From:   "Fabio M. De Francesco" <fmdefrancesco@gmail.com>
To:     Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Cc:     Peter Hurley <peter@hurleysoftware.com>,
        Jiri Slaby <jirislaby@kernel.org>,
        Max Filippov <jcmvbkbc@gmail.com>,
        David Sterba <dsterba@suse.com>,
        Bhaskar Chowdhury <unixbhaskar@gmail.com>,
        Igor Matheus Andrade Torrente <igormtorrente@gmail.com>,
        nick black <dankamongmen@gmail.com>,
        linux-kernel@vger.kernel.org,
        syzbot+5f47a8cea6a12b77a876@syzkaller.appspotmail.com,
        Marco Elver <elver@google.com>
Subject: Re: [PATCH] vt: Fix sleeping functions called from atomic context
Date:   Thu, 18 Nov 2021 09:31:06 +0100
Message-ID: <77983591.hkYTjcaLry@localhost.localdomain>
In-Reply-To: <df266c83-88df-4d1a-5c7e-ea0214f3de3b@i-love.sakura.ne.jp>
References: <20211116144937.19035-1-fmdefrancesco@gmail.com> <YZTDY/h8HcEkq7mO@kroah.com> <df266c83-88df-4d1a-5c7e-ea0214f3de3b@i-love.sakura.ne.jp>
MIME-Version: 1.0
Content-Transfer-Encoding: 7Bit
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wednesday, November 17, 2021 11:51:13 AM CET Tetsuo Handa wrote:
> On 2021/11/17 17:54, Greg Kroah-Hartman wrote:
> > Great, you have a reproducer, so you should be able to duplicate this
> > locally to figure out what is really happening here.
>
> Until commit ac751efa6a0d70f2 ("console: rename acquire/release_console_sem() to
> console_lock/unlock()"), do_con_write() was surely designed to be able to sleep.
> 
> > $ git blame ac751efa6a0d7~1 drivers/tty/vt/vt.c
>
> [...]
> 
> Until that commit, n_hdlc_send_frames() was prepared for being interrupted by signal
> while sleeping.
> 
> $ git blame ac751efa6a0d7~1 drivers/tty/n_hdlc.c
>
> [...]
>
> But as of commit c545b66c6922b002 ("tty: Serialize tcflow() with other tty flow
> control changes"), start_tty() was already holding spinlock.

Hi Tetsuo,

Actually, we don't care of start_tty(). It's not in the path that triggers sleeping in atomic bug.
According to Syzbot report and to my ftrace analysis it's __start_tty() that is called by 
n_tty_ioctl_helper(), and it is this function that acquires a spinlock and disables interrupts. 

I must admit that I've never used git-blame and I'm not sure to understand what you did here :(

Have you had a chance to read my analysis?
 
> $ git blame c545b66c6922b002~1 drivers/tty/tty_io.c
>
> [...]
> 
> Actually, it is commit f9e053dcfc02b0ad ("tty: Serialize tty flow control changes
> with flow_lock") that started calling tty->ops->start(tty) from atomic context.
> 
> $ git blame f9e053dcfc02b~1 drivers/tty/tty_io.c
>
> [...]
> 
> Therefore, I think that bisection will reach f9e053dcfc02b0ad, and I guess that
> this bug was not noticed simply because little people tested n_hdlc driver.
> 
> Well, how to fix? Introduce a new flag for indicating "starting" state (like drivers/block/loop.c uses Lo_* state) ?

I think this is not the correct fix, but I might very well be wrong...

Can you please reply to my last email (the one with the ftrace analysis)?
In the last lines I proposed two alternative solutions, what about them?

Thanks,

Fabio