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#include <sys/cdefs.h>
37__FBSDID("$FreeBSD$");
38
39#define pr_fmt(fmt) "[TTM] " fmt
40
41#include <linux/dma-mapping.h>
42#include <linux/list.h>
43#include <linux/seq_file.h> /* for seq_printf */
44#include <linux/slab.h>
45#include <linux/spinlock.h>
46#include <linux/highmem.h>
47#include <linux/mm_types.h>
48#include <linux/module.h>
49#include <linux/mm.h>
50#include <linux/atomic.h>
51#include <linux/device.h>
52#include <linux/kthread.h>
53#include <drm/ttm/ttm_bo_driver.h>
54#include <drm/ttm/ttm_page_alloc.h>
55#ifdef TTM_HAS_AGP
56#include <asm/agp.h>
57#endif
58
59#define NUM_PAGES_TO_ALLOC		(PAGE_SIZE/sizeof(struct page *))
60#define SMALL_ALLOCATION		4
61#define FREE_ALL_PAGES			(~0U)
62/* times are in msecs */
63#define IS_UNDEFINED			(0)
64#define IS_WC				(1<<1)
65#define IS_UC				(1<<2)
66#define IS_CACHED			(1<<3)
67#define IS_DMA32			(1<<4)
68
69enum pool_type {
70	POOL_IS_UNDEFINED,
71	POOL_IS_WC = IS_WC,
72	POOL_IS_UC = IS_UC,
73	POOL_IS_CACHED = IS_CACHED,
74	POOL_IS_WC_DMA32 = IS_WC | IS_DMA32,
75	POOL_IS_UC_DMA32 = IS_UC | IS_DMA32,
76	POOL_IS_CACHED_DMA32 = IS_CACHED | IS_DMA32,
77};
78/*
79 * The pool structure. There are usually six pools:
80 *  - generic (not restricted to DMA32):
81 *      - write combined, uncached, cached.
82 *  - dma32 (up to 2^32 - so up 4GB):
83 *      - write combined, uncached, cached.
84 * for each 'struct device'. The 'cached' is for pages that are actively used.
85 * The other ones can be shrunk by the shrinker API if neccessary.
86 * @pools: The 'struct device->dma_pools' link.
87 * @type: Type of the pool
88 * @lock: Protects the inuse_list and free_list from concurrnet access. Must be
89 * used with irqsave/irqrestore variants because pool allocator maybe called
90 * from delayed work.
91 * @inuse_list: Pool of pages that are in use. The order is very important and
92 *   it is in the order that the TTM pages that are put back are in.
93 * @free_list: Pool of pages that are free to be used. No order requirements.
94 * @dev: The device that is associated with these pools.
95 * @size: Size used during DMA allocation.
96 * @npages_free: Count of available pages for re-use.
97 * @npages_in_use: Count of pages that are in use.
98 * @nfrees: Stats when pool is shrinking.
99 * @nrefills: Stats when the pool is grown.
100 * @gfp_flags: Flags to pass for alloc_page.
101 * @name: Name of the pool.
102 * @dev_name: Name derieved from dev - similar to how dev_info works.
103 *   Used during shutdown as the dev_info during release is unavailable.
104 */
105struct dma_pool {
106	struct list_head pools; /* The 'struct device->dma_pools link */
107	enum pool_type type;
108	spinlock_t lock;
109	struct list_head inuse_list;
110	struct list_head free_list;
111	struct device *dev;
112	unsigned size;
113	unsigned npages_free;
114	unsigned npages_in_use;
115	unsigned long nfrees; /* Stats when shrunk. */
116	unsigned long nrefills; /* Stats when grown. */
117	gfp_t gfp_flags;
118	char name[13]; /* "cached dma32" */
119	char dev_name[64]; /* Constructed from dev */
120};
121
122/*
123 * The accounting page keeping track of the allocated page along with
124 * the DMA address.
125 * @page_list: The link to the 'page_list' in 'struct dma_pool'.
126 * @vaddr: The virtual address of the page
127 * @dma: The bus address of the page. If the page is not allocated
128 *   via the DMA API, it will be -1.
129 */
130struct dma_page {
131	struct list_head page_list;
132	void *vaddr;
133	struct page *p;
134	dma_addr_t dma;
135};
136
137/*
138 * Limits for the pool. They are handled without locks because only place where
139 * they may change is in sysfs store. They won't have immediate effect anyway
140 * so forcing serialization to access them is pointless.
141 */
142
143struct ttm_pool_opts {
144	unsigned	alloc_size;
145	unsigned	max_size;
146	unsigned	small;
147};
148
149/*
150 * Contains the list of all of the 'struct device' and their corresponding
151 * DMA pools. Guarded by _mutex->lock.
152 * @pools: The link to 'struct ttm_pool_manager->pools'
153 * @dev: The 'struct device' associated with the 'pool'
154 * @pool: The 'struct dma_pool' associated with the 'dev'
155 */
156struct device_pools {
157	struct list_head pools;
158	struct device *dev;
159	struct dma_pool *pool;
160};
161
162/*
163 * struct ttm_pool_manager - Holds memory pools for fast allocation
164 *
165 * @lock: Lock used when adding/removing from pools
166 * @pools: List of 'struct device' and 'struct dma_pool' tuples.
167 * @options: Limits for the pool.
168 * @npools: Total amount of pools in existence.
169 * @shrinker: The structure used by [un|]register_shrinker
170 */
171struct ttm_pool_manager {
172	struct mutex		lock;
173	struct list_head	pools;
174	struct ttm_pool_opts	options;
175	unsigned		npools;
176	struct shrinker		mm_shrink;
177	struct kobject		kobj;
178};
179
180static struct ttm_pool_manager *_manager;
181
182static struct attribute ttm_page_pool_max = {
183	.name = "pool_max_size",
184	.mode = S_IRUGO | S_IWUSR
185};
186static struct attribute ttm_page_pool_small = {
187	.name = "pool_small_allocation",
188	.mode = S_IRUGO | S_IWUSR
189};
190static struct attribute ttm_page_pool_alloc_size = {
191	.name = "pool_allocation_size",
192	.mode = S_IRUGO | S_IWUSR
193};
194
195static struct attribute *ttm_pool_attrs[] = {
196	&ttm_page_pool_max,
197	&ttm_page_pool_small,
198	&ttm_page_pool_alloc_size,
199	NULL
200};
201
202static void ttm_pool_kobj_release(struct kobject *kobj)
203{
204	struct ttm_pool_manager *m =
205		container_of(kobj, struct ttm_pool_manager, kobj);
206	kfree(m);
207}
208
209static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
210			      const char *buffer, size_t size)
211{
212	struct ttm_pool_manager *m =
213		container_of(kobj, struct ttm_pool_manager, kobj);
214	int chars;
215	unsigned val;
216	chars = sscanf(buffer, "%u", &val);
217	if (chars == 0)
218		return size;
219
220	/* Convert kb to number of pages */
221	val = val / (PAGE_SIZE >> 10);
222
223	if (attr == &ttm_page_pool_max)
224		m->options.max_size = val;
225	else if (attr == &ttm_page_pool_small)
226		m->options.small = val;
227	else if (attr == &ttm_page_pool_alloc_size) {
228		if (val > NUM_PAGES_TO_ALLOC*8) {
229			pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
230			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
231			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
232			return size;
233		} else if (val > NUM_PAGES_TO_ALLOC) {
234			pr_warn("Setting allocation size to larger than %lu is not recommended\n",
235				NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
236		}
237		m->options.alloc_size = val;
238	}
239
240	return size;
241}
242
243static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
244			     char *buffer)
245{
246	struct ttm_pool_manager *m =
247		container_of(kobj, struct ttm_pool_manager, kobj);
248	unsigned val = 0;
249
250	if (attr == &ttm_page_pool_max)
251		val = m->options.max_size;
252	else if (attr == &ttm_page_pool_small)
253		val = m->options.small;
254	else if (attr == &ttm_page_pool_alloc_size)
255		val = m->options.alloc_size;
256
257	val = val * (PAGE_SIZE >> 10);
258
259	return snprintf(buffer, PAGE_SIZE, "%u\n", val);
260}
261
262static const struct sysfs_ops ttm_pool_sysfs_ops = {
263	.show = &ttm_pool_show,
264	.store = &ttm_pool_store,
265};
266
267static struct kobj_type ttm_pool_kobj_type = {
268	.release = &ttm_pool_kobj_release,
269	.sysfs_ops = &ttm_pool_sysfs_ops,
270	.default_attrs = ttm_pool_attrs,
271};
272
273#ifndef CONFIG_X86
274static int set_pages_array_wb(struct page **pages, int addrinarray)
275{
276#ifdef TTM_HAS_AGP
277	int i;
278
279	for (i = 0; i < addrinarray; i++)
280		unmap_page_from_agp(pages[i]);
281#endif
282	return 0;
283}
284
285static int set_pages_array_wc(struct page **pages, int addrinarray)
286{
287#ifdef TTM_HAS_AGP
288	int i;
289
290	for (i = 0; i < addrinarray; i++)
291		map_page_into_agp(pages[i]);
292#endif
293	return 0;
294}
295
296static int set_pages_array_uc(struct page **pages, int addrinarray)
297{
298#ifdef TTM_HAS_AGP
299	int i;
300
301	for (i = 0; i < addrinarray; i++)
302		map_page_into_agp(pages[i]);
303#endif
304	return 0;
305}
306#endif /* for !CONFIG_X86 */
307
308static int ttm_set_pages_caching(struct dma_pool *pool,
309				 struct page **pages, unsigned cpages)
310{
311	int r = 0;
312	/* Set page caching */
313	if (pool->type & IS_UC) {
314		r = set_pages_array_uc(pages, cpages);
315		if (r)
316			pr_err("%s: Failed to set %d pages to uc!\n",
317			       pool->dev_name, cpages);
318	}
319	if (pool->type & IS_WC) {
320		r = set_pages_array_wc(pages, cpages);
321		if (r)
322			pr_err("%s: Failed to set %d pages to wc!\n",
323			       pool->dev_name, cpages);
324	}
325	return r;
326}
327
328static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
329{
330	dma_addr_t dma = d_page->dma;
331	dma_free_coherent(pool->dev, pool->size, d_page->vaddr, dma);
332
333	kfree(d_page);
334	d_page = NULL;
335}
336static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
337{
338	struct dma_page *d_page;
339
340	d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
341	if (!d_page)
342		return NULL;
343
344	d_page->vaddr = dma_alloc_coherent(pool->dev, pool->size,
345					   &d_page->dma,
346					   pool->gfp_flags);
347	if (d_page->vaddr)
348		d_page->p = virt_to_page(d_page->vaddr);
349	else {
350		kfree(d_page);
351		d_page = NULL;
352	}
353	return d_page;
354}
355static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
356{
357	enum pool_type type = IS_UNDEFINED;
358
359	if (flags & TTM_PAGE_FLAG_DMA32)
360		type |= IS_DMA32;
361	if (cstate == tt_cached)
362		type |= IS_CACHED;
363	else if (cstate == tt_uncached)
364		type |= IS_UC;
365	else
366		type |= IS_WC;
367
368	return type;
369}
370
371static void ttm_pool_update_free_locked(struct dma_pool *pool,
372					unsigned freed_pages)
373{
374	pool->npages_free -= freed_pages;
375	pool->nfrees += freed_pages;
376
377}
378
379/* set memory back to wb and free the pages. */
380static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
381			      struct page *pages[], unsigned npages)
382{
383	struct dma_page *d_page, *tmp;
384
385	/* Don't set WB on WB page pool. */
386	if (npages && !(pool->type & IS_CACHED) &&
387	    set_pages_array_wb(pages, npages))
388		pr_err("%s: Failed to set %d pages to wb!\n",
389		       pool->dev_name, npages);
390
391	list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
392		list_del(&d_page->page_list);
393		__ttm_dma_free_page(pool, d_page);
394	}
395}
396
397static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
398{
399	/* Don't set WB on WB page pool. */
400	if (!(pool->type & IS_CACHED) && set_pages_array_wb(&d_page->p, 1))
401		pr_err("%s: Failed to set %d pages to wb!\n",
402		       pool->dev_name, 1);
403
404	list_del(&d_page->page_list);
405	__ttm_dma_free_page(pool, d_page);
406}
407
408/*
409 * Free pages from pool.
410 *
411 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
412 * number of pages in one go.
413 *
414 * @pool: to free the pages from
415 * @nr_free: If set to true will free all pages in pool
416 **/
417static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free)
418{
419	unsigned long irq_flags;
420	struct dma_page *dma_p, *tmp;
421	struct page **pages_to_free;
422	struct list_head d_pages;
423	unsigned freed_pages = 0,
424		 npages_to_free = nr_free;
425
426	if (NUM_PAGES_TO_ALLOC < nr_free)
427		npages_to_free = NUM_PAGES_TO_ALLOC;
428#if 0
429	if (nr_free > 1) {
430		pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
431			 pool->dev_name, pool->name, current->pid,
432			 npages_to_free, nr_free);
433	}
434#endif
435	pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
436			GFP_KERNEL);
437
438	if (!pages_to_free) {
439		pr_err("%s: Failed to allocate memory for pool free operation\n",
440		       pool->dev_name);
441		return 0;
442	}
443	INIT_LIST_HEAD(&d_pages);
444restart:
445	spin_lock_irqsave(&pool->lock, irq_flags);
446
447	/* We picking the oldest ones off the list */
448	list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
449					 page_list) {
450		if (freed_pages >= npages_to_free)
451			break;
452
453		/* Move the dma_page from one list to another. */
454		list_move(&dma_p->page_list, &d_pages);
455
456		pages_to_free[freed_pages++] = dma_p->p;
457		/* We can only remove NUM_PAGES_TO_ALLOC at a time. */
458		if (freed_pages >= NUM_PAGES_TO_ALLOC) {
459
460			ttm_pool_update_free_locked(pool, freed_pages);
461			/**
462			 * Because changing page caching is costly
463			 * we unlock the pool to prevent stalling.
464			 */
465			spin_unlock_irqrestore(&pool->lock, irq_flags);
466
467			ttm_dma_pages_put(pool, &d_pages, pages_to_free,
468					  freed_pages);
469
470			INIT_LIST_HEAD(&d_pages);
471
472			if (likely(nr_free != FREE_ALL_PAGES))
473				nr_free -= freed_pages;
474
475			if (NUM_PAGES_TO_ALLOC >= nr_free)
476				npages_to_free = nr_free;
477			else
478				npages_to_free = NUM_PAGES_TO_ALLOC;
479
480			freed_pages = 0;
481
482			/* free all so restart the processing */
483			if (nr_free)
484				goto restart;
485
486			/* Not allowed to fall through or break because
487			 * following context is inside spinlock while we are
488			 * outside here.
489			 */
490			goto out;
491
492		}
493	}
494
495	/* remove range of pages from the pool */
496	if (freed_pages) {
497		ttm_pool_update_free_locked(pool, freed_pages);
498		nr_free -= freed_pages;
499	}
500
501	spin_unlock_irqrestore(&pool->lock, irq_flags);
502
503	if (freed_pages)
504		ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
505out:
506	kfree(pages_to_free);
507	return nr_free;
508}
509
510static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
511{
512	struct device_pools *p;
513	struct dma_pool *pool;
514
515	if (!dev)
516		return;
517
518	mutex_lock(&_manager->lock);
519	list_for_each_entry_reverse(p, &_manager->pools, pools) {
520		if (p->dev != dev)
521			continue;
522		pool = p->pool;
523		if (pool->type != type)
524			continue;
525
526		list_del(&p->pools);
527		kfree(p);
528		_manager->npools--;
529		break;
530	}
531	list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
532		if (pool->type != type)
533			continue;
534		/* Takes a spinlock.. */
535		ttm_dma_page_pool_free(pool, FREE_ALL_PAGES);
536		WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
537		/* This code path is called after _all_ references to the
538		 * struct device has been dropped - so nobody should be
539		 * touching it. In case somebody is trying to _add_ we are
540		 * guarded by the mutex. */
541		list_del(&pool->pools);
542		kfree(pool);
543		break;
544	}
545	mutex_unlock(&_manager->lock);
546}
547
548/*
549 * On free-ing of the 'struct device' this deconstructor is run.
550 * Albeit the pool might have already been freed earlier.
551 */
552static void ttm_dma_pool_release(struct device *dev, void *res)
553{
554	struct dma_pool *pool = *(struct dma_pool **)res;
555
556	if (pool)
557		ttm_dma_free_pool(dev, pool->type);
558}
559
560static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
561{
562	return *(struct dma_pool **)res == match_data;
563}
564
565static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
566					  enum pool_type type)
567{
568	char *n[] = {"wc", "uc", "cached", " dma32", "unknown",};
569	enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_UNDEFINED};
570	struct device_pools *sec_pool = NULL;
571	struct dma_pool *pool = NULL, **ptr;
572	unsigned i;
573	int ret = -ENODEV;
574	char *p;
575
576	if (!dev)
577		return NULL;
578
579	ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
580	if (!ptr)
581		return NULL;
582
583	ret = -ENOMEM;
584
585	pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
586			    dev_to_node(dev));
587	if (!pool)
588		goto err_mem;
589
590	sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
591				dev_to_node(dev));
592	if (!sec_pool)
593		goto err_mem;
594
595	INIT_LIST_HEAD(&sec_pool->pools);
596	sec_pool->dev = dev;
597	sec_pool->pool =  pool;
598
599	INIT_LIST_HEAD(&pool->free_list);
600	INIT_LIST_HEAD(&pool->inuse_list);
601	INIT_LIST_HEAD(&pool->pools);
602	spin_lock_init(&pool->lock);
603	pool->dev = dev;
604	pool->npages_free = pool->npages_in_use = 0;
605	pool->nfrees = 0;
606	pool->gfp_flags = flags;
607	pool->size = PAGE_SIZE;
608	pool->type = type;
609	pool->nrefills = 0;
610	p = pool->name;
611	for (i = 0; i < 5; i++) {
612		if (type & t[i]) {
613			p += snprintf(p, sizeof(pool->name) - (p - pool->name),
614				      "%s", n[i]);
615		}
616	}
617	*p = 0;
618	/* We copy the name for pr_ calls b/c when dma_pool_destroy is called
619	 * - the kobj->name has already been deallocated.*/
620	snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
621		 dev_driver_string(dev), dev_name(dev));
622	mutex_lock(&_manager->lock);
623	/* You can get the dma_pool from either the global: */
624	list_add(&sec_pool->pools, &_manager->pools);
625	_manager->npools++;
626	/* or from 'struct device': */
627	list_add(&pool->pools, &dev->dma_pools);
628	mutex_unlock(&_manager->lock);
629
630	*ptr = pool;
631	devres_add(dev, ptr);
632
633	return pool;
634err_mem:
635	devres_free(ptr);
636	kfree(sec_pool);
637	kfree(pool);
638	return ERR_PTR(ret);
639}
640
641static struct dma_pool *ttm_dma_find_pool(struct device *dev,
642					  enum pool_type type)
643{
644	struct dma_pool *pool, *tmp, *found = NULL;
645
646	if (type == IS_UNDEFINED)
647		return found;
648
649	/* NB: We iterate on the 'struct dev' which has no spinlock, but
650	 * it does have a kref which we have taken. The kref is taken during
651	 * graphic driver loading - in the drm_pci_init it calls either
652	 * pci_dev_get or pci_register_driver which both end up taking a kref
653	 * on 'struct device'.
654	 *
655	 * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
656	 * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
657	 * thing is at that point of time there are no pages associated with the
658	 * driver so this function will not be called.
659	 */
660	list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools) {
661		if (pool->type != type)
662			continue;
663		found = pool;
664		break;
665	}
666	return found;
667}
668
669/*
670 * Free pages the pages that failed to change the caching state. If there
671 * are pages that have changed their caching state already put them to the
672 * pool.
673 */
674static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
675						 struct list_head *d_pages,
676						 struct page **failed_pages,
677						 unsigned cpages)
678{
679	struct dma_page *d_page, *tmp;
680	struct page *p;
681	unsigned i = 0;
682
683	p = failed_pages[0];
684	if (!p)
685		return;
686	/* Find the failed page. */
687	list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
688		if (d_page->p != p)
689			continue;
690		/* .. and then progress over the full list. */
691		list_del(&d_page->page_list);
692		__ttm_dma_free_page(pool, d_page);
693		if (++i < cpages)
694			p = failed_pages[i];
695		else
696			break;
697	}
698
699}
700
701/*
702 * Allocate 'count' pages, and put 'need' number of them on the
703 * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
704 * The full list of pages should also be on 'd_pages'.
705 * We return zero for success, and negative numbers as errors.
706 */
707static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
708					struct list_head *d_pages,
709					unsigned count)
710{
711	struct page **caching_array;
712	struct dma_page *dma_p;
713	struct page *p;
714	int r = 0;
715	unsigned i, cpages;
716	unsigned max_cpages = min(count,
717			(unsigned)(PAGE_SIZE/sizeof(struct page *)));
718
719	/* allocate array for page caching change */
720	caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
721
722	if (!caching_array) {
723		pr_err("%s: Unable to allocate table for new pages\n",
724		       pool->dev_name);
725		return -ENOMEM;
726	}
727
728	if (count > 1) {
729		pr_debug("%s: (%s:%d) Getting %d pages\n",
730			 pool->dev_name, pool->name, current->pid, count);
731	}
732
733	for (i = 0, cpages = 0; i < count; ++i) {
734		dma_p = __ttm_dma_alloc_page(pool);
735		if (!dma_p) {
736			pr_err("%s: Unable to get page %u\n",
737			       pool->dev_name, i);
738
739			/* store already allocated pages in the pool after
740			 * setting the caching state */
741			if (cpages) {
742				r = ttm_set_pages_caching(pool, caching_array,
743							  cpages);
744				if (r)
745					ttm_dma_handle_caching_state_failure(
746						pool, d_pages, caching_array,
747						cpages);
748			}
749			r = -ENOMEM;
750			goto out;
751		}
752		p = dma_p->p;
753#ifdef CONFIG_HIGHMEM
754		/* gfp flags of highmem page should never be dma32 so we
755		 * we should be fine in such case
756		 */
757		if (!PageHighMem(p))
758#endif
759		{
760			caching_array[cpages++] = p;
761			if (cpages == max_cpages) {
762				/* Note: Cannot hold the spinlock */
763				r = ttm_set_pages_caching(pool, caching_array,
764						 cpages);
765				if (r) {
766					ttm_dma_handle_caching_state_failure(
767						pool, d_pages, caching_array,
768						cpages);
769					goto out;
770				}
771				cpages = 0;
772			}
773		}
774		list_add(&dma_p->page_list, d_pages);
775	}
776
777	if (cpages) {
778		r = ttm_set_pages_caching(pool, caching_array, cpages);
779		if (r)
780			ttm_dma_handle_caching_state_failure(pool, d_pages,
781					caching_array, cpages);
782	}
783out:
784	kfree(caching_array);
785	return r;
786}
787
788/*
789 * @return count of pages still required to fulfill the request.
790 */
791static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
792					 unsigned long *irq_flags)
793{
794	unsigned count = _manager->options.small;
795	int r = pool->npages_free;
796
797	if (count > pool->npages_free) {
798		struct list_head d_pages;
799
800		INIT_LIST_HEAD(&d_pages);
801
802		spin_unlock_irqrestore(&pool->lock, *irq_flags);
803
804		/* Returns how many more are neccessary to fulfill the
805		 * request. */
806		r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
807
808		spin_lock_irqsave(&pool->lock, *irq_flags);
809		if (!r) {
810			/* Add the fresh to the end.. */
811			list_splice(&d_pages, &pool->free_list);
812			++pool->nrefills;
813			pool->npages_free += count;
814			r = count;
815		} else {
816			struct dma_page *d_page;
817			unsigned cpages = 0;
818
819			pr_err("%s: Failed to fill %s pool (r:%d)!\n",
820			       pool->dev_name, pool->name, r);
821
822			list_for_each_entry(d_page, &d_pages, page_list) {
823				cpages++;
824			}
825			list_splice_tail(&d_pages, &pool->free_list);
826			pool->npages_free += cpages;
827			r = cpages;
828		}
829	}
830	return r;
831}
832
833/*
834 * @return count of pages still required to fulfill the request.
835 * The populate list is actually a stack (not that is matters as TTM
836 * allocates one page at a time.
837 */
838static int ttm_dma_pool_get_pages(struct dma_pool *pool,
839				  struct ttm_dma_tt *ttm_dma,
840				  unsigned index)
841{
842	struct dma_page *d_page;
843	struct ttm_tt *ttm = &ttm_dma->ttm;
844	unsigned long irq_flags;
845	int count, r = -ENOMEM;
846
847	spin_lock_irqsave(&pool->lock, irq_flags);
848	count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
849	if (count) {
850		d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
851		ttm->pages[index] = d_page->p;
852		ttm_dma->dma_address[index] = d_page->dma;
853		list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
854		r = 0;
855		pool->npages_in_use += 1;
856		pool->npages_free -= 1;
857	}
858	spin_unlock_irqrestore(&pool->lock, irq_flags);
859	return r;
860}
861
862/*
863 * On success pages list will hold count number of correctly
864 * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
865 */
866int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev)
867{
868	struct ttm_tt *ttm = &ttm_dma->ttm;
869	struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
870	struct dma_pool *pool;
871	enum pool_type type;
872	unsigned i;
873	gfp_t gfp_flags;
874	int ret;
875
876	if (ttm->state != tt_unpopulated)
877		return 0;
878
879	type = ttm_to_type(ttm->page_flags, ttm->caching_state);
880	if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
881		gfp_flags = GFP_USER | GFP_DMA32;
882	else
883		gfp_flags = GFP_HIGHUSER;
884	if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
885		gfp_flags |= __GFP_ZERO;
886
887	pool = ttm_dma_find_pool(dev, type);
888	if (!pool) {
889		pool = ttm_dma_pool_init(dev, gfp_flags, type);
890		if (IS_ERR_OR_NULL(pool)) {
891			return -ENOMEM;
892		}
893	}
894
895	INIT_LIST_HEAD(&ttm_dma->pages_list);
896	for (i = 0; i < ttm->num_pages; ++i) {
897		ret = ttm_dma_pool_get_pages(pool, ttm_dma, i);
898		if (ret != 0) {
899			ttm_dma_unpopulate(ttm_dma, dev);
900			return -ENOMEM;
901		}
902
903		ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
904						false, false);
905		if (unlikely(ret != 0)) {
906			ttm_dma_unpopulate(ttm_dma, dev);
907			return -ENOMEM;
908		}
909	}
910
911	if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
912		ret = ttm_tt_swapin(ttm);
913		if (unlikely(ret != 0)) {
914			ttm_dma_unpopulate(ttm_dma, dev);
915			return ret;
916		}
917	}
918
919	ttm->state = tt_unbound;
920	return 0;
921}
922EXPORT_SYMBOL_GPL(ttm_dma_populate);
923
924/* Get good estimation how many pages are free in pools */
925static int ttm_dma_pool_get_num_unused_pages(void)
926{
927	struct device_pools *p;
928	unsigned total = 0;
929
930	mutex_lock(&_manager->lock);
931	list_for_each_entry(p, &_manager->pools, pools)
932		total += p->pool->npages_free;
933	mutex_unlock(&_manager->lock);
934	return total;
935}
936
937/* Put all pages in pages list to correct pool to wait for reuse */
938void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
939{
940	struct ttm_tt *ttm = &ttm_dma->ttm;
941	struct dma_pool *pool;
942	struct dma_page *d_page, *next;
943	enum pool_type type;
944	bool is_cached = false;
945	unsigned count = 0, i, npages = 0;
946	unsigned long irq_flags;
947
948	type = ttm_to_type(ttm->page_flags, ttm->caching_state);
949	pool = ttm_dma_find_pool(dev, type);
950	if (!pool)
951		return;
952
953	is_cached = (ttm_dma_find_pool(pool->dev,
954		     ttm_to_type(ttm->page_flags, tt_cached)) == pool);
955
956	/* make sure pages array match list and count number of pages */
957	list_for_each_entry(d_page, &ttm_dma->pages_list, page_list) {
958		ttm->pages[count] = d_page->p;
959		count++;
960	}
961
962	spin_lock_irqsave(&pool->lock, irq_flags);
963	pool->npages_in_use -= count;
964	if (is_cached) {
965		pool->nfrees += count;
966	} else {
967		pool->npages_free += count;
968		list_splice(&ttm_dma->pages_list, &pool->free_list);
969		npages = count;
970		if (pool->npages_free > _manager->options.max_size) {
971			npages = pool->npages_free - _manager->options.max_size;
972			/* free at least NUM_PAGES_TO_ALLOC number of pages
973			 * to reduce calls to set_memory_wb */
974			if (npages < NUM_PAGES_TO_ALLOC)
975				npages = NUM_PAGES_TO_ALLOC;
976		}
977	}
978	spin_unlock_irqrestore(&pool->lock, irq_flags);
979
980	if (is_cached) {
981		list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list, page_list) {
982			ttm_mem_global_free_page(ttm->glob->mem_glob,
983						 d_page->p);
984			ttm_dma_page_put(pool, d_page);
985		}
986	} else {
987		for (i = 0; i < count; i++) {
988			ttm_mem_global_free_page(ttm->glob->mem_glob,
989						 ttm->pages[i]);
990		}
991	}
992
993	INIT_LIST_HEAD(&ttm_dma->pages_list);
994	for (i = 0; i < ttm->num_pages; i++) {
995		ttm->pages[i] = NULL;
996		ttm_dma->dma_address[i] = 0;
997	}
998
999	/* shrink pool if necessary (only on !is_cached pools)*/
1000	if (npages)
1001		ttm_dma_page_pool_free(pool, npages);
1002	ttm->state = tt_unpopulated;
1003}
1004EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1005
1006/**
1007 * Callback for mm to request pool to reduce number of page held.
1008 */
1009static int ttm_dma_pool_mm_shrink(struct shrinker *shrink,
1010				  struct shrink_control *sc)
1011{
1012	static atomic_t start_pool = ATOMIC_INIT(0);
1013	unsigned idx = 0;
1014	unsigned pool_offset = atomic_add_return(1, &start_pool);
1015	unsigned shrink_pages = sc->nr_to_scan;
1016	struct device_pools *p;
1017
1018	if (list_empty(&_manager->pools))
1019		return 0;
1020
1021	mutex_lock(&_manager->lock);
1022	pool_offset = pool_offset % _manager->npools;
1023	list_for_each_entry(p, &_manager->pools, pools) {
1024		unsigned nr_free;
1025
1026		if (!p->dev)
1027			continue;
1028		if (shrink_pages == 0)
1029			break;
1030		/* Do it in round-robin fashion. */
1031		if (++idx < pool_offset)
1032			continue;
1033		nr_free = shrink_pages;
1034		shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free);
1035		pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1036			 p->pool->dev_name, p->pool->name, current->pid,
1037			 nr_free, shrink_pages);
1038	}
1039	mutex_unlock(&_manager->lock);
1040	/* return estimated number of unused pages in pool */
1041	return ttm_dma_pool_get_num_unused_pages();
1042}
1043
1044static void ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1045{
1046	manager->mm_shrink.shrink = &ttm_dma_pool_mm_shrink;
1047	manager->mm_shrink.seeks = 1;
1048	register_shrinker(&manager->mm_shrink);
1049}
1050
1051static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1052{
1053	unregister_shrinker(&manager->mm_shrink);
1054}
1055
1056int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1057{
1058	int ret = -ENOMEM;
1059
1060	WARN_ON(_manager);
1061
1062	pr_info("Initializing DMA pool allocator\n");
1063
1064	_manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1065	if (!_manager)
1066		goto err;
1067
1068	mutex_init(&_manager->lock);
1069	INIT_LIST_HEAD(&_manager->pools);
1070
1071	_manager->options.max_size = max_pages;
1072	_manager->options.small = SMALL_ALLOCATION;
1073	_manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1074
1075	/* This takes care of auto-freeing the _manager */
1076	ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1077				   &glob->kobj, "dma_pool");
1078	if (unlikely(ret != 0)) {
1079		kobject_put(&_manager->kobj);
1080		goto err;
1081	}
1082	ttm_dma_pool_mm_shrink_init(_manager);
1083	return 0;
1084err:
1085	return ret;
1086}
1087
1088void ttm_dma_page_alloc_fini(void)
1089{
1090	struct device_pools *p, *t;
1091
1092	pr_info("Finalizing DMA pool allocator\n");
1093	ttm_dma_pool_mm_shrink_fini(_manager);
1094
1095	list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1096		dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1097			current->pid);
1098		WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1099			ttm_dma_pool_match, p->pool));
1100		ttm_dma_free_pool(p->dev, p->pool->type);
1101	}
1102	kobject_put(&_manager->kobj);
1103	_manager = NULL;
1104}
1105
1106int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1107{
1108	struct device_pools *p;
1109	struct dma_pool *pool = NULL;
1110	char *h[] = {"pool", "refills", "pages freed", "inuse", "available",
1111		     "name", "virt", "busaddr"};
1112
1113	if (!_manager) {
1114		seq_printf(m, "No pool allocator running.\n");
1115		return 0;
1116	}
1117	seq_printf(m, "%13s %12s %13s %8s %8s %8s\n",
1118		   h[0], h[1], h[2], h[3], h[4], h[5]);
1119	mutex_lock(&_manager->lock);
1120	list_for_each_entry(p, &_manager->pools, pools) {
1121		struct device *dev = p->dev;
1122		if (!dev)
1123			continue;
1124		pool = p->pool;
1125		seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1126				pool->name, pool->nrefills,
1127				pool->nfrees, pool->npages_in_use,
1128				pool->npages_free,
1129				pool->dev_name);
1130	}
1131	mutex_unlock(&_manager->lock);
1132	return 0;
1133}
1134EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1135