src/gallium/drivers/svga/svga_resource_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 #include "util/u_resource.h"
  35
  36 #include "svga_context.h"
  37 #include "svga_screen.h"
  38 #include "svga_resource_buffer.h"
  39 #include "svga_resource_buffer_upload.h"
  40 #include "svga_winsys.h"
  41 #include "svga_debug.h"
  42
  43
  44 /**
  45  * Vertex and index buffers need hardware backing.  Constant buffers
  46  * do not.  No other types of buffers currently supported.
  47  */
  48 static INLINE boolean
  49 svga_buffer_needs_hw_storage(unsigned usage)
  50 {
  51    return usage & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER);
  52 }
  53
  54
  55 /**
  56  * Create a buffer transfer.
  57  *
  58  * Unlike texture DMAs (which are written immediately to the command buffer and
  59  * therefore inherently serialized with other context operations), for buffers
  60  * we try to coalesce multiple range mappings (i.e, multiple calls to this
  61  * function) into a single DMA command, for better efficiency in command
  62  * processing.  This means we need to exercise extra care here to ensure that
  63  * the end result is exactly the same as if one DMA was used for every mapped
  64  * range.
  65  */
  66 static void *
  67 svga_buffer_transfer_map(struct pipe_context *pipe,
  68                          struct pipe_resource *resource,
  69                          unsigned level,
  70                          unsigned usage,
  71                          const struct pipe_box *box,
  72                          struct pipe_transfer **ptransfer)
  73 {
  74    struct svga_context *svga = svga_context(pipe);
  75    struct svga_screen *ss = svga_screen(pipe->screen);
  76    struct svga_buffer *sbuf = svga_buffer(resource);
  77    struct pipe_transfer *transfer;
  78    uint8_t *map;
  79
  80    transfer = CALLOC_STRUCT(pipe_transfer);
  81    if (transfer == NULL) {
  82       return NULL;
  83    }
  84
  85    transfer->resource = resource;
  86    transfer->level = level;
  87    transfer->usage = usage;
  88    transfer->box = *box;
  89
  90    if (usage & PIPE_TRANSFER_WRITE) {
  91       if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
  92          /*
  93           * Flush any pending primitives, finish writing any pending DMA
  94           * commands, and tell the host to discard the buffer contents on
  95           * the next DMA operation.
  96           */
  97
  98          svga_hwtnl_flush_buffer(svga, resource);
  99
 100          if (sbuf->dma.pending) {
 101             svga_buffer_upload_flush(svga, sbuf);
 102
 103             /*
 104              * Instead of flushing the context command buffer, simply discard
 105              * the current hwbuf, and start a new one.
 106              * With GB objects, the map operation takes care of this
 107              * if passed the PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE flag,
 108              * and the old backing store is busy.
 109              */
 110
 111             if (!svga_have_gb_objects(svga))
 112                svga_buffer_destroy_hw_storage(ss, sbuf);
 113          }
 114
 115          sbuf->map.num_ranges = 0;
 116          sbuf->dma.flags.discard = TRUE;
 117       }
 118
 119       if (usage & PIPE_TRANSFER_UNSYNCHRONIZED) {
 120          if (!sbuf->map.num_ranges) {
 121             /*
 122              * No pending ranges to upload so far, so we can tell the host to
 123              * not synchronize on the next DMA command.
 124              */
 125
 126             sbuf->dma.flags.unsynchronized = TRUE;
 127          }
 128       } else {
 129          /*
 130           * Synchronizing, so flush any pending primitives, finish writing any
 131           * pending DMA command, and ensure the next DMA will be done in order.
 132           */
 133
 134          svga_hwtnl_flush_buffer(svga, resource);
 135
 136          if (sbuf->dma.pending) {
 137             svga_buffer_upload_flush(svga, sbuf);
 138
 139             if (svga_buffer_has_hw_storage(sbuf)) {
 140                /*
 141                 * We have a pending DMA upload from a hardware buffer, therefore
 142                 * we need to ensure that the host finishes processing that DMA
 143                 * command before the state tracker can start overwriting the
 144                 * hardware buffer.
 145                 *
 146                 * XXX: This could be avoided by tying the hardware buffer to
 147                 * the transfer (just as done with textures), which would allow
 148                 * overlapping DMAs commands to be queued on the same context
 149                 * buffer. However, due to the likelihood of software vertex
 150                 * processing, it is more convenient to hold on to the hardware
 151                 * buffer, allowing to quickly access the contents from the CPU
 152                 * without having to do a DMA download from the host.
 153                 */
 154
 155                if (usage & PIPE_TRANSFER_DONTBLOCK) {
 156                   /*
 157                    * Flushing the command buffer here will most likely cause
 158                    * the map of the hwbuf below to block, so preemptively
 159                    * return NULL here if DONTBLOCK is set to prevent unnecessary
 160                    * command buffer flushes.
 161                    */
 162
 163                   FREE(transfer);
 164                   return NULL;
 165                }
 166
 167                svga_context_flush(svga, NULL);
 168             }
 169          }
 170
 171          sbuf->dma.flags.unsynchronized = FALSE;
 172       }
 173    }
 174
 175    if (!sbuf->swbuf && !svga_buffer_has_hw_storage(sbuf)) {
 176       if (svga_buffer_create_hw_storage(ss, sbuf) != PIPE_OK) {
 177          /*
 178           * We can't create a hardware buffer big enough, so create a malloc
 179           * buffer instead.
 180           */
 181          if (0) {
 182             debug_printf("%s: failed to allocate %u KB of DMA, "
 183                          "splitting DMA transfers\n",
 184                          __FUNCTION__,
 185                          (sbuf->b.b.width0 + 1023)/1024);
 186          }
 187
 188          sbuf->swbuf = align_malloc(sbuf->b.b.width0, 16);
 189          if (!sbuf->swbuf) {
 190             FREE(transfer);
 191             return NULL;
 192          }
 193       }
 194    }
 195
 196    if (sbuf->swbuf) {
 197       /* User/malloc buffer */
 198       map = sbuf->swbuf;
 199    }
 200    else if (svga_buffer_has_hw_storage(sbuf)) {
 201       boolean retry;
 202
 203       map = svga_buffer_hw_storage_map(svga, sbuf, transfer->usage, &retry);
 204       if (map == NULL && retry) {
 205          /*
 206           * At this point, svga_buffer_get_transfer() has already
 207           * hit the DISCARD_WHOLE_RESOURCE path and flushed HWTNL
 208           * for this buffer.
 209           */
 210          svga_context_flush(svga, NULL);
 211          map = svga_buffer_hw_storage_map(svga, sbuf, transfer->usage, &retry);
 212       }
 213    }
 214    else {
 215       map = NULL;
 216    }
 217
 218    if (map) {
 219       ++sbuf->map.count;
 220       map += transfer->box.x;
 221       *ptransfer = transfer;
 222    } else {
 223       FREE(transfer);
 224    }
 225
 226    return map;
 227 }
 228
 229
 230 static void
 231 svga_buffer_transfer_flush_region( struct pipe_context *pipe,
 232                                    struct pipe_transfer *transfer,
 233                                    const struct pipe_box *box)
 234 {
 235    struct svga_screen *ss = svga_screen(pipe->screen);
 236    struct svga_buffer *sbuf = svga_buffer(transfer->resource);
 237
 238    unsigned offset = transfer->box.x + box->x;
 239    unsigned length = box->width;
 240
 241    assert(transfer->usage & PIPE_TRANSFER_WRITE);
 242    assert(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT);
 243
 244    pipe_mutex_lock(ss->swc_mutex);
 245    svga_buffer_add_range(sbuf, offset, offset + length);
 246    pipe_mutex_unlock(ss->swc_mutex);
 247 }
 248
 249
 250 static void
 251 svga_buffer_transfer_unmap( struct pipe_context *pipe,
 252                             struct pipe_transfer *transfer )
 253 {
 254    struct svga_screen *ss = svga_screen(pipe->screen);
 255    struct svga_context *svga = svga_context(pipe);
 256    struct svga_buffer *sbuf = svga_buffer(transfer->resource);
 257
 258    pipe_mutex_lock(ss->swc_mutex);
 259
 260    assert(sbuf->map.count);
 261    if (sbuf->map.count) {
 262       --sbuf->map.count;
 263    }
 264
 265    if (svga_buffer_has_hw_storage(sbuf)) {
 266       svga_buffer_hw_storage_unmap(svga, sbuf);
 267    }
 268
 269    if (transfer->usage & PIPE_TRANSFER_WRITE) {
 270       if (!(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT)) {
 271          /*
 272           * Mapped range not flushed explicitly, so flush the whole buffer,
 273           * and tell the host to discard the contents when processing the DMA
 274           * command.
 275           */
 276
 277          SVGA_DBG(DEBUG_DMA, "flushing the whole buffer\n");
 278
 279          sbuf->dma.flags.discard = TRUE;
 280
 281          svga_buffer_add_range(sbuf, 0, sbuf->b.b.width0);
 282       }
 283    }
 284
 285    pipe_mutex_unlock(ss->swc_mutex);
 286    FREE(transfer);
 287 }
 288
 289
 290 static void
 291 svga_buffer_destroy( struct pipe_screen *screen,
 292                      struct pipe_resource *buf )
 293 {
 294    struct svga_screen *ss = svga_screen(screen);
 295    struct svga_buffer *sbuf = svga_buffer( buf );
 296
 297    assert(!p_atomic_read(&buf->reference.count));
 298
 299    assert(!sbuf->dma.pending);
 300
 301    if(sbuf->handle)
 302       svga_buffer_destroy_host_surface(ss, sbuf);
 303
 304    if(sbuf->uploaded.buffer)
 305       pipe_resource_reference(&sbuf->uploaded.buffer, NULL);
 306
 307    if(sbuf->hwbuf)
 308       svga_buffer_destroy_hw_storage(ss, sbuf);
 309
 310    if(sbuf->swbuf && !sbuf->user)
 311       align_free(sbuf->swbuf);
 312
 313    ss->total_resource_bytes -= sbuf->size;
 314
 315    FREE(sbuf);
 316 }
 317
 318
 319 struct u_resource_vtbl svga_buffer_vtbl =
 320 {
 321    u_default_resource_get_handle,      /* get_handle */
 322    svga_buffer_destroy,              /* resource_destroy */
 323    svga_buffer_transfer_map,         /* transfer_map */
 324    svga_buffer_transfer_flush_region,  /* transfer_flush_region */
 325    svga_buffer_transfer_unmap,       /* transfer_unmap */
 326    u_default_transfer_inline_write   /* transfer_inline_write */
 327 };
 328
 329
 330
 331 struct pipe_resource *
 332 svga_buffer_create(struct pipe_screen *screen,
 333                    const struct pipe_resource *template)
 334 {
 335    struct svga_screen *ss = svga_screen(screen);
 336    struct svga_buffer *sbuf;
 337
 338    sbuf = CALLOC_STRUCT(svga_buffer);
 339    if(!sbuf)
 340       goto error1;
 341
 342    sbuf->b.b = *template;
 343    sbuf->b.vtbl = &svga_buffer_vtbl;
 344    pipe_reference_init(&sbuf->b.b.reference, 1);
 345    sbuf->b.b.screen = screen;
 346
 347    if(svga_buffer_needs_hw_storage(template->bind)) {
 348       if(svga_buffer_create_host_surface(ss, sbuf) != PIPE_OK)
 349          goto error2;
 350    }
 351    else {
 352       sbuf->swbuf = align_malloc(template->width0, 64);
 353       if(!sbuf->swbuf)
 354          goto error2;
 355    }
 356
 357    debug_reference(&sbuf->b.b.reference,
 358                    (debug_reference_descriptor)debug_describe_resource, 0);
 359
 360    sbuf->size = util_resource_size(template);
 361    ss->total_resource_bytes += sbuf->size;
 362
 363    return &sbuf->b.b;
 364
 365 error2:
 366    FREE(sbuf);
 367 error1:
 368    return NULL;
 369 }
 370
 371 struct pipe_resource *
 372 svga_user_buffer_create(struct pipe_screen *screen,
 373                         void *ptr,
 374                         unsigned bytes,
 375                         unsigned bind)
 376 {
 377    struct svga_buffer *sbuf;
 378
 379    sbuf = CALLOC_STRUCT(svga_buffer);
 380    if(!sbuf)
 381       goto no_sbuf;
 382
 383    pipe_reference_init(&sbuf->b.b.reference, 1);
 384    sbuf->b.vtbl = &svga_buffer_vtbl;
 385    sbuf->b.b.screen = screen;
 386    sbuf->b.b.format = PIPE_FORMAT_R8_UNORM; /* ?? */
 387    sbuf->b.b.usage = PIPE_USAGE_IMMUTABLE;
 388    sbuf->b.b.bind = bind;
 389    sbuf->b.b.width0 = bytes;
 390    sbuf->b.b.height0 = 1;
 391    sbuf->b.b.depth0 = 1;
 392    sbuf->b.b.array_size = 1;
 393
 394    sbuf->swbuf = ptr;
 395    sbuf->user = TRUE;
 396
 397    debug_reference(&sbuf->b.b.reference,
 398                    (debug_reference_descriptor)debug_describe_resource, 0);
 399
 400    return &sbuf->b.b;
 401
 402 no_sbuf:
 403    return NULL;
 404 }
 405
 406
 407