[PATCH] mm: don't account shared file pages in user_reserve_pages

[PATCH] mm: don't account shared file pages in user_reserve_pages

[Roman Gushchin]

Shared file pages are never accounted in memory overcommit code,
so it isn't reasonable to count them in a code that limits the
maximal size of a process in OVERCOMMIT_NONE mode.

If a process has few large file mappings, the consequent attempts
to allocate anonymous memory may unexpectedly fail with -ENOMEM,
while there is free memory and overcommit limit if significantly
larger than the committed amount (as displayed in /proc/meminfo).

The problem is significantly smoothed by commit c9b1d0981fcc
("mm: limit growth of 3% hardcoded other user reserve"),
which limits the impact of this check with 128Mb (tunable via sysctl),
but it can still be a problem on small machines.

Signed-off-by: Roman Gushchin <klamm@yandex-team.ru>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrew Shewmaker <agshew@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
---
 mm/mmap.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/mmap.c b/mm/mmap.c
index 7f684d5..151fadf 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -220,7 +220,7 @@ int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
 	 */
 	if (mm) {
 		reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
-		allowed -= min(mm->total_vm / 32, reserve);
+		allowed -= min((mm->total_vm - mm->shared_vm) / 32, reserve);
 	}
 
 	if (percpu_counter_read_positive(&vm_committed_as) < allowed)
-- 
2.1.0

Re: [PATCH] mm: don't account shared file pages in user_reserve_pages

[Andrew Shewmaker]

On Thu, Jan 29, 2015 at 02:51:27PM +0300, Roman Gushchin wrote:
> Shared file pages are never accounted in memory overcommit code,
> so it isn't reasonable to count them in a code that limits the
> maximal size of a process in OVERCOMMIT_NONE mode.
> 
> If a process has few large file mappings, the consequent attempts
> to allocate anonymous memory may unexpectedly fail with -ENOMEM,
> while there is free memory and overcommit limit if significantly
> larger than the committed amount (as displayed in /proc/meminfo).
> 
> The problem is significantly smoothed by commit c9b1d0981fcc
> ("mm: limit growth of 3% hardcoded other user reserve"),
> which limits the impact of this check with 128Mb (tunable via sysctl),
> but it can still be a problem on small machines.
> 
> Signed-off-by: Roman Gushchin <klamm@yandex-team.ru>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Andrew Shewmaker <agshew@gmail.com>
> Cc: Rik van Riel <riel@redhat.com>
> Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
> ---
>  mm/mmap.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/mm/mmap.c b/mm/mmap.c
> index 7f684d5..151fadf 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -220,7 +220,7 @@ int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
>  	 */
>  	if (mm) {
>  		reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
> -		allowed -= min(mm->total_vm / 32, reserve);
> +		allowed -= min((mm->total_vm - mm->shared_vm) / 32, reserve);
>  	}
>  
>  	if (percpu_counter_read_positive(&vm_committed_as) < allowed)
> -- 
> 2.1.0

You're two patches conflict, don't they? Maybe you should resend
them as a patch series such that they can both be applied?

Does mm->shared_vm include memory that's mapped MAP_ANONYMOUS in
conjunction with MAP_SHARED? If so, then subtracting it could
overcommit the system OVERCOMMIT_NEVER mode.

-Andrew

Re: [PATCH] mm: don't account shared file pages in user_reserve_pages

[Konstantin Khlebnikov]

On 29.01.2015 23:11, Andrew Shewmaker wrote:
> On Thu, Jan 29, 2015 at 02:51:27PM +0300, Roman Gushchin wrote:
>> Shared file pages are never accounted in memory overcommit code,
>> so it isn't reasonable to count them in a code that limits the
>> maximal size of a process in OVERCOMMIT_NONE mode.
>>
>> If a process has few large file mappings, the consequent attempts
>> to allocate anonymous memory may unexpectedly fail with -ENOMEM,
>> while there is free memory and overcommit limit if significantly
>> larger than the committed amount (as displayed in /proc/meminfo).
>>
>> The problem is significantly smoothed by commit c9b1d0981fcc
>> ("mm: limit growth of 3% hardcoded other user reserve"),
>> which limits the impact of this check with 128Mb (tunable via sysctl),
>> but it can still be a problem on small machines.
>>
>> Signed-off-by: Roman Gushchin <klamm@yandex-team.ru>
>> Cc: Andrew Morton <akpm@linux-foundation.org>
>> Cc: Andrew Shewmaker <agshew@gmail.com>
>> Cc: Rik van Riel <riel@redhat.com>
>> Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
>> ---
>>   mm/mmap.c | 2 +-
>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/mm/mmap.c b/mm/mmap.c
>> index 7f684d5..151fadf 100644
>> --- a/mm/mmap.c
>> +++ b/mm/mmap.c
>> @@ -220,7 +220,7 @@ int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
>>   	 */
>>   	if (mm) {
>>   		reserve = sysctl_user_reserve_kbytes >> (PAGE_SHIFT - 10);
>> -		allowed -= min(mm->total_vm / 32, reserve);
>> +		allowed -= min((mm->total_vm - mm->shared_vm) / 32, reserve);
>>   	}
>>
>>   	if (percpu_counter_read_positive(&vm_committed_as) < allowed)
>> --
>> 2.1.0
>
> You're two patches conflict, don't they? Maybe you should resend
> them as a patch series such that they can both be applied?

I think arithmetic overflow is more important. Upper bound 128M
for user reserve makes mis-accounting of shared memory mostly invisible.

>
> Does mm->shared_vm include memory that's mapped MAP_ANONYMOUS in
> conjunction with MAP_SHARED? If so, then subtracting it could
> overcommit the system OVERCOMMIT_NEVER mode.

Yep.

Moreover shared_vm also includes file mappings with MAP_PRIVATE.
It works more likely as "maybe shared", upper bound for "file-rss"
(MM_FILEPAGES).

I think we need here total size of vmas where VM_ACCOUNT is set --
writable private mappings mapped without MAP_NORESERVE or something
like that. But total_vm after limiting with 128Mb gives almost always
the same or similar value. So, let's keep it as is.

-- 
Konstantin

>
> -Andrew
>