FWIW, modifying ept_next_level() and ept_set_entry() to be
{write,read}-once (i.e., reading once and working with a local copy;
or making the entry in a local copy and writing it all at once) seems
to work as well, at least on 64-bit.
But overall, unless there's a measured reason to avoid locks, I think
the best thing for long-term stability is just to have ept_get_entry()
grab the p2m lock.
-George
On Thu, Aug 26, 2010 at 11:35 AM, George Dunlap
<George.Dunlap@xxxxxxxxxxxxx> wrote:
> In the course of doing some fixes for my populate-on-demand testing, I
> found that a Windows Server 2008 VM with 30G static max and 24G ram
> (i.e., booting ballooned) crashed 1-2 times out of ten during boot,
> reporting MMIO errors.
>
> I managed to get a trace of this crash. Strangely enough, the trace
> indicated that the page the NPF occured on was populate-on-demand --
> but that hvm_hap_nested_page_fault() injected a GP anyway.
>
> The only way this would be possible is if the gfn_to_mfn_query() in
> the trace function got a p2m type of p2m_popluate_on_demand, but the
> gfn_to_mfn_current() in hvm_hap_nested_page_fault() got a p2m type of
> p2m_mmio_dm.
>
> Looking at the trace (snippet attached), the failed NPF happened on
> d1v1; but almost simultaneously on d1v0, an NPF fault happened that
> caused a populate-on-demand demand populate. That demand populate
> happened to be of a superpage that was shared with the gpa fault on
> d1v1.
>
> So, the first query on d1v1 (correctly) got a PoD; but the second
> query, instead of either causing the demand-populate, or successfully
> getting the result of d1v0's demand populate, returned failure,
> causing the guest to crash.
>
> I looked in the p2m-ept.c code, and noticed (once again) that
> ept_get_entry() can be called without the p2m lock held. I added
> conditional locks, and am running the test again. The guest has now
> booted 20 times successfully without crashing (whereas before, the
> average was about 2 in 10 crashing).
>
> Looking closely at the code, I can see one potential race:
> * entry starts out PoD, not-present.
> * v0 finds the entry PoD, allocates a page, calls set_p2m_entry(),
> which calls ept_set_entry().
> * v1 begins to walk the pagetable; at some point, it calls
> ept_next_level(), which finds the flags all clear (entry->epte & 7 ==
> 0)
> * v0 ept_set_entry() changes the p2m type from p2m_populate_on_demand
> to p2m_ram_rw
> * v1 ept_next_level() reads entry->avail1 and finds that it is not
> p2m_populate_on_demand, so it returns GUEST_TABLE_MAP_FAILED
> * v0 ept_set_entry() sets the flags to present.
>
> Is there a good reason not to just grab the p2m lock when walking the
> ept tables? We could conceivably do some cleverness to avoid this
> kind of race, but unless there's a significant performance gain, I
> think the simple approach is better.
>
> -George
>
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel
|