------
gets = 391317389
backtracks = 5339175 semantic match passed = 390787523 semantic match miss = 30215 null node hit= 375827612 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 391353001
backtracks = 121122087 semantic match passed = 37599 semantic match miss = 0
null node hit= 261709 skipped node resize = 0

Mon Jun 29 11:55:21 2009
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.28
       Max depth:      7
       Leaves:         276981
       Prefixes:       290451
       Internal nodes: 66813
         1: 34704  2: 13943  3: 9921  4: 4807  5: 2273  6: 1158  7: 5  
9: 1  18: 1
       Pointers: 691604

Null ptrs: 347811
Total size: 18403 kB

Counters:
---------
gets = 391434307
backtracks = 5340855 semantic match passed = 390904256 semantic match miss = 30225 null node hit= 375931780 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 391469942
backtracks = 121157220 semantic match passed = 37619 semantic match miss = 0
null node hit= 261753 skipped node resize = 0

Mon Jun 29 11:55:31 2009
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.28
       Max depth:      7
       Leaves:         276981
       Prefixes:       290451
       Internal nodes: 66813
         1: 34704  2: 13943  3: 9921  4: 4807  5: 2273  6: 1158  7: 5  
9: 1  18: 1
       Pointers: 691604

Null ptrs: 347811
Total size: 18403 kB

Counters:
---------
gets = 391519852
backtracks = 5342208 semantic match passed = 390989658 semantic match miss = 30234 null node hit= 376010537 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 391555492
backtracks = 121181992 semantic match passed = 37625 semantic match miss = 0
null node hit= 261762 skipped node resize = 0

Mon Jun 29 11:55:42 2009
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.28
       Max depth:      7
       Leaves:         276978
       Prefixes:       290447
       Internal nodes: 66812
         1: 34703  2: 13943  3: 9921  4: 4807  5: 2273  6: 1158  7: 5  
9: 1  18: 1
       Pointers: 691602

Null ptrs: 347813
Total size: 18403 kB

Counters:
---------
gets = 391589032
backtracks = 5343757 semantic match passed = 391058601 semantic match miss = 30237 null node hit= 376075389 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 391624673
backtracks = 121202115 semantic match passed = 37628 semantic match miss = 0
null node hit= 261763 skipped node resize = 0

Mon Jun 29 11:55:52 2009
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.28
       Max depth:      7
       Leaves:         276985
       Prefixes:       290455
       Internal nodes: 66814
         1: 34704  2: 13944  3: 9921  4: 4807  5: 2273  6: 1158  7: 5  
9: 1  18: 1
       Pointers: 691608

Null ptrs: 347810
Total size: 18403 kB

Counters:
---------
gets = 391723925
backtracks = 5345934 semantic match passed = 391193292 semantic match miss = 30242 null node hit= 376195655 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 391759578
backtracks = 121241080 semantic match passed = 37640 semantic match miss = 0
null node hit= 261804 skipped node resize = 0

Mon Jun 29 11:56:02 2009
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.28
       Max depth:      7
       Leaves:         276985
       Prefixes:       290455
       Internal nodes: 66814
         1: 34704  2: 13944  3: 9921  4: 4807  5: 2273  6: 1158  7: 5  
9: 1  18: 1
       Pointers: 691608

Null ptrs: 347810
Total size: 18403 kB

Counters:
---------
gets = 391811219
backtracks = 5347635 semantic match passed = 391280357 semantic match miss = 30250 null node hit= 376276182 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 391846880
backtracks = 121265316 semantic match passed = 37648 semantic match miss = 0
null node hit= 261813 skipped node resize = 0

Quoted 80 lines
Quoted 74 lines · Jarek Poplawski

On Sat, Jun 27, 2009 at 09:20:57PM +0200, Jarek Poplawski wrote:
...

Quoted 8 lines

Then these settable thresholds might be more useful here than memory fixes, but here is some idea to try handle this automatically within some limits. The patch below increases inflate_threshold_root (only) up to ~50% of its initial value if needed, and should be able to go back sometimes.

Pawel and Jorge, could you try this? (It applies to 2.6.29 too, with some offsets.)

A tiny adjustment in the last if...

Jarek P.
--- (take 2)

 net/ipv4/fib_trie.c |   23 ++++++++++++++++-------
 1 files changed, 16 insertions(+), 7 deletions(-)
net/ipv4/fib_trie.c +16 -8
index 012cf5a..1dc1bb4 100644
@@ -318,6 +318,7 @@ static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
 static const int halve_threshold_root = 8;
 static const int inflate_threshold_root = 15;
+static int inflate_threshold_root_fix;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -602,7 +603,8 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	/* Keep root node larger  */
 
 	if (!tn->parent)
-		inflate_threshold_use = inflate_threshold_root;
+		inflate_threshold_use = inflate_threshold_root +
+					inflate_threshold_root_fix;
 	else
 		inflate_threshold_use = inflate_threshold;
 
@@ -626,15 +628,22 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	}
 
 	if (max_resize < 0) {
-		if (!tn->parent)
-			pr_warning("Fix inflate_threshold_root."
-				   " Now=%d size=%d bits\n",
-				   inflate_threshold_root, tn->bits);
-		else
+		if (!tn->parent) {
+			if (inflate_threshold_root_fix * 2 <
+			    inflate_threshold_root)
+				inflate_threshold_root_fix++;
+			else
+				pr_warning("Fix inflate_threshold_root."
+					   " Now=%d size=%d bits fix=%d\n",
+					   inflate_threshold_root, tn->bits,
+					   inflate_threshold_root_fix);
+		} else {
 			pr_warning("Fix inflate_threshold."
 				   " Now=%d size=%d bits\n",
 				   inflate_threshold, tn->bits);
-	}
+		}
+	} else if (max_resize > 4 && !tn->parent && inflate_threshold_root_fix)
+		inflate_threshold_root_fix--;
 
 	check_tnode(tn);
 
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#35 lore raw
In reply to Jarek Poplawski
Quoted 11 lines · Jarek Poplawski

Robert, you and Eric pointed at memory problems, so I thought I missed something. But after the second look I see "skipped node resize" should show this, but it's always zero on these reports. So, isn't it possible the current inflate_threshold_root is simply unreachable with some conditions, at least within 10 loops?

Then these settable thresholds might be more useful here than memory fixes, but here is some idea to try handle this automatically within some limits. The patch below increases inflate_threshold_root (only) up to ~50% of its initial value if needed, and should be able to go back sometimes.

Yes we keep the old tnode size and the convergence interval was some of the concerns. That why this checks was added. Still we want to inflate the root node to a very max.

So this approach with halving or doubling tnodes towards the root node was the suggest by Dyntree paper. I asked Stefan (one of the authors) if we could get safe and very offensive settings. But from what I understood there was no easy way to calculate this. So any bright ideas in this area are very welcome. But we should also monitor size of root and average tree depth so we don't take an to defensive approach just to solve this case.

The memory patches and "manual RCU" are also interesting to address the case with PREEMTP's.

Inserts and deletes are also very fast due to the flat tree so I think we can "slow down" this a bit if need to be safe with all PREEMPT's.

Thanks for giving this area energy.

Cheers

					--ro
Quoted 61 lines

Pawel and Jorge, could you try this? (It applies to 2.6.29 too, with some offsets.)

Thanks,
Jarek P.
---

 net/ipv4/fib_trie.c |   23 ++++++++++++++++-------
 1 files changed, 16 insertions(+), 7 deletions(-)
net/ipv4/fib_trie.c +16 -7
index 012cf5a..1dc1bb4 100644
@@ -318,6 +318,7 @@ static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
 static const int halve_threshold_root = 8;
 static const int inflate_threshold_root = 15;
+static int inflate_threshold_root_fix;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -602,7 +603,8 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	/* Keep root node larger  */
 
 	if (!tn->parent)
-		inflate_threshold_use = inflate_threshold_root;
+		inflate_threshold_use = inflate_threshold_root +
+					inflate_threshold_root_fix;
 	else
 		inflate_threshold_use = inflate_threshold;
 
@@ -626,15 +628,22 @@ static struct node *resize(struct trie *t, struct tnode *tn)
 	}
 
 	if (max_resize < 0) {
-		if (!tn->parent)
-			pr_warning("Fix inflate_threshold_root."
-				   " Now=%d size=%d bits\n",
-				   inflate_threshold_root, tn->bits);
-		else
+		if (!tn->parent) {
+			if (inflate_threshold_root_fix * 2 <
+			    inflate_threshold_root)
+				inflate_threshold_root_fix++;
+			else
+				pr_warning("Fix inflate_threshold_root."
+					   " Now=%d size=%d bits fix=%d\n",
+					   inflate_threshold_root, tn->bits,
+					   inflate_threshold_root_fix);
+		} else {
 			pr_warning("Fix inflate_threshold."
 				   " Now=%d size=%d bits\n",
 				   inflate_threshold, tn->bits);
-	}
+		}
+	} else if (max_resize < 5 && !tn->parent && inflate_threshold_root_fix)
+		inflate_threshold_root_fix--;
 
 	check_tnode(tn);
 
#36 lore raw

On Sun, Jun 28, 2009 at 01:04:51PM +0200, Robert Olsson wrote:
...

Quoted 11 lines

Yes we keep the old tnode size and the convergence interval was some of the concerns. That why this checks was added. Still we want to inflate the root node to a very max.

So this approach with halving or doubling tnodes towards the root node was the suggest by Dyntree paper. I asked Stefan (one of the authors) if we could get safe and very offensive settings. But from what I understood there was no easy way to calculate this. So any bright ideas in this area are very welcome. But we should also monitor size of root and average tree depth so we don't take an to defensive approach just to solve this case.

Yes, but with this offensive approach it seems the current level of warnings could be too alarming. Btw., because of a design flaw in my current patch this _fix variable, which should be logically per trie/ table, can be reset by changes of other tables now, but I think it all could be fine tuned more in the future. Of course if there are people interested in testing/reporting this more.

Thanks,
Jarek P.

#37 lore raw
In reply to Robert Olsson

On Sun, Jun 28, 2009 at 01:04:51PM +0200, Robert Olsson wrote:
...

Quoted 2 lines

The memory patches and "manual RCU" are also interesting to address the case with PREEMTP's.

Since 2.6.29 looks like prefered here, and there were a lot of takes in this thread, I attach below a combined all-in-one patch including:
- 2.6.29 -> 2.6.30 preemption disable patch
- 2 RCU vs. preemption fixes from 2.6.31-rc
- "manual RCU" patch to force vfree/kfree before root's resize (take 3)
- "automatic" inflate_threshold_root fix (take 2)

Thanks,
Jarek P.

--- (for 2.6.29.x or even .28 or .27; any testing appreciated)

net/ipv4/fib_trie.c +74 -20
diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
@@ -123,6 +123,7 @@ struct tnode {
 	union {
 		struct rcu_head rcu;
 		struct work_struct work;
+		struct tnode *tnode_free;
 	};
 	struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -315,6 +318,7 @@ static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
 static const int halve_threshold_root = 8;
 static const int inflate_threshold_root = 15;
+static int inflate_threshold_root_fix;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -363,6 +367,17 @@ static void __tnode_vfree(struct work_st
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	size_t size = sizeof(struct tnode) +
+		      (sizeof(struct node *) << tn->bits);
+
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -385,6 +400,24 @@ static inline void tnode_free(struct tno
 		call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+	BUG_ON(IS_LEAF(tn));
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+	struct tnode *tn;
+
+	while ((tn = tnode_free_head)) {
+		tnode_free_head = tn->tnode_free;
+		tn->tnode_free = NULL;
+		__tnode_free(tn);
+	}
+}
+
 static struct leaf *leaf_new(void)
 {
 	struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +528,7 @@ static struct node *resize(struct trie *
 
 	/* No children */
 	if (tn->empty_children == tnode_child_length(tn)) {
-		tnode_free(tn);
+		tnode_free_safe(tn);
 		return NULL;
 	}
 	/* One child */
@@ -509,7 +542,7 @@ static struct node *resize(struct trie *
 
 			/* compress one level */
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 	/*
@@ -581,7 +614,8 @@ static struct node *resize(struct trie *
 	/* Keep root node larger  */
 
 	if (!tn->parent)
-		inflate_threshold_use = inflate_threshold_root;
+		inflate_threshold_use = inflate_threshold_root +
+					inflate_threshold_root_fix;
 	else
 		inflate_threshold_use = inflate_threshold;
 
@@ -605,15 +639,22 @@ static struct node *resize(struct trie *
 	}
 
 	if (max_resize < 0) {
-		if (!tn->parent)
-			pr_warning("Fix inflate_threshold_root."
-				   " Now=%d size=%d bits\n",
-				   inflate_threshold_root, tn->bits);
-		else
+		if (!tn->parent) {
+			if (inflate_threshold_root_fix * 2 <
+			    inflate_threshold_root)
+				inflate_threshold_root_fix++;
+			else
+				pr_warning("Fix inflate_threshold_root."
+					   " Now=%d size=%d bits fix=%d\n",
+					   inflate_threshold_root, tn->bits,
+					   inflate_threshold_root_fix);
+		} else {
 			pr_warning("Fix inflate_threshold."
 				   " Now=%d size=%d bits\n",
 				   inflate_threshold, tn->bits);
-	}
+		}
+	} else if (max_resize > 4 && !tn->parent && inflate_threshold_root_fix)
+		inflate_threshold_root_fix--;
 
 	check_tnode(tn);
 
@@ -670,7 +711,7 @@ static struct node *resize(struct trie *
 			/* compress one level */
 
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 
@@ -756,7 +797,7 @@ static struct tnode *inflate(struct trie
 			put_child(t, tn, 2*i, inode->child[0]);
 			put_child(t, tn, 2*i+1, inode->child[1]);
 
-			tnode_free(inode);
+			tnode_free_safe(inode);
 			continue;
 		}
 
@@ -801,9 +842,9 @@ static struct tnode *inflate(struct trie
 		put_child(t, tn, 2*i, resize(t, left));
 		put_child(t, tn, 2*i+1, resize(t, right));
 
-		tnode_free(inode);
+		tnode_free_safe(inode);
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -885,7 +926,7 @@ static struct tnode *halve(struct trie *
 		put_child(t, newBinNode, 1, right);
 		put_child(t, tn, i/2, resize(t, newBinNode));
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -983,12 +1024,14 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
-	t_key cindex, key = tn->key;
+	t_key cindex, key;
 	struct tnode *tp;
 
+	key = tn->key;
+
 	while (tn != NULL && (tp = node_parent((struct node *)tn)) != NULL) {
 		cindex = tkey_extract_bits(key, tp->pos, tp->bits);
 		wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
@@ -1003,11 +1046,22 @@ static struct node *trie_rebalance(struc
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
-	return (struct node *)tn;
+	return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1209,7 @@ static struct list_head *fib_insert_node
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp);
 done:
 	return fa_head;
 }
@@ -1575,7 +1629,7 @@ static void trie_leaf_remove(struct trie
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 
#38 lore raw

After 18 hours from apply first Jarek patch i have no info about Fix inflate_threshold_root even if i make: "clear ip bgp *" on router
So i change Jarek patch from previous to this new one for test and we will see ...

Regards
Pawel Staszewski

Jarek Poplawski pisze:

Quoted 253 lines · Robert Olsson

...

Quoted 3 lines

The memory patches and "manual RCU" are also interesting to address the case with PREEMTP's.

Since 2.6.29 looks like prefered here, and there were a lot of takes in this thread, I attach below a combined all-in-one patch including:
- 2.6.29 -> 2.6.30 preemption disable patch
- 2 RCU vs. preemption fixes from 2.6.31-rc
- "manual RCU" patch to force vfree/kfree before root's resize (take 3)
- "automatic" inflate_threshold_root fix (take 2)

Thanks,
Jarek P.

--- (for 2.6.29.x or even .28 or .27; any testing appreciated)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c +74 -21
@@ -123,6 +123,7 @@ struct tnode {
 	union {
 		struct rcu_head rcu;
 		struct work_struct work;
+		struct tnode *tnode_free;
 	};
 	struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -315,6 +318,7 @@ static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
 static const int halve_threshold_root = 8;
 static const int inflate_threshold_root = 15;
+static int inflate_threshold_root_fix;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -363,6 +367,17 @@ static void __tnode_vfree(struct work_st
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	size_t size = sizeof(struct tnode) +
+		      (sizeof(struct node *) << tn->bits);
+
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -385,6 +400,24 @@ static inline void tnode_free(struct tno
 		call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+	BUG_ON(IS_LEAF(tn));
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+	struct tnode *tn;
+
+	while ((tn = tnode_free_head)) {
+		tnode_free_head = tn->tnode_free;
+		tn->tnode_free = NULL;
+		__tnode_free(tn);
+	}
+}
+
 static struct leaf *leaf_new(void)
 {
 	struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +528,7 @@ static struct node *resize(struct trie *
 
 	/* No children */
 	if (tn->empty_children == tnode_child_length(tn)) {
-		tnode_free(tn);
+		tnode_free_safe(tn);
 		return NULL;
 	}
 	/* One child */
@@ -509,7 +542,7 @@ static struct node *resize(struct trie *
 
 			/* compress one level */
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 	/*
@@ -581,7 +614,8 @@ static struct node *resize(struct trie *
 	/* Keep root node larger  */
 
 	if (!tn->parent)
-		inflate_threshold_use = inflate_threshold_root;
+		inflate_threshold_use = inflate_threshold_root +
+					inflate_threshold_root_fix;
 	else
 		inflate_threshold_use = inflate_threshold;
 
@@ -605,15 +639,22 @@ static struct node *resize(struct trie *
 	}
 
 	if (max_resize < 0) {
-		if (!tn->parent)
-			pr_warning("Fix inflate_threshold_root."
-				   " Now=%d size=%d bits\n",
-				   inflate_threshold_root, tn->bits);
-		else
+		if (!tn->parent) {
+			if (inflate_threshold_root_fix * 2 <
+			    inflate_threshold_root)
+				inflate_threshold_root_fix++;
+			else
+				pr_warning("Fix inflate_threshold_root."
+					   " Now=%d size=%d bits fix=%d\n",
+					   inflate_threshold_root, tn->bits,
+					   inflate_threshold_root_fix);
+		} else {
 			pr_warning("Fix inflate_threshold."
 				   " Now=%d size=%d bits\n",
 				   inflate_threshold, tn->bits);
-	}
+		}
+	} else if (max_resize > 4 && !tn->parent && inflate_threshold_root_fix)
+		inflate_threshold_root_fix--;
 
 	check_tnode(tn);
 
@@ -670,7 +711,7 @@ static struct node *resize(struct trie *
 			/* compress one level */
 
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 
@@ -756,7 +797,7 @@ static struct tnode *inflate(struct trie
 			put_child(t, tn, 2*i, inode->child[0]);
 			put_child(t, tn, 2*i+1, inode->child[1]);
 
-			tnode_free(inode);
+			tnode_free_safe(inode);
 			continue;
 		}
 
@@ -801,9 +842,9 @@ static struct tnode *inflate(struct trie
 		put_child(t, tn, 2*i, resize(t, left));
 		put_child(t, tn, 2*i+1, resize(t, right));
 
-		tnode_free(inode);
+		tnode_free_safe(inode);
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -885,7 +926,7 @@ static struct tnode *halve(struct trie *
 		put_child(t, newBinNode, 1, right);
 		put_child(t, tn, i/2, resize(t, newBinNode));
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -983,12 +1024,14 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
-	t_key cindex, key = tn->key;
+	t_key cindex, key;
 	struct tnode *tp;
 
+	key = tn->key;
+
 	while (tn != NULL && (tp = node_parent((struct node *)tn)) != NULL) {
 		cindex = tkey_extract_bits(key, tp->pos, tp->bits);
 		wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
@@ -1003,11 +1046,22 @@ static struct node *trie_rebalance(struc
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
-	return (struct node *)tn;
+	return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1209,7 @@ static struct list_head *fib_insert_node
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp);
 done:
 	return fa_head;
 }
@@ -1575,7 +1629,7 @@ static void trie_leaf_remove(struct trie
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#39 lore raw

After apply this patch something is wrong

Traffic is not forwarded
no info in dmesg / no info from bgp
and also i can't connect to bgpd process

I revert kernel to past version with first Jarek patch

Paweł Staszewski pisze:

Quoted 273 lines

After 18 hours from apply first Jarek patch i have no info about Fix inflate_threshold_root even if i make: "clear ip bgp *" on router
So i change Jarek patch from previous to this new one for test and we will see ...

Regards
Pawel Staszewski

Jarek Poplawski pisze:

Quoted 255 lines · Robert Olsson

...

Quoted 3 lines

The memory patches and "manual RCU" are also interesting to address the case with PREEMTP's.

Since 2.6.29 looks like prefered here, and there were a lot of takes in this thread, I attach below a combined all-in-one patch including:
- 2.6.29 -> 2.6.30 preemption disable patch
- 2 RCU vs. preemption fixes from 2.6.31-rc
- "manual RCU" patch to force vfree/kfree before root's resize (take 3)
- "automatic" inflate_threshold_root fix (take 2)

Thanks,
Jarek P.

--- (for 2.6.29.x or even .28 or .27; any testing appreciated)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c 2009-06-28 15:45:02.000000000 +0200 renamed from net/ipv4/fib_trie.c 2009-06-27 20:25:06.000000000 +0200 +59 -16
@@ -123,6 +123,7 @@ struct tnode {
     union {
         struct rcu_head rcu;
         struct work_struct work;
+        struct tnode *tnode_free;
     };
     struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -315,6 +318,7 @@ static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
 static const int halve_threshold_root = 8;
 static const int inflate_threshold_root = 15;
+static int inflate_threshold_root_fix;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -363,6 +367,17 @@ static void __tnode_vfree(struct work_st
     vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+    size_t size = sizeof(struct tnode) +
+              (sizeof(struct node *) << tn->bits);
+
+    if (size <= PAGE_SIZE)
+        kfree(tn);
+    else
+        vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
     struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -385,6 +400,24 @@ static inline void tnode_free(struct tno
         call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+    BUG_ON(IS_LEAF(tn));
+    tn->tnode_free = tnode_free_head;
+    tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+    struct tnode *tn;
+
+    while ((tn = tnode_free_head)) {
+        tnode_free_head = tn->tnode_free;
+        tn->tnode_free = NULL;
+        __tnode_free(tn);
+    }
+}
+
 static struct leaf *leaf_new(void)
 {
     struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +528,7 @@ static struct node *resize(struct trie *
 
     /* No children */
     if (tn->empty_children == tnode_child_length(tn)) {
-        tnode_free(tn);
+        tnode_free_safe(tn);
         return NULL;
     }
     /* One child */
@@ -509,7 +542,7 @@ static struct node *resize(struct trie *
 
             /* compress one level */
             node_set_parent(n, NULL);
-            tnode_free(tn);
+            tnode_free_safe(tn);
             return n;
         }
     /*
@@ -581,7 +614,8 @@ static struct node *resize(struct trie *
     /* Keep root node larger  */
 
     if (!tn->parent)
-        inflate_threshold_use = inflate_threshold_root;
+        inflate_threshold_use = inflate_threshold_root +
+                    inflate_threshold_root_fix;
     else
         inflate_threshold_use = inflate_threshold;
 
@@ -605,15 +639,22 @@ static struct node *resize(struct trie *
     }
 
     if (max_resize < 0) {
-        if (!tn->parent)
-            pr_warning("Fix inflate_threshold_root."
-                   " Now=%d size=%d bits\n",
-                   inflate_threshold_root, tn->bits);
-        else
+        if (!tn->parent) {
+            if (inflate_threshold_root_fix * 2 <
+                inflate_threshold_root)
+                inflate_threshold_root_fix++;
+            else
+                pr_warning("Fix inflate_threshold_root."
+                       " Now=%d size=%d bits fix=%d\n",
+                       inflate_threshold_root, tn->bits,
+                       inflate_threshold_root_fix);
+        } else {
             pr_warning("Fix inflate_threshold."
                    " Now=%d size=%d bits\n",
                    inflate_threshold, tn->bits);
-    }
+        }
+    } else if (max_resize > 4 && !tn->parent && 
 inflate_threshold_root_fix)
+        inflate_threshold_root_fix--;
 
     check_tnode(tn);
@@ -670,7 +711,7 @@ static struct node *resize(struct trie *
             /* compress one level */
 
             node_set_parent(n, NULL);
-            tnode_free(tn);
+            tnode_free_safe(tn);
             return n;
         }
 
@@ -756,7 +797,7 @@ static struct tnode *inflate(struct trie
             put_child(t, tn, 2*i, inode->child[0]);
             put_child(t, tn, 2*i+1, inode->child[1]);
 
-            tnode_free(inode);
+            tnode_free_safe(inode);
             continue;
         }
 
@@ -801,9 +842,9 @@ static struct tnode *inflate(struct trie
         put_child(t, tn, 2*i, resize(t, left));
         put_child(t, tn, 2*i+1, resize(t, right));
 
-        tnode_free(inode);
+        tnode_free_safe(inode);
     }
-    tnode_free(oldtnode);
+    tnode_free_safe(oldtnode);
     return tn;
 nomem:
     {
@@ -885,7 +926,7 @@ static struct tnode *halve(struct trie *
         put_child(t, newBinNode, 1, right);
         put_child(t, tn, i/2, resize(t, newBinNode));
     }
-    tnode_free(oldtnode);
+    tnode_free_safe(oldtnode);
     return tn;
 nomem:
     {
@@ -983,12 +1024,14 @@ fib_find_node(struct trie *t, u32 key)
     return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
     int wasfull;
-    t_key cindex, key = tn->key;
+    t_key cindex, key;
     struct tnode *tp;
 
+    key = tn->key;
+
     while (tn != NULL && (tp = node_parent((struct node *)tn)) != 
 NULL) {
         cindex = tkey_extract_bits(key, tp->pos, tp->bits);
         wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
net/ipv4/fib_trie.c 2009-06-28 15:45:02.000000000 +0200 renamed from net/ipv4/fib_trie.c 2009-06-27 20:25:06.000000000 +0200 +15 -4
@@ -1003,11 +1046,22 @@ static struct node *trie_rebalance(struc
         tn = tp;
     }
 
+    if (tnode_free_head) {
+        synchronize_rcu();
+        tnode_free_flush();
+    }
+
     /* Handle last (top) tnode */
-    if (IS_TNODE(tn))
+    if (IS_TNODE(tn)) {
         tn = (struct tnode *)resize(t, (struct tnode *)tn);
+        rcu_assign_pointer(t->trie, (struct node *)tn);
+        synchronize_rcu();
+        tnode_free_flush();
+    } else {
+        rcu_assign_pointer(t->trie, (struct node *)tn);
+    }
 
-    return (struct node *)tn;
+    return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1209,7 @@ static struct list_head *fib_insert_node
 
     /* Rebalance the trie */
 
-    rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+    trie_rebalance(t, tp);
 done:
     return fa_head;
 }
@@ -1575,7 +1629,7 @@ static void trie_leaf_remove(struct trie
     if (tp) {
         t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
         put_child(t, (struct tnode *)tp, cindex, NULL);
-        rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+        trie_rebalance(t, tp);
     } else
         rcu_assign_pointer(t->trie, NULL);
 
Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#40 lore raw
Quoted 9 lines · Paweł Staszewski

After apply this patch something is wrong

Traffic is not forwarded
no info in dmesg / no info from bgp
and also i can't connect to bgpd process

I revert kernel to past version with first Jarek patch

Thank you very much, Pawel, for trying this. I'm starting to look for the reason. In the meantime try to get some fib_trie stats for Robert.

Jarek P.

Quoted 38 lines

Paweł Staszewski pisze:

Quoted 35 lines

After 18 hours from apply first Jarek patch i have no info about Fix inflate_threshold_root
even if i make: "clear ip bgp *" on router
So i change Jarek patch from previous to this new one for test and we will see ...

Regards
Pawel Staszewski

Jarek Poplawski pisze:

Quoted 22 lines · Robert Olsson

...

Quoted 3 lines

The memory patches and "manual RCU" are also interesting to address the case with PREEMTP's.

Since 2.6.29 looks like prefered here, and there were a lot of takes in this thread, I attach below a combined all-in-one patch including:
- 2.6.29 -> 2.6.30 preemption disable patch
- 2 RCU vs. preemption fixes from 2.6.31-rc
- "manual RCU" patch to force vfree/kfree before root's resize (take 3)
- "automatic" inflate_threshold_root fix (take 2)

Thanks,
Jarek P.

--- (for 2.6.29.x or even .28 or .27; any testing appreciated)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c 2009-06-28 15:45:02.000000000 +0200 renamed from net/ipv4/fib_trie.c 2009-06-27 20:25:06.000000000 +0200 +0 -0

...

#41 lore raw
In reply to Paweł Staszewski

To David Miller:
since among patches tested negatively by Pawel are current 2 fixes from 2.6.31-rc, I hope they weren't sent to -stable yet. Otherwise, please withdraw them until they are tested alone. Thanks.

To Pawel:

Quoted 9 lines · Paweł Staszewski

After apply this patch something is wrong

Traffic is not forwarded
no info in dmesg / no info from bgp
and also i can't connect to bgpd process

I revert kernel to past version with first Jarek patch

Since checking this can take time I attach here a patch with only changes which are currently in 2.6.31-rc. Of course, this part can be broken as well, so it's up to you: if you could try it with caution somewhere it would be very helpful; otherwise don't bother.

It could be applied to 2.6.29 with or without this currently working patch.

Thanks,
Jarek P.
--- (for 2.6.29.x, .28 or .27)

net/ipv4/fib_trie.c +39 -12
diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
@@ -123,6 +123,7 @@ struct tnode {
 	union {
 		struct rcu_head rcu;
 		struct work_struct work;
+		struct tnode *tnode_free;
 	};
 	struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -385,6 +388,24 @@ static inline void tnode_free(struct tno
 		call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+	BUG_ON(IS_LEAF(tn));
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+	struct tnode *tn;
+
+	while ((tn = tnode_free_head)) {
+		tnode_free_head = tn->tnode_free;
+		tn->tnode_free = NULL;
+		tnode_free(tn);
+	}
+}
+
 static struct leaf *leaf_new(void)
 {
 	struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +516,7 @@ static struct node *resize(struct trie *
 
 	/* No children */
 	if (tn->empty_children == tnode_child_length(tn)) {
-		tnode_free(tn);
+		tnode_free_safe(tn);
 		return NULL;
 	}
 	/* One child */
@@ -509,7 +530,7 @@ static struct node *resize(struct trie *
 
 			/* compress one level */
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 	/*
@@ -670,7 +691,7 @@ static struct node *resize(struct trie *
 			/* compress one level */
 
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 
@@ -756,7 +777,7 @@ static struct tnode *inflate(struct trie
 			put_child(t, tn, 2*i, inode->child[0]);
 			put_child(t, tn, 2*i+1, inode->child[1]);
 
-			tnode_free(inode);
+			tnode_free_safe(inode);
 			continue;
 		}
 
@@ -801,9 +822,9 @@ static struct tnode *inflate(struct trie
 		put_child(t, tn, 2*i, resize(t, left));
 		put_child(t, tn, 2*i+1, resize(t, right));
 
-		tnode_free(inode);
+		tnode_free_safe(inode);
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -885,7 +906,7 @@ static struct tnode *halve(struct trie *
 		put_child(t, newBinNode, 1, right);
 		put_child(t, tn, i/2, resize(t, newBinNode));
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -983,12 +1004,14 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
-	t_key cindex, key = tn->key;
+	t_key cindex, key;
 	struct tnode *tp;
 
+	key = tn->key;
+
 	while (tn != NULL && (tp = node_parent((struct node *)tn)) != NULL) {
 		cindex = tkey_extract_bits(key, tp->pos, tp->bits);
 		wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
@@ -998,6 +1021,7 @@ static struct node *trie_rebalance(struc
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
+		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
@@ -1007,7 +1031,10 @@ static struct node *trie_rebalance(struc
 	if (IS_TNODE(tn))
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
 
-	return (struct node *)tn;
+	rcu_assign_pointer(t->trie, (struct node *)tn);
+	tnode_free_flush();
+
+	return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1182,7 @@ static struct list_head *fib_insert_node
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp);
 done:
 	return fa_head;
 }
@@ -1575,7 +1602,7 @@ static void trie_leaf_remove(struct trie
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 
#42 lore raw

Jarek Poplawski pisze:

Quoted 18 lines

To David Miller:
since among patches tested negatively by Pawel are current 2 fixes from 2.6.31-rc, I hope they weren't sent to -stable yet. Otherwise, please withdraw them until they are tested alone. Thanks.

To Pawel:

Quoted 9 lines · Paweł Staszewski

After apply this patch something is wrong

Traffic is not forwarded
no info in dmesg / no info from bgp
and also i can't connect to bgpd process

I revert kernel to past version with first Jarek patch

Quoted 9 lines

Since checking this can take time I attach here a patch with only changes which are currently in 2.6.31-rc. Of course, this part can be broken as well, so it's up to you: if you could try it with caution somewhere it would be very helpful; otherwise don't bother.

It could be applied to 2.6.29 with or without this currently working patch.

Ok.
I applied this patch 15mins ago to 2.6.29.5 and now it's working - traffic is forwarded.

Some fib_triestats
cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.29
       Max depth:      6
       Leaves:         277015
       Prefixes:       290493
       Internal nodes: 67115
         1: 35733  2: 13635  3: 9544  4: 4832  5: 2239  6: 1125  7: 5  
9: 1  18: 1
       Pointers: 686614

Null ptrs: 342485
Total size: 18396 kB

Counters:
---------
gets = 3956301
backtracks = 192497 semantic match passed = 3895955 semantic match miss = 133 null node hit= 4306948 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 3960981
backtracks = 2152441 semantic match passed = 4757 semantic match miss = 0 null node hit= 194997 skipped node resize = 0

Quoted 165 lines

Thanks,
Jarek P.
--- (for 2.6.29.x, .28 or .27)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c +39 -12
@@ -123,6 +123,7 @@ struct tnode {
 	union {
 		struct rcu_head rcu;
 		struct work_struct work;
+		struct tnode *tnode_free;
 	};
 	struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -385,6 +388,24 @@ static inline void tnode_free(struct tno
 		call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+	BUG_ON(IS_LEAF(tn));
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+	struct tnode *tn;
+
+	while ((tn = tnode_free_head)) {
+		tnode_free_head = tn->tnode_free;
+		tn->tnode_free = NULL;
+		tnode_free(tn);
+	}
+}
+
 static struct leaf *leaf_new(void)
 {
 	struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +516,7 @@ static struct node *resize(struct trie *
 
 	/* No children */
 	if (tn->empty_children == tnode_child_length(tn)) {
-		tnode_free(tn);
+		tnode_free_safe(tn);
 		return NULL;
 	}
 	/* One child */
@@ -509,7 +530,7 @@ static struct node *resize(struct trie *
 
 			/* compress one level */
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 	/*
@@ -670,7 +691,7 @@ static struct node *resize(struct trie *
 			/* compress one level */
 
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 
@@ -756,7 +777,7 @@ static struct tnode *inflate(struct trie
 			put_child(t, tn, 2*i, inode->child[0]);
 			put_child(t, tn, 2*i+1, inode->child[1]);
 
-			tnode_free(inode);
+			tnode_free_safe(inode);
 			continue;
 		}
 
@@ -801,9 +822,9 @@ static struct tnode *inflate(struct trie
 		put_child(t, tn, 2*i, resize(t, left));
 		put_child(t, tn, 2*i+1, resize(t, right));
 
-		tnode_free(inode);
+		tnode_free_safe(inode);
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -885,7 +906,7 @@ static struct tnode *halve(struct trie *
 		put_child(t, newBinNode, 1, right);
 		put_child(t, tn, i/2, resize(t, newBinNode));
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -983,12 +1004,14 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
-	t_key cindex, key = tn->key;
+	t_key cindex, key;
 	struct tnode *tp;
 
+	key = tn->key;
+
 	while (tn != NULL && (tp = node_parent((struct node *)tn)) != NULL) {
 		cindex = tkey_extract_bits(key, tp->pos, tp->bits);
 		wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
@@ -998,6 +1021,7 @@ static struct node *trie_rebalance(struc
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
+		tnode_free_flush();
 		if (!tp)
 			break;
 		tn = tp;
@@ -1007,7 +1031,10 @@ static struct node *trie_rebalance(struc
 	if (IS_TNODE(tn))
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
 
-	return (struct node *)tn;
+	rcu_assign_pointer(t->trie, (struct node *)tn);
+	tnode_free_flush();
+
+	return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1182,7 @@ static struct list_head *fib_insert_node
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp);
 done:
 	return fa_head;
 }
@@ -1575,7 +1602,7 @@ static void trie_leaf_remove(struct trie
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 
#43 lore raw

But
With all this patches i have the same problem with CPU load
Every time when route cache entries are purged cpu load is increasing from 1% to 40 / 80% it depends

I see that on 64bit machine when route cache entries are going down i have almost 80% load on each cpu where ethernet card is binded by smp_affinity But on 32bit machine cpu load reported by mpstat is half that on 64bit machine
here is example from 32bit machine ( mpstat + rtstat -k entries )

Linux 2.6.29.5 (TM_02_C1) 06/29/09 _i686_ (2 CPU)

12:36:54     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle        RT CACHE ENTRIES (from rtstat)
12:36:57     all    0.00    0.00    0.00    0.00    1.51   15.08    0.00    0.00   83.42        83346
12:36:58     all    0.00    0.00    0.00    0.00    1.01    7.58    0.00    0.00   91.41        85988
12:36:59     all    0.00    0.00    0.00    0.00    0.50    1.01    0.00    0.00   98.49        89979
12:37:00     all    0.00    0.00    0.50    0.00    0.00    1.51    0.00    0.00   97.99        93652
12:37:01     all    0.00    0.00    0.00    0.00    0.00    2.01    0.00    0.00   97.99        96533
12:37:02     all    0.00    0.00    0.00    0.00    0.51    1.01    0.00    0.00   98.48        99451
12:37:03     all    0.00    0.00    0.00    0.00    0.00    2.49    0.00    0.00   97.51        102018
12:37:04     all    0.00    0.00    0.00    0.00    0.00    1.52    0.00    0.00   98.48        104153
12:37:05     all    0.00    0.00    0.00    0.00    0.00    1.01    0.00    0.00   98.99        105979
12:37:06     all    0.00    0.00    0.00    0.00    0.00    1.01    0.00    0.00   98.99        107684
12:37:07     all    0.00    0.00    0.00    0.00    0.00    1.53    0.00    0.00   98.47        109070
12:37:08     all    0.00    0.00    0.00    0.00    0.00    1.51    0.00    0.00   98.49        110462
12:37:09     all    0.00    0.00    0.00    0.00    0.00    1.52    0.00    0.00   98.48        112301
12:37:10     all    0.00    0.00    0.00    0.00    2.00   20.00    0.00    0.00   78.00        111535
12:37:11     all    0.00    0.00    0.00    0.00    2.49   34.33    0.00    0.00   63.18        108659
12:37:12     all    0.00    0.00    0.00    0.00    3.03   28.28    0.00    0.00   68.69        105534
12:37:13     all    0.00    0.00    0.00    0.00    3.98   30.85    0.00    0.00   65.17        103341
12:37:14     all    0.00    0.00    0.00    0.00    4.50   30.50    0.00    0.00   65.00        101307
12:37:15     all    5.56    0.00    0.00    0.00    1.52   28.79    0.00    0.00   64.14        97435
12:37:16     all   11.39    0.00    0.50    0.00    4.95   30.69    0.00    0.00   52.48        93908
12:37:17     all    1.51    0.00    0.00    0.00    1.01   27.64    0.00    0.00   69.85        90229
12:37:18     all    0.00    0.00    0.00    0.00    2.99   27.36    0.00    0.00   69.65        87030
12:37:19     all    0.00    0.00    0.00    0.00    3.02   29.65    0.00    0.00   67.34        84324
12:37:20     all    0.00    0.00    0.00    0.00    2.99   30.35    0.00    0.00   66.67        82167
12:37:21     all    0.00    0.00    0.00    0.00    1.98   31.68    0.00    0.00   66.34        80121
12:37:22     all    0.00    0.00    0.00    0.00    1.51   30.65    0.00    0.00   67.84        77850
12:37:23     all    0.00    0.00    0.00    0.00    2.50   28.50    0.00    0.00   69.00        76005
12:37:24     all    0.00    0.00    0.00    0.00    1.98   23.27    0.00    0.00   74.75        74538
12:37:25     all    0.00    0.00    0.49    0.00    2.93   22.44    0.00    0.00   74.15        76923
12:37:26     all    0.00    0.00    0.00    0.00    1.51   15.58    0.00    0.00   82.91        79396
12:37:27     all    0.00    0.00    0.00    0.00    0.50    7.96    0.00    0.00   91.54        81835
12:37:28     all    0.00    0.00    0.00    0.00    0.50    3.52    0.00    0.00   95.98        84169
12:37:29     all    0.00    0.00    0.00    0.00    0.00    2.02    0.00    0.00   97.98        87740
12:37:30     all    0.00    0.00    0.00    0.00    0.51    1.52    0.00    0.00   97.98        91152
12:37:31     all    0.00    0.00    0.00    0.00    0.00    1.99    0.00    0.00   98.01        94102
12:37:32     all    0.00    0.00    0.00    0.00    0.00    1.52    0.00    0.00   98.48        97032
12:37:33     all    0.00    0.00    0.00    0.00    0.00    0.50    0.00    0.00   99.50        99685
12:37:34     all    0.00    0.00    0.00    0.00    0.00    1.00    0.00    0.00   99.00        101970
12:37:35     all    0.00    0.00    0.00    0.00    0.50    1.00    0.00    0.00   98.50        103814
12:37:36     all    0.00    0.00    0.00    0.00    0.00    1.52    0.00    0.00   98.48        104793
12:37:37     all    0.00    0.00    0.00    0.00    0.00    1.01    0.00    0.00   98.99        106214
12:37:38     all    0.00    0.00    0.00    0.00    0.50    1.01    0.00    0.00   98.49        107300
12:37:39     all    0.00    0.00    0.00    0.00    0.00   13.00    0.00    0.00   87.00        111951
12:37:40     all    0.00    0.00    0.00    0.00    2.50   29.50    0.00    0.00   68.00        111215
12:37:41     all    0.00    0.00    0.00    0.00    2.01   30.65    0.00    0.00   67.34        108023
12:37:42     all    0.00    0.00    0.00    0.00    2.99   29.85    0.00    0.00   67.16        104751
12:37:43     all    0.00    0.00    0.00    0.00    2.00   31.00    0.00    0.00   67.00        100827
12:37:44     all    0.00    0.00    0.00    0.00    3.00   27.00    0.00    0.00   70.00        97184
12:37:45     all    0.00    0.00    0.00    0.00    2.50   29.00    0.00    0.00   68.50        93904
12:37:46     all    0.00    0.00    0.00    0.00    3.02   30.15    0.00    0.00   66.83        90979
12:37:47     all    0.00    0.00    0.00    0.00    2.49   27.86    0.00    0.00   69.65        88315
12:37:48     all    0.00    0.00    0.00    0.00    2.48   31.19    0.00    0.00   66.34        87777
12:37:49     all    0.00    0.00    0.00    0.00    2.94   32.35    0.00    0.00   64.71        89218
12:37:50     all    0.00    0.00    0.00    0.00    3.00   32.50    0.00    0.00   64.50        85896
12:37:51     all    0.00    0.00    0.00    0.00    2.50   30.00    0.00    0.00   67.50        82712
12:37:52     all    0.50    0.00    0.00    0.00    2.49   30.85    0.00    0.00   66.17        79137
12:37:53     all    0.00    0.00    0.50    0.00    2.00   28.50    0.00    0.00   69.00        75644
12:37:54     all    0.00    0.00    0.00    0.00    2.51   30.65    0.00    0.00   66.83        72843
12:37:55     all    0.00    0.00    0.50    0.00    3.48   28.36    0.00    0.00   67.66        73460

Paweł Staszewski pisze:

Quoted 251 lines

Jarek Poplawski pisze:

Quoted 18 lines

To David Miller:
since among patches tested negatively by Pawel are current 2 fixes from 2.6.31-rc, I hope they weren't sent to -stable yet. Otherwise, please withdraw them until they are tested alone. Thanks.

To Pawel:

Quoted 9 lines · Paweł Staszewski

After apply this patch something is wrong

Traffic is not forwarded
no info in dmesg / no info from bgp
and also i can't connect to bgpd process

I revert kernel to past version with first Jarek patch

Quoted 9 lines

Since checking this can take time I attach here a patch with only changes which are currently in 2.6.31-rc. Of course, this part can be broken as well, so it's up to you: if you could try it with caution somewhere it would be very helpful; otherwise don't bother.

It could be applied to 2.6.29 with or without this currently working patch.

Ok.
I applied this patch 15mins ago to 2.6.29.5 and now it's working - traffic is forwarded.

Some fib_triestats
cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.29
       Max depth:      6
       Leaves:         277015
       Prefixes:       290493
       Internal nodes: 67115
         1: 35733  2: 13635  3: 9544  4: 4832  5: 2239  6: 1125  7: 5  
9: 1  18: 1
       Pointers: 686614

Null ptrs: 342485
Total size: 18396 kB

Counters:
---------
gets = 3956301
backtracks = 192497
semantic match passed = 3895955
semantic match miss = 133
null node hit= 4306948
skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 3960981
backtracks = 2152441
semantic match passed = 4757
semantic match miss = 0
null node hit= 194997
skipped node resize = 0

Quoted 166 lines

Thanks,
Jarek P.
--- (for 2.6.29.x, .28 or .27)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c 2009-06-28 23:06:02.000000000 +0200 renamed from net/ipv4/fib_trie.c 2009-06-27 20:25:06.000000000 +0200 +32 -9
@@ -123,6 +123,7 @@ struct tnode {
     union {
         struct rcu_head rcu;
         struct work_struct work;
+        struct tnode *tnode_free;
     };
     struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -385,6 +388,24 @@ static inline void tnode_free(struct tno
         call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+    BUG_ON(IS_LEAF(tn));
+    tn->tnode_free = tnode_free_head;
+    tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+    struct tnode *tn;
+
+    while ((tn = tnode_free_head)) {
+        tnode_free_head = tn->tnode_free;
+        tn->tnode_free = NULL;
+        tnode_free(tn);
+    }
+}
+
 static struct leaf *leaf_new(void)
 {
     struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +516,7 @@ static struct node *resize(struct trie *
 
     /* No children */
     if (tn->empty_children == tnode_child_length(tn)) {
-        tnode_free(tn);
+        tnode_free_safe(tn);
         return NULL;
     }
     /* One child */
@@ -509,7 +530,7 @@ static struct node *resize(struct trie *
 
             /* compress one level */
             node_set_parent(n, NULL);
-            tnode_free(tn);
+            tnode_free_safe(tn);
             return n;
         }
     /*
@@ -670,7 +691,7 @@ static struct node *resize(struct trie *
             /* compress one level */
 
             node_set_parent(n, NULL);
-            tnode_free(tn);
+            tnode_free_safe(tn);
             return n;
         }
 
@@ -756,7 +777,7 @@ static struct tnode *inflate(struct trie
             put_child(t, tn, 2*i, inode->child[0]);
             put_child(t, tn, 2*i+1, inode->child[1]);
 
-            tnode_free(inode);
+            tnode_free_safe(inode);
             continue;
         }
 
@@ -801,9 +822,9 @@ static struct tnode *inflate(struct trie
         put_child(t, tn, 2*i, resize(t, left));
         put_child(t, tn, 2*i+1, resize(t, right));
 
-        tnode_free(inode);
+        tnode_free_safe(inode);
     }
-    tnode_free(oldtnode);
+    tnode_free_safe(oldtnode);
     return tn;
 nomem:
     {
@@ -885,7 +906,7 @@ static struct tnode *halve(struct trie *
         put_child(t, newBinNode, 1, right);
         put_child(t, tn, i/2, resize(t, newBinNode));
     }
-    tnode_free(oldtnode);
+    tnode_free_safe(oldtnode);
     return tn;
 nomem:
     {
@@ -983,12 +1004,14 @@ fib_find_node(struct trie *t, u32 key)
     return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
     int wasfull;
-    t_key cindex, key = tn->key;
+    t_key cindex, key;
     struct tnode *tp;
 
+    key = tn->key;
+
     while (tn != NULL && (tp = node_parent((struct node *)tn)) != 
 NULL) {
         cindex = tkey_extract_bits(key, tp->pos, tp->bits);
         wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
net/ipv4/fib_trie.c 2009-06-28 23:06:02.000000000 +0200 renamed from net/ipv4/fib_trie.c 2009-06-27 20:25:06.000000000 +0200 +7 -3
@@ -998,6 +1021,7 @@ static struct node *trie_rebalance(struc
                       (struct node *)tn, wasfull);
 
         tp = node_parent((struct node *) tn);
+        tnode_free_flush();
         if (!tp)
             break;
         tn = tp;
@@ -1007,7 +1031,10 @@ static struct node *trie_rebalance(struc
     if (IS_TNODE(tn))
         tn = (struct tnode *)resize(t, (struct tnode *)tn);
 
-    return (struct node *)tn;
+    rcu_assign_pointer(t->trie, (struct node *)tn);
+    tnode_free_flush();
+
+    return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1182,7 @@ static struct list_head *fib_insert_node
 
     /* Rebalance the trie */
 
-    rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+    trie_rebalance(t, tp);
 done:
     return fa_head;
 }
@@ -1575,7 +1602,7 @@ static void trie_leaf_remove(struct trie
     if (tp) {
         t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
         put_child(t, (struct tnode *)tp, cindex, NULL);
-        rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+        trie_rebalance(t, tp);
     } else
         rcu_assign_pointer(t->trie, NULL);
 
Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#44 lore raw
Quoted 12 lines · Paweł Staszewski

But
With all this patches i have the same problem with CPU load Every time when route cache entries are purged cpu load is increasing from 1% to 40 / 80% it depends

I see that on 64bit machine when route cache entries are going down i have almost 80% load on each cpu where ethernet card is binded by smp_affinity
But on 32bit machine cpu load reported by mpstat is half that on 64bit machine
here is example from 32bit machine ( mpstat + rtstat -k entries )

Quoted 6 lines

Linux 2.6.29.5 (TM_02_C1) 06/29/09 _i686_ (2 CPU)

12:36:54 CPU %usr %nice %sys %iowait %irq %soft %steal
%guest %idle RT CACHE ENTRIES (from rtstat)
12:36:57 all 0.00 0.00 0.00 0.00 1.51 15.08 0.00
0.00 83.42 83346

I guess Eric is thinking about this. Btw., two little suggestions:
it should be easier to track if these route cache reports stay in its starting thread ("weird problem"?), and if you could send these stats/logs as attachements or turn off line wrapping, please? ;-)

Thanks,
Jarek P.

#45 lore raw

Jarek Poplawski pisze:

Quoted 22 lines · Paweł Staszewski
Quoted 9 lines

But
With all this patches i have the same problem with CPU load
Every time when route cache entries are purged cpu load is increasing from 1% to 40 / 80% it depends

I see that on 64bit machine when route cache entries are going down i have almost 80% load on each cpu where ethernet card is binded by smp_affinity But on 32bit machine cpu load reported by mpstat is half that on 64bit machine
here is example from 32bit machine ( mpstat + rtstat -k entries )

Quoted 6 lines

Linux 2.6.29.5 (TM_02_C1) 06/29/09 _i686_ (2 CPU)

12:36:54     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle        RT CACHE ENTRIES (from rtstat)
12:36:57     all    0.00    0.00    0.00    0.00    1.51   15.08    0.00  

0.00 83.42 83346

I guess Eric is thinking about this. Btw., two little suggestions:
it should be easier to track if these route cache reports stay in its starting thread ("weird problem"?), and if you could send these stats/logs as attachements or turn off line wrapping, please? ;-)

Quoted 9 lines

Thanks,
Jarek P.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Sorry Jarek for combining problems :)
And yes i will apply next stats in attachements :)

#46 lore raw
In reply to Jarek Poplawski
Quoted 4 lines · Jarek Poplawski

To David Miller:
since among patches tested negatively by Pawel are current 2 fixes from 2.6.31-rc, I hope they weren't sent to -stable yet. Otherwise, please withdraw them until they are tested alone. Thanks.

David, IMHO this fix is needed in net-2.6 even if it doesn't fix the problem reported by Pawel (there could be still something more).

Pawel, I see you decided to test my previous patch, but try to add this one on top.

Thanks,
Jarek P.
------------------->
ipv4: Fix fib_trie rebalancing, part 3

Alas current delaying of freeing old tnodes by RCU in trie_rebalance is still not enough because we can free a top tnode before updating a t->trie pointer.

---

 net/ipv4/fib_trie.c |    3 +++
 1 files changed, 3 insertions(+), 0 deletions(-)
net/ipv4/fib_trie.c +3 -0
index 012cf5a..00a54b2 100644
@@ -1021,6 +1021,9 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
+		if (!tp)
+			rcu_assign_pointer(t->trie, (struct node *)tn);
+
 		tnode_free_flush();
 		if (!tp)
 			break;
#47 lore raw

I apply this patch

fib_triestats in attached file :)

Jarek Poplawski pisze:

Quoted 45 lines · Jarek Poplawski
Quoted 6 lines

To David Miller:
since among patches tested negatively by Pawel are current 2 fixes from 2.6.31-rc, I hope they weren't sent to -stable yet. Otherwise, please withdraw them until they are tested alone. Thanks.

David, IMHO this fix is needed in net-2.6 even if it doesn't fix the problem reported by Pawel (there could be still something more).

Pawel, I see you decided to test my previous patch, but try to add this one on top.

Thanks,
Jarek P.
------------------->
ipv4: Fix fib_trie rebalancing, part 3

Alas current delaying of freeing old tnodes by RCU in trie_rebalance is still not enough because we can free a top tnode before updating a t->trie pointer.

---

 net/ipv4/fib_trie.c |    3 +++
 1 files changed, 3 insertions(+), 0 deletions(-)
net/ipv4/fib_trie.c +3 -1
index 012cf5a..00a54b2 100644
@@ -1021,6 +1021,9 @@ static void trie_rebalance(struct trie *t, struct tnode *tn)
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
+		if (!tp)
+			rcu_assign_pointer(t->trie, (struct node *)tn);
+
 		tnode_free_flush();
 		if (!tp)
 			break;
--
Quoted 6 lines

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

        Aver depth:     2.29
        Max depth:      7
        Leaves:         276909
        Prefixes:       290383
        Internal nodes: 66893
          1: 34715  2: 14024  3: 9889  4: 4833  5: 2275  6: 1150  7: 5  9: 1  18: 1
        Pointers: 691662

Null ptrs: 347861
Total size: 18403 kB

Counters:
---------
gets = 2297579
backtracks = 131491 semantic match passed = 2233070 semantic match miss = 42 null node hit= 2016883 skipped node resize = 0

Local:

        Aver depth:     3.75
        Max depth:      5
        Leaves:         12
        Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 2302102
backtracks = 1545197 semantic match passed = 4536 semantic match miss = 0 null node hit= 192664 skipped node resize = 0

----------------------------------------------------------------

cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

        Aver depth:     2.29
        Max depth:      7
        Leaves:         276904
        Prefixes:       290378
        Internal nodes: 66889
          1: 34711  2: 14023  3: 9890  4: 4833  5: 2275  6: 1150  7: 5  9: 1  18: 1
        Pointers: 691658

Null ptrs: 347866
Total size: 18402 kB

Counters:
---------
gets = 3006945
backtracks = 138787 semantic match passed = 2942047 semantic match miss = 85 null node hit= 2826377 skipped node resize = 0

Local:

        Aver depth:     3.75
        Max depth:      5
        Leaves:         12
        Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 3011504
backtracks = 1796587 semantic match passed = 4577 semantic match miss = 0 null node hit= 192747 skipped node resize = 0

--------------------------------------------------------------

cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

        Aver depth:     2.29
        Max depth:      7
        Leaves:         276904
        Prefixes:       290378
        Internal nodes: 66891
          1: 34710  2: 14025  3: 9892  4: 4832  5: 2275  6: 1150  7: 5  9: 1  18: 1
        Pointers: 691664

Null ptrs: 347870
Total size: 18402 kB

Counters:
---------
gets = 3320633
backtracks = 141904 semantic match passed = 3255585 semantic match miss = 99 null node hit= 3177543 skipped node resize = 0

Local:

        Aver depth:     3.75
        Max depth:      5
        Leaves:         12
        Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 3325226
backtracks = 1904022 semantic match passed = 4601 semantic match miss = 0 null node hit= 192782 skipped node resize = 0

#48 lore raw
Quoted 3 lines · Paweł Staszewski

I apply this patch

fib_triestats in attached file :)

Great! But it would be nice to check if this (accidentally ;-) might fix the previous problem, so I attach below the patch with "manual RCU", which btw. (or even more important) should verify RCU use here.

It should be applied on top of this last "Fix..., part3". And
again: it's quite probable it can fail, so with caution, no hurry (it can wait for quiet time)...

Many thanks,
Jarek P.
--------------------> (synchronize_rcu take 4)

net/ipv4/fib_trie.c +26 -9
diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
@@ -366,6 +366,17 @@ static void __tnode_vfree(struct work_st
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	size_t size = sizeof(struct tnode) +
+		      (sizeof(struct node *) << tn->bits);
+
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +413,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1021,21 +1032,27 @@ static void trie_rebalance(struct trie *
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
-		if (!tp)
+		if (!tp) {
 			rcu_assign_pointer(t->trie, (struct node *)tn);
-
-		tnode_free_flush();
-		if (!tp)
 			break;
+		}
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }
#49 lore raw

Jarek Poplawski pisze:

Quoted 12 lines · Paweł Staszewski
Quoted 5 lines

I apply this patch

fib_triestats in attached file :)

Great! But it would be nice to check if this (accidentally ;-) might fix the previous problem, so I attach below the patch with "manual RCU", which btw. (or even more important) should verify RCU use here.

After this patches all is OK now i don't see Fix inflate_threshold_root. Even if i make "clear ip bgp * "

Before this patches when i make clear ip bgp there was always info in dmesg about "Fix inflate_threshold_root"

Quoted 5 lines

It should be applied on top of this last "Fix..., part3". And
again: it's quite probable it can fail, so with caution, no hurry (it can wait for quiet time)...

After apply this last patch - traffic is not forwarded again :)
i was fast and have only some fib_triestats in attached file before failover switch routers.

This stats are from machine with this last patch that makes kernel to stop forwarding

Quoted 71 lines

Many thanks,
Jarek P.
--------------------> (synchronize_rcu take 4)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c +26 -10
@@ -366,6 +366,17 @@ static void __tnode_vfree(struct work_st
 	vfree(tn);
 }
 
+static void __tnode_free(struct tnode *tn)
+{
+	size_t size = sizeof(struct tnode) +
+		      (sizeof(struct node *) << tn->bits);
+
+	if (size <= PAGE_SIZE)
+		kfree(tn);
+	else
+		vfree(tn);
+}
+
 static void __tnode_free_rcu(struct rcu_head *head)
 {
 	struct tnode *tn = container_of(head, struct tnode, rcu);
@@ -402,7 +413,7 @@ static void tnode_free_flush(void)
 	while ((tn = tnode_free_head)) {
 		tnode_free_head = tn->tnode_free;
 		tn->tnode_free = NULL;
-		tnode_free(tn);
+		__tnode_free(tn);
 	}
 }
 
@@ -1021,21 +1032,27 @@ static void trie_rebalance(struct trie *
 				      (struct node *)tn, wasfull);
 
 		tp = node_parent((struct node *) tn);
-		if (!tp)
+		if (!tp) {
 			rcu_assign_pointer(t->trie, (struct node *)tn);
-
-		tnode_free_flush();
-		if (!tp)
 			break;
+		}
 		tn = tp;
 	}
 
+	if (tnode_free_head) {
+		synchronize_rcu();
+		tnode_free_flush();
+	}
+
 	/* Handle last (top) tnode */
-	if (IS_TNODE(tn))
+	if (IS_TNODE(tn)) {
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
-
-	rcu_assign_pointer(t->trie, (struct node *)tn);
-	tnode_free_flush();
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+		synchronize_rcu();
+		tnode_free_flush();
+	} else {
+		rcu_assign_pointer(t->trie, (struct node *)tn);
+	}
 
 	return;
 }
--
Quoted 6 lines

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

        Aver depth:     3.51
        Max depth:      8
        Leaves:         3089
        Prefixes:       3156
        Internal nodes: 1167
          1: 737  2: 202  3: 150  4: 40  5: 28  6: 8  7: 1  10: 1
        Pointers: 6682

Null ptrs: 2427
Total size: 214 kB

Counters:
---------
gets = 1554240
backtracks = 916511
semantic match passed = 1127691
semantic match miss = 27
null node hit= 1439140
skipped node resize = 0

Local:

        Aver depth:     3.75
        Max depth:      5
        Leaves:         12
        Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 1554767
backtracks = 572694
semantic match passed = 534
semantic match miss = 0
null node hit= 288
skipped node resize = 0

#50 lore raw
Quoted 25 lines · Paweł Staszewski

Jarek Poplawski pisze:

Quoted 12 lines · Paweł Staszewski
Quoted 5 lines

I apply this patch

fib_triestats in attached file :)

Great! But it would be nice to check if this (accidentally ;-) might fix the previous problem, so I attach below the patch with "manual RCU", which btw. (or even more important) should verify RCU use here.

After this patches all is OK now i don't see Fix inflate_threshold_root. Even if i make "clear ip bgp * "

Before this patches when i make clear ip bgp there was always info in dmesg about "Fix inflate_threshold_root"

Quoted 5 lines

It should be applied on top of this last "Fix..., part3". And
again: it's quite probable it can fail, so with caution, no hurry (it can wait for quiet time)...

Quoted 6 lines

After apply this last patch - traffic is not forwarded again :) i was fast and have only some fib_triestats in attached file before failover switch routers.

This stats are from machine with this last patch that makes kernel to stop forwarding

OK, I'll look at it again.

Thanks for testing!
Jarek P.

#51 lore raw
In reply to Jarek Poplawski
Quoted 5 lines · Jarek Poplawski
Quoted 3 lines · Paweł Staszewski

I apply this patch

fib_triestats in attached file :)

Quoted 7 lines

Great! But it would be nice to check if this (accidentally ;-) might fix the previous problem, so I attach below the patch with "manual RCU", which btw. (or even more important) should verify RCU use here.

It should be applied on top of this last "Fix..., part3". And
again: it's quite probable it can fail, so with caution, no hurry (it can wait for quiet time)...

Pawel, here is another try to check what's going on here, so just like before, but this one on top of these 2 last working patches, plus quite time... (Stats aren't necessary; if these are some doubts let me know.)

Thanks,
Jarek P.
--------------------> (synchronize_rcu take 5)

net/ipv4/fib_trie.c +1 -0
diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
@@ -1036,6 +1036,7 @@ static void trie_rebalance(struct trie *
 
 	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();
+	synchronize_rcu();
 
 	return;
 }
#52 lore raw

Jarek Poplawski pisze:

Quoted 31 lines · Jarek Poplawski
Quoted 7 lines · Paweł Staszewski
Quoted 5 lines

I apply this patch

fib_triestats in attached file :)

Quoted 8 lines

Great! But it would be nice to check if this (accidentally ;-) might fix the previous problem, so I attach below the patch with "manual RCU", which btw. (or even more important) should verify RCU use here.

It should be applied on top of this last "Fix..., part3". And
again: it's quite probable it can fail, so with caution, no hurry (it can wait for quiet time)...

Pawel, here is another try to check what's going on here, so just like before, but this one on top of these 2 last working patches, plus quite time... (Stats aren't necessary; if these are some doubts let me know.)

Thanks,
Jarek P.
--------------------> (synchronize_rcu take 5)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c static void trie_rebalance(struct trie * · line 1042

Quoted 7 lines
net/ipv4/fib_trie.c +1 -0
@@ -1036,6 +1036,7 @@ static void trie_rebalance(struct trie *
 
 	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();
+	synchronize_rcu();
 
 	return;
 }

Apply and tested

Traffic is not forwarded after apply this patch.:)

Quoted 7 lines

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#53 lore raw
Quoted 40 lines · Paweł Staszewski

Jarek Poplawski pisze:

Quoted 30 lines · Jarek Poplawski
Quoted 7 lines · Paweł Staszewski
Quoted 5 lines

I apply this patch

fib_triestats in attached file :)

Quoted 8 lines

Great! But it would be nice to check if this (accidentally ;-) might fix the previous problem, so I attach below the patch with "manual RCU", which btw. (or even more important) should verify RCU use here.

It should be applied on top of this last "Fix..., part3". And
again: it's quite probable it can fail, so with caution, no hurry (it can wait for quiet time)...

Pawel, here is another try to check what's going on here, so just like before, but this one on top of these 2 last working patches, plus quite time... (Stats aren't necessary; if these are some doubts let me know.)

Thanks,
Jarek P.
--------------------> (synchronize_rcu take 5)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c static void trie_rebalance(struct trie * · line 1041

Quoted 6 lines
net/ipv4/fib_trie.c +1 -0
@@ -1036,6 +1036,7 @@ static void trie_rebalance(struct trie *
  	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();
+	synchronize_rcu();
  	return;
 }
 

Apply and tested

Quoted 1 line

Traffic is not forwarded after apply this patch.:)

A little comment: these last 2 patches weren't exactly to fix the problem you reported, which should be mostly fixed by the earlier patch.

There is some other bug, which you omit with CONFIG_PREEMPT_NONE (but it's not for sure there is no by effects). So, I'd like to be sure you are willing and can (without too much risk) to do more such tests. Alas I've no way to generate similar conditions so it would simply have to wait for somebody else.

Many thanks again,
Jarek P.

#54 lore raw

Jarek Poplawski pisze:

Quoted 53 lines · Paweł Staszewski
Quoted 45 lines

Jarek Poplawski pisze:

Quoted 33 lines · Jarek Poplawski
Quoted 10 lines · Paweł Staszewski
Quoted 7 lines

I apply this patch

fib_triestats in attached file :)

Quoted 9 lines

Great! But it would be nice to check if this (accidentally ;-) might fix the previous problem, so I attach below the patch with "manual RCU", which btw. (or even more important) should verify RCU use here.

It should be applied on top of this last "Fix..., part3". And
again: it's quite probable it can fail, so with caution, no hurry (it can wait for quiet time)...

Pawel, here is another try to check what's going on here, so just like before, but this one on top of these 2 last working patches, plus quite time... (Stats aren't necessary; if these are some doubts let me know.)

Thanks,
Jarek P.
--------------------> (synchronize_rcu take 5)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c static void trie_rebalance(struct trie * · line 1042

Quoted 7 lines
net/ipv4/fib_trie.c +1 -0
@@ -1036,6 +1036,7 @@ static void trie_rebalance(struct trie *
  	rcu_assign_pointer(t->trie, (struct node *)tn);
 	tnode_free_flush();
+	synchronize_rcu();
  	return;
 }
 
 

Apply and tested

Quoted 2 lines

Traffic is not forwarded after apply this patch.:)

A little comment: these last 2 patches weren't exactly to fix the problem you reported, which should be mostly fixed by the earlier patch.

Quoted 7 lines

There is some other bug, which you omit with CONFIG_PREEMPT_NONE (but it's not for sure there is no by effects). So, I'd like to be sure you are willing and can (without too much risk) to do more such tests. Alas I've no way to generate similar conditions so it would simply have to wait for somebody else.

Yes i can make tests like this.
My network is splited to test clients and other normal clients
so it's really no problem to make testing. - if testing clients working then traffic from normal clients is also switched to this router (but if traffic is not forwarded "like in this case" for testing clients then failover switching them to working router )

and other point to make this tests - is that - it is good to have all in linux kernel networking working well :)

Regards
Paweł Staszewski

Quoted 5 lines

Many thanks again,
Jarek P.

#55 lore raw

On Wed, Jul 01, 2009 at 01:31:09AM +0200, Paweł Staszewski wrote:
...

Quoted 9 lines

Yes i can make tests like this.
My network is splited to test clients and other normal clients so it's really no problem to make testing. - if testing clients working then traffic from normal clients is also switched to this router (but if traffic is not forwarded "like in this case" for testing clients then failover switching them to working router )

and other point to make this tests - is that - it is good to have all in linux kernel networking working well :)

It's extremely nice of you! On the other hand, this type of change was planned to the net-next to fix possible memory problems, which might have happened to you as well. So you'd probably experience this problem in the future (2.6.32) anyway.

So here is the first of 2 patches (the second in a separate message), which should be tested separately, each one applied on top of the 2.6.29.x (vanilla - at least fib_trie.c), after reverting the previous one. So, they are again all-in-one, to eclude any misunderstanding.

Btw., I assume there were no oopses, warnings or lockups after those previous non-working patches - only no routing/forwarding.

Thanks,
Jarek P.
----------> (synchronize take 6 all-in-one for 2.6.29x, .28, or .27)

net/ipv4/fib_trie.c +44 -12
diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
@@ -123,6 +123,7 @@ struct tnode {
 	union {
 		struct rcu_head rcu;
 		struct work_struct work;
+		struct tnode *tnode_free;
 	};
 	struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -385,6 +388,24 @@ static inline void tnode_free(struct tno
 		call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+	BUG_ON(IS_LEAF(tn));
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+	struct tnode *tn;
+
+	while ((tn = tnode_free_head)) {
+		tnode_free_head = tn->tnode_free;
+		tn->tnode_free = NULL;
+		tnode_free(tn);
+	}
+}
+
 static struct leaf *leaf_new(void)
 {
 	struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +516,7 @@ static struct node *resize(struct trie *
 
 	/* No children */
 	if (tn->empty_children == tnode_child_length(tn)) {
-		tnode_free(tn);
+		tnode_free_safe(tn);
 		return NULL;
 	}
 	/* One child */
@@ -509,7 +530,7 @@ static struct node *resize(struct trie *
 
 			/* compress one level */
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 	/*
@@ -670,7 +691,7 @@ static struct node *resize(struct trie *
 			/* compress one level */
 
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 
@@ -756,7 +777,7 @@ static struct tnode *inflate(struct trie
 			put_child(t, tn, 2*i, inode->child[0]);
 			put_child(t, tn, 2*i+1, inode->child[1]);
 
-			tnode_free(inode);
+			tnode_free_safe(inode);
 			continue;
 		}
 
@@ -801,9 +822,9 @@ static struct tnode *inflate(struct trie
 		put_child(t, tn, 2*i, resize(t, left));
 		put_child(t, tn, 2*i+1, resize(t, right));
 
-		tnode_free(inode);
+		tnode_free_safe(inode);
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -885,7 +906,7 @@ static struct tnode *halve(struct trie *
 		put_child(t, newBinNode, 1, right);
 		put_child(t, tn, i/2, resize(t, newBinNode));
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -983,12 +1004,14 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn, bool sync)
 {
 	int wasfull;
-	t_key cindex, key = tn->key;
+	t_key cindex, key;
 	struct tnode *tp;
 
+	key = tn->key;
+
 	while (tn != NULL && (tp = node_parent((struct node *)tn)) != NULL) {
 		cindex = tkey_extract_bits(key, tp->pos, tp->bits);
 		wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
@@ -999,6 +1022,10 @@ static struct node *trie_rebalance(struc
 
 		tp = node_parent((struct node *) tn);
 		if (!tp)
+			rcu_assign_pointer(t->trie, (struct node *)tn);
+
+		//tnode_free_flush();
+		if (!tp)
 			break;
 		tn = tp;
 	}
@@ -1007,7 +1034,12 @@ static struct node *trie_rebalance(struc
 	if (IS_TNODE(tn))
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
 
-	return (struct node *)tn;
+	rcu_assign_pointer(t->trie, (struct node *)tn);
+	if (sync)
+		synchronize_rcu();
+	tnode_free_flush();
+
+	return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1187,7 @@ static struct list_head *fib_insert_node
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp, true);
 done:
 	return fa_head;
 }
@@ -1575,7 +1607,7 @@ static void trie_leaf_remove(struct trie
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp, false);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 

A message that is not in the archive (sent off-list or not yet received): [email protected]

#57 lore raw

Jarek Poplawski pisze:

Quoted 6 lines · Jarek Poplawski
Quoted 4 lines · Paweł Staszewski

...

Quoted 27 lines

It looks like Cc was shortened BTW, but I guess at least Robert is interested in this testing, so I add him back.

Cheers,
Jarek P.

Quoted 20 lines
Quoted 6 lines

Yes i can make tests like this.
My network is splited to test clients and other normal clients
so it's really no problem to make testing. - if testing clients working then traffic from normal clients is also switched to this router (but if traffic is not forwarded "like in this case" for testing clients then failover switching them to working router )

and other point to make this tests - is that - it is good to have all in linux kernel networking working well :)

It's extremely nice of you! On the other hand, this type of change was planned to the net-next to fix possible memory problems, which might have happened to you as well. So you'd probably experience this problem in the future (2.6.32) anyway.

So here is the first of 2 patches (the second in a separate message), which should be tested separately, each one applied on top of the 2.6.29.x (vanilla - at least fib_trie.c), after reverting the previous one. So, they are again all-in-one, to eclude any misunderstanding.

Btw., I assume there were no oopses, warnings or lockups after those previous non-working patches - only no routing/forwarding.

Yes on on previous patches there was / no warnings / no oopses or lockups

But now i apply this patch and i make more testing.
First boot with start of bgpd and - traffic is not forwarded
So i start to search and make only some routes (static without bgpd) thru this host And all is working for this host when i make all by static routes.

So i change a little my bgp configuration and make default route to only one of my iBGP peers and start bgpd process All is working and what is weird is number of routes in kernel table.
Kernel is learning routes from bgpd but very slowly - really very slowly.

In attached file there are some fib_triestats after 5min of traffic.

Without this patch (normally)
total size: reported by fib_triestats in less that 1sec is: "Total size: 35769 kB"

But with this patch
Total size is growing up and in 5 min of traffic it grow to only: "Total size: 1005 kB"

Regards
Paweł Staszewski

Quoted 174 lines
Quoted 159 lines

Thanks,
Jarek P.
----------> (synchronize take 6 all-in-one for 2.6.29x, .28, or .27)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c +43 -11
@@ -123,6 +123,7 @@ struct tnode {
 	union {
 		struct rcu_head rcu;
 		struct work_struct work;
+		struct tnode *tnode_free;
 	};
 	struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -385,6 +388,24 @@ static inline void tnode_free(struct tno
 		call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+	BUG_ON(IS_LEAF(tn));
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+	struct tnode *tn;
+
+	while ((tn = tnode_free_head)) {
+		tnode_free_head = tn->tnode_free;
+		tn->tnode_free = NULL;
+		tnode_free(tn);
+	}
+}
+
 static struct leaf *leaf_new(void)
 {
 	struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +516,7 @@ static struct node *resize(struct trie *
 
 	/* No children */
 	if (tn->empty_children == tnode_child_length(tn)) {
-		tnode_free(tn);
+		tnode_free_safe(tn);
 		return NULL;
 	}
 	/* One child */
@@ -509,7 +530,7 @@ static struct node *resize(struct trie *
 
 			/* compress one level */
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 	/*
@@ -670,7 +691,7 @@ static struct node *resize(struct trie *
 			/* compress one level */
 
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 
@@ -756,7 +777,7 @@ static struct tnode *inflate(struct trie
 			put_child(t, tn, 2*i, inode->child[0]);
 			put_child(t, tn, 2*i+1, inode->child[1]);
 
-			tnode_free(inode);
+			tnode_free_safe(inode);
 			continue;
 		}
 
@@ -801,9 +822,9 @@ static struct tnode *inflate(struct trie
 		put_child(t, tn, 2*i, resize(t, left));
 		put_child(t, tn, 2*i+1, resize(t, right));
 
-		tnode_free(inode);
+		tnode_free_safe(inode);
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -885,7 +906,7 @@ static struct tnode *halve(struct trie *
 		put_child(t, newBinNode, 1, right);
 		put_child(t, tn, i/2, resize(t, newBinNode));
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -983,12 +1004,14 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn, bool sync)
 {
 	int wasfull;
-	t_key cindex, key = tn->key;
+	t_key cindex, key;
 	struct tnode *tp;
 
+	key = tn->key;
+
 	while (tn != NULL && (tp = node_parent((struct node *)tn)) != NULL) {
 		cindex = tkey_extract_bits(key, tp->pos, tp->bits);
 		wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
@@ -999,6 +1022,10 @@ static struct node *trie_rebalance(struc
 
 		tp = node_parent((struct node *) tn);
 		if (!tp)
+			rcu_assign_pointer(t->trie, (struct node *)tn);
+
+		//tnode_free_flush();
+		if (!tp)
 			break;
 		tn = tp;
 	}
@@ -1007,7 +1034,12 @@ static struct node *trie_rebalance(struc
 	if (IS_TNODE(tn))
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
 
-	return (struct node *)tn;
+	rcu_assign_pointer(t->trie, (struct node *)tn);
+	if (sync)
+		synchronize_rcu();
+	tnode_free_flush();
+
+	return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1187,7 @@ static struct list_head *fib_insert_node
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp, true);
 done:
 	return fa_head;
 }

net/ipv4/fib_trie.c static void trie_leaf_remove(struct trie · line 1613

Quoted 8 lines
net/ipv4/fib_trie.c +1 -1
@@ -1575,7 +1607,7 @@ static void trie_leaf_remove(struct trie
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp, false);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#58 lore raw

Paweł Staszewski pisze:

Quoted 45 lines

Jarek Poplawski pisze:

Quoted 6 lines · Jarek Poplawski
Quoted 4 lines · Paweł Staszewski

...

Quoted 27 lines

It looks like Cc was shortened BTW, but I guess at least Robert is interested in this testing, so I add him back.

Cheers,
Jarek P.

Quoted 20 lines
Quoted 6 lines

Yes i can make tests like this.
My network is splited to test clients and other normal clients
so it's really no problem to make testing. - if testing clients working then traffic from normal clients is also switched to this router (but if traffic is not forwarded "like in this case" for testing clients then failover switching them to working router )

and other point to make this tests - is that - it is good to have all in linux kernel networking working well :)

It's extremely nice of you! On the other hand, this type of change was planned to the net-next to fix possible memory problems, which might have happened to you as well. So you'd probably experience this problem in the future (2.6.32) anyway.

So here is the first of 2 patches (the second in a separate message), which should be tested separately, each one applied on top of the 2.6.29.x (vanilla - at least fib_trie.c), after reverting the previous one. So, they are again all-in-one, to eclude any misunderstanding.

Btw., I assume there were no oopses, warnings or lockups after those previous non-working patches - only no routing/forwarding.

Yes on on previous patches there was / no warnings / no oopses or lockups

But now i apply this patch and i make more testing.
First boot with start of bgpd and - traffic is not forwarded
So i start to search and make only some routes (static without bgpd) thru this host And all is working for this host when i make all by static routes.

So i change a little my bgp configuration and make default route to only one of my iBGP peers and start bgpd process All is working and what is weird is number of routes in kernel table.
Kernel is learning routes from bgpd but very slowly - really very slowly.

Quoted 8 lines

In attached file there are some fib_triestats after 5min of traffic.

Without this patch (normally)
total size: reported by fib_triestats in less that 1sec is: "Total size: 35769 kB"

But with this patch
Total size is growing up and in 5 min of traffic it grow to only: "Total size: 1005 kB"

Sorry no attached file.

Quoted 171 lines

Regards
Paweł Staszewski

Quoted 168 lines
Quoted 168 lines

Thanks,
Jarek P.
----------> (synchronize take 6 all-in-one for 2.6.29x, .28, or .27)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c 2009-07-01 05:15:37.000000000 +0000 renamed from net/ipv4/fib_trie.c 2009-06-29 05:30:50.000000000 +0000 +32 -9
@@ -123,6 +123,7 @@ struct tnode {
     union {
         struct rcu_head rcu;
         struct work_struct work;
+        struct tnode *tnode_free;
     };
     struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -385,6 +388,24 @@ static inline void tnode_free(struct tno
         call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+    BUG_ON(IS_LEAF(tn));
+    tn->tnode_free = tnode_free_head;
+    tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+    struct tnode *tn;
+
+    while ((tn = tnode_free_head)) {
+        tnode_free_head = tn->tnode_free;
+        tn->tnode_free = NULL;
+        tnode_free(tn);
+    }
+}
+
 static struct leaf *leaf_new(void)
 {
     struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +516,7 @@ static struct node *resize(struct trie *
 
     /* No children */
     if (tn->empty_children == tnode_child_length(tn)) {
-        tnode_free(tn);
+        tnode_free_safe(tn);
         return NULL;
     }
     /* One child */
@@ -509,7 +530,7 @@ static struct node *resize(struct trie *
 
             /* compress one level */
             node_set_parent(n, NULL);
-            tnode_free(tn);
+            tnode_free_safe(tn);
             return n;
         }
     /*
@@ -670,7 +691,7 @@ static struct node *resize(struct trie *
             /* compress one level */
 
             node_set_parent(n, NULL);
-            tnode_free(tn);
+            tnode_free_safe(tn);
             return n;
         }
 
@@ -756,7 +777,7 @@ static struct tnode *inflate(struct trie
             put_child(t, tn, 2*i, inode->child[0]);
             put_child(t, tn, 2*i+1, inode->child[1]);
 
-            tnode_free(inode);
+            tnode_free_safe(inode);
             continue;
         }
 
@@ -801,9 +822,9 @@ static struct tnode *inflate(struct trie
         put_child(t, tn, 2*i, resize(t, left));
         put_child(t, tn, 2*i+1, resize(t, right));
 
-        tnode_free(inode);
+        tnode_free_safe(inode);
     }
-    tnode_free(oldtnode);
+    tnode_free_safe(oldtnode);
     return tn;
 nomem:
     {
@@ -885,7 +906,7 @@ static struct tnode *halve(struct trie *
         put_child(t, newBinNode, 1, right);
         put_child(t, tn, i/2, resize(t, newBinNode));
     }
-    tnode_free(oldtnode);
+    tnode_free_safe(oldtnode);
     return tn;
 nomem:
     {
@@ -983,12 +1004,14 @@ fib_find_node(struct trie *t, u32 key)
     return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn, bool 
 sync)
 {
     int wasfull;
-    t_key cindex, key = tn->key;
+    t_key cindex, key;
     struct tnode *tp;
 
+    key = tn->key;
+
     while (tn != NULL && (tp = node_parent((struct node *)tn)) != 
 NULL) {
         cindex = tkey_extract_bits(key, tp->pos, tp->bits);
         wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
net/ipv4/fib_trie.c 2009-07-01 05:15:37.000000000 +0000 renamed from net/ipv4/fib_trie.c 2009-06-29 05:30:50.000000000 +0000 +12 -3
@@ -999,6 +1022,10 @@ static struct node *trie_rebalance(struc
 
         tp = node_parent((struct node *) tn);
         if (!tp)
+            rcu_assign_pointer(t->trie, (struct node *)tn);
+
+        //tnode_free_flush();
+        if (!tp)
             break;
         tn = tp;
     }
@@ -1007,7 +1034,12 @@ static struct node *trie_rebalance(struc
     if (IS_TNODE(tn))
         tn = (struct tnode *)resize(t, (struct tnode *)tn);
 
-    return (struct node *)tn;
+    rcu_assign_pointer(t->trie, (struct node *)tn);
+    if (sync)
+        synchronize_rcu();
+    tnode_free_flush();
+
+    return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1187,7 @@ static struct list_head *fib_insert_node
 
     /* Rebalance the trie */
 
-    rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+    trie_rebalance(t, tp, true);
 done:
     return fa_head;
 }
@@ -1575,7 +1607,7 @@ static void trie_leaf_remove(struct trie
     if (tp) {
         t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
         put_child(t, (struct tnode *)tp, cindex, NULL);
-        rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+        trie_rebalance(t, tp, false);
     } else
         rcu_assign_pointer(t->trie, NULL);
 
Quoted 15 lines
Quoted 8 lines
Quoted 1 line
Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

        Aver depth:     3.79
        Max depth:      9
        Leaves:         15518
        Prefixes:       15933
        Internal nodes: 3973
          1: 2260  2: 674  3: 518  4: 268  5: 164  6: 79  7: 5  8: 2  9: 1  10: 2
        Pointers: 29664

Null ptrs: 10174
Total size: 995 kB

Counters:
---------
gets = 17863461
backtracks = 13345457 semantic match passed = 17305229 semantic match miss = 419 null node hit= 17602641 skipped node resize = 0

Local:

        Aver depth:     3.75
        Max depth:      5
        Leaves:         12
        Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 17865423
backtracks = 4964174 semantic match passed = 2126 semantic match miss = 0 null node hit= 1853
skipped node resize = 0

----------- After 30sec ----------------
cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

        Aver depth:     3.79
        Max depth:      9
        Leaves:         15686
        Prefixes:       16111
        Internal nodes: 4002
          1: 2259  2: 679  3: 536  4: 274  5: 165  6: 79  7: 5  8: 2  9: 1  10: 2
        Pointers: 29954

Null ptrs: 10267
Total size: 1005 kB

Counters:
---------
gets = 18042821
backtracks = 13523292 semantic match passed = 17484572 semantic match miss = 419 null node hit= 17799334 skipped node resize = 0

Local:

        Aver depth:     3.75
        Max depth:      5
        Leaves:         12
        Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 18044798
backtracks = 5012942 semantic match passed = 2140 semantic match miss = 0 null node hit= 1865
skipped node resize = 0

#59 lore raw
In reply to Paweł Staszewski

On Wed, Jul 01, 2009 at 11:43:04AM +0200, Paweł Staszewski wrote:
...

Quoted 8 lines

Yes on on previous patches there was / no warnings / no oopses or lockups

But now i apply this patch and i make more testing.
First boot with start of bgpd and - traffic is not forwarded So i start to search and make only some routes (static without bgpd) thru this host
And all is working for this host when i make all by static routes.

Quoted 4 lines

So i change a little my bgp configuration and make default route to only one of my iBGP peers and start bgpd process
All is working and what is weird is number of routes in kernel table. Kernel is learning routes from bgpd but very slowly - really very slowly.

Pawel, this is really very helpful! So, this is (probably) only about timing, not wrong memory freeing. On the other hand this test was only for inserts. Btw., if you didn't start the second test, you can skip it. I have to rethink this.

Many thanks,
Jarek P.

#60 lore raw
Quoted 15 lines · Jarek Poplawski

On Wed, Jul 01, 2009 at 11:43:04AM +0200, Paweł Staszewski wrote:
...

Quoted 8 lines

Yes on on previous patches there was / no warnings / no oopses or lockups

But now i apply this patch and i make more testing.
First boot with start of bgpd and - traffic is not forwarded So i start to search and make only some routes (static without bgpd) thru this host
And all is working for this host when i make all by static routes.

Quoted 4 lines

So i change a little my bgp configuration and make default route to only one of my iBGP peers and start bgpd process
All is working and what is weird is number of routes in kernel table. Kernel is learning routes from bgpd but very slowly - really very slowly.

Quoted 4 lines

Pawel, this is really very helpful! So, this is (probably) only about timing, not wrong memory freeing. On the other hand this test was only for inserts. Btw., if you didn't start the second test, you can skip it. I have to rethink this.

So, after your findings I'm about to recommend sending to -stable 3 patches from net-2.6, with additional lowering of threshold_root settings, but it would be nice if you could give it a try with CONFIG_PREEMPT instead of CONFIG_PREEMPT_NONE (if it doesn't break your other apps!) It is expected to work this time...;-) Maybe a bit slower.

Thanks,
Jarek P.
--------> (all-in-one preempt fixes to apply with vanilla 2.6.29.x)

net/ipv4/fib_trie.c +44 -14
diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
@@ -123,6 +123,7 @@ struct tnode {
 	union {
 		struct rcu_head rcu;
 		struct work_struct work;
+		struct tnode *tnode_free;
 	};
 	struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -313,8 +316,8 @@ static inline void check_tnode(const str
 
 static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
-static const int halve_threshold_root = 8;
-static const int inflate_threshold_root = 15;
+static const int halve_threshold_root = 15;
+static const int inflate_threshold_root = 25;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -385,6 +388,24 @@ static inline void tnode_free(struct tno
 		call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+	BUG_ON(IS_LEAF(tn));
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+	struct tnode *tn;
+
+	while ((tn = tnode_free_head)) {
+		tnode_free_head = tn->tnode_free;
+		tn->tnode_free = NULL;
+		tnode_free(tn);
+	}
+}
+
 static struct leaf *leaf_new(void)
 {
 	struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +516,7 @@ static struct node *resize(struct trie *
 
 	/* No children */
 	if (tn->empty_children == tnode_child_length(tn)) {
-		tnode_free(tn);
+		tnode_free_safe(tn);
 		return NULL;
 	}
 	/* One child */
@@ -509,7 +530,7 @@ static struct node *resize(struct trie *
 
 			/* compress one level */
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 	/*
@@ -670,7 +691,7 @@ static struct node *resize(struct trie *
 			/* compress one level */
 
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 
@@ -756,7 +777,7 @@ static struct tnode *inflate(struct trie
 			put_child(t, tn, 2*i, inode->child[0]);
 			put_child(t, tn, 2*i+1, inode->child[1]);
 
-			tnode_free(inode);
+			tnode_free_safe(inode);
 			continue;
 		}
 
@@ -801,9 +822,9 @@ static struct tnode *inflate(struct trie
 		put_child(t, tn, 2*i, resize(t, left));
 		put_child(t, tn, 2*i+1, resize(t, right));
 
-		tnode_free(inode);
+		tnode_free_safe(inode);
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -885,7 +906,7 @@ static struct tnode *halve(struct trie *
 		put_child(t, newBinNode, 1, right);
 		put_child(t, tn, i/2, resize(t, newBinNode));
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -983,12 +1004,14 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
-	t_key cindex, key = tn->key;
+	t_key cindex, key;
 	struct tnode *tp;
 
+	key = tn->key;
+
 	while (tn != NULL && (tp = node_parent((struct node *)tn)) != NULL) {
 		cindex = tkey_extract_bits(key, tp->pos, tp->bits);
 		wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
@@ -999,6 +1022,10 @@ static struct node *trie_rebalance(struc
 
 		tp = node_parent((struct node *) tn);
 		if (!tp)
+			rcu_assign_pointer(t->trie, (struct node *)tn);
+
+		tnode_free_flush();
+		if (!tp)
 			break;
 		tn = tp;
 	}
@@ -1007,7 +1034,10 @@ static struct node *trie_rebalance(struc
 	if (IS_TNODE(tn))
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
 
-	return (struct node *)tn;
+	rcu_assign_pointer(t->trie, (struct node *)tn);
+	tnode_free_flush();
+
+	return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1185,7 @@ static struct list_head *fib_insert_node
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp);
 done:
 	return fa_head;
 }
@@ -1575,7 +1605,7 @@ static void trie_leaf_remove(struct trie
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 
#61 lore raw

Jarek Poplawski pisze:

Quoted 22 lines · Jarek Poplawski
Quoted 11 lines · Paweł Staszewski

...

Quoted 7 lines

Yes on on previous patches there was / no warnings / no oopses or lockups

But now i apply this patch and i make more testing.
First boot with start of bgpd and - traffic is not forwarded
So i start to search and make only some routes (static without bgpd) thru this host And all is working for this host when i make all by static routes.

Quoted 9 lines
Quoted 4 lines

So i change a little my bgp configuration and make default route to only one of my iBGP peers and start bgpd process All is working and what is weird is number of routes in kernel table.
Kernel is learning routes from bgpd but very slowly - really very slowly.

Pawel, this is really very helpful! So, this is (probably) only about timing, not wrong memory freeing. On the other hand this test was only for inserts. Btw., if you didn't start the second test, you can skip it. I have to rethink this.

Quoted 8 lines

So, after your findings I'm about to recommend sending to -stable 3 patches from net-2.6, with additional lowering of threshold_root settings, but it would be nice if you could give it a try with CONFIG_PREEMPT instead of CONFIG_PREEMPT_NONE (if it doesn't break your other apps!) It is expected to work this time...;-) Maybe a bit slower.

Patch applied to 2.6.29.5 with CONFIG_PREEMPT_NONE And working :)

fib_triestats in attached file

I think I can test it with PREEMPT enabled but first i must make some other tests of my apps that are on server.

Regards
Paweł Staszewski

Quoted 170 lines

Thanks,
Jarek P.
--------> (all-in-one preempt fixes to apply with vanilla 2.6.29.x)

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c +43 -13
@@ -123,6 +123,7 @@ struct tnode {
 	union {
 		struct rcu_head rcu;
 		struct work_struct work;
+		struct tnode *tnode_free;
 	};
 	struct node *child[0];
 };
@@ -161,6 +162,8 @@ static void tnode_put_child_reorg(struct
 static struct node *resize(struct trie *t, struct tnode *tn);
 static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
+/* tnodes to free after resize(); protected by RTNL */
+static struct tnode *tnode_free_head;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -313,8 +316,8 @@ static inline void check_tnode(const str
 
 static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
-static const int halve_threshold_root = 8;
-static const int inflate_threshold_root = 15;
+static const int halve_threshold_root = 15;
+static const int inflate_threshold_root = 25;
 
 
 static void __alias_free_mem(struct rcu_head *head)
@@ -385,6 +388,24 @@ static inline void tnode_free(struct tno
 		call_rcu(&tn->rcu, __tnode_free_rcu);
 }
 
+static void tnode_free_safe(struct tnode *tn)
+{
+	BUG_ON(IS_LEAF(tn));
+	tn->tnode_free = tnode_free_head;
+	tnode_free_head = tn;
+}
+
+static void tnode_free_flush(void)
+{
+	struct tnode *tn;
+
+	while ((tn = tnode_free_head)) {
+		tnode_free_head = tn->tnode_free;
+		tn->tnode_free = NULL;
+		tnode_free(tn);
+	}
+}
+
 static struct leaf *leaf_new(void)
 {
 	struct leaf *l = kmem_cache_alloc(trie_leaf_kmem, GFP_KERNEL);
@@ -495,7 +516,7 @@ static struct node *resize(struct trie *
 
 	/* No children */
 	if (tn->empty_children == tnode_child_length(tn)) {
-		tnode_free(tn);
+		tnode_free_safe(tn);
 		return NULL;
 	}
 	/* One child */
@@ -509,7 +530,7 @@ static struct node *resize(struct trie *
 
 			/* compress one level */
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 	/*
@@ -670,7 +691,7 @@ static struct node *resize(struct trie *
 			/* compress one level */
 
 			node_set_parent(n, NULL);
-			tnode_free(tn);
+			tnode_free_safe(tn);
 			return n;
 		}
 
@@ -756,7 +777,7 @@ static struct tnode *inflate(struct trie
 			put_child(t, tn, 2*i, inode->child[0]);
 			put_child(t, tn, 2*i+1, inode->child[1]);
 
-			tnode_free(inode);
+			tnode_free_safe(inode);
 			continue;
 		}
 
@@ -801,9 +822,9 @@ static struct tnode *inflate(struct trie
 		put_child(t, tn, 2*i, resize(t, left));
 		put_child(t, tn, 2*i+1, resize(t, right));
 
-		tnode_free(inode);
+		tnode_free_safe(inode);
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -885,7 +906,7 @@ static struct tnode *halve(struct trie *
 		put_child(t, newBinNode, 1, right);
 		put_child(t, tn, i/2, resize(t, newBinNode));
 	}
-	tnode_free(oldtnode);
+	tnode_free_safe(oldtnode);
 	return tn;
 nomem:
 	{
@@ -983,12 +1004,14 @@ fib_find_node(struct trie *t, u32 key)
 	return NULL;
 }
 
-static struct node *trie_rebalance(struct trie *t, struct tnode *tn)
+static void trie_rebalance(struct trie *t, struct tnode *tn)
 {
 	int wasfull;
-	t_key cindex, key = tn->key;
+	t_key cindex, key;
 	struct tnode *tp;
 
+	key = tn->key;
+
 	while (tn != NULL && (tp = node_parent((struct node *)tn)) != NULL) {
 		cindex = tkey_extract_bits(key, tp->pos, tp->bits);
 		wasfull = tnode_full(tp, tnode_get_child(tp, cindex));
@@ -999,6 +1022,10 @@ static struct node *trie_rebalance(struc
 
 		tp = node_parent((struct node *) tn);
 		if (!tp)
+			rcu_assign_pointer(t->trie, (struct node *)tn);
+
+		tnode_free_flush();
+		if (!tp)
 			break;
 		tn = tp;
 	}
@@ -1007,7 +1034,10 @@ static struct node *trie_rebalance(struc
 	if (IS_TNODE(tn))
 		tn = (struct tnode *)resize(t, (struct tnode *)tn);
 
-	return (struct node *)tn;
+	rcu_assign_pointer(t->trie, (struct node *)tn);
+	tnode_free_flush();
+
+	return;
 }
 
 /* only used from updater-side */
@@ -1155,7 +1185,7 @@ static struct list_head *fib_insert_node
 
 	/* Rebalance the trie */
 
-	rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+	trie_rebalance(t, tp);
 done:
 	return fa_head;
 }

net/ipv4/fib_trie.c static void trie_leaf_remove(struct trie · line 1611

Quoted 8 lines
net/ipv4/fib_trie.c +1 -1
@@ -1575,7 +1605,7 @@ static void trie_leaf_remove(struct trie
 	if (tp) {
 		t_key cindex = tkey_extract_bits(l->key, tp->pos, tp->bits);
 		put_child(t, (struct tnode *)tp, cindex, NULL);
-		rcu_assign_pointer(t->trie, trie_rebalance(t, tp));
+		trie_rebalance(t, tp);
 	} else
 		rcu_assign_pointer(t->trie, NULL);
 

cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

        Aver depth:     2.44
        Max depth:      6
        Leaves:         277395
        Prefixes:       290874
        Internal nodes: 66711
          1: 32915  2: 14668  3: 10752  4: 4913  5: 2197  6: 895  7: 367  8: 3  17: 1
        Pointers: 595526

Null ptrs: 251421
Total size: 18044 kB

Counters:
---------
gets = 2705388
backtracks = 137797 semantic match passed = 2658993 semantic match miss = 87 null node hit= 1980950 skipped node resize = 0

Local:

        Aver depth:     3.75
        Max depth:      5
        Leaves:         12
        Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 2709741
backtracks = 1584810 semantic match passed = 4417 semantic match miss = 0 null node hit= 192688 skipped node resize = 0

#62 lore raw
Quoted 1 line · Paweł Staszewski

Jarek Poplawski pisze:

...

Quoted 10 lines
Quoted 8 lines

So, after your findings I'm about to recommend sending to -stable 3 patches from net-2.6, with additional lowering of threshold_root settings, but it would be nice if you could give it a try with CONFIG_PREEMPT instead of CONFIG_PREEMPT_NONE (if it doesn't break your other apps!) It is expected to work this time...;-) Maybe a bit slower.

Patch applied to 2.6.29.5 with CONFIG_PREEMPT_NONE And working :)

Hmm... It should, because you tested very similar patch already;-) Sorry if I didn't make it clear.

Quoted 5 lines

fib_triestats in attached file

I think I can test it with PREEMPT enabled but first i must make some other tests of my apps that are on server.

It could probably matter only if you're using some broken out-of-tree patches. Otherwise the kernel is expected to work OK.

Btw., it would be also interesting to check if there is any difference wrt. these route cache problems while PREEMPT is enabled.

Thanks,
Jarek P.

#63 lore raw

Jarek Poplawski pisze:

Quoted 15 lines · Paweł Staszewski
Quoted 3 lines

Jarek Poplawski pisze:

...

Quoted 9 lines
Quoted 9 lines

So, after your findings I'm about to recommend sending to -stable 3 patches from net-2.6, with additional lowering of threshold_root settings, but it would be nice if you could give it a try with CONFIG_PREEMPT instead of CONFIG_PREEMPT_NONE (if it doesn't break your other apps!) It is expected to work this time...;-) Maybe a bit slower.

Quoted 8 lines
Quoted 3 lines

Patch applied to 2.6.29.5 with CONFIG_PREEMPT_NONE And working :)

Hmm... It should, because you tested very similar patch already;-) Sorry if I didn't make it clear.

Yes i know there was almost identical one.
And i see this was without sync rcu :)

Quoted 9 lines
Quoted 4 lines

fib_triestats in attached file

I think I can test it with PREEMPT enabled but first i must make some other tests of my apps that are on server.

It could probably matter only if you're using some broken out-of-tree patches. Otherwise the kernel is expected to work OK.

Im a little confused about using of PREEMPT kernel because of past there was many oopses / lockups :) but yes that was a little long time ago. I will try to make this test today.

Quoted 12 lines

Btw., it would be also interesting to check if there is any difference wrt. these route cache problems while PREEMPT is enabled.

Thanks,
Jarek P.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#64 lore raw
Quoted 19 lines · Paweł Staszewski

Jarek Poplawski pisze:

Quoted 18 lines · Paweł Staszewski
Quoted 3 lines

Jarek Poplawski pisze:

...

Quoted 11 lines
Quoted 8 lines

So, after your findings I'm about to recommend sending to -stable 3 patches from net-2.6, with additional lowering of threshold_root settings, but it would be nice if you could give it a try with CONFIG_PREEMPT instead of CONFIG_PREEMPT_NONE (if it doesn't break your other apps!) It is expected to work this time...;-) Maybe a bit slower.

Patch applied to 2.6.29.5 with CONFIG_PREEMPT_NONE And working :)

Quoted 6 lines
Quoted 4 lines

Hmm... It should, because you tested very similar patch already;-) Sorry if I didn't make it clear.

Yes i know there was almost identical one.
And i see this was without sync rcu :)

Yes, it looks like we can't free memory so simple because of such huge latencies.

Quoted 7 lines
Quoted 7 lines
Quoted 6 lines

fib_triestats in attached file

I think I can test it with PREEMPT enabled but first i must make some
other tests of my apps that are on server.

Quoted 10 lines
Quoted 4 lines

It could probably matter only if you're using some broken out-of-tree patches. Otherwise the kernel is expected to work OK.

Im a little confused about using of PREEMPT kernel because of past there was many oopses / lockups :) but yes that was a little long time ago. I will try to make this test today.

Quoted 2 lines

Btw., it would be also interesting to check if there is any difference wrt. these route cache problems while PREEMPT is enabled.

And you're very right! The place we're fixing is the best example. On the other hand, I hope there is not many such places yet. But if we test/fix it there will be one less...

Jarek P.

#65 lore raw
Quoted 2 lines · Jarek Poplawski

Yes, it looks like we can't free memory so simple because of such huge latencies.

Controlling RCU seems crucial. Insertion of the full BGP table increased from 2 seconds to > 20 min with one synchronize_rcu patches.

And fib_trie "worst case" wrt memory is the root node. So maybe we should monitor changes in root node and use this to control synchronize_rcu.

Didn't Paul suggest something like this?

And with don't find any decent solution we have to add an option for a fixed and pre-allocated root-nod typically for BGP-routers.

Cheers

					--ro
#66 lore raw
Quoted 8 lines · Robert Olsson
Quoted 2 lines · Jarek Poplawski

Yes, it looks like we can't free memory so simple because of such huge latencies.

Controlling RCU seems crucial. Insertion of the full BGP table increased from 2 seconds to > 20 min with one synchronize_rcu patches.

I wish I knew this a few days before. I could imagine a slow down, but it looked like it was stuck. Since these last changes weren't tested on SMP + PREEMPT I thought there is still something broken. (I was mainly interested in this synchronize_rcu at the moment as a preemption test.)

Quoted 4 lines

And fib_trie "worst case" wrt memory is the root node. So maybe we should monitor changes in root node and use this to control synchronize_rcu.

Didn't Paul suggest something like this?

Sure, and it needs testing, but we should send some safe preemption fix for -stable first, don't we?

Quoted 2 lines

And with don't find any decent solution we have to add an option for a fixed and pre-allocated root-nod typically for BGP-routers.

Probably you're right; I'd prefer to see the test results showing a difference vs. simply less aggressive root thresholds. But of course, even if not convinced, I'll respect your choice as the author and maintainer, so feel free to NAK my proposals - I won't get it personally.;-)

Cheers,
Jarek P.

#67 lore raw
Quoted 8 lines · Jarek Poplawski
Quoted 2 lines

Controlling RCU seems crucial. Insertion of the full BGP table increased from 2 seconds to > 20 min with one synchronize_rcu patches.

I wish I knew this a few days before. I could imagine a slow down, but it looked like it was stuck. Since these last changes weren't tested on SMP + PREEMPT I thought there is still something broken. (I was mainly interested in this synchronize_rcu at the moment as a preemption test.)

Honestly this huge slowdown was surprise for me too. I think I sent you a script so you could insert the full table yourself.

Quoted 7 lines
Quoted 4 lines

And fib_trie "worst case" wrt memory is the root node. So maybe we should monitor changes in root node and use this to control synchronize_rcu.

Didn't Paul suggest something like this?

Sure, and it needs testing, but we should send some safe preemption fix for -stable first, don't we?

Yes my hope was that we could combine them... personally I'll need to understand who we can preeemted better in the different configs and most of that this can be handled by "standard" RCU.

Quoted 8 lines
Quoted 2 lines

And with don't find any decent solution we have to add an option for a fixed and pre-allocated root-nod typically for BGP-routers.

Probably you're right; I'd prefer to see the test results showing a difference vs. simply less aggressive root thresholds. But of course, even if not convinced, I'll respect your choice as the author and maintainer, so feel free to NAK my proposals - I won't get it personally.;-)

Thresholds we can change no problem... but very soon I'll people will start routing without the route cache this at least in close to Internet core ,we will need all fib_look performance we can get.

fib_trie was designed for classical RCU and no preempt you see the names i file... so this new and very challenging work to all of us.

First week of vacation and have to fix the roof of the house... it's hot and dirty.

Cheers.

					--ro
#68 lore raw
Quoted 13 lines · Robert Olsson
Quoted 8 lines · Jarek Poplawski
Quoted 2 lines

Controlling RCU seems crucial. Insertion of the full BGP table increased from 2 seconds to > 20 min with one synchronize_rcu patches.

I wish I knew this a few days before. I could imagine a slow down, but it looked like it was stuck. Since these last changes weren't tested on SMP + PREEMPT I thought there is still something broken. (I was mainly interested in this synchronize_rcu at the moment as a preemption test.)

Quoted 2 lines

Honestly this huge slowdown was surprise for me too. I think I sent you a script so you could insert the full table yourself.

I can't remember this script, but I guess my hardware should be suitable for reading it.;-)

Quoted 22 lines
Quoted 8 lines
Quoted 5 lines

And fib_trie "worst case" wrt memory is the root node. So maybe we should monitor changes in root node and use this to control synchronize_rcu.

Didn't Paul suggest something like this?

Sure, and it needs testing, but we should send some safe preemption fix for -stable first, don't we?

Yes my hope was that we could combine them... personally I'll need to understand who we can preeemted better in the different configs and most of that this can be handled by "standard" RCU.

Quoted 8 lines
Quoted 2 lines

And with don't find any decent solution we have to add an option for a fixed and pre-allocated root-nod typically for BGP-routers.

Probably you're right; I'd prefer to see the test results showing a difference vs. simply less aggressive root thresholds. But of course, even if not convinced, I'll respect your choice as the author and maintainer, so feel free to NAK my proposals - I won't get it personally.;-)

Quoted 3 lines

Thresholds we can change no problem... but very soon I'll people will start routing without the route cache this at least in close to Internet core ,we will need all fib_look performance we can get.

I mean changing thresholds as a temporary solution, until we can control memory freeing; and it seems to me, even excluding the root node, there could be a lot of temporary allocations during all those cycles repeated 10 times.

Quoted 3 lines

fib_trie was designed for classical RCU and no preempt you see the names i file... so this new and very challenging work to all of us.

Then it should depend on CONFIG_PREEMPT_NONE, I guess.

Quoted 3 lines

First week of vacation and have to fix the roof of the house... it's hot and dirty.

Have a nice time,
Jarek P.

#69 lore raw
In reply to Jarek Poplawski

Jarek Poplawski pisze:

Quoted 13 lines · Paweł Staszewski
Quoted 12 lines

Jarek Poplawski pisze:

Quoted 9 lines · Paweł Staszewski
Quoted 5 lines

Jarek Poplawski pisze:

...

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

So, after your findings I'm about to recommend sending to -stable 3 patches from net-2.6, with additional lowering of threshold_root settings, but it would be nice if you could give it a try with CONFIG_PREEMPT instead of CONFIG_PREEMPT_NONE (if it doesn't break your other apps!) It is expected to work this time...;-) Maybe a bit slower.

Ok kernel configured with CONFIG_PREEMPT
and all this day work without any problems (with Jarek last patch).

So in attached file trere is fib_tirestats
I dont see any big change of (cpu load or faster/slower routing/propagating routes from bgpd or something else) - in avg there is from 2% to 3% more of CPU load i dont know why but it is - i change from "preempt" to "no preempt" 3 times and check this my "mpstat -P ALL 1 30"
always avg cpu load was from 2 to 3% more compared to "no preempt"

Regards
Paweł Staszewski

Quoted 47 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 2 lines

Patch applied to 2.6.29.5 with CONFIG_PREEMPT_NONE And working :)

Quoted 8 lines
Quoted 5 lines

Hmm... It should, because you tested very similar patch already;-) Sorry if I didn't make it clear.

Yes i know there was almost identical one.
And i see this was without sync rcu :)

Yes, it looks like we can't free memory so simple because of such huge latencies.

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

fib_triestats in attached file

I think I can test it with PREEMPT enabled but first i must make some
other tests of my apps that are on server.

Quoted 13 lines
Quoted 5 lines

It could probably matter only if you're using some broken out-of-tree patches. Otherwise the kernel is expected to work OK.

Im a little confused about using of PREEMPT kernel because of past there was many oopses / lockups :) but yes that was a little long time ago. I will try to make this test today.

Quoted 3 lines

Btw., it would be also interesting to check if there is any difference wrt. these route cache problems while PREEMPT is enabled.

And you're very right! The place we're fixing is the best example. On the other hand, I hope there is not many such places yet. But if we test/fix it there will be one less...

Jarek P.

#70 lore raw

Oh

I forgot - please Jarek give me patch with sync rcu and i will make test on preempt kernel

Thanks
Paweł Staszewski

Paweł Staszewski pisze:

Quoted 79 lines

Jarek Poplawski pisze:

Quoted 10 lines · Paweł Staszewski
Quoted 9 lines

Jarek Poplawski pisze:

Quoted 6 lines · Paweł Staszewski
Quoted 3 lines

Jarek Poplawski pisze:

...

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

So, after your findings I'm about to recommend sending to -stable 3 patches from net-2.6, with additional lowering of threshold_root settings, but it would be nice if you could give it a try with CONFIG_PREEMPT instead of CONFIG_PREEMPT_NONE (if it doesn't break your other apps!) It is expected to work this time...;-) Maybe a bit slower.

Ok kernel configured with CONFIG_PREEMPT
and all this day work without any problems (with Jarek last patch).

So in attached file trere is fib_tirestats
I dont see any big change of (cpu load or faster/slower routing/propagating routes from bgpd or something else) - in avg there is from 2% to 3% more of CPU load i dont know why but it is - i change from "preempt" to "no preempt" 3 times and check this my "mpstat -P ALL 1 30"
always avg cpu load was from 2 to 3% more compared to "no preempt"

Regards
Paweł Staszewski

Quoted 40 lines
Quoted 11 lines
Quoted 8 lines
Quoted 4 lines
Quoted 1 line

Patch applied to 2.6.29.5 with CONFIG_PREEMPT_NONE And working :)

Hmm... It should, because you tested very similar patch already;-) Sorry if I didn't make it clear.

Yes i know there was almost identical one.
And i see this was without sync rcu :)

Yes, it looks like we can't free memory so simple because of such huge latencies.

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

fib_triestats in attached file

I think I can test it with PREEMPT enabled but first i must make some other tests of my apps that are on server.

Quoted 12 lines
Quoted 4 lines

It could probably matter only if you're using some broken out-of-tree patches. Otherwise the kernel is expected to work OK.

Im a little confused about using of PREEMPT kernel because of past there was many oopses / lockups :) but yes that was a little long time ago. I will try to make this test today.

Quoted 3 lines

Btw., it would be also interesting to check if there is any difference wrt. these route cache problems while PREEMPT is enabled.

And you're very right! The place we're fixing is the best example. On the other hand, I hope there is not many such places yet. But if we test/fix it there will be one less...

Jarek P.

Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#71 lore raw
Quoted 4 lines · Paweł Staszewski

Oh

I forgot - please Jarek give me patch with sync rcu and i will make test on preempt kernel

Probably non-preempt kernel might need something like this more, but comparing is always interesting. This patch is based on Paul's suggestion (I hope).

Thanks,
Jarek P.
---> (synchronize take 7; apply on top of the 2.6.29.x with the last

	all-in-one patch, or net-2.6)
 net/ipv4/fib_trie.c |    8 ++++++++
 1 files changed, 8 insertions(+), 0 deletions(-)
net/ipv4/fib_trie.c +8 -0
index 00a54b2..fce8238 100644
@@ -164,6 +164,7 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -393,6 +394,8 @@ static void tnode_free_safe(struct tnode *tn)
 	BUG_ON(IS_LEAF(tn));
 	tn->tnode_free = tnode_free_head;
 	tnode_free_head = tn;
+	tnode_free_size += sizeof(struct tnode) +
+			   (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +407,11 @@ static void tnode_free_flush(void)
 		tn->tnode_free = NULL;
 		tnode_free(tn);
 	}
+
+	if (tnode_free_size >= PAGE_SIZE * 128) {
+		tnode_free_size = 0;
+		synchronize_rcu();
+	}
 }
 
 static struct leaf *leaf_new(void)
#72 lore raw
Quoted 9 lines · Jarek Poplawski
Quoted 4 lines · Paweł Staszewski

Oh

I forgot - please Jarek give me patch with sync rcu and i will make test on preempt kernel

Probably non-preempt kernel might need something like this more, but comparing is always interesting. This patch is based on Paul's suggestion (I hope).

Hold on ;-) Here is something even better... Syncing after 128 pages might be still too slow, so here is a higher initial value, 1000, plus you can change this while testing in:

/sys/module/fib_trie/parameters/sync_pages

It would be interesting to find the lowest acceptable value.

Jarek P.
---> (synchronize take 8; apply on top of the 2.6.29.x with the last
all-in-one patch, or net-2.6)

 net/ipv4/fib_trie.c |   12 ++++++++++++
 1 files changed, 12 insertions(+), 0 deletions(-)
net/ipv4/fib_trie.c +12 -0
index 00a54b2..decc8d0 100644
@@ -71,6 +71,7 @@
 #include <linux/netlink.h>
 #include <linux/init.h>
 #include <linux/list.h>
+#include <linux/moduleparam.h>
 #include <net/net_namespace.h>
 #include <net/ip.h>
 #include <net/protocol.h>
@@ -164,6 +165,10 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
+
+static int sync_pages __read_mostly = 1000;
+module_param(sync_pages, int, 0640);
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -393,6 +398,8 @@ static void tnode_free_safe(struct tnode *tn)
 	BUG_ON(IS_LEAF(tn));
 	tn->tnode_free = tnode_free_head;
 	tnode_free_head = tn;
+	tnode_free_size += sizeof(struct tnode) +
+			   (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +411,11 @@ static void tnode_free_flush(void)
 		tn->tnode_free = NULL;
 		tnode_free(tn);
 	}
+
+	if (tnode_free_size >= PAGE_SIZE * sync_pages) {
+		tnode_free_size = 0;
+		synchronize_rcu();
+	}
 }
 
 static struct leaf *leaf_new(void)
#73 lore raw
Quoted 11 lines · Jarek Poplawski
Quoted 9 lines · Jarek Poplawski
Quoted 4 lines · Paweł Staszewski

Oh

I forgot - please Jarek give me patch with sync rcu and i will make test on preempt kernel

Probably non-preempt kernel might need something like this more, but comparing is always interesting. This patch is based on Paul's suggestion (I hope).

Quoted 7 lines

Hold on ;-) Here is something even better... Syncing after 128 pages might be still too slow, so here is a higher initial value, 1000, plus you can change this while testing in:

/sys/module/fib_trie/parameters/sync_pages

It would be interesting to find the lowest acceptable value.

Looks like a promising approach to me!

Thanx, Paul

Quoted 55 lines

Jarek P.
---> (synchronize take 8; apply on top of the 2.6.29.x with the last

(quoted diff) +0 -0
 	all-in-one patch, or net-2.6)
 
 net/ipv4/fib_trie.c |   12 ++++++++++++
 1 files changed, 12 insertions(+), 0 deletions(-)
net/ipv4/fib_trie.c +12 -1
index 00a54b2..decc8d0 100644
@@ -71,6 +71,7 @@
 #include <linux/netlink.h>
 #include <linux/init.h>
 #include <linux/list.h>
+#include <linux/moduleparam.h>
 #include <net/net_namespace.h>
 #include <net/ip.h>
 #include <net/protocol.h>
@@ -164,6 +165,10 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
+
+static int sync_pages __read_mostly = 1000;
+module_param(sync_pages, int, 0640);
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -393,6 +398,8 @@ static void tnode_free_safe(struct tnode *tn)
 	BUG_ON(IS_LEAF(tn));
 	tn->tnode_free = tnode_free_head;
 	tnode_free_head = tn;
+	tnode_free_size += sizeof(struct tnode) +
+			   (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +411,11 @@ static void tnode_free_flush(void)
 		tn->tnode_free = NULL;
 		tnode_free(tn);
 	}
+
+	if (tnode_free_size >= PAGE_SIZE * sync_pages) {
+		tnode_free_size = 0;
+		synchronize_rcu();
+	}
 }
 
 static struct leaf *leaf_new(void)
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#74 lore raw
Quoted 16 lines · Paul E. McKenney
Quoted 15 lines · Jarek Poplawski
Quoted 9 lines · Jarek Poplawski
Quoted 4 lines · Paweł Staszewski

Oh

I forgot - please Jarek give me patch with sync rcu and i will make test on preempt kernel

Probably non-preempt kernel might need something like this more, but comparing is always interesting. This patch is based on Paul's suggestion (I hope).

Hold on ;-) Here is something even better... Syncing after 128 pages might be still too slow, so here is a higher initial value, 1000, plus you can change this while testing in:

Quoted 7 lines
Quoted 3 lines

/sys/module/fib_trie/parameters/sync_pages

It would be interesting to find the lowest acceptable value.

Looks like a promising approach to me!

Thanx, Paul

Hmm... As a matter of fact, I'm a bit sceptical now: I'm worrying this synchronize_rcu done at the lowest acceptable rate could be actually mostly idle or on the contrary too late. Probably some more complex (per cpu?) accounting would be necessary to really matter here, but on the other hand these problems weren't reported often enough.

Thanks,
Jarek P.

Quoted 54 lines
Quoted 54 lines

---> (synchronize take 8; apply on top of the 2.6.29.x with the last

(quoted diff) +0 -0
 	all-in-one patch, or net-2.6)
 
 net/ipv4/fib_trie.c |   12 ++++++++++++
 1 files changed, 12 insertions(+), 0 deletions(-)
net/ipv4/fib_trie.c +12 -1
index 00a54b2..decc8d0 100644
@@ -71,6 +71,7 @@
 #include <linux/netlink.h>
 #include <linux/init.h>
 #include <linux/list.h>
+#include <linux/moduleparam.h>
 #include <net/net_namespace.h>
 #include <net/ip.h>
 #include <net/protocol.h>
@@ -164,6 +165,10 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
+
+static int sync_pages __read_mostly = 1000;
+module_param(sync_pages, int, 0640);
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -393,6 +398,8 @@ static void tnode_free_safe(struct tnode *tn)
 	BUG_ON(IS_LEAF(tn));
 	tn->tnode_free = tnode_free_head;
 	tnode_free_head = tn;
+	tnode_free_size += sizeof(struct tnode) +
+			   (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +411,11 @@ static void tnode_free_flush(void)
 		tn->tnode_free = NULL;
 		tnode_free(tn);
 	}
+
+	if (tnode_free_size >= PAGE_SIZE * sync_pages) {
+		tnode_free_size = 0;
+		synchronize_rcu();
+	}
 }
 
 static struct leaf *leaf_new(void)
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#75 lore raw

kernel 2.6.29.5 preempt
bgp starts normal and kernel know routes normaly like without patch

Here are some fib_triestats

cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.44
       Max depth:      6
       Leaves:         277888
       Prefixes:       291399
       Internal nodes: 66818
         1: 33080  2: 14584  3: 10788  4: 4911  5: 2185  6: 900  7: 
366  8: 3  17: 1
       Pointers: 595584

Null ptrs: 250879
Total size: 18072 kB

Counters:
---------
gets = 1052940
backtracks = 55985 semantic match passed = 1034114 semantic match miss = 5 null node hit= 534415 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 1057636
backtracks = 1101307 semantic match passed = 4751 semantic match miss = 0 null node hit= 195605 skipped node resize = 0

kernel 2.6.29.5 no-preempt
All is ok like with preempt kernel (andl all working in normal time "routes propagation")

cat /sys/module/fib_trie/parameters/sync_pages 1000

cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.45
       Max depth:      6
       Leaves:         277905
       Prefixes:       291416
       Internal nodes: 66863
         1: 33119  2: 14594  3: 10782  4: 4911  5: 2187  6: 901  7: 
365  8: 3  17: 1
       Pointers: 595654

Null ptrs: 250887
Total size: 18074 kB

Counters:
---------
gets = 1060650
backtracks = 53161 semantic match passed = 1041008 semantic match miss = 12 null node hit= 504478 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 1065517
backtracks = 1095422 semantic match passed = 4954 semantic match miss = 0 null node hit= 195584 skipped node resize = 0

So i make tests with changing sync_pages And

####################################
sync_pages: 64
total size reach maximum in 17sec

Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.43
       Max depth:      6
       Leaves:         271928
       Prefixes:       285435
       Internal nodes: 66185
         1: 32904  2: 14554  3: 10740  4: 4677  5: 2047  6: 901  7: 
361  17: 1
       Pointers: 585224

Null ptrs: 247112
Total size: 17729 kB

Counters:
---------
gets = 5313544
backtracks = 230501 semantic match passed = 5233998 semantic match miss = 61 null node hit= 2757531 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 5332471
backtracks = 4708505 semantic match passed = 19264 semantic match miss = 0 null node hit= 782757 skipped node resize = 0

######################################
sync_pages: 128
Fib trie Total size reach max in 14sec
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.44
       Max depth:      6
       Leaves:         277915
       Prefixes:       291427
       Internal nodes: 66832
         1: 33085  2: 14597  3: 10785  4: 4908  5: 2187  6: 900  7: 
366  8: 3  17: 1
       Pointers: 595638

Null ptrs: 250892
Total size: 18074 kB

Counters:
---------
gets = 6698058
backtracks = 307491 semantic match passed = 6593421 semantic match miss = 66 null node hit= 3498560 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 6721120
backtracks = 5934017 semantic match passed = 23440 semantic match miss = 0 null node hit= 978008 skipped node resize = 0

#########################################
sync_pages: 256
hmm no difference also in 10sec

Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

       Aver depth:     2.44
       Max depth:      6
       Leaves:         277913
       Prefixes:       291425
       Internal nodes: 66829
         1: 33082  2: 14596  3: 10786  4: 4909  5: 2186  6: 900  7: 
366  8: 3  17: 1
       Pointers: 595620

Null ptrs: 250879
Total size: 18073 kB

Counters:
---------
gets = 4637474
backtracks = 188624 semantic match passed = 4577266 semantic match miss = 61 null node hit= 2451890 skipped node resize = 0

Local:

       Aver depth:     3.75
       Max depth:      5
       Leaves:         12
       Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 4651791
backtracks = 3716400 semantic match passed = 14613 semantic match miss = 0 null node hit= 587208 skipped node resize = 0

And with sync_pages higher that 256 time of filling kernel routes is the same approx 10sec.

I make this test bu use:
watch -n1 cat /proc/net/fib_triestat
timer start when Total size was 1kB and stop when Total size reach 18073 kB

Regards
Paweł Staszewski

Jarek Poplawski pisze:

Quoted 97 lines · Paul E. McKenney
Quoted 19 lines · Jarek Poplawski
Quoted 17 lines · Jarek Poplawski
Quoted 11 lines · Paweł Staszewski
Quoted 5 lines

Oh

I forgot - please Jarek give me patch with sync rcu and i will make test on preempt kernel

Probably non-preempt kernel might need something like this more, but comparing is always interesting. This patch is based on Paul's suggestion (I hope).

Hold on ;-) Here is something even better... Syncing after 128 pages might be still too slow, so here is a higher initial value, 1000, plus you can change this while testing in:

Quoted 8 lines
Quoted 4 lines

/sys/module/fib_trie/parameters/sync_pages

It would be interesting to find the lowest acceptable value.

Looks like a promising approach to me!

Thanx, Paul

Hmm... As a matter of fact, I'm a bit sceptical now: I'm worrying this synchronize_rcu done at the lowest acceptable rate could be actually mostly idle or on the contrary too late. Probably some more complex (per cpu?) accounting would be necessary to really matter here, but on the other hand these problems weren't reported often enough.

Thanks,
Jarek P.

Quoted 55 lines
Quoted 55 lines

---> (synchronize take 8; apply on top of the 2.6.29.x with the last

(quoted diff) +0 -0
 	all-in-one patch, or net-2.6)
 
 net/ipv4/fib_trie.c |   12 ++++++++++++
 1 files changed, 12 insertions(+), 0 deletions(-)
net/ipv4/fib_trie.c +12 -1
index 00a54b2..decc8d0 100644
@@ -71,6 +71,7 @@
 #include <linux/netlink.h>
 #include <linux/init.h>
 #include <linux/list.h>
+#include <linux/moduleparam.h>
 #include <net/net_namespace.h>
 #include <net/ip.h>
 #include <net/protocol.h>
@@ -164,6 +165,10 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
+
+static int sync_pages __read_mostly = 1000;
+module_param(sync_pages, int, 0640);
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -393,6 +398,8 @@ static void tnode_free_safe(struct tnode *tn)
 	BUG_ON(IS_LEAF(tn));
 	tn->tnode_free = tnode_free_head;
 	tnode_free_head = tn;
+	tnode_free_size += sizeof(struct tnode) +
+			   (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +411,11 @@ static void tnode_free_flush(void)
 		tn->tnode_free = NULL;
 		tnode_free(tn);
 	}
+
+	if (tnode_free_size >= PAGE_SIZE * sync_pages) {
+		tnode_free_size = 0;
+		synchronize_rcu();
+	}
 }
 
 static struct leaf *leaf_new(void)
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#76 lore raw

On Mon, Jul 06, 2009 at 01:53:49AM +0200, Paweł Staszewski wrote:
...

Quoted 6 lines

So i make tests with changing sync_pages
And

####################################
sync_pages: 64
total size reach maximum in 17sec

...

Quoted 3 lines

######################################
sync_pages: 128
Fib trie Total size reach max in 14sec

...

Quoted 3 lines

#########################################
sync_pages: 256
hmm no difference also in 10sec

14 == 10!? ;-)
...

Quoted 2 lines

And with sync_pages higher that 256 time of filling kernel routes is the same approx 10sec.

Hmm... So, it's better than I expected; syncing after 128 or 256 pages could be quite reasonable. But then it would be interesting to find out if with such a safety we could go back to more aggressive values for possibly better performance. So here is 'the same' patch (so the previous, take 8, should be reverted), but with additional possibility to change:
/sys/module/fib_trie/parameters/inflate_threshold_root

I guess, you could try e.g. if: sync_pages 256, inflate_threshold_root 15 can give faster lookups (or lower cpu loads); with this these inflate warnings could be back btw.; or maybe you'll find something in between like inflate_threshold_root 20 is optimal for you. (I think it should be enough to try this only for PREEMPT_NONE unless you have spare time ;-)

Thanks,
Jarek P.
---> (synchronize take 9; apply on top of the 2.6.29.x with the last
all-in-one patch, or net-2.6)

 net/ipv4/fib_trie.c |   18 ++++++++++++++++--
 1 files changed, 16 insertions(+), 2 deletions(-)
net/ipv4/fib_trie.c +16 -2
index 00a54b2..e8fca11 100644
@@ -71,6 +71,7 @@
 #include <linux/netlink.h>
 #include <linux/init.h>
 #include <linux/list.h>
+#include <linux/moduleparam.h>
 #include <net/net_namespace.h>
 #include <net/ip.h>
 #include <net/protocol.h>
@@ -164,6 +165,10 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
+
+static int sync_pages __read_mostly = 1000;
+module_param(sync_pages, int, 0640);
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -316,9 +321,11 @@ static inline void check_tnode(const struct tnode *tn)
 
 static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
-static const int halve_threshold_root = 15;
-static const int inflate_threshold_root = 25;
 
+static int inflate_threshold_root __read_mostly = 25;
+module_param(inflate_threshold_root, int, 0640);
+
+#define halve_threshold_root	(inflate_threshold_root / 2 + 1)
 
 static void __alias_free_mem(struct rcu_head *head)
 {
@@ -393,6 +400,8 @@ static void tnode_free_safe(struct tnode *tn)
 	BUG_ON(IS_LEAF(tn));
 	tn->tnode_free = tnode_free_head;
 	tnode_free_head = tn;
+	tnode_free_size += sizeof(struct tnode) +
+			   (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +413,11 @@ static void tnode_free_flush(void)
 		tn->tnode_free = NULL;
 		tnode_free(tn);
 	}
+
+	if (tnode_free_size >= PAGE_SIZE * sync_pages) {
+		tnode_free_size = 0;
+		synchronize_rcu();
+	}
 }
 
 static struct leaf *leaf_new(void)
#77 lore raw

Jarek Poplawski pisze:

Quoted 43 lines · Paweł Staszewski

...

Quoted 7 lines

So i make tests with changing sync_pages
And

####################################
sync_pages: 64
total size reach maximum in 17sec

...

Quoted 4 lines

######################################
sync_pages: 128
Fib trie Total size reach max in 14sec

...

Quoted 4 lines

#########################################
sync_pages: 256
hmm no difference also in 10sec

14 == 10!? ;-)
...

Quoted 2 lines

And with sync_pages higher that 256 time of filling kernel routes is the same approx 10sec.

Hmm... So, it's better than I expected; syncing after 128 or 256 pages could be quite reasonable. But then it would be interesting to find out if with such a safety we could go back to more aggressive values for possibly better performance. So here is 'the same' patch (so the previous, take 8, should be reverted), but with additional possibility to change:
/sys/module/fib_trie/parameters/inflate_threshold_root

I guess, you could try e.g. if: sync_pages 256, inflate_threshold_root 15 can give faster lookups (or lower cpu loads); with this these inflate warnings could be back btw.; or maybe you'll find something in between like inflate_threshold_root 20 is optimal for you. (I think it should be enough to try this only for PREEMPT_NONE unless you have spare time ;-)

Quoted 6 lines

Thanks,
Jarek P.
---> (synchronize take 9; apply on top of the 2.6.29.x with the last
all-in-one patch, or net-2.6)

Applied to 2.6.29.5 preempt/no-preempt and tested: - with preempt i make only one test with sync_pages = 256 to check that is working :)

So here are some tests for different sync_pages size.

echo 1 > /sys/module/fib_trie/parameters/sync_pages I stop count after 1minute - total size still rising :)

echo 2 > /sys/module/fib_trie/parameters/sync_pages Total size in fib_triestats reach maximum in 33sec

echo 3 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 31sec

echo 4 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 23sec

echo 8 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 17sec

echo 16 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 14 sec

echo 32 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 14 sec

So i see in prev tests i make something wrong in time counting So i modify test script and make tests again:

echo 64 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 13 sec

echo 128 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 10 sec

echo 256 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 10 sec

And for sync_paqges >256 time for propagating routes is always 10sec.

Also today i have many messages in dmesg like this:
Fix inflate_threshold_root. Now=25 size=11 bits :)

And after tune :
/sys/module/fib_trie/parameters/inflate_threshold_root no more info :)

Regards
Paweł Staszewski

Quoted 68 lines
 net/ipv4/fib_trie.c |   18 ++++++++++++++++--
 1 files changed, 16 insertions(+), 2 deletions(-)
net/ipv4/fib_trie.c +16 -3
index 00a54b2..e8fca11 100644
@@ -71,6 +71,7 @@
 #include <linux/netlink.h>
 #include <linux/init.h>
 #include <linux/list.h>
+#include <linux/moduleparam.h>
 #include <net/net_namespace.h>
 #include <net/ip.h>
 #include <net/protocol.h>
@@ -164,6 +165,10 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
+
+static int sync_pages __read_mostly = 1000;
+module_param(sync_pages, int, 0640);
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -316,9 +321,11 @@ static inline void check_tnode(const struct tnode *tn)
 
 static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
-static const int halve_threshold_root = 15;
-static const int inflate_threshold_root = 25;
 
+static int inflate_threshold_root __read_mostly = 25;
+module_param(inflate_threshold_root, int, 0640);
+
+#define halve_threshold_root	(inflate_threshold_root / 2 + 1)
 
 static void __alias_free_mem(struct rcu_head *head)
 {
@@ -393,6 +400,8 @@ static void tnode_free_safe(struct tnode *tn)
 	BUG_ON(IS_LEAF(tn));
 	tn->tnode_free = tnode_free_head;
 	tnode_free_head = tn;
+	tnode_free_size += sizeof(struct tnode) +
+			   (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +413,11 @@ static void tnode_free_flush(void)
 		tn->tnode_free = NULL;
 		tnode_free(tn);
 	}
+
+	if (tnode_free_size >= PAGE_SIZE * sync_pages) {
+		tnode_free_size = 0;
+		synchronize_rcu();
+	}
 }
 
 static struct leaf *leaf_new(void)
--

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#78 lore raw
Quoted 1 line · Paweł Staszewski

Jarek Poplawski pisze:

...

Quoted 9 lines
Quoted 4 lines

---> (synchronize take 9; apply on top of the 2.6.29.x with the last
all-in-one patch, or net-2.6)

Applied to 2.6.29.5 preempt/no-preempt and tested: - with preempt i make only one test with sync_pages = 256 to check that is working :)

So here are some tests for different sync_pages size.

...

Quoted 6 lines

So i see in prev tests i make something wrong in time counting So i modify test script and make tests again:

echo 64 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 13 sec

Quoted 7 lines

echo 128 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 10 sec

echo 256 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 10 sec

And for sync_paqges >256 time for propagating routes is always 10sec.

So this means sync_pages 128 or 256 is reasonable.

Quoted 4 lines

Also today i have many messages in dmesg like this:
Fix inflate_threshold_root. Now=25 size=11 bits :)

This is something new and a bit surprising to me: the same threshold in previous tests didn't generate this? Do you mean more than:
"Fix inflate_threshold_root. Now=15 size=11 bits" before?

Quoted 3 lines

And after tune :
/sys/module/fib_trie/parameters/inflate_threshold_root no more info :)

With what value?

Pawel, let's say that current defaults are:
inflate_threshold_root 25 sync_pages 256

I'd like you to try to check if e.g.:
inflate_threshold_root 15 sync_pages 256
can give you any visible or subjective difference worth tweaking it at all? (These stats from the next messages don't show this enough.) You don't need to hurry with this...

Thanks,
Jarek P.

#79 lore raw

Jarek Poplawski pisze:

Quoted 35 lines · Paweł Staszewski
Quoted 3 lines

Jarek Poplawski pisze:

...

Quoted 9 lines
Quoted 5 lines

---> (synchronize take 9; apply on top of the 2.6.29.x with the last
all-in-one patch, or net-2.6)

Applied to 2.6.29.5 preempt/no-preempt and tested: - with preempt i make only one test with sync_pages = 256 to check that is working :)

So here are some tests for different sync_pages size.

...

Quoted 6 lines

So i see in prev tests i make something wrong in time counting So i modify test script and make tests again:

echo 64 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 13 sec

Quoted 8 lines

echo 128 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 10 sec

echo 256 > /sys/module/fib_trie/parameters/sync_pages Total size reach max in 10 sec

And for sync_paqges >256 time for propagating routes is always 10sec.

So this means sync_pages 128 or 256 is reasonable.

Quoted 10 lines
Quoted 4 lines

Also today i have many messages in dmesg like this:
Fix inflate_threshold_root. Now=25 size=11 bits :)

This is something new and a bit surprising to me: the same threshold
in previous tests didn't generate this? Do you mean more than: "Fix inflate_threshold_root. Now=15 size=11 bits" before?

Yes. Sorry for that - this info was not all the day but only 5 minutes when i was making tests. This info was reported only when all iBGP peers was down/up fast.

Quoted 8 lines
Quoted 4 lines

And after tune :
/sys/module/fib_trie/parameters/inflate_threshold_root no more info :)

With what value?

When i set 35 as inflate_threshold_root there was no info even if all iBGP peers was down/up.

But i start to search when i have info about "Fix inflate_threshold_root"
And i see that the best is set this to 20 for me i have no info then in normal router operation / without down/up bgp peers many times in short time.

Quoted 10 lines

Pawel, let's say that current defaults are:
inflate_threshold_root 25 sync_pages 256

I'd like you to try to check if e.g.:
inflate_threshold_root 15 sync_pages 256
can give you any visible or subjective difference worth tweaking it at all? (These stats from the next messages don't show this enough.) You don't need to hurry with this...

I will try to make more accurate tests in weekend.

Regards
Paweł Staszewski

Quoted 5 lines

Thanks,
Jarek P.

Tested-by Pawel Staszewski
#80 lore raw
Quoted 2 lines · Paweł Staszewski

Jarek Poplawski pisze:

Quoted 1 line · Paweł Staszewski

...

Quoted 19 lines
Quoted 10 lines
Quoted 4 lines

Also today i have many messages in dmesg like this:
Fix inflate_threshold_root. Now=25 size=11 bits :)

This is something new and a bit surprising to me: the same threshold
in previous tests didn't generate this? Do you mean more than: "Fix inflate_threshold_root. Now=15 size=11 bits" before?

Yes. Sorry for that - this info was not all the day but only 5 minutes when i was making tests.
This info was reported only when all iBGP peers was down/up fast.

Quoted 5 lines
Quoted 4 lines

And after tune :
/sys/module/fib_trie/parameters/inflate_threshold_root no more info :)

Quoted 5 lines
Quoted 3 lines

With what value?

When i set 35 as inflate_threshold_root there was no info even if all iBGP peers was down/up.

So it looks like the patch tested earlier could be still useful; after changing the inflate_threshold_root it seems these warnings should be very rare but there is no reason to alarm users with something they can't fix optimally, anyway.

Thanks,
Jarek P.
--------------------->
ipv4: Fix inflate_threshold_root automatically

During large updates there could be triggered warnings like: "Fix inflate_threshold_root. Now=25 size=11 bits" if inflate() of the root node isn't finished in 10 loops. It should be much rarer now, after changing the threshold from 15 to 25, and a temporary problem, so this patch tries to handle it automatically using a fix variable to increase by one inflate threshold for next root resizes (up to the 35 limit, max fix = 10). The fix variable is decreased when root's inflate() finishes below 7 loops (even if some other, smaller table/ trie is updated -- for simplicity the fix variable is global for now).

---

net/ipv4/fib_trie.c +22 -7
diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
@@ -327,6 +327,8 @@ static const int inflate_threshold = 50;
 static const int halve_threshold_root = 15;
 static const int inflate_threshold_root = 25;
 
+static int inflate_threshold_root_fix;
+#define INFLATE_FIX_MAX 10	/* a comment in resize() */
 
 static void __alias_free_mem(struct rcu_head *head)
 {
@@ -617,7 +619,8 @@ static struct node *resize(struct trie *
 	/* Keep root node larger  */
 
 	if (!tn->parent)
-		inflate_threshold_use = inflate_threshold_root;
+		inflate_threshold_use = inflate_threshold_root +
+					inflate_threshold_root_fix;
 	else
 		inflate_threshold_use = inflate_threshold;
 
@@ -641,15 +644,27 @@ static struct node *resize(struct trie *
 	}
 
 	if (max_resize < 0) {
-		if (!tn->parent)
-			pr_warning("Fix inflate_threshold_root."
-				   " Now=%d size=%d bits\n",
-				   inflate_threshold_root, tn->bits);
-		else
+		if (!tn->parent) {
+			/*
+			 * It was observed that during large updates even
+			 * inflate_threshold_root = 35 might be needed to avoid
+			 * this warning; but it should be temporary, so let's
+			 * try to handle this automatically.
+			 */
+			if (inflate_threshold_root_fix < INFLATE_FIX_MAX)
+				inflate_threshold_root_fix++;
+			else
+				pr_warning("Fix inflate_threshold_root."
+					   " Now=%d size=%d bits fix=%d\n",
+					   inflate_threshold_root, tn->bits,
+					   inflate_threshold_root_fix);
+		} else {
 			pr_warning("Fix inflate_threshold."
 				   " Now=%d size=%d bits\n",
 				   inflate_threshold, tn->bits);
-	}
+		}
+	} else if (max_resize > 3 && !tn->parent && inflate_threshold_root_fix)
+		inflate_threshold_root_fix--;
 
 	check_tnode(tn);
 
#81 lore raw

Jarek Poplawski writes:

Looks good. Maybe we're getting close to some generic solution to take a very optimistic approach wrt thresholds for root node and adjust to settings without the warning. Or maybe now even remove warning totally with stata counter?

Can we even consider some other different strategy for bumping up the root node.

We need all lookup performance we can get when we now try to route without the route cache. And we probably need to evaluate the cost for the multiple lookups again at least for LOCAL and MAIN when we talking routing well at least straight-forward simple routing. (Semantic change)

I think I've got ~6.2 Gbit/s for simplex forwarding using traffic patterns we see in/close to Internet core. This w/o route cache on our hi-end opterons with 8 CPU cores using niu and ixgbe. I'll test again and your patches when I'm back from vacation.

Cheers

					--ro
Quoted 82 lines

So it looks like the patch tested earlier could be still useful; after changing the inflate_threshold_root it seems these warnings should be very rare but there is no reason to alarm users with something they can't fix optimally, anyway.

Thanks,
Jarek P.
--------------------->
ipv4: Fix inflate_threshold_root automatically

During large updates there could be triggered warnings like: "Fix inflate_threshold_root. Now=25 size=11 bits" if inflate() of the root node isn't finished in 10 loops. It should be much rarer now, after changing the threshold from 15 to 25, and a temporary problem, so this patch tries to handle it automatically using a fix variable to increase by one inflate threshold for next root resizes (up to the 35 limit, max fix = 10). The fix variable is decreased when root's inflate() finishes below 7 loops (even if some other, smaller table/ trie is updated -- for simplicity the fix variable is global for now).

---

diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c

net/ipv4/fib_trie.c +22 -7
@@ -327,6 +327,8 @@ static const int inflate_threshold = 50;
 static const int halve_threshold_root = 15;
 static const int inflate_threshold_root = 25;
 
+static int inflate_threshold_root_fix;
+#define INFLATE_FIX_MAX 10	/* a comment in resize() */
 
 static void __alias_free_mem(struct rcu_head *head)
 {
@@ -617,7 +619,8 @@ static struct node *resize(struct trie *
 	/* Keep root node larger  */
 
 	if (!tn->parent)
-		inflate_threshold_use = inflate_threshold_root;
+		inflate_threshold_use = inflate_threshold_root +
+					inflate_threshold_root_fix;
 	else
 		inflate_threshold_use = inflate_threshold;
 
@@ -641,15 +644,27 @@ static struct node *resize(struct trie *
 	}
 
 	if (max_resize < 0) {
-		if (!tn->parent)
-			pr_warning("Fix inflate_threshold_root."
-				   " Now=%d size=%d bits\n",
-				   inflate_threshold_root, tn->bits);
-		else
+		if (!tn->parent) {
+			/*
+			 * It was observed that during large updates even
+			 * inflate_threshold_root = 35 might be needed to avoid
+			 * this warning; but it should be temporary, so let's
+			 * try to handle this automatically.
+			 */
+			if (inflate_threshold_root_fix < INFLATE_FIX_MAX)
+				inflate_threshold_root_fix++;
+			else
+				pr_warning("Fix inflate_threshold_root."
+					   " Now=%d size=%d bits fix=%d\n",
+					   inflate_threshold_root, tn->bits,
+					   inflate_threshold_root_fix);
+		} else {
 			pr_warning("Fix inflate_threshold."
 				   " Now=%d size=%d bits\n",
 				   inflate_threshold, tn->bits);
-	}
+		}
+	} else if (max_resize > 3 && !tn->parent && inflate_threshold_root_fix)
+		inflate_threshold_root_fix--;
 
 	check_tnode(tn);
 
#82 lore raw
Quoted 8 lines · Robert Olsson

Jarek Poplawski writes:

Looks good. Maybe we're getting close to some generic solution to take a very optimistic approach wrt thresholds for root node and adjust to settings without the warning. Or maybe now even remove warning totally with stata counter?

I guess, we could, but maybe let's wait a bit to make sure there is nothing surprising?

Quoted 9 lines

Can we even consider some other different strategy for bumping up the root node.

We need all lookup performance we can get when we now try to route without the route cache. And we probably need to evaluate the cost for the multiple lookups again at least for LOCAL and MAIN when we talking routing well at least straight-forward simple routing. (Semantic change)

Quoted 5 lines

I think I've got ~6.2 Gbit/s for simplex forwarding using traffic patterns we see in/close to Internet core. This w/o route cache on our hi-end opterons with 8 CPU cores using niu and ixgbe. I'll test again and your patches when I'm back from vacation.

Sure, I was mainly aiming at safe defaults (wrt. memory usage), but if tests show there is a better strategy we should go for it.

Thanks,
Jarek P.

#83 lore raw
Quoted 8 lines · Jarek Poplawski
Quoted 3 lines · Robert Olsson

a very optimistic approach wrt thresholds for root node and adjust to settings without the warning. Or maybe now even remove warning totally with stata counter?

I guess, we could, but maybe let's wait a bit to make sure there is nothing surprising?

Yes if Pawel is running it we we'll get reports. I've no chance to upgrade any of our routers now. I've seen this printout in one our routers but we don't do "clear ip bgp *" to often and besides we try to use soft re- configuration inbound.

Quoted 8 lines
Quoted 5 lines

I think I've got ~6.2 Gbit/s for simplex forwarding using traffic patterns we see in/close to Internet core. This w/o route cache on our hi-end opterons with 8 CPU cores using niu and ixgbe. I'll test again and your patches when I'm back from vacation.

Sure, I was mainly aiming at safe defaults (wrt. memory usage), but if tests show there is a better strategy we should go for it.

Routing without route cache is "new" area probably for minority of systems were caching is not possible. Read BGP routers in core.

Yes we should have safe defults. Thanks for all your work.

Cheers

					--ro
maintainer Applied
#84 lore raw
In reply to Jarek Poplawski

From: Jarek Poplawski <[email protected]>
Date: Tue, 14 Jul 2009 21:41:00 +0200

Quoted 12 lines

ipv4: Fix inflate_threshold_root automatically

During large updates there could be triggered warnings like: "Fix inflate_threshold_root. Now=25 size=11 bits" if inflate() of the root node isn't finished in 10 loops. It should be much rarer now, after changing the threshold from 15 to 25, and a temporary problem, so this patch tries to handle it automatically using a fix variable to increase by one inflate threshold for next root resizes (up to the 35 limit, max fix = 10). The fix variable is decreased when root's inflate() finishes below 7 loops (even if some other, smaller table/ trie is updated -- for simplicity the fix variable is global for now).

Quoted 4 lines

Applied.

#85 lore raw
In reply to Jarek Poplawski

Jarek Poplawski pisze:

Quoted 18 lines · Paweł Staszewski

...

Quoted 7 lines

So i make tests with changing sync_pages
And

####################################
sync_pages: 64
total size reach maximum in 17sec

...

Quoted 4 lines

######################################
sync_pages: 128
Fib trie Total size reach max in 14sec

...

Quoted 8 lines
Quoted 4 lines

#########################################
sync_pages: 256
hmm no difference also in 10sec

14 == 10!? ;-)
...

:) i miss one test

Quoted 11 lines
Quoted 2 lines

And with sync_pages higher that 256 time of filling kernel routes is the same approx 10sec.

Hmm... So, it's better than I expected; syncing after 128 or 256 pages could be quite reasonable. But then it would be interesting to find out if with such a safety we could go back to more aggressive values for possibly better performance. So here is 'the same' patch (so the previous, take 8, should be reverted), but with additional possibility to change:
/sys/module/fib_trie/parameters/inflate_threshold_root

Quoted 7 lines

I guess, you could try e.g. if: sync_pages 256, inflate_threshold_root 15 can give faster lookups (or lower cpu loads); with this these inflate warnings could be back btw.; or maybe you'll find something in between like inflate_threshold_root 20 is optimal for you. (I think it should be enough to try this only for PREEMPT_NONE unless you have spare time ;-)

And i can't make good tests with cpu load because of problem that i have from "weird problem" emails
It depend when i make mpstat to check cpu load and for what time because every 15 sec i have 1 do 3 % of cpu and after 15 sec i have almost 40% cpu load for next 15 sec. I try to make mpstat -P ALL 1 60
but after 15 sec of 1 to 3 % cpu load this next higher cpu load if different everytime it balance from 30 to 50%

so i make test shorter when cpu load is 1 to 3 % - "mpstat -P ALL 1 10" output in attached file

Regards
Paweł Staszewski

Quoted 67 lines

Thanks,
Jarek P.
---> (synchronize take 9; apply on top of the 2.6.29.x with the last
all-in-one patch, or net-2.6)

 net/ipv4/fib_trie.c |   18 ++++++++++++++++--
 1 files changed, 16 insertions(+), 2 deletions(-)
net/ipv4/fib_trie.c +16 -3
index 00a54b2..e8fca11 100644
@@ -71,6 +71,7 @@
 #include <linux/netlink.h>
 #include <linux/init.h>
 #include <linux/list.h>
+#include <linux/moduleparam.h>
 #include <net/net_namespace.h>
 #include <net/ip.h>
 #include <net/protocol.h>
@@ -164,6 +165,10 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
+
+static int sync_pages __read_mostly = 1000;
+module_param(sync_pages, int, 0640);
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -316,9 +321,11 @@ static inline void check_tnode(const struct tnode *tn)
 
 static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
-static const int halve_threshold_root = 15;
-static const int inflate_threshold_root = 25;
 
+static int inflate_threshold_root __read_mostly = 25;
+module_param(inflate_threshold_root, int, 0640);
+
+#define halve_threshold_root	(inflate_threshold_root / 2 + 1)
 
 static void __alias_free_mem(struct rcu_head *head)
 {
@@ -393,6 +400,8 @@ static void tnode_free_safe(struct tnode *tn)
 	BUG_ON(IS_LEAF(tn));
 	tn->tnode_free = tnode_free_head;
 	tnode_free_head = tn;
+	tnode_free_size += sizeof(struct tnode) +
+			   (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +413,11 @@ static void tnode_free_flush(void)
 		tn->tnode_free = NULL;
 		tnode_free(tn);
 	}
+
+	if (tnode_free_size >= PAGE_SIZE * sync_pages) {
+		tnode_free_size = 0;
+		synchronize_rcu();
+	}
 }
 
 static struct leaf *leaf_new(void)
--
Quoted 6 lines

To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

sync_pages: 256
inflate_threshold_root: 10
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.00    0.00    0.00    0.00    0.00    0.60    0.00    0.00   99.40
Average:       0    0.00    0.00    0.00    0.00    0.00    0.60    0.00    0.00   99.40
Average:       1    0.00    0.00    0.00    0.00    0.00    0.40    0.00    0.00   99.60
sync_pages: 256
inflate_threshold_root: 15
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.00    0.00    0.10    0.00    0.00    0.70    0.00    0.00   99.20
Average:       0    0.00    0.00    0.00    0.00    0.20    0.80    0.00    0.00   99.00
Average:       1    0.00    0.00    0.20    0.00    0.00    0.61    0.00    0.00   99.19
sync_pages: 256
inflate_threshold_root: 20
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.00    0.00    0.00    0.00    0.10    0.80    0.00    0.00   99.10
Average:       0    0.00    0.00    0.00    0.00    0.00    1.00    0.00    0.00   99.00
Average:       1    0.00    0.00    0.00    0.00    0.00    0.61    0.00    0.00   99.39
sync_pages: 256
inflate_threshold_root: 25
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.00    0.00    0.00    0.00    0.00    0.70    0.00    0.00   99.30
Average:       0    0.00    0.00    0.00    0.00    0.20    1.00    0.00    0.00   98.80
Average:       1    0.00    0.00    0.00    0.00    0.00    0.40    0.00    0.00   99.60
sync_pages: 512
inflate_threshold_root: 10
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.00    0.00    0.10    0.00    0.10    0.60    0.00    0.00   99.20
Average:       0    0.00    0.00    0.20    0.00    0.00    1.00    0.00    0.00   98.80
Average:       1    0.00    0.00    0.00    0.00    0.00    0.40    0.00    0.00   99.60
sync_pages: 512
inflate_threshold_root: 15
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.00    0.00    0.20    0.00    0.00    1.10    0.00    0.00   98.70
Average:       0    0.00    0.00    0.40    0.00    0.00    1.00    0.00    0.00   98.60
Average:       1    0.00    0.00    0.00    0.00    0.00    1.01    0.00    0.00   98.99
sync_pages: 512
inflate_threshold_root: 20
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.00    0.00    0.10    0.00    0.10    1.01    0.00    0.00   98.79
Average:       0    0.00    0.00    0.20    0.00    0.20    1.40    0.00    0.00   98.20
Average:       1    0.00    0.00    0.00    0.00    0.00    0.61    0.00    0.00   99.39
sync_pages: 512
inflate_threshold_root: 25
Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.00    0.00    0.00    0.00    0.10    0.90    0.00    0.00   99.00
Average:       0    0.00    0.00    0.00    0.00    0.00    1.00    0.00    0.00   99.00
Average:       1    0.00    0.00    0.00    0.00    0.20    0.80    0.00    0.00   99.00
#86 lore raw

Paweł Staszewski pisze:

Quoted 51 lines

Jarek Poplawski pisze:

Quoted 18 lines · Paweł Staszewski

...

Quoted 7 lines

So i make tests with changing sync_pages
And

####################################
sync_pages: 64
total size reach maximum in 17sec

...

Quoted 4 lines

######################################
sync_pages: 128
Fib trie Total size reach max in 14sec

...

Quoted 8 lines
Quoted 4 lines

#########################################
sync_pages: 256
hmm no difference also in 10sec

14 == 10!? ;-)
...

:) i miss one test

Quoted 11 lines
Quoted 2 lines

And with sync_pages higher that 256 time of filling kernel routes is the same approx 10sec.

Hmm... So, it's better than I expected; syncing after 128 or 256 pages could be quite reasonable. But then it would be interesting to find out if with such a safety we could go back to more aggressive values for possibly better performance. So here is 'the same' patch (so the previous, take 8, should be reverted), but with additional possibility to change:
/sys/module/fib_trie/parameters/inflate_threshold_root

Quoted 7 lines

I guess, you could try e.g. if: sync_pages 256, inflate_threshold_root 15 can give faster lookups (or lower cpu loads); with this these inflate warnings could be back btw.; or maybe you'll find something in between like inflate_threshold_root 20 is optimal for you. (I think it should be enough to try this only for PREEMPT_NONE unless you have spare time ;-)

And i can't make good tests with cpu load because of problem that i have from "weird problem" emails
It depend when i make mpstat to check cpu load and for what time because every 15 sec i have 1 do 3 % of cpu and after 15 sec i have almost 40% cpu load for next 15 sec. I try to make mpstat -P ALL 1 60
but after 15 sec of 1 to 3 % cpu load this next higher cpu load if different everytime it balance from 30 to 50%

Quoted 6 lines

so i make test shorter when cpu load is 1 to 3 % - "mpstat -P ALL 1 10" output in attached file

Regards
Paweł Staszewski

i forgot to add:
Traffic when i make test was +/- 10Mbit/s in next tests:
eth0: RX: 231.21 Mb/s TX: 287.40 Mb/s eth1: RX: 289.19 Mb/s TX: 231.35 Mb/s

Quoted 77 lines
Quoted 76 lines

Thanks,
Jarek P.
---> (synchronize take 9; apply on top of the 2.6.29.x with the last

      all-in-one patch, or net-2.6)
 net/ipv4/fib_trie.c |   18 ++++++++++++++++--
 1 files changed, 16 insertions(+), 2 deletions(-)
net/ipv4/fib_trie.c +5 -0
index 00a54b2..e8fca11 100644
@@ -71,6 +71,7 @@
 #include <linux/netlink.h>
 #include <linux/init.h>
 #include <linux/list.h>
+#include <linux/moduleparam.h>
 #include <net/net_namespace.h>
 #include <net/ip.h>
 #include <net/protocol.h>
@@ -164,6 +165,10 @@ static struct tnode *inflate(struct trie *t, 
 struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
+
+static int sync_pages __read_mostly = 1000;
+module_param(sync_pages, int, 0640);
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;

static struct kmem_cache *trie_leaf_kmem __read_mostly;

net/ipv4/fib_trie.c +4 -2
@@ -316,9 +321,11 @@ static inline void check_tnode(const struct 
 tnode *tn)
 
 static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
-static const int halve_threshold_root = 15;
-static const int inflate_threshold_root = 25;
 
+static int inflate_threshold_root __read_mostly = 25;
+module_param(inflate_threshold_root, int, 0640);
+
+#define halve_threshold_root    (inflate_threshold_root / 2 + 1)
 
 static void __alias_free_mem(struct rcu_head *head)

{

net/ipv4/fib_trie.c +7 -0
@@ -393,6 +400,8 @@ static void tnode_free_safe(struct tnode *tn)
     BUG_ON(IS_LEAF(tn));
     tn->tnode_free = tnode_free_head;
     tnode_free_head = tn;
+    tnode_free_size += sizeof(struct tnode) +
+               (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +413,11 @@ static void tnode_free_flush(void)
         tn->tnode_free = NULL;
         tnode_free(tn);
     }
+
+    if (tnode_free_size >= PAGE_SIZE * sync_pages) {
+        tnode_free_size = 0;
+        synchronize_rcu();
+    }
 }
 
 static struct leaf *leaf_new(void)
Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Tested-by Pawel Staszewski
#87 lore raw
In reply to Paul E. McKenney
Quoted 16 lines · Paul E. McKenney
Quoted 15 lines · Jarek Poplawski
Quoted 9 lines · Jarek Poplawski
Quoted 4 lines · Paweł Staszewski

Oh

I forgot - please Jarek give me patch with sync rcu and i will make test on preempt kernel

Probably non-preempt kernel might need something like this more, but comparing is always interesting. This patch is based on Paul's suggestion (I hope).

Hold on ;-) Here is something even better... Syncing after 128 pages might be still too slow, so here is a higher initial value, 1000, plus you can change this while testing in:

Quoted 7 lines
Quoted 3 lines

/sys/module/fib_trie/parameters/sync_pages

It would be interesting to find the lowest acceptable value.

Looks like a promising approach to me!

Thanx, Paul

Below is a simpler version of this patch, without the sysfs parameter. (I left the previous version quoted for comparison.) Thanks.

Quoted 50 lines
Quoted 50 lines

Jarek P.
---> (synchronize take 8; apply on top of the 2.6.29.x with the last

(quoted diff) +0 -0
 	all-in-one patch, or net-2.6)
 
 net/ipv4/fib_trie.c |   12 ++++++++++++
 1 files changed, 12 insertions(+), 0 deletions(-)
net/ipv4/fib_trie.c +12 -0
index 00a54b2..decc8d0 100644
@@ -71,6 +71,7 @@
 #include <linux/netlink.h>
 #include <linux/init.h>
 #include <linux/list.h>
+#include <linux/moduleparam.h>
 #include <net/net_namespace.h>
 #include <net/ip.h>
 #include <net/protocol.h>
@@ -164,6 +165,10 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
+
+static int sync_pages __read_mostly = 1000;
+module_param(sync_pages, int, 0640);
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -393,6 +398,8 @@ static void tnode_free_safe(struct tnode *tn)
 	BUG_ON(IS_LEAF(tn));
 	tn->tnode_free = tnode_free_head;
 	tnode_free_head = tn;
+	tnode_free_size += sizeof(struct tnode) +
+			   (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +411,11 @@ static void tnode_free_flush(void)
 		tn->tnode_free = NULL;
 		tnode_free(tn);
 	}
+
+	if (tnode_free_size >= PAGE_SIZE * sync_pages) {
+		tnode_free_size = 0;
+		synchronize_rcu();
+	}
 }
 
Quoted 2 lines

net/ipv4/fib_trie.c static void tnode_free_flush(void)

Quoted 2 lines
net/ipv4/fib_trie.c +0 -1
 static struct leaf *leaf_new(void)
--

------------------------>
ipv4: Use synchronize_rcu() during trie_rebalance()

During trie_rebalance() we free memory after resizing with call_rcu(), but large updates, especially with PREEMPT_NONE configs, can cause memory stresses, so this patch calls synchronize_rcu() in tnode_free_flush() after each sync_pages to guarantee such freeing (especially before resizing the root node).

The value of sync_pages = 128 is based on Pawel Staszewski's tests as the lowest which doesn't hinder updating times. (For testing purposes there was a sysfs module parameter to change it on demand, but it's removed until we're sure it could be really useful.)

The patch is based on suggestions by: Paul E. McKenney <[email protected]>

---

 net/ipv4/fib_trie.c |   15 +++++++++++++++
 1 files changed, 15 insertions(+), 0 deletions(-)
net/ipv4/fib_trie.c +15 -0
index 63c2fa7..58ba9f4 100644
@@ -164,6 +164,14 @@ static struct tnode *inflate(struct trie *t, struct tnode *tn);
 static struct tnode *halve(struct trie *t, struct tnode *tn);
 /* tnodes to free after resize(); protected by RTNL */
 static struct tnode *tnode_free_head;
+static size_t tnode_free_size;
+
+/*
+ * synchronize_rcu after call_rcu for that many pages; it should be especially
+ * useful before resizing the root node with PREEMPT_NONE configs; the value was
+ * obtained experimentally, aiming to avoid visible slowdown.
+ */
+static const int sync_pages = 128;
 
 static struct kmem_cache *fn_alias_kmem __read_mostly;
 static struct kmem_cache *trie_leaf_kmem __read_mostly;
@@ -393,6 +401,8 @@ static void tnode_free_safe(struct tnode *tn)
 	BUG_ON(IS_LEAF(tn));
 	tn->tnode_free = tnode_free_head;
 	tnode_free_head = tn;
+	tnode_free_size += sizeof(struct tnode) +
+			   (sizeof(struct node *) << tn->bits);
 }
 
 static void tnode_free_flush(void)
@@ -404,6 +414,11 @@ static void tnode_free_flush(void)
 		tn->tnode_free = NULL;
 		tnode_free(tn);
 	}
+
+	if (tnode_free_size >= PAGE_SIZE * sync_pages) {
+		tnode_free_size = 0;
+		synchronize_rcu();
+	}
 }
 
 static struct leaf *leaf_new(void)
maintainer Applied
#88 lore raw

From: Jarek Poplawski <[email protected]>
Date: Tue, 14 Jul 2009 20:33:08 +0200

Quoted 13 lines

ipv4: Use synchronize_rcu() during trie_rebalance()

During trie_rebalance() we free memory after resizing with call_rcu(), but large updates, especially with PREEMPT_NONE configs, can cause memory stresses, so this patch calls synchronize_rcu() in tnode_free_flush() after each sync_pages to guarantee such freeing (especially before resizing the root node).

The value of sync_pages = 128 is based on Pawel Staszewski's tests as the lowest which doesn't hinder updating times. (For testing purposes there was a sysfs module parameter to change it on demand, but it's removed until we're sure it could be really useful.)

Quoted 6 lines

The patch is based on suggestions by: Paul E. McKenney <[email protected]>

Applied.

#89 lore raw

PATCH 1/1 ipv4: fib_trie: Use tnode_get_child_rcu() and node_parent_rcu() in lookups

In reply to Paul E. McKenney

While looking for other fib_trie problems reported by Pawel Staszewski I noticed there are a few uses of tnode_get_child() and node_parent() in lookups instead of their rcu versions.

---
(this patch was prepared on top of my 2 today's fib_trie patches)

net/ipv4/fib_trie.c +3 -3
diff -Nurp a/net/ipv4/fib_trie.c b/net/ipv4/fib_trie.c
@@ -1465,7 +1465,7 @@ static int fn_trie_lookup(struct fib_tab
 			cindex = tkey_extract_bits(mask_pfx(key, current_prefix_length),
 						   pos, bits);
 
-		n = tnode_get_child(pn, cindex);
+		n = tnode_get_child_rcu(pn, cindex);
 
 		if (n == NULL) {
 #ifdef CONFIG_IP_FIB_TRIE_STATS
@@ -1600,7 +1600,7 @@ backtrace:
 		if (chopped_off <= pn->bits) {
 			cindex &= ~(1 << (chopped_off-1));
 		} else {
-			struct tnode *parent = node_parent((struct node *) pn);
+			struct tnode *parent = node_parent_rcu((struct node *) pn);
 			if (!parent)
 				goto failed;
 
@@ -1813,7 +1813,7 @@ static struct leaf *trie_firstleaf(struc
 static struct leaf *trie_nextleaf(struct leaf *l)
 {
 	struct node *c = (struct node *) l;
-	struct tnode *p = node_parent(c);
+	struct tnode *p = node_parent_rcu(c);
 
 	if (!p)
 		return NULL;	/* trie with just one leaf */
maintainer Applied
#90 lore raw

Re: [PATCH net-next] ipv4: fib_trie: Use tnode_get_child_rcu() and node_parent_rcu() in lookups

From: Jarek Poplawski <[email protected]>
Date: Tue, 14 Jul 2009 23:20:32 +0200

Quoted 6 lines

While looking for other fib_trie problems reported by Pawel Staszewski I noticed there are a few uses of tnode_get_child() and node_parent() in lookups instead of their rcu versions.

Applied.

#91 lore raw
In reply to Paweł Staszewski

Sorry again no attachement.

Paweł Staszewski pisze:

Quoted 64 lines

Jarek Poplawski pisze:

Quoted 10 lines · Paweł Staszewski
Quoted 9 lines

Jarek Poplawski pisze:

Quoted 6 lines · Paweł Staszewski
Quoted 3 lines

Jarek Poplawski pisze:

...

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

So, after your findings I'm about to recommend sending to -stable 3 patches from net-2.6, with additional lowering of threshold_root settings, but it would be nice if you could give it a try with CONFIG_PREEMPT instead of CONFIG_PREEMPT_NONE (if it doesn't break your other apps!) It is expected to work this time...;-) Maybe a bit slower.

Ok kernel configured with CONFIG_PREEMPT
and all this day work without any problems (with Jarek last patch).

So in attached file trere is fib_tirestats
I dont see any big change of (cpu load or faster/slower routing/propagating routes from bgpd or something else) - in avg there is from 2% to 3% more of CPU load i dont know why but it is - i change from "preempt" to "no preempt" 3 times and check this my "mpstat -P ALL 1 30"
always avg cpu load was from 2 to 3% more compared to "no preempt"

Regards
Paweł Staszewski

Quoted 32 lines
Quoted 11 lines
Quoted 8 lines
Quoted 4 lines
Quoted 1 line

Patch applied to 2.6.29.5 with CONFIG_PREEMPT_NONE And working :)

Hmm... It should, because you tested very similar patch already;-) Sorry if I didn't make it clear.

Yes i know there was almost identical one.
And i see this was without sync rcu :)

Yes, it looks like we can't free memory so simple because of such huge latencies.

Quoted 16 lines
Quoted 8 lines
Quoted 4 lines

fib_triestats in attached file

I think I can test it with PREEMPT enabled but first i must make some other tests of my apps that are on server.

It could probably matter only if you're using some broken out-of-tree patches. Otherwise the kernel is expected to work OK.

Im a little confused about using of PREEMPT kernel because of past there was many oopses / lockups :) but yes that was a little long time ago. I will try to make this test today.

Quoted 3 lines

Btw., it would be also interesting to check if there is any difference wrt. these route cache problems while PREEMPT is enabled.

Quoted 15 lines
Quoted 8 lines

And you're very right! The place we're fixing is the best example. On the other hand, I hope there is not many such places yet. But if we test/fix it there will be one less...

Jarek P.

Signature · 3 lines
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

        Aver depth:     2.44
        Max depth:      6
        Leaves:         277814
        Prefixes:       291306
        Internal nodes: 66420
          1: 32737  2: 14850  3: 10332  4: 4871  5: 2313  6: 942  7: 371  8: 3  17: 1
        Pointers: 599098

Null ptrs: 254865
Total size: 18067 kB

Counters:
---------
gets = 2003686
backtracks = 78789 semantic match passed = 1977687 semantic match miss = 112 null node hit= 1470619 skipped node resize = 0

Local:

        Aver depth:     3.75
        Max depth:      5
        Leaves:         12
        Prefixes:       13

Internal nodes: 10
1: 9 2: 1
Pointers: 22
Null ptrs: 1
Total size: 2 kB

Counters:
---------
gets = 2008497
backtracks = 1417179 semantic match passed = 4823 semantic match miss = 0 null node hit= 197044 skipped node resize = 0

Tested-by Pawel Staszewski Tested-by Pawel Staszewski
#92 lore raw
In reply to Paweł Staszewski

David & Robert,
below are my recommendations for -stable plus one more patch:

On Sun, Jul 05, 2009 at 02:26:54AM +0200, Paweł Staszewski wrote:
...

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

Jarek Poplawski pisze:

...

Quoted 20 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

So, after your findings I'm about to recommend sending to -stable 3 patches from net-2.6, with additional lowering of threshold_root settings, but it would be nice if you could give it a try with CONFIG_PREEMPT instead of CONFIG_PREEMPT_NONE (if it doesn't break your other apps!) It is expected to work this time...;-) Maybe a bit slower.

Ok kernel configured with CONFIG_PREEMPT
and all this day work without any problems (with Jarek last patch).

So in attached file trere is fib_tirestats
I dont see any big change of (cpu load or faster/slower routing/propagating routes from bgpd or something else) - in avg there is from 2% to 3% more of CPU load i dont know why but it is - i change from "preempt" to "no preempt" 3 times and check this my "mpstat -P ALL 1 30"
always avg cpu load was from 2 to 3% more compared to "no preempt"

Quoted 2 lines

Regards
Paweł Staszewski

So after these patches from net-2.6 are tested both for PREEMPT and PREEMPT_NONE I think they should go to -stable:

2.6.30 needs:
-------------

commit e0f7cb8c8cc6cccce28d2ce39ad8c60d23c3799f

Author: Jarek Poplawski <[email protected]>
Date:   Mon Jun 15 02:31:29 2009 -0700

ipv4: Fix fib_trie rebalancing

commit 7b85576d15bf2574b0a451108f59f9ad4170dd3f

Author: Jarek Poplawski <[email protected]>
Date:   Thu Jun 18 00:28:51 2009 -0700

ipv4: Fix fib_trie rebalancing, part 2

commit 008440e3ad4b72f5048d1b1f6f5ed894fdc5ad08

Author: Jarek Poplawski <[email protected]>
Date:   Tue Jun 30 12:47:19 2009 -0700

ipv4: Fix fib_trie rebalancing, part 3

plus the new patch below

ipv4: Fix fib_trie rebalancing, part 4 (root thresholds)

2.6.29 needs:
-------------

this patch from 2.6.30:
commit 3ed18d76d959e5cbfa5d70c8f7ba95476582a556

Author: Robert Olsson <[email protected]>
Date:   Thu May 21 15:20:59 2009 -0700

ipv4: Fix oops with FIB_TRIE

plus above mentionned patches for 2.6.30 (part 1 - 4)

-----------------

David, if possible, please add to all these "Fix... part 1 - 4":

This new patch below is intended only for -stable (and later for net-next), because it doesn't meet rules of the current -rc. Anyway, it's not critical (but it actually fixes a regression from 2.6.22).

Thanks,
Jarek P.
---------------->
ipv4: Fix fib_trie rebalancing, part 4 (root thresholds)

Pawel Staszewski wrote:
<blockquote>
Some time ago i report this:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

and now with 2.6.29 / 2.6.29.1 / 2.6.29.3 and 2.6.30 it back dmesg output:
oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits
...
Fix inflate_threshold_root. Now=15 size=11 bits

cat /proc/net/fib_triestat
Basic info: size of leaf: 40 bytes, size of tnode: 56 bytes. Main:

        Aver depth:     2.28
        Max depth:      6
        Leaves:         276539
        Prefixes:       289922
        Internal nodes: 66762
          1: 35046  2: 13824  3: 9508  4: 4897  5: 2331  6: 1149  7: 5  
9: 1  18: 1
        Pointers: 691228

Null ptrs: 347928
Total size: 35709 kB
</blockquote>

It seems, the current threshold for root resizing is too aggressive, and it causes misleading warnings during big updates, but it might be also responsible for memory problems, especially with non-preempt configs, when RCU freeing is delayed long after call_rcu.

It should be also mentionned that because of non-atomic changes during resizing/rebalancing the current lookup algorithm can miss valid leafs so it's additional argument to shorten these activities even at a cost of a minimally longer searching.

This patch restores values before the patch "[IPV4]: fib_trie root node settings", commit: 965ffea43d4ebe8cd7b9fee78d651268dd7d23c5 from v2.6.22.

Pawel's report:
<blockquote>
I dont see any big change of (cpu load or faster/slower routing/propagating routes from bgpd or something else) - in avg there is from 2% to 3% more of CPU load i dont know why but it is - i change from "preempt" to "no preempt" 3 times and check this my "mpstat -P ALL 1 30"
always avg cpu load was from 2 to 3% more compared to "no preempt" [...]
cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

        Aver depth:     2.44
        Max depth:      6
        Leaves:         277814
        Prefixes:       291306
        Internal nodes: 66420
          1: 32737  2: 14850  3: 10332  4: 4871  5: 2313  6: 942  7: 371  8: 3  17: 1
        Pointers: 599098

Null ptrs: 254865
Total size: 18067 kB
</blockquote>

According to this and other similar reports average depth is slightly increased (~0.2), and root nodes are shorter (log 17 vs. 18), but there is no visible performance decrease. So, until memory handling is improved or added parameters for changing this individually, this patch resets to safer defaults.

---

 net/ipv4/fib_trie.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
net/ipv4/fib_trie.c +2 -2
index 00a54b2..63c2fa7 100644
@@ -316,8 +316,8 @@ static inline void check_tnode(const struct tnode *tn)
 
 static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
-static const int halve_threshold_root = 8;
-static const int inflate_threshold_root = 15;
+static const int halve_threshold_root = 15;
+static const int inflate_threshold_root = 25;
 
 
 static void __alias_free_mem(struct rcu_head *head)
Tested-by Pawel Staszewski Tested-by Pawel Staszewski
#93 lore raw
In reply to Paweł Staszewski

(Take 2: Changelog spelling fixes, sorry.)

David & Robert,
below are my recommendations for -stable plus one more patch:

On Sun, Jul 05, 2009 at 02:26:54AM +0200, Paweł Staszewski wrote:
...

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

Jarek Poplawski pisze:

...

Quoted 20 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

So, after your findings I'm about to recommend sending to -stable 3 patches from net-2.6, with additional lowering of threshold_root settings, but it would be nice if you could give it a try with CONFIG_PREEMPT instead of CONFIG_PREEMPT_NONE (if it doesn't break your other apps!) It is expected to work this time...;-) Maybe a bit slower.

Ok kernel configured with CONFIG_PREEMPT
and all this day work without any problems (with Jarek last patch).

So in attached file trere is fib_tirestats
I dont see any big change of (cpu load or faster/slower routing/propagating routes from bgpd or something else) - in avg there is from 2% to 3% more of CPU load i dont know why but it is - i change from "preempt" to "no preempt" 3 times and check this my "mpstat -P ALL 1 30"
always avg cpu load was from 2 to 3% more compared to "no preempt"

Quoted 2 lines

Regards
Paweł Staszewski

So after these patches from net-2.6 are tested both for PREEMPT and PREEMPT_NONE I think they should go to -stable:

2.6.30 needs:
-------------

commit e0f7cb8c8cc6cccce28d2ce39ad8c60d23c3799f

Author: Jarek Poplawski <[email protected]>
Date:   Mon Jun 15 02:31:29 2009 -0700

ipv4: Fix fib_trie rebalancing

commit 7b85576d15bf2574b0a451108f59f9ad4170dd3f

Author: Jarek Poplawski <[email protected]>
Date:   Thu Jun 18 00:28:51 2009 -0700

ipv4: Fix fib_trie rebalancing, part 2

commit 008440e3ad4b72f5048d1b1f6f5ed894fdc5ad08

Author: Jarek Poplawski <[email protected]>
Date:   Tue Jun 30 12:47:19 2009 -0700

ipv4: Fix fib_trie rebalancing, part 3

plus the new patch below

ipv4: Fix fib_trie rebalancing, part 4 (root thresholds)

2.6.29 needs:
-------------

this patch from 2.6.30:
commit 3ed18d76d959e5cbfa5d70c8f7ba95476582a556

Author: Robert Olsson <[email protected]>
Date:   Thu May 21 15:20:59 2009 -0700

ipv4: Fix oops with FIB_TRIE

plus above mentionned patches for 2.6.30 (part 1 - 4)

-----------------

David, if possible, please add to all these "Fix... part 1 - 4":

This new patch below is intended only for -stable (and later for net-next), because it doesn't meet rules of the current -rc. Anyway, it's not critical (but it actually fixes a regression from 2.6.22).

Thanks,
Jarek P.
---------------->
ipv4: Fix fib_trie rebalancing, part 4 (root thresholds)

Pawel Staszewski wrote:
<blockquote>
Some time ago i report this:
http://bugzilla.kernel.org/show_bug.cgi?id=6648

and now with 2.6.29 / 2.6.29.1 / 2.6.29.3 and 2.6.30 it back dmesg output:
oprofile: using NMI interrupt.
Fix inflate_threshold_root. Now=15 size=11 bits
...
Fix inflate_threshold_root. Now=15 size=11 bits

cat /proc/net/fib_triestat
Basic info: size of leaf: 40 bytes, size of tnode: 56 bytes. Main:

        Aver depth:     2.28
        Max depth:      6
        Leaves:         276539
        Prefixes:       289922
        Internal nodes: 66762
          1: 35046  2: 13824  3: 9508  4: 4897  5: 2331  6: 1149  7: 5  
9: 1  18: 1
        Pointers: 691228

Null ptrs: 347928
Total size: 35709 kB
</blockquote>

It seems, the current threshold for root resizing is too aggressive, and it causes misleading warnings during big updates, but it might be also responsible for memory problems, especially with non-preempt configs, when RCU freeing is delayed long after call_rcu.

It should be also mentioned that because of non-atomic changes during resizing/rebalancing the current lookup algorithm can miss valid leaves so it's additional argument to shorten these activities even at a cost of a minimally longer searching.

This patch restores values before the patch "[IPV4]: fib_trie root node settings", commit: 965ffea43d4ebe8cd7b9fee78d651268dd7d23c5 from v2.6.22.

Pawel's report:
<blockquote>
I dont see any big change of (cpu load or faster/slower routing/propagating routes from bgpd or something else) - in avg there is from 2% to 3% more of CPU load i dont know why but it is - i change from "preempt" to "no preempt" 3 times and check this my "mpstat -P ALL 1 30"
always avg cpu load was from 2 to 3% more compared to "no preempt" [...]
cat /proc/net/fib_triestat
Basic info: size of leaf: 20 bytes, size of tnode: 36 bytes. Main:

        Aver depth:     2.44
        Max depth:      6
        Leaves:         277814
        Prefixes:       291306
        Internal nodes: 66420
          1: 32737  2: 14850  3: 10332  4: 4871  5: 2313  6: 942  7: 371  8: 3  17: 1
        Pointers: 599098

Null ptrs: 254865
Total size: 18067 kB
</blockquote>

According to this and other similar reports average depth is slightly increased (~0.2), and root nodes are shorter (log 17 vs. 18), but there is no visible performance decrease. So, until memory handling is improved or added parameters for changing this individually, this patch resets to safer defaults.

---

 net/ipv4/fib_trie.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
net/ipv4/fib_trie.c +2 -2
index 00a54b2..63c2fa7 100644
@@ -316,8 +316,8 @@ static inline void check_tnode(const struct tnode *tn)
 
 static const int halve_threshold = 25;
 static const int inflate_threshold = 50;
-static const int halve_threshold_root = 8;
-static const int inflate_threshold_root = 15;
+static const int halve_threshold_root = 15;
+static const int inflate_threshold_root = 25;
 
 
 static void __alias_free_mem(struct rcu_head *head)
maintainer
#94 lore raw

From: Jarek Poplawski <[email protected]>
Date: Sun, 5 Jul 2009 15:08:28 +0200

Quoted 3 lines

This new patch below is intended only for -stable (and later for net-next), because it doesn't meet rules of the current -rc. Anyway, it's not critical (but it actually fixes a regression from 2.6.22).

I think if we' re going to toss this into -stable, we should put it into net-2.6 too, and that's what I'm going to do.

Once this makes it's way to Linus I'll work on the -stable submissions.

And I'll make sure to add the tested-by tags, as you mentioned.

Thanks!

#95 lore raw
Quoted 9 lines · David Miller

From: Jarek Poplawski <[email protected]>
Date: Sun, 5 Jul 2009 15:08:28 +0200

Quoted 3 lines

This new patch below is intended only for -stable (and later for net-next), because it doesn't meet rules of the current -rc. Anyway, it's not critical (but it actually fixes a regression from 2.6.22).

I think if we' re going to toss this into -stable, we should put it into net-2.6 too, and that's what I'm going to do.

It's your decision: I don't think this patch is worth any arguing about (de)stabilizing. Btw., since -stable rules are less strict it seems natural such patches with bug fixes should rather go net-next -> -stable way, unless I miss something?

Thanks,
Jarek P.

Tested-by Pawel Staszewski
#96 lore raw
In reply to Paweł Staszewski
Quoted 14 lines · Paweł Staszewski

I apply this patch

fib_triestats in attached file :)

Quoted 10 lines

------------------->
ipv4: Fix fib_trie rebalancing, part 3

Alas current delaying of freeing old tnodes by RCU in trie_rebalance is still not enough because we can free a top tnode before updating a t->trie pointer.

---

David, I guess you could add:

Thanks,
Jarek P.

maintainer
#97 lore raw

From: Jarek Poplawski <[email protected]>
Date: Mon, 29 Jun 2009 10:58:20 +0000

Quoted 12 lines · Paweł Staszewski
Quoted 11 lines

I apply this patch

fib_triestats in attached file :)

Quoted 7 lines

------------------->
ipv4: Fix fib_trie rebalancing, part 3

Alas current delaying of freeing old tnodes by RCU in trie_rebalance is still not enough because we can free a top tnode before updating a t->trie pointer.

Quoted 7 lines
Quoted 3 lines
Quoted 3 lines

---

David, I guess you could add:

Done, and applied, thanks Jarek.

#98 lore raw
Quoted 19 lines · David Miller

From: Jarek Poplawski <[email protected]>
Date: Mon, 29 Jun 2009 10:58:20 +0000

Quoted 16 lines · Paweł Staszewski
Quoted 14 lines

I apply this patch

fib_triestats in attached file :)

Quoted 10 lines

------------------->
ipv4: Fix fib_trie rebalancing, part 3

Alas current delaying of freeing old tnodes by RCU in trie_rebalance is still not enough because we can free a top tnode before updating a t->trie pointer.

---

Quoted 5 lines
Quoted 3 lines

David, I guess you could add:

Done, and applied, thanks Jarek.

Btw., a little comment: there are still some issues while trying to reclaim memory after synchronize_rcu, which means the algorithm is buggy, or RCU use is still buggy, or maybe some timing because of synchronize_rcu. Anyway, fib_trie still seems to be safe only with CONFIG_PREEMPT_NONE, so I have no idea how this should be fixed in -stables (or why people don't report more this BUG in 2.6.30)...

Thanks,
Jarek P.

#99 lore raw
In reply to David Miller
Quoted 21 lines · On Tue, 30 Jun 2009 12:48:49 -0700 (PDT) David Miller <[email protected]>…

From: Jarek Poplawski <[email protected]>
Date: Mon, 29 Jun 2009 10:58:20 +0000

Quoted 18 lines · Paweł Staszewski
Quoted 14 lines

I apply this patch

fib_triestats in attached file :)

Quoted 10 lines

------------------->
ipv4: Fix fib_trie rebalancing, part 3

Alas current delaying of freeing old tnodes by RCU in trie_rebalance is still not enough because we can free a top tnode before updating a t->trie pointer.

---

David, I guess you could add:

Quoted 7 lines
Quoted 1 line

Done, and applied, thanks Jarek.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

This is probably in kernel bugzilla as well, so someone should update:
http://bugzilla.kernel.org/show_bug.cgi?id=6648