Re: [PATCH] mem_event: use wait queue when ring is full

[Olaf Hering <olaf@aepfle.de>]

On Wed, Jan 11, Andres Lagar-Cavilla wrote:

> > mem_event: use wait queue when ring is full
> >
> > This change is based on an idea/patch from Adin Scannell.
> 
> Olaf,
> thanks for the post. We'll have to nack this patch in its current form. It
> hard reboots our host during our testing.

Thats very unfortunate. I have seen such unexpected reboots myself a few
weeks ago. I suspect they were caused by an incorrect debug change which
I had on top of my waitqueue changes. Also the fixes Keir provided a few
weeks ago may have helped.

Is it an otherwise unmodified xen-unstable build, or do you use other
patches as well? Whats your environment and workload anyway in dom0 and
domU?

It would be very good to know why the reboots happen. Perhaps such
failures can not be debugged without special hardware, or a detailed
code review.


I just tested an otherwise unmodified xen-unstable build and did not
encounter reboots while ballooning a single 2G guest up and down. The
guest did just hang after a few iterations, most likely because v7 of my
patch again (or still?) has the math wrong in the ring accounting. I
will check what the issue is. I think v6 was ok in that respect, but I
will double check that older version as well.


> What we did is take this patch, amalgamate it with some bits from our ring
> management approach. We're ready to submit that, along with a description
> of how we test it. It works for us, and it involves wait queue's for
> corner cases.

Now if the patch you just sent out uses wait queues as well, and using
wait queues causes sudden host reboots for reasons not yet known, how is
your patch any better other that the reboots dont appear to happen
anymore?

I did not use anything but paging for testing, perhaps I should also run
some access tests. How should I use tools/tests/xen-access/xen-access.c?

Olaf