Date:   Thu, 11 Nov 2021 10:27:06 +0100
From:   Petr Mladek <pmladek@suse.com>
To:     Peter Zijlstra <peterz@infradead.org>,
        Sultan Alsawaf <sultan@kerneltoast.com>,
        Anton Vorontsov <anton@enomsg.org>,
        Ben Segall <bsegall@google.com>,
        Colin Cross <ccross@android.com>,
        Daniel Bristot de Oliveira <bristot@redhat.com>,
        David Airlie <airlied@linux.ie>,
        Dietmar Eggemann <dietmar.eggemann@arm.com>,
        dri-devel@lists.freedesktop.org, Ingo Molnar <mingo@redhat.com>,
        John Ogness <john.ogness@linutronix.de>,
        Juri Lelli <juri.lelli@redhat.com>,
        Kees Cook <keescook@chromium.org>,
        linux-kernel@vger.kernel.org,
        Maarten Lankhorst <maarten.lankhorst@linux.intel.com>,
        Maxime Ripard <mripard@kernel.org>,
        Mel Gorman <mgorman@suse.de>,
        Sergey Senozhatsky <senozhatsky@chromium.org>,
        Steven Rostedt <rostedt@goodmis.org>,
        Thomas Zimmermann <tzimmermann@suse.de>,
        Tony Luck <tony.luck@intel.com>,
        Vincent Guittot <vincent.guittot@linaro.org>, mkoutny@suse.com
Subject: Re: printk deadlock due to double lock attempt on current CPU's
 runqueue
Message-ID: <YYzh6rxX6tY8/d+L@alley>
References: <YYrU2PdmdNkulWSM@sultan-box.localdomain>
 <20211109213847.GY174703@worktop.programming.kicks-ass.net>
 <YYuS1uNhxWOEX1Ci@phenom.ffwll.local>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <YYuS1uNhxWOEX1Ci@phenom.ffwll.local>
Precedence: bulk

On Wed 2021-11-10 10:37:26, Daniel Vetter wrote:
> On Tue, Nov 09, 2021 at 10:38:47PM +0100, Peter Zijlstra wrote:
> > On Tue, Nov 09, 2021 at 12:06:48PM -0800, Sultan Alsawaf wrote:
> > > Hi,
> > > 
> > > I encountered a printk deadlock on 5.13 which appears to still affect the latest
> > > kernel. The deadlock occurs due to printk being used while having the current
 > > > CPU's runqueue locked, and the underlying framebuffer console attempting to lock
> > > the same runqueue when printk tries to flush the log buffer.
> > 
> > Yes, that's a known 'feature' of some consoles. printk() is in the
> > process of being reworked to not call con->write() from the printk()
> > calling context, which would go a long way towards fixing this.
> 
> And it's getting worse by the year:
> - direct scanout displays (i.e. just a few mmio writes and it will show
>   up) are on the way out at least in laptops, everyone gets self-refresh
>   (dp psr) under software control, so without being able to kick a kthread
>   off nothing shows up except more oopses

printk() tries to avoid this problem by waking processes via
irq_vork(). For example, it is used for waking userspace log
daemons, see wake_up_klogd().

There is one hidden wake_up_process() in up() of @console_sem.
It is "less" critical because it calls the wake_up() only when
there is another process waiting for console_lock(). It must
be a non-printk caller because printk() uses trylock and does
not add waiters. So, there is a non-trivial chance to avoid
the deadlock and the messages.

I wonder if fbdev might use the same trick.

> - because of the impendence mismatch between fbdev and drm-kms we even go
>   ever more this direction for dumb framebuffers, including the firmware
>   boot-up framebuffer simpledrm. This could perhaps be fixed with a new
>   dedicate console driver directly on top of drm-kms, but that's on the
>   wishlist for years and I don't see anyone typing that.
> 
> So yeah for fbcon at least I think we really should throw out direct
> con->write from printk completely.

Yup, that would be great.

Well, I guess that it would be available only for a panic mode when it
will not race with other users?

Best Regards,
Petr