xref: /freebsd/sys/dev/drm2/ttm/ttm_page_alloc.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
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_contig(NULL, 0, req, 1, 0, 0xffffffff,
167 		    PAGE_SIZE, 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 	while (1) {
182 		p = vm_page_alloc(NULL, 0, req);
183 		if (p != NULL)
184 			break;
185 		vm_wait(NULL);
186 	}
187 	pmap_page_set_memattr(p, memattr);
188 	return (p);
189 }
190 
191 static vm_page_t
192 ttm_vm_page_alloc(int flags, enum ttm_caching_state cstate)
193 {
194 	vm_page_t p;
195 	vm_memattr_t memattr;
196 	int req;
197 
198 	memattr = ttm_caching_state_to_vm(cstate);
199 	req = VM_ALLOC_NORMAL | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ;
200 	if ((flags & TTM_PAGE_FLAG_ZERO_ALLOC) != 0)
201 		req |= VM_ALLOC_ZERO;
202 
203 	if ((flags & TTM_PAGE_FLAG_DMA32) != 0)
204 		p = ttm_vm_page_alloc_dma32(req, memattr);
205 	else
206 		p = ttm_vm_page_alloc_any(req, memattr);
207 
208 	if (p != NULL) {
209 		p->oflags &= ~VPO_UNMANAGED;
210 		p->flags |= PG_FICTITIOUS;
211 	}
212 	return (p);
213 }
214 
215 static void ttm_pool_kobj_release(struct ttm_pool_manager *m)
216 {
217 
218 	free(m, M_TTM_POOLMGR);
219 }
220 
221 #if 0
222 /* XXXKIB sysctl */
223 static ssize_t ttm_pool_store(struct ttm_pool_manager *m,
224 		struct attribute *attr, const char *buffer, size_t size)
225 {
226 	int chars;
227 	unsigned val;
228 	chars = sscanf(buffer, "%u", &val);
229 	if (chars == 0)
230 		return size;
231 
232 	/* Convert kb to number of pages */
233 	val = val / (PAGE_SIZE >> 10);
234 
235 	if (attr == &ttm_page_pool_max)
236 		m->options.max_size = val;
237 	else if (attr == &ttm_page_pool_small)
238 		m->options.small = val;
239 	else if (attr == &ttm_page_pool_alloc_size) {
240 		if (val > NUM_PAGES_TO_ALLOC*8) {
241 			pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
242 			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
243 			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
244 			return size;
245 		} else if (val > NUM_PAGES_TO_ALLOC) {
246 			pr_warn("Setting allocation size to larger than %lu is not recommended\n",
247 				NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
248 		}
249 		m->options.alloc_size = val;
250 	}
251 
252 	return size;
253 }
254 
255 static ssize_t ttm_pool_show(struct ttm_pool_manager *m,
256 		struct attribute *attr, char *buffer)
257 {
258 	unsigned val = 0;
259 
260 	if (attr == &ttm_page_pool_max)
261 		val = m->options.max_size;
262 	else if (attr == &ttm_page_pool_small)
263 		val = m->options.small;
264 	else if (attr == &ttm_page_pool_alloc_size)
265 		val = m->options.alloc_size;
266 
267 	val = val * (PAGE_SIZE >> 10);
268 
269 	return snprintf(buffer, PAGE_SIZE, "%u\n", val);
270 }
271 #endif
272 
273 static struct ttm_pool_manager *_manager;
274 
275 static int set_pages_array_wb(vm_page_t *pages, int addrinarray)
276 {
277 #ifdef TTM_HAS_AGP
278 	int i;
279 
280 	for (i = 0; i < addrinarray; i++)
281 		pmap_page_set_memattr(pages[i], VM_MEMATTR_WRITE_BACK);
282 #endif
283 	return 0;
284 }
285 
286 static int set_pages_array_wc(vm_page_t *pages, int addrinarray)
287 {
288 #ifdef TTM_HAS_AGP
289 	int i;
290 
291 	for (i = 0; i < addrinarray; i++)
292 		pmap_page_set_memattr(pages[i], VM_MEMATTR_WRITE_COMBINING);
293 #endif
294 	return 0;
295 }
296 
297 static int set_pages_array_uc(vm_page_t *pages, int addrinarray)
298 {
299 #ifdef TTM_HAS_AGP
300 	int i;
301 
302 	for (i = 0; i < addrinarray; i++)
303 		pmap_page_set_memattr(pages[i], VM_MEMATTR_UNCACHEABLE);
304 #endif
305 	return 0;
306 }
307 
308 /**
309  * Select the right pool or requested caching state and ttm flags. */
310 static struct ttm_page_pool *ttm_get_pool(int flags,
311 		enum ttm_caching_state cstate)
312 {
313 	int pool_index;
314 
315 	if (cstate == tt_cached)
316 		return NULL;
317 
318 	if (cstate == tt_wc)
319 		pool_index = 0x0;
320 	else
321 		pool_index = 0x1;
322 
323 	if (flags & TTM_PAGE_FLAG_DMA32)
324 		pool_index |= 0x2;
325 
326 	return &_manager->pools[pool_index];
327 }
328 
329 /* set memory back to wb and free the pages. */
330 static void ttm_pages_put(vm_page_t *pages, unsigned npages)
331 {
332 	unsigned i;
333 
334 	/* Our VM handles vm memattr automatically on the page free. */
335 	if (set_pages_array_wb(pages, npages))
336 		printf("[TTM] Failed to set %d pages to wb!\n", npages);
337 	for (i = 0; i < npages; ++i)
338 		ttm_vm_page_free(pages[i]);
339 }
340 
341 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
342 		unsigned freed_pages)
343 {
344 	pool->npages -= freed_pages;
345 	pool->nfrees += freed_pages;
346 }
347 
348 /**
349  * Free pages from pool.
350  *
351  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
352  * number of pages in one go.
353  *
354  * @pool: to free the pages from
355  * @free_all: If set to true will free all pages in pool
356  **/
357 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free)
358 {
359 	vm_page_t p, p1;
360 	vm_page_t *pages_to_free;
361 	unsigned freed_pages = 0,
362 		 npages_to_free = nr_free;
363 	unsigned i;
364 
365 	if (NUM_PAGES_TO_ALLOC < nr_free)
366 		npages_to_free = NUM_PAGES_TO_ALLOC;
367 
368 	pages_to_free = malloc(npages_to_free * sizeof(vm_page_t),
369 	    M_TEMP, M_WAITOK | M_ZERO);
370 
371 restart:
372 	mtx_lock(&pool->lock);
373 
374 	TAILQ_FOREACH_REVERSE_SAFE(p, &pool->list, pglist, plinks.q, p1) {
375 		if (freed_pages >= npages_to_free)
376 			break;
377 
378 		pages_to_free[freed_pages++] = p;
379 		/* We can only remove NUM_PAGES_TO_ALLOC at a time. */
380 		if (freed_pages >= NUM_PAGES_TO_ALLOC) {
381 			/* remove range of pages from the pool */
382 			for (i = 0; i < freed_pages; i++)
383 				TAILQ_REMOVE(&pool->list, pages_to_free[i], plinks.q);
384 
385 			ttm_pool_update_free_locked(pool, freed_pages);
386 			/**
387 			 * Because changing page caching is costly
388 			 * we unlock the pool to prevent stalling.
389 			 */
390 			mtx_unlock(&pool->lock);
391 
392 			ttm_pages_put(pages_to_free, freed_pages);
393 			if (likely(nr_free != FREE_ALL_PAGES))
394 				nr_free -= freed_pages;
395 
396 			if (NUM_PAGES_TO_ALLOC >= nr_free)
397 				npages_to_free = nr_free;
398 			else
399 				npages_to_free = NUM_PAGES_TO_ALLOC;
400 
401 			freed_pages = 0;
402 
403 			/* free all so restart the processing */
404 			if (nr_free)
405 				goto restart;
406 
407 			/* Not allowed to fall through or break because
408 			 * following context is inside spinlock while we are
409 			 * outside here.
410 			 */
411 			goto out;
412 
413 		}
414 	}
415 
416 	/* remove range of pages from the pool */
417 	if (freed_pages) {
418 		for (i = 0; i < freed_pages; i++)
419 			TAILQ_REMOVE(&pool->list, pages_to_free[i], plinks.q);
420 
421 		ttm_pool_update_free_locked(pool, freed_pages);
422 		nr_free -= freed_pages;
423 	}
424 
425 	mtx_unlock(&pool->lock);
426 
427 	if (freed_pages)
428 		ttm_pages_put(pages_to_free, freed_pages);
429 out:
430 	free(pages_to_free, M_TEMP);
431 	return nr_free;
432 }
433 
434 /* Get good estimation how many pages are free in pools */
435 static int ttm_pool_get_num_unused_pages(void)
436 {
437 	unsigned i;
438 	int total = 0;
439 	for (i = 0; i < NUM_POOLS; ++i)
440 		total += _manager->pools[i].npages;
441 
442 	return total;
443 }
444 
445 /**
446  * Callback for mm to request pool to reduce number of page held.
447  */
448 static int ttm_pool_mm_shrink(void *arg)
449 {
450 	static unsigned int start_pool = 0;
451 	unsigned i;
452 	unsigned pool_offset = atomic_fetchadd_int(&start_pool, 1);
453 	struct ttm_page_pool *pool;
454 	int shrink_pages = 100; /* XXXKIB */
455 
456 	pool_offset = pool_offset % NUM_POOLS;
457 	/* select start pool in round robin fashion */
458 	for (i = 0; i < NUM_POOLS; ++i) {
459 		unsigned nr_free = shrink_pages;
460 		if (shrink_pages == 0)
461 			break;
462 		pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
463 		shrink_pages = ttm_page_pool_free(pool, nr_free);
464 	}
465 	/* return estimated number of unused pages in pool */
466 	return ttm_pool_get_num_unused_pages();
467 }
468 
469 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
470 {
471 
472 	manager->lowmem_handler = EVENTHANDLER_REGISTER(vm_lowmem,
473 	    ttm_pool_mm_shrink, manager, EVENTHANDLER_PRI_ANY);
474 }
475 
476 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
477 {
478 
479 	EVENTHANDLER_DEREGISTER(vm_lowmem, manager->lowmem_handler);
480 }
481 
482 static int ttm_set_pages_caching(vm_page_t *pages,
483 		enum ttm_caching_state cstate, unsigned cpages)
484 {
485 	int r = 0;
486 	/* Set page caching */
487 	switch (cstate) {
488 	case tt_uncached:
489 		r = set_pages_array_uc(pages, cpages);
490 		if (r)
491 			printf("[TTM] Failed to set %d pages to uc!\n", cpages);
492 		break;
493 	case tt_wc:
494 		r = set_pages_array_wc(pages, cpages);
495 		if (r)
496 			printf("[TTM] Failed to set %d pages to wc!\n", cpages);
497 		break;
498 	default:
499 		break;
500 	}
501 	return r;
502 }
503 
504 /**
505  * Free pages the pages that failed to change the caching state. If there is
506  * any pages that have changed their caching state already put them to the
507  * pool.
508  */
509 static void ttm_handle_caching_state_failure(struct pglist *pages,
510 		int ttm_flags, enum ttm_caching_state cstate,
511 		vm_page_t *failed_pages, unsigned cpages)
512 {
513 	unsigned i;
514 	/* Failed pages have to be freed */
515 	for (i = 0; i < cpages; ++i) {
516 		TAILQ_REMOVE(pages, failed_pages[i], plinks.q);
517 		ttm_vm_page_free(failed_pages[i]);
518 	}
519 }
520 
521 /**
522  * Allocate new pages with correct caching.
523  *
524  * This function is reentrant if caller updates count depending on number of
525  * pages returned in pages array.
526  */
527 static int ttm_alloc_new_pages(struct pglist *pages, int ttm_alloc_flags,
528 		int ttm_flags, enum ttm_caching_state cstate, unsigned count)
529 {
530 	vm_page_t *caching_array;
531 	vm_page_t p;
532 	int r = 0;
533 	unsigned i, cpages;
534 	unsigned max_cpages = min(count,
535 			(unsigned)(PAGE_SIZE/sizeof(vm_page_t)));
536 
537 	/* allocate array for page caching change */
538 	caching_array = malloc(max_cpages * sizeof(vm_page_t), M_TEMP,
539 	    M_WAITOK | M_ZERO);
540 
541 	for (i = 0, cpages = 0; i < count; ++i) {
542 		p = ttm_vm_page_alloc(ttm_alloc_flags, cstate);
543 		if (!p) {
544 			printf("[TTM] Unable to get page %u\n", i);
545 
546 			/* store already allocated pages in the pool after
547 			 * setting the caching state */
548 			if (cpages) {
549 				r = ttm_set_pages_caching(caching_array,
550 							  cstate, cpages);
551 				if (r)
552 					ttm_handle_caching_state_failure(pages,
553 						ttm_flags, cstate,
554 						caching_array, cpages);
555 			}
556 			r = -ENOMEM;
557 			goto out;
558 		}
559 
560 #ifdef CONFIG_HIGHMEM /* KIB: nop */
561 		/* gfp flags of highmem page should never be dma32 so we
562 		 * we should be fine in such case
563 		 */
564 		if (!PageHighMem(p))
565 #endif
566 		{
567 			caching_array[cpages++] = p;
568 			if (cpages == max_cpages) {
569 
570 				r = ttm_set_pages_caching(caching_array,
571 						cstate, cpages);
572 				if (r) {
573 					ttm_handle_caching_state_failure(pages,
574 						ttm_flags, cstate,
575 						caching_array, cpages);
576 					goto out;
577 				}
578 				cpages = 0;
579 			}
580 		}
581 
582 		TAILQ_INSERT_HEAD(pages, p, plinks.q);
583 	}
584 
585 	if (cpages) {
586 		r = ttm_set_pages_caching(caching_array, cstate, cpages);
587 		if (r)
588 			ttm_handle_caching_state_failure(pages,
589 					ttm_flags, cstate,
590 					caching_array, cpages);
591 	}
592 out:
593 	free(caching_array, M_TEMP);
594 
595 	return r;
596 }
597 
598 /**
599  * Fill the given pool if there aren't enough pages and the requested number of
600  * pages is small.
601  */
602 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
603     int ttm_flags, enum ttm_caching_state cstate, unsigned count)
604 {
605 	vm_page_t p;
606 	int r;
607 	unsigned cpages = 0;
608 	/**
609 	 * Only allow one pool fill operation at a time.
610 	 * If pool doesn't have enough pages for the allocation new pages are
611 	 * allocated from outside of pool.
612 	 */
613 	if (pool->fill_lock)
614 		return;
615 
616 	pool->fill_lock = true;
617 
618 	/* If allocation request is small and there are not enough
619 	 * pages in a pool we fill the pool up first. */
620 	if (count < _manager->options.small
621 		&& count > pool->npages) {
622 		struct pglist new_pages;
623 		unsigned alloc_size = _manager->options.alloc_size;
624 
625 		/**
626 		 * Can't change page caching if in irqsave context. We have to
627 		 * drop the pool->lock.
628 		 */
629 		mtx_unlock(&pool->lock);
630 
631 		TAILQ_INIT(&new_pages);
632 		r = ttm_alloc_new_pages(&new_pages, pool->ttm_page_alloc_flags,
633 		    ttm_flags, cstate, alloc_size);
634 		mtx_lock(&pool->lock);
635 
636 		if (!r) {
637 			TAILQ_CONCAT(&pool->list, &new_pages, plinks.q);
638 			++pool->nrefills;
639 			pool->npages += alloc_size;
640 		} else {
641 			printf("[TTM] Failed to fill pool (%p)\n", pool);
642 			/* If we have any pages left put them to the pool. */
643 			TAILQ_FOREACH(p, &pool->list, plinks.q) {
644 				++cpages;
645 			}
646 			TAILQ_CONCAT(&pool->list, &new_pages, plinks.q);
647 			pool->npages += cpages;
648 		}
649 
650 	}
651 	pool->fill_lock = false;
652 }
653 
654 /**
655  * Cut 'count' number of pages from the pool and put them on the return list.
656  *
657  * @return count of pages still required to fulfill the request.
658  */
659 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
660 					struct pglist *pages,
661 					int ttm_flags,
662 					enum ttm_caching_state cstate,
663 					unsigned count)
664 {
665 	vm_page_t p;
666 	unsigned i;
667 
668 	mtx_lock(&pool->lock);
669 	ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count);
670 
671 	if (count >= pool->npages) {
672 		/* take all pages from the pool */
673 		TAILQ_CONCAT(pages, &pool->list, plinks.q);
674 		count -= pool->npages;
675 		pool->npages = 0;
676 		goto out;
677 	}
678 	for (i = 0; i < count; i++) {
679 		p = TAILQ_FIRST(&pool->list);
680 		TAILQ_REMOVE(&pool->list, p, plinks.q);
681 		TAILQ_INSERT_TAIL(pages, p, plinks.q);
682 	}
683 	pool->npages -= count;
684 	count = 0;
685 out:
686 	mtx_unlock(&pool->lock);
687 	return count;
688 }
689 
690 /* Put all pages in pages list to correct pool to wait for reuse */
691 static void ttm_put_pages(vm_page_t *pages, unsigned npages, int flags,
692 			  enum ttm_caching_state cstate)
693 {
694 	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
695 	unsigned i;
696 
697 	if (pool == NULL) {
698 		/* No pool for this memory type so free the pages */
699 		for (i = 0; i < npages; i++) {
700 			if (pages[i]) {
701 				ttm_vm_page_free(pages[i]);
702 				pages[i] = NULL;
703 			}
704 		}
705 		return;
706 	}
707 
708 	mtx_lock(&pool->lock);
709 	for (i = 0; i < npages; i++) {
710 		if (pages[i]) {
711 			TAILQ_INSERT_TAIL(&pool->list, pages[i], plinks.q);
712 			pages[i] = NULL;
713 			pool->npages++;
714 		}
715 	}
716 	/* Check that we don't go over the pool limit */
717 	npages = 0;
718 	if (pool->npages > _manager->options.max_size) {
719 		npages = pool->npages - _manager->options.max_size;
720 		/* free at least NUM_PAGES_TO_ALLOC number of pages
721 		 * to reduce calls to set_memory_wb */
722 		if (npages < NUM_PAGES_TO_ALLOC)
723 			npages = NUM_PAGES_TO_ALLOC;
724 	}
725 	mtx_unlock(&pool->lock);
726 	if (npages)
727 		ttm_page_pool_free(pool, npages);
728 }
729 
730 /*
731  * On success pages list will hold count number of correctly
732  * cached pages.
733  */
734 static int ttm_get_pages(vm_page_t *pages, unsigned npages, int flags,
735 			 enum ttm_caching_state cstate)
736 {
737 	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
738 	struct pglist plist;
739 	vm_page_t p = NULL;
740 	int gfp_flags;
741 	unsigned count;
742 	int r;
743 
744 	/* No pool for cached pages */
745 	if (pool == NULL) {
746 		for (r = 0; r < npages; ++r) {
747 			p = ttm_vm_page_alloc(flags, cstate);
748 			if (!p) {
749 				printf("[TTM] Unable to allocate page\n");
750 				return -ENOMEM;
751 			}
752 			pages[r] = p;
753 		}
754 		return 0;
755 	}
756 
757 	/* combine zero flag to pool flags */
758 	gfp_flags = flags | pool->ttm_page_alloc_flags;
759 
760 	/* First we take pages from the pool */
761 	TAILQ_INIT(&plist);
762 	npages = ttm_page_pool_get_pages(pool, &plist, flags, cstate, npages);
763 	count = 0;
764 	TAILQ_FOREACH(p, &plist, plinks.q) {
765 		pages[count++] = p;
766 	}
767 
768 	/* clear the pages coming from the pool if requested */
769 	if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
770 		TAILQ_FOREACH(p, &plist, plinks.q) {
771 			pmap_zero_page(p);
772 		}
773 	}
774 
775 	/* If pool didn't have enough pages allocate new one. */
776 	if (npages > 0) {
777 		/* ttm_alloc_new_pages doesn't reference pool so we can run
778 		 * multiple requests in parallel.
779 		 **/
780 		TAILQ_INIT(&plist);
781 		r = ttm_alloc_new_pages(&plist, gfp_flags, flags, cstate,
782 		    npages);
783 		TAILQ_FOREACH(p, &plist, plinks.q) {
784 			pages[count++] = p;
785 		}
786 		if (r) {
787 			/* If there is any pages in the list put them back to
788 			 * the pool. */
789 			printf("[TTM] Failed to allocate extra pages for large request\n");
790 			ttm_put_pages(pages, count, flags, cstate);
791 			return r;
792 		}
793 	}
794 
795 	return 0;
796 }
797 
798 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags,
799 				      char *name)
800 {
801 	mtx_init(&pool->lock, "ttmpool", NULL, MTX_DEF);
802 	pool->fill_lock = false;
803 	TAILQ_INIT(&pool->list);
804 	pool->npages = pool->nfrees = 0;
805 	pool->ttm_page_alloc_flags = flags;
806 	pool->name = name;
807 }
808 
809 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
810 {
811 
812 	if (_manager != NULL)
813 		printf("[TTM] manager != NULL\n");
814 	printf("[TTM] Initializing pool allocator\n");
815 
816 	_manager = malloc(sizeof(*_manager), M_TTM_POOLMGR, M_WAITOK | M_ZERO);
817 
818 	ttm_page_pool_init_locked(&_manager->wc_pool, 0, "wc");
819 	ttm_page_pool_init_locked(&_manager->uc_pool, 0, "uc");
820 	ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
821 	    TTM_PAGE_FLAG_DMA32, "wc dma");
822 	ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
823 	    TTM_PAGE_FLAG_DMA32, "uc dma");
824 
825 	_manager->options.max_size = max_pages;
826 	_manager->options.small = SMALL_ALLOCATION;
827 	_manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
828 
829 	refcount_init(&_manager->kobj_ref, 1);
830 	ttm_pool_mm_shrink_init(_manager);
831 
832 	return 0;
833 }
834 
835 void ttm_page_alloc_fini(void)
836 {
837 	int i;
838 
839 	printf("[TTM] Finalizing pool allocator\n");
840 	ttm_pool_mm_shrink_fini(_manager);
841 
842 	for (i = 0; i < NUM_POOLS; ++i)
843 		ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES);
844 
845 	if (refcount_release(&_manager->kobj_ref))
846 		ttm_pool_kobj_release(_manager);
847 	_manager = NULL;
848 }
849 
850 int ttm_pool_populate(struct ttm_tt *ttm)
851 {
852 	struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
853 	unsigned i;
854 	int ret;
855 
856 	if (ttm->state != tt_unpopulated)
857 		return 0;
858 
859 	for (i = 0; i < ttm->num_pages; ++i) {
860 		ret = ttm_get_pages(&ttm->pages[i], 1,
861 				    ttm->page_flags,
862 				    ttm->caching_state);
863 		if (ret != 0) {
864 			ttm_pool_unpopulate(ttm);
865 			return -ENOMEM;
866 		}
867 
868 		ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
869 						false, false);
870 		if (unlikely(ret != 0)) {
871 			ttm_pool_unpopulate(ttm);
872 			return -ENOMEM;
873 		}
874 	}
875 
876 	if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
877 		ret = ttm_tt_swapin(ttm);
878 		if (unlikely(ret != 0)) {
879 			ttm_pool_unpopulate(ttm);
880 			return ret;
881 		}
882 	}
883 
884 	ttm->state = tt_unbound;
885 	return 0;
886 }
887 
888 void ttm_pool_unpopulate(struct ttm_tt *ttm)
889 {
890 	unsigned i;
891 
892 	for (i = 0; i < ttm->num_pages; ++i) {
893 		if (ttm->pages[i]) {
894 			ttm_mem_global_free_page(ttm->glob->mem_glob,
895 						 ttm->pages[i]);
896 			ttm_put_pages(&ttm->pages[i], 1,
897 				      ttm->page_flags,
898 				      ttm->caching_state);
899 		}
900 	}
901 	ttm->state = tt_unpopulated;
902 }
903 
904 #if 0
905 /* XXXKIB sysctl */
906 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
907 {
908 	struct ttm_page_pool *p;
909 	unsigned i;
910 	char *h[] = {"pool", "refills", "pages freed", "size"};
911 	if (!_manager) {
912 		seq_printf(m, "No pool allocator running.\n");
913 		return 0;
914 	}
915 	seq_printf(m, "%6s %12s %13s %8s\n",
916 			h[0], h[1], h[2], h[3]);
917 	for (i = 0; i < NUM_POOLS; ++i) {
918 		p = &_manager->pools[i];
919 
920 		seq_printf(m, "%6s %12ld %13ld %8d\n",
921 				p->name, p->nrefills,
922 				p->nfrees, p->npages);
923 	}
924 	return 0;
925 }
926 #endif
927