Re: [PATCH RFC] ext4: use __GFP_NOFAIL in ext4_free_blocks()

[Nikolay Borisov]

On 02/25/2016 11:08 AM, Michal Hocko wrote:
> On Thu 25-02-16 11:01:32, Nikolay Borisov wrote:
>>
>>
>> On 02/24/2016 07:09 PM, Konstantin Khlebnikov wrote:
>>> This might be unexpected but pages allocated for sbi->s_buddy_cache are
>>> charged to current memory cgroup. So, GFP_NOFS allocation could fail if
>>> current task has been killed by OOM or if current memory cgroup has no
>>> free memory left. Block allocator cannot handle such failures here yet.
>>>
>>> Signed-off-by: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>
>>
>> Adding new users of GFP_NOFAIL is deprecated.
> 
> This is not true. GFP_NOFAIL should be used where the allocation failure
> is no tolleratable and it is much more preferrable to doing an opencoded
> endless loop over page allocator.

In that case the comments in buffered_rmqueue, and the WARN_ON in
__alloc_pages_may_oom and __alloc_pages_slowpath perhaps should be
removed since they are misleading?

> 
>> Where exactly does the
>> block allocator fail, I skimmed the code and failing ext4_mb_load_buddy
>> seems to be handled at all call sites. There are some BUG_ONs but from
>> the comments there I guess they should occur when we try to find a page
>> and not allocate a new one?
> 
> I have posted a similar patch last year:
> http://lkml.kernel.org/r/1438768284-30927-6-git-send-email-mhocko@xxxxxxxxxx
> because I could see emergency reboots when GFP_NOFS allocations were
> allowed to fail.
>