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    dri_bo_get_subdata(intel_obj->buffer, offset, size, data);
 233 }
 234
 235
 236
 237 /**
 238  * Called via glMapBufferARB().
 239  */
 240 static void *
 241 intel_bufferobj_map(GLcontext * ctx,
 242                     GLenum target,
 243                     GLenum access, struct gl_buffer_object *obj)
 244 {
 245    struct intel_context *intel = intel_context(ctx);
 246    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 247    GLboolean read_only = (access == GL_READ_ONLY_ARB);
 248    GLboolean write_only = (access == GL_WRITE_ONLY_ARB);
 249
 250    assert(intel_obj);
 251
 252    if (intel_obj->sys_buffer) {
 253       obj->Pointer = intel_obj->sys_buffer;
 254       return obj->Pointer;
 255    }
 256
 257    /* Flush any existing batchbuffer that might have written to this
 258     * buffer.
 259     */
 260    intelFlush(ctx);
 261
 262    if (intel_obj->region)
 263       intel_bufferobj_cow(intel, intel_obj);
 264
 265    if (intel_obj->buffer == NULL) {
 266       obj->Pointer = NULL;
 267       return NULL;
 268    }
 269
 270    if (write_only && intel->intelScreen->kernel_exec_fencing) {
 271       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
 272       intel_obj->mapped_gtt = GL_TRUE;
 273    } else {
 274       drm_intel_bo_map(intel_obj->buffer, !read_only);
 275       intel_obj->mapped_gtt = GL_FALSE;
 276    }
 277
 278    obj->Pointer = intel_obj->buffer->virtual;
 279    obj->Length = obj->Size;
 280    obj->Offset = 0;
 281
 282    return obj->Pointer;
 283 }
 284
 285 /**
 286  * Called via glMapBufferRange().
 287  *
 288  * The goal of this extension is to allow apps to accumulate their rendering
 289  * at the same time as they accumulate their buffer object.  Without it,
 290  * you'd end up blocking on execution of rendering every time you mapped
 291  * the buffer to put new data in.
 292  *
 293  * We support it in 3 ways: If unsynchronized, then don't bother
 294  * flushing the batchbuffer before mapping the buffer, which can save blocking
 295  * in many cases.  If we would still block, and they allow the whole buffer
 296  * to be invalidated, then just allocate a new buffer to replace the old one.
 297  * If not, and we'd block, and they allow the subrange of the buffer to be
 298  * invalidated, then we can make a new little BO, let them write into that,
 299  * and blit it into the real BO at unmap time.
 300  */
 301 static void *
 302 intel_bufferobj_map_range(GLcontext * ctx,
 303                           GLenum target, GLintptr offset, GLsizeiptr length,
 304                           GLbitfield access, struct gl_buffer_object *obj)
 305 {
 306    struct intel_context *intel = intel_context(ctx);
 307    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 308
 309    assert(intel_obj);
 310
 311    /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
 312     * internally uses our functions directly.
 313     */
 314    obj->Offset = offset;
 315    obj->Length = length;
 316    obj->AccessFlags = access;
 317
 318    if (intel_obj->sys_buffer) {
 319       obj->Pointer = intel_obj->sys_buffer + offset;
 320       return obj->Pointer;
 321    }
 322
 323    if (intel_obj->region)
 324       intel_bufferobj_cow(intel, intel_obj);
 325
 326    /* If the mapping is synchronized with other GL operations, flush
 327     * the batchbuffer so that GEM knows about the buffer access for later
 328     * syncing.
 329     */
 330    if (!(access & GL_MAP_UNSYNCHRONIZED_BIT))
 331       intelFlush(ctx);
 332
 333    if (intel_obj->buffer == NULL) {
 334       obj->Pointer = NULL;
 335       return NULL;
 336    }
 337
 338    /* If the user doesn't care about existing buffer contents and mapping
 339     * would cause us to block, then throw out the old buffer.
 340     */
 341    if (!(access & GL_MAP_UNSYNCHRONIZED_BIT) &&
 342        (access & GL_MAP_INVALIDATE_BUFFER_BIT) &&
 343        drm_intel_bo_busy(intel_obj->buffer)) {
 344       drm_intel_bo_unreference(intel_obj->buffer);
 345       intel_obj->buffer = dri_bo_alloc(intel->bufmgr, "bufferobj",
 346                                        intel_obj->Base.Size, 64);
 347    }
 348
 349    /* If the user is mapping a range of an active buffer object but
 350     * doesn't require the current contents of that range, make a new
 351     * BO, and we'll copy what they put in there out at unmap or
 352     * FlushRange time.
 353     */
 354    if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
 355        drm_intel_bo_busy(intel_obj->buffer)) {
 356       if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
 357          intel_obj->range_map_buffer = _mesa_malloc(length);
 358          obj->Pointer = intel_obj->range_map_buffer;
 359       } else {
 360          intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
 361                                                       "range map",
 362                                                       length, 64);
 363          if (!(access & GL_MAP_READ_BIT) &&
 364              intel->intelScreen->kernel_exec_fencing) {
 365             drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
 366             intel_obj->mapped_gtt = GL_TRUE;
 367          } else {
 368             drm_intel_bo_map(intel_obj->range_map_bo,
 369                              (access & GL_MAP_WRITE_BIT) != 0);
 370             intel_obj->mapped_gtt = GL_FALSE;
 371          }
 372          obj->Pointer = intel_obj->range_map_bo->virtual;
 373       }
 374       return obj->Pointer;
 375    }
 376
 377    if (!(access & GL_MAP_READ_BIT) &&
 378        intel->intelScreen->kernel_exec_fencing) {
 379       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
 380       intel_obj->mapped_gtt = GL_TRUE;
 381    } else {
 382       drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
 383       intel_obj->mapped_gtt = GL_FALSE;
 384    }
 385
 386    obj->Pointer = intel_obj->buffer->virtual + offset;
 387    return obj->Pointer;
 388 }
 389
 390 /* Ideally we'd use a BO to avoid taking up cache space for the temporary
 391  * data, but FlushMappedBufferRange may be followed by further writes to
 392  * the pointer, so we would have to re-map after emitting our blit, which
 393  * would defeat the point.
 394  */
 395 static void
 396 intel_bufferobj_flush_mapped_range(GLcontext *ctx, GLenum target,
 397                                    GLintptr offset, GLsizeiptr length,
 398                                    struct gl_buffer_object *obj)
 399 {
 400    struct intel_context *intel = intel_context(ctx);
 401    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 402    drm_intel_bo *temp_bo;
 403
 404    /* Unless we're in the range map using a temporary system buffer,
 405     * there's no work to do.
 406     */
 407    if (intel_obj->range_map_buffer == NULL)
 408       return;
 409
 410    temp_bo = drm_intel_bo_alloc(intel->bufmgr, "range map flush", length, 64);
 411
 412    drm_intel_bo_subdata(temp_bo, 0, length, intel_obj->range_map_buffer);
 413
 414    intel_emit_linear_blit(intel,
 415                           intel_obj->buffer, obj->Offset + offset,
 416                           temp_bo, 0,
 417                           length);
 418
 419    drm_intel_bo_unreference(temp_bo);
 420 }
 421
 422
 423 /**
 424  * Called via glUnmapBuffer().
 425  */
 426 static GLboolean
 427 intel_bufferobj_unmap(GLcontext * ctx,
 428                       GLenum target, struct gl_buffer_object *obj)
 429 {
 430    struct intel_context *intel = intel_context(ctx);
 431    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 432
 433    assert(intel_obj);
 434    assert(obj->Pointer);
 435    if (intel_obj->sys_buffer != NULL) {
 436       /* always keep the mapping around. */
 437    } else if (intel_obj->range_map_buffer != NULL) {
 438       /* Since we've emitted some blits to buffers that will (likely) be used
 439        * in rendering operations in other cache domains in this batch, emit a
 440        * flush.  Once again, we wish for a domain tracker in libdrm to cover
 441        * usage inside of a batchbuffer.
 442        */
 443       intel_batchbuffer_emit_mi_flush(intel->batch);
 444       free(intel_obj->range_map_buffer);
 445       intel_obj->range_map_buffer = NULL;
 446    } else if (intel_obj->range_map_bo != NULL) {
 447       if (intel_obj->mapped_gtt) {
 448          drm_intel_gem_bo_unmap_gtt(intel_obj->range_map_bo);
 449       } else {
 450          drm_intel_bo_unmap(intel_obj->range_map_bo);
 451       }
 452
 453       intel_emit_linear_blit(intel,
 454                              intel_obj->buffer, obj->Offset,
 455                              intel_obj->range_map_bo, 0,
 456                              obj->Length);
 457
 458       /* Since we've emitted some blits to buffers that will (likely) be used
 459        * in rendering operations in other cache domains in this batch, emit a
 460        * flush.  Once again, we wish for a domain tracker in libdrm to cover
 461        * usage inside of a batchbuffer.
 462        */
 463       intel_batchbuffer_emit_mi_flush(intel->batch);
 464
 465       drm_intel_bo_unreference(intel_obj->range_map_bo);
 466       intel_obj->range_map_bo = NULL;
 467    } else if (intel_obj->buffer != NULL) {
 468       if (intel_obj->mapped_gtt) {
 469          drm_intel_gem_bo_unmap_gtt(intel_obj->buffer);
 470       } else {
 471          drm_intel_bo_unmap(intel_obj->buffer);
 472       }
 473    }
 474    obj->Pointer = NULL;
 475    obj->Offset = 0;
 476    obj->Length = 0;
 477
 478    return GL_TRUE;
 479 }
 480
 481 dri_bo *
 482 intel_bufferobj_buffer(struct intel_context *intel,
 483                        struct intel_buffer_object *intel_obj, GLuint flag)
 484 {
 485    if (intel_obj->region) {
 486       if (flag == INTEL_WRITE_PART)
 487          intel_bufferobj_cow(intel, intel_obj);
 488       else if (flag == INTEL_WRITE_FULL) {
 489          intel_bufferobj_release_region(intel, intel_obj);
 490          intel_bufferobj_alloc_buffer(intel, intel_obj);
 491       }
 492    }
 493
 494    if (intel_obj->buffer == NULL) {
 495       void *sys_buffer = intel_obj->sys_buffer;
 496
 497       /* only one of buffer and sys_buffer could be non-NULL */
 498       intel_bufferobj_alloc_buffer(intel, intel_obj);
 499       intel_obj->sys_buffer = NULL;
 500
 501       intel_bufferobj_subdata(&intel->ctx,
 502                               GL_ARRAY_BUFFER_ARB,
 503                               0,
 504                               intel_obj->Base.Size,
 505                               sys_buffer,
 506                               &intel_obj->Base);
 507       _mesa_free(sys_buffer);
 508       intel_obj->sys_buffer = NULL;
 509    }
 510
 511    return intel_obj->buffer;
 512 }
 513
 514 static void
 515 intel_bufferobj_copy_subdata(GLcontext *ctx,
 516                              struct gl_buffer_object *src,
 517                              struct gl_buffer_object *dst,
 518                              GLintptr read_offset, GLintptr write_offset,
 519                              GLsizeiptr size)
 520 {
 521    struct intel_context *intel = intel_context(ctx);
 522    struct intel_buffer_object *intel_src = intel_buffer_object(src);
 523    struct intel_buffer_object *intel_dst = intel_buffer_object(dst);
 524    drm_intel_bo *src_bo, *dst_bo;
 525
 526    if (size == 0)
 527       return;
 528
 529    /* If we're in system memory, just map and memcpy. */
 530    if (intel_src->sys_buffer || intel_dst->sys_buffer) {
 531       /* The same buffer may be used, but note that regions copied may
 532        * not overlap.
 533        */
 534       if (src == dst) {
 535          char *ptr = intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
 536                                          GL_READ_WRITE, dst);
 537          memcpy(ptr + write_offset, ptr + read_offset, size);
 538          intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
 539       } else {
 540          const char *src_ptr;
 541          char *dst_ptr;
 542
 543          src_ptr =  intel_bufferobj_map(ctx, GL_COPY_READ_BUFFER,
 544                                         GL_READ_ONLY, src);
 545          dst_ptr =  intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
 546                                         GL_WRITE_ONLY, dst);
 547
 548          memcpy(dst_ptr + write_offset, src_ptr + read_offset, size);
 549
 550          intel_bufferobj_unmap(ctx, GL_COPY_READ_BUFFER, src);
 551          intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
 552       }
 553    }
 554
 555    /* Otherwise, we have real BOs, so blit them. */
 556
 557    dst_bo = intel_bufferobj_buffer(intel, intel_dst, INTEL_WRITE_PART);
 558    src_bo = intel_bufferobj_buffer(intel, intel_src, INTEL_READ);
 559
 560    intel_emit_linear_blit(intel,
 561                           dst_bo, write_offset,
 562                           src_bo, read_offset, size);
 563
 564    /* Since we've emitted some blits to buffers that will (likely) be used
 565     * in rendering operations in other cache domains in this batch, emit a
 566     * flush.  Once again, we wish for a domain tracker in libdrm to cover
 567     * usage inside of a batchbuffer.
 568     */
 569    intel_batchbuffer_emit_mi_flush(intel->batch);
 570 }
 571
 572 void
 573 intelInitBufferObjectFuncs(struct dd_function_table *functions)
 574 {
 575    functions->NewBufferObject = intel_bufferobj_alloc;
 576    functions->DeleteBuffer = intel_bufferobj_free;
 577    functions->BufferData = intel_bufferobj_data;
 578    functions->BufferSubData = intel_bufferobj_subdata;
 579    functions->GetBufferSubData = intel_bufferobj_get_subdata;
 580    functions->MapBuffer = intel_bufferobj_map;
 581    functions->MapBufferRange = intel_bufferobj_map_range;
 582    functions->FlushMappedBufferRange = intel_bufferobj_flush_mapped_range;
 583    functions->UnmapBuffer = intel_bufferobj_unmap;
 584    functions->CopyBufferSubData = intel_bufferobj_copy_subdata;
 585 }