Re: [PATCH next v3 1/2] dump_stack: move cpu lock to printk.c

From: John Ogness <john.ogness@linutronix.de>
To: Petr Mladek <pmladek@suse.com>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>,
	Steven Rostedt <rostedt@goodmis.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	linux-kernel@vger.kernel.org,
	Stephen Rothwell <sfr@canb.auug.org.au>,
	Andrew Morton <akpm@linux-foundation.org>,
	Daniel Bristot de Oliveira <bristot@redhat.com>,
	Stephen Boyd <swboyd@chromium.org>,
	Alexander Potapenko <glider@google.com>,
	Peter Zijlstra <peterz@infradead.org>
Subject: Re: [PATCH next v3 1/2] dump_stack: move cpu lock to printk.c
Date: Wed, 16 Jun 2021 15:46:02 +0206	[thread overview]
Message-ID: <877diund1p.fsf@jogness.linutronix.de> (raw)
In-Reply-To: <YMnenOBTUclLld9i@alley>

On 2021-06-16, Petr Mladek <pmladek@suse.com> wrote:
>> With this series version I moved the tracking into a global variable
>> @printk_cpulock_nested, which is fine, except that a boolean is not
>> capable of tracking more than 1 nesting. Which means that
>> __printk_cpu_unlock() would release cpu lock ownership too soon.
>> 
>> Doing this correctly is a simple change:
>> 
>> diff --git a/kernel/printk/printk.c b/kernel/printk/printk.c
>> index e67dc510fa1b..5376216e4f3d 100644
>> --- a/kernel/printk/printk.c
>> +++ b/kernel/printk/printk.c
>> @@ -3535,7 +3535,7 @@ EXPORT_SYMBOL_GPL(kmsg_dump_rewind);
>>  
>>  #ifdef CONFIG_SMP
>>  static atomic_t printk_cpulock_owner = ATOMIC_INIT(-1);
>> -static bool printk_cpulock_nested;
>> +static atomic_t printk_cpulock_nested = ATOMIC_INIT(0);
>>  
>>  /**
>>   * __printk_wait_on_cpu_lock() - Busy wait until the printk cpu-reentrant
>> @@ -3596,7 +3598,7 @@ int __printk_cpu_trylock(void)
>>  
>>  	} else if (old == cpu) {
>>  		/* This CPU is already the owner. */
>> -		printk_cpulock_nested = true;
>> +		atomic_inc(&printk_cpulock_nested);
>>  		return 1;
>>  	}
>>  
>> @@ -3613,8 +3615,8 @@ EXPORT_SYMBOL(__printk_cpu_trylock);
>>   */
>>  void __printk_cpu_unlock(void)
>>  {
>> -	if (printk_cpulock_nested) {
>> -		printk_cpulock_nested = false;
>> +	if (atomic_read(&printk_cpulock_nested)) {
>> +		atomic_dec(&printk_cpulock_nested);
>
> I think about handling printk_cpulock_nested with only one
> atomic operation. Something like:
>
> 	if (atomic_dec_return(&printk_cpulock_level) == 0)
> 		atomic_set_release(&printk_cpulock_owner, -1);
>
> It would require always incremanting the number in lock, e.g.
>
> 	old = atomic_cmpxchg(&printk_cpulock_owner, -1, cpu);
> 	if (old == -1 || old == cpu) {
> 		atomic_inc(&printk_cpulock_level);
> 		return 1;
> 	}

I actually implemented similar code during an internal draft. I later
decided against it, mainly because I prefer to keep the old==-1 and
old==cpu cases separate.

Also note that atomic_dec_return() introduces an unnecessary memory
barrier. If we take your proposed implementation we would use
atomic_dec_return_relaxed() instead.

> But I am not sure if it is really better. Feel free to keep
> your variant.

*sigh* Frankly, I don't care much. My variant saves a few CPU
instructions for the normal case (non-nested), but that probably is not
much of an argument.

For v4 I will keep my variant because it explicitly handles the
non-nested/nested cases separately, which helps when adding the memory
barrier comments in the follow-up patch. In particular, the label
LMM(__printk_cpu_trylock:B), which represents the first moment a new CPU
begins to load/store data, only applies to the old==-1 condition.

John Ogness