Re: [PATCH] xen: xen-pciback: Reset MSI-X state when exposing a device

From: Jan Beulich
Date: Thu Dec 13 2018 - 02:54:57 EST


>>> On 13.12.18 at 04:46, <chao.gao@xxxxxxxxx> wrote:
> On Wed, Dec 12, 2018 at 08:21:39AM -0700, Jan Beulich wrote:
>>>>> On 12.12.18 at 16:18, <chao.gao@xxxxxxxxx> wrote:
>>> On Wed, Dec 12, 2018 at 01:51:01AM -0700, Jan Beulich wrote:
>>>>>>> On 12.12.18 at 08:06, <chao.gao@xxxxxxxxx> wrote:
>>>>> On Wed, Dec 05, 2018 at 09:01:33AM -0500, Boris Ostrovsky wrote:
>>>>>>On 12/5/18 4:32 AM, Roger Pau Monnà wrote:
>>>>>>> On Wed, Dec 05, 2018 at 10:19:17AM +0800, Chao Gao wrote:
>>>>>>>> I find some pass-thru devices don't work any more across guest reboot.
>>>>>>>> Assigning it to another guest also meets the same issue. And the only
>>>>>>>> way to make it work again is un-binding and binding it to pciback.
>>>>>>>> Someone reported this issue one year ago [1]. More detail also can be
>>>>>>>> found in [2].
>>>>>>>>
>>>>>>>> The root-cause is Xen's internal MSI-X state isn't reset properly
>>>>>>>> during reboot or re-assignment. In the above case, Xen set maskall bit
>>>>>>>> to mask all MSI interrupts after it detected a potential security
>>>>>>>> issue. Even after device reset, Xen didn't reset its internal maskall
>>>>>>>> bit. As a result, maskall bit would be set again in next write to
>>>>>>>> MSI-X message control register.
>>>>>>>>
>>>>>>>> Given that PHYSDEVOPS_prepare_msix() also triggers Xen resetting MSI-X
>>>>>>>> internal state of a device, we employ it to fix this issue rather than
>>>>>>>> introducing another dedicated sub-hypercall.
>>>>>>>>
>>>>>>>> Note that PHYSDEVOPS_release_msix() will fail if the mapping between
>>>>>>>> the device's msix and pirq has been created. This limitation prevents
>>>>>>>> us calling this function when detaching a device from a guest during
>>>>>>>> guest shutdown. Thus it is called right before calling
>>>>>>>> PHYSDEVOPS_prepare_msix().
>>>>>>> s/PHYSDEVOPS/PHYSDEVOP/ (no final S). And then I would also drop the
>>>>>>> () at the end of the hypercall name since it's not a function.
>>>>>>>
>>>>>>> I'm also wondering why the release can't be done when the device is
>>>>>>> detached from the guest (or the guest has been shut down). This makes
>>>>>>> me worry about the raciness of the attach/detach procedure: if there's
>>>>>>> a state where pciback assumes the device has been detached from the
>>>>>>> guest, but there are still pirqs bound, an attempt to attach to
>>>>>>> another guest in such state will fail.
>>>>>>
>>>>>>I wonder whether this additional reset functionality could be done out
>>>>>>of xen_pcibk_xenbus_remove(). We first do a (best effort) device reset
>>>>>>and then do the extra things that are not properly done there.
>>>>>
>>>>> No. It cannot be done in xen_pcibk_xenbus_remove() without modifying
>>>>> the handler of PHYSDEVOP_release_msix. To do a successful Xen internal
>>>>> MSI-X state reset, PHYSDEVOP_{release, prepare}_msix should be finished
>>>>> without error. But ATM, xen expects that no msi is bound to pirq when
>>>>> doing PHYSDEVOP_release_msix. Otherwise it fails with error code -EBUSY.
>>>>> However, the expectation isn't guaranteed in xen_pcibk_xenbus_remove().
>>>>> In some cases, if qemu fails to unmap MSIs, MSIs are unmapped by Xen
>>>>> at last minute, which happens after device reset in
>>>>> xen_pcibk_xenbus_remove().
>>>>
>>>>But that may need taking care of: I don't think it is a good idea to have
>>>>anything left from the prior owning domain when the device gets reset.
>>>>I.e. left over IRQ bindings should perhaps be forcibly cleared before
>>>>invoking the reset;
>>>
>>> Agree. How about pciback to track the established IRQ bindings? Then
>>> pciback can clear irq binding before invoking the reset.
>>
>>How would pciback even know of those mappings, when it's qemu
>>who establishes (and manages) them?
>
> I meant to expose some interfaces from pciback. And pciback serves
> as the proxy of IRQ (un)binding APIs.

If at all possible we should avoid having to change more parties (qemu,
libxc, kernel, hypervisor) than really necessary. Remember that such
a bug fix may want backporting, and making sure affected people have
all relevant components updated is increasingly difficult with their
number growing.

>>>>in fact I'd expect this to happen in the course of
>>>>domain destruction, and I'd expect the device reset to come after the
>>>>domain was cleaned up. Perhaps simply an ordering issue in the tool
>>>>stack?
>>>
>>> I don't think reversing the sequences of device reset and domain
>>> destruction would be simple. Furthermore, during device hot-unplug,
>>> device reset is done when the owner is alive. So if we use domain
>>> destruction to enforce all irq binding cleared, in theory, it won't be
>>> applicable to hot-unplug case (if qemu's hot-unplug logic is
>>> compromised).
>>
>>Even in the hot-unplug case the tool stack could issue unbind
>>requests, behind the back of the possibly compromised qemu,
>>once neither the guest nor qemu have access to the device
>>anymore.
>
> But currently, tool stack doesn't know the remaining IRQ bindings.
> If tool stack can maintaine IRQ binding information of a pass-thru
> device (stored in Xenstore?), we can come up with a clean solution
> without modifying linux kernel and Xen.

If there's no way for the tool stack to either find out the bindings
or "blindly" issue unbind requests (accepting them to fail), then a
"wildcard" unbind operation may want adding. Or, perhaps even
better, XEN_DOMCTL_deassign_device could unbind anything left
in place for the specified device. I actually wonder why that's not
already the case.

Jan