Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20220511222910.635307-1-dlunev@chromium.org> <20220512082832.v2.2.I692165059274c30b59bed56940b54a573ccb46e4@changeid>
 <YoQfls6hFcP3kCaH@zeniv-ca.linux.org.uk> <YoQihi4OMjJj2Mj0@zeniv-ca.linux.org.uk>
 <CAJfpegtQUP045X5N8ib1rUTKzSj-giih0eL=jC5-MP7aVgyN_g@mail.gmail.com>
 <CAONX=-do9yvxW2gTak0WGbFVPiLbkM2xH5LReMZkvC-upOUVxg@mail.gmail.com>
 <CAONX=-ehh=uGYAi++oV_uS23mp2yZcrUC+7U5H0rRz8q0h6OeQ@mail.gmail.com> <CAJfpegsPjFMCG-WHbvREZXzHPUd1R2Qa83maiTJbWSua9Kz=hg@mail.gmail.com>
In-Reply-To: <CAJfpegsPjFMCG-WHbvREZXzHPUd1R2Qa83maiTJbWSua9Kz=hg@mail.gmail.com>
From:   Daniil Lunev <dlunev@chromium.org>
Date:   Wed, 25 May 2022 08:44:50 +1000
Message-ID: <CAONX=-d-6Bm9qGQyhmpkMQov+wjgH3+pAMHdSsifM6FpmHNPFg@mail.gmail.com>
Subject: Re: [PATCH v2 2/2] FUSE: Retire superblock on force unmount
To:     Miklos Szeredi <miklos@szeredi.hu>,
        Al Viro <viro@zeniv.linux.org.uk>
Cc:     linux-fsdevel@vger.kernel.org,
        Christoph Hellwig <hch@infradead.org>,
        fuse-devel <fuse-devel@lists.sourceforge.net>,
        "Theodore Ts'o" <tytso@mit.edu>, linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

> Calling bdi_unregister() might be okay, and that should fix this.  I'm
> not familiar enough with that part to say for sure.
> But freeing sb->s_bdi while the superblock is active looks problematic.
Tracing the code, I see that, yes, in general that might not be safe to call
the "bdi_put" function for any FS - because it might have in-flight ops even
with force, where they will routinely access the members of the bdi struct
without locks. However, we do replace the struct with a no_op version,
and specifically for the FUSE case we sever the connection first, so no
in-flight ops can actually be there. And I am not sure if other FS may
need to do this retirement, given they don't have these
user-space/kernel split. It would make sense however to delay the actual
put till  the actual super block destruction, but that would require
introducing extra state tracking to see if the block is or is not registered
anymore. It can be piggybacked on the v1 patch where I have an explicit
state flag, but not on v2.
Miklos, Al, will the following work and be acceptable?
Get v1 patchset[1], in fuse_umount_begin do bdi_unregister and set
the flag, but do not do bdi_put or replacement with no_op. And then in
generic shutdown super if the bdi is not no_op and the 'defunct' flag is
set, skip unregister part.
Thanks,
Daniil

[1]
https://lore.kernel.org/linux-fsdevel/20220511013057.245827-1-dlunev@chromium.org/T/#u