Re: [PATCH] drm/amdgpu: only keep most recent fence for each context

[Yadav, Arvind]

On 4/15/2025 7:53 PM, Christian König wrote:

Am 15.04.25 um 16:18 schrieb Arvind Yadav:

Mesa passes shared bo, fence syncobj to userq_ioctl.
There can be duplicates here or some fences that are old.
This patch is remove duplicates fence and only keep
the most recent fence for each context.

Cc: Alex Deucher <alexander.deuc...@amd.com>
Cc: Christian König <christian.koe...@amd.com>
Cc: Sunil Khatri <sunil.kha...@amd.com>
Cc: Arunpravin Paneer Selvam <arunpravin.paneersel...@amd.com>
Signed-off-by: Arvind Yadav <arvind.ya...@amd.com>
---
  .../gpu/drm/amd/amdgpu/amdgpu_userq_fence.c   | 37 +++++++++++++++++++
  1 file changed, 37 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
index a4953d668972..9a09d545ceb5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c
@@ -25,6 +25,7 @@
  #include <linux/kref.h>
  #include <linux/slab.h>
  #include <linux/dma-fence-unwrap.h>
+#include <linux/sort.h>

#include <drm/drm_exec.h>

  #include <drm/drm_syncobj.h>
@@ -584,6 +585,24 @@ int amdgpu_userq_signal_ioctl(struct drm_device *dev, void 
*data,
  #endif

#ifdef CONFIG_DRM_AMDGPU_NAVI3X_USERQ

+static int fence_cmp(const void *_a, const void *_b)
+{
+       struct dma_fence *a = *(struct dma_fence **)_a;
+       struct dma_fence *b = *(struct dma_fence **)_b;
+
+       if (a->context < b->context)
+               return -1;
+       else if (a->context > b->context)
+               return 1;
+
+       if (dma_fence_is_later(b, a))
+               return 1;
+       else if (dma_fence_is_later(a, b))
+               return -1;
+
+       return 0;
+}
+
  int amdgpu_userq_wait_ioctl(struct drm_device *dev, void *data,
                            struct drm_file *filp)
  {
@@ -840,6 +859,24 @@ int amdgpu_userq_wait_ioctl(struct drm_device *dev, void 
*data,
                        fences[num_fences++] = fence;
                }

+ if (num_fences > 1) {

+                       int j = 0;
+
+                       sort(fences, num_fences, sizeof(*fences), fence_cmp, 
NULL);
+
+                       /*
+                        * Only keep the most recent fence for each context.
+                        */
+                       for (int i = 1; i < num_fences; i++) {
+                               if (fences[i]->context == fences[j]->context)
+                                       dma_fence_put(fences[i]);
+                               else
+                                       fences[++j] = fences[i];
+                       }
+
+                       num_fences = ++j;
+               }
+

We already have the same code in drivers/dma-buf/dma-fence-unwrap.c.

I suggest to export this code from drivers/dma-buf/dma-fence-unwrap.c and just 
use it here as well.

Something like num_fences = dma_fence_shrink_array(fences, num_fences)

Noted.

Regards,
Christian.

                for (i = 0, cnt = 0; i < num_fences; i++) {
                        struct amdgpu_userq_fence_driver *fence_drv;
                        struct amdgpu_userq_fence *userq_fence;