• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /**************************************************************************
2  *
3  * Copyright 2006-2008 Tungsten Graphics, Inc., Cedar Park, TX., USA
4  * All Rights Reserved.
5  * Copyright 2009 VMware, Inc., Palo Alto, CA., USA
6  * All Rights Reserved.
7  *
8  * Permission is hereby granted, free of charge, to any person obtaining a
9  * copy of this software and associated documentation files (the
10  * "Software"), to deal in the Software without restriction, including
11  * without limitation the rights to use, copy, modify, merge, publish,
12  * distribute, sub license, and/or sell copies of the Software, and to
13  * permit persons to whom the Software is furnished to do so, subject to
14  * the following conditions:
15  *
16  * The above copyright notice and this permission notice (including the
17  * next paragraph) shall be included in all copies or substantial portions
18  * of the Software.
19  *
20  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
21  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
23  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
24  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
25  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
26  * USE OR OTHER DEALINGS IN THE SOFTWARE.
27  *
28  **************************************************************************/
29 /*
30  * Authors: Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
31  */
32 
33 #ifdef HAVE_CONFIG_H
34 #include "config.h"
35 #endif
36 
37 #include <drm/psb_ttm_placement_user.h>
38 #include <stdint.h>
39 #include <sys/time.h>
40 #include <errno.h>
41 #include <unistd.h>
42 #include <assert.h>
43 #include <string.h>
44 #include <sys/mman.h>
45 #include <xf86drm.h>
46 #include "wsbm_pool.h"
47 #include "wsbm_fencemgr.h"
48 #include "wsbm_priv.h"
49 #include "wsbm_manager.h"
50 
51 #define WSBM_SLABPOOL_ALLOC_RETRIES 100
52 #define DRMRESTARTCOMMANDWRITE(_fd, _val, _arg, _ret)			\
53 	do {								\
54 		(_ret) = drmCommandWrite(_fd, _val, &(_arg), sizeof(_arg)); \
55 	} while ((_ret) == -EAGAIN || (_ret) == -ERESTART);		\
56 
57 #define DRMRESTARTCOMMANDWRITEREAD(_fd, _val, _arg, _ret)		\
58 	do {								\
59 		(_ret) = drmCommandWriteRead(_fd, _val, &(_arg), sizeof(_arg)); \
60 	} while ((_ret) == -EAGAIN || (_ret) == -ERESTART);		\
61 
62 
63 #ifdef DEBUG_FENCESIGNALED
64 static int createbuffer = 0;
65 static int fencesignaled = 0;
66 #endif
67 
68 struct _WsbmSlab;
69 
70 struct _WsbmSlabBuffer
71 {
72     struct _WsbmKernelBuf kBuf;
73     struct _WsbmBufStorage storage;
74     struct _WsbmCond event;
75 
76     /*
77      * Remains constant after creation.
78      */
79 
80     int isSlabBuffer;
81     struct _WsbmSlab *parent;
82     uint32_t start;
83     void *virtual;
84     unsigned long requestedSize;
85     uint64_t mapHandle;
86 
87     /*
88      * Protected by struct _WsbmSlabSizeHeader::mutex
89      */
90 
91     struct _WsbmListHead head;
92 
93     /*
94      * Protected by this::mutex
95      */
96 
97     struct _WsbmFenceObject *fence;
98     uint32_t fenceType;
99     struct _WsbmAtomic writers;	       /* (Only upping) */
100     int unFenced;
101 };
102 
103 struct _WsbmSlabPool;
104 struct _WsbmSlabKernelBO
105 {
106 
107     /*
108      * Constant at creation
109      */
110 
111     struct _WsbmKernelBuf kBuf;
112     uint32_t pageAlignment;
113     void *virtual;
114     unsigned long actualSize;
115     uint64_t mapHandle;
116 
117     /*
118      * Protected by struct _WsbmSlabCache::mutex
119      */
120 
121     struct _WsbmSlabPool *slabPool;
122     uint32_t proposedPlacement;
123     struct _WsbmListHead timeoutHead;
124     struct _WsbmListHead head;
125     struct timeval timeFreed;
126 };
127 
128 struct _WsbmSlab
129 {
130     struct _WsbmListHead head;
131     struct _WsbmListHead freeBuffers;
132     uint32_t numBuffers;
133     uint32_t numFree;
134     struct _WsbmSlabBuffer *buffers;
135     struct _WsbmSlabSizeHeader *header;
136     struct _WsbmSlabKernelBO *kbo;
137 };
138 
139 struct _WsbmSlabSizeHeader
140 {
141     /*
142      * Constant at creation.
143      */
144     struct _WsbmSlabPool *slabPool;
145     uint32_t bufSize;
146 
147     /*
148      * Protected by this::mutex
149      */
150 
151     struct _WsbmListHead slabs;
152     struct _WsbmListHead freeSlabs;
153     struct _WsbmListHead delayedBuffers;
154     uint32_t numDelayed;
155     struct _WsbmMutex mutex;
156 };
157 
158 struct _WsbmSlabCache
159 {
160     struct timeval slabTimeout;
161     struct timeval checkInterval;
162     struct timeval nextCheck;
163     struct _WsbmListHead timeoutList;
164     struct _WsbmListHead unCached;
165     struct _WsbmListHead cached;
166     struct _WsbmMutex mutex;
167 };
168 
169 struct _WsbmSlabPool
170 {
171     struct _WsbmBufferPool pool;
172 
173     /*
174      * The data of this structure remains constant after
175      * initialization and thus needs no mutex protection.
176      */
177 
178     unsigned int devOffset;
179     struct _WsbmSlabCache *cache;
180     uint32_t proposedPlacement;
181     uint32_t validMask;
182     uint32_t *bucketSizes;
183     uint32_t numBuckets;
184     uint32_t pageSize;
185     int pageAlignment;
186     int maxSlabSize;
187     int desiredNumBuffers;
188     struct _WsbmSlabSizeHeader *headers;
189 };
190 
191 static inline struct _WsbmSlabPool *
slabPoolFromPool(struct _WsbmBufferPool * pool)192 slabPoolFromPool(struct _WsbmBufferPool *pool)
193 {
194     return containerOf(pool, struct _WsbmSlabPool, pool);
195 }
196 
197 static inline struct _WsbmSlabPool *
slabPoolFromBuf(struct _WsbmSlabBuffer * sBuf)198 slabPoolFromBuf(struct _WsbmSlabBuffer *sBuf)
199 {
200     return slabPoolFromPool(sBuf->storage.pool);
201 }
202 
203 static inline struct _WsbmSlabBuffer *
slabBuffer(struct _WsbmBufStorage * buf)204 slabBuffer(struct _WsbmBufStorage *buf)
205 {
206     return containerOf(buf, struct _WsbmSlabBuffer, storage);
207 }
208 
209 /*
210  * FIXME: Perhaps arrange timeout slabs in size buckets for fast
211  * retreival??
212  */
213 
214 static inline int
wsbmTimeAfterEq(struct timeval * arg1,struct timeval * arg2)215 wsbmTimeAfterEq(struct timeval *arg1, struct timeval *arg2)
216 {
217     return ((arg1->tv_sec > arg2->tv_sec) ||
218 	    ((arg1->tv_sec == arg2->tv_sec) &&
219 	     (arg1->tv_usec > arg2->tv_usec)));
220 }
221 
222 static inline void
wsbmTimeAdd(struct timeval * arg,struct timeval * add)223 wsbmTimeAdd(struct timeval *arg, struct timeval *add)
224 {
225     unsigned int sec;
226 
227     arg->tv_sec += add->tv_sec;
228     arg->tv_usec += add->tv_usec;
229     sec = arg->tv_usec / 1000000;
230     arg->tv_sec += sec;
231     arg->tv_usec -= sec * 1000000;
232 }
233 
234 static void
wsbmFreeKernelBO(struct _WsbmSlabKernelBO * kbo)235 wsbmFreeKernelBO(struct _WsbmSlabKernelBO *kbo)
236 {
237     struct ttm_pl_reference_req arg;
238     struct _WsbmSlabPool *slabPool;
239 
240     if (!kbo)
241 	return;
242 
243     slabPool = kbo->slabPool;
244     arg.handle = kbo->kBuf.handle;
245     (void)munmap(kbo->virtual, kbo->actualSize);
246     (void)drmCommandWrite(slabPool->pool.fd,
247 			  slabPool->devOffset + TTM_PL_UNREF, &arg,
248 			  sizeof(arg));
249     free(kbo);
250 }
251 
252 static void
wsbmFreeTimeoutKBOsLocked(struct _WsbmSlabCache * cache,struct timeval * time)253 wsbmFreeTimeoutKBOsLocked(struct _WsbmSlabCache *cache, struct timeval *time)
254 {
255     struct _WsbmListHead *list, *next;
256     struct _WsbmSlabKernelBO *kbo;
257 
258     if (!wsbmTimeAfterEq(time, &cache->nextCheck))
259 	return;
260 
261     WSBMLISTFOREACHSAFE(list, next, &cache->timeoutList) {
262 	kbo = WSBMLISTENTRY(list, struct _WsbmSlabKernelBO, timeoutHead);
263 
264 	if (!wsbmTimeAfterEq(time, &kbo->timeFreed))
265 	    break;
266 
267 	WSBMLISTDELINIT(&kbo->timeoutHead);
268 	WSBMLISTDELINIT(&kbo->head);
269 	wsbmFreeKernelBO(kbo);
270     }
271 
272     cache->nextCheck = *time;
273     wsbmTimeAdd(&cache->nextCheck, &cache->checkInterval);
274 }
275 
276 /*
277  * Add a _SlabKernelBO to the free slab manager.
278  * This means that it is available for reuse, but if it's not
279  * reused in a while, it will be freed.
280  */
281 
282 static void
wsbmSetKernelBOFree(struct _WsbmSlabCache * cache,struct _WsbmSlabKernelBO * kbo)283 wsbmSetKernelBOFree(struct _WsbmSlabCache *cache,
284 		    struct _WsbmSlabKernelBO *kbo)
285 {
286     struct timeval time;
287     struct timeval timeFreed;
288 
289     gettimeofday(&time, NULL);
290     timeFreed = time;
291     WSBM_MUTEX_LOCK(&cache->mutex);
292     wsbmTimeAdd(&timeFreed, &cache->slabTimeout);
293     kbo->timeFreed = timeFreed;
294 
295     if (kbo->kBuf.placement & TTM_PL_FLAG_CACHED)
296 	WSBMLISTADD(&kbo->head, &cache->cached);
297     else
298 	WSBMLISTADD(&kbo->head, &cache->unCached);
299 
300     WSBMLISTADDTAIL(&kbo->timeoutHead, &cache->timeoutList);
301     wsbmFreeTimeoutKBOsLocked(cache, &time);
302 
303     WSBM_MUTEX_UNLOCK(&cache->mutex);
304 }
305 
306 /*
307  * Get a _SlabKernelBO for us to use as storage for a slab.
308  */
309 
310 static struct _WsbmSlabKernelBO *
wsbmAllocKernelBO(struct _WsbmSlabSizeHeader * header)311 wsbmAllocKernelBO(struct _WsbmSlabSizeHeader *header)
312 {
313     struct _WsbmSlabPool *slabPool = header->slabPool;
314     struct _WsbmSlabCache *cache = slabPool->cache;
315     struct _WsbmListHead *list, *head;
316     uint32_t size = header->bufSize * slabPool->desiredNumBuffers;
317     struct _WsbmSlabKernelBO *kbo;
318     struct _WsbmSlabKernelBO *kboTmp;
319     int ret;
320 
321     /*
322      * FIXME: We should perhaps allow some variation in slabsize in order
323      * to efficiently reuse slabs.
324      */
325 
326     size = (size <= (uint32_t) slabPool->maxSlabSize) ? size : (uint32_t) slabPool->maxSlabSize;
327     if (size < header->bufSize)
328 	size = header->bufSize;
329     size = (size + slabPool->pageSize - 1) & ~(slabPool->pageSize - 1);
330     WSBM_MUTEX_LOCK(&cache->mutex);
331 
332     kbo = NULL;
333 
334   retry:
335     head = (slabPool->proposedPlacement & TTM_PL_FLAG_CACHED) ?
336 	&cache->cached : &cache->unCached;
337 
338     WSBMLISTFOREACH(list, head) {
339 	kboTmp = WSBMLISTENTRY(list, struct _WsbmSlabKernelBO, head);
340 
341 	if ((kboTmp->actualSize == size) &&
342 	    (slabPool->pageAlignment == 0 ||
343 	     (kboTmp->pageAlignment % slabPool->pageAlignment) == 0)) {
344 
345 	    if (!kbo)
346 		kbo = kboTmp;
347 
348 	    if ((kbo->proposedPlacement ^ slabPool->proposedPlacement) == 0)
349 		break;
350 
351 	}
352     }
353 
354     if (kbo) {
355 	WSBMLISTDELINIT(&kbo->head);
356 	WSBMLISTDELINIT(&kbo->timeoutHead);
357     }
358 
359     WSBM_MUTEX_UNLOCK(&cache->mutex);
360 
361     if (kbo) {
362 	uint32_t new_mask =
363 	    kbo->proposedPlacement ^ slabPool->proposedPlacement;
364 
365 	ret = 0;
366 	if (new_mask) {
367 	    union ttm_pl_setstatus_arg arg;
368 	    struct ttm_pl_setstatus_req *req = &arg.req;
369 	    struct ttm_pl_rep *rep = &arg.rep;
370 
371 	    req->handle = kbo->kBuf.handle;
372 	    req->set_placement = slabPool->proposedPlacement & new_mask;
373 	    req->clr_placement = ~slabPool->proposedPlacement & new_mask;
374 	    DRMRESTARTCOMMANDWRITEREAD(slabPool->pool.fd,
375 				       slabPool->devOffset + TTM_PL_SETSTATUS,
376 				       arg, ret);
377 	    if (ret == 0) {
378 		kbo->kBuf.gpuOffset = rep->gpu_offset;
379 		kbo->kBuf.placement = rep->placement;
380 	    }
381 	    kbo->proposedPlacement = slabPool->proposedPlacement;
382 	}
383 
384 	if (ret == 0)
385 	    return kbo;
386 
387 	wsbmFreeKernelBO(kbo);
388 	kbo = NULL;
389 	goto retry;
390     }
391 
392     kbo = calloc(1, sizeof(*kbo));
393     if (!kbo)
394 	return NULL;
395 
396     {
397 	union ttm_pl_create_arg arg;
398 
399 	kbo->slabPool = slabPool;
400 	WSBMINITLISTHEAD(&kbo->head);
401 	WSBMINITLISTHEAD(&kbo->timeoutHead);
402 
403 	arg.req.size = size;
404 	arg.req.placement = slabPool->proposedPlacement;
405 	arg.req.page_alignment = slabPool->pageAlignment;
406 
407 	DRMRESTARTCOMMANDWRITEREAD(slabPool->pool.fd,
408 				   slabPool->devOffset + TTM_PL_CREATE,
409 				   arg, ret);
410 	if (ret)
411 	    goto out_err0;
412 
413 	kbo->kBuf.gpuOffset = arg.rep.gpu_offset;
414 	kbo->kBuf.placement = arg.rep.placement;
415 	kbo->kBuf.handle = arg.rep.handle;
416 
417 	kbo->actualSize = arg.rep.bo_size;
418 	kbo->mapHandle = arg.rep.map_handle;
419 	kbo->proposedPlacement = slabPool->proposedPlacement;
420     }
421 
422     kbo->virtual = mmap(0, kbo->actualSize,
423 			PROT_READ | PROT_WRITE, MAP_SHARED,
424 			slabPool->pool.fd, kbo->mapHandle);
425 
426     if (kbo->virtual == MAP_FAILED) {
427 	ret = -errno;
428 	goto out_err1;
429     }
430 
431     return kbo;
432 
433   out_err1:
434     {
435 	struct ttm_pl_reference_req arg = {.handle = kbo->kBuf.handle };
436 
437 	(void)drmCommandWrite(slabPool->pool.fd,
438 			      slabPool->devOffset + TTM_PL_UNREF,
439 			      &arg, sizeof(arg));
440     }
441   out_err0:
442     free(kbo);
443     return NULL;
444 }
445 
446 static int
wsbmAllocSlab(struct _WsbmSlabSizeHeader * header)447 wsbmAllocSlab(struct _WsbmSlabSizeHeader *header)
448 {
449     struct _WsbmSlab *slab;
450     struct _WsbmSlabBuffer *sBuf;
451     uint32_t numBuffers;
452     uint32_t ret;
453     uint32_t i;
454 
455     slab = calloc(1, sizeof(*slab));
456     if (!slab)
457 	return -ENOMEM;
458 
459     slab->kbo = wsbmAllocKernelBO(header);
460     if (!slab->kbo) {
461 	ret = -ENOMEM;
462 	goto out_err0;
463     }
464 
465     numBuffers = slab->kbo->actualSize / header->bufSize;
466 
467     slab->buffers = calloc(numBuffers, sizeof(*slab->buffers));
468     if (!slab->buffers) {
469 	ret = -ENOMEM;
470 	goto out_err1;
471     }
472 
473     WSBMINITLISTHEAD(&slab->head);
474     WSBMINITLISTHEAD(&slab->freeBuffers);
475     slab->numBuffers = numBuffers;
476     slab->numFree = 0;
477     slab->header = header;
478 
479     sBuf = slab->buffers;
480     for (i = 0; i < numBuffers; ++i) {
481 	ret = wsbmBufStorageInit(&sBuf->storage, &header->slabPool->pool);
482 	if (ret)
483 	    goto out_err2;
484 	sBuf->parent = slab;
485 	sBuf->start = i * header->bufSize;
486 	sBuf->virtual = (void *)((uint8_t *) slab->kbo->virtual +
487 				 sBuf->start);
488 	wsbmAtomicSet(&sBuf->writers, 0);
489 	sBuf->isSlabBuffer = 1;
490 	WSBM_COND_INIT(&sBuf->event);
491 	WSBMLISTADDTAIL(&sBuf->head, &slab->freeBuffers);
492 	slab->numFree++;
493 	sBuf++;
494     }
495 
496     WSBMLISTADDTAIL(&slab->head, &header->slabs);
497 
498     return 0;
499 
500   out_err2:
501     sBuf = slab->buffers;
502     for (i = 0; i < numBuffers; ++i) {
503 	if (sBuf->parent == slab) {
504 	    WSBM_COND_FREE(&sBuf->event);
505 	    wsbmBufStorageTakedown(&sBuf->storage);
506 	}
507 	sBuf++;
508     }
509     free(slab->buffers);
510   out_err1:
511     wsbmSetKernelBOFree(header->slabPool->cache, slab->kbo);
512   out_err0:
513     free(slab);
514     return ret;
515 }
516 
517 /*
518  * Delete a buffer from the slab header delayed list and put
519  * it on the slab free list.
520  */
521 
522 static void
wsbmSlabFreeBufferLocked(struct _WsbmSlabBuffer * buf)523 wsbmSlabFreeBufferLocked(struct _WsbmSlabBuffer *buf)
524 {
525     struct _WsbmSlab *slab = buf->parent;
526     struct _WsbmSlabSizeHeader *header = slab->header;
527     struct _WsbmListHead *list = &buf->head;
528 
529     WSBMLISTDEL(list);
530     WSBMLISTADDTAIL(list, &slab->freeBuffers);
531     slab->numFree++;
532 
533     if (slab->head.next == &slab->head)
534 	WSBMLISTADDTAIL(&slab->head, &header->slabs);
535 
536     if (slab->numFree == slab->numBuffers) {
537 	list = &slab->head;
538 	WSBMLISTDEL(list);
539 	WSBMLISTADDTAIL(list, &header->freeSlabs);
540     }
541 
542     if (header->slabs.next == &header->slabs ||
543 	slab->numFree != slab->numBuffers) {
544 
545 	struct _WsbmListHead *next;
546 	struct _WsbmSlabCache *cache = header->slabPool->cache;
547 
548 	WSBMLISTFOREACHSAFE(list, next, &header->freeSlabs) {
549 	    uint32_t i;
550 	    struct _WsbmSlabBuffer *sBuf;
551 
552 	    slab = WSBMLISTENTRY(list, struct _WsbmSlab, head);
553 
554 	    WSBMLISTDELINIT(list);
555 
556 	    sBuf = slab->buffers;
557 	    for (i = 0; i < slab->numBuffers; ++i) {
558 		if (sBuf->parent == slab) {
559 		    WSBM_COND_FREE(&sBuf->event);
560 		    wsbmBufStorageTakedown(&sBuf->storage);
561 		}
562 		sBuf++;
563 	    }
564 	    wsbmSetKernelBOFree(cache, slab->kbo);
565 	    free(slab->buffers);
566 	    free(slab);
567 	}
568     }
569 }
570 
571 static void
wsbmSlabCheckFreeLocked(struct _WsbmSlabSizeHeader * header,int wait)572 wsbmSlabCheckFreeLocked(struct _WsbmSlabSizeHeader *header, int wait)
573 {
574   struct _WsbmListHead *list, *prev, *first, *head;
575     struct _WsbmSlabBuffer *sBuf;
576     struct _WsbmSlab *slab;
577     int firstWasSignaled = 1;
578     int signaled;
579     uint32_t i;
580     int ret;
581 
582     /*
583      * Rerun the freeing test if the youngest tested buffer
584      * was signaled, since there might be more idle buffers
585      * in the delay list.
586      */
587 
588     while (firstWasSignaled) {
589 	firstWasSignaled = 0;
590 	signaled = 0;
591 	first = header->delayedBuffers.next;
592 
593 	/* Only examine the oldest 1/3 of delayed buffers:
594 	 */
595 	if (header->numDelayed > 3) {
596 	    for (i = 0; i < header->numDelayed; i += 3) {
597 		first = first->next;
598 	    }
599 	}
600 
601 	/*
602 	 * No need to take the buffer mutex for each buffer we loop
603 	 * through since we're currently the only user.
604 	 */
605 
606 	head = first->next;
607 	WSBMLISTFOREACHPREVSAFE(list, prev, head) {
608 
609 	    if (list == &header->delayedBuffers)
610 		break;
611 
612 	    sBuf = WSBMLISTENTRY(list, struct _WsbmSlabBuffer, head);
613 
614 	    slab = sBuf->parent;
615 
616 	    if (!signaled) {
617 		if (wait) {
618 		    ret = wsbmFenceFinish(sBuf->fence, sBuf->fenceType, 0);
619 		    if (ret)
620 			break;
621 		    signaled = 1;
622 		    wait = 0;
623 		} else {
624 		    signaled =
625 			wsbmFenceSignaled(sBuf->fence, sBuf->fenceType);
626 #ifdef DEBUG_FENCESIGNALED
627 		    fencesignaled++;
628 #endif
629 		}
630 		if (signaled) {
631 		    if (list == first)
632 			firstWasSignaled = 1;
633 		    wsbmFenceUnreference(&sBuf->fence);
634 		    header->numDelayed--;
635 		    wsbmSlabFreeBufferLocked(sBuf);
636 		} else
637 		    break;
638 	    } else if (wsbmFenceSignaledCached(sBuf->fence, sBuf->fenceType)) {
639 		wsbmFenceUnreference(&sBuf->fence);
640 		header->numDelayed--;
641 		wsbmSlabFreeBufferLocked(sBuf);
642 	    }
643 	}
644     }
645 }
646 
647 static struct _WsbmSlabBuffer *
wsbmSlabAllocBuffer(struct _WsbmSlabSizeHeader * header)648 wsbmSlabAllocBuffer(struct _WsbmSlabSizeHeader *header)
649 {
650     static struct _WsbmSlabBuffer *buf;
651     struct _WsbmSlab *slab;
652     struct _WsbmListHead *list;
653     int count = WSBM_SLABPOOL_ALLOC_RETRIES;
654 
655     WSBM_MUTEX_LOCK(&header->mutex);
656     while (header->slabs.next == &header->slabs && count > 0) {
657 	wsbmSlabCheckFreeLocked(header, 0);
658 	if (header->slabs.next != &header->slabs)
659 	    break;
660 
661 	WSBM_MUTEX_UNLOCK(&header->mutex);
662 	if (count != WSBM_SLABPOOL_ALLOC_RETRIES)
663 	    usleep(1000);
664 	WSBM_MUTEX_LOCK(&header->mutex);
665 	(void)wsbmAllocSlab(header);
666 	count--;
667     }
668 
669     list = header->slabs.next;
670     if (list == &header->slabs) {
671 	WSBM_MUTEX_UNLOCK(&header->mutex);
672 	return NULL;
673     }
674     slab = WSBMLISTENTRY(list, struct _WsbmSlab, head);
675     if (--slab->numFree == 0)
676 	WSBMLISTDELINIT(list);
677 
678     list = slab->freeBuffers.next;
679     WSBMLISTDELINIT(list);
680 
681     WSBM_MUTEX_UNLOCK(&header->mutex);
682     buf = WSBMLISTENTRY(list, struct _WsbmSlabBuffer, head);
683 
684     buf->storage.destroyContainer = NULL;
685 
686 #ifdef DEBUG_FENCESIGNALED
687     createbuffer++;
688 #endif
689     return buf;
690 }
691 
692 static struct _WsbmBufStorage *
pool_create(struct _WsbmBufferPool * pool,unsigned long size,uint32_t placement,unsigned alignment)693 pool_create(struct _WsbmBufferPool *pool, unsigned long size,
694 	    uint32_t placement, unsigned alignment)
695 {
696     struct _WsbmSlabPool *slabPool = slabPoolFromPool(pool);
697     struct _WsbmSlabSizeHeader *header;
698     struct _WsbmSlabBuffer *sBuf;
699     uint32_t i;
700     int ret;
701 
702     /*
703      * FIXME: Check for compatibility.
704      */
705 
706     header = slabPool->headers;
707     for (i = 0; i < slabPool->numBuckets; ++i) {
708 	if (header->bufSize >= size)
709 	    break;
710 	header++;
711     }
712 
713     if (i < slabPool->numBuckets) {
714 	sBuf = wsbmSlabAllocBuffer(header);
715 	return ((sBuf) ? &sBuf->storage : NULL);
716     }
717 
718     /*
719      * Fall back to allocate a buffer object directly from DRM.
720      * and wrap it in a wsbmBO structure.
721      */
722 
723     sBuf = calloc(1, sizeof(*sBuf));
724 
725     if (!sBuf)
726 	return NULL;
727 
728     if (alignment) {
729 	if ((alignment < slabPool->pageSize)
730 	    && (slabPool->pageSize % alignment))
731 	    goto out_err0;
732 	if ((alignment > slabPool->pageSize)
733 	    && (alignment % slabPool->pageSize))
734 	    goto out_err0;
735     }
736 
737     ret = wsbmBufStorageInit(&sBuf->storage, pool);
738     if (ret)
739 	goto out_err0;
740 
741     ret = WSBM_COND_INIT(&sBuf->event);
742     if (ret)
743 	goto out_err1;
744 
745     {
746 	union ttm_pl_create_arg arg;
747 
748 	arg.req.size = size;
749 	arg.req.placement = placement;
750 	arg.req.page_alignment = alignment / slabPool->pageSize;
751 
752 	DRMRESTARTCOMMANDWRITEREAD(pool->fd,
753 				   slabPool->devOffset + TTM_PL_CREATE,
754 				   arg, ret);
755 
756 	if (ret)
757 	    goto out_err2;
758 
759 	sBuf->kBuf.gpuOffset = arg.rep.gpu_offset;
760 	sBuf->kBuf.placement = arg.rep.placement;
761 	sBuf->kBuf.handle = arg.rep.handle;
762 	sBuf->mapHandle = arg.rep.map_handle;
763 	sBuf->requestedSize = size;
764 
765 	sBuf->virtual = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED,
766 			     pool->fd, sBuf->mapHandle);
767 
768 	if (sBuf->virtual == MAP_FAILED)
769 	    goto out_err3;
770     }
771 
772     wsbmAtomicSet(&sBuf->writers, 0);
773     return &sBuf->storage;
774   out_err3:
775     {
776 	struct ttm_pl_reference_req arg;
777 
778 	arg.handle = sBuf->kBuf.handle;
779 	(void)drmCommandWriteRead(pool->fd,
780 				  slabPool->devOffset + TTM_PL_UNREF,
781 				  &arg, sizeof(arg));
782     }
783   out_err2:
784     WSBM_COND_FREE(&sBuf->event);
785   out_err1:
786     wsbmBufStorageTakedown(&sBuf->storage);
787   out_err0:
788     free(sBuf);
789     return NULL;
790 }
791 
792 static void
pool_destroy(struct _WsbmBufStorage ** p_buf)793 pool_destroy(struct _WsbmBufStorage **p_buf)
794 {
795     struct _WsbmBufStorage *buf = *p_buf;
796     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
797     struct _WsbmSlab *slab;
798     struct _WsbmSlabSizeHeader *header;
799 
800     *p_buf = NULL;
801 
802     if (!sBuf->isSlabBuffer) {
803 	struct _WsbmSlabPool *slabPool = slabPoolFromBuf(sBuf);
804 	struct ttm_pl_reference_req arg;
805 
806 	if (sBuf->virtual != NULL) {
807 	    (void)munmap(sBuf->virtual, sBuf->requestedSize);
808 	    sBuf->virtual = NULL;
809 	}
810 
811 	arg.handle = sBuf->kBuf.handle;
812 	(void)drmCommandWrite(slabPool->pool.fd,
813 			      slabPool->devOffset + TTM_PL_UNREF,
814 			      &arg, sizeof(arg));
815 
816 	WSBM_COND_FREE(&sBuf->event);
817 	wsbmBufStorageTakedown(&sBuf->storage);
818 	free(sBuf);
819 	return;
820     }
821 
822     slab = sBuf->parent;
823     header = slab->header;
824 
825     /*
826      * No need to take the buffer mutex below since we're the only user.
827      */
828 
829     WSBM_MUTEX_LOCK(&header->mutex);
830     sBuf->unFenced = 0;
831     wsbmAtomicSet(&sBuf->writers, 0);
832     wsbmAtomicSet(&sBuf->storage.refCount, 1);
833 
834     if (sBuf->fence && !wsbmFenceSignaledCached(sBuf->fence, sBuf->fenceType)) {
835 	WSBMLISTADDTAIL(&sBuf->head, &header->delayedBuffers);
836 	header->numDelayed++;
837     } else {
838 	if (sBuf->fence)
839 	    wsbmFenceUnreference(&sBuf->fence);
840 	wsbmSlabFreeBufferLocked(sBuf);
841     }
842 
843     WSBM_MUTEX_UNLOCK(&header->mutex);
844 }
845 
846 static void
waitIdleLocked(struct _WsbmSlabBuffer * sBuf,int lazy)847 waitIdleLocked(struct _WsbmSlabBuffer *sBuf, int lazy)
848 {
849     struct _WsbmBufStorage *storage = &sBuf->storage;
850 
851     while (sBuf->unFenced || sBuf->fence != NULL) {
852 
853 	if (sBuf->unFenced)
854 	    WSBM_COND_WAIT(&sBuf->event, &storage->mutex);
855 
856 	if (sBuf->fence != NULL) {
857 	    if (!wsbmFenceSignaled(sBuf->fence, sBuf->fenceType)) {
858 		struct _WsbmFenceObject *fence =
859 		    wsbmFenceReference(sBuf->fence);
860 
861 		WSBM_MUTEX_UNLOCK(&storage->mutex);
862 		(void)wsbmFenceFinish(fence, sBuf->fenceType, lazy);
863 		WSBM_MUTEX_LOCK(&storage->mutex);
864 		if (sBuf->fence == fence)
865 		    wsbmFenceUnreference(&sBuf->fence);
866 
867 		wsbmFenceUnreference(&fence);
868 	    } else {
869 		wsbmFenceUnreference(&sBuf->fence);
870 	    }
871 	}
872     }
873 }
874 
875 static int
pool_waitIdle(struct _WsbmBufStorage * buf,int lazy)876 pool_waitIdle(struct _WsbmBufStorage *buf, int lazy)
877 {
878     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
879 
880     WSBM_MUTEX_LOCK(&buf->mutex);
881     waitIdleLocked(sBuf, lazy);
882     WSBM_MUTEX_UNLOCK(&buf->mutex);
883 
884     return 0;
885 }
886 
887 static int
pool_map(struct _WsbmBufStorage * buf,unsigned mode,void ** virtual)888 pool_map(struct _WsbmBufStorage *buf, unsigned mode __attribute__ ((unused)), void **virtual)
889 {
890     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
891 
892     *virtual = sBuf->virtual;
893 
894     return 0;
895 }
896 
897 static void
pool_releaseFromCpu(struct _WsbmBufStorage * buf,unsigned mode)898 pool_releaseFromCpu(struct _WsbmBufStorage *buf, unsigned mode __attribute__ ((unused)))
899 {
900     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
901 
902     if (wsbmAtomicDecZero(&sBuf->writers))
903 	WSBM_COND_BROADCAST(&sBuf->event);
904 }
905 
906 static int
pool_syncForCpu(struct _WsbmBufStorage * buf,unsigned mode)907 pool_syncForCpu(struct _WsbmBufStorage *buf, unsigned mode)
908 {
909     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
910     int ret = 0;
911 
912     WSBM_MUTEX_LOCK(&buf->mutex);
913     if ((mode & WSBM_SYNCCPU_DONT_BLOCK)) {
914 	int signaled;
915 
916 	if (sBuf->unFenced) {
917 	    ret = -EBUSY;
918 	    goto out_unlock;
919 	}
920 
921 	if (sBuf->isSlabBuffer)
922 	    signaled = (sBuf->fence == NULL) ||
923 		wsbmFenceSignaledCached(sBuf->fence, sBuf->fenceType);
924 	else
925 	    signaled = (sBuf->fence == NULL) ||
926 		wsbmFenceSignaled(sBuf->fence, sBuf->fenceType);
927 
928 	ret = 0;
929 	if (signaled) {
930 	    wsbmFenceUnreference(&sBuf->fence);
931 	    wsbmAtomicInc(&sBuf->writers);
932 	} else
933 	    ret = -EBUSY;
934 	goto out_unlock;
935     }
936     waitIdleLocked(sBuf, 0);
937     wsbmAtomicInc(&sBuf->writers);
938   out_unlock:
939     WSBM_MUTEX_UNLOCK(&buf->mutex);
940     return ret;
941 }
942 
943 static void
pool_unmap(struct _WsbmBufStorage * buf)944 pool_unmap(struct _WsbmBufStorage *buf __attribute__ ((unused)))
945 {
946     ;
947 }
948 
949 static unsigned long
pool_poolOffset(struct _WsbmBufStorage * buf)950 pool_poolOffset(struct _WsbmBufStorage *buf)
951 {
952     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
953 
954     return sBuf->start;
955 }
956 
957 static unsigned long
pool_size(struct _WsbmBufStorage * buf)958 pool_size(struct _WsbmBufStorage *buf)
959 {
960     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
961 
962     if (!sBuf->isSlabBuffer)
963 	return sBuf->requestedSize;
964 
965     return sBuf->parent->header->bufSize;
966 }
967 
968 static struct _WsbmKernelBuf *
pool_kernel(struct _WsbmBufStorage * buf)969 pool_kernel(struct _WsbmBufStorage *buf)
970 {
971     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
972 
973     return (sBuf->isSlabBuffer) ? &sBuf->parent->kbo->kBuf : &sBuf->kBuf;
974 }
975 
976 static unsigned long
pool_offset(struct _WsbmBufStorage * buf)977 pool_offset(struct _WsbmBufStorage *buf)
978 {
979     return pool_kernel(buf)->gpuOffset + pool_poolOffset(buf);
980 }
981 
982 static void
pool_fence(struct _WsbmBufStorage * buf,struct _WsbmFenceObject * fence)983 pool_fence(struct _WsbmBufStorage *buf, struct _WsbmFenceObject *fence)
984 {
985     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
986     struct _WsbmKernelBuf *kBuf;
987 
988     WSBM_MUTEX_LOCK(&buf->mutex);
989     if (sBuf->fence)
990 	wsbmFenceUnreference(&sBuf->fence);
991 
992     kBuf = pool_kernel(buf);
993     sBuf->fenceType = kBuf->fence_type_mask;
994     if (!wsbmFenceSignaledCached(fence, sBuf->fenceType))
995 	sBuf->fence = wsbmFenceReference(fence);
996 
997     sBuf->unFenced = 0;
998     WSBM_COND_BROADCAST(&sBuf->event);
999     WSBM_MUTEX_UNLOCK(&buf->mutex);
1000 }
1001 
1002 static int
pool_validate(struct _WsbmBufStorage * buf,uint64_t set_flags,uint64_t clr_flags)1003 pool_validate(struct _WsbmBufStorage *buf,
1004         uint64_t set_flags __attribute__ ((unused)), uint64_t clr_flags __attribute__ ((unused)))
1005 {
1006     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
1007 
1008     WSBM_MUTEX_LOCK(&buf->mutex);
1009     while (wsbmAtomicRead(&sBuf->writers) != 0) {
1010 	WSBM_COND_WAIT(&sBuf->event, &buf->mutex);
1011     }
1012 
1013     sBuf->unFenced = 1;
1014     WSBM_MUTEX_UNLOCK(&buf->mutex);
1015     return 0;
1016 }
1017 
1018 static void
pool_unvalidate(struct _WsbmBufStorage * buf)1019 pool_unvalidate(struct _WsbmBufStorage *buf)
1020 {
1021     struct _WsbmSlabBuffer *sBuf = slabBuffer(buf);
1022 
1023     WSBM_MUTEX_LOCK(&buf->mutex);
1024     if (sBuf->unFenced) {
1025 	sBuf->unFenced = 0;
1026 	WSBM_COND_BROADCAST(&sBuf->event);
1027     }
1028     WSBM_MUTEX_UNLOCK(&buf->mutex);
1029 }
1030 
1031 struct _WsbmSlabCache *
wsbmSlabCacheInit(uint32_t checkIntervalMsec,uint32_t slabTimeoutMsec)1032 wsbmSlabCacheInit(uint32_t checkIntervalMsec, uint32_t slabTimeoutMsec)
1033 {
1034     struct _WsbmSlabCache *tmp;
1035 
1036     tmp = calloc(1, sizeof(*tmp));
1037     if (!tmp)
1038 	return NULL;
1039 
1040     WSBM_MUTEX_INIT(&tmp->mutex);
1041     WSBM_MUTEX_LOCK(&tmp->mutex);
1042     tmp->slabTimeout.tv_usec = slabTimeoutMsec * 1000;
1043     tmp->slabTimeout.tv_sec = tmp->slabTimeout.tv_usec / 1000000;
1044     tmp->slabTimeout.tv_usec -= tmp->slabTimeout.tv_sec * 1000000;
1045 
1046     tmp->checkInterval.tv_usec = checkIntervalMsec * 1000;
1047     tmp->checkInterval.tv_sec = tmp->checkInterval.tv_usec / 1000000;
1048     tmp->checkInterval.tv_usec -= tmp->checkInterval.tv_sec * 1000000;
1049 
1050     gettimeofday(&tmp->nextCheck, NULL);
1051     wsbmTimeAdd(&tmp->nextCheck, &tmp->checkInterval);
1052     WSBMINITLISTHEAD(&tmp->timeoutList);
1053     WSBMINITLISTHEAD(&tmp->unCached);
1054     WSBMINITLISTHEAD(&tmp->cached);
1055     WSBM_MUTEX_UNLOCK(&tmp->mutex);
1056 
1057     return tmp;
1058 }
1059 
1060 void
wsbmSlabCacheFinish(struct _WsbmSlabCache * cache)1061 wsbmSlabCacheFinish(struct _WsbmSlabCache *cache)
1062 {
1063     struct timeval time;
1064 
1065     time = cache->nextCheck;
1066     WSBM_MUTEX_LOCK(&cache->mutex);
1067     wsbmTimeAdd(&time, &cache->checkInterval);
1068     wsbmFreeTimeoutKBOsLocked(cache, &time);
1069     WSBM_MUTEX_UNLOCK(&cache->mutex);
1070 
1071     assert(cache->timeoutList.next == &cache->timeoutList);
1072     assert(cache->unCached.next == &cache->unCached);
1073     assert(cache->cached.next == &cache->cached);
1074 
1075     WSBM_MUTEX_FREE(&cache->mutex);
1076     free(cache);
1077 }
1078 
1079 static void
wsbmInitSizeHeader(struct _WsbmSlabPool * slabPool,uint32_t size,struct _WsbmSlabSizeHeader * header)1080 wsbmInitSizeHeader(struct _WsbmSlabPool *slabPool, uint32_t size,
1081 		   struct _WsbmSlabSizeHeader *header)
1082 {
1083     WSBM_MUTEX_INIT(&header->mutex);
1084     WSBM_MUTEX_LOCK(&header->mutex);
1085 
1086     WSBMINITLISTHEAD(&header->slabs);
1087     WSBMINITLISTHEAD(&header->freeSlabs);
1088     WSBMINITLISTHEAD(&header->delayedBuffers);
1089 
1090     header->numDelayed = 0;
1091     header->slabPool = slabPool;
1092     header->bufSize = size;
1093 
1094     WSBM_MUTEX_UNLOCK(&header->mutex);
1095 }
1096 
1097 static void
wsbmFinishSizeHeader(struct _WsbmSlabSizeHeader * header)1098 wsbmFinishSizeHeader(struct _WsbmSlabSizeHeader *header)
1099 {
1100     struct _WsbmListHead *list, *next;
1101     struct _WsbmSlabBuffer *sBuf;
1102 
1103     WSBM_MUTEX_LOCK(&header->mutex);
1104     WSBMLISTFOREACHSAFE(list, next, &header->delayedBuffers) {
1105 	sBuf = WSBMLISTENTRY(list, struct _WsbmSlabBuffer, head);
1106 
1107 	if (sBuf->fence) {
1108 	    (void)wsbmFenceFinish(sBuf->fence, sBuf->fenceType, 0);
1109 	    wsbmFenceUnreference(&sBuf->fence);
1110 	}
1111 	header->numDelayed--;
1112 	wsbmSlabFreeBufferLocked(sBuf);
1113     }
1114     WSBM_MUTEX_UNLOCK(&header->mutex);
1115     WSBM_MUTEX_FREE(&header->mutex);
1116 }
1117 
1118 static void
pool_takedown(struct _WsbmBufferPool * pool)1119 pool_takedown(struct _WsbmBufferPool *pool)
1120 {
1121     struct _WsbmSlabPool *slabPool = slabPoolFromPool(pool);
1122     unsigned int i;
1123 
1124     for (i = 0; i < slabPool->numBuckets; ++i) {
1125 	wsbmFinishSizeHeader(&slabPool->headers[i]);
1126     }
1127 
1128     free(slabPool->headers);
1129     free(slabPool->bucketSizes);
1130     free(slabPool);
1131 }
1132 
1133 struct _WsbmBufferPool *
wsbmSlabPoolInit(int fd,uint32_t devOffset,uint32_t placement,uint32_t validMask,uint32_t smallestSize,uint32_t numSizes,uint32_t desiredNumBuffers,uint32_t maxSlabSize,uint32_t pageAlignment,struct _WsbmSlabCache * cache)1134 wsbmSlabPoolInit(int fd,
1135 		 uint32_t devOffset,
1136 		 uint32_t placement,
1137 		 uint32_t validMask,
1138 		 uint32_t smallestSize,
1139 		 uint32_t numSizes,
1140 		 uint32_t desiredNumBuffers,
1141 		 uint32_t maxSlabSize,
1142 		 uint32_t pageAlignment, struct _WsbmSlabCache *cache)
1143 {
1144     struct _WsbmBufferPool *pool;
1145     struct _WsbmSlabPool *slabPool;
1146     uint32_t i;
1147 
1148     slabPool = calloc(1, sizeof(*slabPool));
1149     if (!slabPool)
1150 	return NULL;
1151 
1152     pool = &slabPool->pool;
1153 
1154     slabPool->bucketSizes = calloc(numSizes, sizeof(*slabPool->bucketSizes));
1155     if (!slabPool->bucketSizes)
1156 	goto out_err0;
1157 
1158     slabPool->headers = calloc(numSizes, sizeof(*slabPool->headers));
1159     if (!slabPool->headers)
1160 	goto out_err1;
1161 
1162     slabPool->devOffset = devOffset;
1163     slabPool->cache = cache;
1164     slabPool->proposedPlacement = placement;
1165     slabPool->validMask = validMask;
1166     slabPool->numBuckets = numSizes;
1167     slabPool->pageSize = getpagesize();
1168     slabPool->pageAlignment = pageAlignment;
1169     slabPool->maxSlabSize = maxSlabSize;
1170     slabPool->desiredNumBuffers = desiredNumBuffers;
1171 
1172     for (i = 0; i < slabPool->numBuckets; ++i) {
1173 	slabPool->bucketSizes[i] = (smallestSize << i);
1174 	wsbmInitSizeHeader(slabPool, slabPool->bucketSizes[i],
1175 			   &slabPool->headers[i]);
1176     }
1177 
1178     pool->fd = fd;
1179     pool->map = &pool_map;
1180     pool->unmap = &pool_unmap;
1181     pool->destroy = &pool_destroy;
1182     pool->offset = &pool_offset;
1183     pool->poolOffset = &pool_poolOffset;
1184     pool->size = &pool_size;
1185     pool->create = &pool_create;
1186     pool->fence = &pool_fence;
1187     pool->kernel = &pool_kernel;
1188     pool->validate = &pool_validate;
1189     pool->unvalidate = &pool_unvalidate;
1190     pool->waitIdle = &pool_waitIdle;
1191     pool->takeDown = &pool_takedown;
1192     pool->releasefromcpu = &pool_releaseFromCpu;
1193     pool->syncforcpu = &pool_syncForCpu;
1194 
1195     return pool;
1196 
1197   out_err1:
1198     free(slabPool->bucketSizes);
1199   out_err0:
1200     free(slabPool);
1201 
1202     return NULL;
1203 }
1204