src/gallium/drivers/svga/svga_screen_buffer.c

   1 /**********************************************************
   2  * Copyright 2008-2009 VMware, Inc.  All rights reserved.
   3  *
   4  * Permission is hereby granted, free of charge, to any person
   5  * obtaining a copy of this software and associated documentation
   6  * files (the "Software"), to deal in the Software without
   7  * restriction, including without limitation the rights to use, copy,
   8  * modify, merge, publish, distribute, sublicense, and/or sell copies
   9  * of the Software, and to permit persons to whom the Software is
  10  * furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be
  13  * included in all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  16  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  17  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  18  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  19  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  20  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  21  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  **********************************************************/
  25
  26 #include "svga_cmd.h"
  27
  28 #include "pipe/p_state.h"
  29 #include "pipe/p_defines.h"
  30 #include "util/u_inlines.h"
  31 #include "os/os_thread.h"
  32 #include "util/u_math.h"
  33 #include "util/u_memory.h"
  34
  35 #include "svga_context.h"
  36 #include "svga_screen.h"
  37 #include "svga_screen_buffer.h"
  38 #include "svga_winsys.h"
  39 #include "svga_debug.h"
  40
  41
  42 /**
  43  * Vertex and index buffers have to be treated slightly differently from
  44  * regular guest memory regions because the SVGA device sees them as
  45  * surfaces, and the state tracker can create/destroy without the pipe
  46  * driver, therefore we must do the uploads from the vws.
  47  */
  48 static INLINE boolean
  49 svga_buffer_needs_hw_storage(unsigned usage)
  50 {
  51    return usage & (PIPE_BUFFER_USAGE_VERTEX | PIPE_BUFFER_USAGE_INDEX);
  52 }
  53
  54
  55 static INLINE enum pipe_error
  56 svga_buffer_create_host_surface(struct svga_screen *ss,
  57                                 struct svga_buffer *sbuf)
  58 {
  59    if(!sbuf->handle) {
  60       sbuf->key.flags = 0;
  61
  62       sbuf->key.format = SVGA3D_BUFFER;
  63       if(sbuf->base.usage & PIPE_BUFFER_USAGE_VERTEX)
  64          sbuf->key.flags |= SVGA3D_SURFACE_HINT_VERTEXBUFFER;
  65       if(sbuf->base.usage & PIPE_BUFFER_USAGE_INDEX)
  66          sbuf->key.flags |= SVGA3D_SURFACE_HINT_INDEXBUFFER;
  67
  68       sbuf->key.size.width = sbuf->base.size;
  69       sbuf->key.size.height = 1;
  70       sbuf->key.size.depth = 1;
  71
  72       sbuf->key.numFaces = 1;
  73       sbuf->key.numMipLevels = 1;
  74       sbuf->key.cachable = 1;
  75
  76       SVGA_DBG(DEBUG_DMA, "surface_create for buffer sz %d\n", sbuf->base.size);
  77
  78       sbuf->handle = svga_screen_surface_create(ss, &sbuf->key);
  79       if(!sbuf->handle)
  80          return PIPE_ERROR_OUT_OF_MEMORY;
  81
  82       /* Always set the discard flag on the first time the buffer is written
  83        * as svga_screen_surface_create might have passed a recycled host
  84        * buffer.
  85        */
  86       sbuf->dma.flags.discard = TRUE;
  87
  88       SVGA_DBG(DEBUG_DMA, "   --> got sid %p sz %d (buffer)\n", sbuf->handle, sbuf->base.size);
  89    }
  90
  91    return PIPE_OK;
  92 }
  93
  94
  95 static INLINE void
  96 svga_buffer_destroy_host_surface(struct svga_screen *ss,
  97                                  struct svga_buffer *sbuf)
  98 {
  99    if(sbuf->handle) {
 100       SVGA_DBG(DEBUG_DMA, " ungrab sid %p sz %d\n", sbuf->handle, sbuf->base.size);
 101       svga_screen_surface_destroy(ss, &sbuf->key, &sbuf->handle);
 102    }
 103 }
 104
 105
 106 static INLINE void
 107 svga_buffer_destroy_hw_storage(struct svga_screen *ss, struct svga_buffer *sbuf)
 108 {
 109    struct svga_winsys_screen *sws = ss->sws;
 110
 111    assert(!sbuf->map.count);
 112    assert(sbuf->hwbuf);
 113    if(sbuf->hwbuf) {
 114       sws->buffer_destroy(sws, sbuf->hwbuf);
 115       sbuf->hwbuf = NULL;
 116    }
 117 }
 118
 119 struct svga_winsys_buffer *
 120 svga_winsys_buffer_create( struct svga_screen *ss,
 121                            unsigned alignment,
 122                            unsigned usage,
 123                            unsigned size )
 124 {
 125    struct svga_winsys_screen *sws = ss->sws;
 126    struct svga_winsys_buffer *buf;
 127
 128    /* Just try */
 129    buf = sws->buffer_create(sws, alignment, usage, size);
 130    if(!buf) {
 131
 132       SVGA_DBG(DEBUG_DMA|DEBUG_PERF, "flushing screen to find %d bytes GMR\n",
 133                size);
 134
 135       /* Try flushing all pending DMAs */
 136       svga_screen_flush(ss, NULL);
 137       buf = sws->buffer_create(sws, alignment, usage, size);
 138
 139    }
 140
 141    return buf;
 142 }
 143
 144
 145 /**
 146  * Allocate DMA'ble storage for the buffer.
 147  *
 148  * Called before mapping a buffer.
 149  */
 150 static INLINE enum pipe_error
 151 svga_buffer_create_hw_storage(struct svga_screen *ss,
 152                               struct svga_buffer *sbuf)
 153 {
 154    assert(!sbuf->user);
 155
 156    if(!sbuf->hwbuf) {
 157       unsigned alignment = sbuf->base.alignment;
 158       unsigned usage = 0;
 159       unsigned size = sbuf->base.size;
 160
 161       sbuf->hwbuf = svga_winsys_buffer_create(ss, alignment, usage, size);
 162       if(!sbuf->hwbuf)
 163          return PIPE_ERROR_OUT_OF_MEMORY;
 164
 165       assert(!sbuf->dma.pending);
 166    }
 167
 168    return PIPE_OK;
 169 }
 170
 171
 172 /**
 173  * Variant of SVGA3D_BufferDMA which leaves the copy box temporarily in blank.
 174  */
 175 static enum pipe_error
 176 svga_buffer_upload_command(struct svga_context *svga,
 177                            struct svga_buffer *sbuf)
 178 {
 179    struct svga_winsys_context *swc = svga->swc;
 180    struct svga_winsys_buffer *guest = sbuf->hwbuf;
 181    struct svga_winsys_surface *host = sbuf->handle;
 182    SVGA3dTransferType transfer = SVGA3D_WRITE_HOST_VRAM;
 183    SVGA3dCmdSurfaceDMA *cmd;
 184    uint32 numBoxes = sbuf->map.num_ranges;
 185    SVGA3dCopyBox *boxes;
 186    SVGA3dCmdSurfaceDMASuffix *pSuffix;
 187    unsigned region_flags;
 188    unsigned surface_flags;
 189    struct pipe_buffer *dummy;
 190
 191    if(transfer == SVGA3D_WRITE_HOST_VRAM) {
 192       region_flags = PIPE_BUFFER_USAGE_GPU_READ;
 193       surface_flags = PIPE_BUFFER_USAGE_GPU_WRITE;
 194    }
 195    else if(transfer == SVGA3D_READ_HOST_VRAM) {
 196       region_flags = PIPE_BUFFER_USAGE_GPU_WRITE;
 197       surface_flags = PIPE_BUFFER_USAGE_GPU_READ;
 198    }
 199    else {
 200       assert(0);
 201       return PIPE_ERROR_BAD_INPUT;
 202    }
 203
 204    assert(numBoxes);
 205
 206    cmd = SVGA3D_FIFOReserve(swc,
 207                             SVGA_3D_CMD_SURFACE_DMA,
 208                             sizeof *cmd + numBoxes * sizeof *boxes + sizeof *pSuffix,
 209                             2);
 210    if(!cmd)
 211       return PIPE_ERROR_OUT_OF_MEMORY;
 212
 213    swc->region_relocation(swc, &cmd->guest.ptr, guest, 0, region_flags);
 214    cmd->guest.pitch = 0;
 215
 216    swc->surface_relocation(swc, &cmd->host.sid, host, surface_flags);
 217    cmd->host.face = 0;
 218    cmd->host.mipmap = 0;
 219
 220    cmd->transfer = transfer;
 221
 222    sbuf->dma.boxes = (SVGA3dCopyBox *)&cmd[1];
 223    sbuf->dma.svga = svga;
 224
 225    /* Increment reference count */
 226    dummy = NULL;
 227    pipe_buffer_reference(&dummy, &sbuf->base);
 228
 229    pSuffix = (SVGA3dCmdSurfaceDMASuffix *)((uint8_t*)cmd + sizeof *cmd + numBoxes * sizeof *boxes);
 230    pSuffix->suffixSize = sizeof *pSuffix;
 231    pSuffix->maximumOffset = sbuf->base.size;
 232    pSuffix->flags = sbuf->dma.flags;
 233
 234    SVGA_FIFOCommitAll(swc);
 235
 236    sbuf->dma.flags.discard = FALSE;
 237
 238    return PIPE_OK;
 239 }
 240
 241
 242 /**
 243  * Patch up the upload DMA command reserved by svga_buffer_upload_command
 244  * with the final ranges.
 245  */
 246 static void
 247 svga_buffer_upload_flush(struct svga_context *svga,
 248                          struct svga_buffer *sbuf)
 249 {
 250    SVGA3dCopyBox *boxes;
 251    unsigned i;
 252
 253    assert(sbuf->handle);
 254    assert(sbuf->hwbuf);
 255    assert(sbuf->map.num_ranges);
 256    assert(sbuf->dma.svga == svga);
 257    assert(sbuf->dma.boxes);
 258
 259    /*
 260     * Patch the DMA command with the final copy box.
 261     */
 262
 263    SVGA_DBG(DEBUG_DMA, "dma to sid %p\n", sbuf->handle);
 264
 265    boxes = sbuf->dma.boxes;
 266    for(i = 0; i < sbuf->map.num_ranges; ++i) {
 267       SVGA_DBG(DEBUG_DMA, "  bytes %u - %u\n",
 268                sbuf->map.ranges[i].start, sbuf->map.ranges[i].end);
 269
 270       boxes[i].x = sbuf->map.ranges[i].start;
 271       boxes[i].y = 0;
 272       boxes[i].z = 0;
 273       boxes[i].w = sbuf->map.ranges[i].end - sbuf->map.ranges[i].start;
 274       boxes[i].h = 1;
 275       boxes[i].d = 1;
 276       boxes[i].srcx = sbuf->map.ranges[i].start;
 277       boxes[i].srcy = 0;
 278       boxes[i].srcz = 0;
 279    }
 280
 281    sbuf->map.num_ranges = 0;
 282
 283    assert(sbuf->head.prev && sbuf->head.next);
 284    LIST_DEL(&sbuf->head);
 285 #ifdef DEBUG
 286    sbuf->head.next = sbuf->head.prev = NULL;
 287 #endif
 288    sbuf->dma.pending = FALSE;
 289
 290    sbuf->dma.svga = NULL;
 291    sbuf->dma.boxes = NULL;
 292
 293    /* Decrement reference count */
 294    pipe_reference(&(sbuf->base.reference), NULL);
 295    sbuf = NULL;
 296 }
 297
 298
 299 /**
 300  * Note a dirty range.
 301  *
 302  * This function only notes the range down. It doesn't actually emit a DMA
 303  * upload command. That only happens when a context tries to refer to this
 304  * buffer, and the DMA upload command is added to that context's command buffer.
 305  *
 306  * We try to lump as many contiguous DMA transfers together as possible.
 307  */
 308 static void
 309 svga_buffer_add_range(struct svga_buffer *sbuf,
 310                       unsigned start,
 311                       unsigned end)
 312 {
 313    unsigned i;
 314    unsigned nearest_range;
 315    unsigned nearest_dist;
 316
 317    assert(end > start);
 318
 319    if (sbuf->map.num_ranges < SVGA_BUFFER_MAX_RANGES) {
 320       nearest_range = sbuf->map.num_ranges;
 321       nearest_dist = ~0;
 322    } else {
 323       nearest_range = SVGA_BUFFER_MAX_RANGES - 1;
 324       nearest_dist = 0;
 325    }
 326
 327    /*
 328     * Try to grow one of the ranges.
 329     *
 330     * Note that it is not this function task to care about overlapping ranges,
 331     * as the GMR was already given so it is too late to do anything. Situations
 332     * where overlapping ranges may pose a problem should be detected via
 333     * pipe_context::is_buffer_referenced and the context that refers to the
 334     * buffer should be flushed.
 335     */
 336
 337    for(i = 0; i < sbuf->map.num_ranges; ++i) {
 338       int left_dist;
 339       int right_dist;
 340       int dist;
 341
 342       left_dist = start - sbuf->map.ranges[i].end;
 343       right_dist = sbuf->map.ranges[i].start - end;
 344       dist = MAX2(left_dist, right_dist);
 345
 346       if (dist <= 0) {
 347          /*
 348           * Ranges are contiguous or overlapping -- extend this one and return.
 349           */
 350
 351          sbuf->map.ranges[i].start = MIN2(sbuf->map.ranges[i].start, start);
 352          sbuf->map.ranges[i].end   = MAX2(sbuf->map.ranges[i].end,   end);
 353          return;
 354       }
 355       else {
 356          /*
 357           * Discontiguous ranges -- keep track of the nearest range.
 358           */
 359
 360          if (dist < nearest_dist) {
 361             nearest_range = i;
 362             nearest_dist = dist;
 363          }
 364       }
 365    }
 366
 367    /*
 368     * We cannot add a new range to an existing DMA command, so patch-up the
 369     * pending DMA upload and start clean.
 370     */
 371
 372    if(sbuf->dma.pending)
 373       svga_buffer_upload_flush(sbuf->dma.svga, sbuf);
 374
 375    assert(!sbuf->dma.pending);
 376    assert(!sbuf->dma.svga);
 377    assert(!sbuf->dma.boxes);
 378
 379    if (sbuf->map.num_ranges < SVGA_BUFFER_MAX_RANGES) {
 380       /*
 381        * Add a new range.
 382        */
 383
 384       sbuf->map.ranges[sbuf->map.num_ranges].start = start;
 385       sbuf->map.ranges[sbuf->map.num_ranges].end = end;
 386       ++sbuf->map.num_ranges;
 387    } else {
 388       /*
 389        * Everything else failed, so just extend the nearest range.
 390        *
 391        * It is OK to do this because we always keep a local copy of the
 392        * host buffer data, for SW TNL, and the host never modifies the buffer.
 393        */
 394
 395       assert(nearest_range < SVGA_BUFFER_MAX_RANGES);
 396       assert(nearest_range < sbuf->map.num_ranges);
 397       sbuf->map.ranges[nearest_range].start = MIN2(sbuf->map.ranges[nearest_range].start, start);
 398       sbuf->map.ranges[nearest_range].end   = MAX2(sbuf->map.ranges[nearest_range].end,   end);
 399    }
 400 }
 401
 402
 403 static void *
 404 svga_buffer_map_range( struct pipe_screen *screen,
 405                        struct pipe_buffer *buf,
 406                        unsigned offset, unsigned length,
 407                        unsigned usage )
 408 {
 409    struct svga_screen *ss = svga_screen(screen);
 410    struct svga_winsys_screen *sws = ss->sws;
 411    struct svga_buffer *sbuf = svga_buffer( buf );
 412    void *map;
 413
 414    if (!sbuf->swbuf && !sbuf->hwbuf) {
 415       if (svga_buffer_create_hw_storage(ss, sbuf) != PIPE_OK) {
 416          /*
 417           * We can't create a hardware buffer big enough, so create a malloc
 418           * buffer instead.
 419           */
 420
 421          debug_printf("%s: failed to allocate %u KB of DMA, splitting DMA transfers\n",
 422                       __FUNCTION__,
 423                       (sbuf->base.size + 1023)/1024);
 424
 425          sbuf->swbuf = align_malloc(sbuf->base.size, sbuf->base.alignment);
 426       }
 427    }
 428
 429    if (sbuf->swbuf) {
 430       /* User/malloc buffer */
 431       map = sbuf->swbuf;
 432    }
 433    else if (sbuf->hwbuf) {
 434       map = sws->buffer_map(sws, sbuf->hwbuf, usage);
 435    }
 436    else {
 437       map = NULL;
 438    }
 439
 440    if(map) {
 441       pipe_mutex_lock(ss->swc_mutex);
 442
 443       ++sbuf->map.count;
 444
 445       if (usage & PIPE_BUFFER_USAGE_CPU_WRITE) {
 446          assert(sbuf->map.count <= 1);
 447          sbuf->map.writing = TRUE;
 448          if (usage & PIPE_BUFFER_USAGE_FLUSH_EXPLICIT)
 449             sbuf->map.flush_explicit = TRUE;
 450       }
 451
 452       pipe_mutex_unlock(ss->swc_mutex);
 453    }
 454
 455    return map;
 456 }
 457
 458 static void
 459 svga_buffer_flush_mapped_range( struct pipe_screen *screen,
 460                                 struct pipe_buffer *buf,
 461                                 unsigned offset, unsigned length)
 462 {
 463    struct svga_buffer *sbuf = svga_buffer( buf );
 464    struct svga_screen *ss = svga_screen(screen);
 465
 466    pipe_mutex_lock(ss->swc_mutex);
 467    assert(sbuf->map.writing);
 468    if(sbuf->map.writing) {
 469       assert(sbuf->map.flush_explicit);
 470       svga_buffer_add_range(sbuf, offset, offset + length);
 471    }
 472    pipe_mutex_unlock(ss->swc_mutex);
 473 }
 474
 475 static void
 476 svga_buffer_unmap( struct pipe_screen *screen,
 477                    struct pipe_buffer *buf)
 478 {
 479    struct svga_screen *ss = svga_screen(screen);
 480    struct svga_winsys_screen *sws = ss->sws;
 481    struct svga_buffer *sbuf = svga_buffer( buf );
 482
 483    pipe_mutex_lock(ss->swc_mutex);
 484
 485    assert(sbuf->map.count);
 486    if(sbuf->map.count)
 487       --sbuf->map.count;
 488
 489    if(sbuf->hwbuf)
 490       sws->buffer_unmap(sws, sbuf->hwbuf);
 491
 492    if(sbuf->map.writing) {
 493       if(!sbuf->map.flush_explicit) {
 494          /* No mapped range was flushed -- flush the whole buffer */
 495          SVGA_DBG(DEBUG_DMA, "flushing the whole buffer\n");
 496
 497          svga_buffer_add_range(sbuf, 0, sbuf->base.size);
 498       }
 499
 500       sbuf->map.writing = FALSE;
 501       sbuf->map.flush_explicit = FALSE;
 502    }
 503
 504    pipe_mutex_unlock(ss->swc_mutex);
 505 }
 506
 507 static void
 508 svga_buffer_destroy( struct pipe_buffer *buf )
 509 {
 510    struct svga_screen *ss = svga_screen(buf->screen);
 511    struct svga_buffer *sbuf = svga_buffer( buf );
 512
 513    assert(!p_atomic_read(&buf->reference.count));
 514
 515    assert(!sbuf->dma.pending);
 516
 517    if(sbuf->handle)
 518       svga_buffer_destroy_host_surface(ss, sbuf);
 519
 520    if(sbuf->uploaded.buffer)
 521       pipe_buffer_reference(&sbuf->uploaded.buffer, NULL);
 522
 523    if(sbuf->hwbuf)
 524       svga_buffer_destroy_hw_storage(ss, sbuf);
 525
 526    if(sbuf->swbuf && !sbuf->user)
 527       align_free(sbuf->swbuf);
 528
 529    FREE(sbuf);
 530 }
 531
 532 static struct pipe_buffer *
 533 svga_buffer_create(struct pipe_screen *screen,
 534                    unsigned alignment,
 535                    unsigned usage,
 536                    unsigned size)
 537 {
 538    struct svga_screen *ss = svga_screen(screen);
 539    struct svga_buffer *sbuf;
 540
 541    assert(size);
 542    assert(alignment);
 543
 544    sbuf = CALLOC_STRUCT(svga_buffer);
 545    if(!sbuf)
 546       goto error1;
 547
 548    sbuf->magic = SVGA_BUFFER_MAGIC;
 549
 550    pipe_reference_init(&sbuf->base.reference, 1);
 551    sbuf->base.screen = screen;
 552    sbuf->base.alignment = alignment;
 553    sbuf->base.usage = usage;
 554    sbuf->base.size = size;
 555
 556    if(svga_buffer_needs_hw_storage(usage)) {
 557       if(svga_buffer_create_host_surface(ss, sbuf) != PIPE_OK)
 558          goto error2;
 559    }
 560    else {
 561       if(alignment < sizeof(void*))
 562          alignment = sizeof(void*);
 563
 564       usage |= PIPE_BUFFER_USAGE_CPU_READ_WRITE;
 565
 566       sbuf->swbuf = align_malloc(size, alignment);
 567       if(!sbuf->swbuf)
 568          goto error2;
 569    }
 570
 571    return &sbuf->base;
 572
 573 error2:
 574    FREE(sbuf);
 575 error1:
 576    return NULL;
 577 }
 578
 579 static struct pipe_buffer *
 580 svga_user_buffer_create(struct pipe_screen *screen,
 581                         void *ptr,
 582                         unsigned bytes)
 583 {
 584    struct svga_buffer *sbuf;
 585
 586    sbuf = CALLOC_STRUCT(svga_buffer);
 587    if(!sbuf)
 588       goto no_sbuf;
 589
 590    sbuf->magic = SVGA_BUFFER_MAGIC;
 591
 592    sbuf->swbuf = ptr;
 593    sbuf->user = TRUE;
 594
 595    pipe_reference_init(&sbuf->base.reference, 1);
 596    sbuf->base.screen = screen;
 597    sbuf->base.alignment = 1;
 598    sbuf->base.usage = 0;
 599    sbuf->base.size = bytes;
 600
 601    return &sbuf->base;
 602
 603 no_sbuf:
 604    return NULL;
 605 }
 606
 607
 608 void
 609 svga_screen_init_buffer_functions(struct pipe_screen *screen)
 610 {
 611    screen->buffer_create = svga_buffer_create;
 612    screen->user_buffer_create = svga_user_buffer_create;
 613    screen->buffer_map_range = svga_buffer_map_range;
 614    screen->buffer_flush_mapped_range = svga_buffer_flush_mapped_range;
 615    screen->buffer_unmap = svga_buffer_unmap;
 616    screen->buffer_destroy = svga_buffer_destroy;
 617 }
 618
 619
 620 /**
 621  * Copy the contents of the malloc buffer to a hardware buffer.
 622  */
 623 static INLINE enum pipe_error
 624 svga_buffer_update_hw(struct svga_screen *ss, struct svga_buffer *sbuf)
 625 {
 626    assert(!sbuf->user);
 627    if(!sbuf->hwbuf) {
 628       enum pipe_error ret;
 629       void *map;
 630
 631       assert(sbuf->swbuf);
 632       if(!sbuf->swbuf)
 633          return PIPE_ERROR;
 634
 635       ret = svga_buffer_create_hw_storage(ss, sbuf);
 636       if(ret != PIPE_OK)
 637          return ret;
 638
 639       pipe_mutex_lock(ss->swc_mutex);
 640       map = ss->sws->buffer_map(ss->sws, sbuf->hwbuf, PIPE_BUFFER_USAGE_CPU_WRITE);
 641       assert(map);
 642       if(!map) {
 643          pipe_mutex_unlock(ss->swc_mutex);
 644          svga_buffer_destroy_hw_storage(ss, sbuf);
 645          return PIPE_ERROR;
 646       }
 647
 648       memcpy(map, sbuf->swbuf, sbuf->base.size);
 649       ss->sws->buffer_unmap(ss->sws, sbuf->hwbuf);
 650
 651       /* This user/malloc buffer is now indistinguishable from a gpu buffer */
 652       assert(!sbuf->map.count);
 653       if(!sbuf->map.count) {
 654          if(sbuf->user)
 655             sbuf->user = FALSE;
 656          else
 657             align_free(sbuf->swbuf);
 658          sbuf->swbuf = NULL;
 659       }
 660
 661       pipe_mutex_unlock(ss->swc_mutex);
 662    }
 663
 664    return PIPE_OK;
 665 }
 666
 667
 668 /**
 669  * Upload the buffer to the host in a piecewise fashion.
 670  *
 671  * Used when the buffer is too big to fit in the GMR aperture.
 672  */
 673 static INLINE enum pipe_error
 674 svga_buffer_upload_piecewise(struct svga_screen *ss,
 675                              struct svga_context *svga,
 676                              struct svga_buffer *sbuf)
 677 {
 678    struct svga_winsys_screen *sws = ss->sws;
 679    const unsigned alignment = sizeof(void *);
 680    const unsigned usage = 0;
 681    unsigned i;
 682
 683    assert(sbuf->map.num_ranges);
 684    assert(!sbuf->dma.pending);
 685
 686    SVGA_DBG(DEBUG_DMA, "dma to sid %p\n", sbuf->handle);
 687
 688    for (i = 0; i < sbuf->map.num_ranges; ++i) {
 689       struct svga_buffer_range *range = &sbuf->map.ranges[i];
 690       unsigned offset = range->start;
 691       unsigned size = range->end - range->start;
 692
 693       while (offset < range->end) {
 694          struct svga_winsys_buffer *hwbuf;
 695          uint8_t *map;
 696          enum pipe_error ret;
 697
 698          if (offset + size > range->end)
 699             size = range->end - offset;
 700
 701          hwbuf = svga_winsys_buffer_create(ss, alignment, usage, size);
 702          while (!hwbuf) {
 703             size /= 2;
 704             if (!size)
 705                return PIPE_ERROR_OUT_OF_MEMORY;
 706             hwbuf = svga_winsys_buffer_create(ss, alignment, usage, size);
 707          }
 708
 709          SVGA_DBG(DEBUG_DMA, "  bytes %u - %u\n",
 710                   offset, offset + size);
 711
 712          map = sws->buffer_map(sws, hwbuf,
 713                                PIPE_BUFFER_USAGE_CPU_WRITE |
 714                                PIPE_BUFFER_USAGE_DISCARD);
 715          assert(map);
 716          if (map) {
 717             memcpy(map, sbuf->swbuf, size);
 718             sws->buffer_unmap(sws, hwbuf);
 719          }
 720
 721          ret = SVGA3D_BufferDMA(svga->swc,
 722                                 hwbuf, sbuf->handle,
 723                                 SVGA3D_WRITE_HOST_VRAM,
 724                                 size, 0, offset, sbuf->dma.flags);
 725          if(ret != PIPE_OK) {
 726             svga_context_flush(svga, NULL);
 727             ret =  SVGA3D_BufferDMA(svga->swc,
 728                                     hwbuf, sbuf->handle,
 729                                     SVGA3D_WRITE_HOST_VRAM,
 730                                     size, 0, offset, sbuf->dma.flags);
 731             assert(ret == PIPE_OK);
 732          }
 733
 734          sbuf->dma.flags.discard = FALSE;
 735
 736          sws->buffer_destroy(sws, hwbuf);
 737
 738          offset += size;
 739       }
 740    }
 741
 742    sbuf->map.num_ranges = 0;
 743
 744    return PIPE_OK;
 745 }
 746
 747
 748 struct svga_winsys_surface *
 749 svga_buffer_handle(struct svga_context *svga,
 750                    struct pipe_buffer *buf)
 751 {
 752    struct pipe_screen *screen = svga->pipe.screen;
 753    struct svga_screen *ss = svga_screen(screen);
 754    struct svga_buffer *sbuf;
 755    enum pipe_error ret;
 756
 757    if(!buf)
 758       return NULL;
 759
 760    sbuf = svga_buffer(buf);
 761
 762    assert(!sbuf->map.count);
 763    assert(!sbuf->user);
 764
 765    if(!sbuf->handle) {
 766       ret = svga_buffer_create_host_surface(ss, sbuf);
 767       if(ret != PIPE_OK)
 768          return NULL;
 769    }
 770
 771    assert(sbuf->handle);
 772
 773    if (sbuf->map.num_ranges) {
 774       if (!sbuf->dma.pending) {
 775          /*
 776           * No pending DMA upload yet, so insert a DMA upload command now.
 777           */
 778
 779          /*
 780           * Migrate the data from swbuf -> hwbuf if necessary.
 781           */
 782          ret = svga_buffer_update_hw(ss, sbuf);
 783          if (ret == PIPE_OK) {
 784             /*
 785              * Queue a dma command.
 786              */
 787
 788             ret = svga_buffer_upload_command(svga, sbuf);
 789             if (ret == PIPE_ERROR_OUT_OF_MEMORY) {
 790                svga_context_flush(svga, NULL);
 791                ret = svga_buffer_upload_command(svga, sbuf);
 792                assert(ret == PIPE_OK);
 793             }
 794             if (ret == PIPE_OK) {
 795                sbuf->dma.pending = TRUE;
 796                assert(!sbuf->head.prev && !sbuf->head.next);
 797                LIST_ADDTAIL(&sbuf->head, &svga->dirty_buffers);
 798             }
 799          }
 800          else if (ret == PIPE_ERROR_OUT_OF_MEMORY) {
 801             /*
 802              * The buffer is too big to fit in the GMR aperture, so break it in
 803              * smaller pieces.
 804              */
 805             ret = svga_buffer_upload_piecewise(ss, svga, sbuf);
 806          }
 807
 808          if (ret != PIPE_OK) {
 809             /*
 810              * Something unexpected happened above. There is very little that
 811              * we can do other than proceeding while ignoring the dirty ranges.
 812              */
 813             assert(0);
 814             sbuf->map.num_ranges = 0;
 815          }
 816       }
 817       else {
 818          /*
 819           * There a pending dma already. Make sure it is from this context.
 820           */
 821          assert(sbuf->dma.svga == svga);
 822       }
 823    }
 824
 825    assert(!sbuf->map.num_ranges || sbuf->dma.pending);
 826
 827    return sbuf->handle;
 828 }
 829
 830
 831 struct pipe_buffer *
 832 svga_screen_buffer_wrap_surface(struct pipe_screen *screen,
 833                                 enum SVGA3dSurfaceFormat format,
 834                                 struct svga_winsys_surface *srf)
 835 {
 836    struct pipe_buffer *buf;
 837    struct svga_buffer *sbuf;
 838    struct svga_winsys_screen *sws = svga_winsys_screen(screen);
 839
 840    buf = svga_buffer_create(screen, 0, SVGA_BUFFER_USAGE_WRAPPED, 0);
 841    if (!buf)
 842       return NULL;
 843
 844    sbuf = svga_buffer(buf);
 845
 846    /*
 847     * We are not the creator of this surface and therefore we must not
 848     * cache it for reuse. Set the cacheable flag to zero in the key to
 849     * prevent this.
 850     */
 851    sbuf->key.format = format;
 852    sbuf->key.cachable = 0;
 853    sws->surface_reference(sws, &sbuf->handle, srf);
 854
 855    return buf;
 856 }
 857
 858
 859 struct svga_winsys_surface *
 860 svga_screen_buffer_get_winsys_surface(struct pipe_buffer *buffer)
 861 {
 862    struct svga_winsys_screen *sws = svga_winsys_screen(buffer->screen);
 863    struct svga_winsys_surface *vsurf = NULL;
 864
 865    assert(svga_buffer(buffer)->key.cachable == 0);
 866    svga_buffer(buffer)->key.cachable = 0;
 867    sws->surface_reference(sws, &vsurf, svga_buffer(buffer)->handle);
 868    return vsurf;
 869 }
 870
 871 void
 872 svga_context_flush_buffers(struct svga_context *svga)
 873 {
 874    struct list_head *curr, *next;
 875    struct svga_buffer *sbuf;
 876
 877    curr = svga->dirty_buffers.next;
 878    next = curr->next;
 879    while(curr != &svga->dirty_buffers) {
 880       sbuf = LIST_ENTRY(struct svga_buffer, curr, head);
 881
 882       assert(p_atomic_read(&sbuf->base.reference.count) != 0);
 883       assert(sbuf->dma.pending);
 884
 885       svga_buffer_upload_flush(svga, sbuf);
 886
 887       curr = next;
 888       next = curr->next;
 889    }
 890 }