drm/amdgpu: rework context priority handling
authorNirmoy Das <nirmoy.das@amd.com>
Tue, 24 Aug 2021 18:39:18 +0000 (20:39 +0200)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 1 Sep 2021 20:55:11 +0000 (16:55 -0400)
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>
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.h
drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c

index e7a010b..c88c5c6 100644 (file)
@@ -43,14 +43,61 @@ const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM] = {
        [AMDGPU_HW_IP_VCN_JPEG] =       1,
 };
 
+bool amdgpu_ctx_priority_is_valid(int32_t ctx_prio)
+{
+       switch (ctx_prio) {
+       case AMDGPU_CTX_PRIORITY_UNSET:
+       case AMDGPU_CTX_PRIORITY_VERY_LOW:
+       case AMDGPU_CTX_PRIORITY_LOW:
+       case AMDGPU_CTX_PRIORITY_NORMAL:
+       case AMDGPU_CTX_PRIORITY_HIGH:
+       case AMDGPU_CTX_PRIORITY_VERY_HIGH:
+               return true;
+       default:
+               return false;
+       }
+}
+
+static enum drm_sched_priority
+amdgpu_ctx_to_drm_sched_prio(int32_t ctx_prio)
+{
+       switch (ctx_prio) {
+       case AMDGPU_CTX_PRIORITY_UNSET:
+               return DRM_SCHED_PRIORITY_UNSET;
+
+       case AMDGPU_CTX_PRIORITY_VERY_LOW:
+               return DRM_SCHED_PRIORITY_MIN;
+
+       case AMDGPU_CTX_PRIORITY_LOW:
+               return DRM_SCHED_PRIORITY_MIN;
+
+       case AMDGPU_CTX_PRIORITY_NORMAL:
+               return DRM_SCHED_PRIORITY_NORMAL;
+
+       case AMDGPU_CTX_PRIORITY_HIGH:
+               return DRM_SCHED_PRIORITY_HIGH;
+
+       case AMDGPU_CTX_PRIORITY_VERY_HIGH:
+               return DRM_SCHED_PRIORITY_HIGH;
+
+       /* This should not happen as we sanitized userspace provided priority
+        * already, WARN if this happens.
+        */
+       default:
+               WARN(1, "Invalid context priority %d\n", ctx_prio);
+               return DRM_SCHED_PRIORITY_NORMAL;
+       }
+
+}
+
 static int amdgpu_ctx_priority_permit(struct drm_file *filp,
-                                     enum drm_sched_priority priority)
+                                     int32_t priority)
 {
-       if (priority < 0 || priority >= DRM_SCHED_PRIORITY_COUNT)
+       if (!amdgpu_ctx_priority_is_valid(priority))
                return -EINVAL;
 
        /* NORMAL and below are accessible by everyone */
-       if (priority <= DRM_SCHED_PRIORITY_NORMAL)
+       if (priority <= AMDGPU_CTX_PRIORITY_NORMAL)
                return 0;
 
        if (capable(CAP_SYS_NICE))
@@ -62,26 +109,35 @@ static int amdgpu_ctx_priority_permit(struct drm_file *filp,
        return -EACCES;
 }
 
-static enum gfx_pipe_priority amdgpu_ctx_sched_prio_to_compute_prio(enum drm_sched_priority prio)
+static enum gfx_pipe_priority amdgpu_ctx_prio_to_compute_prio(int32_t prio)
 {
        switch (prio) {
-       case DRM_SCHED_PRIORITY_HIGH:
-       case DRM_SCHED_PRIORITY_KERNEL:
+       case AMDGPU_CTX_PRIORITY_HIGH:
+       case AMDGPU_CTX_PRIORITY_VERY_HIGH:
                return AMDGPU_GFX_PIPE_PRIO_HIGH;
        default:
                return AMDGPU_GFX_PIPE_PRIO_NORMAL;
        }
 }
 
-static unsigned int amdgpu_ctx_prio_sched_to_hw(struct amdgpu_device *adev,
-                                                enum drm_sched_priority prio,
-                                                u32 hw_ip)
+static unsigned int amdgpu_ctx_get_hw_prio(struct amdgpu_ctx *ctx, u32 hw_ip)
 {
+       struct amdgpu_device *adev = ctx->adev;
+       int32_t ctx_prio;
        unsigned int hw_prio;
 
-       hw_prio = (hw_ip == AMDGPU_HW_IP_COMPUTE) ?
-                       amdgpu_ctx_sched_prio_to_compute_prio(prio) :
-                       AMDGPU_RING_PRIO_DEFAULT;
+       ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
+                       ctx->init_priority : ctx->override_priority;
+
+       switch (hw_ip) {
+       case AMDGPU_HW_IP_COMPUTE:
+               hw_prio = amdgpu_ctx_prio_to_compute_prio(ctx_prio);
+               break;
+       default:
+               hw_prio = AMDGPU_RING_PRIO_DEFAULT;
+               break;
+       }
+
        hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM);
        if (adev->gpu_sched[hw_ip][hw_prio].num_scheds == 0)
                hw_prio = AMDGPU_RING_PRIO_DEFAULT;
@@ -89,15 +145,17 @@ static unsigned int amdgpu_ctx_prio_sched_to_hw(struct amdgpu_device *adev,
        return hw_prio;
 }
 
+
 static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
-                                  const u32 ring)
+                                 const u32 ring)
 {
        struct amdgpu_device *adev = ctx->adev;
        struct amdgpu_ctx_entity *entity;
        struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
        unsigned num_scheds = 0;
+       int32_t ctx_prio;
        unsigned int hw_prio;
-       enum drm_sched_priority priority;
+       enum drm_sched_priority drm_prio;
        int r;
 
        entity = kzalloc(struct_size(entity, fences, amdgpu_sched_jobs),
@@ -105,10 +163,11 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
        if (!entity)
                return  -ENOMEM;
 
+       ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
+                       ctx->init_priority : ctx->override_priority;
        entity->sequence = 1;
-       priority = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
-                               ctx->init_priority : ctx->override_priority;
-       hw_prio = amdgpu_ctx_prio_sched_to_hw(adev, priority, hw_ip);
+       hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip);
+       drm_prio = amdgpu_ctx_to_drm_sched_prio(ctx_prio);
 
        hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM);
        scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
@@ -124,7 +183,7 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
                num_scheds = 1;
        }
 
-       r = drm_sched_entity_init(&entity->entity, priority, scheds, num_scheds,
+       r = drm_sched_entity_init(&entity->entity, drm_prio, scheds, num_scheds,
                                  &ctx->guilty);
        if (r)
                goto error_free_entity;
@@ -139,7 +198,7 @@ error_free_entity:
 }
 
 static int amdgpu_ctx_init(struct amdgpu_device *adev,
-                          enum drm_sched_priority priority,
+                          int32_t priority,
                           struct drm_file *filp,
                           struct amdgpu_ctx *ctx)
 {
@@ -161,7 +220,7 @@ static int amdgpu_ctx_init(struct amdgpu_device *adev,
        ctx->reset_counter_query = ctx->reset_counter;
        ctx->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
        ctx->init_priority = priority;
-       ctx->override_priority = DRM_SCHED_PRIORITY_UNSET;
+       ctx->override_priority = AMDGPU_CTX_PRIORITY_UNSET;
 
        return 0;
 }
@@ -234,7 +293,7 @@ int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
 static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
                            struct amdgpu_fpriv *fpriv,
                            struct drm_file *filp,
-                           enum drm_sched_priority priority,
+                           int32_t priority,
                            uint32_t *id)
 {
        struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
@@ -397,19 +456,19 @@ int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
 {
        int r;
        uint32_t id;
-       enum drm_sched_priority priority;
+       int32_t priority;
 
        union drm_amdgpu_ctx *args = data;
        struct amdgpu_device *adev = drm_to_adev(dev);
        struct amdgpu_fpriv *fpriv = filp->driver_priv;
 
        id = args->in.ctx_id;
-       r = amdgpu_to_sched_priority(args->in.priority, &priority);
+       priority = args->in.priority;
 
        /* For backwards compatibility reasons, we need to accept
         * ioctls with garbage in the priority field */
-       if (r == -EINVAL)
-               priority = DRM_SCHED_PRIORITY_NORMAL;
+       if (!amdgpu_ctx_priority_is_valid(priority))
+               priority = AMDGPU_CTX_PRIORITY_NORMAL;
 
        switch (args->in.op) {
        case AMDGPU_CTX_OP_ALLOC_CTX:
@@ -515,9 +574,9 @@ struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
 }
 
 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
-                                           struct amdgpu_ctx_entity *aentity,
-                                           int hw_ip,
-                                           enum drm_sched_priority priority)
+                                          struct amdgpu_ctx_entity *aentity,
+                                          int hw_ip,
+                                          int32_t priority)
 {
        struct amdgpu_device *adev = ctx->adev;
        unsigned int hw_prio;
@@ -525,12 +584,12 @@ static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
        unsigned num_scheds;
 
        /* set sw priority */
-       drm_sched_entity_set_priority(&aentity->entity, priority);
+       drm_sched_entity_set_priority(&aentity->entity,
+                                     amdgpu_ctx_to_drm_sched_prio(priority));
 
        /* set hw priority */
        if (hw_ip == AMDGPU_HW_IP_COMPUTE) {
-               hw_prio = amdgpu_ctx_prio_sched_to_hw(adev, priority,
-                                                     AMDGPU_HW_IP_COMPUTE);
+               hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip);
                hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX);
                scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
                num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
@@ -540,14 +599,14 @@ static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
 }
 
 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
-                                 enum drm_sched_priority priority)
+                                 int32_t priority)
 {
-       enum drm_sched_priority ctx_prio;
+       int32_t ctx_prio;
        unsigned i, j;
 
        ctx->override_priority = priority;
 
-       ctx_prio = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
+       ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
                        ctx->init_priority : ctx->override_priority;
        for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
                for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
index 14db16b..a44b8b8 100644 (file)
@@ -47,8 +47,8 @@ struct amdgpu_ctx {
        spinlock_t                      ring_lock;
        struct amdgpu_ctx_entity        *entities[AMDGPU_HW_IP_NUM][AMDGPU_MAX_ENTITY_NUM];
        bool                            preamble_presented;
-       enum drm_sched_priority         init_priority;
-       enum drm_sched_priority         override_priority;
+       int32_t                         init_priority;
+       int32_t                         override_priority;
        struct mutex                    lock;
        atomic_t                        guilty;
        unsigned long                   ras_counter_ce;
@@ -75,8 +75,8 @@ void amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
                                       struct drm_sched_entity *entity,
                                       uint64_t seq);
-void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
-                                 enum drm_sched_priority priority);
+bool amdgpu_ctx_priority_is_valid(int32_t ctx_prio);
+void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx, int32_t ctx_prio);
 
 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
                     struct drm_file *filp);
index b7d861e..e9b4508 100644 (file)
 #include "amdgpu_sched.h"
 #include "amdgpu_vm.h"
 
-int amdgpu_to_sched_priority(int amdgpu_priority,
-                            enum drm_sched_priority *prio)
-{
-       switch (amdgpu_priority) {
-       case AMDGPU_CTX_PRIORITY_VERY_HIGH:
-               *prio = DRM_SCHED_PRIORITY_HIGH;
-               break;
-       case AMDGPU_CTX_PRIORITY_HIGH:
-               *prio = DRM_SCHED_PRIORITY_HIGH;
-               break;
-       case AMDGPU_CTX_PRIORITY_NORMAL:
-               *prio = DRM_SCHED_PRIORITY_NORMAL;
-               break;
-       case AMDGPU_CTX_PRIORITY_LOW:
-       case AMDGPU_CTX_PRIORITY_VERY_LOW:
-               *prio = DRM_SCHED_PRIORITY_MIN;
-               break;
-       case AMDGPU_CTX_PRIORITY_UNSET:
-               *prio = DRM_SCHED_PRIORITY_UNSET;
-               break;
-       default:
-               WARN(1, "Invalid context priority %d\n", amdgpu_priority);
-               return -EINVAL;
-       }
-
-       return 0;
-}
-
 static int amdgpu_sched_process_priority_override(struct amdgpu_device *adev,
                                                  int fd,
-                                                 enum drm_sched_priority priority)
+                                                 int32_t priority)
 {
        struct fd f = fdget(fd);
        struct amdgpu_fpriv *fpriv;
@@ -89,7 +61,7 @@ static int amdgpu_sched_process_priority_override(struct amdgpu_device *adev,
 static int amdgpu_sched_context_priority_override(struct amdgpu_device *adev,
                                                  int fd,
                                                  unsigned ctx_id,
-                                                 enum drm_sched_priority priority)
+                                                 int32_t priority)
 {
        struct fd f = fdget(fd);
        struct amdgpu_fpriv *fpriv;
@@ -124,7 +96,6 @@ int amdgpu_sched_ioctl(struct drm_device *dev, void *data,
 {
        union drm_amdgpu_sched *args = data;
        struct amdgpu_device *adev = drm_to_adev(dev);
-       enum drm_sched_priority priority;
        int r;
 
        /* First check the op, then the op's argument.
@@ -138,21 +109,22 @@ int amdgpu_sched_ioctl(struct drm_device *dev, void *data,
                return -EINVAL;
        }
 
-       r = amdgpu_to_sched_priority(args->in.priority, &priority);
-       if (r)
-               return r;
+       if (!amdgpu_ctx_priority_is_valid(args->in.priority)) {
+               WARN(1, "Invalid context priority %d\n", args->in.priority);
+               return -EINVAL;
+       }
 
        switch (args->in.op) {
        case AMDGPU_SCHED_OP_PROCESS_PRIORITY_OVERRIDE:
                r = amdgpu_sched_process_priority_override(adev,
                                                           args->in.fd,
-                                                          priority);
+                                                          args->in.priority);
                break;
        case AMDGPU_SCHED_OP_CONTEXT_PRIORITY_OVERRIDE:
                r = amdgpu_sched_context_priority_override(adev,
                                                           args->in.fd,
                                                           args->in.ctx_id,
-                                                          priority);
+                                                          args->in.priority);
                break;
        default:
                /* Impossible.