]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/commitdiff
drm/amdgpu: split VM mappings into smaller operations (v3)
authorChristian König <christian.koenig@amd.com>
Mon, 25 Jan 2016 13:27:31 +0000 (14:27 +0100)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 10 Feb 2016 19:17:05 +0000 (14:17 -0500)
If we can't copy entries from the GTT or fill them with one command split
up the mapping operation into multiple ones.

v2: agd: rebase on upstream
v3: squash in Christian's fix

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

index b371a60837ebcc2eadf4db987201d127e13dc245..7b660db3693160b5529450508469b5c845787b1b 100644 (file)
@@ -696,42 +696,32 @@ static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
  *
  * @adev: amdgpu_device pointer
  * @gtt: GART instance to use for mapping
+ * @gtt_flags: flags as they are used for GTT
  * @vm: requested vm
- * @mapping: mapped range and flags to use for the update
+ * @start: start of mapped range
+ * @last: last mapped entry
+ * @flags: flags for the entries
  * @addr: addr to set the area to
- * @gtt_flags: flags as they are used for GTT
  * @fence: optional resulting fence
  *
- * Fill in the page table entries for @mapping.
+ * Fill in the page table entries between @start and @last.
  * Returns 0 for success, -EINVAL for failure.
- *
- * Object have to be reserved and mutex must be locked!
  */
 static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
                                       struct amdgpu_gart *gtt,
                                       uint32_t gtt_flags,
                                       struct amdgpu_vm *vm,
-                                      struct amdgpu_bo_va_mapping *mapping,
-                                      uint64_t addr, struct fence **fence)
+                                      uint64_t start, uint64_t last,
+                                      uint32_t flags, uint64_t addr,
+                                      struct fence **fence)
 {
        struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
        unsigned nptes, ncmds, ndw;
-       uint32_t flags = gtt_flags;
        struct amdgpu_ib *ib;
        struct fence *f = NULL;
        int r;
 
-       /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
-        * but in case of something, we filter the flags in first place
-        */
-       if (!(mapping->flags & AMDGPU_PTE_READABLE))
-               flags &= ~AMDGPU_PTE_READABLE;
-       if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
-               flags &= ~AMDGPU_PTE_WRITEABLE;
-
-       trace_amdgpu_vm_bo_update(mapping);
-
-       nptes = mapping->it.last - mapping->it.start + 1;
+       nptes = last - start + 1;
 
        /*
         * reserve space for one command every (1 << BLOCK_SIZE)
@@ -773,10 +763,8 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 
        ib->length_dw = 0;
 
-       r = amdgpu_vm_update_ptes(adev, gtt, gtt_flags, vm, ib,
-                                 mapping->it.start, mapping->it.last + 1,
-                                 addr + mapping->offset, flags);
-
+       r = amdgpu_vm_update_ptes(adev, gtt, gtt_flags, vm, ib, start,
+                                 last + 1, addr, flags);
        if (r) {
                amdgpu_ib_free(adev, ib);
                kfree(ib);
@@ -806,6 +794,68 @@ error_free:
        return r;
 }
 
+/**
+ * amdgpu_vm_bo_split_mapping - split a mapping into smaller chunks
+ *
+ * @adev: amdgpu_device pointer
+ * @gtt: GART instance to use for mapping
+ * @vm: requested vm
+ * @mapping: mapped range and flags to use for the update
+ * @addr: addr to set the area to
+ * @gtt_flags: flags as they are used for GTT
+ * @fence: optional resulting fence
+ *
+ * Split the mapping into smaller chunks so that each update fits
+ * into a SDMA IB.
+ * Returns 0 for success, -EINVAL for failure.
+ */
+static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
+                                     struct amdgpu_gart *gtt,
+                                     uint32_t gtt_flags,
+                                     struct amdgpu_vm *vm,
+                                     struct amdgpu_bo_va_mapping *mapping,
+                                     uint64_t addr, struct fence **fence)
+{
+       const uint64_t max_size = 64ULL * 1024ULL * 1024ULL / AMDGPU_GPU_PAGE_SIZE;
+
+       uint64_t start = mapping->it.start;
+       uint32_t flags = gtt_flags;
+       int r;
+
+       /* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
+        * but in case of something, we filter the flags in first place
+        */
+       if (!(mapping->flags & AMDGPU_PTE_READABLE))
+               flags &= ~AMDGPU_PTE_READABLE;
+       if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
+               flags &= ~AMDGPU_PTE_WRITEABLE;
+
+       trace_amdgpu_vm_bo_update(mapping);
+
+       addr += mapping->offset;
+
+       if (!gtt || ((gtt == &adev->gart) && (flags == gtt_flags)))
+               return amdgpu_vm_bo_update_mapping(adev, gtt, gtt_flags, vm,
+                                                  start, mapping->it.last,
+                                                  flags, addr, fence);
+
+       while (start != mapping->it.last + 1) {
+               uint64_t last;
+
+               last = min((uint64_t)mapping->it.last, start + max_size);
+               r = amdgpu_vm_bo_update_mapping(adev, gtt, gtt_flags, vm,
+                                               start, last, flags, addr,
+                                               fence);
+               if (r)
+                       return r;
+
+               start = last + 1;
+               addr += max_size;
+       }
+
+       return 0;
+}
+
 /**
  * amdgpu_vm_bo_update - update all BO mappings in the vm page table
  *
@@ -855,8 +905,8 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
        spin_unlock(&vm->status_lock);
 
        list_for_each_entry(mapping, &bo_va->invalids, list) {
-               r = amdgpu_vm_bo_update_mapping(adev, gtt, flags, vm, mapping, addr,
-                                               &bo_va->last_pt_update);
+               r = amdgpu_vm_bo_split_mapping(adev, gtt, flags, vm, mapping, addr,
+                                              &bo_va->last_pt_update);
                if (r)
                        return r;
        }
@@ -902,8 +952,8 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
                        struct amdgpu_bo_va_mapping, list);
                list_del(&mapping->list);
                spin_unlock(&vm->freed_lock);
-               r = amdgpu_vm_bo_update_mapping(adev, NULL, 0, vm, mapping,
-                                               0, NULL);
+               r = amdgpu_vm_bo_split_mapping(adev, NULL, 0, vm, mapping,
+                                              0, NULL);
                kfree(mapping);
                if (r)
                        return r;