pb_buffer_fenced.c revision b06cd4debfc4fb4162b4b45e61ea91948de0a279 
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#endif
42
43#include "pipe/p_compiler.h"
44#include "pipe/p_error.h"
45#include "pipe/p_debug.h"
46#include "pipe/p_winsys.h"
47#include "pipe/p_thread.h"
48#include "pipe/p_util.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#define PIPE_BUFFER_USAGE_CPU_READ_WRITE \
62   ( PIPE_BUFFER_USAGE_CPU_READ | PIPE_BUFFER_USAGE_CPU_WRITE )
63#define PIPE_BUFFER_USAGE_GPU_READ_WRITE \
64   ( PIPE_BUFFER_USAGE_GPU_READ | PIPE_BUFFER_USAGE_GPU_WRITE )
65#define PIPE_BUFFER_USAGE_WRITE \
66   ( PIPE_BUFFER_USAGE_CPU_WRITE | PIPE_BUFFER_USAGE_GPU_WRITE )
67
68
69struct fenced_buffer_list
70{
71   _glthread_Mutex mutex;
72
73   struct pipe_winsys *winsys;
74
75   size_t numDelayed;
76
77   struct list_head delayed;
78};
79
80
81/**
82 * Wrapper around a pipe buffer which adds fencing and reference counting.
83 */
84struct fenced_buffer
85{
86   struct pb_buffer base;
87
88   struct pb_buffer *buffer;
89
90   /* FIXME: protect access with mutex */
91
92   /**
93    * A bitmask of PIPE_BUFFER_USAGE_CPU/GPU_READ/WRITE describing the current
94    * buffer usage.
95    */
96   unsigned flags;
97
98   unsigned mapcount;
99   struct pipe_fence_handle *fence;
100
101   struct list_head head;
102   struct fenced_buffer_list *list;
103};
104
105
106static INLINE struct fenced_buffer *
107fenced_buffer(struct pb_buffer *buf)
108{
109   assert(buf);
110   assert(buf->vtbl == &fenced_buffer_vtbl);
111   return (struct fenced_buffer *)buf;
112}
113
114
115static INLINE void
116_fenced_buffer_add(struct fenced_buffer *fenced_buf)
117{
118   struct fenced_buffer_list *fenced_list = fenced_buf->list;
119
120   assert(fenced_buf->base.base.refcount);
121   assert(fenced_buf->flags & PIPE_BUFFER_USAGE_GPU_READ_WRITE);
122   assert(fenced_buf->fence);
123
124   assert(!fenced_buf->head.prev);
125   assert(!fenced_buf->head.next);
126   LIST_ADDTAIL(&fenced_buf->head, &fenced_list->delayed);
127   ++fenced_list->numDelayed;
128}
129
130
131/**
132 * Actually destroy the buffer.
133 */
134static INLINE void
135_fenced_buffer_destroy(struct fenced_buffer *fenced_buf)
136{
137   struct fenced_buffer_list *fenced_list = fenced_buf->list;
138
139   assert(!fenced_buf->base.base.refcount);
140   assert(!fenced_buf->fence);
141   pb_reference(&fenced_buf->buffer, NULL);
142   FREE(fenced_buf);
143}
144
145
146static INLINE void
147_fenced_buffer_remove(struct fenced_buffer *fenced_buf)
148{
149   struct fenced_buffer_list *fenced_list = fenced_buf->list;
150   struct pipe_winsys *winsys = fenced_list->winsys;
151
152   assert(fenced_buf->fence);
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/**
174 * Free as many fenced buffers from the list head as possible.
175 */
176static void
177_fenced_buffer_list_check_free(struct fenced_buffer_list *fenced_list,
178                               int wait)
179{
180   struct pipe_winsys *winsys = fenced_list->winsys;
181   struct list_head *curr, *next;
182   struct fenced_buffer *fenced_buf;
183   struct pipe_fence_handle *prev_fence = NULL;
184
185   curr = fenced_list->delayed.next;
186   next = curr->next;
187   while(curr != &fenced_list->delayed) {
188      fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
189
190      if(fenced_buf->fence != prev_fence) {
191	 int signaled;
192	 if (wait)
193	    signaled = winsys->fence_finish(winsys, fenced_buf->fence, 0);
194	 else
195	    signaled = winsys->fence_signalled(winsys, fenced_buf->fence, 0);
196	 if (signaled != 0)
197	    break;
198	 prev_fence = fenced_buf->fence;
199      }
200      else {
201	 assert(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0);
202      }
203
204      _fenced_buffer_remove(fenced_buf);
205
206      curr = next;
207      next = curr->next;
208   }
209}
210
211
212/**
213 * Serialize writes, but allow concurrent reads.
214 */
215static INLINE enum pipe_error
216fenced_buffer_serialize(struct fenced_buffer *fenced_buf, unsigned flags)
217{
218   struct fenced_buffer_list *fenced_list = fenced_buf->list;
219   struct pipe_winsys *winsys = fenced_list->winsys;
220
221   /* Allow concurrent reads */
222   if(((fenced_buf->flags | flags) & PIPE_BUFFER_USAGE_WRITE) == 0)
223      return PIPE_OK;
224
225   /* Wait for the CPU to finish */
226   if(fenced_buf->mapcount) {
227      /* FIXME: Use thread conditions variables to signal when mapcount
228       * reaches zero */
229      debug_warning("attemp to write concurrently to buffer");
230      /* XXX: we must not fail here in order to support texture mipmap generation
231      return PIPE_ERROR_RETRY;
232       */
233   }
234
235   /* Wait for the GPU to finish */
236   if(fenced_buf->fence) {
237      if(winsys->fence_finish(winsys, fenced_buf->fence, 0) != 0)
238	 return PIPE_ERROR_RETRY;
239      _fenced_buffer_remove(fenced_buf);
240   }
241
242   return PIPE_OK;
243}
244
245
246static void
247fenced_buffer_destroy(struct pb_buffer *buf)
248{
249   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
250   struct fenced_buffer_list *fenced_list = fenced_buf->list;
251
252   _glthread_LOCK_MUTEX(fenced_list->mutex);
253   assert(fenced_buf->base.base.refcount == 0);
254   if (fenced_buf->fence) {
255      struct pipe_winsys *winsys = fenced_list->winsys;
256      if(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0) {
257	 struct list_head *curr, *prev;
258	 curr = &fenced_buf->head;
259	 prev = curr->prev;
260	 do {
261	    fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
262	    assert(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0);
263	    _fenced_buffer_remove(fenced_buf);
264	    curr = prev;
265	    prev = curr->prev;
266	 } while (curr != &fenced_list->delayed);
267      }
268      else {
269	 /* delay destruction */
270      }
271   }
272   else {
273      _fenced_buffer_destroy(fenced_buf);
274   }
275   _glthread_UNLOCK_MUTEX(fenced_list->mutex);
276}
277
278
279static void *
280fenced_buffer_map(struct pb_buffer *buf,
281                  unsigned flags)
282{
283   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
284   void *map;
285   assert((flags & ~PIPE_BUFFER_USAGE_CPU_READ_WRITE) == 0);
286
287   if(fenced_buffer_serialize(fenced_buf, flags) != PIPE_OK)
288      return NULL;
289
290   map = pb_map(fenced_buf->buffer, flags);
291   if(map)
292      ++fenced_buf->mapcount;
293   fenced_buf->flags |= flags & PIPE_BUFFER_USAGE_CPU_READ_WRITE;
294   return map;
295}
296
297
298static void
299fenced_buffer_unmap(struct pb_buffer *buf)
300{
301   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
302   assert(fenced_buf->mapcount);
303   pb_unmap(fenced_buf->buffer);
304   --fenced_buf->mapcount;
305   if(!fenced_buf->mapcount)
306      fenced_buf->flags &= ~PIPE_BUFFER_USAGE_CPU_READ_WRITE;
307}
308
309
310static void
311fenced_buffer_get_base_buffer(struct pb_buffer *buf,
312                              struct pb_buffer **base_buf,
313                              unsigned *offset)
314{
315   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
316   pb_get_base_buffer(fenced_buf->buffer, base_buf, offset);
317}
318
319
320const struct pb_vtbl
321fenced_buffer_vtbl = {
322      fenced_buffer_destroy,
323      fenced_buffer_map,
324      fenced_buffer_unmap,
325      fenced_buffer_get_base_buffer
326};
327
328
329struct pb_buffer *
330fenced_buffer_create(struct fenced_buffer_list *fenced_list,
331                     struct pb_buffer *buffer)
332{
333   struct fenced_buffer *buf;
334
335   if(!buffer)
336      return NULL;
337
338   buf = CALLOC_STRUCT(fenced_buffer);
339   if(!buf)
340      return NULL;
341
342   buf->base.base.refcount = 1;
343   buf->base.base.alignment = buffer->base.alignment;
344   buf->base.base.usage = buffer->base.usage;
345   buf->base.base.size = buffer->base.size;
346
347   buf->base.vtbl = &fenced_buffer_vtbl;
348   buf->buffer = buffer;
349   buf->list = fenced_list;
350
351   return &buf->base;
352}
353
354
355void
356buffer_fence(struct pb_buffer *buf,
357             struct pipe_fence_handle *fence)
358{
359   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
360   struct fenced_buffer_list *fenced_list = fenced_buf->list;
361   struct pipe_winsys *winsys = fenced_list->winsys;
362   /* FIXME: receive this as a parameter */
363   unsigned flags = fence ? PIPE_BUFFER_USAGE_GPU_READ_WRITE : 0;
364
365   if(fence == fenced_buf->fence) {
366      /* Handle the same fence case specially, not only because it is a fast
367       * path, but mostly to avoid serializing two writes with the same fence,
368       * as that would bring the hardware down to synchronous operation without
369       * any benefit.
370       */
371      fenced_buf->flags |= flags & PIPE_BUFFER_USAGE_GPU_READ_WRITE;
372      return;
373   }
374
375   if(fenced_buffer_serialize(fenced_buf, flags) != PIPE_OK) {
376      /* FIXME: propagate error */
377      (void)0;
378   }
379
380   _glthread_LOCK_MUTEX(fenced_list->mutex);
381   if (fenced_buf->fence)
382      _fenced_buffer_remove(fenced_buf);
383   if (fence) {
384      winsys->fence_reference(winsys, &fenced_buf->fence, fence);
385      fenced_buf->flags |= flags & PIPE_BUFFER_USAGE_GPU_READ_WRITE;
386      _fenced_buffer_add(fenced_buf);
387   }
388   _glthread_UNLOCK_MUTEX(fenced_list->mutex);
389}
390
391
392struct fenced_buffer_list *
393fenced_buffer_list_create(struct pipe_winsys *winsys)
394{
395   struct fenced_buffer_list *fenced_list;
396
397   fenced_list = (struct fenced_buffer_list *)CALLOC(1, sizeof(*fenced_list));
398   if (!fenced_list)
399      return NULL;
400
401   fenced_list->winsys = winsys;
402
403   LIST_INITHEAD(&fenced_list->delayed);
404
405   fenced_list->numDelayed = 0;
406
407   _glthread_INIT_MUTEX(fenced_list->mutex);
408
409   return fenced_list;
410}
411
412
413void
414fenced_buffer_list_check_free(struct fenced_buffer_list *fenced_list,
415                              int wait)
416{
417   _glthread_LOCK_MUTEX(fenced_list->mutex);
418   _fenced_buffer_list_check_free(fenced_list, wait);
419   _glthread_UNLOCK_MUTEX(fenced_list->mutex);
420}
421
422
423void
424fenced_buffer_list_destroy(struct fenced_buffer_list *fenced_list)
425{
426   _glthread_LOCK_MUTEX(fenced_list->mutex);
427
428   /* Wait on outstanding fences */
429   while (fenced_list->numDelayed) {
430      _glthread_UNLOCK_MUTEX(fenced_list->mutex);
431#if defined(PIPE_OS_LINUX)
432      sched_yield();
433#endif
434      _fenced_buffer_list_check_free(fenced_list, 1);
435      _glthread_LOCK_MUTEX(fenced_list->mutex);
436   }
437
438   _glthread_UNLOCK_MUTEX(fenced_list->mutex);
439
440   FREE(fenced_list);
441}
442
443
444