Loading...
1/*
2 * Copyright 2011 (c) Oracle Corp.
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 * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
24 */
25
26/*
27 * A simple DMA pool losely based on dmapool.c. It has certain advantages
28 * over the DMA pools:
29 * - Pool collects resently freed pages for reuse (and hooks up to
30 * the shrinker).
31 * - Tracks currently in use pages
32 * - Tracks whether the page is UC, WB or cached (and reverts to WB
33 * when freed).
34 */
35
36#if defined(CONFIG_SWIOTLB) || defined(CONFIG_INTEL_IOMMU)
37#define pr_fmt(fmt) "[TTM] " fmt
38
39#include <linux/dma-mapping.h>
40#include <linux/list.h>
41#include <linux/seq_file.h> /* for seq_printf */
42#include <linux/slab.h>
43#include <linux/spinlock.h>
44#include <linux/highmem.h>
45#include <linux/mm_types.h>
46#include <linux/module.h>
47#include <linux/mm.h>
48#include <linux/atomic.h>
49#include <linux/device.h>
50#include <linux/kthread.h>
51#include <drm/ttm/ttm_bo_driver.h>
52#include <drm/ttm/ttm_page_alloc.h>
53#ifdef TTM_HAS_AGP
54#include <asm/agp.h>
55#endif
56
57#define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *))
58#define SMALL_ALLOCATION 4
59#define FREE_ALL_PAGES (~0U)
60/* times are in msecs */
61#define IS_UNDEFINED (0)
62#define IS_WC (1<<1)
63#define IS_UC (1<<2)
64#define IS_CACHED (1<<3)
65#define IS_DMA32 (1<<4)
66
67enum pool_type {
68 POOL_IS_UNDEFINED,
69 POOL_IS_WC = IS_WC,
70 POOL_IS_UC = IS_UC,
71 POOL_IS_CACHED = IS_CACHED,
72 POOL_IS_WC_DMA32 = IS_WC | IS_DMA32,
73 POOL_IS_UC_DMA32 = IS_UC | IS_DMA32,
74 POOL_IS_CACHED_DMA32 = IS_CACHED | IS_DMA32,
75};
76/*
77 * The pool structure. There are usually six pools:
78 * - generic (not restricted to DMA32):
79 * - write combined, uncached, cached.
80 * - dma32 (up to 2^32 - so up 4GB):
81 * - write combined, uncached, cached.
82 * for each 'struct device'. The 'cached' is for pages that are actively used.
83 * The other ones can be shrunk by the shrinker API if neccessary.
84 * @pools: The 'struct device->dma_pools' link.
85 * @type: Type of the pool
86 * @lock: Protects the inuse_list and free_list from concurrnet access. Must be
87 * used with irqsave/irqrestore variants because pool allocator maybe called
88 * from delayed work.
89 * @inuse_list: Pool of pages that are in use. The order is very important and
90 * it is in the order that the TTM pages that are put back are in.
91 * @free_list: Pool of pages that are free to be used. No order requirements.
92 * @dev: The device that is associated with these pools.
93 * @size: Size used during DMA allocation.
94 * @npages_free: Count of available pages for re-use.
95 * @npages_in_use: Count of pages that are in use.
96 * @nfrees: Stats when pool is shrinking.
97 * @nrefills: Stats when the pool is grown.
98 * @gfp_flags: Flags to pass for alloc_page.
99 * @name: Name of the pool.
100 * @dev_name: Name derieved from dev - similar to how dev_info works.
101 * Used during shutdown as the dev_info during release is unavailable.
102 */
103struct dma_pool {
104 struct list_head pools; /* The 'struct device->dma_pools link */
105 enum pool_type type;
106 spinlock_t lock;
107 struct list_head inuse_list;
108 struct list_head free_list;
109 struct device *dev;
110 unsigned size;
111 unsigned npages_free;
112 unsigned npages_in_use;
113 unsigned long nfrees; /* Stats when shrunk. */
114 unsigned long nrefills; /* Stats when grown. */
115 gfp_t gfp_flags;
116 char name[13]; /* "cached dma32" */
117 char dev_name[64]; /* Constructed from dev */
118};
119
120/*
121 * The accounting page keeping track of the allocated page along with
122 * the DMA address.
123 * @page_list: The link to the 'page_list' in 'struct dma_pool'.
124 * @vaddr: The virtual address of the page
125 * @dma: The bus address of the page. If the page is not allocated
126 * via the DMA API, it will be -1.
127 */
128struct dma_page {
129 struct list_head page_list;
130 void *vaddr;
131 struct page *p;
132 dma_addr_t dma;
133};
134
135/*
136 * Limits for the pool. They are handled without locks because only place where
137 * they may change is in sysfs store. They won't have immediate effect anyway
138 * so forcing serialization to access them is pointless.
139 */
140
141struct ttm_pool_opts {
142 unsigned alloc_size;
143 unsigned max_size;
144 unsigned small;
145};
146
147/*
148 * Contains the list of all of the 'struct device' and their corresponding
149 * DMA pools. Guarded by _mutex->lock.
150 * @pools: The link to 'struct ttm_pool_manager->pools'
151 * @dev: The 'struct device' associated with the 'pool'
152 * @pool: The 'struct dma_pool' associated with the 'dev'
153 */
154struct device_pools {
155 struct list_head pools;
156 struct device *dev;
157 struct dma_pool *pool;
158};
159
160/*
161 * struct ttm_pool_manager - Holds memory pools for fast allocation
162 *
163 * @lock: Lock used when adding/removing from pools
164 * @pools: List of 'struct device' and 'struct dma_pool' tuples.
165 * @options: Limits for the pool.
166 * @npools: Total amount of pools in existence.
167 * @shrinker: The structure used by [un|]register_shrinker
168 */
169struct ttm_pool_manager {
170 struct mutex lock;
171 struct list_head pools;
172 struct ttm_pool_opts options;
173 unsigned npools;
174 struct shrinker mm_shrink;
175 struct kobject kobj;
176};
177
178static struct ttm_pool_manager *_manager;
179
180static struct attribute ttm_page_pool_max = {
181 .name = "pool_max_size",
182 .mode = S_IRUGO | S_IWUSR
183};
184static struct attribute ttm_page_pool_small = {
185 .name = "pool_small_allocation",
186 .mode = S_IRUGO | S_IWUSR
187};
188static struct attribute ttm_page_pool_alloc_size = {
189 .name = "pool_allocation_size",
190 .mode = S_IRUGO | S_IWUSR
191};
192
193static struct attribute *ttm_pool_attrs[] = {
194 &ttm_page_pool_max,
195 &ttm_page_pool_small,
196 &ttm_page_pool_alloc_size,
197 NULL
198};
199
200static void ttm_pool_kobj_release(struct kobject *kobj)
201{
202 struct ttm_pool_manager *m =
203 container_of(kobj, struct ttm_pool_manager, kobj);
204 kfree(m);
205}
206
207static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
208 const char *buffer, size_t size)
209{
210 struct ttm_pool_manager *m =
211 container_of(kobj, struct ttm_pool_manager, kobj);
212 int chars;
213 unsigned val;
214 chars = sscanf(buffer, "%u", &val);
215 if (chars == 0)
216 return size;
217
218 /* Convert kb to number of pages */
219 val = val / (PAGE_SIZE >> 10);
220
221 if (attr == &ttm_page_pool_max)
222 m->options.max_size = val;
223 else if (attr == &ttm_page_pool_small)
224 m->options.small = val;
225 else if (attr == &ttm_page_pool_alloc_size) {
226 if (val > NUM_PAGES_TO_ALLOC*8) {
227 pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
228 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
229 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
230 return size;
231 } else if (val > NUM_PAGES_TO_ALLOC) {
232 pr_warn("Setting allocation size to larger than %lu is not recommended\n",
233 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
234 }
235 m->options.alloc_size = val;
236 }
237
238 return size;
239}
240
241static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
242 char *buffer)
243{
244 struct ttm_pool_manager *m =
245 container_of(kobj, struct ttm_pool_manager, kobj);
246 unsigned val = 0;
247
248 if (attr == &ttm_page_pool_max)
249 val = m->options.max_size;
250 else if (attr == &ttm_page_pool_small)
251 val = m->options.small;
252 else if (attr == &ttm_page_pool_alloc_size)
253 val = m->options.alloc_size;
254
255 val = val * (PAGE_SIZE >> 10);
256
257 return snprintf(buffer, PAGE_SIZE, "%u\n", val);
258}
259
260static const struct sysfs_ops ttm_pool_sysfs_ops = {
261 .show = &ttm_pool_show,
262 .store = &ttm_pool_store,
263};
264
265static struct kobj_type ttm_pool_kobj_type = {
266 .release = &ttm_pool_kobj_release,
267 .sysfs_ops = &ttm_pool_sysfs_ops,
268 .default_attrs = ttm_pool_attrs,
269};
270
271#ifndef CONFIG_X86
272static int set_pages_array_wb(struct page **pages, int addrinarray)
273{
274#ifdef TTM_HAS_AGP
275 int i;
276
277 for (i = 0; i < addrinarray; i++)
278 unmap_page_from_agp(pages[i]);
279#endif
280 return 0;
281}
282
283static int set_pages_array_wc(struct page **pages, int addrinarray)
284{
285#ifdef TTM_HAS_AGP
286 int i;
287
288 for (i = 0; i < addrinarray; i++)
289 map_page_into_agp(pages[i]);
290#endif
291 return 0;
292}
293
294static int set_pages_array_uc(struct page **pages, int addrinarray)
295{
296#ifdef TTM_HAS_AGP
297 int i;
298
299 for (i = 0; i < addrinarray; i++)
300 map_page_into_agp(pages[i]);
301#endif
302 return 0;
303}
304#endif /* for !CONFIG_X86 */
305
306static int ttm_set_pages_caching(struct dma_pool *pool,
307 struct page **pages, unsigned cpages)
308{
309 int r = 0;
310 /* Set page caching */
311 if (pool->type & IS_UC) {
312 r = set_pages_array_uc(pages, cpages);
313 if (r)
314 pr_err("%s: Failed to set %d pages to uc!\n",
315 pool->dev_name, cpages);
316 }
317 if (pool->type & IS_WC) {
318 r = set_pages_array_wc(pages, cpages);
319 if (r)
320 pr_err("%s: Failed to set %d pages to wc!\n",
321 pool->dev_name, cpages);
322 }
323 return r;
324}
325
326static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
327{
328 dma_addr_t dma = d_page->dma;
329 dma_free_coherent(pool->dev, pool->size, d_page->vaddr, dma);
330
331 kfree(d_page);
332 d_page = NULL;
333}
334static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
335{
336 struct dma_page *d_page;
337
338 d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
339 if (!d_page)
340 return NULL;
341
342 d_page->vaddr = dma_alloc_coherent(pool->dev, pool->size,
343 &d_page->dma,
344 pool->gfp_flags);
345 if (d_page->vaddr) {
346 if (is_vmalloc_addr(d_page->vaddr))
347 d_page->p = vmalloc_to_page(d_page->vaddr);
348 else
349 d_page->p = virt_to_page(d_page->vaddr);
350 } else {
351 kfree(d_page);
352 d_page = NULL;
353 }
354 return d_page;
355}
356static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
357{
358 enum pool_type type = IS_UNDEFINED;
359
360 if (flags & TTM_PAGE_FLAG_DMA32)
361 type |= IS_DMA32;
362 if (cstate == tt_cached)
363 type |= IS_CACHED;
364 else if (cstate == tt_uncached)
365 type |= IS_UC;
366 else
367 type |= IS_WC;
368
369 return type;
370}
371
372static void ttm_pool_update_free_locked(struct dma_pool *pool,
373 unsigned freed_pages)
374{
375 pool->npages_free -= freed_pages;
376 pool->nfrees += freed_pages;
377
378}
379
380/* set memory back to wb and free the pages. */
381static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
382 struct page *pages[], unsigned npages)
383{
384 struct dma_page *d_page, *tmp;
385
386 /* Don't set WB on WB page pool. */
387 if (npages && !(pool->type & IS_CACHED) &&
388 set_pages_array_wb(pages, npages))
389 pr_err("%s: Failed to set %d pages to wb!\n",
390 pool->dev_name, npages);
391
392 list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
393 list_del(&d_page->page_list);
394 __ttm_dma_free_page(pool, d_page);
395 }
396}
397
398static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
399{
400 /* Don't set WB on WB page pool. */
401 if (!(pool->type & IS_CACHED) && set_pages_array_wb(&d_page->p, 1))
402 pr_err("%s: Failed to set %d pages to wb!\n",
403 pool->dev_name, 1);
404
405 list_del(&d_page->page_list);
406 __ttm_dma_free_page(pool, d_page);
407}
408
409/*
410 * Free pages from pool.
411 *
412 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
413 * number of pages in one go.
414 *
415 * @pool: to free the pages from
416 * @nr_free: If set to true will free all pages in pool
417 * @use_static: Safe to use static buffer
418 **/
419static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free,
420 bool use_static)
421{
422 static struct page *static_buf[NUM_PAGES_TO_ALLOC];
423 unsigned long irq_flags;
424 struct dma_page *dma_p, *tmp;
425 struct page **pages_to_free;
426 struct list_head d_pages;
427 unsigned freed_pages = 0,
428 npages_to_free = nr_free;
429
430 if (NUM_PAGES_TO_ALLOC < nr_free)
431 npages_to_free = NUM_PAGES_TO_ALLOC;
432#if 0
433 if (nr_free > 1) {
434 pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
435 pool->dev_name, pool->name, current->pid,
436 npages_to_free, nr_free);
437 }
438#endif
439 if (use_static)
440 pages_to_free = static_buf;
441 else
442 pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
443 GFP_KERNEL);
444
445 if (!pages_to_free) {
446 pr_err("%s: Failed to allocate memory for pool free operation\n",
447 pool->dev_name);
448 return 0;
449 }
450 INIT_LIST_HEAD(&d_pages);
451restart:
452 spin_lock_irqsave(&pool->lock, irq_flags);
453
454 /* We picking the oldest ones off the list */
455 list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
456 page_list) {
457 if (freed_pages >= npages_to_free)
458 break;
459
460 /* Move the dma_page from one list to another. */
461 list_move(&dma_p->page_list, &d_pages);
462
463 pages_to_free[freed_pages++] = dma_p->p;
464 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
465 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
466
467 ttm_pool_update_free_locked(pool, freed_pages);
468 /**
469 * Because changing page caching is costly
470 * we unlock the pool to prevent stalling.
471 */
472 spin_unlock_irqrestore(&pool->lock, irq_flags);
473
474 ttm_dma_pages_put(pool, &d_pages, pages_to_free,
475 freed_pages);
476
477 INIT_LIST_HEAD(&d_pages);
478
479 if (likely(nr_free != FREE_ALL_PAGES))
480 nr_free -= freed_pages;
481
482 if (NUM_PAGES_TO_ALLOC >= nr_free)
483 npages_to_free = nr_free;
484 else
485 npages_to_free = NUM_PAGES_TO_ALLOC;
486
487 freed_pages = 0;
488
489 /* free all so restart the processing */
490 if (nr_free)
491 goto restart;
492
493 /* Not allowed to fall through or break because
494 * following context is inside spinlock while we are
495 * outside here.
496 */
497 goto out;
498
499 }
500 }
501
502 /* remove range of pages from the pool */
503 if (freed_pages) {
504 ttm_pool_update_free_locked(pool, freed_pages);
505 nr_free -= freed_pages;
506 }
507
508 spin_unlock_irqrestore(&pool->lock, irq_flags);
509
510 if (freed_pages)
511 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
512out:
513 if (pages_to_free != static_buf)
514 kfree(pages_to_free);
515 return nr_free;
516}
517
518static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
519{
520 struct device_pools *p;
521 struct dma_pool *pool;
522
523 if (!dev)
524 return;
525
526 mutex_lock(&_manager->lock);
527 list_for_each_entry_reverse(p, &_manager->pools, pools) {
528 if (p->dev != dev)
529 continue;
530 pool = p->pool;
531 if (pool->type != type)
532 continue;
533
534 list_del(&p->pools);
535 kfree(p);
536 _manager->npools--;
537 break;
538 }
539 list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
540 if (pool->type != type)
541 continue;
542 /* Takes a spinlock.. */
543 /* OK to use static buffer since global mutex is held. */
544 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, true);
545 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
546 /* This code path is called after _all_ references to the
547 * struct device has been dropped - so nobody should be
548 * touching it. In case somebody is trying to _add_ we are
549 * guarded by the mutex. */
550 list_del(&pool->pools);
551 kfree(pool);
552 break;
553 }
554 mutex_unlock(&_manager->lock);
555}
556
557/*
558 * On free-ing of the 'struct device' this deconstructor is run.
559 * Albeit the pool might have already been freed earlier.
560 */
561static void ttm_dma_pool_release(struct device *dev, void *res)
562{
563 struct dma_pool *pool = *(struct dma_pool **)res;
564
565 if (pool)
566 ttm_dma_free_pool(dev, pool->type);
567}
568
569static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
570{
571 return *(struct dma_pool **)res == match_data;
572}
573
574static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
575 enum pool_type type)
576{
577 char *n[] = {"wc", "uc", "cached", " dma32", "unknown",};
578 enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_UNDEFINED};
579 struct device_pools *sec_pool = NULL;
580 struct dma_pool *pool = NULL, **ptr;
581 unsigned i;
582 int ret = -ENODEV;
583 char *p;
584
585 if (!dev)
586 return NULL;
587
588 ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
589 if (!ptr)
590 return NULL;
591
592 ret = -ENOMEM;
593
594 pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
595 dev_to_node(dev));
596 if (!pool)
597 goto err_mem;
598
599 sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
600 dev_to_node(dev));
601 if (!sec_pool)
602 goto err_mem;
603
604 INIT_LIST_HEAD(&sec_pool->pools);
605 sec_pool->dev = dev;
606 sec_pool->pool = pool;
607
608 INIT_LIST_HEAD(&pool->free_list);
609 INIT_LIST_HEAD(&pool->inuse_list);
610 INIT_LIST_HEAD(&pool->pools);
611 spin_lock_init(&pool->lock);
612 pool->dev = dev;
613 pool->npages_free = pool->npages_in_use = 0;
614 pool->nfrees = 0;
615 pool->gfp_flags = flags;
616 pool->size = PAGE_SIZE;
617 pool->type = type;
618 pool->nrefills = 0;
619 p = pool->name;
620 for (i = 0; i < 5; i++) {
621 if (type & t[i]) {
622 p += snprintf(p, sizeof(pool->name) - (p - pool->name),
623 "%s", n[i]);
624 }
625 }
626 *p = 0;
627 /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
628 * - the kobj->name has already been deallocated.*/
629 snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
630 dev_driver_string(dev), dev_name(dev));
631 mutex_lock(&_manager->lock);
632 /* You can get the dma_pool from either the global: */
633 list_add(&sec_pool->pools, &_manager->pools);
634 _manager->npools++;
635 /* or from 'struct device': */
636 list_add(&pool->pools, &dev->dma_pools);
637 mutex_unlock(&_manager->lock);
638
639 *ptr = pool;
640 devres_add(dev, ptr);
641
642 return pool;
643err_mem:
644 devres_free(ptr);
645 kfree(sec_pool);
646 kfree(pool);
647 return ERR_PTR(ret);
648}
649
650static struct dma_pool *ttm_dma_find_pool(struct device *dev,
651 enum pool_type type)
652{
653 struct dma_pool *pool, *tmp, *found = NULL;
654
655 if (type == IS_UNDEFINED)
656 return found;
657
658 /* NB: We iterate on the 'struct dev' which has no spinlock, but
659 * it does have a kref which we have taken. The kref is taken during
660 * graphic driver loading - in the drm_pci_init it calls either
661 * pci_dev_get or pci_register_driver which both end up taking a kref
662 * on 'struct device'.
663 *
664 * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
665 * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
666 * thing is at that point of time there are no pages associated with the
667 * driver so this function will not be called.
668 */
669 list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools) {
670 if (pool->type != type)
671 continue;
672 found = pool;
673 break;
674 }
675 return found;
676}
677
678/*
679 * Free pages the pages that failed to change the caching state. If there
680 * are pages that have changed their caching state already put them to the
681 * pool.
682 */
683static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
684 struct list_head *d_pages,
685 struct page **failed_pages,
686 unsigned cpages)
687{
688 struct dma_page *d_page, *tmp;
689 struct page *p;
690 unsigned i = 0;
691
692 p = failed_pages[0];
693 if (!p)
694 return;
695 /* Find the failed page. */
696 list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
697 if (d_page->p != p)
698 continue;
699 /* .. and then progress over the full list. */
700 list_del(&d_page->page_list);
701 __ttm_dma_free_page(pool, d_page);
702 if (++i < cpages)
703 p = failed_pages[i];
704 else
705 break;
706 }
707
708}
709
710/*
711 * Allocate 'count' pages, and put 'need' number of them on the
712 * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
713 * The full list of pages should also be on 'd_pages'.
714 * We return zero for success, and negative numbers as errors.
715 */
716static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
717 struct list_head *d_pages,
718 unsigned count)
719{
720 struct page **caching_array;
721 struct dma_page *dma_p;
722 struct page *p;
723 int r = 0;
724 unsigned i, cpages;
725 unsigned max_cpages = min(count,
726 (unsigned)(PAGE_SIZE/sizeof(struct page *)));
727
728 /* allocate array for page caching change */
729 caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
730
731 if (!caching_array) {
732 pr_err("%s: Unable to allocate table for new pages\n",
733 pool->dev_name);
734 return -ENOMEM;
735 }
736
737 if (count > 1) {
738 pr_debug("%s: (%s:%d) Getting %d pages\n",
739 pool->dev_name, pool->name, current->pid, count);
740 }
741
742 for (i = 0, cpages = 0; i < count; ++i) {
743 dma_p = __ttm_dma_alloc_page(pool);
744 if (!dma_p) {
745 pr_err("%s: Unable to get page %u\n",
746 pool->dev_name, i);
747
748 /* store already allocated pages in the pool after
749 * setting the caching state */
750 if (cpages) {
751 r = ttm_set_pages_caching(pool, caching_array,
752 cpages);
753 if (r)
754 ttm_dma_handle_caching_state_failure(
755 pool, d_pages, caching_array,
756 cpages);
757 }
758 r = -ENOMEM;
759 goto out;
760 }
761 p = dma_p->p;
762#ifdef CONFIG_HIGHMEM
763 /* gfp flags of highmem page should never be dma32 so we
764 * we should be fine in such case
765 */
766 if (!PageHighMem(p))
767#endif
768 {
769 caching_array[cpages++] = p;
770 if (cpages == max_cpages) {
771 /* Note: Cannot hold the spinlock */
772 r = ttm_set_pages_caching(pool, caching_array,
773 cpages);
774 if (r) {
775 ttm_dma_handle_caching_state_failure(
776 pool, d_pages, caching_array,
777 cpages);
778 goto out;
779 }
780 cpages = 0;
781 }
782 }
783 list_add(&dma_p->page_list, d_pages);
784 }
785
786 if (cpages) {
787 r = ttm_set_pages_caching(pool, caching_array, cpages);
788 if (r)
789 ttm_dma_handle_caching_state_failure(pool, d_pages,
790 caching_array, cpages);
791 }
792out:
793 kfree(caching_array);
794 return r;
795}
796
797/*
798 * @return count of pages still required to fulfill the request.
799 */
800static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
801 unsigned long *irq_flags)
802{
803 unsigned count = _manager->options.small;
804 int r = pool->npages_free;
805
806 if (count > pool->npages_free) {
807 struct list_head d_pages;
808
809 INIT_LIST_HEAD(&d_pages);
810
811 spin_unlock_irqrestore(&pool->lock, *irq_flags);
812
813 /* Returns how many more are neccessary to fulfill the
814 * request. */
815 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
816
817 spin_lock_irqsave(&pool->lock, *irq_flags);
818 if (!r) {
819 /* Add the fresh to the end.. */
820 list_splice(&d_pages, &pool->free_list);
821 ++pool->nrefills;
822 pool->npages_free += count;
823 r = count;
824 } else {
825 struct dma_page *d_page;
826 unsigned cpages = 0;
827
828 pr_err("%s: Failed to fill %s pool (r:%d)!\n",
829 pool->dev_name, pool->name, r);
830
831 list_for_each_entry(d_page, &d_pages, page_list) {
832 cpages++;
833 }
834 list_splice_tail(&d_pages, &pool->free_list);
835 pool->npages_free += cpages;
836 r = cpages;
837 }
838 }
839 return r;
840}
841
842/*
843 * @return count of pages still required to fulfill the request.
844 * The populate list is actually a stack (not that is matters as TTM
845 * allocates one page at a time.
846 */
847static int ttm_dma_pool_get_pages(struct dma_pool *pool,
848 struct ttm_dma_tt *ttm_dma,
849 unsigned index)
850{
851 struct dma_page *d_page;
852 struct ttm_tt *ttm = &ttm_dma->ttm;
853 unsigned long irq_flags;
854 int count, r = -ENOMEM;
855
856 spin_lock_irqsave(&pool->lock, irq_flags);
857 count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
858 if (count) {
859 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
860 ttm->pages[index] = d_page->p;
861 ttm_dma->cpu_address[index] = d_page->vaddr;
862 ttm_dma->dma_address[index] = d_page->dma;
863 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
864 r = 0;
865 pool->npages_in_use += 1;
866 pool->npages_free -= 1;
867 }
868 spin_unlock_irqrestore(&pool->lock, irq_flags);
869 return r;
870}
871
872/*
873 * On success pages list will hold count number of correctly
874 * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
875 */
876int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev)
877{
878 struct ttm_tt *ttm = &ttm_dma->ttm;
879 struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
880 struct dma_pool *pool;
881 enum pool_type type;
882 unsigned i;
883 gfp_t gfp_flags;
884 int ret;
885
886 if (ttm->state != tt_unpopulated)
887 return 0;
888
889 type = ttm_to_type(ttm->page_flags, ttm->caching_state);
890 if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
891 gfp_flags = GFP_USER | GFP_DMA32;
892 else
893 gfp_flags = GFP_HIGHUSER;
894 if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
895 gfp_flags |= __GFP_ZERO;
896
897 pool = ttm_dma_find_pool(dev, type);
898 if (!pool) {
899 pool = ttm_dma_pool_init(dev, gfp_flags, type);
900 if (IS_ERR_OR_NULL(pool)) {
901 return -ENOMEM;
902 }
903 }
904
905 INIT_LIST_HEAD(&ttm_dma->pages_list);
906 for (i = 0; i < ttm->num_pages; ++i) {
907 ret = ttm_dma_pool_get_pages(pool, ttm_dma, i);
908 if (ret != 0) {
909 ttm_dma_unpopulate(ttm_dma, dev);
910 return -ENOMEM;
911 }
912
913 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
914 false, false);
915 if (unlikely(ret != 0)) {
916 ttm_dma_unpopulate(ttm_dma, dev);
917 return -ENOMEM;
918 }
919 }
920
921 if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
922 ret = ttm_tt_swapin(ttm);
923 if (unlikely(ret != 0)) {
924 ttm_dma_unpopulate(ttm_dma, dev);
925 return ret;
926 }
927 }
928
929 ttm->state = tt_unbound;
930 return 0;
931}
932EXPORT_SYMBOL_GPL(ttm_dma_populate);
933
934/* Put all pages in pages list to correct pool to wait for reuse */
935void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
936{
937 struct ttm_tt *ttm = &ttm_dma->ttm;
938 struct dma_pool *pool;
939 struct dma_page *d_page, *next;
940 enum pool_type type;
941 bool is_cached = false;
942 unsigned count = 0, i, npages = 0;
943 unsigned long irq_flags;
944
945 type = ttm_to_type(ttm->page_flags, ttm->caching_state);
946 pool = ttm_dma_find_pool(dev, type);
947 if (!pool)
948 return;
949
950 is_cached = (ttm_dma_find_pool(pool->dev,
951 ttm_to_type(ttm->page_flags, tt_cached)) == pool);
952
953 /* make sure pages array match list and count number of pages */
954 list_for_each_entry(d_page, &ttm_dma->pages_list, page_list) {
955 ttm->pages[count] = d_page->p;
956 count++;
957 }
958
959 spin_lock_irqsave(&pool->lock, irq_flags);
960 pool->npages_in_use -= count;
961 if (is_cached) {
962 pool->nfrees += count;
963 } else {
964 pool->npages_free += count;
965 list_splice(&ttm_dma->pages_list, &pool->free_list);
966 /*
967 * Wait to have at at least NUM_PAGES_TO_ALLOC number of pages
968 * to free in order to minimize calls to set_memory_wb().
969 */
970 if (pool->npages_free >= (_manager->options.max_size +
971 NUM_PAGES_TO_ALLOC))
972 npages = pool->npages_free - _manager->options.max_size;
973 }
974 spin_unlock_irqrestore(&pool->lock, irq_flags);
975
976 if (is_cached) {
977 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list, page_list) {
978 ttm_mem_global_free_page(ttm->glob->mem_glob,
979 d_page->p);
980 ttm_dma_page_put(pool, d_page);
981 }
982 } else {
983 for (i = 0; i < count; i++) {
984 ttm_mem_global_free_page(ttm->glob->mem_glob,
985 ttm->pages[i]);
986 }
987 }
988
989 INIT_LIST_HEAD(&ttm_dma->pages_list);
990 for (i = 0; i < ttm->num_pages; i++) {
991 ttm->pages[i] = NULL;
992 ttm_dma->cpu_address[i] = 0;
993 ttm_dma->dma_address[i] = 0;
994 }
995
996 /* shrink pool if necessary (only on !is_cached pools)*/
997 if (npages)
998 ttm_dma_page_pool_free(pool, npages, false);
999 ttm->state = tt_unpopulated;
1000}
1001EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1002
1003/**
1004 * Callback for mm to request pool to reduce number of page held.
1005 *
1006 * XXX: (dchinner) Deadlock warning!
1007 *
1008 * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1009 * shrinkers
1010 */
1011static unsigned long
1012ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1013{
1014 static unsigned start_pool;
1015 unsigned idx = 0;
1016 unsigned pool_offset;
1017 unsigned shrink_pages = sc->nr_to_scan;
1018 struct device_pools *p;
1019 unsigned long freed = 0;
1020
1021 if (list_empty(&_manager->pools))
1022 return SHRINK_STOP;
1023
1024 if (!mutex_trylock(&_manager->lock))
1025 return SHRINK_STOP;
1026 if (!_manager->npools)
1027 goto out;
1028 pool_offset = ++start_pool % _manager->npools;
1029 list_for_each_entry(p, &_manager->pools, pools) {
1030 unsigned nr_free;
1031
1032 if (!p->dev)
1033 continue;
1034 if (shrink_pages == 0)
1035 break;
1036 /* Do it in round-robin fashion. */
1037 if (++idx < pool_offset)
1038 continue;
1039 nr_free = shrink_pages;
1040 /* OK to use static buffer since global mutex is held. */
1041 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free, true);
1042 freed += nr_free - shrink_pages;
1043
1044 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1045 p->pool->dev_name, p->pool->name, current->pid,
1046 nr_free, shrink_pages);
1047 }
1048out:
1049 mutex_unlock(&_manager->lock);
1050 return freed;
1051}
1052
1053static unsigned long
1054ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1055{
1056 struct device_pools *p;
1057 unsigned long count = 0;
1058
1059 if (!mutex_trylock(&_manager->lock))
1060 return 0;
1061 list_for_each_entry(p, &_manager->pools, pools)
1062 count += p->pool->npages_free;
1063 mutex_unlock(&_manager->lock);
1064 return count;
1065}
1066
1067static void ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1068{
1069 manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1070 manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1071 manager->mm_shrink.seeks = 1;
1072 register_shrinker(&manager->mm_shrink);
1073}
1074
1075static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1076{
1077 unregister_shrinker(&manager->mm_shrink);
1078}
1079
1080int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1081{
1082 int ret = -ENOMEM;
1083
1084 WARN_ON(_manager);
1085
1086 pr_info("Initializing DMA pool allocator\n");
1087
1088 _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1089 if (!_manager)
1090 goto err;
1091
1092 mutex_init(&_manager->lock);
1093 INIT_LIST_HEAD(&_manager->pools);
1094
1095 _manager->options.max_size = max_pages;
1096 _manager->options.small = SMALL_ALLOCATION;
1097 _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1098
1099 /* This takes care of auto-freeing the _manager */
1100 ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1101 &glob->kobj, "dma_pool");
1102 if (unlikely(ret != 0)) {
1103 kobject_put(&_manager->kobj);
1104 goto err;
1105 }
1106 ttm_dma_pool_mm_shrink_init(_manager);
1107 return 0;
1108err:
1109 return ret;
1110}
1111
1112void ttm_dma_page_alloc_fini(void)
1113{
1114 struct device_pools *p, *t;
1115
1116 pr_info("Finalizing DMA pool allocator\n");
1117 ttm_dma_pool_mm_shrink_fini(_manager);
1118
1119 list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1120 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1121 current->pid);
1122 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1123 ttm_dma_pool_match, p->pool));
1124 ttm_dma_free_pool(p->dev, p->pool->type);
1125 }
1126 kobject_put(&_manager->kobj);
1127 _manager = NULL;
1128}
1129
1130int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1131{
1132 struct device_pools *p;
1133 struct dma_pool *pool = NULL;
1134 char *h[] = {"pool", "refills", "pages freed", "inuse", "available",
1135 "name", "virt", "busaddr"};
1136
1137 if (!_manager) {
1138 seq_printf(m, "No pool allocator running.\n");
1139 return 0;
1140 }
1141 seq_printf(m, "%13s %12s %13s %8s %8s %8s\n",
1142 h[0], h[1], h[2], h[3], h[4], h[5]);
1143 mutex_lock(&_manager->lock);
1144 list_for_each_entry(p, &_manager->pools, pools) {
1145 struct device *dev = p->dev;
1146 if (!dev)
1147 continue;
1148 pool = p->pool;
1149 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1150 pool->name, pool->nrefills,
1151 pool->nfrees, pool->npages_in_use,
1152 pool->npages_free,
1153 pool->dev_name);
1154 }
1155 mutex_unlock(&_manager->lock);
1156 return 0;
1157}
1158EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1159
1160#endif
1/*
2 * Copyright 2011 (c) Oracle Corp.
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 * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
24 */
25
26/*
27 * A simple DMA pool losely based on dmapool.c. It has certain advantages
28 * over the DMA pools:
29 * - Pool collects resently freed pages for reuse (and hooks up to
30 * the shrinker).
31 * - Tracks currently in use pages
32 * - Tracks whether the page is UC, WB or cached (and reverts to WB
33 * when freed).
34 */
35
36#if defined(CONFIG_SWIOTLB) || defined(CONFIG_INTEL_IOMMU)
37#define pr_fmt(fmt) "[TTM] " fmt
38
39#include <linux/dma-mapping.h>
40#include <linux/list.h>
41#include <linux/seq_file.h> /* for seq_printf */
42#include <linux/slab.h>
43#include <linux/spinlock.h>
44#include <linux/highmem.h>
45#include <linux/mm_types.h>
46#include <linux/module.h>
47#include <linux/mm.h>
48#include <linux/atomic.h>
49#include <linux/device.h>
50#include <linux/kthread.h>
51#include <drm/ttm/ttm_bo_driver.h>
52#include <drm/ttm/ttm_page_alloc.h>
53#if IS_ENABLED(CONFIG_AGP)
54#include <asm/agp.h>
55#endif
56#ifdef CONFIG_X86
57#include <asm/set_memory.h>
58#endif
59
60#define NUM_PAGES_TO_ALLOC (PAGE_SIZE/sizeof(struct page *))
61#define SMALL_ALLOCATION 4
62#define FREE_ALL_PAGES (~0U)
63#define VADDR_FLAG_HUGE_POOL 1UL
64#define VADDR_FLAG_UPDATED_COUNT 2UL
65
66enum pool_type {
67 IS_UNDEFINED = 0,
68 IS_WC = 1 << 1,
69 IS_UC = 1 << 2,
70 IS_CACHED = 1 << 3,
71 IS_DMA32 = 1 << 4,
72 IS_HUGE = 1 << 5
73};
74
75/*
76 * The pool structure. There are up to nine pools:
77 * - generic (not restricted to DMA32):
78 * - write combined, uncached, cached.
79 * - dma32 (up to 2^32 - so up 4GB):
80 * - write combined, uncached, cached.
81 * - huge (not restricted to DMA32):
82 * - write combined, uncached, cached.
83 * for each 'struct device'. The 'cached' is for pages that are actively used.
84 * The other ones can be shrunk by the shrinker API if neccessary.
85 * @pools: The 'struct device->dma_pools' link.
86 * @type: Type of the pool
87 * @lock: Protects the free_list from concurrnet access. Must be
88 * used with irqsave/irqrestore variants because pool allocator maybe called
89 * from delayed work.
90 * @free_list: Pool of pages that are free to be used. No order requirements.
91 * @dev: The device that is associated with these pools.
92 * @size: Size used during DMA allocation.
93 * @npages_free: Count of available pages for re-use.
94 * @npages_in_use: Count of pages that are in use.
95 * @nfrees: Stats when pool is shrinking.
96 * @nrefills: Stats when the pool is grown.
97 * @gfp_flags: Flags to pass for alloc_page.
98 * @name: Name of the pool.
99 * @dev_name: Name derieved from dev - similar to how dev_info works.
100 * Used during shutdown as the dev_info during release is unavailable.
101 */
102struct dma_pool {
103 struct list_head pools; /* The 'struct device->dma_pools link */
104 enum pool_type type;
105 spinlock_t lock;
106 struct list_head free_list;
107 struct device *dev;
108 unsigned size;
109 unsigned npages_free;
110 unsigned npages_in_use;
111 unsigned long nfrees; /* Stats when shrunk. */
112 unsigned long nrefills; /* Stats when grown. */
113 gfp_t gfp_flags;
114 char name[13]; /* "cached dma32" */
115 char dev_name[64]; /* Constructed from dev */
116};
117
118/*
119 * The accounting page keeping track of the allocated page along with
120 * the DMA address.
121 * @page_list: The link to the 'page_list' in 'struct dma_pool'.
122 * @vaddr: The virtual address of the page and a flag if the page belongs to a
123 * huge pool
124 * @dma: The bus address of the page. If the page is not allocated
125 * via the DMA API, it will be -1.
126 */
127struct dma_page {
128 struct list_head page_list;
129 unsigned long vaddr;
130 struct page *p;
131 dma_addr_t dma;
132};
133
134/*
135 * Limits for the pool. They are handled without locks because only place where
136 * they may change is in sysfs store. They won't have immediate effect anyway
137 * so forcing serialization to access them is pointless.
138 */
139
140struct ttm_pool_opts {
141 unsigned alloc_size;
142 unsigned max_size;
143 unsigned small;
144};
145
146/*
147 * Contains the list of all of the 'struct device' and their corresponding
148 * DMA pools. Guarded by _mutex->lock.
149 * @pools: The link to 'struct ttm_pool_manager->pools'
150 * @dev: The 'struct device' associated with the 'pool'
151 * @pool: The 'struct dma_pool' associated with the 'dev'
152 */
153struct device_pools {
154 struct list_head pools;
155 struct device *dev;
156 struct dma_pool *pool;
157};
158
159/*
160 * struct ttm_pool_manager - Holds memory pools for fast allocation
161 *
162 * @lock: Lock used when adding/removing from pools
163 * @pools: List of 'struct device' and 'struct dma_pool' tuples.
164 * @options: Limits for the pool.
165 * @npools: Total amount of pools in existence.
166 * @shrinker: The structure used by [un|]register_shrinker
167 */
168struct ttm_pool_manager {
169 struct mutex lock;
170 struct list_head pools;
171 struct ttm_pool_opts options;
172 unsigned npools;
173 struct shrinker mm_shrink;
174 struct kobject kobj;
175};
176
177static struct ttm_pool_manager *_manager;
178
179static struct attribute ttm_page_pool_max = {
180 .name = "pool_max_size",
181 .mode = S_IRUGO | S_IWUSR
182};
183static struct attribute ttm_page_pool_small = {
184 .name = "pool_small_allocation",
185 .mode = S_IRUGO | S_IWUSR
186};
187static struct attribute ttm_page_pool_alloc_size = {
188 .name = "pool_allocation_size",
189 .mode = S_IRUGO | S_IWUSR
190};
191
192static struct attribute *ttm_pool_attrs[] = {
193 &ttm_page_pool_max,
194 &ttm_page_pool_small,
195 &ttm_page_pool_alloc_size,
196 NULL
197};
198
199static void ttm_pool_kobj_release(struct kobject *kobj)
200{
201 struct ttm_pool_manager *m =
202 container_of(kobj, struct ttm_pool_manager, kobj);
203 kfree(m);
204}
205
206static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
207 const char *buffer, size_t size)
208{
209 struct ttm_pool_manager *m =
210 container_of(kobj, struct ttm_pool_manager, kobj);
211 int chars;
212 unsigned val;
213
214 chars = sscanf(buffer, "%u", &val);
215 if (chars == 0)
216 return size;
217
218 /* Convert kb to number of pages */
219 val = val / (PAGE_SIZE >> 10);
220
221 if (attr == &ttm_page_pool_max) {
222 m->options.max_size = val;
223 } else if (attr == &ttm_page_pool_small) {
224 m->options.small = val;
225 } else if (attr == &ttm_page_pool_alloc_size) {
226 if (val > NUM_PAGES_TO_ALLOC*8) {
227 pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
228 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
229 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
230 return size;
231 } else if (val > NUM_PAGES_TO_ALLOC) {
232 pr_warn("Setting allocation size to larger than %lu is not recommended\n",
233 NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
234 }
235 m->options.alloc_size = val;
236 }
237
238 return size;
239}
240
241static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
242 char *buffer)
243{
244 struct ttm_pool_manager *m =
245 container_of(kobj, struct ttm_pool_manager, kobj);
246 unsigned val = 0;
247
248 if (attr == &ttm_page_pool_max)
249 val = m->options.max_size;
250 else if (attr == &ttm_page_pool_small)
251 val = m->options.small;
252 else if (attr == &ttm_page_pool_alloc_size)
253 val = m->options.alloc_size;
254
255 val = val * (PAGE_SIZE >> 10);
256
257 return snprintf(buffer, PAGE_SIZE, "%u\n", val);
258}
259
260static const struct sysfs_ops ttm_pool_sysfs_ops = {
261 .show = &ttm_pool_show,
262 .store = &ttm_pool_store,
263};
264
265static struct kobj_type ttm_pool_kobj_type = {
266 .release = &ttm_pool_kobj_release,
267 .sysfs_ops = &ttm_pool_sysfs_ops,
268 .default_attrs = ttm_pool_attrs,
269};
270
271#ifndef CONFIG_X86
272static int set_pages_array_wb(struct page **pages, int addrinarray)
273{
274#if IS_ENABLED(CONFIG_AGP)
275 int i;
276
277 for (i = 0; i < addrinarray; i++)
278 unmap_page_from_agp(pages[i]);
279#endif
280 return 0;
281}
282
283static int set_pages_array_wc(struct page **pages, int addrinarray)
284{
285#if IS_ENABLED(CONFIG_AGP)
286 int i;
287
288 for (i = 0; i < addrinarray; i++)
289 map_page_into_agp(pages[i]);
290#endif
291 return 0;
292}
293
294static int set_pages_array_uc(struct page **pages, int addrinarray)
295{
296#if IS_ENABLED(CONFIG_AGP)
297 int i;
298
299 for (i = 0; i < addrinarray; i++)
300 map_page_into_agp(pages[i]);
301#endif
302 return 0;
303}
304#endif /* for !CONFIG_X86 */
305
306static int ttm_set_pages_caching(struct dma_pool *pool,
307 struct page **pages, unsigned cpages)
308{
309 int r = 0;
310 /* Set page caching */
311 if (pool->type & IS_UC) {
312 r = set_pages_array_uc(pages, cpages);
313 if (r)
314 pr_err("%s: Failed to set %d pages to uc!\n",
315 pool->dev_name, cpages);
316 }
317 if (pool->type & IS_WC) {
318 r = set_pages_array_wc(pages, cpages);
319 if (r)
320 pr_err("%s: Failed to set %d pages to wc!\n",
321 pool->dev_name, cpages);
322 }
323 return r;
324}
325
326static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
327{
328 dma_addr_t dma = d_page->dma;
329 d_page->vaddr &= ~VADDR_FLAG_HUGE_POOL;
330 dma_free_coherent(pool->dev, pool->size, (void *)d_page->vaddr, dma);
331
332 kfree(d_page);
333 d_page = NULL;
334}
335static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
336{
337 struct dma_page *d_page;
338 unsigned long attrs = 0;
339 void *vaddr;
340
341 d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
342 if (!d_page)
343 return NULL;
344
345 if (pool->type & IS_HUGE)
346 attrs = DMA_ATTR_NO_WARN;
347
348 vaddr = dma_alloc_attrs(pool->dev, pool->size, &d_page->dma,
349 pool->gfp_flags, attrs);
350 if (vaddr) {
351 if (is_vmalloc_addr(vaddr))
352 d_page->p = vmalloc_to_page(vaddr);
353 else
354 d_page->p = virt_to_page(vaddr);
355 d_page->vaddr = (unsigned long)vaddr;
356 if (pool->type & IS_HUGE)
357 d_page->vaddr |= VADDR_FLAG_HUGE_POOL;
358 } else {
359 kfree(d_page);
360 d_page = NULL;
361 }
362 return d_page;
363}
364static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
365{
366 enum pool_type type = IS_UNDEFINED;
367
368 if (flags & TTM_PAGE_FLAG_DMA32)
369 type |= IS_DMA32;
370 if (cstate == tt_cached)
371 type |= IS_CACHED;
372 else if (cstate == tt_uncached)
373 type |= IS_UC;
374 else
375 type |= IS_WC;
376
377 return type;
378}
379
380static void ttm_pool_update_free_locked(struct dma_pool *pool,
381 unsigned freed_pages)
382{
383 pool->npages_free -= freed_pages;
384 pool->nfrees += freed_pages;
385
386}
387
388/* set memory back to wb and free the pages. */
389static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
390{
391 struct page *page = d_page->p;
392 unsigned i, num_pages;
393
394 /* Don't set WB on WB page pool. */
395 if (!(pool->type & IS_CACHED)) {
396 num_pages = pool->size / PAGE_SIZE;
397 for (i = 0; i < num_pages; ++i, ++page) {
398 if (set_pages_array_wb(&page, 1)) {
399 pr_err("%s: Failed to set %d pages to wb!\n",
400 pool->dev_name, 1);
401 }
402 }
403 }
404
405 list_del(&d_page->page_list);
406 __ttm_dma_free_page(pool, d_page);
407}
408
409static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
410 struct page *pages[], unsigned npages)
411{
412 struct dma_page *d_page, *tmp;
413
414 if (pool->type & IS_HUGE) {
415 list_for_each_entry_safe(d_page, tmp, d_pages, page_list)
416 ttm_dma_page_put(pool, d_page);
417
418 return;
419 }
420
421 /* Don't set WB on WB page pool. */
422 if (npages && !(pool->type & IS_CACHED) &&
423 set_pages_array_wb(pages, npages))
424 pr_err("%s: Failed to set %d pages to wb!\n",
425 pool->dev_name, npages);
426
427 list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
428 list_del(&d_page->page_list);
429 __ttm_dma_free_page(pool, d_page);
430 }
431}
432
433/*
434 * Free pages from pool.
435 *
436 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
437 * number of pages in one go.
438 *
439 * @pool: to free the pages from
440 * @nr_free: If set to true will free all pages in pool
441 * @use_static: Safe to use static buffer
442 **/
443static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free,
444 bool use_static)
445{
446 static struct page *static_buf[NUM_PAGES_TO_ALLOC];
447 unsigned long irq_flags;
448 struct dma_page *dma_p, *tmp;
449 struct page **pages_to_free;
450 struct list_head d_pages;
451 unsigned freed_pages = 0,
452 npages_to_free = nr_free;
453
454 if (NUM_PAGES_TO_ALLOC < nr_free)
455 npages_to_free = NUM_PAGES_TO_ALLOC;
456#if 0
457 if (nr_free > 1) {
458 pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
459 pool->dev_name, pool->name, current->pid,
460 npages_to_free, nr_free);
461 }
462#endif
463 if (use_static)
464 pages_to_free = static_buf;
465 else
466 pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
467 GFP_KERNEL);
468
469 if (!pages_to_free) {
470 pr_debug("%s: Failed to allocate memory for pool free operation\n",
471 pool->dev_name);
472 return 0;
473 }
474 INIT_LIST_HEAD(&d_pages);
475restart:
476 spin_lock_irqsave(&pool->lock, irq_flags);
477
478 /* We picking the oldest ones off the list */
479 list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
480 page_list) {
481 if (freed_pages >= npages_to_free)
482 break;
483
484 /* Move the dma_page from one list to another. */
485 list_move(&dma_p->page_list, &d_pages);
486
487 pages_to_free[freed_pages++] = dma_p->p;
488 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
489 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
490
491 ttm_pool_update_free_locked(pool, freed_pages);
492 /**
493 * Because changing page caching is costly
494 * we unlock the pool to prevent stalling.
495 */
496 spin_unlock_irqrestore(&pool->lock, irq_flags);
497
498 ttm_dma_pages_put(pool, &d_pages, pages_to_free,
499 freed_pages);
500
501 INIT_LIST_HEAD(&d_pages);
502
503 if (likely(nr_free != FREE_ALL_PAGES))
504 nr_free -= freed_pages;
505
506 if (NUM_PAGES_TO_ALLOC >= nr_free)
507 npages_to_free = nr_free;
508 else
509 npages_to_free = NUM_PAGES_TO_ALLOC;
510
511 freed_pages = 0;
512
513 /* free all so restart the processing */
514 if (nr_free)
515 goto restart;
516
517 /* Not allowed to fall through or break because
518 * following context is inside spinlock while we are
519 * outside here.
520 */
521 goto out;
522
523 }
524 }
525
526 /* remove range of pages from the pool */
527 if (freed_pages) {
528 ttm_pool_update_free_locked(pool, freed_pages);
529 nr_free -= freed_pages;
530 }
531
532 spin_unlock_irqrestore(&pool->lock, irq_flags);
533
534 if (freed_pages)
535 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
536out:
537 if (pages_to_free != static_buf)
538 kfree(pages_to_free);
539 return nr_free;
540}
541
542static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
543{
544 struct device_pools *p;
545 struct dma_pool *pool;
546
547 if (!dev)
548 return;
549
550 mutex_lock(&_manager->lock);
551 list_for_each_entry_reverse(p, &_manager->pools, pools) {
552 if (p->dev != dev)
553 continue;
554 pool = p->pool;
555 if (pool->type != type)
556 continue;
557
558 list_del(&p->pools);
559 kfree(p);
560 _manager->npools--;
561 break;
562 }
563 list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
564 if (pool->type != type)
565 continue;
566 /* Takes a spinlock.. */
567 /* OK to use static buffer since global mutex is held. */
568 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, true);
569 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
570 /* This code path is called after _all_ references to the
571 * struct device has been dropped - so nobody should be
572 * touching it. In case somebody is trying to _add_ we are
573 * guarded by the mutex. */
574 list_del(&pool->pools);
575 kfree(pool);
576 break;
577 }
578 mutex_unlock(&_manager->lock);
579}
580
581/*
582 * On free-ing of the 'struct device' this deconstructor is run.
583 * Albeit the pool might have already been freed earlier.
584 */
585static void ttm_dma_pool_release(struct device *dev, void *res)
586{
587 struct dma_pool *pool = *(struct dma_pool **)res;
588
589 if (pool)
590 ttm_dma_free_pool(dev, pool->type);
591}
592
593static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
594{
595 return *(struct dma_pool **)res == match_data;
596}
597
598static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
599 enum pool_type type)
600{
601 const char *n[] = {"wc", "uc", "cached", " dma32", "huge"};
602 enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_HUGE};
603 struct device_pools *sec_pool = NULL;
604 struct dma_pool *pool = NULL, **ptr;
605 unsigned i;
606 int ret = -ENODEV;
607 char *p;
608
609 if (!dev)
610 return NULL;
611
612 ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
613 if (!ptr)
614 return NULL;
615
616 ret = -ENOMEM;
617
618 pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
619 dev_to_node(dev));
620 if (!pool)
621 goto err_mem;
622
623 sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
624 dev_to_node(dev));
625 if (!sec_pool)
626 goto err_mem;
627
628 INIT_LIST_HEAD(&sec_pool->pools);
629 sec_pool->dev = dev;
630 sec_pool->pool = pool;
631
632 INIT_LIST_HEAD(&pool->free_list);
633 INIT_LIST_HEAD(&pool->pools);
634 spin_lock_init(&pool->lock);
635 pool->dev = dev;
636 pool->npages_free = pool->npages_in_use = 0;
637 pool->nfrees = 0;
638 pool->gfp_flags = flags;
639 if (type & IS_HUGE)
640#ifdef CONFIG_TRANSPARENT_HUGEPAGE
641 pool->size = HPAGE_PMD_SIZE;
642#else
643 BUG();
644#endif
645 else
646 pool->size = PAGE_SIZE;
647 pool->type = type;
648 pool->nrefills = 0;
649 p = pool->name;
650 for (i = 0; i < ARRAY_SIZE(t); i++) {
651 if (type & t[i]) {
652 p += snprintf(p, sizeof(pool->name) - (p - pool->name),
653 "%s", n[i]);
654 }
655 }
656 *p = 0;
657 /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
658 * - the kobj->name has already been deallocated.*/
659 snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
660 dev_driver_string(dev), dev_name(dev));
661 mutex_lock(&_manager->lock);
662 /* You can get the dma_pool from either the global: */
663 list_add(&sec_pool->pools, &_manager->pools);
664 _manager->npools++;
665 /* or from 'struct device': */
666 list_add(&pool->pools, &dev->dma_pools);
667 mutex_unlock(&_manager->lock);
668
669 *ptr = pool;
670 devres_add(dev, ptr);
671
672 return pool;
673err_mem:
674 devres_free(ptr);
675 kfree(sec_pool);
676 kfree(pool);
677 return ERR_PTR(ret);
678}
679
680static struct dma_pool *ttm_dma_find_pool(struct device *dev,
681 enum pool_type type)
682{
683 struct dma_pool *pool, *tmp;
684
685 if (type == IS_UNDEFINED)
686 return NULL;
687
688 /* NB: We iterate on the 'struct dev' which has no spinlock, but
689 * it does have a kref which we have taken. The kref is taken during
690 * graphic driver loading - in the drm_pci_init it calls either
691 * pci_dev_get or pci_register_driver which both end up taking a kref
692 * on 'struct device'.
693 *
694 * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
695 * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
696 * thing is at that point of time there are no pages associated with the
697 * driver so this function will not be called.
698 */
699 list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools)
700 if (pool->type == type)
701 return pool;
702 return NULL;
703}
704
705/*
706 * Free pages the pages that failed to change the caching state. If there
707 * are pages that have changed their caching state already put them to the
708 * pool.
709 */
710static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
711 struct list_head *d_pages,
712 struct page **failed_pages,
713 unsigned cpages)
714{
715 struct dma_page *d_page, *tmp;
716 struct page *p;
717 unsigned i = 0;
718
719 p = failed_pages[0];
720 if (!p)
721 return;
722 /* Find the failed page. */
723 list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
724 if (d_page->p != p)
725 continue;
726 /* .. and then progress over the full list. */
727 list_del(&d_page->page_list);
728 __ttm_dma_free_page(pool, d_page);
729 if (++i < cpages)
730 p = failed_pages[i];
731 else
732 break;
733 }
734
735}
736
737/*
738 * Allocate 'count' pages, and put 'need' number of them on the
739 * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
740 * The full list of pages should also be on 'd_pages'.
741 * We return zero for success, and negative numbers as errors.
742 */
743static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
744 struct list_head *d_pages,
745 unsigned count)
746{
747 struct page **caching_array;
748 struct dma_page *dma_p;
749 struct page *p;
750 int r = 0;
751 unsigned i, j, npages, cpages;
752 unsigned max_cpages = min(count,
753 (unsigned)(PAGE_SIZE/sizeof(struct page *)));
754
755 /* allocate array for page caching change */
756 caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
757
758 if (!caching_array) {
759 pr_debug("%s: Unable to allocate table for new pages\n",
760 pool->dev_name);
761 return -ENOMEM;
762 }
763
764 if (count > 1)
765 pr_debug("%s: (%s:%d) Getting %d pages\n",
766 pool->dev_name, pool->name, current->pid, count);
767
768 for (i = 0, cpages = 0; i < count; ++i) {
769 dma_p = __ttm_dma_alloc_page(pool);
770 if (!dma_p) {
771 pr_debug("%s: Unable to get page %u\n",
772 pool->dev_name, i);
773
774 /* store already allocated pages in the pool after
775 * setting the caching state */
776 if (cpages) {
777 r = ttm_set_pages_caching(pool, caching_array,
778 cpages);
779 if (r)
780 ttm_dma_handle_caching_state_failure(
781 pool, d_pages, caching_array,
782 cpages);
783 }
784 r = -ENOMEM;
785 goto out;
786 }
787 p = dma_p->p;
788 list_add(&dma_p->page_list, d_pages);
789
790#ifdef CONFIG_HIGHMEM
791 /* gfp flags of highmem page should never be dma32 so we
792 * we should be fine in such case
793 */
794 if (PageHighMem(p))
795 continue;
796#endif
797
798 npages = pool->size / PAGE_SIZE;
799 for (j = 0; j < npages; ++j) {
800 caching_array[cpages++] = p + j;
801 if (cpages == max_cpages) {
802 /* Note: Cannot hold the spinlock */
803 r = ttm_set_pages_caching(pool, caching_array,
804 cpages);
805 if (r) {
806 ttm_dma_handle_caching_state_failure(
807 pool, d_pages, caching_array,
808 cpages);
809 goto out;
810 }
811 cpages = 0;
812 }
813 }
814 }
815
816 if (cpages) {
817 r = ttm_set_pages_caching(pool, caching_array, cpages);
818 if (r)
819 ttm_dma_handle_caching_state_failure(pool, d_pages,
820 caching_array, cpages);
821 }
822out:
823 kfree(caching_array);
824 return r;
825}
826
827/*
828 * @return count of pages still required to fulfill the request.
829 */
830static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
831 unsigned long *irq_flags)
832{
833 unsigned count = _manager->options.small;
834 int r = pool->npages_free;
835
836 if (count > pool->npages_free) {
837 struct list_head d_pages;
838
839 INIT_LIST_HEAD(&d_pages);
840
841 spin_unlock_irqrestore(&pool->lock, *irq_flags);
842
843 /* Returns how many more are neccessary to fulfill the
844 * request. */
845 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
846
847 spin_lock_irqsave(&pool->lock, *irq_flags);
848 if (!r) {
849 /* Add the fresh to the end.. */
850 list_splice(&d_pages, &pool->free_list);
851 ++pool->nrefills;
852 pool->npages_free += count;
853 r = count;
854 } else {
855 struct dma_page *d_page;
856 unsigned cpages = 0;
857
858 pr_debug("%s: Failed to fill %s pool (r:%d)!\n",
859 pool->dev_name, pool->name, r);
860
861 list_for_each_entry(d_page, &d_pages, page_list) {
862 cpages++;
863 }
864 list_splice_tail(&d_pages, &pool->free_list);
865 pool->npages_free += cpages;
866 r = cpages;
867 }
868 }
869 return r;
870}
871
872/*
873 * The populate list is actually a stack (not that is matters as TTM
874 * allocates one page at a time.
875 * return dma_page pointer if success, otherwise NULL.
876 */
877static struct dma_page *ttm_dma_pool_get_pages(struct dma_pool *pool,
878 struct ttm_dma_tt *ttm_dma,
879 unsigned index)
880{
881 struct dma_page *d_page = NULL;
882 struct ttm_tt *ttm = &ttm_dma->ttm;
883 unsigned long irq_flags;
884 int count;
885
886 spin_lock_irqsave(&pool->lock, irq_flags);
887 count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
888 if (count) {
889 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
890 ttm->pages[index] = d_page->p;
891 ttm_dma->dma_address[index] = d_page->dma;
892 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
893 pool->npages_in_use += 1;
894 pool->npages_free -= 1;
895 }
896 spin_unlock_irqrestore(&pool->lock, irq_flags);
897 return d_page;
898}
899
900static gfp_t ttm_dma_pool_gfp_flags(struct ttm_dma_tt *ttm_dma, bool huge)
901{
902 struct ttm_tt *ttm = &ttm_dma->ttm;
903 gfp_t gfp_flags;
904
905 if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
906 gfp_flags = GFP_USER | GFP_DMA32;
907 else
908 gfp_flags = GFP_HIGHUSER;
909 if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
910 gfp_flags |= __GFP_ZERO;
911
912 if (huge) {
913 gfp_flags |= GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
914 __GFP_KSWAPD_RECLAIM;
915 gfp_flags &= ~__GFP_MOVABLE;
916 gfp_flags &= ~__GFP_COMP;
917 }
918
919 if (ttm->page_flags & TTM_PAGE_FLAG_NO_RETRY)
920 gfp_flags |= __GFP_RETRY_MAYFAIL;
921
922 return gfp_flags;
923}
924
925/*
926 * On success pages list will hold count number of correctly
927 * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
928 */
929int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev,
930 struct ttm_operation_ctx *ctx)
931{
932 struct ttm_tt *ttm = &ttm_dma->ttm;
933 struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
934 unsigned long num_pages = ttm->num_pages;
935 struct dma_pool *pool;
936 struct dma_page *d_page;
937 enum pool_type type;
938 unsigned i;
939 int ret;
940
941 if (ttm->state != tt_unpopulated)
942 return 0;
943
944 if (ttm_check_under_lowerlimit(mem_glob, num_pages, ctx))
945 return -ENOMEM;
946
947 INIT_LIST_HEAD(&ttm_dma->pages_list);
948 i = 0;
949
950 type = ttm_to_type(ttm->page_flags, ttm->caching_state);
951
952#ifdef CONFIG_TRANSPARENT_HUGEPAGE
953 if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
954 goto skip_huge;
955
956 pool = ttm_dma_find_pool(dev, type | IS_HUGE);
957 if (!pool) {
958 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, true);
959
960 pool = ttm_dma_pool_init(dev, gfp_flags, type | IS_HUGE);
961 if (IS_ERR_OR_NULL(pool))
962 goto skip_huge;
963 }
964
965 while (num_pages >= HPAGE_PMD_NR) {
966 unsigned j;
967
968 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
969 if (!d_page)
970 break;
971
972 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
973 pool->size, ctx);
974 if (unlikely(ret != 0)) {
975 ttm_dma_unpopulate(ttm_dma, dev);
976 return -ENOMEM;
977 }
978
979 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
980 for (j = i + 1; j < (i + HPAGE_PMD_NR); ++j) {
981 ttm->pages[j] = ttm->pages[j - 1] + 1;
982 ttm_dma->dma_address[j] = ttm_dma->dma_address[j - 1] +
983 PAGE_SIZE;
984 }
985
986 i += HPAGE_PMD_NR;
987 num_pages -= HPAGE_PMD_NR;
988 }
989
990skip_huge:
991#endif
992
993 pool = ttm_dma_find_pool(dev, type);
994 if (!pool) {
995 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, false);
996
997 pool = ttm_dma_pool_init(dev, gfp_flags, type);
998 if (IS_ERR_OR_NULL(pool))
999 return -ENOMEM;
1000 }
1001
1002 while (num_pages) {
1003 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
1004 if (!d_page) {
1005 ttm_dma_unpopulate(ttm_dma, dev);
1006 return -ENOMEM;
1007 }
1008
1009 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
1010 pool->size, ctx);
1011 if (unlikely(ret != 0)) {
1012 ttm_dma_unpopulate(ttm_dma, dev);
1013 return -ENOMEM;
1014 }
1015
1016 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
1017 ++i;
1018 --num_pages;
1019 }
1020
1021 if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
1022 ret = ttm_tt_swapin(ttm);
1023 if (unlikely(ret != 0)) {
1024 ttm_dma_unpopulate(ttm_dma, dev);
1025 return ret;
1026 }
1027 }
1028
1029 ttm->state = tt_unbound;
1030 return 0;
1031}
1032EXPORT_SYMBOL_GPL(ttm_dma_populate);
1033
1034/* Put all pages in pages list to correct pool to wait for reuse */
1035void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
1036{
1037 struct ttm_tt *ttm = &ttm_dma->ttm;
1038 struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
1039 struct dma_pool *pool;
1040 struct dma_page *d_page, *next;
1041 enum pool_type type;
1042 bool is_cached = false;
1043 unsigned count, i, npages = 0;
1044 unsigned long irq_flags;
1045
1046 type = ttm_to_type(ttm->page_flags, ttm->caching_state);
1047
1048#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1049 pool = ttm_dma_find_pool(dev, type | IS_HUGE);
1050 if (pool) {
1051 count = 0;
1052 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1053 page_list) {
1054 if (!(d_page->vaddr & VADDR_FLAG_HUGE_POOL))
1055 continue;
1056
1057 count++;
1058 if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1059 ttm_mem_global_free_page(mem_glob, d_page->p,
1060 pool->size);
1061 d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1062 }
1063 ttm_dma_page_put(pool, d_page);
1064 }
1065
1066 spin_lock_irqsave(&pool->lock, irq_flags);
1067 pool->npages_in_use -= count;
1068 pool->nfrees += count;
1069 spin_unlock_irqrestore(&pool->lock, irq_flags);
1070 }
1071#endif
1072
1073 pool = ttm_dma_find_pool(dev, type);
1074 if (!pool)
1075 return;
1076
1077 is_cached = (ttm_dma_find_pool(pool->dev,
1078 ttm_to_type(ttm->page_flags, tt_cached)) == pool);
1079
1080 /* make sure pages array match list and count number of pages */
1081 count = 0;
1082 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1083 page_list) {
1084 ttm->pages[count] = d_page->p;
1085 count++;
1086
1087 if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1088 ttm_mem_global_free_page(mem_glob, d_page->p,
1089 pool->size);
1090 d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1091 }
1092
1093 if (is_cached)
1094 ttm_dma_page_put(pool, d_page);
1095 }
1096
1097 spin_lock_irqsave(&pool->lock, irq_flags);
1098 pool->npages_in_use -= count;
1099 if (is_cached) {
1100 pool->nfrees += count;
1101 } else {
1102 pool->npages_free += count;
1103 list_splice(&ttm_dma->pages_list, &pool->free_list);
1104 /*
1105 * Wait to have at at least NUM_PAGES_TO_ALLOC number of pages
1106 * to free in order to minimize calls to set_memory_wb().
1107 */
1108 if (pool->npages_free >= (_manager->options.max_size +
1109 NUM_PAGES_TO_ALLOC))
1110 npages = pool->npages_free - _manager->options.max_size;
1111 }
1112 spin_unlock_irqrestore(&pool->lock, irq_flags);
1113
1114 INIT_LIST_HEAD(&ttm_dma->pages_list);
1115 for (i = 0; i < ttm->num_pages; i++) {
1116 ttm->pages[i] = NULL;
1117 ttm_dma->dma_address[i] = 0;
1118 }
1119
1120 /* shrink pool if necessary (only on !is_cached pools)*/
1121 if (npages)
1122 ttm_dma_page_pool_free(pool, npages, false);
1123 ttm->state = tt_unpopulated;
1124}
1125EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1126
1127/**
1128 * Callback for mm to request pool to reduce number of page held.
1129 *
1130 * XXX: (dchinner) Deadlock warning!
1131 *
1132 * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1133 * shrinkers
1134 */
1135static unsigned long
1136ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1137{
1138 static unsigned start_pool;
1139 unsigned idx = 0;
1140 unsigned pool_offset;
1141 unsigned shrink_pages = sc->nr_to_scan;
1142 struct device_pools *p;
1143 unsigned long freed = 0;
1144
1145 if (list_empty(&_manager->pools))
1146 return SHRINK_STOP;
1147
1148 if (!mutex_trylock(&_manager->lock))
1149 return SHRINK_STOP;
1150 if (!_manager->npools)
1151 goto out;
1152 pool_offset = ++start_pool % _manager->npools;
1153 list_for_each_entry(p, &_manager->pools, pools) {
1154 unsigned nr_free;
1155
1156 if (!p->dev)
1157 continue;
1158 if (shrink_pages == 0)
1159 break;
1160 /* Do it in round-robin fashion. */
1161 if (++idx < pool_offset)
1162 continue;
1163 nr_free = shrink_pages;
1164 /* OK to use static buffer since global mutex is held. */
1165 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free, true);
1166 freed += nr_free - shrink_pages;
1167
1168 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1169 p->pool->dev_name, p->pool->name, current->pid,
1170 nr_free, shrink_pages);
1171 }
1172out:
1173 mutex_unlock(&_manager->lock);
1174 return freed;
1175}
1176
1177static unsigned long
1178ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1179{
1180 struct device_pools *p;
1181 unsigned long count = 0;
1182
1183 if (!mutex_trylock(&_manager->lock))
1184 return 0;
1185 list_for_each_entry(p, &_manager->pools, pools)
1186 count += p->pool->npages_free;
1187 mutex_unlock(&_manager->lock);
1188 return count;
1189}
1190
1191static int ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1192{
1193 manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1194 manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1195 manager->mm_shrink.seeks = 1;
1196 return register_shrinker(&manager->mm_shrink);
1197}
1198
1199static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1200{
1201 unregister_shrinker(&manager->mm_shrink);
1202}
1203
1204int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1205{
1206 int ret;
1207
1208 WARN_ON(_manager);
1209
1210 pr_info("Initializing DMA pool allocator\n");
1211
1212 _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1213 if (!_manager)
1214 return -ENOMEM;
1215
1216 mutex_init(&_manager->lock);
1217 INIT_LIST_HEAD(&_manager->pools);
1218
1219 _manager->options.max_size = max_pages;
1220 _manager->options.small = SMALL_ALLOCATION;
1221 _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1222
1223 /* This takes care of auto-freeing the _manager */
1224 ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1225 &glob->kobj, "dma_pool");
1226 if (unlikely(ret != 0))
1227 goto error;
1228
1229 ret = ttm_dma_pool_mm_shrink_init(_manager);
1230 if (unlikely(ret != 0))
1231 goto error;
1232 return 0;
1233
1234error:
1235 kobject_put(&_manager->kobj);
1236 _manager = NULL;
1237 return ret;
1238}
1239
1240void ttm_dma_page_alloc_fini(void)
1241{
1242 struct device_pools *p, *t;
1243
1244 pr_info("Finalizing DMA pool allocator\n");
1245 ttm_dma_pool_mm_shrink_fini(_manager);
1246
1247 list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1248 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1249 current->pid);
1250 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1251 ttm_dma_pool_match, p->pool));
1252 ttm_dma_free_pool(p->dev, p->pool->type);
1253 }
1254 kobject_put(&_manager->kobj);
1255 _manager = NULL;
1256}
1257
1258int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1259{
1260 struct device_pools *p;
1261 struct dma_pool *pool = NULL;
1262
1263 if (!_manager) {
1264 seq_printf(m, "No pool allocator running.\n");
1265 return 0;
1266 }
1267 seq_printf(m, " pool refills pages freed inuse available name\n");
1268 mutex_lock(&_manager->lock);
1269 list_for_each_entry(p, &_manager->pools, pools) {
1270 struct device *dev = p->dev;
1271 if (!dev)
1272 continue;
1273 pool = p->pool;
1274 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1275 pool->name, pool->nrefills,
1276 pool->nfrees, pool->npages_in_use,
1277 pool->npages_free,
1278 pool->dev_name);
1279 }
1280 mutex_unlock(&_manager->lock);
1281 return 0;
1282}
1283EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1284
1285#endif