projects/linux·blob v5.0 · refs

1707 lines · 44.5 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#define pr_fmt(fmt) "DMA-API: " fmt
21
22#include <linux/sched/task_stack.h>
23#include <linux/scatterlist.h>
24#include <linux/dma-mapping.h>
25#include <linux/sched/task.h>
26#include <linux/stacktrace.h>
27#include <linux/dma-debug.h>
28#include <linux/spinlock.h>
29#include <linux/vmalloc.h>
30#include <linux/debugfs.h>
31#include <linux/uaccess.h>
32#include <linux/export.h>
33#include <linux/device.h>
34#include <linux/types.h>
35#include <linux/sched.h>
36#include <linux/ctype.h>
37#include <linux/list.h>
38#include <linux/slab.h>
39
40#include <asm/sections.h>
41
42#define HASH_SIZE 1024ULL
43#define HASH_FN_SHIFT 13
44#define HASH_FN_MASK (HASH_SIZE - 1)
45
46#define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
47/* If the pool runs out, add this many new entries at once */
48#define DMA_DEBUG_DYNAMIC_ENTRIES (PAGE_SIZE / sizeof(struct dma_debug_entry))
49
50enum {
51 dma_debug_single,
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 *nr_total_entries_dent __read_mostly;
146static struct dentry *filter_dent __read_mostly;
147
148/* per-driver filter related state */
149
150#define NAME_MAX_LEN 64
151
152static char current_driver_name[NAME_MAX_LEN] __read_mostly;
153static struct device_driver *current_driver __read_mostly;
154
155static DEFINE_RWLOCK(driver_name_lock);
156
157static const char *const maperr2str[] = {
158 [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
159 [MAP_ERR_NOT_CHECKED] = "dma map error not checked",
160 [MAP_ERR_CHECKED] = "dma map error checked",
161};
162
163static const char *type2name[5] = { "single", "page",
164 "scather-gather", "coherent",
165 "resource" };
166
167static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
168 "DMA_FROM_DEVICE", "DMA_NONE" };
169
170/*
171 * The access to some variables in this macro is racy. We can't use atomic_t
172 * here because all these variables are exported to debugfs. Some of them even
173 * writeable. This is also the reason why a lock won't help much. But anyway,
174 * the races are no big deal. Here is why:
175 *
176 * error_count: the addition is racy, but the worst thing that can happen is
177 * that we don't count some errors
178 * show_num_errors: the subtraction is racy. Also no big deal because in
179 * worst case this will result in one warning more in the
180 * system log than the user configured. This variable is
181 * writeable via debugfs.
182 */
183static inline void dump_entry_trace(struct dma_debug_entry *entry)
184{
185#ifdef CONFIG_STACKTRACE
186 if (entry) {
187 pr_warning("Mapped at:\n");
188 print_stack_trace(&entry->stacktrace, 0);
189 }
190#endif
191}
192
193static bool driver_filter(struct device *dev)
194{
195 struct device_driver *drv;
196 unsigned long flags;
197 bool ret;
198
199 /* driver filter off */
200 if (likely(!current_driver_name[0]))
201 return true;
202
203 /* driver filter on and initialized */
204 if (current_driver && dev && dev->driver == current_driver)
205 return true;
206
207 /* driver filter on, but we can't filter on a NULL device... */
208 if (!dev)
209 return false;
210
211 if (current_driver || !current_driver_name[0])
212 return false;
213
214 /* driver filter on but not yet initialized */
215 drv = dev->driver;
216 if (!drv)
217 return false;
218
219 /* lock to protect against change of current_driver_name */
220 read_lock_irqsave(&driver_name_lock, flags);
221
222 ret = false;
223 if (drv->name &&
224 strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
225 current_driver = drv;
226 ret = true;
227 }
228
229 read_unlock_irqrestore(&driver_name_lock, flags);
230
231 return ret;
232}
233
234#define err_printk(dev, entry, format, arg...) do { \
235 error_count += 1; \
236 if (driver_filter(dev) && \
237 (show_all_errors || show_num_errors > 0)) { \
238 WARN(1, pr_fmt("%s %s: ") format, \
239 dev ? dev_driver_string(dev) : "NULL", \
240 dev ? dev_name(dev) : "NULL", ## arg); \
241 dump_entry_trace(entry); \
242 } \
243 if (!show_all_errors && show_num_errors > 0) \
244 show_num_errors -= 1; \
245 } while (0);
246
247/*
248 * Hash related functions
249 *
250 * Every DMA-API request is saved into a struct dma_debug_entry. To
251 * have quick access to these structs they are stored into a hash.
252 */
253static int hash_fn(struct dma_debug_entry *entry)
254{
255 /*
256 * Hash function is based on the dma address.
257 * We use bits 20-27 here as the index into the hash
258 */
259 return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
260}
261
262/*
263 * Request exclusive access to a hash bucket for a given dma_debug_entry.
264 */
265static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
266 unsigned long *flags)
267 __acquires(&dma_entry_hash[idx].lock)
268{
269 int idx = hash_fn(entry);
270 unsigned long __flags;
271
272 spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
273 *flags = __flags;
274 return &dma_entry_hash[idx];
275}
276
277/*
278 * Give up exclusive access to the hash bucket
279 */
280static void put_hash_bucket(struct hash_bucket *bucket,
281 unsigned long *flags)
282 __releases(&bucket->lock)
283{
284 unsigned long __flags = *flags;
285
286 spin_unlock_irqrestore(&bucket->lock, __flags);
287}
288
289static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
290{
291 return ((a->dev_addr == b->dev_addr) &&
292 (a->dev == b->dev)) ? true : false;
293}
294
295static bool containing_match(struct dma_debug_entry *a,
296 struct dma_debug_entry *b)
297{
298 if (a->dev != b->dev)
299 return false;
300
301 if ((b->dev_addr <= a->dev_addr) &&
302 ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
303 return true;
304
305 return false;
306}
307
308/*
309 * Search a given entry in the hash bucket list
310 */
311static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
312 struct dma_debug_entry *ref,
313 match_fn match)
314{
315 struct dma_debug_entry *entry, *ret = NULL;
316 int matches = 0, match_lvl, last_lvl = -1;
317
318 list_for_each_entry(entry, &bucket->list, list) {
319 if (!match(ref, entry))
320 continue;
321
322 /*
323 * Some drivers map the same physical address multiple
324 * times. Without a hardware IOMMU this results in the
325 * same device addresses being put into the dma-debug
326 * hash multiple times too. This can result in false
327 * positives being reported. Therefore we implement a
328 * best-fit algorithm here which returns the entry from
329 * the hash which fits best to the reference value
330 * instead of the first-fit.
331 */
332 matches += 1;
333 match_lvl = 0;
334 entry->size == ref->size ? ++match_lvl : 0;
335 entry->type == ref->type ? ++match_lvl : 0;
336 entry->direction == ref->direction ? ++match_lvl : 0;
337 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
338
339 if (match_lvl == 4) {
340 /* perfect-fit - return the result */
341 return entry;
342 } else if (match_lvl > last_lvl) {
343 /*
344 * We found an entry that fits better then the
345 * previous one or it is the 1st match.
346 */
347 last_lvl = match_lvl;
348 ret = entry;
349 }
350 }
351
352 /*
353 * If we have multiple matches but no perfect-fit, just return
354 * NULL.
355 */
356 ret = (matches == 1) ? ret : NULL;
357
358 return ret;
359}
360
361static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
362 struct dma_debug_entry *ref)
363{
364 return __hash_bucket_find(bucket, ref, exact_match);
365}
366
367static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
368 struct dma_debug_entry *ref,
369 unsigned long *flags)
370{
371
372 unsigned int max_range = dma_get_max_seg_size(ref->dev);
373 struct dma_debug_entry *entry, index = *ref;
374 unsigned int range = 0;
375
376 while (range <= max_range) {
377 entry = __hash_bucket_find(*bucket, ref, containing_match);
378
379 if (entry)
380 return entry;
381
382 /*
383 * Nothing found, go back a hash bucket
384 */
385 put_hash_bucket(*bucket, flags);
386 range += (1 << HASH_FN_SHIFT);
387 index.dev_addr -= (1 << HASH_FN_SHIFT);
388 *bucket = get_hash_bucket(&index, flags);
389 }
390
391 return NULL;
392}
393
394/*
395 * Add an entry to a hash bucket
396 */
397static void hash_bucket_add(struct hash_bucket *bucket,
398 struct dma_debug_entry *entry)
399{
400 list_add_tail(&entry->list, &bucket->list);
401}
402
403/*
404 * Remove entry from a hash bucket list
405 */
406static void hash_bucket_del(struct dma_debug_entry *entry)
407{
408 list_del(&entry->list);
409}
410
411static unsigned long long phys_addr(struct dma_debug_entry *entry)
412{
413 if (entry->type == dma_debug_resource)
414 return __pfn_to_phys(entry->pfn) + entry->offset;
415
416 return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
417}
418
419/*
420 * Dump mapping entries for debugging purposes
421 */
422void debug_dma_dump_mappings(struct device *dev)
423{
424 int idx;
425
426 for (idx = 0; idx < HASH_SIZE; idx++) {
427 struct hash_bucket *bucket = &dma_entry_hash[idx];
428 struct dma_debug_entry *entry;
429 unsigned long flags;
430
431 spin_lock_irqsave(&bucket->lock, flags);
432
433 list_for_each_entry(entry, &bucket->list, list) {
434 if (!dev || dev == entry->dev) {
435 dev_info(entry->dev,
436 "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
437 type2name[entry->type], idx,
438 phys_addr(entry), entry->pfn,
439 entry->dev_addr, entry->size,
440 dir2name[entry->direction],
441 maperr2str[entry->map_err_type]);
442 }
443 }
444
445 spin_unlock_irqrestore(&bucket->lock, flags);
446 }
447}
448
449/*
450 * For each mapping (initial cacheline in the case of
451 * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
452 * scatterlist, or the cacheline specified in dma_map_single) insert
453 * into this tree using the cacheline as the key. At
454 * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry. If
455 * the entry already exists at insertion time add a tag as a reference
456 * count for the overlapping mappings. For now, the overlap tracking
457 * just ensures that 'unmaps' balance 'maps' before marking the
458 * cacheline idle, but we should also be flagging overlaps as an API
459 * violation.
460 *
461 * Memory usage is mostly constrained by the maximum number of available
462 * dma-debug entries in that we need a free dma_debug_entry before
463 * inserting into the tree. In the case of dma_map_page and
464 * dma_alloc_coherent there is only one dma_debug_entry and one
465 * dma_active_cacheline entry to track per event. dma_map_sg(), on the
466 * other hand, consumes a single dma_debug_entry, but inserts 'nents'
467 * entries into the tree.
468 *
469 * At any time debug_dma_assert_idle() can be called to trigger a
470 * warning if any cachelines in the given page are in the active set.
471 */
472static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
473static DEFINE_SPINLOCK(radix_lock);
474#define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
475#define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
476#define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
477
478static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
479{
480 return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
481 (entry->offset >> L1_CACHE_SHIFT);
482}
483
484static int active_cacheline_read_overlap(phys_addr_t cln)
485{
486 int overlap = 0, i;
487
488 for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
489 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
490 overlap |= 1 << i;
491 return overlap;
492}
493
494static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
495{
496 int i;
497
498 if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
499 return overlap;
500
501 for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
502 if (overlap & 1 << i)
503 radix_tree_tag_set(&dma_active_cacheline, cln, i);
504 else
505 radix_tree_tag_clear(&dma_active_cacheline, cln, i);
506
507 return overlap;
508}
509
510static void active_cacheline_inc_overlap(phys_addr_t cln)
511{
512 int overlap = active_cacheline_read_overlap(cln);
513
514 overlap = active_cacheline_set_overlap(cln, ++overlap);
515
516 /* If we overflowed the overlap counter then we're potentially
517 * leaking dma-mappings. Otherwise, if maps and unmaps are
518 * balanced then this overflow may cause false negatives in
519 * debug_dma_assert_idle() as the cacheline may be marked idle
520 * prematurely.
521 */
522 WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
523 pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
524 ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
525}
526
527static int active_cacheline_dec_overlap(phys_addr_t cln)
528{
529 int overlap = active_cacheline_read_overlap(cln);
530
531 return active_cacheline_set_overlap(cln, --overlap);
532}
533
534static int active_cacheline_insert(struct dma_debug_entry *entry)
535{
536 phys_addr_t cln = to_cacheline_number(entry);
537 unsigned long flags;
538 int rc;
539
540 /* If the device is not writing memory then we don't have any
541 * concerns about the cpu consuming stale data. This mitigates
542 * legitimate usages of overlapping mappings.
543 */
544 if (entry->direction == DMA_TO_DEVICE)
545 return 0;
546
547 spin_lock_irqsave(&radix_lock, flags);
548 rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
549 if (rc == -EEXIST)
550 active_cacheline_inc_overlap(cln);
551 spin_unlock_irqrestore(&radix_lock, flags);
552
553 return rc;
554}
555
556static void active_cacheline_remove(struct dma_debug_entry *entry)
557{
558 phys_addr_t cln = to_cacheline_number(entry);
559 unsigned long flags;
560
561 /* ...mirror the insert case */
562 if (entry->direction == DMA_TO_DEVICE)
563 return;
564
565 spin_lock_irqsave(&radix_lock, flags);
566 /* since we are counting overlaps the final put of the
567 * cacheline will occur when the overlap count is 0.
568 * active_cacheline_dec_overlap() returns -1 in that case
569 */
570 if (active_cacheline_dec_overlap(cln) < 0)
571 radix_tree_delete(&dma_active_cacheline, cln);
572 spin_unlock_irqrestore(&radix_lock, flags);
573}
574
575/**
576 * debug_dma_assert_idle() - assert that a page is not undergoing dma
577 * @page: page to lookup in the dma_active_cacheline tree
578 *
579 * Place a call to this routine in cases where the cpu touching the page
580 * before the dma completes (page is dma_unmapped) will lead to data
581 * corruption.
582 */
583void debug_dma_assert_idle(struct page *page)
584{
585 static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
586 struct dma_debug_entry *entry = NULL;
587 void **results = (void **) &ents;
588 unsigned int nents, i;
589 unsigned long flags;
590 phys_addr_t cln;
591
592 if (dma_debug_disabled())
593 return;
594
595 if (!page)
596 return;
597
598 cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
599 spin_lock_irqsave(&radix_lock, flags);
600 nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
601 CACHELINES_PER_PAGE);
602 for (i = 0; i < nents; i++) {
603 phys_addr_t ent_cln = to_cacheline_number(ents[i]);
604
605 if (ent_cln == cln) {
606 entry = ents[i];
607 break;
608 } else if (ent_cln >= cln + CACHELINES_PER_PAGE)
609 break;
610 }
611 spin_unlock_irqrestore(&radix_lock, flags);
612
613 if (!entry)
614 return;
615
616 cln = to_cacheline_number(entry);
617 err_printk(entry->dev, entry,
618 "cpu touching an active dma mapped cacheline [cln=%pa]\n",
619 &cln);
620}
621
622/*
623 * Wrapper function for adding an entry to the hash.
624 * This function takes care of locking itself.
625 */
626static void add_dma_entry(struct dma_debug_entry *entry)
627{
628 struct hash_bucket *bucket;
629 unsigned long flags;
630 int rc;
631
632 bucket = get_hash_bucket(entry, &flags);
633 hash_bucket_add(bucket, entry);
634 put_hash_bucket(bucket, &flags);
635
636 rc = active_cacheline_insert(entry);
637 if (rc == -ENOMEM) {
638 pr_err("cacheline tracking ENOMEM, dma-debug disabled\n");
639 global_disable = true;
640 }
641
642 /* TODO: report -EEXIST errors here as overlapping mappings are
643 * not supported by the DMA API
644 */
645}
646
647static int dma_debug_create_entries(gfp_t gfp)
648{
649 struct dma_debug_entry *entry;
650 int i;
651
652 entry = (void *)get_zeroed_page(gfp);
653 if (!entry)
654 return -ENOMEM;
655
656 for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
657 list_add_tail(&entry[i].list, &free_entries);
658
659 num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
660 nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
661
662 return 0;
663}
664
665static struct dma_debug_entry *__dma_entry_alloc(void)
666{
667 struct dma_debug_entry *entry;
668
669 entry = list_entry(free_entries.next, struct dma_debug_entry, list);
670 list_del(&entry->list);
671 memset(entry, 0, sizeof(*entry));
672
673 num_free_entries -= 1;
674 if (num_free_entries < min_free_entries)
675 min_free_entries = num_free_entries;
676
677 return entry;
678}
679
680void __dma_entry_alloc_check_leak(void)
681{
682 u32 tmp = nr_total_entries % nr_prealloc_entries;
683
684 /* Shout each time we tick over some multiple of the initial pool */
685 if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
686 pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
687 nr_total_entries,
688 (nr_total_entries / nr_prealloc_entries));
689 }
690}
691
692/* struct dma_entry allocator
693 *
694 * The next two functions implement the allocator for
695 * struct dma_debug_entries.
696 */
697static struct dma_debug_entry *dma_entry_alloc(void)
698{
699 struct dma_debug_entry *entry;
700 unsigned long flags;
701
702 spin_lock_irqsave(&free_entries_lock, flags);
703 if (num_free_entries == 0) {
704 if (dma_debug_create_entries(GFP_ATOMIC)) {
705 global_disable = true;
706 spin_unlock_irqrestore(&free_entries_lock, flags);
707 pr_err("debugging out of memory - disabling\n");
708 return NULL;
709 }
710 __dma_entry_alloc_check_leak();
711 }
712
713 entry = __dma_entry_alloc();
714
715 spin_unlock_irqrestore(&free_entries_lock, flags);
716
717#ifdef CONFIG_STACKTRACE
718 entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
719 entry->stacktrace.entries = entry->st_entries;
720 entry->stacktrace.skip = 2;
721 save_stack_trace(&entry->stacktrace);
722#endif
723
724 return entry;
725}
726
727static void dma_entry_free(struct dma_debug_entry *entry)
728{
729 unsigned long flags;
730
731 active_cacheline_remove(entry);
732
733 /*
734 * add to beginning of the list - this way the entries are
735 * more likely cache hot when they are reallocated.
736 */
737 spin_lock_irqsave(&free_entries_lock, flags);
738 list_add(&entry->list, &free_entries);
739 num_free_entries += 1;
740 spin_unlock_irqrestore(&free_entries_lock, flags);
741}
742
743/*
744 * DMA-API debugging init code
745 *
746 * The init code does two things:
747 * 1. Initialize core data structures
748 * 2. Preallocate a given number of dma_debug_entry structs
749 */
750
751static ssize_t filter_read(struct file *file, char __user *user_buf,
752 size_t count, loff_t *ppos)
753{
754 char buf[NAME_MAX_LEN + 1];
755 unsigned long flags;
756 int len;
757
758 if (!current_driver_name[0])
759 return 0;
760
761 /*
762 * We can't copy to userspace directly because current_driver_name can
763 * only be read under the driver_name_lock with irqs disabled. So
764 * create a temporary copy first.
765 */
766 read_lock_irqsave(&driver_name_lock, flags);
767 len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
768 read_unlock_irqrestore(&driver_name_lock, flags);
769
770 return simple_read_from_buffer(user_buf, count, ppos, buf, len);
771}
772
773static ssize_t filter_write(struct file *file, const char __user *userbuf,
774 size_t count, loff_t *ppos)
775{
776 char buf[NAME_MAX_LEN];
777 unsigned long flags;
778 size_t len;
779 int i;
780
781 /*
782 * We can't copy from userspace directly. Access to
783 * current_driver_name is protected with a write_lock with irqs
784 * disabled. Since copy_from_user can fault and may sleep we
785 * need to copy to temporary buffer first
786 */
787 len = min(count, (size_t)(NAME_MAX_LEN - 1));
788 if (copy_from_user(buf, userbuf, len))
789 return -EFAULT;
790
791 buf[len] = 0;
792
793 write_lock_irqsave(&driver_name_lock, flags);
794
795 /*
796 * Now handle the string we got from userspace very carefully.
797 * The rules are:
798 * - only use the first token we got
799 * - token delimiter is everything looking like a space
800 * character (' ', '\n', '\t' ...)
801 *
802 */
803 if (!isalnum(buf[0])) {
804 /*
805 * If the first character userspace gave us is not
806 * alphanumerical then assume the filter should be
807 * switched off.
808 */
809 if (current_driver_name[0])
810 pr_info("switching off dma-debug driver filter\n");
811 current_driver_name[0] = 0;
812 current_driver = NULL;
813 goto out_unlock;
814 }
815
816 /*
817 * Now parse out the first token and use it as the name for the
818 * driver to filter for.
819 */
820 for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
821 current_driver_name[i] = buf[i];
822 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
823 break;
824 }
825 current_driver_name[i] = 0;
826 current_driver = NULL;
827
828 pr_info("enable driver filter for driver [%s]\n",
829 current_driver_name);
830
831out_unlock:
832 write_unlock_irqrestore(&driver_name_lock, flags);
833
834 return count;
835}
836
837static const struct file_operations filter_fops = {
838 .read = filter_read,
839 .write = filter_write,
840 .llseek = default_llseek,
841};
842
843static int dma_debug_fs_init(void)
844{
845 dma_debug_dent = debugfs_create_dir("dma-api", NULL);
846 if (!dma_debug_dent) {
847 pr_err("can not create debugfs directory\n");
848 return -ENOMEM;
849 }
850
851 global_disable_dent = debugfs_create_bool("disabled", 0444,
852 dma_debug_dent,
853 &global_disable);
854 if (!global_disable_dent)
855 goto out_err;
856
857 error_count_dent = debugfs_create_u32("error_count", 0444,
858 dma_debug_dent, &error_count);
859 if (!error_count_dent)
860 goto out_err;
861
862 show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
863 dma_debug_dent,
864 &show_all_errors);
865 if (!show_all_errors_dent)
866 goto out_err;
867
868 show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
869 dma_debug_dent,
870 &show_num_errors);
871 if (!show_num_errors_dent)
872 goto out_err;
873
874 num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
875 dma_debug_dent,
876 &num_free_entries);
877 if (!num_free_entries_dent)
878 goto out_err;
879
880 min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
881 dma_debug_dent,
882 &min_free_entries);
883 if (!min_free_entries_dent)
884 goto out_err;
885
886 nr_total_entries_dent = debugfs_create_u32("nr_total_entries", 0444,
887 dma_debug_dent,
888 &nr_total_entries);
889 if (!nr_total_entries_dent)
890 goto out_err;
891
892 filter_dent = debugfs_create_file("driver_filter", 0644,
893 dma_debug_dent, NULL, &filter_fops);
894 if (!filter_dent)
895 goto out_err;
896
897 return 0;
898
899out_err:
900 debugfs_remove_recursive(dma_debug_dent);
901
902 return -ENOMEM;
903}
904
905static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
906{
907 struct dma_debug_entry *entry;
908 unsigned long flags;
909 int count = 0, i;
910
911 for (i = 0; i < HASH_SIZE; ++i) {
912 spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
913 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
914 if (entry->dev == dev) {
915 count += 1;
916 *out_entry = entry;
917 }
918 }
919 spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
920 }
921
922 return count;
923}
924
925static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
926{
927 struct device *dev = data;
928 struct dma_debug_entry *uninitialized_var(entry);
929 int count;
930
931 if (dma_debug_disabled())
932 return 0;
933
934 switch (action) {
935 case BUS_NOTIFY_UNBOUND_DRIVER:
936 count = device_dma_allocations(dev, &entry);
937 if (count == 0)
938 break;
939 err_printk(dev, entry, "device driver has pending "
940 "DMA allocations while released from device "
941 "[count=%d]\n"
942 "One of leaked entries details: "
943 "[device address=0x%016llx] [size=%llu bytes] "
944 "[mapped with %s] [mapped as %s]\n",
945 count, entry->dev_addr, entry->size,
946 dir2name[entry->direction], type2name[entry->type]);
947 break;
948 default:
949 break;
950 }
951
952 return 0;
953}
954
955void dma_debug_add_bus(struct bus_type *bus)
956{
957 struct notifier_block *nb;
958
959 if (dma_debug_disabled())
960 return;
961
962 nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
963 if (nb == NULL) {
964 pr_err("dma_debug_add_bus: out of memory\n");
965 return;
966 }
967
968 nb->notifier_call = dma_debug_device_change;
969
970 bus_register_notifier(bus, nb);
971}
972
973static int dma_debug_init(void)
974{
975 int i, nr_pages;
976
977 /* Do not use dma_debug_initialized here, since we really want to be
978 * called to set dma_debug_initialized
979 */
980 if (global_disable)
981 return 0;
982
983 for (i = 0; i < HASH_SIZE; ++i) {
984 INIT_LIST_HEAD(&dma_entry_hash[i].list);
985 spin_lock_init(&dma_entry_hash[i].lock);
986 }
987
988 if (dma_debug_fs_init() != 0) {
989 pr_err("error creating debugfs entries - disabling\n");
990 global_disable = true;
991
992 return 0;
993 }
994
995 nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
996 for (i = 0; i < nr_pages; ++i)
997 dma_debug_create_entries(GFP_KERNEL);
998 if (num_free_entries >= nr_prealloc_entries) {
999 pr_info("preallocated %d debug entries\n", nr_total_entries);
1000 } else if (num_free_entries > 0) {
1001 pr_warn("%d debug entries requested but only %d allocated\n",
1002 nr_prealloc_entries, nr_total_entries);
1003 } else {
1004 pr_err("debugging out of memory error - disabled\n");
1005 global_disable = true;
1006
1007 return 0;
1008 }
1009 min_free_entries = num_free_entries;
1010
1011 dma_debug_initialized = true;
1012
1013 pr_info("debugging enabled by kernel config\n");
1014 return 0;
1015}
1016core_initcall(dma_debug_init);
1017
1018static __init int dma_debug_cmdline(char *str)
1019{
1020 if (!str)
1021 return -EINVAL;
1022
1023 if (strncmp(str, "off", 3) == 0) {
1024 pr_info("debugging disabled on kernel command line\n");
1025 global_disable = true;
1026 }
1027
1028 return 0;
1029}
1030
1031static __init int dma_debug_entries_cmdline(char *str)
1032{
1033 if (!str)
1034 return -EINVAL;
1035 if (!get_option(&str, &nr_prealloc_entries))
1036 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
1037 return 0;
1038}
1039
1040__setup("dma_debug=", dma_debug_cmdline);
1041__setup("dma_debug_entries=", dma_debug_entries_cmdline);
1042
1043static void check_unmap(struct dma_debug_entry *ref)
1044{
1045 struct dma_debug_entry *entry;
1046 struct hash_bucket *bucket;
1047 unsigned long flags;
1048
1049 bucket = get_hash_bucket(ref, &flags);
1050 entry = bucket_find_exact(bucket, ref);
1051
1052 if (!entry) {
1053 /* must drop lock before calling dma_mapping_error */
1054 put_hash_bucket(bucket, &flags);
1055
1056 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1057 err_printk(ref->dev, NULL,
1058 "device driver tries to free an "
1059 "invalid DMA memory address\n");
1060 } else {
1061 err_printk(ref->dev, NULL,
1062 "device driver tries to free DMA "
1063 "memory it has not allocated [device "
1064 "address=0x%016llx] [size=%llu bytes]\n",
1065 ref->dev_addr, ref->size);
1066 }
1067 return;
1068 }
1069
1070 if (ref->size != entry->size) {
1071 err_printk(ref->dev, entry, "device driver frees "
1072 "DMA memory with different size "
1073 "[device address=0x%016llx] [map size=%llu bytes] "
1074 "[unmap size=%llu bytes]\n",
1075 ref->dev_addr, entry->size, ref->size);
1076 }
1077
1078 if (ref->type != entry->type) {
1079 err_printk(ref->dev, entry, "device driver frees "
1080 "DMA memory with wrong function "
1081 "[device address=0x%016llx] [size=%llu bytes] "
1082 "[mapped as %s] [unmapped as %s]\n",
1083 ref->dev_addr, ref->size,
1084 type2name[entry->type], type2name[ref->type]);
1085 } else if ((entry->type == dma_debug_coherent) &&
1086 (phys_addr(ref) != phys_addr(entry))) {
1087 err_printk(ref->dev, entry, "device driver frees "
1088 "DMA memory with different CPU address "
1089 "[device address=0x%016llx] [size=%llu bytes] "
1090 "[cpu alloc address=0x%016llx] "
1091 "[cpu free address=0x%016llx]",
1092 ref->dev_addr, ref->size,
1093 phys_addr(entry),
1094 phys_addr(ref));
1095 }
1096
1097 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1098 ref->sg_call_ents != entry->sg_call_ents) {
1099 err_printk(ref->dev, entry, "device driver frees "
1100 "DMA sg list with different entry count "
1101 "[map count=%d] [unmap count=%d]\n",
1102 entry->sg_call_ents, ref->sg_call_ents);
1103 }
1104
1105 /*
1106 * This may be no bug in reality - but most implementations of the
1107 * DMA API don't handle this properly, so check for it here
1108 */
1109 if (ref->direction != entry->direction) {
1110 err_printk(ref->dev, entry, "device driver frees "
1111 "DMA memory with different direction "
1112 "[device address=0x%016llx] [size=%llu bytes] "
1113 "[mapped with %s] [unmapped with %s]\n",
1114 ref->dev_addr, ref->size,
1115 dir2name[entry->direction],
1116 dir2name[ref->direction]);
1117 }
1118
1119 /*
1120 * Drivers should use dma_mapping_error() to check the returned
1121 * addresses of dma_map_single() and dma_map_page().
1122 * If not, print this warning message. See Documentation/DMA-API.txt.
1123 */
1124 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1125 err_printk(ref->dev, entry,
1126 "device driver failed to check map error"
1127 "[device address=0x%016llx] [size=%llu bytes] "
1128 "[mapped as %s]",
1129 ref->dev_addr, ref->size,
1130 type2name[entry->type]);
1131 }
1132
1133 hash_bucket_del(entry);
1134 dma_entry_free(entry);
1135
1136 put_hash_bucket(bucket, &flags);
1137}
1138
1139static void check_for_stack(struct device *dev,
1140 struct page *page, size_t offset)
1141{
1142 void *addr;
1143 struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1144
1145 if (!stack_vm_area) {
1146 /* Stack is direct-mapped. */
1147 if (PageHighMem(page))
1148 return;
1149 addr = page_address(page) + offset;
1150 if (object_is_on_stack(addr))
1151 err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1152 } else {
1153 /* Stack is vmalloced. */
1154 int i;
1155
1156 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1157 if (page != stack_vm_area->pages[i])
1158 continue;
1159
1160 addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1161 err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1162 break;
1163 }
1164 }
1165}
1166
1167static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1168{
1169 unsigned long a1 = (unsigned long)addr;
1170 unsigned long b1 = a1 + len;
1171 unsigned long a2 = (unsigned long)start;
1172 unsigned long b2 = (unsigned long)end;
1173
1174 return !(b1 <= a2 || a1 >= b2);
1175}
1176
1177static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1178{
1179 if (overlap(addr, len, _stext, _etext) ||
1180 overlap(addr, len, __start_rodata, __end_rodata))
1181 err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1182}
1183
1184static void check_sync(struct device *dev,
1185 struct dma_debug_entry *ref,
1186 bool to_cpu)
1187{
1188 struct dma_debug_entry *entry;
1189 struct hash_bucket *bucket;
1190 unsigned long flags;
1191
1192 bucket = get_hash_bucket(ref, &flags);
1193
1194 entry = bucket_find_contain(&bucket, ref, &flags);
1195
1196 if (!entry) {
1197 err_printk(dev, NULL, "device driver tries "
1198 "to sync DMA memory it has not allocated "
1199 "[device address=0x%016llx] [size=%llu bytes]\n",
1200 (unsigned long long)ref->dev_addr, ref->size);
1201 goto out;
1202 }
1203
1204 if (ref->size > entry->size) {
1205 err_printk(dev, entry, "device driver syncs"
1206 " DMA memory outside allocated range "
1207 "[device address=0x%016llx] "
1208 "[allocation size=%llu bytes] "
1209 "[sync offset+size=%llu]\n",
1210 entry->dev_addr, entry->size,
1211 ref->size);
1212 }
1213
1214 if (entry->direction == DMA_BIDIRECTIONAL)
1215 goto out;
1216
1217 if (ref->direction != entry->direction) {
1218 err_printk(dev, entry, "device driver syncs "
1219 "DMA memory with different direction "
1220 "[device address=0x%016llx] [size=%llu bytes] "
1221 "[mapped with %s] [synced with %s]\n",
1222 (unsigned long long)ref->dev_addr, entry->size,
1223 dir2name[entry->direction],
1224 dir2name[ref->direction]);
1225 }
1226
1227 if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1228 !(ref->direction == DMA_TO_DEVICE))
1229 err_printk(dev, entry, "device driver syncs "
1230 "device read-only DMA memory for cpu "
1231 "[device address=0x%016llx] [size=%llu bytes] "
1232 "[mapped with %s] [synced with %s]\n",
1233 (unsigned long long)ref->dev_addr, entry->size,
1234 dir2name[entry->direction],
1235 dir2name[ref->direction]);
1236
1237 if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1238 !(ref->direction == DMA_FROM_DEVICE))
1239 err_printk(dev, entry, "device driver syncs "
1240 "device write-only DMA memory to device "
1241 "[device address=0x%016llx] [size=%llu bytes] "
1242 "[mapped with %s] [synced with %s]\n",
1243 (unsigned long long)ref->dev_addr, entry->size,
1244 dir2name[entry->direction],
1245 dir2name[ref->direction]);
1246
1247 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1248 ref->sg_call_ents != entry->sg_call_ents) {
1249 err_printk(ref->dev, entry, "device driver syncs "
1250 "DMA sg list with different entry count "
1251 "[map count=%d] [sync count=%d]\n",
1252 entry->sg_call_ents, ref->sg_call_ents);
1253 }
1254
1255out:
1256 put_hash_bucket(bucket, &flags);
1257}
1258
1259static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1260{
1261#ifdef CONFIG_DMA_API_DEBUG_SG
1262 unsigned int max_seg = dma_get_max_seg_size(dev);
1263 u64 start, end, boundary = dma_get_seg_boundary(dev);
1264
1265 /*
1266 * Either the driver forgot to set dma_parms appropriately, or
1267 * whoever generated the list forgot to check them.
1268 */
1269 if (sg->length > max_seg)
1270 err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1271 sg->length, max_seg);
1272 /*
1273 * In some cases this could potentially be the DMA API
1274 * implementation's fault, but it would usually imply that
1275 * the scatterlist was built inappropriately to begin with.
1276 */
1277 start = sg_dma_address(sg);
1278 end = start + sg_dma_len(sg) - 1;
1279 if ((start ^ end) & ~boundary)
1280 err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1281 start, end, boundary);
1282#endif
1283}
1284
1285void debug_dma_map_single(struct device *dev, const void *addr,
1286 unsigned long len)
1287{
1288 if (unlikely(dma_debug_disabled()))
1289 return;
1290
1291 if (!virt_addr_valid(addr))
1292 err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1293 addr, len);
1294
1295 if (is_vmalloc_addr(addr))
1296 err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1297 addr, len);
1298}
1299EXPORT_SYMBOL(debug_dma_map_single);
1300
1301void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1302 size_t size, int direction, dma_addr_t dma_addr)
1303{
1304 struct dma_debug_entry *entry;
1305
1306 if (unlikely(dma_debug_disabled()))
1307 return;
1308
1309 if (dma_mapping_error(dev, dma_addr))
1310 return;
1311
1312 entry = dma_entry_alloc();
1313 if (!entry)
1314 return;
1315
1316 entry->dev = dev;
1317 entry->type = dma_debug_single;
1318 entry->pfn = page_to_pfn(page);
1319 entry->offset = offset,
1320 entry->dev_addr = dma_addr;
1321 entry->size = size;
1322 entry->direction = direction;
1323 entry->map_err_type = MAP_ERR_NOT_CHECKED;
1324
1325 check_for_stack(dev, page, offset);
1326
1327 if (!PageHighMem(page)) {
1328 void *addr = page_address(page) + offset;
1329
1330 check_for_illegal_area(dev, addr, size);
1331 }
1332
1333 add_dma_entry(entry);
1334}
1335EXPORT_SYMBOL(debug_dma_map_page);
1336
1337void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1338{
1339 struct dma_debug_entry ref;
1340 struct dma_debug_entry *entry;
1341 struct hash_bucket *bucket;
1342 unsigned long flags;
1343
1344 if (unlikely(dma_debug_disabled()))
1345 return;
1346
1347 ref.dev = dev;
1348 ref.dev_addr = dma_addr;
1349 bucket = get_hash_bucket(&ref, &flags);
1350
1351 list_for_each_entry(entry, &bucket->list, list) {
1352 if (!exact_match(&ref, entry))
1353 continue;
1354
1355 /*
1356 * The same physical address can be mapped multiple
1357 * times. Without a hardware IOMMU this results in the
1358 * same device addresses being put into the dma-debug
1359 * hash multiple times too. This can result in false
1360 * positives being reported. Therefore we implement a
1361 * best-fit algorithm here which updates the first entry
1362 * from the hash which fits the reference value and is
1363 * not currently listed as being checked.
1364 */
1365 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1366 entry->map_err_type = MAP_ERR_CHECKED;
1367 break;
1368 }
1369 }
1370
1371 put_hash_bucket(bucket, &flags);
1372}
1373EXPORT_SYMBOL(debug_dma_mapping_error);
1374
1375void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1376 size_t size, int direction)
1377{
1378 struct dma_debug_entry ref = {
1379 .type = dma_debug_single,
1380 .dev = dev,
1381 .dev_addr = addr,
1382 .size = size,
1383 .direction = direction,
1384 };
1385
1386 if (unlikely(dma_debug_disabled()))
1387 return;
1388 check_unmap(&ref);
1389}
1390EXPORT_SYMBOL(debug_dma_unmap_page);
1391
1392void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1393 int nents, int mapped_ents, int direction)
1394{
1395 struct dma_debug_entry *entry;
1396 struct scatterlist *s;
1397 int i;
1398
1399 if (unlikely(dma_debug_disabled()))
1400 return;
1401
1402 for_each_sg(sg, s, mapped_ents, i) {
1403 entry = dma_entry_alloc();
1404 if (!entry)
1405 return;
1406
1407 entry->type = dma_debug_sg;
1408 entry->dev = dev;
1409 entry->pfn = page_to_pfn(sg_page(s));
1410 entry->offset = s->offset,
1411 entry->size = sg_dma_len(s);
1412 entry->dev_addr = sg_dma_address(s);
1413 entry->direction = direction;
1414 entry->sg_call_ents = nents;
1415 entry->sg_mapped_ents = mapped_ents;
1416
1417 check_for_stack(dev, sg_page(s), s->offset);
1418
1419 if (!PageHighMem(sg_page(s))) {
1420 check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1421 }
1422
1423 check_sg_segment(dev, s);
1424
1425 add_dma_entry(entry);
1426 }
1427}
1428EXPORT_SYMBOL(debug_dma_map_sg);
1429
1430static int get_nr_mapped_entries(struct device *dev,
1431 struct dma_debug_entry *ref)
1432{
1433 struct dma_debug_entry *entry;
1434 struct hash_bucket *bucket;
1435 unsigned long flags;
1436 int mapped_ents;
1437
1438 bucket = get_hash_bucket(ref, &flags);
1439 entry = bucket_find_exact(bucket, ref);
1440 mapped_ents = 0;
1441
1442 if (entry)
1443 mapped_ents = entry->sg_mapped_ents;
1444 put_hash_bucket(bucket, &flags);
1445
1446 return mapped_ents;
1447}
1448
1449void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1450 int nelems, int dir)
1451{
1452 struct scatterlist *s;
1453 int mapped_ents = 0, i;
1454
1455 if (unlikely(dma_debug_disabled()))
1456 return;
1457
1458 for_each_sg(sglist, s, nelems, i) {
1459
1460 struct dma_debug_entry ref = {
1461 .type = dma_debug_sg,
1462 .dev = dev,
1463 .pfn = page_to_pfn(sg_page(s)),
1464 .offset = s->offset,
1465 .dev_addr = sg_dma_address(s),
1466 .size = sg_dma_len(s),
1467 .direction = dir,
1468 .sg_call_ents = nelems,
1469 };
1470
1471 if (mapped_ents && i >= mapped_ents)
1472 break;
1473
1474 if (!i)
1475 mapped_ents = get_nr_mapped_entries(dev, &ref);
1476
1477 check_unmap(&ref);
1478 }
1479}
1480EXPORT_SYMBOL(debug_dma_unmap_sg);
1481
1482void debug_dma_alloc_coherent(struct device *dev, size_t size,
1483 dma_addr_t dma_addr, void *virt)
1484{
1485 struct dma_debug_entry *entry;
1486
1487 if (unlikely(dma_debug_disabled()))
1488 return;
1489
1490 if (unlikely(virt == NULL))
1491 return;
1492
1493 /* handle vmalloc and linear addresses */
1494 if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1495 return;
1496
1497 entry = dma_entry_alloc();
1498 if (!entry)
1499 return;
1500
1501 entry->type = dma_debug_coherent;
1502 entry->dev = dev;
1503 entry->offset = offset_in_page(virt);
1504 entry->size = size;
1505 entry->dev_addr = dma_addr;
1506 entry->direction = DMA_BIDIRECTIONAL;
1507
1508 if (is_vmalloc_addr(virt))
1509 entry->pfn = vmalloc_to_pfn(virt);
1510 else
1511 entry->pfn = page_to_pfn(virt_to_page(virt));
1512
1513 add_dma_entry(entry);
1514}
1515
1516void debug_dma_free_coherent(struct device *dev, size_t size,
1517 void *virt, dma_addr_t addr)
1518{
1519 struct dma_debug_entry ref = {
1520 .type = dma_debug_coherent,
1521 .dev = dev,
1522 .offset = offset_in_page(virt),
1523 .dev_addr = addr,
1524 .size = size,
1525 .direction = DMA_BIDIRECTIONAL,
1526 };
1527
1528 /* handle vmalloc and linear addresses */
1529 if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1530 return;
1531
1532 if (is_vmalloc_addr(virt))
1533 ref.pfn = vmalloc_to_pfn(virt);
1534 else
1535 ref.pfn = page_to_pfn(virt_to_page(virt));
1536
1537 if (unlikely(dma_debug_disabled()))
1538 return;
1539
1540 check_unmap(&ref);
1541}
1542
1543void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1544 int direction, dma_addr_t dma_addr)
1545{
1546 struct dma_debug_entry *entry;
1547
1548 if (unlikely(dma_debug_disabled()))
1549 return;
1550
1551 entry = dma_entry_alloc();
1552 if (!entry)
1553 return;
1554
1555 entry->type = dma_debug_resource;
1556 entry->dev = dev;
1557 entry->pfn = PHYS_PFN(addr);
1558 entry->offset = offset_in_page(addr);
1559 entry->size = size;
1560 entry->dev_addr = dma_addr;
1561 entry->direction = direction;
1562 entry->map_err_type = MAP_ERR_NOT_CHECKED;
1563
1564 add_dma_entry(entry);
1565}
1566EXPORT_SYMBOL(debug_dma_map_resource);
1567
1568void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1569 size_t size, int direction)
1570{
1571 struct dma_debug_entry ref = {
1572 .type = dma_debug_resource,
1573 .dev = dev,
1574 .dev_addr = dma_addr,
1575 .size = size,
1576 .direction = direction,
1577 };
1578
1579 if (unlikely(dma_debug_disabled()))
1580 return;
1581
1582 check_unmap(&ref);
1583}
1584EXPORT_SYMBOL(debug_dma_unmap_resource);
1585
1586void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1587 size_t size, int direction)
1588{
1589 struct dma_debug_entry ref;
1590
1591 if (unlikely(dma_debug_disabled()))
1592 return;
1593
1594 ref.type = dma_debug_single;
1595 ref.dev = dev;
1596 ref.dev_addr = dma_handle;
1597 ref.size = size;
1598 ref.direction = direction;
1599 ref.sg_call_ents = 0;
1600
1601 check_sync(dev, &ref, true);
1602}
1603EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1604
1605void debug_dma_sync_single_for_device(struct device *dev,
1606 dma_addr_t dma_handle, size_t size,
1607 int direction)
1608{
1609 struct dma_debug_entry ref;
1610
1611 if (unlikely(dma_debug_disabled()))
1612 return;
1613
1614 ref.type = dma_debug_single;
1615 ref.dev = dev;
1616 ref.dev_addr = dma_handle;
1617 ref.size = size;
1618 ref.direction = direction;
1619 ref.sg_call_ents = 0;
1620
1621 check_sync(dev, &ref, false);
1622}
1623EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1624
1625void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1626 int nelems, int direction)
1627{
1628 struct scatterlist *s;
1629 int mapped_ents = 0, i;
1630
1631 if (unlikely(dma_debug_disabled()))
1632 return;
1633
1634 for_each_sg(sg, s, nelems, i) {
1635
1636 struct dma_debug_entry ref = {
1637 .type = dma_debug_sg,
1638 .dev = dev,
1639 .pfn = page_to_pfn(sg_page(s)),
1640 .offset = s->offset,
1641 .dev_addr = sg_dma_address(s),
1642 .size = sg_dma_len(s),
1643 .direction = direction,
1644 .sg_call_ents = nelems,
1645 };
1646
1647 if (!i)
1648 mapped_ents = get_nr_mapped_entries(dev, &ref);
1649
1650 if (i >= mapped_ents)
1651 break;
1652
1653 check_sync(dev, &ref, true);
1654 }
1655}
1656EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1657
1658void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1659 int nelems, int direction)
1660{
1661 struct scatterlist *s;
1662 int mapped_ents = 0, i;
1663
1664 if (unlikely(dma_debug_disabled()))
1665 return;
1666
1667 for_each_sg(sg, s, nelems, i) {
1668
1669 struct dma_debug_entry ref = {
1670 .type = dma_debug_sg,
1671 .dev = dev,
1672 .pfn = page_to_pfn(sg_page(s)),
1673 .offset = s->offset,
1674 .dev_addr = sg_dma_address(s),
1675 .size = sg_dma_len(s),
1676 .direction = direction,
1677 .sg_call_ents = nelems,
1678 };
1679 if (!i)
1680 mapped_ents = get_nr_mapped_entries(dev, &ref);
1681
1682 if (i >= mapped_ents)
1683 break;
1684
1685 check_sync(dev, &ref, false);
1686 }
1687}
1688EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1689
1690static int __init dma_debug_driver_setup(char *str)
1691{
1692 int i;
1693
1694 for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1695 current_driver_name[i] = *str;
1696 if (*str == 0)
1697 break;
1698 }
1699
1700 if (current_driver_name[0])
1701 pr_info("enable driver filter for driver [%s]\n",
1702 current_driver_name);
1703
1704
1705 return 1;
1706}
1707__setup("dma_debug_driver=", dma_debug_driver_setup);