On Wed, 2015-04-01 at 12:12 +0800, Chen Fan wrote: > On 03/25/2015 10:41 AM, Alex Williamson wrote: > > On Wed, 2015-03-25 at 09:53 +0800, Chen Fan wrote: > >> On 03/16/2015 10:09 PM, Alex Williamson wrote: > >>> On Mon, 2015-03-16 at 15:35 +0800, Chen Fan wrote: > >>>> On 03/16/2015 11:52 AM, Alex Williamson wrote: > >>>>> On Mon, 2015-03-16 at 11:05 +0800, Chen Fan wrote: > >>>>>> On 03/14/2015 06:34 AM, Alex Williamson wrote: > >>>>>>> On Thu, 2015-03-12 at 18:23 +0800, Chen Fan wrote: > >>>>>>>> when the vfio device encounters an uncorrectable error in host, > >>>>>>>> the vfio_pci driver will signal the eventfd registered by this > >>>>>>>> vfio device, the results in the qemu eventfd handler getting > >>>>>>>> invoked. > >>>>>>>> > >>>>>>>> this patch is to pass the error to guest and have the guest driver > >>>>>>>> recover from the error. > >>>>>>> What is going to be the typical recovery mechanism for the guest? I'm > >>>>>>> concerned that the topology of the device in the guest doesn't > >>>>>>> necessarily match the topology of the device in the host, so if the > >>>>>>> guest were to attempt a bus reset to recover a device, for instance, > >>>>>>> what happens? > >>>>>> the recovery mechanism is that when guest got an aer error from a > >>>>>> device, > >>>>>> guest will clean the corresponding status bit in device register. and > >>>>>> for > >>>>>> need reset device, the guest aer driver would reset all devices under > >>>>>> bus. > >>>>> Sorry, I'm still confused, how does the guest aer driver reset all > >>>>> devices under a bus? Are we talking about function-level, device > >>>>> specific reset mechanisms or secondary bus resets? If the guest is > >>>>> performing secondary bus resets, what guarantee do they have that it > >>>>> will translate to a physical secondary bus reset? vfio may only do an > >>>>> FLR when the bus is reset or it may not be able to do anything depending > >>>>> on the available function-level resets and physical and virtual topology > >>>>> of the device. Thanks, > >>>> in general, functions depends on the corresponding device driver > >>>> behaviors > >>>> to do the recovery. e.g: implemented the error_detect, slot_reset > >>>> callbacks. > >>>> and for link reset, it usually do secondary bus reset. > >>>> > >>>> and do we must require to the physical secondary bus reset for vfio > >>>> device > >>>> as bus reset? > >>> That depends on how the guest driver attempts recovery, doesn't it? > >>> There are only a very limited number of cases where a secondary bus > >>> reset initiated by the guest will translate to a secondary bus reset of > >>> the physical device (iirc, single function device without FLR). In most > >>> cases, it will at best be translated to an FLR. VFIO really only does > >>> bus resets on VM reset because that's the only time we know that it's ok > >>> to reset multiple devices. If the guest driver is depending on a > >>> secondary bus reset to put the device into a recoverable state and we're > >>> not able to provide that, then we're actually reducing containment of > >>> the error by exposing AER to the guest and allowing it to attempt > >>> recovery. So in practice, I'm afraid we're risking the integrity of the > >>> VM by exposing AER to the guest and making it think that it can perform > >>> recovery operations that are not effective. Thanks, > >> I also have seen that if device without FLR, it seems can do hot reset > >> by ioctl VFIO_DEVICE_PCI_HOT_RESET to reset the physical slot or bus > >> in vfio_pci_reset. does it satisfy the recovery issues that you said? > > The hot reset interface can only be used when a) the user (QEMU) owns > > all of the devices on the bus and b) we know we're resetting all of the > > devices. That mostly limits its use to VM reset. I think that on a > > secondary bus reset, we don't know the scope of the reset at the QEMU > > vfio driver, so we only make use of reset methods with a function-level > > scope. That would only result in a secondary bus reset if that's the > > reset mechanism used by the host kernel's PCI code (pci_reset_function), > > which is limited to single function devices on a secondary bus, with no > > other reset mechanisms. The host reset is also only available in some > > configurations, for instance if we have a dual-port NIC where each > > function is a separate IOMMU group, then we clearly cannot do a hot > > reset unless both functions are assigned to the same VM _and_ appear to > > the guest on the same virtual bus. So even if we could know the scope > > of the reset in the QEMU vfio driver, we can only make use of it under > > very strict guest configurations. Thanks, > Hi Alex, > > have you some idea or scenario to fix/escape this issue?
Hi Chen, I expect there are two major components to this. The first is that QEMU/vfio-pci needs to enforce that a bus reset is possible for the host and guest topology when guest AER handling is specified for a device. That means that everything affected by the bus reset needs to be exposed to the guest in a compatible way. For instance, if a bus reset affects devices from multiple groups, the guest needs to not only own all of those groups, but they also need to be exposed to the guest such that the virtual bus layout reflects the extent of the reset for the physical bus. This also implies that guest AER handling cannot be the default since it will impose significant configuration restrictions on device assignment. This seems like a difficult configuration enforcement to make, but maybe there are simplifying assumptions that can help. For instance the devices need to be exposed as PCIe therefore we won't have multiple slots in use on a bus and I think we can therefore mostly ignore hotplug since we can only hotplug at a slot granularity. That may also imply that we should simply enforce a 1:1 mapping of physical functions to virtual functions. At least one function from each group affected by a reset must be exposed to the guest. The second issue is that individual QEMU PCI devices have no callback for a bus reset. QEMU/vfio-pci currently has the DeviceClass.reset callback, which we assume to be a function-level reset. We also register with qemu_register_reset() for a VM reset, which is the only point currently that we know we can do a reset affecting multiple devices. Infrastructure will need to be added to QEMU/PCI to expose the link down/RST signal to devices on a bus to trigger a multi-device reset in vfio-pci. Hopefully I'm not missing something, but I think both of those changes are going to be required before we can have anything remotely supportable for guest-based AER error handle. This pretty complicated for the user and also for libvirt to figure out. At a minimum libvirt would need to support a new guest-based AER handling flag for devices. We probably need to determine whether this is unique to vfio-pci or a generic PCIDevice option. Thanks, Alex
