xref: /linux/drivers/gpu/drm/ttm/ttm_pool.c (revision 021bc4b9)
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 #include <linux/debugfs.h>
37 #include <linux/highmem.h>
38 #include <linux/sched/mm.h>
39 
40 #ifdef CONFIG_X86
41 #include <asm/set_memory.h>
42 #endif
43 
44 #include <drm/ttm/ttm_pool.h>
45 #include <drm/ttm/ttm_tt.h>
46 #include <drm/ttm/ttm_bo.h>
47 
48 #include "ttm_module.h"
49 
50 /**
51  * struct ttm_pool_dma - Helper object for coherent DMA mappings
52  *
53  * @addr: original DMA address returned for the mapping
54  * @vaddr: original vaddr return for the mapping and order in the lower bits
55  */
56 struct ttm_pool_dma {
57 	dma_addr_t addr;
58 	unsigned long vaddr;
59 };
60 
61 static unsigned long page_pool_size;
62 
63 MODULE_PARM_DESC(page_pool_size, "Number of pages in the WC/UC/DMA pool");
64 module_param(page_pool_size, ulong, 0644);
65 
66 static atomic_long_t allocated_pages;
67 
68 static struct ttm_pool_type global_write_combined[NR_PAGE_ORDERS];
69 static struct ttm_pool_type global_uncached[NR_PAGE_ORDERS];
70 
71 static struct ttm_pool_type global_dma32_write_combined[NR_PAGE_ORDERS];
72 static struct ttm_pool_type global_dma32_uncached[NR_PAGE_ORDERS];
73 
74 static spinlock_t shrinker_lock;
75 static struct list_head shrinker_list;
76 static struct shrinker *mm_shrinker;
77 static DECLARE_RWSEM(pool_shrink_rwsem);
78 
79 /* Allocate pages of size 1 << order with the given gfp_flags */
80 static struct page *ttm_pool_alloc_page(struct ttm_pool *pool, gfp_t gfp_flags,
81 					unsigned int order)
82 {
83 	unsigned long attr = DMA_ATTR_FORCE_CONTIGUOUS;
84 	struct ttm_pool_dma *dma;
85 	struct page *p;
86 	void *vaddr;
87 
88 	/* Don't set the __GFP_COMP flag for higher order allocations.
89 	 * Mapping pages directly into an userspace process and calling
90 	 * put_page() on a TTM allocated page is illegal.
91 	 */
92 	if (order)
93 		gfp_flags |= __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN |
94 			__GFP_KSWAPD_RECLAIM;
95 
96 	if (!pool->use_dma_alloc) {
97 		p = alloc_pages_node(pool->nid, gfp_flags, order);
98 		if (p)
99 			p->private = order;
100 		return p;
101 	}
102 
103 	dma = kmalloc(sizeof(*dma), GFP_KERNEL);
104 	if (!dma)
105 		return NULL;
106 
107 	if (order)
108 		attr |= DMA_ATTR_NO_WARN;
109 
110 	vaddr = dma_alloc_attrs(pool->dev, (1ULL << order) * PAGE_SIZE,
111 				&dma->addr, gfp_flags, attr);
112 	if (!vaddr)
113 		goto error_free;
114 
115 	/* TODO: This is an illegal abuse of the DMA API, but we need to rework
116 	 * TTM page fault handling and extend the DMA API to clean this up.
117 	 */
118 	if (is_vmalloc_addr(vaddr))
119 		p = vmalloc_to_page(vaddr);
120 	else
121 		p = virt_to_page(vaddr);
122 
123 	dma->vaddr = (unsigned long)vaddr | order;
124 	p->private = (unsigned long)dma;
125 	return p;
126 
127 error_free:
128 	kfree(dma);
129 	return NULL;
130 }
131 
132 /* Reset the caching and pages of size 1 << order */
133 static void ttm_pool_free_page(struct ttm_pool *pool, enum ttm_caching caching,
134 			       unsigned int order, struct page *p)
135 {
136 	unsigned long attr = DMA_ATTR_FORCE_CONTIGUOUS;
137 	struct ttm_pool_dma *dma;
138 	void *vaddr;
139 
140 #ifdef CONFIG_X86
141 	/* We don't care that set_pages_wb is inefficient here. This is only
142 	 * used when we have to shrink and CPU overhead is irrelevant then.
143 	 */
144 	if (caching != ttm_cached && !PageHighMem(p))
145 		set_pages_wb(p, 1 << order);
146 #endif
147 
148 	if (!pool || !pool->use_dma_alloc) {
149 		__free_pages(p, order);
150 		return;
151 	}
152 
153 	if (order)
154 		attr |= DMA_ATTR_NO_WARN;
155 
156 	dma = (void *)p->private;
157 	vaddr = (void *)(dma->vaddr & PAGE_MASK);
158 	dma_free_attrs(pool->dev, (1UL << order) * PAGE_SIZE, vaddr, dma->addr,
159 		       attr);
160 	kfree(dma);
161 }
162 
163 /* Apply a new caching to an array of pages */
164 static int ttm_pool_apply_caching(struct page **first, struct page **last,
165 				  enum ttm_caching caching)
166 {
167 #ifdef CONFIG_X86
168 	unsigned int num_pages = last - first;
169 
170 	if (!num_pages)
171 		return 0;
172 
173 	switch (caching) {
174 	case ttm_cached:
175 		break;
176 	case ttm_write_combined:
177 		return set_pages_array_wc(first, num_pages);
178 	case ttm_uncached:
179 		return set_pages_array_uc(first, num_pages);
180 	}
181 #endif
182 	return 0;
183 }
184 
185 /* Map pages of 1 << order size and fill the DMA address array  */
186 static int ttm_pool_map(struct ttm_pool *pool, unsigned int order,
187 			struct page *p, dma_addr_t **dma_addr)
188 {
189 	dma_addr_t addr;
190 	unsigned int i;
191 
192 	if (pool->use_dma_alloc) {
193 		struct ttm_pool_dma *dma = (void *)p->private;
194 
195 		addr = dma->addr;
196 	} else {
197 		size_t size = (1ULL << order) * PAGE_SIZE;
198 
199 		addr = dma_map_page(pool->dev, p, 0, size, DMA_BIDIRECTIONAL);
200 		if (dma_mapping_error(pool->dev, addr))
201 			return -EFAULT;
202 	}
203 
204 	for (i = 1 << order; i ; --i) {
205 		*(*dma_addr)++ = addr;
206 		addr += PAGE_SIZE;
207 	}
208 
209 	return 0;
210 }
211 
212 /* Unmap pages of 1 << order size */
213 static void ttm_pool_unmap(struct ttm_pool *pool, dma_addr_t dma_addr,
214 			   unsigned int num_pages)
215 {
216 	/* Unmapped while freeing the page */
217 	if (pool->use_dma_alloc)
218 		return;
219 
220 	dma_unmap_page(pool->dev, dma_addr, (long)num_pages << PAGE_SHIFT,
221 		       DMA_BIDIRECTIONAL);
222 }
223 
224 /* Give pages into a specific pool_type */
225 static void ttm_pool_type_give(struct ttm_pool_type *pt, struct page *p)
226 {
227 	unsigned int i, num_pages = 1 << pt->order;
228 
229 	for (i = 0; i < num_pages; ++i) {
230 		if (PageHighMem(p))
231 			clear_highpage(p + i);
232 		else
233 			clear_page(page_address(p + i));
234 	}
235 
236 	spin_lock(&pt->lock);
237 	list_add(&p->lru, &pt->pages);
238 	spin_unlock(&pt->lock);
239 	atomic_long_add(1 << pt->order, &allocated_pages);
240 }
241 
242 /* Take pages from a specific pool_type, return NULL when nothing available */
243 static struct page *ttm_pool_type_take(struct ttm_pool_type *pt)
244 {
245 	struct page *p;
246 
247 	spin_lock(&pt->lock);
248 	p = list_first_entry_or_null(&pt->pages, typeof(*p), lru);
249 	if (p) {
250 		atomic_long_sub(1 << pt->order, &allocated_pages);
251 		list_del(&p->lru);
252 	}
253 	spin_unlock(&pt->lock);
254 
255 	return p;
256 }
257 
258 /* Initialize and add a pool type to the global shrinker list */
259 static void ttm_pool_type_init(struct ttm_pool_type *pt, struct ttm_pool *pool,
260 			       enum ttm_caching caching, unsigned int order)
261 {
262 	pt->pool = pool;
263 	pt->caching = caching;
264 	pt->order = order;
265 	spin_lock_init(&pt->lock);
266 	INIT_LIST_HEAD(&pt->pages);
267 
268 	spin_lock(&shrinker_lock);
269 	list_add_tail(&pt->shrinker_list, &shrinker_list);
270 	spin_unlock(&shrinker_lock);
271 }
272 
273 /* Remove a pool_type from the global shrinker list and free all pages */
274 static void ttm_pool_type_fini(struct ttm_pool_type *pt)
275 {
276 	struct page *p;
277 
278 	spin_lock(&shrinker_lock);
279 	list_del(&pt->shrinker_list);
280 	spin_unlock(&shrinker_lock);
281 
282 	while ((p = ttm_pool_type_take(pt)))
283 		ttm_pool_free_page(pt->pool, pt->caching, pt->order, p);
284 }
285 
286 /* Return the pool_type to use for the given caching and order */
287 static struct ttm_pool_type *ttm_pool_select_type(struct ttm_pool *pool,
288 						  enum ttm_caching caching,
289 						  unsigned int order)
290 {
291 	if (pool->use_dma_alloc || pool->nid != NUMA_NO_NODE)
292 		return &pool->caching[caching].orders[order];
293 
294 #ifdef CONFIG_X86
295 	switch (caching) {
296 	case ttm_write_combined:
297 		if (pool->use_dma32)
298 			return &global_dma32_write_combined[order];
299 
300 		return &global_write_combined[order];
301 	case ttm_uncached:
302 		if (pool->use_dma32)
303 			return &global_dma32_uncached[order];
304 
305 		return &global_uncached[order];
306 	default:
307 		break;
308 	}
309 #endif
310 
311 	return NULL;
312 }
313 
314 /* Free pages using the global shrinker list */
315 static unsigned int ttm_pool_shrink(void)
316 {
317 	struct ttm_pool_type *pt;
318 	unsigned int num_pages;
319 	struct page *p;
320 
321 	down_read(&pool_shrink_rwsem);
322 	spin_lock(&shrinker_lock);
323 	pt = list_first_entry(&shrinker_list, typeof(*pt), shrinker_list);
324 	list_move_tail(&pt->shrinker_list, &shrinker_list);
325 	spin_unlock(&shrinker_lock);
326 
327 	p = ttm_pool_type_take(pt);
328 	if (p) {
329 		ttm_pool_free_page(pt->pool, pt->caching, pt->order, p);
330 		num_pages = 1 << pt->order;
331 	} else {
332 		num_pages = 0;
333 	}
334 	up_read(&pool_shrink_rwsem);
335 
336 	return num_pages;
337 }
338 
339 /* Return the allocation order based for a page */
340 static unsigned int ttm_pool_page_order(struct ttm_pool *pool, struct page *p)
341 {
342 	if (pool->use_dma_alloc) {
343 		struct ttm_pool_dma *dma = (void *)p->private;
344 
345 		return dma->vaddr & ~PAGE_MASK;
346 	}
347 
348 	return p->private;
349 }
350 
351 /* Called when we got a page, either from a pool or newly allocated */
352 static int ttm_pool_page_allocated(struct ttm_pool *pool, unsigned int order,
353 				   struct page *p, dma_addr_t **dma_addr,
354 				   unsigned long *num_pages,
355 				   struct page ***pages)
356 {
357 	unsigned int i;
358 	int r;
359 
360 	if (*dma_addr) {
361 		r = ttm_pool_map(pool, order, p, dma_addr);
362 		if (r)
363 			return r;
364 	}
365 
366 	*num_pages -= 1 << order;
367 	for (i = 1 << order; i; --i, ++(*pages), ++p)
368 		**pages = p;
369 
370 	return 0;
371 }
372 
373 /**
374  * ttm_pool_free_range() - Free a range of TTM pages
375  * @pool: The pool used for allocating.
376  * @tt: The struct ttm_tt holding the page pointers.
377  * @caching: The page caching mode used by the range.
378  * @start_page: index for first page to free.
379  * @end_page: index for last page to free + 1.
380  *
381  * During allocation the ttm_tt page-vector may be populated with ranges of
382  * pages with different attributes if allocation hit an error without being
383  * able to completely fulfill the allocation. This function can be used
384  * to free these individual ranges.
385  */
386 static void ttm_pool_free_range(struct ttm_pool *pool, struct ttm_tt *tt,
387 				enum ttm_caching caching,
388 				pgoff_t start_page, pgoff_t end_page)
389 {
390 	struct page **pages = tt->pages;
391 	unsigned int order;
392 	pgoff_t i, nr;
393 
394 	for (i = start_page; i < end_page; i += nr, pages += nr) {
395 		struct ttm_pool_type *pt = NULL;
396 
397 		order = ttm_pool_page_order(pool, *pages);
398 		nr = (1UL << order);
399 		if (tt->dma_address)
400 			ttm_pool_unmap(pool, tt->dma_address[i], nr);
401 
402 		pt = ttm_pool_select_type(pool, caching, order);
403 		if (pt)
404 			ttm_pool_type_give(pt, *pages);
405 		else
406 			ttm_pool_free_page(pool, caching, order, *pages);
407 	}
408 }
409 
410 /**
411  * ttm_pool_alloc - Fill a ttm_tt object
412  *
413  * @pool: ttm_pool to use
414  * @tt: ttm_tt object to fill
415  * @ctx: operation context
416  *
417  * Fill the ttm_tt object with pages and also make sure to DMA map them when
418  * necessary.
419  *
420  * Returns: 0 on successe, negative error code otherwise.
421  */
422 int ttm_pool_alloc(struct ttm_pool *pool, struct ttm_tt *tt,
423 		   struct ttm_operation_ctx *ctx)
424 {
425 	pgoff_t num_pages = tt->num_pages;
426 	dma_addr_t *dma_addr = tt->dma_address;
427 	struct page **caching = tt->pages;
428 	struct page **pages = tt->pages;
429 	enum ttm_caching page_caching;
430 	gfp_t gfp_flags = GFP_USER;
431 	pgoff_t caching_divide;
432 	unsigned int order;
433 	struct page *p;
434 	int r;
435 
436 	WARN_ON(!num_pages || ttm_tt_is_populated(tt));
437 	WARN_ON(dma_addr && !pool->dev);
438 
439 	if (tt->page_flags & TTM_TT_FLAG_ZERO_ALLOC)
440 		gfp_flags |= __GFP_ZERO;
441 
442 	if (ctx->gfp_retry_mayfail)
443 		gfp_flags |= __GFP_RETRY_MAYFAIL;
444 
445 	if (pool->use_dma32)
446 		gfp_flags |= GFP_DMA32;
447 	else
448 		gfp_flags |= GFP_HIGHUSER;
449 
450 	for (order = min_t(unsigned int, MAX_PAGE_ORDER, __fls(num_pages));
451 	     num_pages;
452 	     order = min_t(unsigned int, order, __fls(num_pages))) {
453 		struct ttm_pool_type *pt;
454 
455 		page_caching = tt->caching;
456 		pt = ttm_pool_select_type(pool, tt->caching, order);
457 		p = pt ? ttm_pool_type_take(pt) : NULL;
458 		if (p) {
459 			r = ttm_pool_apply_caching(caching, pages,
460 						   tt->caching);
461 			if (r)
462 				goto error_free_page;
463 
464 			caching = pages;
465 			do {
466 				r = ttm_pool_page_allocated(pool, order, p,
467 							    &dma_addr,
468 							    &num_pages,
469 							    &pages);
470 				if (r)
471 					goto error_free_page;
472 
473 				caching = pages;
474 				if (num_pages < (1 << order))
475 					break;
476 
477 				p = ttm_pool_type_take(pt);
478 			} while (p);
479 		}
480 
481 		page_caching = ttm_cached;
482 		while (num_pages >= (1 << order) &&
483 		       (p = ttm_pool_alloc_page(pool, gfp_flags, order))) {
484 
485 			if (PageHighMem(p)) {
486 				r = ttm_pool_apply_caching(caching, pages,
487 							   tt->caching);
488 				if (r)
489 					goto error_free_page;
490 				caching = pages;
491 			}
492 			r = ttm_pool_page_allocated(pool, order, p, &dma_addr,
493 						    &num_pages, &pages);
494 			if (r)
495 				goto error_free_page;
496 			if (PageHighMem(p))
497 				caching = pages;
498 		}
499 
500 		if (!p) {
501 			if (order) {
502 				--order;
503 				continue;
504 			}
505 			r = -ENOMEM;
506 			goto error_free_all;
507 		}
508 	}
509 
510 	r = ttm_pool_apply_caching(caching, pages, tt->caching);
511 	if (r)
512 		goto error_free_all;
513 
514 	return 0;
515 
516 error_free_page:
517 	ttm_pool_free_page(pool, page_caching, order, p);
518 
519 error_free_all:
520 	num_pages = tt->num_pages - num_pages;
521 	caching_divide = caching - tt->pages;
522 	ttm_pool_free_range(pool, tt, tt->caching, 0, caching_divide);
523 	ttm_pool_free_range(pool, tt, ttm_cached, caching_divide, num_pages);
524 
525 	return r;
526 }
527 EXPORT_SYMBOL(ttm_pool_alloc);
528 
529 /**
530  * ttm_pool_free - Free the backing pages from a ttm_tt object
531  *
532  * @pool: Pool to give pages back to.
533  * @tt: ttm_tt object to unpopulate
534  *
535  * Give the packing pages back to a pool or free them
536  */
537 void ttm_pool_free(struct ttm_pool *pool, struct ttm_tt *tt)
538 {
539 	ttm_pool_free_range(pool, tt, tt->caching, 0, tt->num_pages);
540 
541 	while (atomic_long_read(&allocated_pages) > page_pool_size)
542 		ttm_pool_shrink();
543 }
544 EXPORT_SYMBOL(ttm_pool_free);
545 
546 /**
547  * ttm_pool_init - Initialize a pool
548  *
549  * @pool: the pool to initialize
550  * @dev: device for DMA allocations and mappings
551  * @nid: NUMA node to use for allocations
552  * @use_dma_alloc: true if coherent DMA alloc should be used
553  * @use_dma32: true if GFP_DMA32 should be used
554  *
555  * Initialize the pool and its pool types.
556  */
557 void ttm_pool_init(struct ttm_pool *pool, struct device *dev,
558 		   int nid, bool use_dma_alloc, bool use_dma32)
559 {
560 	unsigned int i, j;
561 
562 	WARN_ON(!dev && use_dma_alloc);
563 
564 	pool->dev = dev;
565 	pool->nid = nid;
566 	pool->use_dma_alloc = use_dma_alloc;
567 	pool->use_dma32 = use_dma32;
568 
569 	if (use_dma_alloc || nid != NUMA_NO_NODE) {
570 		for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i)
571 			for (j = 0; j < NR_PAGE_ORDERS; ++j)
572 				ttm_pool_type_init(&pool->caching[i].orders[j],
573 						   pool, i, j);
574 	}
575 }
576 EXPORT_SYMBOL(ttm_pool_init);
577 
578 /**
579  * ttm_pool_synchronize_shrinkers - Wait for all running shrinkers to complete.
580  *
581  * This is useful to guarantee that all shrinker invocations have seen an
582  * update, before freeing memory, similar to rcu.
583  */
584 static void ttm_pool_synchronize_shrinkers(void)
585 {
586 	down_write(&pool_shrink_rwsem);
587 	up_write(&pool_shrink_rwsem);
588 }
589 
590 /**
591  * ttm_pool_fini - Cleanup a pool
592  *
593  * @pool: the pool to clean up
594  *
595  * Free all pages in the pool and unregister the types from the global
596  * shrinker.
597  */
598 void ttm_pool_fini(struct ttm_pool *pool)
599 {
600 	unsigned int i, j;
601 
602 	if (pool->use_dma_alloc || pool->nid != NUMA_NO_NODE) {
603 		for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i)
604 			for (j = 0; j < NR_PAGE_ORDERS; ++j)
605 				ttm_pool_type_fini(&pool->caching[i].orders[j]);
606 	}
607 
608 	/* We removed the pool types from the LRU, but we need to also make sure
609 	 * that no shrinker is concurrently freeing pages from the pool.
610 	 */
611 	ttm_pool_synchronize_shrinkers();
612 }
613 EXPORT_SYMBOL(ttm_pool_fini);
614 
615 /* As long as pages are available make sure to release at least one */
616 static unsigned long ttm_pool_shrinker_scan(struct shrinker *shrink,
617 					    struct shrink_control *sc)
618 {
619 	unsigned long num_freed = 0;
620 
621 	do
622 		num_freed += ttm_pool_shrink();
623 	while (!num_freed && atomic_long_read(&allocated_pages));
624 
625 	return num_freed;
626 }
627 
628 /* Return the number of pages available or SHRINK_EMPTY if we have none */
629 static unsigned long ttm_pool_shrinker_count(struct shrinker *shrink,
630 					     struct shrink_control *sc)
631 {
632 	unsigned long num_pages = atomic_long_read(&allocated_pages);
633 
634 	return num_pages ? num_pages : SHRINK_EMPTY;
635 }
636 
637 #ifdef CONFIG_DEBUG_FS
638 /* Count the number of pages available in a pool_type */
639 static unsigned int ttm_pool_type_count(struct ttm_pool_type *pt)
640 {
641 	unsigned int count = 0;
642 	struct page *p;
643 
644 	spin_lock(&pt->lock);
645 	/* Only used for debugfs, the overhead doesn't matter */
646 	list_for_each_entry(p, &pt->pages, lru)
647 		++count;
648 	spin_unlock(&pt->lock);
649 
650 	return count;
651 }
652 
653 /* Print a nice header for the order */
654 static void ttm_pool_debugfs_header(struct seq_file *m)
655 {
656 	unsigned int i;
657 
658 	seq_puts(m, "\t ");
659 	for (i = 0; i < NR_PAGE_ORDERS; ++i)
660 		seq_printf(m, " ---%2u---", i);
661 	seq_puts(m, "\n");
662 }
663 
664 /* Dump information about the different pool types */
665 static void ttm_pool_debugfs_orders(struct ttm_pool_type *pt,
666 				    struct seq_file *m)
667 {
668 	unsigned int i;
669 
670 	for (i = 0; i < NR_PAGE_ORDERS; ++i)
671 		seq_printf(m, " %8u", ttm_pool_type_count(&pt[i]));
672 	seq_puts(m, "\n");
673 }
674 
675 /* Dump the total amount of allocated pages */
676 static void ttm_pool_debugfs_footer(struct seq_file *m)
677 {
678 	seq_printf(m, "\ntotal\t: %8lu of %8lu\n",
679 		   atomic_long_read(&allocated_pages), page_pool_size);
680 }
681 
682 /* Dump the information for the global pools */
683 static int ttm_pool_debugfs_globals_show(struct seq_file *m, void *data)
684 {
685 	ttm_pool_debugfs_header(m);
686 
687 	spin_lock(&shrinker_lock);
688 	seq_puts(m, "wc\t:");
689 	ttm_pool_debugfs_orders(global_write_combined, m);
690 	seq_puts(m, "uc\t:");
691 	ttm_pool_debugfs_orders(global_uncached, m);
692 	seq_puts(m, "wc 32\t:");
693 	ttm_pool_debugfs_orders(global_dma32_write_combined, m);
694 	seq_puts(m, "uc 32\t:");
695 	ttm_pool_debugfs_orders(global_dma32_uncached, m);
696 	spin_unlock(&shrinker_lock);
697 
698 	ttm_pool_debugfs_footer(m);
699 
700 	return 0;
701 }
702 DEFINE_SHOW_ATTRIBUTE(ttm_pool_debugfs_globals);
703 
704 /**
705  * ttm_pool_debugfs - Debugfs dump function for a pool
706  *
707  * @pool: the pool to dump the information for
708  * @m: seq_file to dump to
709  *
710  * Make a debugfs dump with the per pool and global information.
711  */
712 int ttm_pool_debugfs(struct ttm_pool *pool, struct seq_file *m)
713 {
714 	unsigned int i;
715 
716 	if (!pool->use_dma_alloc) {
717 		seq_puts(m, "unused\n");
718 		return 0;
719 	}
720 
721 	ttm_pool_debugfs_header(m);
722 
723 	spin_lock(&shrinker_lock);
724 	for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i) {
725 		seq_puts(m, "DMA ");
726 		switch (i) {
727 		case ttm_cached:
728 			seq_puts(m, "\t:");
729 			break;
730 		case ttm_write_combined:
731 			seq_puts(m, "wc\t:");
732 			break;
733 		case ttm_uncached:
734 			seq_puts(m, "uc\t:");
735 			break;
736 		}
737 		ttm_pool_debugfs_orders(pool->caching[i].orders, m);
738 	}
739 	spin_unlock(&shrinker_lock);
740 
741 	ttm_pool_debugfs_footer(m);
742 	return 0;
743 }
744 EXPORT_SYMBOL(ttm_pool_debugfs);
745 
746 /* Test the shrinker functions and dump the result */
747 static int ttm_pool_debugfs_shrink_show(struct seq_file *m, void *data)
748 {
749 	struct shrink_control sc = { .gfp_mask = GFP_NOFS };
750 
751 	fs_reclaim_acquire(GFP_KERNEL);
752 	seq_printf(m, "%lu/%lu\n", ttm_pool_shrinker_count(mm_shrinker, &sc),
753 		   ttm_pool_shrinker_scan(mm_shrinker, &sc));
754 	fs_reclaim_release(GFP_KERNEL);
755 
756 	return 0;
757 }
758 DEFINE_SHOW_ATTRIBUTE(ttm_pool_debugfs_shrink);
759 
760 #endif
761 
762 /**
763  * ttm_pool_mgr_init - Initialize globals
764  *
765  * @num_pages: default number of pages
766  *
767  * Initialize the global locks and lists for the MM shrinker.
768  */
769 int ttm_pool_mgr_init(unsigned long num_pages)
770 {
771 	unsigned int i;
772 
773 	if (!page_pool_size)
774 		page_pool_size = num_pages;
775 
776 	spin_lock_init(&shrinker_lock);
777 	INIT_LIST_HEAD(&shrinker_list);
778 
779 	for (i = 0; i < NR_PAGE_ORDERS; ++i) {
780 		ttm_pool_type_init(&global_write_combined[i], NULL,
781 				   ttm_write_combined, i);
782 		ttm_pool_type_init(&global_uncached[i], NULL, ttm_uncached, i);
783 
784 		ttm_pool_type_init(&global_dma32_write_combined[i], NULL,
785 				   ttm_write_combined, i);
786 		ttm_pool_type_init(&global_dma32_uncached[i], NULL,
787 				   ttm_uncached, i);
788 	}
789 
790 #ifdef CONFIG_DEBUG_FS
791 	debugfs_create_file("page_pool", 0444, ttm_debugfs_root, NULL,
792 			    &ttm_pool_debugfs_globals_fops);
793 	debugfs_create_file("page_pool_shrink", 0400, ttm_debugfs_root, NULL,
794 			    &ttm_pool_debugfs_shrink_fops);
795 #endif
796 
797 	mm_shrinker = shrinker_alloc(0, "drm-ttm_pool");
798 	if (!mm_shrinker)
799 		return -ENOMEM;
800 
801 	mm_shrinker->count_objects = ttm_pool_shrinker_count;
802 	mm_shrinker->scan_objects = ttm_pool_shrinker_scan;
803 	mm_shrinker->seeks = 1;
804 
805 	shrinker_register(mm_shrinker);
806 
807 	return 0;
808 }
809 
810 /**
811  * ttm_pool_mgr_fini - Finalize globals
812  *
813  * Cleanup the global pools and unregister the MM shrinker.
814  */
815 void ttm_pool_mgr_fini(void)
816 {
817 	unsigned int i;
818 
819 	for (i = 0; i < NR_PAGE_ORDERS; ++i) {
820 		ttm_pool_type_fini(&global_write_combined[i]);
821 		ttm_pool_type_fini(&global_uncached[i]);
822 
823 		ttm_pool_type_fini(&global_dma32_write_combined[i]);
824 		ttm_pool_type_fini(&global_dma32_uncached[i]);
825 	}
826 
827 	shrinker_free(mm_shrinker);
828 	WARN_ON(!list_empty(&shrinker_list));
829 }
830