idr_init_base(), implemented by commit 6ce711f275 ("idr: Make 1-based
IDRs more efficient"), let us set an arbitrary base other than
idr_init(), which uses base 0.
Since, for this IDR, no ID < 1 is ever requested, using
idr_init_base(&idr, 1) avoids unnecessary tree walks.
Signed-off-by: Danilo Krummrich <dakr@redhat.com>
Signed-off-by: Maxime Ripard <maxime@cerno.tech>
Link: https://lore.kernel.org/r/20220701185303.284082-2-dakr@redhat.com
Save the original stable pstate on ctx init and restore
it on ctx fini so that we restore a manually selected
stable pstate on ctx exit.
v2: fix init order (Alex)
v3: don't add new variable to ctx struct (Evan)
Fixes: c65b364c52 ("drm/amdgpu/ctx: only reset stable pstate if the user changed it (v2)")
Reviewed-by: Evan Quan <evan.quan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Starting from SIENNA CICHLID asic supports two gfx pipes, enabling
two graphics queues, 1 on each pipe, pipe0 queue0 would be the normal
piority queue and pipe1 queue0 would be the high priority queue
Only one queue per pipe is visble to SPI, SPI looks at the priority
value assigned to CP_GFX_HQD_QUEUE_PRIORITY from each of the queue's
HQD/MQD.
Create contexts applying AMDGPU_CTX_PRIORITY_HIGH which submits job
to the high priority queue on GFX pipe1. There would be starvation
of LP workload if HP workload is always available.
v2:
- remove unnecessary check(Nirmoy)
- make pipe1 hardware support a separate patch(Nirmoy)
- remove duplicate code(Shashank)
- add CSA support for second gfx pipe(Alex)
v3(Christian):
- fix incorrect indentation
- merge COMPUTE and GFX switch cases as both calls the same function.
v4:
- rebase w/ latest code base
Signed-off-by: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Convert fdinfo format to one documented in drm-usage-stats.rst.
It turned out that the existing implementation was actually completely
nonsense. The calculated percentages indeed represented the usage of the
engine, but with varying time slices.
So 10% usage for application A could mean something completely different
than 10% usage for application B.
Completely nuke that and just use the now standardized nanosecond
interface.
v2: drop the documentation change for now, nuke percentage calculation
v3: only account for each hw_ip, move the time_spend to the ctx mgr.
v4: move general ctx changes into separate patch, rework the fdinfo to
ctx_mgr interface so that all usages are calculated at once, drop
some unecessary and dangerous refcount dance.
v5: add one more comment how we calculate the time spend
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Shashank Sharma <shashank.sharma@amd.com>
Cc: Daniel Vetter <daniel@ffwll.ch>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Let each context have a pointer to the ctx manager and properly
initialize the adev pointer inside the context manager.
Reduce the BUG_ON() in amdgpu_ctx_add_fence() into a WARN_ON() and
directly return the sequence number instead of writing into a parmeter.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Shashank Sharma <shashank.sharma@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Check if the requested stable pstate matches the current one before
changing it. This avoids changing the stable pstate on context
destroy if the user never changed it in the first place via the
IOCTL.
v2: compare the current and requested rather than setting a flag (Lijo)
Fixes: 8cda7a4f96 ("drm/amdgpu/UAPI: add new CTX OP to get/set stable pstates")
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Since we removed the context lock we need to make sure that not two threads
are trying to install an entity at the same time.
Signed-off-by: Christian König <christian.koenig@amd.com>
Fixes: 461fa7b0ac ("drm/amdgpu: remove ctx->lock")
Reviewed-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Clang static analysis reports this problem
amdgpu_ctx.c:616:26: warning: Assigned value is garbage
or undefined
args->out.pstate.flags = stable_pstate;
^ ~~~~~~~~~~~~~
amdgpu_ctx_stable_pstate can fail without setting
stable_pstate. So check.
Fixes: 8cda7a4f96 ("drm/amdgpu/UAPI: add new CTX OP to get/set stable pstates")
Signed-off-by: Tom Rix <trix@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
KMD reports a warning on holding a lock from drm_syncobj_find_fence,
when running amdgpu_test case “syncobj timeline test”.
ctx->lock was designed to prevent concurrent "amdgpu_ctx_wait_prev_fence"
calls and avoid dead reservation lock from GPU reset. since no reservation
lock is held in latest GPU reset any more, ctx->lock can be simply removed
and concurrent "amdgpu_ctx_wait_prev_fence" call also can be prevented by
PD root bo reservation lock.
call stacks:
=================
//hold lock
amdgpu_cs_ioctl->amdgpu_cs_parser_init->mutex_lock(&parser->ctx->lock);
…
//report warning
amdgpu_cs_dependencies->amdgpu_cs_process_syncobj_timeline_in_dep \
->amdgpu_syncobj_lookup_and_add_to_sync -> drm_syncobj_find_fence \
-> lockdep_assert_none_held_once
…
amdgpu_cs_ioctl->amdgpu_cs_parser_fini->mutex_unlock(&parser->ctx->lock);
Signed-off-by: Ken Xue <Ken.Xue@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add a new CTX ioctl operation to set stable pstates for profiling.
When creating traces for tools like RGP or using SPM or doing
performance profiling, it's required to enable a special
stable profiling power state on the GPU. These profiling
states set fixed clocks and disable certain other power
features like powergating which may impact the results.
Historically, these profiling pstates were enabled via sysfs,
but this adds an interface to enable it via the CTX ioctl
from the application. Since the power state is global
only one application can set it at a time, so if multiple
applications try and use it only the first will get it,
the ioctl will return -EBUSY for others. The sysfs interface
will override whatever has been set by this interface.
Mesa MR: https://gitlab.freedesktop.org/mesa/drm/-/merge_requests/207
v2: don't default r = 0;
v3: rebase on Evan's PM cleanup
Reviewed-by: Evan Quan <evan.quan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Currently AMDGPU_RING_PRIO_MAX is redefinition of a
max gfx hwip priority, this won't work well when we will
have a hwip with different set of priorities than gfx.
Also, HW ring priorities are different from ring priorities.
Create a global enum for ring priority levels which each
HWIP can use to define its own priority levels.
Signed-off-by: Nirmoy Das <nirmoy.das@amd.com>
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
To get a hardware queue priority for a context, we are currently
mapping AMDGPU_CTX_PRIORITY_* to DRM_SCHED_PRIORITY_* and then
to hardware queue priority, which is not the right way to do that
as DRM_SCHED_PRIORITY_* is software scheduler's priority and it is
independent from a hardware queue priority.
Use userspace provided context priority, AMDGPU_CTX_PRIORITY_* to
map a context to proper hardware queue priority.
Signed-off-by: Nirmoy Das <nirmoy.das@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
On Context Query2 IOCTL return the correctable and
uncorrectable errors in O(1) fashion, from cached
values, and schedule a delayed work function to
calculate and cache them for the next such IOCTL.
v2: Cancel pending delayed work at ras_fini().
v3: Remove conditionals when dealing with delayed
work manipulation as they're inherently racy.
Cc: Alexander Deucher <Alexander.Deucher@amd.com>
Cc: Christian König <christian.koenig@amd.com>
Cc: John Clements <john.clements@amd.com>
Cc: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Luben Tuikov <luben.tuikov@amd.com>
Reviewed-by: Alexander Deucher <Alexander.Deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
On QUERY2 IOCTL don't query counts of correctable
and uncorrectable errors, since when RAS is
enabled and supported on Vega20 server boards,
this takes insurmountably long time, in O(n^3),
which slows the system down to the point of it
being unusable when we have GUI up.
Fixes: ae363a212b ("drm/amdgpu: Add a new flag to AMDGPU_CTX_OP_QUERY_STATE2")
Cc: Alexander Deucher <Alexander.Deucher@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Luben Tuikov <luben.tuikov@amd.com>
Reviewed-by: Alexander Deucher <Alexander.Deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The proper metric for fence utilization over several
contexts is an harmonic mean, but such calculation is
prohibitive in kernel space, so the code approximates it.
Because the approximation diverges when one context has a
very small ratio compared with the other context, this change
filter out ratios smaller that 0.01%
v2: make the fence calculation static and initialize variables
within that function
v3: Fix warnings (Alex)
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: David M Nieto <david.nieto@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210513174539.27409-2-david.nieto@amd.com
Christian König
Dave Airlie
Guchun Chen
Thomas Zimmermann
Danilo Krummrich
Alex Deucher
Arunpravin Paneer Selvam
Weiguo Li
Andrey Grodzovsky
Tom Rix
Ken Xue
Satyajit Sahu
Nirmoy Das
Luben Tuikov
David M Nieto