Loading...
Note: File does not exist in v3.1.
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 d_page->p = virt_to_page(d_page->vaddr);
347 else {
348 kfree(d_page);
349 d_page = NULL;
350 }
351 return d_page;
352}
353static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
354{
355 enum pool_type type = IS_UNDEFINED;
356
357 if (flags & TTM_PAGE_FLAG_DMA32)
358 type |= IS_DMA32;
359 if (cstate == tt_cached)
360 type |= IS_CACHED;
361 else if (cstate == tt_uncached)
362 type |= IS_UC;
363 else
364 type |= IS_WC;
365
366 return type;
367}
368
369static void ttm_pool_update_free_locked(struct dma_pool *pool,
370 unsigned freed_pages)
371{
372 pool->npages_free -= freed_pages;
373 pool->nfrees += freed_pages;
374
375}
376
377/* set memory back to wb and free the pages. */
378static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
379 struct page *pages[], unsigned npages)
380{
381 struct dma_page *d_page, *tmp;
382
383 /* Don't set WB on WB page pool. */
384 if (npages && !(pool->type & IS_CACHED) &&
385 set_pages_array_wb(pages, npages))
386 pr_err("%s: Failed to set %d pages to wb!\n",
387 pool->dev_name, npages);
388
389 list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
390 list_del(&d_page->page_list);
391 __ttm_dma_free_page(pool, d_page);
392 }
393}
394
395static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
396{
397 /* Don't set WB on WB page pool. */
398 if (!(pool->type & IS_CACHED) && set_pages_array_wb(&d_page->p, 1))
399 pr_err("%s: Failed to set %d pages to wb!\n",
400 pool->dev_name, 1);
401
402 list_del(&d_page->page_list);
403 __ttm_dma_free_page(pool, d_page);
404}
405
406/*
407 * Free pages from pool.
408 *
409 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
410 * number of pages in one go.
411 *
412 * @pool: to free the pages from
413 * @nr_free: If set to true will free all pages in pool
414 **/
415static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free)
416{
417 unsigned long irq_flags;
418 struct dma_page *dma_p, *tmp;
419 struct page **pages_to_free;
420 struct list_head d_pages;
421 unsigned freed_pages = 0,
422 npages_to_free = nr_free;
423
424 if (NUM_PAGES_TO_ALLOC < nr_free)
425 npages_to_free = NUM_PAGES_TO_ALLOC;
426#if 0
427 if (nr_free > 1) {
428 pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
429 pool->dev_name, pool->name, current->pid,
430 npages_to_free, nr_free);
431 }
432#endif
433 pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
434 GFP_KERNEL);
435
436 if (!pages_to_free) {
437 pr_err("%s: Failed to allocate memory for pool free operation\n",
438 pool->dev_name);
439 return 0;
440 }
441 INIT_LIST_HEAD(&d_pages);
442restart:
443 spin_lock_irqsave(&pool->lock, irq_flags);
444
445 /* We picking the oldest ones off the list */
446 list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
447 page_list) {
448 if (freed_pages >= npages_to_free)
449 break;
450
451 /* Move the dma_page from one list to another. */
452 list_move(&dma_p->page_list, &d_pages);
453
454 pages_to_free[freed_pages++] = dma_p->p;
455 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
456 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
457
458 ttm_pool_update_free_locked(pool, freed_pages);
459 /**
460 * Because changing page caching is costly
461 * we unlock the pool to prevent stalling.
462 */
463 spin_unlock_irqrestore(&pool->lock, irq_flags);
464
465 ttm_dma_pages_put(pool, &d_pages, pages_to_free,
466 freed_pages);
467
468 INIT_LIST_HEAD(&d_pages);
469
470 if (likely(nr_free != FREE_ALL_PAGES))
471 nr_free -= freed_pages;
472
473 if (NUM_PAGES_TO_ALLOC >= nr_free)
474 npages_to_free = nr_free;
475 else
476 npages_to_free = NUM_PAGES_TO_ALLOC;
477
478 freed_pages = 0;
479
480 /* free all so restart the processing */
481 if (nr_free)
482 goto restart;
483
484 /* Not allowed to fall through or break because
485 * following context is inside spinlock while we are
486 * outside here.
487 */
488 goto out;
489
490 }
491 }
492
493 /* remove range of pages from the pool */
494 if (freed_pages) {
495 ttm_pool_update_free_locked(pool, freed_pages);
496 nr_free -= freed_pages;
497 }
498
499 spin_unlock_irqrestore(&pool->lock, irq_flags);
500
501 if (freed_pages)
502 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
503out:
504 kfree(pages_to_free);
505 return nr_free;
506}
507
508static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
509{
510 struct device_pools *p;
511 struct dma_pool *pool;
512
513 if (!dev)
514 return;
515
516 mutex_lock(&_manager->lock);
517 list_for_each_entry_reverse(p, &_manager->pools, pools) {
518 if (p->dev != dev)
519 continue;
520 pool = p->pool;
521 if (pool->type != type)
522 continue;
523
524 list_del(&p->pools);
525 kfree(p);
526 _manager->npools--;
527 break;
528 }
529 list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
530 if (pool->type != type)
531 continue;
532 /* Takes a spinlock.. */
533 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES);
534 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
535 /* This code path is called after _all_ references to the
536 * struct device has been dropped - so nobody should be
537 * touching it. In case somebody is trying to _add_ we are
538 * guarded by the mutex. */
539 list_del(&pool->pools);
540 kfree(pool);
541 break;
542 }
543 mutex_unlock(&_manager->lock);
544}
545
546/*
547 * On free-ing of the 'struct device' this deconstructor is run.
548 * Albeit the pool might have already been freed earlier.
549 */
550static void ttm_dma_pool_release(struct device *dev, void *res)
551{
552 struct dma_pool *pool = *(struct dma_pool **)res;
553
554 if (pool)
555 ttm_dma_free_pool(dev, pool->type);
556}
557
558static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
559{
560 return *(struct dma_pool **)res == match_data;
561}
562
563static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
564 enum pool_type type)
565{
566 char *n[] = {"wc", "uc", "cached", " dma32", "unknown",};
567 enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_UNDEFINED};
568 struct device_pools *sec_pool = NULL;
569 struct dma_pool *pool = NULL, **ptr;
570 unsigned i;
571 int ret = -ENODEV;
572 char *p;
573
574 if (!dev)
575 return NULL;
576
577 ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
578 if (!ptr)
579 return NULL;
580
581 ret = -ENOMEM;
582
583 pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
584 dev_to_node(dev));
585 if (!pool)
586 goto err_mem;
587
588 sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
589 dev_to_node(dev));
590 if (!sec_pool)
591 goto err_mem;
592
593 INIT_LIST_HEAD(&sec_pool->pools);
594 sec_pool->dev = dev;
595 sec_pool->pool = pool;
596
597 INIT_LIST_HEAD(&pool->free_list);
598 INIT_LIST_HEAD(&pool->inuse_list);
599 INIT_LIST_HEAD(&pool->pools);
600 spin_lock_init(&pool->lock);
601 pool->dev = dev;
602 pool->npages_free = pool->npages_in_use = 0;
603 pool->nfrees = 0;
604 pool->gfp_flags = flags;
605 pool->size = PAGE_SIZE;
606 pool->type = type;
607 pool->nrefills = 0;
608 p = pool->name;
609 for (i = 0; i < 5; i++) {
610 if (type & t[i]) {
611 p += snprintf(p, sizeof(pool->name) - (p - pool->name),
612 "%s", n[i]);
613 }
614 }
615 *p = 0;
616 /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
617 * - the kobj->name has already been deallocated.*/
618 snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
619 dev_driver_string(dev), dev_name(dev));
620 mutex_lock(&_manager->lock);
621 /* You can get the dma_pool from either the global: */
622 list_add(&sec_pool->pools, &_manager->pools);
623 _manager->npools++;
624 /* or from 'struct device': */
625 list_add(&pool->pools, &dev->dma_pools);
626 mutex_unlock(&_manager->lock);
627
628 *ptr = pool;
629 devres_add(dev, ptr);
630
631 return pool;
632err_mem:
633 devres_free(ptr);
634 kfree(sec_pool);
635 kfree(pool);
636 return ERR_PTR(ret);
637}
638
639static struct dma_pool *ttm_dma_find_pool(struct device *dev,
640 enum pool_type type)
641{
642 struct dma_pool *pool, *tmp, *found = NULL;
643
644 if (type == IS_UNDEFINED)
645 return found;
646
647 /* NB: We iterate on the 'struct dev' which has no spinlock, but
648 * it does have a kref which we have taken. The kref is taken during
649 * graphic driver loading - in the drm_pci_init it calls either
650 * pci_dev_get or pci_register_driver which both end up taking a kref
651 * on 'struct device'.
652 *
653 * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
654 * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
655 * thing is at that point of time there are no pages associated with the
656 * driver so this function will not be called.
657 */
658 list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools) {
659 if (pool->type != type)
660 continue;
661 found = pool;
662 break;
663 }
664 return found;
665}
666
667/*
668 * Free pages the pages that failed to change the caching state. If there
669 * are pages that have changed their caching state already put them to the
670 * pool.
671 */
672static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
673 struct list_head *d_pages,
674 struct page **failed_pages,
675 unsigned cpages)
676{
677 struct dma_page *d_page, *tmp;
678 struct page *p;
679 unsigned i = 0;
680
681 p = failed_pages[0];
682 if (!p)
683 return;
684 /* Find the failed page. */
685 list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
686 if (d_page->p != p)
687 continue;
688 /* .. and then progress over the full list. */
689 list_del(&d_page->page_list);
690 __ttm_dma_free_page(pool, d_page);
691 if (++i < cpages)
692 p = failed_pages[i];
693 else
694 break;
695 }
696
697}
698
699/*
700 * Allocate 'count' pages, and put 'need' number of them on the
701 * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
702 * The full list of pages should also be on 'd_pages'.
703 * We return zero for success, and negative numbers as errors.
704 */
705static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
706 struct list_head *d_pages,
707 unsigned count)
708{
709 struct page **caching_array;
710 struct dma_page *dma_p;
711 struct page *p;
712 int r = 0;
713 unsigned i, cpages;
714 unsigned max_cpages = min(count,
715 (unsigned)(PAGE_SIZE/sizeof(struct page *)));
716
717 /* allocate array for page caching change */
718 caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
719
720 if (!caching_array) {
721 pr_err("%s: Unable to allocate table for new pages\n",
722 pool->dev_name);
723 return -ENOMEM;
724 }
725
726 if (count > 1) {
727 pr_debug("%s: (%s:%d) Getting %d pages\n",
728 pool->dev_name, pool->name, current->pid, count);
729 }
730
731 for (i = 0, cpages = 0; i < count; ++i) {
732 dma_p = __ttm_dma_alloc_page(pool);
733 if (!dma_p) {
734 pr_err("%s: Unable to get page %u\n",
735 pool->dev_name, i);
736
737 /* store already allocated pages in the pool after
738 * setting the caching state */
739 if (cpages) {
740 r = ttm_set_pages_caching(pool, caching_array,
741 cpages);
742 if (r)
743 ttm_dma_handle_caching_state_failure(
744 pool, d_pages, caching_array,
745 cpages);
746 }
747 r = -ENOMEM;
748 goto out;
749 }
750 p = dma_p->p;
751#ifdef CONFIG_HIGHMEM
752 /* gfp flags of highmem page should never be dma32 so we
753 * we should be fine in such case
754 */
755 if (!PageHighMem(p))
756#endif
757 {
758 caching_array[cpages++] = p;
759 if (cpages == max_cpages) {
760 /* Note: Cannot hold the spinlock */
761 r = ttm_set_pages_caching(pool, caching_array,
762 cpages);
763 if (r) {
764 ttm_dma_handle_caching_state_failure(
765 pool, d_pages, caching_array,
766 cpages);
767 goto out;
768 }
769 cpages = 0;
770 }
771 }
772 list_add(&dma_p->page_list, d_pages);
773 }
774
775 if (cpages) {
776 r = ttm_set_pages_caching(pool, caching_array, cpages);
777 if (r)
778 ttm_dma_handle_caching_state_failure(pool, d_pages,
779 caching_array, cpages);
780 }
781out:
782 kfree(caching_array);
783 return r;
784}
785
786/*
787 * @return count of pages still required to fulfill the request.
788 */
789static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
790 unsigned long *irq_flags)
791{
792 unsigned count = _manager->options.small;
793 int r = pool->npages_free;
794
795 if (count > pool->npages_free) {
796 struct list_head d_pages;
797
798 INIT_LIST_HEAD(&d_pages);
799
800 spin_unlock_irqrestore(&pool->lock, *irq_flags);
801
802 /* Returns how many more are neccessary to fulfill the
803 * request. */
804 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
805
806 spin_lock_irqsave(&pool->lock, *irq_flags);
807 if (!r) {
808 /* Add the fresh to the end.. */
809 list_splice(&d_pages, &pool->free_list);
810 ++pool->nrefills;
811 pool->npages_free += count;
812 r = count;
813 } else {
814 struct dma_page *d_page;
815 unsigned cpages = 0;
816
817 pr_err("%s: Failed to fill %s pool (r:%d)!\n",
818 pool->dev_name, pool->name, r);
819
820 list_for_each_entry(d_page, &d_pages, page_list) {
821 cpages++;
822 }
823 list_splice_tail(&d_pages, &pool->free_list);
824 pool->npages_free += cpages;
825 r = cpages;
826 }
827 }
828 return r;
829}
830
831/*
832 * @return count of pages still required to fulfill the request.
833 * The populate list is actually a stack (not that is matters as TTM
834 * allocates one page at a time.
835 */
836static int ttm_dma_pool_get_pages(struct dma_pool *pool,
837 struct ttm_dma_tt *ttm_dma,
838 unsigned index)
839{
840 struct dma_page *d_page;
841 struct ttm_tt *ttm = &ttm_dma->ttm;
842 unsigned long irq_flags;
843 int count, r = -ENOMEM;
844
845 spin_lock_irqsave(&pool->lock, irq_flags);
846 count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
847 if (count) {
848 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
849 ttm->pages[index] = d_page->p;
850 ttm_dma->dma_address[index] = d_page->dma;
851 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
852 r = 0;
853 pool->npages_in_use += 1;
854 pool->npages_free -= 1;
855 }
856 spin_unlock_irqrestore(&pool->lock, irq_flags);
857 return r;
858}
859
860/*
861 * On success pages list will hold count number of correctly
862 * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
863 */
864int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev)
865{
866 struct ttm_tt *ttm = &ttm_dma->ttm;
867 struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
868 struct dma_pool *pool;
869 enum pool_type type;
870 unsigned i;
871 gfp_t gfp_flags;
872 int ret;
873
874 if (ttm->state != tt_unpopulated)
875 return 0;
876
877 type = ttm_to_type(ttm->page_flags, ttm->caching_state);
878 if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
879 gfp_flags = GFP_USER | GFP_DMA32;
880 else
881 gfp_flags = GFP_HIGHUSER;
882 if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
883 gfp_flags |= __GFP_ZERO;
884
885 pool = ttm_dma_find_pool(dev, type);
886 if (!pool) {
887 pool = ttm_dma_pool_init(dev, gfp_flags, type);
888 if (IS_ERR_OR_NULL(pool)) {
889 return -ENOMEM;
890 }
891 }
892
893 INIT_LIST_HEAD(&ttm_dma->pages_list);
894 for (i = 0; i < ttm->num_pages; ++i) {
895 ret = ttm_dma_pool_get_pages(pool, ttm_dma, i);
896 if (ret != 0) {
897 ttm_dma_unpopulate(ttm_dma, dev);
898 return -ENOMEM;
899 }
900
901 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
902 false, false);
903 if (unlikely(ret != 0)) {
904 ttm_dma_unpopulate(ttm_dma, dev);
905 return -ENOMEM;
906 }
907 }
908
909 if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
910 ret = ttm_tt_swapin(ttm);
911 if (unlikely(ret != 0)) {
912 ttm_dma_unpopulate(ttm_dma, dev);
913 return ret;
914 }
915 }
916
917 ttm->state = tt_unbound;
918 return 0;
919}
920EXPORT_SYMBOL_GPL(ttm_dma_populate);
921
922/* Put all pages in pages list to correct pool to wait for reuse */
923void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
924{
925 struct ttm_tt *ttm = &ttm_dma->ttm;
926 struct dma_pool *pool;
927 struct dma_page *d_page, *next;
928 enum pool_type type;
929 bool is_cached = false;
930 unsigned count = 0, i, npages = 0;
931 unsigned long irq_flags;
932
933 type = ttm_to_type(ttm->page_flags, ttm->caching_state);
934 pool = ttm_dma_find_pool(dev, type);
935 if (!pool)
936 return;
937
938 is_cached = (ttm_dma_find_pool(pool->dev,
939 ttm_to_type(ttm->page_flags, tt_cached)) == pool);
940
941 /* make sure pages array match list and count number of pages */
942 list_for_each_entry(d_page, &ttm_dma->pages_list, page_list) {
943 ttm->pages[count] = d_page->p;
944 count++;
945 }
946
947 spin_lock_irqsave(&pool->lock, irq_flags);
948 pool->npages_in_use -= count;
949 if (is_cached) {
950 pool->nfrees += count;
951 } else {
952 pool->npages_free += count;
953 list_splice(&ttm_dma->pages_list, &pool->free_list);
954 npages = count;
955 if (pool->npages_free > _manager->options.max_size) {
956 npages = pool->npages_free - _manager->options.max_size;
957 /* free at least NUM_PAGES_TO_ALLOC number of pages
958 * to reduce calls to set_memory_wb */
959 if (npages < NUM_PAGES_TO_ALLOC)
960 npages = NUM_PAGES_TO_ALLOC;
961 }
962 }
963 spin_unlock_irqrestore(&pool->lock, irq_flags);
964
965 if (is_cached) {
966 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list, page_list) {
967 ttm_mem_global_free_page(ttm->glob->mem_glob,
968 d_page->p);
969 ttm_dma_page_put(pool, d_page);
970 }
971 } else {
972 for (i = 0; i < count; i++) {
973 ttm_mem_global_free_page(ttm->glob->mem_glob,
974 ttm->pages[i]);
975 }
976 }
977
978 INIT_LIST_HEAD(&ttm_dma->pages_list);
979 for (i = 0; i < ttm->num_pages; i++) {
980 ttm->pages[i] = NULL;
981 ttm_dma->dma_address[i] = 0;
982 }
983
984 /* shrink pool if necessary (only on !is_cached pools)*/
985 if (npages)
986 ttm_dma_page_pool_free(pool, npages);
987 ttm->state = tt_unpopulated;
988}
989EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
990
991/**
992 * Callback for mm to request pool to reduce number of page held.
993 *
994 * XXX: (dchinner) Deadlock warning!
995 *
996 * ttm_dma_page_pool_free() does GFP_KERNEL memory allocation, and so attention
997 * needs to be paid to sc->gfp_mask to determine if this can be done or not.
998 * GFP_KERNEL memory allocation in a GFP_ATOMIC reclaim context woul dbe really
999 * bad.
1000 *
1001 * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1002 * shrinkers
1003 */
1004static unsigned long
1005ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1006{
1007 static atomic_t start_pool = ATOMIC_INIT(0);
1008 unsigned idx = 0;
1009 unsigned pool_offset = atomic_add_return(1, &start_pool);
1010 unsigned shrink_pages = sc->nr_to_scan;
1011 struct device_pools *p;
1012 unsigned long freed = 0;
1013
1014 if (list_empty(&_manager->pools))
1015 return SHRINK_STOP;
1016
1017 mutex_lock(&_manager->lock);
1018 pool_offset = pool_offset % _manager->npools;
1019 list_for_each_entry(p, &_manager->pools, pools) {
1020 unsigned nr_free;
1021
1022 if (!p->dev)
1023 continue;
1024 if (shrink_pages == 0)
1025 break;
1026 /* Do it in round-robin fashion. */
1027 if (++idx < pool_offset)
1028 continue;
1029 nr_free = shrink_pages;
1030 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free);
1031 freed += nr_free - shrink_pages;
1032
1033 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1034 p->pool->dev_name, p->pool->name, current->pid,
1035 nr_free, shrink_pages);
1036 }
1037 mutex_unlock(&_manager->lock);
1038 return freed;
1039}
1040
1041static unsigned long
1042ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1043{
1044 struct device_pools *p;
1045 unsigned long count = 0;
1046
1047 mutex_lock(&_manager->lock);
1048 list_for_each_entry(p, &_manager->pools, pools)
1049 count += p->pool->npages_free;
1050 mutex_unlock(&_manager->lock);
1051 return count;
1052}
1053
1054static void ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1055{
1056 manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1057 manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1058 manager->mm_shrink.seeks = 1;
1059 register_shrinker(&manager->mm_shrink);
1060}
1061
1062static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1063{
1064 unregister_shrinker(&manager->mm_shrink);
1065}
1066
1067int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1068{
1069 int ret = -ENOMEM;
1070
1071 WARN_ON(_manager);
1072
1073 pr_info("Initializing DMA pool allocator\n");
1074
1075 _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1076 if (!_manager)
1077 goto err;
1078
1079 mutex_init(&_manager->lock);
1080 INIT_LIST_HEAD(&_manager->pools);
1081
1082 _manager->options.max_size = max_pages;
1083 _manager->options.small = SMALL_ALLOCATION;
1084 _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1085
1086 /* This takes care of auto-freeing the _manager */
1087 ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1088 &glob->kobj, "dma_pool");
1089 if (unlikely(ret != 0)) {
1090 kobject_put(&_manager->kobj);
1091 goto err;
1092 }
1093 ttm_dma_pool_mm_shrink_init(_manager);
1094 return 0;
1095err:
1096 return ret;
1097}
1098
1099void ttm_dma_page_alloc_fini(void)
1100{
1101 struct device_pools *p, *t;
1102
1103 pr_info("Finalizing DMA pool allocator\n");
1104 ttm_dma_pool_mm_shrink_fini(_manager);
1105
1106 list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1107 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1108 current->pid);
1109 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1110 ttm_dma_pool_match, p->pool));
1111 ttm_dma_free_pool(p->dev, p->pool->type);
1112 }
1113 kobject_put(&_manager->kobj);
1114 _manager = NULL;
1115}
1116
1117int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1118{
1119 struct device_pools *p;
1120 struct dma_pool *pool = NULL;
1121 char *h[] = {"pool", "refills", "pages freed", "inuse", "available",
1122 "name", "virt", "busaddr"};
1123
1124 if (!_manager) {
1125 seq_printf(m, "No pool allocator running.\n");
1126 return 0;
1127 }
1128 seq_printf(m, "%13s %12s %13s %8s %8s %8s\n",
1129 h[0], h[1], h[2], h[3], h[4], h[5]);
1130 mutex_lock(&_manager->lock);
1131 list_for_each_entry(p, &_manager->pools, pools) {
1132 struct device *dev = p->dev;
1133 if (!dev)
1134 continue;
1135 pool = p->pool;
1136 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1137 pool->name, pool->nrefills,
1138 pool->nfrees, pool->npages_in_use,
1139 pool->npages_free,
1140 pool->dev_name);
1141 }
1142 mutex_unlock(&_manager->lock);
1143 return 0;
1144}
1145EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1146
1147#endif