1/*
2 * Copyright 2011 (c) Oracle Corp.
3
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sub license,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the
12 * next paragraph) shall be included in all copies or substantial portions
13 * of the Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
22 *
23 * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
24 */
25
26/*
27 * A simple DMA pool losely based on dmapool.c. It has certain advantages
28 * over the DMA pools:
29 * - Pool collects resently freed pages for reuse (and hooks up to
30 *   the shrinker).
31 * - Tracks currently in use pages
32 * - Tracks whether the page is UC, WB or cached (and reverts to WB
33 *   when freed).
34 */
35
36#if defined(CONFIG_SWIOTLB) || defined(CONFIG_INTEL_IOMMU)
37#define pr_fmt(fmt) "[TTM] " fmt
38
39#include <linux/dma-mapping.h>
40#include <linux/list.h>
41#include <linux/seq_file.h> /* for seq_printf */
42#include <linux/slab.h>
43#include <linux/spinlock.h>
44#include <linux/highmem.h>
45#include <linux/mm_types.h>
46#include <linux/module.h>
47#include <linux/mm.h>
48#include <linux/atomic.h>
49#include <linux/device.h>
50#include <linux/kthread.h>
51#include <drm/ttm/ttm_bo_driver.h>
52#include <drm/ttm/ttm_page_alloc.h>
53#ifdef TTM_HAS_AGP
54#include <asm/agp.h>
55#endif
56
57#define NUM_PAGES_TO_ALLOC		(PAGE_SIZE/sizeof(struct page *))
58#define SMALL_ALLOCATION		4
59#define FREE_ALL_PAGES			(~0U)
60/* times are in msecs */
61#define IS_UNDEFINED			(0)
62#define IS_WC				(1<<1)
63#define IS_UC				(1<<2)
64#define IS_CACHED			(1<<3)
65#define IS_DMA32			(1<<4)
66
67enum pool_type {
68	POOL_IS_UNDEFINED,
69	POOL_IS_WC = IS_WC,
70	POOL_IS_UC = IS_UC,
71	POOL_IS_CACHED = IS_CACHED,
72	POOL_IS_WC_DMA32 = IS_WC | IS_DMA32,
73	POOL_IS_UC_DMA32 = IS_UC | IS_DMA32,
74	POOL_IS_CACHED_DMA32 = IS_CACHED | IS_DMA32,
75};
76/*
77 * The pool structure. There are usually six pools:
78 *  - generic (not restricted to DMA32):
79 *      - write combined, uncached, cached.
80 *  - dma32 (up to 2^32 - so up 4GB):
81 *      - write combined, uncached, cached.
82 * for each 'struct device'. The 'cached' is for pages that are actively used.
83 * The other ones can be shrunk by the shrinker API if neccessary.
84 * @pools: The 'struct device->dma_pools' link.
85 * @type: Type of the pool
86 * @lock: Protects the inuse_list and free_list from concurrnet access. Must be
87 * used with irqsave/irqrestore variants because pool allocator maybe called
88 * from delayed work.
89 * @inuse_list: Pool of pages that are in use. The order is very important and
90 *   it is in the order that the TTM pages that are put back are in.
91 * @free_list: Pool of pages that are free to be used. No order requirements.
92 * @dev: The device that is associated with these pools.
93 * @size: Size used during DMA allocation.
94 * @npages_free: Count of available pages for re-use.
95 * @npages_in_use: Count of pages that are in use.
96 * @nfrees: Stats when pool is shrinking.
97 * @nrefills: Stats when the pool is grown.
98 * @gfp_flags: Flags to pass for alloc_page.
99 * @name: Name of the pool.
100 * @dev_name: Name derieved from dev - similar to how dev_info works.
101 *   Used during shutdown as the dev_info during release is unavailable.
102 */
103struct dma_pool {
104	struct list_head pools; /* The 'struct device->dma_pools link */
105	enum pool_type type;
106	spinlock_t lock;
107	struct list_head inuse_list;
108	struct list_head free_list;
109	struct device *dev;
110	unsigned size;
111	unsigned npages_free;
112	unsigned npages_in_use;
113	unsigned long nfrees; /* Stats when shrunk. */
114	unsigned long nrefills; /* Stats when grown. */
115	gfp_t gfp_flags;
116	char name[13]; /* "cached dma32" */
117	char dev_name[64]; /* Constructed from dev */
118};
119
120/*
121 * The accounting page keeping track of the allocated page along with
122 * the DMA address.
123 * @page_list: The link to the 'page_list' in 'struct dma_pool'.
124 * @vaddr: The virtual address of the page
125 * @dma: The bus address of the page. If the page is not allocated
126 *   via the DMA API, it will be -1.
127 */
128struct dma_page {
129	struct list_head page_list;
130	void *vaddr;
131	struct page *p;
132	dma_addr_t dma;
133};
134
135/*
136 * Limits for the pool. They are handled without locks because only place where
137 * they may change is in sysfs store. They won't have immediate effect anyway
138 * so forcing serialization to access them is pointless.
139 */
140
141struct ttm_pool_opts {
142	unsigned	alloc_size;
143	unsigned	max_size;
144	unsigned	small;
145};
146
147/*
148 * Contains the list of all of the 'struct device' and their corresponding
149 * DMA pools. Guarded by _mutex->lock.
150 * @pools: The link to 'struct ttm_pool_manager->pools'
151 * @dev: The 'struct device' associated with the 'pool'
152 * @pool: The 'struct dma_pool' associated with the 'dev'
153 */
154struct device_pools {
155	struct list_head pools;
156	struct device *dev;
157	struct dma_pool *pool;
158};
159
160/*
161 * struct ttm_pool_manager - Holds memory pools for fast allocation
162 *
163 * @lock: Lock used when adding/removing from pools
164 * @pools: List of 'struct device' and 'struct dma_pool' tuples.
165 * @options: Limits for the pool.
166 * @npools: Total amount of pools in existence.
167 * @shrinker: The structure used by [un|]register_shrinker
168 */
169struct ttm_pool_manager {
170	struct mutex		lock;
171	struct list_head	pools;
172	struct ttm_pool_opts	options;
173	unsigned		npools;
174	struct shrinker		mm_shrink;
175	struct kobject		kobj;
176};
177
178static struct ttm_pool_manager *_manager;
179
180static struct attribute ttm_page_pool_max = {
181	.name = "pool_max_size",
182	.mode = S_IRUGO | S_IWUSR
183};
184static struct attribute ttm_page_pool_small = {
185	.name = "pool_small_allocation",
186	.mode = S_IRUGO | S_IWUSR
187};
188static struct attribute ttm_page_pool_alloc_size = {
189	.name = "pool_allocation_size",
190	.mode = S_IRUGO | S_IWUSR
191};
192
193static struct attribute *ttm_pool_attrs[] = {
194	&ttm_page_pool_max,
195	&ttm_page_pool_small,
196	&ttm_page_pool_alloc_size,
197	NULL
198};
199
200static void ttm_pool_kobj_release(struct kobject *kobj)
201{
202	struct ttm_pool_manager *m =
203		container_of(kobj, struct ttm_pool_manager, kobj);
204	kfree(m);
205}
206
207static ssize_t ttm_pool_store(struct kobject *kobj, struct attribute *attr,
208			      const char *buffer, size_t size)
209{
210	struct ttm_pool_manager *m =
211		container_of(kobj, struct ttm_pool_manager, kobj);
212	int chars;
213	unsigned val;
214	chars = sscanf(buffer, "%u", &val);
215	if (chars == 0)
216		return size;
217
218	/* Convert kb to number of pages */
219	val = val / (PAGE_SIZE >> 10);
220
221	if (attr == &ttm_page_pool_max)
222		m->options.max_size = val;
223	else if (attr == &ttm_page_pool_small)
224		m->options.small = val;
225	else if (attr == &ttm_page_pool_alloc_size) {
226		if (val > NUM_PAGES_TO_ALLOC*8) {
227			pr_err("Setting allocation size to %lu is not allowed. Recommended size is %lu\n",
228			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 7),
229			       NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
230			return size;
231		} else if (val > NUM_PAGES_TO_ALLOC) {
232			pr_warn("Setting allocation size to larger than %lu is not recommended\n",
233				NUM_PAGES_TO_ALLOC*(PAGE_SIZE >> 10));
234		}
235		m->options.alloc_size = val;
236	}
237
238	return size;
239}
240
241static ssize_t ttm_pool_show(struct kobject *kobj, struct attribute *attr,
242			     char *buffer)
243{
244	struct ttm_pool_manager *m =
245		container_of(kobj, struct ttm_pool_manager, kobj);
246	unsigned val = 0;
247
248	if (attr == &ttm_page_pool_max)
249		val = m->options.max_size;
250	else if (attr == &ttm_page_pool_small)
251		val = m->options.small;
252	else if (attr == &ttm_page_pool_alloc_size)
253		val = m->options.alloc_size;
254
255	val = val * (PAGE_SIZE >> 10);
256
257	return snprintf(buffer, PAGE_SIZE, "%u\n", val);
258}
259
260static const struct sysfs_ops ttm_pool_sysfs_ops = {
261	.show = &ttm_pool_show,
262	.store = &ttm_pool_store,
263};
264
265static struct kobj_type ttm_pool_kobj_type = {
266	.release = &ttm_pool_kobj_release,
267	.sysfs_ops = &ttm_pool_sysfs_ops,
268	.default_attrs = ttm_pool_attrs,
269};
270
271#ifndef CONFIG_X86
272static int set_pages_array_wb(struct page **pages, int addrinarray)
273{
274#ifdef TTM_HAS_AGP
275	int i;
276
277	for (i = 0; i < addrinarray; i++)
278		unmap_page_from_agp(pages[i]);
279#endif
280	return 0;
281}
282
283static int set_pages_array_wc(struct page **pages, int addrinarray)
284{
285#ifdef TTM_HAS_AGP
286	int i;
287
288	for (i = 0; i < addrinarray; i++)
289		map_page_into_agp(pages[i]);
290#endif
291	return 0;
292}
293
294static int set_pages_array_uc(struct page **pages, int addrinarray)
295{
296#ifdef TTM_HAS_AGP
297	int i;
298
299	for (i = 0; i < addrinarray; i++)
300		map_page_into_agp(pages[i]);
301#endif
302	return 0;
303}
304#endif /* for !CONFIG_X86 */
305
306static int ttm_set_pages_caching(struct dma_pool *pool,
307				 struct page **pages, unsigned cpages)
308{
309	int r = 0;
310	/* Set page caching */
311	if (pool->type & IS_UC) {
312		r = set_pages_array_uc(pages, cpages);
313		if (r)
314			pr_err("%s: Failed to set %d pages to uc!\n",
315			       pool->dev_name, cpages);
316	}
317	if (pool->type & IS_WC) {
318		r = set_pages_array_wc(pages, cpages);
319		if (r)
320			pr_err("%s: Failed to set %d pages to wc!\n",
321			       pool->dev_name, cpages);
322	}
323	return r;
324}
325
326static void __ttm_dma_free_page(struct dma_pool *pool, struct dma_page *d_page)
327{
328	dma_addr_t dma = d_page->dma;
329	dma_free_coherent(pool->dev, pool->size, d_page->vaddr, dma);
330
331	kfree(d_page);
332	d_page = NULL;
333}
334static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
335{
336	struct dma_page *d_page;
337
338	d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
339	if (!d_page)
340		return NULL;
341
342	d_page->vaddr = dma_alloc_coherent(pool->dev, pool->size,
343					   &d_page->dma,
344					   pool->gfp_flags);
345	if (d_page->vaddr) {
346		if (is_vmalloc_addr(d_page->vaddr))
347			d_page->p = vmalloc_to_page(d_page->vaddr);
348		else
349			d_page->p = virt_to_page(d_page->vaddr);
350	} else {
351		kfree(d_page);
352		d_page = NULL;
353	}
354	return d_page;
355}
356static enum pool_type ttm_to_type(int flags, enum ttm_caching_state cstate)
357{
358	enum pool_type type = IS_UNDEFINED;
359
360	if (flags & TTM_PAGE_FLAG_DMA32)
361		type |= IS_DMA32;
362	if (cstate == tt_cached)
363		type |= IS_CACHED;
364	else if (cstate == tt_uncached)
365		type |= IS_UC;
366	else
367		type |= IS_WC;
368
369	return type;
370}
371
372static void ttm_pool_update_free_locked(struct dma_pool *pool,
373					unsigned freed_pages)
374{
375	pool->npages_free -= freed_pages;
376	pool->nfrees += freed_pages;
377
378}
379
380/* set memory back to wb and free the pages. */
381static void ttm_dma_pages_put(struct dma_pool *pool, struct list_head *d_pages,
382			      struct page *pages[], unsigned npages)
383{
384	struct dma_page *d_page, *tmp;
385
386	/* Don't set WB on WB page pool. */
387	if (npages && !(pool->type & IS_CACHED) &&
388	    set_pages_array_wb(pages, npages))
389		pr_err("%s: Failed to set %d pages to wb!\n",
390		       pool->dev_name, npages);
391
392	list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
393		list_del(&d_page->page_list);
394		__ttm_dma_free_page(pool, d_page);
395	}
396}
397
398static void ttm_dma_page_put(struct dma_pool *pool, struct dma_page *d_page)
399{
400	/* Don't set WB on WB page pool. */
401	if (!(pool->type & IS_CACHED) && set_pages_array_wb(&d_page->p, 1))
402		pr_err("%s: Failed to set %d pages to wb!\n",
403		       pool->dev_name, 1);
404
405	list_del(&d_page->page_list);
406	__ttm_dma_free_page(pool, d_page);
407}
408
409/*
410 * Free pages from pool.
411 *
412 * To prevent hogging the ttm_swap process we only free NUM_PAGES_TO_ALLOC
413 * number of pages in one go.
414 *
415 * @pool: to free the pages from
416 * @nr_free: If set to true will free all pages in pool
417 * @use_static: Safe to use static buffer
418 **/
419static unsigned ttm_dma_page_pool_free(struct dma_pool *pool, unsigned nr_free,
420				       bool use_static)
421{
422	static struct page *static_buf[NUM_PAGES_TO_ALLOC];
423	unsigned long irq_flags;
424	struct dma_page *dma_p, *tmp;
425	struct page **pages_to_free;
426	struct list_head d_pages;
427	unsigned freed_pages = 0,
428		 npages_to_free = nr_free;
429
430	if (NUM_PAGES_TO_ALLOC < nr_free)
431		npages_to_free = NUM_PAGES_TO_ALLOC;
432#if 0
433	if (nr_free > 1) {
434		pr_debug("%s: (%s:%d) Attempting to free %d (%d) pages\n",
435			 pool->dev_name, pool->name, current->pid,
436			 npages_to_free, nr_free);
437	}
438#endif
439	if (use_static)
440		pages_to_free = static_buf;
441	else
442		pages_to_free = kmalloc(npages_to_free * sizeof(struct page *),
443					GFP_KERNEL);
444
445	if (!pages_to_free) {
446		pr_err("%s: Failed to allocate memory for pool free operation\n",
447		       pool->dev_name);
448		return 0;
449	}
450	INIT_LIST_HEAD(&d_pages);
451restart:
452	spin_lock_irqsave(&pool->lock, irq_flags);
453
454	/* We picking the oldest ones off the list */
455	list_for_each_entry_safe_reverse(dma_p, tmp, &pool->free_list,
456					 page_list) {
457		if (freed_pages >= npages_to_free)
458			break;
459
460		/* Move the dma_page from one list to another. */
461		list_move(&dma_p->page_list, &d_pages);
462
463		pages_to_free[freed_pages++] = dma_p->p;
464		/* We can only remove NUM_PAGES_TO_ALLOC at a time. */
465		if (freed_pages >= NUM_PAGES_TO_ALLOC) {
466
467			ttm_pool_update_free_locked(pool, freed_pages);
468			/**
469			 * Because changing page caching is costly
470			 * we unlock the pool to prevent stalling.
471			 */
472			spin_unlock_irqrestore(&pool->lock, irq_flags);
473
474			ttm_dma_pages_put(pool, &d_pages, pages_to_free,
475					  freed_pages);
476
477			INIT_LIST_HEAD(&d_pages);
478
479			if (likely(nr_free != FREE_ALL_PAGES))
480				nr_free -= freed_pages;
481
482			if (NUM_PAGES_TO_ALLOC >= nr_free)
483				npages_to_free = nr_free;
484			else
485				npages_to_free = NUM_PAGES_TO_ALLOC;
486
487			freed_pages = 0;
488
489			/* free all so restart the processing */
490			if (nr_free)
491				goto restart;
492
493			/* Not allowed to fall through or break because
494			 * following context is inside spinlock while we are
495			 * outside here.
496			 */
497			goto out;
498
499		}
500	}
501
502	/* remove range of pages from the pool */
503	if (freed_pages) {
504		ttm_pool_update_free_locked(pool, freed_pages);
505		nr_free -= freed_pages;
506	}
507
508	spin_unlock_irqrestore(&pool->lock, irq_flags);
509
510	if (freed_pages)
511		ttm_dma_pages_put(pool, &d_pages, pages_to_free, freed_pages);
512out:
513	if (pages_to_free != static_buf)
514		kfree(pages_to_free);
515	return nr_free;
516}
517
518static void ttm_dma_free_pool(struct device *dev, enum pool_type type)
519{
520	struct device_pools *p;
521	struct dma_pool *pool;
522
523	if (!dev)
524		return;
525
526	mutex_lock(&_manager->lock);
527	list_for_each_entry_reverse(p, &_manager->pools, pools) {
528		if (p->dev != dev)
529			continue;
530		pool = p->pool;
531		if (pool->type != type)
532			continue;
533
534		list_del(&p->pools);
535		kfree(p);
536		_manager->npools--;
537		break;
538	}
539	list_for_each_entry_reverse(pool, &dev->dma_pools, pools) {
540		if (pool->type != type)
541			continue;
542		/* Takes a spinlock.. */
543		/* OK to use static buffer since global mutex is held. */
544		ttm_dma_page_pool_free(pool, FREE_ALL_PAGES, true);
545		WARN_ON(((pool->npages_in_use + pool->npages_free) != 0));
546		/* This code path is called after _all_ references to the
547		 * struct device has been dropped - so nobody should be
548		 * touching it. In case somebody is trying to _add_ we are
549		 * guarded by the mutex. */
550		list_del(&pool->pools);
551		kfree(pool);
552		break;
553	}
554	mutex_unlock(&_manager->lock);
555}
556
557/*
558 * On free-ing of the 'struct device' this deconstructor is run.
559 * Albeit the pool might have already been freed earlier.
560 */
561static void ttm_dma_pool_release(struct device *dev, void *res)
562{
563	struct dma_pool *pool = *(struct dma_pool **)res;
564
565	if (pool)
566		ttm_dma_free_pool(dev, pool->type);
567}
568
569static int ttm_dma_pool_match(struct device *dev, void *res, void *match_data)
570{
571	return *(struct dma_pool **)res == match_data;
572}
573
574static struct dma_pool *ttm_dma_pool_init(struct device *dev, gfp_t flags,
575					  enum pool_type type)
576{
577	char *n[] = {"wc", "uc", "cached", " dma32", "unknown",};
578	enum pool_type t[] = {IS_WC, IS_UC, IS_CACHED, IS_DMA32, IS_UNDEFINED};
579	struct device_pools *sec_pool = NULL;
580	struct dma_pool *pool = NULL, **ptr;
581	unsigned i;
582	int ret = -ENODEV;
583	char *p;
584
585	if (!dev)
586		return NULL;
587
588	ptr = devres_alloc(ttm_dma_pool_release, sizeof(*ptr), GFP_KERNEL);
589	if (!ptr)
590		return NULL;
591
592	ret = -ENOMEM;
593
594	pool = kmalloc_node(sizeof(struct dma_pool), GFP_KERNEL,
595			    dev_to_node(dev));
596	if (!pool)
597		goto err_mem;
598
599	sec_pool = kmalloc_node(sizeof(struct device_pools), GFP_KERNEL,
600				dev_to_node(dev));
601	if (!sec_pool)
602		goto err_mem;
603
604	INIT_LIST_HEAD(&sec_pool->pools);
605	sec_pool->dev = dev;
606	sec_pool->pool =  pool;
607
608	INIT_LIST_HEAD(&pool->free_list);
609	INIT_LIST_HEAD(&pool->inuse_list);
610	INIT_LIST_HEAD(&pool->pools);
611	spin_lock_init(&pool->lock);
612	pool->dev = dev;
613	pool->npages_free = pool->npages_in_use = 0;
614	pool->nfrees = 0;
615	pool->gfp_flags = flags;
616	pool->size = PAGE_SIZE;
617	pool->type = type;
618	pool->nrefills = 0;
619	p = pool->name;
620	for (i = 0; i < 5; i++) {
621		if (type & t[i]) {
622			p += snprintf(p, sizeof(pool->name) - (p - pool->name),
623				      "%s", n[i]);
624		}
625	}
626	*p = 0;
627	/* We copy the name for pr_ calls b/c when dma_pool_destroy is called
628	 * - the kobj->name has already been deallocated.*/
629	snprintf(pool->dev_name, sizeof(pool->dev_name), "%s %s",
630		 dev_driver_string(dev), dev_name(dev));
631	mutex_lock(&_manager->lock);
632	/* You can get the dma_pool from either the global: */
633	list_add(&sec_pool->pools, &_manager->pools);
634	_manager->npools++;
635	/* or from 'struct device': */
636	list_add(&pool->pools, &dev->dma_pools);
637	mutex_unlock(&_manager->lock);
638
639	*ptr = pool;
640	devres_add(dev, ptr);
641
642	return pool;
643err_mem:
644	devres_free(ptr);
645	kfree(sec_pool);
646	kfree(pool);
647	return ERR_PTR(ret);
648}
649
650static struct dma_pool *ttm_dma_find_pool(struct device *dev,
651					  enum pool_type type)
652{
653	struct dma_pool *pool, *tmp, *found = NULL;
654
655	if (type == IS_UNDEFINED)
656		return found;
657
658	/* NB: We iterate on the 'struct dev' which has no spinlock, but
659	 * it does have a kref which we have taken. The kref is taken during
660	 * graphic driver loading - in the drm_pci_init it calls either
661	 * pci_dev_get or pci_register_driver which both end up taking a kref
662	 * on 'struct device'.
663	 *
664	 * On teardown, the graphic drivers end up quiescing the TTM (put_pages)
665	 * and calls the dev_res deconstructors: ttm_dma_pool_release. The nice
666	 * thing is at that point of time there are no pages associated with the
667	 * driver so this function will not be called.
668	 */
669	list_for_each_entry_safe(pool, tmp, &dev->dma_pools, pools) {
670		if (pool->type != type)
671			continue;
672		found = pool;
673		break;
674	}
675	return found;
676}
677
678/*
679 * Free pages the pages that failed to change the caching state. If there
680 * are pages that have changed their caching state already put them to the
681 * pool.
682 */
683static void ttm_dma_handle_caching_state_failure(struct dma_pool *pool,
684						 struct list_head *d_pages,
685						 struct page **failed_pages,
686						 unsigned cpages)
687{
688	struct dma_page *d_page, *tmp;
689	struct page *p;
690	unsigned i = 0;
691
692	p = failed_pages[0];
693	if (!p)
694		return;
695	/* Find the failed page. */
696	list_for_each_entry_safe(d_page, tmp, d_pages, page_list) {
697		if (d_page->p != p)
698			continue;
699		/* .. and then progress over the full list. */
700		list_del(&d_page->page_list);
701		__ttm_dma_free_page(pool, d_page);
702		if (++i < cpages)
703			p = failed_pages[i];
704		else
705			break;
706	}
707
708}
709
710/*
711 * Allocate 'count' pages, and put 'need' number of them on the
712 * 'pages' and as well on the 'dma_address' starting at 'dma_offset' offset.
713 * The full list of pages should also be on 'd_pages'.
714 * We return zero for success, and negative numbers as errors.
715 */
716static int ttm_dma_pool_alloc_new_pages(struct dma_pool *pool,
717					struct list_head *d_pages,
718					unsigned count)
719{
720	struct page **caching_array;
721	struct dma_page *dma_p;
722	struct page *p;
723	int r = 0;
724	unsigned i, cpages;
725	unsigned max_cpages = min(count,
726			(unsigned)(PAGE_SIZE/sizeof(struct page *)));
727
728	/* allocate array for page caching change */
729	caching_array = kmalloc(max_cpages*sizeof(struct page *), GFP_KERNEL);
730
731	if (!caching_array) {
732		pr_err("%s: Unable to allocate table for new pages\n",
733		       pool->dev_name);
734		return -ENOMEM;
735	}
736
737	if (count > 1) {
738		pr_debug("%s: (%s:%d) Getting %d pages\n",
739			 pool->dev_name, pool->name, current->pid, count);
740	}
741
742	for (i = 0, cpages = 0; i < count; ++i) {
743		dma_p = __ttm_dma_alloc_page(pool);
744		if (!dma_p) {
745			pr_err("%s: Unable to get page %u\n",
746			       pool->dev_name, i);
747
748			/* store already allocated pages in the pool after
749			 * setting the caching state */
750			if (cpages) {
751				r = ttm_set_pages_caching(pool, caching_array,
752							  cpages);
753				if (r)
754					ttm_dma_handle_caching_state_failure(
755						pool, d_pages, caching_array,
756						cpages);
757			}
758			r = -ENOMEM;
759			goto out;
760		}
761		p = dma_p->p;
762#ifdef CONFIG_HIGHMEM
763		/* gfp flags of highmem page should never be dma32 so we
764		 * we should be fine in such case
765		 */
766		if (!PageHighMem(p))
767#endif
768		{
769			caching_array[cpages++] = p;
770			if (cpages == max_cpages) {
771				/* Note: Cannot hold the spinlock */
772				r = ttm_set_pages_caching(pool, caching_array,
773						 cpages);
774				if (r) {
775					ttm_dma_handle_caching_state_failure(
776						pool, d_pages, caching_array,
777						cpages);
778					goto out;
779				}
780				cpages = 0;
781			}
782		}
783		list_add(&dma_p->page_list, d_pages);
784	}
785
786	if (cpages) {
787		r = ttm_set_pages_caching(pool, caching_array, cpages);
788		if (r)
789			ttm_dma_handle_caching_state_failure(pool, d_pages,
790					caching_array, cpages);
791	}
792out:
793	kfree(caching_array);
794	return r;
795}
796
797/*
798 * @return count of pages still required to fulfill the request.
799 */
800static int ttm_dma_page_pool_fill_locked(struct dma_pool *pool,
801					 unsigned long *irq_flags)
802{
803	unsigned count = _manager->options.small;
804	int r = pool->npages_free;
805
806	if (count > pool->npages_free) {
807		struct list_head d_pages;
808
809		INIT_LIST_HEAD(&d_pages);
810
811		spin_unlock_irqrestore(&pool->lock, *irq_flags);
812
813		/* Returns how many more are neccessary to fulfill the
814		 * request. */
815		r = ttm_dma_pool_alloc_new_pages(pool, &d_pages, count);
816
817		spin_lock_irqsave(&pool->lock, *irq_flags);
818		if (!r) {
819			/* Add the fresh to the end.. */
820			list_splice(&d_pages, &pool->free_list);
821			++pool->nrefills;
822			pool->npages_free += count;
823			r = count;
824		} else {
825			struct dma_page *d_page;
826			unsigned cpages = 0;
827
828			pr_err("%s: Failed to fill %s pool (r:%d)!\n",
829			       pool->dev_name, pool->name, r);
830
831			list_for_each_entry(d_page, &d_pages, page_list) {
832				cpages++;
833			}
834			list_splice_tail(&d_pages, &pool->free_list);
835			pool->npages_free += cpages;
836			r = cpages;
837		}
838	}
839	return r;
840}
841
842/*
843 * @return count of pages still required to fulfill the request.
844 * The populate list is actually a stack (not that is matters as TTM
845 * allocates one page at a time.
846 */
847static int ttm_dma_pool_get_pages(struct dma_pool *pool,
848				  struct ttm_dma_tt *ttm_dma,
849				  unsigned index)
850{
851	struct dma_page *d_page;
852	struct ttm_tt *ttm = &ttm_dma->ttm;
853	unsigned long irq_flags;
854	int count, r = -ENOMEM;
855
856	spin_lock_irqsave(&pool->lock, irq_flags);
857	count = ttm_dma_page_pool_fill_locked(pool, &irq_flags);
858	if (count) {
859		d_page = list_first_entry(&pool->free_list, struct dma_page, page_list);
860		ttm->pages[index] = d_page->p;
861		ttm_dma->cpu_address[index] = d_page->vaddr;
862		ttm_dma->dma_address[index] = d_page->dma;
863		list_move_tail(&d_page->page_list, &ttm_dma->pages_list);
864		r = 0;
865		pool->npages_in_use += 1;
866		pool->npages_free -= 1;
867	}
868	spin_unlock_irqrestore(&pool->lock, irq_flags);
869	return r;
870}
871
872/*
873 * On success pages list will hold count number of correctly
874 * cached pages. On failure will hold the negative return value (-ENOMEM, etc).
875 */
876int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev)
877{
878	struct ttm_tt *ttm = &ttm_dma->ttm;
879	struct ttm_mem_global *mem_glob = ttm->glob->mem_glob;
880	struct dma_pool *pool;
881	enum pool_type type;
882	unsigned i;
883	gfp_t gfp_flags;
884	int ret;
885
886	if (ttm->state != tt_unpopulated)
887		return 0;
888
889	type = ttm_to_type(ttm->page_flags, ttm->caching_state);
890	if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
891		gfp_flags = GFP_USER | GFP_DMA32;
892	else
893		gfp_flags = GFP_HIGHUSER;
894	if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
895		gfp_flags |= __GFP_ZERO;
896
897	pool = ttm_dma_find_pool(dev, type);
898	if (!pool) {
899		pool = ttm_dma_pool_init(dev, gfp_flags, type);
900		if (IS_ERR_OR_NULL(pool)) {
901			return -ENOMEM;
902		}
903	}
904
905	INIT_LIST_HEAD(&ttm_dma->pages_list);
906	for (i = 0; i < ttm->num_pages; ++i) {
907		ret = ttm_dma_pool_get_pages(pool, ttm_dma, i);
908		if (ret != 0) {
909			ttm_dma_unpopulate(ttm_dma, dev);
910			return -ENOMEM;
911		}
912
913		ret = ttm_mem_global_alloc_page(mem_glob, ttm->pages[i],
914						false, false);
915		if (unlikely(ret != 0)) {
916			ttm_dma_unpopulate(ttm_dma, dev);
917			return -ENOMEM;
918		}
919	}
920
921	if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
922		ret = ttm_tt_swapin(ttm);
923		if (unlikely(ret != 0)) {
924			ttm_dma_unpopulate(ttm_dma, dev);
925			return ret;
926		}
927	}
928
929	ttm->state = tt_unbound;
930	return 0;
931}
932EXPORT_SYMBOL_GPL(ttm_dma_populate);
933
934/* Put all pages in pages list to correct pool to wait for reuse */
935void ttm_dma_unpopulate(struct ttm_dma_tt *ttm_dma, struct device *dev)
936{
937	struct ttm_tt *ttm = &ttm_dma->ttm;
938	struct dma_pool *pool;
939	struct dma_page *d_page, *next;
940	enum pool_type type;
941	bool is_cached = false;
942	unsigned count = 0, i, npages = 0;
943	unsigned long irq_flags;
944
945	type = ttm_to_type(ttm->page_flags, ttm->caching_state);
946	pool = ttm_dma_find_pool(dev, type);
947	if (!pool)
948		return;
949
950	is_cached = (ttm_dma_find_pool(pool->dev,
951		     ttm_to_type(ttm->page_flags, tt_cached)) == pool);
952
953	/* make sure pages array match list and count number of pages */
954	list_for_each_entry(d_page, &ttm_dma->pages_list, page_list) {
955		ttm->pages[count] = d_page->p;
956		count++;
957	}
958
959	spin_lock_irqsave(&pool->lock, irq_flags);
960	pool->npages_in_use -= count;
961	if (is_cached) {
962		pool->nfrees += count;
963	} else {
964		pool->npages_free += count;
965		list_splice(&ttm_dma->pages_list, &pool->free_list);
966		/*
967		 * Wait to have at at least NUM_PAGES_TO_ALLOC number of pages
968		 * to free in order to minimize calls to set_memory_wb().
969		 */
970		if (pool->npages_free >= (_manager->options.max_size +
971					  NUM_PAGES_TO_ALLOC))
972			npages = pool->npages_free - _manager->options.max_size;
973	}
974	spin_unlock_irqrestore(&pool->lock, irq_flags);
975
976	if (is_cached) {
977		list_for_each_entry_safe(d_page, next, &ttm_dma->pages_list, page_list) {
978			ttm_mem_global_free_page(ttm->glob->mem_glob,
979						 d_page->p);
980			ttm_dma_page_put(pool, d_page);
981		}
982	} else {
983		for (i = 0; i < count; i++) {
984			ttm_mem_global_free_page(ttm->glob->mem_glob,
985						 ttm->pages[i]);
986		}
987	}
988
989	INIT_LIST_HEAD(&ttm_dma->pages_list);
990	for (i = 0; i < ttm->num_pages; i++) {
991		ttm->pages[i] = NULL;
992		ttm_dma->cpu_address[i] = 0;
993		ttm_dma->dma_address[i] = 0;
994	}
995
996	/* shrink pool if necessary (only on !is_cached pools)*/
997	if (npages)
998		ttm_dma_page_pool_free(pool, npages, false);
999	ttm->state = tt_unpopulated;
1000}
1001EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
1002
1003/**
1004 * Callback for mm to request pool to reduce number of page held.
1005 *
1006 * XXX: (dchinner) Deadlock warning!
1007 *
1008 * I'm getting sadder as I hear more pathetical whimpers about needing per-pool
1009 * shrinkers
1010 */
1011static unsigned long
1012ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1013{
1014	static unsigned start_pool;
1015	unsigned idx = 0;
1016	unsigned pool_offset;
1017	unsigned shrink_pages = sc->nr_to_scan;
1018	struct device_pools *p;
1019	unsigned long freed = 0;
1020
1021	if (list_empty(&_manager->pools))
1022		return SHRINK_STOP;
1023
1024	if (!mutex_trylock(&_manager->lock))
1025		return SHRINK_STOP;
1026	if (!_manager->npools)
1027		goto out;
1028	pool_offset = ++start_pool % _manager->npools;
1029	list_for_each_entry(p, &_manager->pools, pools) {
1030		unsigned nr_free;
1031
1032		if (!p->dev)
1033			continue;
1034		if (shrink_pages == 0)
1035			break;
1036		/* Do it in round-robin fashion. */
1037		if (++idx < pool_offset)
1038			continue;
1039		nr_free = shrink_pages;
1040		/* OK to use static buffer since global mutex is held. */
1041		shrink_pages = ttm_dma_page_pool_free(p->pool, nr_free, true);
1042		freed += nr_free - shrink_pages;
1043
1044		pr_debug("%s: (%s:%d) Asked to shrink %d, have %d more to go\n",
1045			 p->pool->dev_name, p->pool->name, current->pid,
1046			 nr_free, shrink_pages);
1047	}
1048out:
1049	mutex_unlock(&_manager->lock);
1050	return freed;
1051}
1052
1053static unsigned long
1054ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1055{
1056	struct device_pools *p;
1057	unsigned long count = 0;
1058
1059	if (!mutex_trylock(&_manager->lock))
1060		return 0;
1061	list_for_each_entry(p, &_manager->pools, pools)
1062		count += p->pool->npages_free;
1063	mutex_unlock(&_manager->lock);
1064	return count;
1065}
1066
1067static void ttm_dma_pool_mm_shrink_init(struct ttm_pool_manager *manager)
1068{
1069	manager->mm_shrink.count_objects = ttm_dma_pool_shrink_count;
1070	manager->mm_shrink.scan_objects = &ttm_dma_pool_shrink_scan;
1071	manager->mm_shrink.seeks = 1;
1072	register_shrinker(&manager->mm_shrink);
1073}
1074
1075static void ttm_dma_pool_mm_shrink_fini(struct ttm_pool_manager *manager)
1076{
1077	unregister_shrinker(&manager->mm_shrink);
1078}
1079
1080int ttm_dma_page_alloc_init(struct ttm_mem_global *glob, unsigned max_pages)
1081{
1082	int ret = -ENOMEM;
1083
1084	WARN_ON(_manager);
1085
1086	pr_info("Initializing DMA pool allocator\n");
1087
1088	_manager = kzalloc(sizeof(*_manager), GFP_KERNEL);
1089	if (!_manager)
1090		goto err;
1091
1092	mutex_init(&_manager->lock);
1093	INIT_LIST_HEAD(&_manager->pools);
1094
1095	_manager->options.max_size = max_pages;
1096	_manager->options.small = SMALL_ALLOCATION;
1097	_manager->options.alloc_size = NUM_PAGES_TO_ALLOC;
1098
1099	/* This takes care of auto-freeing the _manager */
1100	ret = kobject_init_and_add(&_manager->kobj, &ttm_pool_kobj_type,
1101				   &glob->kobj, "dma_pool");
1102	if (unlikely(ret != 0)) {
1103		kobject_put(&_manager->kobj);
1104		goto err;
1105	}
1106	ttm_dma_pool_mm_shrink_init(_manager);
1107	return 0;
1108err:
1109	return ret;
1110}
1111
1112void ttm_dma_page_alloc_fini(void)
1113{
1114	struct device_pools *p, *t;
1115
1116	pr_info("Finalizing DMA pool allocator\n");
1117	ttm_dma_pool_mm_shrink_fini(_manager);
1118
1119	list_for_each_entry_safe_reverse(p, t, &_manager->pools, pools) {
1120		dev_dbg(p->dev, "(%s:%d) Freeing.\n", p->pool->name,
1121			current->pid);
1122		WARN_ON(devres_destroy(p->dev, ttm_dma_pool_release,
1123			ttm_dma_pool_match, p->pool));
1124		ttm_dma_free_pool(p->dev, p->pool->type);
1125	}
1126	kobject_put(&_manager->kobj);
1127	_manager = NULL;
1128}
1129
1130int ttm_dma_page_alloc_debugfs(struct seq_file *m, void *data)
1131{
1132	struct device_pools *p;
1133	struct dma_pool *pool = NULL;
1134	char *h[] = {"pool", "refills", "pages freed", "inuse", "available",
1135		     "name", "virt", "busaddr"};
1136
1137	if (!_manager) {
1138		seq_printf(m, "No pool allocator running.\n");
1139		return 0;
1140	}
1141	seq_printf(m, "%13s %12s %13s %8s %8s %8s\n",
1142		   h[0], h[1], h[2], h[3], h[4], h[5]);
1143	mutex_lock(&_manager->lock);
1144	list_for_each_entry(p, &_manager->pools, pools) {
1145		struct device *dev = p->dev;
1146		if (!dev)
1147			continue;
1148		pool = p->pool;
1149		seq_printf(m, "%13s %12ld %13ld %8d %8d %8s\n",
1150				pool->name, pool->nrefills,
1151				pool->nfrees, pool->npages_in_use,
1152				pool->npages_free,
1153				pool->dev_name);
1154	}
1155	mutex_unlock(&_manager->lock);
1156	return 0;
1157}
1158EXPORT_SYMBOL_GPL(ttm_dma_page_alloc_debugfs);
1159
1160#endif
1161