mm: hugepages can cause negative commitlimit

If the total size of hugepages allocated on a system is over half of the total memory size, commitlimit becomes a negative number.

What happens in fs/proc/meminfo.c is this calculation:

allowed = ((totalram_pages - hugetlb_total_pages())
* sysctl_overcommit_ratio / 100) + total_swap_pages;

The problem is that hugetlb_total_pages() is larger than totalram_pages resulting in a negative number. Since allowed is an unsigned long the negative shows up as a big number.

A similar calculation occurs in __vm_enough_memory() in mm/mmap.c.

A symptom of this problem is that /proc/meminfo prints a very large CommitLimit number.

CommitLimit: 737869762947802600 kB

To reproduce the problem reserve over half of memory as hugepages. For example "default_hugepagesz=1G hugepagesz=1G hugepages=64 Then look at /proc/meminfo "CommitLimit:" to see if it is too big.

The fix is to not subtract hugetlb_total_pages(). When hugepages are allocated totalram_pages is decremented so there is no need to subtract out hugetlb_total_pages() a second time.

Reported-by Russ Anderson <[email protected]>

Signed-off-by Russ Anderson <[email protected]>

---

Example of "CommitLimit:" being too big.

uv1-sys:~ # cat /proc/meminfo

MemTotal:       32395508 kB
MemFree:        32029276 kB
Buffers:            8656 kB
Cached:            89548 kB
SwapCached:            0 kB
Active:            55336 kB
Inactive:          73916 kB
Active(anon):      31220 kB
Inactive(anon):       36 kB
Active(file):      24116 kB
Inactive(file):    73880 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:             0 kB
SwapFree:              0 kB
Dirty:              1692 kB
Writeback:             0 kB
AnonPages:         31132 kB
Mapped:            15668 kB
Shmem:               152 kB
Slab:              70256 kB
SReclaimable:      17148 kB
SUnreclaim:        53108 kB
KernelStack:        6536 kB
PageTables:         3704 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:    737869762947802600 kB
Committed_AS:     394044 kB
VmallocTotal:   34359738367 kB
VmallocUsed:      713960 kB
VmallocChunk:   34325764204 kB
HardwareCorrupted:     0 kB
HugePages_Total:      32
HugePages_Free:       32
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:    1048576 kB
DirectMap4k:       16384 kB
DirectMap2M:     2064384 kB
DirectMap1G:    65011712 kB

 fs/proc/meminfo.c |    2 +-
 mm/mmap.c         |    3 +--
 2 files changed, 2 insertions(+), 3 deletions(-)

Index: linux/fs/proc/meminfo.c
===================================================================

fs/proc/meminfo.c 2011-05-18 08:53:00.568784147 -0500 renamed from fs/proc/meminfo.c 2011-05-17 16:03:50.935658801 -0500 +1 -1

@@ -36,7 +36,7 @@ static int meminfo_proc_show(struct seq_
       si_meminfo(&i);
       si_swapinfo(&i);
       committed = percpu_counter_read_positive(&vm_committed_as);
-       allowed = ((totalram_pages - hugetlb_total_pages())
+       allowed = (totalram_pages
               * sysctl_overcommit_ratio / 100) + total_swap_pages;
 
       cached = global_page_state(NR_FILE_PAGES) -

Index: linux/mm/mmap.c
===================================================================

mm/mmap.c 2011-05-18 08:54:34.912222405 -0500 renamed from mm/mmap.c 2011-05-17 16:03:51.727658828 -0500 +1 -3

@@ -167,8 +167,7 @@ int __vm_enough_memory(struct mm_struct
               goto error;
       }
 
-       allowed = (totalram_pages - hugetlb_total_pages())
-               * sysctl_overcommit_ratio / 100;
+       allowed = totalram_pages * sysctl_overcommit_ratio / 100;
       /*
        * Leave the last 3% for root
        */
--

Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc [email protected]

Howdy,

Quoted 112 lines · Russ Anderson

If the total size of hugepages allocated on a system is over half of the total memory size, commitlimit becomes a negative number.

What happens in fs/proc/meminfo.c is this calculation:

allowed = ((totalram_pages - hugetlb_total_pages())
* sysctl_overcommit_ratio / 100) + total_swap_pages;

The problem is that hugetlb_total_pages() is larger than totalram_pages resulting in a negative number. Since allowed is an unsigned long the negative shows up as a big number.

A similar calculation occurs in __vm_enough_memory() in mm/mmap.c.

A symptom of this problem is that /proc/meminfo prints a very large CommitLimit number.

CommitLimit: 737869762947802600 kB

To reproduce the problem reserve over half of memory as hugepages. For example "default_hugepagesz=1G hugepagesz=1G hugepages=64 Then look at /proc/meminfo "CommitLimit:" to see if it is too big.

The fix is to not subtract hugetlb_total_pages(). When hugepages are allocated totalram_pages is decremented so there is no need to subtract out hugetlb_total_pages() a second time.

Reported-by Russ Anderson <[email protected]>

Signed-off-by Russ Anderson <[email protected]>

---

Example of "CommitLimit:" being too big.

uv1-sys:~ # cat /proc/meminfo

MemTotal:       32395508 kB
MemFree:        32029276 kB
Buffers:            8656 kB
Cached:            89548 kB
SwapCached:            0 kB
Active:            55336 kB
Inactive:          73916 kB
Active(anon):      31220 kB
Inactive(anon):       36 kB
Active(file):      24116 kB
Inactive(file):    73880 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:             0 kB
SwapFree:              0 kB
Dirty:              1692 kB
Writeback:             0 kB
AnonPages:         31132 kB
Mapped:            15668 kB
Shmem:               152 kB
Slab:              70256 kB
SReclaimable:      17148 kB
SUnreclaim:        53108 kB
KernelStack:        6536 kB
PageTables:         3704 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:    737869762947802600 kB
Committed_AS:     394044 kB
VmallocTotal:   34359738367 kB
VmallocUsed:      713960 kB
VmallocChunk:   34325764204 kB
HardwareCorrupted:     0 kB
HugePages_Total:      32
HugePages_Free:       32
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:    1048576 kB
DirectMap4k:       16384 kB
DirectMap2M:     2064384 kB
DirectMap1G:    65011712 kB

 fs/proc/meminfo.c |    2 +-
 mm/mmap.c         |    3 +--
 2 files changed, 2 insertions(+), 3 deletions(-)

Index: linux/fs/proc/meminfo.c
===================================================================

fs/proc/meminfo.c 2011-05-18 08:53:00.568784147 -0500 renamed from fs/proc/meminfo.c 2011-05-17 16:03:50.935658801 -0500 +1 -1

@@ -36,7 +36,7 @@ static int meminfo_proc_show(struct seq_
       si_meminfo(&i);
       si_swapinfo(&i);
       committed = percpu_counter_read_positive(&vm_committed_as);
-       allowed = ((totalram_pages - hugetlb_total_pages())
+       allowed = (totalram_pages
               * sysctl_overcommit_ratio / 100) + total_swap_pages;
 
       cached = global_page_state(NR_FILE_PAGES) -

Index: linux/mm/mmap.c
===================================================================

mm/mmap.c 2011-05-18 08:54:34.912222405 -0500 renamed from mm/mmap.c 2011-05-17 16:03:51.727658828 -0500 +1 -3

@@ -167,8 +167,7 @@ int __vm_enough_memory(struct mm_struct
               goto error;
       }
 
-       allowed = (totalram_pages - hugetlb_total_pages())
-               * sysctl_overcommit_ratio / 100;
+       allowed = totalram_pages * sysctl_overcommit_ratio / 100;
       /*
        * Leave the last 3% for root
        */
--

Quoted 2 lines

Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc [email protected]

I'm afraid this will introduce a bug on how accurate kernel will account memory for overcommitment limits.

totalram_pages is not decremented as hugepages are allocated. Since

hugepages are reserved, hugetlb_total_pages() has to be accounted and subtracted from totalram_pages in order to render an accurate number of remaining pages available to the general memory workload commitment.

I've tried to reproduce your findings on my boxes, without success, unfortunately.

I'll keep chasing to hit this behaviour, though.

Cheers!
--aquini

Quoted 91 lines · Rafael Aquini

Howdy,

Quoted 87 lines · Russ Anderson

If the total size of hugepages allocated on a system is over half of the total memory size, commitlimit becomes a negative number.

What happens in fs/proc/meminfo.c is this calculation:

allowed = ((totalram_pages - hugetlb_total_pages())
* sysctl_overcommit_ratio / 100) + total_swap_pages;

The problem is that hugetlb_total_pages() is larger than totalram_pages resulting in a negative number. Since allowed is an unsigned long the negative shows up as a big number.

A similar calculation occurs in __vm_enough_memory() in mm/mmap.c.

A symptom of this problem is that /proc/meminfo prints a very large CommitLimit number.

CommitLimit: 737869762947802600 kB

To reproduce the problem reserve over half of memory as hugepages. For example "default_hugepagesz=1G hugepagesz=1G hugepages=64 Then look at /proc/meminfo "CommitLimit:" to see if it is too big.

The fix is to not subtract hugetlb_total_pages(). When hugepages are allocated totalram_pages is decremented so there is no need to subtract out hugetlb_total_pages() a second time.

Reported-by Russ Anderson <[email protected]>

Signed-off-by Russ Anderson <[email protected]>

---

Example of "CommitLimit:" being too big.

uv1-sys:~ # cat /proc/meminfo

MemTotal:       32395508 kB
MemFree:        32029276 kB
Buffers:            8656 kB
Cached:            89548 kB
SwapCached:            0 kB
Active:            55336 kB
Inactive:          73916 kB
Active(anon):      31220 kB
Inactive(anon):       36 kB
Active(file):      24116 kB
Inactive(file):    73880 kB
Unevictable:           0 kB
Mlocked:               0 kB
SwapTotal:             0 kB
SwapFree:              0 kB
Dirty:              1692 kB
Writeback:             0 kB
AnonPages:         31132 kB
Mapped:            15668 kB
Shmem:               152 kB
Slab:              70256 kB
SReclaimable:      17148 kB
SUnreclaim:        53108 kB
KernelStack:        6536 kB
PageTables:         3704 kB
NFS_Unstable:          0 kB
Bounce:                0 kB
WritebackTmp:          0 kB
CommitLimit:    737869762947802600 kB
Committed_AS:     394044 kB
VmallocTotal:   34359738367 kB
VmallocUsed:      713960 kB
VmallocChunk:   34325764204 kB
HardwareCorrupted:     0 kB
HugePages_Total:      32
HugePages_Free:       32
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:    1048576 kB
DirectMap4k:       16384 kB
DirectMap2M:     2064384 kB
DirectMap1G:    65011712 kB

 fs/proc/meminfo.c |    2 +-
 mm/mmap.c         |    3 +--
 2 files changed, 2 insertions(+), 3 deletions(-)

Index: linux/fs/proc/meminfo.c
===================================================================
--- linux.orig/fs/proc/meminfo.c 2011-05-17 16:03:50.935658801

-0500

Quoted 32 lines

Quoted 25 lines

+++ linux/fs/proc/meminfo.c 2011-05-18 08:53:00.568784147 -0500

fs/proc/meminfo.c +1 -1

@@ -36,7 +36,7 @@ static int meminfo_proc_show(struct seq_
       si_meminfo(&i);
       si_swapinfo(&i);
       committed = percpu_counter_read_positive(&vm_committed_as);
-       allowed = ((totalram_pages - hugetlb_total_pages())
+       allowed = (totalram_pages
               * sysctl_overcommit_ratio / 100) + total_swap_pages;
 
       cached = global_page_state(NR_FILE_PAGES) -

Index: linux/mm/mmap.c
===================================================================

mm/mmap.c 2011-05-18 08:54:34.912222405 -0500 renamed from mm/mmap.c 2011-05-17 16:03:51.727658828 -0500 +1 -3

@@ -167,8 +167,7 @@ int __vm_enough_memory(struct mm_struct
               goto error;
       }
 
-       allowed = (totalram_pages - hugetlb_total_pages())
-               * sysctl_overcommit_ratio / 100;
+       allowed = totalram_pages * sysctl_overcommit_ratio / 100;
       /*
        * Leave the last 3% for root
        */
--

Quoted 2 lines

Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc [email protected]

I'm afraid this will introduce a bug on how accurate kernel will account memory for overcommitment limits.

Quoted 1 line

totalram_pages is not decremented as hugepages are allocated. Since

Are you running on x86? It decrements totalram_pages on a x86_64 test system. Perhaps different architectures allocate hugepages differently.

The way it was verified was putting a printk in to print totalram_pages and hugetlb_total_pages. First the system was booted without any huge pages. The next boot one huge page was allocated. The next boot more hugepages allocated. Each time totalram_pages was reduced by the nuber of huge pages allocated, with totalram_pages + hugetlb_total_pages equaling the original number of pages.

That behavior is also consistent with allocating over half of memory resulting in CommitLimit going negative (as is shown in the above output).

Here is some data. Each represents a boot using 1G hugepages.
0 hugepages : totalram_pages 16519867 hugetlb_total_pages 0
1 hugepages : totalram_pages 16257723 hugetlb_total_pages 262144 2 hugepages : totalram_pages 15995578 hugetlb_total_pages 524288
31 hugepages : totalram_pages 8393403 hugetlb_total_pages 8126464 32 hugepages : totalram_pages 8131258 hugetlb_total_pages 8388608

Quoted 6 lines

hugepages are reserved, hugetlb_total_pages() has to be accounted and subtracted from totalram_pages in order to render an accurate number of remaining pages available to the general memory workload commitment.

I've tried to reproduce your findings on my boxes, without success, unfortunately.

Put a printk in meminfo_proc_show() to print totalram_pages and hugetlb_total_pages(). Add "default_hugepagesz=1G hugepagesz=1G hugepages=64"
to the boot line (varying the number of hugepages).

Quoted 4 lines

I'll keep chasing to hit this behaviour, though.

Cheers!
--aquini

--
Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc [email protected]

Quoted 24 lines · Russ Anderson

The way it was verified was putting a printk in to print totalram_pages and hugetlb_total_pages. First the system was booted without any huge pages. The next boot one huge page was allocated. The next boot more hugepages allocated. Each time totalram_pages was reduced by the nuber of huge pages allocated, with totalram_pages + hugetlb_total_pages equaling the original number of pages.

That behavior is also consistent with allocating over half of memory resulting in CommitLimit going negative (as is shown in the above output).

Here is some data. Each represents a boot using 1G hugepages.
0 hugepages : totalram_pages 16519867 hugetlb_total_pages 0
1 hugepages : totalram_pages 16257723 hugetlb_total_pages 262144 2 hugepages : totalram_pages 15995578 hugetlb_total_pages 524288
31 hugepages : totalram_pages 8393403 hugetlb_total_pages 8126464 32 hugepages : totalram_pages 8131258 hugetlb_total_pages 8388608

Quoted 4 lines

hugepages are reserved, hugetlb_total_pages() has to be accounted and subtracted from totalram_pages in order to render an accurate number of remaining pages available to the general memory workload commitment.

Quoted 9 lines

Quoted 2 lines

I've tried to reproduce your findings on my boxes, without success, unfortunately.

Put a printk in meminfo_proc_show() to print totalram_pages and hugetlb_total_pages(). Add "default_hugepagesz=1G hugepagesz=1G hugepages=64"
to the boot line (varying the number of hugepages).

Quoted 1 line

I'll keep chasing to hit this behaviour, though.

I got what I was doing different, and you are partially right. Checking mm/hugetlb.c:
1811 static int __init hugetlb_nrpages_setup(char *s) 1812 {
....
1834 /*
1835 * Global state is always initialized later in hugetlb_init.
1836 * But we need to allocate >= MAX_ORDER hstates here early to
still
1837 * use the bootmem allocator.
1838 */
1839 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1840 hugetlb_hstate_alloc_pages(parsed_hstate);
1841
1842 last_mhp = mhp;
1843
1844 return 1;
1845 }
1846 __setup("hugepages=", hugetlb_nrpages_setup);

I realize this issue you've reported only happens when you're using oversized hugepages. As their order are always >= MAX_ORDER, they got pages early allocated from bootmem allocator. So, these pages are not accounted for totalram_pages.

Although your patch covers a fix for the proposed case, it only works for scenarios where oversized hugepages are allocated on boot. I think it will, unfortunately, cause a bug for the remaining scenarios.

Cheers!
--aquini

OK, I see your point. The root problem is hugepages allocated at boot are subtracted from totalram_pages but hugepages allocated at run time are not. Correct me if I've mistate it or are other conditions.

By "allocated at run time" I mean "echo 1 > /proc/sys/vm/nr_hugepages". That allocation will not change totalram_pages but will change hugetlb_total_pages().

How best to fix this inconsistency? Should totalram_pages include or exclude hugepages? What are the implications?

Quoted 10 lines · On Thu, 19 May 2011 17:11:01 -0500 Russ Anderson <[email protected]> wrote:

OK, I see your point. The root problem is hugepages allocated at boot are subtracted from totalram_pages but hugepages allocated at run time are not. Correct me if I've mistate it or are other conditions.

By "allocated at run time" I mean "echo 1 > /proc/sys/vm/nr_hugepages". That allocation will not change totalram_pages but will change hugetlb_total_pages().

How best to fix this inconsistency? Should totalram_pages include or exclude hugepages? What are the implications?

The problem is that hugetlb_total_pages() is trying to account for two different things, while totalram_pages accounts for only one of those things, yes?

One fix would be to stop accounting for huge pages in totalram_pages altogether. That might break other things so careful checking would be needed.

Or we stop accounting for the boot-time allocated huge pages in hugetlb_total_pages(). Split the two things apart altogether and account for boot-time allocated and runtime-allocated pages separately. This souds saner to me - it reflects what's actually happening in the kernel.

Quoted 18 lines · Andrew Morton

Quoted 10 lines · On Thu, 19 May 2011 17:11:01 -0500 Russ Anderson <[email protected]> wrote:

OK, I see your point. The root problem is hugepages allocated at boot are subtracted from totalram_pages but hugepages allocated at run time are not. Correct me if I've mistate it or are other conditions.

By "allocated at run time" I mean "echo 1 > /proc/sys/vm/nr_hugepages". That allocation will not change totalram_pages but will change hugetlb_total_pages().

How best to fix this inconsistency? Should totalram_pages include or exclude hugepages? What are the implications?

The problem is that hugetlb_total_pages() is trying to account for two different things, while totalram_pages accounts for only one of those things, yes?

Quoted 8 lines

One fix would be to stop accounting for huge pages in totalram_pages altogether. That might break other things so careful checking would be needed.

Or we stop accounting for the boot-time allocated huge pages in hugetlb_total_pages(). Split the two things apart altogether and account for boot-time allocated and runtime-allocated pages separately. This souds saner to me - it reflects what's actually happening in the kernel.

Perhaps we can just reinstate the # of pages "stealed" at early boot allocation later, when hugetlb_init() calls gather_bootmem_prealloc()

mm/hugetlb.c +0 -0

index 8ee3bd8..d606c9c 100644

Quoted 36 lines · Rafael Aquini

Quoted 18 lines · Andrew Morton

Quoted 10 lines · On Thu, 19 May 2011 17:11:01 -0500 Russ Anderson <[email protected]> wrote:

OK, I see your point. The root problem is hugepages allocated at boot are subtracted from totalram_pages but hugepages allocated at run time are not. Correct me if I've mistate it or are other conditions.

By "allocated at run time" I mean "echo 1 > /proc/sys/vm/nr_hugepages". That allocation will not change totalram_pages but will change hugetlb_total_pages().

How best to fix this inconsistency? Should totalram_pages include or exclude hugepages? What are the implications?

The problem is that hugetlb_total_pages() is trying to account for two different things, while totalram_pages accounts for only one of those things, yes?

Quoted 8 lines

One fix would be to stop accounting for huge pages in totalram_pages altogether. That might break other things so careful checking would be needed.

Or we stop accounting for the boot-time allocated huge pages in hugetlb_total_pages(). Split the two things apart altogether and account for boot-time allocated and runtime-allocated pages separately. This souds saner to me - it reflects what's actually happening in the kernel.

Perhaps we can just reinstate the # of pages "stealed" at early boot allocation later, when hugetlb_init() calls gather_bootmem_prealloc()

mm/hugetlb.c +0 -0

index 8ee3bd8..d606c9c 100644

mm/hugetlb.c static void __init gather_bootmem_prealloc(void) · line 1116

Quoted 6 lines

mm/hugetlb.c +1 -0

@@ -1111,6 +1111,7 @@ static void __init gather_bootmem_prealloc(void)
                WARN_ON(page_count(page) != 1);
                prep_compound_huge_page(page, h->order);
                prep_new_huge_page(h, page, page_to_nid(page));
+               totalram_pages += 1 << h->order;
        }
 }

Howdy Russ,

Were you able to confirm if that proposed change fix the issue you've reported?

Although I've tested it with usual size hugepages and it did not messed things up, I'm not able to test it with GB hugepages, as I do not have any proc with "pdpe1gb" flag available.

Thanks in advance!
Cheers!

Signature · 1 line

Rafael Aquini <[email protected]>

Quoted 36 lines · Rafael Aquini

Quoted 18 lines · Andrew Morton

Quoted 10 lines · On Thu, 19 May 2011 17:11:01 -0500 Russ Anderson <[email protected]> wrote:

OK, I see your point. The root problem is hugepages allocated at boot are subtracted from totalram_pages but hugepages allocated at run time are not. Correct me if I've mistate it or are other conditions.

By "allocated at run time" I mean "echo 1 > /proc/sys/vm/nr_hugepages". That allocation will not change totalram_pages but will change hugetlb_total_pages().

How best to fix this inconsistency? Should totalram_pages include or exclude hugepages? What are the implications?

The problem is that hugetlb_total_pages() is trying to account for two different things, while totalram_pages accounts for only one of those things, yes?

Quoted 8 lines

One fix would be to stop accounting for huge pages in totalram_pages altogether. That might break other things so careful checking would be needed.

Or we stop accounting for the boot-time allocated huge pages in hugetlb_total_pages(). Split the two things apart altogether and account for boot-time allocated and runtime-allocated pages separately. This souds saner to me - it reflects what's actually happening in the kernel.

Perhaps we can just reinstate the # of pages "stealed" at early boot allocation later, when hugetlb_init() calls gather_bootmem_prealloc()

mm/hugetlb.c +0 -0

index 8ee3bd8..d606c9c 100644

mm/hugetlb.c static void __init gather_bootmem_prealloc(void) · line 1116

Quoted 6 lines

mm/hugetlb.c +1 -0

@@ -1111,6 +1111,7 @@ static void __init gather_bootmem_prealloc(void)
                WARN_ON(page_count(page) != 1);
                prep_compound_huge_page(page, h->order);
                prep_new_huge_page(h, page, page_to_nid(page));
+               totalram_pages += 1 << h->order;
        }
 }

Howdy Russ,

Were you able to confirm if that proposed change fix the issue you've reported?

Although I've tested it with usual size hugepages and it did not messed things up, I'm not able to test it with GB hugepages, as I do not have any proc with "pdpe1gb" flag available.

Thanks in advance!
Cheers!

Signature · 1 line

Rafael Aquini <[email protected]>

When 1GB hugepages are allocated on a system, free(1) reports less available memory than what really is installed in the box. Also, if the total size of hugepages allocated on a system is over half of the total memory size, CommitLimit becomes a negative number.

The problem is that gigantic hugepages (order > MAX_ORDER) can only be allocated at boot with bootmem, thus its frames are not accounted to 'totalram_pages'. However, they are accounted to hugetlb_total_pages()

What happens to turn CommitLimit into a negative number is this calculation, in fs/proc/meminfo.c:

allowed = ((totalram_pages - hugetlb_total_pages())
* sysctl_overcommit_ratio / 100) + total_swap_pages;

A similar calculation occurs in __vm_enough_memory() in mm/mmap.c.

Also, every vm statistic which depends on 'totalram_pages' will render confusing values, as if system were 'missing' some part of its memory.

Reported-by Russ Anderson <[email protected]>

Signed-off-by Rafael Aquini <[email protected]>

---
 mm/hugetlb.c |    8 ++++++++
 1 files changed, 8 insertions(+), 0 deletions(-)

mm/hugetlb.c +8 -0

index f33bb31..c67dd0f 100644
@@ -1111,6 +1111,14 @@ static void __init gather_bootmem_prealloc(void)
 		WARN_ON(page_count(page) != 1);
 		prep_compound_huge_page(page, h->order);
 		prep_new_huge_page(h, page, page_to_nid(page));
+
+		/* if we had gigantic hugepages allocated at boot time,
+		 * we need to reinstate the 'stolen' pages to totalram_pages,
+		 * in order to fix confusing memory reports from free(1)
+		 * and another side-effects, like CommitLimit going negative.
+		 */
+		if (h->order > (MAX_ORDER - 1))
+			totalram_pages += 1 << h->order;
 	}
 }
 

Signature · 1 line

1.7.4.4

When 1GB hugepages are allocated on a system, free(1) reports less available memory than what really is installed in the box. Also, if the total size of hugepages allocated on a system is over half of the total memory size, CommitLimit becomes a negative number.

The problem is that gigantic hugepages (order > MAX_ORDER) can only be allocated at boot with bootmem, thus its frames are not accounted to 'totalram_pages'. However, they are accounted to hugetlb_total_pages()

What happens to turn CommitLimit into a negative number is this calculation, in fs/proc/meminfo.c:

allowed = ((totalram_pages - hugetlb_total_pages())
* sysctl_overcommit_ratio / 100) + total_swap_pages;

A similar calculation occurs in __vm_enough_memory() in mm/mmap.c.

Also, every vm statistic which depends on 'totalram_pages' will render confusing values, as if system were 'missing' some part of its memory.

Quoted 17 lines · On Thu, 2 Jun 2011 23:55:57 -0300 Rafael Aquini <[email protected]> wrote:

When 1GB hugepages are allocated on a system, free(1) reports less available memory than what really is installed in the box. Also, if the total size of hugepages allocated on a system is over half of the total memory size, CommitLimit becomes a negative number.

The problem is that gigantic hugepages (order > MAX_ORDER) can only be allocated at boot with bootmem, thus its frames are not accounted to 'totalram_pages'. However, they are accounted to hugetlb_total_pages()

What happens to turn CommitLimit into a negative number is this calculation, in fs/proc/meminfo.c:

allowed = ((totalram_pages - hugetlb_total_pages())
* sysctl_overcommit_ratio / 100) + total_swap_pages;

Quoted 4 lines

A similar calculation occurs in __vm_enough_memory() in mm/mmap.c.

Also, every vm statistic which depends on 'totalram_pages' will render confusing values, as if system were 'missing' some part of its memory.

Is this bug serious enough to justify backporting the fix into -stable kernels?

Howdy Andrew,

Sorry, for this late reply.

Quoted 20 lines · Andrew Morton

Quoted 17 lines · On Thu, 2 Jun 2011 23:55:57 -0300 Rafael Aquini <[email protected]> wrote:

When 1GB hugepages are allocated on a system, free(1) reports less available memory than what really is installed in the box. Also, if the total size of hugepages allocated on a system is over half of the total memory size, CommitLimit becomes a negative number.

The problem is that gigantic hugepages (order > MAX_ORDER) can only be allocated at boot with bootmem, thus its frames are not accounted to 'totalram_pages'. However, they are accounted to hugetlb_total_pages()

What happens to turn CommitLimit into a negative number is this calculation, in fs/proc/meminfo.c:

allowed = ((totalram_pages - hugetlb_total_pages())
* sysctl_overcommit_ratio / 100) + total_swap_pages;

Quoted 7 lines

Quoted 4 lines

A similar calculation occurs in __vm_enough_memory() in mm/mmap.c.

Also, every vm statistic which depends on 'totalram_pages' will render confusing values, as if system were 'missing' some part of its memory.

Is this bug serious enough to justify backporting the fix into -stable kernels?

Despite not having testing it, I can think the following scenario as troublesome:
When gigantic hugepages are allocated and sysctl_overcommit_memory == OVERCOMMIT_NEVER. In a such situation, __vm_enough_memory() goes through the mentioned 'allowed' calculation and might end up mistakenly returning -ENOMEM, thus forcing the system to start reclaiming pages earlier than it would be ususal, and this could cause detrimental impact to overall system's performance, depending on the workload.

Besides the aforementioned scenario, I can only think of this causing annoyances with memory reports from /proc/meminfo and free(1).

Yes, it fixes the inconsistency in reporting totalram_pages.

There seems to be another issue. 1G hugepages can be allocated at boot time, but cannot be allocated at run time. "default_hugepagesz=1G hugepagesz=1G hugepages=1" on the boot line works. With "default_hugepagesz=1G hugepagesz=1G" the command "echo 1 > /proc/sys/vm/nr_hugepages" fails.

uv4-sys:~ # echo 1 > /proc/sys/vm/nr_hugepages
-bash: echo: write error: Invalid argument