On 2019/09/11 22:52, Hillf Danton wrote:
>
> On Wed, 11 Sep 2019 19:07:34 +0900
>>
>> But I guess that there is a problem.
>
> Not a new one. (see commit 7dea19f9ee63)
>
>> Setting PF_MEMALLOC_NOIO causes
>> current_gfp_context() to mask __GFP_IO | __GFP_FS, but the OOM killer cannot
>> be invoked when __GFP_FS is masked. As a result, any userspace thread which
>> has PF_MEMALLOC_NOIO cannot invoke the OOM killer.
>
> Correct.
>
>> If the userspace thread
>> which uses PF_MEMALLOC_NOIO is involved in memory reclaiming activities,
>> the memory reclaiming activities won't be able to make forward progress when
>> the userspace thread triggered e.g. a page fault. Can the "userspace components
>> that can run in the IO path" survive without any memory allocation?
>
> Good question.
>
> It can be solved without oom killer involved because user should be
> aware of the risk of PF_MEMALLOC_NOIO if they ask for the convenience.
> OTOH we are able to control any abuse of it as you worry, knowing that
> the combination of __GFP_FS and oom killer can not get more than 50 users
> works done, and we have to pay as much attention as we can to the decisions
> they make. In case of PF_MEMALLOC_NOIO, we simply fail the allocation
> rather than killing a random victim.
According to commit c288983dddf71421 ("mm/page_alloc.c: make sure OOM victim can
try allocations with no watermarks once"), memory allocation failure from a page
fault results in invocation of the OOM killer via pagefault_out_of_memory() which
after all kills a random victim.
>
>
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -3854,6 +3854,8 @@ __alloc_pages_may_oom(gfp_t gfp_mask, un
> * out_of_memory). Once filesystems are ready to handle allocation
> * failures more gracefully we should just bail out here.
> */
> + if (current->flags & PF_MEMALLOC_NOIO)
> + goto out;
>
> /* The OOM killer may not free memory on a specific node */
> if (gfp_mask & __GFP_THISNODE)
>
>
