On Mon, Nov 02, 2020 at 10:09:28AM +0100, Andrea Righi wrote:
> On Mon, Nov 02, 2020 at 09:56:58AM +0100, Pavel Machek wrote:
> > Hi!
> >
> > > > > I'm getting the following lockdep splat (see below).
> > > > >
> > > > > Apparently this warning starts to be reported after applying:
> > > > >
> > > > > e918188611f0 ("locking: More accurate annotations for read_lock()")
> > > > >
> > > > > It looks like a false positive to me, but it made me think a bit and
> > > > > IIUC there can be still a potential deadlock, even if the deadlock
> > > > > scenario is a bit different than what lockdep is showing.
> > > > >
> > > > > In the assumption that read-locks are recursive only in_interrupt()
> > > > > context (as stated in e918188611f0), the following scenario can still
> > > > > happen:
> > > > >
> > > > > CPU0 CPU1
> > > > > ---- ----
> > > > > read_lock(&trig->leddev_list_lock);
> > > > >
> > > > > write_lock(&trig->leddev_list_lock);
> > > > > <soft-irq>
> > > > > kbd_bh()
> > > > > -> read_lock(&trig->leddev_list_lock);
> > > > >
> > > > > *** DEADLOCK ***
> > > > >
> > > > > The write-lock is waiting on CPU1 and the second read_lock() on CPU0
> > > > > would be blocked by the write-lock *waiter* on CPU1 => deadlock.
> > > > >
> > > >
> > > > No, this is not a deadlock, as a write-lock waiter only blocks
> > > > *non-recursive* readers, so since the read_lock() in kbd_bh() is called
> > > > in soft-irq (which in_interrupt() returns true), so it's a recursive
> > > > reader and won't get blocked by the write-lock waiter.
> > >
> > > That's right, I was missing that in_interrupt() returns true also from
> > > soft-irq context.
> > >
> > > > > In that case we could prevent this deadlock condition using a
> > > > > workqueue
> > > > > to call kbd_propagate_led_state() instead of calling it directly from
> > > > > kbd_bh() (even if lockdep would still report the false positive).
> > > > >
> > > >
> > > > The deadlock senario reported by the following splat is:
> > > >
> > > >
> > > > CPU 0: CPU 1:
> > > > CPU 2:
> > > > ----- -----
> > > > -----
> > > > led_trigger_event():
> > > > read_lock(&trig->leddev_list_lock);
> > > > <work queue processing>
> > > > ata_hsm_qs_complete():
> > > >
> > > > spin_lock_irqsave(&host->lock);
> > > >
> > > > write_lock(&trig->leddev_list_lock);
> > > > ata_port_freeze():
> > > > ata_do_link_abort():
> > > > ata_qc_complete():
> > > > ledtrig_disk_activity():
> > > >
> > > > led_trigger_blink_oneshot():
> > > >
> > > > read_lock(&trig->leddev_list_lock);
> > > > // ^ not in
> > > > in_interrupt() context, so could get blocked by CPU 2
> > > > <interrupt>
> > > > ata_bmdma_interrupt():
> > > > spin_lock_irqsave(&host->lock);
> > > >
> > > > , where CPU 0 is blocked by CPU 1 because of the spin_lock_irqsave() in
> > > > ata_bmdma_interrupt() and CPU 1 is blocked by CPU 2 because of the
> > > > read_lock() in led_trigger_blink_oneshot() and CPU 2 is blocked by CPU 0
> > > > because of an arbitrary writer on &trig->leddev_list_lock.
> > > >
> > > > So I don't think it's false positive, but I might miss something
> > > > obvious, because I don't know what the code here actually does ;-)
> > >
> > > With the CPU2 part it all makes sense now and lockdep was right. :)
> > >
> > > At this point I think we could just schedule a separate work to do the
> > > led trigger and avoid calling it with host->lock held and that should
> > > prevent the deadlock. I'll send a patch to do that.
> >
> > Let's... not do that, unless we have no choice.
> >
> > Would it help if leddev_list_lock used _irqsave() locking?
>
> Using read_lock_irqsave/irqrestore() in led_trigger_event() would be
> enough to prevent the deadlock. If it's an acceptable solution I can
> send a patch (already tested it and lockdep doesn't complain :)).
Any comment on
https://lore.kernel.org/lkml/20201102104152.GG9930@xps-13-7390/?
Thanks,
-Andrea
