Home | History | Annotate | Download | only in i965
      1 /*
      2  * Copyright 2003 VMware, Inc.
      3  * All Rights Reserved.
      4  *
      5  * Permission is hereby granted, free of charge, to any person obtaining a
      6  * copy of this software and associated documentation files (the
      7  * "Software"), to deal in the Software without restriction, including
      8  * without limitation the rights to use, copy, modify, merge, publish,
      9  * distribute, sublicense, and/or sell copies of the Software, and to
     10  * permit persons to whom the Software is furnished to do so, subject to
     11  * the following conditions:
     12  *
     13  * The above copyright notice and this permission notice (including the
     14  * next paragraph) shall be included in all copies or substantial portions
     15  * of the Software.
     16  *
     17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
     18  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
     19  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
     20  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
     21  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
     22  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
     23  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
     24  */
     25 
     26 /**
     27  * @file intel_buffer_objects.c
     28  *
     29  * This provides core GL buffer object functionality.
     30  */
     31 
     32 #include "main/imports.h"
     33 #include "main/mtypes.h"
     34 #include "main/macros.h"
     35 #include "main/bufferobj.h"
     36 
     37 #include "brw_context.h"
     38 #include "intel_blit.h"
     39 #include "intel_buffer_objects.h"
     40 #include "intel_batchbuffer.h"
     41 
     42 /**
     43  * Map a buffer object; issue performance warnings if mapping causes stalls.
     44  *
     45  * This matches the drm_intel_bo_map API, but takes an additional human-readable
     46  * name for the buffer object to use in the performance debug message.
     47  */
     48 int
     49 brw_bo_map(struct brw_context *brw,
     50            drm_intel_bo *bo, int write_enable,
     51            const char *bo_name)
     52 {
     53    if (likely(!brw->perf_debug) || !drm_intel_bo_busy(bo))
     54       return drm_intel_bo_map(bo, write_enable);
     55 
     56    double start_time = get_time();
     57 
     58    int ret = drm_intel_bo_map(bo, write_enable);
     59 
     60    perf_debug("CPU mapping a busy %s BO stalled and took %.03f ms.\n",
     61               bo_name, (get_time() - start_time) * 1000);
     62 
     63    return ret;
     64 }
     65 
     66 int
     67 brw_bo_map_gtt(struct brw_context *brw, drm_intel_bo *bo, const char *bo_name)
     68 {
     69    if (likely(!brw->perf_debug) || !drm_intel_bo_busy(bo))
     70       return drm_intel_gem_bo_map_gtt(bo);
     71 
     72    double start_time = get_time();
     73 
     74    int ret = drm_intel_gem_bo_map_gtt(bo);
     75 
     76    perf_debug("GTT mapping a busy %s BO stalled and took %.03f ms.\n",
     77               bo_name, (get_time() - start_time) * 1000);
     78 
     79    return ret;
     80 }
     81 
     82 static void
     83 mark_buffer_gpu_usage(struct intel_buffer_object *intel_obj,
     84                                uint32_t offset, uint32_t size)
     85 {
     86    intel_obj->gpu_active_start = MIN2(intel_obj->gpu_active_start, offset);
     87    intel_obj->gpu_active_end = MAX2(intel_obj->gpu_active_end, offset + size);
     88 }
     89 
     90 static void
     91 mark_buffer_inactive(struct intel_buffer_object *intel_obj)
     92 {
     93    intel_obj->gpu_active_start = ~0;
     94    intel_obj->gpu_active_end = 0;
     95 }
     96 
     97 /** Allocates a new drm_intel_bo to store the data for the buffer object. */
     98 static void
     99 alloc_buffer_object(struct brw_context *brw,
    100                     struct intel_buffer_object *intel_obj)
    101 {
    102    intel_obj->buffer = drm_intel_bo_alloc(brw->bufmgr, "bufferobj",
    103 					  intel_obj->Base.Size, 64);
    104 
    105    /* the buffer might be bound as a uniform buffer, need to update it
    106     */
    107    if (intel_obj->Base.UsageHistory & USAGE_UNIFORM_BUFFER)
    108       brw->ctx.NewDriverState |= BRW_NEW_UNIFORM_BUFFER;
    109    if (intel_obj->Base.UsageHistory & USAGE_SHADER_STORAGE_BUFFER)
    110       brw->ctx.NewDriverState |= BRW_NEW_UNIFORM_BUFFER;
    111    if (intel_obj->Base.UsageHistory & USAGE_TEXTURE_BUFFER)
    112       brw->ctx.NewDriverState |= BRW_NEW_TEXTURE_BUFFER;
    113    if (intel_obj->Base.UsageHistory & USAGE_ATOMIC_COUNTER_BUFFER)
    114       brw->ctx.NewDriverState |= BRW_NEW_ATOMIC_BUFFER;
    115 
    116    mark_buffer_inactive(intel_obj);
    117 }
    118 
    119 static void
    120 release_buffer(struct intel_buffer_object *intel_obj)
    121 {
    122    drm_intel_bo_unreference(intel_obj->buffer);
    123    intel_obj->buffer = NULL;
    124 }
    125 
    126 /**
    127  * The NewBufferObject() driver hook.
    128  *
    129  * Allocates a new intel_buffer_object structure and initializes it.
    130  *
    131  * There is some duplication between mesa's bufferobjects and our
    132  * bufmgr buffers.  Both have an integer handle and a hashtable to
    133  * lookup an opaque structure.  It would be nice if the handles and
    134  * internal structure where somehow shared.
    135  */
    136 static struct gl_buffer_object *
    137 brw_new_buffer_object(struct gl_context * ctx, GLuint name)
    138 {
    139    struct intel_buffer_object *obj = CALLOC_STRUCT(intel_buffer_object);
    140    if (!obj) {
    141       _mesa_error_no_memory(__func__);
    142    }
    143 
    144    _mesa_initialize_buffer_object(ctx, &obj->Base, name);
    145 
    146    obj->buffer = NULL;
    147 
    148    return &obj->Base;
    149 }
    150 
    151 /**
    152  * The DeleteBuffer() driver hook.
    153  *
    154  * Deletes a single OpenGL buffer object.  Used by glDeleteBuffers().
    155  */
    156 static void
    157 brw_delete_buffer(struct gl_context * ctx, struct gl_buffer_object *obj)
    158 {
    159    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
    160 
    161    assert(intel_obj);
    162 
    163    /* Buffer objects are automatically unmapped when deleting according
    164     * to the spec, but Mesa doesn't do UnmapBuffer for us at context destroy
    165     * (though it does if you call glDeleteBuffers)
    166     */
    167    _mesa_buffer_unmap_all_mappings(ctx, obj);
    168 
    169    drm_intel_bo_unreference(intel_obj->buffer);
    170    _mesa_delete_buffer_object(ctx, obj);
    171 }
    172 
    173 
    174 /**
    175  * The BufferData() driver hook.
    176  *
    177  * Implements glBufferData(), which recreates a buffer object's data store
    178  * and populates it with the given data, if present.
    179  *
    180  * Any data that was previously stored in the buffer object is lost.
    181  *
    182  * \return true for success, false if out of memory
    183  */
    184 static GLboolean
    185 brw_buffer_data(struct gl_context *ctx,
    186                 GLenum target,
    187                 GLsizeiptrARB size,
    188                 const GLvoid *data,
    189                 GLenum usage,
    190                 GLbitfield storageFlags,
    191                 struct gl_buffer_object *obj)
    192 {
    193    struct brw_context *brw = brw_context(ctx);
    194    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
    195 
    196    /* Part of the ABI, but this function doesn't use it.
    197     */
    198    (void) target;
    199 
    200    intel_obj->Base.Size = size;
    201    intel_obj->Base.Usage = usage;
    202    intel_obj->Base.StorageFlags = storageFlags;
    203 
    204    assert(!obj->Mappings[MAP_USER].Pointer); /* Mesa should have unmapped it */
    205    assert(!obj->Mappings[MAP_INTERNAL].Pointer);
    206 
    207    if (intel_obj->buffer != NULL)
    208       release_buffer(intel_obj);
    209 
    210    if (size != 0) {
    211       alloc_buffer_object(brw, intel_obj);
    212       if (!intel_obj->buffer)
    213          return false;
    214 
    215       if (data != NULL)
    216 	 drm_intel_bo_subdata(intel_obj->buffer, 0, size, data);
    217    }
    218 
    219    return true;
    220 }
    221 
    222 
    223 /**
    224  * The BufferSubData() driver hook.
    225  *
    226  * Implements glBufferSubData(), which replaces a portion of the data in a
    227  * buffer object.
    228  *
    229  * If the data range specified by (size + offset) extends beyond the end of
    230  * the buffer or if data is NULL, no copy is performed.
    231  */
    232 static void
    233 brw_buffer_subdata(struct gl_context *ctx,
    234                    GLintptrARB offset,
    235                    GLsizeiptrARB size,
    236                    const GLvoid *data,
    237                    struct gl_buffer_object *obj)
    238 {
    239    struct brw_context *brw = brw_context(ctx);
    240    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
    241    bool busy;
    242 
    243    if (size == 0)
    244       return;
    245 
    246    assert(intel_obj);
    247 
    248    /* See if we can unsynchronized write the data into the user's BO. This
    249     * avoids GPU stalls in unfortunately common user patterns (uploading
    250     * sequentially into a BO, with draw calls in between each upload).
    251     *
    252     * Once we've hit this path, we mark this GL BO as preferring stalling to
    253     * blits, so that we can hopefully hit this path again in the future
    254     * (otherwise, an app that might occasionally stall but mostly not will end
    255     * up with blitting all the time, at the cost of bandwidth)
    256     */
    257    if (offset + size <= intel_obj->gpu_active_start ||
    258        intel_obj->gpu_active_end <= offset) {
    259       if (brw->has_llc) {
    260          drm_intel_gem_bo_map_unsynchronized(intel_obj->buffer);
    261          memcpy(intel_obj->buffer->virtual + offset, data, size);
    262          drm_intel_bo_unmap(intel_obj->buffer);
    263 
    264          if (intel_obj->gpu_active_end > intel_obj->gpu_active_start)
    265             intel_obj->prefer_stall_to_blit = true;
    266          return;
    267       } else {
    268          perf_debug("BufferSubData could be unsynchronized, but !LLC doesn't support it yet\n");
    269       }
    270    }
    271 
    272    busy =
    273       drm_intel_bo_busy(intel_obj->buffer) ||
    274       drm_intel_bo_references(brw->batch.bo, intel_obj->buffer);
    275 
    276    if (busy) {
    277       if (size == intel_obj->Base.Size) {
    278 	 /* Replace the current busy bo so the subdata doesn't stall. */
    279 	 drm_intel_bo_unreference(intel_obj->buffer);
    280 	 alloc_buffer_object(brw, intel_obj);
    281       } else if (!intel_obj->prefer_stall_to_blit) {
    282          perf_debug("Using a blit copy to avoid stalling on "
    283                     "glBufferSubData(%ld, %ld) (%ldkb) to a busy "
    284                     "(%d-%d) buffer object.\n",
    285                     (long)offset, (long)offset + size, (long)(size/1024),
    286                     intel_obj->gpu_active_start,
    287                     intel_obj->gpu_active_end);
    288 	 drm_intel_bo *temp_bo =
    289 	    drm_intel_bo_alloc(brw->bufmgr, "subdata temp", size, 64);
    290 
    291 	 drm_intel_bo_subdata(temp_bo, 0, size, data);
    292 
    293 	 intel_emit_linear_blit(brw,
    294 				intel_obj->buffer, offset,
    295 				temp_bo, 0,
    296 				size);
    297 
    298 	 drm_intel_bo_unreference(temp_bo);
    299          return;
    300       } else {
    301          perf_debug("Stalling on glBufferSubData(%ld, %ld) (%ldkb) to a busy "
    302                     "(%d-%d) buffer object.  Use glMapBufferRange() to "
    303                     "avoid this.\n",
    304                     (long)offset, (long)offset + size, (long)(size/1024),
    305                     intel_obj->gpu_active_start,
    306                     intel_obj->gpu_active_end);
    307          intel_batchbuffer_flush(brw);
    308       }
    309    }
    310 
    311    drm_intel_bo_subdata(intel_obj->buffer, offset, size, data);
    312    mark_buffer_inactive(intel_obj);
    313 }
    314 
    315 
    316 /**
    317  * The GetBufferSubData() driver hook.
    318  *
    319  * Implements glGetBufferSubData(), which copies a subrange of a buffer
    320  * object into user memory.
    321  */
    322 static void
    323 brw_get_buffer_subdata(struct gl_context *ctx,
    324                        GLintptrARB offset,
    325                        GLsizeiptrARB size,
    326                        GLvoid *data,
    327                        struct gl_buffer_object *obj)
    328 {
    329    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
    330    struct brw_context *brw = brw_context(ctx);
    331 
    332    assert(intel_obj);
    333    if (drm_intel_bo_references(brw->batch.bo, intel_obj->buffer)) {
    334       intel_batchbuffer_flush(brw);
    335    }
    336    drm_intel_bo_get_subdata(intel_obj->buffer, offset, size, data);
    337 
    338    mark_buffer_inactive(intel_obj);
    339 }
    340 
    341 
    342 /**
    343  * The MapBufferRange() driver hook.
    344  *
    345  * This implements both glMapBufferRange() and glMapBuffer().
    346  *
    347  * The goal of this extension is to allow apps to accumulate their rendering
    348  * at the same time as they accumulate their buffer object.  Without it,
    349  * you'd end up blocking on execution of rendering every time you mapped
    350  * the buffer to put new data in.
    351  *
    352  * We support it in 3 ways: If unsynchronized, then don't bother
    353  * flushing the batchbuffer before mapping the buffer, which can save blocking
    354  * in many cases.  If we would still block, and they allow the whole buffer
    355  * to be invalidated, then just allocate a new buffer to replace the old one.
    356  * If not, and we'd block, and they allow the subrange of the buffer to be
    357  * invalidated, then we can make a new little BO, let them write into that,
    358  * and blit it into the real BO at unmap time.
    359  */
    360 static void *
    361 brw_map_buffer_range(struct gl_context *ctx,
    362                      GLintptr offset, GLsizeiptr length,
    363                      GLbitfield access, struct gl_buffer_object *obj,
    364                      gl_map_buffer_index index)
    365 {
    366    struct brw_context *brw = brw_context(ctx);
    367    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
    368 
    369    assert(intel_obj);
    370 
    371    /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
    372     * internally uses our functions directly.
    373     */
    374    obj->Mappings[index].Offset = offset;
    375    obj->Mappings[index].Length = length;
    376    obj->Mappings[index].AccessFlags = access;
    377 
    378    if (intel_obj->buffer == NULL) {
    379       obj->Mappings[index].Pointer = NULL;
    380       return NULL;
    381    }
    382 
    383    /* If the access is synchronized (like a normal buffer mapping), then get
    384     * things flushed out so the later mapping syncs appropriately through GEM.
    385     * If the user doesn't care about existing buffer contents and mapping would
    386     * cause us to block, then throw out the old buffer.
    387     *
    388     * If they set INVALIDATE_BUFFER, we can pitch the current contents to
    389     * achieve the required synchronization.
    390     */
    391    if (!(access & GL_MAP_UNSYNCHRONIZED_BIT)) {
    392       if (drm_intel_bo_references(brw->batch.bo, intel_obj->buffer)) {
    393 	 if (access & GL_MAP_INVALIDATE_BUFFER_BIT) {
    394 	    drm_intel_bo_unreference(intel_obj->buffer);
    395 	    alloc_buffer_object(brw, intel_obj);
    396 	 } else {
    397             perf_debug("Stalling on the GPU for mapping a busy buffer "
    398                        "object\n");
    399 	    intel_batchbuffer_flush(brw);
    400 	 }
    401       } else if (drm_intel_bo_busy(intel_obj->buffer) &&
    402 		 (access & GL_MAP_INVALIDATE_BUFFER_BIT)) {
    403 	 drm_intel_bo_unreference(intel_obj->buffer);
    404 	 alloc_buffer_object(brw, intel_obj);
    405       }
    406    }
    407 
    408    /* If the user is mapping a range of an active buffer object but
    409     * doesn't require the current contents of that range, make a new
    410     * BO, and we'll copy what they put in there out at unmap or
    411     * FlushRange time.
    412     *
    413     * That is, unless they're looking for a persistent mapping -- we would
    414     * need to do blits in the MemoryBarrier call, and it's easier to just do a
    415     * GPU stall and do a mapping.
    416     */
    417    if (!(access & (GL_MAP_UNSYNCHRONIZED_BIT | GL_MAP_PERSISTENT_BIT)) &&
    418        (access & GL_MAP_INVALIDATE_RANGE_BIT) &&
    419        drm_intel_bo_busy(intel_obj->buffer)) {
    420       /* Ensure that the base alignment of the allocation meets the alignment
    421        * guarantees the driver has advertised to the application.
    422        */
    423       const unsigned alignment = ctx->Const.MinMapBufferAlignment;
    424 
    425       intel_obj->map_extra[index] = (uintptr_t) offset % alignment;
    426       intel_obj->range_map_bo[index] = drm_intel_bo_alloc(brw->bufmgr,
    427                                                           "BO blit temp",
    428                                                           length +
    429                                                           intel_obj->map_extra[index],
    430                                                           alignment);
    431       if (brw->has_llc) {
    432          brw_bo_map(brw, intel_obj->range_map_bo[index],
    433                     (access & GL_MAP_WRITE_BIT) != 0, "range-map");
    434       } else {
    435          drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo[index]);
    436       }
    437       obj->Mappings[index].Pointer =
    438          intel_obj->range_map_bo[index]->virtual + intel_obj->map_extra[index];
    439       return obj->Mappings[index].Pointer;
    440    }
    441 
    442    if (access & GL_MAP_UNSYNCHRONIZED_BIT) {
    443       if (!brw->has_llc && brw->perf_debug &&
    444           drm_intel_bo_busy(intel_obj->buffer)) {
    445          perf_debug("MapBufferRange with GL_MAP_UNSYNCHRONIZED_BIT stalling (it's actually synchronized on non-LLC platforms)\n");
    446       }
    447       drm_intel_gem_bo_map_unsynchronized(intel_obj->buffer);
    448    } else if (!brw->has_llc && (!(access & GL_MAP_READ_BIT) ||
    449                               (access & GL_MAP_PERSISTENT_BIT))) {
    450       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
    451       mark_buffer_inactive(intel_obj);
    452    } else {
    453       brw_bo_map(brw, intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0,
    454                  "MapBufferRange");
    455       mark_buffer_inactive(intel_obj);
    456    }
    457 
    458    obj->Mappings[index].Pointer = intel_obj->buffer->virtual + offset;
    459    return obj->Mappings[index].Pointer;
    460 }
    461 
    462 /**
    463  * The FlushMappedBufferRange() driver hook.
    464  *
    465  * Implements glFlushMappedBufferRange(), which signifies that modifications
    466  * have been made to a range of a mapped buffer, and it should be flushed.
    467  *
    468  * This is only used for buffers mapped with GL_MAP_FLUSH_EXPLICIT_BIT.
    469  *
    470  * Ideally we'd use a BO to avoid taking up cache space for the temporary
    471  * data, but FlushMappedBufferRange may be followed by further writes to
    472  * the pointer, so we would have to re-map after emitting our blit, which
    473  * would defeat the point.
    474  */
    475 static void
    476 brw_flush_mapped_buffer_range(struct gl_context *ctx,
    477                               GLintptr offset, GLsizeiptr length,
    478                               struct gl_buffer_object *obj,
    479                               gl_map_buffer_index index)
    480 {
    481    struct brw_context *brw = brw_context(ctx);
    482    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
    483 
    484    assert(obj->Mappings[index].AccessFlags & GL_MAP_FLUSH_EXPLICIT_BIT);
    485 
    486    /* If we gave a direct mapping of the buffer instead of using a temporary,
    487     * then there's nothing to do.
    488     */
    489    if (intel_obj->range_map_bo[index] == NULL)
    490       return;
    491 
    492    if (length == 0)
    493       return;
    494 
    495    /* Note that we're not unmapping our buffer while executing the blit.  We
    496     * need to have a mapping still at the end of this call, since the user
    497     * gets to make further modifications and glFlushMappedBufferRange() calls.
    498     * This is safe, because:
    499     *
    500     * - On LLC platforms, we're using a CPU mapping that's coherent with the
    501     *   GPU (except for the render caches), so the kernel doesn't need to do
    502     *   any flushing work for us except for what happens at batch exec time
    503     *   anyway.
    504     *
    505     * - On non-LLC platforms, we're using a GTT mapping that writes directly
    506     *   to system memory (except for the chipset cache that gets flushed at
    507     *   batch exec time).
    508     *
    509     * In both cases we don't need to stall for the previous blit to complete
    510     * so we can re-map (and we definitely don't want to, since that would be
    511     * slow): If the user edits a part of their buffer that's previously been
    512     * blitted, then our lack of synchoronization is fine, because either
    513     * they'll get some too-new data in the first blit and not do another blit
    514     * of that area (but in that case the results are undefined), or they'll do
    515     * another blit of that area and the complete newer data will land the
    516     * second time.
    517     */
    518    intel_emit_linear_blit(brw,
    519 			  intel_obj->buffer,
    520                           obj->Mappings[index].Offset + offset,
    521 			  intel_obj->range_map_bo[index],
    522                           intel_obj->map_extra[index] + offset,
    523 			  length);
    524    mark_buffer_gpu_usage(intel_obj,
    525                          obj->Mappings[index].Offset + offset,
    526                          length);
    527 }
    528 
    529 
    530 /**
    531  * The UnmapBuffer() driver hook.
    532  *
    533  * Implements glUnmapBuffer().
    534  */
    535 static GLboolean
    536 brw_unmap_buffer(struct gl_context *ctx,
    537                  struct gl_buffer_object *obj,
    538                  gl_map_buffer_index index)
    539 {
    540    struct brw_context *brw = brw_context(ctx);
    541    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
    542 
    543    assert(intel_obj);
    544    assert(obj->Mappings[index].Pointer);
    545    if (intel_obj->range_map_bo[index] != NULL) {
    546       drm_intel_bo_unmap(intel_obj->range_map_bo[index]);
    547 
    548       if (!(obj->Mappings[index].AccessFlags & GL_MAP_FLUSH_EXPLICIT_BIT)) {
    549          intel_emit_linear_blit(brw,
    550                                 intel_obj->buffer, obj->Mappings[index].Offset,
    551                                 intel_obj->range_map_bo[index],
    552                                 intel_obj->map_extra[index],
    553                                 obj->Mappings[index].Length);
    554          mark_buffer_gpu_usage(intel_obj, obj->Mappings[index].Offset,
    555                                obj->Mappings[index].Length);
    556       }
    557 
    558       /* Since we've emitted some blits to buffers that will (likely) be used
    559        * in rendering operations in other cache domains in this batch, emit a
    560        * flush.  Once again, we wish for a domain tracker in libdrm to cover
    561        * usage inside of a batchbuffer.
    562        */
    563       brw_emit_mi_flush(brw);
    564 
    565       drm_intel_bo_unreference(intel_obj->range_map_bo[index]);
    566       intel_obj->range_map_bo[index] = NULL;
    567    } else if (intel_obj->buffer != NULL) {
    568       drm_intel_bo_unmap(intel_obj->buffer);
    569    }
    570    obj->Mappings[index].Pointer = NULL;
    571    obj->Mappings[index].Offset = 0;
    572    obj->Mappings[index].Length = 0;
    573 
    574    return true;
    575 }
    576 
    577 /**
    578  * Gets a pointer to the object's BO, and marks the given range as being used
    579  * on the GPU.
    580  *
    581  * Anywhere that uses buffer objects in the pipeline should be using this to
    582  * mark the range of the buffer that is being accessed by the pipeline.
    583  */
    584 drm_intel_bo *
    585 intel_bufferobj_buffer(struct brw_context *brw,
    586                        struct intel_buffer_object *intel_obj,
    587                        uint32_t offset, uint32_t size)
    588 {
    589    /* This is needed so that things like transform feedback and texture buffer
    590     * objects that need a BO but don't want to check that they exist for
    591     * draw-time validation can just always get a BO from a GL buffer object.
    592     */
    593    if (intel_obj->buffer == NULL)
    594       alloc_buffer_object(brw, intel_obj);
    595 
    596    mark_buffer_gpu_usage(intel_obj, offset, size);
    597 
    598    return intel_obj->buffer;
    599 }
    600 
    601 /**
    602  * The CopyBufferSubData() driver hook.
    603  *
    604  * Implements glCopyBufferSubData(), which copies a portion of one buffer
    605  * object's data to another.  Independent source and destination offsets
    606  * are allowed.
    607  */
    608 static void
    609 brw_copy_buffer_subdata(struct gl_context *ctx,
    610                         struct gl_buffer_object *src,
    611                         struct gl_buffer_object *dst,
    612                         GLintptr read_offset, GLintptr write_offset,
    613                         GLsizeiptr size)
    614 {
    615    struct brw_context *brw = brw_context(ctx);
    616    struct intel_buffer_object *intel_src = intel_buffer_object(src);
    617    struct intel_buffer_object *intel_dst = intel_buffer_object(dst);
    618    drm_intel_bo *src_bo, *dst_bo;
    619 
    620    if (size == 0)
    621       return;
    622 
    623    dst_bo = intel_bufferobj_buffer(brw, intel_dst, write_offset, size);
    624    src_bo = intel_bufferobj_buffer(brw, intel_src, read_offset, size);
    625 
    626    intel_emit_linear_blit(brw,
    627 			  dst_bo, write_offset,
    628 			  src_bo, read_offset, size);
    629 
    630    /* Since we've emitted some blits to buffers that will (likely) be used
    631     * in rendering operations in other cache domains in this batch, emit a
    632     * flush.  Once again, we wish for a domain tracker in libdrm to cover
    633     * usage inside of a batchbuffer.
    634     */
    635    brw_emit_mi_flush(brw);
    636 }
    637 
    638 void
    639 intelInitBufferObjectFuncs(struct dd_function_table *functions)
    640 {
    641    functions->NewBufferObject = brw_new_buffer_object;
    642    functions->DeleteBuffer = brw_delete_buffer;
    643    functions->BufferData = brw_buffer_data;
    644    functions->BufferSubData = brw_buffer_subdata;
    645    functions->GetBufferSubData = brw_get_buffer_subdata;
    646    functions->MapBufferRange = brw_map_buffer_range;
    647    functions->FlushMappedBufferRange = brw_flush_mapped_buffer_range;
    648    functions->UnmapBuffer = brw_unmap_buffer;
    649    functions->CopyBufferSubData = brw_copy_buffer_subdata;
    650 }
    651