Re: [PATCH v3] mm: slub: avoid wake up kswapd in set_track_prepare

[Vlastimil Babka <vbabka@suse.cz>]

On 8/27/25 10:00, Vlastimil Babka wrote:
> On 8/27/25 07:17, Harry Yoo wrote:
>> On Mon, Aug 25, 2025 at 05:42:52PM +0200, Vlastimil Babka wrote:
>>> On 8/25/25 14:40, Matthew Wilcox wrote:
>>> > On Mon, Aug 25, 2025 at 08:17:37PM +0800, yangshiguang1011@163.com wrote:
>>> >> Avoid deadlock caused by implicitly waking up kswapd by
>>> >> passing in allocation flags.
>>> > [...]
>>> >> +	/* Preemption is disabled in ___slab_alloc() */
>>> >> +	gfp_flags &= ~(__GFP_DIRECT_RECLAIM);
>>> > 
>>> > If you don't mean __GFP_KSWAPD_RECLAIM here, the explanation needs to
>>> > be better.
>>> 
>>> It was suggested by Harry here:
>>> https://lore.kernel.org/all/aKKhUoUkRNDkFYYb@harry
>>> 
>>> I think the comment is enough? Disabling preemption means we can't direct
>>> reclaim, but we can wake up kswapd. If the slab caller context is such that
>>> we can't, __GFP_KSWAPD_RECLAIM already won't be in the gfp_flags.
>> 
>> To make it a little bit more verbose, this ^^ explanation can be added to the
>> changelog?
>> 
>>> But I think we should mask our also __GFP_NOFAIL and add __GFP_NOWARN?
>> 
>> That sounds good.
>> 
>>> (we should get some common helpers for these kinds of gfp flag manipulations
>>> already)
>> 
>> Any ideas for its name?
>> 
>> gfp_dont_try_too_hard(),
>> gfp_adjust_lightweight(),
>> gfp_adjust_mayfail(),
>> ...
>> 
>> I'm not good at naming :/
> 
> Looks like there's already gfp_nested_mask() for these purposes. I'm not
> sure if it should be allowing GFP_ATOMIC (thus __GFP_HIGH) as it does
> though. Seems to contradict the comment about not exhausing reserves. Wonder
> if that was raised during review...

Looks like I mentioned it but inconclusively.

https://lore.kernel.org/linux-xfs/1fb5b8f3-d8c7-4350-888a-ad8f4d54bc66@suse.cz/

Anyway that's orthogonal to using the helper here right now.

> The masking out of __GFP_DIRECT_RECLAIM is specific to the slab case so we
> don't need a helper for that (unless we find other users). It could be then
> e.g. gfp_nested_mask_noblock() ?
>