treewide: kmalloc() -> kmalloc_array()
[linux-2.6-block.git] / drivers / gpu / drm / ttm / ttm_page_alloc_dma.c
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
66 enum 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  */
102 struct 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  */
127 struct 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
140 struct 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  */
153 struct 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  */
168 struct 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
177 static struct ttm_pool_manager *_manager;
178
179 static struct attribute ttm_page_pool_max = {
180         .name = "pool_max_size",
181         .mode = S_IRUGO | S_IWUSR
182 };
183 static struct attribute ttm_page_pool_small = {
184         .name = "pool_small_allocation",
185         .mode = S_IRUGO | S_IWUSR
186 };
187 static struct attribute ttm_page_pool_alloc_size = {
188         .name = "pool_allocation_size",
189         .mode = S_IRUGO | S_IWUSR
190 };
191
192 static 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
199 static 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
206 static 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
241 static 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
260 static const struct sysfs_ops ttm_pool_sysfs_ops = {
261         .show = &ttm_pool_show,
262         .store = &ttm_pool_store,
263 };
264
265 static 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
272 static 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
283 static 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
294 static 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
306 static 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
326 static 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 }
335 static 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 }
364 static 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
380 static 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. */
389 static 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
409 static 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  **/
443 static 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_array(npages_to_free,
467                                               sizeof(struct page *),
468                                               GFP_KERNEL);
469
470         if (!pages_to_free) {
471                 pr_debug("%s: Failed to allocate memory for pool free operation\n",
472                        pool->dev_name);
473                 return 0;
474         }
475         INIT_LIST_HEAD(&d_pages);
476 restart:
477         spin_lock_irqsave(&pool->lock, irq_flags);
478
479         /* We picking the oldest ones off the list */
480         list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
481                                          page_list) {
482                 if (freed_pages >= npages_to_free)
483                         break;
484
485                 /* Move the dma_page from one list to another. */
486                 list_move(&dma_p->page_list, &d_pages);
487
488                 pages_to_free[freed_pages++] = dma_p->p;
489                 /* We can only remove NUM_PAGES_TO_ALLOC at a time. */
490                 if (freed_pages >= NUM_PAGES_TO_ALLOC) {
491
492                         ttm_pool_update_free_locked(pool, freed_pages);
493                         /**
494                          * Because changing page caching is costly
495                          * we unlock the pool to prevent stalling.
496                          */
497                         spin_unlock_irqrestore(&pool->lock, irq_flags);
498
499                         ttm_dma_pages_put(pool, &d_pages, pages_to_free,
500                                           freed_pages);
501
502                         INIT_LIST_HEAD(&d_pages);
503
504                         if (likely(nr_free != FREE_ALL_PAGES))
505                                 nr_free -= freed_pages;
506
507                         if (NUM_PAGES_TO_ALLOC >= nr_free)
508                                 npages_to_free = nr_free;
509                         else
510                                 npages_to_free = NUM_PAGES_TO_ALLOC;
511
512                         freed_pages = 0;
513
514                         /* free all so restart the processing */
515                         if (nr_free)
516                                 goto restart;
517
518                         /* Not allowed to fall through or break because
519                          * following context is inside spinlock while we are
520                          * outside here.
521                          */
522                         goto out;
523
524                 }
525         }
526
527         /* remove range of pages from the pool */
528         if (freed_pages) {
529                 ttm_pool_update_free_locked(pool, freed_pages);
530                 nr_free -= freed_pages;
531         }
532
533         spin_unlock_irqrestore(&pool->lock, irq_flags);
534
535         if (freed_pages)
536                 ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
537 out:
538         if (pages_to_free != static_buf)
539                 kfree(pages_to_free);
540         return nr_free;
541 }
542
543 static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
544 {
545         struct device_pools *p;
546         struct dma_pool *pool;
547
548         if (!dev)
549                 return;
550
551         mutex_lock(&_manager->lock);
552         list_for_each_entry_reverse(p, &_manager->pools, pools) {
553                 if (p->dev != dev)
554                         continue;
555                 pool = p->pool;
556                 if (pool->type != type)
557                         continue;
558
559                 list_del(&p->pools);
560                 kfree(p);
561                 _manager->npools--;
562                 break;
563         }
564         list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
565                 if (pool->type != type)
566                         continue;
567                 /* Takes a spinlock.. */
568                 /* OK to use static buffer since global mutex is held. */
569                 ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, true);
570                 WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
571                 /* This code path is called after _all_ references to the
572                  * struct device has been dropped - so nobody should be
573                  * touching it. In case somebody is trying to _add_ we are
574                  * guarded by the mutex. */
575                 list_del(&pool->pools);
576                 kfree(pool);
577                 break;
578         }
579         mutex_unlock(&_manager->lock);
580 }
581
582 /*
583  * On free-ing of the 'struct device' this deconstructor is run.
584  * Albeit the pool might have already been freed earlier.
585  */
586 static void ttm_dma_pool_release(struct device *dev, void *res)
587 {
588         struct dma_pool *pool = *(struct dma_pool **)res;
589
590         if (pool)
591                 ttm_dma_free_pool(dev, pool->type);
592 }
593
594 static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
595 {
596         return *(struct dma_pool **)res == match_data;
597 }
598
599 static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
600                                           enum pool_type type)
601 {
602         const char *n[] = {"wc", "uc", "cached", " dma32", "huge"};
603         enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_HUGE};
604         struct device_pools *sec_pool = NULL;
605         struct dma_pool *pool = NULL, **ptr;
606         unsigned i;
607         int ret = -ENODEV;
608         char *p;
609
610         if (!dev)
611                 return NULL;
612
613         ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
614         if (!ptr)
615                 return NULL;
616
617         ret = -ENOMEM;
618
619         pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
620                             dev_to_node(dev));
621         if (!pool)
622                 goto err_mem;
623
624         sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
625                                 dev_to_node(dev));
626         if (!sec_pool)
627                 goto err_mem;
628
629         INIT_LIST_HEAD(&sec_pool->pools);
630         sec_pool->dev = dev;
631         sec_pool->pool =  pool;
632
633         INIT_LIST_HEAD(&pool->free_list);
634         INIT_LIST_HEAD(&pool->pools);
635         spin_lock_init(&pool->lock);
636         pool->dev = dev;
637         pool->npages_free = pool->npages_in_use = 0;
638         pool->nfrees = 0;
639         pool->gfp_flags = flags;
640         if (type & IS_HUGE)
641 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
642                 pool->size = HPAGE_PMD_SIZE;
643 #else
644                 BUG();
645 #endif
646         else
647                 pool->size = PAGE_SIZE;
648         pool->type = type;
649         pool->nrefills = 0;
650         p = pool->name;
651         for (i = 0; i < ARRAY_SIZE(t); i++) {
652                 if (type & t[i]) {
653                         p += snprintf(p, sizeof(pool->name) - (p - pool->name),
654                                       "%s", n[i]);
655                 }
656         }
657         *p = 0;
658         /* We copy the name for pr_ calls b/c when dma_pool_destroy is called
659          * - the kobj->name has already been deallocated.*/
660         snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
661                  dev_driver_string(dev), dev_name(dev));
662         mutex_lock(&_manager->lock);
663         /* You can get the dma_pool from either the global: */
664         list_add(&sec_pool->pools, &_manager->pools);
665         _manager->npools++;
666         /* or from 'struct device': */
667         list_add(&pool->pools, &dev->dma_pools);
668         mutex_unlock(&_manager->lock);
669
670         *ptr = pool;
671         devres_add(dev, ptr);
672
673         return pool;
674 err_mem:
675         devres_free(ptr);
676         kfree(sec_pool);
677         kfree(pool);
678         return ERR_PTR(ret);
679 }
680
681 static struct dma_pool *ttm_dma_find_pool(struct device *dev,
682                                           enum pool_type type)
683 {
684         struct dma_pool *pool, *tmp;
685
686         if (type == IS_UNDEFINED)
687                 return NULL;
688
689         /* NB: We iterate on the 'struct dev' which has no spinlock, but
690          * it does have a kref which we have taken. The kref is taken during
691          * graphic driver loading - in the drm_pci_init it calls either
692          * pci_dev_get or pci_register_driver which both end up taking a kref
693          * on 'struct device'.
694          *
695          * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
696          * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
697          * thing is at that point of time there are no pages associated with the
698          * driver so this function will not be called.
699          */
700         list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools)
701                 if (pool->type == type)
702                         return pool;
703         return NULL;
704 }
705
706 /*
707  * Free pages the pages that failed to change the caching state. If there
708  * are pages that have changed their caching state already put them to the
709  * pool.
710  */
711 static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
712                                                  struct list_head *d_pages,
713                                                  struct page **failed_pages,
714                                                  unsigned cpages)
715 {
716         struct dma_page *d_page, *tmp;
717         struct page *p;
718         unsigned i = 0;
719
720         p = failed_pages[0];
721         if (!p)
722                 return;
723         /* Find the failed page. */
724         list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
725                 if (d_page->p != p)
726                         continue;
727                 /* .. and then progress over the full list. */
728                 list_del(&d_page->page_list);
729                 __ttm_dma_free_page(pool, d_page);
730                 if (++i < cpages)
731                         p = failed_pages[i];
732                 else
733                         break;
734         }
735
736 }
737
738 /*
739  * Allocate 'count' pages, and put 'need' number of them on the
740  * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
741  * The full list of pages should also be on 'd_pages'.
742  * We return zero for success, and negative numbers as errors.
743  */
744 static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
745                                         struct list_head *d_pages,
746                                         unsigned count)
747 {
748         struct page **caching_array;
749         struct dma_page *dma_p;
750         struct page *p;
751         int r = 0;
752         unsigned i, j, npages, cpages;
753         unsigned max_cpages = min(count,
754                         (unsigned)(PAGE_SIZE/sizeof(struct page *)));
755
756         /* allocate array for page caching change */
757         caching_array = kmalloc_array(max_cpages, sizeof(struct page *),
758                                       GFP_KERNEL);
759
760         if (!caching_array) {
761                 pr_debug("%s: Unable to allocate table for new pages\n",
762                        pool->dev_name);
763                 return -ENOMEM;
764         }
765
766         if (count > 1)
767                 pr_debug("%s: (%s:%d) Getting %d pages\n",
768                          pool->dev_name, pool->name, current->pid, count);
769
770         for (i = 0, cpages = 0; i < count; ++i) {
771                 dma_p = __ttm_dma_alloc_page(pool);
772                 if (!dma_p) {
773                         pr_debug("%s: Unable to get page %u\n",
774                                  pool->dev_name, i);
775
776                         /* store already allocated pages in the pool after
777                          * setting the caching state */
778                         if (cpages) {
779                                 r = ttm_set_pages_caching(pool, caching_array,
780                                                           cpages);
781                                 if (r)
782                                         ttm_dma_handle_caching_state_failure(
783                                                 pool, d_pages, caching_array,
784                                                 cpages);
785                         }
786                         r = -ENOMEM;
787                         goto out;
788                 }
789                 p = dma_p->p;
790                 list_add(&dma_p->page_list, d_pages);
791
792 #ifdef CONFIG_HIGHMEM
793                 /* gfp flags of highmem page should never be dma32 so we
794                  * we should be fine in such case
795                  */
796                 if (PageHighMem(p))
797                         continue;
798 #endif
799
800                 npages = pool->size / PAGE_SIZE;
801                 for (j = 0; j < npages; ++j) {
802                         caching_array[cpages++] = p + j;
803                         if (cpages == max_cpages) {
804                                 /* Note: Cannot hold the spinlock */
805                                 r = ttm_set_pages_caching(pool, caching_array,
806                                                           cpages);
807                                 if (r) {
808                                         ttm_dma_handle_caching_state_failure(
809                                              pool, d_pages, caching_array,
810                                              cpages);
811                                         goto out;
812                                 }
813                                 cpages = 0;
814                         }
815                 }
816         }
817
818         if (cpages) {
819                 r = ttm_set_pages_caching(pool, caching_array, cpages);
820                 if (r)
821                         ttm_dma_handle_caching_state_failure(pool, d_pages,
822                                         caching_array, cpages);
823         }
824 out:
825         kfree(caching_array);
826         return r;
827 }
828
829 /*
830  * @return count of pages still required to fulfill the request.
831  */
832 static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
833                                          unsigned long *irq_flags)
834 {
835         unsigned count = _manager->options.small;
836         int r = pool->npages_free;
837
838         if (count > pool->npages_free) {
839                 struct list_head d_pages;
840
841                 INIT_LIST_HEAD(&d_pages);
842
843                 spin_unlock_irqrestore(&pool->lock, *irq_flags);
844
845                 /* Returns how many more are neccessary to fulfill the
846                  * request. */
847                 r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
848
849                 spin_lock_irqsave(&pool->lock, *irq_flags);
850                 if (!r) {
851                         /* Add the fresh to the end.. */
852                         list_splice(&d_pages, &pool->free_list);
853                         ++pool->nrefills;
854                         pool->npages_free += count;
855                         r = count;
856                 } else {
857                         struct dma_page *d_page;
858                         unsigned cpages = 0;
859
860                         pr_debug("%s: Failed to fill %s pool (r:%d)!\n",
861                                  pool->dev_name, pool->name, r);
862
863                         list_for_each_entry(d_page, &d_pages, page_list) {
864                                 cpages++;
865                         }
866                         list_splice_tail(&d_pages, &pool->free_list);
867                         pool->npages_free += cpages;
868                         r = cpages;
869                 }
870         }
871         return r;
872 }
873
874 /*
875  * The populate list is actually a stack (not that is matters as TTM
876  * allocates one page at a time.
877  * return dma_page pointer if success, otherwise NULL.
878  */
879 static struct dma_page *ttm_dma_pool_get_pages(struct dma_pool *pool,
880                                   struct ttm_dma_tt *ttm_dma,
881                                   unsigned index)
882 {
883         struct dma_page *d_page = NULL;
884         struct ttm_tt *ttm = &ttm_dma->ttm;
885         unsigned long irq_flags;
886         int count;
887
888         spin_lock_irqsave(&pool->lock, irq_flags);
889         count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
890         if (count) {
891                 d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
892                 ttm->pages[index] = d_page->p;
893                 ttm_dma->dma_address[index] = d_page->dma;
894                 list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
895                 pool->npages_in_use += 1;
896                 pool->npages_free -= 1;
897         }
898         spin_unlock_irqrestore(&pool->lock, irq_flags);
899         return d_page;
900 }
901
902 static gfp_t ttm_dma_pool_gfp_flags(struct ttm_dma_tt *ttm_dma, bool huge)
903 {
904         struct ttm_tt *ttm = &ttm_dma->ttm;
905         gfp_t gfp_flags;
906
907         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
908                 gfp_flags = GFP_USER | GFP_DMA32;
909         else
910                 gfp_flags = GFP_HIGHUSER;
911         if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
912                 gfp_flags |= __GFP_ZERO;
913
914         if (huge) {
915                 gfp_flags |= GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
916                         __GFP_KSWAPD_RECLAIM;
917                 gfp_flags &= ~__GFP_MOVABLE;
918                 gfp_flags &= ~__GFP_COMP;
919         }
920
921         if (ttm->page_flags & TTM_PAGE_FLAG_NO_RETRY)
922                 gfp_flags |= __GFP_RETRY_MAYFAIL;
923
924         return gfp_flags;
925 }
926
927 /*
928  * On success pages list will hold count number of correctly
929  * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
930  */
931 int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev,
932                         struct ttm_operation_ctx *ctx)
933 {
934         struct ttm_tt *ttm = &ttm_dma->ttm;
935         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
936         unsigned long num_pages = ttm->num_pages;
937         struct dma_pool *pool;
938         struct dma_page *d_page;
939         enum pool_type type;
940         unsigned i;
941         int ret;
942
943         if (ttm->state != tt_unpopulated)
944                 return 0;
945
946         if (ttm_check_under_lowerlimit(mem_glob, num_pages, ctx))
947                 return -ENOMEM;
948
949         INIT_LIST_HEAD(&ttm_dma->pages_list);
950         i = 0;
951
952         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
953
954 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
955         if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
956                 goto skip_huge;
957
958         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
959         if (!pool) {
960                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, true);
961
962                 pool = ttm_dma_pool_init(dev, gfp_flags, type | IS_HUGE);
963                 if (IS_ERR_OR_NULL(pool))
964                         goto skip_huge;
965         }
966
967         while (num_pages >= HPAGE_PMD_NR) {
968                 unsigned j;
969
970                 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
971                 if (!d_page)
972                         break;
973
974                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
975                                                 pool->size, ctx);
976                 if (unlikely(ret != 0)) {
977                         ttm_dma_unpopulate(ttm_dma, dev);
978                         return -ENOMEM;
979                 }
980
981                 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
982                 for (j = i + 1; j < (i + HPAGE_PMD_NR); ++j) {
983                         ttm->pages[j] = ttm->pages[j - 1] + 1;
984                         ttm_dma->dma_address[j] = ttm_dma->dma_address[j - 1] +
985                                 PAGE_SIZE;
986                 }
987
988                 i += HPAGE_PMD_NR;
989                 num_pages -= HPAGE_PMD_NR;
990         }
991
992 skip_huge:
993 #endif
994
995         pool = ttm_dma_find_pool(dev, type);
996         if (!pool) {
997                 gfp_t gfp_flags = ttm_dma_pool_gfp_flags(ttm_dma, false);
998
999                 pool = ttm_dma_pool_init(dev, gfp_flags, type);
1000                 if (IS_ERR_OR_NULL(pool))
1001                         return -ENOMEM;
1002         }
1003
1004         while (num_pages) {
1005                 d_page = ttm_dma_pool_get_pages(pool, ttm_dma, i);
1006                 if (!d_page) {
1007                         ttm_dma_unpopulate(ttm_dma, dev);
1008                         return -ENOMEM;
1009                 }
1010
1011                 ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
1012                                                 pool->size, ctx);
1013                 if (unlikely(ret != 0)) {
1014                         ttm_dma_unpopulate(ttm_dma, dev);
1015                         return -ENOMEM;
1016                 }
1017
1018                 d_page->vaddr |= VADDR_FLAG_UPDATED_COUNT;
1019                 ++i;
1020                 --num_pages;
1021         }
1022
1023         if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
1024                 ret = ttm_tt_swapin(ttm);
1025                 if (unlikely(ret != 0)) {
1026                         ttm_dma_unpopulate(ttm_dma, dev);
1027                         return ret;
1028                 }
1029         }
1030
1031         ttm->state = tt_unbound;
1032         return 0;
1033 }
1034 EXPORT_SYMBOL_GPL(ttm_dma_populate);
1035
1036 /* Put all pages in pages list to correct pool to wait for reuse */
1037 void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
1038 {
1039         struct ttm_tt *ttm = &ttm_dma->ttm;
1040         struct ttm_mem_global *mem_glob = ttm->bdev->glob->mem_glob;
1041         struct dma_pool *pool;
1042         struct dma_page *d_page, *next;
1043         enum pool_type type;
1044         bool is_cached = false;
1045         unsigned count, i, npages = 0;
1046         unsigned long irq_flags;
1047
1048         type = ttm_to_type(ttm->page_flags, ttm->caching_state);
1049
1050 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1051         pool = ttm_dma_find_pool(dev, type | IS_HUGE);
1052         if (pool) {
1053                 count = 0;
1054                 list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1055                                          page_list) {
1056                         if (!(d_page->vaddr & VADDR_FLAG_HUGE_POOL))
1057                                 continue;
1058
1059                         count++;
1060                         if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1061                                 ttm_mem_global_free_page(mem_glob, d_page->p,
1062                                                          pool->size);
1063                                 d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1064                         }
1065                         ttm_dma_page_put(pool, d_page);
1066                 }
1067
1068                 spin_lock_irqsave(&pool->lock, irq_flags);
1069                 pool->npages_in_use -= count;
1070                 pool->nfrees += count;
1071                 spin_unlock_irqrestore(&pool->lock, irq_flags);
1072         }
1073 #endif
1074
1075         pool = ttm_dma_find_pool(dev, type);
1076         if (!pool)
1077                 return;
1078
1079         is_cached = (ttm_dma_find_pool(pool->dev,
1080                      ttm_to_type(ttm->page_flags, tt_cached)) == pool);
1081
1082         /* make sure pages array match list and count number of pages */
1083         count = 0;
1084         list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list,
1085                                  page_list) {
1086                 ttm->pages[count] = d_page->p;
1087                 count++;
1088
1089                 if (d_page->vaddr & VADDR_FLAG_UPDATED_COUNT) {
1090                         ttm_mem_global_free_page(mem_glob, d_page->p,
1091                                                  pool->size);
1092                         d_page->vaddr &= ~VADDR_FLAG_UPDATED_COUNT;
1093                 }
1094
1095                 if (is_cached)
1096                         ttm_dma_page_put(pool, d_page);
1097         }
1098
1099         spin_lock_irqsave(&pool->lock, irq_flags);
1100         pool->npages_in_use -= count;
1101         if (is_cached) {
1102                 pool->nfrees += count;
1103         } else {
1104                 pool->npages_free += count;
1105                 list_splice(&ttm_dma->pages_list, &pool->free_list);
1106                 /*
1107                  * Wait to have at at least NUM_PAGES_TO_ALLOC number of pages
1108                  * to free in order to minimize calls to set_memory_wb().
1109                  */
1110                 if (pool->npages_free >= (_manager->options.max_size +
1111                                           NUM_PAGES_TO_ALLOC))
1112                         npages = pool->npages_free - _manager->options.max_size;
1113         }
1114         spin_unlock_irqrestore(&pool->lock, irq_flags);
1115
1116         INIT_LIST_HEAD(&ttm_dma->pages_list);
1117         for (i = 0; i < ttm->num_pages; i++) {
1118                 ttm->pages[i] = NULL;
1119                 ttm_dma->dma_address[i] = 0;
1120         }
1121
1122         /* shrink pool if necessary (only on !is_cached pools)*/
1123         if (npages)
1124                 ttm_dma_page_pool_free(pool, npages, false);
1125         ttm->state = tt_unpopulated;
1126 }
1127 EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1128
1129 /**
1130  * Callback for mm to request pool to reduce number of page held.
1131  *
1132  * XXX: (dchinner) Deadlock warning!
1133  *
1134  * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1135  * shrinkers
1136  */
1137 static unsigned long
1138 ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1139 {
1140         static unsigned start_pool;
1141         unsigned idx = 0;
1142         unsigned pool_offset;
1143         unsigned shrink_pages = sc->nr_to_scan;
1144         struct device_pools *p;
1145         unsigned long freed = 0;
1146
1147         if (list_empty(&_manager->pools))
1148                 return SHRINK_STOP;
1149
1150         if (!mutex_trylock(&_manager->lock))
1151                 return SHRINK_STOP;
1152         if (!_manager->npools)
1153                 goto out;
1154         pool_offset = ++start_pool % _manager->npools;
1155         list_for_each_entry(p, &_manager->pools, pools) {
1156                 unsigned nr_free;
1157
1158                 if (!p->dev)
1159                         continue;
1160                 if (shrink_pages == 0)
1161                         break;
1162                 /* Do it in round-robin fashion. */
1163                 if (++idx < pool_offset)
1164                         continue;
1165                 nr_free = shrink_pages;
1166                 /* OK to use static buffer since global mutex is held. */
1167                 shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free, true);
1168                 freed += nr_free - shrink_pages;
1169
1170                 pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1171                          p->pool->dev_name, p->pool->name, current->pid,
1172                          nr_free, shrink_pages);
1173         }
1174 out:
1175         mutex_unlock(&_manager->lock);
1176         return freed;
1177 }
1178
1179 static unsigned long
1180 ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1181 {
1182         struct device_pools *p;
1183         unsigned long count = 0;
1184
1185         if (!mutex_trylock(&_manager->lock))
1186                 return 0;
1187         list_for_each_entry(p, &_manager->pools, pools)
1188                 count += p->pool->npages_free;
1189         mutex_unlock(&_manager->lock);
1190         return count;
1191 }
1192
1193 static int ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1194 {
1195         manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1196         manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1197         manager->mm_shrink.seeks = 1;
1198         return register_shrinker(&manager->mm_shrink);
1199 }
1200
1201 static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1202 {
1203         unregister_shrinker(&manager->mm_shrink);
1204 }
1205
1206 int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1207 {
1208         int ret;
1209
1210         WARN_ON(_manager);
1211
1212         pr_info("Initializing DMA pool allocator\n");
1213
1214         _manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1215         if (!_manager)
1216                 return -ENOMEM;
1217
1218         mutex_init(&_manager->lock);
1219         INIT_LIST_HEAD(&_manager->pools);
1220
1221         _manager->options.max_size = max_pages;
1222         _manager->options.small = SMALL_ALLOCATION;
1223         _manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1224
1225         /* This takes care of auto-freeing the _manager */
1226         ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1227                                    &glob->kobj, "dma_pool");
1228         if (unlikely(ret != 0))
1229                 goto error;
1230
1231         ret = ttm_dma_pool_mm_shrink_init(_manager);
1232         if (unlikely(ret != 0))
1233                 goto error;
1234         return 0;
1235
1236 error:
1237         kobject_put(&_manager->kobj);
1238         _manager = NULL;
1239         return ret;
1240 }
1241
1242 void ttm_dma_page_alloc_fini(void)
1243 {
1244         struct device_pools *p, *t;
1245
1246         pr_info("Finalizing DMA pool allocator\n");
1247         ttm_dma_pool_mm_shrink_fini(_manager);
1248
1249         list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1250                 dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1251                         current->pid);
1252                 WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1253                         ttm_dma_pool_match, p->pool));
1254                 ttm_dma_free_pool(p->dev, p->pool->type);
1255         }
1256         kobject_put(&_manager->kobj);
1257         _manager = NULL;
1258 }
1259
1260 int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1261 {
1262         struct device_pools *p;
1263         struct dma_pool *pool = NULL;
1264
1265         if (!_manager) {
1266                 seq_printf(m, "No pool allocator running.\n");
1267                 return 0;
1268         }
1269         seq_printf(m, "         pool      refills   pages freed    inuse available     name\n");
1270         mutex_lock(&_manager->lock);
1271         list_for_each_entry(p, &_manager->pools, pools) {
1272                 struct device *dev = p->dev;
1273                 if (!dev)
1274                         continue;
1275                 pool = p->pool;
1276                 seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1277                                 pool->name, pool->nrefills,
1278                                 pool->nfrees, pool->npages_in_use,
1279                                 pool->npages_free,
1280                                 pool->dev_name);
1281         }
1282         mutex_unlock(&_manager->lock);
1283         return 0;
1284 }
1285 EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1286
1287 #endif