GNU Linux-libre 4.9.309-gnu1
[releases.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/vmalloc.h>
26 #include <linux/debugfs.h>
27 #include <linux/uaccess.h>
28 #include <linux/export.h>
29 #include <linux/device.h>
30 #include <linux/types.h>
31 #include <linux/sched.h>
32 #include <linux/ctype.h>
33 #include <linux/list.h>
34 #include <linux/slab.h>
35
36 #include <asm/sections.h>
37
38 #define HASH_SIZE       1024ULL
39 #define HASH_FN_SHIFT   13
40 #define HASH_FN_MASK    (HASH_SIZE - 1)
41
42 enum {
43         dma_debug_single,
44         dma_debug_page,
45         dma_debug_sg,
46         dma_debug_coherent,
47         dma_debug_resource,
48 };
49
50 enum map_err_types {
51         MAP_ERR_CHECK_NOT_APPLICABLE,
52         MAP_ERR_NOT_CHECKED,
53         MAP_ERR_CHECKED,
54 };
55
56 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
57
58 /**
59  * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
60  * @list: node on pre-allocated free_entries list
61  * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
62  * @type: single, page, sg, coherent
63  * @pfn: page frame of the start address
64  * @offset: offset of mapping relative to pfn
65  * @size: length of the mapping
66  * @direction: enum dma_data_direction
67  * @sg_call_ents: 'nents' from dma_map_sg
68  * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
69  * @map_err_type: track whether dma_mapping_error() was checked
70  * @stacktrace: support backtraces when a violation is detected
71  */
72 struct dma_debug_entry {
73         struct list_head list;
74         struct device    *dev;
75         int              type;
76         unsigned long    pfn;
77         size_t           offset;
78         u64              dev_addr;
79         u64              size;
80         int              direction;
81         int              sg_call_ents;
82         int              sg_mapped_ents;
83         enum map_err_types  map_err_type;
84 #ifdef CONFIG_STACKTRACE
85         struct           stack_trace stacktrace;
86         unsigned long    st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
87 #endif
88 };
89
90 typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
91
92 struct hash_bucket {
93         struct list_head list;
94         spinlock_t lock;
95 } ____cacheline_aligned_in_smp;
96
97 /* Hash list to save the allocated dma addresses */
98 static struct hash_bucket dma_entry_hash[HASH_SIZE];
99 /* List of pre-allocated dma_debug_entry's */
100 static LIST_HEAD(free_entries);
101 /* Lock for the list above */
102 static DEFINE_SPINLOCK(free_entries_lock);
103
104 /* Global disable flag - will be set in case of an error */
105 static bool global_disable __read_mostly;
106
107 /* Early initialization disable flag, set at the end of dma_debug_init */
108 static bool dma_debug_initialized __read_mostly;
109
110 static inline bool dma_debug_disabled(void)
111 {
112         return global_disable || !dma_debug_initialized;
113 }
114
115 /* Global error count */
116 static u32 error_count;
117
118 /* Global error show enable*/
119 static u32 show_all_errors __read_mostly;
120 /* Number of errors to show */
121 static u32 show_num_errors = 1;
122
123 static u32 num_free_entries;
124 static u32 min_free_entries;
125 static u32 nr_total_entries;
126
127 /* number of preallocated entries requested by kernel cmdline */
128 static u32 req_entries;
129
130 /* debugfs dentry's for the stuff above */
131 static struct dentry *dma_debug_dent        __read_mostly;
132 static struct dentry *global_disable_dent   __read_mostly;
133 static struct dentry *error_count_dent      __read_mostly;
134 static struct dentry *show_all_errors_dent  __read_mostly;
135 static struct dentry *show_num_errors_dent  __read_mostly;
136 static struct dentry *num_free_entries_dent __read_mostly;
137 static struct dentry *min_free_entries_dent __read_mostly;
138 static struct dentry *filter_dent           __read_mostly;
139
140 /* per-driver filter related state */
141
142 #define NAME_MAX_LEN    64
143
144 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
145 static struct device_driver *current_driver                    __read_mostly;
146
147 static DEFINE_RWLOCK(driver_name_lock);
148
149 static const char *const maperr2str[] = {
150         [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
151         [MAP_ERR_NOT_CHECKED] = "dma map error not checked",
152         [MAP_ERR_CHECKED] = "dma map error checked",
153 };
154
155 static const char *type2name[5] = { "single", "page",
156                                     "scather-gather", "coherent",
157                                     "resource" };
158
159 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
160                                    "DMA_FROM_DEVICE", "DMA_NONE" };
161
162 /*
163  * The access to some variables in this macro is racy. We can't use atomic_t
164  * here because all these variables are exported to debugfs. Some of them even
165  * writeable. This is also the reason why a lock won't help much. But anyway,
166  * the races are no big deal. Here is why:
167  *
168  *   error_count: the addition is racy, but the worst thing that can happen is
169  *                that we don't count some errors
170  *   show_num_errors: the subtraction is racy. Also no big deal because in
171  *                    worst case this will result in one warning more in the
172  *                    system log than the user configured. This variable is
173  *                    writeable via debugfs.
174  */
175 static inline void dump_entry_trace(struct dma_debug_entry *entry)
176 {
177 #ifdef CONFIG_STACKTRACE
178         if (entry) {
179                 pr_warning("Mapped at:\n");
180                 print_stack_trace(&entry->stacktrace, 0);
181         }
182 #endif
183 }
184
185 static bool driver_filter(struct device *dev)
186 {
187         struct device_driver *drv;
188         unsigned long flags;
189         bool ret;
190
191         /* driver filter off */
192         if (likely(!current_driver_name[0]))
193                 return true;
194
195         /* driver filter on and initialized */
196         if (current_driver && dev && dev->driver == current_driver)
197                 return true;
198
199         /* driver filter on, but we can't filter on a NULL device... */
200         if (!dev)
201                 return false;
202
203         if (current_driver || !current_driver_name[0])
204                 return false;
205
206         /* driver filter on but not yet initialized */
207         drv = dev->driver;
208         if (!drv)
209                 return false;
210
211         /* lock to protect against change of current_driver_name */
212         read_lock_irqsave(&driver_name_lock, flags);
213
214         ret = false;
215         if (drv->name &&
216             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
217                 current_driver = drv;
218                 ret = true;
219         }
220
221         read_unlock_irqrestore(&driver_name_lock, flags);
222
223         return ret;
224 }
225
226 #define err_printk(dev, entry, format, arg...) do {                     \
227                 error_count += 1;                                       \
228                 if (driver_filter(dev) &&                               \
229                     (show_all_errors || show_num_errors > 0)) {         \
230                         WARN(1, "%s %s: " format,                       \
231                              dev ? dev_driver_string(dev) : "NULL",     \
232                              dev ? dev_name(dev) : "NULL", ## arg);     \
233                         dump_entry_trace(entry);                        \
234                 }                                                       \
235                 if (!show_all_errors && show_num_errors > 0)            \
236                         show_num_errors -= 1;                           \
237         } while (0);
238
239 /*
240  * Hash related functions
241  *
242  * Every DMA-API request is saved into a struct dma_debug_entry. To
243  * have quick access to these structs they are stored into a hash.
244  */
245 static int hash_fn(struct dma_debug_entry *entry)
246 {
247         /*
248          * Hash function is based on the dma address.
249          * We use bits 20-27 here as the index into the hash
250          */
251         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
252 }
253
254 /*
255  * Request exclusive access to a hash bucket for a given dma_debug_entry.
256  */
257 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
258                                            unsigned long *flags)
259         __acquires(&dma_entry_hash[idx].lock)
260 {
261         int idx = hash_fn(entry);
262         unsigned long __flags;
263
264         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
265         *flags = __flags;
266         return &dma_entry_hash[idx];
267 }
268
269 /*
270  * Give up exclusive access to the hash bucket
271  */
272 static void put_hash_bucket(struct hash_bucket *bucket,
273                             unsigned long *flags)
274         __releases(&bucket->lock)
275 {
276         unsigned long __flags = *flags;
277
278         spin_unlock_irqrestore(&bucket->lock, __flags);
279 }
280
281 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
282 {
283         return ((a->dev_addr == b->dev_addr) &&
284                 (a->dev == b->dev)) ? true : false;
285 }
286
287 static bool containing_match(struct dma_debug_entry *a,
288                              struct dma_debug_entry *b)
289 {
290         if (a->dev != b->dev)
291                 return false;
292
293         if ((b->dev_addr <= a->dev_addr) &&
294             ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
295                 return true;
296
297         return false;
298 }
299
300 /*
301  * Search a given entry in the hash bucket list
302  */
303 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
304                                                   struct dma_debug_entry *ref,
305                                                   match_fn match)
306 {
307         struct dma_debug_entry *entry, *ret = NULL;
308         int matches = 0, match_lvl, last_lvl = -1;
309
310         list_for_each_entry(entry, &bucket->list, list) {
311                 if (!match(ref, entry))
312                         continue;
313
314                 /*
315                  * Some drivers map the same physical address multiple
316                  * times. Without a hardware IOMMU this results in the
317                  * same device addresses being put into the dma-debug
318                  * hash multiple times too. This can result in false
319                  * positives being reported. Therefore we implement a
320                  * best-fit algorithm here which returns the entry from
321                  * the hash which fits best to the reference value
322                  * instead of the first-fit.
323                  */
324                 matches += 1;
325                 match_lvl = 0;
326                 entry->size         == ref->size         ? ++match_lvl : 0;
327                 entry->type         == ref->type         ? ++match_lvl : 0;
328                 entry->direction    == ref->direction    ? ++match_lvl : 0;
329                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
330
331                 if (match_lvl == 4) {
332                         /* perfect-fit - return the result */
333                         return entry;
334                 } else if (match_lvl > last_lvl) {
335                         /*
336                          * We found an entry that fits better then the
337                          * previous one or it is the 1st match.
338                          */
339                         last_lvl = match_lvl;
340                         ret      = entry;
341                 }
342         }
343
344         /*
345          * If we have multiple matches but no perfect-fit, just return
346          * NULL.
347          */
348         ret = (matches == 1) ? ret : NULL;
349
350         return ret;
351 }
352
353 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
354                                                  struct dma_debug_entry *ref)
355 {
356         return __hash_bucket_find(bucket, ref, exact_match);
357 }
358
359 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
360                                                    struct dma_debug_entry *ref,
361                                                    unsigned long *flags)
362 {
363
364         unsigned int max_range = dma_get_max_seg_size(ref->dev);
365         struct dma_debug_entry *entry, index = *ref;
366         unsigned int range = 0;
367
368         while (range <= max_range) {
369                 entry = __hash_bucket_find(*bucket, ref, containing_match);
370
371                 if (entry)
372                         return entry;
373
374                 /*
375                  * Nothing found, go back a hash bucket
376                  */
377                 put_hash_bucket(*bucket, flags);
378                 range          += (1 << HASH_FN_SHIFT);
379                 index.dev_addr -= (1 << HASH_FN_SHIFT);
380                 *bucket = get_hash_bucket(&index, flags);
381         }
382
383         return NULL;
384 }
385
386 /*
387  * Add an entry to a hash bucket
388  */
389 static void hash_bucket_add(struct hash_bucket *bucket,
390                             struct dma_debug_entry *entry)
391 {
392         list_add_tail(&entry->list, &bucket->list);
393 }
394
395 /*
396  * Remove entry from a hash bucket list
397  */
398 static void hash_bucket_del(struct dma_debug_entry *entry)
399 {
400         list_del(&entry->list);
401 }
402
403 static unsigned long long phys_addr(struct dma_debug_entry *entry)
404 {
405         if (entry->type == dma_debug_resource)
406                 return __pfn_to_phys(entry->pfn) + entry->offset;
407
408         return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
409 }
410
411 /*
412  * Dump mapping entries for debugging purposes
413  */
414 void debug_dma_dump_mappings(struct device *dev)
415 {
416         int idx;
417
418         for (idx = 0; idx < HASH_SIZE; idx++) {
419                 struct hash_bucket *bucket = &dma_entry_hash[idx];
420                 struct dma_debug_entry *entry;
421                 unsigned long flags;
422
423                 spin_lock_irqsave(&bucket->lock, flags);
424
425                 list_for_each_entry(entry, &bucket->list, list) {
426                         if (!dev || dev == entry->dev) {
427                                 dev_info(entry->dev,
428                                          "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
429                                          type2name[entry->type], idx,
430                                          phys_addr(entry), entry->pfn,
431                                          entry->dev_addr, entry->size,
432                                          dir2name[entry->direction],
433                                          maperr2str[entry->map_err_type]);
434                         }
435                 }
436
437                 spin_unlock_irqrestore(&bucket->lock, flags);
438                 cond_resched();
439         }
440 }
441 EXPORT_SYMBOL(debug_dma_dump_mappings);
442
443 /*
444  * For each mapping (initial cacheline in the case of
445  * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
446  * scatterlist, or the cacheline specified in dma_map_single) insert
447  * into this tree using the cacheline as the key. At
448  * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
449  * the entry already exists at insertion time add a tag as a reference
450  * count for the overlapping mappings.  For now, the overlap tracking
451  * just ensures that 'unmaps' balance 'maps' before marking the
452  * cacheline idle, but we should also be flagging overlaps as an API
453  * violation.
454  *
455  * Memory usage is mostly constrained by the maximum number of available
456  * dma-debug entries in that we need a free dma_debug_entry before
457  * inserting into the tree.  In the case of dma_map_page and
458  * dma_alloc_coherent there is only one dma_debug_entry and one
459  * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
460  * other hand, consumes a single dma_debug_entry, but inserts 'nents'
461  * entries into the tree.
462  *
463  * At any time debug_dma_assert_idle() can be called to trigger a
464  * warning if any cachelines in the given page are in the active set.
465  */
466 static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
467 static DEFINE_SPINLOCK(radix_lock);
468 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
469 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
470 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
471
472 static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
473 {
474         return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
475                 (entry->offset >> L1_CACHE_SHIFT);
476 }
477
478 static int active_cacheline_read_overlap(phys_addr_t cln)
479 {
480         int overlap = 0, i;
481
482         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
483                 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
484                         overlap |= 1 << i;
485         return overlap;
486 }
487
488 static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
489 {
490         int i;
491
492         if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
493                 return overlap;
494
495         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
496                 if (overlap & 1 << i)
497                         radix_tree_tag_set(&dma_active_cacheline, cln, i);
498                 else
499                         radix_tree_tag_clear(&dma_active_cacheline, cln, i);
500
501         return overlap;
502 }
503
504 static void active_cacheline_inc_overlap(phys_addr_t cln)
505 {
506         int overlap = active_cacheline_read_overlap(cln);
507
508         overlap = active_cacheline_set_overlap(cln, ++overlap);
509
510         /* If we overflowed the overlap counter then we're potentially
511          * leaking dma-mappings.  Otherwise, if maps and unmaps are
512          * balanced then this overflow may cause false negatives in
513          * debug_dma_assert_idle() as the cacheline may be marked idle
514          * prematurely.
515          */
516         WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
517                   "DMA-API: exceeded %d overlapping mappings of cacheline %pa\n",
518                   ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
519 }
520
521 static int active_cacheline_dec_overlap(phys_addr_t cln)
522 {
523         int overlap = active_cacheline_read_overlap(cln);
524
525         return active_cacheline_set_overlap(cln, --overlap);
526 }
527
528 static int active_cacheline_insert(struct dma_debug_entry *entry)
529 {
530         phys_addr_t cln = to_cacheline_number(entry);
531         unsigned long flags;
532         int rc;
533
534         /* If the device is not writing memory then we don't have any
535          * concerns about the cpu consuming stale data.  This mitigates
536          * legitimate usages of overlapping mappings.
537          */
538         if (entry->direction == DMA_TO_DEVICE)
539                 return 0;
540
541         spin_lock_irqsave(&radix_lock, flags);
542         rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
543         if (rc == -EEXIST)
544                 active_cacheline_inc_overlap(cln);
545         spin_unlock_irqrestore(&radix_lock, flags);
546
547         return rc;
548 }
549
550 static void active_cacheline_remove(struct dma_debug_entry *entry)
551 {
552         phys_addr_t cln = to_cacheline_number(entry);
553         unsigned long flags;
554
555         /* ...mirror the insert case */
556         if (entry->direction == DMA_TO_DEVICE)
557                 return;
558
559         spin_lock_irqsave(&radix_lock, flags);
560         /* since we are counting overlaps the final put of the
561          * cacheline will occur when the overlap count is 0.
562          * active_cacheline_dec_overlap() returns -1 in that case
563          */
564         if (active_cacheline_dec_overlap(cln) < 0)
565                 radix_tree_delete(&dma_active_cacheline, cln);
566         spin_unlock_irqrestore(&radix_lock, flags);
567 }
568
569 /**
570  * debug_dma_assert_idle() - assert that a page is not undergoing dma
571  * @page: page to lookup in the dma_active_cacheline tree
572  *
573  * Place a call to this routine in cases where the cpu touching the page
574  * before the dma completes (page is dma_unmapped) will lead to data
575  * corruption.
576  */
577 void debug_dma_assert_idle(struct page *page)
578 {
579         static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
580         struct dma_debug_entry *entry = NULL;
581         void **results = (void **) &ents;
582         unsigned int nents, i;
583         unsigned long flags;
584         phys_addr_t cln;
585
586         if (dma_debug_disabled())
587                 return;
588
589         if (!page)
590                 return;
591
592         cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
593         spin_lock_irqsave(&radix_lock, flags);
594         nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
595                                        CACHELINES_PER_PAGE);
596         for (i = 0; i < nents; i++) {
597                 phys_addr_t ent_cln = to_cacheline_number(ents[i]);
598
599                 if (ent_cln == cln) {
600                         entry = ents[i];
601                         break;
602                 } else if (ent_cln >= cln + CACHELINES_PER_PAGE)
603                         break;
604         }
605         spin_unlock_irqrestore(&radix_lock, flags);
606
607         if (!entry)
608                 return;
609
610         cln = to_cacheline_number(entry);
611         err_printk(entry->dev, entry,
612                    "DMA-API: cpu touching an active dma mapped cacheline [cln=%pa]\n",
613                    &cln);
614 }
615
616 /*
617  * Wrapper function for adding an entry to the hash.
618  * This function takes care of locking itself.
619  */
620 static void add_dma_entry(struct dma_debug_entry *entry)
621 {
622         struct hash_bucket *bucket;
623         unsigned long flags;
624         int rc;
625
626         bucket = get_hash_bucket(entry, &flags);
627         hash_bucket_add(bucket, entry);
628         put_hash_bucket(bucket, &flags);
629
630         rc = active_cacheline_insert(entry);
631         if (rc == -ENOMEM) {
632                 pr_err("DMA-API: cacheline tracking ENOMEM, dma-debug disabled\n");
633                 global_disable = true;
634         }
635
636         /* TODO: report -EEXIST errors here as overlapping mappings are
637          * not supported by the DMA API
638          */
639 }
640
641 static struct dma_debug_entry *__dma_entry_alloc(void)
642 {
643         struct dma_debug_entry *entry;
644
645         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
646         list_del(&entry->list);
647         memset(entry, 0, sizeof(*entry));
648
649         num_free_entries -= 1;
650         if (num_free_entries < min_free_entries)
651                 min_free_entries = num_free_entries;
652
653         return entry;
654 }
655
656 /* struct dma_entry allocator
657  *
658  * The next two functions implement the allocator for
659  * struct dma_debug_entries.
660  */
661 static struct dma_debug_entry *dma_entry_alloc(void)
662 {
663         struct dma_debug_entry *entry;
664         unsigned long flags;
665
666         spin_lock_irqsave(&free_entries_lock, flags);
667
668         if (list_empty(&free_entries)) {
669                 global_disable = true;
670                 spin_unlock_irqrestore(&free_entries_lock, flags);
671                 pr_err("DMA-API: debugging out of memory - disabling\n");
672                 return NULL;
673         }
674
675         entry = __dma_entry_alloc();
676
677         spin_unlock_irqrestore(&free_entries_lock, flags);
678
679 #ifdef CONFIG_STACKTRACE
680         entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
681         entry->stacktrace.entries = entry->st_entries;
682         entry->stacktrace.skip = 2;
683         save_stack_trace(&entry->stacktrace);
684 #endif
685
686         return entry;
687 }
688
689 static void dma_entry_free(struct dma_debug_entry *entry)
690 {
691         unsigned long flags;
692
693         active_cacheline_remove(entry);
694
695         /*
696          * add to beginning of the list - this way the entries are
697          * more likely cache hot when they are reallocated.
698          */
699         spin_lock_irqsave(&free_entries_lock, flags);
700         list_add(&entry->list, &free_entries);
701         num_free_entries += 1;
702         spin_unlock_irqrestore(&free_entries_lock, flags);
703 }
704
705 int dma_debug_resize_entries(u32 num_entries)
706 {
707         int i, delta, ret = 0;
708         unsigned long flags;
709         struct dma_debug_entry *entry;
710         LIST_HEAD(tmp);
711
712         spin_lock_irqsave(&free_entries_lock, flags);
713
714         if (nr_total_entries < num_entries) {
715                 delta = num_entries - nr_total_entries;
716
717                 spin_unlock_irqrestore(&free_entries_lock, flags);
718
719                 for (i = 0; i < delta; i++) {
720                         entry = kzalloc(sizeof(*entry), GFP_KERNEL);
721                         if (!entry)
722                                 break;
723
724                         list_add_tail(&entry->list, &tmp);
725                 }
726
727                 spin_lock_irqsave(&free_entries_lock, flags);
728
729                 list_splice(&tmp, &free_entries);
730                 nr_total_entries += i;
731                 num_free_entries += i;
732         } else {
733                 delta = nr_total_entries - num_entries;
734
735                 for (i = 0; i < delta && !list_empty(&free_entries); i++) {
736                         entry = __dma_entry_alloc();
737                         kfree(entry);
738                 }
739
740                 nr_total_entries -= i;
741         }
742
743         if (nr_total_entries != num_entries)
744                 ret = 1;
745
746         spin_unlock_irqrestore(&free_entries_lock, flags);
747
748         return ret;
749 }
750 EXPORT_SYMBOL(dma_debug_resize_entries);
751
752 /*
753  * DMA-API debugging init code
754  *
755  * The init code does two things:
756  *   1. Initialize core data structures
757  *   2. Preallocate a given number of dma_debug_entry structs
758  */
759
760 static int prealloc_memory(u32 num_entries)
761 {
762         struct dma_debug_entry *entry, *next_entry;
763         int i;
764
765         for (i = 0; i < num_entries; ++i) {
766                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
767                 if (!entry)
768                         goto out_err;
769
770                 list_add_tail(&entry->list, &free_entries);
771         }
772
773         num_free_entries = num_entries;
774         min_free_entries = num_entries;
775
776         pr_info("DMA-API: preallocated %d debug entries\n", num_entries);
777
778         return 0;
779
780 out_err:
781
782         list_for_each_entry_safe(entry, next_entry, &free_entries, list) {
783                 list_del(&entry->list);
784                 kfree(entry);
785         }
786
787         return -ENOMEM;
788 }
789
790 static ssize_t filter_read(struct file *file, char __user *user_buf,
791                            size_t count, loff_t *ppos)
792 {
793         char buf[NAME_MAX_LEN + 1];
794         unsigned long flags;
795         int len;
796
797         if (!current_driver_name[0])
798                 return 0;
799
800         /*
801          * We can't copy to userspace directly because current_driver_name can
802          * only be read under the driver_name_lock with irqs disabled. So
803          * create a temporary copy first.
804          */
805         read_lock_irqsave(&driver_name_lock, flags);
806         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
807         read_unlock_irqrestore(&driver_name_lock, flags);
808
809         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
810 }
811
812 static ssize_t filter_write(struct file *file, const char __user *userbuf,
813                             size_t count, loff_t *ppos)
814 {
815         char buf[NAME_MAX_LEN];
816         unsigned long flags;
817         size_t len;
818         int i;
819
820         /*
821          * We can't copy from userspace directly. Access to
822          * current_driver_name is protected with a write_lock with irqs
823          * disabled. Since copy_from_user can fault and may sleep we
824          * need to copy to temporary buffer first
825          */
826         len = min(count, (size_t)(NAME_MAX_LEN - 1));
827         if (copy_from_user(buf, userbuf, len))
828                 return -EFAULT;
829
830         buf[len] = 0;
831
832         write_lock_irqsave(&driver_name_lock, flags);
833
834         /*
835          * Now handle the string we got from userspace very carefully.
836          * The rules are:
837          *         - only use the first token we got
838          *         - token delimiter is everything looking like a space
839          *           character (' ', '\n', '\t' ...)
840          *
841          */
842         if (!isalnum(buf[0])) {
843                 /*
844                  * If the first character userspace gave us is not
845                  * alphanumerical then assume the filter should be
846                  * switched off.
847                  */
848                 if (current_driver_name[0])
849                         pr_info("DMA-API: switching off dma-debug driver filter\n");
850                 current_driver_name[0] = 0;
851                 current_driver = NULL;
852                 goto out_unlock;
853         }
854
855         /*
856          * Now parse out the first token and use it as the name for the
857          * driver to filter for.
858          */
859         for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
860                 current_driver_name[i] = buf[i];
861                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
862                         break;
863         }
864         current_driver_name[i] = 0;
865         current_driver = NULL;
866
867         pr_info("DMA-API: enable driver filter for driver [%s]\n",
868                 current_driver_name);
869
870 out_unlock:
871         write_unlock_irqrestore(&driver_name_lock, flags);
872
873         return count;
874 }
875
876 static const struct file_operations filter_fops = {
877         .read  = filter_read,
878         .write = filter_write,
879         .llseek = default_llseek,
880 };
881
882 static int dma_debug_fs_init(void)
883 {
884         dma_debug_dent = debugfs_create_dir("dma-api", NULL);
885         if (!dma_debug_dent) {
886                 pr_err("DMA-API: can not create debugfs directory\n");
887                 return -ENOMEM;
888         }
889
890         global_disable_dent = debugfs_create_bool("disabled", 0444,
891                         dma_debug_dent,
892                         &global_disable);
893         if (!global_disable_dent)
894                 goto out_err;
895
896         error_count_dent = debugfs_create_u32("error_count", 0444,
897                         dma_debug_dent, &error_count);
898         if (!error_count_dent)
899                 goto out_err;
900
901         show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
902                         dma_debug_dent,
903                         &show_all_errors);
904         if (!show_all_errors_dent)
905                 goto out_err;
906
907         show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
908                         dma_debug_dent,
909                         &show_num_errors);
910         if (!show_num_errors_dent)
911                 goto out_err;
912
913         num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
914                         dma_debug_dent,
915                         &num_free_entries);
916         if (!num_free_entries_dent)
917                 goto out_err;
918
919         min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
920                         dma_debug_dent,
921                         &min_free_entries);
922         if (!min_free_entries_dent)
923                 goto out_err;
924
925         filter_dent = debugfs_create_file("driver_filter", 0644,
926                                           dma_debug_dent, NULL, &filter_fops);
927         if (!filter_dent)
928                 goto out_err;
929
930         return 0;
931
932 out_err:
933         debugfs_remove_recursive(dma_debug_dent);
934
935         return -ENOMEM;
936 }
937
938 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
939 {
940         struct dma_debug_entry *entry;
941         unsigned long flags;
942         int count = 0, i;
943
944         local_irq_save(flags);
945
946         for (i = 0; i < HASH_SIZE; ++i) {
947                 spin_lock(&dma_entry_hash[i].lock);
948                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
949                         if (entry->dev == dev) {
950                                 count += 1;
951                                 *out_entry = entry;
952                         }
953                 }
954                 spin_unlock(&dma_entry_hash[i].lock);
955         }
956
957         local_irq_restore(flags);
958
959         return count;
960 }
961
962 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
963 {
964         struct device *dev = data;
965         struct dma_debug_entry *uninitialized_var(entry);
966         int count;
967
968         if (dma_debug_disabled())
969                 return 0;
970
971         switch (action) {
972         case BUS_NOTIFY_UNBOUND_DRIVER:
973                 count = device_dma_allocations(dev, &entry);
974                 if (count == 0)
975                         break;
976                 err_printk(dev, entry, "DMA-API: device driver has pending "
977                                 "DMA allocations while released from device "
978                                 "[count=%d]\n"
979                                 "One of leaked entries details: "
980                                 "[device address=0x%016llx] [size=%llu bytes] "
981                                 "[mapped with %s] [mapped as %s]\n",
982                         count, entry->dev_addr, entry->size,
983                         dir2name[entry->direction], type2name[entry->type]);
984                 break;
985         default:
986                 break;
987         }
988
989         return 0;
990 }
991
992 void dma_debug_add_bus(struct bus_type *bus)
993 {
994         struct notifier_block *nb;
995
996         if (dma_debug_disabled())
997                 return;
998
999         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
1000         if (nb == NULL) {
1001                 pr_err("dma_debug_add_bus: out of memory\n");
1002                 return;
1003         }
1004
1005         nb->notifier_call = dma_debug_device_change;
1006
1007         bus_register_notifier(bus, nb);
1008 }
1009
1010 /*
1011  * Let the architectures decide how many entries should be preallocated.
1012  */
1013 void dma_debug_init(u32 num_entries)
1014 {
1015         int i;
1016
1017         /* Do not use dma_debug_initialized here, since we really want to be
1018          * called to set dma_debug_initialized
1019          */
1020         if (global_disable)
1021                 return;
1022
1023         for (i = 0; i < HASH_SIZE; ++i) {
1024                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
1025                 spin_lock_init(&dma_entry_hash[i].lock);
1026         }
1027
1028         if (dma_debug_fs_init() != 0) {
1029                 pr_err("DMA-API: error creating debugfs entries - disabling\n");
1030                 global_disable = true;
1031
1032                 return;
1033         }
1034
1035         if (req_entries)
1036                 num_entries = req_entries;
1037
1038         if (prealloc_memory(num_entries) != 0) {
1039                 pr_err("DMA-API: debugging out of memory error - disabled\n");
1040                 global_disable = true;
1041
1042                 return;
1043         }
1044
1045         nr_total_entries = num_free_entries;
1046
1047         dma_debug_initialized = true;
1048
1049         pr_info("DMA-API: debugging enabled by kernel config\n");
1050 }
1051
1052 static __init int dma_debug_cmdline(char *str)
1053 {
1054         if (!str)
1055                 return -EINVAL;
1056
1057         if (strncmp(str, "off", 3) == 0) {
1058                 pr_info("DMA-API: debugging disabled on kernel command line\n");
1059                 global_disable = true;
1060         }
1061
1062         return 0;
1063 }
1064
1065 static __init int dma_debug_entries_cmdline(char *str)
1066 {
1067         int res;
1068
1069         if (!str)
1070                 return -EINVAL;
1071
1072         res = get_option(&str, &req_entries);
1073
1074         if (!res)
1075                 req_entries = 0;
1076
1077         return 0;
1078 }
1079
1080 __setup("dma_debug=", dma_debug_cmdline);
1081 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
1082
1083 static void check_unmap(struct dma_debug_entry *ref)
1084 {
1085         struct dma_debug_entry *entry;
1086         struct hash_bucket *bucket;
1087         unsigned long flags;
1088
1089         bucket = get_hash_bucket(ref, &flags);
1090         entry = bucket_find_exact(bucket, ref);
1091
1092         if (!entry) {
1093                 /* must drop lock before calling dma_mapping_error */
1094                 put_hash_bucket(bucket, &flags);
1095
1096                 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1097                         err_printk(ref->dev, NULL,
1098                                    "DMA-API: device driver tries to free an "
1099                                    "invalid DMA memory address\n");
1100                 } else {
1101                         err_printk(ref->dev, NULL,
1102                                    "DMA-API: device driver tries to free DMA "
1103                                    "memory it has not allocated [device "
1104                                    "address=0x%016llx] [size=%llu bytes]\n",
1105                                    ref->dev_addr, ref->size);
1106                 }
1107                 return;
1108         }
1109
1110         if (ref->size != entry->size) {
1111                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1112                            "DMA memory with different size "
1113                            "[device address=0x%016llx] [map size=%llu bytes] "
1114                            "[unmap size=%llu bytes]\n",
1115                            ref->dev_addr, entry->size, ref->size);
1116         }
1117
1118         if (ref->type != entry->type) {
1119                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1120                            "DMA memory with wrong function "
1121                            "[device address=0x%016llx] [size=%llu bytes] "
1122                            "[mapped as %s] [unmapped as %s]\n",
1123                            ref->dev_addr, ref->size,
1124                            type2name[entry->type], type2name[ref->type]);
1125         } else if ((entry->type == dma_debug_coherent) &&
1126                    (phys_addr(ref) != phys_addr(entry))) {
1127                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1128                            "DMA memory with different CPU address "
1129                            "[device address=0x%016llx] [size=%llu bytes] "
1130                            "[cpu alloc address=0x%016llx] "
1131                            "[cpu free address=0x%016llx]",
1132                            ref->dev_addr, ref->size,
1133                            phys_addr(entry),
1134                            phys_addr(ref));
1135         }
1136
1137         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1138             ref->sg_call_ents != entry->sg_call_ents) {
1139                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1140                            "DMA sg list with different entry count "
1141                            "[map count=%d] [unmap count=%d]\n",
1142                            entry->sg_call_ents, ref->sg_call_ents);
1143         }
1144
1145         /*
1146          * This may be no bug in reality - but most implementations of the
1147          * DMA API don't handle this properly, so check for it here
1148          */
1149         if (ref->direction != entry->direction) {
1150                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1151                            "DMA memory with different direction "
1152                            "[device address=0x%016llx] [size=%llu bytes] "
1153                            "[mapped with %s] [unmapped with %s]\n",
1154                            ref->dev_addr, ref->size,
1155                            dir2name[entry->direction],
1156                            dir2name[ref->direction]);
1157         }
1158
1159         if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1160                 err_printk(ref->dev, entry,
1161                            "DMA-API: device driver failed to check map error"
1162                            "[device address=0x%016llx] [size=%llu bytes] "
1163                            "[mapped as %s]",
1164                            ref->dev_addr, ref->size,
1165                            type2name[entry->type]);
1166         }
1167
1168         hash_bucket_del(entry);
1169         dma_entry_free(entry);
1170
1171         put_hash_bucket(bucket, &flags);
1172 }
1173
1174 static void check_for_stack(struct device *dev,
1175                             struct page *page, size_t offset)
1176 {
1177         void *addr;
1178         struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1179
1180         if (!stack_vm_area) {
1181                 /* Stack is direct-mapped. */
1182                 if (PageHighMem(page))
1183                         return;
1184                 addr = page_address(page) + offset;
1185                 if (object_is_on_stack(addr))
1186                         err_printk(dev, NULL, "DMA-API: device driver maps memory from stack [addr=%p]\n", addr);
1187         } else {
1188                 /* Stack is vmalloced. */
1189                 int i;
1190
1191                 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1192                         if (page != stack_vm_area->pages[i])
1193                                 continue;
1194
1195                         addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1196                         err_printk(dev, NULL, "DMA-API: device driver maps memory from stack [probable addr=%p]\n", addr);
1197                         break;
1198                 }
1199         }
1200 }
1201
1202 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1203 {
1204         unsigned long a1 = (unsigned long)addr;
1205         unsigned long b1 = a1 + len;
1206         unsigned long a2 = (unsigned long)start;
1207         unsigned long b2 = (unsigned long)end;
1208
1209         return !(b1 <= a2 || a1 >= b2);
1210 }
1211
1212 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1213 {
1214         if (overlap(addr, len, _stext, _etext) ||
1215             overlap(addr, len, __start_rodata, __end_rodata))
1216                 err_printk(dev, NULL, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1217 }
1218
1219 static void check_sync(struct device *dev,
1220                        struct dma_debug_entry *ref,
1221                        bool to_cpu)
1222 {
1223         struct dma_debug_entry *entry;
1224         struct hash_bucket *bucket;
1225         unsigned long flags;
1226
1227         bucket = get_hash_bucket(ref, &flags);
1228
1229         entry = bucket_find_contain(&bucket, ref, &flags);
1230
1231         if (!entry) {
1232                 err_printk(dev, NULL, "DMA-API: device driver tries "
1233                                 "to sync DMA memory it has not allocated "
1234                                 "[device address=0x%016llx] [size=%llu bytes]\n",
1235                                 (unsigned long long)ref->dev_addr, ref->size);
1236                 goto out;
1237         }
1238
1239         if (ref->size > entry->size) {
1240                 err_printk(dev, entry, "DMA-API: device driver syncs"
1241                                 " DMA memory outside allocated range "
1242                                 "[device address=0x%016llx] "
1243                                 "[allocation size=%llu bytes] "
1244                                 "[sync offset+size=%llu]\n",
1245                                 entry->dev_addr, entry->size,
1246                                 ref->size);
1247         }
1248
1249         if (entry->direction == DMA_BIDIRECTIONAL)
1250                 goto out;
1251
1252         if (ref->direction != entry->direction) {
1253                 err_printk(dev, entry, "DMA-API: device driver syncs "
1254                                 "DMA memory with different direction "
1255                                 "[device address=0x%016llx] [size=%llu bytes] "
1256                                 "[mapped with %s] [synced with %s]\n",
1257                                 (unsigned long long)ref->dev_addr, entry->size,
1258                                 dir2name[entry->direction],
1259                                 dir2name[ref->direction]);
1260         }
1261
1262         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1263                       !(ref->direction == DMA_TO_DEVICE))
1264                 err_printk(dev, entry, "DMA-API: device driver syncs "
1265                                 "device read-only DMA memory for cpu "
1266                                 "[device address=0x%016llx] [size=%llu bytes] "
1267                                 "[mapped with %s] [synced with %s]\n",
1268                                 (unsigned long long)ref->dev_addr, entry->size,
1269                                 dir2name[entry->direction],
1270                                 dir2name[ref->direction]);
1271
1272         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1273                        !(ref->direction == DMA_FROM_DEVICE))
1274                 err_printk(dev, entry, "DMA-API: device driver syncs "
1275                                 "device write-only DMA memory to device "
1276                                 "[device address=0x%016llx] [size=%llu bytes] "
1277                                 "[mapped with %s] [synced with %s]\n",
1278                                 (unsigned long long)ref->dev_addr, entry->size,
1279                                 dir2name[entry->direction],
1280                                 dir2name[ref->direction]);
1281
1282         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1283             ref->sg_call_ents != entry->sg_call_ents) {
1284                 err_printk(ref->dev, entry, "DMA-API: device driver syncs "
1285                            "DMA sg list with different entry count "
1286                            "[map count=%d] [sync count=%d]\n",
1287                            entry->sg_call_ents, ref->sg_call_ents);
1288         }
1289
1290 out:
1291         put_hash_bucket(bucket, &flags);
1292 }
1293
1294 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1295                         size_t size, int direction, dma_addr_t dma_addr,
1296                         bool map_single)
1297 {
1298         struct dma_debug_entry *entry;
1299
1300         if (unlikely(dma_debug_disabled()))
1301                 return;
1302
1303         if (dma_mapping_error(dev, dma_addr))
1304                 return;
1305
1306         entry = dma_entry_alloc();
1307         if (!entry)
1308                 return;
1309
1310         entry->dev       = dev;
1311         entry->type      = dma_debug_page;
1312         entry->pfn       = page_to_pfn(page);
1313         entry->offset    = offset,
1314         entry->dev_addr  = dma_addr;
1315         entry->size      = size;
1316         entry->direction = direction;
1317         entry->map_err_type = MAP_ERR_NOT_CHECKED;
1318
1319         if (map_single)
1320                 entry->type = dma_debug_single;
1321
1322         check_for_stack(dev, page, offset);
1323
1324         if (!PageHighMem(page)) {
1325                 void *addr = page_address(page) + offset;
1326
1327                 check_for_illegal_area(dev, addr, size);
1328         }
1329
1330         add_dma_entry(entry);
1331 }
1332 EXPORT_SYMBOL(debug_dma_map_page);
1333
1334 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1335 {
1336         struct dma_debug_entry ref;
1337         struct dma_debug_entry *entry;
1338         struct hash_bucket *bucket;
1339         unsigned long flags;
1340
1341         if (unlikely(dma_debug_disabled()))
1342                 return;
1343
1344         ref.dev = dev;
1345         ref.dev_addr = dma_addr;
1346         bucket = get_hash_bucket(&ref, &flags);
1347
1348         list_for_each_entry(entry, &bucket->list, list) {
1349                 if (!exact_match(&ref, entry))
1350                         continue;
1351
1352                 /*
1353                  * The same physical address can be mapped multiple
1354                  * times. Without a hardware IOMMU this results in the
1355                  * same device addresses being put into the dma-debug
1356                  * hash multiple times too. This can result in false
1357                  * positives being reported. Therefore we implement a
1358                  * best-fit algorithm here which updates the first entry
1359                  * from the hash which fits the reference value and is
1360                  * not currently listed as being checked.
1361                  */
1362                 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1363                         entry->map_err_type = MAP_ERR_CHECKED;
1364                         break;
1365                 }
1366         }
1367
1368         put_hash_bucket(bucket, &flags);
1369 }
1370 EXPORT_SYMBOL(debug_dma_mapping_error);
1371
1372 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1373                           size_t size, int direction, bool map_single)
1374 {
1375         struct dma_debug_entry ref = {
1376                 .type           = dma_debug_page,
1377                 .dev            = dev,
1378                 .dev_addr       = addr,
1379                 .size           = size,
1380                 .direction      = direction,
1381         };
1382
1383         if (unlikely(dma_debug_disabled()))
1384                 return;
1385
1386         if (map_single)
1387                 ref.type = dma_debug_single;
1388
1389         check_unmap(&ref);
1390 }
1391 EXPORT_SYMBOL(debug_dma_unmap_page);
1392
1393 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1394                       int nents, int mapped_ents, int direction)
1395 {
1396         struct dma_debug_entry *entry;
1397         struct scatterlist *s;
1398         int i;
1399
1400         if (unlikely(dma_debug_disabled()))
1401                 return;
1402
1403         for_each_sg(sg, s, mapped_ents, i) {
1404                 entry = dma_entry_alloc();
1405                 if (!entry)
1406                         return;
1407
1408                 entry->type           = dma_debug_sg;
1409                 entry->dev            = dev;
1410                 entry->pfn            = page_to_pfn(sg_page(s));
1411                 entry->offset         = s->offset,
1412                 entry->size           = sg_dma_len(s);
1413                 entry->dev_addr       = sg_dma_address(s);
1414                 entry->direction      = direction;
1415                 entry->sg_call_ents   = nents;
1416                 entry->sg_mapped_ents = mapped_ents;
1417
1418                 check_for_stack(dev, sg_page(s), s->offset);
1419
1420                 if (!PageHighMem(sg_page(s))) {
1421                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1422                 }
1423
1424                 add_dma_entry(entry);
1425         }
1426 }
1427 EXPORT_SYMBOL(debug_dma_map_sg);
1428
1429 static int get_nr_mapped_entries(struct device *dev,
1430                                  struct dma_debug_entry *ref)
1431 {
1432         struct dma_debug_entry *entry;
1433         struct hash_bucket *bucket;
1434         unsigned long flags;
1435         int mapped_ents;
1436
1437         bucket       = get_hash_bucket(ref, &flags);
1438         entry        = bucket_find_exact(bucket, ref);
1439         mapped_ents  = 0;
1440
1441         if (entry)
1442                 mapped_ents = entry->sg_mapped_ents;
1443         put_hash_bucket(bucket, &flags);
1444
1445         return mapped_ents;
1446 }
1447
1448 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1449                         int nelems, int dir)
1450 {
1451         struct scatterlist *s;
1452         int mapped_ents = 0, i;
1453
1454         if (unlikely(dma_debug_disabled()))
1455                 return;
1456
1457         for_each_sg(sglist, s, nelems, i) {
1458
1459                 struct dma_debug_entry ref = {
1460                         .type           = dma_debug_sg,
1461                         .dev            = dev,
1462                         .pfn            = page_to_pfn(sg_page(s)),
1463                         .offset         = s->offset,
1464                         .dev_addr       = sg_dma_address(s),
1465                         .size           = sg_dma_len(s),
1466                         .direction      = dir,
1467                         .sg_call_ents   = nelems,
1468                 };
1469
1470                 if (mapped_ents && i >= mapped_ents)
1471                         break;
1472
1473                 if (!i)
1474                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1475
1476                 check_unmap(&ref);
1477         }
1478 }
1479 EXPORT_SYMBOL(debug_dma_unmap_sg);
1480
1481 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1482                               dma_addr_t dma_addr, void *virt)
1483 {
1484         struct dma_debug_entry *entry;
1485
1486         if (unlikely(dma_debug_disabled()))
1487                 return;
1488
1489         if (unlikely(virt == NULL))
1490                 return;
1491
1492         entry = dma_entry_alloc();
1493         if (!entry)
1494                 return;
1495
1496         entry->type      = dma_debug_coherent;
1497         entry->dev       = dev;
1498         entry->pfn       = page_to_pfn(virt_to_page(virt));
1499         entry->offset    = (size_t) virt & ~PAGE_MASK;
1500         entry->size      = size;
1501         entry->dev_addr  = dma_addr;
1502         entry->direction = DMA_BIDIRECTIONAL;
1503
1504         add_dma_entry(entry);
1505 }
1506 EXPORT_SYMBOL(debug_dma_alloc_coherent);
1507
1508 void debug_dma_free_coherent(struct device *dev, size_t size,
1509                          void *virt, dma_addr_t addr)
1510 {
1511         struct dma_debug_entry ref = {
1512                 .type           = dma_debug_coherent,
1513                 .dev            = dev,
1514                 .pfn            = page_to_pfn(virt_to_page(virt)),
1515                 .offset         = (size_t) virt & ~PAGE_MASK,
1516                 .dev_addr       = addr,
1517                 .size           = size,
1518                 .direction      = DMA_BIDIRECTIONAL,
1519         };
1520
1521         if (unlikely(dma_debug_disabled()))
1522                 return;
1523
1524         check_unmap(&ref);
1525 }
1526 EXPORT_SYMBOL(debug_dma_free_coherent);
1527
1528 void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1529                             int direction, dma_addr_t dma_addr)
1530 {
1531         struct dma_debug_entry *entry;
1532
1533         if (unlikely(dma_debug_disabled()))
1534                 return;
1535
1536         entry = dma_entry_alloc();
1537         if (!entry)
1538                 return;
1539
1540         entry->type             = dma_debug_resource;
1541         entry->dev              = dev;
1542         entry->pfn              = PHYS_PFN(addr);
1543         entry->offset           = offset_in_page(addr);
1544         entry->size             = size;
1545         entry->dev_addr         = dma_addr;
1546         entry->direction        = direction;
1547         entry->map_err_type     = MAP_ERR_NOT_CHECKED;
1548
1549         add_dma_entry(entry);
1550 }
1551 EXPORT_SYMBOL(debug_dma_map_resource);
1552
1553 void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1554                               size_t size, int direction)
1555 {
1556         struct dma_debug_entry ref = {
1557                 .type           = dma_debug_resource,
1558                 .dev            = dev,
1559                 .dev_addr       = dma_addr,
1560                 .size           = size,
1561                 .direction      = direction,
1562         };
1563
1564         if (unlikely(dma_debug_disabled()))
1565                 return;
1566
1567         check_unmap(&ref);
1568 }
1569 EXPORT_SYMBOL(debug_dma_unmap_resource);
1570
1571 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1572                                    size_t size, int direction)
1573 {
1574         struct dma_debug_entry ref;
1575
1576         if (unlikely(dma_debug_disabled()))
1577                 return;
1578
1579         ref.type         = dma_debug_single;
1580         ref.dev          = dev;
1581         ref.dev_addr     = dma_handle;
1582         ref.size         = size;
1583         ref.direction    = direction;
1584         ref.sg_call_ents = 0;
1585
1586         check_sync(dev, &ref, true);
1587 }
1588 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1589
1590 void debug_dma_sync_single_for_device(struct device *dev,
1591                                       dma_addr_t dma_handle, size_t size,
1592                                       int direction)
1593 {
1594         struct dma_debug_entry ref;
1595
1596         if (unlikely(dma_debug_disabled()))
1597                 return;
1598
1599         ref.type         = dma_debug_single;
1600         ref.dev          = dev;
1601         ref.dev_addr     = dma_handle;
1602         ref.size         = size;
1603         ref.direction    = direction;
1604         ref.sg_call_ents = 0;
1605
1606         check_sync(dev, &ref, false);
1607 }
1608 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1609
1610 void debug_dma_sync_single_range_for_cpu(struct device *dev,
1611                                          dma_addr_t dma_handle,
1612                                          unsigned long offset, size_t size,
1613                                          int direction)
1614 {
1615         struct dma_debug_entry ref;
1616
1617         if (unlikely(dma_debug_disabled()))
1618                 return;
1619
1620         ref.type         = dma_debug_single;
1621         ref.dev          = dev;
1622         ref.dev_addr     = dma_handle;
1623         ref.size         = offset + size;
1624         ref.direction    = direction;
1625         ref.sg_call_ents = 0;
1626
1627         check_sync(dev, &ref, true);
1628 }
1629 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
1630
1631 void debug_dma_sync_single_range_for_device(struct device *dev,
1632                                             dma_addr_t dma_handle,
1633                                             unsigned long offset,
1634                                             size_t size, int direction)
1635 {
1636         struct dma_debug_entry ref;
1637
1638         if (unlikely(dma_debug_disabled()))
1639                 return;
1640
1641         ref.type         = dma_debug_single;
1642         ref.dev          = dev;
1643         ref.dev_addr     = dma_handle;
1644         ref.size         = offset + size;
1645         ref.direction    = direction;
1646         ref.sg_call_ents = 0;
1647
1648         check_sync(dev, &ref, false);
1649 }
1650 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
1651
1652 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1653                                int nelems, int direction)
1654 {
1655         struct scatterlist *s;
1656         int mapped_ents = 0, i;
1657
1658         if (unlikely(dma_debug_disabled()))
1659                 return;
1660
1661         for_each_sg(sg, s, nelems, i) {
1662
1663                 struct dma_debug_entry ref = {
1664                         .type           = dma_debug_sg,
1665                         .dev            = dev,
1666                         .pfn            = page_to_pfn(sg_page(s)),
1667                         .offset         = s->offset,
1668                         .dev_addr       = sg_dma_address(s),
1669                         .size           = sg_dma_len(s),
1670                         .direction      = direction,
1671                         .sg_call_ents   = nelems,
1672                 };
1673
1674                 if (!i)
1675                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1676
1677                 if (i >= mapped_ents)
1678                         break;
1679
1680                 check_sync(dev, &ref, true);
1681         }
1682 }
1683 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1684
1685 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1686                                   int nelems, int direction)
1687 {
1688         struct scatterlist *s;
1689         int mapped_ents = 0, i;
1690
1691         if (unlikely(dma_debug_disabled()))
1692                 return;
1693
1694         for_each_sg(sg, s, nelems, i) {
1695
1696                 struct dma_debug_entry ref = {
1697                         .type           = dma_debug_sg,
1698                         .dev            = dev,
1699                         .pfn            = page_to_pfn(sg_page(s)),
1700                         .offset         = s->offset,
1701                         .dev_addr       = sg_dma_address(s),
1702                         .size           = sg_dma_len(s),
1703                         .direction      = direction,
1704                         .sg_call_ents   = nelems,
1705                 };
1706                 if (!i)
1707                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1708
1709                 if (i >= mapped_ents)
1710                         break;
1711
1712                 check_sync(dev, &ref, false);
1713         }
1714 }
1715 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1716
1717 static int __init dma_debug_driver_setup(char *str)
1718 {
1719         int i;
1720
1721         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1722                 current_driver_name[i] = *str;
1723                 if (*str == 0)
1724                         break;
1725         }
1726
1727         if (current_driver_name[0])
1728                 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1729                         current_driver_name);
1730
1731
1732         return 1;
1733 }
1734 __setup("dma_debug_driver=", dma_debug_driver_setup);