Re: [PATCH 3/3] pciehp: Fix race condition handling surprise link-down
From: Bjorn Helgaas
Date: Thu Dec 08 2016 - 12:50:19 EST
On Thu, Dec 08, 2016 at 12:20:58PM -0500, Keith Busch wrote:
> On Thu, Dec 08, 2016 at 09:11:58AM -0600, Bjorn Helgaas wrote:
> > On Wed, Dec 07, 2016 at 07:04:33PM -0500, Keith Busch wrote:
> > >
> > > It currently looks safe to nest the p_slot->lock under the
> > > p_slot->hotplug_lock if that is you recommendation.
> > I'm not making a recommendation; that would take a lot more thought
> > than I've given this.
> > There are at least three locks in pciehp:
> > struct controller.ctrl_lock
> > struct slot.lock
> > struct slot.hotplug_lock
> > There shouldn't really be any performance paths in pciehp, so I'm
> > pretty doubtful that we need such complicated locking.
> They are protecting different things, but I agree it looks like room
> for simplification exists.
If we can't simplify this immediately, can we add a comment about what
the different locks protect so people have a hint about which one to
use? For example, it looks like this patch might benefit from that
> > > Alternatively we could fix this if we used an ordered work queue for
> > > the slot's work, but that is a significantly more complex change.
> > You mean we can currently execute things from the work queue in a
> > different order than they were enqueued? That sounds ... difficult to
> > analyze, to say the least.
> The events are dequeued in order, but they don't wait for the previous
> to complete, so pciehp's current work queue can have multiple events
> executing in parallel. That's part of why rapid pciehp slot events are
> a little more difficult to follow, and I think we may even be unsafely
> relying on the order the mutexes are obtained from these work events.
Hmm. I certainly did not expect multiple events executing in
parallel. That sounds like a pretty serious issue to me.
> Partly unrelated, we could process surprise removals significantly
> faster (microseconds vs. seconds), with the limited pci access series
> here, giving fewer simultaneously executing events to consider:
> Do you have any concerns with that series?
I'm dragging my feet because I want the removal process to become
simpler to understand, not more complicated, and we're exposing more
issues that I didn't even know about.
> > I don't know much about work queues, and Documentation/workqueue.txt
> > doesn't mention alloc_ordered_workqueue(). Is that what you are
> > referring to?
> Yes, the alloc_ordered_workqueue is what I had in mind, though switching
> to that is not as simple as calling the different API. I am looking into
> that for longer term, but for the incremental fix, do you think we can
> go forward with Raj's proposal?
I'd like to at least see a consistent locking strategy for protecting
p_slot->state. All the existing updates are protected by
p_slot->lock, but the one Raj is adding is protected by