src/gallium/drivers/svga/svga_resource_buffer_upload.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_resource_buffer.h"
  38 #include "svga_resource_buffer_upload.h"
  39 #include "svga_winsys.h"
  40 #include "svga_debug.h"
  41
  42
  43 /**
  44  * Allocate a winsys_buffer (ie. DMA, aka GMR memory).
  45  *
  46  * It will flush and retry in case the first attempt to create a DMA buffer
  47  * fails, so it should not be called from any function involved in flushing
  48  * to avoid recursion.
  49  */
  50 struct svga_winsys_buffer *
  51 svga_winsys_buffer_create( struct svga_context *svga,
  52                            unsigned alignment,
  53                            unsigned usage,
  54                            unsigned size )
  55 {
  56    struct svga_screen *svgascreen = svga_screen(svga->pipe.screen);
  57    struct svga_winsys_screen *sws = svgascreen->sws;
  58    struct svga_winsys_buffer *buf;
  59
  60    /* Just try */
  61    buf = sws->buffer_create(sws, alignment, usage, size);
  62    if(!buf) {
  63
  64       SVGA_DBG(DEBUG_DMA|DEBUG_PERF, "flushing screen to find %d bytes GMR\n",
  65                size);
  66
  67       /* Try flushing all pending DMAs */
  68       svga_context_flush(svga, NULL);
  69       buf = sws->buffer_create(sws, alignment, usage, size);
  70    }
  71
  72    return buf;
  73 }
  74
  75
  76 void
  77 svga_buffer_destroy_hw_storage(struct svga_screen *ss, struct svga_buffer *sbuf)
  78 {
  79    struct svga_winsys_screen *sws = ss->sws;
  80
  81    assert(!sbuf->map.count);
  82    assert(sbuf->hwbuf);
  83    if(sbuf->hwbuf) {
  84       sws->buffer_destroy(sws, sbuf->hwbuf);
  85       sbuf->hwbuf = NULL;
  86    }
  87 }
  88
  89
  90
  91 /**
  92  * Allocate DMA'ble storage for the buffer.
  93  *
  94  * Called before mapping a buffer.
  95  */
  96 enum pipe_error
  97 svga_buffer_create_hw_storage(struct svga_screen *ss,
  98                               struct svga_buffer *sbuf)
  99 {
 100    assert(!sbuf->user);
 101
 102    if(!sbuf->hwbuf) {
 103       struct svga_winsys_screen *sws = ss->sws;
 104       unsigned alignment = 16;
 105       unsigned usage = 0;
 106       unsigned size = sbuf->b.b.width0;
 107
 108       sbuf->hwbuf = sws->buffer_create(sws, alignment, usage, size);
 109       if(!sbuf->hwbuf)
 110          return PIPE_ERROR_OUT_OF_MEMORY;
 111
 112       assert(!sbuf->dma.pending);
 113    }
 114
 115    return PIPE_OK;
 116 }
 117
 118
 119
 120 enum pipe_error
 121 svga_buffer_create_host_surface(struct svga_screen *ss,
 122                                 struct svga_buffer *sbuf)
 123 {
 124    if(!sbuf->handle) {
 125       sbuf->key.flags = 0;
 126
 127       sbuf->key.format = SVGA3D_BUFFER;
 128       if(sbuf->b.b.bind & PIPE_BIND_VERTEX_BUFFER)
 129          sbuf->key.flags |= SVGA3D_SURFACE_HINT_VERTEXBUFFER;
 130       if(sbuf->b.b.bind & PIPE_BIND_INDEX_BUFFER)
 131          sbuf->key.flags |= SVGA3D_SURFACE_HINT_INDEXBUFFER;
 132
 133       sbuf->key.size.width = sbuf->b.b.width0;
 134       sbuf->key.size.height = 1;
 135       sbuf->key.size.depth = 1;
 136
 137       sbuf->key.numFaces = 1;
 138       sbuf->key.numMipLevels = 1;
 139       sbuf->key.cachable = 1;
 140
 141       SVGA_DBG(DEBUG_DMA, "surface_create for buffer sz %d\n", sbuf->b.b.width0);
 142
 143       sbuf->handle = svga_screen_surface_create(ss, &sbuf->key);
 144       if(!sbuf->handle)
 145          return PIPE_ERROR_OUT_OF_MEMORY;
 146
 147       /* Always set the discard flag on the first time the buffer is written
 148        * as svga_screen_surface_create might have passed a recycled host
 149        * buffer.
 150        */
 151       sbuf->dma.flags.discard = TRUE;
 152
 153       SVGA_DBG(DEBUG_DMA, "   --> got sid %p sz %d (buffer)\n", sbuf->handle, sbuf->b.b.width0);
 154    }
 155
 156    return PIPE_OK;
 157 }
 158
 159
 160 void
 161 svga_buffer_destroy_host_surface(struct svga_screen *ss,
 162                                  struct svga_buffer *sbuf)
 163 {
 164    if(sbuf->handle) {
 165       SVGA_DBG(DEBUG_DMA, " ungrab sid %p sz %d\n", sbuf->handle, sbuf->b.b.width0);
 166       svga_screen_surface_destroy(ss, &sbuf->key, &sbuf->handle);
 167    }
 168 }
 169
 170
 171 /**
 172  * Variant of SVGA3D_BufferDMA which leaves the copy box temporarily in blank.
 173  */
 174 static enum pipe_error
 175 svga_buffer_upload_command(struct svga_context *svga,
 176                            struct svga_buffer *sbuf)
 177 {
 178    struct svga_winsys_context *swc = svga->swc;
 179    struct svga_winsys_buffer *guest = sbuf->hwbuf;
 180    struct svga_winsys_surface *host = sbuf->handle;
 181    SVGA3dTransferType transfer = SVGA3D_WRITE_HOST_VRAM;
 182    SVGA3dCmdSurfaceDMA *cmd;
 183    uint32 numBoxes = sbuf->map.num_ranges;
 184    SVGA3dCopyBox *boxes;
 185    SVGA3dCmdSurfaceDMASuffix *pSuffix;
 186    unsigned region_flags;
 187    unsigned surface_flags;
 188    struct pipe_resource *dummy;
 189
 190    if(transfer == SVGA3D_WRITE_HOST_VRAM) {
 191       region_flags = SVGA_RELOC_READ;
 192       surface_flags = SVGA_RELOC_WRITE;
 193    }
 194    else if(transfer == SVGA3D_READ_HOST_VRAM) {
 195       region_flags = SVGA_RELOC_WRITE;
 196       surface_flags = SVGA_RELOC_READ;
 197    }
 198    else {
 199       assert(0);
 200       return PIPE_ERROR_BAD_INPUT;
 201    }
 202
 203    assert(numBoxes);
 204
 205    cmd = SVGA3D_FIFOReserve(swc,
 206                             SVGA_3D_CMD_SURFACE_DMA,
 207                             sizeof *cmd + numBoxes * sizeof *boxes + sizeof *pSuffix,
 208                             2);
 209    if(!cmd)
 210       return PIPE_ERROR_OUT_OF_MEMORY;
 211
 212    swc->region_relocation(swc, &cmd->guest.ptr, guest, 0, region_flags);
 213    cmd->guest.pitch = 0;
 214
 215    swc->surface_relocation(swc, &cmd->host.sid, host, surface_flags);
 216    cmd->host.face = 0;
 217    cmd->host.mipmap = 0;
 218
 219    cmd->transfer = transfer;
 220
 221    sbuf->dma.boxes = (SVGA3dCopyBox *)&cmd[1];
 222    sbuf->dma.svga = svga;
 223
 224    /* Increment reference count */
 225    dummy = NULL;
 226    pipe_resource_reference(&dummy, &sbuf->b.b);
 227
 228    pSuffix = (SVGA3dCmdSurfaceDMASuffix *)((uint8_t*)cmd + sizeof *cmd + numBoxes * sizeof *boxes);
 229    pSuffix->suffixSize = sizeof *pSuffix;
 230    pSuffix->maximumOffset = sbuf->b.b.width0;
 231    pSuffix->flags = sbuf->dma.flags;
 232
 233    SVGA_FIFOCommitAll(swc);
 234
 235    sbuf->dma.flags.discard = FALSE;
 236
 237    return PIPE_OK;
 238 }
 239
 240
 241 /**
 242  * Patch up the upload DMA command reserved by svga_buffer_upload_command
 243  * with the final ranges.
 244  */
 245 static void
 246 svga_buffer_upload_flush(struct svga_context *svga,
 247                          struct svga_buffer *sbuf)
 248 {
 249    SVGA3dCopyBox *boxes;
 250    unsigned i;
 251
 252    assert(sbuf->handle);
 253    assert(sbuf->hwbuf);
 254    assert(sbuf->map.num_ranges);
 255    assert(sbuf->dma.svga == svga);
 256    assert(sbuf->dma.boxes);
 257
 258    /*
 259     * Patch the DMA command with the final copy box.
 260     */
 261
 262    SVGA_DBG(DEBUG_DMA, "dma to sid %p\n", sbuf->handle);
 263
 264    boxes = sbuf->dma.boxes;
 265    for(i = 0; i < sbuf->map.num_ranges; ++i) {
 266       SVGA_DBG(DEBUG_DMA, "  bytes %u - %u\n",
 267                sbuf->map.ranges[i].start, sbuf->map.ranges[i].end);
 268
 269       boxes[i].x = sbuf->map.ranges[i].start;
 270       boxes[i].y = 0;
 271       boxes[i].z = 0;
 272       boxes[i].w = sbuf->map.ranges[i].end - sbuf->map.ranges[i].start;
 273       boxes[i].h = 1;
 274       boxes[i].d = 1;
 275       boxes[i].srcx = sbuf->map.ranges[i].start;
 276       boxes[i].srcy = 0;
 277       boxes[i].srcz = 0;
 278    }
 279
 280    sbuf->map.num_ranges = 0;
 281
 282    assert(sbuf->head.prev && sbuf->head.next);
 283    LIST_DEL(&sbuf->head);
 284 #ifdef DEBUG
 285    sbuf->head.next = sbuf->head.prev = NULL;
 286 #endif
 287    sbuf->dma.pending = FALSE;
 288
 289    sbuf->dma.svga = NULL;
 290    sbuf->dma.boxes = NULL;
 291
 292    /* Decrement reference count */
 293    pipe_reference(&(sbuf->b.b.reference), NULL);
 294    sbuf = NULL;
 295 }
 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 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_resource_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
 404 /**
 405  * Copy the contents of the malloc buffer to a hardware buffer.
 406  */
 407 static INLINE enum pipe_error
 408 svga_buffer_update_hw(struct svga_screen *ss, struct svga_buffer *sbuf)
 409 {
 410    assert(!sbuf->user);
 411    if(!sbuf->hwbuf) {
 412       enum pipe_error ret;
 413       void *map;
 414
 415       assert(sbuf->swbuf);
 416       if(!sbuf->swbuf)
 417          return PIPE_ERROR;
 418
 419       ret = svga_buffer_create_hw_storage(ss, sbuf);
 420       if(ret != PIPE_OK)
 421          return ret;
 422
 423       pipe_mutex_lock(ss->swc_mutex);
 424       map = ss->sws->buffer_map(ss->sws, sbuf->hwbuf, PIPE_TRANSFER_WRITE);
 425       assert(map);
 426       if(!map) {
 427          pipe_mutex_unlock(ss->swc_mutex);
 428          svga_buffer_destroy_hw_storage(ss, sbuf);
 429          return PIPE_ERROR;
 430       }
 431
 432       memcpy(map, sbuf->swbuf, sbuf->b.b.width0);
 433       ss->sws->buffer_unmap(ss->sws, sbuf->hwbuf);
 434
 435       /* This user/malloc buffer is now indistinguishable from a gpu buffer */
 436       assert(!sbuf->map.count);
 437       if(!sbuf->map.count) {
 438          if(sbuf->user)
 439             sbuf->user = FALSE;
 440          else
 441             align_free(sbuf->swbuf);
 442          sbuf->swbuf = NULL;
 443       }
 444
 445       pipe_mutex_unlock(ss->swc_mutex);
 446    }
 447
 448    return PIPE_OK;
 449 }
 450
 451
 452 /**
 453  * Upload the buffer to the host in a piecewise fashion.
 454  *
 455  * Used when the buffer is too big to fit in the GMR aperture.
 456  */
 457 static INLINE enum pipe_error
 458 svga_buffer_upload_piecewise(struct svga_screen *ss,
 459                              struct svga_context *svga,
 460                              struct svga_buffer *sbuf)
 461 {
 462    struct svga_winsys_screen *sws = ss->sws;
 463    const unsigned alignment = sizeof(void *);
 464    const unsigned usage = 0;
 465    unsigned i;
 466
 467    assert(sbuf->map.num_ranges);
 468    assert(!sbuf->dma.pending);
 469
 470    SVGA_DBG(DEBUG_DMA, "dma to sid %p\n", sbuf->handle);
 471
 472    for (i = 0; i < sbuf->map.num_ranges; ++i) {
 473       struct svga_buffer_range *range = &sbuf->map.ranges[i];
 474       unsigned offset = range->start;
 475       unsigned size = range->end - range->start;
 476
 477       while (offset < range->end) {
 478          struct svga_winsys_buffer *hwbuf;
 479          uint8_t *map;
 480          enum pipe_error ret;
 481
 482          if (offset + size > range->end)
 483             size = range->end - offset;
 484
 485          hwbuf = sws->buffer_create(sws, alignment, usage, size);
 486          while (!hwbuf) {
 487             size /= 2;
 488             if (!size)
 489                return PIPE_ERROR_OUT_OF_MEMORY;
 490             hwbuf = sws->buffer_create(sws, alignment, usage, size);
 491          }
 492
 493          SVGA_DBG(DEBUG_DMA, "  bytes %u - %u\n",
 494                   offset, offset + size);
 495
 496          map = sws->buffer_map(sws, hwbuf,
 497                                PIPE_TRANSFER_WRITE |
 498                                PIPE_TRANSFER_DISCARD);
 499          assert(map);
 500          if (map) {
 501             memcpy(map, sbuf->swbuf, size);
 502             sws->buffer_unmap(sws, hwbuf);
 503          }
 504
 505          ret = SVGA3D_BufferDMA(svga->swc,
 506                                 hwbuf, sbuf->handle,
 507                                 SVGA3D_WRITE_HOST_VRAM,
 508                                 size, 0, offset, sbuf->dma.flags);
 509          if(ret != PIPE_OK) {
 510             svga_context_flush(svga, NULL);
 511             ret =  SVGA3D_BufferDMA(svga->swc,
 512                                     hwbuf, sbuf->handle,
 513                                     SVGA3D_WRITE_HOST_VRAM,
 514                                     size, 0, offset, sbuf->dma.flags);
 515             assert(ret == PIPE_OK);
 516          }
 517
 518          sbuf->dma.flags.discard = FALSE;
 519
 520          sws->buffer_destroy(sws, hwbuf);
 521
 522          offset += size;
 523       }
 524    }
 525
 526    sbuf->map.num_ranges = 0;
 527
 528    return PIPE_OK;
 529 }
 530
 531
 532
 533
 534 /* Get (or create/upload) the winsys surface handle so that we can
 535  * refer to this buffer in fifo commands.
 536  */
 537 struct svga_winsys_surface *
 538 svga_buffer_handle(struct svga_context *svga,
 539                    struct pipe_resource *buf)
 540 {
 541    struct pipe_screen *screen = svga->pipe.screen;
 542    struct svga_screen *ss = svga_screen(screen);
 543    struct svga_buffer *sbuf;
 544    enum pipe_error ret;
 545
 546    if(!buf)
 547       return NULL;
 548
 549    sbuf = svga_buffer(buf);
 550
 551    assert(!sbuf->map.count);
 552    assert(!sbuf->user);
 553
 554    if(!sbuf->handle) {
 555       ret = svga_buffer_create_host_surface(ss, sbuf);
 556       if(ret != PIPE_OK)
 557          return NULL;
 558    }
 559
 560    assert(sbuf->handle);
 561
 562    if (sbuf->map.num_ranges) {
 563       if (!sbuf->dma.pending) {
 564          /*
 565           * No pending DMA upload yet, so insert a DMA upload command now.
 566           */
 567
 568          /*
 569           * Migrate the data from swbuf -> hwbuf if necessary.
 570           */
 571          ret = svga_buffer_update_hw(ss, sbuf);
 572          if (ret == PIPE_OK) {
 573             /*
 574              * Queue a dma command.
 575              */
 576
 577             ret = svga_buffer_upload_command(svga, sbuf);
 578             if (ret == PIPE_ERROR_OUT_OF_MEMORY) {
 579                svga_context_flush(svga, NULL);
 580                ret = svga_buffer_upload_command(svga, sbuf);
 581                assert(ret == PIPE_OK);
 582             }
 583             if (ret == PIPE_OK) {
 584                sbuf->dma.pending = TRUE;
 585                assert(!sbuf->head.prev && !sbuf->head.next);
 586                LIST_ADDTAIL(&sbuf->head, &svga->dirty_buffers);
 587             }
 588          }
 589          else if (ret == PIPE_ERROR_OUT_OF_MEMORY) {
 590             /*
 591              * The buffer is too big to fit in the GMR aperture, so break it in
 592              * smaller pieces.
 593              */
 594             ret = svga_buffer_upload_piecewise(ss, svga, sbuf);
 595          }
 596
 597          if (ret != PIPE_OK) {
 598             /*
 599              * Something unexpected happened above. There is very little that
 600              * we can do other than proceeding while ignoring the dirty ranges.
 601              */
 602             assert(0);
 603             sbuf->map.num_ranges = 0;
 604          }
 605       }
 606       else {
 607          /*
 608           * There a pending dma already. Make sure it is from this context.
 609           */
 610          assert(sbuf->dma.svga == svga);
 611       }
 612    }
 613
 614    assert(!sbuf->map.num_ranges || sbuf->dma.pending);
 615
 616    return sbuf->handle;
 617 }
 618
 619
 620
 621 void
 622 svga_context_flush_buffers(struct svga_context *svga)
 623 {
 624    struct list_head *curr, *next;
 625    struct svga_buffer *sbuf;
 626
 627    curr = svga->dirty_buffers.next;
 628    next = curr->next;
 629    while(curr != &svga->dirty_buffers) {
 630       sbuf = LIST_ENTRY(struct svga_buffer, curr, head);
 631
 632       assert(p_atomic_read(&sbuf->b.b.reference.count) != 0);
 633       assert(sbuf->dma.pending);
 634
 635       svga_buffer_upload_flush(svga, sbuf);
 636
 637       curr = next;
 638       next = curr->next;
 639    }
 640 }