Re: [Qemu-devel] [RFC PATCH] tests: Add a device_add/del HMP test

["Dr. David Alan Gilbert" <dgilbert@redhat.com>]

* Markus Armbruster (armbru@redhat.com) wrote:
> Thomas Huth <thuth@redhat.com> writes:
> 
> > People tend to forget to mark internal devices with "user_creatable = false
> > or hotpluggable = false, and these devices can crash QEMU if added via the
> > HMP monitor. So let's add a test to run through all devices and that tries
> > to add them blindly (without arguments) to see whether this could crash the
> > QEMU instance.
> >
> > Signed-off-by: Thomas Huth <thuth@redhat.com>
> > ---
> >  I've marked the patch as RFC since not all of the required device bug
> >  fixes have been merged yet (so this patch can not be included yet without
> >  breaking "make check"). It's also sad that "macio-oldworld" currently
> >  has to be blacklisted - I tried to find a fix for that device,  but I was
> >  not able to spot the exact problem so far. So help for fixing that device
> >  is very very welcome! The crash can be reproduced like this:
> >
> >  $ qemu-system-ppc64 -nographic -S -monitor stdio -serial none
> >  QEMU 2.10.50 monitor - type 'help' for more information
> >  (qemu) device_add macio-oldworld,id=x
> >  (qemu) device_del x
> >  (qemu) **
> >  ERROR:qom/object.c:1611:object_get_canonical_path_component:
> >   assertion failed: (obj->parent != NULL)
> >  Aborted (core dumped)
> >
> >  tests/test-hmp.c | 60 +++++++++++++++++++++++++++++++++++++++++++++++++++++++-
> >  1 file changed, 59 insertions(+), 1 deletion(-)
> >
> > diff --git a/tests/test-hmp.c b/tests/test-hmp.c
> > index 7a38cdc..e8a25c4 100644
> > --- a/tests/test-hmp.c
> > +++ b/tests/test-hmp.c
> > @@ -28,7 +28,6 @@ static const char *hmp_cmds[] = {
> >      "commit all",
> >      "cpu-add 1",
> >      "cpu 0",
> > -    "device_add ?",
> >      "device_add usb-mouse,id=mouse1",
> >      "mouse_button 7",
> >      "mouse_move 10 10",
> > @@ -116,6 +115,64 @@ static void test_info_commands(void)
> >      g_free(info_buf);
> >  }
> >  
> > +/*
> > + * People tend to forget to mark internal devices with "user_creatable = false
> > + * and these devices can crash QEMU if added via the HMP monitor. So let's run
> > + * through all devices and try to add them blindly (without arguments) to see
> > + * whether this could crash the QEMU instance.
> > + */
> > +static void test_device_add_commands(void)
> > +{
> > +    char *resp, *devices, *devices_buf, *endp;
> > +
> > +    devices = devices_buf = hmp("device_add help");
> > +
> > +    while (*devices) {
> > +        /* Ignore header lines etc. */
> > +        if (strncmp(devices, "name \"", 6)) {
> > +            devices = strchr(devices, '\n');
> > +            if (!devices) {
> > +                break;
> > +            }
> > +            devices += 1;
> > +            continue;
> > +        }
> > +        /* Extract the device name, ignore parameters and description */
> > +        devices += 6;
> > +        endp = strchr(devices, '"');
> > +        g_assert(endp != NULL);
> > +        *endp = '\0';
> > +        /* Check whether it is blacklisted... */
> > +        if (g_str_equal("macio-oldworld", devices)) {
> > +            devices = strchr(endp + 1, '\n');
> > +            if (!devices) {
> > +                break;
> > +            }
> > +            devices += 1;
> > +            continue;
> > +        }
> > +        /* Now run the device_add + device_del commands */
> > +        if (verbose) {
> > +            fprintf(stderr, "\tdevice_add %s,id=%s\n", devices, devices);
> > +        }
> > +        resp = hmp("device_add %s,id=%s", devices, devices);
> > +        g_free(resp);
> > +        if (verbose) {
> > +            fprintf(stderr, "\tdevice_del %s\n", devices);
> > +        }
> > +        resp = hmp("device_del %s", devices);
> > +        g_free(resp);
> > +        /* And move forward to the next line */
> > +        devices = strchr(endp + 1, '\n');
> > +        if (!devices) {
> > +            break;
> > +        }
> > +        devices += 1;
> > +    }
> > +
> > +    g_free(devices_buf);
> > +}
> > +
> >  static void test_machine(gconstpointer data)
> >  {
> >      const char *machine = data;
> > @@ -125,6 +182,7 @@ static void test_machine(gconstpointer data)
> >      qtest_start(args);
> >  
> >      test_info_commands();
> > +    test_device_add_commands();
> >      test_commands();
> >  
> >      qtest_end();
> 
> This finds devices by parsing output of HMP help.  I think you should
> use introspection instead, like device-introspect-test does.  You may
> want to extract common utility code from there.
> 
> The actual device_add and device_del also use HMP.  Failures are
> ignored.  A few device_add failures I'd expect:
> 
> * There is no suitable bus.
> 
> * There are suitable buses, but the default one is full.
> 
> * Mandatory parameters are missing, such as device backend.
> 
> * The bus doesn't support hot plug (some other bus might).
> 
> * The device supports only cold plug with -device, not hot plug with
>   device_add.
> 
> I'm afraid the test only tests one of these common failure modes for
> many devices.
> 
> device_del failures I'd expect:
> 
> * The device doesn't exist, because it hasn't completed hot plug, yet.
>   In some cases such as ACPI PCI hot plug, hot plug may require guest
>   cooperation, which may take unbounded time.  device_add merely kicks
>   off the hot plug then.  I can't remember how to poll for completion.
>   I also can't remember why we don't send a QMP event.
> 
>   The hot plug usually completes quickly, but it may take its own sweet
>   time, or not complete at all, say because the guest doesn't support
>   ACPI, or just doesn't feel like plugging right now.
> 
>   The test needs to set up a guest that cooperates.  I guess that
>   involves libqos; yet another thing I've forgotten.
> 
> * Same for device_del.  You should wait for the DEVICE_DELETED event
>   with a suitable timeout.

Yes, I think that's an interesting problem - although the test
might still be worth it just to see how many issues it finds;   I'm
curious how many devices actually work with no parameters though,
most seem to fail.

If I'm reading the code right it's creating the device with the same
name as the device;  I wonder if that always works?
But still,  it should mean that if the previous hotplug hasn't really
happened then you can move onto the next add.

> We could improve device_add to cold plug before the machine starts,
> i.e. between start with -S and the first cont.  We could similarly
> improve device_del to cold plug.  Together, that would let you sidestep
> the hot plug complications.
> 
> I guess this test is a case of "if it was easy, we would've done it
> already"...

Still maybe it's worth it as a start.

Dave

--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK