Re: [Xen-devel] Migration memory corruption - PV backends need to quiesce

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 30/06/14 09:38, Ian Campbell wrote:
> On Fri, 2014-06-27 at 17:51 +0100, Andrew Cooper wrote:
>> Hello,
>>
>> After a long time fixing my own memory corruption bugs with migration
>> v2, I have finally tracked down (what I really really hope is) the last
>> of the corruption.
>>
>>
>> There appears to be a systematic problem affecting all PV drivers,
>> whereby a non-quiescent backend can cause memory corruption in the VM.
>>
>> Active grant mapped pages are only reflected in the dirty bitmap after
>> the grant has been unmapped, as mapping the ring read-only would be
>> catastrophic to performance, and remapping as read-only when logdirty is
>> enabled is (as far as I understand) impossible, as Xen doesn't track the
>> PTEs pointing at granted frames.
>>
>> PV backend drivers hold their mappings of the rings (and persistently
>> granted frames) open until the domain is destroyed, which is after the
>> memory image has been sent.  Therefore, any requests which are processed
>> after the migration code sending the ring frame on its first pass will
>> not be reflected in the resumed domain, as this frame will never be
>> marked as dirty in Xen.
>>
>> Furthermore, as the migration code uses memcpy() on the frames, it is
>> possible that a backed update intersects with the copy, and a corrupt
>> descriptor appears on the resumed side.
>>
>> In addition, after the domain has been paused, the backend might still
>> process requests.  The migration code excepts the guest be completely
>> quiesced after it has been suspended, so will only check the dirty
>> bitmap once.  Any requests which get processed and completed might still
>> be missed by the migration code.
>>
>> From a heavily instrumented Xen and migration code, I am fairly sure I
>> have confirmed that all pages corrupted on migration are a result of
>> still-active grant maps, grant copies which complete after domain
>> suspend, or the xenstore ring which xenstored has a magic mapping of,
>> and will never be reflected in the dirty bitmap.
>>
>>
>> Overall, it would appear that there needs to be a hook for all PV
>> drivers to force quiescence.  In particular, a backend must guarantee to
>> unmap all active grant maps (so the frames get properly reflected in the
>> dirty bitmap), and never process subsequent requests (so no new frames
>> appear dirty in the bitmap after the guest has been paused).
>>
>> Thoughts/comments?
> I thought PV drivers were already (supposed to be) handling this in the
> frontend.
>
> For reasons of checkpoint performance I think Linux's net and blkfront
> are handling this on resume rather than on suspend by tearing down on
> resume and then requeueing any outstanding I/O after they reattach to
> the new backend. In the blkfront case this is explicit, whereas IIRC
> netfront just discards any active requests and relies on L3+
> retransmition to get the job done. (see netfront_resume and
> blkfront_resume/blkif_recover).
>
> Part of the tear down and reconnect should involve invalidating any
> inflight descriptors, whether or not they were partially completed or
> have corrupted replies in them etc. This ought to be happening before
> the new backend sees the ring at all.
>
> Can you give an example of an instance of the corruption which you've
> seen?

The intersection of requests processed by the backends and memcpy() in
the migration code means that the migration code might send a frame
which is actively being written by the backend at the same time.

As these changes are not reflected in the dirty bitmap (or indeed
reflected only after the migration code has stopped looking for further
changes), the corrupt frames never get resent.

>
> Do you have a theory why this wasn't seen with the migration v1 code?

It most certainly is.  This is not a problem unique to migration v2, and
I can prove its presence in the legacy migration code given a bit of
extra debugging.

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel