Re: [PATCH] mm/failslab: By default, do not fail allocations with direct reclaim only

[Akinobu Mita]

2019å5æ20æ(æ) 13:49 Nicolas Boichat <drinkcat@xxxxxxxxxxxx>:
>
> When failslab was originally written, the intention of the
> "ignore-gfp-wait" flag default value ("N") was to fail
> GFP_ATOMIC allocations. Those were defined as (__GFP_HIGH),
> and the code would test for __GFP_WAIT (0x10u).
>
> However, since then, __GFP_WAIT was replaced by __GFP_RECLAIM
> (___GFP_DIRECT_RECLAIM|___GFP_KSWAPD_RECLAIM), and GFP_ATOMIC is
> now defined as (__GFP_HIGH|__GFP_ATOMIC|__GFP_KSWAPD_RECLAIM).
>
> This means that when the flag is false, almost no allocation
> ever fails (as even GFP_ATOMIC allocations contain
> __GFP_KSWAPD_RECLAIM).
>
> Restore the original intent of the code, by ignoring calls
> that directly reclaim only (___GFP_DIRECT_RECLAIM), and thus,
> failing GFP_ATOMIC calls again by default.
>
> Fixes: 71baba4b92dc1fa1 ("mm, page_alloc: rename __GFP_WAIT to __GFP_RECLAIM")
> Signed-off-by: Nicolas Boichat <drinkcat@xxxxxxxxxxxx>

Good catch.

Reviewed-by: Akinobu Mita <akinobu.mita@xxxxxxxxx>

> ---
>  mm/failslab.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/failslab.c b/mm/failslab.c
> index ec5aad211c5be97..33efcb60e633c0a 100644
> --- a/mm/failslab.c
> +++ b/mm/failslab.c
> @@ -23,7 +23,8 @@ bool __should_failslab(struct kmem_cache *s, gfp_t gfpflags)
>         if (gfpflags & __GFP_NOFAIL)
>                 return false;
>
> -       if (failslab.ignore_gfp_reclaim && (gfpflags & __GFP_RECLAIM))
> +       if (failslab.ignore_gfp_reclaim &&
> +                       (gfpflags & ___GFP_DIRECT_RECLAIM))
>                 return false;

Should we use __GFP_DIRECT_RECLAIM instead of ___GFP_DIRECT_RECLAIM?
Because I found the following comment in gfp.h

/* Plain integer GFP bitmasks. Do not use this directly. */