Re: [PATCH RFC 02/19] slab: handle pfmemalloc slabs properly with sheaves

[Chris Mason <clm@meta.com>]

On 10/29/25 11:00 AM, Vlastimil Babka wrote:
> On 10/24/25 16:21, Chris Mason wrote:
>> On Thu, 23 Oct 2025 15:52:24 +0200 Vlastimil Babka <vbabka@suse.cz> wrote:
>>> @@ -5497,7 +5528,7 @@ int kmem_cache_refill_sheaf(struct kmem_cache *s, gfp_t gfp,
>>>
>>>  	if (likely(sheaf->capacity >= size)) {
>>>  		if (likely(sheaf->capacity == s->sheaf_capacity))
>>> -			return refill_sheaf(s, sheaf, gfp);
>>> +			return __prefill_sheaf_pfmemalloc(s, sheaf, gfp);
>>>
>>>  		if (!__kmem_cache_alloc_bulk(s, gfp, sheaf->capacity - sheaf->size,
>>>  					     &sheaf->objects[sheaf->size])) {
>>                                              ^^^
>>
>> In kmem_cache_refill_sheaf(), does the oversize sheaf path (when
>> sheaf->capacity != s->sheaf_capacity) need __GFP_NOMEMALLOC too?
>>
>> The commit message says "When refilling sheaves, use __GFP_NOMEMALLOC
>> to override any pfmemalloc context", and the normal capacity path now
>> calls __prefill_sheaf_pfmemalloc() which adds __GFP_NOMEMALLOC.
>>
>> But this oversize path still calls __kmem_cache_alloc_bulk() with the
>> raw gfp flags. If the calling context is pfmemalloc-enabled (e.g.,
>> during swap or network operations), could pfmemalloc objects be
>> allocated into the sheaf? Those objects would then be returned via
> 
> Yes.
> 
>> kmem_cache_alloc_from_sheaf() to potentially non-pfmemalloc callers.
> 
> The assumption is the caller will use the prefilled sheaf for its purposes
> and not pass it to other callers. The reason for caring about pfmemalloc and
> setting sheaf->pfmemalloc is only to recognize them when the prefilled sheaf
> is returned - so that it's flushed+freed and not attached as pcs->spare -
> that would then be available to other non-pfmemalloc callers.
> 
> But we always flush oversize sheaves when those are returned, so it's not
> necessary to also track pfmemalloc for them. I'll add a comment about it.

Oh I see, this makes sense.  Thanks!

-chris