drm/amdkfd: Force perfmon clock on for profiler PMC requests

Counters and SQ/PC sampling read 0 at DPM AUTO on gfx11/gfx12 because
RLC_CGTT_MGCG_OVERRIDE.PERFMON_CLOCK_STATE is gated, and the existing
workaround (STABLE_STD) needs root and pins clocks.

Add a refcounted amdgpu_gfx_perfmon_clk_override() that forces the
perfmon clock on per XCC, keeps GFXOFF disallowed while held, and
restores the previous state on last release, resume and reset.
perfmon_clk_mutex serializes it with the GFX clock gating sequences
and the soc21/soc24 stable pstate hook.

Use it from KFD_IOC_PROFILER_PMC, after binding the process to the GPU
and under reset_domain->sem:
- lock=1 (device mode, CAP_PERFMON): held until unlock or exit.
- lock=0, perfcount_enable=1 (dispatch mode): no CAP_PERFMON needed,
  held until lock=0, perfcount_enable=0 or process exit.

No DPM change, no new ioctl, no struct change. Bump
KFD_IOC_PROFILER_VERSION_NUM to 2.

v2:
- Take GFXOFF outside perfmon_clk_mutex to break a pm.mutex lock cycle
- Replace the racy defer() with amdgpu_gfx_perfmon_clk_update(), and hold
  perfmon_clk_mutex across the GFX clock gating sequences
- Bind the process to the GPU before forcing the perfmon clock on
- Take reset_domain->sem before profiler_lock
- Release a dispatch-mode hold with lock=0, perfcount_enable=0

v3: Add lockdep_assert_held() in amdgpu_gfx_perfmon_clk_program().

Proposed userspace:
https://github.com/ROCm/rocm-systems/pull/14521

Signed-off-by: Apurv Mishra <Apurv.Mishra@amd.com>
Tested-by: Jason Tang <Jason.Tang@amd.com>
Reviewed-by: James Zhu <James.Zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

1 files changed, 10 insertions, 1 deletions

diff --git a/include/uapi/linux/kfd_ioctl.h b/include/uapi/linux/kfd_ioctl.h
index 564faefe6978..3d608d6b0ab8 100644
--- a/include/uapi/linux/kfd_ioctl.h
+++ b/include/uapi/linux/kfd_ioctl.h
@@ -1558,7 +1558,16 @@ struct kfd_ioctl_dbg_trap_args {
 	};
 };
 
-#define KFD_IOC_PROFILER_VERSION_NUM 1
+/*
+ * KFD_IOC_PROFILER_VERSION_NUM history:
+ * - 1 - initial version
+ * - 2 - KFD_IOC_PROFILER_PMC lock also forces the perfmon clock on until the
+ *       lock is released or the process exits. lock == 0 with
+ *       perfcount_enable == 1 forces it on without CAP_PERFMON and without
+ *       locking, until lock == 0 with perfcount_enable == 0 or the process
+ *       exits
+ */
+#define KFD_IOC_PROFILER_VERSION_NUM 2
 enum kfd_profiler_ops {
 	KFD_IOC_PROFILER_PMC = 0,
 	KFD_IOC_PROFILER_VERSION = 2,