RE: [PATCH] tests/core_hotunplug: add *-with-load subtests exercising GPU workload during hotplug

Hi Janusz,

> -----Original Message-----
> From: Janusz Krzysztofik <[email protected]>
> Sent: Friday, August 21, 2026 3:13 PM
> To: Gote, Nitin R <[email protected]>; [email protected]
> Cc: Auld, Matthew <[email protected]>; Kamil Konieczny
> <[email protected]>
> Subject: Re: [PATCH] tests/core_hotunplug: add *-with-load subtests exercising
> GPU workload during hotplug
> 
> Hi Nitin,
> 
> Good job, however, I think we still need some improvements.
> 
> On Fri, 2026-08-21 at 12:28 +0530, Nitin Gote wrote:
> > The six fd-holding hot* subtests keep a DRM fd open across
> > unbind/unplug but leave the GPU idle, so the kernel hotplug path is
> > never exercised against an active workload.
> >
> > Add six new dedicated *-with-load subtests that run a GPU spinner
> > across the unbind/unplug sequence so the device is removed while a
> > workload is actively using it.
> >
> > The workload helpers use an explicit chipset dispatch block in
> > workload_start()/stop() so that other vendors can contribute support
> > by adding an else-if branch. New subtests are gated on
> > workload_available(), which currently covers DRIVER_XE and
> > DRIVER_INTEL, and returns false (SKIP) on any other driver.
> >
> > igt_spin_free() is avoided in workload_stop() since the device may be
> > gone by then -- igt_spin_end() writes the stop condition via a
> > userspace mmap and lets the DRM fd close reclaim kernel-side resources.
> >
> > Validated on BMG and DG2 with a KASAN-enabled kernel:
> > all six new subtests pass with no KASAN reports.
> >
> > v4: Rebase to latest igt.
> >     Also the failure seen with the new core_hotunplug "with-load"
> >     IGT subtests is fixed by kmd patch:
> >     https://patchwork.freedesktop.org/series/171989/
> >
> > v3: Extend struct gpu_workload with a union of vendor-named members
> >     to match the chipset dispatch in workload_start()/stop(). (Janusz)
> >
> > v2: Pass the hotunplug priv struct to workload_start()/stop()
> >     and dispatch via an explicit if/else-if chipset block to invite
> >     other vendors to contribute. (Janusz)
> >
> > Cc: Matthew Auld <[email protected]>
> > Cc: Janusz Krzysztofik <[email protected]>
> > Cc: Kamil Konieczny <[email protected]>
> > Reviewed-by: Janusz Krzysztofik <[email protected]>
> > Signed-off-by: Nitin Gote <[email protected]>
> > ---
> >  tests/core_hotunplug.c | 294
> > +++++++++++++++++++++++++++++++++++++++++
> >  1 file changed, 294 insertions(+)
> >
> > diff --git a/tests/core_hotunplug.c b/tests/core_hotunplug.c index
> > 33de65369..b11a4231b 100644
> > --- a/tests/core_hotunplug.c
> > +++ b/tests/core_hotunplug.c
> > @@ -80,6 +80,34 @@
> >   * SUBTEST: unplug-rescan
> >   * Description: Check if a device believed to be closed can be cleanly
> >   *		unplugged, then restored
> > + *
> > + * SUBTEST: hotunbind-rebind-with-load
> > + * Description: Check if the driver can be cleanly unbound from an open device
> > + *		with a background GPU workload in flight, then released and
> > + *		rebound
> > + *
> > + * SUBTEST: hotunplug-rescan-with-load
> > + * Description: Check if an open device with a background GPU workload in
> > + *		flight can be cleanly unplugged, then released and restored
> > + *
> > + * SUBTEST: hotrebind-with-load
> > + * Description: Check if the driver can be cleanly rebound to a device with a
> > + *		still open hot unbound driver instance while a background GPU
> > + *		workload is in flight
> > + *
> > + * SUBTEST: hotreplug-with-load
> > + * Description: Check if a hot unplugged and still open device can be cleanly
> > + *		restored while a background GPU workload is in flight
> > + *
> > + * SUBTEST: hotrebind-lateclose-with-load
> > + * Description: Check if a hot unbound driver instance still open after hot
> > + *		rebind with a background GPU workload in flight can be cleanly
> > + *		released
> > + *
> > + * SUBTEST: hotreplug-lateclose-with-load
> > + * Description: Check if an instance of a still open while hot replugged
> > + *		device with a background GPU workload in flight can be cleanly
> > + *		released
> >   */
> >
> >  IGT_TEST_DESCRIPTION("Examine behavior of a driver on device hot
> > unplug"); @@ -550,6 +578,56 @@ static void post_healthcheck(struct
> hotunplug *priv)
> >  	cleanup(priv);
> >  }
> >
> > +/* GPU workload helpers */
> > +
> > +struct gpu_workload {
> > +	union {
> > +		struct {
> > +			igt_spin_t *spin;
> > +			uint64_t ahnd;
> > +		} intel; /* Xe and i915 */
> > +	};
> > +};
> > +
> > +static bool workload_available(int chipset) {
> > +	return chipset == DRIVER_XE || chipset == DRIVER_INTEL;
> 
> NIT: I think a switch (chipset) { case DRIVER_XE: ... } construct, here
> and consequently below in workload_start/stop, could be more effective and
> more easily extendable with potential future additions from other vendors, but
> that can be refactored later, on a first such addition.
> 

> > +}
> > +
> > +static void workload_start(struct hotunplug *priv, int fd,
> > +			   struct gpu_workload *w)
> > +{
> > +	if (priv->chipset == DRIVER_XE || priv->chipset == DRIVER_INTEL) {
> > +		local_debug("%s\n", "starting GPU spinner");
> > +		priv->failure = "GPU workload start failure!";
> > +		w->intel.ahnd = intel_allocator_open(fd, 0,
> INTEL_ALLOCATOR_RELOC);
> > +		w->intel.spin = igt_spin_new(fd, .ahnd = w->intel.ahnd);
> > +		priv->failure = NULL;
> > +	}
> > +}
> > +
> > +static void workload_stop(struct hotunplug *priv, struct gpu_workload
> > +*w) {
> > +	if (priv->chipset == DRIVER_XE || priv->chipset == DRIVER_INTEL) {
> > +		if (!w->intel.spin)
> > +			return;
> 
> Since igt_fail() can be called from inside igt_spin_new(), workload_start() can fail
> leaving an open ahnd but no spin.  We need to handle that case.
> 
> > +
> > +		/*
> > +		 * The device may be gone (unbind/unplug), so avoid
> > +		 * igt_spin_free(): on Xe it would wait forever on a syncobj
> > +		 * that never signals; on i915 it would touch the dead fd via
> > +		 * gem_munmap()/gem_close(). Instead, end the spinner via a
> > +		 * userspace mmap write and let the DRM fd close reclaim the
> > +		 * kernel-side resources.
> > +		 */
> > +		local_debug("%s\n", "stopping GPU spinner");
> > +		igt_spin_end(w->intel.spin);
> > +		put_ahnd(w->intel.ahnd);
> > +		w->intel.spin = NULL;
> > +		w->intel.ahnd = 0;
> > +	}
> > +}
> > +
> >  /* Subtests */
> >
> >  static void unbind_rebind(struct hotunplug *priv) @@ -664,6 +742,150
> > @@ static void hotreplug_lateclose(struct hotunplug *priv)
> >  	igt_assert_f(healthcheck(priv, false), "%s\n", priv->failure);  }
> >
> > +static void hotunbind_rebind_with_load(struct hotunplug *priv) {
> > +	struct gpu_workload w = { 0 };
> > +
> > +	pre_check(priv);
> > +
> > +	igt_require_f(workload_available(priv->chipset),
> > +		      "No GPU workload support for this driver\n");
> > +
> > +	priv->fd.drm = local_drm_open_driver(false, "", " for hot unbind");
> > +
> > +	workload_start(priv, priv->fd.drm, &w);
> > +
> > +	driver_unbind(priv, "hot ", 0);
> 
> Since igt_fail() can be called from driver_unbind(), as well as from other functions
> called between workload_start() and workload_stop() in other subtests below, we
> need to address those cases to be on the safe side.
> We could e.g.:
> - embed struct gpu_workload inside struct hotunplug,
> - extend cleanup() (or recover()) function with another workload_stop().
> 

Thank you for the improvement suggestions. I have noted all three comments.
As this patch has been merged in between, I will address these review comments in a separate patch.

- Nitin

> Thanks,
> Janusz
> 
>