radv: make the GDS/GDS OA buffer objects resident

GDS is used for NGG queries/streamout (GFX10+ only) and the BOs were
only added to the graphics queue because compute doesn't need them.
Though, the kernel emits a GDS switch when a queue submission doesn't
use GDS. That means that submitting jobs on the compute queue without
GDS can reset the state of the graphics queue and lead to GPU hangs.

The only viable solution for now is to make the GDS BOs resident to
avoid resetting the state between queues. This shouldn't introduce
more syncs between queues because GDS BOs are similar for both.

This fixes a GPU hang with Warhammer Chaosbane during loading time and
possibly some spurious random GPU hangs. Note that this GPU hang was
workarounded on the Steam side with RADV_DEBUG=nongg.

Cc: mesa-stable
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19466>
(cherry picked from commit 26c8fedc1bb12fa8f3d6c646308f4b46756d77c7)

diff --git a/.pick_status.json b/.pick_status.json
index 6c28032..ca9b4b7 100644 (file)
--- a/.pick_status.json
+++ b/.pick_status.json
@@ -1282,7 +1282,7 @@
         "description": "radv: make the GDS/GDS OA buffer objects resident",
         "nominated": true,
         "nomination_type": 0,
-        "resolution": 0,
+        "resolution": 1,
         "main_sha": null,
         "because_sha": null
     },

diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
index a8d7fbc..3c52aca 100644 (file)
--- a/src/amd/vulkan/radv_device.c
+++ b/src/amd/vulkan/radv_device.c
@@ -3033,10 +3033,14 @@ radv_queue_state_finish(struct radv_queue_state *queue, struct radeon_winsys *ws
       ws->buffer_destroy(ws, queue->task_rings_bo);
    if (queue->attr_ring_bo)
       ws->buffer_destroy(ws, queue->attr_ring_bo);
-   if (queue->gds_bo)
+   if (queue->gds_bo) {
+      ws->buffer_make_resident(ws, queue->gds_bo, false);
       ws->buffer_destroy(ws, queue->gds_bo);
-   if (queue->gds_oa_bo)
+   }
+   if (queue->gds_oa_bo) {
+      ws->buffer_make_resident(ws, queue->gds_oa_bo, false);
       ws->buffer_destroy(ws, queue->gds_oa_bo);
+   }
    if (queue->compute_scratch_bo)
       ws->buffer_destroy(ws, queue->compute_scratch_bo);
 }
@@ -4710,6 +4714,13 @@ radv_update_preamble_cs(struct radv_queue_state *queue, struct radv_device *devi
                                  RADV_BO_PRIORITY_SCRATCH, 0, &gds_bo);
       if (result != VK_SUCCESS)
          goto fail;
+
+      /* Add the GDS BO to our global BO list to prevent the kernel to emit a GDS switch and reset
+       * the state when a compute queue is used.
+       */
+      result = device->ws->buffer_make_resident(ws, gds_bo, true);
+      if (result != VK_SUCCESS)
+         goto fail;
    }
 
    if (!queue->ring_info.gds_oa && needs->gds_oa) {
@@ -4719,6 +4730,13 @@ radv_update_preamble_cs(struct radv_queue_state *queue, struct radv_device *devi
                                  RADV_BO_PRIORITY_SCRATCH, 0, &gds_oa_bo);
       if (result != VK_SUCCESS)
          goto fail;
+
+      /* Add the GDS OA BO to our global BO list to prevent the kernel to emit a GDS switch and
+       * reset the state when a compute queue is used.
+       */
+      result = device->ws->buffer_make_resident(ws, gds_oa_bo, true);
+      if (result != VK_SUCCESS)
+         goto fail;
    }
 
    /* Re-initialize the descriptor BO when any ring BOs changed.
@@ -4847,11 +4865,6 @@ radv_update_preamble_cs(struct radv_queue_state *queue, struct radv_device *devi
          break;
       }
 
-      if (gds_bo)
-         radv_cs_add_buffer(ws, cs, gds_bo);
-      if (gds_oa_bo)
-         radv_cs_add_buffer(ws, cs, gds_oa_bo);
-
       if (i < 2) {
          /* The two initial preambles have a cache flush at the beginning. */
          const enum amd_gfx_level gfx_level = device->physical_device->rad_info.gfx_level;
@@ -4946,10 +4959,14 @@ fail:
       ws->buffer_destroy(ws, task_rings_bo);
    if (attr_ring_bo && attr_ring_bo != queue->attr_ring_bo)
       ws->buffer_destroy(ws, attr_ring_bo);
-   if (gds_bo && gds_bo != queue->gds_bo)
+   if (gds_bo && gds_bo != queue->gds_bo) {
+      ws->buffer_make_resident(ws, queue->gds_bo, false);
       ws->buffer_destroy(ws, gds_bo);
-   if (gds_oa_bo && gds_oa_bo != queue->gds_oa_bo)
+   }
+   if (gds_oa_bo && gds_oa_bo != queue->gds_oa_bo) {
+      ws->buffer_make_resident(ws, queue->gds_oa_bo, false);
       ws->buffer_destroy(ws, gds_oa_bo);
+   }
 
    return vk_error(queue, result);
 }

diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
index 61ade92..02aba37 100644 (file)
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
@@ -486,7 +486,7 @@ radv_amdgpu_winsys_bo_create(struct radeon_winsys *_ws, uint64_t size, unsigned
       request.flags |= AMDGPU_GEM_CREATE_CPU_GTT_USWC;
    if (!(flags & RADEON_FLAG_IMPLICIT_SYNC))
       request.flags |= AMDGPU_GEM_CREATE_EXPLICIT_SYNC;
-   if (flags & RADEON_FLAG_NO_INTERPROCESS_SHARING &&
+   if ((initial_domain & RADEON_DOMAIN_VRAM_GTT) && (flags & RADEON_FLAG_NO_INTERPROCESS_SHARING) &&
        ((ws->perftest & RADV_PERFTEST_LOCAL_BOS) || (flags & RADEON_FLAG_PREFER_LOCAL_BO))) {
       bo->base.is_local = true;
       request.flags |= AMDGPU_GEM_CREATE_VM_ALWAYS_VALID;