[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] S3 crash with VTD Queue Invalidation enabled

>>> On 05.06.13 at 15:54, Ben Guthro <ben@xxxxxxxxxx> wrote:
> On Wed, Jun 5, 2013 at 4:24 AM, Jan Beulich <JBeulich@xxxxxxxx> wrote:
>>>>> On 04.06.13 at 23:09, Ben Guthro <ben@xxxxxxxxxx> wrote:
>>> On Tue, Jun 4, 2013 at 3:49 PM, Ben Guthro <ben@xxxxxxxxxx> wrote:
>>>> On Tue, Jun 4, 2013 at 3:20 PM, Ben Guthro <ben@xxxxxxxxxx> wrote:
>>>>> On Tue, Jun 4, 2013 at 10:01 AM, Jan Beulich <JBeulich@xxxxxxxx> wrote:
>>>>>>>>> On 04.06.13 at 14:25, Ben Guthro <ben@xxxxxxxxxx> wrote:
>>>>>>> On Tue, Jun 4, 2013 at 4:54 AM, Jan Beulich <JBeulich@xxxxxxxx> wrote:
>>>>>>>> Is this perhaps having some similarity with
>>>>>>>> http://lists.xen.org/archives/html/xen-devel/2013-04/msg00343.html? 
>>>>>>>> We're clearly running single-CPU only here and there...
>>>>>>> We certainly should be, as we have gone through the
>>>>>>> disable_nonboot_cpus() by this point - and I can verify that from the
>>>>>>> logs.
>>>>>> I'm much more tending towards the connection here, noting that
>>>>>> Andrew's original thread didn't really lead anywhere (i.e. we still
>>>>>> don't know what the panic he saw was actually caused by).
>>>>> I'm starting to think you're on to something here.
>>>> hmm - maybe not.
>>>> I get the same crash with "maxcpus=1"
>>>>> I've put a bunch of trace throughout the functions in qinval.c
>>>>> It seems that everything is functioning properly, up until we go
>>>>> through the disable_nonboot_cpus() path.
>>>>> Prior to this, I see the qinval.c functions being executed on all
>>>>> cpus, and both drhd units
>>>>> Afterward, it gets stuck in queue_invalidate_wait on the first drhd
>>>>> unit.. and eventually panics.
>>>>> I'm not exactly sure what to make of this yet.
>>> querying status of the hardware all seems to be working correctly...
>>> it just doesn't work with querying the QINVAL_STAT_DONE state, as far
>>> as I can tell.
>>> Other register state is:
>>> (XEN)  VER = 10
>>> (XEN)  CAP = c0000020e60262
>>> (XEN)  n_fault_reg = 1
>>> (XEN)  fault_recording_offset = 200
>>> (XEN)  fault_recording_reg_l = 0
>>> (XEN)  fault_recording_reg_h = 0
>>> (XEN)  ECAP = f0101a
>>> (XEN)  GCMD = 0
>>> (XEN)  GSTS = c7000000


#define DMA_GCMD_QIE    (((u64)1) << 26)


#define DMA_GSTS_QIES   (((u64)1) <<26)

this means qinval is still enabled at this point.

>>> (XEN)  RTADDR = 137a31000
>>> (XEN)  CCMD = 800000000000000
>>> (XEN)  FSTS = 0
>>> (XEN)  FECTL = 0
>>> (XEN)  FEDATA = 4128
>>> (XEN)  FEADDR = fee0000c
>>> (XEN)  FEUADDR = 0
>>> (with code lifted from print_iommu_regs() )
>>> None of this looks suspicious to my untrained eye - but I'm including
>>> it here in case someone else sees something I don't.
>> Xiantao, you certainly will want to give some advice here. I won't
>> be able to look into this more deeply right away.
> Thanks Jan. Xiantao - I'd appreciate any insight you may have.
> One curious thing I have found, that seems buggy to me, is that
> {dis,en}able_qinval() is being called prior to the platform quirks
> being executed.
> It appears they are being called through iommu_{en,dis}able_x2apic_IR()

That's because this setup needs to happen when interrupt (i.e.
APIC) initialization is happening, not when the IOMMUs get set
up (which is a process that assumes interrupts can already be

In effect we have

lapic_suspend() -> iommu_disable_x2apic_IR() ->


iommu_suspend() -> vtd_suspend() -> disable_qinval()

but the latter tail call only when !iommu_intremap, and you
have been running with interrupt remapping enabled, so only
the former code path would result in qinval getting disabled,
which is after the point of the hang.

Depending on whether ATS is in use, more than one invalidation
can be done in the processing here - could you therefore check
whether there's any sign of ATS use ("iommu=verbose" should
make you see respective messages), and if so see whether
disabling it ("ats=off") makes a difference?


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.