On Tue, May 07, 2024 at 08:14:18PM GMT, Daniel P. Berrangé wrote:
> On Tue, May 07, 2024 at 04:56:00PM +0000, Andrea Bolognani wrote:
> > On Fri, Mar 22, 2024 at 06:56:08PM GMT, Peter Krempa wrote:
> > > +static void
> > > +testPipeFeeder(void *opaque)
> > > +{
> > > + /* feed more than observed buffer size which was historically 128k in the
> > > + * test this was adapted from */
> > > + size_t emptyspace = 140 * 1024;
> >
> > This test seems to fail consistently at least on ppc64le, among other
> > less common architectures. This can be seen both in Debian[1] and
> > Fedora[2]. It runs for a while, then it hits the timeout gets
> > terminated by meson.
> >
> > I've reproduced it locally and this is the output:
> >
> > # LIBVIRT_DEBUG=1 VIR_TEST_DEBUG=1 VIR_TEST_RANGE=71 ./tests/virshtest
> > ninja: no work to do.
> > TEST: virshtest
> > 71) read-big-pipe
> > ... 2024-05-07 16:43:17.099+0000: 69735: info : libvirt version:
> > 10.4.0
> > 2024-05-07 16:43:17.099+0000: 69735: debug : virThreadJobSet:96 :
> > Thread 69735 is now running job testPipeFeeder
> > 2024-05-07 16:43:17.099+0000: 69734: debug : virCommandRunAsync:2657
> > : About to run LANG=C /root/libvirt/build/tools/virsh --connect
> > test:///default 'define /tmp/libvirt_virshtest_XUTXGN2/pipe ; list
> > --all'
> > 2024-05-07 16:43:17.099+0000: 69735: debug : virThreadJobClear:121 :
> > Thread 69735 finished job testPipeFeeder with ret=0
> > 2024-05-07 16:43:17.099+0000: 69734: debug : virCommandRunAsync:2659
> > : Command result 0, with PID 69736
> >
> > I've bumped the size of emptyspace to 1024*1024 and that causes the
> > test to pass. 1023*1024 doesn't. Could it be something about the
> > fifo's capacity being different across architectures?
>
> I had multiple builds fail in Fedora, but today a ppc64
> build magically passed. So even on ppc64 it is racy :-(
>
> The virFileReadAll method reads in BUFSIZ chunks but that
> hasn't changed in years. The new test is pretty trivial,
> and I struggle to see why changing the buffer size would
> affect it, given the old test this replaced did largely
> the same thing.
Maybe the shell writes to the fifo differently than libvirt does?
Also "read XML from a fifo" and "read a really big XML file" were two
separate test cases before. Some unexpected interaction between
buffer size and fifo handling, perhaps?
--
Andrea Bolognani / Red Hat / Virtualization
_______________________________________________
Devel mailing list -- devel@xxxxxxxxxxxxxxxxx
To unsubscribe send an email to devel-leave@xxxxxxxxxxxxxxxxx
