On Thu, Nov 16, 2023 at 04:47:48PM +0530, Ani Sinha wrote: > > > > On 16-Nov-2023, at 2:25 PM, Daniel P. Berrangé <berra...@redhat.com> wrote: > > > > On Thu, Nov 16, 2023 at 05:14:43PM +1000, Nicholas Piggin wrote: > >> On Thu Nov 16, 2023 at 1:55 PM AEST, Ani Sinha wrote: > >>> > >>> > >>>> On 16-Nov-2023, at 6:45 AM, Nicholas Piggin <npig...@gmail.com> wrote: > >>>> > >>>> On Thu Nov 16, 2023 at 3:22 AM AEST, Daniel P. Berrangé wrote: > >>>>> On Wed, Nov 15, 2023 at 01:14:53PM +0000, Daniel P. Berrangé wrote: > >>>>> > >>>>> After an absolutely horrendous hours long debugging session I think > >>>>> I figured out the problem. The QEMU process is blocking in > >>>>> > >>>>> qemu_chr_write_buffer > >>>>> > >>>>> spinning in the loop on EAGAIN. > >>>> > >>>> Great work. > >>>> > >>>> Why does this make the gdb socket give an empty response? Something > >>>> just times out? > >>>> > >>>>> > >>>>> The Python Machine() class has passed one of a pre-created socketpair > >>>>> FDs for the serial port chardev. The guest is trying to write to this > >>>>> and blocking. Nothing in the Machine() class is reading from the > >>>>> other end of the serial port console. > >>>>> > >>>>> > >>>>> Before John's change, the serial port uses a chardev in server mode > >>>>> and crucially 'wait=off', and the Machine() class never opened the > >>>>> console socket unless the test case wanted to read from it. > >>>>> > >>>>> IOW, QEMU had a background job setting there waiting for a connection > >>>>> that would never come. > >>>>> > >>>>> As a result when QEMU started executing the guest, all the serial port > >>>>> writes get sent into to the void. > >>>>> > >>>>> > >>>>> So John's patch has had a semantic change in behaviour, because the > >>>>> console socket is permanently open, and thus socket buffers are liable > >>>>> to fill up. > >>>>> > >>>>> As a demo I increased the socket buffers to 1MB and everything then > >>>>> succeeded. > >>>>> > >>>>> @@ -357,6 +360,10 @@ def _pre_launch(self) -> None: > >>>>> > >>>>> if self._console_set: > >>>>> self._cons_sock_pair = socket.socketpair() > >>>>> + self._cons_sock_pair[0].setsockopt(socket.SOL_SOCKET, > >>>>> socket.SO_SNDBUF, 1024*1024); > >>>>> + self._cons_sock_pair[0].setsockopt(socket.SOL_SOCKET, > >>>>> socket.SO_RCVBUF, 1024*1024); > >>>>> + self._cons_sock_pair[1].setsockopt(socket.SOL_SOCKET, > >>>>> socket.SO_SNDBUF, 1024*1024); > >>>>> + self._cons_sock_pair[1].setsockopt(socket.SOL_SOCKET, > >>>>> socket.SO_RCVBUF, 1024*1024); > >>>>> os.set_inheritable(self._cons_sock_pair[0].fileno(), True) > >>>>> > >>>>> # NOTE: Make sure any opened resources are *definitely* freed in > >>>> > >>>> So perhaps ppc64 fails just because it prints more to the console in > >>>> early > >>>> boot than other targets? > >>>> > >>>>> The Machine class doesn't know if anything will ever use the console, > >>>>> so as is the change is unsafe. > >>>>> > >>>>> The original goal of John's change was to guarantee we capture early > >>>>> boot messages as some test need that. > >>>>> > >>>>> I think we need to be able to have a flag to say whether the caller > >>>>> needs > >>>>> an "early console" facility, and only use the pre-opened FD passing for > >>>>> that case. Tests we need early console will have to ask for that > >>>>> guarantee > >>>>> explicitly. > >>>> > >>>> The below patch makes this test work. Maybe as a quick fix it is > >>>> better than disabling the test. > >>>> > >>>> I guess we still have a problem if a test invokes vm.launch() > >>>> directly without subsequently waiting for a console pattern or > >>>> doing something with the console as you say. Your suggesstion is > >>>> add something like vm.launch(console=True) ? > >>> > >>> I think what he is saying is to add a new property for QEMUMachine() with > >>> which the test can explicitly tell the machine init code that it is going > >>> to drain the console logs. By default it can be false. When tests use > >>> console_drainer, they can set the property to true and inspect the early > >>> console logs after draining it. > >> > >> Hmm... well we do have QEMUMachine.set_console already. Is this enough? > >> If the test case is not going to drain or interact with the console > >> then it could set it to false. Or am I missing something? > > > > Yeah, set_console is enough - i missed that that exists. > > > > Thus problem is more specific. It hits when a test calls > > set_console(True), but then fails to read from the console. > > So then it is a test issue in that the test requests console to be enabled > but does not look at the console o/p.
Yes, any test cases affected by this are broken already, and need to drain the console. We might easily miss if we only test with x86_64 QEMU and not other arches, as eg ppc firmware appears more chatty on serial port than seabios With regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
