[PATCH i-g-t] tests/intel/xe_wedged: Introduce a new test for Xe device wedged state

[Rodrigo Vivi]

On Wed, Mar 13, 2024 at 03:07:44PM -0500, Lucas De Marchi wrote:
> On Wed, Mar 13, 2024 at 03:55:28PM -0400, Rodrigo Vivi wrote:
> > Let's inject a gt_reset failure that will put Xe device in the
> > new wedged state, then we confirm the IOCTL is blocked and we
> > reload the driver to get back to a clean state for other test
> > execution, since wedged state in Xe is a final state that can only
> > be cleared with a module reload.
> > 
> > This new test case is entirely based on xe_uevent provided by
> > Himal.
> 
> /me confused... I don't see any uevent handling here.

the uevent part is gone, but the failure injection came from there.

> 
> > 
> > Cc:  Himal Prasad Ghimiray <himal.prasad.ghimiray at intel.com>
> > Signed-off-by: Rodrigo Vivi <rodrigo.vivi at intel.com>
> > ---
> > tests/intel/xe_wedged.c | 91 +++++++++++++++++++++++++++++++++++++++++
> > tests/meson.build       |  1 +
> > 2 files changed, 92 insertions(+)
> > create mode 100644 tests/intel/xe_wedged.c
> > 
> > diff --git a/tests/intel/xe_wedged.c b/tests/intel/xe_wedged.c
> > new file mode 100644
> > index 000000000..f767e2511
> > --- /dev/null
> > +++ b/tests/intel/xe_wedged.c
> > @@ -0,0 +1,91 @@
> > +// SPDX-License-Identifier: MIT
> > +/*
> > + * Copyright © 2024 Intel Corporation
> > + */
> > +
> > +/**
> > + * TEST: cause fake gt reset failure which put Xe device in wedged state
> > + * Category: Software building block
> > + * Sub-category: driver
> > + * Functionality: wedged
> > + * Test category: functionality test
> > + */
> > +
> > +#include "igt.h"
> > +#include "igt_kmod.h"
> > +
> > +#include "xe/xe_ioctl.h"
> > +
> > +static void force_wedged(int fd)
> > +{
> > +	igt_debugfs_write(fd, "fail_gt_reset/probability", "100");
> > +	igt_debugfs_write(fd, "fail_gt_reset/times", "2");
> > +
> > +	xe_force_gt_reset(fd, 0);
> 
> humn... do we have to check the writes above did anything?

unfortunately the debugfs_write is a void return...

we could read it back, but I don't believe it brings anything...

>  I also don't
> see the kernel side, but if it just resets normally, the test would
> still pass afaics.

nope, if the reset works normally without injecting the failure and
declaring the gt busted, then we would fail below

igt_assert_eq(simple_ioctl(fd), 0);
force_busted(fd);
igt_assert_neq(simple_ioctl(fd), 0);
fd = rebind_xe(fd);
igt_assert_eq(simple_ioctl(fd), 0);

notice that the middle one is != 0,


but I'm considering to change that to

igt_assert_eq(simple_ioctl(fd), -ECANCELED);

for clarity.

> 
> > +	sleep(1);
> > +}
> > +
> > +static int reload_xe(int fd)
> > +{
> > +	int error;
> > +
> > +	drm_close_driver(fd);
> > +	igt_xe_driver_unload();
> 
> 
> what if we are running on e.g. MTL with a DG2 and want to debug one of
> them? Rather than re-loading the module and possibly causing unrelated
> issues (if e.g. module removal from the other card crashes), why not
> just unbind the module from the card under test?
> 
> i.e. the equivalent in C of:
> 
> rebind() {
> 	pci_slot=$1
> 	echo -n "0000:$pci_slot" > /sys/bus/pci/drivers/$driver/unbind
> 	echo -n "0000:$pci_slot" > /sys/bus/pci/drivers/$driver/bind
> }

Thanks, that indeed is a better choice.

the only caveat is that I need to close the main fd client for a proper
exit before we can rebind cleanly. But I could finally get that working.

> 
> Lucas De Marchi
> 
> > +
> > +	error = igt_xe_driver_load(NULL);
> > +
> > +	igt_assert_eq(error, 0);
> > +
> > +	/* driver is ready, check if it's bound */
> > +	fd = __drm_open_driver(DRIVER_XE);
> > +	igt_fail_on_f(fd < 0, "Cannot open the xe DRM driver while reloading xe after wedged\n");
> > +	return fd;
> > +}
> > +
> > +static int simple_ioctl(int fd)
> > +{
> > +	int ret;
> > +
> > +	struct drm_xe_vm_create create = {
> > +		.extensions = 0,
> > +		.flags = 0,
> > +	};
> > +
> > +	ret = igt_ioctl(fd, DRM_IOCTL_XE_VM_CREATE, &create);
> > +
> > +	if (ret == 0)
> > +		xe_vm_destroy(fd, create.vm_id);
> > +
> > +	return ret;
> > +}
> > +
> > +/**
> > + * SUBTEST: basic-wedged
> > + * Description: Force Xe device wedged after injecting a failure in GT reset
> > + */
> > +igt_main
> > +{
> > +	int fd;
> > +
> > +	igt_fixture {
> > +		fd = drm_open_driver(DRIVER_XE);
> > +		igt_require(igt_debugfs_exists(fd, "fail_gt_reset/probability",
> > +					       O_RDWR));
> > +	}
> > +
> > +	igt_subtest("basic-wedged") {
> > +		igt_assert_eq(simple_ioctl(fd), 0);
> > +		force_wedged(fd);
> > +		igt_assert_neq(simple_ioctl(fd), 0);
> > +		fd = reload_xe(fd);
> > +		igt_assert_eq(simple_ioctl(fd), 0);
> > +	}
> > +
> > +	igt_fixture {
> > +		if (igt_debugfs_exists(fd, "fail_gt_reset/probability", O_RDWR)) {
> > +			igt_debugfs_write(fd, "fail_gt_reset/probability", "0");
> > +			igt_debugfs_write(fd, "fail_gt_reset/times", "1");
> > +		}
> > +		drm_close_driver(fd);
> > +	}
> > +}
> > diff --git a/tests/meson.build b/tests/meson.build
> > index a856510fc..e590d4348 100644
> > --- a/tests/meson.build
> > +++ b/tests/meson.build
> > @@ -312,6 +312,7 @@ intel_xe_progs = [
> > 	'xe_render_copy',
> > 	'xe_vm',
> > 	'xe_waitfence',
> > +	'xe_wedged',
> > 	'xe_spin_batch',
> > 	'xe_sysfs_defaults',
> > 	'xe_sysfs_scheduler',
> > -- 
> > 2.44.0
> >