[lkml]   [2019]   [Jul]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] kernel/printk: prevent deadlock at calling kmsg_dump from NMI context
On 15.07.2019 10:54, Petr Mladek wrote:
> On Mon 2019-07-15 11:33:38, Sergey Senozhatsky wrote:
>> On (07/13/19 17:03), Konstantin Khlebnikov wrote:
>>>> We call kmsg_dump(KMSG_DUMP_PANIC) after smp_send_stop()
>>> Indeed, panic is especially handled and looks fine.
> Just to get a picture. What other situations are we talking about,
> please?
> oops_exit() is one candidate. The other callers seem to be working
> in normal context.

Oops in NMI mostly. Also I've screwed up several times with NMI watchdog
and dumping log at setting taint.

>>> Sanity check in my patch could be relaxed:
>>> if (WARN_ON_ONCE(reason != KMSG_DUMP_PANIC && in_nmi()))
>>> return;
>> How critical kmsg_dump() is? We deadlock only if NMI->kmsg_dump()
>> happens on the CPU which already holds the logbuf_lock; in any
>> other case logbuf_lock is owned by another CPU which is expected
>> to unlock it eventually. So it doesn't look like disabling all
>> NMI->kmsg_dump() is the only way to fix it.
>> When we lock logbuf_lock we increment per-CPU printk_context
>> (PRINTK_SAFE_CONTEXT_MASK bits); when we unlock logbuf_lock
>> we decrement printk_context. Thus we always can tell if the
>> logbuf_lock was locked on the very same CPU - this_cpu printk_context
>> has PRINTK_SAFE_CONTEXT_MASK bits sets - and we are about to deadlock
>> in a nested context (NMI), or the lock was locked on another CPU and
>> it's "safe" to spin on logbuf_lock and wait for logbuf_lock to become
>> available.
> This sounds familiar. I think that we did not consider it safe in the
> end, see the commit 03fc7f9c99c1e7ae29 ("printk/nmi: Prevent deadlock
> when accessing the main log buffer in NMI").
> If the problem is only with Oops then the 2nd propose might be
> acceptable. The system will either try to continue or it will end
> up in panic() anyway.
> Well, WARN() looks like an overkill, especially if there is only
> one possible caller that prints the stack anyway. I would personally
> do not print any message and do just:
> /*
> * Prevent deadlock on logbuf_lock. It is safe only
> * in panic() after smp_send_stop() and resetting
> * the lock.
> */
> if (in_nmi() && reason != KMSG_DUMP_PANIC)
> return;

WARN_ON_ONCE is useful timesaver in debugging.
It's better to know when happens something that shouldn't.

> Best Regards,
> Petr

 \ /
  Last update: 2019-07-15 10:13    [W:0.059 / U:7.624 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site