Re: [RFC] drm/msm: Add UABI to request perfcntr usage

Antonino Maniscalco <antomani103@xxxxxxxxx> · Fri, 13 Dec 2024 18:10:27 +0100

On 12/13/24 5:50 PM, Akhil P Oommen wrote:
On 12/12/2024 9:44 PM, Antonino Maniscalco wrote:
On 12/12/24 4:58 PM, Akhil P Oommen wrote:
On 12/5/2024 10:24 PM, Rob Clark wrote:
From: Rob Clark <robdclark@xxxxxxxxxxxx>

Performance counter usage falls into two categories:

1. Local usage, where the counter configuration, start, and end read
     happen within (locally to) a single SUBMIT.  In this case, there is
     no dependency on counter configuration or values between submits,
and
     in fact counters are normally cleared on context switches, making it
     impossible to rely on cross-submit state.

2. Global usage, where a single privilaged daemon/process is sampling
     counter values across all processes for profiling.

The two categories are mutually exclusive.  While you can have many
processes making local counter usage, you cannot combine global and
local usage without the two stepping on each others feet (by changing
counter configuration).

For global counter usage, there is already a SYSPROF param (since global
counter usage requires disabling counter clearing on context switch).
This patch adds a REQ_CNTRS param to request local counter usage.  If
one or more processes has requested counter usage, then a SYSPROF
request will fail with -EBUSY.  And if SYSPROF is active, then REQ_CNTRS
will fail with -EBUSY, maintaining the mutual exclusivity.

This is purely an advisory interface to help coordinate userspace.
There is no real means of enforcement, but the worst that can happen if
userspace ignores a REQ_CNTRS failure is that you'll get nonsense
profiling data.

Signed-off-by: Rob Clark <robdclark@xxxxxxxxxxxx>
---
kgsl takes a different approach, which involves a lot more UABI for
assigning counters to different processes.  But I think by taking
advantage of the fact that mesa (freedreno+turnip) reconfigure the
counters they need in each SUBMIT, for their respective gl/vk perf-
counter extensions, we can take this simpler approach.

KGSL's approach is preemption and ifpc safe (also whatever HW changes
that will come up in future generations). How will we ensure that here?

I have plans to bring up IFPC support in near future. Also, I brought up
this point during preemption series. But from the responses, I felt that
profiling was not considered a serious usecase. Still I wonder how the
perfcounter extensions work accurately with preemption.

So back then I implemented the postamble IB to clear perf counters and
that gets disabled when sysprof (so global usage) is happening. The
kernel is oblivious to "Local isage" of profiling but in that case
really what we want to do is disable preemption which in my
understanding can be done from userspace with a PKT. In my understanding
this had us covered for all usecases.

I think this wasn't mentioned at that time. Which UMD PKT in a6x+ did
you mean?

Ah, I thought it wasmentioned, sorry.
The packet I was referring to is:
	<doc> Make next dword 1 to disable preemption, 0 to re-enable it. </doc>
	<value name="CP_PREEMPT_DISABLE" value="0x6c" variants="A6XX"/>

BTW you mentioned wanting to look into IFPC. Since I too wanted to look 
into implementing it wonder if you could let me know when you planned on 
working on it.


-Akhil.


So what would you expect instead we should do kernel side to make
profiling preemption safe?


-Akhil


   drivers/gpu/drm/msm/adreno/adreno_gpu.c |  2 +
   drivers/gpu/drm/msm/msm_drv.c           |  5 ++-
   drivers/gpu/drm/msm/msm_gpu.c           |  1 +
   drivers/gpu/drm/msm/msm_gpu.h           | 29 +++++++++++++-
   drivers/gpu/drm/msm/msm_submitqueue.c   | 52 ++++++++++++++++++++++++-
   include/uapi/drm/msm_drm.h              |  1 +
   6 files changed, 85 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.c b/drivers/gpu/
drm/msm/adreno/adreno_gpu.c
index 31bbf2c83de4..f688e37059b8 100644
--- a/drivers/gpu/drm/msm/adreno/adreno_gpu.c
+++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.c
@@ -441,6 +441,8 @@ int adreno_set_param(struct msm_gpu *gpu, struct
msm_file_private *ctx,
           if (!capable(CAP_SYS_ADMIN))
               return UERR(EPERM, drm, "invalid permissions");
           return msm_file_private_set_sysprof(ctx, gpu, value);
+    case MSM_PARAM_REQ_CNTRS:
+        return msm_file_private_request_counters(ctx, gpu, value);
       default:
           return UERR(EINVAL, drm, "%s: invalid param: %u", gpu-
name, param);
       }
diff --git a/drivers/gpu/drm/msm/msm_drv.c b/drivers/gpu/drm/msm/
msm_drv.c
index 6416d2cb4efc..bf8314ff4a25 100644
--- a/drivers/gpu/drm/msm/msm_drv.c
+++ b/drivers/gpu/drm/msm/msm_drv.c
@@ -377,9 +377,12 @@ static void msm_postclose(struct drm_device
*dev, struct drm_file *file)
        * It is not possible to set sysprof param to non-zero if gpu
        * is not initialized:
        */
-    if (priv->gpu)
+    if (ctx->sysprof)
           msm_file_private_set_sysprof(ctx, priv->gpu, 0);
   +    if (ctx->counters_requested)
+        msm_file_private_request_counters(ctx, priv->gpu, 0);
+
       context_close(ctx);
   }
   diff --git a/drivers/gpu/drm/msm/msm_gpu.c b/drivers/gpu/drm/msm/
msm_gpu.c
index 82f204f3bb8f..013b59ca3bb1 100644
--- a/drivers/gpu/drm/msm/msm_gpu.c
+++ b/drivers/gpu/drm/msm/msm_gpu.c
@@ -991,6 +991,7 @@ int msm_gpu_init(struct drm_device *drm, struct
platform_device *pdev,
       gpu->nr_rings = nr_rings;
         refcount_set(&gpu->sysprof_active, 1);
+    refcount_set(&gpu->local_counters_active, 1);
         return 0;
   diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/
msm_gpu.h
index e25009150579..83c61e523b1b 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/drivers/gpu/drm/msm/msm_gpu.h
@@ -195,12 +195,28 @@ struct msm_gpu {
       int nr_rings;
         /**
-     * sysprof_active:
+     * @sysprof_active:
        *
-     * The count of contexts that have enabled system profiling.
+     * The count of contexts that have enabled system profiling plus
one.
+     *
+     * Note: refcount_t does not like 0->1 transitions.. we want to
keep
+     * the under/overflow checks that refcount_t provides, but allow
+     * multiple on/off transitions so we track the logical value
plus one.)
        */
       refcount_t sysprof_active;
   +    /**
+     * @local_counters_active:
+     *
+     * The count of contexts that have requested local (intra-submit)
+     * performance counter usage plus one.
+     *
+     * Note: refcount_t does not like 0->1 transitions.. we want to
keep
+     * the under/overflow checks that refcount_t provides, but allow
+     * multiple on/off transitions so we track the logical value
plus one.)
+     */
+    refcount_t local_counters_active;
+
       /**
        * lock:
        *
@@ -383,6 +399,13 @@ struct msm_file_private {
        */
       int sysprof;
   +    /**
+     * @counters_requested:
+     *
+     * Has the context requested local perfcntr usage.
+     */
+    bool counters_requested;
+
       /**
        * comm: Overridden task comm, see MSM_PARAM_COMM
        *
@@ -626,6 +649,8 @@ void msm_submitqueue_destroy(struct kref *kref);
     int msm_file_private_set_sysprof(struct msm_file_private *ctx,
                    struct msm_gpu *gpu, int sysprof);
+int msm_file_private_request_counters(struct msm_file_private *ctx,
+                      struct msm_gpu *gpu, int reqcntrs);
   void __msm_file_private_destroy(struct kref *kref);
     static inline void msm_file_private_put(struct msm_file_private
*ctx)
diff --git a/drivers/gpu/drm/msm/msm_submitqueue.c b/drivers/gpu/drm/
msm/msm_submitqueue.c
index 7fed1de63b5d..1e1e21e6f7ae 100644
--- a/drivers/gpu/drm/msm/msm_submitqueue.c
+++ b/drivers/gpu/drm/msm/msm_submitqueue.c
@@ -10,6 +10,15 @@
   int msm_file_private_set_sysprof(struct msm_file_private *ctx,
                    struct msm_gpu *gpu, int sysprof)
   {
+    int ret = 0;
+
+    mutex_lock(&gpu->lock);
+
+    if (sysprof && (refcount_read(&gpu->local_counters_active) > 1)) {
+        ret = UERR(EBUSY, gpu->dev, "Local counter usage active");
+        goto out_unlock;
+    }
+
       /*
        * Since pm_runtime and sysprof_active are both refcounts, we
        * call apply the new value first, and then unwind the previous
@@ -18,7 +27,8 @@ int msm_file_private_set_sysprof(struct
msm_file_private *ctx,
         switch (sysprof) {
       default:
-        return UERR(EINVAL, gpu->dev, "Invalid sysprof: %d", sysprof);
+        ret = UERR(EINVAL, gpu->dev, "Invalid sysprof: %d", sysprof);
+        goto out_unlock;
       case 2:
           pm_runtime_get_sync(&gpu->pdev->dev);
           fallthrough;
@@ -43,7 +53,45 @@ int msm_file_private_set_sysprof(struct
msm_file_private *ctx,
         ctx->sysprof = sysprof;
   -    return 0;
+out_unlock:
+    mutex_unlock(&gpu->lock);
+
+    return ret;
+}
+
+int msm_file_private_request_counters(struct msm_file_private *ctx,
+                      struct msm_gpu *gpu, int reqctrs)
+{
+    int ret = 0;
+
+    mutex_lock(&gpu->lock);
+
+    if (reqctrs && (refcount_read(&gpu->sysprof_active) > 1)) {
+        ret = UERR(EBUSY, gpu->dev, "System profiling active");
+        goto out_unlock;
+    }
+
+    if (reqctrs) {
+        if (ctx->counters_requested) {
+            ret = UERR(EINVAL, gpu->dev, "Already requested");
+            goto out_unlock;
+        }
+
+        ctx->counters_requested = true;
+        refcount_inc(&gpu->local_counters_active);
+    } else {
+        if (!ctx->counters_requested) {
+            ret = UERR(EINVAL, gpu->dev, "Not requested");
+            goto out_unlock;
+        }
+        refcount_dec(&gpu->local_counters_active);
+        ctx->counters_requested = false;
+    }
+
+out_unlock:
+    mutex_unlock(&gpu->lock);
+
+    return ret;
   }
     void __msm_file_private_destroy(struct kref *kref)
diff --git a/include/uapi/drm/msm_drm.h b/include/uapi/drm/msm_drm.h
index 2342cb90857e..ae7fb355e4a1 100644
--- a/include/uapi/drm/msm_drm.h
+++ b/include/uapi/drm/msm_drm.h
@@ -91,6 +91,7 @@ struct drm_msm_timespec {
   #define MSM_PARAM_UBWC_SWIZZLE 0x12 /* RO */
   #define MSM_PARAM_MACROTILE_MODE 0x13 /* RO */
   #define MSM_PARAM_UCHE_TRAP_BASE 0x14 /* RO */
+#define MSM_PARAM_REQ_CNTRS  0x15 /* WO: request "local" (intra-
submit) perfcntr usage  */
     /* For backwards compat.  The original support for preemption was
based on
    * a single ring per priority level so # of priority levels equals
the #



Best regards,



Best regards,
--
Antonino Maniscalco <antomani103@xxxxxxxxx>






