Re: [PATCH] tmpfs: don't interrupt fallocate with EINTR

From: Christian Brauner <brauner@kernel.org>
To: Mikulas Patocka <mpatocka@redhat.com>
Cc: Hugh Dickins <hughd@google.com>,
	 Alexander Viro <viro@zeniv.linux.org.uk>,
	Jan Kara <jack@suse.cz>,
	linux-fsdevel@vger.kernel.org,  linux-mm@kvack.org
Subject: Re: [PATCH] tmpfs: don't interrupt fallocate with EINTR
Date: Thu, 7 Mar 2024 11:47:01 +0100	[thread overview]
Message-ID: <20240307-entkriminalisierung-wankt-1187288fbb27@brauner> (raw)
In-Reply-To: <a3834f3-8beb-988a-e387-5c8d31e013f@redhat.com>

On Tue, Mar 05, 2024 at 03:03:53PM +0100, Mikulas Patocka wrote:
> 
> 
> On Tue, 5 Mar 2024, Christian Brauner wrote:
> 
> > On Tue, Mar 05, 2024 at 10:34:26AM +0100, Mikulas Patocka wrote:
> > > 
> > > 
> > > On Tue, 5 Mar 2024, Christian Brauner wrote:
> > > 
> > > > On Mon, Mar 04, 2024 at 07:43:39PM +0100, Mikulas Patocka wrote:
> > > > > 
> > > > > Index: linux-2.6/mm/shmem.c
> > > > > ===================================================================
> > > > > --- linux-2.6.orig/mm/shmem.c	2024-01-18 19:18:31.000000000 +0100
> > > > > +++ linux-2.6/mm/shmem.c	2024-03-04 19:05:25.000000000 +0100
> > > > > @@ -3143,7 +3143,7 @@ static long shmem_fallocate(struct file
> > > > >  		 * Good, the fallocate(2) manpage permits EINTR: we may have
> > > > >  		 * been interrupted because we are using up too much memory.
> > > > >  		 */
> > > > > -		if (signal_pending(current))
> > > > > +		if (fatal_signal_pending(current))
> > > > 
> > > > I think that's likely wrong and probably would cause regressions as
> > > > there may be users relying on this?
> > > 
> > > ext4 fallocate doesn't return -EINTR. So, userspace code can't rely on it.
> > 
> > I'm confused what does this have to do with ext4 since this is about
> > tmpfs.
> 
> You said that applications may rely on -EINTR and I said they don't 
> because ext4 doesn't return -EINTR.
> 
> > Also note, that fallocate(2) documents EINTR as a valid return
> > value. And fwiw, the manpage also states that "EINTR  A signal was
> > caught during execution; see signal(7)." not a "fatal signal".
> 
> Yes, but how should the userspace use the fallocate call reliably? Block 
> all the signals around the call to fallocate? What to do if I use some 
> library that calls fallocate and retries on EINTR?
> 
> > Aside from that. If a user sends SIGUSR1 then with the code as it is now
> > that fallocate call will be interrupted. With your change that SIGUSR1
> > won't do anything anymore. Instead userspace would need to send SIGKILL.
> > So userspace that uses SIGUSR1 will suddenly hang.
> 
> It will survive one SIGUSR, but it hangs if the signal is being sent at a 
> periodic interval.
> 
> A quick search shows that people are already adding loops when fallocate 
> returns EINTR. All these loops will livelock when a signal is repeatedly 
> being delivered: 
> https://forge.chapril.org/hardcoresushi/libgocryptfs/commit/8518d6d7bde33fdc7ef5bcb7c3c7709404392ad8?style=unified&whitespace= 
> https://postgrespro.com/media/maillist-attaches/pgsql-hackers/2022/07/1/20220701154105.jjfutmngoedgiad3@alvherre.pgsql/v2-0001-retry-ftruncate.patch 
> https://lists.nongnu.org/archive/html/qemu-devel/2015-02/msg01116.html
> 
> Here, Postgres developers hit the same problem with retrying (they have 
> 5ms timer):
> https://www.postgresql.org/message-id/CA%2BhUKGKS2Radu-1Ewhe1-LEj19C-3XAQ7wnkQMb4e9E9q9ZXSg%40mail.gmail.com

All fair points.