[Mesa-dev] [PATCH 09/26] winsys/amdpgu: retry after ENOMEM to work around a GDS memory management bug

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

If GDS/GWS/OA resources are being used by IBs that are currently busy,
the kernel driver returns -ENOMEM instead of waiting until those resources
are idle.
---
 src/gallium/winsys/amdgpu/drm/amdgpu_cs.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
index 1438b1ffe76..edb90c1a734 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
@@ -1596,8 +1596,13 @@ void amdgpu_cs_submit_ib(void *job, int thread_index)
          chunks[num_chunks].chunk_data = (uintptr_t)&cs->ib[IB_PARALLEL_COMPUTE];
          num_chunks++;
 
-         r = amdgpu_cs_submit_raw2(ws->dev, acs->ctx->ctx, bo_list,
-                                   num_chunks, chunks, NULL);
+         /* The memory manager can return -ENOMEM for GDS when all GDS resources
+          * are busy. The workaround is to wait until they are idle.
+          */
+         while ((r = amdgpu_cs_submit_raw2(ws->dev, acs->ctx->ctx, bo_list,
+                                           num_chunks, chunks, NULL)) == -ENOMEM)
+            usleep(2000);
+
          if (r)
             goto finalize;
 
@@ -1643,8 +1648,12 @@ void amdgpu_cs_submit_ib(void *job, int thread_index)
 
       assert(num_chunks <= ARRAY_SIZE(chunks));
 
-      r = amdgpu_cs_submit_raw2(ws->dev, acs->ctx->ctx, bo_list,
-                                num_chunks, chunks, &seq_no);
+      /* The memory manager can return -ENOMEM for GDS when all GDS resources
+       * are busy. The workaround is to wait until they are idle.
+       */
+      while ((r = amdgpu_cs_submit_raw2(ws->dev, acs->ctx->ctx, bo_list,
+                                        num_chunks, chunks, &seq_no)) == -ENOMEM)
+         usleep(2000);
    }
 finalize:
 
-- 
2.17.1