drm/amdgpu: use BOs GART instance for mapping addresses v4
authorChristian König <christian.koenig@amd.com>
Mon, 30 Nov 2015 13:19:26 +0000 (14:19 +0100)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 10 Feb 2016 19:17:05 +0000 (14:17 -0500)
That allows the VM code to use GART BOs from other driver instances.

v2: don't use copy optimization for foreign GARTs, that won't work.
v3: some more comment cleanups
v4: agd: rebase on upstream

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 ae3b275..b371a60 100644 (file)
@@ -276,31 +276,34 @@ struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
  * amdgpu_vm_update_pages - helper to call the right asic function
  *
  * @adev: amdgpu_device pointer
+ * @gtt: GART instance to use for mapping
+ * @gtt_flags: GTT hw access flags
  * @ib: indirect buffer to fill with commands
  * @pe: addr of the page entry
  * @addr: dst addr to write into pe
  * @count: number of page entries to update
  * @incr: increase next addr by incr bytes
  * @flags: hw access flags
- * @gtt_flags: GTT hw access flags
  *
  * Traces the parameters and calls the right asic functions
  * to setup the page table using the DMA.
  */
 static void amdgpu_vm_update_pages(struct amdgpu_device *adev,
+                                  struct amdgpu_gart *gtt,
+                                  uint32_t gtt_flags,
                                   struct amdgpu_ib *ib,
                                   uint64_t pe, uint64_t addr,
                                   unsigned count, uint32_t incr,
-                                  uint32_t flags, uint32_t gtt_flags)
+                                  uint32_t flags)
 {
        trace_amdgpu_vm_set_page(pe, addr, count, incr, flags);
 
-       if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
-               uint64_t src = adev->gart.table_addr + (addr >> 12) * 8;
+       if ((gtt == &adev->gart) && (flags == gtt_flags)) {
+               uint64_t src = gtt->table_addr + (addr >> 12) * 8;
                amdgpu_vm_copy_pte(adev, ib, pe, src, count);
 
-       } else if (flags & AMDGPU_PTE_SYSTEM) {
-               dma_addr_t *pages_addr = adev->gart.pages_addr;
+       } else if (gtt) {
+               dma_addr_t *pages_addr = gtt->pages_addr;
                amdgpu_vm_write_pte(adev, ib, pages_addr, pe, addr,
                                    count, incr, flags);
 
@@ -362,7 +365,8 @@ static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
 
        ib->length_dw = 0;
 
-       amdgpu_vm_update_pages(adev, ib, addr, 0, entries, 0, 0, 0);
+       amdgpu_vm_update_pages(adev, NULL, 0, ib, addr, 0, entries, 0, 0);
+
        amdgpu_vm_pad_ib(adev, ib);
        WARN_ON(ib->length_dw > 64);
        r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
@@ -475,9 +479,10 @@ int amdgpu_vm_update_page_directory(struct amdgpu_device *adev,
                    ((last_pt + incr * count) != pt)) {
 
                        if (count) {
-                               amdgpu_vm_update_pages(adev, ib, last_pde,
-                                                      last_pt, count, incr,
-                                                      AMDGPU_PTE_VALID, 0);
+                               amdgpu_vm_update_pages(adev, NULL, 0, ib,
+                                                      last_pde, last_pt,
+                                                      count, incr,
+                                                      AMDGPU_PTE_VALID);
                        }
 
                        count = 1;
@@ -489,8 +494,8 @@ int amdgpu_vm_update_page_directory(struct amdgpu_device *adev,
        }
 
        if (count)
-               amdgpu_vm_update_pages(adev, ib, last_pde, last_pt, count,
-                                      incr, AMDGPU_PTE_VALID, 0);
+               amdgpu_vm_update_pages(adev, NULL, 0, ib, last_pde, last_pt,
+                                      count, incr, AMDGPU_PTE_VALID);
 
        if (ib->length_dw != 0) {
                amdgpu_vm_pad_ib(adev, ib);
@@ -526,20 +531,22 @@ error_free:
  * amdgpu_vm_frag_ptes - add fragment information to PTEs
  *
  * @adev: amdgpu_device pointer
+ * @gtt: GART instance to use for mapping
+ * @gtt_flags: GTT hw mapping flags
  * @ib: IB for the update
  * @pe_start: first PTE to handle
  * @pe_end: last PTE to handle
  * @addr: addr those PTEs should point to
  * @flags: hw mapping flags
- * @gtt_flags: GTT hw mapping flags
  *
  * Global and local mutex must be locked!
  */
 static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
+                               struct amdgpu_gart *gtt,
+                               uint32_t gtt_flags,
                                struct amdgpu_ib *ib,
                                uint64_t pe_start, uint64_t pe_end,
-                               uint64_t addr, uint32_t flags,
-                               uint32_t gtt_flags)
+                               uint64_t addr, uint32_t flags)
 {
        /**
         * The MC L1 TLB supports variable sized pages, based on a fragment
@@ -570,35 +577,34 @@ static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
        unsigned count;
 
        /* system pages are non continuously */
-       if ((flags & AMDGPU_PTE_SYSTEM) || !(flags & AMDGPU_PTE_VALID) ||
-           (frag_start >= frag_end)) {
+       if (gtt || !(flags & AMDGPU_PTE_VALID) || (frag_start >= frag_end)) {
 
                count = (pe_end - pe_start) / 8;
-               amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
-                                      AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
+               amdgpu_vm_update_pages(adev, gtt, gtt_flags, ib, pe_start,
+                                      addr, count, AMDGPU_GPU_PAGE_SIZE,
+                                      flags);
                return;
        }
 
        /* handle the 4K area at the beginning */
        if (pe_start != frag_start) {
                count = (frag_start - pe_start) / 8;
-               amdgpu_vm_update_pages(adev, ib, pe_start, addr, count,
-                                      AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
+               amdgpu_vm_update_pages(adev, NULL, 0, ib, pe_start, addr,
+                                      count, AMDGPU_GPU_PAGE_SIZE, flags);
                addr += AMDGPU_GPU_PAGE_SIZE * count;
        }
 
        /* handle the area in the middle */
        count = (frag_end - frag_start) / 8;
-       amdgpu_vm_update_pages(adev, ib, frag_start, addr, count,
-                              AMDGPU_GPU_PAGE_SIZE, flags | frag_flags,
-                              gtt_flags);
+       amdgpu_vm_update_pages(adev, NULL, 0, ib, frag_start, addr, count,
+                              AMDGPU_GPU_PAGE_SIZE, flags | frag_flags);
 
        /* handle the 4K area at the end */
        if (frag_end != pe_end) {
                addr += AMDGPU_GPU_PAGE_SIZE * count;
                count = (pe_end - frag_end) / 8;
-               amdgpu_vm_update_pages(adev, ib, frag_end, addr, count,
-                                      AMDGPU_GPU_PAGE_SIZE, flags, gtt_flags);
+               amdgpu_vm_update_pages(adev, NULL, 0, ib, frag_end, addr,
+                                      count, AMDGPU_GPU_PAGE_SIZE, flags);
        }
 }
 
@@ -606,6 +612,8 @@ static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
  * amdgpu_vm_update_ptes - make sure that page tables are valid
  *
  * @adev: amdgpu_device pointer
+ * @gtt: GART instance to use for mapping
+ * @gtt_flags: GTT hw mapping flags
  * @vm: requested vm
  * @start: start of GPU address range
  * @end: end of GPU address range
@@ -617,11 +625,12 @@ static void amdgpu_vm_frag_ptes(struct amdgpu_device *adev,
  * Global and local mutex must be locked!
  */
 static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
+                                struct amdgpu_gart *gtt,
+                                uint32_t gtt_flags,
                                 struct amdgpu_vm *vm,
                                 struct amdgpu_ib *ib,
                                 uint64_t start, uint64_t end,
-                                uint64_t dst, uint32_t flags,
-                                uint32_t gtt_flags)
+                                uint64_t dst, uint32_t flags)
 {
        uint64_t mask = AMDGPU_VM_PTE_COUNT - 1;
        uint64_t last_pte = ~0, last_dst = ~0;
@@ -657,10 +666,9 @@ static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
                if ((last_pte + 8 * count) != pte) {
 
                        if (count) {
-                               amdgpu_vm_frag_ptes(adev, ib, last_pte,
-                                                   last_pte + 8 * count,
-                                                   last_dst, flags,
-                                                   gtt_flags);
+                               amdgpu_vm_frag_ptes(adev, gtt, gtt_flags, ib,
+                                                   last_pte, last_pte + 8 * count,
+                                                   last_dst, flags);
                        }
 
                        count = nptes;
@@ -675,9 +683,9 @@ static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
        }
 
        if (count) {
-               amdgpu_vm_frag_ptes(adev, ib, last_pte,
-                                   last_pte + 8 * count,
-                                   last_dst, flags, gtt_flags);
+               amdgpu_vm_frag_ptes(adev, gtt, gtt_flags, ib,
+                                   last_pte, last_pte + 8 * count,
+                                   last_dst, flags);
        }
 
        return 0;
@@ -687,6 +695,7 @@ static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
  * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
  *
  * @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
@@ -699,10 +708,11 @@ static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
  * 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, uint32_t gtt_flags,
-                                      struct fence **fence)
+                                      uint64_t addr, struct fence **fence)
 {
        struct amdgpu_ring *ring = adev->vm_manager.vm_pte_funcs_ring;
        unsigned nptes, ncmds, ndw;
@@ -732,11 +742,11 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
        /* padding, etc. */
        ndw = 64;
 
-       if ((flags & AMDGPU_PTE_SYSTEM) && (flags == gtt_flags)) {
+       if ((gtt == &adev->gart) && (flags == gtt_flags)) {
                /* only copy commands needed */
                ndw += ncmds * 7;
 
-       } else if (flags & AMDGPU_PTE_SYSTEM) {
+       } else if (gtt) {
                /* header for write data commands */
                ndw += ncmds * 4;
 
@@ -763,9 +773,9 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 
        ib->length_dw = 0;
 
-       r = amdgpu_vm_update_ptes(adev, vm, ib, mapping->it.start,
-                                 mapping->it.last + 1, addr + mapping->offset,
-                                 flags, gtt_flags);
+       r = amdgpu_vm_update_ptes(adev, gtt, gtt_flags, vm, ib,
+                                 mapping->it.start, mapping->it.last + 1,
+                                 addr + mapping->offset, flags);
 
        if (r) {
                amdgpu_ib_free(adev, ib);
@@ -814,14 +824,25 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
 {
        struct amdgpu_vm *vm = bo_va->vm;
        struct amdgpu_bo_va_mapping *mapping;
+       struct amdgpu_gart *gtt = NULL;
        uint32_t flags;
        uint64_t addr;
        int r;
 
        if (mem) {
                addr = (u64)mem->start << PAGE_SHIFT;
-               if (mem->mem_type != TTM_PL_TT)
+               switch (mem->mem_type) {
+               case TTM_PL_TT:
+                       gtt = &bo_va->bo->adev->gart;
+                       break;
+
+               case TTM_PL_VRAM:
                        addr += adev->vm_manager.vram_base_offset;
+                       break;
+
+               default:
+                       break;
+               }
        } else {
                addr = 0;
        }
@@ -834,8 +855,8 @@ 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, vm, mapping, addr,
-                                               flags, &bo_va->last_pt_update);
+               r = amdgpu_vm_bo_update_mapping(adev, gtt, flags, vm, mapping, addr,
+                                               &bo_va->last_pt_update);
                if (r)
                        return r;
        }
@@ -881,7 +902,8 @@ 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, vm, mapping, 0, 0, NULL);
+               r = amdgpu_vm_bo_update_mapping(adev, NULL, 0, vm, mapping,
+                                               0, NULL);
                kfree(mapping);
                if (r)
                        return r;