Received-SPF: pass (google.com: domain of linux-kernel+bounces-168419-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Date: Sat, 4 May 2024 00:39:00 +0100
From: Al Viro <viro@zeniv.linux.org.uk>
To: Linus Torvalds <torvalds@linux-foundation.org>
Cc: keescook@chromium.org, axboe@kernel.dk, brauner@kernel.org,
	christian.koenig@amd.com, dri-devel@lists.freedesktop.org,
	io-uring@vger.kernel.org, jack@suse.cz, laura@labbott.name,
	linaro-mm-sig@lists.linaro.org, linux-fsdevel@vger.kernel.org,
	linux-kernel@vger.kernel.org, linux-media@vger.kernel.org,
	minhquangbui99@gmail.com, sumit.semwal@linaro.org,
	syzbot+045b454ab35fd82a35fb@syzkaller.appspotmail.com,
	syzkaller-bugs@googlegroups.com
Subject: Re: [PATCH] epoll: try to be a _bit_ better about file lifetimes
Message-ID: <20240503233900.GG2118490@ZenIV>
References: <202405031110.6F47982593@keescook>
 <20240503211129.679762-2-torvalds@linux-foundation.org>
 <20240503212428.GY2118490@ZenIV>
 <CAHk-=wjpsTEkHgo1uev3xGJ2bQXYShaRf3GPEqDWNgUuKx0JFw@mail.gmail.com>
 <20240503214531.GB2118490@ZenIV>
 <CAHk-=wgC+QpveKCJpeqsaORu7htoNNKA8mp+d9mvJEXmSKjhbw@mail.gmail.com>
 <20240503220145.GD2118490@ZenIV>
 <20240503220744.GE2118490@ZenIV>
 <CAHk-=whULchE1i5LA2Fa=ZndSAzPXGWh_e5+a=YV3qT1BEST7w@mail.gmail.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAHk-=whULchE1i5LA2Fa=ZndSAzPXGWh_e5+a=YV3qT1BEST7w@mail.gmail.com>
Sender: Al Viro <viro@ftp.linux.org.uk>

On Fri, May 03, 2024 at 04:16:15PM -0700, Linus Torvalds wrote:
> On Fri, 3 May 2024 at 15:07, Al Viro <viro@zeniv.linux.org.uk> wrote:
> >
> > Suppose your program calls select() on a pipe and dmabuf, sees data to be read
> > from pipe, reads it, closes both pipe and dmabuf and exits.
> >
> > Would you expect that dmabuf file would stick around for hell knows how long
> > after that?  I would certainly be very surprised by running into that...
> 
> Why?
> 
> That's the _point_ of refcounts. They make the thing they refcount
> stay around until it's no longer referenced.
> 
> Now, I agree that dmabuf's are a bit odd in how they use a 'struct
> file' *as* their refcount, but hey, it's a specialty use. Unusual
> perhaps, but not exactly wrong.
> 
> I suspect that if you saw a dmabuf just have its own 'refcount_t' and
> stay around until it was done, you wouldn't bat an eye at it, and it's
> really just the "it uses a struct file for counting" that you are
> reacting to.

*IF* those files are on purely internal filesystem, that's probably
OK; do that with something on something mountable (char device,
sysfs file, etc.) and you have a problem with filesystem staying
busy.

I'm really unfamiliar with the subsystem; it might be OK with all
objects that use that for ->poll(), but that's definitely not a good
thing to see in ->poll() instance in general.  And code gets copied,
so there really should be a big fat comment about the reasons why
it's OK in this particular case.

Said that, it seems that a better approach might be to have
their ->release() cancel callbacks and drop fence references.
Note that they *do* have refcounts - on fences.  The file
(well, dmabuf, really) is pinned only to protect against the
situation when pending callback is still around.  And Kees'
observation about multiple fences is also interesting - we don't
get extra fput(), but only because we get events only from one
fence, which does look fishy...