Re: [PATCH 4/9] drm/xe: Add a new debug focused configfs group

["Summers, Stuart" <stuart.summers@intel.com>]

On Thu, 2026-05-07 at 15:58 -0700, Matt Roper wrote:
> On Mon, May 04, 2026 at 04:04:11PM -0300, Gustavo Sousa wrote:
> > "Summers, Stuart" <stuart.summers@intel.com> writes:
> > 
> > > On Mon, 2026-05-04 at 14:28 -0300, Gustavo Sousa wrote:
> > > > "Summers, Stuart" <stuart.summers@intel.com> writes:
> > > > 
> > > > > On Mon, 2026-05-04 at 12:42 -0300, Gustavo Sousa wrote:
> > > > > > Stuart Summers <stuart.summers@intel.com> writes:
> > > > > > 
> > > > > > > Add the skeleton code for a new debug specific configfs
> > > > > > > group.
> > > > > > > 
> > > > > > > Signed-off-by: Stuart Summers <stuart.summers@intel.com>
> > > > > > > Assisted-by: Copilot:claude-opus-4.7
> > > > > > > ---
> > > > > > >  drivers/gpu/drm/xe/Makefile            |  2 ++
> > > > > > >  drivers/gpu/drm/xe/xe_configfs.c       |  4 ++++
> > > > > > >  drivers/gpu/drm/xe/xe_configfs_debug.c | 14
> > > > > > > ++++++++++++++
> > > > > > >  drivers/gpu/drm/xe/xe_configfs_debug.h |  8 ++++++++
> > > > > > >  drivers/gpu/drm/xe/xe_configfs_types.h |  7 +++++++
> > > > > > >  5 files changed, 35 insertions(+)
> > > > > > >  create mode 100644
> > > > > > > drivers/gpu/drm/xe/xe_configfs_debug.c
> > > > > > >  create mode 100644
> > > > > > > drivers/gpu/drm/xe/xe_configfs_debug.h
> > > > > > > 
> > > > > > > diff --git a/drivers/gpu/drm/xe/Makefile
> > > > > > > b/drivers/gpu/drm/xe/Makefile
> > > > > > > index 09661f079d03..b58667b0b18e 100644
> > > > > > > --- a/drivers/gpu/drm/xe/Makefile
> > > > > > > +++ b/drivers/gpu/drm/xe/Makefile
> > > > > > > @@ -161,6 +161,8 @@ xe-$(CONFIG_HWMON) += xe_hwmon.o
> > > > > > >  
> > > > > > >  xe-$(CONFIG_PERF_EVENTS) += xe_pmu.o
> > > > > > >  xe-$(CONFIG_CONFIGFS_FS) += xe_configfs.o
> > > > > > > +xe_debug_configfs_obj-$(CONFIG_DRM_XE_DEBUG) :=
> > > > > > > xe_configfs_debug.o
> > > > > > > +xe-$(CONFIG_CONFIGFS_FS) += $(xe_debug_configfs_obj-y)
> > > > > > 
> > > > > > What are the downsides of keeping this available for non-
> > > > > > debug
> > > > > > builds?
> > > > > > 
> > > > > > I imagine that the debug configfs could be useful for
> > > > > > someone not
> > > > > > involved in kernel development working together with the
> > > > > > developers
> > > > > > to
> > > > > > debug issues without needing to rebuild their kernel.
> > > > > 
> > > > > So the focal parameter we had in i915 was enable_rc6.
> > > > > Something
> > > > > like
> > > > > this is what we explicitly don't want to expose to a general
> > > > > user
> > > > > as it
> > > > > can have unforeseen consequences (like having very high power
> > > > > usage).
> > > > > But, having the ability to disable rc6 for debug purposes as
> > > > > a
> > > > > developer is extremely useful at narrowing down certain types
> > > > > of
> > > > > bugs
> > > > > at the hardware/software interface.
> > > > 
> > > > We could probably keep the debug stuff available but cause a
> > > > taint if
> > > > they ever get used?
> > > 
> > > So the tainting is really just for our CI I think. This won't
> > > necessarily impact end users - at least on the client side - who
> > > might
> > > just ignore it (or not even notice it).
> > 
> > I *think* tainting is not only for CI (someone could correct me
> > here).
> > I see that our driver adds a taint when the user is applying some
> > parameters (module parameters or configfs) that alters the
> > "officially
> > supported" behavior of the driver.
> > 
> > Some examples I could find are usages of
> > module_param_named_unsafe() in
> > the module param front, and setup_configfs_post_ctx_restore_bb(),
> > setup_configfs_mid_ctx_restore_bb() in configfs.
> 
> Yeah, there are various reasons to taint the kernel (hardware,
> software,
> or behavioral/usage) and they indicate different things; some are
> more
> serious than others.  See Documentation/admin-guide/tainted-
> kernels.rst
> for details.  The general idea though is that you've entered a
> dangerous/risky situation and we can no longer trust the system to
> behave properly going forward.  I believe the kernel community will
> often ignore or de-prioritize bug reports that arise from tainted
> kernels because there isn't really any guarantee of proper behavior
> or
> support from the community at that point.
> 
> Some of our configfs entries in Xe are "official" and end-user
> focused
> (e.g., SR-IOV configuration settings and such) so those shouldn't
> cause
> any taint.  But others are low-level debugging knobs that are
> effectively handing the user a loaded gun to shoot themselves in the
> foot with.  E.g., the "workaround batchbuffer" configfs stuff is very
> useful for hardware debugging sessions by expert driver developers,
> but
> if used improperly makes it it very easy to mess up both hardware and
> software internal state in ways that make the system unstable. 
> Tainting
> on these developer-focused debug facilities is important because we
> can't trust any bug reports that show up after someone has used those
> settings.  With i915 module parameters we got into situations where a
> bunch of websites and wikis posted horribly inaccurate suggestions
> and
> users would blindly copy those settings without realizing that they
> were
> deliberately breaking their systems, and some of the debug knobs
> we're
> making available in configfs are even more risky.  Honestly, it might
> be
> best to even stick them behind CONFIG_EXPERT and/or other build
> settings.

So this was the reason I used CONFIG_DRM_XE_DEBUG... Should we have
these behind multiple? Or maybe we should force build CONFIG_EXPORT if
CONFIG_DRM_XE_DEBUG is set?

I don't really think tainting the kernel adds much here IMO. Yeah it
does tell the user they're doing something that could cause issues, but
they're already running in a debug mode if they got to this point, so
IMO they already know what they're doing (or should).

Thanks,
Stuart

> 
> 
> Matt
> 
> > 
> > > 
> > > > 
> > > > The scenario I had in mind is someone who is not familiar with
> > > > building
> > > > their own kernel working together with the development
> > > > community on a
> > > > hard-to-reproduce issue and some of the debug configfs
> > > > attributes
> > > > being
> > > > useful to gather more relevant data.
> > > 
> > > Yeah but IMO we really should be able to instruct such a user how
> > > to do
> > > that - the steps there aren't too difficult if they're already
> > > installing kernels etc.
> > > 
> > > A lot of our debug hooks are already held under various Kconfig
> > > options
> > > anyway like DRM_XE_DEBUG, DRM_XE_DEBUG_GUC, DRM_XE_DEBUG_VM, etc.
> > > And a
> > > lot of the information under those Kconfigs are also critical for
> > > different debugs.
> > 
> > Are those due to performance or resource usage implications if they
> > are
> > kept enabled?  If so, is this the same case here?
> > 
> > > 
> > > I'd really prefer to keep these under DRM_XE_DEBUG (or some other
> > > Kconfig) if possible so we can maximize the ease of adding new
> > > parameters as specific debugs come up.
> > 
> > But why would keeping them under a build flag make it easier to add
> > new
> > parameters as opposed to always keeping the knobs available for
> > use?
> > 
> > --
> > Gustavo Sousa
> > 
> > > 
> > > Thanks,
> > > Stuart
> > > 
> > > > 
> > > > --
> > > > Gustavo Sousa
> > > > 
> > > > > 
> > > > > That's just one example, but there are a bunch of things that
> > > > > could
> > > > > fit
> > > > > this category and I expect we'll have even more use cases
> > > > > over
> > > > > time. 
> > > > > 
> > > > > Right now we have to hold these types of changes in a side
> > > > > branch
> > > > > not
> > > > > exposed externally and I really want to converge our
> > > > > processes so
> > > > > we
> > > > > can more easily make these kinds of changes and let us debug
> > > > > directly
> > > > > out of drm-tip.
> > > > > 
> > > > > > 
> > > > > > >  
> > > > > > >  # graphics virtualization (SR-IOV) support
> > > > > > >  xe-y += \
> > > > > > > diff --git a/drivers/gpu/drm/xe/xe_configfs.c
> > > > > > > b/drivers/gpu/drm/xe/xe_configfs.c
> > > > > > > index 12b7fe65446d..85df8ce5cf2a 100644
> > > > > > > --- a/drivers/gpu/drm/xe/xe_configfs.c
> > > > > > > +++ b/drivers/gpu/drm/xe/xe_configfs.c
> > > > > > > @@ -1006,6 +1006,10 @@ static struct config_group
> > > > > > > *xe_config_make_device_group(struct config_group *gro
> > > > > > >                 config_group_init_type_name(&dev->sriov,
> > > > > > > "sriov",
> > > > > > > &xe_config_sriov_type);
> > > > > > >                 configfs_add_default_group(&dev->sriov,
> > > > > > > &dev-
> > > > > > > > group);
> > > > > > >         }
> > > > > > > +#if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
> > > > > > > +       config_group_init_type_name(&dev->debug, "debug",
> > > > > > > &xe_configfs_debug_type);
> > > > > > > +       configfs_add_default_group(&dev->debug, &dev-
> > > > > > > >group);
> > > > > > > +#endif
> > > > > > 
> > > > > > I guess there isn't much we can do about it today, but,
> > > > > > based on
> > > > > > what
> > > > > > I
> > > > > > read in configfs documentation a while ago, I suspect it
> > > > > > wasn't
> > > > > > really
> > > > > > designed with arbitrary hierarchies of config items in
> > > > > > mind.  It
> > > > > > would
> > > > > > be nice if there was an API for that though.
> > > > > > 
> > > > > > Basically, a config group is expected to be a directory to
> > > > > > contain a
> > > > > > set
> > > > > > of config items of the same type, which is not the case on
> > > > > > our
> > > > > > current
> > > > > > usage (i.e. the "sriov" and now the "debug" group).
> > > > > > 
> > > > > > I guess we are able to get our desired behavior (basically
> > > > > > just a
> > > > > > subdirectory behaving as a config item) by not implementing
> > > > > > the
> > > > > > "make_item" hook.  Not sure if that's an expected use of
> > > > > > the API
> > > > > > or
> > > > > > if
> > > > > > that's more of a hack on our side and could come back to
> > > > > > bite us
> > > > > > in
> > > > > > the
> > > > > > future.
> > > > > 
> > > > > Hm.. ok that's good feedback. I do like bucketing these in
> > > > > groups
> > > > > and
> > > > > subdirectories. It also makes it easy to show or not show
> > > > > based on
> > > > > the
> > > > > Kconfig. But let me go through and make sure there isn't
> > > > > something
> > > > > more
> > > > > suited here. Happy to hear any other feedback as well here.
> > > > > 
> > > > > Thanks,
> > > > > Stuart
> > > > > 
> > > > > > 
> > > > > > --
> > > > > > Gustavo Sousa
> > > > > > 
> > > > > > >  
> > > > > > >         mutex_init(&dev->lock);
> > > > > > >  
> > > > > > > diff --git a/drivers/gpu/drm/xe/xe_configfs_debug.c
> > > > > > > b/drivers/gpu/drm/xe/xe_configfs_debug.c
> > > > > > > new file mode 100644
> > > > > > > index 000000000000..45617282cec5
> > > > > > > --- /dev/null
> > > > > > > +++ b/drivers/gpu/drm/xe/xe_configfs_debug.c
> > > > > > > @@ -0,0 +1,14 @@
> > > > > > > +// SPDX-License-Identifier: MIT
> > > > > > > +/*
> > > > > > > + * Copyright © 2026 Intel Corporation
> > > > > > > + */
> > > > > > > +
> > > > > > > +#include <linux/configfs.h>
> > > > > > > +#include <linux/module.h>
> > > > > > > +
> > > > > > > +#include "xe_configfs_debug.h"
> > > > > > > +#include "xe_configfs_types.h"
> > > > > > > +
> > > > > > > +const struct config_item_type xe_configfs_debug_type = {
> > > > > > > +       .ct_owner       = THIS_MODULE,
> > > > > > > +};
> > > > > > > diff --git a/drivers/gpu/drm/xe/xe_configfs_debug.h
> > > > > > > b/drivers/gpu/drm/xe/xe_configfs_debug.h
> > > > > > > new file mode 100644
> > > > > > > index 000000000000..01170dc2f97e
> > > > > > > --- /dev/null
> > > > > > > +++ b/drivers/gpu/drm/xe/xe_configfs_debug.h
> > > > > > > @@ -0,0 +1,8 @@
> > > > > > > +/* SPDX-License-Identifier: MIT */
> > > > > > > +/*
> > > > > > > + * Copyright © 2026 Intel Corporation
> > > > > > > + */
> > > > > > > +#ifndef _XE_CONFIGFS_DEBUG_H_
> > > > > > > +#define _XE_CONFIGFS_DEBUG_H_
> > > > > > > +
> > > > > > > +#endif /* _XE_CONFIGFS_DEBUG_H_ */
> > > > > > > diff --git a/drivers/gpu/drm/xe/xe_configfs_types.h
> > > > > > > b/drivers/gpu/drm/xe/xe_configfs_types.h
> > > > > > > index 935097aafa96..c9d94a3c26a7 100644
> > > > > > > --- a/drivers/gpu/drm/xe/xe_configfs_types.h
> > > > > > > +++ b/drivers/gpu/drm/xe/xe_configfs_types.h
> > > > > > > @@ -24,6 +24,9 @@ struct wa_bb {
> > > > > > >  struct xe_config_group_device {
> > > > > > >         struct config_group group;
> > > > > > >         struct config_group sriov;
> > > > > > > +#if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
> > > > > > > +       struct config_group debug;
> > > > > > > +#endif
> > > > > > >  
> > > > > > >         struct xe_config_device {
> > > > > > >                 struct wa_bb
> > > > > > > ctx_restore_mid_bb[XE_ENGINE_CLASS_MAX];
> > > > > > > @@ -56,4 +59,8 @@ static inline struct xe_config_device
> > > > > > > *xe_configfs_to_device(struct config_item
> > > > > > >         return &xe_configfs_to_group_device(item)-
> > > > > > > >config;
> > > > > > >  }
> > > > > > >  
> > > > > > > +#if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
> > > > > > > +extern const struct config_item_type
> > > > > > > xe_configfs_debug_type;
> > > > > > > +#endif
> > > > > > > +
> > > > > > >  #endif /* _XE_CONFIGFS_TYPES_H_ */
> > > > > > > -- 
> > > > > > > 2.43.0
>