Re: [ofa-general] [POSSIBLE BUG] use of tasklet_unlock inipath_no_bufs_available
From: Ralph Campbell
Date: Mon Jun 25 2007 - 16:37:50 EST
This was fixed by a patch that Arthur Jones sent out to
general@xxxxxxxxxxxxxxxxxxxxx
Tue Jun 19 16:42:09 PDT 2007
[PATCH 17/28] IB/ipath - wait for PIO available interrupt
I imagine that it is working its way into Roland's git tree
for Linus.
On Mon, 2007-06-25 at 15:33 -0400, Steven Rostedt wrote:
> As some of you know, lately I've been trying to get rid of tasklets. In
> doing so, I've come across this usage of tasklet_unlock.
>
> The only user of tasklet_unlock in the kernel outside of softirq.c is
> ipath_no_bufs_available in drivers/inifiniband/hw/ipath/ipath_ruc.c
>
> Here's the offending code:
>
> void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev)
> {
> unsigned long flags;
>
> spin_lock_irqsave(&dev->pending_lock, flags);
> if (list_empty(&qp->piowait))
> list_add_tail(&qp->piowait, &dev->piowait);
> spin_unlock_irqrestore(&dev->pending_lock, flags);
> /*
> * Note that as soon as want_buffer() is called and
> * possibly before it returns, ipath_ib_piobufavail()
> * could be called. If we are still in the tasklet function,
> * tasklet_hi_schedule() will not call us until the next time
> * tasklet_hi_schedule() is called.
> * We clear the tasklet flag now since we are committing to return
> * from the tasklet function.
> */
> clear_bit(IPATH_S_BUSY, &qp->s_flags);
> tasklet_unlock(&qp->s_task);
> want_buffer(dev->dd);
> dev->n_piowait++;
> }
>
>
> As the comment states, it looks like it's trying to prevent a race where
> the want_buffer can allow for ipath_ib_piobufavail be called which would
> schedule this tasklet again. But since the tasklet is running, it would
> simply be skipped if it were to schedule on another CPU. And this would
> mean that the tasklet would need to wait for it to be scheduled again
> before doing the work.
>
> Is my above analysis correct?
>
> Now for the BUG.
>
> Lets say this situation does happen. Lets look at the code.
>
> softirq.c: tasklet_hi_action
>
> if (tasklet_trylock(t)) {
> if (!atomic_read(&t->count)) {
> if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
> BUG();
> t->func(t->data);
> tasklet_unlock(t);
> continue;
> }
> tasklet_unlock(t);
> }
>
> The race being prevented is the failure of the tasklet_trylock running
> on another CPU. The call to tasklet_unlock in ipath_no_bufs_available is
> letting the other CPU succeed, and the comment suggests that this is OK
> because this function will be exiting shortly. But what it doesn't take
> into consideration is the above "tasklet_unlock" called again in
> tasklet_hi_action.
>
> So while the tasklet function is allowed to run on another CPU, we are
> unlocking the tasklet on this CPU. So now this tasklet function is no
> longer protected from being reentrant. There is now no guarantee that
> the tasklet function would only be running on one CPU.
>
> What's worse, we also add the chance of hitting the above BUG(). If the
> tasklet gets scheduled again, takes an interrupt before doing the
> tast_and_clear, another CPU runs the tasklet and clears the
> TASKLET_STATE_SCHED, when the first instance comes back from the
> interrupt, it will hit the BUG.
>
> So, does all this make sense, or am I full of crap. Still, I think
> tasklet_unlock and tasklet_trylock should not be exported for anyone
> else to use besides softirq.c and perhaps the ipath code needs to find a
> better way around this.
>
> -- Steve
>
>
> _______________________________________________
> general mailing list
> general@xxxxxxxxxxxxxxxxxxxxx
> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
>
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/