1/*
2 * Copyright © 2011 Marek Olšák <maraeo@gmail.com>
3 * Copyright © 2015 Advanced Micro Devices, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining
7 * a copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
15 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
16 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
17 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
18 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * The above copyright notice and this permission notice (including the
24 * next paragraph) shall be included in all copies or substantial portions
25 * of the Software.
26 */
27
28#include "amdgpu_cs.h"
29
30#include "util/hash_table.h"
31#include "util/os_time.h"
32#include "util/u_hash_table.h"
33#include "frontend/drm_driver.h"
34#include "drm-uapi/amdgpu_drm.h"
35#include <xf86drm.h>
36#include <stdio.h>
37#include <inttypes.h>
38
39#ifndef AMDGPU_VA_RANGE_HIGH
40#define AMDGPU_VA_RANGE_HIGH	0x2
41#endif
42
43/* Set to 1 for verbose output showing committed sparse buffer ranges. */
44#define DEBUG_SPARSE_COMMITS 0
45
46struct amdgpu_sparse_backing_chunk {
47   uint32_t begin, end;
48};
49
50static bool amdgpu_bo_wait(struct radeon_winsys *rws,
51                           struct pb_buffer *_buf, uint64_t timeout,
52                           unsigned usage)
53{
54   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
55   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
56   int64_t abs_timeout = 0;
57
58   if (timeout == 0) {
59      if (p_atomic_read(&bo->num_active_ioctls))
60         return false;
61
62   } else {
63      abs_timeout = os_time_get_absolute_timeout(timeout);
64
65      /* Wait if any ioctl is being submitted with this buffer. */
66      if (!os_wait_until_zero_abs_timeout(&bo->num_active_ioctls, abs_timeout))
67         return false;
68   }
69
70   if (bo->bo && bo->u.real.is_shared) {
71      /* We can't use user fences for shared buffers, because user fences
72       * are local to this process only. If we want to wait for all buffer
73       * uses in all processes, we have to use amdgpu_bo_wait_for_idle.
74       */
75      bool buffer_busy = true;
76      int r;
77
78      r = amdgpu_bo_wait_for_idle(bo->bo, timeout, &buffer_busy);
79      if (r)
80         fprintf(stderr, "%s: amdgpu_bo_wait_for_idle failed %i\n", __func__,
81                 r);
82      return !buffer_busy;
83   }
84
85   if (timeout == 0) {
86      unsigned idle_fences;
87      bool buffer_idle;
88
89      simple_mtx_lock(&ws->bo_fence_lock);
90
91      for (idle_fences = 0; idle_fences < bo->num_fences; ++idle_fences) {
92         if (!amdgpu_fence_wait(bo->fences[idle_fences], 0, false))
93            break;
94      }
95
96      /* Release the idle fences to avoid checking them again later. */
97      for (unsigned i = 0; i < idle_fences; ++i)
98         amdgpu_fence_reference(&bo->fences[i], NULL);
99
100      memmove(&bo->fences[0], &bo->fences[idle_fences],
101              (bo->num_fences - idle_fences) * sizeof(*bo->fences));
102      bo->num_fences -= idle_fences;
103
104      buffer_idle = !bo->num_fences;
105      simple_mtx_unlock(&ws->bo_fence_lock);
106
107      return buffer_idle;
108   } else {
109      bool buffer_idle = true;
110
111      simple_mtx_lock(&ws->bo_fence_lock);
112      while (bo->num_fences && buffer_idle) {
113         struct pipe_fence_handle *fence = NULL;
114         bool fence_idle = false;
115
116         amdgpu_fence_reference(&fence, bo->fences[0]);
117
118         /* Wait for the fence. */
119         simple_mtx_unlock(&ws->bo_fence_lock);
120         if (amdgpu_fence_wait(fence, abs_timeout, true))
121            fence_idle = true;
122         else
123            buffer_idle = false;
124         simple_mtx_lock(&ws->bo_fence_lock);
125
126         /* Release an idle fence to avoid checking it again later, keeping in
127          * mind that the fence array may have been modified by other threads.
128          */
129         if (fence_idle && bo->num_fences && bo->fences[0] == fence) {
130            amdgpu_fence_reference(&bo->fences[0], NULL);
131            memmove(&bo->fences[0], &bo->fences[1],
132                    (bo->num_fences - 1) * sizeof(*bo->fences));
133            bo->num_fences--;
134         }
135
136         amdgpu_fence_reference(&fence, NULL);
137      }
138      simple_mtx_unlock(&ws->bo_fence_lock);
139
140      return buffer_idle;
141   }
142}
143
144static enum radeon_bo_domain amdgpu_bo_get_initial_domain(
145      struct pb_buffer *buf)
146{
147   return ((struct amdgpu_winsys_bo*)buf)->base.placement;
148}
149
150static enum radeon_bo_flag amdgpu_bo_get_flags(
151      struct pb_buffer *buf)
152{
153   return ((struct amdgpu_winsys_bo*)buf)->base.usage;
154}
155
156static void amdgpu_bo_remove_fences(struct amdgpu_winsys_bo *bo)
157{
158   for (unsigned i = 0; i < bo->num_fences; ++i)
159      amdgpu_fence_reference(&bo->fences[i], NULL);
160
161   FREE(bo->fences);
162   bo->num_fences = 0;
163   bo->max_fences = 0;
164}
165
166void amdgpu_bo_destroy(struct amdgpu_winsys *ws, struct pb_buffer *_buf)
167{
168   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
169   struct amdgpu_screen_winsys *sws_iter;
170
171   assert(bo->bo && "must not be called for slab entries");
172
173   if (!bo->u.real.is_user_ptr && bo->u.real.cpu_ptr) {
174      bo->u.real.cpu_ptr = NULL;
175      amdgpu_bo_unmap(&ws->dummy_ws.base, &bo->base);
176   }
177   assert(bo->u.real.is_user_ptr || bo->u.real.map_count == 0);
178
179#if DEBUG
180   if (ws->debug_all_bos) {
181      simple_mtx_lock(&ws->global_bo_list_lock);
182      list_del(&bo->u.real.global_list_item);
183      ws->num_buffers--;
184      simple_mtx_unlock(&ws->global_bo_list_lock);
185   }
186#endif
187
188   /* Close all KMS handles retrieved for other DRM file descriptions */
189   simple_mtx_lock(&ws->sws_list_lock);
190   for (sws_iter = ws->sws_list; sws_iter; sws_iter = sws_iter->next) {
191      struct hash_entry *entry;
192
193      if (!sws_iter->kms_handles)
194         continue;
195
196      entry = _mesa_hash_table_search(sws_iter->kms_handles, bo);
197      if (entry) {
198         struct drm_gem_close args = { .handle = (uintptr_t)entry->data };
199
200         drmIoctl(sws_iter->fd, DRM_IOCTL_GEM_CLOSE, &args);
201         _mesa_hash_table_remove(sws_iter->kms_handles, entry);
202      }
203   }
204   simple_mtx_unlock(&ws->sws_list_lock);
205
206   simple_mtx_lock(&ws->bo_export_table_lock);
207   _mesa_hash_table_remove_key(ws->bo_export_table, bo->bo);
208   simple_mtx_unlock(&ws->bo_export_table_lock);
209
210   if (bo->base.placement & RADEON_DOMAIN_VRAM_GTT) {
211      amdgpu_bo_va_op(bo->bo, 0, bo->base.size, bo->va, 0, AMDGPU_VA_OP_UNMAP);
212      amdgpu_va_range_free(bo->u.real.va_handle);
213   }
214   amdgpu_bo_free(bo->bo);
215
216   amdgpu_bo_remove_fences(bo);
217
218   if (bo->base.placement & RADEON_DOMAIN_VRAM)
219      ws->allocated_vram -= align64(bo->base.size, ws->info.gart_page_size);
220   else if (bo->base.placement & RADEON_DOMAIN_GTT)
221      ws->allocated_gtt -= align64(bo->base.size, ws->info.gart_page_size);
222
223   simple_mtx_destroy(&bo->lock);
224   FREE(bo);
225}
226
227static void amdgpu_bo_destroy_or_cache(struct radeon_winsys *rws, struct pb_buffer *_buf)
228{
229   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
230   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
231
232   assert(bo->bo); /* slab buffers have a separate vtbl */
233
234   if (bo->u.real.use_reusable_pool)
235      pb_cache_add_buffer(bo->cache_entry);
236   else
237      amdgpu_bo_destroy(ws, _buf);
238}
239
240static void amdgpu_clean_up_buffer_managers(struct amdgpu_winsys *ws)
241{
242   for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++)
243      pb_slabs_reclaim(&ws->bo_slabs[i]);
244
245   pb_cache_release_all_buffers(&ws->bo_cache);
246}
247
248static bool amdgpu_bo_do_map(struct radeon_winsys *rws, struct amdgpu_winsys_bo *bo, void **cpu)
249{
250   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
251
252   assert(!(bo->base.usage & RADEON_FLAG_SPARSE) && bo->bo && !bo->u.real.is_user_ptr);
253   int r = amdgpu_bo_cpu_map(bo->bo, cpu);
254   if (r) {
255      /* Clean up buffer managers and try again. */
256      amdgpu_clean_up_buffer_managers(ws);
257      r = amdgpu_bo_cpu_map(bo->bo, cpu);
258      if (r)
259         return false;
260   }
261
262   if (p_atomic_inc_return(&bo->u.real.map_count) == 1) {
263      if (bo->base.placement & RADEON_DOMAIN_VRAM)
264         ws->mapped_vram += bo->base.size;
265      else if (bo->base.placement & RADEON_DOMAIN_GTT)
266         ws->mapped_gtt += bo->base.size;
267      ws->num_mapped_buffers++;
268   }
269
270   return true;
271}
272
273void *amdgpu_bo_map(struct radeon_winsys *rws,
274                    struct pb_buffer *buf,
275                    struct radeon_cmdbuf *rcs,
276                    enum pipe_map_flags usage)
277{
278   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
279   struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf;
280   struct amdgpu_winsys_bo *real;
281   struct amdgpu_cs *cs = rcs ? amdgpu_cs(rcs) : NULL;
282
283   assert(!(bo->base.usage & RADEON_FLAG_SPARSE));
284
285   /* If it's not unsynchronized bo_map, flush CS if needed and then wait. */
286   if (!(usage & PIPE_MAP_UNSYNCHRONIZED)) {
287      /* DONTBLOCK doesn't make sense with UNSYNCHRONIZED. */
288      if (usage & PIPE_MAP_DONTBLOCK) {
289         if (!(usage & PIPE_MAP_WRITE)) {
290            /* Mapping for read.
291             *
292             * Since we are mapping for read, we don't need to wait
293             * if the GPU is using the buffer for read too
294             * (neither one is changing it).
295             *
296             * Only check whether the buffer is being used for write. */
297            if (cs && amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo,
298                                                               RADEON_USAGE_WRITE)) {
299               cs->flush_cs(cs->flush_data,
300			    RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
301               return NULL;
302            }
303
304            if (!amdgpu_bo_wait(rws, (struct pb_buffer*)bo, 0,
305                                RADEON_USAGE_WRITE)) {
306               return NULL;
307            }
308         } else {
309            if (cs && amdgpu_bo_is_referenced_by_cs(cs, bo)) {
310               cs->flush_cs(cs->flush_data,
311			    RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
312               return NULL;
313            }
314
315            if (!amdgpu_bo_wait(rws, (struct pb_buffer*)bo, 0,
316                                RADEON_USAGE_READWRITE)) {
317               return NULL;
318            }
319         }
320      } else {
321         uint64_t time = os_time_get_nano();
322
323         if (!(usage & PIPE_MAP_WRITE)) {
324            /* Mapping for read.
325             *
326             * Since we are mapping for read, we don't need to wait
327             * if the GPU is using the buffer for read too
328             * (neither one is changing it).
329             *
330             * Only check whether the buffer is being used for write. */
331            if (cs) {
332               if (amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo,
333                                                            RADEON_USAGE_WRITE)) {
334                  cs->flush_cs(cs->flush_data,
335			       RADEON_FLUSH_START_NEXT_GFX_IB_NOW, NULL);
336               } else {
337                  /* Try to avoid busy-waiting in amdgpu_bo_wait. */
338                  if (p_atomic_read(&bo->num_active_ioctls))
339                     amdgpu_cs_sync_flush(rcs);
340               }
341            }
342
343            amdgpu_bo_wait(rws, (struct pb_buffer*)bo, PIPE_TIMEOUT_INFINITE,
344                           RADEON_USAGE_WRITE);
345         } else {
346            /* Mapping for write. */
347            if (cs) {
348               if (amdgpu_bo_is_referenced_by_cs(cs, bo)) {
349                  cs->flush_cs(cs->flush_data,
350			       RADEON_FLUSH_START_NEXT_GFX_IB_NOW, NULL);
351               } else {
352                  /* Try to avoid busy-waiting in amdgpu_bo_wait. */
353                  if (p_atomic_read(&bo->num_active_ioctls))
354                     amdgpu_cs_sync_flush(rcs);
355               }
356            }
357
358            amdgpu_bo_wait(rws, (struct pb_buffer*)bo, PIPE_TIMEOUT_INFINITE,
359                           RADEON_USAGE_READWRITE);
360         }
361
362         ws->buffer_wait_time += os_time_get_nano() - time;
363      }
364   }
365
366   /* Buffer synchronization has been checked, now actually map the buffer. */
367   void *cpu = NULL;
368   uint64_t offset = 0;
369
370   if (bo->bo) {
371      real = bo;
372   } else {
373      real = bo->u.slab.real;
374      offset = bo->va - real->va;
375   }
376
377   if (usage & RADEON_MAP_TEMPORARY) {
378      if (real->u.real.is_user_ptr) {
379         cpu = real->u.real.cpu_ptr;
380      } else {
381         if (!amdgpu_bo_do_map(rws, real, &cpu))
382            return NULL;
383      }
384   } else {
385      cpu = p_atomic_read(&real->u.real.cpu_ptr);
386      if (!cpu) {
387         simple_mtx_lock(&real->lock);
388         /* Must re-check due to the possibility of a race. Re-check need not
389          * be atomic thanks to the lock. */
390         cpu = real->u.real.cpu_ptr;
391         if (!cpu) {
392            if (!amdgpu_bo_do_map(rws, real, &cpu)) {
393               simple_mtx_unlock(&real->lock);
394               return NULL;
395            }
396            p_atomic_set(&real->u.real.cpu_ptr, cpu);
397         }
398         simple_mtx_unlock(&real->lock);
399      }
400   }
401
402   return (uint8_t*)cpu + offset;
403}
404
405void amdgpu_bo_unmap(struct radeon_winsys *rws, struct pb_buffer *buf)
406{
407   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
408   struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf;
409   struct amdgpu_winsys_bo *real;
410
411   assert(!(bo->base.usage & RADEON_FLAG_SPARSE));
412
413   real = bo->bo ? bo : bo->u.slab.real;
414
415   if (real->u.real.is_user_ptr)
416      return;
417
418   assert(real->u.real.map_count != 0 && "too many unmaps");
419   if (p_atomic_dec_zero(&real->u.real.map_count)) {
420      assert(!real->u.real.cpu_ptr &&
421             "too many unmaps or forgot RADEON_MAP_TEMPORARY flag");
422
423      if (real->base.placement & RADEON_DOMAIN_VRAM)
424         ws->mapped_vram -= real->base.size;
425      else if (real->base.placement & RADEON_DOMAIN_GTT)
426         ws->mapped_gtt -= real->base.size;
427      ws->num_mapped_buffers--;
428   }
429
430   amdgpu_bo_cpu_unmap(real->bo);
431}
432
433static const struct pb_vtbl amdgpu_winsys_bo_vtbl = {
434   /* Cast to void* because one of the function parameters is a struct pointer instead of void*. */
435   (void*)amdgpu_bo_destroy_or_cache
436   /* other functions are never called */
437};
438
439static void amdgpu_add_buffer_to_global_list(struct amdgpu_winsys *ws, struct amdgpu_winsys_bo *bo)
440{
441#if DEBUG
442   assert(bo->bo);
443
444   if (ws->debug_all_bos) {
445      simple_mtx_lock(&ws->global_bo_list_lock);
446      list_addtail(&bo->u.real.global_list_item, &ws->global_bo_list);
447      ws->num_buffers++;
448      simple_mtx_unlock(&ws->global_bo_list_lock);
449   }
450#endif
451}
452
453static unsigned amdgpu_get_optimal_alignment(struct amdgpu_winsys *ws,
454                                             uint64_t size, unsigned alignment)
455{
456   /* Increase the alignment for faster address translation and better memory
457    * access pattern.
458    */
459   if (size >= ws->info.pte_fragment_size) {
460      alignment = MAX2(alignment, ws->info.pte_fragment_size);
461   } else if (size) {
462      unsigned msb = util_last_bit(size);
463
464      alignment = MAX2(alignment, 1u << (msb - 1));
465   }
466   return alignment;
467}
468
469static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws,
470                                                 uint64_t size,
471                                                 unsigned alignment,
472                                                 enum radeon_bo_domain initial_domain,
473                                                 unsigned flags,
474                                                 int heap)
475{
476   struct amdgpu_bo_alloc_request request = {0};
477   amdgpu_bo_handle buf_handle;
478   uint64_t va = 0;
479   struct amdgpu_winsys_bo *bo;
480   amdgpu_va_handle va_handle = NULL;
481   int r;
482   bool init_pb_cache;
483
484   /* VRAM or GTT must be specified, but not both at the same time. */
485   assert(util_bitcount(initial_domain & (RADEON_DOMAIN_VRAM_GTT |
486                                          RADEON_DOMAIN_GDS |
487                                          RADEON_DOMAIN_OA)) == 1);
488
489   alignment = amdgpu_get_optimal_alignment(ws, size, alignment);
490
491   init_pb_cache = heap >= 0 && (flags & RADEON_FLAG_NO_INTERPROCESS_SHARING);
492
493   bo = CALLOC(1, sizeof(struct amdgpu_winsys_bo) +
494                  init_pb_cache * sizeof(struct pb_cache_entry));
495   if (!bo) {
496      return NULL;
497   }
498
499   if (init_pb_cache) {
500      bo->u.real.use_reusable_pool = true;
501      pb_cache_init_entry(&ws->bo_cache, bo->cache_entry, &bo->base,
502                          heap);
503   }
504   request.alloc_size = size;
505   request.phys_alignment = alignment;
506
507   if (initial_domain & RADEON_DOMAIN_VRAM) {
508      request.preferred_heap |= AMDGPU_GEM_DOMAIN_VRAM;
509
510      /* Since VRAM and GTT have almost the same performance on APUs, we could
511       * just set GTT. However, in order to decrease GTT(RAM) usage, which is
512       * shared with the OS, allow VRAM placements too. The idea is not to use
513       * VRAM usefully, but to use it so that it's not unused and wasted.
514       */
515      if (!ws->info.has_dedicated_vram)
516         request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT;
517   }
518
519   if (initial_domain & RADEON_DOMAIN_GTT)
520      request.preferred_heap |= AMDGPU_GEM_DOMAIN_GTT;
521   if (initial_domain & RADEON_DOMAIN_GDS)
522      request.preferred_heap |= AMDGPU_GEM_DOMAIN_GDS;
523   if (initial_domain & RADEON_DOMAIN_OA)
524      request.preferred_heap |= AMDGPU_GEM_DOMAIN_OA;
525
526   if (flags & RADEON_FLAG_NO_CPU_ACCESS)
527      request.flags |= AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
528   if (flags & RADEON_FLAG_GTT_WC)
529      request.flags |= AMDGPU_GEM_CREATE_CPU_GTT_USWC;
530
531   if (flags & RADEON_FLAG_DISCARDABLE &&
532       ws->info.drm_minor >= 47)
533      request.flags |= AMDGPU_GEM_CREATE_DISCARDABLE;
534
535   if (ws->zero_all_vram_allocs &&
536       (request.preferred_heap & AMDGPU_GEM_DOMAIN_VRAM))
537      request.flags |= AMDGPU_GEM_CREATE_VRAM_CLEARED;
538
539   if ((flags & RADEON_FLAG_ENCRYPTED) &&
540       ws->info.has_tmz_support) {
541      request.flags |= AMDGPU_GEM_CREATE_ENCRYPTED;
542
543      if (!(flags & RADEON_FLAG_DRIVER_INTERNAL)) {
544         struct amdgpu_screen_winsys *sws_iter;
545         simple_mtx_lock(&ws->sws_list_lock);
546         for (sws_iter = ws->sws_list; sws_iter; sws_iter = sws_iter->next) {
547            *((bool*) &sws_iter->base.uses_secure_bos) = true;
548         }
549         simple_mtx_unlock(&ws->sws_list_lock);
550      }
551   }
552
553   r = amdgpu_bo_alloc(ws->dev, &request, &buf_handle);
554   if (r) {
555      fprintf(stderr, "amdgpu: Failed to allocate a buffer:\n");
556      fprintf(stderr, "amdgpu:    size      : %"PRIu64" bytes\n", size);
557      fprintf(stderr, "amdgpu:    alignment : %u bytes\n", alignment);
558      fprintf(stderr, "amdgpu:    domains   : %u\n", initial_domain);
559      fprintf(stderr, "amdgpu:    flags   : %" PRIx64 "\n", request.flags);
560      goto error_bo_alloc;
561   }
562
563   if (initial_domain & RADEON_DOMAIN_VRAM_GTT) {
564      unsigned va_gap_size = ws->check_vm ? MAX2(4 * alignment, 64 * 1024) : 0;
565
566      r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
567                                size + va_gap_size, alignment,
568                                0, &va, &va_handle,
569                                (flags & RADEON_FLAG_32BIT ? AMDGPU_VA_RANGE_32_BIT : 0) |
570                                AMDGPU_VA_RANGE_HIGH);
571      if (r)
572         goto error_va_alloc;
573
574      unsigned vm_flags = AMDGPU_VM_PAGE_READABLE |
575                          AMDGPU_VM_PAGE_EXECUTABLE;
576
577      if (!(flags & RADEON_FLAG_READ_ONLY))
578         vm_flags |= AMDGPU_VM_PAGE_WRITEABLE;
579
580      if (flags & RADEON_FLAG_GL2_BYPASS)
581         vm_flags |= AMDGPU_VM_MTYPE_UC;
582
583      if (flags & RADEON_FLAG_MALL_NOALLOC &&
584          ws->info.drm_minor >= 47)
585         vm_flags |= AMDGPU_VM_PAGE_NOALLOC;
586
587      r = amdgpu_bo_va_op_raw(ws->dev, buf_handle, 0, size, va, vm_flags,
588			   AMDGPU_VA_OP_MAP);
589      if (r)
590         goto error_va_map;
591   }
592
593   simple_mtx_init(&bo->lock, mtx_plain);
594   pipe_reference_init(&bo->base.reference, 1);
595   bo->base.alignment_log2 = util_logbase2(alignment);
596   bo->base.size = size;
597   bo->base.vtbl = &amdgpu_winsys_bo_vtbl;
598   bo->bo = buf_handle;
599   bo->va = va;
600   bo->u.real.va_handle = va_handle;
601   bo->base.placement = initial_domain;
602   bo->base.usage = flags;
603   bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1);
604
605   if (initial_domain & RADEON_DOMAIN_VRAM)
606      ws->allocated_vram += align64(size, ws->info.gart_page_size);
607   else if (initial_domain & RADEON_DOMAIN_GTT)
608      ws->allocated_gtt += align64(size, ws->info.gart_page_size);
609
610   amdgpu_bo_export(bo->bo, amdgpu_bo_handle_type_kms, &bo->u.real.kms_handle);
611
612   amdgpu_add_buffer_to_global_list(ws, bo);
613
614   return bo;
615
616error_va_map:
617   amdgpu_va_range_free(va_handle);
618
619error_va_alloc:
620   amdgpu_bo_free(buf_handle);
621
622error_bo_alloc:
623   FREE(bo);
624   return NULL;
625}
626
627bool amdgpu_bo_can_reclaim(struct amdgpu_winsys *ws, struct pb_buffer *_buf)
628{
629   return amdgpu_bo_wait(&ws->dummy_ws.base, _buf, 0, RADEON_USAGE_READWRITE);
630}
631
632bool amdgpu_bo_can_reclaim_slab(void *priv, struct pb_slab_entry *entry)
633{
634   struct amdgpu_winsys_bo *bo = container_of(entry, struct amdgpu_winsys_bo, u.slab.entry);
635
636   return amdgpu_bo_can_reclaim(priv, &bo->base);
637}
638
639static struct pb_slabs *get_slabs(struct amdgpu_winsys *ws, uint64_t size)
640{
641   /* Find the correct slab allocator for the given size. */
642   for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
643      struct pb_slabs *slabs = &ws->bo_slabs[i];
644
645      if (size <= 1 << (slabs->min_order + slabs->num_orders - 1))
646         return slabs;
647   }
648
649   assert(0);
650   return NULL;
651}
652
653static unsigned get_slab_wasted_size(struct amdgpu_winsys *ws, struct amdgpu_winsys_bo *bo)
654{
655   assert(bo->base.size <= bo->u.slab.entry.entry_size);
656   assert(bo->base.size < (1 << bo->base.alignment_log2) ||
657          bo->base.size < 1 << ws->bo_slabs[0].min_order ||
658          bo->base.size > bo->u.slab.entry.entry_size / 2);
659   return bo->u.slab.entry.entry_size - bo->base.size;
660}
661
662static void amdgpu_bo_slab_destroy(struct radeon_winsys *rws, struct pb_buffer *_buf)
663{
664   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
665   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
666   struct pb_slabs *slabs;
667
668   assert(!bo->bo);
669
670   slabs = get_slabs(ws, bo->base.size);
671
672   if (bo->base.placement & RADEON_DOMAIN_VRAM)
673      ws->slab_wasted_vram -= get_slab_wasted_size(ws, bo);
674   else
675      ws->slab_wasted_gtt -= get_slab_wasted_size(ws, bo);
676
677   pb_slab_free(slabs, &bo->u.slab.entry);
678}
679
680static const struct pb_vtbl amdgpu_winsys_bo_slab_vtbl = {
681   /* Cast to void* because one of the function parameters is a struct pointer instead of void*. */
682   (void*)amdgpu_bo_slab_destroy
683   /* other functions are never called */
684};
685
686/* Return the power of two size of a slab entry matching the input size. */
687static unsigned get_slab_pot_entry_size(struct amdgpu_winsys *ws, unsigned size)
688{
689   unsigned entry_size = util_next_power_of_two(size);
690   unsigned min_entry_size = 1 << ws->bo_slabs[0].min_order;
691
692   return MAX2(entry_size, min_entry_size);
693}
694
695/* Return the slab entry alignment. */
696static unsigned get_slab_entry_alignment(struct amdgpu_winsys *ws, unsigned size)
697{
698   unsigned entry_size = get_slab_pot_entry_size(ws, size);
699
700   if (size <= entry_size * 3 / 4)
701      return entry_size / 4;
702
703   return entry_size;
704}
705
706struct pb_slab *amdgpu_bo_slab_alloc(void *priv, unsigned heap, unsigned entry_size,
707                                     unsigned group_index)
708{
709   struct amdgpu_winsys *ws = priv;
710   struct amdgpu_slab *slab = CALLOC_STRUCT(amdgpu_slab);
711   enum radeon_bo_domain domains = radeon_domain_from_heap(heap);
712   enum radeon_bo_flag flags = radeon_flags_from_heap(heap);
713   uint32_t base_id;
714   unsigned slab_size = 0;
715
716   if (!slab)
717      return NULL;
718
719   /* Determine the slab buffer size. */
720   for (unsigned i = 0; i < NUM_SLAB_ALLOCATORS; i++) {
721      unsigned max_entry_size = 1 << (ws->bo_slabs[i].min_order + ws->bo_slabs[i].num_orders - 1);
722
723      if (entry_size <= max_entry_size) {
724         /* The slab size is twice the size of the largest possible entry. */
725         slab_size = max_entry_size * 2;
726
727         if (!util_is_power_of_two_nonzero(entry_size)) {
728            assert(util_is_power_of_two_nonzero(entry_size * 4 / 3));
729
730            /* If the entry size is 3/4 of a power of two, we would waste space and not gain
731             * anything if we allocated only twice the power of two for the backing buffer:
732             *   2 * 3/4 = 1.5 usable with buffer size 2
733             *
734             * Allocating 5 times the entry size leads us to the next power of two and results
735             * in a much better memory utilization:
736             *   5 * 3/4 = 3.75 usable with buffer size 4
737             */
738            if (entry_size * 5 > slab_size)
739               slab_size = util_next_power_of_two(entry_size * 5);
740         }
741
742         /* The largest slab should have the same size as the PTE fragment
743          * size to get faster address translation.
744          */
745         if (i == NUM_SLAB_ALLOCATORS - 1 &&
746             slab_size < ws->info.pte_fragment_size)
747            slab_size = ws->info.pte_fragment_size;
748         break;
749      }
750   }
751   assert(slab_size != 0);
752
753   slab->buffer = amdgpu_winsys_bo(amdgpu_bo_create(ws,
754                                                    slab_size, slab_size,
755                                                    domains, flags));
756   if (!slab->buffer)
757      goto fail;
758
759   slab_size = slab->buffer->base.size;
760
761   slab->base.num_entries = slab_size / entry_size;
762   slab->base.num_free = slab->base.num_entries;
763   slab->entry_size = entry_size;
764   slab->entries = CALLOC(slab->base.num_entries, sizeof(*slab->entries));
765   if (!slab->entries)
766      goto fail_buffer;
767
768   list_inithead(&slab->base.free);
769
770   base_id = __sync_fetch_and_add(&ws->next_bo_unique_id, slab->base.num_entries);
771
772   for (unsigned i = 0; i < slab->base.num_entries; ++i) {
773      struct amdgpu_winsys_bo *bo = &slab->entries[i];
774
775      simple_mtx_init(&bo->lock, mtx_plain);
776      bo->base.alignment_log2 = util_logbase2(get_slab_entry_alignment(ws, entry_size));
777      bo->base.size = entry_size;
778      bo->base.vtbl = &amdgpu_winsys_bo_slab_vtbl;
779      bo->va = slab->buffer->va + i * entry_size;
780      bo->base.placement = domains;
781      bo->unique_id = base_id + i;
782      bo->u.slab.entry.slab = &slab->base;
783      bo->u.slab.entry.group_index = group_index;
784      bo->u.slab.entry.entry_size = entry_size;
785
786      if (slab->buffer->bo) {
787         /* The slab is not suballocated. */
788         bo->u.slab.real = slab->buffer;
789      } else {
790         /* The slab is allocated out of a bigger slab. */
791         bo->u.slab.real = slab->buffer->u.slab.real;
792         assert(bo->u.slab.real->bo);
793      }
794
795      list_addtail(&bo->u.slab.entry.head, &slab->base.free);
796   }
797
798   /* Wasted alignment due to slabs with 3/4 allocations being aligned to a power of two. */
799   assert(slab->base.num_entries * entry_size <= slab_size);
800   if (domains & RADEON_DOMAIN_VRAM)
801      ws->slab_wasted_vram += slab_size - slab->base.num_entries * entry_size;
802   else
803      ws->slab_wasted_gtt += slab_size - slab->base.num_entries * entry_size;
804
805   return &slab->base;
806
807fail_buffer:
808   amdgpu_winsys_bo_reference(ws, &slab->buffer, NULL);
809fail:
810   FREE(slab);
811   return NULL;
812}
813
814void amdgpu_bo_slab_free(struct amdgpu_winsys *ws, struct pb_slab *pslab)
815{
816   struct amdgpu_slab *slab = amdgpu_slab(pslab);
817   unsigned slab_size = slab->buffer->base.size;
818
819   assert(slab->base.num_entries * slab->entry_size <= slab_size);
820   if (slab->buffer->base.placement & RADEON_DOMAIN_VRAM)
821      ws->slab_wasted_vram -= slab_size - slab->base.num_entries * slab->entry_size;
822   else
823      ws->slab_wasted_gtt -= slab_size - slab->base.num_entries * slab->entry_size;
824
825   for (unsigned i = 0; i < slab->base.num_entries; ++i) {
826      amdgpu_bo_remove_fences(&slab->entries[i]);
827      simple_mtx_destroy(&slab->entries[i].lock);
828   }
829
830   FREE(slab->entries);
831   amdgpu_winsys_bo_reference(ws, &slab->buffer, NULL);
832   FREE(slab);
833}
834
835#if DEBUG_SPARSE_COMMITS
836static void
837sparse_dump(struct amdgpu_winsys_bo *bo, const char *func)
838{
839   fprintf(stderr, "%s: %p (size=%"PRIu64", num_va_pages=%u) @ %s\n"
840                   "Commitments:\n",
841           __func__, bo, bo->base.size, bo->u.sparse.num_va_pages, func);
842
843   struct amdgpu_sparse_backing *span_backing = NULL;
844   uint32_t span_first_backing_page = 0;
845   uint32_t span_first_va_page = 0;
846   uint32_t va_page = 0;
847
848   for (;;) {
849      struct amdgpu_sparse_backing *backing = 0;
850      uint32_t backing_page = 0;
851
852      if (va_page < bo->u.sparse.num_va_pages) {
853         backing = bo->u.sparse.commitments[va_page].backing;
854         backing_page = bo->u.sparse.commitments[va_page].page;
855      }
856
857      if (span_backing &&
858          (backing != span_backing ||
859           backing_page != span_first_backing_page + (va_page - span_first_va_page))) {
860         fprintf(stderr, " %u..%u: backing=%p:%u..%u\n",
861                 span_first_va_page, va_page - 1, span_backing,
862                 span_first_backing_page,
863                 span_first_backing_page + (va_page - span_first_va_page) - 1);
864
865         span_backing = NULL;
866      }
867
868      if (va_page >= bo->u.sparse.num_va_pages)
869         break;
870
871      if (backing && !span_backing) {
872         span_backing = backing;
873         span_first_backing_page = backing_page;
874         span_first_va_page = va_page;
875      }
876
877      va_page++;
878   }
879
880   fprintf(stderr, "Backing:\n");
881
882   list_for_each_entry(struct amdgpu_sparse_backing, backing, &bo->u.sparse.backing, list) {
883      fprintf(stderr, " %p (size=%"PRIu64")\n", backing, backing->bo->base.size);
884      for (unsigned i = 0; i < backing->num_chunks; ++i)
885         fprintf(stderr, "   %u..%u\n", backing->chunks[i].begin, backing->chunks[i].end);
886   }
887}
888#endif
889
890/*
891 * Attempt to allocate the given number of backing pages. Fewer pages may be
892 * allocated (depending on the fragmentation of existing backing buffers),
893 * which will be reflected by a change to *pnum_pages.
894 */
895static struct amdgpu_sparse_backing *
896sparse_backing_alloc(struct amdgpu_winsys *ws, struct amdgpu_winsys_bo *bo,
897                     uint32_t *pstart_page, uint32_t *pnum_pages)
898{
899   struct amdgpu_sparse_backing *best_backing;
900   unsigned best_idx;
901   uint32_t best_num_pages;
902
903   best_backing = NULL;
904   best_idx = 0;
905   best_num_pages = 0;
906
907   /* This is a very simple and inefficient best-fit algorithm. */
908   list_for_each_entry(struct amdgpu_sparse_backing, backing, &bo->u.sparse.backing, list) {
909      for (unsigned idx = 0; idx < backing->num_chunks; ++idx) {
910         uint32_t cur_num_pages = backing->chunks[idx].end - backing->chunks[idx].begin;
911         if ((best_num_pages < *pnum_pages && cur_num_pages > best_num_pages) ||
912            (best_num_pages > *pnum_pages && cur_num_pages < best_num_pages)) {
913            best_backing = backing;
914            best_idx = idx;
915            best_num_pages = cur_num_pages;
916         }
917      }
918   }
919
920   /* Allocate a new backing buffer if necessary. */
921   if (!best_backing) {
922      struct pb_buffer *buf;
923      uint64_t size;
924      uint32_t pages;
925
926      best_backing = CALLOC_STRUCT(amdgpu_sparse_backing);
927      if (!best_backing)
928         return NULL;
929
930      best_backing->max_chunks = 4;
931      best_backing->chunks = CALLOC(best_backing->max_chunks,
932                                    sizeof(*best_backing->chunks));
933      if (!best_backing->chunks) {
934         FREE(best_backing);
935         return NULL;
936      }
937
938      assert(bo->u.sparse.num_backing_pages < DIV_ROUND_UP(bo->base.size, RADEON_SPARSE_PAGE_SIZE));
939
940      size = MIN3(bo->base.size / 16,
941                  8 * 1024 * 1024,
942                  bo->base.size - (uint64_t)bo->u.sparse.num_backing_pages * RADEON_SPARSE_PAGE_SIZE);
943      size = MAX2(size, RADEON_SPARSE_PAGE_SIZE);
944
945      buf = amdgpu_bo_create(ws, size, RADEON_SPARSE_PAGE_SIZE,
946                             bo->base.placement,
947                             (bo->base.usage & ~RADEON_FLAG_SPARSE &
948                              /* Set the interprocess sharing flag to disable pb_cache because
949                               * amdgpu_bo_wait doesn't wait for active CS jobs.
950                               */
951                              ~RADEON_FLAG_NO_INTERPROCESS_SHARING) | RADEON_FLAG_NO_SUBALLOC);
952      if (!buf) {
953         FREE(best_backing->chunks);
954         FREE(best_backing);
955         return NULL;
956      }
957
958      /* We might have gotten a bigger buffer than requested via caching. */
959      pages = buf->size / RADEON_SPARSE_PAGE_SIZE;
960
961      best_backing->bo = amdgpu_winsys_bo(buf);
962      best_backing->num_chunks = 1;
963      best_backing->chunks[0].begin = 0;
964      best_backing->chunks[0].end = pages;
965
966      list_add(&best_backing->list, &bo->u.sparse.backing);
967      bo->u.sparse.num_backing_pages += pages;
968
969      best_idx = 0;
970      best_num_pages = pages;
971   }
972
973   *pnum_pages = MIN2(*pnum_pages, best_num_pages);
974   *pstart_page = best_backing->chunks[best_idx].begin;
975   best_backing->chunks[best_idx].begin += *pnum_pages;
976
977   if (best_backing->chunks[best_idx].begin >= best_backing->chunks[best_idx].end) {
978      memmove(&best_backing->chunks[best_idx], &best_backing->chunks[best_idx + 1],
979              sizeof(*best_backing->chunks) * (best_backing->num_chunks - best_idx - 1));
980      best_backing->num_chunks--;
981   }
982
983   return best_backing;
984}
985
986static void
987sparse_free_backing_buffer(struct amdgpu_winsys *ws, struct amdgpu_winsys_bo *bo,
988                           struct amdgpu_sparse_backing *backing)
989{
990   bo->u.sparse.num_backing_pages -= backing->bo->base.size / RADEON_SPARSE_PAGE_SIZE;
991
992   simple_mtx_lock(&ws->bo_fence_lock);
993   amdgpu_add_fences(backing->bo, bo->num_fences, bo->fences);
994   simple_mtx_unlock(&ws->bo_fence_lock);
995
996   list_del(&backing->list);
997   amdgpu_winsys_bo_reference(ws, &backing->bo, NULL);
998   FREE(backing->chunks);
999   FREE(backing);
1000}
1001
1002/*
1003 * Return a range of pages from the given backing buffer back into the
1004 * free structure.
1005 */
1006static bool
1007sparse_backing_free(struct amdgpu_winsys *ws, struct amdgpu_winsys_bo *bo,
1008                    struct amdgpu_sparse_backing *backing,
1009                    uint32_t start_page, uint32_t num_pages)
1010{
1011   uint32_t end_page = start_page + num_pages;
1012   unsigned low = 0;
1013   unsigned high = backing->num_chunks;
1014
1015   /* Find the first chunk with begin >= start_page. */
1016   while (low < high) {
1017      unsigned mid = low + (high - low) / 2;
1018
1019      if (backing->chunks[mid].begin >= start_page)
1020         high = mid;
1021      else
1022         low = mid + 1;
1023   }
1024
1025   assert(low >= backing->num_chunks || end_page <= backing->chunks[low].begin);
1026   assert(low == 0 || backing->chunks[low - 1].end <= start_page);
1027
1028   if (low > 0 && backing->chunks[low - 1].end == start_page) {
1029      backing->chunks[low - 1].end = end_page;
1030
1031      if (low < backing->num_chunks && end_page == backing->chunks[low].begin) {
1032         backing->chunks[low - 1].end = backing->chunks[low].end;
1033         memmove(&backing->chunks[low], &backing->chunks[low + 1],
1034                 sizeof(*backing->chunks) * (backing->num_chunks - low - 1));
1035         backing->num_chunks--;
1036      }
1037   } else if (low < backing->num_chunks && end_page == backing->chunks[low].begin) {
1038      backing->chunks[low].begin = start_page;
1039   } else {
1040      if (backing->num_chunks >= backing->max_chunks) {
1041         unsigned new_max_chunks = 2 * backing->max_chunks;
1042         struct amdgpu_sparse_backing_chunk *new_chunks =
1043            REALLOC(backing->chunks,
1044                    sizeof(*backing->chunks) * backing->max_chunks,
1045                    sizeof(*backing->chunks) * new_max_chunks);
1046         if (!new_chunks)
1047            return false;
1048
1049         backing->max_chunks = new_max_chunks;
1050         backing->chunks = new_chunks;
1051      }
1052
1053      memmove(&backing->chunks[low + 1], &backing->chunks[low],
1054              sizeof(*backing->chunks) * (backing->num_chunks - low));
1055      backing->chunks[low].begin = start_page;
1056      backing->chunks[low].end = end_page;
1057      backing->num_chunks++;
1058   }
1059
1060   if (backing->num_chunks == 1 && backing->chunks[0].begin == 0 &&
1061       backing->chunks[0].end == backing->bo->base.size / RADEON_SPARSE_PAGE_SIZE)
1062      sparse_free_backing_buffer(ws, bo, backing);
1063
1064   return true;
1065}
1066
1067static void amdgpu_bo_sparse_destroy(struct radeon_winsys *rws, struct pb_buffer *_buf)
1068{
1069   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
1070   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
1071   int r;
1072
1073   assert(!bo->bo && bo->base.usage & RADEON_FLAG_SPARSE);
1074
1075   r = amdgpu_bo_va_op_raw(ws->dev, NULL, 0,
1076                           (uint64_t)bo->u.sparse.num_va_pages * RADEON_SPARSE_PAGE_SIZE,
1077                           bo->va, 0, AMDGPU_VA_OP_CLEAR);
1078   if (r) {
1079      fprintf(stderr, "amdgpu: clearing PRT VA region on destroy failed (%d)\n", r);
1080   }
1081
1082   while (!list_is_empty(&bo->u.sparse.backing)) {
1083      sparse_free_backing_buffer(ws, bo,
1084                                 container_of(bo->u.sparse.backing.next,
1085                                              struct amdgpu_sparse_backing, list));
1086   }
1087
1088   amdgpu_va_range_free(bo->u.sparse.va_handle);
1089   FREE(bo->u.sparse.commitments);
1090   simple_mtx_destroy(&bo->lock);
1091   FREE(bo);
1092}
1093
1094static const struct pb_vtbl amdgpu_winsys_bo_sparse_vtbl = {
1095   /* Cast to void* because one of the function parameters is a struct pointer instead of void*. */
1096   (void*)amdgpu_bo_sparse_destroy
1097   /* other functions are never called */
1098};
1099
1100static struct pb_buffer *
1101amdgpu_bo_sparse_create(struct amdgpu_winsys *ws, uint64_t size,
1102                        enum radeon_bo_domain domain,
1103                        enum radeon_bo_flag flags)
1104{
1105   struct amdgpu_winsys_bo *bo;
1106   uint64_t map_size;
1107   uint64_t va_gap_size;
1108   int r;
1109
1110   /* We use 32-bit page numbers; refuse to attempt allocating sparse buffers
1111    * that exceed this limit. This is not really a restriction: we don't have
1112    * that much virtual address space anyway.
1113    */
1114   if (size > (uint64_t)INT32_MAX * RADEON_SPARSE_PAGE_SIZE)
1115      return NULL;
1116
1117   bo = CALLOC_STRUCT(amdgpu_winsys_bo);
1118   if (!bo)
1119      return NULL;
1120
1121   simple_mtx_init(&bo->lock, mtx_plain);
1122   pipe_reference_init(&bo->base.reference, 1);
1123   bo->base.alignment_log2 = util_logbase2(RADEON_SPARSE_PAGE_SIZE);
1124   bo->base.size = size;
1125   bo->base.vtbl = &amdgpu_winsys_bo_sparse_vtbl;
1126   bo->base.placement = domain;
1127   bo->unique_id =  __sync_fetch_and_add(&ws->next_bo_unique_id, 1);
1128   bo->base.usage = flags;
1129
1130   bo->u.sparse.num_va_pages = DIV_ROUND_UP(size, RADEON_SPARSE_PAGE_SIZE);
1131   bo->u.sparse.commitments = CALLOC(bo->u.sparse.num_va_pages,
1132                                     sizeof(*bo->u.sparse.commitments));
1133   if (!bo->u.sparse.commitments)
1134      goto error_alloc_commitments;
1135
1136   list_inithead(&bo->u.sparse.backing);
1137
1138   /* For simplicity, we always map a multiple of the page size. */
1139   map_size = align64(size, RADEON_SPARSE_PAGE_SIZE);
1140   va_gap_size = ws->check_vm ? 4 * RADEON_SPARSE_PAGE_SIZE : 0;
1141   r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
1142                             map_size + va_gap_size, RADEON_SPARSE_PAGE_SIZE,
1143                             0, &bo->va, &bo->u.sparse.va_handle,
1144			     AMDGPU_VA_RANGE_HIGH);
1145   if (r)
1146      goto error_va_alloc;
1147
1148   r = amdgpu_bo_va_op_raw(ws->dev, NULL, 0, map_size, bo->va,
1149                           AMDGPU_VM_PAGE_PRT, AMDGPU_VA_OP_MAP);
1150   if (r)
1151      goto error_va_map;
1152
1153   return &bo->base;
1154
1155error_va_map:
1156   amdgpu_va_range_free(bo->u.sparse.va_handle);
1157error_va_alloc:
1158   FREE(bo->u.sparse.commitments);
1159error_alloc_commitments:
1160   simple_mtx_destroy(&bo->lock);
1161   FREE(bo);
1162   return NULL;
1163}
1164
1165static bool
1166amdgpu_bo_sparse_commit(struct radeon_winsys *rws, struct pb_buffer *buf,
1167                        uint64_t offset, uint64_t size, bool commit)
1168{
1169   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
1170   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(buf);
1171   struct amdgpu_sparse_commitment *comm;
1172   uint32_t va_page, end_va_page;
1173   bool ok = true;
1174   int r;
1175
1176   assert(bo->base.usage & RADEON_FLAG_SPARSE);
1177   assert(offset % RADEON_SPARSE_PAGE_SIZE == 0);
1178   assert(offset <= bo->base.size);
1179   assert(size <= bo->base.size - offset);
1180   assert(size % RADEON_SPARSE_PAGE_SIZE == 0 || offset + size == bo->base.size);
1181
1182   comm = bo->u.sparse.commitments;
1183   va_page = offset / RADEON_SPARSE_PAGE_SIZE;
1184   end_va_page = va_page + DIV_ROUND_UP(size, RADEON_SPARSE_PAGE_SIZE);
1185
1186   simple_mtx_lock(&bo->lock);
1187
1188#if DEBUG_SPARSE_COMMITS
1189   sparse_dump(bo, __func__);
1190#endif
1191
1192   if (commit) {
1193      while (va_page < end_va_page) {
1194         uint32_t span_va_page;
1195
1196         /* Skip pages that are already committed. */
1197         if (comm[va_page].backing) {
1198            va_page++;
1199            continue;
1200         }
1201
1202         /* Determine length of uncommitted span. */
1203         span_va_page = va_page;
1204         while (va_page < end_va_page && !comm[va_page].backing)
1205            va_page++;
1206
1207         /* Fill the uncommitted span with chunks of backing memory. */
1208         while (span_va_page < va_page) {
1209            struct amdgpu_sparse_backing *backing;
1210            uint32_t backing_start, backing_size;
1211
1212            backing_size = va_page - span_va_page;
1213            backing = sparse_backing_alloc(ws, bo, &backing_start, &backing_size);
1214            if (!backing) {
1215               ok = false;
1216               goto out;
1217            }
1218
1219            r = amdgpu_bo_va_op_raw(ws->dev, backing->bo->bo,
1220                                    (uint64_t)backing_start * RADEON_SPARSE_PAGE_SIZE,
1221                                    (uint64_t)backing_size * RADEON_SPARSE_PAGE_SIZE,
1222                                    bo->va + (uint64_t)span_va_page * RADEON_SPARSE_PAGE_SIZE,
1223                                    AMDGPU_VM_PAGE_READABLE |
1224                                    AMDGPU_VM_PAGE_WRITEABLE |
1225                                    AMDGPU_VM_PAGE_EXECUTABLE,
1226                                    AMDGPU_VA_OP_REPLACE);
1227            if (r) {
1228               ok = sparse_backing_free(ws, bo, backing, backing_start, backing_size);
1229               assert(ok && "sufficient memory should already be allocated");
1230
1231               ok = false;
1232               goto out;
1233            }
1234
1235            while (backing_size) {
1236               comm[span_va_page].backing = backing;
1237               comm[span_va_page].page = backing_start;
1238               span_va_page++;
1239               backing_start++;
1240               backing_size--;
1241            }
1242         }
1243      }
1244   } else {
1245      r = amdgpu_bo_va_op_raw(ws->dev, NULL, 0,
1246                              (uint64_t)(end_va_page - va_page) * RADEON_SPARSE_PAGE_SIZE,
1247                              bo->va + (uint64_t)va_page * RADEON_SPARSE_PAGE_SIZE,
1248                              AMDGPU_VM_PAGE_PRT, AMDGPU_VA_OP_REPLACE);
1249      if (r) {
1250         ok = false;
1251         goto out;
1252      }
1253
1254      while (va_page < end_va_page) {
1255         struct amdgpu_sparse_backing *backing;
1256         uint32_t backing_start;
1257         uint32_t span_pages;
1258
1259         /* Skip pages that are already uncommitted. */
1260         if (!comm[va_page].backing) {
1261            va_page++;
1262            continue;
1263         }
1264
1265         /* Group contiguous spans of pages. */
1266         backing = comm[va_page].backing;
1267         backing_start = comm[va_page].page;
1268         comm[va_page].backing = NULL;
1269
1270         span_pages = 1;
1271         va_page++;
1272
1273         while (va_page < end_va_page &&
1274                comm[va_page].backing == backing &&
1275                comm[va_page].page == backing_start + span_pages) {
1276            comm[va_page].backing = NULL;
1277            va_page++;
1278            span_pages++;
1279         }
1280
1281         if (!sparse_backing_free(ws, bo, backing, backing_start, span_pages)) {
1282            /* Couldn't allocate tracking data structures, so we have to leak */
1283            fprintf(stderr, "amdgpu: leaking PRT backing memory\n");
1284            ok = false;
1285         }
1286      }
1287   }
1288out:
1289
1290   simple_mtx_unlock(&bo->lock);
1291
1292   return ok;
1293}
1294
1295static void amdgpu_buffer_get_metadata(struct radeon_winsys *rws,
1296                                       struct pb_buffer *_buf,
1297                                       struct radeon_bo_metadata *md,
1298                                       struct radeon_surf *surf)
1299{
1300   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
1301   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
1302   struct amdgpu_bo_info info = {0};
1303   int r;
1304
1305   assert(bo->bo && "must not be called for slab entries");
1306
1307   r = amdgpu_bo_query_info(bo->bo, &info);
1308   if (r)
1309      return;
1310
1311   ac_surface_set_bo_metadata(&ws->info, surf, info.metadata.tiling_info,
1312                              &md->mode);
1313
1314   md->size_metadata = info.metadata.size_metadata;
1315   memcpy(md->metadata, info.metadata.umd_metadata, sizeof(md->metadata));
1316}
1317
1318static void amdgpu_buffer_set_metadata(struct radeon_winsys *rws,
1319                                       struct pb_buffer *_buf,
1320                                       struct radeon_bo_metadata *md,
1321                                       struct radeon_surf *surf)
1322{
1323   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
1324   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(_buf);
1325   struct amdgpu_bo_metadata metadata = {0};
1326
1327   assert(bo->bo && "must not be called for slab entries");
1328
1329   ac_surface_get_bo_metadata(&ws->info, surf, &metadata.tiling_info);
1330
1331   metadata.size_metadata = md->size_metadata;
1332   memcpy(metadata.umd_metadata, md->metadata, sizeof(md->metadata));
1333
1334   amdgpu_bo_set_metadata(bo->bo, &metadata);
1335}
1336
1337struct pb_buffer *
1338amdgpu_bo_create(struct amdgpu_winsys *ws,
1339                 uint64_t size,
1340                 unsigned alignment,
1341                 enum radeon_bo_domain domain,
1342                 enum radeon_bo_flag flags)
1343{
1344   struct amdgpu_winsys_bo *bo;
1345
1346   radeon_canonicalize_bo_flags(&domain, &flags);
1347
1348   /* Handle sparse buffers first. */
1349   if (flags & RADEON_FLAG_SPARSE) {
1350      assert(RADEON_SPARSE_PAGE_SIZE % alignment == 0);
1351
1352      return amdgpu_bo_sparse_create(ws, size, domain, flags);
1353   }
1354
1355   struct pb_slabs *last_slab = &ws->bo_slabs[NUM_SLAB_ALLOCATORS - 1];
1356   unsigned max_slab_entry_size = 1 << (last_slab->min_order + last_slab->num_orders - 1);
1357   int heap = radeon_get_heap_index(domain, flags);
1358
1359   /* Sub-allocate small buffers from slabs. */
1360   if (heap >= 0 && size <= max_slab_entry_size) {
1361      struct pb_slab_entry *entry;
1362      unsigned alloc_size = size;
1363
1364      /* Always use slabs for sizes less than 4 KB because the kernel aligns
1365       * everything to 4 KB.
1366       */
1367      if (size < alignment && alignment <= 4 * 1024)
1368         alloc_size = alignment;
1369
1370      if (alignment > get_slab_entry_alignment(ws, alloc_size)) {
1371         /* 3/4 allocations can return too small alignment. Try again with a power of two
1372          * allocation size.
1373          */
1374         unsigned pot_size = get_slab_pot_entry_size(ws, alloc_size);
1375
1376         if (alignment <= pot_size) {
1377            /* This size works but wastes some memory to fulfil the alignment. */
1378            alloc_size = pot_size;
1379         } else {
1380            goto no_slab; /* can't fulfil alignment requirements */
1381         }
1382      }
1383
1384      struct pb_slabs *slabs = get_slabs(ws, alloc_size);
1385      entry = pb_slab_alloc(slabs, alloc_size, heap);
1386      if (!entry) {
1387         /* Clean up buffer managers and try again. */
1388         amdgpu_clean_up_buffer_managers(ws);
1389
1390         entry = pb_slab_alloc(slabs, alloc_size, heap);
1391      }
1392      if (!entry)
1393         return NULL;
1394
1395      bo = container_of(entry, struct amdgpu_winsys_bo, u.slab.entry);
1396      pipe_reference_init(&bo->base.reference, 1);
1397      bo->base.size = size;
1398      assert(alignment <= 1 << bo->base.alignment_log2);
1399
1400      if (domain & RADEON_DOMAIN_VRAM)
1401         ws->slab_wasted_vram += get_slab_wasted_size(ws, bo);
1402      else
1403         ws->slab_wasted_gtt += get_slab_wasted_size(ws, bo);
1404
1405      return &bo->base;
1406   }
1407no_slab:
1408
1409   /* Align size to page size. This is the minimum alignment for normal
1410    * BOs. Aligning this here helps the cached bufmgr. Especially small BOs,
1411    * like constant/uniform buffers, can benefit from better and more reuse.
1412    */
1413   if (domain & RADEON_DOMAIN_VRAM_GTT) {
1414      size = align64(size, ws->info.gart_page_size);
1415      alignment = align(alignment, ws->info.gart_page_size);
1416   }
1417
1418   bool use_reusable_pool = flags & RADEON_FLAG_NO_INTERPROCESS_SHARING &&
1419                            !(flags & RADEON_FLAG_DISCARDABLE);
1420
1421   if (use_reusable_pool) {
1422       /* RADEON_FLAG_NO_SUBALLOC is irrelevant for the cache. */
1423       heap = radeon_get_heap_index(domain, flags & ~RADEON_FLAG_NO_SUBALLOC);
1424       assert(heap >= 0 && heap < RADEON_NUM_HEAPS);
1425
1426       /* Get a buffer from the cache. */
1427       bo = (struct amdgpu_winsys_bo*)
1428            pb_cache_reclaim_buffer(&ws->bo_cache, size, alignment, 0, heap);
1429       if (bo)
1430          return &bo->base;
1431   }
1432
1433   /* Create a new one. */
1434   bo = amdgpu_create_bo(ws, size, alignment, domain, flags, heap);
1435   if (!bo) {
1436      /* Clean up buffer managers and try again. */
1437      amdgpu_clean_up_buffer_managers(ws);
1438
1439      bo = amdgpu_create_bo(ws, size, alignment, domain, flags, heap);
1440      if (!bo)
1441         return NULL;
1442   }
1443
1444   return &bo->base;
1445}
1446
1447static struct pb_buffer *
1448amdgpu_buffer_create(struct radeon_winsys *ws,
1449                     uint64_t size,
1450                     unsigned alignment,
1451                     enum radeon_bo_domain domain,
1452                     enum radeon_bo_flag flags)
1453{
1454   struct pb_buffer * res = amdgpu_bo_create(amdgpu_winsys(ws), size, alignment, domain,
1455                           flags);
1456   return res;
1457}
1458
1459static struct pb_buffer *amdgpu_bo_from_handle(struct radeon_winsys *rws,
1460                                               struct winsys_handle *whandle,
1461                                               unsigned vm_alignment,
1462                                               bool is_prime_linear_buffer)
1463{
1464   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
1465   struct amdgpu_winsys_bo *bo = NULL;
1466   enum amdgpu_bo_handle_type type;
1467   struct amdgpu_bo_import_result result = {0};
1468   uint64_t va;
1469   amdgpu_va_handle va_handle = NULL;
1470   struct amdgpu_bo_info info = {0};
1471   enum radeon_bo_domain initial = 0;
1472   enum radeon_bo_flag flags = 0;
1473   int r;
1474
1475   switch (whandle->type) {
1476   case WINSYS_HANDLE_TYPE_SHARED:
1477      type = amdgpu_bo_handle_type_gem_flink_name;
1478      break;
1479   case WINSYS_HANDLE_TYPE_FD:
1480      type = amdgpu_bo_handle_type_dma_buf_fd;
1481      break;
1482   default:
1483      return NULL;
1484   }
1485
1486   r = amdgpu_bo_import(ws->dev, type, whandle->handle, &result);
1487   if (r)
1488      return NULL;
1489
1490   simple_mtx_lock(&ws->bo_export_table_lock);
1491   bo = util_hash_table_get(ws->bo_export_table, result.buf_handle);
1492
1493   /* If the amdgpu_winsys_bo instance already exists, bump the reference
1494    * counter and return it.
1495    */
1496   if (bo) {
1497      p_atomic_inc(&bo->base.reference.count);
1498      simple_mtx_unlock(&ws->bo_export_table_lock);
1499
1500      /* Release the buffer handle, because we don't need it anymore.
1501       * This function is returning an existing buffer, which has its own
1502       * handle.
1503       */
1504      amdgpu_bo_free(result.buf_handle);
1505      return &bo->base;
1506   }
1507
1508   /* Get initial domains. */
1509   r = amdgpu_bo_query_info(result.buf_handle, &info);
1510   if (r)
1511      goto error;
1512
1513   r = amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
1514                             result.alloc_size,
1515                             amdgpu_get_optimal_alignment(ws, result.alloc_size,
1516                                                          vm_alignment),
1517                             0, &va, &va_handle, AMDGPU_VA_RANGE_HIGH);
1518   if (r)
1519      goto error;
1520
1521   bo = CALLOC_STRUCT(amdgpu_winsys_bo);
1522   if (!bo)
1523      goto error;
1524
1525   r = amdgpu_bo_va_op_raw(ws->dev, result.buf_handle, 0, result.alloc_size, va,
1526                           AMDGPU_VM_PAGE_READABLE | AMDGPU_VM_PAGE_WRITEABLE |
1527                           AMDGPU_VM_PAGE_EXECUTABLE |
1528                           (is_prime_linear_buffer ? AMDGPU_VM_MTYPE_UC : 0),
1529                           AMDGPU_VA_OP_MAP);
1530   if (r)
1531      goto error;
1532
1533   if (info.preferred_heap & AMDGPU_GEM_DOMAIN_VRAM)
1534      initial |= RADEON_DOMAIN_VRAM;
1535   if (info.preferred_heap & AMDGPU_GEM_DOMAIN_GTT)
1536      initial |= RADEON_DOMAIN_GTT;
1537   if (info.alloc_flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
1538      flags |= RADEON_FLAG_NO_CPU_ACCESS;
1539   if (info.alloc_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
1540      flags |= RADEON_FLAG_GTT_WC;
1541   if (info.alloc_flags & AMDGPU_GEM_CREATE_ENCRYPTED) {
1542      /* Imports are always possible even if the importer isn't using TMZ.
1543       * For instance libweston needs to import the buffer to be able to determine
1544       * if it can be used for scanout.
1545       */
1546      flags |= RADEON_FLAG_ENCRYPTED;
1547      *((bool*)&rws->uses_secure_bos) = true;
1548   }
1549
1550   /* Initialize the structure. */
1551   simple_mtx_init(&bo->lock, mtx_plain);
1552   pipe_reference_init(&bo->base.reference, 1);
1553   bo->base.alignment_log2 = util_logbase2(info.phys_alignment);
1554   bo->bo = result.buf_handle;
1555   bo->base.size = result.alloc_size;
1556   bo->base.vtbl = &amdgpu_winsys_bo_vtbl;
1557   bo->va = va;
1558   bo->u.real.va_handle = va_handle;
1559   bo->base.placement = initial;
1560   bo->base.usage = flags;
1561   bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1);
1562   bo->u.real.is_shared = true;
1563
1564   if (bo->base.placement & RADEON_DOMAIN_VRAM)
1565      ws->allocated_vram += align64(bo->base.size, ws->info.gart_page_size);
1566   else if (bo->base.placement & RADEON_DOMAIN_GTT)
1567      ws->allocated_gtt += align64(bo->base.size, ws->info.gart_page_size);
1568
1569   amdgpu_bo_export(bo->bo, amdgpu_bo_handle_type_kms, &bo->u.real.kms_handle);
1570
1571   amdgpu_add_buffer_to_global_list(ws, bo);
1572
1573   _mesa_hash_table_insert(ws->bo_export_table, bo->bo, bo);
1574   simple_mtx_unlock(&ws->bo_export_table_lock);
1575
1576   return &bo->base;
1577
1578error:
1579   simple_mtx_unlock(&ws->bo_export_table_lock);
1580   if (bo)
1581      FREE(bo);
1582   if (va_handle)
1583      amdgpu_va_range_free(va_handle);
1584   amdgpu_bo_free(result.buf_handle);
1585   return NULL;
1586}
1587
1588static bool amdgpu_bo_get_handle(struct radeon_winsys *rws,
1589                                 struct pb_buffer *buffer,
1590                                 struct winsys_handle *whandle)
1591{
1592   struct amdgpu_screen_winsys *sws = amdgpu_screen_winsys(rws);
1593   struct amdgpu_winsys *ws = amdgpu_winsys(rws);
1594   struct amdgpu_winsys_bo *bo = amdgpu_winsys_bo(buffer);
1595   enum amdgpu_bo_handle_type type;
1596   struct hash_entry *entry;
1597   int r;
1598
1599   /* Don't allow exports of slab entries and sparse buffers. */
1600   if (!bo->bo)
1601      return false;
1602
1603   bo->u.real.use_reusable_pool = false;
1604
1605   switch (whandle->type) {
1606   case WINSYS_HANDLE_TYPE_SHARED:
1607      type = amdgpu_bo_handle_type_gem_flink_name;
1608      break;
1609   case WINSYS_HANDLE_TYPE_KMS:
1610      if (sws->fd == ws->fd) {
1611         whandle->handle = bo->u.real.kms_handle;
1612
1613         if (bo->u.real.is_shared)
1614            return true;
1615
1616         goto hash_table_set;
1617      }
1618
1619      simple_mtx_lock(&ws->sws_list_lock);
1620      entry = _mesa_hash_table_search(sws->kms_handles, bo);
1621      simple_mtx_unlock(&ws->sws_list_lock);
1622      if (entry) {
1623         whandle->handle = (uintptr_t)entry->data;
1624         return true;
1625      }
1626      FALLTHROUGH;
1627   case WINSYS_HANDLE_TYPE_FD:
1628      type = amdgpu_bo_handle_type_dma_buf_fd;
1629      break;
1630   default:
1631      return false;
1632   }
1633
1634   r = amdgpu_bo_export(bo->bo, type, &whandle->handle);
1635   if (r)
1636      return false;
1637
1638   if (whandle->type == WINSYS_HANDLE_TYPE_KMS) {
1639      int dma_fd = whandle->handle;
1640
1641      r = drmPrimeFDToHandle(sws->fd, dma_fd, &whandle->handle);
1642      close(dma_fd);
1643
1644      if (r)
1645         return false;
1646
1647      simple_mtx_lock(&ws->sws_list_lock);
1648      _mesa_hash_table_insert_pre_hashed(sws->kms_handles,
1649                                         bo->u.real.kms_handle, bo,
1650                                         (void*)(uintptr_t)whandle->handle);
1651      simple_mtx_unlock(&ws->sws_list_lock);
1652   }
1653
1654 hash_table_set:
1655   simple_mtx_lock(&ws->bo_export_table_lock);
1656   _mesa_hash_table_insert(ws->bo_export_table, bo->bo, bo);
1657   simple_mtx_unlock(&ws->bo_export_table_lock);
1658
1659   bo->u.real.is_shared = true;
1660   return true;
1661}
1662
1663static struct pb_buffer *amdgpu_bo_from_ptr(struct radeon_winsys *rws,
1664					    void *pointer, uint64_t size,
1665					    enum radeon_bo_flag flags)
1666{
1667    struct amdgpu_winsys *ws = amdgpu_winsys(rws);
1668    amdgpu_bo_handle buf_handle;
1669    struct amdgpu_winsys_bo *bo;
1670    uint64_t va;
1671    amdgpu_va_handle va_handle;
1672    /* Avoid failure when the size is not page aligned */
1673    uint64_t aligned_size = align64(size, ws->info.gart_page_size);
1674
1675    bo = CALLOC_STRUCT(amdgpu_winsys_bo);
1676    if (!bo)
1677        return NULL;
1678
1679    if (amdgpu_create_bo_from_user_mem(ws->dev, pointer,
1680                                       aligned_size, &buf_handle))
1681        goto error;
1682
1683    if (amdgpu_va_range_alloc(ws->dev, amdgpu_gpu_va_range_general,
1684                              aligned_size,
1685                              amdgpu_get_optimal_alignment(ws, aligned_size,
1686                                                           ws->info.gart_page_size),
1687                              0, &va, &va_handle, AMDGPU_VA_RANGE_HIGH))
1688        goto error_va_alloc;
1689
1690    if (amdgpu_bo_va_op(buf_handle, 0, aligned_size, va, 0, AMDGPU_VA_OP_MAP))
1691        goto error_va_map;
1692
1693    /* Initialize it. */
1694    bo->u.real.is_user_ptr = true;
1695    pipe_reference_init(&bo->base.reference, 1);
1696    simple_mtx_init(&bo->lock, mtx_plain);
1697    bo->bo = buf_handle;
1698    bo->base.alignment_log2 = 0;
1699    bo->base.size = size;
1700    bo->base.vtbl = &amdgpu_winsys_bo_vtbl;
1701    bo->u.real.cpu_ptr = pointer;
1702    bo->va = va;
1703    bo->u.real.va_handle = va_handle;
1704    bo->base.placement = RADEON_DOMAIN_GTT;
1705    bo->unique_id = __sync_fetch_and_add(&ws->next_bo_unique_id, 1);
1706
1707    ws->allocated_gtt += aligned_size;
1708
1709    amdgpu_add_buffer_to_global_list(ws, bo);
1710
1711    amdgpu_bo_export(bo->bo, amdgpu_bo_handle_type_kms, &bo->u.real.kms_handle);
1712
1713    return (struct pb_buffer*)bo;
1714
1715error_va_map:
1716    amdgpu_va_range_free(va_handle);
1717
1718error_va_alloc:
1719    amdgpu_bo_free(buf_handle);
1720
1721error:
1722    FREE(bo);
1723    return NULL;
1724}
1725
1726static bool amdgpu_bo_is_user_ptr(struct pb_buffer *buf)
1727{
1728   struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf;
1729
1730   return bo->bo ? bo->u.real.is_user_ptr : false;
1731}
1732
1733static bool amdgpu_bo_is_suballocated(struct pb_buffer *buf)
1734{
1735   struct amdgpu_winsys_bo *bo = (struct amdgpu_winsys_bo*)buf;
1736
1737   return !bo->bo && !(bo->base.usage & RADEON_FLAG_SPARSE);
1738}
1739
1740static uint64_t amdgpu_bo_get_va(struct pb_buffer *buf)
1741{
1742   return ((struct amdgpu_winsys_bo*)buf)->va;
1743}
1744
1745void amdgpu_bo_init_functions(struct amdgpu_screen_winsys *ws)
1746{
1747   ws->base.buffer_set_metadata = amdgpu_buffer_set_metadata;
1748   ws->base.buffer_get_metadata = amdgpu_buffer_get_metadata;
1749   ws->base.buffer_map = amdgpu_bo_map;
1750   ws->base.buffer_unmap = amdgpu_bo_unmap;
1751   ws->base.buffer_wait = amdgpu_bo_wait;
1752   ws->base.buffer_create = amdgpu_buffer_create;
1753   ws->base.buffer_from_handle = amdgpu_bo_from_handle;
1754   ws->base.buffer_from_ptr = amdgpu_bo_from_ptr;
1755   ws->base.buffer_is_user_ptr = amdgpu_bo_is_user_ptr;
1756   ws->base.buffer_is_suballocated = amdgpu_bo_is_suballocated;
1757   ws->base.buffer_get_handle = amdgpu_bo_get_handle;
1758   ws->base.buffer_commit = amdgpu_bo_sparse_commit;
1759   ws->base.buffer_get_virtual_address = amdgpu_bo_get_va;
1760   ws->base.buffer_get_initial_domain = amdgpu_bo_get_initial_domain;
1761   ws->base.buffer_get_flags = amdgpu_bo_get_flags;
1762}
1763