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