Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Mon, 10 Feb 2020 14:51:18 -0800
From:   Andrew Morton <akpm@linux-foundation.org>
To:     Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc:     Petr Mladek <pmladek@suse.com>,
        Peter Zijlstra <peterz@infradead.org>,
        linux-kernel@vger.kernel.org, Steven Rostedt <rostedt@goodmis.org>,
        Sergey Senozhatsky <sergey.senozhatsky@gmail.com>,
        Dmitry Monakhov <dmtrmonakhov@yandex-team.ru>
Subject: Re: [PATCH] kernel/watchdog: flush all printk nmi buffers when
 hardlockup detected
Message-Id: <20200210145118.1d80e248c9206aeafd5baae6@linux-foundation.org>
In-Reply-To: <158132813726.1980.17382047082627699898.stgit@buzz>
References: <158132813726.1980.17382047082627699898.stgit@buzz>
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Mon, 10 Feb 2020 12:48:57 +0300 Konstantin Khlebnikov <khlebnikov@yandex-team.ru> wrote:

> In NMI context printk() could save messages into per-cpu buffers and
> schedule flush by irq_work when IRQ are unblocked. This means message
> about hardlockup appears in kernel log only when/if lockup is gone.

I think I understand what this means.  The hard lockup detector runs at
NMI time but if it detects a lockup within IRQ context it cannot call
printk, because it's within NMI context, where synchronous printk
doesn't work.  Yes?

> Comment in irq_work_queue_on() states that remote IPI aren't NMI safe
> thus printk() cannot schedule flush work to another cpu.
> 
> This patch adds simple atomic counter of detected hardlockups and
> flushes all per-cpu printk buffers in context softlockup watchdog
> at any other cpu when it sees changes of this counter.

And I think this works because the softlockup detector runs within irq
context?

>
> ...
>
> --- a/kernel/watchdog.c
> +++ b/kernel/watchdog.c
> @@ -92,6 +92,26 @@ static int __init hardlockup_all_cpu_backtrace_setup(char *str)
>  }
>  __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
>  # endif /* CONFIG_SMP */
> +
> +atomic_t hardlockup_detected = ATOMIC_INIT(0);
> +
> +static inline void flush_hardlockup_messages(void)

I don't think this needs to be inlined?

> +{
> +	static atomic_t flushed = ATOMIC_INIT(0);
> +
> +	/* flush messages from hard lockup detector */
> +	if (atomic_read(&hardlockup_detected) != atomic_read(&flushed)) {
> +		atomic_set(&flushed, atomic_read(&hardlockup_detected));
> +		printk_safe_flush();
> +	}
> +}

Could we add some explanatory comments here?  Explain to the reader why
this code exists, what purpose it serves?  Basically a micro version of
the above changelog.

>
> ...
>