Re: [RFC] Fix SMP brokenness for PF_FREEZE and make freezing usable for other purposes

[Kirill Korotaev <dev@sw.ru>]

>>>-static inline int freezing(struct task_struct *p)
>>>-{
>>>-	return p->flags & PF_FREEZE;
>>>-}
>>>+#if defined(CONFIG_PM) || defined(CONFIG_MIGRATE)
>>
>><<<< why not to make a single option CONFIG_REFRIGERATOR? It looks to be a
>>more robust way, since there are multiple users of it.
> 
> 
> Yes. That may be better. We can do that once the migration code is 
> finished and when we know what kind of CONFIG_XXX the migration code 
> really needs.
> 
> 
>>>+#ifdef CONFIG_PM
>>
>><<<< is it intentionaly? or you just lost CONFIG_MIGRATE?
> It is intentional. freeze_processes and thaw_processes are only needed 
> for suspend. One only needs to freeze a couple of processes for process 
> migration.
But PM and your migrate code can be not the only users of it.

>><<<< I still think this refrigerator is racy with freeze_processes():
>><<<< scenarios:
>><<<< scenario 1
>><<<<
>>task1 -> freeze_processes():			taskXXX ->refrigerator()
>>  checked (task->flags & PF_FROZEN) == 0	cur->flags |= PF_FROZEN
>>						clear TIF_FREEZE
>>						<sleep on thaw>
>>  set TIF_FREEZING
>>						clear PF_FROZEN
>>
>><<<< so the task awakes with TIF_FREEZE flag set!!!
> 
> 
> Hmm... If we wait to clear both flags until after the completion 
> notification then we do not have the race right? But then we need to move 
> the signal recalc since it tests for TIF_FREEZE too.
It is almost ok, but it is still not fine :)

look what happens if you call freeze/unfreeze in a loop:

refrigerator:
awakes

freezer:
check PF_FROZEN, it is still set, skips task and thinks it is finished 
freezing.

refrigerator:
clears PF_FROZEN and TIF_FREEZE and returns.

I think you can fix this by moving PF_FROZEN check and set in both 
places under siglock.

Kirill


>><<<< scenario 2
>><<<< look at error path in freeze_processes (on timeout), it is broken as
>>well. You need to wakeup tasks there...
> 
> 
> Ok. How about this additional patch? This still requires that process 
> freezing does not immediately occurr again after the completion 
> handler. All of this is iffy due to not having a real lock protecting all 
> these values and we may still need to add some barriers.
> 
> Index: linux-2.6.12/kernel/power/process.c
> ===================================================================
> --- linux-2.6.12.orig/kernel/power/process.c	2005-06-28 06:34:52.000000000 +0000
> +++ linux-2.6.12/kernel/power/process.c	2005-06-28 06:40:28.000000000 +0000
> @@ -47,12 +47,13 @@ int freeze_processes(void)
>  			unsigned long flags;
>  			if (!freezeable(p))
>  				continue;
> -			if ((p->flags & PF_FROZEN) ||
> -			    (p->state == TASK_TRACED) ||
> +			if ((p->state == TASK_TRACED) ||
>  			    (p->state == TASK_STOPPED))
>  				continue;
>  
>  			set_thread_flag(TIF_FREEZE);
> +			if (p->flags & PF_FROZEN)
> +				continue;
>  			spin_lock_irqsave(&p->sighand->siglock, flags);
>  			signal_wake_up(p, 0);
>  			spin_unlock_irqrestore(&p->sighand->siglock, flags);
> @@ -63,6 +64,8 @@ int freeze_processes(void)
>  		if (time_after(jiffies, start_time + TIMEOUT)) {
>  			printk( "\n" );
>  			printk(KERN_ERR " stopping tasks failed (%d tasks remaining)\n", todo );
> +			complete_all(&thaw);
> +			up(&freezer_sem);
>  			return todo;
>  		}
>  	} while(todo);
> Index: linux-2.6.12/kernel/sched.c
> ===================================================================
> --- linux-2.6.12.orig/kernel/sched.c	2005-06-28 06:34:52.000000000 +0000
> +++ linux-2.6.12/kernel/sched.c	2005-06-28 06:37:36.000000000 +0000
> @@ -5210,13 +5210,13 @@ DECLARE_COMPLETION(thaw);
>  void refrigerator(void)
>  {
>  	current->flags |= PF_FROZEN;
> +	wait_for_completion(&thaw);
>  	clear_thread_flag(TIF_FREEZE);
> +	current->flags &= ~PF_FROZEN;
>  	/* A fake signal 0 may have been sent. Recalculate sigpending */
>  	spin_lock_irq(&current->sighand->siglock);
>  	recalc_sigpending();
>  	spin_unlock_irq(&current->sighand->siglock);
> -	wait_for_completion(&thaw);
> -	current->flags &= ~PF_FROZEN;
>  }
>  EXPORT_SYMBOL(refrigerator);
>  #endif
> 


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"aart@kvack.org"> aart@kvack.org </a>