projects/linux·blob v4.19 · refs

1773 lines · 45.5 KiB

1/*
2 * Copyright (C) 2008 Advanced Micro Devices, Inc.
3 *
4 * Author: Joerg Roedel <[email protected]>
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 */
19
20#include <linux/sched/task_stack.h>
21#include <linux/scatterlist.h>
22#include <linux/dma-mapping.h>
23#include <linux/sched/task.h>
24#include <linux/stacktrace.h>
25#include <linux/dma-debug.h>
26#include <linux/spinlock.h>
27#include <linux/vmalloc.h>
28#include <linux/debugfs.h>
29#include <linux/uaccess.h>
30#include <linux/export.h>
31#include <linux/device.h>
32#include <linux/types.h>
33#include <linux/sched.h>
34#include <linux/ctype.h>
35#include <linux/list.h>
36#include <linux/slab.h>
37
38#include <asm/sections.h>
39
40#define HASH_SIZE 1024ULL
41#define HASH_FN_SHIFT 13
42#define HASH_FN_MASK (HASH_SIZE - 1)
43
44/* allow architectures to override this if absolutely required */
45#ifndef PREALLOC_DMA_DEBUG_ENTRIES
46#define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
47#endif
48
49enum {
50 dma_debug_single,
51 dma_debug_page,
52 dma_debug_sg,
53 dma_debug_coherent,
54 dma_debug_resource,
55};
56
57enum map_err_types {
58 MAP_ERR_CHECK_NOT_APPLICABLE,
59 MAP_ERR_NOT_CHECKED,
60 MAP_ERR_CHECKED,
61};
62
63#define DMA_DEBUG_STACKTRACE_ENTRIES 5
64
65/**
66 * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
67 * @list: node on pre-allocated free_entries list
68 * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
69 * @type: single, page, sg, coherent
70 * @pfn: page frame of the start address
71 * @offset: offset of mapping relative to pfn
72 * @size: length of the mapping
73 * @direction: enum dma_data_direction
74 * @sg_call_ents: 'nents' from dma_map_sg
75 * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
76 * @map_err_type: track whether dma_mapping_error() was checked
77 * @stacktrace: support backtraces when a violation is detected
78 */
79struct dma_debug_entry {
80 struct list_head list;
81 struct device *dev;
82 int type;
83 unsigned long pfn;
84 size_t offset;
85 u64 dev_addr;
86 u64 size;
87 int direction;
88 int sg_call_ents;
89 int sg_mapped_ents;
90 enum map_err_types map_err_type;
91#ifdef CONFIG_STACKTRACE
92 struct stack_trace stacktrace;
93 unsigned long st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
94#endif
95};
96
97typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
98
99struct hash_bucket {
100 struct list_head list;
101 spinlock_t lock;
102} ____cacheline_aligned_in_smp;
103
104/* Hash list to save the allocated dma addresses */
105static struct hash_bucket dma_entry_hash[HASH_SIZE];
106/* List of pre-allocated dma_debug_entry's */
107static LIST_HEAD(free_entries);
108/* Lock for the list above */
109static DEFINE_SPINLOCK(free_entries_lock);
110
111/* Global disable flag - will be set in case of an error */
112static bool global_disable __read_mostly;
113
114/* Early initialization disable flag, set at the end of dma_debug_init */
115static bool dma_debug_initialized __read_mostly;
116
117static inline bool dma_debug_disabled(void)
118{
119 return global_disable || !dma_debug_initialized;
120}
121
122/* Global error count */
123static u32 error_count;
124
125/* Global error show enable*/
126static u32 show_all_errors __read_mostly;
127/* Number of errors to show */
128static u32 show_num_errors = 1;
129
130static u32 num_free_entries;
131static u32 min_free_entries;
132static u32 nr_total_entries;
133
134/* number of preallocated entries requested by kernel cmdline */
135static u32 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
136
137/* debugfs dentry's for the stuff above */
138static struct dentry *dma_debug_dent __read_mostly;
139static struct dentry *global_disable_dent __read_mostly;
140static struct dentry *error_count_dent __read_mostly;
141static struct dentry *show_all_errors_dent __read_mostly;
142static struct dentry *show_num_errors_dent __read_mostly;
143static struct dentry *num_free_entries_dent __read_mostly;
144static struct dentry *min_free_entries_dent __read_mostly;
145static struct dentry *filter_dent __read_mostly;
146
147/* per-driver filter related state */
148
149#define NAME_MAX_LEN 64
150
151static char current_driver_name[NAME_MAX_LEN] __read_mostly;
152static struct device_driver *current_driver __read_mostly;
153
154static DEFINE_RWLOCK(driver_name_lock);
155
156static const char *const maperr2str[] = {
157 [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
158 [MAP_ERR_NOT_CHECKED] = "dma map error not checked",
159 [MAP_ERR_CHECKED] = "dma map error checked",
160};
161
162static const char *type2name[5] = { "single", "page",
163 "scather-gather", "coherent",
164 "resource" };
165
166static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
167 "DMA_FROM_DEVICE", "DMA_NONE" };
168
169/*
170 * The access to some variables in this macro is racy. We can't use atomic_t
171 * here because all these variables are exported to debugfs. Some of them even
172 * writeable. This is also the reason why a lock won't help much. But anyway,
173 * the races are no big deal. Here is why:
174 *
175 * error_count: the addition is racy, but the worst thing that can happen is
176 * that we don't count some errors
177 * show_num_errors: the subtraction is racy. Also no big deal because in
178 * worst case this will result in one warning more in the
179 * system log than the user configured. This variable is
180 * writeable via debugfs.
181 */
182static inline void dump_entry_trace(struct dma_debug_entry *entry)
183{
184#ifdef CONFIG_STACKTRACE
185 if (entry) {
186 pr_warning("Mapped at:\n");
187 print_stack_trace(&entry->stacktrace, 0);
188 }
189#endif
190}
191
192static bool driver_filter(struct device *dev)
193{
194 struct device_driver *drv;
195 unsigned long flags;
196 bool ret;
197
198 /* driver filter off */
199 if (likely(!current_driver_name[0]))
200 return true;
201
202 /* driver filter on and initialized */
203 if (current_driver && dev && dev->driver == current_driver)
204 return true;
205
206 /* driver filter on, but we can't filter on a NULL device... */
207 if (!dev)
208 return false;
209
210 if (current_driver || !current_driver_name[0])
211 return false;
212
213 /* driver filter on but not yet initialized */
214 drv = dev->driver;
215 if (!drv)
216 return false;
217
218 /* lock to protect against change of current_driver_name */
219 read_lock_irqsave(&driver_name_lock, flags);
220
221 ret = false;
222 if (drv->name &&
223 strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
224 current_driver = drv;
225 ret = true;
226 }
227
228 read_unlock_irqrestore(&driver_name_lock, flags);
229
230 return ret;
231}
232
233#define err_printk(dev, entry, format, arg...) do { \
234 error_count += 1; \
235 if (driver_filter(dev) && \
236 (show_all_errors || show_num_errors > 0)) { \
237 WARN(1, "%s %s: " format, \
238 dev ? dev_driver_string(dev) : "NULL", \
239 dev ? dev_name(dev) : "NULL", ## arg); \
240 dump_entry_trace(entry); \
241 } \
242 if (!show_all_errors && show_num_errors > 0) \
243 show_num_errors -= 1; \
244 } while (0);
245
246/*
247 * Hash related functions
248 *
249 * Every DMA-API request is saved into a struct dma_debug_entry. To
250 * have quick access to these structs they are stored into a hash.
251 */
252static int hash_fn(struct dma_debug_entry *entry)
253{
254 /*
255 * Hash function is based on the dma address.
256 * We use bits 20-27 here as the index into the hash
257 */
258 return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
259}
260
261/*
262 * Request exclusive access to a hash bucket for a given dma_debug_entry.
263 */
264static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
265 unsigned long *flags)
266 __acquires(&dma_entry_hash[idx].lock)
267{
268 int idx = hash_fn(entry);
269 unsigned long __flags;
270
271 spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
272 *flags = __flags;
273 return &dma_entry_hash[idx];
274}
275
276/*
277 * Give up exclusive access to the hash bucket
278 */
279static void put_hash_bucket(struct hash_bucket *bucket,
280 unsigned long *flags)
281 __releases(&bucket->lock)
282{
283 unsigned long __flags = *flags;
284
285 spin_unlock_irqrestore(&bucket->lock, __flags);
286}
287
288static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
289{
290 return ((a->dev_addr == b->dev_addr) &&
291 (a->dev == b->dev)) ? true : false;
292}
293
294static bool containing_match(struct dma_debug_entry *a,
295 struct dma_debug_entry *b)
296{
297 if (a->dev != b->dev)
298 return false;
299
300 if ((b->dev_addr <= a->dev_addr) &&
301 ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
302 return true;
303
304 return false;
305}
306
307/*
308 * Search a given entry in the hash bucket list
309 */
310static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
311 struct dma_debug_entry *ref,
312 match_fn match)
313{
314 struct dma_debug_entry *entry, *ret = NULL;
315 int matches = 0, match_lvl, last_lvl = -1;
316
317 list_for_each_entry(entry, &bucket->list, list) {
318 if (!match(ref, entry))
319 continue;
320
321 /*
322 * Some drivers map the same physical address multiple
323 * times. Without a hardware IOMMU this results in the
324 * same device addresses being put into the dma-debug
325 * hash multiple times too. This can result in false
326 * positives being reported. Therefore we implement a
327 * best-fit algorithm here which returns the entry from
328 * the hash which fits best to the reference value
329 * instead of the first-fit.
330 */
331 matches += 1;
332 match_lvl = 0;
333 entry->size == ref->size ? ++match_lvl : 0;
334 entry->type == ref->type ? ++match_lvl : 0;
335 entry->direction == ref->direction ? ++match_lvl : 0;
336 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
337
338 if (match_lvl == 4) {
339 /* perfect-fit - return the result */
340 return entry;
341 } else if (match_lvl > last_lvl) {
342 /*
343 * We found an entry that fits better then the
344 * previous one or it is the 1st match.
345 */
346 last_lvl = match_lvl;
347 ret = entry;
348 }
349 }
350
351 /*
352 * If we have multiple matches but no perfect-fit, just return
353 * NULL.
354 */
355 ret = (matches == 1) ? ret : NULL;
356
357 return ret;
358}
359
360static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
361 struct dma_debug_entry *ref)
362{
363 return __hash_bucket_find(bucket, ref, exact_match);
364}
365
366static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
367 struct dma_debug_entry *ref,
368 unsigned long *flags)
369{
370
371 unsigned int max_range = dma_get_max_seg_size(ref->dev);
372 struct dma_debug_entry *entry, index = *ref;
373 unsigned int range = 0;
374
375 while (range <= max_range) {
376 entry = __hash_bucket_find(*bucket, ref, containing_match);
377
378 if (entry)
379 return entry;
380
381 /*
382 * Nothing found, go back a hash bucket
383 */
384 put_hash_bucket(*bucket, flags);
385 range += (1 << HASH_FN_SHIFT);
386 index.dev_addr -= (1 << HASH_FN_SHIFT);
387 *bucket = get_hash_bucket(&index, flags);
388 }
389
390 return NULL;
391}
392
393/*
394 * Add an entry to a hash bucket
395 */
396static void hash_bucket_add(struct hash_bucket *bucket,
397 struct dma_debug_entry *entry)
398{
399 list_add_tail(&entry->list, &bucket->list);
400}
401
402/*
403 * Remove entry from a hash bucket list
404 */
405static void hash_bucket_del(struct dma_debug_entry *entry)
406{
407 list_del(&entry->list);
408}
409
410static unsigned long long phys_addr(struct dma_debug_entry *entry)
411{
412 if (entry->type == dma_debug_resource)
413 return __pfn_to_phys(entry->pfn) + entry->offset;
414
415 return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
416}
417
418/*
419 * Dump mapping entries for debugging purposes
420 */
421void debug_dma_dump_mappings(struct device *dev)
422{
423 int idx;
424
425 for (idx = 0; idx < HASH_SIZE; idx++) {
426 struct hash_bucket *bucket = &dma_entry_hash[idx];
427 struct dma_debug_entry *entry;
428 unsigned long flags;
429
430 spin_lock_irqsave(&bucket->lock, flags);
431
432 list_for_each_entry(entry, &bucket->list, list) {
433 if (!dev || dev == entry->dev) {
434 dev_info(entry->dev,
435 "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
436 type2name[entry->type], idx,
437 phys_addr(entry), entry->pfn,
438 entry->dev_addr, entry->size,
439 dir2name[entry->direction],
440 maperr2str[entry->map_err_type]);
441 }
442 }
443
444 spin_unlock_irqrestore(&bucket->lock, flags);
445 }
446}
447
448/*
449 * For each mapping (initial cacheline in the case of
450 * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
451 * scatterlist, or the cacheline specified in dma_map_single) insert
452 * into this tree using the cacheline as the key. At
453 * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry. If
454 * the entry already exists at insertion time add a tag as a reference
455 * count for the overlapping mappings. For now, the overlap tracking
456 * just ensures that 'unmaps' balance 'maps' before marking the
457 * cacheline idle, but we should also be flagging overlaps as an API
458 * violation.
459 *
460 * Memory usage is mostly constrained by the maximum number of available
461 * dma-debug entries in that we need a free dma_debug_entry before
462 * inserting into the tree. In the case of dma_map_page and
463 * dma_alloc_coherent there is only one dma_debug_entry and one
464 * dma_active_cacheline entry to track per event. dma_map_sg(), on the
465 * other hand, consumes a single dma_debug_entry, but inserts 'nents'
466 * entries into the tree.
467 *
468 * At any time debug_dma_assert_idle() can be called to trigger a
469 * warning if any cachelines in the given page are in the active set.
470 */
471static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
472static DEFINE_SPINLOCK(radix_lock);
473#define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
474#define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
475#define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
476
477static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
478{
479 return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
480 (entry->offset >> L1_CACHE_SHIFT);
481}
482
483static int active_cacheline_read_overlap(phys_addr_t cln)
484{
485 int overlap = 0, i;
486
487 for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
488 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
489 overlap |= 1 << i;
490 return overlap;
491}
492
493static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
494{
495 int i;
496
497 if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
498 return overlap;
499
500 for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
501 if (overlap & 1 << i)
502 radix_tree_tag_set(&dma_active_cacheline, cln, i);
503 else
504 radix_tree_tag_clear(&dma_active_cacheline, cln, i);
505
506 return overlap;
507}
508
509static void active_cacheline_inc_overlap(phys_addr_t cln)
510{
511 int overlap = active_cacheline_read_overlap(cln);
512
513 overlap = active_cacheline_set_overlap(cln, ++overlap);
514
515 /* If we overflowed the overlap counter then we're potentially
516 * leaking dma-mappings. Otherwise, if maps and unmaps are
517 * balanced then this overflow may cause false negatives in
518 * debug_dma_assert_idle() as the cacheline may be marked idle
519 * prematurely.
520 */
521 WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
522 "DMA-API: exceeded %d overlapping mappings of cacheline %pa\n",
523 ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
524}
525
526static int active_cacheline_dec_overlap(phys_addr_t cln)
527{
528 int overlap = active_cacheline_read_overlap(cln);
529
530 return active_cacheline_set_overlap(cln, --overlap);
531}
532
533static int active_cacheline_insert(struct dma_debug_entry *entry)
534{
535 phys_addr_t cln = to_cacheline_number(entry);
536 unsigned long flags;
537 int rc;
538
539 /* If the device is not writing memory then we don't have any
540 * concerns about the cpu consuming stale data. This mitigates
541 * legitimate usages of overlapping mappings.
542 */
543 if (entry->direction == DMA_TO_DEVICE)
544 return 0;
545
546 spin_lock_irqsave(&radix_lock, flags);
547 rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
548 if (rc == -EEXIST)
549 active_cacheline_inc_overlap(cln);
550 spin_unlock_irqrestore(&radix_lock, flags);
551
552 return rc;
553}
554
555static void active_cacheline_remove(struct dma_debug_entry *entry)
556{
557 phys_addr_t cln = to_cacheline_number(entry);
558 unsigned long flags;
559
560 /* ...mirror the insert case */
561 if (entry->direction == DMA_TO_DEVICE)
562 return;
563
564 spin_lock_irqsave(&radix_lock, flags);
565 /* since we are counting overlaps the final put of the
566 * cacheline will occur when the overlap count is 0.
567 * active_cacheline_dec_overlap() returns -1 in that case
568 */
569 if (active_cacheline_dec_overlap(cln) < 0)
570 radix_tree_delete(&dma_active_cacheline, cln);
571 spin_unlock_irqrestore(&radix_lock, flags);
572}
573
574/**
575 * debug_dma_assert_idle() - assert that a page is not undergoing dma
576 * @page: page to lookup in the dma_active_cacheline tree
577 *
578 * Place a call to this routine in cases where the cpu touching the page
579 * before the dma completes (page is dma_unmapped) will lead to data
580 * corruption.
581 */
582void debug_dma_assert_idle(struct page *page)
583{
584 static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
585 struct dma_debug_entry *entry = NULL;
586 void **results = (void **) &ents;
587 unsigned int nents, i;
588 unsigned long flags;
589 phys_addr_t cln;
590
591 if (dma_debug_disabled())
592 return;
593
594 if (!page)
595 return;
596
597 cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
598 spin_lock_irqsave(&radix_lock, flags);
599 nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
600 CACHELINES_PER_PAGE);
601 for (i = 0; i < nents; i++) {
602 phys_addr_t ent_cln = to_cacheline_number(ents[i]);
603
604 if (ent_cln == cln) {
605 entry = ents[i];
606 break;
607 } else if (ent_cln >= cln + CACHELINES_PER_PAGE)
608 break;
609 }
610 spin_unlock_irqrestore(&radix_lock, flags);
611
612 if (!entry)
613 return;
614
615 cln = to_cacheline_number(entry);
616 err_printk(entry->dev, entry,
617 "DMA-API: cpu touching an active dma mapped cacheline [cln=%pa]\n",
618 &cln);
619}
620
621/*
622 * Wrapper function for adding an entry to the hash.
623 * This function takes care of locking itself.
624 */
625static void add_dma_entry(struct dma_debug_entry *entry)
626{
627 struct hash_bucket *bucket;
628 unsigned long flags;
629 int rc;
630
631 bucket = get_hash_bucket(entry, &flags);
632 hash_bucket_add(bucket, entry);
633 put_hash_bucket(bucket, &flags);
634
635 rc = active_cacheline_insert(entry);
636 if (rc == -ENOMEM) {
637 pr_err("DMA-API: cacheline tracking ENOMEM, dma-debug disabled\n");
638 global_disable = true;
639 }
640
641 /* TODO: report -EEXIST errors here as overlapping mappings are
642 * not supported by the DMA API
643 */
644}
645
646static struct dma_debug_entry *__dma_entry_alloc(void)
647{
648 struct dma_debug_entry *entry;
649
650 entry = list_entry(free_entries.next, struct dma_debug_entry, list);
651 list_del(&entry->list);
652 memset(entry, 0, sizeof(*entry));
653
654 num_free_entries -= 1;
655 if (num_free_entries < min_free_entries)
656 min_free_entries = num_free_entries;
657
658 return entry;
659}
660
661/* struct dma_entry allocator
662 *
663 * The next two functions implement the allocator for
664 * struct dma_debug_entries.
665 */
666static struct dma_debug_entry *dma_entry_alloc(void)
667{
668 struct dma_debug_entry *entry;
669 unsigned long flags;
670
671 spin_lock_irqsave(&free_entries_lock, flags);
672
673 if (list_empty(&free_entries)) {
674 global_disable = true;
675 spin_unlock_irqrestore(&free_entries_lock, flags);
676 pr_err("DMA-API: debugging out of memory - disabling\n");
677 return NULL;
678 }
679
680 entry = __dma_entry_alloc();
681
682 spin_unlock_irqrestore(&free_entries_lock, flags);
683
684#ifdef CONFIG_STACKTRACE
685 entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
686 entry->stacktrace.entries = entry->st_entries;
687 entry->stacktrace.skip = 2;
688 save_stack_trace(&entry->stacktrace);
689#endif
690
691 return entry;
692}
693
694static void dma_entry_free(struct dma_debug_entry *entry)
695{
696 unsigned long flags;
697
698 active_cacheline_remove(entry);
699
700 /*
701 * add to beginning of the list - this way the entries are
702 * more likely cache hot when they are reallocated.
703 */
704 spin_lock_irqsave(&free_entries_lock, flags);
705 list_add(&entry->list, &free_entries);
706 num_free_entries += 1;
707 spin_unlock_irqrestore(&free_entries_lock, flags);
708}
709
710int dma_debug_resize_entries(u32 num_entries)
711{
712 int i, delta, ret = 0;
713 unsigned long flags;
714 struct dma_debug_entry *entry;
715 LIST_HEAD(tmp);
716
717 spin_lock_irqsave(&free_entries_lock, flags);
718
719 if (nr_total_entries < num_entries) {
720 delta = num_entries - nr_total_entries;
721
722 spin_unlock_irqrestore(&free_entries_lock, flags);
723
724 for (i = 0; i < delta; i++) {
725 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
726 if (!entry)
727 break;
728
729 list_add_tail(&entry->list, &tmp);
730 }
731
732 spin_lock_irqsave(&free_entries_lock, flags);
733
734 list_splice(&tmp, &free_entries);
735 nr_total_entries += i;
736 num_free_entries += i;
737 } else {
738 delta = nr_total_entries - num_entries;
739
740 for (i = 0; i < delta && !list_empty(&free_entries); i++) {
741 entry = __dma_entry_alloc();
742 kfree(entry);
743 }
744
745 nr_total_entries -= i;
746 }
747
748 if (nr_total_entries != num_entries)
749 ret = 1;
750
751 spin_unlock_irqrestore(&free_entries_lock, flags);
752
753 return ret;
754}
755
756/*
757 * DMA-API debugging init code
758 *
759 * The init code does two things:
760 * 1. Initialize core data structures
761 * 2. Preallocate a given number of dma_debug_entry structs
762 */
763
764static int prealloc_memory(u32 num_entries)
765{
766 struct dma_debug_entry *entry, *next_entry;
767 int i;
768
769 for (i = 0; i < num_entries; ++i) {
770 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
771 if (!entry)
772 goto out_err;
773
774 list_add_tail(&entry->list, &free_entries);
775 }
776
777 num_free_entries = num_entries;
778 min_free_entries = num_entries;
779
780 pr_info("DMA-API: preallocated %d debug entries\n", num_entries);
781
782 return 0;
783
784out_err:
785
786 list_for_each_entry_safe(entry, next_entry, &free_entries, list) {
787 list_del(&entry->list);
788 kfree(entry);
789 }
790
791 return -ENOMEM;
792}
793
794static ssize_t filter_read(struct file *file, char __user *user_buf,
795 size_t count, loff_t *ppos)
796{
797 char buf[NAME_MAX_LEN + 1];
798 unsigned long flags;
799 int len;
800
801 if (!current_driver_name[0])
802 return 0;
803
804 /*
805 * We can't copy to userspace directly because current_driver_name can
806 * only be read under the driver_name_lock with irqs disabled. So
807 * create a temporary copy first.
808 */
809 read_lock_irqsave(&driver_name_lock, flags);
810 len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
811 read_unlock_irqrestore(&driver_name_lock, flags);
812
813 return simple_read_from_buffer(user_buf, count, ppos, buf, len);
814}
815
816static ssize_t filter_write(struct file *file, const char __user *userbuf,
817 size_t count, loff_t *ppos)
818{
819 char buf[NAME_MAX_LEN];
820 unsigned long flags;
821 size_t len;
822 int i;
823
824 /*
825 * We can't copy from userspace directly. Access to
826 * current_driver_name is protected with a write_lock with irqs
827 * disabled. Since copy_from_user can fault and may sleep we
828 * need to copy to temporary buffer first
829 */
830 len = min(count, (size_t)(NAME_MAX_LEN - 1));
831 if (copy_from_user(buf, userbuf, len))
832 return -EFAULT;
833
834 buf[len] = 0;
835
836 write_lock_irqsave(&driver_name_lock, flags);
837
838 /*
839 * Now handle the string we got from userspace very carefully.
840 * The rules are:
841 * - only use the first token we got
842 * - token delimiter is everything looking like a space
843 * character (' ', '\n', '\t' ...)
844 *
845 */
846 if (!isalnum(buf[0])) {
847 /*
848 * If the first character userspace gave us is not
849 * alphanumerical then assume the filter should be
850 * switched off.
851 */
852 if (current_driver_name[0])
853 pr_info("DMA-API: switching off dma-debug driver filter\n");
854 current_driver_name[0] = 0;
855 current_driver = NULL;
856 goto out_unlock;
857 }
858
859 /*
860 * Now parse out the first token and use it as the name for the
861 * driver to filter for.
862 */
863 for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
864 current_driver_name[i] = buf[i];
865 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
866 break;
867 }
868 current_driver_name[i] = 0;
869 current_driver = NULL;
870
871 pr_info("DMA-API: enable driver filter for driver [%s]\n",
872 current_driver_name);
873
874out_unlock:
875 write_unlock_irqrestore(&driver_name_lock, flags);
876
877 return count;
878}
879
880static const struct file_operations filter_fops = {
881 .read = filter_read,
882 .write = filter_write,
883 .llseek = default_llseek,
884};
885
886static int dma_debug_fs_init(void)
887{
888 dma_debug_dent = debugfs_create_dir("dma-api", NULL);
889 if (!dma_debug_dent) {
890 pr_err("DMA-API: can not create debugfs directory\n");
891 return -ENOMEM;
892 }
893
894 global_disable_dent = debugfs_create_bool("disabled", 0444,
895 dma_debug_dent,
896 &global_disable);
897 if (!global_disable_dent)
898 goto out_err;
899
900 error_count_dent = debugfs_create_u32("error_count", 0444,
901 dma_debug_dent, &error_count);
902 if (!error_count_dent)
903 goto out_err;
904
905 show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
906 dma_debug_dent,
907 &show_all_errors);
908 if (!show_all_errors_dent)
909 goto out_err;
910
911 show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
912 dma_debug_dent,
913 &show_num_errors);
914 if (!show_num_errors_dent)
915 goto out_err;
916
917 num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
918 dma_debug_dent,
919 &num_free_entries);
920 if (!num_free_entries_dent)
921 goto out_err;
922
923 min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
924 dma_debug_dent,
925 &min_free_entries);
926 if (!min_free_entries_dent)
927 goto out_err;
928
929 filter_dent = debugfs_create_file("driver_filter", 0644,
930 dma_debug_dent, NULL, &filter_fops);
931 if (!filter_dent)
932 goto out_err;
933
934 return 0;
935
936out_err:
937 debugfs_remove_recursive(dma_debug_dent);
938
939 return -ENOMEM;
940}
941
942static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
943{
944 struct dma_debug_entry *entry;
945 unsigned long flags;
946 int count = 0, i;
947
948 for (i = 0; i < HASH_SIZE; ++i) {
949 spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
950 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
951 if (entry->dev == dev) {
952 count += 1;
953 *out_entry = entry;
954 }
955 }
956 spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
957 }
958
959 return count;
960}
961
962static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
963{
964 struct device *dev = data;
965 struct dma_debug_entry *uninitialized_var(entry);
966 int count;
967
968 if (dma_debug_disabled())
969 return 0;
970
971 switch (action) {
972 case BUS_NOTIFY_UNBOUND_DRIVER:
973 count = device_dma_allocations(dev, &entry);
974 if (count == 0)
975 break;
976 err_printk(dev, entry, "DMA-API: device driver has pending "
977 "DMA allocations while released from device "
978 "[count=%d]\n"
979 "One of leaked entries details: "
980 "[device address=0x%016llx] [size=%llu bytes] "
981 "[mapped with %s] [mapped as %s]\n",
982 count, entry->dev_addr, entry->size,
983 dir2name[entry->direction], type2name[entry->type]);
984 break;
985 default:
986 break;
987 }
988
989 return 0;
990}
991
992void dma_debug_add_bus(struct bus_type *bus)
993{
994 struct notifier_block *nb;
995
996 if (dma_debug_disabled())
997 return;
998
999 nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
1000 if (nb == NULL) {
1001 pr_err("dma_debug_add_bus: out of memory\n");
1002 return;
1003 }
1004
1005 nb->notifier_call = dma_debug_device_change;
1006
1007 bus_register_notifier(bus, nb);
1008}
1009
1010static int dma_debug_init(void)
1011{
1012 int i;
1013
1014 /* Do not use dma_debug_initialized here, since we really want to be
1015 * called to set dma_debug_initialized
1016 */
1017 if (global_disable)
1018 return 0;
1019
1020 for (i = 0; i < HASH_SIZE; ++i) {
1021 INIT_LIST_HEAD(&dma_entry_hash[i].list);
1022 spin_lock_init(&dma_entry_hash[i].lock);
1023 }
1024
1025 if (dma_debug_fs_init() != 0) {
1026 pr_err("DMA-API: error creating debugfs entries - disabling\n");
1027 global_disable = true;
1028
1029 return 0;
1030 }
1031
1032 if (prealloc_memory(nr_prealloc_entries) != 0) {
1033 pr_err("DMA-API: debugging out of memory error - disabled\n");
1034 global_disable = true;
1035
1036 return 0;
1037 }
1038
1039 nr_total_entries = num_free_entries;
1040
1041 dma_debug_initialized = true;
1042
1043 pr_info("DMA-API: debugging enabled by kernel config\n");
1044 return 0;
1045}
1046core_initcall(dma_debug_init);
1047
1048static __init int dma_debug_cmdline(char *str)
1049{
1050 if (!str)
1051 return -EINVAL;
1052
1053 if (strncmp(str, "off", 3) == 0) {
1054 pr_info("DMA-API: debugging disabled on kernel command line\n");
1055 global_disable = true;
1056 }
1057
1058 return 0;
1059}
1060
1061static __init int dma_debug_entries_cmdline(char *str)
1062{
1063 if (!str)
1064 return -EINVAL;
1065 if (!get_option(&str, &nr_prealloc_entries))
1066 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
1067 return 0;
1068}
1069
1070__setup("dma_debug=", dma_debug_cmdline);
1071__setup("dma_debug_entries=", dma_debug_entries_cmdline);
1072
1073static void check_unmap(struct dma_debug_entry *ref)
1074{
1075 struct dma_debug_entry *entry;
1076 struct hash_bucket *bucket;
1077 unsigned long flags;
1078
1079 bucket = get_hash_bucket(ref, &flags);
1080 entry = bucket_find_exact(bucket, ref);
1081
1082 if (!entry) {
1083 /* must drop lock before calling dma_mapping_error */
1084 put_hash_bucket(bucket, &flags);
1085
1086 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1087 err_printk(ref->dev, NULL,
1088 "DMA-API: device driver tries to free an "
1089 "invalid DMA memory address\n");
1090 } else {
1091 err_printk(ref->dev, NULL,
1092 "DMA-API: device driver tries to free DMA "
1093 "memory it has not allocated [device "
1094 "address=0x%016llx] [size=%llu bytes]\n",
1095 ref->dev_addr, ref->size);
1096 }
1097 return;
1098 }
1099
1100 if (ref->size != entry->size) {
1101 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1102 "DMA memory with different size "
1103 "[device address=0x%016llx] [map size=%llu bytes] "
1104 "[unmap size=%llu bytes]\n",
1105 ref->dev_addr, entry->size, ref->size);
1106 }
1107
1108 if (ref->type != entry->type) {
1109 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1110 "DMA memory with wrong function "
1111 "[device address=0x%016llx] [size=%llu bytes] "
1112 "[mapped as %s] [unmapped as %s]\n",
1113 ref->dev_addr, ref->size,
1114 type2name[entry->type], type2name[ref->type]);
1115 } else if ((entry->type == dma_debug_coherent) &&
1116 (phys_addr(ref) != phys_addr(entry))) {
1117 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1118 "DMA memory with different CPU address "
1119 "[device address=0x%016llx] [size=%llu bytes] "
1120 "[cpu alloc address=0x%016llx] "
1121 "[cpu free address=0x%016llx]",
1122 ref->dev_addr, ref->size,
1123 phys_addr(entry),
1124 phys_addr(ref));
1125 }
1126
1127 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1128 ref->sg_call_ents != entry->sg_call_ents) {
1129 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1130 "DMA sg list with different entry count "
1131 "[map count=%d] [unmap count=%d]\n",
1132 entry->sg_call_ents, ref->sg_call_ents);
1133 }
1134
1135 /*
1136 * This may be no bug in reality - but most implementations of the
1137 * DMA API don't handle this properly, so check for it here
1138 */
1139 if (ref->direction != entry->direction) {
1140 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1141 "DMA memory with different direction "
1142 "[device address=0x%016llx] [size=%llu bytes] "
1143 "[mapped with %s] [unmapped with %s]\n",
1144 ref->dev_addr, ref->size,
1145 dir2name[entry->direction],
1146 dir2name[ref->direction]);
1147 }
1148
1149 /*
1150 * Drivers should use dma_mapping_error() to check the returned
1151 * addresses of dma_map_single() and dma_map_page().
1152 * If not, print this warning message. See Documentation/DMA-API.txt.
1153 */
1154 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1155 err_printk(ref->dev, entry,
1156 "DMA-API: device driver failed to check map error"
1157 "[device address=0x%016llx] [size=%llu bytes] "
1158 "[mapped as %s]",
1159 ref->dev_addr, ref->size,
1160 type2name[entry->type]);
1161 }
1162
1163 hash_bucket_del(entry);
1164 dma_entry_free(entry);
1165
1166 put_hash_bucket(bucket, &flags);
1167}
1168
1169static void check_for_stack(struct device *dev,
1170 struct page *page, size_t offset)
1171{
1172 void *addr;
1173 struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1174
1175 if (!stack_vm_area) {
1176 /* Stack is direct-mapped. */
1177 if (PageHighMem(page))
1178 return;
1179 addr = page_address(page) + offset;
1180 if (object_is_on_stack(addr))
1181 err_printk(dev, NULL, "DMA-API: device driver maps memory from stack [addr=%p]\n", addr);
1182 } else {
1183 /* Stack is vmalloced. */
1184 int i;
1185
1186 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1187 if (page != stack_vm_area->pages[i])
1188 continue;
1189
1190 addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1191 err_printk(dev, NULL, "DMA-API: device driver maps memory from stack [probable addr=%p]\n", addr);
1192 break;
1193 }
1194 }
1195}
1196
1197static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1198{
1199 unsigned long a1 = (unsigned long)addr;
1200 unsigned long b1 = a1 + len;
1201 unsigned long a2 = (unsigned long)start;
1202 unsigned long b2 = (unsigned long)end;
1203
1204 return !(b1 <= a2 || a1 >= b2);
1205}
1206
1207static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1208{
1209 if (overlap(addr, len, _stext, _etext) ||
1210 overlap(addr, len, __start_rodata, __end_rodata))
1211 err_printk(dev, NULL, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1212}
1213
1214static void check_sync(struct device *dev,
1215 struct dma_debug_entry *ref,
1216 bool to_cpu)
1217{
1218 struct dma_debug_entry *entry;
1219 struct hash_bucket *bucket;
1220 unsigned long flags;
1221
1222 bucket = get_hash_bucket(ref, &flags);
1223
1224 entry = bucket_find_contain(&bucket, ref, &flags);
1225
1226 if (!entry) {
1227 err_printk(dev, NULL, "DMA-API: device driver tries "
1228 "to sync DMA memory it has not allocated "
1229 "[device address=0x%016llx] [size=%llu bytes]\n",
1230 (unsigned long long)ref->dev_addr, ref->size);
1231 goto out;
1232 }
1233
1234 if (ref->size > entry->size) {
1235 err_printk(dev, entry, "DMA-API: device driver syncs"
1236 " DMA memory outside allocated range "
1237 "[device address=0x%016llx] "
1238 "[allocation size=%llu bytes] "
1239 "[sync offset+size=%llu]\n",
1240 entry->dev_addr, entry->size,
1241 ref->size);
1242 }
1243
1244 if (entry->direction == DMA_BIDIRECTIONAL)
1245 goto out;
1246
1247 if (ref->direction != entry->direction) {
1248 err_printk(dev, entry, "DMA-API: device driver syncs "
1249 "DMA memory with different direction "
1250 "[device address=0x%016llx] [size=%llu bytes] "
1251 "[mapped with %s] [synced with %s]\n",
1252 (unsigned long long)ref->dev_addr, entry->size,
1253 dir2name[entry->direction],
1254 dir2name[ref->direction]);
1255 }
1256
1257 if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1258 !(ref->direction == DMA_TO_DEVICE))
1259 err_printk(dev, entry, "DMA-API: device driver syncs "
1260 "device read-only DMA memory for cpu "
1261 "[device address=0x%016llx] [size=%llu bytes] "
1262 "[mapped with %s] [synced with %s]\n",
1263 (unsigned long long)ref->dev_addr, entry->size,
1264 dir2name[entry->direction],
1265 dir2name[ref->direction]);
1266
1267 if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1268 !(ref->direction == DMA_FROM_DEVICE))
1269 err_printk(dev, entry, "DMA-API: device driver syncs "
1270 "device write-only DMA memory to device "
1271 "[device address=0x%016llx] [size=%llu bytes] "
1272 "[mapped with %s] [synced with %s]\n",
1273 (unsigned long long)ref->dev_addr, entry->size,
1274 dir2name[entry->direction],
1275 dir2name[ref->direction]);
1276
1277 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1278 ref->sg_call_ents != entry->sg_call_ents) {
1279 err_printk(ref->dev, entry, "DMA-API: device driver syncs "
1280 "DMA sg list with different entry count "
1281 "[map count=%d] [sync count=%d]\n",
1282 entry->sg_call_ents, ref->sg_call_ents);
1283 }
1284
1285out:
1286 put_hash_bucket(bucket, &flags);
1287}
1288
1289static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1290{
1291#ifdef CONFIG_DMA_API_DEBUG_SG
1292 unsigned int max_seg = dma_get_max_seg_size(dev);
1293 u64 start, end, boundary = dma_get_seg_boundary(dev);
1294
1295 /*
1296 * Either the driver forgot to set dma_parms appropriately, or
1297 * whoever generated the list forgot to check them.
1298 */
1299 if (sg->length > max_seg)
1300 err_printk(dev, NULL, "DMA-API: mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1301 sg->length, max_seg);
1302 /*
1303 * In some cases this could potentially be the DMA API
1304 * implementation's fault, but it would usually imply that
1305 * the scatterlist was built inappropriately to begin with.
1306 */
1307 start = sg_dma_address(sg);
1308 end = start + sg_dma_len(sg) - 1;
1309 if ((start ^ end) & ~boundary)
1310 err_printk(dev, NULL, "DMA-API: mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1311 start, end, boundary);
1312#endif
1313}
1314
1315void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1316 size_t size, int direction, dma_addr_t dma_addr,
1317 bool map_single)
1318{
1319 struct dma_debug_entry *entry;
1320
1321 if (unlikely(dma_debug_disabled()))
1322 return;
1323
1324 if (dma_mapping_error(dev, dma_addr))
1325 return;
1326
1327 entry = dma_entry_alloc();
1328 if (!entry)
1329 return;
1330
1331 entry->dev = dev;
1332 entry->type = dma_debug_page;
1333 entry->pfn = page_to_pfn(page);
1334 entry->offset = offset,
1335 entry->dev_addr = dma_addr;
1336 entry->size = size;
1337 entry->direction = direction;
1338 entry->map_err_type = MAP_ERR_NOT_CHECKED;
1339
1340 if (map_single)
1341 entry->type = dma_debug_single;
1342
1343 check_for_stack(dev, page, offset);
1344
1345 if (!PageHighMem(page)) {
1346 void *addr = page_address(page) + offset;
1347
1348 check_for_illegal_area(dev, addr, size);
1349 }
1350
1351 add_dma_entry(entry);
1352}
1353EXPORT_SYMBOL(debug_dma_map_page);
1354
1355void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1356{
1357 struct dma_debug_entry ref;
1358 struct dma_debug_entry *entry;
1359 struct hash_bucket *bucket;
1360 unsigned long flags;
1361
1362 if (unlikely(dma_debug_disabled()))
1363 return;
1364
1365 ref.dev = dev;
1366 ref.dev_addr = dma_addr;
1367 bucket = get_hash_bucket(&ref, &flags);
1368
1369 list_for_each_entry(entry, &bucket->list, list) {
1370 if (!exact_match(&ref, entry))
1371 continue;
1372
1373 /*
1374 * The same physical address can be mapped multiple
1375 * times. Without a hardware IOMMU this results in the
1376 * same device addresses being put into the dma-debug
1377 * hash multiple times too. This can result in false
1378 * positives being reported. Therefore we implement a
1379 * best-fit algorithm here which updates the first entry
1380 * from the hash which fits the reference value and is
1381 * not currently listed as being checked.
1382 */
1383 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1384 entry->map_err_type = MAP_ERR_CHECKED;
1385 break;
1386 }
1387 }
1388
1389 put_hash_bucket(bucket, &flags);
1390}
1391EXPORT_SYMBOL(debug_dma_mapping_error);
1392
1393void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1394 size_t size, int direction, bool map_single)
1395{
1396 struct dma_debug_entry ref = {
1397 .type = dma_debug_page,
1398 .dev = dev,
1399 .dev_addr = addr,
1400 .size = size,
1401 .direction = direction,
1402 };
1403
1404 if (unlikely(dma_debug_disabled()))
1405 return;
1406
1407 if (map_single)
1408 ref.type = dma_debug_single;
1409
1410 check_unmap(&ref);
1411}
1412EXPORT_SYMBOL(debug_dma_unmap_page);
1413
1414void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1415 int nents, int mapped_ents, int direction)
1416{
1417 struct dma_debug_entry *entry;
1418 struct scatterlist *s;
1419 int i;
1420
1421 if (unlikely(dma_debug_disabled()))
1422 return;
1423
1424 for_each_sg(sg, s, mapped_ents, i) {
1425 entry = dma_entry_alloc();
1426 if (!entry)
1427 return;
1428
1429 entry->type = dma_debug_sg;
1430 entry->dev = dev;
1431 entry->pfn = page_to_pfn(sg_page(s));
1432 entry->offset = s->offset,
1433 entry->size = sg_dma_len(s);
1434 entry->dev_addr = sg_dma_address(s);
1435 entry->direction = direction;
1436 entry->sg_call_ents = nents;
1437 entry->sg_mapped_ents = mapped_ents;
1438
1439 check_for_stack(dev, sg_page(s), s->offset);
1440
1441 if (!PageHighMem(sg_page(s))) {
1442 check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1443 }
1444
1445 check_sg_segment(dev, s);
1446
1447 add_dma_entry(entry);
1448 }
1449}
1450EXPORT_SYMBOL(debug_dma_map_sg);
1451
1452static int get_nr_mapped_entries(struct device *dev,
1453 struct dma_debug_entry *ref)
1454{
1455 struct dma_debug_entry *entry;
1456 struct hash_bucket *bucket;
1457 unsigned long flags;
1458 int mapped_ents;
1459
1460 bucket = get_hash_bucket(ref, &flags);
1461 entry = bucket_find_exact(bucket, ref);
1462 mapped_ents = 0;
1463
1464 if (entry)
1465 mapped_ents = entry->sg_mapped_ents;
1466 put_hash_bucket(bucket, &flags);
1467
1468 return mapped_ents;
1469}
1470
1471void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1472 int nelems, int dir)
1473{
1474 struct scatterlist *s;
1475 int mapped_ents = 0, i;
1476
1477 if (unlikely(dma_debug_disabled()))
1478 return;
1479
1480 for_each_sg(sglist, s, nelems, i) {
1481
1482 struct dma_debug_entry ref = {
1483 .type = dma_debug_sg,
1484 .dev = dev,
1485 .pfn = page_to_pfn(sg_page(s)),
1486 .offset = s->offset,
1487 .dev_addr = sg_dma_address(s),
1488 .size = sg_dma_len(s),
1489 .direction = dir,
1490 .sg_call_ents = nelems,
1491 };
1492
1493 if (mapped_ents && i >= mapped_ents)
1494 break;
1495
1496 if (!i)
1497 mapped_ents = get_nr_mapped_entries(dev, &ref);
1498
1499 check_unmap(&ref);
1500 }
1501}
1502EXPORT_SYMBOL(debug_dma_unmap_sg);
1503
1504void debug_dma_alloc_coherent(struct device *dev, size_t size,
1505 dma_addr_t dma_addr, void *virt)
1506{
1507 struct dma_debug_entry *entry;
1508
1509 if (unlikely(dma_debug_disabled()))
1510 return;
1511
1512 if (unlikely(virt == NULL))
1513 return;
1514
1515 /* handle vmalloc and linear addresses */
1516 if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1517 return;
1518
1519 entry = dma_entry_alloc();
1520 if (!entry)
1521 return;
1522
1523 entry->type = dma_debug_coherent;
1524 entry->dev = dev;
1525 entry->offset = offset_in_page(virt);
1526 entry->size = size;
1527 entry->dev_addr = dma_addr;
1528 entry->direction = DMA_BIDIRECTIONAL;
1529
1530 if (is_vmalloc_addr(virt))
1531 entry->pfn = vmalloc_to_pfn(virt);
1532 else
1533 entry->pfn = page_to_pfn(virt_to_page(virt));
1534
1535 add_dma_entry(entry);
1536}
1537EXPORT_SYMBOL(debug_dma_alloc_coherent);
1538
1539void debug_dma_free_coherent(struct device *dev, size_t size,
1540 void *virt, dma_addr_t addr)
1541{
1542 struct dma_debug_entry ref = {
1543 .type = dma_debug_coherent,
1544 .dev = dev,
1545 .offset = offset_in_page(virt),
1546 .dev_addr = addr,
1547 .size = size,
1548 .direction = DMA_BIDIRECTIONAL,
1549 };
1550
1551 /* handle vmalloc and linear addresses */
1552 if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1553 return;
1554
1555 if (is_vmalloc_addr(virt))
1556 ref.pfn = vmalloc_to_pfn(virt);
1557 else
1558 ref.pfn = page_to_pfn(virt_to_page(virt));
1559
1560 if (unlikely(dma_debug_disabled()))
1561 return;
1562
1563 check_unmap(&ref);
1564}
1565EXPORT_SYMBOL(debug_dma_free_coherent);
1566
1567void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1568 int direction, dma_addr_t dma_addr)
1569{
1570 struct dma_debug_entry *entry;
1571
1572 if (unlikely(dma_debug_disabled()))
1573 return;
1574
1575 entry = dma_entry_alloc();
1576 if (!entry)
1577 return;
1578
1579 entry->type = dma_debug_resource;
1580 entry->dev = dev;
1581 entry->pfn = PHYS_PFN(addr);
1582 entry->offset = offset_in_page(addr);
1583 entry->size = size;
1584 entry->dev_addr = dma_addr;
1585 entry->direction = direction;
1586 entry->map_err_type = MAP_ERR_NOT_CHECKED;
1587
1588 add_dma_entry(entry);
1589}
1590EXPORT_SYMBOL(debug_dma_map_resource);
1591
1592void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1593 size_t size, int direction)
1594{
1595 struct dma_debug_entry ref = {
1596 .type = dma_debug_resource,
1597 .dev = dev,
1598 .dev_addr = dma_addr,
1599 .size = size,
1600 .direction = direction,
1601 };
1602
1603 if (unlikely(dma_debug_disabled()))
1604 return;
1605
1606 check_unmap(&ref);
1607}
1608EXPORT_SYMBOL(debug_dma_unmap_resource);
1609
1610void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1611 size_t size, int direction)
1612{
1613 struct dma_debug_entry ref;
1614
1615 if (unlikely(dma_debug_disabled()))
1616 return;
1617
1618 ref.type = dma_debug_single;
1619 ref.dev = dev;
1620 ref.dev_addr = dma_handle;
1621 ref.size = size;
1622 ref.direction = direction;
1623 ref.sg_call_ents = 0;
1624
1625 check_sync(dev, &ref, true);
1626}
1627EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1628
1629void debug_dma_sync_single_for_device(struct device *dev,
1630 dma_addr_t dma_handle, size_t size,
1631 int direction)
1632{
1633 struct dma_debug_entry ref;
1634
1635 if (unlikely(dma_debug_disabled()))
1636 return;
1637
1638 ref.type = dma_debug_single;
1639 ref.dev = dev;
1640 ref.dev_addr = dma_handle;
1641 ref.size = size;
1642 ref.direction = direction;
1643 ref.sg_call_ents = 0;
1644
1645 check_sync(dev, &ref, false);
1646}
1647EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1648
1649void debug_dma_sync_single_range_for_cpu(struct device *dev,
1650 dma_addr_t dma_handle,
1651 unsigned long offset, size_t size,
1652 int direction)
1653{
1654 struct dma_debug_entry ref;
1655
1656 if (unlikely(dma_debug_disabled()))
1657 return;
1658
1659 ref.type = dma_debug_single;
1660 ref.dev = dev;
1661 ref.dev_addr = dma_handle;
1662 ref.size = offset + size;
1663 ref.direction = direction;
1664 ref.sg_call_ents = 0;
1665
1666 check_sync(dev, &ref, true);
1667}
1668EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
1669
1670void debug_dma_sync_single_range_for_device(struct device *dev,
1671 dma_addr_t dma_handle,
1672 unsigned long offset,
1673 size_t size, int direction)
1674{
1675 struct dma_debug_entry ref;
1676
1677 if (unlikely(dma_debug_disabled()))
1678 return;
1679
1680 ref.type = dma_debug_single;
1681 ref.dev = dev;
1682 ref.dev_addr = dma_handle;
1683 ref.size = offset + size;
1684 ref.direction = direction;
1685 ref.sg_call_ents = 0;
1686
1687 check_sync(dev, &ref, false);
1688}
1689EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
1690
1691void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1692 int nelems, int direction)
1693{
1694 struct scatterlist *s;
1695 int mapped_ents = 0, i;
1696
1697 if (unlikely(dma_debug_disabled()))
1698 return;
1699
1700 for_each_sg(sg, s, nelems, i) {
1701
1702 struct dma_debug_entry ref = {
1703 .type = dma_debug_sg,
1704 .dev = dev,
1705 .pfn = page_to_pfn(sg_page(s)),
1706 .offset = s->offset,
1707 .dev_addr = sg_dma_address(s),
1708 .size = sg_dma_len(s),
1709 .direction = direction,
1710 .sg_call_ents = nelems,
1711 };
1712
1713 if (!i)
1714 mapped_ents = get_nr_mapped_entries(dev, &ref);
1715
1716 if (i >= mapped_ents)
1717 break;
1718
1719 check_sync(dev, &ref, true);
1720 }
1721}
1722EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1723
1724void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1725 int nelems, int direction)
1726{
1727 struct scatterlist *s;
1728 int mapped_ents = 0, i;
1729
1730 if (unlikely(dma_debug_disabled()))
1731 return;
1732
1733 for_each_sg(sg, s, nelems, i) {
1734
1735 struct dma_debug_entry ref = {
1736 .type = dma_debug_sg,
1737 .dev = dev,
1738 .pfn = page_to_pfn(sg_page(s)),
1739 .offset = s->offset,
1740 .dev_addr = sg_dma_address(s),
1741 .size = sg_dma_len(s),
1742 .direction = direction,
1743 .sg_call_ents = nelems,
1744 };
1745 if (!i)
1746 mapped_ents = get_nr_mapped_entries(dev, &ref);
1747
1748 if (i >= mapped_ents)
1749 break;
1750
1751 check_sync(dev, &ref, false);
1752 }
1753}
1754EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1755
1756static int __init dma_debug_driver_setup(char *str)
1757{
1758 int i;
1759
1760 for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1761 current_driver_name[i] = *str;
1762 if (*str == 0)
1763 break;
1764 }
1765
1766 if (current_driver_name[0])
1767 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1768 current_driver_name);
1769
1770
1771 return 1;
1772}
1773__setup("dma_debug_driver=", dma_debug_driver_setup);