Call drm_gem_fb_begin_cpu_access() and drm_gem_fb_end_cpu_access()
around cursor image updates. Imported buffers might have to be
synchronized for CPU access before they can be used.
Ignore errors from drm_gem_fb_begin_cpu_access(). These errors can
often be transitory. The cursor image will be updated on the next
frame. Meanwhile display a white square where the cursor would be.
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Jocelyn Falempe <jfalempe@redhat.com>>
Link: https://patch.msgid.link/20251126094626.41985-4-tzimmermann@suse.de
Copy the ARGB4444 cursor buffer to system memory if it is located in
I/O memory. While this cannot happen with ast's native GEM objects, an
imported buffer object might be on the external device's I/O memory.
If the cursor buffer is located in system memory continue to use it
directly.
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Jocelyn Falempe <jfalempe@redhat.com>>
Link: https://patch.msgid.link/20251126094626.41985-3-tzimmermann@suse.de
Move the format conversion of the cursor framebuffer into the new
helper ast_cursor_plane_get_argb4444(). It returns a buffer in system
memory, which the atomic_update handler copies to video memory.
The returned buffer is either the GEM buffer itself, or a temporary
copy within the plane in ARGB4444 format.
As a small change, list supported formats explicitly in the switch
statement. Do not assume ARGB8888 input by default. The cursor
framebuffer knows its format, so should we.
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Jocelyn Falempe <jfalempe@redhat.com>
Link: https://patch.msgid.link/20251126094626.41985-2-tzimmermann@suse.de
All of the necessary building blocks are now in place to support SR-IOV
VF migration.
Flip the enable/disable logic to match VF code and disable the feature
only for platforms that don't meet the necessary prerequisites.
To allow more testing and experiments, on DEBUG builds any missing
prerequisites will be ignored.
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
Link: https://patch.msgid.link/20251127093934.1462188-2-michal.winiarski@intel.com
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
During some IGT tests (e.g. xe_pm@s2idle-exec-after, xe_pm@s2idle-mocs)
sink disconnects across suspend/resume, reconnecting later during resume
at some point. Hence during resume, where the driver is restoring the
pre-suspend mode, all the AUX transfers to the sink are expected to
fail.
Switch error message to KMS debug message of failed FEC enabling.
Signed-off-by: Michał Grzelak <michal.grzelak@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Imre Deak <imre.deak@intel.com>
Link: https://patch.msgid.link/20251117082046.4190705-1-michal.grzelak@intel.com
Currently the user can write anything into the drm.panic_screen modparam,
either at runtime via sysfs, or as a kernel boot time argument. Invalid
strings will be silently accepted and ignored at use time by defaulting to
the 'user' panic mode.
Let instead add some validation in order to have immediate feedback when
something has been mistyped, or not compiled in.
For example during kernel boot:
Booting kernel: `bsod' invalid for parameter `drm.panic_screen'
Or at runtime:
# echo -n bsod > /sys/module/drm/parameters/panic_screen
-bash: echo: write error: Invalid argument
Change of behavior is that when invalid mode is attempted to be
configured, currently the code will default to the 'user' mode, while with
this change the code will ignore it, and default to the mode set at kernel
build time via CONFIG_DRM_PANIC_SCREEN.
While at it lets also fix the module parameter description to include all
compiled in modes.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
Cc: Jocelyn Falempe <jfalempe@redhat.com>
Cc: Javier Martinez Canillas <javierm@redhat.com>
Reviewed-by: Javier Martinez Canillas <javierm@redhat.com>
Reviewed-by: Jocelyn Falempe <jfalempe@redhat.com>
Signed-off-by: Tvrtko Ursulin <tursulin@ursulin.net>
Link: https://lore.kernel.org/r/20251127090349.92717-1-tvrtko.ursulin@igalia.com
If the group is already assigned a slot but was idle before this job
submission, we need to make sure the priority rotation happens in the
future. Extract the existing logic living in group_schedule_locked()
and call this new sched_resume_tick() helper from the "group is
assigned a slot" path.
v2:
- Add R-b
v3:
- Re-use queue_mask to clear the bit
- Collect R-b
Fixes: de85488138 ("drm/panthor: Add the scheduler logical block")
Reviewed-by: Steven Price <steven.price@arm.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Link: https://patch.msgid.link/20251128094839.3856402-8-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
When we have multiple active groups with the same priority, we need to
keep ticking for the priority rotation to take place. If we don't do
that, we might starve slots with lower priorities.
It's annoying to deal with that in tick_ctx_update_resched_target(),
so let's add a ::stop_tick field to the tick context which is
initialized to true, and downgraded to false as soon as we detect
something that requires to tick to happen. This way we can complement
the current logic with extra conditions if needed.
v2:
- Add R-b
v3:
- Drop panthor_sched_tick_ctx::min_priority (no longer relevant)
- Collect R-b
Fixes: de85488138 ("drm/panthor: Add the scheduler logical block")
Reviewed-by: Steven Price <steven.price@arm.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Link: https://patch.msgid.link/20251128094839.3856402-7-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
We have a few paths where we schedule the tick work immediately without
changing the resched_target. If the tick was stopped, this would lead
to a remaining_jiffies that's always > 0, and it wouldn't force a full
tick in that case. Add extra checks to cover that case properly.
v2:
- Fix typo
- Simplify the code as suggested by Steve
v3:
- Collect R-b
Fixes: de85488138 ("drm/panthor: Add the scheduler logical block")
Reviewed-by: Steven Price <steven.price@arm.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Link: https://patch.msgid.link/20251128094839.3856402-6-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
When rotating group priorities, we want the group with the
highest priority to go back to the end of the queue, and all
other active groups to get their priority bumped, otherwise
some groups will never get a chance to run with the highest
priority. This implies moving the rotation itself to
tick_work(), and only dealing with old group ordering in
tick_ctx_insert_old_group().
v2:
- Add R-b
- Fix the commit message
v3:
- Drop the full_tick argument in tick_ctx_init()
- Collect R-b
Fixes: de85488138 ("drm/panthor: Add the scheduler logical block")
Reviewed-by: Steven Price <steven.price@arm.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Link: https://patch.msgid.link/20251128094839.3856402-5-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Not only this only works once, because of how extract events work
(event is enabled if the req and ack bit differ, and it's signalled
by the FW by setting identical req and ack, to re-enable the event,
we need to toggle the bit, which we never do). But more importantly,
we never do anything with this event, so we're better off dropping it
when programming the CS slot.
v2:
- Add R-b
v3:
- Collect R-b
Reviewed-by: Steven Price <steven.price@arm.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Link: https://patch.msgid.link/20251128094839.3856402-3-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Groups are only moved out of the runnable lists when
panthor_group_stop() is called or when they run out of jobs.
What should not happen though is having one group added to one of
the runnable list after reset.in_progress has been set to true, but
that's not something we can easily check, so let's just drop the
WARN_ON() in panthor_sched_pre_reset().
v2:
- Adjust explanation in commit message
v3:
- Collect R-b
v4:
- No changes
Reviewed-by: Liviu Dudau <liviu.dudau@arm.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Link: https://patch.msgid.link/20251128084841.3804658-7-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
There's no reason for panthor_vm_[un]map_pages() to fail unless the
drm_gpuvm state and the page table are out of sync, so let's reflect that
by making panthor_vm_unmap_pages() a void function and adding
WARN_ON()s in various places. We also try to recover from those
unexpected mismatch by checking for already unmapped ranges and skipping
them. But there's only so much we can do to try and cope with such
SW bugs, so when we see a mismatch, we flag the VM unusable and disable
the AS to avoid further GPU accesses to the memory.
It could be that the as_disable() call fails because the MMU unit is
stuck, in which case the whole GPU is frozen, and only a GPU reset can
unblock things. Ater the reset, the VM will be seen as unusable and
any attempt to re-use it will fail, so we should be covered for any
use-after-unmap issues.
v2:
- Fix double unlock
v3:
- Collect R-b
v4:
- No changes
Reviewed-by: Steven Price <steven.price@arm.com>
Link: https://patch.msgid.link/20251128084841.3804658-6-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Move the lock/flush_mem operations around the gpuvm_sm_[un]map() calls
so we can implement true atomic page updates, where any access in the
locked range done by the GPU has to wait for the page table updates
to land before proceeding.
This is needed for vkQueueBindSparse(), so we can replace the dummy
page mapped over the entire object by actual BO backed pages in an atomic
way. But it's also useful to avoid "AS_ACTIVE bit stuck" failures in
the sm_[un]map() path, leading to gpuvm state inconsistencies.
v2:
- Adjust to match the two new preliminary patches
v3:
- Collect R-b
v4:
- No changes
Reviewed-by: Steven Price <steven.price@arm.com>
Link: https://patch.msgid.link/20251128084841.3804658-5-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
We have seen a few cases where the whole memory subsystem is blocked
and flush operations never complete. When that happens, we want to:
- schedule a reset, so we can recover from this situation
- in the reset path, we need to reset the pending_reqs so we can send
new commands after the reset
- if more panthor_gpu_flush_caches() operations are queued after
the timeout, we skip them and return -EIO directly to avoid needless
waits (the memory block won't miraculously work again)
Note that we drop the WARN_ON()s because these hangs can be triggered
with buggy GPU jobs created by the UMD, and there's no way we can
prevent it. We do keep the error messages though.
v2:
- New patch
v3:
- Collect R-b
- Explicitly mention the fact we dropped the WARN_ON()s in the commit
message
v4:
- No changes
Fixes: 5cd894e258 ("drm/panthor: Add the GPU logical block")
Reviewed-by: Steven Price <steven.price@arm.com>
Link: https://patch.msgid.link/20251128084841.3804658-4-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
There's currently no situation where we want to issue a command to an
AS and not wait for this command to complete. The wait is either
explicitly done (LOCK, UNLOCK) or it's missing (UPDATE). So let's
turn write_cmd() into as_send_cmd_and_wait() that has the wait after
a command is sent.
v2:
- New patch
v3:
- Collect R-b
v4:
- No changes
Reviewed-by: Steven Price <steven.price@arm.com>
Link: https://patch.msgid.link/20251128084841.3804658-2-boris.brezillon@collabora.com
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
This commit prevents the possibility of a use after free issue in the
GROUP_CREATE ioctl function, which arose as pointer to the group is
accessed in that ioctl function after storing it in the Xarray.
A malicious userspace can second guess the handle of a group and try
to call GROUP_DESTROY ioctl from another thread around the same time
as GROUP_CREATE ioctl.
To prevent the use after free exploit, this commit uses a mark on an
entry of group pool Xarray which is added just before returning from
the GROUP_CREATE ioctl function. The mark is checked for all ioctls
that specify the group handle and so userspace won't be abe to delete
a group that isn't marked yet.
v2: Add R-bs and fixes tags
Fixes: de85488138 ("drm/panthor: Add the scheduler logical block")
Co-developed-by: Boris Brezillon <boris.brezillon@collabora.com>
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Signed-off-by: Akash Goel <akash.goel@arm.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Link: https://patch.msgid.link/20251127164912.3788155-1-akash.goel@arm.com
This reverts commit 5a9b0c7418.
The switch from AUX interrupts to pollign was very hand-wavy.
Yes, there have been some situations in CI on a few platforms
where the AUX hardware seemingly forgets to signal the timeout,
but those have been happening after we switched to polling as
well. So I don't think we have any conclusive evidence that
polling actually helps here.
Someone really should root cause the actual problem, and see
if there is a proper workaround we could implemnt (eg. disabling
clock gating/etc.). In the meantime just go back to using the
interrupt for AUX completion.
If the hardware fails to signal the timeout we will just hit
the wait_event_timeout() software timeout instead. I suppose
we could try to tune the software timeout to more closely
match the expected hardware timeout. Might need to use
wait_event_hrtimeout() or something to avoid jiffies
granularity issues...
The AUX polling is also a hinderance towards using poll_timeout_us()
because we have a very long timeout, but would need a fairly short
polling interval to keep AUX transfer reasonably fast. Someone would
need to come up with good numbers in a somewhat scientific way.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patch.msgid.link/20251119185310.10428-3-ville.syrjala@linux.intel.com
Acked-by: Jani Nikula <jani.nikula@intel.com>
Another (somewhat expensive) drm_format_info() call has
appeared in intel_plane_can_async_flip(). That one may get
called several times per commit so we need to get rid of
it.
Fortunately most callers already have the framebuffer at
hand, so we can just grab the format info from there.
The one exception is intel_plane_format_mod_supported_async()
where we have to do the lookup. But that only gets called
(a bunch of times) during driver init to build the
IN_FORMATS_ASYNC blob, and afterwards there is no runtime
cost.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patch.msgid.link/20251112233030.24117-4-ville.syrjala@linux.intel.com
Reviewed-by: Jouni Högander <jouni.hogander@intel.com>
Since old kernel versions wouldn't expose the IN_FORMATS_ASYNC blob,
userspace can't really use the absence of the blob to determine
that async flips aren't supported. Thus it seems better to always
expose the blob on all planes, whether they support async flips
or not. The blob will simply not indicate any format+modifier
combinations as supported on planes that aren't async flip capable.
Currently we expose the blob for all skl+ universal planes (even
though we implement async flips only for the first plane on each
pipe), and i9xx primary planes (for ilk+ we have async flips support,
for pre-ilk we do not). Complete the full set by also expsosing
the blob on pre-skl sprite planes, and cursors.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patch.msgid.link/20251112233030.24117-3-ville.syrjala@linux.intel.com
Reviewed-by: Jouni Högander <jouni.hogander@intel.com>
If power state is retained between suspend/resume cycle, we don't need
to perform full GT re-initialization. Introduce runtime helpers for GT
which greatly reduce suspend/resume delay.
v2: Drop redundant xe_gt_sanitize() and xe_guc_ct_stop() (Daniele)
Use runtime naming for guc helpers (Daniele)
v3: Drop redundant logging, add kernel doc (Michal)
Use runtime naming for ct helpers (Michal)
v4: Fix tags (Rodrigo)
v5: Include host_l2_vram workaround (Daniele)
Reuse xe_guc_submit_enable/disable() helpers (Daniele)
Co-developed-by: Riana Tauro <riana.tauro@intel.com>
Signed-off-by: Riana Tauro <riana.tauro@intel.com>
Signed-off-by: Raag Jadav <raag.jadav@intel.com>
Acked-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Link: https://patch.msgid.link/20251030122357.128825-5-raag.jadav@intel.com
Introduce pause/unpause() helpers which stop/start further runs of
submission tasks on given GuC and can be called from PF context. This
is in preparation of usecases where we simply need to stop/start the
scheduler without losing GuC state and don't require dealing with VF
migration.
v7: Reword commit message (Matthew Brost)
Signed-off-by: Raag Jadav <raag.jadav@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Link: https://patch.msgid.link/20251030122357.128825-3-raag.jadav@intel.com
Previously, VRAM TTM resource manager debugfs entries (vram0_mm / vram1_mm)
were created globally in the XE debugfs root directory. But technically,
each tile has an associated VRAM TTM manager, which it can own.
Let's create VRAM memory manager debugfs entries directly under each tile's
debugfs directory for better alignment with the per-tile memory layout.
Signed-off-by: Piotr Piórkowski <piotr.piorkowski@intel.com>
Reviewed-by: Stuart Summers <stuart.summers@intel.com>
Link: https://patch.msgid.link/20251127073643.144379-1-piotr.piorkowski@intel.com
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
