Re: [PATCH] mm/page_alloc: Let GFP_ATOMIC order-0 allocs access highatomic reserves

[Vlastimil Babka <vbabka@suse.cz>]

On 10/11/24 14:07, Matt Fleming wrote:
> From: Matt Fleming <mfleming@cloudflare.com>
> 
> Under memory pressure it's possible for GFP_ATOMIC order-0 allocations
> to fail even though free pages are available in the highatomic reserves.
> GFP_ATOMIC allocations cannot trigger unreserve_highatomic_pageblock()
> since it's only run from reclaim.
> 
> Given that such allocations will pass the watermarks in
> __zone_watermark_unusable_free(), it makes sense to fallback to
> highatomic reserves the same way that ALLOC_OOM can.
> 
> This fixes order-0 page allocation failures observed on Cloudflare's
> fleet when handling network packets:
> 
>   kswapd1: page allocation failure: order:0, mode:0x820(GFP_ATOMIC),
>   nodemask=(null),cpuset=/,mems_allowed=0-7
>   CPU: 10 PID: 696 Comm: kswapd1 Kdump: loaded Tainted: G           O 6.6.43-CUSTOM #1
>   Hardware name: MACHINE
>   Call Trace:
>    <IRQ>
>    dump_stack_lvl+0x3c/0x50
>    warn_alloc+0x13a/0x1c0
>    __alloc_pages_slowpath.constprop.0+0xc9d/0xd10
>    __alloc_pages+0x327/0x340
>    __napi_alloc_skb+0x16d/0x1f0
>    bnxt_rx_page_skb+0x96/0x1b0 [bnxt_en]
>    bnxt_rx_pkt+0x201/0x15e0 [bnxt_en]
>    __bnxt_poll_work+0x156/0x2b0 [bnxt_en]
>    bnxt_poll+0xd9/0x1c0 [bnxt_en]
>    __napi_poll+0x2b/0x1b0
>    bpf_trampoline_6442524138+0x7d/0x1000
>    __napi_poll+0x5/0x1b0
>    net_rx_action+0x342/0x740
>    handle_softirqs+0xcf/0x2b0
>    irq_exit_rcu+0x6c/0x90
>    sysvec_apic_timer_interrupt+0x72/0x90
>    </IRQ>
> 
> Suggested-by: Vlastimil Babka <vbabka@suse.cz>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Mel Gorman <mgorman@techsingularity.net>
> Cc: linux-mm@kvack.org
> Link: https://lore.kernel.org/all/CAGis_TWzSu=P7QJmjD58WWiu3zjMTVKSzdOwWE8ORaGytzWJwQ@mail.gmail.com/
> Signed-off-by: Matt Fleming <mfleming@cloudflare.com>

Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
(but a comment should be updated, see below)

I think we could add Cc: stable and I believe the commit that broke it was:

Fixes: 1d91df85f399 ("mm/page_alloc: handle a missing case for
memalloc_nocma_{save/restore} APIs")

because it was where an order > 0 condition was introduced to allow
allocation from MIGRATE_HIGHATOMIC

commit eb2e2b425c69 ("mm/page_alloc: explicitly record high-order atomic
allocations in alloc_flags") realized there's a gap for OOM (even if
changelog doesn't mention it) but we should allow the order-0 atomic
allocations to fallback as well.

> ---
>  mm/page_alloc.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 8afab64814dc..0c4c359f5ba7 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -2898,7 +2898,7 @@ struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
>  			 * failing a high-order atomic allocation in the
>  			 * future.
>  			 */

We should also update the comment above to reflect this is no longer just
for the OOM case?

> -			if (!page && (alloc_flags & ALLOC_OOM))
> +			if (!page && (alloc_flags & (ALLOC_OOM|ALLOC_NON_BLOCK)))
>  				page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
>  
>  			if (!page) {