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

   1 /**************************************************************************
   2  *
   3  * Copyright 2006 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 /* Provide additional functionality on top of bufmgr buffers:
  29  *   - 2d semantics and blit operations
  30  *   - refcounting of buffers for multiple images in a buffer.
  31  *   - refcounting of buffer mappings.
  32  *   - some logic for moving the buffers to the best memory pools for
  33  *     given operations.
  34  *
  35  * Most of this is to make it easier to implement the fixed-layout
  36  * mipmap tree required by intel hardware in the face of GL's
  37  * programming interface where each image can be specifed in random
  38  * order and it isn't clear what layout the tree should have until the
  39  * last moment.
  40  */
  41
  42 #include <sys/ioctl.h>
  43 #include <errno.h>
  44
  45 #include "intel_context.h"
  46 #include "intel_regions.h"
  47 #include "intel_blit.h"
  48 #include "intel_buffer_objects.h"
  49 #include "intel_bufmgr.h"
  50 #include "intel_batchbuffer.h"
  51 #include "intel_chipset.h"
  52
  53 #define FILE_DEBUG_FLAG DEBUG_REGION
  54
  55 /* This should be set to the maximum backtrace size desired.
  56  * Set it to 0 to disable backtrace debugging.
  57  */
  58 #define DEBUG_BACKTRACE_SIZE 0
  59
  60 #if DEBUG_BACKTRACE_SIZE == 0
  61 /* Use the standard debug output */
  62 #define _DBG(...) DBG(__VA_ARGS__)
  63 #else
  64 /* Use backtracing debug output */
  65 #define _DBG(...) {debug_backtrace(); DBG(__VA_ARGS__);}
  66
  67 /* Backtracing debug support */
  68 #include <execinfo.h>
  69
  70 static void
  71 debug_backtrace(void)
  72 {
  73    void *trace[DEBUG_BACKTRACE_SIZE];
  74    char **strings = NULL;
  75    int traceSize;
  76    register int i;
  77
  78    traceSize = backtrace(trace, DEBUG_BACKTRACE_SIZE);
  79    strings = backtrace_symbols(trace, traceSize);
  80    if (strings == NULL) {
  81       DBG("no backtrace:");
  82       return;
  83    }
  84
  85    /* Spit out all the strings with a colon separator.  Ignore
  86     * the first, since we don't really care about the call
  87     * to debug_backtrace() itself.  Skip until the final "/" in
  88     * the trace to avoid really long lines.
  89     */
  90    for (i = 1; i < traceSize; i++) {
  91       char *p = strings[i], *slash = strings[i];
  92       while (*p) {
  93          if (*p++ == '/') {
  94             slash = p;
  95          }
  96       }
  97
  98       DBG("%s:", slash);
  99    }
 100
 101    /* Free up the memory, and we're done */
 102    free(strings);
 103 }
 104
 105 #endif
 106
 107
 108
 109 /* XXX: Thread safety?
 110  */
 111 GLubyte *
 112 intel_region_map(struct intel_context *intel, struct intel_region *region)
 113 {
 114    intelFlush(&intel->ctx);
 115
 116    _DBG("%s %p\n", __FUNCTION__, region);
 117    if (!region->map_refcount++) {
 118       if (region->pbo)
 119          intel_region_cow(intel, region);
 120
 121       if (region->tiling != I915_TILING_NONE &&
 122           intel->intelScreen->kernel_exec_fencing)
 123          drm_intel_gem_bo_map_gtt(region->buffer);
 124       else
 125          dri_bo_map(region->buffer, GL_TRUE);
 126       region->map = region->buffer->virtual;
 127    }
 128
 129    return region->map;
 130 }
 131
 132 void
 133 intel_region_unmap(struct intel_context *intel, struct intel_region *region)
 134 {
 135    _DBG("%s %p\n", __FUNCTION__, region);
 136    if (!--region->map_refcount) {
 137       if (region->tiling != I915_TILING_NONE &&
 138           intel->intelScreen->kernel_exec_fencing)
 139          drm_intel_gem_bo_unmap_gtt(region->buffer);
 140       else
 141          dri_bo_unmap(region->buffer);
 142       region->map = NULL;
 143    }
 144 }
 145
 146 static struct intel_region *
 147 intel_region_alloc_internal(struct intel_context *intel,
 148                             GLuint cpp,
 149                             GLuint width, GLuint height, GLuint pitch,
 150                             dri_bo *buffer)
 151 {
 152    struct intel_region *region;
 153
 154    if (buffer == NULL) {
 155       _DBG("%s <-- NULL\n", __FUNCTION__);
 156       return NULL;
 157    }
 158
 159    region = calloc(sizeof(*region), 1);
 160    region->cpp = cpp;
 161    region->width = width;
 162    region->height = height;
 163    region->pitch = pitch;
 164    region->refcount = 1;
 165    region->buffer = buffer;
 166
 167    /* Default to no tiling */
 168    region->tiling = I915_TILING_NONE;
 169    region->bit_6_swizzle = I915_BIT_6_SWIZZLE_NONE;
 170
 171    _DBG("%s <-- %p\n", __FUNCTION__, region);
 172    return region;
 173 }
 174
 175 struct intel_region *
 176 intel_region_alloc(struct intel_context *intel,
 177                    uint32_t tiling,
 178                    GLuint cpp, GLuint width, GLuint height, GLuint pitch,
 179                    GLboolean expect_accelerated_upload)
 180 {
 181    dri_bo *buffer;
 182    struct intel_region *region;
 183
 184    if (tiling == I915_TILING_X)
 185       height = ALIGN(height, 8);
 186    else if (tiling == I915_TILING_Y)
 187       height = ALIGN(height, 32);
 188
 189    if (expect_accelerated_upload) {
 190       buffer = drm_intel_bo_alloc_for_render(intel->bufmgr, "region",
 191                                              pitch * cpp * height, 64);
 192    } else {
 193       buffer = drm_intel_bo_alloc(intel->bufmgr, "region",
 194                                   pitch * cpp * height, 64);
 195    }
 196
 197    region = intel_region_alloc_internal(intel, cpp, width, height,
 198                                         pitch, buffer);
 199
 200    if (tiling != I915_TILING_NONE) {
 201       assert(((pitch * cpp) & 127) == 0);
 202       drm_intel_bo_set_tiling(buffer, &tiling, pitch * cpp);
 203       drm_intel_bo_get_tiling(buffer, &region->tiling, &region->bit_6_swizzle);
 204    }
 205
 206    return region;
 207 }
 208
 209 struct intel_region *
 210 intel_region_alloc_for_handle(struct intel_context *intel,
 211                               GLuint cpp,
 212                               GLuint width, GLuint height, GLuint pitch,
 213                               GLuint handle, const char *name)
 214 {
 215    struct intel_region *region;
 216    dri_bo *buffer;
 217    int ret;
 218
 219    buffer = intel_bo_gem_create_from_name(intel->bufmgr, name, handle);
 220
 221    region = intel_region_alloc_internal(intel, cpp,
 222                                         width, height, pitch, buffer);
 223    if (region == NULL)
 224       return region;
 225
 226    ret = dri_bo_get_tiling(region->buffer, &region->tiling,
 227                            &region->bit_6_swizzle);
 228    if (ret != 0) {
 229       fprintf(stderr, "Couldn't get tiling of buffer %d (%s): %s\n",
 230               handle, name, strerror(-ret));
 231       intel_region_release(&region);
 232       return NULL;
 233    }
 234
 235    return region;
 236 }
 237
 238 void
 239 intel_region_reference(struct intel_region **dst, struct intel_region *src)
 240 {
 241    if (src)
 242       _DBG("%s %p %d\n", __FUNCTION__, src, src->refcount);
 243
 244    assert(*dst == NULL);
 245    if (src) {
 246       src->refcount++;
 247       *dst = src;
 248    }
 249 }
 250
 251 void
 252 intel_region_release(struct intel_region **region_handle)
 253 {
 254    struct intel_region *region = *region_handle;
 255
 256    if (region == NULL) {
 257       _DBG("%s NULL\n", __FUNCTION__);
 258       return;
 259    }
 260
 261    _DBG("%s %p %d\n", __FUNCTION__, region, region->refcount - 1);
 262
 263    ASSERT(region->refcount > 0);
 264    region->refcount--;
 265
 266    if (region->refcount == 0) {
 267       assert(region->map_refcount == 0);
 268
 269       if (region->pbo)
 270          region->pbo->region = NULL;
 271       region->pbo = NULL;
 272       dri_bo_unreference(region->buffer);
 273
 274       if (region->classic_map != NULL) {
 275          drmUnmap(region->classic_map,
 276                         region->pitch * region->cpp * region->height);
 277       }
 278
 279       free(region);
 280    }
 281    *region_handle = NULL;
 282 }
 283
 284 /*
 285  * XXX Move this into core Mesa?
 286  */
 287 void
 288 _mesa_copy_rect(GLubyte * dst,
 289                 GLuint cpp,
 290                 GLuint dst_pitch,
 291                 GLuint dst_x,
 292                 GLuint dst_y,
 293                 GLuint width,
 294                 GLuint height,
 295                 const GLubyte * src,
 296                 GLuint src_pitch, GLuint src_x, GLuint src_y)
 297 {
 298    GLuint i;
 299
 300    dst_pitch *= cpp;
 301    src_pitch *= cpp;
 302    dst += dst_x * cpp;
 303    src += src_x * cpp;
 304    dst += dst_y * dst_pitch;
 305    src += src_y * dst_pitch;
 306    width *= cpp;
 307
 308    if (width == dst_pitch && width == src_pitch)
 309       memcpy(dst, src, height * width);
 310    else {
 311       for (i = 0; i < height; i++) {
 312          memcpy(dst, src, width);
 313          dst += dst_pitch;
 314          src += src_pitch;
 315       }
 316    }
 317 }
 318
 319
 320 /* Upload data to a rectangular sub-region.  Lots of choices how to do this:
 321  *
 322  * - memcpy by span to current destination
 323  * - upload data as new buffer and blit
 324  *
 325  * Currently always memcpy.
 326  */
 327 void
 328 intel_region_data(struct intel_context *intel,
 329                   struct intel_region *dst,
 330                   GLuint dst_offset,
 331                   GLuint dstx, GLuint dsty,
 332                   const void *src, GLuint src_pitch,
 333                   GLuint srcx, GLuint srcy, GLuint width, GLuint height)
 334 {
 335    _DBG("%s\n", __FUNCTION__);
 336
 337    if (intel == NULL)
 338       return;
 339
 340    if (dst->pbo) {
 341       if (dstx == 0 &&
 342           dsty == 0 && width == dst->pitch && height == dst->height)
 343          intel_region_release_pbo(intel, dst);
 344       else
 345          intel_region_cow(intel, dst);
 346    }
 347
 348    LOCK_HARDWARE(intel);
 349    _mesa_copy_rect(intel_region_map(intel, dst) + dst_offset,
 350                    dst->cpp,
 351                    dst->pitch,
 352                    dstx, dsty, width, height, src, src_pitch, srcx, srcy);
 353
 354    intel_region_unmap(intel, dst);
 355    UNLOCK_HARDWARE(intel);
 356 }
 357
 358 /* Copy rectangular sub-regions. Need better logic about when to
 359  * push buffers into AGP - will currently do so whenever possible.
 360  */
 361 GLboolean
 362 intel_region_copy(struct intel_context *intel,
 363                   struct intel_region *dst,
 364                   GLuint dst_offset,
 365                   GLuint dstx, GLuint dsty,
 366                   struct intel_region *src,
 367                   GLuint src_offset,
 368                   GLuint srcx, GLuint srcy, GLuint width, GLuint height,
 369                   GLenum logicop)
 370 {
 371    _DBG("%s\n", __FUNCTION__);
 372
 373    if (intel == NULL)
 374       return GL_FALSE;
 375
 376    if (dst->pbo) {
 377       if (dstx == 0 &&
 378           dsty == 0 && width == dst->pitch && height == dst->height)
 379          intel_region_release_pbo(intel, dst);
 380       else
 381          intel_region_cow(intel, dst);
 382    }
 383
 384    assert(src->cpp == dst->cpp);
 385
 386    return intelEmitCopyBlit(intel,
 387                             dst->cpp,
 388                             src->pitch, src->buffer, src_offset, src->tiling,
 389                             dst->pitch, dst->buffer, dst_offset, dst->tiling,
 390                             srcx, srcy, dstx, dsty, width, height,
 391                             logicop);
 392 }
 393
 394 /* Attach to a pbo, discarding our data.  Effectively zero-copy upload
 395  * the pbo's data.
 396  */
 397 void
 398 intel_region_attach_pbo(struct intel_context *intel,
 399                         struct intel_region *region,
 400                         struct intel_buffer_object *pbo)
 401 {
 402    dri_bo *buffer;
 403
 404    if (region->pbo == pbo)
 405       return;
 406
 407    _DBG("%s %p %p\n", __FUNCTION__, region, pbo);
 408
 409    /* If there is already a pbo attached, break the cow tie now.
 410     * Don't call intel_region_release_pbo() as that would
 411     * unnecessarily allocate a new buffer we would have to immediately
 412     * discard.
 413     */
 414    if (region->pbo) {
 415       region->pbo->region = NULL;
 416       region->pbo = NULL;
 417    }
 418
 419    if (region->buffer) {
 420       dri_bo_unreference(region->buffer);
 421       region->buffer = NULL;
 422    }
 423
 424    /* make sure pbo has a buffer of its own */
 425    buffer = intel_bufferobj_buffer(intel, pbo, INTEL_WRITE_FULL);
 426
 427    region->pbo = pbo;
 428    region->pbo->region = region;
 429    dri_bo_reference(buffer);
 430    region->buffer = buffer;
 431 }
 432
 433
 434 /* Break the COW tie to the pbo and allocate a new buffer.
 435  * The pbo gets to keep the data.
 436  */
 437 void
 438 intel_region_release_pbo(struct intel_context *intel,
 439                          struct intel_region *region)
 440 {
 441    _DBG("%s %p\n", __FUNCTION__, region);
 442    assert(region->buffer == region->pbo->buffer);
 443    region->pbo->region = NULL;
 444    region->pbo = NULL;
 445    dri_bo_unreference(region->buffer);
 446    region->buffer = NULL;
 447
 448    region->buffer = dri_bo_alloc(intel->bufmgr, "region",
 449                                  region->pitch * region->cpp * region->height,
 450                                  64);
 451 }
 452
 453 /* Break the COW tie to the pbo.  Both the pbo and the region end up
 454  * with a copy of the data.
 455  */
 456 void
 457 intel_region_cow(struct intel_context *intel, struct intel_region *region)
 458 {
 459    struct intel_buffer_object *pbo = region->pbo;
 460    GLboolean ok;
 461
 462    intel_region_release_pbo(intel, region);
 463
 464    assert(region->cpp * region->pitch * region->height == pbo->Base.Size);
 465
 466    _DBG("%s %p (%d bytes)\n", __FUNCTION__, region, pbo->Base.Size);
 467
 468    /* Now blit from the texture buffer to the new buffer:
 469     */
 470
 471    LOCK_HARDWARE(intel);
 472    ok = intelEmitCopyBlit(intel,
 473                           region->cpp,
 474                           region->pitch, pbo->buffer, 0, region->tiling,
 475                           region->pitch, region->buffer, 0, region->tiling,
 476                           0, 0, 0, 0,
 477                           region->pitch, region->height,
 478                           GL_COPY);
 479    assert(ok);
 480    UNLOCK_HARDWARE(intel);
 481 }
 482
 483 dri_bo *
 484 intel_region_buffer(struct intel_context *intel,
 485                     struct intel_region *region, GLuint flag)
 486 {
 487    if (region->pbo) {
 488       if (flag == INTEL_WRITE_PART)
 489          intel_region_cow(intel, region);
 490       else if (flag == INTEL_WRITE_FULL)
 491          intel_region_release_pbo(intel, region);
 492    }
 493
 494    return region->buffer;
 495 }
 496
 497 static struct intel_region *
 498 intel_recreate_static(struct intel_context *intel,
 499                       const char *name,
 500                       struct intel_region *region,
 501                       intelRegion *region_desc)
 502 {
 503    intelScreenPrivate *intelScreen = intel->intelScreen;
 504    int ret;
 505
 506    if (region == NULL) {
 507       region = calloc(sizeof(*region), 1);
 508       region->refcount = 1;
 509       _DBG("%s creating new region %p\n", __FUNCTION__, region);
 510    }
 511    else {
 512       _DBG("%s %p\n", __FUNCTION__, region);
 513    }
 514
 515    if (intel->ctx.Visual.rgbBits == 24)
 516       region->cpp = 4;
 517    else
 518       region->cpp = intel->ctx.Visual.rgbBits / 8;
 519    region->pitch = intelScreen->pitch;
 520    region->width = intelScreen->width;
 521    region->height = intelScreen->height;
 522
 523    if (region->buffer != NULL) {
 524       dri_bo_unreference(region->buffer);
 525       region->buffer = NULL;
 526    }
 527
 528    if (intel->ttm) {
 529       assert(region_desc->bo_handle != -1);
 530       region->buffer = intel_bo_gem_create_from_name(intel->bufmgr,
 531                                                      name,
 532                                                      region_desc->bo_handle);
 533
 534       ret = dri_bo_get_tiling(region->buffer, &region->tiling,
 535                               &region->bit_6_swizzle);
 536       if (ret != 0) {
 537          fprintf(stderr, "Couldn't get tiling of buffer %d (%s): %s\n",
 538                  region_desc->bo_handle, name, strerror(-ret));
 539          intel_region_release(&region);
 540          return NULL;
 541       }
 542    } else {
 543       if (region->classic_map != NULL) {
 544          drmUnmap(region->classic_map,
 545                   region->pitch * region->cpp * region->height);
 546          region->classic_map = NULL;
 547       }
 548       ret = drmMap(intel->driFd, region_desc->handle,
 549                    region->pitch * region->cpp * region->height,
 550                    &region->classic_map);
 551       if (ret != 0) {
 552          fprintf(stderr, "Failed to drmMap %s buffer\n", name);
 553          free(region);
 554          return NULL;
 555       }
 556
 557       region->buffer = intel_bo_fake_alloc_static(intel->bufmgr,
 558                                                   name,
 559                                                   region_desc->offset,
 560                                                   region->pitch * region->cpp *
 561                                                   region->height,
 562                                                   region->classic_map);
 563
 564       /* The sarea just gives us a boolean for whether it's tiled or not,
 565        * instead of which tiling mode it is.  Guess.
 566        */
 567       if (region_desc->tiled) {
 568          if (IS_965(intel->intelScreen->deviceID) &&
 569              region_desc == &intelScreen->depth)
 570             region->tiling = I915_TILING_Y;
 571          else
 572             region->tiling = I915_TILING_X;
 573       } else {
 574          region->tiling = I915_TILING_NONE;
 575       }
 576
 577       region->bit_6_swizzle = I915_BIT_6_SWIZZLE_NONE;
 578    }
 579
 580    assert(region->buffer != NULL);
 581
 582    return region;
 583 }
 584
 585 /**
 586  * Create intel_region structs to describe the static front, back, and depth
 587  * buffers created by the xserver.
 588  *
 589  * Although FBO's mean we now no longer use these as render targets in
 590  * all circumstances, they won't go away until the back and depth
 591  * buffers become private, and the front buffer will remain even then.
 592  *
 593  * Note that these don't allocate video memory, just describe
 594  * allocations alread made by the X server.
 595  */
 596 void
 597 intel_recreate_static_regions(struct intel_context *intel)
 598 {
 599    intelScreenPrivate *intelScreen = intel->intelScreen;
 600
 601    intel->front_region =
 602       intel_recreate_static(intel, "front",
 603                             intel->front_region,
 604                             &intelScreen->front);
 605
 606    intel->back_region =
 607       intel_recreate_static(intel, "back",
 608                             intel->back_region,
 609                             &intelScreen->back);
 610
 611    /* Still assumes front.cpp == depth.cpp.  We can kill this when we move to
 612     * private buffers.
 613     */
 614    intel->depth_region =
 615       intel_recreate_static(intel, "depth",
 616                             intel->depth_region,
 617                             &intelScreen->depth);
 618 }