src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 /**
  29  * \file
  30  * Implementation of fenced buffers.
  31  *
  32  * \author José Fonseca <jrfonseca-at-tungstengraphics-dot-com>
  33  * \author Thomas Hellström <thomas-at-tungstengraphics-dot-com>
  34  */
  35
  36
  37 #include "pipe/p_config.h"
  38
  39 #if defined(PIPE_OS_LINUX)
  40 #include <unistd.h>
  41 #include <sched.h>
  42 #endif
  43
  44 #include "pipe/p_compiler.h"
  45 #include "pipe/p_error.h"
  46 #include "pipe/p_debug.h"
  47 #include "pipe/p_winsys.h"
  48 #include "pipe/p_thread.h"
  49 #include "util/u_memory.h"
  50 #include "util/u_double_list.h"
  51
  52 #include "pb_buffer.h"
  53 #include "pb_buffer_fenced.h"
  54
  55
  56
  57 /**
  58  * Convenience macro (type safe).
  59  */
  60 #define SUPER(__derived) (&(__derived)->base)
  61
  62 #define PIPE_BUFFER_USAGE_CPU_READ_WRITE \
  63    ( PIPE_BUFFER_USAGE_CPU_READ | PIPE_BUFFER_USAGE_CPU_WRITE )
  64 #define PIPE_BUFFER_USAGE_GPU_READ_WRITE \
  65    ( PIPE_BUFFER_USAGE_GPU_READ | PIPE_BUFFER_USAGE_GPU_WRITE )
  66 #define PIPE_BUFFER_USAGE_WRITE \
  67    ( PIPE_BUFFER_USAGE_CPU_WRITE | PIPE_BUFFER_USAGE_GPU_WRITE )
  68
  69
  70 struct fenced_buffer_list
  71 {
  72    pipe_mutex mutex;
  73
  74    struct pipe_winsys *winsys;
  75
  76    size_t numDelayed;
  77
  78    struct list_head delayed;
  79 };
  80
  81
  82 /**
  83  * Wrapper around a pipe buffer which adds fencing and reference counting.
  84  */
  85 struct fenced_buffer
  86 {
  87    struct pb_buffer base;
  88
  89    struct pb_buffer *buffer;
  90
  91    /* FIXME: protect access with mutex */
  92
  93    /**
  94     * A bitmask of PIPE_BUFFER_USAGE_CPU/GPU_READ/WRITE describing the current
  95     * buffer usage.
  96     */
  97    unsigned flags;
  98
  99    unsigned mapcount;
 100    struct pipe_fence_handle *fence;
 101
 102    struct list_head head;
 103    struct fenced_buffer_list *list;
 104 };
 105
 106
 107 static INLINE struct fenced_buffer *
 108 fenced_buffer(struct pb_buffer *buf)
 109 {
 110    assert(buf);
 111    assert(buf->vtbl == &fenced_buffer_vtbl);
 112    return (struct fenced_buffer *)buf;
 113 }
 114
 115
 116 static INLINE void
 117 _fenced_buffer_add(struct fenced_buffer *fenced_buf)
 118 {
 119    struct fenced_buffer_list *fenced_list = fenced_buf->list;
 120
 121    assert(fenced_buf->base.base.refcount);
 122    assert(fenced_buf->flags & PIPE_BUFFER_USAGE_GPU_READ_WRITE);
 123    assert(fenced_buf->fence);
 124
 125    assert(!fenced_buf->head.prev);
 126    assert(!fenced_buf->head.next);
 127    LIST_ADDTAIL(&fenced_buf->head, &fenced_list->delayed);
 128    ++fenced_list->numDelayed;
 129 }
 130
 131
 132 /**
 133  * Actually destroy the buffer.
 134  */
 135 static INLINE void
 136 _fenced_buffer_destroy(struct fenced_buffer *fenced_buf)
 137 {
 138    assert(!fenced_buf->base.base.refcount);
 139    assert(!fenced_buf->fence);
 140    pb_reference(&fenced_buf->buffer, NULL);
 141    FREE(fenced_buf);
 142 }
 143
 144
 145 static INLINE void
 146 _fenced_buffer_remove(struct fenced_buffer_list *fenced_list,
 147                       struct fenced_buffer *fenced_buf)
 148 {
 149    struct pipe_winsys *winsys = fenced_list->winsys;
 150
 151    assert(fenced_buf->fence);
 152    assert(fenced_buf->list == fenced_list);
 153
 154    winsys->fence_reference(winsys, &fenced_buf->fence, NULL);
 155    fenced_buf->flags &= ~PIPE_BUFFER_USAGE_GPU_READ_WRITE;
 156
 157    assert(fenced_buf->head.prev);
 158    assert(fenced_buf->head.next);
 159    LIST_DEL(&fenced_buf->head);
 160 #ifdef DEBUG
 161    fenced_buf->head.prev = NULL;
 162    fenced_buf->head.next = NULL;
 163 #endif
 164
 165    assert(fenced_list->numDelayed);
 166    --fenced_list->numDelayed;
 167
 168    if(!fenced_buf->base.base.refcount)
 169       _fenced_buffer_destroy(fenced_buf);
 170 }
 171
 172
 173 static INLINE enum pipe_error
 174 _fenced_buffer_finish(struct fenced_buffer *fenced_buf)
 175 {
 176    struct fenced_buffer_list *fenced_list = fenced_buf->list;
 177    struct pipe_winsys *winsys = fenced_list->winsys;
 178
 179 #if 0
 180    debug_warning("waiting for GPU");
 181 #endif
 182
 183    assert(fenced_buf->fence);
 184    if(fenced_buf->fence) {
 185       if(winsys->fence_finish(winsys, fenced_buf->fence, 0) != 0) {
 186          return PIPE_ERROR;
 187       }
 188       /* Remove from the fenced list */
 189       /* TODO: remove consequents */
 190       _fenced_buffer_remove(fenced_list, fenced_buf);
 191    }
 192
 193    fenced_buf->flags &= ~PIPE_BUFFER_USAGE_GPU_READ_WRITE;
 194    return PIPE_OK;
 195 }
 196
 197
 198 /**
 199  * Free as many fenced buffers from the list head as possible.
 200  */
 201 static void
 202 _fenced_buffer_list_check_free(struct fenced_buffer_list *fenced_list,
 203                                int wait)
 204 {
 205    struct pipe_winsys *winsys = fenced_list->winsys;
 206    struct list_head *curr, *next;
 207    struct fenced_buffer *fenced_buf;
 208    struct pipe_fence_handle *prev_fence = NULL;
 209
 210    curr = fenced_list->delayed.next;
 211    next = curr->next;
 212    while(curr != &fenced_list->delayed) {
 213       fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
 214
 215       if(fenced_buf->fence != prev_fence) {
 216          int signaled;
 217          if (wait)
 218             signaled = winsys->fence_finish(winsys, fenced_buf->fence, 0);
 219          else
 220             signaled = winsys->fence_signalled(winsys, fenced_buf->fence, 0);
 221          if (signaled != 0)
 222             break;
 223          prev_fence = fenced_buf->fence;
 224       }
 225       else {
 226          assert(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0);
 227       }
 228
 229       _fenced_buffer_remove(fenced_list, fenced_buf);
 230
 231       curr = next;
 232       next = curr->next;
 233    }
 234 }
 235
 236
 237 static void
 238 fenced_buffer_destroy(struct pb_buffer *buf)
 239 {
 240    struct fenced_buffer *fenced_buf = fenced_buffer(buf);
 241    struct fenced_buffer_list *fenced_list = fenced_buf->list;
 242
 243    pipe_mutex_lock(fenced_list->mutex);
 244    assert(fenced_buf->base.base.refcount == 0);
 245    if (fenced_buf->fence) {
 246       struct pipe_winsys *winsys = fenced_list->winsys;
 247       if(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0) {
 248          struct list_head *curr, *prev;
 249          curr = &fenced_buf->head;
 250          prev = curr->prev;
 251          do {
 252             fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
 253             assert(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0);
 254             _fenced_buffer_remove(fenced_list, fenced_buf);
 255             curr = prev;
 256             prev = curr->prev;
 257          } while (curr != &fenced_list->delayed);
 258       }
 259       else {
 260          /* delay destruction */
 261       }
 262    }
 263    else {
 264       _fenced_buffer_destroy(fenced_buf);
 265    }
 266    pipe_mutex_unlock(fenced_list->mutex);
 267 }
 268
 269
 270 static void *
 271 fenced_buffer_map(struct pb_buffer *buf,
 272                   unsigned flags)
 273 {
 274    struct fenced_buffer *fenced_buf = fenced_buffer(buf);
 275    void *map;
 276
 277    assert(!(flags & ~PIPE_BUFFER_USAGE_CPU_READ_WRITE));
 278    flags &= PIPE_BUFFER_USAGE_CPU_READ_WRITE;
 279
 280    /* Check for GPU read/write access */
 281    if(fenced_buf->flags & PIPE_BUFFER_USAGE_GPU_WRITE) {
 282       /* Wait for the GPU to finish writing */
 283       _fenced_buffer_finish(fenced_buf);
 284    }
 285
 286 #if 0
 287    /* Check for CPU write access (read is OK) */
 288    if(fenced_buf->flags & PIPE_BUFFER_USAGE_CPU_READ_WRITE) {
 289       /* this is legal -- just for debugging */
 290       debug_warning("concurrent CPU writes");
 291    }
 292 #endif
 293
 294    map = pb_map(fenced_buf->buffer, flags);
 295    if(map) {
 296       ++fenced_buf->mapcount;
 297       fenced_buf->flags |= flags;
 298    }
 299
 300    return map;
 301 }
 302
 303
 304 static void
 305 fenced_buffer_unmap(struct pb_buffer *buf)
 306 {
 307    struct fenced_buffer *fenced_buf = fenced_buffer(buf);
 308    assert(fenced_buf->mapcount);
 309    if(fenced_buf->mapcount) {
 310       pb_unmap(fenced_buf->buffer);
 311       --fenced_buf->mapcount;
 312       if(!fenced_buf->mapcount)
 313          fenced_buf->flags &= ~PIPE_BUFFER_USAGE_CPU_READ_WRITE;
 314    }
 315 }
 316
 317
 318 static void
 319 fenced_buffer_get_base_buffer(struct pb_buffer *buf,
 320                               struct pb_buffer **base_buf,
 321                               unsigned *offset)
 322 {
 323    struct fenced_buffer *fenced_buf = fenced_buffer(buf);
 324    pb_get_base_buffer(fenced_buf->buffer, base_buf, offset);
 325 }
 326
 327
 328 const struct pb_vtbl
 329 fenced_buffer_vtbl = {
 330       fenced_buffer_destroy,
 331       fenced_buffer_map,
 332       fenced_buffer_unmap,
 333       fenced_buffer_get_base_buffer
 334 };
 335
 336
 337 struct pb_buffer *
 338 fenced_buffer_create(struct fenced_buffer_list *fenced_list,
 339                      struct pb_buffer *buffer)
 340 {
 341    struct fenced_buffer *buf;
 342
 343    if(!buffer)
 344       return NULL;
 345
 346    buf = CALLOC_STRUCT(fenced_buffer);
 347    if(!buf) {
 348       pb_reference(&buffer, NULL);
 349       return NULL;
 350    }
 351
 352    buf->base.base.refcount = 1;
 353    buf->base.base.alignment = buffer->base.alignment;
 354    buf->base.base.usage = buffer->base.usage;
 355    buf->base.base.size = buffer->base.size;
 356
 357    buf->base.vtbl = &fenced_buffer_vtbl;
 358    buf->buffer = buffer;
 359    buf->list = fenced_list;
 360
 361    return &buf->base;
 362 }
 363
 364
 365 void
 366 buffer_fence(struct pb_buffer *buf,
 367              struct pipe_fence_handle *fence)
 368 {
 369    struct fenced_buffer *fenced_buf;
 370    struct fenced_buffer_list *fenced_list;
 371    struct pipe_winsys *winsys;
 372    /* FIXME: receive this as a parameter */
 373    unsigned flags = fence ? PIPE_BUFFER_USAGE_GPU_READ_WRITE : 0;
 374
 375    /* This is a public function, so be extra cautious with the buffer passed,
 376     * as happens frequently to receive null buffers, or pointer to buffers
 377     * other than fenced buffers. */
 378    assert(buf);
 379    if(!buf)
 380       return;
 381    assert(buf->vtbl == &fenced_buffer_vtbl);
 382    if(buf->vtbl != &fenced_buffer_vtbl)
 383       return;
 384
 385    fenced_buf = fenced_buffer(buf);
 386    fenced_list = fenced_buf->list;
 387    winsys = fenced_list->winsys;
 388
 389    if(!fence || fence == fenced_buf->fence) {
 390       /* Handle the same fence case specially, not only because it is a fast
 391        * path, but mostly to avoid serializing two writes with the same fence,
 392        * as that would bring the hardware down to synchronous operation without
 393        * any benefit.
 394        */
 395       fenced_buf->flags |= flags & PIPE_BUFFER_USAGE_GPU_READ_WRITE;
 396       return;
 397    }
 398
 399    pipe_mutex_lock(fenced_list->mutex);
 400    if (fenced_buf->fence)
 401       _fenced_buffer_remove(fenced_list, fenced_buf);
 402    if (fence) {
 403       winsys->fence_reference(winsys, &fenced_buf->fence, fence);
 404       fenced_buf->flags |= flags & PIPE_BUFFER_USAGE_GPU_READ_WRITE;
 405       _fenced_buffer_add(fenced_buf);
 406    }
 407    pipe_mutex_unlock(fenced_list->mutex);
 408 }
 409
 410
 411 struct fenced_buffer_list *
 412 fenced_buffer_list_create(struct pipe_winsys *winsys)
 413 {
 414    struct fenced_buffer_list *fenced_list;
 415
 416    fenced_list = CALLOC_STRUCT(fenced_buffer_list);
 417    if (!fenced_list)
 418       return NULL;
 419
 420    fenced_list->winsys = winsys;
 421
 422    LIST_INITHEAD(&fenced_list->delayed);
 423
 424    fenced_list->numDelayed = 0;
 425
 426    pipe_mutex_init(fenced_list->mutex);
 427
 428    return fenced_list;
 429 }
 430
 431
 432 void
 433 fenced_buffer_list_check_free(struct fenced_buffer_list *fenced_list,
 434                               int wait)
 435 {
 436    pipe_mutex_lock(fenced_list->mutex);
 437    _fenced_buffer_list_check_free(fenced_list, wait);
 438    pipe_mutex_unlock(fenced_list->mutex);
 439 }
 440
 441
 442 void
 443 fenced_buffer_list_destroy(struct fenced_buffer_list *fenced_list)
 444 {
 445    pipe_mutex_lock(fenced_list->mutex);
 446
 447    /* Wait on outstanding fences */
 448    while (fenced_list->numDelayed) {
 449       pipe_mutex_unlock(fenced_list->mutex);
 450 #if defined(PIPE_OS_LINUX)
 451       sched_yield();
 452 #endif
 453       _fenced_buffer_list_check_free(fenced_list, 1);
 454       pipe_mutex_lock(fenced_list->mutex);
 455    }
 456
 457    pipe_mutex_unlock(fenced_list->mutex);
 458
 459    FREE(fenced_list);
 460 }
 461
 462