Re: [Qemu-devel] [PATCH 15/18] Boot Linux Console Test: add a test for aarch64 + virt

["Daniel P. Berrangé" <berrange@redhat.com>]

On Fri, Jun 07, 2019 at 03:58:57PM -0300, Eduardo Habkost wrote:
> CCing Daniel, who wrote commit 6ab3fc32ea64.
> 
> On Fri, Jun 07, 2019 at 11:44:32AM -0400, Cleber Rosa wrote:
> > On Fri, Jun 07, 2019 at 12:42:14AM -0300, Eduardo Habkost wrote:
> > > On Fri, Jun 07, 2019 at 12:26:48AM -0300, Eduardo Habkost wrote:
> > > > On Fri, Feb 01, 2019 at 11:10:31AM -0500, Cleber Rosa wrote:
> > > > > 
> > > > > 
> > > > > On 1/31/19 4:26 PM, Cleber Rosa wrote:
> > > > > > 
> > > > > > 
> > > > > > On 1/31/19 3:21 PM, Cleber Rosa wrote:
> > > > > >>
> > > > > >>
> > > > > >> On 1/31/19 3:02 PM, Wainer dos Santos Moschetta wrote:
> > > > > >>>
> > > > > >>> On 01/17/2019 04:56 PM, Cleber Rosa wrote:
> > > > > >>>> Just like the previous tests, boots a Linux kernel on a aarch64 target
> > > > > >>>> using the virt machine.
> > > > > >>>>
> > > > > >>>> One special option added is the CPU type, given that the kernel
> > > > > >>>> selected fails to boot on the virt machine's default CPU (cortex-a15).
> > > > > >>>>
> > > > > >>>> Signed-off-by: Cleber Rosa <crosa@redhat.com>
> > > > > >>>> ---
> > > > > >>>>   .travis.yml                            |  2 +-
> > > > > >>>>   tests/acceptance/boot_linux_console.py | 20 ++++++++++++++++++++
> > > > > >>>>   2 files changed, 21 insertions(+), 1 deletion(-)
> > > > > >>>>
> > > > > >>>> diff --git a/.travis.yml b/.travis.yml
> > > > > >>>> index 54100eea5a..595e8c0b6c 100644
> > > > > >>>> --- a/.travis.yml
> > > > > >>>> +++ b/.travis.yml
> > > > > >>>> @@ -187,7 +187,7 @@ matrix:
> > > > > >>>>         # Acceptance (Functional) tests
> > > > > >>>>       - env:
> > > > > >>>> -        - CONFIG="--python=/usr/bin/python3
> > > > > >>>> --target-list=x86_64-softmmu,mips-softmmu,mips64el-softmmu,ppc64-softmmu"
> > > > > >>>> +        - CONFIG="--python=/usr/bin/python3
> > > > > >>>> --target-list=x86_64-softmmu,mips-softmmu,mips64el-softmmu,ppc64-softmmu,aarch64-softmmu"
> > > > > >>>>
> > > > > >>>>           - TEST_CMD="make check-acceptance"
> > > > > >>>>         addons:
> > > > > >>>>           apt:
> > > > > >>>> diff --git a/tests/acceptance/boot_linux_console.py
> > > > > >>>> b/tests/acceptance/boot_linux_console.py
> > > > > >>>> index f3ccd23a7a..107700b517 100644
> > > > > >>>> --- a/tests/acceptance/boot_linux_console.py
> > > > > >>>> +++ b/tests/acceptance/boot_linux_console.py
> > > > > >>>> @@ -138,3 +138,23 @@ class BootLinuxConsole(Test):
> > > > > >>>>           self.vm.launch()
> > > > > >>>>           console_pattern = 'Kernel command line: %s' %
> > > > > >>>> kernel_command_line
> > > > > >>>>           self.wait_for_console_pattern(console_pattern)
> > > > > >>>> +
> > > > > >>>> +    def test_aarch64_virt(self):
> > > > > >>>
> > > > > >>> That test case fails on my system (Fedora 29 x86_64). Avocado seems
> > > > > >>> unable to kill the VM so it  reaches the timeout.
> > > > > >>>
> > > > > >>> I compiled QEMU with default configuration:
> > > > > >>>
> > > > > >>> $ configure --python=/usr/bin/python3 --target-list=x86_64-softmmu
> > > > > >>> --target-list=x86_64-softmmu,mips-softmmu,mips64el-softmmu,ppc64-softmmu,aarch64-softmmu)
> > > > > >>>
> > > > > >>>
> > > > > >>> Follows a snippet of the Avocado's job.log file:
> > > > > >>> ----
> > > > > >>> 2019-01-31 14:41:34,912 test             L0602 INFO | START
> > > > > >>> 07-/root/src/qemu/tests/acceptance/boot_linux_console.py:BootLinuxConsole.test_aarch64_virt
> > > > > >>>
> > > > > >>> 2019-01-31 14:41:34,912 test             L0298 DEBUG| DATA
> > > > > >>> (filename=output.expected) => NOT FOUND (data sources: variant, test, file)
> > > > > >>> 2019-01-31 14:41:34,913 parameters       L0146 DEBUG| PARAMS (key=arch,
> > > > > >>> path=*, default=aarch64) => 'aarch64'
> > > > > >>> 2019-01-31 14:41:34,913 parameters       L0146 DEBUG| PARAMS
> > > > > >>> (key=qemu_bin, path=*, default=aarch64-softmmu/qemu-system-aarch64) =>
> > > > > >>> 'aarch64-softmmu/qemu-system-aarch64'
> > > > > >>> 2019-01-31 14:41:34,915 download         L0070 INFO | Fetching
> > > > > >>> https://sjc.edge.kernel.org/fedora-buffet/fedora/linux/releases/29/Server/aarch64/os/images/pxeboot/vmlinuz
> > > > > >>> -> /var/lib/avocado/data/cache/by_name/vmlinuz.3upct2pr
> > > > > >>> 2019-01-31 14:41:35,490 download         L0054 DEBUG| Retrieved URL
> > > > > >>> "https://sjc.edge.kernel.org/fedora-buffet/fedora/linux/releases/29/Server/aarch64/os/images/pxeboot/vmlinuz":
> > > > > >>> content-length 8623423, date: "Thu, 31 Jan 2019 19:41:35 GMT",
> > > > > >>> last-modified: "Sun, 21 Oct 2018 00:43:09 GMT"
> > > > > >>> 2019-01-31 14:41:41,765 qemu             L0317 DEBUG| VM launch command:
> > > > > >>> 'aarch64-softmmu/qemu-system-aarch64 -chardev
> > > > > >>> socket,id=mon,path=/var/tmp/tmpizirkcud/qemu-32609-monitor.sock -mon
> > > > > >>> chardev=mon,mode=control -display none -vga none -machine virt -chardev
> > > > > >>> socket,id=console,path=/var/tmp/tmpizirkcud/qemu-32609-console.sock,server,nowait
> > > > > >>> -serial chardev:console -cpu cortex-a53 -kernel
> > > > > >>> /var/lib/avocado/data/cache/by_name/vmlinuz -append console=ttyAMA0'
> > > > > >>> 2019-01-31 14:41:41,779 qmp              L0167 DEBUG| >>> {'execute':
> > > > > >>> 'qmp_capabilities'}
> > > > > >>> 2019-01-31 14:41:41,931 qmp              L0175 DEBUG| <<< {'return': {}}
> > > > > >>> 2019-01-31 14:41:42,830 boot_linux_conso L0041 DEBUG| [    0.000000]
> > > > > >>> Booting Linux on physical CPU 0x0000000000 [0x410fd034]
> > > > > >>>
> > > > > >>> (...)
> > > > > >>>
> > > > > >>> 2019-01-31 14:41:42,833 boot_linux_conso L0041 DEBUG| [    0.000000]
> > > > > >>> Policy zone: DMA32
> > > > > >>> 2019-01-31 14:41:42,833 boot_linux_conso L0041 DEBUG| [    0.000000]
> > > > > >>> Kernel command line: console=ttyAMA0
> > > > > >>> 2019-01-31 14:41:42,833 qmp              L0167 DEBUG| >>> {'execute':
> > > > > >>> 'quit'}
> > > > > >>
> > > > > >> Here, a QMP response like "<<< {'return': {}}" would be expected.
> > > > > >>
> > > > > >> Since I can not reproduce this on my system (or on Travis-CI jobs I've
> > > > > >> sent), can you tell me on top of which commit you've applied these patches?
> > > > > >>
> > > > > > 
> > > > > > I spoke too soon:
> > > > > > 
> > > > > > https://travis-ci.org/clebergnu/qemu/jobs/487121425#L3033
> > > > > > 
> > > > > > This looks like a recent regression, and I'm guessing it's not on the
> > > > > > test's side.  I'll try to bisect it and let you know.
> > > > > > 
> > > > > 
> > > > > On a fresh environment, I am able to get this reproduced on every 2 of
> > > > > runs, more or less.  When I hit it, I attached GDB to it, and the
> > > > > backtrace shows:
> > > > > 
> > > > > Thread debugging using libthread_db enabled]
> > > > > Using host libthread_db library "/lib64/libthread_db.so.1".
> > > > > warning: Loadable section ".note.gnu.property" outside of ELF segments
> > > > > warning: Loadable section ".note.gnu.property" outside of ELF segments
> > > > > __lll_lock_wait () at ../sysdeps/unix/sysv/linux/x86_64/lowlevellock.S:103
> > > > > 103     2:      movl    %edx, %eax
> > > > > (gdb) bt
> > > > > #0  __lll_lock_wait () at
> > > > > ../sysdeps/unix/sysv/linux/x86_64/lowlevellock.S:103
> > > > > #1  0x00007fc6ba1a2e09 in __GI___pthread_mutex_lock
> > > > > (mutex=mutex@entry=0x5615a233d020 <qemu_global_mutex>) at
> > > > > ../nptl/pthread_mutex_lock.c:80
> > > > > #2  0x00005615a1bb7593 in qemu_mutex_lock_impl (mutex=0x5615a233d020
> > > > > <qemu_global_mutex>, file=0x5615a1db2d4c "util/main-loop.c", line=236)
> > > > > at util/qemu-thread-posix.c:66
> > > > > #3  0x00005615a171125e in qemu_mutex_lock_iothread_impl
> > > > > (file=file@entry=0x5615a1db2d4c "util/main-loop.c", line=line@entry=236)
> > > > > at /home/cleber/src/qemu/cpus.c:1849
> > > > > #4  0x00005615a1bb415d in os_host_main_loop_wait (timeout=<optimized
> > > > > out>) at util/main-loop.c:236
> > > > > #5  main_loop_wait (nonblocking=<optimized out>) at util/main-loop.c:497
> > > > > #6  0x00005615a18fdd39 in main_loop () at vl.c:1928
> > > > > #7  0x00005615a16c9ee9 in main (argc=<optimized out>, argv=<optimized
> > > > > out>, envp=<optimized out>) at vl.c:4665
> > > > 
> > > > Tip: run "thread apply all bt" so you can get a backtrace of all
> > > > threads.
> > > > 
> > > > 
> > > > > 
> > > > > Running it with `taskset -c 1` prevents this issue from happening, which
> > > > > AFAICT, contributes even further towards this being a QEMU race condition.
> > > > > 
> > > > > I'm CC'ing Peter and Claudio (listed maintainers of aarch64), as this
> > > > > seems to limited to that target.  Any tips on what to do here?
> > > > 
> > > > I am hitting this on Travis, too, and I finally could reproduce
> > > > it locally,
> > > > 
> > > > The guest is still writing on the serial console, but nobody is
> > > > reading the data on the other side.  A VCPU thread is stuck
> > > > inside the EAGAIN/nanosleep loop at qemu_chr_write_buffer(),
> > > > holding the QEMU global lock.
> > > 
> > > Experimental fix below.
> > > 
> > > Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
> > > ---
> > >  python/qemu/__init__.py | 12 ++++++++----
> > >  1 file changed, 8 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/python/qemu/__init__.py b/python/qemu/__init__.py
> > > index 81d9657ec0..4a691f34da 100644
> > > --- a/python/qemu/__init__.py
> > > +++ b/python/qemu/__init__.py
> > > @@ -274,10 +274,6 @@ class QEMUMachine(object):
> > >  
> > >          self._qemu_log_path = None
> > >  
> > > -        if self._console_socket is not None:
> > > -            self._console_socket.close()
> > > -            self._console_socket = None
> > > -
> > >          if self._temp_dir is not None:
> > >              shutil.rmtree(self._temp_dir)
> > >              self._temp_dir = None
> > > @@ -336,6 +332,14 @@ class QEMUMachine(object):
> > >          """
> > >          Terminate the VM and clean up
> > >          """
> > > +
> > > +        # If we keep the console socket open, we may deadlock waiting
> > > +        # for QEMU to exit, while QEMU is waiting for the socket to
> > > +        # become writeable.
> > > +        if self._console_socket is not None:
> > > +            self._console_socket.close()
> > > +            self._console_socket = None
> > > +
> > 
> > Right, this is somewhat equivalent to the following "workaround":
> > 
> >    https://github.com/clebergnu/qemu/commit/e1713f3b91972ad57c089f276c54db3f3fa63423
> > 
> > I could not tell at the moment, though, if closing (or shutting down)
> > the console socket was the appropriate fix.
> > 
> > What I see is that Rick's commit pointed to by Lazlo is dated from
> > 2016, and it says that a virtio-console fix is necessary.  I'd imagine
> > that, if a fix was ever proposed, it'd have been incorporated in the
> > F29 kernel used in this test... and that this workaround/fix would
> > not be the best solution.
> 
> If I understood this correctly, fixing the guest driver wouldn't
> help here.  The commit mentioned by Laszlo (6ab3fc32ea64 "hw:
> replace most use of qemu_chr_fe_write with qemu_chr_fe_write_all")
> fixes data loss bugs but introduces the potential for deadlocks
> like the one we are seeing.
> 
> Unless we replace the existing ~30 qemu_chr_fe_write_all() calls
> in device code, we need to make sure we are constantly reading
> data from the console socket.

Yes, you must *always* have something reading from the chardev backend.
This really sucks, but it is preferrable to letting data end up in
/dev/null.

If this is being a problem for tests then consider it motivation to
fix the root cause problem and make the devices properly do async
I/O for chardevs.  Only a handful of them properly do this right
now.


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 :|