Date:   Wed, 12 Dec 2018 08:21:39 -0700
From:   "Jan Beulich" <JBeulich@...e.com>
To:     "Chao Gao" <chao.gao@...el.com>
Cc:     "Roger Pau Monne" <roger.pau@...rix.com>,
        "Jia-Ju Bai" <baijiaju1990@...il.com>,
        "Stefano Stabellini" <sstabellini@...nel.org>,
        "xen-devel" <xen-devel@...ts.xenproject.org>,
        "Boris Ostrovsky" <boris.ostrovsky@...cle.com>,
        "Juergen Gross" <jgross@...e.com>, <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] xen: xen-pciback: Reset MSI-X state when exposing
 a device

>>> On 12.12.18 at 16:18, <chao.gao@...el.com> wrote:
> On Wed, Dec 12, 2018 at 01:51:01AM -0700, Jan Beulich wrote:
>>>>> On 12.12.18 at 08:06, <chao.gao@...el.com> wrote:
>>> On Wed, Dec 05, 2018 at 09:01:33AM -0500, Boris Ostrovsky wrote:
>>>>On 12/5/18 4:32 AM, Roger Pau Monné 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?

>>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.

Jan

Powered by blists - more mailing lists