src/mesa/drivers/dri/intel/intel_buffer_objects.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * 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 above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28
  29 #include "main/imports.h"
  30 #include "main/mtypes.h"
  31 #include "main/macros.h"
  32 #include "main/bufferobj.h"
  33
  34 #include "intel_context.h"
  35 #include "intel_blit.h"
  36 #include "intel_buffer_objects.h"
  37 #include "intel_batchbuffer.h"
  38 #include "intel_regions.h"
  39
  40 static GLboolean
  41 intel_bufferobj_unmap(GLcontext * ctx,
  42                       GLenum target, struct gl_buffer_object *obj);
  43
  44 /** Allocates a new dri_bo to store the data for the buffer object. */
  45 static void
  46 intel_bufferobj_alloc_buffer(struct intel_context *intel,
  47                              struct intel_buffer_object *intel_obj)
  48 {
  49    intel_obj->buffer = dri_bo_alloc(intel->bufmgr, "bufferobj",
  50                                     intel_obj->Base.Size, 64);
  51 }
  52
  53 /**
  54  * There is some duplication between mesa's bufferobjects and our
  55  * bufmgr buffers.  Both have an integer handle and a hashtable to
  56  * lookup an opaque structure.  It would be nice if the handles and
  57  * internal structure where somehow shared.
  58  */
  59 static struct gl_buffer_object *
  60 intel_bufferobj_alloc(GLcontext * ctx, GLuint name, GLenum target)
  61 {
  62    struct intel_buffer_object *obj = CALLOC_STRUCT(intel_buffer_object);
  63
  64    _mesa_initialize_buffer_object(&obj->Base, name, target);
  65
  66    obj->buffer = NULL;
  67
  68    return &obj->Base;
  69 }
  70
  71 /* Break the COW tie to the region.  The region gets to keep the data.
  72  */
  73 void
  74 intel_bufferobj_release_region(struct intel_context *intel,
  75                                struct intel_buffer_object *intel_obj)
  76 {
  77    assert(intel_obj->region->buffer == intel_obj->buffer);
  78    intel_obj->region->pbo = NULL;
  79    intel_obj->region = NULL;
  80
  81    dri_bo_unreference(intel_obj->buffer);
  82    intel_obj->buffer = NULL;
  83 }
  84
  85 /* Break the COW tie to the region.  Both the pbo and the region end
  86  * up with a copy of the data.
  87  */
  88 void
  89 intel_bufferobj_cow(struct intel_context *intel,
  90                     struct intel_buffer_object *intel_obj)
  91 {
  92    assert(intel_obj->region);
  93    intel_region_cow(intel, intel_obj->region);
  94 }
  95
  96
  97 /**
  98  * Deallocate/free a vertex/pixel buffer object.
  99  * Called via glDeleteBuffersARB().
 100  */
 101 static void
 102 intel_bufferobj_free(GLcontext * ctx, struct gl_buffer_object *obj)
 103 {
 104    struct intel_context *intel = intel_context(ctx);
 105    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 106
 107    assert(intel_obj);
 108
 109    /* Buffer objects are automatically unmapped when deleting according
 110     * to the spec, but Mesa doesn't do UnmapBuffer for us at context destroy
 111     * (though it does if you call glDeleteBuffers)
 112     */
 113    if (obj->Pointer)
 114       intel_bufferobj_unmap(ctx, 0, obj);
 115
 116    _mesa_free(intel_obj->sys_buffer);
 117    if (intel_obj->region) {
 118       intel_bufferobj_release_region(intel, intel_obj);
 119    }
 120    else if (intel_obj->buffer) {
 121       dri_bo_unreference(intel_obj->buffer);
 122    }
 123
 124    _mesa_free(intel_obj);
 125 }
 126
 127
 128
 129 /**
 130  * Allocate space for and store data in a buffer object.  Any data that was
 131  * previously stored in the buffer object is lost.  If data is NULL,
 132  * memory will be allocated, but no copy will occur.
 133  * Called via ctx->Driver.BufferData().
 134  * \return GL_TRUE for success, GL_FALSE if out of memory
 135  */
 136 static GLboolean
 137 intel_bufferobj_data(GLcontext * ctx,
 138                      GLenum target,
 139                      GLsizeiptrARB size,
 140                      const GLvoid * data,
 141                      GLenum usage, struct gl_buffer_object *obj)
 142 {
 143    struct intel_context *intel = intel_context(ctx);
 144    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 145
 146    intel_obj->Base.Size = size;
 147    intel_obj->Base.Usage = usage;
 148
 149    assert(!obj->Pointer); /* Mesa should have unmapped it */
 150
 151    if (intel_obj->region)
 152       intel_bufferobj_release_region(intel, intel_obj);
 153
 154    if (intel_obj->buffer != NULL) {
 155       dri_bo_unreference(intel_obj->buffer);
 156       intel_obj->buffer = NULL;
 157    }
 158    _mesa_free(intel_obj->sys_buffer);
 159    intel_obj->sys_buffer = NULL;
 160
 161    if (size != 0) {
 162 #ifdef I915
 163       /* On pre-965, stick VBOs in system memory, as we're always doing swtnl
 164        * with their contents anyway.
 165        */
 166       if (target == GL_ARRAY_BUFFER || target == GL_ELEMENT_ARRAY_BUFFER) {
 167          intel_obj->sys_buffer = _mesa_malloc(size);
 168          if (intel_obj->sys_buffer != NULL) {
 169             if (data != NULL)
 170                memcpy(intel_obj->sys_buffer, data, size);
 171             return GL_TRUE;
 172          }
 173       }
 174 #endif
 175       intel_bufferobj_alloc_buffer(intel, intel_obj);
 176       if (!intel_obj->buffer)
 177          return GL_FALSE;
 178
 179       if (data != NULL)
 180          dri_bo_subdata(intel_obj->buffer, 0, size, data);
 181    }
 182
 183    return GL_TRUE;
 184 }
 185
 186
 187 /**
 188  * Replace data in a subrange of buffer object.  If the data range
 189  * specified by size + offset extends beyond the end of the buffer or
 190  * if data is NULL, no copy is performed.
 191  * Called via glBufferSubDataARB().
 192  */
 193 static void
 194 intel_bufferobj_subdata(GLcontext * ctx,
 195                         GLenum target,
 196                         GLintptrARB offset,
 197                         GLsizeiptrARB size,
 198                         const GLvoid * data, struct gl_buffer_object *obj)
 199 {
 200    struct intel_context *intel = intel_context(ctx);
 201    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 202
 203    assert(intel_obj);
 204
 205    if (intel_obj->region)
 206       intel_bufferobj_cow(intel, intel_obj);
 207
 208    if (intel_obj->sys_buffer)
 209       memcpy((char *)intel_obj->sys_buffer + offset, data, size);
 210    else {
 211       /* Flush any existing batchbuffer that might reference this data. */
 212       intelFlush(ctx);
 213
 214       dri_bo_subdata(intel_obj->buffer, offset, size, data);
 215    }
 216 }
 217
 218
 219 /**
 220  * Called via glGetBufferSubDataARB().
 221  */
 222 static void
 223 intel_bufferobj_get_subdata(GLcontext * ctx,
 224                             GLenum target,
 225                             GLintptrARB offset,
 226                             GLsizeiptrARB size,
 227                             GLvoid * data, struct gl_buffer_object *obj)
 228 {
 229    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 230
 231    assert(intel_obj);
 232    if (intel_obj->sys_buffer)
 233       memcpy(data, (char *)intel_obj->sys_buffer + offset, size);
 234    else
 235       dri_bo_get_subdata(intel_obj->buffer, offset, size, data);
 236 }
 237
 238
 239
 240 /**
 241  * Called via glMapBufferARB().
 242  */
 243 static void *
 244 intel_bufferobj_map(GLcontext * ctx,
 245                     GLenum target,
 246                     GLenum access, struct gl_buffer_object *obj)
 247 {
 248    struct intel_context *intel = intel_context(ctx);
 249    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 250    GLboolean read_only = (access == GL_READ_ONLY_ARB);
 251    GLboolean write_only = (access == GL_WRITE_ONLY_ARB);
 252
 253    assert(intel_obj);
 254
 255    if (intel_obj->sys_buffer) {
 256       obj->Pointer = intel_obj->sys_buffer;
 257       return obj->Pointer;
 258    }
 259
 260    /* Flush any existing batchbuffer that might have written to this
 261     * buffer.
 262     */
 263    intelFlush(ctx);
 264
 265    if (intel_obj->region)
 266       intel_bufferobj_cow(intel, intel_obj);
 267
 268    if (intel_obj->buffer == NULL) {
 269       obj->Pointer = NULL;
 270       return NULL;
 271    }
 272
 273    if (write_only && intel->intelScreen->kernel_exec_fencing) {
 274       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
 275       intel_obj->mapped_gtt = GL_TRUE;
 276    } else {
 277       drm_intel_bo_map(intel_obj->buffer, !read_only);
 278       intel_obj->mapped_gtt = GL_FALSE;
 279    }
 280
 281    obj->Pointer = intel_obj->buffer->virtual;
 282    obj->Length = obj->Size;
 283    obj->Offset = 0;
 284
 285    return obj->Pointer;
 286 }
 287
 288 /**
 289  * Called via glMapBufferRange().
 290  *
 291  * The goal of this extension is to allow apps to accumulate their rendering
 292  * at the same time as they accumulate their buffer object.  Without it,
 293  * you'd end up blocking on execution of rendering every time you mapped
 294  * the buffer to put new data in.
 295  *
 296  * We support it in 3 ways: If unsynchronized, then don't bother
 297  * flushing the batchbuffer before mapping the buffer, which can save blocking
 298  * in many cases.  If we would still block, and they allow the whole buffer
 299  * to be invalidated, then just allocate a new buffer to replace the old one.
 300  * If not, and we'd block, and they allow the subrange of the buffer to be
 301  * invalidated, then we can make a new little BO, let them write into that,
 302  * and blit it into the real BO at unmap time.
 303  */
 304 static void *
 305 intel_bufferobj_map_range(GLcontext * ctx,
 306                           GLenum target, GLintptr offset, GLsizeiptr length,
 307                           GLbitfield access, struct gl_buffer_object *obj)
 308 {
 309    struct intel_context *intel = intel_context(ctx);
 310    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 311
 312    assert(intel_obj);
 313
 314    /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
 315     * internally uses our functions directly.
 316     */
 317    obj->Offset = offset;
 318    obj->Length = length;
 319    obj->AccessFlags = access;
 320
 321    if (intel_obj->sys_buffer) {
 322       obj->Pointer = intel_obj->sys_buffer + offset;
 323       return obj->Pointer;
 324    }
 325
 326    if (intel_obj->region)
 327       intel_bufferobj_cow(intel, intel_obj);
 328
 329    /* If the mapping is synchronized with other GL operations, flush
 330     * the batchbuffer so that GEM knows about the buffer access for later
 331     * syncing.
 332     */
 333    if (!(access & GL_MAP_UNSYNCHRONIZED_BIT))
 334       intelFlush(ctx);
 335
 336    if (intel_obj->buffer == NULL) {
 337       obj->Pointer = NULL;
 338       return NULL;
 339    }
 340
 341    /* If the user doesn't care about existing buffer contents and mapping
 342     * would cause us to block, then throw out the old buffer.
 343     */
 344    if (!(access & GL_MAP_UNSYNCHRONIZED_BIT) &&
 345        (access & GL_MAP_INVALIDATE_BUFFER_BIT) &&
 346        drm_intel_bo_busy(intel_obj->buffer)) {
 347       drm_intel_bo_unreference(intel_obj->buffer);
 348       intel_obj->buffer = dri_bo_alloc(intel->bufmgr, "bufferobj",
 349                                        intel_obj->Base.Size, 64);
 350    }
 351
 352    /* If the user is mapping a range of an active buffer object but
 353     * doesn't require the current contents of that range, make a new
 354     * BO, and we'll copy what they put in there out at unmap or
 355     * FlushRange time.
 356     */
 357    if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
 358        drm_intel_bo_busy(intel_obj->buffer)) {
 359       if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
 360          intel_obj->range_map_buffer = _mesa_malloc(length);
 361          obj->Pointer = intel_obj->range_map_buffer;
 362       } else {
 363          intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
 364                                                       "range map",
 365                                                       length, 64);
 366          if (!(access & GL_MAP_READ_BIT) &&
 367              intel->intelScreen->kernel_exec_fencing) {
 368             drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
 369             intel_obj->mapped_gtt = GL_TRUE;
 370          } else {
 371             drm_intel_bo_map(intel_obj->range_map_bo,
 372                              (access & GL_MAP_WRITE_BIT) != 0);
 373             intel_obj->mapped_gtt = GL_FALSE;
 374          }
 375          obj->Pointer = intel_obj->range_map_bo->virtual;
 376       }
 377       return obj->Pointer;
 378    }
 379
 380    if (!(access & GL_MAP_READ_BIT) &&
 381        intel->intelScreen->kernel_exec_fencing) {
 382       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
 383       intel_obj->mapped_gtt = GL_TRUE;
 384    } else {
 385       drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
 386       intel_obj->mapped_gtt = GL_FALSE;
 387    }
 388
 389    obj->Pointer = intel_obj->buffer->virtual + offset;
 390    return obj->Pointer;
 391 }
 392
 393 /* Ideally we'd use a BO to avoid taking up cache space for the temporary
 394  * data, but FlushMappedBufferRange may be followed by further writes to
 395  * the pointer, so we would have to re-map after emitting our blit, which
 396  * would defeat the point.
 397  */
 398 static void
 399 intel_bufferobj_flush_mapped_range(GLcontext *ctx, GLenum target,
 400                                    GLintptr offset, GLsizeiptr length,
 401                                    struct gl_buffer_object *obj)
 402 {
 403    struct intel_context *intel = intel_context(ctx);
 404    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 405    drm_intel_bo *temp_bo;
 406
 407    /* Unless we're in the range map using a temporary system buffer,
 408     * there's no work to do.
 409     */
 410    if (intel_obj->range_map_buffer == NULL)
 411       return;
 412
 413    temp_bo = drm_intel_bo_alloc(intel->bufmgr, "range map flush", length, 64);
 414
 415    drm_intel_bo_subdata(temp_bo, 0, length, intel_obj->range_map_buffer);
 416
 417    intel_emit_linear_blit(intel,
 418                           intel_obj->buffer, obj->Offset + offset,
 419                           temp_bo, 0,
 420                           length);
 421
 422    drm_intel_bo_unreference(temp_bo);
 423 }
 424
 425
 426 /**
 427  * Called via glUnmapBuffer().
 428  */
 429 static GLboolean
 430 intel_bufferobj_unmap(GLcontext * ctx,
 431                       GLenum target, struct gl_buffer_object *obj)
 432 {
 433    struct intel_context *intel = intel_context(ctx);
 434    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 435
 436    assert(intel_obj);
 437    assert(obj->Pointer);
 438    if (intel_obj->sys_buffer != NULL) {
 439       /* always keep the mapping around. */
 440    } else if (intel_obj->range_map_buffer != NULL) {
 441       /* Since we've emitted some blits to buffers that will (likely) be used
 442        * in rendering operations in other cache domains in this batch, emit a
 443        * flush.  Once again, we wish for a domain tracker in libdrm to cover
 444        * usage inside of a batchbuffer.
 445        */
 446       intel_batchbuffer_emit_mi_flush(intel->batch);
 447       free(intel_obj->range_map_buffer);
 448       intel_obj->range_map_buffer = NULL;
 449    } else if (intel_obj->range_map_bo != NULL) {
 450       if (intel_obj->mapped_gtt) {
 451          drm_intel_gem_bo_unmap_gtt(intel_obj->range_map_bo);
 452       } else {
 453          drm_intel_bo_unmap(intel_obj->range_map_bo);
 454       }
 455
 456       intel_emit_linear_blit(intel,
 457                              intel_obj->buffer, obj->Offset,
 458                              intel_obj->range_map_bo, 0,
 459                              obj->Length);
 460
 461       /* Since we've emitted some blits to buffers that will (likely) be used
 462        * in rendering operations in other cache domains in this batch, emit a
 463        * flush.  Once again, we wish for a domain tracker in libdrm to cover
 464        * usage inside of a batchbuffer.
 465        */
 466       intel_batchbuffer_emit_mi_flush(intel->batch);
 467
 468       drm_intel_bo_unreference(intel_obj->range_map_bo);
 469       intel_obj->range_map_bo = NULL;
 470    } else if (intel_obj->buffer != NULL) {
 471       if (intel_obj->mapped_gtt) {
 472          drm_intel_gem_bo_unmap_gtt(intel_obj->buffer);
 473       } else {
 474          drm_intel_bo_unmap(intel_obj->buffer);
 475       }
 476    }
 477    obj->Pointer = NULL;
 478    obj->Offset = 0;
 479    obj->Length = 0;
 480
 481    return GL_TRUE;
 482 }
 483
 484 dri_bo *
 485 intel_bufferobj_buffer(struct intel_context *intel,
 486                        struct intel_buffer_object *intel_obj, GLuint flag)
 487 {
 488    if (intel_obj->region) {
 489       if (flag == INTEL_WRITE_PART)
 490          intel_bufferobj_cow(intel, intel_obj);
 491       else if (flag == INTEL_WRITE_FULL) {
 492          intel_bufferobj_release_region(intel, intel_obj);
 493          intel_bufferobj_alloc_buffer(intel, intel_obj);
 494       }
 495    }
 496
 497    if (intel_obj->buffer == NULL) {
 498       void *sys_buffer = intel_obj->sys_buffer;
 499
 500       /* only one of buffer and sys_buffer could be non-NULL */
 501       intel_bufferobj_alloc_buffer(intel, intel_obj);
 502       intel_obj->sys_buffer = NULL;
 503
 504       intel_bufferobj_subdata(&intel->ctx,
 505                               GL_ARRAY_BUFFER_ARB,
 506                               0,
 507                               intel_obj->Base.Size,
 508                               sys_buffer,
 509                               &intel_obj->Base);
 510       _mesa_free(sys_buffer);
 511       intel_obj->sys_buffer = NULL;
 512    }
 513
 514    return intel_obj->buffer;
 515 }
 516
 517 static void
 518 intel_bufferobj_copy_subdata(GLcontext *ctx,
 519                              struct gl_buffer_object *src,
 520                              struct gl_buffer_object *dst,
 521                              GLintptr read_offset, GLintptr write_offset,
 522                              GLsizeiptr size)
 523 {
 524    struct intel_context *intel = intel_context(ctx);
 525    struct intel_buffer_object *intel_src = intel_buffer_object(src);
 526    struct intel_buffer_object *intel_dst = intel_buffer_object(dst);
 527    drm_intel_bo *src_bo, *dst_bo;
 528
 529    if (size == 0)
 530       return;
 531
 532    /* If we're in system memory, just map and memcpy. */
 533    if (intel_src->sys_buffer || intel_dst->sys_buffer) {
 534       /* The same buffer may be used, but note that regions copied may
 535        * not overlap.
 536        */
 537       if (src == dst) {
 538          char *ptr = intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
 539                                          GL_READ_WRITE, dst);
 540          memcpy(ptr + write_offset, ptr + read_offset, size);
 541          intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
 542       } else {
 543          const char *src_ptr;
 544          char *dst_ptr;
 545
 546          src_ptr =  intel_bufferobj_map(ctx, GL_COPY_READ_BUFFER,
 547                                         GL_READ_ONLY, src);
 548          dst_ptr =  intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
 549                                         GL_WRITE_ONLY, dst);
 550
 551          memcpy(dst_ptr + write_offset, src_ptr + read_offset, size);
 552
 553          intel_bufferobj_unmap(ctx, GL_COPY_READ_BUFFER, src);
 554          intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
 555       }
 556    }
 557
 558    /* Otherwise, we have real BOs, so blit them. */
 559
 560    dst_bo = intel_bufferobj_buffer(intel, intel_dst, INTEL_WRITE_PART);
 561    src_bo = intel_bufferobj_buffer(intel, intel_src, INTEL_READ);
 562
 563    intel_emit_linear_blit(intel,
 564                           dst_bo, write_offset,
 565                           src_bo, read_offset, size);
 566
 567    /* Since we've emitted some blits to buffers that will (likely) be used
 568     * in rendering operations in other cache domains in this batch, emit a
 569     * flush.  Once again, we wish for a domain tracker in libdrm to cover
 570     * usage inside of a batchbuffer.
 571     */
 572    intel_batchbuffer_emit_mi_flush(intel->batch);
 573 }
 574
 575 void
 576 intelInitBufferObjectFuncs(struct dd_function_table *functions)
 577 {
 578    functions->NewBufferObject = intel_bufferobj_alloc;
 579    functions->DeleteBuffer = intel_bufferobj_free;
 580    functions->BufferData = intel_bufferobj_data;
 581    functions->BufferSubData = intel_bufferobj_subdata;
 582    functions->GetBufferSubData = intel_bufferobj_get_subdata;
 583    functions->MapBuffer = intel_bufferobj_map;
 584    functions->MapBufferRange = intel_bufferobj_map_range;
 585    functions->FlushMappedBufferRange = intel_bufferobj_flush_mapped_range;
 586    functions->UnmapBuffer = intel_bufferobj_unmap;
 587    functions->CopyBufferSubData = intel_bufferobj_copy_subdata;
 588 }