Message-ID: <3ad45074-72c6-8844-ba00-caa95e11ea7c@oracle.com>
Date:   Tue, 15 Feb 2022 15:24:50 -0800
From:   Stephen Brennan <stephen.s.brennan@...cle.com>
To:     Petr Mladek <pmladek@...e.com>
Cc:     Sergey Senozhatsky <senozhatsky@...omium.org>,
        linux-kernel@...r.kernel.org, Steven Rostedt <rostedt@...dmis.org>,
        John Ogness <john.ogness@...utronix.de>
Subject: Re: [PATCH v4 0/4] printk: reduce deadlocks during panic

On 2/14/22 05:54, Petr Mladek wrote:
> On Thu 2022-02-10 12:06:44, Stephen Brennan wrote:
>> On 2/10/22 01:22, Petr Mladek wrote:
>>> On Wed 2022-02-02 09:18:17, Stephen Brennan wrote:
>>>> When a caller writes heavily to the kernel log (e.g. writing to
>>>> /dev/kmsg in a loop) while another panics, there's currently a high
>>>> likelihood of a deadlock (see patch 2 for the full description of this
>>>> deadlock).
>>>>
>>>> The principle fix is to disable the optimistic spin once panic_cpu is
>>>> set, so the panic CPU doesn't spin waiting for a halted CPU to hand over
>>>> the console_sem.
>>>>
>>>> However, this exposed us to a livelock situation, where the panic CPU
>>>> holds the console_sem, and another CPU could fill up the log buffer
>>>> faster than the consoles could drain it, preventing the panic from
>>>> progressing and halting the other CPUs. To avoid this, patch 3 adds a
>>>> mechanism to suppress printk (from non-panic-CPU) during panic, if we
>>>> reach a threshold of dropped messages.
>>>>
>>>> A major goal with all of these patches is to try to decrease the
>>>> likelihood that another CPU is holding the console_sem when we halt it
>>>> in panic(). This reduces the odds of needing to break locks and
>>>> potentially encountering further deadlocks with the console drivers.
>>>>
>>>> To test, I use the following script, kmsg_panic.sh:
>>>>
>>>>       #!/bin/bash
>>>>       date
>>>>       # 991 chars (based on log buffer size):
>>>>       chars="$(printf 'a%.0s' {1..991})"
>>>>       while :; do
>>>>           echo $chars > /dev/kmsg
>>>>       done &
>>>>       echo c > /proc/sysrq-trigger &
>>>>       date
>>>>       exit
>>>>
>>>> I defined a hang as any time the system did not reboot to a login prompt
>>>> on the serial console within 60 seconds. Here are the statistics on
>>>> hangs using this script, before and after the patch.
>>>>
>>>> before:  776 hangs / 1484 trials - 52.3%
>>>> after :    0 hangs /  15k trials -  0.0%
>>>>
>>>> Stephen Brennan (4):
>>>>     printk: Add panic_in_progress helper
>>>>     printk: disable optimistic spin during panic
>>>>     printk: Avoid livelock with heavy printk during panic
>>>>     printk: Drop console_sem during panic
>>>>
>>>>    kernel/printk/printk.c | 55 +++++++++++++++++++++++++++++++++++++++++-
>>>>    1 file changed, 54 insertions(+), 1 deletion(-)
>>>
>>> For the entire patchset:
>>>
>>> Reviewed-by: Petr Mladek <pmladek@...e.com>
>>>
>>> It looks ready for linux-next from my POV. I am going to push it early
>>> next week unless anyone complains in the meantime.
> 
> The patchset is committed in printk/linux.git, branch for-5.18-panic-deadlocks.

Thank you!

> 
>> Thank you Petr! It occurs to me that some of this could be stable-worthy,
>> depending on your feelings on it. Patches 1-3 resolve real bugs on customer
>> systems, and they'd apply back a decent way. 1-2 apply all the way back to
>> 4.14, and 3 would apply with some minor changes. I suppose the question is
>> whether they are simple enough. Patch 4 is useful but I don't have a real
>> reproducer for a bug it fixes, so I wouldn't say it's stable worthy.
> 
> Good question. If you saw these deadlocks on customer systems in
> the real life then it might be worth it.
> 
> I newer saw them. But they hard to debug and report. Also they are
> visible only when CPUs are stopped by NMI. And the default
> smp_send_stop() tries to stop CPUs using normal IRQ first.

Some of our default configurations like to ship with 
crash_kexec_post_notifiers=1, so I see the aggressive 
crash_smp_send_stop() code path a lot internally.

And interestingly, Hyper-V enables crash_kexec_post_notifiers in certain 
cases, without so much as a peep in the kernel log, even if kdump is not 
enabled... So folks running on Azure or otherwise running Hyper-V guests 
would be exposed to this. I actually just saw some patches related to 
this which you reviewed, so I guess you're aware of that :)

https://lore.kernel.org/lkml/YgDZX4PYwhrA1+Ct@MiWiFi-R3L-srv/

As somebody who mainly browses subsystem-focused mailing lists I didn't 
see this at all, as these printk/panic related patches only seem to go 
to LKML. I guess I need to check out the new lore+lei system so I can 
keep track of relevant work.

> 
> Anyway, the patches 1,2,4 are pretty straightforward and should be
> safe. Feel free to send them to stable.

Very reasonable. The customer systems experiencing the issue would be 
resolved by patches 1-2 alone. I did not have a customer experiencing 
any sort of livelock (and it would be a pretty ridiculous situation, 
most printk dies down quickly). I see the value of patch 4 for 
architectures without these NMI so I can send it too.

> 
> 3rd patch is a heuristic. It tries to prevent livelock and the cost
> is a possible loss of information. I am not 100% sure that it will
> do the right thing in all situations. I would wait one or two release
> cycles before we backport it to older stable releases.

Fair enough!

> 
>> Of course we have the logbuf_lock in 5.10 and previous, and if a CPU is
>> halted holding that lock, then printk hangs even before the optimistic
>> spinning. I have patches which reinitialize those locks after the CPUs are
>> halted if necessary. I think they are reasonable for stable - printk is
>> guaranteed to hang without doing this, so in the worst case you trade a hang
>> during a panic, with some other sort of printk log buffer bug during a
>> panic. But in the common case, you eliminate the hang. I can send that patch
>> to linux-stable as well.
> 
> The main problem is that the locks can be safely re-initialized only
> when the other CPUs were stopped using NMI. Otherwise, there is
> a risk of double unlock. Such a patch would need to be arch-dependent.

That's a very good point, thank you for mentioning it.

> 
> Also stable people do not like much solutions that were not used
> in the mainline. So, it might be a waste of time.

True. I may try my hand at it simply because I would much rather see 
work be done for all via -stable rather than in vendor-specific trees.

Thanks,
Stephen

> 
> Best Regards,
> Petr

Powered by blists - more mailing lists