Merge tag 'drm-next-2020-12-24' of git://anongit.freedesktop.org/drm/drm
[linux-2.6-microblaze.git] / drivers / gpu / drm / ttm / ttm_pool.c
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2020 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions 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 NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  *
23  * Authors: Christian König
24  */
25
26 /* Pooling of allocated pages is necessary because changing the caching
27  * attributes on x86 of the linear mapping requires a costly cross CPU TLB
28  * invalidate for those addresses.
29  *
30  * Additional to that allocations from the DMA coherent API are pooled as well
31  * cause they are rather slow compared to alloc_pages+map.
32  */
33
34 #include <linux/module.h>
35 #include <linux/dma-mapping.h>
36
37 #ifdef CONFIG_X86
38 #include <asm/set_memory.h>
39 #endif
40
41 #include <drm/ttm/ttm_pool.h>
42 #include <drm/ttm/ttm_bo_driver.h>
43 #include <drm/ttm/ttm_tt.h>
44
45 /**
46  * struct ttm_pool_dma - Helper object for coherent DMA mappings
47  *
48  * @addr: original DMA address returned for the mapping
49  * @vaddr: original vaddr return for the mapping and order in the lower bits
50  */
51 struct ttm_pool_dma {
52         dma_addr_t addr;
53         unsigned long vaddr;
54 };
55
56 static unsigned long page_pool_size;
57
58 MODULE_PARM_DESC(page_pool_size, "Number of pages in the WC/UC/DMA pool");
59 module_param(page_pool_size, ulong, 0644);
60
61 static atomic_long_t allocated_pages;
62
63 static struct ttm_pool_type global_write_combined[MAX_ORDER];
64 static struct ttm_pool_type global_uncached[MAX_ORDER];
65
66 static struct ttm_pool_type global_dma32_write_combined[MAX_ORDER];
67 static struct ttm_pool_type global_dma32_uncached[MAX_ORDER];
68
69 static spinlock_t shrinker_lock;
70 static struct list_head shrinker_list;
71 static struct shrinker mm_shrinker;
72
73 /* Allocate pages of size 1 << order with the given gfp_flags */
74 static struct page *ttm_pool_alloc_page(struct ttm_pool *pool, gfp_t gfp_flags,
75                                         unsigned int order)
76 {
77         unsigned long attr = DMA_ATTR_FORCE_CONTIGUOUS;
78         struct ttm_pool_dma *dma;
79         struct page *p;
80         void *vaddr;
81
82         if (order) {
83                 gfp_flags |= GFP_TRANSHUGE_LIGHT | __GFP_NORETRY |
84                         __GFP_KSWAPD_RECLAIM;
85                 gfp_flags &= ~__GFP_MOVABLE;
86                 gfp_flags &= ~__GFP_COMP;
87         }
88
89         if (!pool->use_dma_alloc) {
90                 p = alloc_pages(gfp_flags, order);
91                 if (p)
92                         p->private = order;
93                 return p;
94         }
95
96         dma = kmalloc(sizeof(*dma), GFP_KERNEL);
97         if (!dma)
98                 return NULL;
99
100         if (order)
101                 attr |= DMA_ATTR_NO_WARN;
102
103         vaddr = dma_alloc_attrs(pool->dev, (1ULL << order) * PAGE_SIZE,
104                                 &dma->addr, gfp_flags, attr);
105         if (!vaddr)
106                 goto error_free;
107
108         /* TODO: This is an illegal abuse of the DMA API, but we need to rework
109          * TTM page fault handling and extend the DMA API to clean this up.
110          */
111         if (is_vmalloc_addr(vaddr))
112                 p = vmalloc_to_page(vaddr);
113         else
114                 p = virt_to_page(vaddr);
115
116         dma->vaddr = (unsigned long)vaddr | order;
117         p->private = (unsigned long)dma;
118         return p;
119
120 error_free:
121         kfree(dma);
122         return NULL;
123 }
124
125 /* Reset the caching and pages of size 1 << order */
126 static void ttm_pool_free_page(struct ttm_pool *pool, enum ttm_caching caching,
127                                unsigned int order, struct page *p)
128 {
129         unsigned long attr = DMA_ATTR_FORCE_CONTIGUOUS;
130         struct ttm_pool_dma *dma;
131         void *vaddr;
132
133 #ifdef CONFIG_X86
134         /* We don't care that set_pages_wb is inefficient here. This is only
135          * used when we have to shrink and CPU overhead is irrelevant then.
136          */
137         if (caching != ttm_cached && !PageHighMem(p))
138                 set_pages_wb(p, 1 << order);
139 #endif
140
141         if (!pool || !pool->use_dma_alloc) {
142                 __free_pages(p, order);
143                 return;
144         }
145
146         if (order)
147                 attr |= DMA_ATTR_NO_WARN;
148
149         dma = (void *)p->private;
150         vaddr = (void *)(dma->vaddr & PAGE_MASK);
151         dma_free_attrs(pool->dev, (1UL << order) * PAGE_SIZE, vaddr, dma->addr,
152                        attr);
153         kfree(dma);
154 }
155
156 /* Apply a new caching to an array of pages */
157 static int ttm_pool_apply_caching(struct page **first, struct page **last,
158                                   enum ttm_caching caching)
159 {
160 #ifdef CONFIG_X86
161         unsigned int num_pages = last - first;
162
163         if (!num_pages)
164                 return 0;
165
166         switch (caching) {
167         case ttm_cached:
168                 break;
169         case ttm_write_combined:
170                 return set_pages_array_wc(first, num_pages);
171         case ttm_uncached:
172                 return set_pages_array_uc(first, num_pages);
173         }
174 #endif
175         return 0;
176 }
177
178 /* Map pages of 1 << order size and fill the DMA address array  */
179 static int ttm_pool_map(struct ttm_pool *pool, unsigned int order,
180                         struct page *p, dma_addr_t **dma_addr)
181 {
182         dma_addr_t addr;
183         unsigned int i;
184
185         if (pool->use_dma_alloc) {
186                 struct ttm_pool_dma *dma = (void *)p->private;
187
188                 addr = dma->addr;
189         } else {
190                 size_t size = (1ULL << order) * PAGE_SIZE;
191
192                 addr = dma_map_page(pool->dev, p, 0, size, DMA_BIDIRECTIONAL);
193                 if (dma_mapping_error(pool->dev, **dma_addr))
194                         return -EFAULT;
195         }
196
197         for (i = 1 << order; i ; --i) {
198                 *(*dma_addr)++ = addr;
199                 addr += PAGE_SIZE;
200         }
201
202         return 0;
203 }
204
205 /* Unmap pages of 1 << order size */
206 static void ttm_pool_unmap(struct ttm_pool *pool, dma_addr_t dma_addr,
207                            unsigned int num_pages)
208 {
209         /* Unmapped while freeing the page */
210         if (pool->use_dma_alloc)
211                 return;
212
213         dma_unmap_page(pool->dev, dma_addr, (long)num_pages << PAGE_SHIFT,
214                        DMA_BIDIRECTIONAL);
215 }
216
217 /* Give pages into a specific pool_type */
218 static void ttm_pool_type_give(struct ttm_pool_type *pt, struct page *p)
219 {
220         spin_lock(&pt->lock);
221         list_add(&p->lru, &pt->pages);
222         spin_unlock(&pt->lock);
223         atomic_long_add(1 << pt->order, &allocated_pages);
224 }
225
226 /* Take pages from a specific pool_type, return NULL when nothing available */
227 static struct page *ttm_pool_type_take(struct ttm_pool_type *pt)
228 {
229         struct page *p;
230
231         spin_lock(&pt->lock);
232         p = list_first_entry_or_null(&pt->pages, typeof(*p), lru);
233         if (p) {
234                 atomic_long_sub(1 << pt->order, &allocated_pages);
235                 list_del(&p->lru);
236         }
237         spin_unlock(&pt->lock);
238
239         return p;
240 }
241
242 /* Initialize and add a pool type to the global shrinker list */
243 static void ttm_pool_type_init(struct ttm_pool_type *pt, struct ttm_pool *pool,
244                                enum ttm_caching caching, unsigned int order)
245 {
246         pt->pool = pool;
247         pt->caching = caching;
248         pt->order = order;
249         spin_lock_init(&pt->lock);
250         INIT_LIST_HEAD(&pt->pages);
251
252         spin_lock(&shrinker_lock);
253         list_add_tail(&pt->shrinker_list, &shrinker_list);
254         spin_unlock(&shrinker_lock);
255 }
256
257 /* Remove a pool_type from the global shrinker list and free all pages */
258 static void ttm_pool_type_fini(struct ttm_pool_type *pt)
259 {
260         struct page *p, *tmp;
261
262         spin_lock(&shrinker_lock);
263         list_del(&pt->shrinker_list);
264         spin_unlock(&shrinker_lock);
265
266         list_for_each_entry_safe(p, tmp, &pt->pages, lru)
267                 ttm_pool_free_page(pt->pool, pt->caching, pt->order, p);
268 }
269
270 /* Return the pool_type to use for the given caching and order */
271 static struct ttm_pool_type *ttm_pool_select_type(struct ttm_pool *pool,
272                                                   enum ttm_caching caching,
273                                                   unsigned int order)
274 {
275         if (pool->use_dma_alloc)
276                 return &pool->caching[caching].orders[order];
277
278 #ifdef CONFIG_X86
279         switch (caching) {
280         case ttm_write_combined:
281                 if (pool->use_dma32)
282                         return &global_dma32_write_combined[order];
283
284                 return &global_write_combined[order];
285         case ttm_uncached:
286                 if (pool->use_dma32)
287                         return &global_dma32_uncached[order];
288
289                 return &global_uncached[order];
290         default:
291                 break;
292         }
293 #endif
294
295         return NULL;
296 }
297
298 /* Free pages using the global shrinker list */
299 static unsigned int ttm_pool_shrink(void)
300 {
301         struct ttm_pool_type *pt;
302         unsigned int num_freed;
303         struct page *p;
304
305         spin_lock(&shrinker_lock);
306         pt = list_first_entry(&shrinker_list, typeof(*pt), shrinker_list);
307
308         p = ttm_pool_type_take(pt);
309         if (p) {
310                 ttm_pool_free_page(pt->pool, pt->caching, pt->order, p);
311                 num_freed = 1 << pt->order;
312         } else {
313                 num_freed = 0;
314         }
315
316         list_move_tail(&pt->shrinker_list, &shrinker_list);
317         spin_unlock(&shrinker_lock);
318
319         return num_freed;
320 }
321
322 /* Return the allocation order based for a page */
323 static unsigned int ttm_pool_page_order(struct ttm_pool *pool, struct page *p)
324 {
325         if (pool->use_dma_alloc) {
326                 struct ttm_pool_dma *dma = (void *)p->private;
327
328                 return dma->vaddr & ~PAGE_MASK;
329         }
330
331         return p->private;
332 }
333
334 /**
335  * ttm_pool_alloc - Fill a ttm_tt object
336  *
337  * @pool: ttm_pool to use
338  * @tt: ttm_tt object to fill
339  * @ctx: operation context
340  *
341  * Fill the ttm_tt object with pages and also make sure to DMA map them when
342  * necessary.
343  *
344  * Returns: 0 on successe, negative error code otherwise.
345  */
346 int ttm_pool_alloc(struct ttm_pool *pool, struct ttm_tt *tt,
347                    struct ttm_operation_ctx *ctx)
348 {
349         unsigned long num_pages = tt->num_pages;
350         dma_addr_t *dma_addr = tt->dma_address;
351         struct page **caching = tt->pages;
352         struct page **pages = tt->pages;
353         gfp_t gfp_flags = GFP_USER;
354         unsigned int i, order;
355         struct page *p;
356         int r;
357
358         WARN_ON(!num_pages || ttm_tt_is_populated(tt));
359         WARN_ON(dma_addr && !pool->dev);
360
361         if (tt->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
362                 gfp_flags |= __GFP_ZERO;
363
364         if (ctx->gfp_retry_mayfail)
365                 gfp_flags |= __GFP_RETRY_MAYFAIL;
366
367         if (pool->use_dma32)
368                 gfp_flags |= GFP_DMA32;
369         else
370                 gfp_flags |= GFP_HIGHUSER;
371
372         for (order = min(MAX_ORDER - 1UL, __fls(num_pages)); num_pages;
373              order = min_t(unsigned int, order, __fls(num_pages))) {
374                 bool apply_caching = false;
375                 struct ttm_pool_type *pt;
376
377                 pt = ttm_pool_select_type(pool, tt->caching, order);
378                 p = pt ? ttm_pool_type_take(pt) : NULL;
379                 if (p) {
380                         apply_caching = true;
381                 } else {
382                         p = ttm_pool_alloc_page(pool, gfp_flags, order);
383                         if (p && PageHighMem(p))
384                                 apply_caching = true;
385                 }
386
387                 if (!p) {
388                         if (order) {
389                                 --order;
390                                 continue;
391                         }
392                         r = -ENOMEM;
393                         goto error_free_all;
394                 }
395
396                 if (apply_caching) {
397                         r = ttm_pool_apply_caching(caching, pages,
398                                                    tt->caching);
399                         if (r)
400                                 goto error_free_page;
401                         caching = pages + (1 << order);
402                 }
403
404                 r = ttm_mem_global_alloc_page(&ttm_mem_glob, p,
405                                               (1 << order) * PAGE_SIZE,
406                                               ctx);
407                 if (r)
408                         goto error_free_page;
409
410                 if (dma_addr) {
411                         r = ttm_pool_map(pool, order, p, &dma_addr);
412                         if (r)
413                                 goto error_global_free;
414                 }
415
416                 num_pages -= 1 << order;
417                 for (i = 1 << order; i; --i)
418                         *(pages++) = p++;
419         }
420
421         r = ttm_pool_apply_caching(caching, pages, tt->caching);
422         if (r)
423                 goto error_free_all;
424
425         return 0;
426
427 error_global_free:
428         ttm_mem_global_free_page(&ttm_mem_glob, p, (1 << order) * PAGE_SIZE);
429
430 error_free_page:
431         ttm_pool_free_page(pool, tt->caching, order, p);
432
433 error_free_all:
434         num_pages = tt->num_pages - num_pages;
435         for (i = 0; i < num_pages; ) {
436                 order = ttm_pool_page_order(pool, tt->pages[i]);
437                 ttm_pool_free_page(pool, tt->caching, order, tt->pages[i]);
438                 i += 1 << order;
439         }
440
441         return r;
442 }
443 EXPORT_SYMBOL(ttm_pool_alloc);
444
445 /**
446  * ttm_pool_free - Free the backing pages from a ttm_tt object
447  *
448  * @pool: Pool to give pages back to.
449  * @tt: ttm_tt object to unpopulate
450  *
451  * Give the packing pages back to a pool or free them
452  */
453 void ttm_pool_free(struct ttm_pool *pool, struct ttm_tt *tt)
454 {
455         unsigned int i;
456
457         for (i = 0; i < tt->num_pages; ) {
458                 struct page *p = tt->pages[i];
459                 unsigned int order, num_pages;
460                 struct ttm_pool_type *pt;
461
462                 order = ttm_pool_page_order(pool, p);
463                 num_pages = 1ULL << order;
464                 ttm_mem_global_free_page(&ttm_mem_glob, p,
465                                          num_pages * PAGE_SIZE);
466                 if (tt->dma_address)
467                         ttm_pool_unmap(pool, tt->dma_address[i], num_pages);
468
469                 pt = ttm_pool_select_type(pool, tt->caching, order);
470                 if (pt)
471                         ttm_pool_type_give(pt, tt->pages[i]);
472                 else
473                         ttm_pool_free_page(pool, tt->caching, order,
474                                            tt->pages[i]);
475
476                 i += num_pages;
477         }
478
479         while (atomic_long_read(&allocated_pages) > page_pool_size)
480                 ttm_pool_shrink();
481 }
482 EXPORT_SYMBOL(ttm_pool_free);
483
484 /**
485  * ttm_pool_init - Initialize a pool
486  *
487  * @pool: the pool to initialize
488  * @dev: device for DMA allocations and mappings
489  * @use_dma_alloc: true if coherent DMA alloc should be used
490  * @use_dma32: true if GFP_DMA32 should be used
491  *
492  * Initialize the pool and its pool types.
493  */
494 void ttm_pool_init(struct ttm_pool *pool, struct device *dev,
495                    bool use_dma_alloc, bool use_dma32)
496 {
497         unsigned int i, j;
498
499         WARN_ON(!dev && use_dma_alloc);
500
501         pool->dev = dev;
502         pool->use_dma_alloc = use_dma_alloc;
503         pool->use_dma32 = use_dma32;
504
505         for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i)
506                 for (j = 0; j < MAX_ORDER; ++j)
507                         ttm_pool_type_init(&pool->caching[i].orders[j],
508                                            pool, i, j);
509 }
510 EXPORT_SYMBOL(ttm_pool_init);
511
512 /**
513  * ttm_pool_fini - Cleanup a pool
514  *
515  * @pool: the pool to clean up
516  *
517  * Free all pages in the pool and unregister the types from the global
518  * shrinker.
519  */
520 void ttm_pool_fini(struct ttm_pool *pool)
521 {
522         unsigned int i, j;
523
524         for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i)
525                 for (j = 0; j < MAX_ORDER; ++j)
526                         ttm_pool_type_fini(&pool->caching[i].orders[j]);
527 }
528 EXPORT_SYMBOL(ttm_pool_fini);
529
530 #ifdef CONFIG_DEBUG_FS
531 /* Count the number of pages available in a pool_type */
532 static unsigned int ttm_pool_type_count(struct ttm_pool_type *pt)
533 {
534         unsigned int count = 0;
535         struct page *p;
536
537         spin_lock(&pt->lock);
538         /* Only used for debugfs, the overhead doesn't matter */
539         list_for_each_entry(p, &pt->pages, lru)
540                 ++count;
541         spin_unlock(&pt->lock);
542
543         return count;
544 }
545
546 /* Dump information about the different pool types */
547 static void ttm_pool_debugfs_orders(struct ttm_pool_type *pt,
548                                     struct seq_file *m)
549 {
550         unsigned int i;
551
552         for (i = 0; i < MAX_ORDER; ++i)
553                 seq_printf(m, " %8u", ttm_pool_type_count(&pt[i]));
554         seq_puts(m, "\n");
555 }
556
557 /**
558  * ttm_pool_debugfs - Debugfs dump function for a pool
559  *
560  * @pool: the pool to dump the information for
561  * @m: seq_file to dump to
562  *
563  * Make a debugfs dump with the per pool and global information.
564  */
565 int ttm_pool_debugfs(struct ttm_pool *pool, struct seq_file *m)
566 {
567         unsigned int i;
568
569         spin_lock(&shrinker_lock);
570
571         seq_puts(m, "\t ");
572         for (i = 0; i < MAX_ORDER; ++i)
573                 seq_printf(m, " ---%2u---", i);
574         seq_puts(m, "\n");
575
576         seq_puts(m, "wc\t:");
577         ttm_pool_debugfs_orders(global_write_combined, m);
578         seq_puts(m, "uc\t:");
579         ttm_pool_debugfs_orders(global_uncached, m);
580
581         seq_puts(m, "wc 32\t:");
582         ttm_pool_debugfs_orders(global_dma32_write_combined, m);
583         seq_puts(m, "uc 32\t:");
584         ttm_pool_debugfs_orders(global_dma32_uncached, m);
585
586         for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i) {
587                 seq_puts(m, "DMA ");
588                 switch (i) {
589                 case ttm_cached:
590                         seq_puts(m, "\t:");
591                         break;
592                 case ttm_write_combined:
593                         seq_puts(m, "wc\t:");
594                         break;
595                 case ttm_uncached:
596                         seq_puts(m, "uc\t:");
597                         break;
598                 }
599                 ttm_pool_debugfs_orders(pool->caching[i].orders, m);
600         }
601
602         seq_printf(m, "\ntotal\t: %8lu of %8lu\n",
603                    atomic_long_read(&allocated_pages), page_pool_size);
604
605         spin_unlock(&shrinker_lock);
606
607         return 0;
608 }
609 EXPORT_SYMBOL(ttm_pool_debugfs);
610
611 #endif
612
613 /* As long as pages are available make sure to release at least one */
614 static unsigned long ttm_pool_shrinker_scan(struct shrinker *shrink,
615                                             struct shrink_control *sc)
616 {
617         unsigned long num_freed = 0;
618
619         do
620                 num_freed += ttm_pool_shrink();
621         while (!num_freed && atomic_long_read(&allocated_pages));
622
623         return num_freed;
624 }
625
626 /* Return the number of pages available or SHRINK_EMPTY if we have none */
627 static unsigned long ttm_pool_shrinker_count(struct shrinker *shrink,
628                                              struct shrink_control *sc)
629 {
630         unsigned long num_pages = atomic_long_read(&allocated_pages);
631
632         return num_pages ? num_pages : SHRINK_EMPTY;
633 }
634
635 /**
636  * ttm_pool_mgr_init - Initialize globals
637  *
638  * @num_pages: default number of pages
639  *
640  * Initialize the global locks and lists for the MM shrinker.
641  */
642 int ttm_pool_mgr_init(unsigned long num_pages)
643 {
644         unsigned int i;
645
646         if (!page_pool_size)
647                 page_pool_size = num_pages;
648
649         spin_lock_init(&shrinker_lock);
650         INIT_LIST_HEAD(&shrinker_list);
651
652         for (i = 0; i < MAX_ORDER; ++i) {
653                 ttm_pool_type_init(&global_write_combined[i], NULL,
654                                    ttm_write_combined, i);
655                 ttm_pool_type_init(&global_uncached[i], NULL, ttm_uncached, i);
656
657                 ttm_pool_type_init(&global_dma32_write_combined[i], NULL,
658                                    ttm_write_combined, i);
659                 ttm_pool_type_init(&global_dma32_uncached[i], NULL,
660                                    ttm_uncached, i);
661         }
662
663         mm_shrinker.count_objects = ttm_pool_shrinker_count;
664         mm_shrinker.scan_objects = ttm_pool_shrinker_scan;
665         mm_shrinker.seeks = 1;
666         return register_shrinker(&mm_shrinker);
667 }
668
669 /**
670  * ttm_pool_mgr_fini - Finalize globals
671  *
672  * Cleanup the global pools and unregister the MM shrinker.
673  */
674 void ttm_pool_mgr_fini(void)
675 {
676         unsigned int i;
677
678         for (i = 0; i < MAX_ORDER; ++i) {
679                 ttm_pool_type_fini(&global_write_combined[i]);
680                 ttm_pool_type_fini(&global_uncached[i]);
681
682                 ttm_pool_type_fini(&global_dma32_write_combined[i]);
683                 ttm_pool_type_fini(&global_dma32_uncached[i]);
684         }
685
686         unregister_shrinker(&mm_shrinker);
687         WARN_ON(!list_empty(&shrinker_list));
688 }