Merge commit 'origin/master' into gallium-map-range
[mesa.git] / 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 Jose 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) || defined(PIPE_OS_BSD)
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 "util/u_debug.h"
47 #include "pipe/p_thread.h"
48 #include "util/u_memory.h"
49 #include "util/u_double_list.h"
50
51 #include "pb_buffer.h"
52 #include "pb_buffer_fenced.h"
53
54
55
56 /**
57 * Convenience macro (type safe).
58 */
59 #define SUPER(__derived) (&(__derived)->base)
60
61
62 struct fenced_buffer_list
63 {
64 pipe_mutex mutex;
65
66 struct pb_fence_ops *ops;
67
68 size_t numDelayed;
69 struct list_head delayed;
70
71 #ifdef DEBUG
72 size_t numUnfenced;
73 struct list_head unfenced;
74 #endif
75 };
76
77
78 /**
79 * Wrapper around a pipe buffer which adds fencing and reference counting.
80 */
81 struct fenced_buffer
82 {
83 struct pb_buffer base;
84
85 struct pb_buffer *buffer;
86
87 /* FIXME: protect access with mutex */
88
89 /**
90 * A bitmask of PIPE_BUFFER_USAGE_CPU/GPU_READ/WRITE describing the current
91 * buffer usage.
92 */
93 unsigned flags;
94
95 unsigned mapcount;
96 struct pb_validate *vl;
97 unsigned validation_flags;
98 struct pipe_fence_handle *fence;
99
100 struct list_head head;
101 struct fenced_buffer_list *list;
102 };
103
104
105 static INLINE struct fenced_buffer *
106 fenced_buffer(struct pb_buffer *buf)
107 {
108 assert(buf);
109 return (struct fenced_buffer *)buf;
110 }
111
112
113 static INLINE void
114 _fenced_buffer_add(struct fenced_buffer *fenced_buf)
115 {
116 struct fenced_buffer_list *fenced_list = fenced_buf->list;
117
118 assert(fenced_buf->base.base.refcount);
119 assert(fenced_buf->flags & PIPE_BUFFER_USAGE_GPU_READ_WRITE);
120 assert(fenced_buf->fence);
121
122 #ifdef DEBUG
123 LIST_DEL(&fenced_buf->head);
124 assert(fenced_list->numUnfenced);
125 --fenced_list->numUnfenced;
126 #endif
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 struct fenced_buffer_list *fenced_list = fenced_buf->list;
139
140 assert(!fenced_buf->base.base.refcount);
141 assert(!fenced_buf->fence);
142 #ifdef DEBUG
143 assert(fenced_buf->head.prev);
144 assert(fenced_buf->head.next);
145 LIST_DEL(&fenced_buf->head);
146 assert(fenced_list->numUnfenced);
147 --fenced_list->numUnfenced;
148 #else
149 (void)fenced_list;
150 #endif
151 pb_reference(&fenced_buf->buffer, NULL);
152 FREE(fenced_buf);
153 }
154
155
156 static INLINE void
157 _fenced_buffer_remove(struct fenced_buffer_list *fenced_list,
158 struct fenced_buffer *fenced_buf)
159 {
160 struct pb_fence_ops *ops = fenced_list->ops;
161
162 assert(fenced_buf->fence);
163 assert(fenced_buf->list == fenced_list);
164
165 ops->fence_reference(ops, &fenced_buf->fence, NULL);
166 fenced_buf->flags &= ~PIPE_BUFFER_USAGE_GPU_READ_WRITE;
167
168 assert(fenced_buf->head.prev);
169 assert(fenced_buf->head.next);
170
171 LIST_DEL(&fenced_buf->head);
172 assert(fenced_list->numDelayed);
173 --fenced_list->numDelayed;
174
175 #ifdef DEBUG
176 LIST_ADDTAIL(&fenced_buf->head, &fenced_list->unfenced);
177 ++fenced_list->numUnfenced;
178 #endif
179
180 if(!fenced_buf->base.base.refcount)
181 _fenced_buffer_destroy(fenced_buf);
182 }
183
184
185 static INLINE enum pipe_error
186 _fenced_buffer_finish(struct fenced_buffer *fenced_buf)
187 {
188 struct fenced_buffer_list *fenced_list = fenced_buf->list;
189 struct pb_fence_ops *ops = fenced_list->ops;
190
191 #if 0
192 debug_warning("waiting for GPU");
193 #endif
194
195 assert(fenced_buf->fence);
196 if(fenced_buf->fence) {
197 if(ops->fence_finish(ops, fenced_buf->fence, 0) != 0) {
198 return PIPE_ERROR;
199 }
200 /* Remove from the fenced list */
201 /* TODO: remove consequents */
202 _fenced_buffer_remove(fenced_list, fenced_buf);
203 }
204
205 fenced_buf->flags &= ~PIPE_BUFFER_USAGE_GPU_READ_WRITE;
206 return PIPE_OK;
207 }
208
209
210 /**
211 * Free as many fenced buffers from the list head as possible.
212 */
213 static void
214 _fenced_buffer_list_check_free(struct fenced_buffer_list *fenced_list,
215 int wait)
216 {
217 struct pb_fence_ops *ops = fenced_list->ops;
218 struct list_head *curr, *next;
219 struct fenced_buffer *fenced_buf;
220 struct pipe_fence_handle *prev_fence = NULL;
221
222 curr = fenced_list->delayed.next;
223 next = curr->next;
224 while(curr != &fenced_list->delayed) {
225 fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
226
227 if(fenced_buf->fence != prev_fence) {
228 int signaled;
229 if (wait)
230 signaled = ops->fence_finish(ops, fenced_buf->fence, 0);
231 else
232 signaled = ops->fence_signalled(ops, fenced_buf->fence, 0);
233 if (signaled != 0)
234 break;
235 prev_fence = fenced_buf->fence;
236 }
237 else {
238 assert(ops->fence_signalled(ops, fenced_buf->fence, 0) == 0);
239 }
240
241 _fenced_buffer_remove(fenced_list, fenced_buf);
242
243 curr = next;
244 next = curr->next;
245 }
246 }
247
248
249 static void
250 fenced_buffer_destroy(struct pb_buffer *buf)
251 {
252 struct fenced_buffer *fenced_buf = fenced_buffer(buf);
253 struct fenced_buffer_list *fenced_list = fenced_buf->list;
254
255 pipe_mutex_lock(fenced_list->mutex);
256 assert(fenced_buf->base.base.refcount == 0);
257 if (fenced_buf->fence) {
258 struct pb_fence_ops *ops = fenced_list->ops;
259 if(ops->fence_signalled(ops, fenced_buf->fence, 0) == 0) {
260 struct list_head *curr, *prev;
261 curr = &fenced_buf->head;
262 prev = curr->prev;
263 do {
264 fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
265 assert(ops->fence_signalled(ops, fenced_buf->fence, 0) == 0);
266 _fenced_buffer_remove(fenced_list, fenced_buf);
267 curr = prev;
268 prev = curr->prev;
269 } while (curr != &fenced_list->delayed);
270 }
271 else {
272 /* delay destruction */
273 }
274 }
275 else {
276 _fenced_buffer_destroy(fenced_buf);
277 }
278 pipe_mutex_unlock(fenced_list->mutex);
279 }
280
281
282 static void *
283 fenced_buffer_map(struct pb_buffer *buf,
284 unsigned flags)
285 {
286 struct fenced_buffer *fenced_buf = fenced_buffer(buf);
287 void *map;
288
289 assert(flags & PIPE_BUFFER_USAGE_CPU_READ_WRITE);
290 assert(!(flags & ~PIPE_BUFFER_USAGE_CPU_READ_WRITE));
291 flags &= PIPE_BUFFER_USAGE_CPU_READ_WRITE;
292
293 /* Check for GPU read/write access */
294 if(fenced_buf->flags & PIPE_BUFFER_USAGE_GPU_WRITE) {
295 /* Wait for the GPU to finish writing */
296 _fenced_buffer_finish(fenced_buf);
297 }
298
299 #if 0
300 /* Check for CPU write access (read is OK) */
301 if(fenced_buf->flags & PIPE_BUFFER_USAGE_CPU_READ_WRITE) {
302 /* this is legal -- just for debugging */
303 debug_warning("concurrent CPU writes");
304 }
305 #endif
306
307 map = pb_map(fenced_buf->buffer, flags);
308 if(map) {
309 ++fenced_buf->mapcount;
310 fenced_buf->flags |= flags;
311 }
312
313 return map;
314 }
315
316
317 static void
318 fenced_buffer_unmap(struct pb_buffer *buf)
319 {
320 struct fenced_buffer *fenced_buf = fenced_buffer(buf);
321 assert(fenced_buf->mapcount);
322 if(fenced_buf->mapcount) {
323 pb_unmap(fenced_buf->buffer);
324 --fenced_buf->mapcount;
325 if(!fenced_buf->mapcount)
326 fenced_buf->flags &= ~PIPE_BUFFER_USAGE_CPU_READ_WRITE;
327 }
328 }
329
330
331 static enum pipe_error
332 fenced_buffer_validate(struct pb_buffer *buf,
333 struct pb_validate *vl,
334 unsigned flags)
335 {
336 struct fenced_buffer *fenced_buf = fenced_buffer(buf);
337 enum pipe_error ret;
338
339 if(!vl) {
340 /* invalidate */
341 fenced_buf->vl = NULL;
342 fenced_buf->validation_flags = 0;
343 return PIPE_OK;
344 }
345
346 assert(flags & PIPE_BUFFER_USAGE_GPU_READ_WRITE);
347 assert(!(flags & ~PIPE_BUFFER_USAGE_GPU_READ_WRITE));
348 flags &= PIPE_BUFFER_USAGE_GPU_READ_WRITE;
349
350 /* Buffer cannot be validated in two different lists */
351 if(fenced_buf->vl && fenced_buf->vl != vl)
352 return PIPE_ERROR_RETRY;
353
354 /* Do not validate if buffer is still mapped */
355 if(fenced_buf->flags & PIPE_BUFFER_USAGE_CPU_READ_WRITE) {
356 /* TODO: wait for the thread that mapped the buffer to unmap it */
357 return PIPE_ERROR_RETRY;
358 }
359
360 if(fenced_buf->vl == vl &&
361 (fenced_buf->validation_flags & flags) == flags) {
362 /* Nothing to do -- buffer already validated */
363 return PIPE_OK;
364 }
365
366 /* Final sanity checking */
367 assert(!(fenced_buf->flags & PIPE_BUFFER_USAGE_CPU_READ_WRITE));
368 assert(!fenced_buf->mapcount);
369
370 ret = pb_validate(fenced_buf->buffer, vl, flags);
371 if (ret != PIPE_OK)
372 return ret;
373
374 fenced_buf->vl = vl;
375 fenced_buf->validation_flags |= flags;
376
377 return PIPE_OK;
378 }
379
380
381 static void
382 fenced_buffer_fence(struct pb_buffer *buf,
383 struct pipe_fence_handle *fence)
384 {
385 struct fenced_buffer *fenced_buf;
386 struct fenced_buffer_list *fenced_list;
387 struct pb_fence_ops *ops;
388
389 fenced_buf = fenced_buffer(buf);
390 fenced_list = fenced_buf->list;
391 ops = fenced_list->ops;
392
393 if(fence == fenced_buf->fence) {
394 /* Nothing to do */
395 return;
396 }
397
398 assert(fenced_buf->vl);
399 assert(fenced_buf->validation_flags);
400
401 pipe_mutex_lock(fenced_list->mutex);
402 if (fenced_buf->fence)
403 _fenced_buffer_remove(fenced_list, fenced_buf);
404 if (fence) {
405 ops->fence_reference(ops, &fenced_buf->fence, fence);
406 fenced_buf->flags |= fenced_buf->validation_flags;
407 _fenced_buffer_add(fenced_buf);
408 }
409 pipe_mutex_unlock(fenced_list->mutex);
410
411 pb_fence(fenced_buf->buffer, fence);
412
413 fenced_buf->vl = NULL;
414 fenced_buf->validation_flags = 0;
415 }
416
417
418 static void
419 fenced_buffer_get_base_buffer(struct pb_buffer *buf,
420 struct pb_buffer **base_buf,
421 unsigned *offset)
422 {
423 struct fenced_buffer *fenced_buf = fenced_buffer(buf);
424 pb_get_base_buffer(fenced_buf->buffer, base_buf, offset);
425 }
426
427
428 static const struct pb_vtbl
429 fenced_buffer_vtbl = {
430 fenced_buffer_destroy,
431 fenced_buffer_map,
432 fenced_buffer_unmap,
433 fenced_buffer_validate,
434 fenced_buffer_fence,
435 fenced_buffer_get_base_buffer
436 };
437
438
439 struct pb_buffer *
440 fenced_buffer_create(struct fenced_buffer_list *fenced_list,
441 struct pb_buffer *buffer)
442 {
443 struct fenced_buffer *buf;
444
445 if(!buffer)
446 return NULL;
447
448 buf = CALLOC_STRUCT(fenced_buffer);
449 if(!buf) {
450 pb_reference(&buffer, NULL);
451 return NULL;
452 }
453
454 buf->base.base.refcount = 1;
455 buf->base.base.alignment = buffer->base.alignment;
456 buf->base.base.usage = buffer->base.usage;
457 buf->base.base.size = buffer->base.size;
458
459 buf->base.vtbl = &fenced_buffer_vtbl;
460 buf->buffer = buffer;
461 buf->list = fenced_list;
462
463 #ifdef DEBUG
464 pipe_mutex_lock(fenced_list->mutex);
465 LIST_ADDTAIL(&buf->head, &fenced_list->unfenced);
466 ++fenced_list->numUnfenced;
467 pipe_mutex_unlock(fenced_list->mutex);
468 #endif
469
470 return &buf->base;
471 }
472
473
474 struct fenced_buffer_list *
475 fenced_buffer_list_create(struct pb_fence_ops *ops)
476 {
477 struct fenced_buffer_list *fenced_list;
478
479 fenced_list = CALLOC_STRUCT(fenced_buffer_list);
480 if (!fenced_list)
481 return NULL;
482
483 fenced_list->ops = ops;
484
485 LIST_INITHEAD(&fenced_list->delayed);
486 fenced_list->numDelayed = 0;
487
488 #ifdef DEBUG
489 LIST_INITHEAD(&fenced_list->unfenced);
490 fenced_list->numUnfenced = 0;
491 #endif
492
493 pipe_mutex_init(fenced_list->mutex);
494
495 return fenced_list;
496 }
497
498
499 void
500 fenced_buffer_list_check_free(struct fenced_buffer_list *fenced_list,
501 int wait)
502 {
503 pipe_mutex_lock(fenced_list->mutex);
504 _fenced_buffer_list_check_free(fenced_list, wait);
505 pipe_mutex_unlock(fenced_list->mutex);
506 }
507
508
509 #ifdef DEBUG
510 void
511 fenced_buffer_list_dump(struct fenced_buffer_list *fenced_list)
512 {
513 struct pb_fence_ops *ops = fenced_list->ops;
514 struct list_head *curr, *next;
515 struct fenced_buffer *fenced_buf;
516
517 pipe_mutex_lock(fenced_list->mutex);
518
519 debug_printf("%10s %7s %10s %s\n",
520 "buffer", "refcount", "fence", "signalled");
521
522 curr = fenced_list->unfenced.next;
523 next = curr->next;
524 while(curr != &fenced_list->unfenced) {
525 fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
526 assert(!fenced_buf->fence);
527 debug_printf("%10p %7u\n",
528 fenced_buf,
529 fenced_buf->base.base.refcount);
530 curr = next;
531 next = curr->next;
532 }
533
534 curr = fenced_list->delayed.next;
535 next = curr->next;
536 while(curr != &fenced_list->delayed) {
537 int signaled;
538 fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
539 signaled = ops->fence_signalled(ops, fenced_buf->fence, 0);
540 debug_printf("%10p %7u %10p %s\n",
541 fenced_buf,
542 fenced_buf->base.base.refcount,
543 fenced_buf->fence,
544 signaled == 0 ? "y" : "n");
545 curr = next;
546 next = curr->next;
547 }
548
549 pipe_mutex_unlock(fenced_list->mutex);
550 }
551 #endif
552
553
554 void
555 fenced_buffer_list_destroy(struct fenced_buffer_list *fenced_list)
556 {
557 pipe_mutex_lock(fenced_list->mutex);
558
559 /* Wait on outstanding fences */
560 while (fenced_list->numDelayed) {
561 pipe_mutex_unlock(fenced_list->mutex);
562 #if defined(PIPE_OS_LINUX) || defined(PIPE_OS_BSD)
563 sched_yield();
564 #endif
565 _fenced_buffer_list_check_free(fenced_list, 1);
566 pipe_mutex_lock(fenced_list->mutex);
567 }
568
569 #ifdef DEBUG
570 //assert(!fenced_list->numUnfenced);
571 #endif
572
573 pipe_mutex_unlock(fenced_list->mutex);
574
575 fenced_list->ops->destroy(fenced_list->ops);
576
577 FREE(fenced_list);
578 }
579
580