MIME-Version: 1.0
References: <1638952658-20285-1-git-send-email-huangzhaoyang@gmail.com>
 <2868725.1638995206@warthog.procyon.org.uk> <CAGWkznHcuiwPPMZE95nYG=EFkM8NmLUQZooS5+a+GigP50qksg@mail.gmail.com>
 <73896.1640098820@warthog.procyon.org.uk>
In-Reply-To: <73896.1640098820@warthog.procyon.org.uk>
From:   Zhaoyang Huang <huangzhaoyang@gmail.com>
Date:   Wed, 22 Dec 2021 16:02:18 +0800
Message-ID: <CAGWkznGgyS5VrcuYkWR_7sbDOkr0k2mDNUwF6F6N-Y_3GGtoJA@mail.gmail.com>
Subject: Re: [PATCH] fs: judging context via current_is_kswapd instead of gfp_flag
To:     David Howells <dhowells@redhat.com>
Cc:     Marc Dionne <marc.dionne@auristor.com>,
        Zhaoyang Huang <zhaoyang.huang@unisoc.com>,
        linux-cachefs@redhat.com, LKML <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Dec 21, 2021 at 11:01 PM David Howells <dhowells@redhat.com> wrote:
>
> Zhaoyang Huang <huangzhaoyang@gmail.com> wrote:
>
> > > > -             if (!(gfp_flags & __GFP_DIRECT_RECLAIM) || !(gfp_flags & __GFP_FS))
> > > > +             if (current_is_kswapd() || !(gfp_flags & __GFP_FS))
> > > >                       return false;
> > > >               wait_on_page_fscache(page);
> > ...
> > If the gfp flag here is used for judging kswapd context, I think the
> > answer is yes as kswapd applied __GFP_DIRECT_RECLAIM.
>
> Now I come to look at applying it, I'm not sure whether this change is right.
>
> I don't know if kswapd has anything to do with it.  The check is to see if
> we're allowed to block at this point - and wait is just for the completion of
> a DIO write to disk.
>
> It would seem like gfpflags_allow_blocking() is the right thing to call - and
> that should use current_is_kswapd() if appropriate.
>
> David
According to my understanding, this check is redundant according to
current vmscan logic. For the allocation which deny
__GFP_DIRECT_RECLAIM could NOT have the context reach here as there is
no synchronous reclaiming. while kswapd also has __GFP_DIRECT_RECLAIM
set and would also block on the page's release.
>