xref: /freebsd/sys/dev/drm2/ttm/ttm_page_alloc.c (revision 7ef62cebc2f965b0f640263e179276928885e33d)
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 <sys/eventhandler.h>
48 #include <vm/vm_pageout.h>
49 
50 #define NUM_PAGES_TO_ALLOC		(PAGE_SIZE/sizeof(vm_page_t))
51 #define SMALL_ALLOCATION		16
52 #define FREE_ALL_PAGES			(~0U)
53 /* times are in msecs */
54 #define PAGE_FREE_INTERVAL		1000
55 
56 /**
57  * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
58  *
59  * @lock: Protects the shared pool from concurrnet access. Must be used with
60  * irqsave/irqrestore variants because pool allocator maybe called from
61  * delayed work.
62  * @fill_lock: Prevent concurrent calls to fill.
63  * @list: Pool of free uc/wc pages for fast reuse.
64  * @gfp_flags: Flags to pass for alloc_page.
65  * @npages: Number of pages in pool.
66  */
67 struct ttm_page_pool {
68 	struct mtx		lock;
69 	bool			fill_lock;
70 	bool			dma32;
71 	struct pglist		list;
72 	int			ttm_page_alloc_flags;
73 	unsigned		npages;
74 	char			*name;
75 	unsigned long		nfrees;
76 	unsigned long		nrefills;
77 };
78 
79 /**
80  * Limits for the pool. They are handled without locks because only place where
81  * they may change is in sysfs store. They won't have immediate effect anyway
82  * so forcing serialization to access them is pointless.
83  */
84 
85 struct ttm_pool_opts {
86 	unsigned	alloc_size;
87 	unsigned	max_size;
88 	unsigned	small;
89 };
90 
91 #define NUM_POOLS 4
92 
93 /**
94  * struct ttm_pool_manager - Holds memory pools for fst allocation
95  *
96  * Manager is read only object for pool code so it doesn't need locking.
97  *
98  * @free_interval: minimum number of jiffies between freeing pages from pool.
99  * @page_alloc_inited: reference counting for pool allocation.
100  * @work: Work that is used to shrink the pool. Work is only run when there is
101  * some pages to free.
102  * @small_allocation: Limit in number of pages what is small allocation.
103  *
104  * @pools: All pool objects in use.
105  **/
106 struct ttm_pool_manager {
107 	unsigned int kobj_ref;
108 	eventhandler_tag lowmem_handler;
109 	struct ttm_pool_opts	options;
110 
111 	union {
112 		struct ttm_page_pool	u_pools[NUM_POOLS];
113 		struct _utag {
114 			struct ttm_page_pool	u_wc_pool;
115 			struct ttm_page_pool	u_uc_pool;
116 			struct ttm_page_pool	u_wc_pool_dma32;
117 			struct ttm_page_pool	u_uc_pool_dma32;
118 		} _ut;
119 	} _u;
120 };
121 
122 #define	pools _u.u_pools
123 #define	wc_pool _u._ut.u_wc_pool
124 #define	uc_pool _u._ut.u_uc_pool
125 #define	wc_pool_dma32 _u._ut.u_wc_pool_dma32
126 #define	uc_pool_dma32 _u._ut.u_uc_pool_dma32
127 
128 MALLOC_DEFINE(M_TTM_POOLMGR, "ttm_poolmgr", "TTM Pool Manager");
129 
130 static void
131 ttm_vm_page_free(vm_page_t m)
132 {
133 
134 	KASSERT(m->object == NULL, ("ttm page %p is owned", m));
135 	KASSERT(vm_page_wired(m), ("ttm lost wire %p", m));
136 	KASSERT((m->flags & PG_FICTITIOUS) != 0, ("ttm lost fictitious %p", m));
137 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("ttm got unmanaged %p", m));
138 	m->flags &= ~PG_FICTITIOUS;
139 	m->oflags |= VPO_UNMANAGED;
140 	vm_page_unwire_noq(m);
141 	vm_page_free(m);
142 }
143 
144 static vm_memattr_t
145 ttm_caching_state_to_vm(enum ttm_caching_state cstate)
146 {
147 
148 	switch (cstate) {
149 	case tt_uncached:
150 		return (VM_MEMATTR_UNCACHEABLE);
151 	case tt_wc:
152 		return (VM_MEMATTR_WRITE_COMBINING);
153 	case tt_cached:
154 		return (VM_MEMATTR_WRITE_BACK);
155 	}
156 	panic("caching state %d\n", cstate);
157 }
158 
159 static vm_page_t
160 ttm_vm_page_alloc_dma32(int req, vm_memattr_t memattr)
161 {
162 	vm_page_t p;
163 	int tries;
164 
165 	for (tries = 0; ; tries++) {
166 		p = vm_page_alloc_noobj_contig(req, 1, 0, 0xffffffff, PAGE_SIZE,
167 		    0, memattr);
168 		if (p != NULL || tries > 2)
169 			return (p);
170 		if (!vm_page_reclaim_contig(req, 1, 0, 0xffffffff,
171 		    PAGE_SIZE, 0))
172 			vm_wait(NULL);
173 	}
174 }
175 
176 static vm_page_t
177 ttm_vm_page_alloc_any(int req, vm_memattr_t memattr)
178 {
179 	vm_page_t p;
180 
181 	p = vm_page_alloc_noobj(req | VM_ALLOC_WAITOK);
182 	pmap_page_set_memattr(p, memattr);
183 	return (p);
184 }
185 
186 static vm_page_t
187 ttm_vm_page_alloc(int flags, enum ttm_caching_state cstate)
188 {
189 	vm_page_t p;
190 	vm_memattr_t memattr;
191 	int req;
192 
193 	memattr = ttm_caching_state_to_vm(cstate);
194 	req = VM_ALLOC_WIRED;
195 	if ((flags & TTM_PAGE_FLAG_ZERO_ALLOC) != 0)
196 		req |= VM_ALLOC_ZERO;
197 
198 	if ((flags & TTM_PAGE_FLAG_DMA32) != 0)
199 		p = ttm_vm_page_alloc_dma32(req, memattr);
200 	else
201 		p = ttm_vm_page_alloc_any(req, memattr);
202 
203 	if (p != NULL) {
204 		p->oflags &= ~VPO_UNMANAGED;
205 		p->flags |= PG_FICTITIOUS;
206 	}
207 	return (p);
208 }
209 
210 static void ttm_pool_kobj_release(struct ttm_pool_manager *m)
211 {
212 
213 	free(m, M_TTM_POOLMGR);
214 }
215 
216 #if 0
217 /* XXXKIB sysctl */
218 static ssize_t ttm_pool_store(struct ttm_pool_manager *m,
219 		struct attribute *attr, const char *buffer, size_t size)
220 {
221 	int chars;
222 	unsigned val;
223 	chars = sscanf(buffer, "%u", &val);
224 	if (chars == 0)
225 		return size;
226 
227 	/* Convert kb to number of pages */
228 	val = val / (PAGE_SIZE >> 10);
229 
230 	if (attr == &ttm_page_pool_max)
231 		m->options.max_size = val;
232 	else if (attr == &ttm_page_pool_small)
233 		m->options.small = val;
234 	else if (attr == &ttm_page_pool_alloc_size) {
235 		if (val > NUM_PAGES_TO_ALLOC*8) {
236 			pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
237 			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
238 			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
239 			return size;
240 		} else if (val > NUM_PAGES_TO_ALLOC) {
241 			pr_warn("Setting allocation size to larger than %lu is not recommended\n",
242 				NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
243 		}
244 		m->options.alloc_size = val;
245 	}
246 
247 	return size;
248 }
249 
250 static ssize_t ttm_pool_show(struct ttm_pool_manager *m,
251 		struct attribute *attr, char *buffer)
252 {
253 	unsigned val = 0;
254 
255 	if (attr == &ttm_page_pool_max)
256 		val = m->options.max_size;
257 	else if (attr == &ttm_page_pool_small)
258 		val = m->options.small;
259 	else if (attr == &ttm_page_pool_alloc_size)
260 		val = m->options.alloc_size;
261 
262 	val = val * (PAGE_SIZE >> 10);
263 
264 	return snprintf(buffer, PAGE_SIZE, "%u\n", val);
265 }
266 #endif
267 
268 static struct ttm_pool_manager *_manager;
269 
270 static int set_pages_array_wb(vm_page_t *pages, int addrinarray)
271 {
272 #ifdef TTM_HAS_AGP
273 	int i;
274 
275 	for (i = 0; i < addrinarray; i++)
276 		pmap_page_set_memattr(pages[i], VM_MEMATTR_WRITE_BACK);
277 #endif
278 	return 0;
279 }
280 
281 static int set_pages_array_wc(vm_page_t *pages, int addrinarray)
282 {
283 #ifdef TTM_HAS_AGP
284 	int i;
285 
286 	for (i = 0; i < addrinarray; i++)
287 		pmap_page_set_memattr(pages[i], VM_MEMATTR_WRITE_COMBINING);
288 #endif
289 	return 0;
290 }
291 
292 static int set_pages_array_uc(vm_page_t *pages, int addrinarray)
293 {
294 #ifdef TTM_HAS_AGP
295 	int i;
296 
297 	for (i = 0; i < addrinarray; i++)
298 		pmap_page_set_memattr(pages[i], VM_MEMATTR_UNCACHEABLE);
299 #endif
300 	return 0;
301 }
302 
303 /**
304  * Select the right pool or requested caching state and ttm flags. */
305 static struct ttm_page_pool *ttm_get_pool(int flags,
306 		enum ttm_caching_state cstate)
307 {
308 	int pool_index;
309 
310 	if (cstate == tt_cached)
311 		return NULL;
312 
313 	if (cstate == tt_wc)
314 		pool_index = 0x0;
315 	else
316 		pool_index = 0x1;
317 
318 	if (flags & TTM_PAGE_FLAG_DMA32)
319 		pool_index |= 0x2;
320 
321 	return &_manager->pools[pool_index];
322 }
323 
324 /* set memory back to wb and free the pages. */
325 static void ttm_pages_put(vm_page_t *pages, unsigned npages)
326 {
327 	unsigned i;
328 
329 	/* Our VM handles vm memattr automatically on the page free. */
330 	if (set_pages_array_wb(pages, npages))
331 		printf("[TTM] Failed to set %d pages to wb!\n", npages);
332 	for (i = 0; i < npages; ++i)
333 		ttm_vm_page_free(pages[i]);
334 }
335 
336 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
337 		unsigned freed_pages)
338 {
339 	pool->npages -= freed_pages;
340 	pool->nfrees += freed_pages;
341 }
342 
343 /**
344  * Free pages from pool.
345  *
346  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
347  * number of pages in one go.
348  *
349  * @pool: to free the pages from
350  * @free_all: If set to true will free all pages in pool
351  **/
352 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free)
353 {
354 	vm_page_t p, p1;
355 	vm_page_t *pages_to_free;
356 	unsigned freed_pages = 0,
357 		 npages_to_free = nr_free;
358 	unsigned i;
359 
360 	if (NUM_PAGES_TO_ALLOC < nr_free)
361 		npages_to_free = NUM_PAGES_TO_ALLOC;
362 
363 	pages_to_free = malloc(npages_to_free * sizeof(vm_page_t),
364 	    M_TEMP, M_WAITOK | M_ZERO);
365 
366 restart:
367 	mtx_lock(&pool->lock);
368 
369 	TAILQ_FOREACH_REVERSE_SAFE(p, &pool->list, pglist, plinks.q, p1) {
370 		if (freed_pages >= npages_to_free)
371 			break;
372 
373 		pages_to_free[freed_pages++] = p;
374 		/* We can only remove NUM_PAGES_TO_ALLOC at a time. */
375 		if (freed_pages >= NUM_PAGES_TO_ALLOC) {
376 			/* remove range of pages from the pool */
377 			for (i = 0; i < freed_pages; i++)
378 				TAILQ_REMOVE(&pool->list, pages_to_free[i], plinks.q);
379 
380 			ttm_pool_update_free_locked(pool, freed_pages);
381 			/**
382 			 * Because changing page caching is costly
383 			 * we unlock the pool to prevent stalling.
384 			 */
385 			mtx_unlock(&pool->lock);
386 
387 			ttm_pages_put(pages_to_free, freed_pages);
388 			if (likely(nr_free != FREE_ALL_PAGES))
389 				nr_free -= freed_pages;
390 
391 			if (NUM_PAGES_TO_ALLOC >= nr_free)
392 				npages_to_free = nr_free;
393 			else
394 				npages_to_free = NUM_PAGES_TO_ALLOC;
395 
396 			freed_pages = 0;
397 
398 			/* free all so restart the processing */
399 			if (nr_free)
400 				goto restart;
401 
402 			/* Not allowed to fall through or break because
403 			 * following context is inside spinlock while we are
404 			 * outside here.
405 			 */
406 			goto out;
407 
408 		}
409 	}
410 
411 	/* remove range of pages from the pool */
412 	if (freed_pages) {
413 		for (i = 0; i < freed_pages; i++)
414 			TAILQ_REMOVE(&pool->list, pages_to_free[i], plinks.q);
415 
416 		ttm_pool_update_free_locked(pool, freed_pages);
417 		nr_free -= freed_pages;
418 	}
419 
420 	mtx_unlock(&pool->lock);
421 
422 	if (freed_pages)
423 		ttm_pages_put(pages_to_free, freed_pages);
424 out:
425 	free(pages_to_free, M_TEMP);
426 	return nr_free;
427 }
428 
429 /* Get good estimation how many pages are free in pools */
430 static int ttm_pool_get_num_unused_pages(void)
431 {
432 	unsigned i;
433 	int total = 0;
434 	for (i = 0; i < NUM_POOLS; ++i)
435 		total += _manager->pools[i].npages;
436 
437 	return total;
438 }
439 
440 /**
441  * Callback for mm to request pool to reduce number of page held.
442  */
443 static int ttm_pool_mm_shrink(void *arg)
444 {
445 	static unsigned int start_pool = 0;
446 	unsigned i;
447 	unsigned pool_offset = atomic_fetchadd_int(&start_pool, 1);
448 	struct ttm_page_pool *pool;
449 	int shrink_pages = 100; /* XXXKIB */
450 
451 	pool_offset = pool_offset % NUM_POOLS;
452 	/* select start pool in round robin fashion */
453 	for (i = 0; i < NUM_POOLS; ++i) {
454 		unsigned nr_free = shrink_pages;
455 		if (shrink_pages == 0)
456 			break;
457 		pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
458 		shrink_pages = ttm_page_pool_free(pool, nr_free);
459 	}
460 	/* return estimated number of unused pages in pool */
461 	return ttm_pool_get_num_unused_pages();
462 }
463 
464 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
465 {
466 
467 	manager->lowmem_handler = EVENTHANDLER_REGISTER(vm_lowmem,
468 	    ttm_pool_mm_shrink, manager, EVENTHANDLER_PRI_ANY);
469 }
470 
471 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
472 {
473 
474 	EVENTHANDLER_DEREGISTER(vm_lowmem, manager->lowmem_handler);
475 }
476 
477 static int ttm_set_pages_caching(vm_page_t *pages,
478 		enum ttm_caching_state cstate, unsigned cpages)
479 {
480 	int r = 0;
481 	/* Set page caching */
482 	switch (cstate) {
483 	case tt_uncached:
484 		r = set_pages_array_uc(pages, cpages);
485 		if (r)
486 			printf("[TTM] Failed to set %d pages to uc!\n", cpages);
487 		break;
488 	case tt_wc:
489 		r = set_pages_array_wc(pages, cpages);
490 		if (r)
491 			printf("[TTM] Failed to set %d pages to wc!\n", cpages);
492 		break;
493 	default:
494 		break;
495 	}
496 	return r;
497 }
498 
499 /**
500  * Free pages the pages that failed to change the caching state. If there is
501  * any pages that have changed their caching state already put them to the
502  * pool.
503  */
504 static void ttm_handle_caching_state_failure(struct pglist *pages,
505 		int ttm_flags, enum ttm_caching_state cstate,
506 		vm_page_t *failed_pages, unsigned cpages)
507 {
508 	unsigned i;
509 	/* Failed pages have to be freed */
510 	for (i = 0; i < cpages; ++i) {
511 		TAILQ_REMOVE(pages, failed_pages[i], plinks.q);
512 		ttm_vm_page_free(failed_pages[i]);
513 	}
514 }
515 
516 /**
517  * Allocate new pages with correct caching.
518  *
519  * This function is reentrant if caller updates count depending on number of
520  * pages returned in pages array.
521  */
522 static int ttm_alloc_new_pages(struct pglist *pages, int ttm_alloc_flags,
523 		int ttm_flags, enum ttm_caching_state cstate, unsigned count)
524 {
525 	vm_page_t *caching_array;
526 	vm_page_t p;
527 	int r = 0;
528 	unsigned i, cpages;
529 	unsigned max_cpages = min(count,
530 			(unsigned)(PAGE_SIZE/sizeof(vm_page_t)));
531 
532 	/* allocate array for page caching change */
533 	caching_array = malloc(max_cpages * sizeof(vm_page_t), M_TEMP,
534 	    M_WAITOK | M_ZERO);
535 
536 	for (i = 0, cpages = 0; i < count; ++i) {
537 		p = ttm_vm_page_alloc(ttm_alloc_flags, cstate);
538 		if (!p) {
539 			printf("[TTM] Unable to get page %u\n", i);
540 
541 			/* store already allocated pages in the pool after
542 			 * setting the caching state */
543 			if (cpages) {
544 				r = ttm_set_pages_caching(caching_array,
545 							  cstate, cpages);
546 				if (r)
547 					ttm_handle_caching_state_failure(pages,
548 						ttm_flags, cstate,
549 						caching_array, cpages);
550 			}
551 			r = -ENOMEM;
552 			goto out;
553 		}
554 
555 #ifdef CONFIG_HIGHMEM /* KIB: nop */
556 		/* gfp flags of highmem page should never be dma32 so we
557 		 * we should be fine in such case
558 		 */
559 		if (!PageHighMem(p))
560 #endif
561 		{
562 			caching_array[cpages++] = p;
563 			if (cpages == max_cpages) {
564 
565 				r = ttm_set_pages_caching(caching_array,
566 						cstate, cpages);
567 				if (r) {
568 					ttm_handle_caching_state_failure(pages,
569 						ttm_flags, cstate,
570 						caching_array, cpages);
571 					goto out;
572 				}
573 				cpages = 0;
574 			}
575 		}
576 
577 		TAILQ_INSERT_HEAD(pages, p, plinks.q);
578 	}
579 
580 	if (cpages) {
581 		r = ttm_set_pages_caching(caching_array, cstate, cpages);
582 		if (r)
583 			ttm_handle_caching_state_failure(pages,
584 					ttm_flags, cstate,
585 					caching_array, cpages);
586 	}
587 out:
588 	free(caching_array, M_TEMP);
589 
590 	return r;
591 }
592 
593 /**
594  * Fill the given pool if there aren't enough pages and the requested number of
595  * pages is small.
596  */
597 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
598     int ttm_flags, enum ttm_caching_state cstate, unsigned count)
599 {
600 	vm_page_t p;
601 	int r;
602 	unsigned cpages = 0;
603 	/**
604 	 * Only allow one pool fill operation at a time.
605 	 * If pool doesn't have enough pages for the allocation new pages are
606 	 * allocated from outside of pool.
607 	 */
608 	if (pool->fill_lock)
609 		return;
610 
611 	pool->fill_lock = true;
612 
613 	/* If allocation request is small and there are not enough
614 	 * pages in a pool we fill the pool up first. */
615 	if (count < _manager->options.small
616 		&& count > pool->npages) {
617 		struct pglist new_pages;
618 		unsigned alloc_size = _manager->options.alloc_size;
619 
620 		/**
621 		 * Can't change page caching if in irqsave context. We have to
622 		 * drop the pool->lock.
623 		 */
624 		mtx_unlock(&pool->lock);
625 
626 		TAILQ_INIT(&new_pages);
627 		r = ttm_alloc_new_pages(&new_pages, pool->ttm_page_alloc_flags,
628 		    ttm_flags, cstate, alloc_size);
629 		mtx_lock(&pool->lock);
630 
631 		if (!r) {
632 			TAILQ_CONCAT(&pool->list, &new_pages, plinks.q);
633 			++pool->nrefills;
634 			pool->npages += alloc_size;
635 		} else {
636 			printf("[TTM] Failed to fill pool (%p)\n", pool);
637 			/* If we have any pages left put them to the pool. */
638 			TAILQ_FOREACH(p, &pool->list, plinks.q) {
639 				++cpages;
640 			}
641 			TAILQ_CONCAT(&pool->list, &new_pages, plinks.q);
642 			pool->npages += cpages;
643 		}
644 
645 	}
646 	pool->fill_lock = false;
647 }
648 
649 /**
650  * Cut 'count' number of pages from the pool and put them on the return list.
651  *
652  * @return count of pages still required to fulfill the request.
653  */
654 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
655 					struct pglist *pages,
656 					int ttm_flags,
657 					enum ttm_caching_state cstate,
658 					unsigned count)
659 {
660 	vm_page_t p;
661 	unsigned i;
662 
663 	mtx_lock(&pool->lock);
664 	ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count);
665 
666 	if (count >= pool->npages) {
667 		/* take all pages from the pool */
668 		TAILQ_CONCAT(pages, &pool->list, plinks.q);
669 		count -= pool->npages;
670 		pool->npages = 0;
671 		goto out;
672 	}
673 	for (i = 0; i < count; i++) {
674 		p = TAILQ_FIRST(&pool->list);
675 		TAILQ_REMOVE(&pool->list, p, plinks.q);
676 		TAILQ_INSERT_TAIL(pages, p, plinks.q);
677 	}
678 	pool->npages -= count;
679 	count = 0;
680 out:
681 	mtx_unlock(&pool->lock);
682 	return count;
683 }
684 
685 /* Put all pages in pages list to correct pool to wait for reuse */
686 static void ttm_put_pages(vm_page_t *pages, unsigned npages, int flags,
687 			  enum ttm_caching_state cstate)
688 {
689 	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
690 	unsigned i;
691 
692 	if (pool == NULL) {
693 		/* No pool for this memory type so free the pages */
694 		for (i = 0; i < npages; i++) {
695 			if (pages[i]) {
696 				ttm_vm_page_free(pages[i]);
697 				pages[i] = NULL;
698 			}
699 		}
700 		return;
701 	}
702 
703 	mtx_lock(&pool->lock);
704 	for (i = 0; i < npages; i++) {
705 		if (pages[i]) {
706 			TAILQ_INSERT_TAIL(&pool->list, pages[i], plinks.q);
707 			pages[i] = NULL;
708 			pool->npages++;
709 		}
710 	}
711 	/* Check that we don't go over the pool limit */
712 	npages = 0;
713 	if (pool->npages > _manager->options.max_size) {
714 		npages = pool->npages - _manager->options.max_size;
715 		/* free at least NUM_PAGES_TO_ALLOC number of pages
716 		 * to reduce calls to set_memory_wb */
717 		if (npages < NUM_PAGES_TO_ALLOC)
718 			npages = NUM_PAGES_TO_ALLOC;
719 	}
720 	mtx_unlock(&pool->lock);
721 	if (npages)
722 		ttm_page_pool_free(pool, npages);
723 }
724 
725 /*
726  * On success pages list will hold count number of correctly
727  * cached pages.
728  */
729 static int ttm_get_pages(vm_page_t *pages, unsigned npages, int flags,
730 			 enum ttm_caching_state cstate)
731 {
732 	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
733 	struct pglist plist;
734 	vm_page_t p = NULL;
735 	int gfp_flags;
736 	unsigned count;
737 	int r;
738 
739 	/* No pool for cached pages */
740 	if (pool == NULL) {
741 		for (r = 0; r < npages; ++r) {
742 			p = ttm_vm_page_alloc(flags, cstate);
743 			if (!p) {
744 				printf("[TTM] Unable to allocate page\n");
745 				return -ENOMEM;
746 			}
747 			pages[r] = p;
748 		}
749 		return 0;
750 	}
751 
752 	/* combine zero flag to pool flags */
753 	gfp_flags = flags | pool->ttm_page_alloc_flags;
754 
755 	/* First we take pages from the pool */
756 	TAILQ_INIT(&plist);
757 	npages = ttm_page_pool_get_pages(pool, &plist, flags, cstate, npages);
758 	count = 0;
759 	TAILQ_FOREACH(p, &plist, plinks.q) {
760 		pages[count++] = p;
761 	}
762 
763 	/* clear the pages coming from the pool if requested */
764 	if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
765 		TAILQ_FOREACH(p, &plist, plinks.q) {
766 			pmap_zero_page(p);
767 		}
768 	}
769 
770 	/* If pool didn't have enough pages allocate new one. */
771 	if (npages > 0) {
772 		/* ttm_alloc_new_pages doesn't reference pool so we can run
773 		 * multiple requests in parallel.
774 		 **/
775 		TAILQ_INIT(&plist);
776 		r = ttm_alloc_new_pages(&plist, gfp_flags, flags, cstate,
777 		    npages);
778 		TAILQ_FOREACH(p, &plist, plinks.q) {
779 			pages[count++] = p;
780 		}
781 		if (r) {
782 			/* If there is any pages in the list put them back to
783 			 * the pool. */
784 			printf("[TTM] Failed to allocate extra pages for large request\n");
785 			ttm_put_pages(pages, count, flags, cstate);
786 			return r;
787 		}
788 	}
789 
790 	return 0;
791 }
792 
793 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags,
794 				      char *name)
795 {
796 	mtx_init(&pool->lock, "ttmpool", NULL, MTX_DEF);
797 	pool->fill_lock = false;
798 	TAILQ_INIT(&pool->list);
799 	pool->npages = pool->nfrees = 0;
800 	pool->ttm_page_alloc_flags = flags;
801 	pool->name = name;
802 }
803 
804 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
805 {
806 
807 	if (_manager != NULL)
808 		printf("[TTM] manager != NULL\n");
809 	printf("[TTM] Initializing pool allocator\n");
810 
811 	_manager = malloc(sizeof(*_manager), M_TTM_POOLMGR, M_WAITOK | M_ZERO);
812 
813 	ttm_page_pool_init_locked(&_manager->wc_pool, 0, "wc");
814 	ttm_page_pool_init_locked(&_manager->uc_pool, 0, "uc");
815 	ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
816 	    TTM_PAGE_FLAG_DMA32, "wc dma");
817 	ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
818 	    TTM_PAGE_FLAG_DMA32, "uc dma");
819 
820 	_manager->options.max_size = max_pages;
821 	_manager->options.small = SMALL_ALLOCATION;
822 	_manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
823 
824 	refcount_init(&_manager->kobj_ref, 1);
825 	ttm_pool_mm_shrink_init(_manager);
826 
827 	return 0;
828 }
829 
830 void ttm_page_alloc_fini(void)
831 {
832 	int i;
833 
834 	printf("[TTM] Finalizing pool allocator\n");
835 	ttm_pool_mm_shrink_fini(_manager);
836 
837 	for (i = 0; i < NUM_POOLS; ++i)
838 		ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES);
839 
840 	if (refcount_release(&_manager->kobj_ref))
841 		ttm_pool_kobj_release(_manager);
842 	_manager = NULL;
843 }
844 
845 int ttm_pool_populate(struct ttm_tt *ttm)
846 {
847 	struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
848 	unsigned i;
849 	int ret;
850 
851 	if (ttm->state != tt_unpopulated)
852 		return 0;
853 
854 	for (i = 0; i < ttm->num_pages; ++i) {
855 		ret = ttm_get_pages(&ttm->pages[i], 1,
856 				    ttm->page_flags,
857 				    ttm->caching_state);
858 		if (ret != 0) {
859 			ttm_pool_unpopulate(ttm);
860 			return -ENOMEM;
861 		}
862 
863 		ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
864 						false, false);
865 		if (unlikely(ret != 0)) {
866 			ttm_pool_unpopulate(ttm);
867 			return -ENOMEM;
868 		}
869 	}
870 
871 	if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
872 		ret = ttm_tt_swapin(ttm);
873 		if (unlikely(ret != 0)) {
874 			ttm_pool_unpopulate(ttm);
875 			return ret;
876 		}
877 	}
878 
879 	ttm->state = tt_unbound;
880 	return 0;
881 }
882 
883 void ttm_pool_unpopulate(struct ttm_tt *ttm)
884 {
885 	unsigned i;
886 
887 	for (i = 0; i < ttm->num_pages; ++i) {
888 		if (ttm->pages[i]) {
889 			ttm_mem_global_free_page(ttm->glob->mem_glob,
890 						 ttm->pages[i]);
891 			ttm_put_pages(&ttm->pages[i], 1,
892 				      ttm->page_flags,
893 				      ttm->caching_state);
894 		}
895 	}
896 	ttm->state = tt_unpopulated;
897 }
898 
899 #if 0
900 /* XXXKIB sysctl */
901 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
902 {
903 	struct ttm_page_pool *p;
904 	unsigned i;
905 	char *h[] = {"pool", "refills", "pages freed", "size"};
906 	if (!_manager) {
907 		seq_printf(m, "No pool allocator running.\n");
908 		return 0;
909 	}
910 	seq_printf(m, "%6s %12s %13s %8s\n",
911 			h[0], h[1], h[2], h[3]);
912 	for (i = 0; i < NUM_POOLS; ++i) {
913 		p = &_manager->pools[i];
914 
915 		seq_printf(m, "%6s %12ld %13ld %8d\n",
916 				p->name, p->nrefills,
917 				p->nfrees, p->npages);
918 	}
919 	return 0;
920 }
921 #endif
922