Merge branch 'topic/ste_dma40' into for-linus
[cascardo/linux.git] / lib / dma-debug.c
1 /*
2  * Copyright (C) 2008 Advanced Micro Devices, Inc.
3  *
4  * Author: Joerg Roedel <joerg.roedel@amd.com>
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/scatterlist.h>
21 #include <linux/dma-mapping.h>
22 #include <linux/stacktrace.h>
23 #include <linux/dma-debug.h>
24 #include <linux/spinlock.h>
25 #include <linux/debugfs.h>
26 #include <linux/uaccess.h>
27 #include <linux/export.h>
28 #include <linux/device.h>
29 #include <linux/types.h>
30 #include <linux/sched.h>
31 #include <linux/ctype.h>
32 #include <linux/list.h>
33 #include <linux/slab.h>
34
35 #include <asm/sections.h>
36
37 #define HASH_SIZE       1024ULL
38 #define HASH_FN_SHIFT   13
39 #define HASH_FN_MASK    (HASH_SIZE - 1)
40
41 enum {
42         dma_debug_single,
43         dma_debug_page,
44         dma_debug_sg,
45         dma_debug_coherent,
46         dma_debug_resource,
47 };
48
49 enum map_err_types {
50         MAP_ERR_CHECK_NOT_APPLICABLE,
51         MAP_ERR_NOT_CHECKED,
52         MAP_ERR_CHECKED,
53 };
54
55 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
56
57 /**
58  * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
59  * @list: node on pre-allocated free_entries list
60  * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
61  * @type: single, page, sg, coherent
62  * @pfn: page frame of the start address
63  * @offset: offset of mapping relative to pfn
64  * @size: length of the mapping
65  * @direction: enum dma_data_direction
66  * @sg_call_ents: 'nents' from dma_map_sg
67  * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
68  * @map_err_type: track whether dma_mapping_error() was checked
69  * @stacktrace: support backtraces when a violation is detected
70  */
71 struct dma_debug_entry {
72         struct list_head list;
73         struct device    *dev;
74         int              type;
75         unsigned long    pfn;
76         size_t           offset;
77         u64              dev_addr;
78         u64              size;
79         int              direction;
80         int              sg_call_ents;
81         int              sg_mapped_ents;
82         enum map_err_types  map_err_type;
83 #ifdef CONFIG_STACKTRACE
84         struct           stack_trace stacktrace;
85         unsigned long    st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
86 #endif
87 };
88
89 typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
90
91 struct hash_bucket {
92         struct list_head list;
93         spinlock_t lock;
94 } ____cacheline_aligned_in_smp;
95
96 /* Hash list to save the allocated dma addresses */
97 static struct hash_bucket dma_entry_hash[HASH_SIZE];
98 /* List of pre-allocated dma_debug_entry's */
99 static LIST_HEAD(free_entries);
100 /* Lock for the list above */
101 static DEFINE_SPINLOCK(free_entries_lock);
102
103 /* Global disable flag - will be set in case of an error */
104 static bool global_disable __read_mostly;
105
106 /* Early initialization disable flag, set at the end of dma_debug_init */
107 static bool dma_debug_initialized __read_mostly;
108
109 static inline bool dma_debug_disabled(void)
110 {
111         return global_disable || !dma_debug_initialized;
112 }
113
114 /* Global error count */
115 static u32 error_count;
116
117 /* Global error show enable*/
118 static u32 show_all_errors __read_mostly;
119 /* Number of errors to show */
120 static u32 show_num_errors = 1;
121
122 static u32 num_free_entries;
123 static u32 min_free_entries;
124 static u32 nr_total_entries;
125
126 /* number of preallocated entries requested by kernel cmdline */
127 static u32 req_entries;
128
129 /* debugfs dentry's for the stuff above */
130 static struct dentry *dma_debug_dent        __read_mostly;
131 static struct dentry *global_disable_dent   __read_mostly;
132 static struct dentry *error_count_dent      __read_mostly;
133 static struct dentry *show_all_errors_dent  __read_mostly;
134 static struct dentry *show_num_errors_dent  __read_mostly;
135 static struct dentry *num_free_entries_dent __read_mostly;
136 static struct dentry *min_free_entries_dent __read_mostly;
137 static struct dentry *filter_dent           __read_mostly;
138
139 /* per-driver filter related state */
140
141 #define NAME_MAX_LEN    64
142
143 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
144 static struct device_driver *current_driver                    __read_mostly;
145
146 static DEFINE_RWLOCK(driver_name_lock);
147
148 static const char *const maperr2str[] = {
149         [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
150         [MAP_ERR_NOT_CHECKED] = "dma map error not checked",
151         [MAP_ERR_CHECKED] = "dma map error checked",
152 };
153
154 static const char *type2name[5] = { "single", "page",
155                                     "scather-gather", "coherent",
156                                     "resource" };
157
158 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
159                                    "DMA_FROM_DEVICE", "DMA_NONE" };
160
161 /*
162  * The access to some variables in this macro is racy. We can't use atomic_t
163  * here because all these variables are exported to debugfs. Some of them even
164  * writeable. This is also the reason why a lock won't help much. But anyway,
165  * the races are no big deal. Here is why:
166  *
167  *   error_count: the addition is racy, but the worst thing that can happen is
168  *                that we don't count some errors
169  *   show_num_errors: the subtraction is racy. Also no big deal because in
170  *                    worst case this will result in one warning more in the
171  *                    system log than the user configured. This variable is
172  *                    writeable via debugfs.
173  */
174 static inline void dump_entry_trace(struct dma_debug_entry *entry)
175 {
176 #ifdef CONFIG_STACKTRACE
177         if (entry) {
178                 pr_warning("Mapped at:\n");
179                 print_stack_trace(&entry->stacktrace, 0);
180         }
181 #endif
182 }
183
184 static bool driver_filter(struct device *dev)
185 {
186         struct device_driver *drv;
187         unsigned long flags;
188         bool ret;
189
190         /* driver filter off */
191         if (likely(!current_driver_name[0]))
192                 return true;
193
194         /* driver filter on and initialized */
195         if (current_driver && dev && dev->driver == current_driver)
196                 return true;
197
198         /* driver filter on, but we can't filter on a NULL device... */
199         if (!dev)
200                 return false;
201
202         if (current_driver || !current_driver_name[0])
203                 return false;
204
205         /* driver filter on but not yet initialized */
206         drv = dev->driver;
207         if (!drv)
208                 return false;
209
210         /* lock to protect against change of current_driver_name */
211         read_lock_irqsave(&driver_name_lock, flags);
212
213         ret = false;
214         if (drv->name &&
215             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
216                 current_driver = drv;
217                 ret = true;
218         }
219
220         read_unlock_irqrestore(&driver_name_lock, flags);
221
222         return ret;
223 }
224
225 #define err_printk(dev, entry, format, arg...) do {                     \
226                 error_count += 1;                                       \
227                 if (driver_filter(dev) &&                               \
228                     (show_all_errors || show_num_errors > 0)) {         \
229                         WARN(1, "%s %s: " format,                       \
230                              dev ? dev_driver_string(dev) : "NULL",     \
231                              dev ? dev_name(dev) : "NULL", ## arg);     \
232                         dump_entry_trace(entry);                        \
233                 }                                                       \
234                 if (!show_all_errors && show_num_errors > 0)            \
235                         show_num_errors -= 1;                           \
236         } while (0);
237
238 /*
239  * Hash related functions
240  *
241  * Every DMA-API request is saved into a struct dma_debug_entry. To
242  * have quick access to these structs they are stored into a hash.
243  */
244 static int hash_fn(struct dma_debug_entry *entry)
245 {
246         /*
247          * Hash function is based on the dma address.
248          * We use bits 20-27 here as the index into the hash
249          */
250         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
251 }
252
253 /*
254  * Request exclusive access to a hash bucket for a given dma_debug_entry.
255  */
256 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
257                                            unsigned long *flags)
258         __acquires(&dma_entry_hash[idx].lock)
259 {
260         int idx = hash_fn(entry);
261         unsigned long __flags;
262
263         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
264         *flags = __flags;
265         return &dma_entry_hash[idx];
266 }
267
268 /*
269  * Give up exclusive access to the hash bucket
270  */
271 static void put_hash_bucket(struct hash_bucket *bucket,
272                             unsigned long *flags)
273         __releases(&bucket->lock)
274 {
275         unsigned long __flags = *flags;
276
277         spin_unlock_irqrestore(&bucket->lock, __flags);
278 }
279
280 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
281 {
282         return ((a->dev_addr == b->dev_addr) &&
283                 (a->dev == b->dev)) ? true : false;
284 }
285
286 static bool containing_match(struct dma_debug_entry *a,
287                              struct dma_debug_entry *b)
288 {
289         if (a->dev != b->dev)
290                 return false;
291
292         if ((b->dev_addr <= a->dev_addr) &&
293             ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
294                 return true;
295
296         return false;
297 }
298
299 /*
300  * Search a given entry in the hash bucket list
301  */
302 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
303                                                   struct dma_debug_entry *ref,
304                                                   match_fn match)
305 {
306         struct dma_debug_entry *entry, *ret = NULL;
307         int matches = 0, match_lvl, last_lvl = -1;
308
309         list_for_each_entry(entry, &bucket->list, list) {
310                 if (!match(ref, entry))
311                         continue;
312
313                 /*
314                  * Some drivers map the same physical address multiple
315                  * times. Without a hardware IOMMU this results in the
316                  * same device addresses being put into the dma-debug
317                  * hash multiple times too. This can result in false
318                  * positives being reported. Therefore we implement a
319                  * best-fit algorithm here which returns the entry from
320                  * the hash which fits best to the reference value
321                  * instead of the first-fit.
322                  */
323                 matches += 1;
324                 match_lvl = 0;
325                 entry->size         == ref->size         ? ++match_lvl : 0;
326                 entry->type         == ref->type         ? ++match_lvl : 0;
327                 entry->direction    == ref->direction    ? ++match_lvl : 0;
328                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
329
330                 if (match_lvl == 4) {
331                         /* perfect-fit - return the result */
332                         return entry;
333                 } else if (match_lvl > last_lvl) {
334                         /*
335                          * We found an entry that fits better then the
336                          * previous one or it is the 1st match.
337                          */
338                         last_lvl = match_lvl;
339                         ret      = entry;
340                 }
341         }
342
343         /*
344          * If we have multiple matches but no perfect-fit, just return
345          * NULL.
346          */
347         ret = (matches == 1) ? ret : NULL;
348
349         return ret;
350 }
351
352 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
353                                                  struct dma_debug_entry *ref)
354 {
355         return __hash_bucket_find(bucket, ref, exact_match);
356 }
357
358 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
359                                                    struct dma_debug_entry *ref,
360                                                    unsigned long *flags)
361 {
362
363         unsigned int max_range = dma_get_max_seg_size(ref->dev);
364         struct dma_debug_entry *entry, index = *ref;
365         unsigned int range = 0;
366
367         while (range <= max_range) {
368                 entry = __hash_bucket_find(*bucket, ref, containing_match);
369
370                 if (entry)
371                         return entry;
372
373                 /*
374                  * Nothing found, go back a hash bucket
375                  */
376                 put_hash_bucket(*bucket, flags);
377                 range          += (1 << HASH_FN_SHIFT);
378                 index.dev_addr -= (1 << HASH_FN_SHIFT);
379                 *bucket = get_hash_bucket(&index, flags);
380         }
381
382         return NULL;
383 }
384
385 /*
386  * Add an entry to a hash bucket
387  */
388 static void hash_bucket_add(struct hash_bucket *bucket,
389                             struct dma_debug_entry *entry)
390 {
391         list_add_tail(&entry->list, &bucket->list);
392 }
393
394 /*
395  * Remove entry from a hash bucket list
396  */
397 static void hash_bucket_del(struct dma_debug_entry *entry)
398 {
399         list_del(&entry->list);
400 }
401
402 static unsigned long long phys_addr(struct dma_debug_entry *entry)
403 {
404         if (entry->type == dma_debug_resource)
405                 return __pfn_to_phys(entry->pfn) + entry->offset;
406
407         return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
408 }
409
410 /*
411  * Dump mapping entries for debugging purposes
412  */
413 void debug_dma_dump_mappings(struct device *dev)
414 {
415         int idx;
416
417         for (idx = 0; idx < HASH_SIZE; idx++) {
418                 struct hash_bucket *bucket = &dma_entry_hash[idx];
419                 struct dma_debug_entry *entry;
420                 unsigned long flags;
421
422                 spin_lock_irqsave(&bucket->lock, flags);
423
424                 list_for_each_entry(entry, &bucket->list, list) {
425                         if (!dev || dev == entry->dev) {
426                                 dev_info(entry->dev,
427                                          "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
428                                          type2name[entry->type], idx,
429                                          phys_addr(entry), entry->pfn,
430                                          entry->dev_addr, entry->size,
431                                          dir2name[entry->direction],
432                                          maperr2str[entry->map_err_type]);
433                         }
434                 }
435
436                 spin_unlock_irqrestore(&bucket->lock, flags);
437         }
438 }
439 EXPORT_SYMBOL(debug_dma_dump_mappings);
440
441 /*
442  * For each mapping (initial cacheline in the case of
443  * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
444  * scatterlist, or the cacheline specified in dma_map_single) insert
445  * into this tree using the cacheline as the key. At
446  * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
447  * the entry already exists at insertion time add a tag as a reference
448  * count for the overlapping mappings.  For now, the overlap tracking
449  * just ensures that 'unmaps' balance 'maps' before marking the
450  * cacheline idle, but we should also be flagging overlaps as an API
451  * violation.
452  *
453  * Memory usage is mostly constrained by the maximum number of available
454  * dma-debug entries in that we need a free dma_debug_entry before
455  * inserting into the tree.  In the case of dma_map_page and
456  * dma_alloc_coherent there is only one dma_debug_entry and one
457  * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
458  * other hand, consumes a single dma_debug_entry, but inserts 'nents'
459  * entries into the tree.
460  *
461  * At any time debug_dma_assert_idle() can be called to trigger a
462  * warning if any cachelines in the given page are in the active set.
463  */
464 static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
465 static DEFINE_SPINLOCK(radix_lock);
466 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
467 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
468 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
469
470 static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
471 {
472         return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
473                 (entry->offset >> L1_CACHE_SHIFT);
474 }
475
476 static int active_cacheline_read_overlap(phys_addr_t cln)
477 {
478         int overlap = 0, i;
479
480         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
481                 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
482                         overlap |= 1 << i;
483         return overlap;
484 }
485
486 static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
487 {
488         int i;
489
490         if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
491                 return overlap;
492
493         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
494                 if (overlap & 1 << i)
495                         radix_tree_tag_set(&dma_active_cacheline, cln, i);
496                 else
497                         radix_tree_tag_clear(&dma_active_cacheline, cln, i);
498
499         return overlap;
500 }
501
502 static void active_cacheline_inc_overlap(phys_addr_t cln)
503 {
504         int overlap = active_cacheline_read_overlap(cln);
505
506         overlap = active_cacheline_set_overlap(cln, ++overlap);
507
508         /* If we overflowed the overlap counter then we're potentially
509          * leaking dma-mappings.  Otherwise, if maps and unmaps are
510          * balanced then this overflow may cause false negatives in
511          * debug_dma_assert_idle() as the cacheline may be marked idle
512          * prematurely.
513          */
514         WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
515                   "DMA-API: exceeded %d overlapping mappings of cacheline %pa\n",
516                   ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
517 }
518
519 static int active_cacheline_dec_overlap(phys_addr_t cln)
520 {
521         int overlap = active_cacheline_read_overlap(cln);
522
523         return active_cacheline_set_overlap(cln, --overlap);
524 }
525
526 static int active_cacheline_insert(struct dma_debug_entry *entry)
527 {
528         phys_addr_t cln = to_cacheline_number(entry);
529         unsigned long flags;
530         int rc;
531
532         /* If the device is not writing memory then we don't have any
533          * concerns about the cpu consuming stale data.  This mitigates
534          * legitimate usages of overlapping mappings.
535          */
536         if (entry->direction == DMA_TO_DEVICE)
537                 return 0;
538
539         spin_lock_irqsave(&radix_lock, flags);
540         rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
541         if (rc == -EEXIST)
542                 active_cacheline_inc_overlap(cln);
543         spin_unlock_irqrestore(&radix_lock, flags);
544
545         return rc;
546 }
547
548 static void active_cacheline_remove(struct dma_debug_entry *entry)
549 {
550         phys_addr_t cln = to_cacheline_number(entry);
551         unsigned long flags;
552
553         /* ...mirror the insert case */
554         if (entry->direction == DMA_TO_DEVICE)
555                 return;
556
557         spin_lock_irqsave(&radix_lock, flags);
558         /* since we are counting overlaps the final put of the
559          * cacheline will occur when the overlap count is 0.
560          * active_cacheline_dec_overlap() returns -1 in that case
561          */
562         if (active_cacheline_dec_overlap(cln) < 0)
563                 radix_tree_delete(&dma_active_cacheline, cln);
564         spin_unlock_irqrestore(&radix_lock, flags);
565 }
566
567 /**
568  * debug_dma_assert_idle() - assert that a page is not undergoing dma
569  * @page: page to lookup in the dma_active_cacheline tree
570  *
571  * Place a call to this routine in cases where the cpu touching the page
572  * before the dma completes (page is dma_unmapped) will lead to data
573  * corruption.
574  */
575 void debug_dma_assert_idle(struct page *page)
576 {
577         static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
578         struct dma_debug_entry *entry = NULL;
579         void **results = (void **) &ents;
580         unsigned int nents, i;
581         unsigned long flags;
582         phys_addr_t cln;
583
584         if (dma_debug_disabled())
585                 return;
586
587         if (!page)
588                 return;
589
590         cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
591         spin_lock_irqsave(&radix_lock, flags);
592         nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
593                                        CACHELINES_PER_PAGE);
594         for (i = 0; i < nents; i++) {
595                 phys_addr_t ent_cln = to_cacheline_number(ents[i]);
596
597                 if (ent_cln == cln) {
598                         entry = ents[i];
599                         break;
600                 } else if (ent_cln >= cln + CACHELINES_PER_PAGE)
601                         break;
602         }
603         spin_unlock_irqrestore(&radix_lock, flags);
604
605         if (!entry)
606                 return;
607
608         cln = to_cacheline_number(entry);
609         err_printk(entry->dev, entry,
610                    "DMA-API: cpu touching an active dma mapped cacheline [cln=%pa]\n",
611                    &cln);
612 }
613
614 /*
615  * Wrapper function for adding an entry to the hash.
616  * This function takes care of locking itself.
617  */
618 static void add_dma_entry(struct dma_debug_entry *entry)
619 {
620         struct hash_bucket *bucket;
621         unsigned long flags;
622         int rc;
623
624         bucket = get_hash_bucket(entry, &flags);
625         hash_bucket_add(bucket, entry);
626         put_hash_bucket(bucket, &flags);
627
628         rc = active_cacheline_insert(entry);
629         if (rc == -ENOMEM) {
630                 pr_err("DMA-API: cacheline tracking ENOMEM, dma-debug disabled\n");
631                 global_disable = true;
632         }
633
634         /* TODO: report -EEXIST errors here as overlapping mappings are
635          * not supported by the DMA API
636          */
637 }
638
639 static struct dma_debug_entry *__dma_entry_alloc(void)
640 {
641         struct dma_debug_entry *entry;
642
643         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
644         list_del(&entry->list);
645         memset(entry, 0, sizeof(*entry));
646
647         num_free_entries -= 1;
648         if (num_free_entries < min_free_entries)
649                 min_free_entries = num_free_entries;
650
651         return entry;
652 }
653
654 /* struct dma_entry allocator
655  *
656  * The next two functions implement the allocator for
657  * struct dma_debug_entries.
658  */
659 static struct dma_debug_entry *dma_entry_alloc(void)
660 {
661         struct dma_debug_entry *entry;
662         unsigned long flags;
663
664         spin_lock_irqsave(&free_entries_lock, flags);
665
666         if (list_empty(&free_entries)) {
667                 global_disable = true;
668                 spin_unlock_irqrestore(&free_entries_lock, flags);
669                 pr_err("DMA-API: debugging out of memory - disabling\n");
670                 return NULL;
671         }
672
673         entry = __dma_entry_alloc();
674
675         spin_unlock_irqrestore(&free_entries_lock, flags);
676
677 #ifdef CONFIG_STACKTRACE
678         entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
679         entry->stacktrace.entries = entry->st_entries;
680         entry->stacktrace.skip = 2;
681         save_stack_trace(&entry->stacktrace);
682 #endif
683
684         return entry;
685 }
686
687 static void dma_entry_free(struct dma_debug_entry *entry)
688 {
689         unsigned long flags;
690
691         active_cacheline_remove(entry);
692
693         /*
694          * add to beginning of the list - this way the entries are
695          * more likely cache hot when they are reallocated.
696          */
697         spin_lock_irqsave(&free_entries_lock, flags);
698         list_add(&entry->list, &free_entries);
699         num_free_entries += 1;
700         spin_unlock_irqrestore(&free_entries_lock, flags);
701 }
702
703 int dma_debug_resize_entries(u32 num_entries)
704 {
705         int i, delta, ret = 0;
706         unsigned long flags;
707         struct dma_debug_entry *entry;
708         LIST_HEAD(tmp);
709
710         spin_lock_irqsave(&free_entries_lock, flags);
711
712         if (nr_total_entries < num_entries) {
713                 delta = num_entries - nr_total_entries;
714
715                 spin_unlock_irqrestore(&free_entries_lock, flags);
716
717                 for (i = 0; i < delta; i++) {
718                         entry = kzalloc(sizeof(*entry), GFP_KERNEL);
719                         if (!entry)
720                                 break;
721
722                         list_add_tail(&entry->list, &tmp);
723                 }
724
725                 spin_lock_irqsave(&free_entries_lock, flags);
726
727                 list_splice(&tmp, &free_entries);
728                 nr_total_entries += i;
729                 num_free_entries += i;
730         } else {
731                 delta = nr_total_entries - num_entries;
732
733                 for (i = 0; i < delta && !list_empty(&free_entries); i++) {
734                         entry = __dma_entry_alloc();
735                         kfree(entry);
736                 }
737
738                 nr_total_entries -= i;
739         }
740
741         if (nr_total_entries != num_entries)
742                 ret = 1;
743
744         spin_unlock_irqrestore(&free_entries_lock, flags);
745
746         return ret;
747 }
748 EXPORT_SYMBOL(dma_debug_resize_entries);
749
750 /*
751  * DMA-API debugging init code
752  *
753  * The init code does two things:
754  *   1. Initialize core data structures
755  *   2. Preallocate a given number of dma_debug_entry structs
756  */
757
758 static int prealloc_memory(u32 num_entries)
759 {
760         struct dma_debug_entry *entry, *next_entry;
761         int i;
762
763         for (i = 0; i < num_entries; ++i) {
764                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
765                 if (!entry)
766                         goto out_err;
767
768                 list_add_tail(&entry->list, &free_entries);
769         }
770
771         num_free_entries = num_entries;
772         min_free_entries = num_entries;
773
774         pr_info("DMA-API: preallocated %d debug entries\n", num_entries);
775
776         return 0;
777
778 out_err:
779
780         list_for_each_entry_safe(entry, next_entry, &free_entries, list) {
781                 list_del(&entry->list);
782                 kfree(entry);
783         }
784
785         return -ENOMEM;
786 }
787
788 static ssize_t filter_read(struct file *file, char __user *user_buf,
789                            size_t count, loff_t *ppos)
790 {
791         char buf[NAME_MAX_LEN + 1];
792         unsigned long flags;
793         int len;
794
795         if (!current_driver_name[0])
796                 return 0;
797
798         /*
799          * We can't copy to userspace directly because current_driver_name can
800          * only be read under the driver_name_lock with irqs disabled. So
801          * create a temporary copy first.
802          */
803         read_lock_irqsave(&driver_name_lock, flags);
804         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
805         read_unlock_irqrestore(&driver_name_lock, flags);
806
807         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
808 }
809
810 static ssize_t filter_write(struct file *file, const char __user *userbuf,
811                             size_t count, loff_t *ppos)
812 {
813         char buf[NAME_MAX_LEN];
814         unsigned long flags;
815         size_t len;
816         int i;
817
818         /*
819          * We can't copy from userspace directly. Access to
820          * current_driver_name is protected with a write_lock with irqs
821          * disabled. Since copy_from_user can fault and may sleep we
822          * need to copy to temporary buffer first
823          */
824         len = min(count, (size_t)(NAME_MAX_LEN - 1));
825         if (copy_from_user(buf, userbuf, len))
826                 return -EFAULT;
827
828         buf[len] = 0;
829
830         write_lock_irqsave(&driver_name_lock, flags);
831
832         /*
833          * Now handle the string we got from userspace very carefully.
834          * The rules are:
835          *         - only use the first token we got
836          *         - token delimiter is everything looking like a space
837          *           character (' ', '\n', '\t' ...)
838          *
839          */
840         if (!isalnum(buf[0])) {
841                 /*
842                  * If the first character userspace gave us is not
843                  * alphanumerical then assume the filter should be
844                  * switched off.
845                  */
846                 if (current_driver_name[0])
847                         pr_info("DMA-API: switching off dma-debug driver filter\n");
848                 current_driver_name[0] = 0;
849                 current_driver = NULL;
850                 goto out_unlock;
851         }
852
853         /*
854          * Now parse out the first token and use it as the name for the
855          * driver to filter for.
856          */
857         for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
858                 current_driver_name[i] = buf[i];
859                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
860                         break;
861         }
862         current_driver_name[i] = 0;
863         current_driver = NULL;
864
865         pr_info("DMA-API: enable driver filter for driver [%s]\n",
866                 current_driver_name);
867
868 out_unlock:
869         write_unlock_irqrestore(&driver_name_lock, flags);
870
871         return count;
872 }
873
874 static const struct file_operations filter_fops = {
875         .read  = filter_read,
876         .write = filter_write,
877         .llseek = default_llseek,
878 };
879
880 static int dma_debug_fs_init(void)
881 {
882         dma_debug_dent = debugfs_create_dir("dma-api", NULL);
883         if (!dma_debug_dent) {
884                 pr_err("DMA-API: can not create debugfs directory\n");
885                 return -ENOMEM;
886         }
887
888         global_disable_dent = debugfs_create_bool("disabled", 0444,
889                         dma_debug_dent,
890                         &global_disable);
891         if (!global_disable_dent)
892                 goto out_err;
893
894         error_count_dent = debugfs_create_u32("error_count", 0444,
895                         dma_debug_dent, &error_count);
896         if (!error_count_dent)
897                 goto out_err;
898
899         show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
900                         dma_debug_dent,
901                         &show_all_errors);
902         if (!show_all_errors_dent)
903                 goto out_err;
904
905         show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
906                         dma_debug_dent,
907                         &show_num_errors);
908         if (!show_num_errors_dent)
909                 goto out_err;
910
911         num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
912                         dma_debug_dent,
913                         &num_free_entries);
914         if (!num_free_entries_dent)
915                 goto out_err;
916
917         min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
918                         dma_debug_dent,
919                         &min_free_entries);
920         if (!min_free_entries_dent)
921                 goto out_err;
922
923         filter_dent = debugfs_create_file("driver_filter", 0644,
924                                           dma_debug_dent, NULL, &filter_fops);
925         if (!filter_dent)
926                 goto out_err;
927
928         return 0;
929
930 out_err:
931         debugfs_remove_recursive(dma_debug_dent);
932
933         return -ENOMEM;
934 }
935
936 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
937 {
938         struct dma_debug_entry *entry;
939         unsigned long flags;
940         int count = 0, i;
941
942         local_irq_save(flags);
943
944         for (i = 0; i < HASH_SIZE; ++i) {
945                 spin_lock(&dma_entry_hash[i].lock);
946                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
947                         if (entry->dev == dev) {
948                                 count += 1;
949                                 *out_entry = entry;
950                         }
951                 }
952                 spin_unlock(&dma_entry_hash[i].lock);
953         }
954
955         local_irq_restore(flags);
956
957         return count;
958 }
959
960 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
961 {
962         struct device *dev = data;
963         struct dma_debug_entry *uninitialized_var(entry);
964         int count;
965
966         if (dma_debug_disabled())
967                 return 0;
968
969         switch (action) {
970         case BUS_NOTIFY_UNBOUND_DRIVER:
971                 count = device_dma_allocations(dev, &entry);
972                 if (count == 0)
973                         break;
974                 err_printk(dev, entry, "DMA-API: device driver has pending "
975                                 "DMA allocations while released from device "
976                                 "[count=%d]\n"
977                                 "One of leaked entries details: "
978                                 "[device address=0x%016llx] [size=%llu bytes] "
979                                 "[mapped with %s] [mapped as %s]\n",
980                         count, entry->dev_addr, entry->size,
981                         dir2name[entry->direction], type2name[entry->type]);
982                 break;
983         default:
984                 break;
985         }
986
987         return 0;
988 }
989
990 void dma_debug_add_bus(struct bus_type *bus)
991 {
992         struct notifier_block *nb;
993
994         if (dma_debug_disabled())
995                 return;
996
997         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
998         if (nb == NULL) {
999                 pr_err("dma_debug_add_bus: out of memory\n");
1000                 return;
1001         }
1002
1003         nb->notifier_call = dma_debug_device_change;
1004
1005         bus_register_notifier(bus, nb);
1006 }
1007
1008 /*
1009  * Let the architectures decide how many entries should be preallocated.
1010  */
1011 void dma_debug_init(u32 num_entries)
1012 {
1013         int i;
1014
1015         /* Do not use dma_debug_initialized here, since we really want to be
1016          * called to set dma_debug_initialized
1017          */
1018         if (global_disable)
1019                 return;
1020
1021         for (i = 0; i < HASH_SIZE; ++i) {
1022                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
1023                 spin_lock_init(&dma_entry_hash[i].lock);
1024         }
1025
1026         if (dma_debug_fs_init() != 0) {
1027                 pr_err("DMA-API: error creating debugfs entries - disabling\n");
1028                 global_disable = true;
1029
1030                 return;
1031         }
1032
1033         if (req_entries)
1034                 num_entries = req_entries;
1035
1036         if (prealloc_memory(num_entries) != 0) {
1037                 pr_err("DMA-API: debugging out of memory error - disabled\n");
1038                 global_disable = true;
1039
1040                 return;
1041         }
1042
1043         nr_total_entries = num_free_entries;
1044
1045         dma_debug_initialized = true;
1046
1047         pr_info("DMA-API: debugging enabled by kernel config\n");
1048 }
1049
1050 static __init int dma_debug_cmdline(char *str)
1051 {
1052         if (!str)
1053                 return -EINVAL;
1054
1055         if (strncmp(str, "off", 3) == 0) {
1056                 pr_info("DMA-API: debugging disabled on kernel command line\n");
1057                 global_disable = true;
1058         }
1059
1060         return 0;
1061 }
1062
1063 static __init int dma_debug_entries_cmdline(char *str)
1064 {
1065         int res;
1066
1067         if (!str)
1068                 return -EINVAL;
1069
1070         res = get_option(&str, &req_entries);
1071
1072         if (!res)
1073                 req_entries = 0;
1074
1075         return 0;
1076 }
1077
1078 __setup("dma_debug=", dma_debug_cmdline);
1079 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
1080
1081 static void check_unmap(struct dma_debug_entry *ref)
1082 {
1083         struct dma_debug_entry *entry;
1084         struct hash_bucket *bucket;
1085         unsigned long flags;
1086
1087         bucket = get_hash_bucket(ref, &flags);
1088         entry = bucket_find_exact(bucket, ref);
1089
1090         if (!entry) {
1091                 /* must drop lock before calling dma_mapping_error */
1092                 put_hash_bucket(bucket, &flags);
1093
1094                 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1095                         err_printk(ref->dev, NULL,
1096                                    "DMA-API: device driver tries to free an "
1097                                    "invalid DMA memory address\n");
1098                 } else {
1099                         err_printk(ref->dev, NULL,
1100                                    "DMA-API: device driver tries to free DMA "
1101                                    "memory it has not allocated [device "
1102                                    "address=0x%016llx] [size=%llu bytes]\n",
1103                                    ref->dev_addr, ref->size);
1104                 }
1105                 return;
1106         }
1107
1108         if (ref->size != entry->size) {
1109                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1110                            "DMA memory with different size "
1111                            "[device address=0x%016llx] [map size=%llu bytes] "
1112                            "[unmap size=%llu bytes]\n",
1113                            ref->dev_addr, entry->size, ref->size);
1114         }
1115
1116         if (ref->type != entry->type) {
1117                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1118                            "DMA memory with wrong function "
1119                            "[device address=0x%016llx] [size=%llu bytes] "
1120                            "[mapped as %s] [unmapped as %s]\n",
1121                            ref->dev_addr, ref->size,
1122                            type2name[entry->type], type2name[ref->type]);
1123         } else if ((entry->type == dma_debug_coherent) &&
1124                    (phys_addr(ref) != phys_addr(entry))) {
1125                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1126                            "DMA memory with different CPU address "
1127                            "[device address=0x%016llx] [size=%llu bytes] "
1128                            "[cpu alloc address=0x%016llx] "
1129                            "[cpu free address=0x%016llx]",
1130                            ref->dev_addr, ref->size,
1131                            phys_addr(entry),
1132                            phys_addr(ref));
1133         }
1134
1135         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1136             ref->sg_call_ents != entry->sg_call_ents) {
1137                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1138                            "DMA sg list with different entry count "
1139                            "[map count=%d] [unmap count=%d]\n",
1140                            entry->sg_call_ents, ref->sg_call_ents);
1141         }
1142
1143         /*
1144          * This may be no bug in reality - but most implementations of the
1145          * DMA API don't handle this properly, so check for it here
1146          */
1147         if (ref->direction != entry->direction) {
1148                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1149                            "DMA memory with different direction "
1150                            "[device address=0x%016llx] [size=%llu bytes] "
1151                            "[mapped with %s] [unmapped with %s]\n",
1152                            ref->dev_addr, ref->size,
1153                            dir2name[entry->direction],
1154                            dir2name[ref->direction]);
1155         }
1156
1157         if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1158                 err_printk(ref->dev, entry,
1159                            "DMA-API: device driver failed to check map error"
1160                            "[device address=0x%016llx] [size=%llu bytes] "
1161                            "[mapped as %s]",
1162                            ref->dev_addr, ref->size,
1163                            type2name[entry->type]);
1164         }
1165
1166         hash_bucket_del(entry);
1167         dma_entry_free(entry);
1168
1169         put_hash_bucket(bucket, &flags);
1170 }
1171
1172 static void check_for_stack(struct device *dev, void *addr)
1173 {
1174         if (object_is_on_stack(addr))
1175                 err_printk(dev, NULL, "DMA-API: device driver maps memory from "
1176                                 "stack [addr=%p]\n", addr);
1177 }
1178
1179 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1180 {
1181         unsigned long a1 = (unsigned long)addr;
1182         unsigned long b1 = a1 + len;
1183         unsigned long a2 = (unsigned long)start;
1184         unsigned long b2 = (unsigned long)end;
1185
1186         return !(b1 <= a2 || a1 >= b2);
1187 }
1188
1189 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1190 {
1191         if (overlap(addr, len, _stext, _etext) ||
1192             overlap(addr, len, __start_rodata, __end_rodata))
1193                 err_printk(dev, NULL, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1194 }
1195
1196 static void check_sync(struct device *dev,
1197                        struct dma_debug_entry *ref,
1198                        bool to_cpu)
1199 {
1200         struct dma_debug_entry *entry;
1201         struct hash_bucket *bucket;
1202         unsigned long flags;
1203
1204         bucket = get_hash_bucket(ref, &flags);
1205
1206         entry = bucket_find_contain(&bucket, ref, &flags);
1207
1208         if (!entry) {
1209                 err_printk(dev, NULL, "DMA-API: device driver tries "
1210                                 "to sync DMA memory it has not allocated "
1211                                 "[device address=0x%016llx] [size=%llu bytes]\n",
1212                                 (unsigned long long)ref->dev_addr, ref->size);
1213                 goto out;
1214         }
1215
1216         if (ref->size > entry->size) {
1217                 err_printk(dev, entry, "DMA-API: device driver syncs"
1218                                 " DMA memory outside allocated range "
1219                                 "[device address=0x%016llx] "
1220                                 "[allocation size=%llu bytes] "
1221                                 "[sync offset+size=%llu]\n",
1222                                 entry->dev_addr, entry->size,
1223                                 ref->size);
1224         }
1225
1226         if (entry->direction == DMA_BIDIRECTIONAL)
1227                 goto out;
1228
1229         if (ref->direction != entry->direction) {
1230                 err_printk(dev, entry, "DMA-API: device driver syncs "
1231                                 "DMA memory with different direction "
1232                                 "[device address=0x%016llx] [size=%llu bytes] "
1233                                 "[mapped with %s] [synced with %s]\n",
1234                                 (unsigned long long)ref->dev_addr, entry->size,
1235                                 dir2name[entry->direction],
1236                                 dir2name[ref->direction]);
1237         }
1238
1239         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1240                       !(ref->direction == DMA_TO_DEVICE))
1241                 err_printk(dev, entry, "DMA-API: device driver syncs "
1242                                 "device read-only DMA memory for cpu "
1243                                 "[device address=0x%016llx] [size=%llu bytes] "
1244                                 "[mapped with %s] [synced with %s]\n",
1245                                 (unsigned long long)ref->dev_addr, entry->size,
1246                                 dir2name[entry->direction],
1247                                 dir2name[ref->direction]);
1248
1249         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1250                        !(ref->direction == DMA_FROM_DEVICE))
1251                 err_printk(dev, entry, "DMA-API: device driver syncs "
1252                                 "device write-only DMA memory to device "
1253                                 "[device address=0x%016llx] [size=%llu bytes] "
1254                                 "[mapped with %s] [synced with %s]\n",
1255                                 (unsigned long long)ref->dev_addr, entry->size,
1256                                 dir2name[entry->direction],
1257                                 dir2name[ref->direction]);
1258
1259         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1260             ref->sg_call_ents != entry->sg_call_ents) {
1261                 err_printk(ref->dev, entry, "DMA-API: device driver syncs "
1262                            "DMA sg list with different entry count "
1263                            "[map count=%d] [sync count=%d]\n",
1264                            entry->sg_call_ents, ref->sg_call_ents);
1265         }
1266
1267 out:
1268         put_hash_bucket(bucket, &flags);
1269 }
1270
1271 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1272                         size_t size, int direction, dma_addr_t dma_addr,
1273                         bool map_single)
1274 {
1275         struct dma_debug_entry *entry;
1276
1277         if (unlikely(dma_debug_disabled()))
1278                 return;
1279
1280         if (dma_mapping_error(dev, dma_addr))
1281                 return;
1282
1283         entry = dma_entry_alloc();
1284         if (!entry)
1285                 return;
1286
1287         entry->dev       = dev;
1288         entry->type      = dma_debug_page;
1289         entry->pfn       = page_to_pfn(page);
1290         entry->offset    = offset,
1291         entry->dev_addr  = dma_addr;
1292         entry->size      = size;
1293         entry->direction = direction;
1294         entry->map_err_type = MAP_ERR_NOT_CHECKED;
1295
1296         if (map_single)
1297                 entry->type = dma_debug_single;
1298
1299         if (!PageHighMem(page)) {
1300                 void *addr = page_address(page) + offset;
1301
1302                 check_for_stack(dev, addr);
1303                 check_for_illegal_area(dev, addr, size);
1304         }
1305
1306         add_dma_entry(entry);
1307 }
1308 EXPORT_SYMBOL(debug_dma_map_page);
1309
1310 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1311 {
1312         struct dma_debug_entry ref;
1313         struct dma_debug_entry *entry;
1314         struct hash_bucket *bucket;
1315         unsigned long flags;
1316
1317         if (unlikely(dma_debug_disabled()))
1318                 return;
1319
1320         ref.dev = dev;
1321         ref.dev_addr = dma_addr;
1322         bucket = get_hash_bucket(&ref, &flags);
1323
1324         list_for_each_entry(entry, &bucket->list, list) {
1325                 if (!exact_match(&ref, entry))
1326                         continue;
1327
1328                 /*
1329                  * The same physical address can be mapped multiple
1330                  * times. Without a hardware IOMMU this results in the
1331                  * same device addresses being put into the dma-debug
1332                  * hash multiple times too. This can result in false
1333                  * positives being reported. Therefore we implement a
1334                  * best-fit algorithm here which updates the first entry
1335                  * from the hash which fits the reference value and is
1336                  * not currently listed as being checked.
1337                  */
1338                 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1339                         entry->map_err_type = MAP_ERR_CHECKED;
1340                         break;
1341                 }
1342         }
1343
1344         put_hash_bucket(bucket, &flags);
1345 }
1346 EXPORT_SYMBOL(debug_dma_mapping_error);
1347
1348 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1349                           size_t size, int direction, bool map_single)
1350 {
1351         struct dma_debug_entry ref = {
1352                 .type           = dma_debug_page,
1353                 .dev            = dev,
1354                 .dev_addr       = addr,
1355                 .size           = size,
1356                 .direction      = direction,
1357         };
1358
1359         if (unlikely(dma_debug_disabled()))
1360                 return;
1361
1362         if (map_single)
1363                 ref.type = dma_debug_single;
1364
1365         check_unmap(&ref);
1366 }
1367 EXPORT_SYMBOL(debug_dma_unmap_page);
1368
1369 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1370                       int nents, int mapped_ents, int direction)
1371 {
1372         struct dma_debug_entry *entry;
1373         struct scatterlist *s;
1374         int i;
1375
1376         if (unlikely(dma_debug_disabled()))
1377                 return;
1378
1379         for_each_sg(sg, s, mapped_ents, i) {
1380                 entry = dma_entry_alloc();
1381                 if (!entry)
1382                         return;
1383
1384                 entry->type           = dma_debug_sg;
1385                 entry->dev            = dev;
1386                 entry->pfn            = page_to_pfn(sg_page(s));
1387                 entry->offset         = s->offset,
1388                 entry->size           = sg_dma_len(s);
1389                 entry->dev_addr       = sg_dma_address(s);
1390                 entry->direction      = direction;
1391                 entry->sg_call_ents   = nents;
1392                 entry->sg_mapped_ents = mapped_ents;
1393
1394                 if (!PageHighMem(sg_page(s))) {
1395                         check_for_stack(dev, sg_virt(s));
1396                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1397                 }
1398
1399                 add_dma_entry(entry);
1400         }
1401 }
1402 EXPORT_SYMBOL(debug_dma_map_sg);
1403
1404 static int get_nr_mapped_entries(struct device *dev,
1405                                  struct dma_debug_entry *ref)
1406 {
1407         struct dma_debug_entry *entry;
1408         struct hash_bucket *bucket;
1409         unsigned long flags;
1410         int mapped_ents;
1411
1412         bucket       = get_hash_bucket(ref, &flags);
1413         entry        = bucket_find_exact(bucket, ref);
1414         mapped_ents  = 0;
1415
1416         if (entry)
1417                 mapped_ents = entry->sg_mapped_ents;
1418         put_hash_bucket(bucket, &flags);
1419
1420         return mapped_ents;
1421 }
1422
1423 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1424                         int nelems, int dir)
1425 {
1426         struct scatterlist *s;
1427         int mapped_ents = 0, i;
1428
1429         if (unlikely(dma_debug_disabled()))
1430                 return;
1431
1432         for_each_sg(sglist, s, nelems, i) {
1433
1434                 struct dma_debug_entry ref = {
1435                         .type           = dma_debug_sg,
1436                         .dev            = dev,
1437                         .pfn            = page_to_pfn(sg_page(s)),
1438                         .offset         = s->offset,
1439                         .dev_addr       = sg_dma_address(s),
1440                         .size           = sg_dma_len(s),
1441                         .direction      = dir,
1442                         .sg_call_ents   = nelems,
1443                 };
1444
1445                 if (mapped_ents && i >= mapped_ents)
1446                         break;
1447
1448                 if (!i)
1449                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1450
1451                 check_unmap(&ref);
1452         }
1453 }
1454 EXPORT_SYMBOL(debug_dma_unmap_sg);
1455
1456 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1457                               dma_addr_t dma_addr, void *virt)
1458 {
1459         struct dma_debug_entry *entry;
1460
1461         if (unlikely(dma_debug_disabled()))
1462                 return;
1463
1464         if (unlikely(virt == NULL))
1465                 return;
1466
1467         entry = dma_entry_alloc();
1468         if (!entry)
1469                 return;
1470
1471         entry->type      = dma_debug_coherent;
1472         entry->dev       = dev;
1473         entry->pfn       = page_to_pfn(virt_to_page(virt));
1474         entry->offset    = (size_t) virt & ~PAGE_MASK;
1475         entry->size      = size;
1476         entry->dev_addr  = dma_addr;
1477         entry->direction = DMA_BIDIRECTIONAL;
1478
1479         add_dma_entry(entry);
1480 }
1481 EXPORT_SYMBOL(debug_dma_alloc_coherent);
1482
1483 void debug_dma_free_coherent(struct device *dev, size_t size,
1484                          void *virt, dma_addr_t addr)
1485 {
1486         struct dma_debug_entry ref = {
1487                 .type           = dma_debug_coherent,
1488                 .dev            = dev,
1489                 .pfn            = page_to_pfn(virt_to_page(virt)),
1490                 .offset         = (size_t) virt & ~PAGE_MASK,
1491                 .dev_addr       = addr,
1492                 .size           = size,
1493                 .direction      = DMA_BIDIRECTIONAL,
1494         };
1495
1496         if (unlikely(dma_debug_disabled()))
1497                 return;
1498
1499         check_unmap(&ref);
1500 }
1501 EXPORT_SYMBOL(debug_dma_free_coherent);
1502
1503 void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1504                             int direction, dma_addr_t dma_addr)
1505 {
1506         struct dma_debug_entry *entry;
1507
1508         if (unlikely(dma_debug_disabled()))
1509                 return;
1510
1511         entry = dma_entry_alloc();
1512         if (!entry)
1513                 return;
1514
1515         entry->type             = dma_debug_resource;
1516         entry->dev              = dev;
1517         entry->pfn              = PHYS_PFN(addr);
1518         entry->offset           = offset_in_page(addr);
1519         entry->size             = size;
1520         entry->dev_addr         = dma_addr;
1521         entry->direction        = direction;
1522         entry->map_err_type     = MAP_ERR_NOT_CHECKED;
1523
1524         add_dma_entry(entry);
1525 }
1526 EXPORT_SYMBOL(debug_dma_map_resource);
1527
1528 void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1529                               size_t size, int direction)
1530 {
1531         struct dma_debug_entry ref = {
1532                 .type           = dma_debug_resource,
1533                 .dev            = dev,
1534                 .dev_addr       = dma_addr,
1535                 .size           = size,
1536                 .direction      = direction,
1537         };
1538
1539         if (unlikely(dma_debug_disabled()))
1540                 return;
1541
1542         check_unmap(&ref);
1543 }
1544 EXPORT_SYMBOL(debug_dma_unmap_resource);
1545
1546 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1547                                    size_t size, int direction)
1548 {
1549         struct dma_debug_entry ref;
1550
1551         if (unlikely(dma_debug_disabled()))
1552                 return;
1553
1554         ref.type         = dma_debug_single;
1555         ref.dev          = dev;
1556         ref.dev_addr     = dma_handle;
1557         ref.size         = size;
1558         ref.direction    = direction;
1559         ref.sg_call_ents = 0;
1560
1561         check_sync(dev, &ref, true);
1562 }
1563 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1564
1565 void debug_dma_sync_single_for_device(struct device *dev,
1566                                       dma_addr_t dma_handle, size_t size,
1567                                       int direction)
1568 {
1569         struct dma_debug_entry ref;
1570
1571         if (unlikely(dma_debug_disabled()))
1572                 return;
1573
1574         ref.type         = dma_debug_single;
1575         ref.dev          = dev;
1576         ref.dev_addr     = dma_handle;
1577         ref.size         = size;
1578         ref.direction    = direction;
1579         ref.sg_call_ents = 0;
1580
1581         check_sync(dev, &ref, false);
1582 }
1583 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1584
1585 void debug_dma_sync_single_range_for_cpu(struct device *dev,
1586                                          dma_addr_t dma_handle,
1587                                          unsigned long offset, size_t size,
1588                                          int direction)
1589 {
1590         struct dma_debug_entry ref;
1591
1592         if (unlikely(dma_debug_disabled()))
1593                 return;
1594
1595         ref.type         = dma_debug_single;
1596         ref.dev          = dev;
1597         ref.dev_addr     = dma_handle;
1598         ref.size         = offset + size;
1599         ref.direction    = direction;
1600         ref.sg_call_ents = 0;
1601
1602         check_sync(dev, &ref, true);
1603 }
1604 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
1605
1606 void debug_dma_sync_single_range_for_device(struct device *dev,
1607                                             dma_addr_t dma_handle,
1608                                             unsigned long offset,
1609                                             size_t size, int direction)
1610 {
1611         struct dma_debug_entry ref;
1612
1613         if (unlikely(dma_debug_disabled()))
1614                 return;
1615
1616         ref.type         = dma_debug_single;
1617         ref.dev          = dev;
1618         ref.dev_addr     = dma_handle;
1619         ref.size         = offset + size;
1620         ref.direction    = direction;
1621         ref.sg_call_ents = 0;
1622
1623         check_sync(dev, &ref, false);
1624 }
1625 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
1626
1627 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1628                                int nelems, int direction)
1629 {
1630         struct scatterlist *s;
1631         int mapped_ents = 0, i;
1632
1633         if (unlikely(dma_debug_disabled()))
1634                 return;
1635
1636         for_each_sg(sg, s, nelems, i) {
1637
1638                 struct dma_debug_entry ref = {
1639                         .type           = dma_debug_sg,
1640                         .dev            = dev,
1641                         .pfn            = page_to_pfn(sg_page(s)),
1642                         .offset         = s->offset,
1643                         .dev_addr       = sg_dma_address(s),
1644                         .size           = sg_dma_len(s),
1645                         .direction      = direction,
1646                         .sg_call_ents   = nelems,
1647                 };
1648
1649                 if (!i)
1650                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1651
1652                 if (i >= mapped_ents)
1653                         break;
1654
1655                 check_sync(dev, &ref, true);
1656         }
1657 }
1658 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1659
1660 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1661                                   int nelems, int direction)
1662 {
1663         struct scatterlist *s;
1664         int mapped_ents = 0, i;
1665
1666         if (unlikely(dma_debug_disabled()))
1667                 return;
1668
1669         for_each_sg(sg, s, nelems, i) {
1670
1671                 struct dma_debug_entry ref = {
1672                         .type           = dma_debug_sg,
1673                         .dev            = dev,
1674                         .pfn            = page_to_pfn(sg_page(s)),
1675                         .offset         = s->offset,
1676                         .dev_addr       = sg_dma_address(s),
1677                         .size           = sg_dma_len(s),
1678                         .direction      = direction,
1679                         .sg_call_ents   = nelems,
1680                 };
1681                 if (!i)
1682                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1683
1684                 if (i >= mapped_ents)
1685                         break;
1686
1687                 check_sync(dev, &ref, false);
1688         }
1689 }
1690 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1691
1692 static int __init dma_debug_driver_setup(char *str)
1693 {
1694         int i;
1695
1696         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1697                 current_driver_name[i] = *str;
1698                 if (*str == 0)
1699                         break;
1700         }
1701
1702         if (current_driver_name[0])
1703                 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1704                         current_driver_name);
1705
1706
1707         return 1;
1708 }
1709 __setup("dma_debug_driver=", dma_debug_driver_setup);