All of the MediaTek SoCs supported by Panfrost can completely cut power
to the GPU during full system sleep without any user-noticeable delay
in the resume operation, as shown by measurements taken on multiple
MediaTek SoCs (MT8183/86/92/95).
As an example, for MT8195 - a "before" with only runtime PM operations
(so, without turning on/off regulators), and an "after" executing both
the system sleep .resume() handler and .runtime_resume() (so the time
refers to T_Resume + T_Runtime_Resume):
Average Panfrost-only system sleep resume time, before: ~33500ns
Average Panfrost-only system sleep resume time, after: ~336200ns
Keep in mind that this additional ~308200 nanoseconds delay happens only
in resume from a full system suspend, and not in runtime PM operations,
hence it is acceptable.
Measurements were also taken on MT8186, showing a delay of ~312000 ns.
Testing of this happened on all of the aforementioned MediaTek SoCs, but:
MT8183 got tested only by KernelCI with <=10 suspend/resume cycles
MT8186, MT8192, MT8195 were tested manually with over 100 suspend/resume
cycles with GNOME DE (Mutter + Wayland).
Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Signed-off-by: Steven Price <steven.price@arm.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231109102543.42971-7-angelogioacchino.delregno@collabora.com
All of the MediaTek SoCs supported by Panfrost can switch the clocks
off and on during system sleep to save some power without any user
experience penalty.
Measurements taken on multiple MediaTek SoCs (MT8183/8186/8192/8195)
show that adding this will not prolong the time that is required to
resume the system in any meaningful way.
As an example, for MT8195 - a "before" with only runtime PM operations
(so, without turning on/off GPU clocks), and an "after" executing both
the system sleep .resume() handler and .runtime_resume() (so the time
refers to T_Resume + T_Runtime_Resume):
Average Panfrost-only system sleep resume time, before: ~28000ns
Average Panfrost-only system sleep resume time, after: ~33500ns
Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Signed-off-by: Steven Price <steven.price@arm.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231109102543.42971-5-angelogioacchino.delregno@collabora.com
Currently, the GPU is being internally powered off for runtime suspend
and turned back on for runtime resume through commands sent to it, but
note that the GPU doesn't need to be clocked during the poweroff state,
hence it is possible to save some power on selected platforms.
Add suspend and resume handlers for full system sleep and then add
a new panfrost_gpu_pm enumeration and a pm_features variable in the
panfrost_compatible structure: BIT(GPU_PM_CLK_DIS) will be used to
enable this power saving technique only on SoCs that are able to
safely use it.
Note that this was implemented only for the system sleep case and not
for runtime PM because testing on one of my MediaTek platforms showed
issues when turning on and off clocks aggressively (in PM runtime)
resulting in a full system lockup.
Doing this only for full system sleep never showed issues during my
testing by suspending and resuming the system continuously for more
than 100 cycles.
Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Signed-off-by: Steven Price <steven.price@arm.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231109102543.42971-4-angelogioacchino.delregno@collabora.com
In many cases, soft reset takes more than 1 microsecond, but definitely
less than 10; moreover in the poweron flow, tilers, shaders and l2 will
become ready (each) in less than 10 microseconds as well.
Even in the cases (at least on my platforms, rarely) in which those take
more than 10 microseconds, it's very unlikely to see both soft reset and
poweron to take more than 70 microseconds.
Shorten the polling delay to 10 microseconds to consistently reduce the
runtime resume time of the GPU.
As an indicative example, measurements taken on a MediaTek MT8195 SoC
Average runtime resume time in nanoseconds before this commit:
GDM, user selection up/down: 88435ns
GDM, Text Entry (typing user/password): 91489ns
GNOME Desktop, idling, GKRELLM running: 73200ns
After this commit:
GDM: user selection up/down: 26690ns
GDM: Text Entry (typing user/password): 27917ns
GNOME Desktop, idling, GKRELLM running: 25304ns
Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Signed-off-by: Steven Price <steven.price@arm.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231109102543.42971-3-angelogioacchino.delregno@collabora.com
Even though soft reset should ideally never fail, during development of
some power management features I managed to get some bits wrong: this
resulted in GPU soft reset failures, where the GPU was never able to
recover, not even after suspend/resume cycles, meaning that the only
way to get functionality back was to reboot the machine.
Perform a hard reset after a soft reset failure to be able to recover
the GPU during runtime (so, without any machine reboot).
Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Signed-off-by: Steven Price <steven.price@arm.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231109102543.42971-2-angelogioacchino.delregno@collabora.com
The layout of the registers {TILER,SHADER,L2}_PWROFF_LO, used to request
powering off cores, is the same as the {TILER,SHADER,L2}_PWRON_LO ones:
this means that in order to request poweroff of cores, we are supposed
to write a bitmask of cores that should be powered off!
This means that the panfrost_gpu_power_off() function has always been
doing nothing.
Fix powering off the GPU by writing a bitmask of the cores to poweroff
to the relevant PWROFF_LO registers and then check that the transition
(from ON to OFF) has finished by polling the relevant PWRTRANS_LO
registers.
While at it, in order to avoid code duplication, move the core mask
logic from panfrost_gpu_power_on() to a new panfrost_get_core_mask()
function, used in both poweron and poweroff.
Fixes: f3ba91228e ("drm/panfrost: Add initial panfrost driver")
Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Reviewed-by: Steven Price <steven.price@arm.com>
Signed-off-by: Steven Price <steven.price@arm.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231102141507.73481-1-angelogioacchino.delregno@collabora.com
Currently, job flow control is implemented simply by limiting the number
of jobs in flight. Therefore, a scheduler is initialized with a credit
limit that corresponds to the number of jobs which can be sent to the
hardware.
This implies that for each job, drivers need to account for the maximum
job size possible in order to not overflow the ring buffer.
However, there are drivers, such as Nouveau, where the job size has a
rather large range. For such drivers it can easily happen that job
submissions not even filling the ring by 1% can block subsequent
submissions, which, in the worst case, can lead to the ring run dry.
In order to overcome this issue, allow for tracking the actual job size
instead of the number of jobs. Therefore, add a field to track a job's
credit count, which represents the number of credits a job contributes
to the scheduler's credit limit.
Signed-off-by: Danilo Krummrich <dakr@redhat.com>
Reviewed-by: Luben Tuikov <ltuikov89@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231110001638.71750-1-dakr@redhat.com
Both the drm_buddy and drm_mm tests have been converted from selftest to
kunit recently.
However, a significant portion of them are trying to exert some part of
their API over a huge number of iterations and with random variations of
their parameters. They are thus more a way to discover new bugs than
actual unit tests.
This is fine in itself but leads to very slow runtime (up to 25s for
some drm_test_mm_reserve and drm_test_mm_insert on a Ryzen 7950x while
running the tests in qemu) which make them a poor fit for kunit.
Let's remove those tests from the drm_mm and drm_buddy test suites for
now, and if it's ever needed we can always create proper unit tests for
them later on.
This made the entire DRM tests execution time (as of v6.6-rc1) come from
65s to less than .5s on a Ryzen 7950x system when running under qemu,
and from 9 minutes to about 4s on a RaspberryPi4.
Acked-by: Daniel Vetter <daniel@ffwll.ch>
Suggested-by: Daniel Vetter <daniel@ffwll.ch>
Link: https://lore.kernel.org/r/20231025132428.723672-1-mripard@kernel.org
Signed-off-by: Maxime Ripard <mripard@kernel.org>
Don't call drm_sched_select_entity() in drm_sched_run_job_queue(). In fact,
rename __drm_sched_run_job_queue() to just drm_sched_run_job_queue(), and let
it do just that, schedule the work item for execution.
The problem is that drm_sched_run_job_queue() calls drm_sched_select_entity()
to determine if the scheduler has an entity ready in one of its run-queues,
and in the case of the Round-Robin (RR) scheduling, the function
drm_sched_rq_select_entity_rr() does just that, selects the _next_ entity
which is ready, sets up the run-queue and completion and returns that
entity. The FIFO scheduling algorithm is unaffected.
Now, since drm_sched_run_job_work() also calls drm_sched_select_entity(), then
in the case of RR scheduling, that would result in drm_sched_select_entity()
having been called twice, which may result in skipping a ready entity if more
than one entity is ready. This commit fixes this by eliminating the call to
drm_sched_select_entity() from drm_sched_run_job_queue(), and leaves it only
in drm_sched_run_job_work().
v2: Rebased on top of Tvrtko's renames series of patches. (Luben)
Add fixes-tag. (Tvrtko)
Signed-off-by: Luben Tuikov <ltuikov89@gmail.com>
Fixes: f7fe64ad0f ("drm/sched: Split free_job into own work item")
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Danilo Krummrich <dakr@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231107041020.10035-2-ltuikov89@gmail.com
The previous patch exposed the accumulated amount of active time per
client for each V3D queue. But this doesn't provide a global notion of
the GPU usage.
Therefore, provide the accumulated amount of active time for each V3D
queue (BIN, RENDER, CSD, TFU and CACHE_CLEAN), considering all the jobs
submitted to the queue, independent of the client.
This data is exposed through the sysfs interface, so that if the
interface is queried at two different points of time the usage percentage
of each of the queues can be calculated.
Co-developed-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Signed-off-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Signed-off-by: Maíra Canal <mcanal@igalia.com>
Acked-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Reviewed-by: Melissa Wen <mwen@igalia.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230905213416.1290219-3-mcanal@igalia.com
This patch exposes the accumulated amount of active time per client
through the fdinfo infrastructure. The amount of active time is exposed
for each V3D queue: BIN, RENDER, CSD, TFU and CACHE_CLEAN.
In order to calculate the amount of active time per client, a CPU clock
is used through the function local_clock(). The point where the jobs has
started is marked and is finally compared with the time that the job had
finished.
Moreover, the number of jobs submitted to each queue is also exposed on
fdinfo through the identifier "v3d-jobs-<queue>".
Co-developed-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Signed-off-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Signed-off-by: Maíra Canal <mcanal@igalia.com>
Acked-by: Jose Maria Casanova Crespo <jmcasanova@igalia.com>
Reviewed-by: Melissa Wen <mwen@igalia.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230905213416.1290219-3-mcanal@igalia.com
It is better not to lose precision and not revert to 1 MiB size
granularity for every size greater than 1 MiB.
Sizes in KiB should not be so troublesome to read (and in fact machine
parsing is I expect the norm here), they align with other api like
/proc/meminfo, and they allow writing tests for the interface without
having to embed drm.ko implementation knowledge into them. (Like knowing
that minimum buffer size one can use for successful verification has to be
1MiB aligned, and on top account for any pre-existing memory utilisation
outside of driver's control.)
But probably even more importantly I think that it is just better to show
the accurate sizes and not arbitrary lose precision for a little bit of a
stretched use case of eyeballing fdinfo text directly.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Rob Clark <robdclark@gmail.com>
Cc: Adrián Larumbe <adrian.larumbe@collabora.com>
Cc: steven.price@arm.com
Reviewed-by: Steven Price <steven.price@arm.com>
Link: https://lore.kernel.org/r/20230927133843.247957-2-tvrtko.ursulin@linux.intel.com
Signed-off-by: Maxime Ripard <mripard@kernel.org>
The current name makes it sound like helper will free a queue, while what
it does is it enqueues the free job worker.
Rename it to drm_sched_run_free_queue to align with existing
drm_sched_run_job_queue.
Despite that creating an illusion there are two queues, while in reality
there is only one, at least it creates a consistent naming for the two
enqueuing helpers.
At the same time simplify the "if done" helper by dropping the suffix and
adding a double underscore prefix to the one which just enqueues.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Luben Tuikov <ltuikov89@gmail.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231102105538.391648-4-tvrtko.ursulin@linux.intel.com
Reviewed-by: Luben Tuikov <ltuikov89@gmail.com>
Signed-off-by: Luben Tuikov <ltuikov89@gmail.com>
Add support to partially execute a slice which is resized to zero.
Executing a zero size slice in a BO should mean that there is no DMA
transfers involved but you should still configure doorbell and semaphores.
For example consider a BO of size 18K and it is sliced into 3 6K slices
and user calls partial execute ioctl with resize as 10K.
slice 0 - size is 6k and offset is 0, so resize of 10K will not cut short
this slice hence we send the entire slice for execution.
slice 1 - size is 6k and offset is 6k, so resize of 10K will cut short this
slice and only the first 4k should be DMA along with configuring
doorbell and semaphores.
slice 2 - size is 6k and offset is 12k, so resize of 10k will cut short
this slice and no DMA transfer would be involved but we should
would configure doorbell and semaphores.
This change begs to change the behavior of 0 resize. Currently, 0 resize
partial execute ioctl behaves exactly like execute ioctl i.e. no resize.
After this patch all the slice in BO should behave exactly like slice 2 in
above example.
Refactor copy_partial_exec_reqs() to make it more readable and less
complex.
Signed-off-by: Pranjal Ramajor Asha Kanojiya <quic_pkanojiy@quicinc.com>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Signed-off-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Reviewed-by: Stanislaw Gruszka <stanislaw.gruszka@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231027164330.11978-1-quic_jhugo@quicinc.com
Current wrapper is right-sized to the message being transferred;
however, this is smaller than the structure defining message wrappers
since the trailing element is a union of message/transfer headers of
various sizes (8 and 32 bytes on 32-bit system where issue was
reported). Using the smaller header with a small message
(wire_trans_dma_xfer is 24 bytes including header) ends up being smaller
than a wrapper with the larger header. There are no accesses outside of
the defined size, however they are possible if the larger union member
is referenced.
Abort messages are outside of hot-path and changing the wrapper struct
would require a larger rewrite, so having the memory allocated to the
message be 8 bytes too big is acceptable.
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202310182253.bcb9JcyJ-lkp@intel.com/
Signed-off-by: Carl Vanderlip <quic_carlv@quicinc.com>
Reviewed-by: Pranjal Ramajor Asha Kanojiya <quic_pkanojiy@quicinc.com>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Signed-off-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Reviewed-by: Stanislaw Gruszka <stanislaw.gruszka@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231027180810.4873-1-quic_jhugo@quicinc.com
The simple-framebuffer device tree bindings document the power-domains
property, so make sure that simplefb supports it. This ensures that the
power domains remain enabled as long as simplefb is active.
v2: - remove unnecessary call to simplefb_detach_genpds() since that's
already done automatically by devres
- fix crash if power-domains property is missing in DT
Signed-off-by: Thierry Reding <treding@nvidia.com>
Reviewed-by: Hans de Goede <hdegoede@redhat.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231101172017.3872242-3-thierry.reding@gmail.com
Rather than call free_job and run_job in same work item have a dedicated
work item for each. This aligns with the design and intended use of work
queues.
v2:
- Test for DMA_FENCE_FLAG_TIMESTAMP_BIT before setting
timestamp in free_job() work item (Danilo)
v3:
- Drop forward dec of drm_sched_select_entity (Boris)
- Return in drm_sched_run_job_work if entity NULL (Boris)
v4:
- Replace dequeue with peek and invert logic (Luben)
- Wrap to 100 lines (Luben)
- Update comments for *_queue / *_queue_if_ready functions (Luben)
v5:
- Drop peek argument, blindly reinit idle (Luben)
- s/drm_sched_free_job_queue_if_ready/drm_sched_free_job_queue_if_done (Luben)
- Update work_run_job & work_free_job kernel doc (Luben)
v6:
- Do not move drm_sched_select_entity in file (Luben)
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Link: https://lore.kernel.org/r/20231031032439.1558703-4-matthew.brost@intel.com
Reviewed-by: Luben Tuikov <ltuikov89@gmail.com>
Signed-off-by: Luben Tuikov <ltuikov89@gmail.com>
In Xe, the new Intel GPU driver, a choice has made to have a 1 to 1
mapping between a drm_gpu_scheduler and drm_sched_entity. At first this
seems a bit odd but let us explain the reasoning below.
1. In Xe the submission order from multiple drm_sched_entity is not
guaranteed to be the same completion even if targeting the same hardware
engine. This is because in Xe we have a firmware scheduler, the GuC,
which allowed to reorder, timeslice, and preempt submissions. If a using
shared drm_gpu_scheduler across multiple drm_sched_entity, the TDR falls
apart as the TDR expects submission order == completion order. Using a
dedicated drm_gpu_scheduler per drm_sched_entity solve this problem.
2. In Xe submissions are done via programming a ring buffer (circular
buffer), a drm_gpu_scheduler provides a limit on number of jobs, if the
limit of number jobs is set to RING_SIZE / MAX_SIZE_PER_JOB we get flow
control on the ring for free.
A problem with this design is currently a drm_gpu_scheduler uses a
kthread for submission / job cleanup. This doesn't scale if a large
number of drm_gpu_scheduler are used. To work around the scaling issue,
use a worker rather than kthread for submission / job cleanup.
v2:
- (Rob Clark) Fix msm build
- Pass in run work queue
v3:
- (Boris) don't have loop in worker
v4:
- (Tvrtko) break out submit ready, stop, start helpers into own patch
v5:
- (Boris) default to ordered work queue
v6:
- (Luben / checkpatch) fix alignment in msm_ringbuffer.c
- (Luben) s/drm_sched_submit_queue/drm_sched_wqueue_enqueue
- (Luben) Update comment for drm_sched_wqueue_enqueue
- (Luben) Positive check for submit_wq in drm_sched_init
- (Luben) s/alloc_submit_wq/own_submit_wq
v7:
- (Luben) s/drm_sched_wqueue_enqueue/drm_sched_run_job_queue
v8:
- (Luben) Adjust var names / comments
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Luben Tuikov <luben.tuikov@amd.com>
Link: https://lore.kernel.org/r/20231031032439.1558703-3-matthew.brost@intel.com
Signed-off-by: Luben Tuikov <ltuikov89@gmail.com>
