xref: /freebsd/sys/dev/drm2/ttm/ttm_page_alloc.c (revision 0957b409)
1 /*
2  * Copyright (c) Red Hat Inc.
3 
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sub license,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the
12  * next paragraph) shall be included in all copies or substantial portions
13  * of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21  * DEALINGS IN THE SOFTWARE.
22  *
23  * Authors: Dave Airlie <airlied@redhat.com>
24  *          Jerome Glisse <jglisse@redhat.com>
25  *          Pauli Nieminen <suokkos@gmail.com>
26  */
27 /*
28  * Copyright (c) 2013 The FreeBSD Foundation
29  * All rights reserved.
30  *
31  * Portions of this software were developed by Konstantin Belousov
32  * <kib@FreeBSD.org> under sponsorship from the FreeBSD Foundation.
33  */
34 
35 /* simple list based uncached page pool
36  * - Pool collects resently freed pages for reuse
37  * - Use page->lru to keep a free list
38  * - doesn't track currently in use pages
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include <dev/drm2/drmP.h>
45 #include <dev/drm2/ttm/ttm_bo_driver.h>
46 #include <dev/drm2/ttm/ttm_page_alloc.h>
47 #include <vm/vm_pageout.h>
48 
49 #define NUM_PAGES_TO_ALLOC		(PAGE_SIZE/sizeof(vm_page_t))
50 #define SMALL_ALLOCATION		16
51 #define FREE_ALL_PAGES			(~0U)
52 /* times are in msecs */
53 #define PAGE_FREE_INTERVAL		1000
54 
55 /**
56  * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
57  *
58  * @lock: Protects the shared pool from concurrnet access. Must be used with
59  * irqsave/irqrestore variants because pool allocator maybe called from
60  * delayed work.
61  * @fill_lock: Prevent concurrent calls to fill.
62  * @list: Pool of free uc/wc pages for fast reuse.
63  * @gfp_flags: Flags to pass for alloc_page.
64  * @npages: Number of pages in pool.
65  */
66 struct ttm_page_pool {
67 	struct mtx		lock;
68 	bool			fill_lock;
69 	bool			dma32;
70 	struct pglist		list;
71 	int			ttm_page_alloc_flags;
72 	unsigned		npages;
73 	char			*name;
74 	unsigned long		nfrees;
75 	unsigned long		nrefills;
76 };
77 
78 /**
79  * Limits for the pool. They are handled without locks because only place where
80  * they may change is in sysfs store. They won't have immediate effect anyway
81  * so forcing serialization to access them is pointless.
82  */
83 
84 struct ttm_pool_opts {
85 	unsigned	alloc_size;
86 	unsigned	max_size;
87 	unsigned	small;
88 };
89 
90 #define NUM_POOLS 4
91 
92 /**
93  * struct ttm_pool_manager - Holds memory pools for fst allocation
94  *
95  * Manager is read only object for pool code so it doesn't need locking.
96  *
97  * @free_interval: minimum number of jiffies between freeing pages from pool.
98  * @page_alloc_inited: reference counting for pool allocation.
99  * @work: Work that is used to shrink the pool. Work is only run when there is
100  * some pages to free.
101  * @small_allocation: Limit in number of pages what is small allocation.
102  *
103  * @pools: All pool objects in use.
104  **/
105 struct ttm_pool_manager {
106 	unsigned int kobj_ref;
107 	eventhandler_tag lowmem_handler;
108 	struct ttm_pool_opts	options;
109 
110 	union {
111 		struct ttm_page_pool	u_pools[NUM_POOLS];
112 		struct _utag {
113 			struct ttm_page_pool	u_wc_pool;
114 			struct ttm_page_pool	u_uc_pool;
115 			struct ttm_page_pool	u_wc_pool_dma32;
116 			struct ttm_page_pool	u_uc_pool_dma32;
117 		} _ut;
118 	} _u;
119 };
120 
121 #define	pools _u.u_pools
122 #define	wc_pool _u._ut.u_wc_pool
123 #define	uc_pool _u._ut.u_uc_pool
124 #define	wc_pool_dma32 _u._ut.u_wc_pool_dma32
125 #define	uc_pool_dma32 _u._ut.u_uc_pool_dma32
126 
127 MALLOC_DEFINE(M_TTM_POOLMGR, "ttm_poolmgr", "TTM Pool Manager");
128 
129 static void
130 ttm_vm_page_free(vm_page_t m)
131 {
132 
133 	KASSERT(m->object == NULL, ("ttm page %p is owned", m));
134 	KASSERT(m->wire_count == 1, ("ttm lost wire %p", m));
135 	KASSERT((m->flags & PG_FICTITIOUS) != 0, ("ttm lost fictitious %p", m));
136 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("ttm got unmanaged %p", m));
137 	m->flags &= ~PG_FICTITIOUS;
138 	m->oflags |= VPO_UNMANAGED;
139 	vm_page_unwire(m, PQ_NONE);
140 	vm_page_free(m);
141 }
142 
143 static vm_memattr_t
144 ttm_caching_state_to_vm(enum ttm_caching_state cstate)
145 {
146 
147 	switch (cstate) {
148 	case tt_uncached:
149 		return (VM_MEMATTR_UNCACHEABLE);
150 	case tt_wc:
151 		return (VM_MEMATTR_WRITE_COMBINING);
152 	case tt_cached:
153 		return (VM_MEMATTR_WRITE_BACK);
154 	}
155 	panic("caching state %d\n", cstate);
156 }
157 
158 static vm_page_t
159 ttm_vm_page_alloc_dma32(int req, vm_memattr_t memattr)
160 {
161 	vm_page_t p;
162 	int tries;
163 
164 	for (tries = 0; ; tries++) {
165 		p = vm_page_alloc_contig(NULL, 0, req, 1, 0, 0xffffffff,
166 		    PAGE_SIZE, 0, memattr);
167 		if (p != NULL || tries > 2)
168 			return (p);
169 		if (!vm_page_reclaim_contig(req, 1, 0, 0xffffffff,
170 		    PAGE_SIZE, 0))
171 			vm_wait(NULL);
172 	}
173 }
174 
175 static vm_page_t
176 ttm_vm_page_alloc_any(int req, vm_memattr_t memattr)
177 {
178 	vm_page_t p;
179 
180 	while (1) {
181 		p = vm_page_alloc(NULL, 0, req);
182 		if (p != NULL)
183 			break;
184 		vm_wait(NULL);
185 	}
186 	pmap_page_set_memattr(p, memattr);
187 	return (p);
188 }
189 
190 static vm_page_t
191 ttm_vm_page_alloc(int flags, enum ttm_caching_state cstate)
192 {
193 	vm_page_t p;
194 	vm_memattr_t memattr;
195 	int req;
196 
197 	memattr = ttm_caching_state_to_vm(cstate);
198 	req = VM_ALLOC_NORMAL | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ;
199 	if ((flags & TTM_PAGE_FLAG_ZERO_ALLOC) != 0)
200 		req |= VM_ALLOC_ZERO;
201 
202 	if ((flags & TTM_PAGE_FLAG_DMA32) != 0)
203 		p = ttm_vm_page_alloc_dma32(req, memattr);
204 	else
205 		p = ttm_vm_page_alloc_any(req, memattr);
206 
207 	if (p != NULL) {
208 		p->oflags &= ~VPO_UNMANAGED;
209 		p->flags |= PG_FICTITIOUS;
210 	}
211 	return (p);
212 }
213 
214 static void ttm_pool_kobj_release(struct ttm_pool_manager *m)
215 {
216 
217 	free(m, M_TTM_POOLMGR);
218 }
219 
220 #if 0
221 /* XXXKIB sysctl */
222 static ssize_t ttm_pool_store(struct ttm_pool_manager *m,
223 		struct attribute *attr, const char *buffer, size_t size)
224 {
225 	int chars;
226 	unsigned val;
227 	chars = sscanf(buffer, "%u", &val);
228 	if (chars == 0)
229 		return size;
230 
231 	/* Convert kb to number of pages */
232 	val = val / (PAGE_SIZE >> 10);
233 
234 	if (attr == &ttm_page_pool_max)
235 		m->options.max_size = val;
236 	else if (attr == &ttm_page_pool_small)
237 		m->options.small = val;
238 	else if (attr == &ttm_page_pool_alloc_size) {
239 		if (val > NUM_PAGES_TO_ALLOC*8) {
240 			pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
241 			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
242 			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
243 			return size;
244 		} else if (val > NUM_PAGES_TO_ALLOC) {
245 			pr_warn("Setting allocation size to larger than %lu is not recommended\n",
246 				NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
247 		}
248 		m->options.alloc_size = val;
249 	}
250 
251 	return size;
252 }
253 
254 static ssize_t ttm_pool_show(struct ttm_pool_manager *m,
255 		struct attribute *attr, char *buffer)
256 {
257 	unsigned val = 0;
258 
259 	if (attr == &ttm_page_pool_max)
260 		val = m->options.max_size;
261 	else if (attr == &ttm_page_pool_small)
262 		val = m->options.small;
263 	else if (attr == &ttm_page_pool_alloc_size)
264 		val = m->options.alloc_size;
265 
266 	val = val * (PAGE_SIZE >> 10);
267 
268 	return snprintf(buffer, PAGE_SIZE, "%u\n", val);
269 }
270 #endif
271 
272 static struct ttm_pool_manager *_manager;
273 
274 static int set_pages_array_wb(vm_page_t *pages, int addrinarray)
275 {
276 #ifdef TTM_HAS_AGP
277 	int i;
278 
279 	for (i = 0; i < addrinarray; i++)
280 		pmap_page_set_memattr(pages[i], VM_MEMATTR_WRITE_BACK);
281 #endif
282 	return 0;
283 }
284 
285 static int set_pages_array_wc(vm_page_t *pages, int addrinarray)
286 {
287 #ifdef TTM_HAS_AGP
288 	int i;
289 
290 	for (i = 0; i < addrinarray; i++)
291 		pmap_page_set_memattr(pages[i], VM_MEMATTR_WRITE_COMBINING);
292 #endif
293 	return 0;
294 }
295 
296 static int set_pages_array_uc(vm_page_t *pages, int addrinarray)
297 {
298 #ifdef TTM_HAS_AGP
299 	int i;
300 
301 	for (i = 0; i < addrinarray; i++)
302 		pmap_page_set_memattr(pages[i], VM_MEMATTR_UNCACHEABLE);
303 #endif
304 	return 0;
305 }
306 
307 /**
308  * Select the right pool or requested caching state and ttm flags. */
309 static struct ttm_page_pool *ttm_get_pool(int flags,
310 		enum ttm_caching_state cstate)
311 {
312 	int pool_index;
313 
314 	if (cstate == tt_cached)
315 		return NULL;
316 
317 	if (cstate == tt_wc)
318 		pool_index = 0x0;
319 	else
320 		pool_index = 0x1;
321 
322 	if (flags & TTM_PAGE_FLAG_DMA32)
323 		pool_index |= 0x2;
324 
325 	return &_manager->pools[pool_index];
326 }
327 
328 /* set memory back to wb and free the pages. */
329 static void ttm_pages_put(vm_page_t *pages, unsigned npages)
330 {
331 	unsigned i;
332 
333 	/* Our VM handles vm memattr automatically on the page free. */
334 	if (set_pages_array_wb(pages, npages))
335 		printf("[TTM] Failed to set %d pages to wb!\n", npages);
336 	for (i = 0; i < npages; ++i)
337 		ttm_vm_page_free(pages[i]);
338 }
339 
340 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
341 		unsigned freed_pages)
342 {
343 	pool->npages -= freed_pages;
344 	pool->nfrees += freed_pages;
345 }
346 
347 /**
348  * Free pages from pool.
349  *
350  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
351  * number of pages in one go.
352  *
353  * @pool: to free the pages from
354  * @free_all: If set to true will free all pages in pool
355  **/
356 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free)
357 {
358 	vm_page_t p, p1;
359 	vm_page_t *pages_to_free;
360 	unsigned freed_pages = 0,
361 		 npages_to_free = nr_free;
362 	unsigned i;
363 
364 	if (NUM_PAGES_TO_ALLOC < nr_free)
365 		npages_to_free = NUM_PAGES_TO_ALLOC;
366 
367 	pages_to_free = malloc(npages_to_free * sizeof(vm_page_t),
368 	    M_TEMP, M_WAITOK | M_ZERO);
369 
370 restart:
371 	mtx_lock(&pool->lock);
372 
373 	TAILQ_FOREACH_REVERSE_SAFE(p, &pool->list, pglist, plinks.q, p1) {
374 		if (freed_pages >= npages_to_free)
375 			break;
376 
377 		pages_to_free[freed_pages++] = p;
378 		/* We can only remove NUM_PAGES_TO_ALLOC at a time. */
379 		if (freed_pages >= NUM_PAGES_TO_ALLOC) {
380 			/* remove range of pages from the pool */
381 			for (i = 0; i < freed_pages; i++)
382 				TAILQ_REMOVE(&pool->list, pages_to_free[i], plinks.q);
383 
384 			ttm_pool_update_free_locked(pool, freed_pages);
385 			/**
386 			 * Because changing page caching is costly
387 			 * we unlock the pool to prevent stalling.
388 			 */
389 			mtx_unlock(&pool->lock);
390 
391 			ttm_pages_put(pages_to_free, freed_pages);
392 			if (likely(nr_free != FREE_ALL_PAGES))
393 				nr_free -= freed_pages;
394 
395 			if (NUM_PAGES_TO_ALLOC >= nr_free)
396 				npages_to_free = nr_free;
397 			else
398 				npages_to_free = NUM_PAGES_TO_ALLOC;
399 
400 			freed_pages = 0;
401 
402 			/* free all so restart the processing */
403 			if (nr_free)
404 				goto restart;
405 
406 			/* Not allowed to fall through or break because
407 			 * following context is inside spinlock while we are
408 			 * outside here.
409 			 */
410 			goto out;
411 
412 		}
413 	}
414 
415 	/* remove range of pages from the pool */
416 	if (freed_pages) {
417 		for (i = 0; i < freed_pages; i++)
418 			TAILQ_REMOVE(&pool->list, pages_to_free[i], plinks.q);
419 
420 		ttm_pool_update_free_locked(pool, freed_pages);
421 		nr_free -= freed_pages;
422 	}
423 
424 	mtx_unlock(&pool->lock);
425 
426 	if (freed_pages)
427 		ttm_pages_put(pages_to_free, freed_pages);
428 out:
429 	free(pages_to_free, M_TEMP);
430 	return nr_free;
431 }
432 
433 /* Get good estimation how many pages are free in pools */
434 static int ttm_pool_get_num_unused_pages(void)
435 {
436 	unsigned i;
437 	int total = 0;
438 	for (i = 0; i < NUM_POOLS; ++i)
439 		total += _manager->pools[i].npages;
440 
441 	return total;
442 }
443 
444 /**
445  * Callback for mm to request pool to reduce number of page held.
446  */
447 static int ttm_pool_mm_shrink(void *arg)
448 {
449 	static unsigned int start_pool = 0;
450 	unsigned i;
451 	unsigned pool_offset = atomic_fetchadd_int(&start_pool, 1);
452 	struct ttm_page_pool *pool;
453 	int shrink_pages = 100; /* XXXKIB */
454 
455 	pool_offset = pool_offset % NUM_POOLS;
456 	/* select start pool in round robin fashion */
457 	for (i = 0; i < NUM_POOLS; ++i) {
458 		unsigned nr_free = shrink_pages;
459 		if (shrink_pages == 0)
460 			break;
461 		pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
462 		shrink_pages = ttm_page_pool_free(pool, nr_free);
463 	}
464 	/* return estimated number of unused pages in pool */
465 	return ttm_pool_get_num_unused_pages();
466 }
467 
468 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
469 {
470 
471 	manager->lowmem_handler = EVENTHANDLER_REGISTER(vm_lowmem,
472 	    ttm_pool_mm_shrink, manager, EVENTHANDLER_PRI_ANY);
473 }
474 
475 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
476 {
477 
478 	EVENTHANDLER_DEREGISTER(vm_lowmem, manager->lowmem_handler);
479 }
480 
481 static int ttm_set_pages_caching(vm_page_t *pages,
482 		enum ttm_caching_state cstate, unsigned cpages)
483 {
484 	int r = 0;
485 	/* Set page caching */
486 	switch (cstate) {
487 	case tt_uncached:
488 		r = set_pages_array_uc(pages, cpages);
489 		if (r)
490 			printf("[TTM] Failed to set %d pages to uc!\n", cpages);
491 		break;
492 	case tt_wc:
493 		r = set_pages_array_wc(pages, cpages);
494 		if (r)
495 			printf("[TTM] Failed to set %d pages to wc!\n", cpages);
496 		break;
497 	default:
498 		break;
499 	}
500 	return r;
501 }
502 
503 /**
504  * Free pages the pages that failed to change the caching state. If there is
505  * any pages that have changed their caching state already put them to the
506  * pool.
507  */
508 static void ttm_handle_caching_state_failure(struct pglist *pages,
509 		int ttm_flags, enum ttm_caching_state cstate,
510 		vm_page_t *failed_pages, unsigned cpages)
511 {
512 	unsigned i;
513 	/* Failed pages have to be freed */
514 	for (i = 0; i < cpages; ++i) {
515 		TAILQ_REMOVE(pages, failed_pages[i], plinks.q);
516 		ttm_vm_page_free(failed_pages[i]);
517 	}
518 }
519 
520 /**
521  * Allocate new pages with correct caching.
522  *
523  * This function is reentrant if caller updates count depending on number of
524  * pages returned in pages array.
525  */
526 static int ttm_alloc_new_pages(struct pglist *pages, int ttm_alloc_flags,
527 		int ttm_flags, enum ttm_caching_state cstate, unsigned count)
528 {
529 	vm_page_t *caching_array;
530 	vm_page_t p;
531 	int r = 0;
532 	unsigned i, cpages;
533 	unsigned max_cpages = min(count,
534 			(unsigned)(PAGE_SIZE/sizeof(vm_page_t)));
535 
536 	/* allocate array for page caching change */
537 	caching_array = malloc(max_cpages * sizeof(vm_page_t), M_TEMP,
538 	    M_WAITOK | M_ZERO);
539 
540 	for (i = 0, cpages = 0; i < count; ++i) {
541 		p = ttm_vm_page_alloc(ttm_alloc_flags, cstate);
542 		if (!p) {
543 			printf("[TTM] Unable to get page %u\n", i);
544 
545 			/* store already allocated pages in the pool after
546 			 * setting the caching state */
547 			if (cpages) {
548 				r = ttm_set_pages_caching(caching_array,
549 							  cstate, cpages);
550 				if (r)
551 					ttm_handle_caching_state_failure(pages,
552 						ttm_flags, cstate,
553 						caching_array, cpages);
554 			}
555 			r = -ENOMEM;
556 			goto out;
557 		}
558 
559 #ifdef CONFIG_HIGHMEM /* KIB: nop */
560 		/* gfp flags of highmem page should never be dma32 so we
561 		 * we should be fine in such case
562 		 */
563 		if (!PageHighMem(p))
564 #endif
565 		{
566 			caching_array[cpages++] = p;
567 			if (cpages == max_cpages) {
568 
569 				r = ttm_set_pages_caching(caching_array,
570 						cstate, cpages);
571 				if (r) {
572 					ttm_handle_caching_state_failure(pages,
573 						ttm_flags, cstate,
574 						caching_array, cpages);
575 					goto out;
576 				}
577 				cpages = 0;
578 			}
579 		}
580 
581 		TAILQ_INSERT_HEAD(pages, p, plinks.q);
582 	}
583 
584 	if (cpages) {
585 		r = ttm_set_pages_caching(caching_array, cstate, cpages);
586 		if (r)
587 			ttm_handle_caching_state_failure(pages,
588 					ttm_flags, cstate,
589 					caching_array, cpages);
590 	}
591 out:
592 	free(caching_array, M_TEMP);
593 
594 	return r;
595 }
596 
597 /**
598  * Fill the given pool if there aren't enough pages and the requested number of
599  * pages is small.
600  */
601 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
602     int ttm_flags, enum ttm_caching_state cstate, unsigned count)
603 {
604 	vm_page_t p;
605 	int r;
606 	unsigned cpages = 0;
607 	/**
608 	 * Only allow one pool fill operation at a time.
609 	 * If pool doesn't have enough pages for the allocation new pages are
610 	 * allocated from outside of pool.
611 	 */
612 	if (pool->fill_lock)
613 		return;
614 
615 	pool->fill_lock = true;
616 
617 	/* If allocation request is small and there are not enough
618 	 * pages in a pool we fill the pool up first. */
619 	if (count < _manager->options.small
620 		&& count > pool->npages) {
621 		struct pglist new_pages;
622 		unsigned alloc_size = _manager->options.alloc_size;
623 
624 		/**
625 		 * Can't change page caching if in irqsave context. We have to
626 		 * drop the pool->lock.
627 		 */
628 		mtx_unlock(&pool->lock);
629 
630 		TAILQ_INIT(&new_pages);
631 		r = ttm_alloc_new_pages(&new_pages, pool->ttm_page_alloc_flags,
632 		    ttm_flags, cstate, alloc_size);
633 		mtx_lock(&pool->lock);
634 
635 		if (!r) {
636 			TAILQ_CONCAT(&pool->list, &new_pages, plinks.q);
637 			++pool->nrefills;
638 			pool->npages += alloc_size;
639 		} else {
640 			printf("[TTM] Failed to fill pool (%p)\n", pool);
641 			/* If we have any pages left put them to the pool. */
642 			TAILQ_FOREACH(p, &pool->list, plinks.q) {
643 				++cpages;
644 			}
645 			TAILQ_CONCAT(&pool->list, &new_pages, plinks.q);
646 			pool->npages += cpages;
647 		}
648 
649 	}
650 	pool->fill_lock = false;
651 }
652 
653 /**
654  * Cut 'count' number of pages from the pool and put them on the return list.
655  *
656  * @return count of pages still required to fulfill the request.
657  */
658 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
659 					struct pglist *pages,
660 					int ttm_flags,
661 					enum ttm_caching_state cstate,
662 					unsigned count)
663 {
664 	vm_page_t p;
665 	unsigned i;
666 
667 	mtx_lock(&pool->lock);
668 	ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count);
669 
670 	if (count >= pool->npages) {
671 		/* take all pages from the pool */
672 		TAILQ_CONCAT(pages, &pool->list, plinks.q);
673 		count -= pool->npages;
674 		pool->npages = 0;
675 		goto out;
676 	}
677 	for (i = 0; i < count; i++) {
678 		p = TAILQ_FIRST(&pool->list);
679 		TAILQ_REMOVE(&pool->list, p, plinks.q);
680 		TAILQ_INSERT_TAIL(pages, p, plinks.q);
681 	}
682 	pool->npages -= count;
683 	count = 0;
684 out:
685 	mtx_unlock(&pool->lock);
686 	return count;
687 }
688 
689 /* Put all pages in pages list to correct pool to wait for reuse */
690 static void ttm_put_pages(vm_page_t *pages, unsigned npages, int flags,
691 			  enum ttm_caching_state cstate)
692 {
693 	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
694 	unsigned i;
695 
696 	if (pool == NULL) {
697 		/* No pool for this memory type so free the pages */
698 		for (i = 0; i < npages; i++) {
699 			if (pages[i]) {
700 				ttm_vm_page_free(pages[i]);
701 				pages[i] = NULL;
702 			}
703 		}
704 		return;
705 	}
706 
707 	mtx_lock(&pool->lock);
708 	for (i = 0; i < npages; i++) {
709 		if (pages[i]) {
710 			TAILQ_INSERT_TAIL(&pool->list, pages[i], plinks.q);
711 			pages[i] = NULL;
712 			pool->npages++;
713 		}
714 	}
715 	/* Check that we don't go over the pool limit */
716 	npages = 0;
717 	if (pool->npages > _manager->options.max_size) {
718 		npages = pool->npages - _manager->options.max_size;
719 		/* free at least NUM_PAGES_TO_ALLOC number of pages
720 		 * to reduce calls to set_memory_wb */
721 		if (npages < NUM_PAGES_TO_ALLOC)
722 			npages = NUM_PAGES_TO_ALLOC;
723 	}
724 	mtx_unlock(&pool->lock);
725 	if (npages)
726 		ttm_page_pool_free(pool, npages);
727 }
728 
729 /*
730  * On success pages list will hold count number of correctly
731  * cached pages.
732  */
733 static int ttm_get_pages(vm_page_t *pages, unsigned npages, int flags,
734 			 enum ttm_caching_state cstate)
735 {
736 	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
737 	struct pglist plist;
738 	vm_page_t p = NULL;
739 	int gfp_flags;
740 	unsigned count;
741 	int r;
742 
743 	/* No pool for cached pages */
744 	if (pool == NULL) {
745 		for (r = 0; r < npages; ++r) {
746 			p = ttm_vm_page_alloc(flags, cstate);
747 			if (!p) {
748 				printf("[TTM] Unable to allocate page\n");
749 				return -ENOMEM;
750 			}
751 			pages[r] = p;
752 		}
753 		return 0;
754 	}
755 
756 	/* combine zero flag to pool flags */
757 	gfp_flags = flags | pool->ttm_page_alloc_flags;
758 
759 	/* First we take pages from the pool */
760 	TAILQ_INIT(&plist);
761 	npages = ttm_page_pool_get_pages(pool, &plist, flags, cstate, npages);
762 	count = 0;
763 	TAILQ_FOREACH(p, &plist, plinks.q) {
764 		pages[count++] = p;
765 	}
766 
767 	/* clear the pages coming from the pool if requested */
768 	if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
769 		TAILQ_FOREACH(p, &plist, plinks.q) {
770 			pmap_zero_page(p);
771 		}
772 	}
773 
774 	/* If pool didn't have enough pages allocate new one. */
775 	if (npages > 0) {
776 		/* ttm_alloc_new_pages doesn't reference pool so we can run
777 		 * multiple requests in parallel.
778 		 **/
779 		TAILQ_INIT(&plist);
780 		r = ttm_alloc_new_pages(&plist, gfp_flags, flags, cstate,
781 		    npages);
782 		TAILQ_FOREACH(p, &plist, plinks.q) {
783 			pages[count++] = p;
784 		}
785 		if (r) {
786 			/* If there is any pages in the list put them back to
787 			 * the pool. */
788 			printf("[TTM] Failed to allocate extra pages for large request\n");
789 			ttm_put_pages(pages, count, flags, cstate);
790 			return r;
791 		}
792 	}
793 
794 	return 0;
795 }
796 
797 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags,
798 				      char *name)
799 {
800 	mtx_init(&pool->lock, "ttmpool", NULL, MTX_DEF);
801 	pool->fill_lock = false;
802 	TAILQ_INIT(&pool->list);
803 	pool->npages = pool->nfrees = 0;
804 	pool->ttm_page_alloc_flags = flags;
805 	pool->name = name;
806 }
807 
808 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
809 {
810 
811 	if (_manager != NULL)
812 		printf("[TTM] manager != NULL\n");
813 	printf("[TTM] Initializing pool allocator\n");
814 
815 	_manager = malloc(sizeof(*_manager), M_TTM_POOLMGR, M_WAITOK | M_ZERO);
816 
817 	ttm_page_pool_init_locked(&_manager->wc_pool, 0, "wc");
818 	ttm_page_pool_init_locked(&_manager->uc_pool, 0, "uc");
819 	ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
820 	    TTM_PAGE_FLAG_DMA32, "wc dma");
821 	ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
822 	    TTM_PAGE_FLAG_DMA32, "uc dma");
823 
824 	_manager->options.max_size = max_pages;
825 	_manager->options.small = SMALL_ALLOCATION;
826 	_manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
827 
828 	refcount_init(&_manager->kobj_ref, 1);
829 	ttm_pool_mm_shrink_init(_manager);
830 
831 	return 0;
832 }
833 
834 void ttm_page_alloc_fini(void)
835 {
836 	int i;
837 
838 	printf("[TTM] Finalizing pool allocator\n");
839 	ttm_pool_mm_shrink_fini(_manager);
840 
841 	for (i = 0; i < NUM_POOLS; ++i)
842 		ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES);
843 
844 	if (refcount_release(&_manager->kobj_ref))
845 		ttm_pool_kobj_release(_manager);
846 	_manager = NULL;
847 }
848 
849 int ttm_pool_populate(struct ttm_tt *ttm)
850 {
851 	struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
852 	unsigned i;
853 	int ret;
854 
855 	if (ttm->state != tt_unpopulated)
856 		return 0;
857 
858 	for (i = 0; i < ttm->num_pages; ++i) {
859 		ret = ttm_get_pages(&ttm->pages[i], 1,
860 				    ttm->page_flags,
861 				    ttm->caching_state);
862 		if (ret != 0) {
863 			ttm_pool_unpopulate(ttm);
864 			return -ENOMEM;
865 		}
866 
867 		ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
868 						false, false);
869 		if (unlikely(ret != 0)) {
870 			ttm_pool_unpopulate(ttm);
871 			return -ENOMEM;
872 		}
873 	}
874 
875 	if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
876 		ret = ttm_tt_swapin(ttm);
877 		if (unlikely(ret != 0)) {
878 			ttm_pool_unpopulate(ttm);
879 			return ret;
880 		}
881 	}
882 
883 	ttm->state = tt_unbound;
884 	return 0;
885 }
886 
887 void ttm_pool_unpopulate(struct ttm_tt *ttm)
888 {
889 	unsigned i;
890 
891 	for (i = 0; i < ttm->num_pages; ++i) {
892 		if (ttm->pages[i]) {
893 			ttm_mem_global_free_page(ttm->glob->mem_glob,
894 						 ttm->pages[i]);
895 			ttm_put_pages(&ttm->pages[i], 1,
896 				      ttm->page_flags,
897 				      ttm->caching_state);
898 		}
899 	}
900 	ttm->state = tt_unpopulated;
901 }
902 
903 #if 0
904 /* XXXKIB sysctl */
905 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
906 {
907 	struct ttm_page_pool *p;
908 	unsigned i;
909 	char *h[] = {"pool", "refills", "pages freed", "size"};
910 	if (!_manager) {
911 		seq_printf(m, "No pool allocator running.\n");
912 		return 0;
913 	}
914 	seq_printf(m, "%6s %12s %13s %8s\n",
915 			h[0], h[1], h[2], h[3]);
916 	for (i = 0; i < NUM_POOLS; ++i) {
917 		p = &_manager->pools[i];
918 
919 		seq_printf(m, "%6s %12ld %13ld %8d\n",
920 				p->name, p->nrefills,
921 				p->nfrees, p->npages);
922 	}
923 	return 0;
924 }
925 #endif
926