Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20200703061350.94474-1-songmuchun@bytedance.com>
In-Reply-To: <20200703061350.94474-1-songmuchun@bytedance.com>
From:   Pekka Enberg <penberg@gmail.com>
Date:   Fri, 3 Jul 2020 09:34:24 +0300
Message-ID: <CAOJsxLGc-o1Os7HAoQmhd1DNxGWVrfbw2R8NroTQg0wHNyWX1w@mail.gmail.com>
Subject: Re: [PATCH RESEND] mm/page_alloc: skip setting nodemask when we are
 in interrupt
To:     Muchun Song <songmuchun@bytedance.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>, david@redhat.com,
        "linux-mm@kvack.org" <linux-mm@kvack.org>,
        LKML <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Fri, Jul 3, 2020 at 9:14 AM Muchun Song <songmuchun@bytedance.com> wrote:
>
> When we are in the interrupt context, it is irrelevant to the
> current task context. If we use current task's mems_allowed, we
> can fair to alloc pages in the fast path and fall back to slow
> path memory allocation when the current node(which is the current
> task mems_allowed) does not have enough memory to allocate. In
> this case, it slows down the memory allocation speed of interrupt
> context. So we can skip setting the nodemask to allow any node
> to allocate memory, so that fast path allocation can success.
>
> Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> ---
>  mm/page_alloc.c | 8 +++++---
>  1 file changed, 5 insertions(+), 3 deletions(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index b48336e20bdcd..a6c36cd557d1d 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -4726,10 +4726,12 @@ static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
>
>         if (cpusets_enabled()) {
>                 *alloc_mask |= __GFP_HARDWALL;
> -               if (!ac->nodemask)
> -                       ac->nodemask = &cpuset_current_mems_allowed;
> -               else
> +               if (!ac->nodemask) {
> +                       if (!in_interrupt())
> +                               ac->nodemask = &cpuset_current_mems_allowed;

If !ac->nodemask and in_interrupt() the ALLOC_CPUSET flag is not set,
which by-passes the __cpuset_zone_allowed() check for allocations.
This works fine because in the case if in_interrupt() the function
allows allocation on any zone/node.

> +               } else {
>                         *alloc_flags |= ALLOC_CPUSET;
> +               }
>         }

However, if you write the condition as follows:

        if (cpusets_enabled()) {
                *alloc_mask |= __GFP_HARDWALL;
                if (!in_interrupt() && !ac->nodemask)
                        ac->nodemask = &cpuset_current_mems_allowed;
                else
                        *alloc_flags |= ALLOC_CPUSET;
        }

then the code is future-proof in case of __cpuset_zone_allowed() is
one day extended to support IRQ context too (it probably should
eventually respect IRQ SMP affinity).

>
>         fs_reclaim_acquire(gfp_mask);
> --
> 2.11.0
>
>