Skip to contentSkip to conversation

x86/mm: Fix incorrect for loop count calculation in sync_global_pgds

PATCH Superseded Baoquan He 5 participants 10 messages May 1, 2017 – May 2, 2017

Subsystems (2)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Baoquan He

author
#0 lore raw

PATCH 1/1 x86/mm: Fix incorrect for loop count calculation in sync_global_pgds

Jeff Moyer reported that on his system with two memory regions 0~64G and 1T~1T+192G, and kernel option "memmap=192G!1024G" added, enabling kaslr will make system hang intermittently during boot. While adding 'nokaslr' won't.

This is because the for loop count calculation in sync_global_pgds is not correct. When a mapping area crosses pgd entries, we should calculate the starting address of region which next pgd covers and assign it to next for loop count, but not add PGDIR_SIZE directly. The old code works right only if the mapping area is times of PGDIR_SIZE, otherwize the end region could be skipped so that it can't be synchronized to all other processes from kernel pgd init_mm.pgd.

In Jeff's system, emulated pmem area [1024G, 1216G) is smaller than PGDIR_SIZE. While 'nokaslr' works because PAGE_OFFSET is 1T aligned, it makes this area be mapped inside one pgd entry. With kaslr enabled, this area could cross two pgd entries, then the next pgd entry won't be synced to all other processes. That is why we saw empty PGD.

Fix it in this patch.

The back trace is pasted as below:

[    9.988867] IP: memcpy_erms+0x6/0x10
[    9.988868] PGD 0
[    9.988868]
[    9.988870] Oops: 0000 [#1] SMP
[    9.988871] Modules linked in: isci(E) mgag200(E+) drm_kms_helper(E) syscopyarea(E) sysfillrect(E) sysimgblt(E) fb_sys_fops(E) igb(E) ahci(E) ttm(E) libsas(E) libahci(E) scsi_transport_sas(E) ptp(E) pps_core(E) nd_pmem(E) dca(E) drm(E) i2c_algo_bit(E) libata(E) crc32c_intel(E) nd_btt(E) i2c_core(E) dm_mirror(E) dm_region_hash(E) dm_log(E) dm_mod(E)
[    9.988886] CPU: 0 PID: 442 Comm: systemd-udevd Tainted: G            E   4.11.0-rc5+ #43
[    9.988887] Hardware name: Intel Corporation LH Pass/SVRBD-ROW_P, BIOS SE5C600.86B.02.01.SP06.050920141054 05/09/2014
[    9.988888] task: ffff9267dc2f8000 task.stack: ffffba92c783c000
[    9.988890] RIP: 0010:memcpy_erms+0x6/0x10
[    9.988891] RSP: 0018:ffffba92c783f9b8 EFLAGS: 00010286
[    9.988892] RAX: ffff925f19e27000 RBX: 0000000000000000 RCX: 0000000000001000
[    9.988893] RDX: 0000000000001000 RSI: ffff9387bfff0000 RDI: ffff925f19e27000
[    9.988893] RBP: ffffba92c783fa38 R08: 0000000000000000 R09: 0000000017ffff80
[    9.988894] R10: 0000000000000000 R11: ffff9387bfff0000 R12: ffff925fde811ed8
[    9.988895] R13: 0000002fffff0000 R14: 0000000000001000 R15: ffff925f19e27000
[    9.988896] FS:  00007f1ee18e68c0(0000) GS:ffff925fdec00000(0000) knlGS:0000000000000000
[    9.988896] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[    9.988897] CR2: ffff9387bfff0000 CR3: 000000081ba28000 CR4: 00000000001406f0
[    9.988897] Call Trace:
[    9.988902]  ? pmem_do_bvec+0x93/0x290 [nd_pmem]
[    9.988904]  ? radix_tree_node_alloc.constprop.20+0x85/0xc0
[    9.988905]  ? radix_tree_node_alloc.constprop.20+0x85/0xc0
[    9.988907]  pmem_rw_page+0x3a/0x60 [nd_pmem]
[    9.988909]  bdev_read_page+0x81/0xb0
[    9.988911]  do_mpage_readpage+0x56f/0x770
[    9.988912]  ? I_BDEV+0x20/0x20
[    9.988915]  ? lru_cache_add+0xe/0x10
[    9.988917]  mpage_readpages+0x148/0x1e0
[    9.988917]  ? I_BDEV+0x20/0x20
[    9.988918]  ? I_BDEV+0x20/0x20
[    9.988921]  ? alloc_pages_current+0x88/0x120
[    9.988923]  blkdev_readpages+0x1d/0x20
[    9.988924]  __do_page_cache_readahead+0x1ce/0x2c0
[    9.988926]  force_page_cache_readahead+0xa2/0x100
[    9.988927]  page_cache_sync_readahead+0x3f/0x50
[    9.988930]  generic_file_read_iter+0x60d/0x8c0
[    9.988931]  blkdev_read_iter+0x37/0x40
[    9.988933]  __vfs_read+0xe0/0x150
[    9.988934]  vfs_read+0x8c/0x130
[    9.988936]  SyS_read+0x55/0xc0
[    9.988939]  entry_SYSCALL_64_fastpath+0x1a/0xa9
[    9.988940] RIP: 0033:0x7f1ee0822480
[    9.988941] RSP: 002b:00007ffcf9e741f8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000
[    9.988942] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f1ee0822480
[    9.988943] RDX: 0000000000000040 RSI: 0000561b7e1aabc8 RDI: 0000000000000008
[    9.988943] RBP: 0000561b7e1a86a0 R08: 0000000000000005 R09: 0000000000000068
[    9.988944] R10: 00007ffcf9e73f80 R11: 0000000000000246 R12: 0000000000000000
[    9.988945] R13: 0000000000000001 R14: 0000561b7e1a61b0 R15: 0000561b7e1a55e0
[    9.988946] Code: ff 90 90 90 90 eb 1e 0f 1f 00 48 89 f8 48 89 d1 48 c1 e9 03 83 e2 07 f3 48 a5 89 d1 f3 a4 c3 66 0f 1f 44 00 00 48 89 f8 48 89 d1 <f3> a4 c3 0f 1f 80 00 00 00 00 48 89 f8 48 83 fa 20 72 7e 40 38
[    9.988962] RIP: memcpy_erms+0x6/0x10 RSP: ffffba92c783f9b8
[    9.988962] CR2: ffff9387bfff0000
[    9.989022] ---[ end trace fe34c0fc0fe685ab ]---
[    9.998690] Kernel panic - not syncing: Fatal exception
[   10.004708] Kernel Offset: 0x11000000 from 0xffffffff81000000 (relocation range: 0xffffffff80000000-0xffffffffbfffffff)
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Kees Cook <[email protected]>
Cc Thomas Garnier <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Yasuaki Ishimatsu <[email protected]>
Cc Jinbum Park <[email protected]>
Cc Dave Hansen <[email protected]>
Cc "Kirill A. Shutemov" <[email protected]>
Cc Yinghai Lu <[email protected]>
Cc Dan Williams <[email protected]>
Cc Dave Young <[email protected]>
---
 arch/x86/mm/init_64.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/init_64.c +4 -2
index 15173d3..dbf4f00 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-	unsigned long address;
+	unsigned long address, address_next;
 
-	for (address = start; address <= end; address += PGDIR_SIZE) {
+	for (address = start; address <= end; address = address_next) {
 		const pgd_t *pgd_ref = pgd_offset_k(address);
 		struct page *page;
 
+		address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+
 		if (pgd_none(*pgd_ref))
 			continue;
 
Signature · 1 line
2.5.5
Acked-by on v1 1/1
#1 lore raw
Quoted 14 lines · Baoquan He

Jeff Moyer reported that on his system with two memory regions 0~64G and 1T~1T+192G, and kernel option "memmap=192G!1024G" added, enabling kaslr will make system hang intermittently during boot. While adding 'nokaslr' won't.

This is because the for loop count calculation in sync_global_pgds is not correct. When a mapping area crosses pgd entries, we should calculate the starting address of region which next pgd covers and assign it to next for loop count, but not add PGDIR_SIZE directly. The old code works right only if the mapping area is times of PGDIR_SIZE, otherwize the end region could be skipped so that it can't be synchronized to all other processes from kernel pgd init_mm.pgd.

Quoted 5 lines

In Jeff's system, emulated pmem area [1024G, 1216G) is smaller than PGDIR_SIZE. While 'nokaslr' works because PAGE_OFFSET is 1T aligned, it makes this area be mapped inside one pgd entry. With kaslr enabled, this area could cross two pgd entries, then the next pgd entry won't be synced to all other processes. That is why we saw empty PGD.

Make a lot of sense. Thanks a lot for investigating this issue!

Quoted 104 lines

Fix it in this patch.

The back trace is pasted as below:

[    9.988867] IP: memcpy_erms+0x6/0x10
[    9.988868] PGD 0
[    9.988868]
[    9.988870] Oops: 0000 [#1] SMP
[    9.988871] Modules linked in: isci(E) mgag200(E+) drm_kms_helper(E) syscopyarea(E) sysfillrect(E) sysimgblt(E) fb_sys_fops(E) igb(E) ahci(E) ttm(E) libsas(E) libahci(E) scsi_transport_sas(E) ptp(E) pps_core(E) nd_pmem(E) dca(E) drm(E) i2c_algo_bit(E) libata(E) crc32c_intel(E) nd_btt(E) i2c_core(E) dm_mirror(E) dm_region_hash(E) dm_log(E) dm_mod(E)
[    9.988886] CPU: 0 PID: 442 Comm: systemd-udevd Tainted: G            E   4.11.0-rc5+ #43
[    9.988887] Hardware name: Intel Corporation LH Pass/SVRBD-ROW_P, BIOS SE5C600.86B.02.01.SP06.050920141054 05/09/2014
[    9.988888] task: ffff9267dc2f8000 task.stack: ffffba92c783c000
[    9.988890] RIP: 0010:memcpy_erms+0x6/0x10
[    9.988891] RSP: 0018:ffffba92c783f9b8 EFLAGS: 00010286
[    9.988892] RAX: ffff925f19e27000 RBX: 0000000000000000 RCX: 0000000000001000
[    9.988893] RDX: 0000000000001000 RSI: ffff9387bfff0000 RDI: ffff925f19e27000
[    9.988893] RBP: ffffba92c783fa38 R08: 0000000000000000 R09: 0000000017ffff80
[    9.988894] R10: 0000000000000000 R11: ffff9387bfff0000 R12: ffff925fde811ed8
[    9.988895] R13: 0000002fffff0000 R14: 0000000000001000 R15: ffff925f19e27000
[    9.988896] FS:  00007f1ee18e68c0(0000) GS:ffff925fdec00000(0000) knlGS:0000000000000000
[    9.988896] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[    9.988897] CR2: ffff9387bfff0000 CR3: 000000081ba28000 CR4: 00000000001406f0
[    9.988897] Call Trace:
[    9.988902]  ? pmem_do_bvec+0x93/0x290 [nd_pmem]
[    9.988904]  ? radix_tree_node_alloc.constprop.20+0x85/0xc0
[    9.988905]  ? radix_tree_node_alloc.constprop.20+0x85/0xc0
[    9.988907]  pmem_rw_page+0x3a/0x60 [nd_pmem]
[    9.988909]  bdev_read_page+0x81/0xb0
[    9.988911]  do_mpage_readpage+0x56f/0x770
[    9.988912]  ? I_BDEV+0x20/0x20
[    9.988915]  ? lru_cache_add+0xe/0x10
[    9.988917]  mpage_readpages+0x148/0x1e0
[    9.988917]  ? I_BDEV+0x20/0x20
[    9.988918]  ? I_BDEV+0x20/0x20
[    9.988921]  ? alloc_pages_current+0x88/0x120
[    9.988923]  blkdev_readpages+0x1d/0x20
[    9.988924]  __do_page_cache_readahead+0x1ce/0x2c0
[    9.988926]  force_page_cache_readahead+0xa2/0x100
[    9.988927]  page_cache_sync_readahead+0x3f/0x50
[    9.988930]  generic_file_read_iter+0x60d/0x8c0
[    9.988931]  blkdev_read_iter+0x37/0x40
[    9.988933]  __vfs_read+0xe0/0x150
[    9.988934]  vfs_read+0x8c/0x130
[    9.988936]  SyS_read+0x55/0xc0
[    9.988939]  entry_SYSCALL_64_fastpath+0x1a/0xa9
[    9.988940] RIP: 0033:0x7f1ee0822480
[    9.988941] RSP: 002b:00007ffcf9e741f8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000
[    9.988942] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f1ee0822480
[    9.988943] RDX: 0000000000000040 RSI: 0000561b7e1aabc8 RDI: 0000000000000008
[    9.988943] RBP: 0000561b7e1a86a0 R08: 0000000000000005 R09: 0000000000000068
[    9.988944] R10: 00007ffcf9e73f80 R11: 0000000000000246 R12: 0000000000000000
[    9.988945] R13: 0000000000000001 R14: 0000561b7e1a61b0 R15: 0000561b7e1a55e0
[    9.988946] Code: ff 90 90 90 90 eb 1e 0f 1f 00 48 89 f8 48 89 d1 48 c1 e9 03 83 e2 07 f3 48 a5 89 d1 f3 a4 c3 66 0f 1f 44 00 00 48 89 f8 48 89 d1 <f3> a4 c3 0f 1f 80 00 00 00 00 48 89 f8 48 83 fa 20 72 7e 40 38
[    9.988962] RIP: memcpy_erms+0x6/0x10 RSP: ffffba92c783f9b8
[    9.988962] CR2: ffff9387bfff0000
[    9.989022] ---[ end trace fe34c0fc0fe685ab ]---
[    9.998690] Kernel panic - not syncing: Fatal exception
[   10.004708] Kernel Offset: 0x11000000 from 0xffffffff81000000 (relocation range: 0xffffffff80000000-0xffffffffbfffffff)
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Kees Cook <[email protected]>
Cc Thomas Garnier <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Yasuaki Ishimatsu <[email protected]>
Cc Jinbum Park <[email protected]>
Cc Dave Hansen <[email protected]>
Cc "Kirill A. Shutemov" <[email protected]>
Cc Yinghai Lu <[email protected]>
Cc Dan Williams <[email protected]>
Cc Dave Young <[email protected]>
---
 arch/x86/mm/init_64.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/init_64.c +4 -3
index 15173d3..dbf4f00 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-       unsigned long address;
+       unsigned long address, address_next;
 
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (address = start; address <= end; address = address_next) {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+
                if (pgd_none(*pgd_ref))
                        continue;
 
--

2.5.5

Signature · 1 line
Thomas
#2 lore raw
In reply to Baoquan He
Quoted 94 lines · Baoquan He

Jeff Moyer reported that on his system with two memory regions 0~64G and 1T~1T+192G, and kernel option "memmap=192G!1024G" added, enabling kaslr will make system hang intermittently during boot. While adding 'nokaslr' won't.

This is because the for loop count calculation in sync_global_pgds is not correct. When a mapping area crosses pgd entries, we should calculate the starting address of region which next pgd covers and assign it to next for loop count, but not add PGDIR_SIZE directly. The old code works right only if the mapping area is times of PGDIR_SIZE, otherwize the end region could be skipped so that it can't be synchronized to all other processes from kernel pgd init_mm.pgd.

In Jeff's system, emulated pmem area [1024G, 1216G) is smaller than PGDIR_SIZE. While 'nokaslr' works because PAGE_OFFSET is 1T aligned, it makes this area be mapped inside one pgd entry. With kaslr enabled, this area could cross two pgd entries, then the next pgd entry won't be synced to all other processes. That is why we saw empty PGD.

Fix it in this patch.

The back trace is pasted as below:

[    9.988867] IP: memcpy_erms+0x6/0x10
[    9.988868] PGD 0
[    9.988868]
[    9.988870] Oops: 0000 [#1] SMP
[    9.988871] Modules linked in: isci(E) mgag200(E+) drm_kms_helper(E) syscopyarea(E) sysfillrect(E) sysimgblt(E) fb_sys_fops(E) igb(E) ahci(E) ttm(E) libsas(E) libahci(E) scsi_transport_sas(E) ptp(E) pps_core(E) nd_pmem(E) dca(E) drm(E) i2c_algo_bit(E) libata(E) crc32c_intel(E) nd_btt(E) i2c_core(E) dm_mirror(E) dm_region_hash(E) dm_log(E) dm_mod(E)
[    9.988886] CPU: 0 PID: 442 Comm: systemd-udevd Tainted: G            E   4.11.0-rc5+ #43
[    9.988887] Hardware name: Intel Corporation LH Pass/SVRBD-ROW_P, BIOS SE5C600.86B.02.01.SP06.050920141054 05/09/2014
[    9.988888] task: ffff9267dc2f8000 task.stack: ffffba92c783c000
[    9.988890] RIP: 0010:memcpy_erms+0x6/0x10
[    9.988891] RSP: 0018:ffffba92c783f9b8 EFLAGS: 00010286
[    9.988892] RAX: ffff925f19e27000 RBX: 0000000000000000 RCX: 0000000000001000
[    9.988893] RDX: 0000000000001000 RSI: ffff9387bfff0000 RDI: ffff925f19e27000
[    9.988893] RBP: ffffba92c783fa38 R08: 0000000000000000 R09: 0000000017ffff80
[    9.988894] R10: 0000000000000000 R11: ffff9387bfff0000 R12: ffff925fde811ed8
[    9.988895] R13: 0000002fffff0000 R14: 0000000000001000 R15: ffff925f19e27000
[    9.988896] FS:  00007f1ee18e68c0(0000) GS:ffff925fdec00000(0000) knlGS:0000000000000000
[    9.988896] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[    9.988897] CR2: ffff9387bfff0000 CR3: 000000081ba28000 CR4: 00000000001406f0
[    9.988897] Call Trace:
[    9.988902]  ? pmem_do_bvec+0x93/0x290 [nd_pmem]
[    9.988904]  ? radix_tree_node_alloc.constprop.20+0x85/0xc0
[    9.988905]  ? radix_tree_node_alloc.constprop.20+0x85/0xc0
[    9.988907]  pmem_rw_page+0x3a/0x60 [nd_pmem]
[    9.988909]  bdev_read_page+0x81/0xb0
[    9.988911]  do_mpage_readpage+0x56f/0x770
[    9.988912]  ? I_BDEV+0x20/0x20
[    9.988915]  ? lru_cache_add+0xe/0x10
[    9.988917]  mpage_readpages+0x148/0x1e0
[    9.988917]  ? I_BDEV+0x20/0x20
[    9.988918]  ? I_BDEV+0x20/0x20
[    9.988921]  ? alloc_pages_current+0x88/0x120
[    9.988923]  blkdev_readpages+0x1d/0x20
[    9.988924]  __do_page_cache_readahead+0x1ce/0x2c0
[    9.988926]  force_page_cache_readahead+0xa2/0x100
[    9.988927]  page_cache_sync_readahead+0x3f/0x50
[    9.988930]  generic_file_read_iter+0x60d/0x8c0
[    9.988931]  blkdev_read_iter+0x37/0x40
[    9.988933]  __vfs_read+0xe0/0x150
[    9.988934]  vfs_read+0x8c/0x130
[    9.988936]  SyS_read+0x55/0xc0
[    9.988939]  entry_SYSCALL_64_fastpath+0x1a/0xa9
[    9.988940] RIP: 0033:0x7f1ee0822480
[    9.988941] RSP: 002b:00007ffcf9e741f8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000
[    9.988942] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f1ee0822480
[    9.988943] RDX: 0000000000000040 RSI: 0000561b7e1aabc8 RDI: 0000000000000008
[    9.988943] RBP: 0000561b7e1a86a0 R08: 0000000000000005 R09: 0000000000000068
[    9.988944] R10: 00007ffcf9e73f80 R11: 0000000000000246 R12: 0000000000000000
[    9.988945] R13: 0000000000000001 R14: 0000561b7e1a61b0 R15: 0000561b7e1a55e0
[    9.988946] Code: ff 90 90 90 90 eb 1e 0f 1f 00 48 89 f8 48 89 d1 48 c1 e9 03 83 e2 07 f3 48 a5 89 d1 f3 a4 c3 66 0f 1f 44 00 00 48 89 f8 48 89 d1 <f3> a4 c3 0f 1f 80 00 00 00 00 48 89 f8 48 83 fa 20 72 7e 40 38
[    9.988962] RIP: memcpy_erms+0x6/0x10 RSP: ffffba92c783f9b8
[    9.988962] CR2: ffff9387bfff0000
[    9.989022] ---[ end trace fe34c0fc0fe685ab ]---
[    9.998690] Kernel panic - not syncing: Fatal exception
[   10.004708] Kernel Offset: 0x11000000 from 0xffffffff81000000 (relocation range: 0xffffffff80000000-0xffffffffbfffffff)
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Kees Cook <[email protected]>
Cc Thomas Garnier <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Yasuaki Ishimatsu <[email protected]>
Cc Jinbum Park <[email protected]>
Cc Dave Hansen <[email protected]>
Cc "Kirill A. Shutemov" <[email protected]>
Cc Yinghai Lu <[email protected]>
Cc Dan Williams <[email protected]>
Cc Dave Young <[email protected]>
Quoted 1 line

---

Good catch!

Quoted 11 lines
 arch/x86/mm/init_64.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/init_64.c +0 -0
index 15173d3..dbf4f00 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {

arch/x86/mm/init_64.c __setup("noexec32=", nonx32_setup); · line 104

Quoted 10 lines
arch/x86/mm/init_64.c +4 -2
-       unsigned long address;
+       unsigned long address, address_next;
 
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (address = start; address <= end; address = address_next) {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+

Let's change this to put the next address calculation in the for loop directly and use the ALIGN macro. Something like:

for (address = start; address <= end; address = ALIGN(address + 1, PGDIR_SIZE))

author
#3 lore raw
Quoted 20 lines · Dan Williams
Quoted 19 lines · Baoquan He
 arch/x86/mm/init_64.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/init_64.c +2 -2
index 15173d3..dbf4f00 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-       unsigned long address;
+       unsigned long address, address_next;
 
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (address = start; address <= end; address = address_next) {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
Quoted 7 lines

arch/x86/mm/init_64.c __setup("noexec32=", nonx32_setup); · line 104

Quoted 2 lines
arch/x86/mm/init_64.c +2 -0
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+

Let's change this to put the next address calculation in the for loop directly and use the ALIGN macro. Something like:

for (address = start; address <= end; address = ALIGN(address + 1, PGDIR_SIZE))

Hi Dan,

Good idea!

Do you think below change is OK for you? Taking out the initialization can make the for loop line be shorter than 80 char.

arch/x86/mm/init_64.c +2 -2
index 15173d3..0840311 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-       unsigned long address;
+       unsigned long address = start;
 
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (; address <= end; address = ALIGN(address + 1, PGDIR_SIZE))
{
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;

+ address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+

                if (pgd_none(*pgd_ref))
                        continue;
#4 lore raw
Quoted 27 lines · Baoquan He
Quoted 26 lines · Dan Williams
Quoted 11 lines · Baoquan He
 arch/x86/mm/init_64.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/init_64.c +0 -0
index 15173d3..dbf4f00 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {

arch/x86/mm/init_64.c __setup("noexec32=", nonx32_setup); · line 104

Quoted 10 lines
arch/x86/mm/init_64.c +4 -2
-       unsigned long address;
+       unsigned long address, address_next;
 
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (address = start; address <= end; address = address_next) {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+

Let's change this to put the next address calculation in the for loop directly and use the ALIGN macro. Something like:

Quoted 9 lines
Quoted 1 line

for (address = start; address <= end; address = ALIGN(address + 1, PGDIR_SIZE))

Hi Dan,

Good idea!

Do you think below change is OK for you? Taking out the initialization can make the for loop line be shorter than 80 char.

I would just wrap the "address = ALIGN(address + 1, PGDIR_SIZE)" if it doesn't fit.

Quoted 10 lines
arch/x86/mm/init_64.c +0 -1
index 15173d3..0840311 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-       unsigned long address;

arch/x86/mm/init_64.c __setup("noexec32=", nonx32_setup); · line 105

Quoted 10 lines
arch/x86/mm/init_64.c +4 -1
+       unsigned long address = start;
 
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (; address <= end; address = ALIGN(address + 1, PGDIR_SIZE))
 {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+

This gets deleted of course.

author
#5 lore raw
Quoted 32 lines · Dan Williams
Quoted 31 lines · Baoquan He
Quoted 20 lines · Dan Williams
Quoted 19 lines · Baoquan He
 arch/x86/mm/init_64.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/init_64.c +2 -2
index 15173d3..dbf4f00 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-       unsigned long address;
+       unsigned long address, address_next;
 
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (address = start; address <= end; address = address_next) {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
Quoted 7 lines

arch/x86/mm/init_64.c __setup("noexec32=", nonx32_setup); · line 104

Quoted 2 lines
arch/x86/mm/init_64.c +2 -0
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+

Let's change this to put the next address calculation in the for loop directly and use the ALIGN macro. Something like:

for (address = start; address <= end; address = ALIGN(address + 1, PGDIR_SIZE))

Hi Dan,

Quoted 8 lines
Quoted 5 lines

Good idea!

Do you think below change is OK for you? Taking out the initialization can make the for loop line be shorter than 80 char.

I would just wrap the "address = ALIGN(address + 1, PGDIR_SIZE)" if it doesn't fit.

OK, it's fine, will do wrapping.

Quoted 19 lines
Quoted 19 lines
arch/x86/mm/init_64.c +2 -2
index 15173d3..0840311 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-       unsigned long address;
+       unsigned long address = start;
 
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (; address <= end; address = ALIGN(address + 1, PGDIR_SIZE))
 {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
Quoted 4 lines

arch/x86/mm/init_64.c __setup("noexec32=", nonx32_setup); · line 105

Quoted 2 lines
arch/x86/mm/init_64.c +2 -0
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+

This gets deleted of course.

Sure, forget deleting it. while I am also testing on Jeff's system, that code is right, otherwise compiling won't pass.

Will repost after the pgd crossing case seen and passed.

Thanks
Baoquan

Acked-by on v1 1/1
#6 lore raw
In reply to Baoquan He
Quoted 109 lines · Baoquan He

Jeff Moyer reported that on his system with two memory regions 0~64G and 1T~1T+192G, and kernel option "memmap=192G!1024G" added, enabling kaslr will make system hang intermittently during boot. While adding 'nokaslr' won't.

This is because the for loop count calculation in sync_global_pgds is not correct. When a mapping area crosses pgd entries, we should calculate the starting address of region which next pgd covers and assign it to next for loop count, but not add PGDIR_SIZE directly. The old code works right only if the mapping area is times of PGDIR_SIZE, otherwize the end region could be skipped so that it can't be synchronized to all other processes from kernel pgd init_mm.pgd.

In Jeff's system, emulated pmem area [1024G, 1216G) is smaller than PGDIR_SIZE. While 'nokaslr' works because PAGE_OFFSET is 1T aligned, it makes this area be mapped inside one pgd entry. With kaslr enabled, this area could cross two pgd entries, then the next pgd entry won't be synced to all other processes. That is why we saw empty PGD.

Fix it in this patch.

The back trace is pasted as below:

[    9.988867] IP: memcpy_erms+0x6/0x10
[    9.988868] PGD 0
[    9.988868]
[    9.988870] Oops: 0000 [#1] SMP
[    9.988871] Modules linked in: isci(E) mgag200(E+) drm_kms_helper(E) syscopyarea(E) sysfillrect(E) sysimgblt(E) fb_sys_fops(E) igb(E) ahci(E) ttm(E) libsas(E) libahci(E) scsi_transport_sas(E) ptp(E) pps_core(E) nd_pmem(E) dca(E) drm(E) i2c_algo_bit(E) libata(E) crc32c_intel(E) nd_btt(E) i2c_core(E) dm_mirror(E) dm_region_hash(E) dm_log(E) dm_mod(E)
[    9.988886] CPU: 0 PID: 442 Comm: systemd-udevd Tainted: G            E   4.11.0-rc5+ #43
[    9.988887] Hardware name: Intel Corporation LH Pass/SVRBD-ROW_P, BIOS SE5C600.86B.02.01.SP06.050920141054 05/09/2014
[    9.988888] task: ffff9267dc2f8000 task.stack: ffffba92c783c000
[    9.988890] RIP: 0010:memcpy_erms+0x6/0x10
[    9.988891] RSP: 0018:ffffba92c783f9b8 EFLAGS: 00010286
[    9.988892] RAX: ffff925f19e27000 RBX: 0000000000000000 RCX: 0000000000001000
[    9.988893] RDX: 0000000000001000 RSI: ffff9387bfff0000 RDI: ffff925f19e27000
[    9.988893] RBP: ffffba92c783fa38 R08: 0000000000000000 R09: 0000000017ffff80
[    9.988894] R10: 0000000000000000 R11: ffff9387bfff0000 R12: ffff925fde811ed8
[    9.988895] R13: 0000002fffff0000 R14: 0000000000001000 R15: ffff925f19e27000
[    9.988896] FS:  00007f1ee18e68c0(0000) GS:ffff925fdec00000(0000) knlGS:0000000000000000
[    9.988896] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[    9.988897] CR2: ffff9387bfff0000 CR3: 000000081ba28000 CR4: 00000000001406f0
[    9.988897] Call Trace:
[    9.988902]  ? pmem_do_bvec+0x93/0x290 [nd_pmem]
[    9.988904]  ? radix_tree_node_alloc.constprop.20+0x85/0xc0
[    9.988905]  ? radix_tree_node_alloc.constprop.20+0x85/0xc0
[    9.988907]  pmem_rw_page+0x3a/0x60 [nd_pmem]
[    9.988909]  bdev_read_page+0x81/0xb0
[    9.988911]  do_mpage_readpage+0x56f/0x770
[    9.988912]  ? I_BDEV+0x20/0x20
[    9.988915]  ? lru_cache_add+0xe/0x10
[    9.988917]  mpage_readpages+0x148/0x1e0
[    9.988917]  ? I_BDEV+0x20/0x20
[    9.988918]  ? I_BDEV+0x20/0x20
[    9.988921]  ? alloc_pages_current+0x88/0x120
[    9.988923]  blkdev_readpages+0x1d/0x20
[    9.988924]  __do_page_cache_readahead+0x1ce/0x2c0
[    9.988926]  force_page_cache_readahead+0xa2/0x100
[    9.988927]  page_cache_sync_readahead+0x3f/0x50
[    9.988930]  generic_file_read_iter+0x60d/0x8c0
[    9.988931]  blkdev_read_iter+0x37/0x40
[    9.988933]  __vfs_read+0xe0/0x150
[    9.988934]  vfs_read+0x8c/0x130
[    9.988936]  SyS_read+0x55/0xc0
[    9.988939]  entry_SYSCALL_64_fastpath+0x1a/0xa9
[    9.988940] RIP: 0033:0x7f1ee0822480
[    9.988941] RSP: 002b:00007ffcf9e741f8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000
[    9.988942] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f1ee0822480
[    9.988943] RDX: 0000000000000040 RSI: 0000561b7e1aabc8 RDI: 0000000000000008
[    9.988943] RBP: 0000561b7e1a86a0 R08: 0000000000000005 R09: 0000000000000068
[    9.988944] R10: 00007ffcf9e73f80 R11: 0000000000000246 R12: 0000000000000000
[    9.988945] R13: 0000000000000001 R14: 0000561b7e1a61b0 R15: 0000561b7e1a55e0
[    9.988946] Code: ff 90 90 90 90 eb 1e 0f 1f 00 48 89 f8 48 89 d1 48 c1 e9 03 83 e2 07 f3 48 a5 89 d1 f3 a4 c3 66 0f 1f 44 00 00 48 89 f8 48 89 d1 <f3> a4 c3 0f 1f 80 00 00 00 00 48 89 f8 48 83 fa 20 72 7e 40 38
[    9.988962] RIP: memcpy_erms+0x6/0x10 RSP: ffffba92c783f9b8
[    9.988962] CR2: ffff9387bfff0000
[    9.989022] ---[ end trace fe34c0fc0fe685ab ]---
[    9.998690] Kernel panic - not syncing: Fatal exception
[   10.004708] Kernel Offset: 0x11000000 from 0xffffffff81000000 (relocation range: 0xffffffff80000000-0xffffffffbfffffff)
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Kees Cook <[email protected]>
Cc Thomas Garnier <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Yasuaki Ishimatsu <[email protected]>
Cc Jinbum Park <[email protected]>
Cc Dave Hansen <[email protected]>
Cc "Kirill A. Shutemov" <[email protected]>
Cc Yinghai Lu <[email protected]>
Cc Dan Williams <[email protected]>
Cc Dave Young <[email protected]>
---
 arch/x86/mm/init_64.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/init_64.c +1 -1
index 15173d3..dbf4f00 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-       unsigned long address;
+       unsigned long address, address_next;
 

arch/x86/mm/init_64.c __setup("noexec32=", nonx32_setup); · line 107

Quoted 10 lines
arch/x86/mm/init_64.c +3 -1
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (address = start; address <= end; address = address_next) {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+
                if (pgd_none(*pgd_ref))
                        continue;
 

This one is better than V2.

It would better if could rename address to addr as Ingo suggested.

author
#7 lore raw
Quoted 28 lines · Yinghai Lu
Quoted 14 lines · Baoquan He
 arch/x86/mm/init_64.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/init_64.c +1 -1
index 15173d3..dbf4f00 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-       unsigned long address;
+       unsigned long address, address_next;
 

arch/x86/mm/init_64.c __setup("noexec32=", nonx32_setup); · line 107

Quoted 10 lines
arch/x86/mm/init_64.c +3 -1
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (address = start; address <= end; address = address_next) {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+
                if (pgd_none(*pgd_ref))
                        continue;
 

This one is better than V2.

Quoted 1 line

It would better if could rename address to addr as Ingo suggested.

Thanks for your checking and suggestion, Yinghai.

Both v1 and v2 are fine to me. As you said, code in v1 is easily understood, while v2 code is more compact, less line. The line of v1 is not more than 80. Maybe Ingo can help choose one which he likes better.

Thanks
Baoquan

#8 lore raw
Quoted 29 lines · Baoquan He
Quoted 28 lines · Yinghai Lu
Quoted 14 lines · Baoquan He
 arch/x86/mm/init_64.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/init_64.c +1 -1
index 15173d3..dbf4f00 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-       unsigned long address;
+       unsigned long address, address_next;
 

arch/x86/mm/init_64.c __setup("noexec32=", nonx32_setup); · line 107

Quoted 10 lines
arch/x86/mm/init_64.c +3 -1
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (address = start; address <= end; address = address_next) {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+
                if (pgd_none(*pgd_ref))
                        continue;
 

This one is better than V2.

Quoted 7 lines
Quoted 1 line

It would better if could rename address to addr as Ingo suggested.

Thanks for your checking and suggestion, Yinghai.

Both v1 and v2 are fine to me. As you said, code in v1 is easily understood, while v2 code is more compact, less line. The line of v1 is not more than 80. Maybe Ingo can help choose one which he likes better.

Let's do the variant I suggested - that makes the loop self-contained ('continue' would work as-is, etc.) and makes the code all around more readable.

Thanks,

Ingo

author
#9 lore raw
Quoted 34 lines · Ingo Molnar
Quoted 31 lines · Baoquan He
Quoted 29 lines · Yinghai Lu
Quoted 14 lines · Baoquan He
 arch/x86/mm/init_64.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/mm/init_64.c +1 -1
index 15173d3..dbf4f00 100644
@@ -94,12 +94,14 @@ __setup("noexec32=", nonx32_setup);
  */
 void sync_global_pgds(unsigned long start, unsigned long end)
 {
-       unsigned long address;
+       unsigned long address, address_next;
 

arch/x86/mm/init_64.c __setup("noexec32=", nonx32_setup); · line 107

Quoted 10 lines
arch/x86/mm/init_64.c +3 -1
-       for (address = start; address <= end; address += PGDIR_SIZE) {
+       for (address = start; address <= end; address = address_next) {
                const pgd_t *pgd_ref = pgd_offset_k(address);
                struct page *page;
 
+               address_next = (address & PGDIR_MASK) + PGDIR_SIZE;
+
                if (pgd_none(*pgd_ref))
                        continue;
 

This one is better than V2.

It would better if could rename address to addr as Ingo suggested.

Quoted 8 lines
Quoted 5 lines

Thanks for your checking and suggestion, Yinghai.

Both v1 and v2 are fine to me. As you said, code in v1 is easily understood, while v2 code is more compact, less line. The line of v1 is not more than 80. Maybe Ingo can help choose one which he likes better.

Let's do the variant I suggested - that makes the loop self-contained ('continue' would work as-is, etc.) and makes the code all around more readable.

Sure, let me change as you said, will post after testing passed.

Thanks a lot!