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