Re: [Question] Hooks for scheduler tracing (CFS)

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Ingo Molnar (mingo@elte.hu) wrote:
> 
> * Ankita Garg <ankita@in.ibm.com> wrote:
> 
> >         local_irq_save(flags);
> >         buf = _stp_chan->buf[smp_processor_id()];
> >         if (unlikely(buf->offset + length > _stp_chan->subbuf_size))
> >                 length = relay_switch_subbuf(buf, length);
> >         memcpy(buf->data + buf->offset, data, length);
> >         buf->offset += length;
> >         local_irq_restore(flags);
> 
> oh, what a fine piece of s^H^H :-/ Who in their right mind calls this 
> from _tracing_ code:
> 
>                 smp_mb();
>                 if (waitqueue_active(&buf->read_wait))
>                         /*
>                          * Calling wake_up_interruptible() from here
>                          * will deadlock if we happen to be logging
>                          * from the scheduler (trying to re-grab
>                          * rq->lock), so defer it.
>                          */
>                         __mod_timer(&buf->timer, jiffies + 1);
> 
> and the comment is utter rubbish: __mod_timer() can lock up just as 
> much. Just use an adaptive-polling method to drive the draining of the 
> relay buffer, instead of mucking with timers from within the tracing 
> code. Whoever implemented this has absolutely zero clue i have to say 
> ...
> 
> the smp_mb() is rubbish too.
> 
> could you try the patch below, does it fix the problem?
> 

Hi Ingo,

I did not have this problem with LTTng, since I only touch atomic
variables in tracing code. However, I iterate on a rcu list of active
traces in a timer interrupt to see if subbuffers must be read and, if
yes, I send a wakeup to my user-space daemon. I had to change the
protection around this rcu list read from preempt disable to mutex lock
in this timer because try_to_wakeup is called in it, which takes a
spinlock.

Note: I don't use relay code at my tracing site.

I guess they might have to switch to such an asynchronous delivery
system if they want to do this properly. Simply put, your polling
solution is exactly what I do, but I check a flag set by the writer
instead of waking up the readers unconditionally.

Mathieu

> 	Ingo
> 
> ------------------------------------->
> Subject: relay: fix timer madness
> From: Ingo Molnar <mingo@elte.hu>
> 
> remove timer calls (!!!) from deep within the tracing infrastructure.
> This was totally bogus code that can cause lockups and worse.
> Poll the buffer every 2 jiffies for now.
> 
> Signed-off-by: Ingo Molnar <mingo@elte.hu>
> ---
>  kernel/relay.c |   14 +++++---------
>  1 file changed, 5 insertions(+), 9 deletions(-)
> 
> Index: linux-rt-rebase.q/kernel/relay.c
> ===================================================================
> --- linux-rt-rebase.q.orig/kernel/relay.c
> +++ linux-rt-rebase.q/kernel/relay.c
> @@ -319,6 +319,10 @@ static void wakeup_readers(unsigned long
>  {
>  	struct rchan_buf *buf = (struct rchan_buf *)data;
>  	wake_up_interruptible(&buf->read_wait);
> +	/*
> +	 * Stupid polling for now:
> +	 */
> +	mod_timer(&buf->timer, jiffies + 1);
>  }
>  
>  /**
> @@ -336,6 +340,7 @@ static void __relay_reset(struct rchan_b
>  		init_waitqueue_head(&buf->read_wait);
>  		kref_init(&buf->kref);
>  		setup_timer(&buf->timer, wakeup_readers, (unsigned long)buf);
> +		mod_timer(&buf->timer, jiffies + 1);
>  	} else
>  		del_timer_sync(&buf->timer);
>  
> @@ -604,15 +609,6 @@ size_t relay_switch_subbuf(struct rchan_
>  		buf->subbufs_produced++;
>  		buf->dentry->d_inode->i_size += buf->chan->subbuf_size -
>  			buf->padding[old_subbuf];
> -		smp_mb();
> -		if (waitqueue_active(&buf->read_wait))
> -			/*
> -			 * Calling wake_up_interruptible() from here
> -			 * will deadlock if we happen to be logging
> -			 * from the scheduler (trying to re-grab
> -			 * rq->lock), so defer it.
> -			 */
> -			__mod_timer(&buf->timer, jiffies + 1);
>  	}
>  
>  	old = buf->data;

-- 
Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/