amdgpu,radeon: add full_reset_only param to ctx_query_reset_status

Using this boolean the caller tells if it wants to ignore resets
fixed by a soft recovery.

When true, amdgpu can skip the call to libdrm if no cs has been
rejected (since only full gpu reset cause cs rejections).

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10179>

src/gallium/drivers/r600/r600_pipe_common.c

@@ -487,7 +487,7 @@ static enum pipe_reset_status r600_get_reset_status(struct pipe_context *ctx)
 {
 	struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 
-	return rctx->ws->ctx_query_reset_status(rctx->ctx, NULL);
+	return rctx->ws->ctx_query_reset_status(rctx->ctx, false, NULL);
 }
 
 static void r600_set_debug_callback(struct pipe_context *ctx,

src/gallium/drivers/radeon/radeon_winsys.h

@@ -492,6 +492,7 @@ struct radeon_winsys {
     * Query a GPU reset status.
     */
    enum pipe_reset_status (*ctx_query_reset_status)(struct radeon_winsys_ctx *ctx,
+                                                    bool full_reset_only,
                                                     bool *needs_reset);
 
    /**

src/gallium/drivers/radeonsi/si_pipe.c

@@ -355,7 +355,7 @@ static enum pipe_reset_status si_get_reset_status(struct pipe_context *ctx)
       return PIPE_NO_RESET;
 
    bool needs_reset;
-   enum pipe_reset_status status = sctx->ws->ctx_query_reset_status(sctx->ctx, &needs_reset);
+   enum pipe_reset_status status = sctx->ws->ctx_query_reset_status(sctx->ctx, false, &needs_reset);
 
    if (status != PIPE_NO_RESET && needs_reset && !(sctx->context_flags & SI_CONTEXT_FLAG_AUX)) {
       /* Call the gallium frontend to set a no-op API dispatch. */
@@ -744,12 +744,11 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen, unsign
 
       /* Check if the aux_context needs to be recreated */
       struct si_context *saux = (struct si_context *)sscreen->aux_context;
-      bool needs_reset;
 
       simple_mtx_lock(&sscreen->aux_context_lock);
       enum pipe_reset_status status = sctx->ws->ctx_query_reset_status(
-         saux->ctx, &needs_reset);
+         saux->ctx, true, NULL);
-      if (status != PIPE_NO_RESET && needs_reset) {
+      if (status != PIPE_NO_RESET) {
          /* We lost the aux_context, create a new one */
          struct u_log_context *aux_log = (saux)->log;
          sscreen->aux_context->set_log_context(sscreen->aux_context, NULL);

src/gallium/winsys/amdgpu/drm/amdgpu_cs.c

@@ -334,7 +334,8 @@ static void amdgpu_ctx_destroy(struct radeon_winsys_ctx *rwctx)
 }
 
 static enum pipe_reset_status
-amdgpu_ctx_query_reset_status(struct radeon_winsys_ctx *rwctx, bool *needs_reset)
+amdgpu_ctx_query_reset_status(struct radeon_winsys_ctx *rwctx, bool full_reset_only,
+                              bool *needs_reset)
 {
    struct amdgpu_ctx *ctx = (struct amdgpu_ctx*)rwctx;
    int r;
@@ -346,6 +347,14 @@ amdgpu_ctx_query_reset_status(struct radeon_winsys_ctx *rwctx, bool *needs_reset
    if (ctx->ws->info.drm_minor >= 24) {
       uint64_t flags;
 
+      if (full_reset_only &&
+          ctx->initial_num_total_rejected_cs == ctx->ws->num_total_rejected_cs) {
+         /* If the caller is only interested in full reset (= wants to ignore soft
+          * recoveries), we can use the rejected cs count as a quick first check.
+          */
+         return PIPE_NO_RESET;
+      }
+
       r = amdgpu_cs_query_reset_state2(ctx->ctx, &flags);
       if (r) {
          fprintf(stderr, "amdgpu: amdgpu_cs_query_reset_state failed. (%i)\n", r);

src/gallium/winsys/radeon/drm/radeon_drm_cs.c

@@ -87,7 +87,8 @@ static void radeon_drm_ctx_destroy(struct radeon_winsys_ctx *ctx)
 }
 
 static enum pipe_reset_status
-radeon_drm_ctx_query_reset_status(struct radeon_winsys_ctx *rctx, bool *needs_reset)
+radeon_drm_ctx_query_reset_status(struct radeon_winsys_ctx *rctx, bool full_reset_only,
+                                  bool *needs_reset)
 {
    struct radeon_ctx *ctx = (struct radeon_ctx*)rctx;