Re: [PATCH v2 07/17] debugobjects: Move printk out of db lock critical sections

From: Waiman Long
Date: Thu Nov 22 2018 - 15:29:42 EST


On 11/22/2018 11:02 AM, Petr Mladek wrote:
> On Thu 2018-11-22 11:04:22, Sergey Senozhatsky wrote:
>> On (11/21/18 11:49), Waiman Long wrote:
>> [..]
>>>> case ODEBUG_STATE_ACTIVE:
>>>> - debug_print_object(obj, "init");
>>>> state = obj->state;
>>>> raw_spin_unlock_irqrestore(&db->lock, flags);
>>>> + debug_print_object(obj, "init");
>>>> debug_object_fixup(descr->fixup_init, addr, state);
>>>> return;
>>>>
>>>> case ODEBUG_STATE_DESTROYED:
>>>> - debug_print_object(obj, "init");
>>>> + debug_printobj = true;
>>>> break;
>>>> default:
>>>> break;
>>>> }
>>>>
>>>> raw_spin_unlock_irqrestore(&db->lock, flags);
>>>> + if (debug_chkstack)
>>>> + debug_object_is_on_stack(addr, onstack);
>>>> + if (debug_printobj)
>>>> + debug_print_object(obj, "init");
>>>>
>> [..]
>>> As a side note, one of the test systems that I used generated a
>>> debugobjects splat in the bootup process and the system hanged
>>> afterward. Applying this patch alone fix the hanging problem and the
>>> system booted up successfully. So it is not really a good idea to call
>>> printk() while holding a raw spinlock.
> Please, was the system hang reproducible? I wonder if it was a
> deadlock described by Sergey below.

Yes, it is 100% reproducible on the testing system that I used.

> The commit message is right. printk() might take too long and
> cause softlockup or livelock. But it does not explain why
> the system could competely hang.
>
> Also note that prinkt() should not longer block a single process
> indefinitely thanks to the commit dbdda842fe96f8932 ("printk:
> Add console owner and waiter logic to load balance console writes").

The problem might have been caused by the fact that IRQ was also
disabled in the lock critical section.

>> Some serial consoles call mod_timer(). So what we could have with the
>> debug objects enabled was
>>
>> mod_timer()
>> lock_timer_base()
>> debug_activate()
>> printk()
>> call_console_drivers()
>> foo_console()
>> mod_timer()
>> lock_timer_base() << deadlock
> Anyway, I wonder what was the primary motivation for this patch.
> Was it the system hang? Or was it lockdep report about nesting
> two terminal locks: db->lock, pool_lock with logbuf_lock?

The primary motivation was to make sure that printk() won't be called
while holding either db->lock or pool_lock in the debugobjects code. In
the determination of which locks can be made terminal, I focused on
local spinlocks that won't cause boundary to an unrelated subsystem as
it will greatly complicate the analysis.

I didn't realize that it fixed a hang problem that I was seeing until I
did bisection to find out that it was caused by the patch that cause the
debugobjects splat in the first place a few days ago. I was comparing
the performance status of the pre and post patched kernel. The pre-patch
kernel failed to boot in the one of my test systems, but the
post-patched kernel could. I narrowed it down to this patch as the fix
for the hang problem.

Cheers,
Longman