sched/headers: Prepare for new header dependencies before moving code to <linux/sched...
[platform/kernel/linux-exynos.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/sched/task.h>
23 #include <linux/stacktrace.h>
24 #include <linux/dma-debug.h>
25 #include <linux/spinlock.h>
26 #include <linux/vmalloc.h>
27 #include <linux/debugfs.h>
28 #include <linux/uaccess.h>
29 #include <linux/export.h>
30 #include <linux/device.h>
31 #include <linux/types.h>
32 #include <linux/sched.h>
33 #include <linux/ctype.h>
34 #include <linux/list.h>
35 #include <linux/slab.h>
36
37 #include <asm/sections.h>
38
39 #define HASH_SIZE       1024ULL
40 #define HASH_FN_SHIFT   13
41 #define HASH_FN_MASK    (HASH_SIZE - 1)
42
43 enum {
44         dma_debug_single,
45         dma_debug_page,
46         dma_debug_sg,
47         dma_debug_coherent,
48         dma_debug_resource,
49 };
50
51 enum map_err_types {
52         MAP_ERR_CHECK_NOT_APPLICABLE,
53         MAP_ERR_NOT_CHECKED,
54         MAP_ERR_CHECKED,
55 };
56
57 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
58
59 /**
60  * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
61  * @list: node on pre-allocated free_entries list
62  * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
63  * @type: single, page, sg, coherent
64  * @pfn: page frame of the start address
65  * @offset: offset of mapping relative to pfn
66  * @size: length of the mapping
67  * @direction: enum dma_data_direction
68  * @sg_call_ents: 'nents' from dma_map_sg
69  * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
70  * @map_err_type: track whether dma_mapping_error() was checked
71  * @stacktrace: support backtraces when a violation is detected
72  */
73 struct dma_debug_entry {
74         struct list_head list;
75         struct device    *dev;
76         int              type;
77         unsigned long    pfn;
78         size_t           offset;
79         u64              dev_addr;
80         u64              size;
81         int              direction;
82         int              sg_call_ents;
83         int              sg_mapped_ents;
84         enum map_err_types  map_err_type;
85 #ifdef CONFIG_STACKTRACE
86         struct           stack_trace stacktrace;
87         unsigned long    st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
88 #endif
89 };
90
91 typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
92
93 struct hash_bucket {
94         struct list_head list;
95         spinlock_t lock;
96 } ____cacheline_aligned_in_smp;
97
98 /* Hash list to save the allocated dma addresses */
99 static struct hash_bucket dma_entry_hash[HASH_SIZE];
100 /* List of pre-allocated dma_debug_entry's */
101 static LIST_HEAD(free_entries);
102 /* Lock for the list above */
103 static DEFINE_SPINLOCK(free_entries_lock);
104
105 /* Global disable flag - will be set in case of an error */
106 static bool global_disable __read_mostly;
107
108 /* Early initialization disable flag, set at the end of dma_debug_init */
109 static bool dma_debug_initialized __read_mostly;
110
111 static inline bool dma_debug_disabled(void)
112 {
113         return global_disable || !dma_debug_initialized;
114 }
115
116 /* Global error count */
117 static u32 error_count;
118
119 /* Global error show enable*/
120 static u32 show_all_errors __read_mostly;
121 /* Number of errors to show */
122 static u32 show_num_errors = 1;
123
124 static u32 num_free_entries;
125 static u32 min_free_entries;
126 static u32 nr_total_entries;
127
128 /* number of preallocated entries requested by kernel cmdline */
129 static u32 req_entries;
130
131 /* debugfs dentry's for the stuff above */
132 static struct dentry *dma_debug_dent        __read_mostly;
133 static struct dentry *global_disable_dent   __read_mostly;
134 static struct dentry *error_count_dent      __read_mostly;
135 static struct dentry *show_all_errors_dent  __read_mostly;
136 static struct dentry *show_num_errors_dent  __read_mostly;
137 static struct dentry *num_free_entries_dent __read_mostly;
138 static struct dentry *min_free_entries_dent __read_mostly;
139 static struct dentry *filter_dent           __read_mostly;
140
141 /* per-driver filter related state */
142
143 #define NAME_MAX_LEN    64
144
145 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
146 static struct device_driver *current_driver                    __read_mostly;
147
148 static DEFINE_RWLOCK(driver_name_lock);
149
150 static const char *const maperr2str[] = {
151         [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
152         [MAP_ERR_NOT_CHECKED] = "dma map error not checked",
153         [MAP_ERR_CHECKED] = "dma map error checked",
154 };
155
156 static const char *type2name[5] = { "single", "page",
157                                     "scather-gather", "coherent",
158                                     "resource" };
159
160 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
161                                    "DMA_FROM_DEVICE", "DMA_NONE" };
162
163 /*
164  * The access to some variables in this macro is racy. We can't use atomic_t
165  * here because all these variables are exported to debugfs. Some of them even
166  * writeable. This is also the reason why a lock won't help much. But anyway,
167  * the races are no big deal. Here is why:
168  *
169  *   error_count: the addition is racy, but the worst thing that can happen is
170  *                that we don't count some errors
171  *   show_num_errors: the subtraction is racy. Also no big deal because in
172  *                    worst case this will result in one warning more in the
173  *                    system log than the user configured. This variable is
174  *                    writeable via debugfs.
175  */
176 static inline void dump_entry_trace(struct dma_debug_entry *entry)
177 {
178 #ifdef CONFIG_STACKTRACE
179         if (entry) {
180                 pr_warning("Mapped at:\n");
181                 print_stack_trace(&entry->stacktrace, 0);
182         }
183 #endif
184 }
185
186 static bool driver_filter(struct device *dev)
187 {
188         struct device_driver *drv;
189         unsigned long flags;
190         bool ret;
191
192         /* driver filter off */
193         if (likely(!current_driver_name[0]))
194                 return true;
195
196         /* driver filter on and initialized */
197         if (current_driver && dev && dev->driver == current_driver)
198                 return true;
199
200         /* driver filter on, but we can't filter on a NULL device... */
201         if (!dev)
202                 return false;
203
204         if (current_driver || !current_driver_name[0])
205                 return false;
206
207         /* driver filter on but not yet initialized */
208         drv = dev->driver;
209         if (!drv)
210                 return false;
211
212         /* lock to protect against change of current_driver_name */
213         read_lock_irqsave(&driver_name_lock, flags);
214
215         ret = false;
216         if (drv->name &&
217             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
218                 current_driver = drv;
219                 ret = true;
220         }
221
222         read_unlock_irqrestore(&driver_name_lock, flags);
223
224         return ret;
225 }
226
227 #define err_printk(dev, entry, format, arg...) do {                     \
228                 error_count += 1;                                       \
229                 if (driver_filter(dev) &&                               \
230                     (show_all_errors || show_num_errors > 0)) {         \
231                         WARN(1, "%s %s: " format,                       \
232                              dev ? dev_driver_string(dev) : "NULL",     \
233                              dev ? dev_name(dev) : "NULL", ## arg);     \
234                         dump_entry_trace(entry);                        \
235                 }                                                       \
236                 if (!show_all_errors && show_num_errors > 0)            \
237                         show_num_errors -= 1;                           \
238         } while (0);
239
240 /*
241  * Hash related functions
242  *
243  * Every DMA-API request is saved into a struct dma_debug_entry. To
244  * have quick access to these structs they are stored into a hash.
245  */
246 static int hash_fn(struct dma_debug_entry *entry)
247 {
248         /*
249          * Hash function is based on the dma address.
250          * We use bits 20-27 here as the index into the hash
251          */
252         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
253 }
254
255 /*
256  * Request exclusive access to a hash bucket for a given dma_debug_entry.
257  */
258 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
259                                            unsigned long *flags)
260         __acquires(&dma_entry_hash[idx].lock)
261 {
262         int idx = hash_fn(entry);
263         unsigned long __flags;
264
265         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
266         *flags = __flags;
267         return &dma_entry_hash[idx];
268 }
269
270 /*
271  * Give up exclusive access to the hash bucket
272  */
273 static void put_hash_bucket(struct hash_bucket *bucket,
274                             unsigned long *flags)
275         __releases(&bucket->lock)
276 {
277         unsigned long __flags = *flags;
278
279         spin_unlock_irqrestore(&bucket->lock, __flags);
280 }
281
282 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
283 {
284         return ((a->dev_addr == b->dev_addr) &&
285                 (a->dev == b->dev)) ? true : false;
286 }
287
288 static bool containing_match(struct dma_debug_entry *a,
289                              struct dma_debug_entry *b)
290 {
291         if (a->dev != b->dev)
292                 return false;
293
294         if ((b->dev_addr <= a->dev_addr) &&
295             ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
296                 return true;
297
298         return false;
299 }
300
301 /*
302  * Search a given entry in the hash bucket list
303  */
304 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
305                                                   struct dma_debug_entry *ref,
306                                                   match_fn match)
307 {
308         struct dma_debug_entry *entry, *ret = NULL;
309         int matches = 0, match_lvl, last_lvl = -1;
310
311         list_for_each_entry(entry, &bucket->list, list) {
312                 if (!match(ref, entry))
313                         continue;
314
315                 /*
316                  * Some drivers map the same physical address multiple
317                  * times. Without a hardware IOMMU this results in the
318                  * same device addresses being put into the dma-debug
319                  * hash multiple times too. This can result in false
320                  * positives being reported. Therefore we implement a
321                  * best-fit algorithm here which returns the entry from
322                  * the hash which fits best to the reference value
323                  * instead of the first-fit.
324                  */
325                 matches += 1;
326                 match_lvl = 0;
327                 entry->size         == ref->size         ? ++match_lvl : 0;
328                 entry->type         == ref->type         ? ++match_lvl : 0;
329                 entry->direction    == ref->direction    ? ++match_lvl : 0;
330                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
331
332                 if (match_lvl == 4) {
333                         /* perfect-fit - return the result */
334                         return entry;
335                 } else if (match_lvl > last_lvl) {
336                         /*
337                          * We found an entry that fits better then the
338                          * previous one or it is the 1st match.
339                          */
340                         last_lvl = match_lvl;
341                         ret      = entry;
342                 }
343         }
344
345         /*
346          * If we have multiple matches but no perfect-fit, just return
347          * NULL.
348          */
349         ret = (matches == 1) ? ret : NULL;
350
351         return ret;
352 }
353
354 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
355                                                  struct dma_debug_entry *ref)
356 {
357         return __hash_bucket_find(bucket, ref, exact_match);
358 }
359
360 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
361                                                    struct dma_debug_entry *ref,
362                                                    unsigned long *flags)
363 {
364
365         unsigned int max_range = dma_get_max_seg_size(ref->dev);
366         struct dma_debug_entry *entry, index = *ref;
367         unsigned int range = 0;
368
369         while (range <= max_range) {
370                 entry = __hash_bucket_find(*bucket, ref, containing_match);
371
372                 if (entry)
373                         return entry;
374
375                 /*
376                  * Nothing found, go back a hash bucket
377                  */
378                 put_hash_bucket(*bucket, flags);
379                 range          += (1 << HASH_FN_SHIFT);
380                 index.dev_addr -= (1 << HASH_FN_SHIFT);
381                 *bucket = get_hash_bucket(&index, flags);
382         }
383
384         return NULL;
385 }
386
387 /*
388  * Add an entry to a hash bucket
389  */
390 static void hash_bucket_add(struct hash_bucket *bucket,
391                             struct dma_debug_entry *entry)
392 {
393         list_add_tail(&entry->list, &bucket->list);
394 }
395
396 /*
397  * Remove entry from a hash bucket list
398  */
399 static void hash_bucket_del(struct dma_debug_entry *entry)
400 {
401         list_del(&entry->list);
402 }
403
404 static unsigned long long phys_addr(struct dma_debug_entry *entry)
405 {
406         if (entry->type == dma_debug_resource)
407                 return __pfn_to_phys(entry->pfn) + entry->offset;
408
409         return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
410 }
411
412 /*
413  * Dump mapping entries for debugging purposes
414  */
415 void debug_dma_dump_mappings(struct device *dev)
416 {
417         int idx;
418
419         for (idx = 0; idx < HASH_SIZE; idx++) {
420                 struct hash_bucket *bucket = &dma_entry_hash[idx];
421                 struct dma_debug_entry *entry;
422                 unsigned long flags;
423
424                 spin_lock_irqsave(&bucket->lock, flags);
425
426                 list_for_each_entry(entry, &bucket->list, list) {
427                         if (!dev || dev == entry->dev) {
428                                 dev_info(entry->dev,
429                                          "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
430                                          type2name[entry->type], idx,
431                                          phys_addr(entry), entry->pfn,
432                                          entry->dev_addr, entry->size,
433                                          dir2name[entry->direction],
434                                          maperr2str[entry->map_err_type]);
435                         }
436                 }
437
438                 spin_unlock_irqrestore(&bucket->lock, flags);
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         /*
1160          * Drivers should use dma_mapping_error() to check the returned
1161          * addresses of dma_map_single() and dma_map_page().
1162          * If not, print this warning message. See Documentation/DMA-API.txt.
1163          */
1164         if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1165                 err_printk(ref->dev, entry,
1166                            "DMA-API: device driver failed to check map error"
1167                            "[device address=0x%016llx] [size=%llu bytes] "
1168                            "[mapped as %s]",
1169                            ref->dev_addr, ref->size,
1170                            type2name[entry->type]);
1171         }
1172
1173         hash_bucket_del(entry);
1174         dma_entry_free(entry);
1175
1176         put_hash_bucket(bucket, &flags);
1177 }
1178
1179 static void check_for_stack(struct device *dev,
1180                             struct page *page, size_t offset)
1181 {
1182         void *addr;
1183         struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1184
1185         if (!stack_vm_area) {
1186                 /* Stack is direct-mapped. */
1187                 if (PageHighMem(page))
1188                         return;
1189                 addr = page_address(page) + offset;
1190                 if (object_is_on_stack(addr))
1191                         err_printk(dev, NULL, "DMA-API: device driver maps memory from stack [addr=%p]\n", addr);
1192         } else {
1193                 /* Stack is vmalloced. */
1194                 int i;
1195
1196                 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1197                         if (page != stack_vm_area->pages[i])
1198                                 continue;
1199
1200                         addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1201                         err_printk(dev, NULL, "DMA-API: device driver maps memory from stack [probable addr=%p]\n", addr);
1202                         break;
1203                 }
1204         }
1205 }
1206
1207 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1208 {
1209         unsigned long a1 = (unsigned long)addr;
1210         unsigned long b1 = a1 + len;
1211         unsigned long a2 = (unsigned long)start;
1212         unsigned long b2 = (unsigned long)end;
1213
1214         return !(b1 <= a2 || a1 >= b2);
1215 }
1216
1217 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1218 {
1219         if (overlap(addr, len, _stext, _etext) ||
1220             overlap(addr, len, __start_rodata, __end_rodata))
1221                 err_printk(dev, NULL, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1222 }
1223
1224 static void check_sync(struct device *dev,
1225                        struct dma_debug_entry *ref,
1226                        bool to_cpu)
1227 {
1228         struct dma_debug_entry *entry;
1229         struct hash_bucket *bucket;
1230         unsigned long flags;
1231
1232         bucket = get_hash_bucket(ref, &flags);
1233
1234         entry = bucket_find_contain(&bucket, ref, &flags);
1235
1236         if (!entry) {
1237                 err_printk(dev, NULL, "DMA-API: device driver tries "
1238                                 "to sync DMA memory it has not allocated "
1239                                 "[device address=0x%016llx] [size=%llu bytes]\n",
1240                                 (unsigned long long)ref->dev_addr, ref->size);
1241                 goto out;
1242         }
1243
1244         if (ref->size > entry->size) {
1245                 err_printk(dev, entry, "DMA-API: device driver syncs"
1246                                 " DMA memory outside allocated range "
1247                                 "[device address=0x%016llx] "
1248                                 "[allocation size=%llu bytes] "
1249                                 "[sync offset+size=%llu]\n",
1250                                 entry->dev_addr, entry->size,
1251                                 ref->size);
1252         }
1253
1254         if (entry->direction == DMA_BIDIRECTIONAL)
1255                 goto out;
1256
1257         if (ref->direction != entry->direction) {
1258                 err_printk(dev, entry, "DMA-API: device driver syncs "
1259                                 "DMA memory with different direction "
1260                                 "[device address=0x%016llx] [size=%llu bytes] "
1261                                 "[mapped with %s] [synced with %s]\n",
1262                                 (unsigned long long)ref->dev_addr, entry->size,
1263                                 dir2name[entry->direction],
1264                                 dir2name[ref->direction]);
1265         }
1266
1267         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1268                       !(ref->direction == DMA_TO_DEVICE))
1269                 err_printk(dev, entry, "DMA-API: device driver syncs "
1270                                 "device read-only DMA memory for cpu "
1271                                 "[device address=0x%016llx] [size=%llu bytes] "
1272                                 "[mapped with %s] [synced with %s]\n",
1273                                 (unsigned long long)ref->dev_addr, entry->size,
1274                                 dir2name[entry->direction],
1275                                 dir2name[ref->direction]);
1276
1277         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1278                        !(ref->direction == DMA_FROM_DEVICE))
1279                 err_printk(dev, entry, "DMA-API: device driver syncs "
1280                                 "device write-only DMA memory to device "
1281                                 "[device address=0x%016llx] [size=%llu bytes] "
1282                                 "[mapped with %s] [synced with %s]\n",
1283                                 (unsigned long long)ref->dev_addr, entry->size,
1284                                 dir2name[entry->direction],
1285                                 dir2name[ref->direction]);
1286
1287         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1288             ref->sg_call_ents != entry->sg_call_ents) {
1289                 err_printk(ref->dev, entry, "DMA-API: device driver syncs "
1290                            "DMA sg list with different entry count "
1291                            "[map count=%d] [sync count=%d]\n",
1292                            entry->sg_call_ents, ref->sg_call_ents);
1293         }
1294
1295 out:
1296         put_hash_bucket(bucket, &flags);
1297 }
1298
1299 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1300                         size_t size, int direction, dma_addr_t dma_addr,
1301                         bool map_single)
1302 {
1303         struct dma_debug_entry *entry;
1304
1305         if (unlikely(dma_debug_disabled()))
1306                 return;
1307
1308         if (dma_mapping_error(dev, dma_addr))
1309                 return;
1310
1311         entry = dma_entry_alloc();
1312         if (!entry)
1313                 return;
1314
1315         entry->dev       = dev;
1316         entry->type      = dma_debug_page;
1317         entry->pfn       = page_to_pfn(page);
1318         entry->offset    = offset,
1319         entry->dev_addr  = dma_addr;
1320         entry->size      = size;
1321         entry->direction = direction;
1322         entry->map_err_type = MAP_ERR_NOT_CHECKED;
1323
1324         if (map_single)
1325                 entry->type = dma_debug_single;
1326
1327         check_for_stack(dev, page, offset);
1328
1329         if (!PageHighMem(page)) {
1330                 void *addr = page_address(page) + offset;
1331
1332                 check_for_illegal_area(dev, addr, size);
1333         }
1334
1335         add_dma_entry(entry);
1336 }
1337 EXPORT_SYMBOL(debug_dma_map_page);
1338
1339 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1340 {
1341         struct dma_debug_entry ref;
1342         struct dma_debug_entry *entry;
1343         struct hash_bucket *bucket;
1344         unsigned long flags;
1345
1346         if (unlikely(dma_debug_disabled()))
1347                 return;
1348
1349         ref.dev = dev;
1350         ref.dev_addr = dma_addr;
1351         bucket = get_hash_bucket(&ref, &flags);
1352
1353         list_for_each_entry(entry, &bucket->list, list) {
1354                 if (!exact_match(&ref, entry))
1355                         continue;
1356
1357                 /*
1358                  * The same physical address can be mapped multiple
1359                  * times. Without a hardware IOMMU this results in the
1360                  * same device addresses being put into the dma-debug
1361                  * hash multiple times too. This can result in false
1362                  * positives being reported. Therefore we implement a
1363                  * best-fit algorithm here which updates the first entry
1364                  * from the hash which fits the reference value and is
1365                  * not currently listed as being checked.
1366                  */
1367                 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1368                         entry->map_err_type = MAP_ERR_CHECKED;
1369                         break;
1370                 }
1371         }
1372
1373         put_hash_bucket(bucket, &flags);
1374 }
1375 EXPORT_SYMBOL(debug_dma_mapping_error);
1376
1377 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1378                           size_t size, int direction, bool map_single)
1379 {
1380         struct dma_debug_entry ref = {
1381                 .type           = dma_debug_page,
1382                 .dev            = dev,
1383                 .dev_addr       = addr,
1384                 .size           = size,
1385                 .direction      = direction,
1386         };
1387
1388         if (unlikely(dma_debug_disabled()))
1389                 return;
1390
1391         if (map_single)
1392                 ref.type = dma_debug_single;
1393
1394         check_unmap(&ref);
1395 }
1396 EXPORT_SYMBOL(debug_dma_unmap_page);
1397
1398 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1399                       int nents, int mapped_ents, int direction)
1400 {
1401         struct dma_debug_entry *entry;
1402         struct scatterlist *s;
1403         int i;
1404
1405         if (unlikely(dma_debug_disabled()))
1406                 return;
1407
1408         for_each_sg(sg, s, mapped_ents, i) {
1409                 entry = dma_entry_alloc();
1410                 if (!entry)
1411                         return;
1412
1413                 entry->type           = dma_debug_sg;
1414                 entry->dev            = dev;
1415                 entry->pfn            = page_to_pfn(sg_page(s));
1416                 entry->offset         = s->offset,
1417                 entry->size           = sg_dma_len(s);
1418                 entry->dev_addr       = sg_dma_address(s);
1419                 entry->direction      = direction;
1420                 entry->sg_call_ents   = nents;
1421                 entry->sg_mapped_ents = mapped_ents;
1422
1423                 check_for_stack(dev, sg_page(s), s->offset);
1424
1425                 if (!PageHighMem(sg_page(s))) {
1426                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1427                 }
1428
1429                 add_dma_entry(entry);
1430         }
1431 }
1432 EXPORT_SYMBOL(debug_dma_map_sg);
1433
1434 static int get_nr_mapped_entries(struct device *dev,
1435                                  struct dma_debug_entry *ref)
1436 {
1437         struct dma_debug_entry *entry;
1438         struct hash_bucket *bucket;
1439         unsigned long flags;
1440         int mapped_ents;
1441
1442         bucket       = get_hash_bucket(ref, &flags);
1443         entry        = bucket_find_exact(bucket, ref);
1444         mapped_ents  = 0;
1445
1446         if (entry)
1447                 mapped_ents = entry->sg_mapped_ents;
1448         put_hash_bucket(bucket, &flags);
1449
1450         return mapped_ents;
1451 }
1452
1453 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1454                         int nelems, int dir)
1455 {
1456         struct scatterlist *s;
1457         int mapped_ents = 0, i;
1458
1459         if (unlikely(dma_debug_disabled()))
1460                 return;
1461
1462         for_each_sg(sglist, s, nelems, i) {
1463
1464                 struct dma_debug_entry ref = {
1465                         .type           = dma_debug_sg,
1466                         .dev            = dev,
1467                         .pfn            = page_to_pfn(sg_page(s)),
1468                         .offset         = s->offset,
1469                         .dev_addr       = sg_dma_address(s),
1470                         .size           = sg_dma_len(s),
1471                         .direction      = dir,
1472                         .sg_call_ents   = nelems,
1473                 };
1474
1475                 if (mapped_ents && i >= mapped_ents)
1476                         break;
1477
1478                 if (!i)
1479                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1480
1481                 check_unmap(&ref);
1482         }
1483 }
1484 EXPORT_SYMBOL(debug_dma_unmap_sg);
1485
1486 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1487                               dma_addr_t dma_addr, void *virt)
1488 {
1489         struct dma_debug_entry *entry;
1490
1491         if (unlikely(dma_debug_disabled()))
1492                 return;
1493
1494         if (unlikely(virt == NULL))
1495                 return;
1496
1497         entry = dma_entry_alloc();
1498         if (!entry)
1499                 return;
1500
1501         entry->type      = dma_debug_coherent;
1502         entry->dev       = dev;
1503         entry->pfn       = page_to_pfn(virt_to_page(virt));
1504         entry->offset    = (size_t) virt & ~PAGE_MASK;
1505         entry->size      = size;
1506         entry->dev_addr  = dma_addr;
1507         entry->direction = DMA_BIDIRECTIONAL;
1508
1509         add_dma_entry(entry);
1510 }
1511 EXPORT_SYMBOL(debug_dma_alloc_coherent);
1512
1513 void debug_dma_free_coherent(struct device *dev, size_t size,
1514                          void *virt, dma_addr_t addr)
1515 {
1516         struct dma_debug_entry ref = {
1517                 .type           = dma_debug_coherent,
1518                 .dev            = dev,
1519                 .pfn            = page_to_pfn(virt_to_page(virt)),
1520                 .offset         = (size_t) virt & ~PAGE_MASK,
1521                 .dev_addr       = addr,
1522                 .size           = size,
1523                 .direction      = DMA_BIDIRECTIONAL,
1524         };
1525
1526         if (unlikely(dma_debug_disabled()))
1527                 return;
1528
1529         check_unmap(&ref);
1530 }
1531 EXPORT_SYMBOL(debug_dma_free_coherent);
1532
1533 void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1534                             int direction, dma_addr_t dma_addr)
1535 {
1536         struct dma_debug_entry *entry;
1537
1538         if (unlikely(dma_debug_disabled()))
1539                 return;
1540
1541         entry = dma_entry_alloc();
1542         if (!entry)
1543                 return;
1544
1545         entry->type             = dma_debug_resource;
1546         entry->dev              = dev;
1547         entry->pfn              = PHYS_PFN(addr);
1548         entry->offset           = offset_in_page(addr);
1549         entry->size             = size;
1550         entry->dev_addr         = dma_addr;
1551         entry->direction        = direction;
1552         entry->map_err_type     = MAP_ERR_NOT_CHECKED;
1553
1554         add_dma_entry(entry);
1555 }
1556 EXPORT_SYMBOL(debug_dma_map_resource);
1557
1558 void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1559                               size_t size, int direction)
1560 {
1561         struct dma_debug_entry ref = {
1562                 .type           = dma_debug_resource,
1563                 .dev            = dev,
1564                 .dev_addr       = dma_addr,
1565                 .size           = size,
1566                 .direction      = direction,
1567         };
1568
1569         if (unlikely(dma_debug_disabled()))
1570                 return;
1571
1572         check_unmap(&ref);
1573 }
1574 EXPORT_SYMBOL(debug_dma_unmap_resource);
1575
1576 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1577                                    size_t size, int direction)
1578 {
1579         struct dma_debug_entry ref;
1580
1581         if (unlikely(dma_debug_disabled()))
1582                 return;
1583
1584         ref.type         = dma_debug_single;
1585         ref.dev          = dev;
1586         ref.dev_addr     = dma_handle;
1587         ref.size         = size;
1588         ref.direction    = direction;
1589         ref.sg_call_ents = 0;
1590
1591         check_sync(dev, &ref, true);
1592 }
1593 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1594
1595 void debug_dma_sync_single_for_device(struct device *dev,
1596                                       dma_addr_t dma_handle, size_t size,
1597                                       int direction)
1598 {
1599         struct dma_debug_entry ref;
1600
1601         if (unlikely(dma_debug_disabled()))
1602                 return;
1603
1604         ref.type         = dma_debug_single;
1605         ref.dev          = dev;
1606         ref.dev_addr     = dma_handle;
1607         ref.size         = size;
1608         ref.direction    = direction;
1609         ref.sg_call_ents = 0;
1610
1611         check_sync(dev, &ref, false);
1612 }
1613 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1614
1615 void debug_dma_sync_single_range_for_cpu(struct device *dev,
1616                                          dma_addr_t dma_handle,
1617                                          unsigned long offset, size_t size,
1618                                          int direction)
1619 {
1620         struct dma_debug_entry ref;
1621
1622         if (unlikely(dma_debug_disabled()))
1623                 return;
1624
1625         ref.type         = dma_debug_single;
1626         ref.dev          = dev;
1627         ref.dev_addr     = dma_handle;
1628         ref.size         = offset + size;
1629         ref.direction    = direction;
1630         ref.sg_call_ents = 0;
1631
1632         check_sync(dev, &ref, true);
1633 }
1634 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
1635
1636 void debug_dma_sync_single_range_for_device(struct device *dev,
1637                                             dma_addr_t dma_handle,
1638                                             unsigned long offset,
1639                                             size_t size, int direction)
1640 {
1641         struct dma_debug_entry ref;
1642
1643         if (unlikely(dma_debug_disabled()))
1644                 return;
1645
1646         ref.type         = dma_debug_single;
1647         ref.dev          = dev;
1648         ref.dev_addr     = dma_handle;
1649         ref.size         = offset + size;
1650         ref.direction    = direction;
1651         ref.sg_call_ents = 0;
1652
1653         check_sync(dev, &ref, false);
1654 }
1655 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
1656
1657 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1658                                int nelems, int direction)
1659 {
1660         struct scatterlist *s;
1661         int mapped_ents = 0, i;
1662
1663         if (unlikely(dma_debug_disabled()))
1664                 return;
1665
1666         for_each_sg(sg, s, nelems, i) {
1667
1668                 struct dma_debug_entry ref = {
1669                         .type           = dma_debug_sg,
1670                         .dev            = dev,
1671                         .pfn            = page_to_pfn(sg_page(s)),
1672                         .offset         = s->offset,
1673                         .dev_addr       = sg_dma_address(s),
1674                         .size           = sg_dma_len(s),
1675                         .direction      = direction,
1676                         .sg_call_ents   = nelems,
1677                 };
1678
1679                 if (!i)
1680                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1681
1682                 if (i >= mapped_ents)
1683                         break;
1684
1685                 check_sync(dev, &ref, true);
1686         }
1687 }
1688 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1689
1690 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1691                                   int nelems, int direction)
1692 {
1693         struct scatterlist *s;
1694         int mapped_ents = 0, i;
1695
1696         if (unlikely(dma_debug_disabled()))
1697                 return;
1698
1699         for_each_sg(sg, s, nelems, i) {
1700
1701                 struct dma_debug_entry ref = {
1702                         .type           = dma_debug_sg,
1703                         .dev            = dev,
1704                         .pfn            = page_to_pfn(sg_page(s)),
1705                         .offset         = s->offset,
1706                         .dev_addr       = sg_dma_address(s),
1707                         .size           = sg_dma_len(s),
1708                         .direction      = direction,
1709                         .sg_call_ents   = nelems,
1710                 };
1711                 if (!i)
1712                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1713
1714                 if (i >= mapped_ents)
1715                         break;
1716
1717                 check_sync(dev, &ref, false);
1718         }
1719 }
1720 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1721
1722 static int __init dma_debug_driver_setup(char *str)
1723 {
1724         int i;
1725
1726         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1727                 current_driver_name[i] = *str;
1728                 if (*str == 0)
1729                         break;
1730         }
1731
1732         if (current_driver_name[0])
1733                 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1734                         current_driver_name);
1735
1736
1737         return 1;
1738 }
1739 __setup("dma_debug_driver=", dma_debug_driver_setup);