xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/ttm/ttm_page_alloc.c (revision bdc22b2e01993381dcefeff2bc9b56ca75a4235c)
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 /* simple list based uncached page pool
29  * - Pool collects resently freed pages for reuse
30  * - Use page->lru to keep a free list
31  * - doesn't track currently in use pages
32  */
33 
34 #define pr_fmt(fmt) "[TTM] " fmt
35 
36 #include <linux/list.h>
37 #include <linux/spinlock.h>
38 #include <linux/highmem.h>
39 #include <linux/mm_types.h>
40 #include <linux/module.h>
41 #include <linux/mm.h>
42 #include <linux/seq_file.h> /* for seq_printf */
43 #include <linux/slab.h>
44 #include <linux/dma-mapping.h>
45 
46 #include <linux/atomic.h>
47 
48 #include <drm/ttm/ttm_bo_driver.h>
49 #include <drm/ttm/ttm_page_alloc.h>
50 
51 #ifdef TTM_HAS_AGP
52 #include <asm/agp.h>
53 #endif
54 
55 #define NUM_PAGES_TO_ALLOC		(PAGE_SIZE/sizeof(struct page *))
56 #define SMALL_ALLOCATION		16
57 #define FREE_ALL_PAGES			(~0U)
58 /* times are in msecs */
59 #define PAGE_FREE_INTERVAL		1000
60 
61 /**
62  * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
63  *
64  * @lock: Protects the shared pool from concurrnet access. Must be used with
65  * irqsave/irqrestore variants because pool allocator maybe called from
66  * delayed work.
67  * @fill_lock: Prevent concurrent calls to fill.
68  * @list: Pool of free uc/wc pages for fast reuse.
69  * @gfp_flags: Flags to pass for alloc_page.
70  * @npages: Number of pages in pool.
71  */
72 struct ttm_page_pool {
73 	spinlock_t		lock;
74 	bool			fill_lock;
75 	struct list_head	list;
76 	gfp_t			gfp_flags;
77 	unsigned		npages;
78 	char			*name;
79 	unsigned long		nfrees;
80 	unsigned long		nrefills;
81 };
82 
83 /**
84  * Limits for the pool. They are handled without locks because only place where
85  * they may change is in sysfs store. They won't have immediate effect anyway
86  * so forcing serialization to access them is pointless.
87  */
88 
89 struct ttm_pool_opts {
90 	unsigned	alloc_size;
91 	unsigned	max_size;
92 	unsigned	small;
93 };
94 
95 #define NUM_POOLS 4
96 
97 /**
98  * struct ttm_pool_manager - Holds memory pools for fst allocation
99  *
100  * Manager is read only object for pool code so it doesn't need locking.
101  *
102  * @free_interval: minimum number of jiffies between freeing pages from pool.
103  * @page_alloc_inited: reference counting for pool allocation.
104  * @work: Work that is used to shrink the pool. Work is only run when there is
105  * some pages to free.
106  * @small_allocation: Limit in number of pages what is small allocation.
107  *
108  * @pools: All pool objects in use.
109  **/
110 struct ttm_pool_manager {
111 	struct kobject		kobj;
112 	struct shrinker		mm_shrink;
113 	struct ttm_pool_opts	options;
114 
115 	union {
116 		struct ttm_page_pool	pools[NUM_POOLS];
117 		struct {
118 			struct ttm_page_pool	wc_pool;
119 			struct ttm_page_pool	uc_pool;
120 			struct ttm_page_pool	wc_pool_dma32;
121 			struct ttm_page_pool	uc_pool_dma32;
122 		} ;
123 	};
124 };
125 
126 static struct attribute ttm_page_pool_max = {
127 	.name = "pool_max_size",
128 	.mode = S_IRUGO | S_IWUSR
129 };
130 static struct attribute ttm_page_pool_small = {
131 	.name = "pool_small_allocation",
132 	.mode = S_IRUGO | S_IWUSR
133 };
134 static struct attribute ttm_page_pool_alloc_size = {
135 	.name = "pool_allocation_size",
136 	.mode = S_IRUGO | S_IWUSR
137 };
138 
139 static struct attribute *ttm_pool_attrs[] = {
140 	&ttm_page_pool_max,
141 	&ttm_page_pool_small,
142 	&ttm_page_pool_alloc_size,
143 	NULL
144 };
145 
146 static void ttm_pool_kobj_release(struct kobject *kobj)
147 {
148 	struct ttm_pool_manager *m =
149 		container_of(kobj, struct ttm_pool_manager, kobj);
150 	kfree(m);
151 }
152 
153 static ssize_t ttm_pool_store(struct kobject *kobj,
154 		struct attribute *attr, const char *buffer, size_t size)
155 {
156 	struct ttm_pool_manager *m =
157 		container_of(kobj, struct ttm_pool_manager, kobj);
158 	int chars;
159 	unsigned val;
160 	chars = sscanf(buffer, "%u", &val);
161 	if (chars == 0)
162 		return size;
163 
164 	/* Convert kb to number of pages */
165 	val = val / (PAGE_SIZE >> 10);
166 
167 	if (attr == &ttm_page_pool_max)
168 		m->options.max_size = val;
169 	else if (attr == &ttm_page_pool_small)
170 		m->options.small = val;
171 	else if (attr == &ttm_page_pool_alloc_size) {
172 		if (val > NUM_PAGES_TO_ALLOC*8) {
173 			pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
174 			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
175 			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
176 			return size;
177 		} else if (val > NUM_PAGES_TO_ALLOC) {
178 			pr_warn("Setting allocation size to larger than %lu is not recommended\n",
179 				NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
180 		}
181 		m->options.alloc_size = val;
182 	}
183 
184 	return size;
185 }
186 
187 static ssize_t ttm_pool_show(struct kobject *kobj,
188 		struct attribute *attr, char *buffer)
189 {
190 	struct ttm_pool_manager *m =
191 		container_of(kobj, struct ttm_pool_manager, kobj);
192 	unsigned val = 0;
193 
194 	if (attr == &ttm_page_pool_max)
195 		val = m->options.max_size;
196 	else if (attr == &ttm_page_pool_small)
197 		val = m->options.small;
198 	else if (attr == &ttm_page_pool_alloc_size)
199 		val = m->options.alloc_size;
200 
201 	val = val * (PAGE_SIZE >> 10);
202 
203 	return snprintf(buffer, PAGE_SIZE, "%u\n", val);
204 }
205 
206 static const struct sysfs_ops ttm_pool_sysfs_ops = {
207 	.show = &ttm_pool_show,
208 	.store = &ttm_pool_store,
209 };
210 
211 static struct kobj_type ttm_pool_kobj_type = {
212 	.release = &ttm_pool_kobj_release,
213 	.sysfs_ops = &ttm_pool_sysfs_ops,
214 	.default_attrs = ttm_pool_attrs,
215 };
216 
217 static struct ttm_pool_manager *_manager;
218 
219 #ifndef CONFIG_X86
220 static int set_pages_array_wb(struct page **pages, int addrinarray)
221 {
222 #ifdef TTM_HAS_AGP
223 	int i;
224 
225 	for (i = 0; i < addrinarray; i++)
226 		unmap_page_from_agp(pages[i]);
227 #endif
228 	return 0;
229 }
230 
231 static int set_pages_array_wc(struct page **pages, int addrinarray)
232 {
233 #ifdef TTM_HAS_AGP
234 	int i;
235 
236 	for (i = 0; i < addrinarray; i++)
237 		map_page_into_agp(pages[i]);
238 #endif
239 	return 0;
240 }
241 
242 static int set_pages_array_uc(struct page **pages, int addrinarray)
243 {
244 #ifdef TTM_HAS_AGP
245 	int i;
246 
247 	for (i = 0; i < addrinarray; i++)
248 		map_page_into_agp(pages[i]);
249 #endif
250 	return 0;
251 }
252 #endif
253 
254 /**
255  * Select the right pool or requested caching state and ttm flags. */
256 static struct ttm_page_pool *ttm_get_pool(int flags,
257 		enum ttm_caching_state cstate)
258 {
259 	int pool_index;
260 
261 	if (cstate == tt_cached)
262 		return NULL;
263 
264 	if (cstate == tt_wc)
265 		pool_index = 0x0;
266 	else
267 		pool_index = 0x1;
268 
269 	if (flags & TTM_PAGE_FLAG_DMA32)
270 		pool_index |= 0x2;
271 
272 	return &_manager->pools[pool_index];
273 }
274 
275 /* set memory back to wb and free the pages. */
276 static void ttm_pages_put(struct page *pages[], unsigned npages)
277 {
278 	unsigned i;
279 	if (set_pages_array_wb(pages, npages))
280 		pr_err("Failed to set %d pages to wb!\n", npages);
281 	for (i = 0; i < npages; ++i)
282 		__free_page(pages[i]);
283 }
284 
285 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
286 		unsigned freed_pages)
287 {
288 	pool->npages -= freed_pages;
289 	pool->nfrees += freed_pages;
290 }
291 
292 /**
293  * Free pages from pool.
294  *
295  * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
296  * number of pages in one go.
297  *
298  * @pool: to free the pages from
299  * @free_all: If set to true will free all pages in pool
300  **/
301 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free)
302 {
303 	unsigned long irq_flags;
304 	struct page *p;
305 	struct page **pages_to_free;
306 	unsigned freed_pages = 0,
307 		 npages_to_free = nr_free;
308 
309 	if (NUM_PAGES_TO_ALLOC < nr_free)
310 		npages_to_free = NUM_PAGES_TO_ALLOC;
311 
312 	pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
313 			GFP_KERNEL);
314 	if (!pages_to_free) {
315 		pr_err("Failed to allocate memory for pool free operation\n");
316 		return 0;
317 	}
318 
319 restart:
320 	spin_lock_irqsave(&pool->lock, irq_flags);
321 
322 	list_for_each_entry_reverse(p, &pool->list, lru) {
323 		if (freed_pages >= npages_to_free)
324 			break;
325 
326 		pages_to_free[freed_pages++] = p;
327 		/* We can only remove NUM_PAGES_TO_ALLOC at a time. */
328 		if (freed_pages >= NUM_PAGES_TO_ALLOC) {
329 			/* remove range of pages from the pool */
330 			__list_del(p->lru.prev, &pool->list);
331 
332 			ttm_pool_update_free_locked(pool, freed_pages);
333 			/**
334 			 * Because changing page caching is costly
335 			 * we unlock the pool to prevent stalling.
336 			 */
337 			spin_unlock_irqrestore(&pool->lock, irq_flags);
338 
339 			ttm_pages_put(pages_to_free, freed_pages);
340 			if (likely(nr_free != FREE_ALL_PAGES))
341 				nr_free -= freed_pages;
342 
343 			if (NUM_PAGES_TO_ALLOC >= nr_free)
344 				npages_to_free = nr_free;
345 			else
346 				npages_to_free = NUM_PAGES_TO_ALLOC;
347 
348 			freed_pages = 0;
349 
350 			/* free all so restart the processing */
351 			if (nr_free)
352 				goto restart;
353 
354 			/* Not allowed to fall through or break because
355 			 * following context is inside spinlock while we are
356 			 * outside here.
357 			 */
358 			goto out;
359 
360 		}
361 	}
362 
363 	/* remove range of pages from the pool */
364 	if (freed_pages) {
365 		__list_del(&p->lru, &pool->list);
366 
367 		ttm_pool_update_free_locked(pool, freed_pages);
368 		nr_free -= freed_pages;
369 	}
370 
371 	spin_unlock_irqrestore(&pool->lock, irq_flags);
372 
373 	if (freed_pages)
374 		ttm_pages_put(pages_to_free, freed_pages);
375 out:
376 	kfree(pages_to_free);
377 	return nr_free;
378 }
379 
380 /**
381  * Callback for mm to request pool to reduce number of page held.
382  *
383  * XXX: (dchinner) Deadlock warning!
384  *
385  * ttm_page_pool_free() does memory allocation using GFP_KERNEL.  that means
386  * this can deadlock when called a sc->gfp_mask that is not equal to
387  * GFP_KERNEL.
388  *
389  * This code is crying out for a shrinker per pool....
390  */
391 static unsigned long
392 ttm_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
393 {
394 	static atomic_t start_pool = ATOMIC_INIT(0);
395 	unsigned i;
396 	unsigned pool_offset = atomic_add_return(1, &start_pool);
397 	struct ttm_page_pool *pool;
398 	int shrink_pages = sc->nr_to_scan;
399 	unsigned long freed = 0;
400 
401 	pool_offset = pool_offset % NUM_POOLS;
402 	/* select start pool in round robin fashion */
403 	for (i = 0; i < NUM_POOLS; ++i) {
404 		unsigned nr_free = shrink_pages;
405 		if (shrink_pages == 0)
406 			break;
407 		pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
408 		shrink_pages = ttm_page_pool_free(pool, nr_free);
409 		freed += nr_free - shrink_pages;
410 	}
411 	return freed;
412 }
413 
414 
415 static unsigned long
416 ttm_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
417 {
418 	unsigned i;
419 	unsigned long count = 0;
420 
421 	for (i = 0; i < NUM_POOLS; ++i)
422 		count += _manager->pools[i].npages;
423 
424 	return count;
425 }
426 
427 static void ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
428 {
429 	manager->mm_shrink.count_objects = ttm_pool_shrink_count;
430 	manager->mm_shrink.scan_objects = ttm_pool_shrink_scan;
431 	manager->mm_shrink.seeks = 1;
432 	register_shrinker(&manager->mm_shrink);
433 }
434 
435 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
436 {
437 	unregister_shrinker(&manager->mm_shrink);
438 }
439 
440 static int ttm_set_pages_caching(struct page **pages,
441 		enum ttm_caching_state cstate, unsigned cpages)
442 {
443 	int r = 0;
444 	/* Set page caching */
445 	switch (cstate) {
446 	case tt_uncached:
447 		r = set_pages_array_uc(pages, cpages);
448 		if (r)
449 			pr_err("Failed to set %d pages to uc!\n", cpages);
450 		break;
451 	case tt_wc:
452 		r = set_pages_array_wc(pages, cpages);
453 		if (r)
454 			pr_err("Failed to set %d pages to wc!\n", cpages);
455 		break;
456 	default:
457 		break;
458 	}
459 	return r;
460 }
461 
462 /**
463  * Free pages the pages that failed to change the caching state. If there is
464  * any pages that have changed their caching state already put them to the
465  * pool.
466  */
467 static void ttm_handle_caching_state_failure(struct list_head *pages,
468 		int ttm_flags, enum ttm_caching_state cstate,
469 		struct page **failed_pages, unsigned cpages)
470 {
471 	unsigned i;
472 	/* Failed pages have to be freed */
473 	for (i = 0; i < cpages; ++i) {
474 		list_del(&failed_pages[i]->lru);
475 		__free_page(failed_pages[i]);
476 	}
477 }
478 
479 /**
480  * Allocate new pages with correct caching.
481  *
482  * This function is reentrant if caller updates count depending on number of
483  * pages returned in pages array.
484  */
485 static int ttm_alloc_new_pages(struct list_head *pages, gfp_t gfp_flags,
486 		int ttm_flags, enum ttm_caching_state cstate, unsigned count)
487 {
488 	struct page **caching_array;
489 	struct page *p;
490 	int r = 0;
491 	unsigned i, cpages;
492 	unsigned max_cpages = min(count,
493 			(unsigned)(PAGE_SIZE/sizeof(struct page *)));
494 
495 	/* allocate array for page caching change */
496 	caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
497 
498 	if (!caching_array) {
499 		pr_err("Unable to allocate table for new pages\n");
500 		return -ENOMEM;
501 	}
502 
503 	for (i = 0, cpages = 0; i < count; ++i) {
504 		p = alloc_page(gfp_flags);
505 
506 		if (!p) {
507 			pr_err("Unable to get page %u\n", i);
508 
509 			/* store already allocated pages in the pool after
510 			 * setting the caching state */
511 			if (cpages) {
512 				r = ttm_set_pages_caching(caching_array,
513 							  cstate, cpages);
514 				if (r)
515 					ttm_handle_caching_state_failure(pages,
516 						ttm_flags, cstate,
517 						caching_array, cpages);
518 			}
519 			r = -ENOMEM;
520 			goto out;
521 		}
522 
523 #ifdef CONFIG_HIGHMEM
524 		/* gfp flags of highmem page should never be dma32 so we
525 		 * we should be fine in such case
526 		 */
527 		if (!PageHighMem(p))
528 #endif
529 		{
530 			caching_array[cpages++] = p;
531 			if (cpages == max_cpages) {
532 
533 				r = ttm_set_pages_caching(caching_array,
534 						cstate, cpages);
535 				if (r) {
536 					ttm_handle_caching_state_failure(pages,
537 						ttm_flags, cstate,
538 						caching_array, cpages);
539 					goto out;
540 				}
541 				cpages = 0;
542 			}
543 		}
544 
545 		list_add(&p->lru, pages);
546 	}
547 
548 	if (cpages) {
549 		r = ttm_set_pages_caching(caching_array, cstate, cpages);
550 		if (r)
551 			ttm_handle_caching_state_failure(pages,
552 					ttm_flags, cstate,
553 					caching_array, cpages);
554 	}
555 out:
556 	kfree(caching_array);
557 
558 	return r;
559 }
560 
561 /**
562  * Fill the given pool if there aren't enough pages and the requested number of
563  * pages is small.
564  */
565 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool,
566 		int ttm_flags, enum ttm_caching_state cstate, unsigned count,
567 		unsigned long *irq_flags)
568 {
569 	struct page *p;
570 	int r;
571 	unsigned cpages = 0;
572 	/**
573 	 * Only allow one pool fill operation at a time.
574 	 * If pool doesn't have enough pages for the allocation new pages are
575 	 * allocated from outside of pool.
576 	 */
577 	if (pool->fill_lock)
578 		return;
579 
580 	pool->fill_lock = true;
581 
582 	/* If allocation request is small and there are not enough
583 	 * pages in a pool we fill the pool up first. */
584 	if (count < _manager->options.small
585 		&& count > pool->npages) {
586 		struct list_head new_pages;
587 		unsigned alloc_size = _manager->options.alloc_size;
588 
589 		/**
590 		 * Can't change page caching if in irqsave context. We have to
591 		 * drop the pool->lock.
592 		 */
593 		spin_unlock_irqrestore(&pool->lock, *irq_flags);
594 
595 		INIT_LIST_HEAD(&new_pages);
596 		r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
597 				cstate,	alloc_size);
598 		spin_lock_irqsave(&pool->lock, *irq_flags);
599 
600 		if (!r) {
601 			list_splice(&new_pages, &pool->list);
602 			++pool->nrefills;
603 			pool->npages += alloc_size;
604 		} else {
605 			pr_err("Failed to fill pool (%p)\n", pool);
606 			/* If we have any pages left put them to the pool. */
607 			list_for_each_entry(p, &pool->list, lru) {
608 				++cpages;
609 			}
610 			list_splice(&new_pages, &pool->list);
611 			pool->npages += cpages;
612 		}
613 
614 	}
615 	pool->fill_lock = false;
616 }
617 
618 /**
619  * Cut 'count' number of pages from the pool and put them on the return list.
620  *
621  * @return count of pages still required to fulfill the request.
622  */
623 static unsigned ttm_page_pool_get_pages(struct ttm_page_pool *pool,
624 					struct list_head *pages,
625 					int ttm_flags,
626 					enum ttm_caching_state cstate,
627 					unsigned count)
628 {
629 	unsigned long irq_flags;
630 	struct list_head *p;
631 	unsigned i;
632 
633 	spin_lock_irqsave(&pool->lock, irq_flags);
634 	ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count, &irq_flags);
635 
636 	if (count >= pool->npages) {
637 		/* take all pages from the pool */
638 		list_splice_init(&pool->list, pages);
639 		count -= pool->npages;
640 		pool->npages = 0;
641 		goto out;
642 	}
643 	/* find the last pages to include for requested number of pages. Split
644 	 * pool to begin and halve it to reduce search space. */
645 	if (count <= pool->npages/2) {
646 		i = 0;
647 		list_for_each(p, &pool->list) {
648 			if (++i == count)
649 				break;
650 		}
651 	} else {
652 		i = pool->npages + 1;
653 		list_for_each_prev(p, &pool->list) {
654 			if (--i == count)
655 				break;
656 		}
657 	}
658 	/* Cut 'count' number of pages from the pool */
659 	list_cut_position(pages, &pool->list, p);
660 	pool->npages -= count;
661 	count = 0;
662 out:
663 	spin_unlock_irqrestore(&pool->lock, irq_flags);
664 	return count;
665 }
666 
667 /* Put all pages in pages list to correct pool to wait for reuse */
668 static void ttm_put_pages(struct page **pages, unsigned npages, int flags,
669 			  enum ttm_caching_state cstate)
670 {
671 	unsigned long irq_flags;
672 	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
673 	unsigned i;
674 
675 	if (pool == NULL) {
676 		/* No pool for this memory type so free the pages */
677 		for (i = 0; i < npages; i++) {
678 			if (pages[i]) {
679 				if (page_count(pages[i]) != 1)
680 					pr_err("Erroneous page count. Leaking pages.\n");
681 				__free_page(pages[i]);
682 				pages[i] = NULL;
683 			}
684 		}
685 		return;
686 	}
687 
688 	spin_lock_irqsave(&pool->lock, irq_flags);
689 	for (i = 0; i < npages; i++) {
690 		if (pages[i]) {
691 			if (page_count(pages[i]) != 1)
692 				pr_err("Erroneous page count. Leaking pages.\n");
693 			list_add_tail(&pages[i]->lru, &pool->list);
694 			pages[i] = NULL;
695 			pool->npages++;
696 		}
697 	}
698 	/* Check that we don't go over the pool limit */
699 	npages = 0;
700 	if (pool->npages > _manager->options.max_size) {
701 		npages = pool->npages - _manager->options.max_size;
702 		/* free at least NUM_PAGES_TO_ALLOC number of pages
703 		 * to reduce calls to set_memory_wb */
704 		if (npages < NUM_PAGES_TO_ALLOC)
705 			npages = NUM_PAGES_TO_ALLOC;
706 	}
707 	spin_unlock_irqrestore(&pool->lock, irq_flags);
708 	if (npages)
709 		ttm_page_pool_free(pool, npages);
710 }
711 
712 /*
713  * On success pages list will hold count number of correctly
714  * cached pages.
715  */
716 static int ttm_get_pages(struct page **pages, unsigned npages, int flags,
717 			 enum ttm_caching_state cstate)
718 {
719 	struct ttm_page_pool *pool = ttm_get_pool(flags, cstate);
720 	struct list_head plist;
721 	struct page *p = NULL;
722 	gfp_t gfp_flags = GFP_USER;
723 	unsigned count;
724 	int r;
725 
726 	/* set zero flag for page allocation if required */
727 	if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
728 		gfp_flags |= __GFP_ZERO;
729 
730 	/* No pool for cached pages */
731 	if (pool == NULL) {
732 		if (flags & TTM_PAGE_FLAG_DMA32)
733 			gfp_flags |= GFP_DMA32;
734 		else
735 			gfp_flags |= GFP_HIGHUSER;
736 
737 		for (r = 0; r < npages; ++r) {
738 			p = alloc_page(gfp_flags);
739 			if (!p) {
740 
741 				pr_err("Unable to allocate page\n");
742 				return -ENOMEM;
743 			}
744 
745 			pages[r] = p;
746 		}
747 		return 0;
748 	}
749 
750 	/* combine zero flag to pool flags */
751 	gfp_flags |= pool->gfp_flags;
752 
753 	/* First we take pages from the pool */
754 	INIT_LIST_HEAD(&plist);
755 	npages = ttm_page_pool_get_pages(pool, &plist, flags, cstate, npages);
756 	count = 0;
757 	list_for_each_entry(p, &plist, lru) {
758 		pages[count++] = p;
759 	}
760 
761 	/* clear the pages coming from the pool if requested */
762 	if (flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
763 		list_for_each_entry(p, &plist, lru) {
764 			if (PageHighMem(p))
765 				clear_highpage(p);
766 			else
767 				clear_page(page_address(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 		INIT_LIST_HEAD(&plist);
777 		r = ttm_alloc_new_pages(&plist, gfp_flags, flags, cstate, npages);
778 		list_for_each_entry(p, &plist, lru) {
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 			pr_err("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 	spin_lock_init(&pool->lock);
797 	pool->fill_lock = false;
798 	INIT_LIST_HEAD(&pool->list);
799 	pool->npages = pool->nfrees = 0;
800 	pool->gfp_flags = flags;
801 	pool->name = name;
802 }
803 
804 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
805 {
806 	int ret;
807 
808 	WARN_ON(_manager);
809 
810 	pr_info("Initializing pool allocator\n");
811 
812 	_manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
813 
814 	ttm_page_pool_init_locked(&_manager->wc_pool, GFP_HIGHUSER, "wc");
815 
816 	ttm_page_pool_init_locked(&_manager->uc_pool, GFP_HIGHUSER, "uc");
817 
818 	ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
819 				  GFP_USER | GFP_DMA32, "wc dma");
820 
821 	ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
822 				  GFP_USER | GFP_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 	ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
829 				   &glob->kobj, "pool");
830 	if (unlikely(ret != 0)) {
831 		kobject_put(&_manager->kobj);
832 		_manager = NULL;
833 		return ret;
834 	}
835 
836 	ttm_pool_mm_shrink_init(_manager);
837 
838 	return 0;
839 }
840 
841 void ttm_page_alloc_fini(void)
842 {
843 	int i;
844 
845 	pr_info("Finalizing pool allocator\n");
846 	ttm_pool_mm_shrink_fini(_manager);
847 
848 	for (i = 0; i < NUM_POOLS; ++i)
849 		ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES);
850 
851 	kobject_put(&_manager->kobj);
852 	_manager = NULL;
853 }
854 
855 int ttm_pool_populate(struct ttm_tt *ttm)
856 {
857 	struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
858 	unsigned i;
859 	int ret;
860 
861 	if (ttm->state != tt_unpopulated)
862 		return 0;
863 
864 	for (i = 0; i < ttm->num_pages; ++i) {
865 		ret = ttm_get_pages(&ttm->pages[i], 1,
866 				    ttm->page_flags,
867 				    ttm->caching_state);
868 		if (ret != 0) {
869 			ttm_pool_unpopulate(ttm);
870 			return -ENOMEM;
871 		}
872 
873 		ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
874 						false, false);
875 		if (unlikely(ret != 0)) {
876 			ttm_pool_unpopulate(ttm);
877 			return -ENOMEM;
878 		}
879 	}
880 
881 	if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
882 		ret = ttm_tt_swapin(ttm);
883 		if (unlikely(ret != 0)) {
884 			ttm_pool_unpopulate(ttm);
885 			return ret;
886 		}
887 	}
888 
889 	ttm->state = tt_unbound;
890 	return 0;
891 }
892 EXPORT_SYMBOL(ttm_pool_populate);
893 
894 void ttm_pool_unpopulate(struct ttm_tt *ttm)
895 {
896 	unsigned i;
897 
898 	for (i = 0; i < ttm->num_pages; ++i) {
899 		if (ttm->pages[i]) {
900 			ttm_mem_global_free_page(ttm->glob->mem_glob,
901 						 ttm->pages[i]);
902 			ttm_put_pages(&ttm->pages[i], 1,
903 				      ttm->page_flags,
904 				      ttm->caching_state);
905 		}
906 	}
907 	ttm->state = tt_unpopulated;
908 }
909 EXPORT_SYMBOL(ttm_pool_unpopulate);
910 
911 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
912 {
913 	struct ttm_page_pool *p;
914 	unsigned i;
915 	char *h[] = {"pool", "refills", "pages freed", "size"};
916 	if (!_manager) {
917 		seq_printf(m, "No pool allocator running.\n");
918 		return 0;
919 	}
920 	seq_printf(m, "%6s %12s %13s %8s\n",
921 			h[0], h[1], h[2], h[3]);
922 	for (i = 0; i < NUM_POOLS; ++i) {
923 		p = &_manager->pools[i];
924 
925 		seq_printf(m, "%6s %12ld %13ld %8d\n",
926 				p->name, p->nrefills,
927 				p->nfrees, p->npages);
928 	}
929 	return 0;
930 }
931 EXPORT_SYMBOL(ttm_page_alloc_debugfs);
932