radeonsi: always use RADEON_USAGE_DISALLOW_SLOW_REPLY

All places are fine with getting a false negative as long as buffer_wait
returns quickly. This can improve performance.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32877>

src/gallium/drivers/radeonsi/gfx11_query.c

@@ -62,7 +62,8 @@ static bool gfx11_alloc_query_buffer(struct si_context *sctx)
       qbuf = list_first_entry(&sctx->shader_query_buffers, struct gfx11_sh_query_buffer, list);
       if (!qbuf->refcount &&
           !si_cs_is_buffer_referenced(sctx, qbuf->buf->buf, RADEON_USAGE_READWRITE) &&
-          sctx->ws->buffer_wait(sctx->ws, qbuf->buf->buf, 0, RADEON_USAGE_READWRITE)) {
+          sctx->ws->buffer_wait(sctx->ws, qbuf->buf->buf, 0,
+                                RADEON_USAGE_READWRITE | RADEON_USAGE_DISALLOW_SLOW_REPLY)) {
          /* Can immediately re-use the oldest buffer */
          list_del(&qbuf->list);
       } else {

src/gallium/drivers/radeonsi/si_buffer.c

@@ -256,7 +256,8 @@ static bool si_invalidate_buffer(struct si_context *sctx, struct si_resource *bu
 
    /* Check if mapping this buffer would cause waiting for the GPU. */
    if (si_cs_is_buffer_referenced(sctx, buf->buf, RADEON_USAGE_READWRITE) ||
-       !sctx->ws->buffer_wait(sctx->ws, buf->buf, 0, RADEON_USAGE_READWRITE)) {
+       !sctx->ws->buffer_wait(sctx->ws, buf->buf, 0,
+                              RADEON_USAGE_READWRITE | RADEON_USAGE_DISALLOW_SLOW_REPLY)) {
       /* Reallocate the buffer in the same pipe_resource. */
       si_alloc_resource(sctx->screen, buf);
       si_rebind_buffer(sctx, &buf->b.b);
@@ -399,7 +400,8 @@ static void *si_buffer_transfer_map(struct pipe_context *ctx, struct pipe_resour
       if (buf->flags & (RADEON_FLAG_SPARSE | RADEON_FLAG_NO_CPU_ACCESS) ||
           force_discard_range ||
           si_cs_is_buffer_referenced(sctx, buf->buf, RADEON_USAGE_READWRITE) ||
-          !sctx->ws->buffer_wait(sctx->ws, buf->buf, 0, RADEON_USAGE_READWRITE)) {
+          !sctx->ws->buffer_wait(sctx->ws, buf->buf, 0,
+                                 RADEON_USAGE_READWRITE | RADEON_USAGE_DISALLOW_SLOW_REPLY)) {
          /* Do a wait-free write-only transfer using a temporary buffer. */
          struct u_upload_mgr *uploader;
          struct si_resource *staging = NULL;

src/gallium/drivers/radeonsi/si_pipe.c

@@ -908,7 +908,8 @@ static bool si_is_resource_busy(struct pipe_screen *screen, struct pipe_resource
                            /* If mapping for write, we need to wait for all reads and writes.
                             * If mapping for read, we only need to wait for writes.
                             */
-                           usage & PIPE_MAP_WRITE ? RADEON_USAGE_READWRITE : RADEON_USAGE_WRITE);
+                           (usage & PIPE_MAP_WRITE ? RADEON_USAGE_READWRITE : RADEON_USAGE_WRITE) |
+                           RADEON_USAGE_DISALLOW_SLOW_REPLY);
 }
 
 static struct pipe_context *si_pipe_create_context(struct pipe_screen *screen, void *priv,

src/gallium/drivers/radeonsi/si_query.c

@@ -518,7 +518,8 @@ void si_query_buffer_reset(struct si_context *sctx, struct si_query_buffer *buff
 
    /* Discard even the oldest buffer if it can't be mapped without a stall. */
    if (si_cs_is_buffer_referenced(sctx, buffer->buf->buf, RADEON_USAGE_READWRITE) ||
-       !sctx->ws->buffer_wait(sctx->ws, buffer->buf->buf, 0, RADEON_USAGE_READWRITE)) {
+       !sctx->ws->buffer_wait(sctx->ws, buffer->buf->buf, 0,
+                              RADEON_USAGE_READWRITE | RADEON_USAGE_DISALLOW_SLOW_REPLY)) {
       si_resource_reference(&buffer->buf, NULL);
    }
 }

src/gallium/drivers/radeonsi/si_texture.c

@@ -2064,7 +2064,8 @@ static void *si_texture_transfer_map(struct pipe_context *ctx, struct pipe_resou
             tex->buffer.domains & RADEON_DOMAIN_VRAM || tex->buffer.flags & RADEON_FLAG_GTT_WC;
       /* Write & linear only: */
       else if (si_cs_is_buffer_referenced(sctx, tex->buffer.buf, RADEON_USAGE_READWRITE) ||
-               !sctx->ws->buffer_wait(sctx->ws, tex->buffer.buf, 0, RADEON_USAGE_READWRITE)) {
+               !sctx->ws->buffer_wait(sctx->ws, tex->buffer.buf, 0,
+                                      RADEON_USAGE_READWRITE | RADEON_USAGE_DISALLOW_SLOW_REPLY)) {
          /* It's busy. */
          if (si_can_invalidate_texture(sctx->screen, tex, usage, box))
             si_texture_invalidate_storage(sctx, tex);