1 /* $NetBSD: ttm_page_alloc.c,v 1.3 2021/12/18 23:45:44 riastradh Exp $ */
2
3 /*
4 * Copyright (c) Red Hat Inc.
5
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sub license,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the
14 * next paragraph) shall be included in all copies or substantial portions
15 * of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
22 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
23 * DEALINGS IN THE SOFTWARE.
24 *
25 * Authors: Dave Airlie <airlied@redhat.com>
26 * Jerome Glisse <jglisse@redhat.com>
27 * Pauli Nieminen <suokkos@gmail.com>
28 */
29
30 /* simple list based uncached page pool
31 * - Pool collects resently freed pages for reuse
32 * - Use page->lru to keep a free list
33 * - doesn't track currently in use pages
34 */
35
36 #include <sys/cdefs.h>
37 __KERNEL_RCSID(0, "$NetBSD: ttm_page_alloc.c,v 1.3 2021/12/18 23:45:44 riastradh Exp $");
38
39 #define pr_fmt(fmt) "[TTM] " fmt
40
41 #include <linux/list.h>
42 #include <linux/spinlock.h>
43 #include <linux/highmem.h>
44 #include <linux/mm_types.h>
45 #include <linux/module.h>
46 #include <linux/mm.h>
47 #include <linux/seq_file.h> /* for seq_printf */
48 #include <linux/slab.h>
49 #include <linux/dma-mapping.h>
50
51 #include <linux/atomic.h>
52
53 #include <drm/ttm/ttm_bo_driver.h>
54 #include <drm/ttm/ttm_page_alloc.h>
55 #include <drm/ttm/ttm_set_memory.h>
56
57 #define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *))
58 #define SMALL_ALLOCATION 16
59 #define FREE_ALL_PAGES (~0U)
60 /* times are in msecs */
61 #define PAGE_FREE_INTERVAL 1000
62
63 /**
64 * struct ttm_page_pool - Pool to reuse recently allocated uc/wc pages.
65 *
66 * @lock: Protects the shared pool from concurrnet access. Must be used with
67 * irqsave/irqrestore variants because pool allocator maybe called from
68 * delayed work.
69 * @fill_lock: Prevent concurrent calls to fill.
70 * @list: Pool of free uc/wc pages for fast reuse.
71 * @gfp_flags: Flags to pass for alloc_page.
72 * @npages: Number of pages in pool.
73 */
74 struct ttm_page_pool {
75 spinlock_t lock;
76 bool fill_lock;
77 struct list_head list;
78 gfp_t gfp_flags;
79 unsigned npages;
80 char *name;
81 unsigned long nfrees;
82 unsigned long nrefills;
83 unsigned int order;
84 };
85
86 /**
87 * Limits for the pool. They are handled without locks because only place where
88 * they may change is in sysfs store. They won't have immediate effect anyway
89 * so forcing serialization to access them is pointless.
90 */
91
92 struct ttm_pool_opts {
93 unsigned alloc_size;
94 unsigned max_size;
95 unsigned small;
96 };
97
98 #define NUM_POOLS 6
99
100 /**
101 * struct ttm_pool_manager - Holds memory pools for fst allocation
102 *
103 * Manager is read only object for pool code so it doesn't need locking.
104 *
105 * @free_interval: minimum number of jiffies between freeing pages from pool.
106 * @page_alloc_inited: reference counting for pool allocation.
107 * @work: Work that is used to shrink the pool. Work is only run when there is
108 * some pages to free.
109 * @small_allocation: Limit in number of pages what is small allocation.
110 *
111 * @pools: All pool objects in use.
112 **/
113 struct ttm_pool_manager {
114 struct kobject kobj;
115 struct shrinker mm_shrink;
116 struct ttm_pool_opts options;
117
118 union {
119 struct ttm_page_pool pools[NUM_POOLS];
120 struct {
121 struct ttm_page_pool wc_pool;
122 struct ttm_page_pool uc_pool;
123 struct ttm_page_pool wc_pool_dma32;
124 struct ttm_page_pool uc_pool_dma32;
125 struct ttm_page_pool wc_pool_huge;
126 struct ttm_page_pool uc_pool_huge;
127 } ;
128 };
129 };
130
131 static struct attribute ttm_page_pool_max = {
132 .name = "pool_max_size",
133 .mode = S_IRUGO | S_IWUSR
134 };
135 static struct attribute ttm_page_pool_small = {
136 .name = "pool_small_allocation",
137 .mode = S_IRUGO | S_IWUSR
138 };
139 static struct attribute ttm_page_pool_alloc_size = {
140 .name = "pool_allocation_size",
141 .mode = S_IRUGO | S_IWUSR
142 };
143
144 static struct attribute *ttm_pool_attrs[] = {
145 &ttm_page_pool_max,
146 &ttm_page_pool_small,
147 &ttm_page_pool_alloc_size,
148 NULL
149 };
150
ttm_pool_kobj_release(struct kobject * kobj)151 static void ttm_pool_kobj_release(struct kobject *kobj)
152 {
153 struct ttm_pool_manager *m =
154 container_of(kobj, struct ttm_pool_manager, kobj);
155 kfree(m);
156 }
157
ttm_pool_store(struct kobject * kobj,struct attribute * attr,const char * buffer,size_t size)158 static ssize_t ttm_pool_store(struct kobject *kobj,
159 struct attribute *attr, const char *buffer, size_t size)
160 {
161 struct ttm_pool_manager *m =
162 container_of(kobj, struct ttm_pool_manager, kobj);
163 int chars;
164 unsigned val;
165 chars = sscanf(buffer, "%u", &val);
166 if (chars == 0)
167 return size;
168
169 /* Convert kb to number of pages */
170 val = val / (PAGE_SIZE >> 10);
171
172 if (attr == &ttm_page_pool_max)
173 m->options.max_size = val;
174 else if (attr == &ttm_page_pool_small)
175 m->options.small = val;
176 else if (attr == &ttm_page_pool_alloc_size) {
177 if (val > NUM_PAGES_TO_ALLOC*8) {
178 pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
179 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
180 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
181 return size;
182 } else if (val > NUM_PAGES_TO_ALLOC) {
183 pr_warn("Setting allocation size to larger than %lu is not recommended\n",
184 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
185 }
186 m->options.alloc_size = val;
187 }
188
189 return size;
190 }
191
ttm_pool_show(struct kobject * kobj,struct attribute * attr,char * buffer)192 static ssize_t ttm_pool_show(struct kobject *kobj,
193 struct attribute *attr, char *buffer)
194 {
195 struct ttm_pool_manager *m =
196 container_of(kobj, struct ttm_pool_manager, kobj);
197 unsigned val = 0;
198
199 if (attr == &ttm_page_pool_max)
200 val = m->options.max_size;
201 else if (attr == &ttm_page_pool_small)
202 val = m->options.small;
203 else if (attr == &ttm_page_pool_alloc_size)
204 val = m->options.alloc_size;
205
206 val = val * (PAGE_SIZE >> 10);
207
208 return snprintf(buffer, PAGE_SIZE, "%u\n", val);
209 }
210
211 static const struct sysfs_ops ttm_pool_sysfs_ops = {
212 .show = &ttm_pool_show,
213 .store = &ttm_pool_store,
214 };
215
216 static struct kobj_type ttm_pool_kobj_type = {
217 .release = &ttm_pool_kobj_release,
218 .sysfs_ops = &ttm_pool_sysfs_ops,
219 .default_attrs = ttm_pool_attrs,
220 };
221
222 static struct ttm_pool_manager *_manager;
223
224 /**
225 * Select the right pool or requested caching state and ttm flags. */
ttm_get_pool(int flags,bool huge,enum ttm_caching_state cstate)226 static struct ttm_page_pool *ttm_get_pool(int flags, bool huge,
227 enum ttm_caching_state cstate)
228 {
229 int pool_index;
230
231 if (cstate == tt_cached)
232 return NULL;
233
234 if (cstate == tt_wc)
235 pool_index = 0x0;
236 else
237 pool_index = 0x1;
238
239 if (flags & TTM_PAGE_FLAG_DMA32) {
240 if (huge)
241 return NULL;
242 pool_index |= 0x2;
243
244 } else if (huge) {
245 pool_index |= 0x4;
246 }
247
248 return &_manager->pools[pool_index];
249 }
250
251 /* set memory back to wb and free the pages. */
ttm_pages_put(struct page * pages[],unsigned npages,unsigned int order)252 static void ttm_pages_put(struct page *pages[], unsigned npages,
253 unsigned int order)
254 {
255 unsigned int i, pages_nr = (1 << order);
256
257 if (order == 0) {
258 if (ttm_set_pages_array_wb(pages, npages))
259 pr_err("Failed to set %d pages to wb!\n", npages);
260 }
261
262 for (i = 0; i < npages; ++i) {
263 if (order > 0) {
264 if (ttm_set_pages_wb(pages[i], pages_nr))
265 pr_err("Failed to set %d pages to wb!\n", pages_nr);
266 }
267 __free_pages(pages[i], order);
268 }
269 }
270
ttm_pool_update_free_locked(struct ttm_page_pool * pool,unsigned freed_pages)271 static void ttm_pool_update_free_locked(struct ttm_page_pool *pool,
272 unsigned freed_pages)
273 {
274 pool->npages -= freed_pages;
275 pool->nfrees += freed_pages;
276 }
277
278 /**
279 * Free pages from pool.
280 *
281 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
282 * number of pages in one go.
283 *
284 * @pool: to free the pages from
285 * @free_all: If set to true will free all pages in pool
286 * @use_static: Safe to use static buffer
287 **/
ttm_page_pool_free(struct ttm_page_pool * pool,unsigned nr_free,bool use_static)288 static int ttm_page_pool_free(struct ttm_page_pool *pool, unsigned nr_free,
289 bool use_static)
290 {
291 static struct page *static_buf[NUM_PAGES_TO_ALLOC];
292 unsigned long irq_flags;
293 struct page *p;
294 struct page **pages_to_free;
295 unsigned freed_pages = 0,
296 npages_to_free = nr_free;
297
298 if (NUM_PAGES_TO_ALLOC < nr_free)
299 npages_to_free = NUM_PAGES_TO_ALLOC;
300
301 if (use_static)
302 pages_to_free = static_buf;
303 else
304 pages_to_free = kmalloc_array(npages_to_free,
305 sizeof(struct page *),
306 GFP_KERNEL);
307 if (!pages_to_free) {
308 pr_debug("Failed to allocate memory for pool free operation\n");
309 return 0;
310 }
311
312 restart:
313 spin_lock_irqsave(&pool->lock, irq_flags);
314
315 list_for_each_entry_reverse(p, &pool->list, lru) {
316 if (freed_pages >= npages_to_free)
317 break;
318
319 pages_to_free[freed_pages++] = p;
320 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
321 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
322 /* remove range of pages from the pool */
323 __list_del(p->lru.prev, &pool->list);
324
325 ttm_pool_update_free_locked(pool, freed_pages);
326 /**
327 * Because changing page caching is costly
328 * we unlock the pool to prevent stalling.
329 */
330 spin_unlock_irqrestore(&pool->lock, irq_flags);
331
332 ttm_pages_put(pages_to_free, freed_pages, pool->order);
333 if (likely(nr_free != FREE_ALL_PAGES))
334 nr_free -= freed_pages;
335
336 if (NUM_PAGES_TO_ALLOC >= nr_free)
337 npages_to_free = nr_free;
338 else
339 npages_to_free = NUM_PAGES_TO_ALLOC;
340
341 freed_pages = 0;
342
343 /* free all so restart the processing */
344 if (nr_free)
345 goto restart;
346
347 /* Not allowed to fall through or break because
348 * following context is inside spinlock while we are
349 * outside here.
350 */
351 goto out;
352
353 }
354 }
355
356 /* remove range of pages from the pool */
357 if (freed_pages) {
358 __list_del(&p->lru, &pool->list);
359
360 ttm_pool_update_free_locked(pool, freed_pages);
361 nr_free -= freed_pages;
362 }
363
364 spin_unlock_irqrestore(&pool->lock, irq_flags);
365
366 if (freed_pages)
367 ttm_pages_put(pages_to_free, freed_pages, pool->order);
368 out:
369 if (pages_to_free != static_buf)
370 kfree(pages_to_free);
371 return nr_free;
372 }
373
374 /**
375 * Callback for mm to request pool to reduce number of page held.
376 *
377 * XXX: (dchinner) Deadlock warning!
378 *
379 * This code is crying out for a shrinker per pool....
380 */
381 static unsigned long
ttm_pool_shrink_scan(struct shrinker * shrink,struct shrink_control * sc)382 ttm_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
383 {
384 static DEFINE_MUTEX(lock);
385 static unsigned start_pool;
386 unsigned i;
387 unsigned pool_offset;
388 struct ttm_page_pool *pool;
389 int shrink_pages = sc->nr_to_scan;
390 unsigned long freed = 0;
391 unsigned int nr_free_pool;
392
393 if (!mutex_trylock(&lock))
394 return SHRINK_STOP;
395 pool_offset = ++start_pool % NUM_POOLS;
396 /* select start pool in round robin fashion */
397 for (i = 0; i < NUM_POOLS; ++i) {
398 unsigned nr_free = shrink_pages;
399 unsigned page_nr;
400
401 if (shrink_pages == 0)
402 break;
403
404 pool = &_manager->pools[(i + pool_offset)%NUM_POOLS];
405 page_nr = (1 << pool->order);
406 /* OK to use static buffer since global mutex is held. */
407 nr_free_pool = roundup(nr_free, page_nr) >> pool->order;
408 shrink_pages = ttm_page_pool_free(pool, nr_free_pool, true);
409 freed += (nr_free_pool - shrink_pages) << pool->order;
410 if (freed >= sc->nr_to_scan)
411 break;
412 shrink_pages <<= pool->order;
413 }
414 mutex_unlock(&lock);
415 return freed;
416 }
417
418
419 static unsigned long
ttm_pool_shrink_count(struct shrinker * shrink,struct shrink_control * sc)420 ttm_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
421 {
422 unsigned i;
423 unsigned long count = 0;
424 struct ttm_page_pool *pool;
425
426 for (i = 0; i < NUM_POOLS; ++i) {
427 pool = &_manager->pools[i];
428 count += (pool->npages << pool->order);
429 }
430
431 return count;
432 }
433
ttm_pool_mm_shrink_init(struct ttm_pool_manager * manager)434 static int ttm_pool_mm_shrink_init(struct ttm_pool_manager *manager)
435 {
436 manager->mm_shrink.count_objects = ttm_pool_shrink_count;
437 manager->mm_shrink.scan_objects = ttm_pool_shrink_scan;
438 manager->mm_shrink.seeks = 1;
439 return register_shrinker(&manager->mm_shrink);
440 }
441
ttm_pool_mm_shrink_fini(struct ttm_pool_manager * manager)442 static void ttm_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
443 {
444 unregister_shrinker(&manager->mm_shrink);
445 }
446
ttm_set_pages_caching(struct page ** pages,enum ttm_caching_state cstate,unsigned cpages)447 static int ttm_set_pages_caching(struct page **pages,
448 enum ttm_caching_state cstate, unsigned cpages)
449 {
450 int r = 0;
451 /* Set page caching */
452 switch (cstate) {
453 case tt_uncached:
454 r = ttm_set_pages_array_uc(pages, cpages);
455 if (r)
456 pr_err("Failed to set %d pages to uc!\n", cpages);
457 break;
458 case tt_wc:
459 r = ttm_set_pages_array_wc(pages, cpages);
460 if (r)
461 pr_err("Failed to set %d pages to wc!\n", cpages);
462 break;
463 default:
464 break;
465 }
466 return r;
467 }
468
469 /**
470 * Free pages the pages that failed to change the caching state. If there is
471 * any pages that have changed their caching state already put them to the
472 * pool.
473 */
ttm_handle_caching_state_failure(struct list_head * pages,int ttm_flags,enum ttm_caching_state cstate,struct page ** failed_pages,unsigned cpages)474 static void ttm_handle_caching_state_failure(struct list_head *pages,
475 int ttm_flags, enum ttm_caching_state cstate,
476 struct page **failed_pages, unsigned cpages)
477 {
478 unsigned i;
479 /* Failed pages have to be freed */
480 for (i = 0; i < cpages; ++i) {
481 list_del(&failed_pages[i]->lru);
482 __free_page(failed_pages[i]);
483 }
484 }
485
486 /**
487 * Allocate new pages with correct caching.
488 *
489 * This function is reentrant if caller updates count depending on number of
490 * pages returned in pages array.
491 */
ttm_alloc_new_pages(struct list_head * pages,gfp_t gfp_flags,int ttm_flags,enum ttm_caching_state cstate,unsigned count,unsigned order)492 static int ttm_alloc_new_pages(struct list_head *pages, gfp_t gfp_flags,
493 int ttm_flags, enum ttm_caching_state cstate,
494 unsigned count, unsigned order)
495 {
496 struct page **caching_array;
497 struct page *p;
498 int r = 0;
499 unsigned i, j, cpages;
500 unsigned npages = 1 << order;
501 unsigned max_cpages = min(count << order, (unsigned)NUM_PAGES_TO_ALLOC);
502
503 /* allocate array for page caching change */
504 caching_array = kmalloc_array(max_cpages, sizeof(struct page *),
505 GFP_KERNEL);
506
507 if (!caching_array) {
508 pr_debug("Unable to allocate table for new pages\n");
509 return -ENOMEM;
510 }
511
512 for (i = 0, cpages = 0; i < count; ++i) {
513 p = alloc_pages(gfp_flags, order);
514
515 if (!p) {
516 pr_debug("Unable to get page %u\n", i);
517
518 /* store already allocated pages in the pool after
519 * setting the caching state */
520 if (cpages) {
521 r = ttm_set_pages_caching(caching_array,
522 cstate, cpages);
523 if (r)
524 ttm_handle_caching_state_failure(pages,
525 ttm_flags, cstate,
526 caching_array, cpages);
527 }
528 r = -ENOMEM;
529 goto out;
530 }
531
532 list_add(&p->lru, pages);
533
534 #ifdef CONFIG_HIGHMEM
535 /* gfp flags of highmem page should never be dma32 so we
536 * we should be fine in such case
537 */
538 if (PageHighMem(p))
539 continue;
540
541 #endif
542 for (j = 0; j < npages; ++j) {
543 caching_array[cpages++] = p++;
544 if (cpages == max_cpages) {
545
546 r = ttm_set_pages_caching(caching_array,
547 cstate, cpages);
548 if (r) {
549 ttm_handle_caching_state_failure(pages,
550 ttm_flags, cstate,
551 caching_array, cpages);
552 goto out;
553 }
554 cpages = 0;
555 }
556 }
557 }
558
559 if (cpages) {
560 r = ttm_set_pages_caching(caching_array, cstate, cpages);
561 if (r)
562 ttm_handle_caching_state_failure(pages,
563 ttm_flags, cstate,
564 caching_array, cpages);
565 }
566 out:
567 kfree(caching_array);
568
569 return r;
570 }
571
572 /**
573 * Fill the given pool if there aren't enough pages and the requested number of
574 * pages is small.
575 */
ttm_page_pool_fill_locked(struct ttm_page_pool * pool,int ttm_flags,enum ttm_caching_state cstate,unsigned count,unsigned long * irq_flags)576 static void ttm_page_pool_fill_locked(struct ttm_page_pool *pool, int ttm_flags,
577 enum ttm_caching_state cstate,
578 unsigned count, unsigned long *irq_flags)
579 {
580 struct page *p;
581 int r;
582 unsigned cpages = 0;
583 /**
584 * Only allow one pool fill operation at a time.
585 * If pool doesn't have enough pages for the allocation new pages are
586 * allocated from outside of pool.
587 */
588 if (pool->fill_lock)
589 return;
590
591 pool->fill_lock = true;
592
593 /* If allocation request is small and there are not enough
594 * pages in a pool we fill the pool up first. */
595 if (count < _manager->options.small
596 && count > pool->npages) {
597 struct list_head new_pages;
598 unsigned alloc_size = _manager->options.alloc_size;
599
600 /**
601 * Can't change page caching if in irqsave context. We have to
602 * drop the pool->lock.
603 */
604 spin_unlock_irqrestore(&pool->lock, *irq_flags);
605
606 INIT_LIST_HEAD(&new_pages);
607 r = ttm_alloc_new_pages(&new_pages, pool->gfp_flags, ttm_flags,
608 cstate, alloc_size, 0);
609 spin_lock_irqsave(&pool->lock, *irq_flags);
610
611 if (!r) {
612 list_splice(&new_pages, &pool->list);
613 ++pool->nrefills;
614 pool->npages += alloc_size;
615 } else {
616 pr_debug("Failed to fill pool (%p)\n", pool);
617 /* If we have any pages left put them to the pool. */
618 list_for_each_entry(p, &new_pages, lru) {
619 ++cpages;
620 }
621 list_splice(&new_pages, &pool->list);
622 pool->npages += cpages;
623 }
624
625 }
626 pool->fill_lock = false;
627 }
628
629 /**
630 * Allocate pages from the pool and put them on the return list.
631 *
632 * @return zero for success or negative error code.
633 */
ttm_page_pool_get_pages(struct ttm_page_pool * pool,struct list_head * pages,int ttm_flags,enum ttm_caching_state cstate,unsigned count,unsigned order)634 static int ttm_page_pool_get_pages(struct ttm_page_pool *pool,
635 struct list_head *pages,
636 int ttm_flags,
637 enum ttm_caching_state cstate,
638 unsigned count, unsigned order)
639 {
640 unsigned long irq_flags;
641 struct list_head *p;
642 unsigned i;
643 int r = 0;
644
645 spin_lock_irqsave(&pool->lock, irq_flags);
646 if (!order)
647 ttm_page_pool_fill_locked(pool, ttm_flags, cstate, count,
648 &irq_flags);
649
650 if (count >= pool->npages) {
651 /* take all pages from the pool */
652 list_splice_init(&pool->list, pages);
653 count -= pool->npages;
654 pool->npages = 0;
655 goto out;
656 }
657 /* find the last pages to include for requested number of pages. Split
658 * pool to begin and halve it to reduce search space. */
659 if (count <= pool->npages/2) {
660 i = 0;
661 list_for_each(p, &pool->list) {
662 if (++i == count)
663 break;
664 }
665 } else {
666 i = pool->npages + 1;
667 list_for_each_prev(p, &pool->list) {
668 if (--i == count)
669 break;
670 }
671 }
672 /* Cut 'count' number of pages from the pool */
673 list_cut_position(pages, &pool->list, p);
674 pool->npages -= count;
675 count = 0;
676 out:
677 spin_unlock_irqrestore(&pool->lock, irq_flags);
678
679 /* clear the pages coming from the pool if requested */
680 if (ttm_flags & TTM_PAGE_FLAG_ZERO_ALLOC) {
681 struct page *page;
682
683 list_for_each_entry(page, pages, lru) {
684 if (PageHighMem(page))
685 clear_highpage(page);
686 else
687 clear_page(page_address(page));
688 }
689 }
690
691 /* If pool didn't have enough pages allocate new one. */
692 if (count) {
693 gfp_t gfp_flags = pool->gfp_flags;
694
695 /* set zero flag for page allocation if required */
696 if (ttm_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
697 gfp_flags |= __GFP_ZERO;
698
699 if (ttm_flags & TTM_PAGE_FLAG_NO_RETRY)
700 gfp_flags |= __GFP_RETRY_MAYFAIL;
701
702 /* ttm_alloc_new_pages doesn't reference pool so we can run
703 * multiple requests in parallel.
704 **/
705 r = ttm_alloc_new_pages(pages, gfp_flags, ttm_flags, cstate,
706 count, order);
707 }
708
709 return r;
710 }
711
712 /* Put all pages in pages list to correct pool to wait for reuse */
ttm_put_pages(struct page ** pages,unsigned npages,int flags,enum ttm_caching_state cstate)713 static void ttm_put_pages(struct page **pages, unsigned npages, int flags,
714 enum ttm_caching_state cstate)
715 {
716 struct ttm_page_pool *pool = ttm_get_pool(flags, false, cstate);
717 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
718 struct ttm_page_pool *huge = ttm_get_pool(flags, true, cstate);
719 #endif
720 unsigned long irq_flags;
721 unsigned i;
722
723 if (pool == NULL) {
724 /* No pool for this memory type so free the pages */
725 i = 0;
726 while (i < npages) {
727 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
728 struct page *p = pages[i];
729 #endif
730 unsigned order = 0, j;
731
732 if (!pages[i]) {
733 ++i;
734 continue;
735 }
736
737 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
738 if (!(flags & TTM_PAGE_FLAG_DMA32) &&
739 (npages - i) >= HPAGE_PMD_NR) {
740 for (j = 1; j < HPAGE_PMD_NR; ++j)
741 if (++p != pages[i + j])
742 break;
743
744 if (j == HPAGE_PMD_NR)
745 order = HPAGE_PMD_ORDER;
746 }
747 #endif
748
749 if (page_count(pages[i]) != 1)
750 pr_err("Erroneous page count. Leaking pages.\n");
751 __free_pages(pages[i], order);
752
753 j = 1 << order;
754 while (j) {
755 pages[i++] = NULL;
756 --j;
757 }
758 }
759 return;
760 }
761
762 i = 0;
763 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
764 if (huge) {
765 unsigned max_size, n2free;
766
767 spin_lock_irqsave(&huge->lock, irq_flags);
768 while ((npages - i) >= HPAGE_PMD_NR) {
769 struct page *p = pages[i];
770 unsigned j;
771
772 if (!p)
773 break;
774
775 for (j = 1; j < HPAGE_PMD_NR; ++j)
776 if (++p != pages[i + j])
777 break;
778
779 if (j != HPAGE_PMD_NR)
780 break;
781
782 list_add_tail(&pages[i]->lru, &huge->list);
783
784 for (j = 0; j < HPAGE_PMD_NR; ++j)
785 pages[i++] = NULL;
786 huge->npages++;
787 }
788
789 /* Check that we don't go over the pool limit */
790 max_size = _manager->options.max_size;
791 max_size /= HPAGE_PMD_NR;
792 if (huge->npages > max_size)
793 n2free = huge->npages - max_size;
794 else
795 n2free = 0;
796 spin_unlock_irqrestore(&huge->lock, irq_flags);
797 if (n2free)
798 ttm_page_pool_free(huge, n2free, false);
799 }
800 #endif
801
802 spin_lock_irqsave(&pool->lock, irq_flags);
803 while (i < npages) {
804 if (pages[i]) {
805 if (page_count(pages[i]) != 1)
806 pr_err("Erroneous page count. Leaking pages.\n");
807 list_add_tail(&pages[i]->lru, &pool->list);
808 pages[i] = NULL;
809 pool->npages++;
810 }
811 ++i;
812 }
813 /* Check that we don't go over the pool limit */
814 npages = 0;
815 if (pool->npages > _manager->options.max_size) {
816 npages = pool->npages - _manager->options.max_size;
817 /* free at least NUM_PAGES_TO_ALLOC number of pages
818 * to reduce calls to set_memory_wb */
819 if (npages < NUM_PAGES_TO_ALLOC)
820 npages = NUM_PAGES_TO_ALLOC;
821 }
822 spin_unlock_irqrestore(&pool->lock, irq_flags);
823 if (npages)
824 ttm_page_pool_free(pool, npages, false);
825 }
826
827 /*
828 * On success pages list will hold count number of correctly
829 * cached pages.
830 */
ttm_get_pages(struct page ** pages,unsigned npages,int flags,enum ttm_caching_state cstate)831 static int ttm_get_pages(struct page **pages, unsigned npages, int flags,
832 enum ttm_caching_state cstate)
833 {
834 struct ttm_page_pool *pool = ttm_get_pool(flags, false, cstate);
835 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
836 struct ttm_page_pool *huge = ttm_get_pool(flags, true, cstate);
837 #endif
838 struct list_head plist;
839 struct page *p = NULL;
840 unsigned count, first;
841 int r;
842
843 /* No pool for cached pages */
844 if (pool == NULL) {
845 gfp_t gfp_flags = GFP_USER;
846 unsigned i;
847 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
848 unsigned j;
849 #endif
850
851 /* set zero flag for page allocation if required */
852 if (flags & TTM_PAGE_FLAG_ZERO_ALLOC)
853 gfp_flags |= __GFP_ZERO;
854
855 if (flags & TTM_PAGE_FLAG_NO_RETRY)
856 gfp_flags |= __GFP_RETRY_MAYFAIL;
857
858 if (flags & TTM_PAGE_FLAG_DMA32)
859 gfp_flags |= GFP_DMA32;
860 else
861 gfp_flags |= GFP_HIGHUSER;
862
863 i = 0;
864 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
865 if (!(gfp_flags & GFP_DMA32)) {
866 while (npages >= HPAGE_PMD_NR) {
867 gfp_t huge_flags = gfp_flags;
868
869 huge_flags |= GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
870 __GFP_KSWAPD_RECLAIM;
871 huge_flags &= ~__GFP_MOVABLE;
872 huge_flags &= ~__GFP_COMP;
873 p = alloc_pages(huge_flags, HPAGE_PMD_ORDER);
874 if (!p)
875 break;
876
877 for (j = 0; j < HPAGE_PMD_NR; ++j)
878 pages[i++] = p++;
879
880 npages -= HPAGE_PMD_NR;
881 }
882 }
883 #endif
884
885 first = i;
886 while (npages) {
887 p = alloc_page(gfp_flags);
888 if (!p) {
889 pr_debug("Unable to allocate page\n");
890 return -ENOMEM;
891 }
892
893 /* Swap the pages if we detect consecutive order */
894 if (i > first && pages[i - 1] == p - 1)
895 swap(p, pages[i - 1]);
896
897 pages[i++] = p;
898 --npages;
899 }
900 return 0;
901 }
902
903 count = 0;
904
905 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
906 if (huge && npages >= HPAGE_PMD_NR) {
907 INIT_LIST_HEAD(&plist);
908 ttm_page_pool_get_pages(huge, &plist, flags, cstate,
909 npages / HPAGE_PMD_NR,
910 HPAGE_PMD_ORDER);
911
912 list_for_each_entry(p, &plist, lru) {
913 unsigned j;
914
915 for (j = 0; j < HPAGE_PMD_NR; ++j)
916 pages[count++] = &p[j];
917 }
918 }
919 #endif
920
921 INIT_LIST_HEAD(&plist);
922 r = ttm_page_pool_get_pages(pool, &plist, flags, cstate,
923 npages - count, 0);
924
925 first = count;
926 list_for_each_entry(p, &plist, lru) {
927 struct page *tmp = p;
928
929 /* Swap the pages if we detect consecutive order */
930 if (count > first && pages[count - 1] == tmp - 1)
931 swap(tmp, pages[count - 1]);
932 pages[count++] = tmp;
933 }
934
935 if (r) {
936 /* If there is any pages in the list put them back to
937 * the pool.
938 */
939 pr_debug("Failed to allocate extra pages for large request\n");
940 ttm_put_pages(pages, count, flags, cstate);
941 return r;
942 }
943
944 return 0;
945 }
946
ttm_page_pool_init_locked(struct ttm_page_pool * pool,gfp_t flags,char * name,unsigned int order)947 static void ttm_page_pool_init_locked(struct ttm_page_pool *pool, gfp_t flags,
948 char *name, unsigned int order)
949 {
950 spin_lock_init(&pool->lock);
951 pool->fill_lock = false;
952 INIT_LIST_HEAD(&pool->list);
953 pool->npages = pool->nfrees = 0;
954 pool->gfp_flags = flags;
955 pool->name = name;
956 pool->order = order;
957 }
958
ttm_page_alloc_init(struct ttm_mem_global * glob,unsigned max_pages)959 int ttm_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
960 {
961 int ret;
962 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
963 unsigned order = HPAGE_PMD_ORDER;
964 #else
965 unsigned order = 0;
966 #endif
967
968 WARN_ON(_manager);
969
970 pr_info("Initializing pool allocator\n");
971
972 _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
973 if (!_manager)
974 return -ENOMEM;
975
976 ttm_page_pool_init_locked(&_manager->wc_pool, GFP_HIGHUSER, "wc", 0);
977
978 ttm_page_pool_init_locked(&_manager->uc_pool, GFP_HIGHUSER, "uc", 0);
979
980 ttm_page_pool_init_locked(&_manager->wc_pool_dma32,
981 GFP_USER | GFP_DMA32, "wc dma", 0);
982
983 ttm_page_pool_init_locked(&_manager->uc_pool_dma32,
984 GFP_USER | GFP_DMA32, "uc dma", 0);
985
986 ttm_page_pool_init_locked(&_manager->wc_pool_huge,
987 (GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
988 __GFP_KSWAPD_RECLAIM) &
989 ~(__GFP_MOVABLE | __GFP_COMP),
990 "wc huge", order);
991
992 ttm_page_pool_init_locked(&_manager->uc_pool_huge,
993 (GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
994 __GFP_KSWAPD_RECLAIM) &
995 ~(__GFP_MOVABLE | __GFP_COMP)
996 , "uc huge", order);
997
998 _manager->options.max_size = max_pages;
999 _manager->options.small = SMALL_ALLOCATION;
1000 _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1001
1002 ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1003 &glob->kobj, "pool");
1004 if (unlikely(ret != 0))
1005 goto error;
1006
1007 ret = ttm_pool_mm_shrink_init(_manager);
1008 if (unlikely(ret != 0))
1009 goto error;
1010 return 0;
1011
1012 error:
1013 kobject_put(&_manager->kobj);
1014 _manager = NULL;
1015 return ret;
1016 }
1017
ttm_page_alloc_fini(void)1018 void ttm_page_alloc_fini(void)
1019 {
1020 int i;
1021
1022 pr_info("Finalizing pool allocator\n");
1023 ttm_pool_mm_shrink_fini(_manager);
1024
1025 /* OK to use static buffer since global mutex is no longer used. */
1026 for (i = 0; i < NUM_POOLS; ++i)
1027 ttm_page_pool_free(&_manager->pools[i], FREE_ALL_PAGES, true);
1028
1029 kobject_put(&_manager->kobj);
1030 _manager = NULL;
1031 }
1032
1033 static void
ttm_pool_unpopulate_helper(struct ttm_tt * ttm,unsigned mem_count_update)1034 ttm_pool_unpopulate_helper(struct ttm_tt *ttm, unsigned mem_count_update)
1035 {
1036 struct ttm_mem_global *mem_glob = &ttm_mem_glob;
1037 unsigned i;
1038
1039 if (mem_count_update == 0)
1040 goto put_pages;
1041
1042 for (i = 0; i < mem_count_update; ++i) {
1043 if (!ttm->pages[i])
1044 continue;
1045
1046 ttm_mem_global_free_page(mem_glob, ttm->pages[i], PAGE_SIZE);
1047 }
1048
1049 put_pages:
1050 ttm_put_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
1051 ttm->caching_state);
1052 ttm->state = tt_unpopulated;
1053 }
1054
ttm_pool_populate(struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)1055 int ttm_pool_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1056 {
1057 struct ttm_mem_global *mem_glob = &ttm_mem_glob;
1058 unsigned i;
1059 int ret;
1060
1061 if (ttm->state != tt_unpopulated)
1062 return 0;
1063
1064 if (ttm_check_under_lowerlimit(mem_glob, ttm->num_pages, ctx))
1065 return -ENOMEM;
1066
1067 ret = ttm_get_pages(ttm->pages, ttm->num_pages, ttm->page_flags,
1068 ttm->caching_state);
1069 if (unlikely(ret != 0)) {
1070 ttm_pool_unpopulate_helper(ttm, 0);
1071 return ret;
1072 }
1073
1074 for (i = 0; i < ttm->num_pages; ++i) {
1075 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
1076 PAGE_SIZE, ctx);
1077 if (unlikely(ret != 0)) {
1078 ttm_pool_unpopulate_helper(ttm, i);
1079 return -ENOMEM;
1080 }
1081 }
1082
1083 if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
1084 ret = ttm_tt_swapin(ttm);
1085 if (unlikely(ret != 0)) {
1086 ttm_pool_unpopulate(ttm);
1087 return ret;
1088 }
1089 }
1090
1091 ttm->state = tt_unbound;
1092 return 0;
1093 }
1094 EXPORT_SYMBOL(ttm_pool_populate);
1095
ttm_pool_unpopulate(struct ttm_tt * ttm)1096 void ttm_pool_unpopulate(struct ttm_tt *ttm)
1097 {
1098 ttm_pool_unpopulate_helper(ttm, ttm->num_pages);
1099 }
1100 EXPORT_SYMBOL(ttm_pool_unpopulate);
1101
ttm_populate_and_map_pages(struct device * dev,struct ttm_dma_tt * tt,struct ttm_operation_ctx * ctx)1102 int ttm_populate_and_map_pages(struct device *dev, struct ttm_dma_tt *tt,
1103 struct ttm_operation_ctx *ctx)
1104 {
1105 unsigned i, j;
1106 int r;
1107
1108 r = ttm_pool_populate(&tt->ttm, ctx);
1109 if (r)
1110 return r;
1111
1112 for (i = 0; i < tt->ttm.num_pages; ++i) {
1113 struct page *p = tt->ttm.pages[i];
1114 size_t num_pages = 1;
1115
1116 for (j = i + 1; j < tt->ttm.num_pages; ++j) {
1117 if (++p != tt->ttm.pages[j])
1118 break;
1119
1120 ++num_pages;
1121 }
1122
1123 tt->dma_address[i] = dma_map_page(dev, tt->ttm.pages[i],
1124 0, num_pages * PAGE_SIZE,
1125 DMA_BIDIRECTIONAL);
1126 if (dma_mapping_error(dev, tt->dma_address[i])) {
1127 while (i--) {
1128 dma_unmap_page(dev, tt->dma_address[i],
1129 PAGE_SIZE, DMA_BIDIRECTIONAL);
1130 tt->dma_address[i] = 0;
1131 }
1132 ttm_pool_unpopulate(&tt->ttm);
1133 return -EFAULT;
1134 }
1135
1136 for (j = 1; j < num_pages; ++j) {
1137 tt->dma_address[i + 1] = tt->dma_address[i] + PAGE_SIZE;
1138 ++i;
1139 }
1140 }
1141 return 0;
1142 }
1143 EXPORT_SYMBOL(ttm_populate_and_map_pages);
1144
ttm_unmap_and_unpopulate_pages(struct device * dev,struct ttm_dma_tt * tt)1145 void ttm_unmap_and_unpopulate_pages(struct device *dev, struct ttm_dma_tt *tt)
1146 {
1147 unsigned i, j;
1148
1149 for (i = 0; i < tt->ttm.num_pages;) {
1150 struct page *p = tt->ttm.pages[i];
1151 size_t num_pages = 1;
1152
1153 if (!tt->dma_address[i] || !tt->ttm.pages[i]) {
1154 ++i;
1155 continue;
1156 }
1157
1158 for (j = i + 1; j < tt->ttm.num_pages; ++j) {
1159 if (++p != tt->ttm.pages[j])
1160 break;
1161
1162 ++num_pages;
1163 }
1164
1165 dma_unmap_page(dev, tt->dma_address[i], num_pages * PAGE_SIZE,
1166 DMA_BIDIRECTIONAL);
1167
1168 i += num_pages;
1169 }
1170 ttm_pool_unpopulate(&tt->ttm);
1171 }
1172 EXPORT_SYMBOL(ttm_unmap_and_unpopulate_pages);
1173
ttm_page_alloc_debugfs(struct seq_file * m,void * data)1174 int ttm_page_alloc_debugfs(struct seq_file *m, void *data)
1175 {
1176 struct ttm_page_pool *p;
1177 unsigned i;
1178 char *h[] = {"pool", "refills", "pages freed", "size"};
1179 if (!_manager) {
1180 seq_printf(m, "No pool allocator running.\n");
1181 return 0;
1182 }
1183 seq_printf(m, "%7s %12s %13s %8s\n",
1184 h[0], h[1], h[2], h[3]);
1185 for (i = 0; i < NUM_POOLS; ++i) {
1186 p = &_manager->pools[i];
1187
1188 seq_printf(m, "%7s %12ld %13ld %8d\n",
1189 p->name, p->nrefills,
1190 p->nfrees, p->npages);
1191 }
1192 return 0;
1193 }
1194 EXPORT_SYMBOL(ttm_page_alloc_debugfs);
1195