aboutsummaryrefslogtreecommitdiffstats
path: root/src/mesa/drivers/dri/i965/intel_buffer_objects.c
blob: 452e6d33c070c75a99c6acc3f7be425e85410b96 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
/*
 * Copyright 2003 VMware, Inc.
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the
 * "Software"), to deal in the Software without restriction, including
 * without limitation the rights to use, copy, modify, merge, publish,
 * distribute, sublicense, and/or sell copies of the Software, and to
 * permit persons to whom the Software is furnished to do so, subject to
 * the following conditions:
 *
 * The above copyright notice and this permission notice (including the
 * next paragraph) shall be included in all copies or substantial portions
 * of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 */

/**
 * @file intel_buffer_objects.c
 *
 * This provides core GL buffer object functionality.
 */

#include "main/imports.h"
#include "main/mtypes.h"
#include "main/macros.h"
#include "main/streaming-load-memcpy.h"
#include "main/bufferobj.h"
#include "x86/common_x86_asm.h"

#include "brw_context.h"
#include "brw_blorp.h"
#include "intel_buffer_objects.h"
#include "intel_batchbuffer.h"
#include "intel_tiled_memcpy.h"

static void
mark_buffer_gpu_usage(struct intel_buffer_object *intel_obj,
                               uint32_t offset, uint32_t size)
{
   intel_obj->gpu_active_start = MIN2(intel_obj->gpu_active_start, offset);
   intel_obj->gpu_active_end = MAX2(intel_obj->gpu_active_end, offset + size);
}

static void
mark_buffer_inactive(struct intel_buffer_object *intel_obj)
{
   intel_obj->gpu_active_start = ~0;
   intel_obj->gpu_active_end = 0;
}

static void
mark_buffer_valid_data(struct intel_buffer_object *intel_obj,
                       uint32_t offset, uint32_t size)
{
   intel_obj->valid_data_start = MIN2(intel_obj->valid_data_start, offset);
   intel_obj->valid_data_end = MAX2(intel_obj->valid_data_end, offset + size);
}

static void
mark_buffer_invalid(struct intel_buffer_object *intel_obj)
{
   intel_obj->valid_data_start = ~0;
   intel_obj->valid_data_end = 0;
}

/** Allocates a new brw_bo to store the data for the buffer object. */
static void
alloc_buffer_object(struct brw_context *brw,
                    struct intel_buffer_object *intel_obj)
{
   const struct gl_context *ctx = &brw->ctx;

   uint64_t size = intel_obj->Base.Size;
   if (ctx->Const.RobustAccess) {
      /* Pad out buffer objects with an extra 2kB (half a page).
       *
       * When pushing UBOs, we need to safeguard against 3DSTATE_CONSTANT_*
       * reading out of bounds memory.  The application might bind a UBO that's
       * smaller than what the program expects.  Ideally, we'd bind an extra
       * push buffer containing zeros, but we have a limited number of those,
       * so it's not always viable.  Our only safe option is to pad all buffer
       * objects by the maximum push data length, so that it will never read
       * past the end of a BO.
       *
       * This is unfortunate, but it should result in at most 1 extra page,
       * which probably isn't too terrible.
       */
      size += 64 * 32; /* max read length of 64 256-bit units */
   }
   intel_obj->buffer =
      brw_bo_alloc(brw->bufmgr, "bufferobj", size, BRW_MEMZONE_OTHER);

   /* the buffer might be bound as a uniform buffer, need to update it
    */
   if (intel_obj->Base.UsageHistory & USAGE_UNIFORM_BUFFER)
      brw->ctx.NewDriverState |= BRW_NEW_UNIFORM_BUFFER;
   if (intel_obj->Base.UsageHistory & USAGE_SHADER_STORAGE_BUFFER)
      brw->ctx.NewDriverState |= BRW_NEW_UNIFORM_BUFFER;
   if (intel_obj->Base.UsageHistory & USAGE_TEXTURE_BUFFER)
      brw->ctx.NewDriverState |= BRW_NEW_TEXTURE_BUFFER;
   if (intel_obj->Base.UsageHistory & USAGE_ATOMIC_COUNTER_BUFFER)
      brw->ctx.NewDriverState |= BRW_NEW_UNIFORM_BUFFER;

   mark_buffer_inactive(intel_obj);
   mark_buffer_invalid(intel_obj);
}

static void
release_buffer(struct intel_buffer_object *intel_obj)
{
   brw_bo_unreference(intel_obj->buffer);
   intel_obj->buffer = NULL;
}

/**
 * The NewBufferObject() driver hook.
 *
 * Allocates a new intel_buffer_object structure and initializes it.
 *
 * There is some duplication between mesa's bufferobjects and our
 * bufmgr buffers.  Both have an integer handle and a hashtable to
 * lookup an opaque structure.  It would be nice if the handles and
 * internal structure where somehow shared.
 */
static struct gl_buffer_object *
brw_new_buffer_object(struct gl_context * ctx, GLuint name)
{
   struct intel_buffer_object *obj = CALLOC_STRUCT(intel_buffer_object);
   if (!obj) {
      _mesa_error_no_memory(__func__);
      return NULL;
   }

   _mesa_initialize_buffer_object(ctx, &obj->Base, name);

   obj->buffer = NULL;

   return &obj->Base;
}

/**
 * The DeleteBuffer() driver hook.
 *
 * Deletes a single OpenGL buffer object.  Used by glDeleteBuffers().
 */
static void
brw_delete_buffer(struct gl_context * ctx, struct gl_buffer_object *obj)
{
   struct intel_buffer_object *intel_obj = intel_buffer_object(obj);

   assert(intel_obj);

   /* Buffer objects are automatically unmapped when deleting according
    * to the spec, but Mesa doesn't do UnmapBuffer for us at context destroy
    * (though it does if you call glDeleteBuffers)
    */
   _mesa_buffer_unmap_all_mappings(ctx, obj);

   brw_bo_unreference(intel_obj->buffer);
   _mesa_delete_buffer_object(ctx, obj);
}


/**
 * The BufferData() driver hook.
 *
 * Implements glBufferData(), which recreates a buffer object's data store
 * and populates it with the given data, if present.
 *
 * Any data that was previously stored in the buffer object is lost.
 *
 * \return true for success, false if out of memory
 */
static GLboolean
brw_buffer_data(struct gl_context *ctx,
                GLenum target,
                GLsizeiptrARB size,
                const GLvoid *data,
                GLenum usage,
                GLbitfield storageFlags,
                struct gl_buffer_object *obj)
{
   struct brw_context *brw = brw_context(ctx);
   struct intel_buffer_object *intel_obj = intel_buffer_object(obj);

   /* Part of the ABI, but this function doesn't use it.
    */
   (void) target;

   intel_obj->Base.Size = size;
   intel_obj->Base.Usage = usage;
   intel_obj->Base.StorageFlags = storageFlags;

   assert(!obj->Mappings[MAP_USER].Pointer); /* Mesa should have unmapped it */
   assert(!obj->Mappings[MAP_INTERNAL].Pointer);

   if (intel_obj->buffer != NULL)
      release_buffer(intel_obj);

   if (size != 0) {
      alloc_buffer_object(brw, intel_obj);
      if (!intel_obj->buffer)
         return false;

      if (data != NULL) {
         brw_bo_subdata(intel_obj->buffer, 0, size, data);
         mark_buffer_valid_data(intel_obj, 0, size);
      }
   }

   return true;
}


/**
 * The BufferSubData() driver hook.
 *
 * Implements glBufferSubData(), which replaces a portion of the data in a
 * buffer object.
 *
 * If the data range specified by (size + offset) extends beyond the end of
 * the buffer or if data is NULL, no copy is performed.
 */
static void
brw_buffer_subdata(struct gl_context *ctx,
                   GLintptrARB offset,
                   GLsizeiptrARB size,
                   const GLvoid *data,
                   struct gl_buffer_object *obj)
{
   struct brw_context *brw = brw_context(ctx);
   struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
   bool busy;

   if (size == 0)
      return;

   assert(intel_obj);

   /* See if we can unsynchronized write the data into the user's BO. This
    * avoids GPU stalls in unfortunately common user patterns (uploading
    * sequentially into a BO, with draw calls in between each upload).
    *
    * Once we've hit this path, we mark this GL BO as preferring stalling to
    * blits, so that we can hopefully hit this path again in the future
    * (otherwise, an app that might occasionally stall but mostly not will end
    * up with blitting all the time, at the cost of bandwidth)
    */
   if (offset + size <= intel_obj->gpu_active_start ||
       intel_obj->gpu_active_end <= offset ||
       offset + size <= intel_obj->valid_data_start ||
       intel_obj->valid_data_end <= offset) {
      void *map = brw_bo_map(brw, intel_obj->buffer, MAP_WRITE | MAP_ASYNC);
      memcpy(map + offset, data, size);
      brw_bo_unmap(intel_obj->buffer);

      if (intel_obj->gpu_active_end > intel_obj->gpu_active_start)
         intel_obj->prefer_stall_to_blit = true;

      mark_buffer_valid_data(intel_obj, offset, size);
      return;
   }

   busy =
      brw_bo_busy(intel_obj->buffer) ||
      brw_batch_references(&brw->batch, intel_obj->buffer);

   if (busy) {
      if (size == intel_obj->Base.Size ||
          (intel_obj->valid_data_start >= offset &&
           intel_obj->valid_data_end <= offset + size)) {
         /* Replace the current busy bo so the subdata doesn't stall. */
         brw_bo_unreference(intel_obj->buffer);
         alloc_buffer_object(brw, intel_obj);
      } else if (!intel_obj->prefer_stall_to_blit) {
         perf_debug("Using a blit copy to avoid stalling on "
                    "glBufferSubData(%ld, %ld) (%ldkb) to a busy "
                    "(%d-%d) / valid (%d-%d) buffer object.\n",
                    (long)offset, (long)offset + size, (long)(size/1024),
                    intel_obj->gpu_active_start,
                    intel_obj->gpu_active_end,
                    intel_obj->valid_data_start,
                    intel_obj->valid_data_end);
         struct brw_bo *temp_bo =
            brw_bo_alloc(brw->bufmgr, "subdata temp", size, BRW_MEMZONE_OTHER);

         brw_bo_subdata(temp_bo, 0, size, data);

         brw_blorp_copy_buffers(brw,
                                temp_bo, 0,
                                intel_obj->buffer, offset,
                                size);
         brw_emit_mi_flush(brw);

         brw_bo_unreference(temp_bo);
         mark_buffer_valid_data(intel_obj, offset, size);
         return;
      } else {
         perf_debug("Stalling on glBufferSubData(%ld, %ld) (%ldkb) to a busy "
                    "(%d-%d) buffer object.  Use glMapBufferRange() to "
                    "avoid this.\n",
                    (long)offset, (long)offset + size, (long)(size/1024),
                    intel_obj->gpu_active_start,
                    intel_obj->gpu_active_end);
         intel_batchbuffer_flush(brw);
      }
   }

   brw_bo_subdata(intel_obj->buffer, offset, size, data);
   mark_buffer_inactive(intel_obj);
   mark_buffer_valid_data(intel_obj, offset, size);
}


/**
 * The GetBufferSubData() driver hook.
 *
 * Implements glGetBufferSubData(), which copies a subrange of a buffer
 * object into user memory.
 */
static void
brw_get_buffer_subdata(struct gl_context *ctx,
                       GLintptrARB offset,
                       GLsizeiptrARB size,
                       GLvoid *data,
                       struct gl_buffer_object *obj)
{
   struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
   struct brw_context *brw = brw_context(ctx);

   assert(intel_obj);
   if (brw_batch_references(&brw->batch, intel_obj->buffer)) {
      intel_batchbuffer_flush(brw);
   }

   unsigned int map_flags = MAP_READ;
   mem_copy_fn memcpy_fn = memcpy;
#ifdef USE_SSE41
   if (!intel_obj->buffer->cache_coherent && cpu_has_sse4_1) {
      /* Rather than acquire a new WB mmaping of the buffer object and pull
       * it into the CPU cache, keep using the WC mmap that we have for writes,
       * and use the magic movntd instructions instead.
       */
      map_flags |= MAP_COHERENT;
      memcpy_fn = (mem_copy_fn) _mesa_streaming_load_memcpy;
   }
#endif

   void *map = brw_bo_map(brw, intel_obj->buffer, map_flags);
   if (unlikely(!map)) {
      _mesa_error_no_memory(__func__);
      return;
   }
   memcpy_fn(data, map + offset, size);
   brw_bo_unmap(intel_obj->buffer);

   mark_buffer_inactive(intel_obj);
}


/**
 * The MapBufferRange() driver hook.
 *
 * This implements both glMapBufferRange() and glMapBuffer().
 *
 * The goal of this extension is to allow apps to accumulate their rendering
 * at the same time as they accumulate their buffer object.  Without it,
 * you'd end up blocking on execution of rendering every time you mapped
 * the buffer to put new data in.
 *
 * We support it in 3 ways: If unsynchronized, then don't bother
 * flushing the batchbuffer before mapping the buffer, which can save blocking
 * in many cases.  If we would still block, and they allow the whole buffer
 * to be invalidated, then just allocate a new buffer to replace the old one.
 * If not, and we'd block, and they allow the subrange of the buffer to be
 * invalidated, then we can make a new little BO, let them write into that,
 * and blit it into the real BO at unmap time.
 */
static void *
brw_map_buffer_range(struct gl_context *ctx,
                     GLintptr offset, GLsizeiptr length,
                     GLbitfield access, struct gl_buffer_object *obj,
                     gl_map_buffer_index index)
{
   struct brw_context *brw = brw_context(ctx);
   struct intel_buffer_object *intel_obj = intel_buffer_object(obj);

   assert(intel_obj);

   STATIC_ASSERT(GL_MAP_UNSYNCHRONIZED_BIT == MAP_ASYNC);
   STATIC_ASSERT(GL_MAP_WRITE_BIT == MAP_WRITE);
   STATIC_ASSERT(GL_MAP_READ_BIT == MAP_READ);
   STATIC_ASSERT(GL_MAP_PERSISTENT_BIT == MAP_PERSISTENT);
   STATIC_ASSERT(GL_MAP_COHERENT_BIT == MAP_COHERENT);
   assert((access & MAP_INTERNAL_MASK) == 0);

   /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
    * internally uses our functions directly.
    */
   obj->Mappings[index].Offset = offset;
   obj->Mappings[index].Length = length;
   obj->Mappings[index].AccessFlags = access;

   if (intel_obj->buffer == NULL) {
      obj->Mappings[index].Pointer = NULL;
      return NULL;
   }

   /* If the access is synchronized (like a normal buffer mapping), then get
    * things flushed out so the later mapping syncs appropriately through GEM.
    * If the user doesn't care about existing buffer contents and mapping would
    * cause us to block, then throw out the old buffer.
    *
    * If they set INVALIDATE_BUFFER, we can pitch the current contents to
    * achieve the required synchronization.
    */
   if (!(access & GL_MAP_UNSYNCHRONIZED_BIT)) {
      if (brw_batch_references(&brw->batch, intel_obj->buffer)) {
         if (access & GL_MAP_INVALIDATE_BUFFER_BIT) {
            brw_bo_unreference(intel_obj->buffer);
            alloc_buffer_object(brw, intel_obj);
         } else {
            perf_debug("Stalling on the GPU for mapping a busy buffer "
                       "object\n");
            intel_batchbuffer_flush(brw);
         }
      } else if (brw_bo_busy(intel_obj->buffer) &&
                 (access & GL_MAP_INVALIDATE_BUFFER_BIT)) {
         brw_bo_unreference(intel_obj->buffer);
         alloc_buffer_object(brw, intel_obj);
      }
   }

   if (access & MAP_WRITE)
      mark_buffer_valid_data(intel_obj, offset, length);

   /* If the user is mapping a range of an active buffer object but
    * doesn't require the current contents of that range, make a new
    * BO, and we'll copy what they put in there out at unmap or
    * FlushRange time.
    *
    * That is, unless they're looking for a persistent mapping -- we would
    * need to do blits in the MemoryBarrier call, and it's easier to just do a
    * GPU stall and do a mapping.
    */
   if (!(access & (GL_MAP_UNSYNCHRONIZED_BIT | GL_MAP_PERSISTENT_BIT)) &&
       (access & GL_MAP_INVALIDATE_RANGE_BIT) &&
       brw_bo_busy(intel_obj->buffer)) {
      /* Ensure that the base alignment of the allocation meets the alignment
       * guarantees the driver has advertised to the application.
       */
      const unsigned alignment = ctx->Const.MinMapBufferAlignment;

      intel_obj->map_extra[index] = (uintptr_t) offset % alignment;
      intel_obj->range_map_bo[index] =
         brw_bo_alloc(brw->bufmgr, "BO blit temp",
                      length + intel_obj->map_extra[index],
                      BRW_MEMZONE_OTHER);
      void *map = brw_bo_map(brw, intel_obj->range_map_bo[index], access);
      obj->Mappings[index].Pointer = map + intel_obj->map_extra[index];
      return obj->Mappings[index].Pointer;
   }

   void *map = brw_bo_map(brw, intel_obj->buffer, access);
   if (!(access & GL_MAP_UNSYNCHRONIZED_BIT)) {
      mark_buffer_inactive(intel_obj);
   }

   obj->Mappings[index].Pointer = map + offset;
   return obj->Mappings[index].Pointer;
}

/**
 * The FlushMappedBufferRange() driver hook.
 *
 * Implements glFlushMappedBufferRange(), which signifies that modifications
 * have been made to a range of a mapped buffer, and it should be flushed.
 *
 * This is only used for buffers mapped with GL_MAP_FLUSH_EXPLICIT_BIT.
 *
 * Ideally we'd use a BO to avoid taking up cache space for the temporary
 * data, but FlushMappedBufferRange may be followed by further writes to
 * the pointer, so we would have to re-map after emitting our blit, which
 * would defeat the point.
 */
static void
brw_flush_mapped_buffer_range(struct gl_context *ctx,
                              GLintptr offset, GLsizeiptr length,
                              struct gl_buffer_object *obj,
                              gl_map_buffer_index index)
{
   struct brw_context *brw = brw_context(ctx);
   struct intel_buffer_object *intel_obj = intel_buffer_object(obj);

   assert(obj->Mappings[index].AccessFlags & GL_MAP_FLUSH_EXPLICIT_BIT);

   /* If we gave a direct mapping of the buffer instead of using a temporary,
    * then there's nothing to do.
    */
   if (intel_obj->range_map_bo[index] == NULL)
      return;

   if (length == 0)
      return;

   /* Note that we're not unmapping our buffer while executing the blit.  We
    * need to have a mapping still at the end of this call, since the user
    * gets to make further modifications and glFlushMappedBufferRange() calls.
    * This is safe, because:
    *
    * - On LLC platforms, we're using a CPU mapping that's coherent with the
    *   GPU (except for the render caches), so the kernel doesn't need to do
    *   any flushing work for us except for what happens at batch exec time
    *   anyway.
    *
    * - On non-LLC platforms, we're using a GTT mapping that writes directly
    *   to system memory (except for the chipset cache that gets flushed at
    *   batch exec time).
    *
    * In both cases we don't need to stall for the previous blit to complete
    * so we can re-map (and we definitely don't want to, since that would be
    * slow): If the user edits a part of their buffer that's previously been
    * blitted, then our lack of synchoronization is fine, because either
    * they'll get some too-new data in the first blit and not do another blit
    * of that area (but in that case the results are undefined), or they'll do
    * another blit of that area and the complete newer data will land the
    * second time.
    */
   brw_blorp_copy_buffers(brw,
                          intel_obj->range_map_bo[index],
                          intel_obj->map_extra[index] + offset,
                          intel_obj->buffer,
                          obj->Mappings[index].Offset + offset,
                          length);
   mark_buffer_gpu_usage(intel_obj,
                         obj->Mappings[index].Offset + offset,
                         length);
   brw_emit_mi_flush(brw);
}


/**
 * The UnmapBuffer() driver hook.
 *
 * Implements glUnmapBuffer().
 */
static GLboolean
brw_unmap_buffer(struct gl_context *ctx,
                 struct gl_buffer_object *obj,
                 gl_map_buffer_index index)
{
   struct brw_context *brw = brw_context(ctx);
   struct intel_buffer_object *intel_obj = intel_buffer_object(obj);

   assert(intel_obj);
   assert(obj->Mappings[index].Pointer);
   if (intel_obj->range_map_bo[index] != NULL) {
      brw_bo_unmap(intel_obj->range_map_bo[index]);

      if (!(obj->Mappings[index].AccessFlags & GL_MAP_FLUSH_EXPLICIT_BIT)) {
         brw_blorp_copy_buffers(brw,
                                intel_obj->range_map_bo[index],
                                intel_obj->map_extra[index],
                                intel_obj->buffer, obj->Mappings[index].Offset,
                                obj->Mappings[index].Length);
         mark_buffer_gpu_usage(intel_obj, obj->Mappings[index].Offset,
                               obj->Mappings[index].Length);
         brw_emit_mi_flush(brw);
      }

      /* Since we've emitted some blits to buffers that will (likely) be used
       * in rendering operations in other cache domains in this batch, emit a
       * flush.  Once again, we wish for a domain tracker in libdrm to cover
       * usage inside of a batchbuffer.
       */

      brw_bo_unreference(intel_obj->range_map_bo[index]);
      intel_obj->range_map_bo[index] = NULL;
   } else if (intel_obj->buffer != NULL) {
      brw_bo_unmap(intel_obj->buffer);
   }
   obj->Mappings[index].Pointer = NULL;
   obj->Mappings[index].Offset = 0;
   obj->Mappings[index].Length = 0;

   return true;
}

/**
 * Gets a pointer to the object's BO, and marks the given range as being used
 * on the GPU.
 *
 * Anywhere that uses buffer objects in the pipeline should be using this to
 * mark the range of the buffer that is being accessed by the pipeline.
 */
struct brw_bo *
intel_bufferobj_buffer(struct brw_context *brw,
                       struct intel_buffer_object *intel_obj,
                       uint32_t offset, uint32_t size, bool write)
{
   /* This is needed so that things like transform feedback and texture buffer
    * objects that need a BO but don't want to check that they exist for
    * draw-time validation can just always get a BO from a GL buffer object.
    */
   if (intel_obj->buffer == NULL)
      alloc_buffer_object(brw, intel_obj);

   mark_buffer_gpu_usage(intel_obj, offset, size);

   /* If writing, (conservatively) mark this section as having valid data. */
   if (write)
      mark_buffer_valid_data(intel_obj, offset, size);

   return intel_obj->buffer;
}

/**
 * The CopyBufferSubData() driver hook.
 *
 * Implements glCopyBufferSubData(), which copies a portion of one buffer
 * object's data to another.  Independent source and destination offsets
 * are allowed.
 */
static void
brw_copy_buffer_subdata(struct gl_context *ctx,
                        struct gl_buffer_object *src,
                        struct gl_buffer_object *dst,
                        GLintptr read_offset, GLintptr write_offset,
                        GLsizeiptr size)
{
   struct brw_context *brw = brw_context(ctx);
   struct intel_buffer_object *intel_src = intel_buffer_object(src);
   struct intel_buffer_object *intel_dst = intel_buffer_object(dst);
   struct brw_bo *src_bo, *dst_bo;

   if (size == 0)
      return;

   dst_bo = intel_bufferobj_buffer(brw, intel_dst, write_offset, size, true);
   src_bo = intel_bufferobj_buffer(brw, intel_src, read_offset, size, false);

   brw_blorp_copy_buffers(brw,
                          src_bo, read_offset,
                          dst_bo, write_offset, size);

   /* Since we've emitted some blits to buffers that will (likely) be used
    * in rendering operations in other cache domains in this batch, emit a
    * flush.  Once again, we wish for a domain tracker in libdrm to cover
    * usage inside of a batchbuffer.
    */
   brw_emit_mi_flush(brw);
}

void
intelInitBufferObjectFuncs(struct dd_function_table *functions)
{
   functions->NewBufferObject = brw_new_buffer_object;
   functions->DeleteBuffer = brw_delete_buffer;
   functions->BufferData = brw_buffer_data;
   functions->BufferSubData = brw_buffer_subdata;
   functions->GetBufferSubData = brw_get_buffer_subdata;
   functions->MapBufferRange = brw_map_buffer_range;
   functions->FlushMappedBufferRange = brw_flush_mapped_buffer_range;
   functions->UnmapBuffer = brw_unmap_buffer;
   functions->CopyBufferSubData = brw_copy_buffer_subdata;
}