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    if(!sbuf->hwbuf) {
 155       unsigned alignment = sbuf->base.alignment;
 156       unsigned usage = 0;
 157       unsigned size = sbuf->base.size;
 158
 159       sbuf->hwbuf = svga_winsys_buffer_create(ss, alignment, usage, size);
 160       if(!sbuf->hwbuf)
 161          return PIPE_ERROR_OUT_OF_MEMORY;
 162
 163       assert(!sbuf->dma.pending);
 164    }
 165
 166    return PIPE_OK;
 167 }
 168
 169
 170 /**
 171  * Variant of SVGA3D_BufferDMA which leaves the copy box temporarily in blank.
 172  */
 173 static enum pipe_error
 174 svga_buffer_upload_command(struct svga_context *svga,
 175                            struct svga_buffer *sbuf)
 176 {
 177    struct svga_winsys_context *swc = svga->swc;
 178    struct svga_winsys_buffer *guest = sbuf->hwbuf;
 179    struct svga_winsys_surface *host = sbuf->handle;
 180    SVGA3dTransferType transfer = SVGA3D_WRITE_HOST_VRAM;
 181    SVGA3dCmdSurfaceDMA *cmd;
 182    uint32 numBoxes = sbuf->map.num_ranges;
 183    SVGA3dCopyBox *boxes;
 184    SVGA3dCmdSurfaceDMASuffix *pSuffix;
 185    unsigned region_flags;
 186    unsigned surface_flags;
 187    struct pipe_buffer *dummy;
 188
 189    if(transfer == SVGA3D_WRITE_HOST_VRAM) {
 190       region_flags = PIPE_BUFFER_USAGE_GPU_READ;
 191       surface_flags = PIPE_BUFFER_USAGE_GPU_WRITE;
 192    }
 193    else if(transfer == SVGA3D_READ_HOST_VRAM) {
 194       region_flags = PIPE_BUFFER_USAGE_GPU_WRITE;
 195       surface_flags = PIPE_BUFFER_USAGE_GPU_READ;
 196    }
 197    else {
 198       assert(0);
 199       return PIPE_ERROR_BAD_INPUT;
 200    }
 201
 202    assert(numBoxes);
 203
 204    cmd = SVGA3D_FIFOReserve(swc,
 205                             SVGA_3D_CMD_SURFACE_DMA,
 206                             sizeof *cmd + numBoxes * sizeof *boxes + sizeof *pSuffix,
 207                             2);
 208    if(!cmd)
 209       return PIPE_ERROR_OUT_OF_MEMORY;
 210
 211    swc->region_relocation(swc, &cmd->guest.ptr, guest, 0, region_flags);
 212    cmd->guest.pitch = 0;
 213
 214    swc->surface_relocation(swc, &cmd->host.sid, host, surface_flags);
 215    cmd->host.face = 0;
 216    cmd->host.mipmap = 0;
 217
 218    cmd->transfer = transfer;
 219
 220    sbuf->dma.boxes = (SVGA3dCopyBox *)&cmd[1];
 221    sbuf->dma.svga = svga;
 222
 223    /* Increment reference count */
 224    dummy = NULL;
 225    pipe_buffer_reference(&dummy, &sbuf->base);
 226
 227    pSuffix = (SVGA3dCmdSurfaceDMASuffix *)((uint8_t*)cmd + sizeof *cmd + numBoxes * sizeof *boxes);
 228    pSuffix->suffixSize = sizeof *pSuffix;
 229    pSuffix->maximumOffset = sbuf->base.size;
 230    pSuffix->flags = sbuf->dma.flags;
 231
 232    SVGA_FIFOCommitAll(swc);
 233
 234    sbuf->dma.flags.discard = FALSE;
 235
 236    return PIPE_OK;
 237 }
 238
 239
 240 /**
 241  * Patch up the upload DMA command reserved by svga_buffer_upload_command
 242  * with the final ranges.
 243  */
 244 static void
 245 svga_buffer_upload_flush(struct svga_context *svga,
 246                          struct svga_buffer *sbuf)
 247 {
 248    SVGA3dCopyBox *boxes;
 249    unsigned i;
 250
 251    assert(sbuf->handle);
 252    assert(sbuf->hwbuf);
 253    assert(sbuf->map.num_ranges);
 254    assert(sbuf->dma.svga == svga);
 255    assert(sbuf->dma.boxes);
 256
 257    /*
 258     * Patch the DMA command with the final copy box.
 259     */
 260
 261    SVGA_DBG(DEBUG_DMA, "dma to sid %p\n", sbuf->handle);
 262
 263    boxes = sbuf->dma.boxes;
 264    for(i = 0; i < sbuf->map.num_ranges; ++i) {
 265       SVGA_DBG(DEBUG_DMA, "  bytes %u - %u\n",
 266                sbuf->map.ranges[i].start, sbuf->map.ranges[i].end);
 267
 268       boxes[i].x = sbuf->map.ranges[i].start;
 269       boxes[i].y = 0;
 270       boxes[i].z = 0;
 271       boxes[i].w = sbuf->map.ranges[i].end - sbuf->map.ranges[i].start;
 272       boxes[i].h = 1;
 273       boxes[i].d = 1;
 274       boxes[i].srcx = sbuf->map.ranges[i].start;
 275       boxes[i].srcy = 0;
 276       boxes[i].srcz = 0;
 277    }
 278
 279    sbuf->map.num_ranges = 0;
 280
 281    assert(sbuf->head.prev && sbuf->head.next);
 282    LIST_DEL(&sbuf->head);
 283 #ifdef DEBUG
 284    sbuf->head.next = sbuf->head.prev = NULL;
 285 #endif
 286    sbuf->dma.pending = FALSE;
 287
 288    sbuf->dma.svga = NULL;
 289    sbuf->dma.boxes = NULL;
 290
 291    /* Decrement reference count */
 292    pipe_reference(&(sbuf->base.reference), NULL);
 293    sbuf = NULL;
 294 }
 295
 296
 297 /**
 298  * Note a dirty range.
 299  *
 300  * This function only notes the range down. It doesn't actually emit a DMA
 301  * upload command. That only happens when a context tries to refer to this
 302  * buffer, and the DMA upload command is added to that context's command buffer.
 303  *
 304  * We try to lump as many contiguous DMA transfers together as possible.
 305  */
 306 static void
 307 svga_buffer_add_range(struct svga_buffer *sbuf,
 308                       unsigned start,
 309                       unsigned end)
 310 {
 311    unsigned i;
 312    unsigned nearest_range;
 313    unsigned nearest_dist;
 314
 315    assert(end > start);
 316
 317    if (sbuf->map.num_ranges < SVGA_BUFFER_MAX_RANGES) {
 318       nearest_range = sbuf->map.num_ranges;
 319       nearest_dist = ~0;
 320    } else {
 321       nearest_range = SVGA_BUFFER_MAX_RANGES - 1;
 322       nearest_dist = 0;
 323    }
 324
 325    /*
 326     * Try to grow one of the ranges.
 327     *
 328     * Note that it is not this function task to care about overlapping ranges,
 329     * as the GMR was already given so it is too late to do anything. Situations
 330     * where overlapping ranges may pose a problem should be detected via
 331     * pipe_context::is_buffer_referenced and the context that refers to the
 332     * buffer should be flushed.
 333     */
 334
 335    for(i = 0; i < sbuf->map.num_ranges; ++i) {
 336       int left_dist;
 337       int right_dist;
 338       int dist;
 339
 340       left_dist = start - sbuf->map.ranges[i].end;
 341       right_dist = sbuf->map.ranges[i].start - end;
 342       dist = MAX2(left_dist, right_dist);
 343
 344       if (dist <= 0) {
 345          /*
 346           * Ranges are contiguous or overlapping -- extend this one and return.
 347           */
 348
 349          sbuf->map.ranges[i].start = MIN2(sbuf->map.ranges[i].start, start);
 350          sbuf->map.ranges[i].end   = MAX2(sbuf->map.ranges[i].end,   end);
 351          return;
 352       }
 353       else {
 354          /*
 355           * Discontiguous ranges -- keep track of the nearest range.
 356           */
 357
 358          if (dist < nearest_dist) {
 359             nearest_range = i;
 360             nearest_dist = dist;
 361          }
 362       }
 363    }
 364
 365    /*
 366     * We cannot add a new range to an existing DMA command, so patch-up the
 367     * pending DMA upload and start clean.
 368     */
 369
 370    if(sbuf->dma.pending)
 371       svga_buffer_upload_flush(sbuf->dma.svga, sbuf);
 372
 373    assert(!sbuf->dma.pending);
 374    assert(!sbuf->dma.svga);
 375    assert(!sbuf->dma.boxes);
 376
 377    if (sbuf->map.num_ranges < SVGA_BUFFER_MAX_RANGES) {
 378       /*
 379        * Add a new range.
 380        */
 381
 382       sbuf->map.ranges[sbuf->map.num_ranges].start = start;
 383       sbuf->map.ranges[sbuf->map.num_ranges].end = end;
 384       ++sbuf->map.num_ranges;
 385    } else {
 386       /*
 387        * Everything else failed, so just extend the nearest range.
 388        *
 389        * It is OK to do this because we always keep a local copy of the
 390        * host buffer data, for SW TNL, and the host never modifies the buffer.
 391        */
 392
 393       assert(nearest_range < SVGA_BUFFER_MAX_RANGES);
 394       assert(nearest_range < sbuf->map.num_ranges);
 395       sbuf->map.ranges[nearest_range].start = MIN2(sbuf->map.ranges[nearest_range].start, start);
 396       sbuf->map.ranges[nearest_range].end   = MAX2(sbuf->map.ranges[nearest_range].end,   end);
 397    }
 398 }
 399
 400
 401 static void *
 402 svga_buffer_map_range( struct pipe_screen *screen,
 403                        struct pipe_buffer *buf,
 404                        unsigned offset, unsigned length,
 405                        unsigned usage )
 406 {
 407    struct svga_screen *ss = svga_screen(screen);
 408    struct svga_winsys_screen *sws = ss->sws;
 409    struct svga_buffer *sbuf = svga_buffer( buf );
 410    void *map;
 411
 412    if (!sbuf->swbuf && !sbuf->hwbuf) {
 413       if (svga_buffer_create_hw_storage(ss, sbuf) != PIPE_OK) {
 414          /*
 415           * We can't create a hardware buffer big enough, so create a malloc
 416           * buffer instead.
 417           */
 418
 419          debug_printf("%s: failed to allocate %u KB of DMA, splitting DMA transfers\n",
 420                       __FUNCTION__,
 421                       (sbuf->base.size + 1023)/1024);
 422
 423          sbuf->swbuf = align_malloc(sbuf->base.size, sbuf->base.alignment);
 424       }
 425    }
 426
 427    if (sbuf->swbuf) {
 428       /* User/malloc buffer */
 429       map = sbuf->swbuf;
 430    }
 431    else if (sbuf->hwbuf) {
 432       map = sws->buffer_map(sws, sbuf->hwbuf, usage);
 433    }
 434    else {
 435       map = NULL;
 436    }
 437
 438    if(map) {
 439       pipe_mutex_lock(ss->swc_mutex);
 440
 441       ++sbuf->map.count;
 442
 443       if (usage & PIPE_BUFFER_USAGE_CPU_WRITE) {
 444          assert(sbuf->map.count <= 1);
 445          sbuf->map.writing = TRUE;
 446          if (usage & PIPE_BUFFER_USAGE_FLUSH_EXPLICIT)
 447             sbuf->map.flush_explicit = TRUE;
 448       }
 449
 450       pipe_mutex_unlock(ss->swc_mutex);
 451    }
 452
 453    return map;
 454 }
 455
 456 static void
 457 svga_buffer_flush_mapped_range( struct pipe_screen *screen,
 458                                 struct pipe_buffer *buf,
 459                                 unsigned offset, unsigned length)
 460 {
 461    struct svga_buffer *sbuf = svga_buffer( buf );
 462    struct svga_screen *ss = svga_screen(screen);
 463
 464    pipe_mutex_lock(ss->swc_mutex);
 465    assert(sbuf->map.writing);
 466    if(sbuf->map.writing) {
 467       assert(sbuf->map.flush_explicit);
 468       svga_buffer_add_range(sbuf, offset, offset + length);
 469    }
 470    pipe_mutex_unlock(ss->swc_mutex);
 471 }
 472
 473 static void
 474 svga_buffer_unmap( struct pipe_screen *screen,
 475                    struct pipe_buffer *buf)
 476 {
 477    struct svga_screen *ss = svga_screen(screen);
 478    struct svga_winsys_screen *sws = ss->sws;
 479    struct svga_buffer *sbuf = svga_buffer( buf );
 480
 481    pipe_mutex_lock(ss->swc_mutex);
 482
 483    assert(sbuf->map.count);
 484    if(sbuf->map.count)
 485       --sbuf->map.count;
 486
 487    if(sbuf->hwbuf)
 488       sws->buffer_unmap(sws, sbuf->hwbuf);
 489
 490    if(sbuf->map.writing) {
 491       if(!sbuf->map.flush_explicit) {
 492          /* No mapped range was flushed -- flush the whole buffer */
 493          SVGA_DBG(DEBUG_DMA, "flushing the whole buffer\n");
 494
 495          svga_buffer_add_range(sbuf, 0, sbuf->base.size);
 496       }
 497
 498       sbuf->map.writing = FALSE;
 499       sbuf->map.flush_explicit = FALSE;
 500    }
 501
 502    pipe_mutex_unlock(ss->swc_mutex);
 503 }
 504
 505 static void
 506 svga_buffer_destroy( struct pipe_buffer *buf )
 507 {
 508    struct svga_screen *ss = svga_screen(buf->screen);
 509    struct svga_buffer *sbuf = svga_buffer( buf );
 510
 511    assert(!p_atomic_read(&buf->reference.count));
 512
 513    assert(!sbuf->dma.pending);
 514
 515    if(sbuf->handle)
 516       svga_buffer_destroy_host_surface(ss, sbuf);
 517
 518    if(sbuf->hwbuf)
 519       svga_buffer_destroy_hw_storage(ss, sbuf);
 520
 521    if(sbuf->swbuf && !sbuf->user)
 522       align_free(sbuf->swbuf);
 523
 524    FREE(sbuf);
 525 }
 526
 527 static struct pipe_buffer *
 528 svga_buffer_create(struct pipe_screen *screen,
 529                    unsigned alignment,
 530                    unsigned usage,
 531                    unsigned size)
 532 {
 533    struct svga_screen *ss = svga_screen(screen);
 534    struct svga_buffer *sbuf;
 535
 536    assert(size);
 537    assert(alignment);
 538
 539    sbuf = CALLOC_STRUCT(svga_buffer);
 540    if(!sbuf)
 541       goto error1;
 542
 543    sbuf->magic = SVGA_BUFFER_MAGIC;
 544
 545    pipe_reference_init(&sbuf->base.reference, 1);
 546    sbuf->base.screen = screen;
 547    sbuf->base.alignment = alignment;
 548    sbuf->base.usage = usage;
 549    sbuf->base.size = size;
 550
 551    if(svga_buffer_needs_hw_storage(usage)) {
 552       if(svga_buffer_create_host_surface(ss, sbuf) != PIPE_OK)
 553          goto error2;
 554    }
 555    else {
 556       if(alignment < sizeof(void*))
 557          alignment = sizeof(void*);
 558
 559       usage |= PIPE_BUFFER_USAGE_CPU_READ_WRITE;
 560
 561       sbuf->swbuf = align_malloc(size, alignment);
 562       if(!sbuf->swbuf)
 563          goto error2;
 564    }
 565
 566    return &sbuf->base;
 567
 568 error2:
 569    FREE(sbuf);
 570 error1:
 571    return NULL;
 572 }
 573
 574 static struct pipe_buffer *
 575 svga_user_buffer_create(struct pipe_screen *screen,
 576                         void *ptr,
 577                         unsigned bytes)
 578 {
 579    struct svga_buffer *sbuf;
 580
 581    sbuf = CALLOC_STRUCT(svga_buffer);
 582    if(!sbuf)
 583       goto no_sbuf;
 584
 585    sbuf->magic = SVGA_BUFFER_MAGIC;
 586
 587    sbuf->swbuf = ptr;
 588    sbuf->user = TRUE;
 589
 590    pipe_reference_init(&sbuf->base.reference, 1);
 591    sbuf->base.screen = screen;
 592    sbuf->base.alignment = 1;
 593    sbuf->base.usage = 0;
 594    sbuf->base.size = bytes;
 595
 596    return &sbuf->base;
 597
 598 no_sbuf:
 599    return NULL;
 600 }
 601
 602
 603 void
 604 svga_screen_init_buffer_functions(struct pipe_screen *screen)
 605 {
 606    screen->buffer_create = svga_buffer_create;
 607    screen->user_buffer_create = svga_user_buffer_create;
 608    screen->buffer_map_range = svga_buffer_map_range;
 609    screen->buffer_flush_mapped_range = svga_buffer_flush_mapped_range;
 610    screen->buffer_unmap = svga_buffer_unmap;
 611    screen->buffer_destroy = svga_buffer_destroy;
 612 }
 613
 614
 615 /**
 616  * Copy the contents of the user buffer / malloc buffer to a hardware buffer.
 617  */
 618 static INLINE enum pipe_error
 619 svga_buffer_update_hw(struct svga_screen *ss, struct svga_buffer *sbuf)
 620 {
 621    if(!sbuf->hwbuf) {
 622       enum pipe_error ret;
 623       void *map;
 624
 625       assert(sbuf->swbuf);
 626       if(!sbuf->swbuf)
 627          return PIPE_ERROR;
 628
 629       ret = svga_buffer_create_hw_storage(ss, sbuf);
 630       if(ret != PIPE_OK)
 631          return ret;
 632
 633       pipe_mutex_lock(ss->swc_mutex);
 634       map = ss->sws->buffer_map(ss->sws, sbuf->hwbuf, PIPE_BUFFER_USAGE_CPU_WRITE);
 635       assert(map);
 636       if(!map) {
 637          pipe_mutex_unlock(ss->swc_mutex);
 638          svga_buffer_destroy_hw_storage(ss, sbuf);
 639          return PIPE_ERROR;
 640       }
 641
 642       memcpy(map, sbuf->swbuf, sbuf->base.size);
 643       ss->sws->buffer_unmap(ss->sws, sbuf->hwbuf);
 644
 645       /* This user/malloc buffer is now indistinguishable from a gpu buffer */
 646       assert(!sbuf->map.count);
 647       if(!sbuf->map.count) {
 648          if(sbuf->user)
 649             sbuf->user = FALSE;
 650          else
 651             align_free(sbuf->swbuf);
 652          sbuf->swbuf = NULL;
 653       }
 654
 655       pipe_mutex_unlock(ss->swc_mutex);
 656    }
 657
 658    return PIPE_OK;
 659 }
 660
 661
 662 /**
 663  * Upload the buffer to the host in a piecewise fashion.
 664  *
 665  * Used when the buffer is too big to fit in the GMR aperture.
 666  */
 667 static INLINE enum pipe_error
 668 svga_buffer_upload_piecewise(struct svga_screen *ss,
 669                              struct svga_context *svga,
 670                              struct svga_buffer *sbuf)
 671 {
 672    struct svga_winsys_screen *sws = ss->sws;
 673    const unsigned alignment = sizeof(void *);
 674    const unsigned usage = 0;
 675    unsigned i;
 676
 677    assert(sbuf->map.num_ranges);
 678    assert(!sbuf->dma.pending);
 679
 680    SVGA_DBG(DEBUG_DMA, "dma to sid %p\n", sbuf->handle);
 681
 682    for (i = 0; i < sbuf->map.num_ranges; ++i) {
 683       struct svga_buffer_range *range = &sbuf->map.ranges[i];
 684       unsigned offset = range->start;
 685       unsigned size = range->end - range->start;
 686
 687       while (offset < range->end) {
 688          struct svga_winsys_buffer *hwbuf;
 689          uint8_t *map;
 690          enum pipe_error ret;
 691
 692          if (offset + size > range->end)
 693             size = range->end - offset;
 694
 695          hwbuf = svga_winsys_buffer_create(ss, alignment, usage, size);
 696          while (!hwbuf) {
 697             size /= 2;
 698             if (!size)
 699                return PIPE_ERROR_OUT_OF_MEMORY;
 700             hwbuf = svga_winsys_buffer_create(ss, alignment, usage, size);
 701          }
 702
 703          SVGA_DBG(DEBUG_DMA, "  bytes %u - %u\n",
 704                   offset, offset + size);
 705
 706          map = sws->buffer_map(sws, hwbuf,
 707                                PIPE_BUFFER_USAGE_CPU_WRITE |
 708                                PIPE_BUFFER_USAGE_DISCARD);
 709          assert(map);
 710          if (map) {
 711             memcpy(map, sbuf->swbuf, size);
 712             sws->buffer_unmap(sws, hwbuf);
 713          }
 714
 715          ret = SVGA3D_BufferDMA(svga->swc,
 716                                 hwbuf, sbuf->handle,
 717                                 SVGA3D_WRITE_HOST_VRAM,
 718                                 size, offset, sbuf->dma.flags);
 719          if(ret != PIPE_OK) {
 720             svga_context_flush(svga, NULL);
 721             ret =  SVGA3D_BufferDMA(svga->swc,
 722                                     hwbuf, sbuf->handle,
 723                                     SVGA3D_WRITE_HOST_VRAM,
 724                                     size, offset, sbuf->dma.flags);
 725             assert(ret == PIPE_OK);
 726          }
 727
 728          sbuf->dma.flags.discard = FALSE;
 729
 730          sws->buffer_destroy(sws, hwbuf);
 731
 732          offset += size;
 733       }
 734    }
 735
 736    sbuf->map.num_ranges = 0;
 737
 738    return PIPE_OK;
 739 }
 740
 741
 742 struct svga_winsys_surface *
 743 svga_buffer_handle(struct svga_context *svga,
 744                    struct pipe_buffer *buf)
 745 {
 746    struct pipe_screen *screen = svga->pipe.screen;
 747    struct svga_screen *ss = svga_screen(screen);
 748    struct svga_buffer *sbuf;
 749    enum pipe_error ret;
 750
 751    if(!buf)
 752       return NULL;
 753
 754    sbuf = svga_buffer(buf);
 755
 756    assert(!sbuf->map.count);
 757
 758    if(!sbuf->handle) {
 759       ret = svga_buffer_create_host_surface(ss, sbuf);
 760       if(ret != PIPE_OK)
 761          return NULL;
 762    }
 763
 764    assert(sbuf->handle);
 765
 766    if (sbuf->map.num_ranges) {
 767       if (!sbuf->dma.pending) {
 768          /*
 769           * No pending DMA upload yet, so insert a DMA upload command now.
 770           */
 771
 772          /*
 773           * Migrate the data from swbuf -> hwbuf if necessary.
 774           */
 775          ret = svga_buffer_update_hw(ss, sbuf);
 776          if (ret == PIPE_OK) {
 777             /*
 778              * Queue a dma command.
 779              */
 780
 781             ret = svga_buffer_upload_command(svga, sbuf);
 782             if (ret == PIPE_ERROR_OUT_OF_MEMORY) {
 783                svga_context_flush(svga, NULL);
 784                ret = svga_buffer_upload_command(svga, sbuf);
 785                assert(ret == PIPE_OK);
 786             }
 787             if (ret == PIPE_OK) {
 788                sbuf->dma.pending = TRUE;
 789                assert(!sbuf->head.prev && !sbuf->head.next);
 790                LIST_ADDTAIL(&sbuf->head, &svga->dirty_buffers);
 791             }
 792          }
 793          else if (ret == PIPE_ERROR_OUT_OF_MEMORY) {
 794             /*
 795              * The buffer is too big to fit in the GMR aperture, so break it in
 796              * smaller pieces.
 797              */
 798             ret = svga_buffer_upload_piecewise(ss, svga, sbuf);
 799          }
 800
 801          if (ret != PIPE_OK) {
 802             /*
 803              * Something unexpected happened above. There is very little that
 804              * we can do other than proceeding while ignoring the dirty ranges.
 805              */
 806             assert(0);
 807             sbuf->map.num_ranges = 0;
 808          }
 809       }
 810       else {
 811          /*
 812           * There a pending dma already. Make sure it is from this context.
 813           */
 814          assert(sbuf->dma.svga == svga);
 815       }
 816    }
 817
 818    assert(!sbuf->map.num_ranges || sbuf->dma.pending);
 819
 820    return sbuf->handle;
 821 }
 822
 823
 824 struct pipe_buffer *
 825 svga_screen_buffer_wrap_surface(struct pipe_screen *screen,
 826                                 enum SVGA3dSurfaceFormat format,
 827                                 struct svga_winsys_surface *srf)
 828 {
 829    struct pipe_buffer *buf;
 830    struct svga_buffer *sbuf;
 831    struct svga_winsys_screen *sws = svga_winsys_screen(screen);
 832
 833    buf = svga_buffer_create(screen, 0, SVGA_BUFFER_USAGE_WRAPPED, 0);
 834    if (!buf)
 835       return NULL;
 836
 837    sbuf = svga_buffer(buf);
 838
 839    /*
 840     * We are not the creator of this surface and therefore we must not
 841     * cache it for reuse. Set the cacheable flag to zero in the key to
 842     * prevent this.
 843     */
 844    sbuf->key.format = format;
 845    sbuf->key.cachable = 0;
 846    sws->surface_reference(sws, &sbuf->handle, srf);
 847
 848    return buf;
 849 }
 850
 851
 852 struct svga_winsys_surface *
 853 svga_screen_buffer_get_winsys_surface(struct pipe_buffer *buffer)
 854 {
 855    struct svga_winsys_screen *sws = svga_winsys_screen(buffer->screen);
 856    struct svga_winsys_surface *vsurf = NULL;
 857
 858    assert(svga_buffer(buffer)->key.cachable == 0);
 859    svga_buffer(buffer)->key.cachable = 0;
 860    sws->surface_reference(sws, &vsurf, svga_buffer(buffer)->handle);
 861    return vsurf;
 862 }
 863
 864 void
 865 svga_context_flush_buffers(struct svga_context *svga)
 866 {
 867    struct list_head *curr, *next;
 868    struct svga_buffer *sbuf;
 869
 870    curr = svga->dirty_buffers.next;
 871    next = curr->next;
 872    while(curr != &svga->dirty_buffers) {
 873       sbuf = LIST_ENTRY(struct svga_buffer, curr, head);
 874
 875       assert(p_atomic_read(&sbuf->base.reference.count) != 0);
 876       assert(sbuf->dma.pending);
 877
 878       svga_buffer_upload_flush(svga, sbuf);
 879
 880       curr = next;
 881       next = curr->next;
 882    }
 883 }