[Mesa-dev] [PATCH] winsys/radeon: Use CPU page size instead of hardcoding 4096 bytes

[Michel Dänzer]

From: Michel Dänzer <michel.daenzer at amd.com>

Fixes GPUVM conflicts with non-4K page size.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92738

Cc: mesa-stable at lists.freedesktop.org
Signed-off-by: Michel Dänzer <michel.daenzer at amd.com>
---
 src/gallium/winsys/radeon/drm/radeon_drm_bo.c | 29 ++++++++++++++++++---------
 1 file changed, 19 insertions(+), 10 deletions(-)

diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_bo.c b/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
index 2878c8f..776a6a8 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
@@ -76,6 +76,9 @@ struct radeon_bomgr {
     bool va;
     uint64_t va_offset;
     struct list_head va_holes;
+
+    /* BO size alignment */
+    long size_align;
 };
 
 static inline struct radeon_bomgr *radeon_bomgr(struct pb_manager *mgr)
@@ -188,8 +191,11 @@ static uint64_t radeon_bomgr_find_va(struct radeon_bomgr *mgr, uint64_t size, ui
     struct radeon_bo_va_hole *hole, *n;
     uint64_t offset = 0, waste = 0;
 
+    /* Start of VM address range only needs to be aligned to the GPU page
+     * size, which is always 4K
+     */
     alignment = MAX2(alignment, 4096);
-    size = align(size, 4096);
+    size = align(size, mgr->size_align);
 
     pipe_mutex_lock(mgr->bo_va_mutex);
     /* first look for a hole */
@@ -246,7 +252,7 @@ static void radeon_bomgr_free_va(struct radeon_bomgr *mgr, uint64_t va, uint64_t
 {
     struct radeon_bo_va_hole *hole;
 
-    size = align(size, 4096);
+    size = align(size, mgr->size_align);
 
     pipe_mutex_lock(mgr->bo_va_mutex);
     if ((va + size) == mgr->va_offset) {
@@ -357,9 +363,9 @@ static void radeon_bo_destroy(struct pb_buffer *_buf)
     pipe_mutex_destroy(bo->map_mutex);
 
     if (bo->initial_domain & RADEON_DOMAIN_VRAM)
-        bo->rws->allocated_vram -= align(bo->base.size, 4096);
+        bo->rws->allocated_vram -= align(bo->base.size, mgr->size_align);
     else if (bo->initial_domain & RADEON_DOMAIN_GTT)
-        bo->rws->allocated_gtt -= align(bo->base.size, 4096);
+        bo->rws->allocated_gtt -= align(bo->base.size, mgr->size_align);
     FREE(bo);
 }
 
@@ -644,9 +650,9 @@ static struct pb_buffer *radeon_bomgr_create_bo(struct pb_manager *_mgr,
     }
 
     if (rdesc->initial_domains & RADEON_DOMAIN_VRAM)
-        rws->allocated_vram += align(size, 4096);
+        rws->allocated_vram += align(size, mgr->size_align);
     else if (rdesc->initial_domains & RADEON_DOMAIN_GTT)
-        rws->allocated_gtt += align(size, 4096);
+        rws->allocated_gtt += align(size, mgr->size_align);
 
     return &bo->base;
 }
@@ -720,6 +726,9 @@ struct pb_manager *radeon_bomgr_create(struct radeon_drm_winsys *rws)
     mgr->va_offset = rws->va_start;
     list_inithead(&mgr->va_holes);
 
+    /* TTM aligns the BO size to the CPU page size */
+    mgr->size_align = sysconf(_SC_PAGESIZE);
+
     return &mgr->base;
 }
 
@@ -882,7 +891,7 @@ radeon_winsys_bo_create(struct radeon_winsys *rws,
      * BOs. Aligning this here helps the cached bufmgr. Especially small BOs,
      * like constant/uniform buffers, can benefit from better and more reuse.
      */
-    size = align(size, 4096);
+    size = align(size, mgr->size_align);
 
     /* Only set one usage bit each for domains and flags, or the cache manager
      * might consider different sets of domains / flags compatible
@@ -993,7 +1002,7 @@ static struct pb_buffer *radeon_winsys_bo_from_ptr(struct radeon_winsys *rws,
         pipe_mutex_unlock(mgr->bo_handles_mutex);
     }
 
-    ws->allocated_gtt += align(bo->base.size, 4096);
+    ws->allocated_gtt += align(bo->base.size, mgr->size_align);
 
     return (struct pb_buffer*)bo;
 }
@@ -1130,9 +1139,9 @@ done:
     bo->initial_domain = radeon_bo_get_initial_domain((void*)bo);
 
     if (bo->initial_domain & RADEON_DOMAIN_VRAM)
-        ws->allocated_vram += align(bo->base.size, 4096);
+        ws->allocated_vram += align(bo->base.size, mgr->size_align);
     else if (bo->initial_domain & RADEON_DOMAIN_GTT)
-        ws->allocated_gtt += align(bo->base.size, 4096);
+        ws->allocated_gtt += align(bo->base.size, mgr->size_align);
 
     return (struct pb_buffer*)bo;
 
-- 
2.6.2