On Fri, 2014-01-10 at 08:50 -0700, Bjorn Helgaas wrote: > On Tue, Jan 07, 2014 at 04:45:19PM -0700, Alex Williamson wrote: > > On Tue, 2014-01-07 at 16:15 -0700, Bjorn Helgaas wrote: > > > Hi Alex, > > > > > > Sorry for the delay in looking at this. > > > > > > On Mon, Dec 16, 2013 at 3:14 PM, Alex Williamson > > > <alex.williamson@xxxxxxxxxx> wrote: > > > > When doing a function/slot/bus reset PCI grabs the device_lock for > > > > each device to block things like suspend and driver probes, which is > > > > all well and good, but call paths exist where this lock may already be > > > > held. This creates an opportunity for deadlock. For instance, vfio > > > > allows userspace to issue resets so long as it owns the device(s). > > > > If a driver unbind .remove callback races with userspace issuing a > > > > reset, we have a deadlock as userspace gets stuck waiting on > > > > device_lock while another thread has device_lock and waits for .remove > > > > to complete. > > > > > > Are you talking about vfio_pci_remove() (the vfio_pci_driver .remove() > > > method) racing with vfio_pci_ioctl()? > > > > Yes, for instance if the admin does something like attempt to unbind the > > device from vfio-pci while it's in use. This can also happen indirectly > > if the device is going away, such as a PF driver attempting to remove > > its VFs. > > > > > Or maybe it's vfio_pci_release (the vfio_pci_ops .release() method), > > > since it looks like you want to use pci_try_reset_function() there and > > > in vfio_pci_ioctl()? > > > > That one too. If any reset races with pci_driver.remove, whether it be > > from ioctl or my own release callback, we'll hit a deadlock. > > > > > Either way, aren't there at least potentially more locking issues than > > > just the reset problem? Seems like any ioctl that might take the > > > device_lock could have the same problem. > > > > I don't know of any others, but our QE folks hit this one pretty > > regularly. It can all be explained away as user error, but a user > > should not be able to cause a kernel deadlock, which is why the user > > exposed interfaces are converted to use this try-lock interface. > > > > > How do you make sure there's > > > no userspace owner of the device before you release the device or > > > remove the driver? > > > > Userspace has a file descriptor for the device, so we know through > > reference counting when the device becomes unused. When our > > pci_driver.remove() callback happens we block until the user releases > > the device, so perhaps it more accurate to describe it as a nested lock > > problem than a race, but the paths are asynchronous so even if we tested > > the lock in advance there's a potential race. > > If I understand correctly, the deadlock happens when A (userspace) has a file > descriptor for the device, and B waits in this path: > > driver_detach > device_lock # take device_lock > __device_release_driver > pci_device_remove # pci_bus_type.remove > vfio_pci_remove # pci_driver .remove > vfio_del_group_dev > wait_event(vfio.release_q, !vfio_dev_present) # wait (holding device_lock) > > Now B is stuck until A gives up the file descriptor. I haven't traced that > path, but I assume giving up the file descriptor involves the > vfio_device_put() ... vfio_device_release() path to wake up B. > > If A tries to acquire device_lock for any reason, we deadlock because A > is waiting for B to release the lock, and B is waiting for A to release the > file descriptor. Correct > You're fixing the case where A does a VFIO_DEVICE_RESET ioctl that calls > pci_reset_function(). It looks like a VFIO_DEVICE_SET_IRQS ioctl will have > the same problem because it can acquire device_lock in > vfio_pci_set_err_trigger(). There's no reason for that path to use device_lock, I've already posted a separate patch to remove it there and will include it in my v3.14 pull request. > It makes me uneasy that B (in the kernel) waits indefinitely for A > (userspace) to do something. If A is malicious and simply does nothing, it > will cause all other paths that acquire device_lock to hang, e.g., suspend > and hibernate, AER reporting (e.g., report_error_detected()), and various > USB events. Are there USB events that would ever grab the PCI dev's device_lock? Blocking until userspace releases the device is certainly not an ideal situation, but releasing a userspace owned device is not as simple as a kernel owned device. We can notify the user, but there's still no guarantee the device will be released. We can try to take back the device, but we don't have any way to revoke the mmaps. Even if the wait in B was not indefinite, I think we'd still run into device_lock problems. Using the same paths you outline above, assume that B eventually kills the userspace process and moves on. There's still a window where userspace can get stuck on device_lock and I'm not sure what's going to happen if the user process gets killed while it's stuck in the kernel waiting on device_lock. The only other alternative I see is to release device_lock while we wait at B and re-acquire, but that makes me just as uneasy. Using try_lock seems like the best short term approach to prioritize kernel-base locks over user requested locks. Thanks, Alex -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
