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       dri_bo_subdata(intel_obj->buffer, offset, size, data);
 212 }
 213
 214
 215 /**
 216  * Called via glGetBufferSubDataARB().
 217  */
 218 static void
 219 intel_bufferobj_get_subdata(GLcontext * ctx,
 220                             GLenum target,
 221                             GLintptrARB offset,
 222                             GLsizeiptrARB size,
 223                             GLvoid * data, struct gl_buffer_object *obj)
 224 {
 225    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 226
 227    assert(intel_obj);
 228    dri_bo_get_subdata(intel_obj->buffer, offset, size, data);
 229 }
 230
 231
 232
 233 /**
 234  * Called via glMapBufferARB().
 235  */
 236 static void *
 237 intel_bufferobj_map(GLcontext * ctx,
 238                     GLenum target,
 239                     GLenum access, struct gl_buffer_object *obj)
 240 {
 241    struct intel_context *intel = intel_context(ctx);
 242    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 243    GLboolean read_only = (access == GL_READ_ONLY_ARB);
 244    GLboolean write_only = (access == GL_WRITE_ONLY_ARB);
 245
 246    assert(intel_obj);
 247
 248    if (intel_obj->sys_buffer) {
 249       obj->Pointer = intel_obj->sys_buffer;
 250       return obj->Pointer;
 251    }
 252
 253    /* Flush any existing batchbuffer that might have written to this
 254     * buffer.
 255     */
 256    intelFlush(ctx);
 257
 258    if (intel_obj->region)
 259       intel_bufferobj_cow(intel, intel_obj);
 260
 261    if (intel_obj->buffer == NULL) {
 262       obj->Pointer = NULL;
 263       return NULL;
 264    }
 265
 266    if (write_only && intel->intelScreen->kernel_exec_fencing) {
 267       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
 268       intel_obj->mapped_gtt = GL_TRUE;
 269    } else {
 270       drm_intel_bo_map(intel_obj->buffer, !read_only);
 271       intel_obj->mapped_gtt = GL_FALSE;
 272    }
 273
 274    obj->Pointer = intel_obj->buffer->virtual;
 275    obj->Length = obj->Size;
 276    obj->Offset = 0;
 277
 278    return obj->Pointer;
 279 }
 280
 281 /**
 282  * Called via glMapBufferRange().
 283  *
 284  * The goal of this extension is to allow apps to accumulate their rendering
 285  * at the same time as they accumulate their buffer object.  Without it,
 286  * you'd end up blocking on execution of rendering every time you mapped
 287  * the buffer to put new data in.
 288  *
 289  * We support it in 3 ways: If unsynchronized, then don't bother
 290  * flushing the batchbuffer before mapping the buffer, which can save blocking
 291  * in many cases.  If we would still block, and they allow the whole buffer
 292  * to be invalidated, then just allocate a new buffer to replace the old one.
 293  * If not, and we'd block, and they allow the subrange of the buffer to be
 294  * invalidated, then we can make a new little BO, let them write into that,
 295  * and blit it into the real BO at unmap time.
 296  */
 297 static void *
 298 intel_bufferobj_map_range(GLcontext * ctx,
 299                           GLenum target, GLintptr offset, GLsizeiptr length,
 300                           GLbitfield access, struct gl_buffer_object *obj)
 301 {
 302    struct intel_context *intel = intel_context(ctx);
 303    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 304
 305    assert(intel_obj);
 306
 307    /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
 308     * internally uses our functions directly.
 309     */
 310    obj->Offset = offset;
 311    obj->Length = length;
 312    obj->AccessFlags = access;
 313
 314    if (intel_obj->sys_buffer) {
 315       obj->Pointer = intel_obj->sys_buffer + offset;
 316       return obj->Pointer;
 317    }
 318
 319    if (intel_obj->region)
 320       intel_bufferobj_cow(intel, intel_obj);
 321
 322    /* If the mapping is synchronized with other GL operations, flush
 323     * the batchbuffer so that GEM knows about the buffer access for later
 324     * syncing.
 325     */
 326    if (!(access & GL_MAP_UNSYNCHRONIZED_BIT))
 327       intelFlush(ctx);
 328
 329    if (intel_obj->buffer == NULL) {
 330       obj->Pointer = NULL;
 331       return NULL;
 332    }
 333
 334    /* If the user doesn't care about existing buffer contents and mapping
 335     * would cause us to block, then throw out the old buffer.
 336     */
 337    if (!(access & GL_MAP_UNSYNCHRONIZED_BIT) &&
 338        (access & GL_MAP_INVALIDATE_BUFFER_BIT) &&
 339        drm_intel_bo_busy(intel_obj->buffer)) {
 340       drm_intel_bo_unreference(intel_obj->buffer);
 341       intel_obj->buffer = dri_bo_alloc(intel->bufmgr, "bufferobj",
 342                                        intel_obj->Base.Size, 64);
 343    }
 344
 345    /* If the user is mapping a range of an active buffer object but
 346     * doesn't require the current contents of that range, make a new
 347     * BO, and we'll copy what they put in there out at unmap or
 348     * FlushRange time.
 349     */
 350    if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
 351        drm_intel_bo_busy(intel_obj->buffer)) {
 352       if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
 353          intel_obj->range_map_buffer = _mesa_malloc(length);
 354          obj->Pointer = intel_obj->range_map_buffer;
 355       } else {
 356          intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
 357                                                       "range map",
 358                                                       length, 64);
 359          if (!(access & GL_MAP_READ_BIT) &&
 360              intel->intelScreen->kernel_exec_fencing) {
 361             drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
 362             intel_obj->mapped_gtt = GL_TRUE;
 363          } else {
 364             drm_intel_bo_map(intel_obj->range_map_bo,
 365                              (access & GL_MAP_WRITE_BIT) != 0);
 366             intel_obj->mapped_gtt = GL_FALSE;
 367          }
 368          obj->Pointer = intel_obj->range_map_bo->virtual;
 369       }
 370       return obj->Pointer;
 371    }
 372
 373    if (!(access & GL_MAP_READ_BIT) &&
 374        intel->intelScreen->kernel_exec_fencing) {
 375       drm_intel_gem_bo_map_gtt(intel_obj->buffer);
 376       intel_obj->mapped_gtt = GL_TRUE;
 377    } else {
 378       drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
 379       intel_obj->mapped_gtt = GL_FALSE;
 380    }
 381
 382    obj->Pointer = intel_obj->buffer->virtual + offset;
 383    return obj->Pointer;
 384 }
 385
 386 /* Ideally we'd use a BO to avoid taking up cache space for the temporary
 387  * data, but FlushMappedBufferRange may be followed by further writes to
 388  * the pointer, so we would have to re-map after emitting our blit, which
 389  * would defeat the point.
 390  */
 391 static void
 392 intel_bufferobj_flush_mapped_range(GLcontext *ctx, GLenum target,
 393                                    GLintptr offset, GLsizeiptr length,
 394                                    struct gl_buffer_object *obj)
 395 {
 396    struct intel_context *intel = intel_context(ctx);
 397    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 398    drm_intel_bo *temp_bo;
 399
 400    /* Unless we're in the range map using a temporary system buffer,
 401     * there's no work to do.
 402     */
 403    if (intel_obj->range_map_buffer == NULL)
 404       return;
 405
 406    temp_bo = drm_intel_bo_alloc(intel->bufmgr, "range map flush", length, 64);
 407
 408    drm_intel_bo_subdata(temp_bo, 0, length, intel_obj->range_map_buffer);
 409
 410    intel_emit_linear_blit(intel,
 411                           intel_obj->buffer, obj->Offset + offset,
 412                           temp_bo, 0,
 413                           length);
 414
 415    drm_intel_bo_unreference(temp_bo);
 416 }
 417
 418
 419 /**
 420  * Called via glUnmapBuffer().
 421  */
 422 static GLboolean
 423 intel_bufferobj_unmap(GLcontext * ctx,
 424                       GLenum target, struct gl_buffer_object *obj)
 425 {
 426    struct intel_context *intel = intel_context(ctx);
 427    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 428
 429    assert(intel_obj);
 430    assert(obj->Pointer);
 431    if (intel_obj->sys_buffer != NULL) {
 432       /* always keep the mapping around. */
 433    } else if (intel_obj->range_map_buffer != NULL) {
 434       /* Since we've emitted some blits to buffers that will (likely) be used
 435        * in rendering operations in other cache domains in this batch, emit a
 436        * flush.  Once again, we wish for a domain tracker in libdrm to cover
 437        * usage inside of a batchbuffer.
 438        */
 439       intel_batchbuffer_emit_mi_flush(intel->batch);
 440       free(intel_obj->range_map_buffer);
 441       intel_obj->range_map_buffer = NULL;
 442    } else if (intel_obj->range_map_bo != NULL) {
 443       if (intel_obj->mapped_gtt) {
 444          drm_intel_gem_bo_unmap_gtt(intel_obj->range_map_bo);
 445       } else {
 446          drm_intel_bo_unmap(intel_obj->range_map_bo);
 447       }
 448
 449       intel_emit_linear_blit(intel,
 450                              intel_obj->buffer, obj->Offset,
 451                              intel_obj->range_map_bo, 0,
 452                              obj->Length);
 453
 454       /* Since we've emitted some blits to buffers that will (likely) be used
 455        * in rendering operations in other cache domains in this batch, emit a
 456        * flush.  Once again, we wish for a domain tracker in libdrm to cover
 457        * usage inside of a batchbuffer.
 458        */
 459       intel_batchbuffer_emit_mi_flush(intel->batch);
 460
 461       drm_intel_bo_unreference(intel_obj->range_map_bo);
 462       intel_obj->range_map_bo = NULL;
 463    } else if (intel_obj->buffer != NULL) {
 464       if (intel_obj->mapped_gtt) {
 465          drm_intel_gem_bo_unmap_gtt(intel_obj->buffer);
 466       } else {
 467          drm_intel_bo_unmap(intel_obj->buffer);
 468       }
 469    }
 470    obj->Pointer = NULL;
 471    obj->Offset = 0;
 472    obj->Length = 0;
 473
 474    return GL_TRUE;
 475 }
 476
 477 dri_bo *
 478 intel_bufferobj_buffer(struct intel_context *intel,
 479                        struct intel_buffer_object *intel_obj, GLuint flag)
 480 {
 481    if (intel_obj->region) {
 482       if (flag == INTEL_WRITE_PART)
 483          intel_bufferobj_cow(intel, intel_obj);
 484       else if (flag == INTEL_WRITE_FULL) {
 485          intel_bufferobj_release_region(intel, intel_obj);
 486          intel_bufferobj_alloc_buffer(intel, intel_obj);
 487       }
 488    }
 489
 490    if (intel_obj->buffer == NULL) {
 491       void *sys_buffer = intel_obj->sys_buffer;
 492
 493       /* only one of buffer and sys_buffer could be non-NULL */
 494       intel_bufferobj_alloc_buffer(intel, intel_obj);
 495       intel_obj->sys_buffer = NULL;
 496
 497       intel_bufferobj_subdata(&intel->ctx,
 498                               GL_ARRAY_BUFFER_ARB,
 499                               0,
 500                               intel_obj->Base.Size,
 501                               sys_buffer,
 502                               &intel_obj->Base);
 503       _mesa_free(sys_buffer);
 504       intel_obj->sys_buffer = NULL;
 505    }
 506
 507    return intel_obj->buffer;
 508 }
 509
 510 static void
 511 intel_bufferobj_copy_subdata(GLcontext *ctx,
 512                              struct gl_buffer_object *src,
 513                              struct gl_buffer_object *dst,
 514                              GLintptr read_offset, GLintptr write_offset,
 515                              GLsizeiptr size)
 516 {
 517    struct intel_context *intel = intel_context(ctx);
 518    struct intel_buffer_object *intel_src = intel_buffer_object(src);
 519    struct intel_buffer_object *intel_dst = intel_buffer_object(dst);
 520    drm_intel_bo *src_bo, *dst_bo;
 521
 522    if (size == 0)
 523       return;
 524
 525    /* If we're in system memory, just map and memcpy. */
 526    if (intel_src->sys_buffer || intel_dst->sys_buffer) {
 527       /* The same buffer may be used, but note that regions copied may
 528        * not overlap.
 529        */
 530       if (src == dst) {
 531          char *ptr = intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
 532                                          GL_READ_WRITE, dst);
 533          memcpy(ptr + write_offset, ptr + read_offset, size);
 534          intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
 535       } else {
 536          const char *src_ptr;
 537          char *dst_ptr;
 538
 539          src_ptr =  intel_bufferobj_map(ctx, GL_COPY_READ_BUFFER,
 540                                         GL_READ_ONLY, src);
 541          dst_ptr =  intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
 542                                         GL_WRITE_ONLY, dst);
 543
 544          memcpy(dst_ptr + write_offset, src_ptr + read_offset, size);
 545
 546          intel_bufferobj_unmap(ctx, GL_COPY_READ_BUFFER, src);
 547          intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
 548       }
 549    }
 550
 551    /* Otherwise, we have real BOs, so blit them. */
 552
 553    dst_bo = intel_bufferobj_buffer(intel, intel_dst, INTEL_WRITE_PART);
 554    src_bo = intel_bufferobj_buffer(intel, intel_src, INTEL_READ);
 555
 556    intel_emit_linear_blit(intel,
 557                           dst_bo, write_offset,
 558                           src_bo, read_offset, size);
 559
 560    /* Since we've emitted some blits to buffers that will (likely) be used
 561     * in rendering operations in other cache domains in this batch, emit a
 562     * flush.  Once again, we wish for a domain tracker in libdrm to cover
 563     * usage inside of a batchbuffer.
 564     */
 565    intel_batchbuffer_emit_mi_flush(intel->batch);
 566 }
 567
 568 void
 569 intelInitBufferObjectFuncs(struct dd_function_table *functions)
 570 {
 571    functions->NewBufferObject = intel_bufferobj_alloc;
 572    functions->DeleteBuffer = intel_bufferobj_free;
 573    functions->BufferData = intel_bufferobj_data;
 574    functions->BufferSubData = intel_bufferobj_subdata;
 575    functions->GetBufferSubData = intel_bufferobj_get_subdata;
 576    functions->MapBuffer = intel_bufferobj_map;
 577    functions->MapBufferRange = intel_bufferobj_map_range;
 578    functions->FlushMappedBufferRange = intel_bufferobj_flush_mapped_range;
 579    functions->UnmapBuffer = intel_bufferobj_unmap;
 580    functions->CopyBufferSubData = intel_bufferobj_copy_subdata;
 581 }