Merge tag 'tty-4.18-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/tty
[sfrench/cifs-2.6.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(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);
475 restart:
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);
536 out:
537         if (pages_to_free != static_buf)
538                 kfree(pages_to_free);
539         return nr_free;
540 }
541
542 static 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  */
585 static 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
593 static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
594 {
595         return *(struct dma_pool **)res == match_data;
596 }
597
598 static 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;
673 err_mem:
674         devres_free(ptr);
675         kfree(sec_pool);
676         kfree(pool);
677         return ERR_PTR(ret);
678 }
679
680 static 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  */
710 static 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  */
743 static 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         }
822 out:
823         kfree(caching_array);
824         return r;
825 }
826
827 /*
828  * @return count of pages still required to fulfill the request.
829  */
830 static 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  */
877 static 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
900 static 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  */
929 int 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
990 skip_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 }
1032 EXPORT_SYMBOL_GPL(ttm_dma_populate);
1033
1034 /* Put all pages in pages list to correct pool to wait for reuse */
1035 void 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 }
1125 EXPORT_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  */
1135 static unsigned long
1136 ttm_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         }
1172 out:
1173         mutex_unlock(&_manager->lock);
1174         return freed;
1175 }
1176
1177 static unsigned long
1178 ttm_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
1191 static 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
1199 static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1200 {
1201         unregister_shrinker(&manager->mm_shrink);
1202 }
1203
1204 int 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
1234 error:
1235         kobject_put(&_manager->kobj);
1236         _manager = NULL;
1237         return ret;
1238 }
1239
1240 void 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
1258 int 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 }
1283 EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1284
1285 #endif