xref: /freebsd/sys/dev/drm2/ttm/ttm_page_alloc.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
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 #include <dev/drm2/drmP.h>
43 #include <dev/drm2/ttm/ttm_bo_driver.h>
44 #include <dev/drm2/ttm/ttm_page_alloc.h>
45 #include <sys/eventhandler.h>
46 #include <vm/vm_pageout.h>
47 
48 #define NUM_PAGES_TO_ALLOC		(PAGE_SIZE/sizeof(vm_page_t))
49 #define SMALL_ALLOCATION		16
50 #define FREE_ALL_PAGES			(~0U)
51 /* times are in msecs */
52 #define PAGE_FREE_INTERVAL		1000
53 
54 /**
55  * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
56  *
57  * @lock: Protects the shared pool from concurrnet access. Must be used with
58  * irqsave/irqrestore variants because pool allocator maybe called from
59  * delayed work.
60  * @fill_lock: Prevent concurrent calls to fill.
61  * @list: Pool of free uc/wc pages for fast reuse.
62  * @gfp_flags: Flags to pass for alloc_page.
63  * @npages: Number of pages in pool.
64  */
65 struct ttm_page_pool {
66 	struct mtx		lock;
67 	bool			fill_lock;
68 	bool			dma32;
69 	struct pglist		list;
70 	int			ttm_page_alloc_flags;
71 	unsigned		npages;
72 	char			*name;
73 	unsigned long		nfrees;
74 	unsigned long		nrefills;
75 };
76 
77 /**
78  * Limits for the pool. They are handled without locks because only place where
79  * they may change is in sysfs store. They won't have immediate effect anyway
80  * so forcing serialization to access them is pointless.
81  */
82 
83 struct ttm_pool_opts {
84 	unsigned	alloc_size;
85 	unsigned	max_size;
86 	unsigned	small;
87 };
88 
89 #define NUM_POOLS 4
90 
91 /**
92  * struct ttm_pool_manager - Holds memory pools for fst allocation
93  *
94  * Manager is read only object for pool code so it doesn't need locking.
95  *
96  * @free_interval: minimum number of jiffies between freeing pages from pool.
97  * @page_alloc_inited: reference counting for pool allocation.
98  * @work: Work that is used to shrink the pool. Work is only run when there is
99  * some pages to free.
100  * @small_allocation: Limit in number of pages what is small allocation.
101  *
102  * @pools: All pool objects in use.
103  **/
104 struct ttm_pool_manager {
105 	unsigned int kobj_ref;
106 	eventhandler_tag lowmem_handler;
107 	struct ttm_pool_opts	options;
108 
109 	union {
110 		struct ttm_page_pool	u_pools[NUM_POOLS];
111 		struct _utag {
112 			struct ttm_page_pool	u_wc_pool;
113 			struct ttm_page_pool	u_uc_pool;
114 			struct ttm_page_pool	u_wc_pool_dma32;
115 			struct ttm_page_pool	u_uc_pool_dma32;
116 		} _ut;
117 	} _u;
118 };
119 
120 #define	pools _u.u_pools
121 #define	wc_pool _u._ut.u_wc_pool
122 #define	uc_pool _u._ut.u_uc_pool
123 #define	wc_pool_dma32 _u._ut.u_wc_pool_dma32
124 #define	uc_pool_dma32 _u._ut.u_uc_pool_dma32
125 
126 MALLOC_DEFINE(M_TTM_POOLMGR, "ttm_poolmgr", "TTM Pool Manager");
127 
128 static void
129 ttm_vm_page_free(vm_page_t m)
130 {
131 
132 	KASSERT(m->object == NULL, ("ttm page %p is owned", m));
133 	KASSERT(vm_page_wired(m), ("ttm lost wire %p", m));
134 	KASSERT((m->flags & PG_FICTITIOUS) != 0, ("ttm lost fictitious %p", m));
135 	KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("ttm got unmanaged %p", m));
136 	m->flags &= ~PG_FICTITIOUS;
137 	m->oflags |= VPO_UNMANAGED;
138 	vm_page_unwire_noq(m);
139 	vm_page_free(m);
140 }
141 
142 static vm_memattr_t
143 ttm_caching_state_to_vm(enum ttm_caching_state cstate)
144 {
145 
146 	switch (cstate) {
147 	case tt_uncached:
148 		return (VM_MEMATTR_UNCACHEABLE);
149 	case tt_wc:
150 		return (VM_MEMATTR_WRITE_COMBINING);
151 	case tt_cached:
152 		return (VM_MEMATTR_WRITE_BACK);
153 	}
154 	panic("caching state %d\n", cstate);
155 }
156 
157 static vm_page_t
158 ttm_vm_page_alloc_dma32(int req, vm_memattr_t memattr)
159 {
160 	vm_page_t p;
161 	int err, tries;
162 
163 	for (tries = 0; ; tries++) {
164 		p = vm_page_alloc_noobj_contig(req, 1, 0, 0xffffffff, PAGE_SIZE,
165 		    0, memattr);
166 		if (p != NULL || tries > 2)
167 			return (p);
168 		err = vm_page_reclaim_contig(req, 1, 0, 0xffffffff,
169 		    PAGE_SIZE, 0);
170 		if (err == ENOMEM)
171 			vm_wait(NULL);
172 		else if (err != 0)
173 			return (NULL);
174 	}
175 }
176 
177 static vm_page_t
178 ttm_vm_page_alloc_any(int req, vm_memattr_t memattr)
179 {
180 	vm_page_t p;
181 
182 	p = vm_page_alloc_noobj(req | VM_ALLOC_WAITOK);
183 	pmap_page_set_memattr(p, memattr);
184 	return (p);
185 }
186 
187 static vm_page_t
188 ttm_vm_page_alloc(int flags, enum ttm_caching_state cstate)
189 {
190 	vm_page_t p;
191 	vm_memattr_t memattr;
192 	int req;
193 
194 	memattr = ttm_caching_state_to_vm(cstate);
195 	req = VM_ALLOC_WIRED;
196 	if ((flags & TTM_PAGE_FLAG_ZERO_ALLOC) != 0)
197 		req |= VM_ALLOC_ZERO;
198 
199 	if ((flags & TTM_PAGE_FLAG_DMA32) != 0)
200 		p = ttm_vm_page_alloc_dma32(req, memattr);
201 	else
202 		p = ttm_vm_page_alloc_any(req, memattr);
203 
204 	if (p != NULL) {
205 		p->oflags &= ~VPO_UNMANAGED;
206 		p->flags |= PG_FICTITIOUS;
207 	}
208 	return (p);
209 }
210 
211 static void ttm_pool_kobj_release(struct ttm_pool_manager *m)
212 {
213 
214 	free(m, M_TTM_POOLMGR);
215 }
216 
217 #if 0
218 /* XXXKIB sysctl */
219 static ssize_t ttm_pool_store(struct ttm_pool_manager *m,
220 		struct attribute *attr, const char *buffer, size_t size)
221 {
222 	int chars;
223 	unsigned val;
224 	chars = sscanf(buffer, "%u", &val);
225 	if (chars == 0)
226 		return size;
227 
228 	/* Convert kb to number of pages */
229 	val = val / (PAGE_SIZE >> 10);
230 
231 	if (attr == &ttm_page_pool_max)
232 		m->options.max_size = val;
233 	else if (attr == &ttm_page_pool_small)
234 		m->options.small = val;
235 	else if (attr == &ttm_page_pool_alloc_size) {
236 		if (val > NUM_PAGES_TO_ALLOC*8) {
237 			pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
238 			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
239 			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
240 			return size;
241 		} else if (val > NUM_PAGES_TO_ALLOC) {
242 			pr_warn("Setting allocation size to larger than %lu is not recommended\n",
243 				NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
244 		}
245 		m->options.alloc_size = val;
246 	}
247 
248 	return size;
249 }
250 
251 static ssize_t ttm_pool_show(struct ttm_pool_manager *m,
252 		struct attribute *attr, char *buffer)
253 {
254 	unsigned val = 0;
255 
256 	if (attr == &ttm_page_pool_max)
257 		val = m->options.max_size;
258 	else if (attr == &ttm_page_pool_small)
259 		val = m->options.small;
260 	else if (attr == &ttm_page_pool_alloc_size)
261 		val = m->options.alloc_size;
262 
263 	val = val * (PAGE_SIZE >> 10);
264 
265 	return snprintf(buffer, PAGE_SIZE, "%u\n", val);
266 }
267 #endif
268 
269 static struct ttm_pool_manager *_manager;
270 
271 static int set_pages_array_wb(vm_page_t *pages, int addrinarray)
272 {
273 #ifdef TTM_HAS_AGP
274 	int i;
275 
276 	for (i = 0; i < addrinarray; i++)
277 		pmap_page_set_memattr(pages[i], VM_MEMATTR_WRITE_BACK);
278 #endif
279 	return 0;
280 }
281 
282 static int set_pages_array_wc(vm_page_t *pages, int addrinarray)
283 {
284 #ifdef TTM_HAS_AGP
285 	int i;
286 
287 	for (i = 0; i < addrinarray; i++)
288 		pmap_page_set_memattr(pages[i], VM_MEMATTR_WRITE_COMBINING);
289 #endif
290 	return 0;
291 }
292 
293 static int set_pages_array_uc(vm_page_t *pages, int addrinarray)
294 {
295 #ifdef TTM_HAS_AGP
296 	int i;
297 
298 	for (i = 0; i < addrinarray; i++)
299 		pmap_page_set_memattr(pages[i], VM_MEMATTR_UNCACHEABLE);
300 #endif
301 	return 0;
302 }
303 
304 /**
305  * Select the right pool or requested caching state and ttm flags. */
306 static struct ttm_page_pool *ttm_get_pool(int flags,
307 		enum ttm_caching_state cstate)
308 {
309 	int pool_index;
310 
311 	if (cstate == tt_cached)
312 		return NULL;
313 
314 	if (cstate == tt_wc)
315 		pool_index = 0x0;
316 	else
317 		pool_index = 0x1;
318 
319 	if (flags & TTM_PAGE_FLAG_DMA32)
320 		pool_index |= 0x2;
321 
322 	return &_manager->pools[pool_index];
323 }
324 
325 /* set memory back to wb and free the pages. */
326 static void ttm_pages_put(vm_page_t *pages, unsigned npages)
327 {
328 	unsigned i;
329 
330 	/* Our VM handles vm memattr automatically on the page free. */
331 	if (set_pages_array_wb(pages, npages))
332 		printf("[TTM] Failed to set %d pages to wb!\n", npages);
333 	for (i = 0; i < npages; ++i)
334 		ttm_vm_page_free(pages[i]);
335 }
336 
337 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
338 		unsigned freed_pages)
339 {
340 	pool->npages -= freed_pages;
341 	pool->nfrees += freed_pages;
342 }
343 
344 /**
345  * Free pages from pool.
346  *
347  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
348  * number of pages in one go.
349  *
350  * @pool: to free the pages from
351  * @free_all: If set to true will free all pages in pool
352  **/
353 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free)
354 {
355 	vm_page_t p, p1;
356 	vm_page_t *pages_to_free;
357 	unsigned freed_pages = 0,
358 		 npages_to_free = nr_free;
359 	unsigned i;
360 
361 	if (NUM_PAGES_TO_ALLOC < nr_free)
362 		npages_to_free = NUM_PAGES_TO_ALLOC;
363 
364 	pages_to_free = malloc(npages_to_free * sizeof(vm_page_t),
365 	    M_TEMP, M_WAITOK | M_ZERO);
366 
367 restart:
368 	mtx_lock(&pool->lock);
369 
370 	TAILQ_FOREACH_REVERSE_SAFE(p, &pool->list, pglist, plinks.q, p1) {
371 		if (freed_pages >= npages_to_free)
372 			break;
373 
374 		pages_to_free[freed_pages++] = p;
375 		/* We can only remove NUM_PAGES_TO_ALLOC at a time. */
376 		if (freed_pages >= NUM_PAGES_TO_ALLOC) {
377 			/* remove range of pages from the pool */
378 			for (i = 0; i < freed_pages; i++)
379 				TAILQ_REMOVE(&pool->list, pages_to_free[i], plinks.q);
380 
381 			ttm_pool_update_free_locked(pool, freed_pages);
382 			/**
383 			 * Because changing page caching is costly
384 			 * we unlock the pool to prevent stalling.
385 			 */
386 			mtx_unlock(&pool->lock);
387 
388 			ttm_pages_put(pages_to_free, freed_pages);
389 			if (likely(nr_free != FREE_ALL_PAGES))
390 				nr_free -= freed_pages;
391 
392 			if (NUM_PAGES_TO_ALLOC >= nr_free)
393 				npages_to_free = nr_free;
394 			else
395 				npages_to_free = NUM_PAGES_TO_ALLOC;
396 
397 			freed_pages = 0;
398 
399 			/* free all so restart the processing */
400 			if (nr_free)
401 				goto restart;
402 
403 			/* Not allowed to fall through or break because
404 			 * following context is inside spinlock while we are
405 			 * outside here.
406 			 */
407 			goto out;
408 
409 		}
410 	}
411 
412 	/* remove range of pages from the pool */
413 	if (freed_pages) {
414 		for (i = 0; i < freed_pages; i++)
415 			TAILQ_REMOVE(&pool->list, pages_to_free[i], plinks.q);
416 
417 		ttm_pool_update_free_locked(pool, freed_pages);
418 		nr_free -= freed_pages;
419 	}
420 
421 	mtx_unlock(&pool->lock);
422 
423 	if (freed_pages)
424 		ttm_pages_put(pages_to_free, freed_pages);
425 out:
426 	free(pages_to_free, M_TEMP);
427 	return nr_free;
428 }
429 
430 /* Get good estimation how many pages are free in pools */
431 static int ttm_pool_get_num_unused_pages(void)
432 {
433 	unsigned i;
434 	int total = 0;
435 	for (i = 0; i < NUM_POOLS; ++i)
436 		total += _manager->pools[i].npages;
437 
438 	return total;
439 }
440 
441 /**
442  * Callback for mm to request pool to reduce number of page held.
443  */
444 static int ttm_pool_mm_shrink(void *arg)
445 {
446 	static unsigned int start_pool = 0;
447 	unsigned i;
448 	unsigned pool_offset = atomic_fetchadd_int(&start_pool, 1);
449 	struct ttm_page_pool *pool;
450 	int shrink_pages = 100; /* XXXKIB */
451 
452 	pool_offset = pool_offset % NUM_POOLS;
453 	/* select start pool in round robin fashion */
454 	for (i = 0; i < NUM_POOLS; ++i) {
455 		unsigned nr_free = shrink_pages;
456 		if (shrink_pages == 0)
457 			break;
458 		pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
459 		shrink_pages = ttm_page_pool_free(pool, nr_free);
460 	}
461 	/* return estimated number of unused pages in pool */
462 	return ttm_pool_get_num_unused_pages();
463 }
464 
465 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
466 {
467 
468 	manager->lowmem_handler = EVENTHANDLER_REGISTER(vm_lowmem,
469 	    ttm_pool_mm_shrink, manager, EVENTHANDLER_PRI_ANY);
470 }
471 
472 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
473 {
474 
475 	EVENTHANDLER_DEREGISTER(vm_lowmem, manager->lowmem_handler);
476 }
477 
478 static int ttm_set_pages_caching(vm_page_t *pages,
479 		enum ttm_caching_state cstate, unsigned cpages)
480 {
481 	int r = 0;
482 	/* Set page caching */
483 	switch (cstate) {
484 	case tt_uncached:
485 		r = set_pages_array_uc(pages, cpages);
486 		if (r)
487 			printf("[TTM] Failed to set %d pages to uc!\n", cpages);
488 		break;
489 	case tt_wc:
490 		r = set_pages_array_wc(pages, cpages);
491 		if (r)
492 			printf("[TTM] Failed to set %d pages to wc!\n", cpages);
493 		break;
494 	default:
495 		break;
496 	}
497 	return r;
498 }
499 
500 /**
501  * Free pages the pages that failed to change the caching state. If there is
502  * any pages that have changed their caching state already put them to the
503  * pool.
504  */
505 static void ttm_handle_caching_state_failure(struct pglist *pages,
506 		int ttm_flags, enum ttm_caching_state cstate,
507 		vm_page_t *failed_pages, unsigned cpages)
508 {
509 	unsigned i;
510 	/* Failed pages have to be freed */
511 	for (i = 0; i < cpages; ++i) {
512 		TAILQ_REMOVE(pages, failed_pages[i], plinks.q);
513 		ttm_vm_page_free(failed_pages[i]);
514 	}
515 }
516 
517 /**
518  * Allocate new pages with correct caching.
519  *
520  * This function is reentrant if caller updates count depending on number of
521  * pages returned in pages array.
522  */
523 static int ttm_alloc_new_pages(struct pglist *pages, int ttm_alloc_flags,
524 		int ttm_flags, enum ttm_caching_state cstate, unsigned count)
525 {
526 	vm_page_t *caching_array;
527 	vm_page_t p;
528 	int r = 0;
529 	unsigned i, cpages;
530 	unsigned max_cpages = min(count,
531 			(unsigned)(PAGE_SIZE/sizeof(vm_page_t)));
532 
533 	/* allocate array for page caching change */
534 	caching_array = malloc(max_cpages * sizeof(vm_page_t), M_TEMP,
535 	    M_WAITOK | M_ZERO);
536 
537 	for (i = 0, cpages = 0; i < count; ++i) {
538 		p = ttm_vm_page_alloc(ttm_alloc_flags, cstate);
539 		if (!p) {
540 			printf("[TTM] Unable to get page %u\n", i);
541 
542 			/* store already allocated pages in the pool after
543 			 * setting the caching state */
544 			if (cpages) {
545 				r = ttm_set_pages_caching(caching_array,
546 							  cstate, cpages);
547 				if (r)
548 					ttm_handle_caching_state_failure(pages,
549 						ttm_flags, cstate,
550 						caching_array, cpages);
551 			}
552 			r = -ENOMEM;
553 			goto out;
554 		}
555 
556 #ifdef CONFIG_HIGHMEM /* KIB: nop */
557 		/* gfp flags of highmem page should never be dma32 so we
558 		 * we should be fine in such case
559 		 */
560 		if (!PageHighMem(p))
561 #endif
562 		{
563 			caching_array[cpages++] = p;
564 			if (cpages == max_cpages) {
565 
566 				r = ttm_set_pages_caching(caching_array,
567 						cstate, cpages);
568 				if (r) {
569 					ttm_handle_caching_state_failure(pages,
570 						ttm_flags, cstate,
571 						caching_array, cpages);
572 					goto out;
573 				}
574 				cpages = 0;
575 			}
576 		}
577 
578 		TAILQ_INSERT_HEAD(pages, p, plinks.q);
579 	}
580 
581 	if (cpages) {
582 		r = ttm_set_pages_caching(caching_array, cstate, cpages);
583 		if (r)
584 			ttm_handle_caching_state_failure(pages,
585 					ttm_flags, cstate,
586 					caching_array, cpages);
587 	}
588 out:
589 	free(caching_array, M_TEMP);
590 
591 	return r;
592 }
593 
594 /**
595  * Fill the given pool if there aren't enough pages and the requested number of
596  * pages is small.
597  */
598 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
599     int ttm_flags, enum ttm_caching_state cstate, unsigned count)
600 {
601 	vm_page_t p;
602 	int r;
603 	unsigned cpages = 0;
604 	/**
605 	 * Only allow one pool fill operation at a time.
606 	 * If pool doesn't have enough pages for the allocation new pages are
607 	 * allocated from outside of pool.
608 	 */
609 	if (pool->fill_lock)
610 		return;
611 
612 	pool->fill_lock = true;
613 
614 	/* If allocation request is small and there are not enough
615 	 * pages in a pool we fill the pool up first. */
616 	if (count < _manager->options.small
617 		&& count > pool->npages) {
618 		struct pglist new_pages;
619 		unsigned alloc_size = _manager->options.alloc_size;
620 
621 		/**
622 		 * Can't change page caching if in irqsave context. We have to
623 		 * drop the pool->lock.
624 		 */
625 		mtx_unlock(&pool->lock);
626 
627 		TAILQ_INIT(&new_pages);
628 		r = ttm_alloc_new_pages(&new_pages, pool->ttm_page_alloc_flags,
629 		    ttm_flags, cstate, alloc_size);
630 		mtx_lock(&pool->lock);
631 
632 		if (!r) {
633 			TAILQ_CONCAT(&pool->list, &new_pages, plinks.q);
634 			++pool->nrefills;
635 			pool->npages += alloc_size;
636 		} else {
637 			printf("[TTM] Failed to fill pool (%p)\n", pool);
638 			/* If we have any pages left put them to the pool. */
639 			TAILQ_FOREACH(p, &pool->list, plinks.q) {
640 				++cpages;
641 			}
642 			TAILQ_CONCAT(&pool->list, &new_pages, plinks.q);
643 			pool->npages += cpages;
644 		}
645 
646 	}
647 	pool->fill_lock = false;
648 }
649 
650 /**
651  * Cut 'count' number of pages from the pool and put them on the return list.
652  *
653  * @return count of pages still required to fulfill the request.
654  */
655 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
656 					struct pglist *pages,
657 					int ttm_flags,
658 					enum ttm_caching_state cstate,
659 					unsigned count)
660 {
661 	vm_page_t p;
662 	unsigned i;
663 
664 	mtx_lock(&pool->lock);
665 	ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count);
666 
667 	if (count >= pool->npages) {
668 		/* take all pages from the pool */
669 		TAILQ_CONCAT(pages, &pool->list, plinks.q);
670 		count -= pool->npages;
671 		pool->npages = 0;
672 		goto out;
673 	}
674 	for (i = 0; i < count; i++) {
675 		p = TAILQ_FIRST(&pool->list);
676 		TAILQ_REMOVE(&pool->list, p, plinks.q);
677 		TAILQ_INSERT_TAIL(pages, p, plinks.q);
678 	}
679 	pool->npages -= count;
680 	count = 0;
681 out:
682 	mtx_unlock(&pool->lock);
683 	return count;
684 }
685 
686 /* Put all pages in pages list to correct pool to wait for reuse */
687 static void ttm_put_pages(vm_page_t *pages, unsigned npages, int flags,
688 			  enum ttm_caching_state cstate)
689 {
690 	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
691 	unsigned i;
692 
693 	if (pool == NULL) {
694 		/* No pool for this memory type so free the pages */
695 		for (i = 0; i < npages; i++) {
696 			if (pages[i]) {
697 				ttm_vm_page_free(pages[i]);
698 				pages[i] = NULL;
699 			}
700 		}
701 		return;
702 	}
703 
704 	mtx_lock(&pool->lock);
705 	for (i = 0; i < npages; i++) {
706 		if (pages[i]) {
707 			TAILQ_INSERT_TAIL(&pool->list, pages[i], plinks.q);
708 			pages[i] = NULL;
709 			pool->npages++;
710 		}
711 	}
712 	/* Check that we don't go over the pool limit */
713 	npages = 0;
714 	if (pool->npages > _manager->options.max_size) {
715 		npages = pool->npages - _manager->options.max_size;
716 		/* free at least NUM_PAGES_TO_ALLOC number of pages
717 		 * to reduce calls to set_memory_wb */
718 		if (npages < NUM_PAGES_TO_ALLOC)
719 			npages = NUM_PAGES_TO_ALLOC;
720 	}
721 	mtx_unlock(&pool->lock);
722 	if (npages)
723 		ttm_page_pool_free(pool, npages);
724 }
725 
726 /*
727  * On success pages list will hold count number of correctly
728  * cached pages.
729  */
730 static int ttm_get_pages(vm_page_t *pages, unsigned npages, int flags,
731 			 enum ttm_caching_state cstate)
732 {
733 	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
734 	struct pglist plist;
735 	vm_page_t p = NULL;
736 	int gfp_flags;
737 	unsigned count;
738 	int r;
739 
740 	/* No pool for cached pages */
741 	if (pool == NULL) {
742 		for (r = 0; r < npages; ++r) {
743 			p = ttm_vm_page_alloc(flags, cstate);
744 			if (!p) {
745 				printf("[TTM] Unable to allocate page\n");
746 				return -ENOMEM;
747 			}
748 			pages[r] = p;
749 		}
750 		return 0;
751 	}
752 
753 	/* combine zero flag to pool flags */
754 	gfp_flags = flags | pool->ttm_page_alloc_flags;
755 
756 	/* First we take pages from the pool */
757 	TAILQ_INIT(&plist);
758 	npages = ttm_page_pool_get_pages(pool, &plist, flags, cstate, npages);
759 	count = 0;
760 	TAILQ_FOREACH(p, &plist, plinks.q) {
761 		pages[count++] = p;
762 	}
763 
764 	/* clear the pages coming from the pool if requested */
765 	if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
766 		TAILQ_FOREACH(p, &plist, plinks.q) {
767 			pmap_zero_page(p);
768 		}
769 	}
770 
771 	/* If pool didn't have enough pages allocate new one. */
772 	if (npages > 0) {
773 		/* ttm_alloc_new_pages doesn't reference pool so we can run
774 		 * multiple requests in parallel.
775 		 **/
776 		TAILQ_INIT(&plist);
777 		r = ttm_alloc_new_pages(&plist, gfp_flags, flags, cstate,
778 		    npages);
779 		TAILQ_FOREACH(p, &plist, plinks.q) {
780 			pages[count++] = p;
781 		}
782 		if (r) {
783 			/* If there is any pages in the list put them back to
784 			 * the pool. */
785 			printf("[TTM] Failed to allocate extra pages for large request\n");
786 			ttm_put_pages(pages, count, flags, cstate);
787 			return r;
788 		}
789 	}
790 
791 	return 0;
792 }
793 
794 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, int flags,
795 				      char *name)
796 {
797 	mtx_init(&pool->lock, "ttmpool", NULL, MTX_DEF);
798 	pool->fill_lock = false;
799 	TAILQ_INIT(&pool->list);
800 	pool->npages = pool->nfrees = 0;
801 	pool->ttm_page_alloc_flags = flags;
802 	pool->name = name;
803 }
804 
805 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
806 {
807 
808 	if (_manager != NULL)
809 		printf("[TTM] manager != NULL\n");
810 	printf("[TTM] Initializing pool allocator\n");
811 
812 	_manager = malloc(sizeof(*_manager), M_TTM_POOLMGR, M_WAITOK | M_ZERO);
813 
814 	ttm_page_pool_init_locked(&_manager->wc_pool, 0, "wc");
815 	ttm_page_pool_init_locked(&_manager->uc_pool, 0, "uc");
816 	ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
817 	    TTM_PAGE_FLAG_DMA32, "wc dma");
818 	ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
819 	    TTM_PAGE_FLAG_DMA32, "uc dma");
820 
821 	_manager->options.max_size = max_pages;
822 	_manager->options.small = SMALL_ALLOCATION;
823 	_manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
824 
825 	refcount_init(&_manager->kobj_ref, 1);
826 	ttm_pool_mm_shrink_init(_manager);
827 
828 	return 0;
829 }
830 
831 void ttm_page_alloc_fini(void)
832 {
833 	int i;
834 
835 	printf("[TTM] Finalizing pool allocator\n");
836 	ttm_pool_mm_shrink_fini(_manager);
837 
838 	for (i = 0; i < NUM_POOLS; ++i)
839 		ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES);
840 
841 	if (refcount_release(&_manager->kobj_ref))
842 		ttm_pool_kobj_release(_manager);
843 	_manager = NULL;
844 }
845 
846 int ttm_pool_populate(struct ttm_tt *ttm)
847 {
848 	struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
849 	unsigned i;
850 	int ret;
851 
852 	if (ttm->state != tt_unpopulated)
853 		return 0;
854 
855 	for (i = 0; i < ttm->num_pages; ++i) {
856 		ret = ttm_get_pages(&ttm->pages[i], 1,
857 				    ttm->page_flags,
858 				    ttm->caching_state);
859 		if (ret != 0) {
860 			ttm_pool_unpopulate(ttm);
861 			return -ENOMEM;
862 		}
863 
864 		ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
865 						false, false);
866 		if (unlikely(ret != 0)) {
867 			ttm_pool_unpopulate(ttm);
868 			return -ENOMEM;
869 		}
870 	}
871 
872 	if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
873 		ret = ttm_tt_swapin(ttm);
874 		if (unlikely(ret != 0)) {
875 			ttm_pool_unpopulate(ttm);
876 			return ret;
877 		}
878 	}
879 
880 	ttm->state = tt_unbound;
881 	return 0;
882 }
883 
884 void ttm_pool_unpopulate(struct ttm_tt *ttm)
885 {
886 	unsigned i;
887 
888 	for (i = 0; i < ttm->num_pages; ++i) {
889 		if (ttm->pages[i]) {
890 			ttm_mem_global_free_page(ttm->glob->mem_glob,
891 						 ttm->pages[i]);
892 			ttm_put_pages(&ttm->pages[i], 1,
893 				      ttm->page_flags,
894 				      ttm->caching_state);
895 		}
896 	}
897 	ttm->state = tt_unpopulated;
898 }
899 
900 #if 0
901 /* XXXKIB sysctl */
902 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
903 {
904 	struct ttm_page_pool *p;
905 	unsigned i;
906 	char *h[] = {"pool", "refills", "pages freed", "size"};
907 	if (!_manager) {
908 		seq_printf(m, "No pool allocator running.\n");
909 		return 0;
910 	}
911 	seq_printf(m, "%6s %12s %13s %8s\n",
912 			h[0], h[1], h[2], h[3]);
913 	for (i = 0; i < NUM_POOLS; ++i) {
914 		p = &_manager->pools[i];
915 
916 		seq_printf(m, "%6s %12ld %13ld %8d\n",
917 				p->name, p->nrefills,
918 				p->nfrees, p->npages);
919 	}
920 	return 0;
921 }
922 #endif
923