Re: Endless loop in qcow2_alloc_cluster_offset

[Kevin Wolf <kwolf@redhat.com>]

Am 07.12.2009 15:16, schrieb Jan Kiszka:
>> Likely not. What I did was nothing special, and I did not noticed such a
>> crash in the last months.
> 
> And now it happened again (qemu-kvm head, during kernel installation
> from network onto local qcow2-disk). Any clever idea how to proceed with
> this?

I still haven't seen this and I still have no theory on what could be
happening here. I'm just trying to write down what I think must happen
to get into this situation. Maybe you can point at something I'm missing
or maybe it helps you to have a sudden inspiration.

The crash happens because we have a loop in the s->cluster_allocs list.
A loop can only be created by inserting an object twice. The only insert
to this list happens in qcow2_alloc_cluster_offset (though an earlier
call than that of the stack trace).

There is only one relevant caller of this function, qcow_aio_write_cb.
Part of it is a call to run_dependent_requests which removes the request
from s->cluster_allocs. So after the QLIST_REMOVE in
run_dependent_requests the request can't be contained in the list, but
at the call of qcow2_alloc_cluster_offset it must be contained again. It
must be added somewhere in between these two calls.

In qcow_aio_write_cb there isn't much happening between these calls. The
only thing that could somehow become dangerous is the
qcow_aio_write_cb(req, 0); for queued requests in run_dependent_requests.

> I could try to run the step in a loop, hopefully retriggering it once in
> a (likely longer) while. But then we need some good instrumentation first.

I can't explain what exactly would be going wrong there, but if my
thoughts are right so far, I think that moving this into a Bottom Half
would help. So if you can reproduce it in a loop this could be worth a try.

I'd certainly prefer to understand the problem first, but thinking about
AIO is the perfect way to make your brain hurt...

Kevin