Re: [PATCH v1 0/6] testing/next (netbsd stuff)

[Peter Maydell]

On Thu, 7 Nov 2019 at 17:54, Kamil Rytarowski <address@hidden> wrote:
>
> On 07.11.2019 18:46, Peter Maydell wrote:
> > On Mon, 4 Nov 2019 at 17:39, Alex Bennée <address@hidden> wrote:
> >>
> >> Hi,
> >>
> >> As we approach hard-freeze I'm trying to temper what comes in through
> >> the testing/next tree. However it would be nice to get the NetBSD upto
> >> speed with the other NetBSDs. Although the serial install is working
> >> well for me this has had a rocky road so if others could also give it
> >> a good testing that would be great. I've also disabled one of the
> >> regular failing tests for non-Linux targets. There are other tests
> >> that still fail however including the tests/test-aio-multithread which
> >> asserts in the async utils around about 20% of the time:
> >>
> >>   assertion "QSLIST_EMPTY(&ctx->scheduled_coroutines)" failed: file
> >>     "/home/qemu/qemu-test.nS1czd/src/util/async.c", line 279, function
> >>     "aio_ctx_finalize"
> >
> > This is unrelated to your NetBSD update in this series -- it's
> > one of the persistent intermittents I see on the BSDs:
> > https://lore.kernel.org/qemu-devel/20190916153312.GD25552@stefanha-x1.localdomain/t/
> >
> > (though the failure rate I see is I think <20%, but I haven't
> > really carefully measured it.)

> Does this patch rely on AIO API in the kernel? If so than this is
> unreliable as of today on NetBSD. We plan to fix it, but there is no
> expected time of accomplishment.

No, we use our own AIO implementation which puts fds into non-blocking
mode and uses a thread which polls them to identify when they're
ready to actually perform IO (plus a lot of coroutine magic).

thanks
-- PMM