dma-debug: Put all hash-chain locks into the same lock class
[linux-2.6.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/device.h>
28 #include <linux/types.h>
29 #include <linux/sched.h>
30 #include <linux/ctype.h>
31 #include <linux/list.h>
32 #include <linux/slab.h>
33
34 #include <asm/sections.h>
35
36 #define HASH_SIZE       1024ULL
37 #define HASH_FN_SHIFT   13
38 #define HASH_FN_MASK    (HASH_SIZE - 1)
39
40 enum {
41         dma_debug_single,
42         dma_debug_page,
43         dma_debug_sg,
44         dma_debug_coherent,
45 };
46
47 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
48
49 struct dma_debug_entry {
50         struct list_head list;
51         struct device    *dev;
52         int              type;
53         phys_addr_t      paddr;
54         u64              dev_addr;
55         u64              size;
56         int              direction;
57         int              sg_call_ents;
58         int              sg_mapped_ents;
59 #ifdef CONFIG_STACKTRACE
60         struct           stack_trace stacktrace;
61         unsigned long    st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
62 #endif
63 };
64
65 struct hash_bucket {
66         struct list_head list;
67         spinlock_t lock;
68 } ____cacheline_aligned_in_smp;
69
70 /* Hash list to save the allocated dma addresses */
71 static struct hash_bucket dma_entry_hash[HASH_SIZE];
72 /* List of pre-allocated dma_debug_entry's */
73 static LIST_HEAD(free_entries);
74 /* Lock for the list above */
75 static DEFINE_SPINLOCK(free_entries_lock);
76
77 /* Global disable flag - will be set in case of an error */
78 static bool global_disable __read_mostly;
79
80 /* Global error count */
81 static u32 error_count;
82
83 /* Global error show enable*/
84 static u32 show_all_errors __read_mostly;
85 /* Number of errors to show */
86 static u32 show_num_errors = 1;
87
88 static u32 num_free_entries;
89 static u32 min_free_entries;
90 static u32 nr_total_entries;
91
92 /* number of preallocated entries requested by kernel cmdline */
93 static u32 req_entries;
94
95 /* debugfs dentry's for the stuff above */
96 static struct dentry *dma_debug_dent        __read_mostly;
97 static struct dentry *global_disable_dent   __read_mostly;
98 static struct dentry *error_count_dent      __read_mostly;
99 static struct dentry *show_all_errors_dent  __read_mostly;
100 static struct dentry *show_num_errors_dent  __read_mostly;
101 static struct dentry *num_free_entries_dent __read_mostly;
102 static struct dentry *min_free_entries_dent __read_mostly;
103 static struct dentry *filter_dent           __read_mostly;
104
105 /* per-driver filter related state */
106
107 #define NAME_MAX_LEN    64
108
109 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
110 static struct device_driver *current_driver                    __read_mostly;
111
112 static DEFINE_RWLOCK(driver_name_lock);
113
114 static const char *type2name[4] = { "single", "page",
115                                     "scather-gather", "coherent" };
116
117 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
118                                    "DMA_FROM_DEVICE", "DMA_NONE" };
119
120 /* little merge helper - remove it after the merge window */
121 #ifndef BUS_NOTIFY_UNBOUND_DRIVER
122 #define BUS_NOTIFY_UNBOUND_DRIVER 0x0005
123 #endif
124
125 /*
126  * The access to some variables in this macro is racy. We can't use atomic_t
127  * here because all these variables are exported to debugfs. Some of them even
128  * writeable. This is also the reason why a lock won't help much. But anyway,
129  * the races are no big deal. Here is why:
130  *
131  *   error_count: the addition is racy, but the worst thing that can happen is
132  *                that we don't count some errors
133  *   show_num_errors: the subtraction is racy. Also no big deal because in
134  *                    worst case this will result in one warning more in the
135  *                    system log than the user configured. This variable is
136  *                    writeable via debugfs.
137  */
138 static inline void dump_entry_trace(struct dma_debug_entry *entry)
139 {
140 #ifdef CONFIG_STACKTRACE
141         if (entry) {
142                 pr_warning("Mapped at:\n");
143                 print_stack_trace(&entry->stacktrace, 0);
144         }
145 #endif
146 }
147
148 static bool driver_filter(struct device *dev)
149 {
150         struct device_driver *drv;
151         unsigned long flags;
152         bool ret;
153
154         /* driver filter off */
155         if (likely(!current_driver_name[0]))
156                 return true;
157
158         /* driver filter on and initialized */
159         if (current_driver && dev->driver == current_driver)
160                 return true;
161
162         if (current_driver || !current_driver_name[0])
163                 return false;
164
165         /* driver filter on but not yet initialized */
166         drv = get_driver(dev->driver);
167         if (!drv)
168                 return false;
169
170         /* lock to protect against change of current_driver_name */
171         read_lock_irqsave(&driver_name_lock, flags);
172
173         ret = false;
174         if (drv->name &&
175             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
176                 current_driver = drv;
177                 ret = true;
178         }
179
180         read_unlock_irqrestore(&driver_name_lock, flags);
181         put_driver(drv);
182
183         return ret;
184 }
185
186 #define err_printk(dev, entry, format, arg...) do {             \
187                 error_count += 1;                               \
188                 if (driver_filter(dev) &&                       \
189                     (show_all_errors || show_num_errors > 0)) { \
190                         WARN(1, "%s %s: " format,               \
191                              dev_driver_string(dev),            \
192                              dev_name(dev) , ## arg);           \
193                         dump_entry_trace(entry);                \
194                 }                                               \
195                 if (!show_all_errors && show_num_errors > 0)    \
196                         show_num_errors -= 1;                   \
197         } while (0);
198
199 /*
200  * Hash related functions
201  *
202  * Every DMA-API request is saved into a struct dma_debug_entry. To
203  * have quick access to these structs they are stored into a hash.
204  */
205 static int hash_fn(struct dma_debug_entry *entry)
206 {
207         /*
208          * Hash function is based on the dma address.
209          * We use bits 20-27 here as the index into the hash
210          */
211         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
212 }
213
214 /*
215  * Request exclusive access to a hash bucket for a given dma_debug_entry.
216  */
217 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
218                                            unsigned long *flags)
219 {
220         int idx = hash_fn(entry);
221         unsigned long __flags;
222
223         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
224         *flags = __flags;
225         return &dma_entry_hash[idx];
226 }
227
228 /*
229  * Give up exclusive access to the hash bucket
230  */
231 static void put_hash_bucket(struct hash_bucket *bucket,
232                             unsigned long *flags)
233 {
234         unsigned long __flags = *flags;
235
236         spin_unlock_irqrestore(&bucket->lock, __flags);
237 }
238
239 /*
240  * Search a given entry in the hash bucket list
241  */
242 static struct dma_debug_entry *hash_bucket_find(struct hash_bucket *bucket,
243                                                 struct dma_debug_entry *ref)
244 {
245         struct dma_debug_entry *entry, *ret = NULL;
246         int matches = 0, match_lvl, last_lvl = 0;
247
248         list_for_each_entry(entry, &bucket->list, list) {
249                 if ((entry->dev_addr != ref->dev_addr) ||
250                     (entry->dev != ref->dev))
251                         continue;
252
253                 /*
254                  * Some drivers map the same physical address multiple
255                  * times. Without a hardware IOMMU this results in the
256                  * same device addresses being put into the dma-debug
257                  * hash multiple times too. This can result in false
258                  * positives being reported. Therfore we implement a
259                  * best-fit algorithm here which returns the entry from
260                  * the hash which fits best to the reference value
261                  * instead of the first-fit.
262                  */
263                 matches += 1;
264                 match_lvl = 0;
265                 entry->size         == ref->size         ? ++match_lvl : 0;
266                 entry->type         == ref->type         ? ++match_lvl : 0;
267                 entry->direction    == ref->direction    ? ++match_lvl : 0;
268                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
269
270                 if (match_lvl == 4) {
271                         /* perfect-fit - return the result */
272                         return entry;
273                 } else if (match_lvl > last_lvl) {
274                         /*
275                          * We found an entry that fits better then the
276                          * previous one
277                          */
278                         last_lvl = match_lvl;
279                         ret      = entry;
280                 }
281         }
282
283         /*
284          * If we have multiple matches but no perfect-fit, just return
285          * NULL.
286          */
287         ret = (matches == 1) ? ret : NULL;
288
289         return ret;
290 }
291
292 /*
293  * Add an entry to a hash bucket
294  */
295 static void hash_bucket_add(struct hash_bucket *bucket,
296                             struct dma_debug_entry *entry)
297 {
298         list_add_tail(&entry->list, &bucket->list);
299 }
300
301 /*
302  * Remove entry from a hash bucket list
303  */
304 static void hash_bucket_del(struct dma_debug_entry *entry)
305 {
306         list_del(&entry->list);
307 }
308
309 /*
310  * Dump mapping entries for debugging purposes
311  */
312 void debug_dma_dump_mappings(struct device *dev)
313 {
314         int idx;
315
316         for (idx = 0; idx < HASH_SIZE; idx++) {
317                 struct hash_bucket *bucket = &dma_entry_hash[idx];
318                 struct dma_debug_entry *entry;
319                 unsigned long flags;
320
321                 spin_lock_irqsave(&bucket->lock, flags);
322
323                 list_for_each_entry(entry, &bucket->list, list) {
324                         if (!dev || dev == entry->dev) {
325                                 dev_info(entry->dev,
326                                          "%s idx %d P=%Lx D=%Lx L=%Lx %s\n",
327                                          type2name[entry->type], idx,
328                                          (unsigned long long)entry->paddr,
329                                          entry->dev_addr, entry->size,
330                                          dir2name[entry->direction]);
331                         }
332                 }
333
334                 spin_unlock_irqrestore(&bucket->lock, flags);
335         }
336 }
337 EXPORT_SYMBOL(debug_dma_dump_mappings);
338
339 /*
340  * Wrapper function for adding an entry to the hash.
341  * This function takes care of locking itself.
342  */
343 static void add_dma_entry(struct dma_debug_entry *entry)
344 {
345         struct hash_bucket *bucket;
346         unsigned long flags;
347
348         bucket = get_hash_bucket(entry, &flags);
349         hash_bucket_add(bucket, entry);
350         put_hash_bucket(bucket, &flags);
351 }
352
353 static struct dma_debug_entry *__dma_entry_alloc(void)
354 {
355         struct dma_debug_entry *entry;
356
357         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
358         list_del(&entry->list);
359         memset(entry, 0, sizeof(*entry));
360
361         num_free_entries -= 1;
362         if (num_free_entries < min_free_entries)
363                 min_free_entries = num_free_entries;
364
365         return entry;
366 }
367
368 /* struct dma_entry allocator
369  *
370  * The next two functions implement the allocator for
371  * struct dma_debug_entries.
372  */
373 static struct dma_debug_entry *dma_entry_alloc(void)
374 {
375         struct dma_debug_entry *entry = NULL;
376         unsigned long flags;
377
378         spin_lock_irqsave(&free_entries_lock, flags);
379
380         if (list_empty(&free_entries)) {
381                 pr_err("DMA-API: debugging out of memory - disabling\n");
382                 global_disable = true;
383                 goto out;
384         }
385
386         entry = __dma_entry_alloc();
387
388 #ifdef CONFIG_STACKTRACE
389         entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
390         entry->stacktrace.entries = entry->st_entries;
391         entry->stacktrace.skip = 2;
392         save_stack_trace(&entry->stacktrace);
393 #endif
394
395 out:
396         spin_unlock_irqrestore(&free_entries_lock, flags);
397
398         return entry;
399 }
400
401 static void dma_entry_free(struct dma_debug_entry *entry)
402 {
403         unsigned long flags;
404
405         /*
406          * add to beginning of the list - this way the entries are
407          * more likely cache hot when they are reallocated.
408          */
409         spin_lock_irqsave(&free_entries_lock, flags);
410         list_add(&entry->list, &free_entries);
411         num_free_entries += 1;
412         spin_unlock_irqrestore(&free_entries_lock, flags);
413 }
414
415 int dma_debug_resize_entries(u32 num_entries)
416 {
417         int i, delta, ret = 0;
418         unsigned long flags;
419         struct dma_debug_entry *entry;
420         LIST_HEAD(tmp);
421
422         spin_lock_irqsave(&free_entries_lock, flags);
423
424         if (nr_total_entries < num_entries) {
425                 delta = num_entries - nr_total_entries;
426
427                 spin_unlock_irqrestore(&free_entries_lock, flags);
428
429                 for (i = 0; i < delta; i++) {
430                         entry = kzalloc(sizeof(*entry), GFP_KERNEL);
431                         if (!entry)
432                                 break;
433
434                         list_add_tail(&entry->list, &tmp);
435                 }
436
437                 spin_lock_irqsave(&free_entries_lock, flags);
438
439                 list_splice(&tmp, &free_entries);
440                 nr_total_entries += i;
441                 num_free_entries += i;
442         } else {
443                 delta = nr_total_entries - num_entries;
444
445                 for (i = 0; i < delta && !list_empty(&free_entries); i++) {
446                         entry = __dma_entry_alloc();
447                         kfree(entry);
448                 }
449
450                 nr_total_entries -= i;
451         }
452
453         if (nr_total_entries != num_entries)
454                 ret = 1;
455
456         spin_unlock_irqrestore(&free_entries_lock, flags);
457
458         return ret;
459 }
460 EXPORT_SYMBOL(dma_debug_resize_entries);
461
462 /*
463  * DMA-API debugging init code
464  *
465  * The init code does two things:
466  *   1. Initialize core data structures
467  *   2. Preallocate a given number of dma_debug_entry structs
468  */
469
470 static int prealloc_memory(u32 num_entries)
471 {
472         struct dma_debug_entry *entry, *next_entry;
473         int i;
474
475         for (i = 0; i < num_entries; ++i) {
476                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
477                 if (!entry)
478                         goto out_err;
479
480                 list_add_tail(&entry->list, &free_entries);
481         }
482
483         num_free_entries = num_entries;
484         min_free_entries = num_entries;
485
486         pr_info("DMA-API: preallocated %d debug entries\n", num_entries);
487
488         return 0;
489
490 out_err:
491
492         list_for_each_entry_safe(entry, next_entry, &free_entries, list) {
493                 list_del(&entry->list);
494                 kfree(entry);
495         }
496
497         return -ENOMEM;
498 }
499
500 static ssize_t filter_read(struct file *file, char __user *user_buf,
501                            size_t count, loff_t *ppos)
502 {
503         char buf[NAME_MAX_LEN + 1];
504         unsigned long flags;
505         int len;
506
507         if (!current_driver_name[0])
508                 return 0;
509
510         /*
511          * We can't copy to userspace directly because current_driver_name can
512          * only be read under the driver_name_lock with irqs disabled. So
513          * create a temporary copy first.
514          */
515         read_lock_irqsave(&driver_name_lock, flags);
516         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
517         read_unlock_irqrestore(&driver_name_lock, flags);
518
519         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
520 }
521
522 static ssize_t filter_write(struct file *file, const char __user *userbuf,
523                             size_t count, loff_t *ppos)
524 {
525         char buf[NAME_MAX_LEN];
526         unsigned long flags;
527         size_t len;
528         int i;
529
530         /*
531          * We can't copy from userspace directly. Access to
532          * current_driver_name is protected with a write_lock with irqs
533          * disabled. Since copy_from_user can fault and may sleep we
534          * need to copy to temporary buffer first
535          */
536         len = min(count, (size_t)(NAME_MAX_LEN - 1));
537         if (copy_from_user(buf, userbuf, len))
538                 return -EFAULT;
539
540         buf[len] = 0;
541
542         write_lock_irqsave(&driver_name_lock, flags);
543
544         /*
545          * Now handle the string we got from userspace very carefully.
546          * The rules are:
547          *         - only use the first token we got
548          *         - token delimiter is everything looking like a space
549          *           character (' ', '\n', '\t' ...)
550          *
551          */
552         if (!isalnum(buf[0])) {
553                 /*
554                  * If the first character userspace gave us is not
555                  * alphanumerical then assume the filter should be
556                  * switched off.
557                  */
558                 if (current_driver_name[0])
559                         pr_info("DMA-API: switching off dma-debug driver filter\n");
560                 current_driver_name[0] = 0;
561                 current_driver = NULL;
562                 goto out_unlock;
563         }
564
565         /*
566          * Now parse out the first token and use it as the name for the
567          * driver to filter for.
568          */
569         for (i = 0; i < NAME_MAX_LEN; ++i) {
570                 current_driver_name[i] = buf[i];
571                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
572                         break;
573         }
574         current_driver_name[i] = 0;
575         current_driver = NULL;
576
577         pr_info("DMA-API: enable driver filter for driver [%s]\n",
578                 current_driver_name);
579
580 out_unlock:
581         write_unlock_irqrestore(&driver_name_lock, flags);
582
583         return count;
584 }
585
586 const struct file_operations filter_fops = {
587         .read  = filter_read,
588         .write = filter_write,
589 };
590
591 static int dma_debug_fs_init(void)
592 {
593         dma_debug_dent = debugfs_create_dir("dma-api", NULL);
594         if (!dma_debug_dent) {
595                 pr_err("DMA-API: can not create debugfs directory\n");
596                 return -ENOMEM;
597         }
598
599         global_disable_dent = debugfs_create_bool("disabled", 0444,
600                         dma_debug_dent,
601                         (u32 *)&global_disable);
602         if (!global_disable_dent)
603                 goto out_err;
604
605         error_count_dent = debugfs_create_u32("error_count", 0444,
606                         dma_debug_dent, &error_count);
607         if (!error_count_dent)
608                 goto out_err;
609
610         show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
611                         dma_debug_dent,
612                         &show_all_errors);
613         if (!show_all_errors_dent)
614                 goto out_err;
615
616         show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
617                         dma_debug_dent,
618                         &show_num_errors);
619         if (!show_num_errors_dent)
620                 goto out_err;
621
622         num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
623                         dma_debug_dent,
624                         &num_free_entries);
625         if (!num_free_entries_dent)
626                 goto out_err;
627
628         min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
629                         dma_debug_dent,
630                         &min_free_entries);
631         if (!min_free_entries_dent)
632                 goto out_err;
633
634         filter_dent = debugfs_create_file("driver_filter", 0644,
635                                           dma_debug_dent, NULL, &filter_fops);
636         if (!filter_dent)
637                 goto out_err;
638
639         return 0;
640
641 out_err:
642         debugfs_remove_recursive(dma_debug_dent);
643
644         return -ENOMEM;
645 }
646
647 static int device_dma_allocations(struct device *dev)
648 {
649         struct dma_debug_entry *entry;
650         unsigned long flags;
651         int count = 0, i;
652
653         local_irq_save(flags);
654
655         for (i = 0; i < HASH_SIZE; ++i) {
656                 spin_lock(&dma_entry_hash[i].lock);
657                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
658                         if (entry->dev == dev)
659                                 count += 1;
660                 }
661                 spin_unlock(&dma_entry_hash[i].lock);
662         }
663
664         local_irq_restore(flags);
665
666         return count;
667 }
668
669 static int dma_debug_device_change(struct notifier_block *nb,
670                                     unsigned long action, void *data)
671 {
672         struct device *dev = data;
673         int count;
674
675
676         switch (action) {
677         case BUS_NOTIFY_UNBOUND_DRIVER:
678                 count = device_dma_allocations(dev);
679                 if (count == 0)
680                         break;
681                 err_printk(dev, NULL, "DMA-API: device driver has pending "
682                                 "DMA allocations while released from device "
683                                 "[count=%d]\n", count);
684                 break;
685         default:
686                 break;
687         }
688
689         return 0;
690 }
691
692 void dma_debug_add_bus(struct bus_type *bus)
693 {
694         struct notifier_block *nb;
695
696         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
697         if (nb == NULL) {
698                 pr_err("dma_debug_add_bus: out of memory\n");
699                 return;
700         }
701
702         nb->notifier_call = dma_debug_device_change;
703
704         bus_register_notifier(bus, nb);
705 }
706
707 /*
708  * Let the architectures decide how many entries should be preallocated.
709  */
710 void dma_debug_init(u32 num_entries)
711 {
712         int i;
713
714         if (global_disable)
715                 return;
716
717         for (i = 0; i < HASH_SIZE; ++i) {
718                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
719                 spin_lock_init(&dma_entry_hash[i].lock);
720         }
721
722         if (dma_debug_fs_init() != 0) {
723                 pr_err("DMA-API: error creating debugfs entries - disabling\n");
724                 global_disable = true;
725
726                 return;
727         }
728
729         if (req_entries)
730                 num_entries = req_entries;
731
732         if (prealloc_memory(num_entries) != 0) {
733                 pr_err("DMA-API: debugging out of memory error - disabled\n");
734                 global_disable = true;
735
736                 return;
737         }
738
739         nr_total_entries = num_free_entries;
740
741         pr_info("DMA-API: debugging enabled by kernel config\n");
742 }
743
744 static __init int dma_debug_cmdline(char *str)
745 {
746         if (!str)
747                 return -EINVAL;
748
749         if (strncmp(str, "off", 3) == 0) {
750                 pr_info("DMA-API: debugging disabled on kernel command line\n");
751                 global_disable = true;
752         }
753
754         return 0;
755 }
756
757 static __init int dma_debug_entries_cmdline(char *str)
758 {
759         int res;
760
761         if (!str)
762                 return -EINVAL;
763
764         res = get_option(&str, &req_entries);
765
766         if (!res)
767                 req_entries = 0;
768
769         return 0;
770 }
771
772 __setup("dma_debug=", dma_debug_cmdline);
773 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
774
775 static void check_unmap(struct dma_debug_entry *ref)
776 {
777         struct dma_debug_entry *entry;
778         struct hash_bucket *bucket;
779         unsigned long flags;
780
781         if (dma_mapping_error(ref->dev, ref->dev_addr)) {
782                 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
783                            "to free an invalid DMA memory address\n");
784                 return;
785         }
786
787         bucket = get_hash_bucket(ref, &flags);
788         entry = hash_bucket_find(bucket, ref);
789
790         if (!entry) {
791                 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
792                            "to free DMA memory it has not allocated "
793                            "[device address=0x%016llx] [size=%llu bytes]\n",
794                            ref->dev_addr, ref->size);
795                 goto out;
796         }
797
798         if (ref->size != entry->size) {
799                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
800                            "DMA memory with different size "
801                            "[device address=0x%016llx] [map size=%llu bytes] "
802                            "[unmap size=%llu bytes]\n",
803                            ref->dev_addr, entry->size, ref->size);
804         }
805
806         if (ref->type != entry->type) {
807                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
808                            "DMA memory with wrong function "
809                            "[device address=0x%016llx] [size=%llu bytes] "
810                            "[mapped as %s] [unmapped as %s]\n",
811                            ref->dev_addr, ref->size,
812                            type2name[entry->type], type2name[ref->type]);
813         } else if ((entry->type == dma_debug_coherent) &&
814                    (ref->paddr != entry->paddr)) {
815                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
816                            "DMA memory with different CPU address "
817                            "[device address=0x%016llx] [size=%llu bytes] "
818                            "[cpu alloc address=%p] [cpu free address=%p]",
819                            ref->dev_addr, ref->size,
820                            (void *)entry->paddr, (void *)ref->paddr);
821         }
822
823         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
824             ref->sg_call_ents != entry->sg_call_ents) {
825                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
826                            "DMA sg list with different entry count "
827                            "[map count=%d] [unmap count=%d]\n",
828                            entry->sg_call_ents, ref->sg_call_ents);
829         }
830
831         /*
832          * This may be no bug in reality - but most implementations of the
833          * DMA API don't handle this properly, so check for it here
834          */
835         if (ref->direction != entry->direction) {
836                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
837                            "DMA memory with different direction "
838                            "[device address=0x%016llx] [size=%llu bytes] "
839                            "[mapped with %s] [unmapped with %s]\n",
840                            ref->dev_addr, ref->size,
841                            dir2name[entry->direction],
842                            dir2name[ref->direction]);
843         }
844
845         hash_bucket_del(entry);
846         dma_entry_free(entry);
847
848 out:
849         put_hash_bucket(bucket, &flags);
850 }
851
852 static void check_for_stack(struct device *dev, void *addr)
853 {
854         if (object_is_on_stack(addr))
855                 err_printk(dev, NULL, "DMA-API: device driver maps memory from"
856                                 "stack [addr=%p]\n", addr);
857 }
858
859 static inline bool overlap(void *addr, u64 size, void *start, void *end)
860 {
861         void *addr2 = (char *)addr + size;
862
863         return ((addr >= start && addr < end) ||
864                 (addr2 >= start && addr2 < end) ||
865                 ((addr < start) && (addr2 > end)));
866 }
867
868 static void check_for_illegal_area(struct device *dev, void *addr, u64 size)
869 {
870         if (overlap(addr, size, _text, _etext) ||
871             overlap(addr, size, __start_rodata, __end_rodata))
872                 err_printk(dev, NULL, "DMA-API: device driver maps "
873                                 "memory from kernel text or rodata "
874                                 "[addr=%p] [size=%llu]\n", addr, size);
875 }
876
877 static void check_sync(struct device *dev,
878                        struct dma_debug_entry *ref,
879                        bool to_cpu)
880 {
881         struct dma_debug_entry *entry;
882         struct hash_bucket *bucket;
883         unsigned long flags;
884
885         bucket = get_hash_bucket(ref, &flags);
886
887         entry = hash_bucket_find(bucket, ref);
888
889         if (!entry) {
890                 err_printk(dev, NULL, "DMA-API: device driver tries "
891                                 "to sync DMA memory it has not allocated "
892                                 "[device address=0x%016llx] [size=%llu bytes]\n",
893                                 (unsigned long long)ref->dev_addr, ref->size);
894                 goto out;
895         }
896
897         if (ref->size > entry->size) {
898                 err_printk(dev, entry, "DMA-API: device driver syncs"
899                                 " DMA memory outside allocated range "
900                                 "[device address=0x%016llx] "
901                                 "[allocation size=%llu bytes] "
902                                 "[sync offset+size=%llu]\n",
903                                 entry->dev_addr, entry->size,
904                                 ref->size);
905         }
906
907         if (ref->direction != entry->direction) {
908                 err_printk(dev, entry, "DMA-API: device driver syncs "
909                                 "DMA memory with different direction "
910                                 "[device address=0x%016llx] [size=%llu bytes] "
911                                 "[mapped with %s] [synced with %s]\n",
912                                 (unsigned long long)ref->dev_addr, entry->size,
913                                 dir2name[entry->direction],
914                                 dir2name[ref->direction]);
915         }
916
917         if (entry->direction == DMA_BIDIRECTIONAL)
918                 goto out;
919
920         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
921                       !(ref->direction == DMA_TO_DEVICE))
922                 err_printk(dev, entry, "DMA-API: device driver syncs "
923                                 "device read-only DMA memory for cpu "
924                                 "[device address=0x%016llx] [size=%llu bytes] "
925                                 "[mapped with %s] [synced with %s]\n",
926                                 (unsigned long long)ref->dev_addr, entry->size,
927                                 dir2name[entry->direction],
928                                 dir2name[ref->direction]);
929
930         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
931                        !(ref->direction == DMA_FROM_DEVICE))
932                 err_printk(dev, entry, "DMA-API: device driver syncs "
933                                 "device write-only DMA memory to device "
934                                 "[device address=0x%016llx] [size=%llu bytes] "
935                                 "[mapped with %s] [synced with %s]\n",
936                                 (unsigned long long)ref->dev_addr, entry->size,
937                                 dir2name[entry->direction],
938                                 dir2name[ref->direction]);
939
940 out:
941         put_hash_bucket(bucket, &flags);
942
943 }
944
945 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
946                         size_t size, int direction, dma_addr_t dma_addr,
947                         bool map_single)
948 {
949         struct dma_debug_entry *entry;
950
951         if (unlikely(global_disable))
952                 return;
953
954         if (unlikely(dma_mapping_error(dev, dma_addr)))
955                 return;
956
957         entry = dma_entry_alloc();
958         if (!entry)
959                 return;
960
961         entry->dev       = dev;
962         entry->type      = dma_debug_page;
963         entry->paddr     = page_to_phys(page) + offset;
964         entry->dev_addr  = dma_addr;
965         entry->size      = size;
966         entry->direction = direction;
967
968         if (map_single)
969                 entry->type = dma_debug_single;
970
971         if (!PageHighMem(page)) {
972                 void *addr = ((char *)page_address(page)) + offset;
973                 check_for_stack(dev, addr);
974                 check_for_illegal_area(dev, addr, size);
975         }
976
977         add_dma_entry(entry);
978 }
979 EXPORT_SYMBOL(debug_dma_map_page);
980
981 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
982                           size_t size, int direction, bool map_single)
983 {
984         struct dma_debug_entry ref = {
985                 .type           = dma_debug_page,
986                 .dev            = dev,
987                 .dev_addr       = addr,
988                 .size           = size,
989                 .direction      = direction,
990         };
991
992         if (unlikely(global_disable))
993                 return;
994
995         if (map_single)
996                 ref.type = dma_debug_single;
997
998         check_unmap(&ref);
999 }
1000 EXPORT_SYMBOL(debug_dma_unmap_page);
1001
1002 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1003                       int nents, int mapped_ents, int direction)
1004 {
1005         struct dma_debug_entry *entry;
1006         struct scatterlist *s;
1007         int i;
1008
1009         if (unlikely(global_disable))
1010                 return;
1011
1012         for_each_sg(sg, s, mapped_ents, i) {
1013                 entry = dma_entry_alloc();
1014                 if (!entry)
1015                         return;
1016
1017                 entry->type           = dma_debug_sg;
1018                 entry->dev            = dev;
1019                 entry->paddr          = sg_phys(s);
1020                 entry->size           = sg_dma_len(s);
1021                 entry->dev_addr       = sg_dma_address(s);
1022                 entry->direction      = direction;
1023                 entry->sg_call_ents   = nents;
1024                 entry->sg_mapped_ents = mapped_ents;
1025
1026                 if (!PageHighMem(sg_page(s))) {
1027                         check_for_stack(dev, sg_virt(s));
1028                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1029                 }
1030
1031                 add_dma_entry(entry);
1032         }
1033 }
1034 EXPORT_SYMBOL(debug_dma_map_sg);
1035
1036 static int get_nr_mapped_entries(struct device *dev,
1037                                  struct dma_debug_entry *ref)
1038 {
1039         struct dma_debug_entry *entry;
1040         struct hash_bucket *bucket;
1041         unsigned long flags;
1042         int mapped_ents;
1043
1044         bucket       = get_hash_bucket(ref, &flags);
1045         entry        = hash_bucket_find(bucket, ref);
1046         mapped_ents  = 0;
1047
1048         if (entry)
1049                 mapped_ents = entry->sg_mapped_ents;
1050         put_hash_bucket(bucket, &flags);
1051
1052         return mapped_ents;
1053 }
1054
1055 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1056                         int nelems, int dir)
1057 {
1058         struct scatterlist *s;
1059         int mapped_ents = 0, i;
1060
1061         if (unlikely(global_disable))
1062                 return;
1063
1064         for_each_sg(sglist, s, nelems, i) {
1065
1066                 struct dma_debug_entry ref = {
1067                         .type           = dma_debug_sg,
1068                         .dev            = dev,
1069                         .paddr          = sg_phys(s),
1070                         .dev_addr       = sg_dma_address(s),
1071                         .size           = sg_dma_len(s),
1072                         .direction      = dir,
1073                         .sg_call_ents   = nelems,
1074                 };
1075
1076                 if (mapped_ents && i >= mapped_ents)
1077                         break;
1078
1079                 if (!i)
1080                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1081
1082                 check_unmap(&ref);
1083         }
1084 }
1085 EXPORT_SYMBOL(debug_dma_unmap_sg);
1086
1087 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1088                               dma_addr_t dma_addr, void *virt)
1089 {
1090         struct dma_debug_entry *entry;
1091
1092         if (unlikely(global_disable))
1093                 return;
1094
1095         if (unlikely(virt == NULL))
1096                 return;
1097
1098         entry = dma_entry_alloc();
1099         if (!entry)
1100                 return;
1101
1102         entry->type      = dma_debug_coherent;
1103         entry->dev       = dev;
1104         entry->paddr     = virt_to_phys(virt);
1105         entry->size      = size;
1106         entry->dev_addr  = dma_addr;
1107         entry->direction = DMA_BIDIRECTIONAL;
1108
1109         add_dma_entry(entry);
1110 }
1111 EXPORT_SYMBOL(debug_dma_alloc_coherent);
1112
1113 void debug_dma_free_coherent(struct device *dev, size_t size,
1114                          void *virt, dma_addr_t addr)
1115 {
1116         struct dma_debug_entry ref = {
1117                 .type           = dma_debug_coherent,
1118                 .dev            = dev,
1119                 .paddr          = virt_to_phys(virt),
1120                 .dev_addr       = addr,
1121                 .size           = size,
1122                 .direction      = DMA_BIDIRECTIONAL,
1123         };
1124
1125         if (unlikely(global_disable))
1126                 return;
1127
1128         check_unmap(&ref);
1129 }
1130 EXPORT_SYMBOL(debug_dma_free_coherent);
1131
1132 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1133                                    size_t size, int direction)
1134 {
1135         struct dma_debug_entry ref;
1136
1137         if (unlikely(global_disable))
1138                 return;
1139
1140         ref.type         = dma_debug_single;
1141         ref.dev          = dev;
1142         ref.dev_addr     = dma_handle;
1143         ref.size         = size;
1144         ref.direction    = direction;
1145         ref.sg_call_ents = 0;
1146
1147         check_sync(dev, &ref, true);
1148 }
1149 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1150
1151 void debug_dma_sync_single_for_device(struct device *dev,
1152                                       dma_addr_t dma_handle, size_t size,
1153                                       int direction)
1154 {
1155         struct dma_debug_entry ref;
1156
1157         if (unlikely(global_disable))
1158                 return;
1159
1160         ref.type         = dma_debug_single;
1161         ref.dev          = dev;
1162         ref.dev_addr     = dma_handle;
1163         ref.size         = size;
1164         ref.direction    = direction;
1165         ref.sg_call_ents = 0;
1166
1167         check_sync(dev, &ref, false);
1168 }
1169 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1170
1171 void debug_dma_sync_single_range_for_cpu(struct device *dev,
1172                                          dma_addr_t dma_handle,
1173                                          unsigned long offset, size_t size,
1174                                          int direction)
1175 {
1176         struct dma_debug_entry ref;
1177
1178         if (unlikely(global_disable))
1179                 return;
1180
1181         ref.type         = dma_debug_single;
1182         ref.dev          = dev;
1183         ref.dev_addr     = dma_handle;
1184         ref.size         = offset + size;
1185         ref.direction    = direction;
1186         ref.sg_call_ents = 0;
1187
1188         check_sync(dev, &ref, true);
1189 }
1190 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
1191
1192 void debug_dma_sync_single_range_for_device(struct device *dev,
1193                                             dma_addr_t dma_handle,
1194                                             unsigned long offset,
1195                                             size_t size, int direction)
1196 {
1197         struct dma_debug_entry ref;
1198
1199         if (unlikely(global_disable))
1200                 return;
1201
1202         ref.type         = dma_debug_single;
1203         ref.dev          = dev;
1204         ref.dev_addr     = dma_handle;
1205         ref.size         = offset + size;
1206         ref.direction    = direction;
1207         ref.sg_call_ents = 0;
1208
1209         check_sync(dev, &ref, false);
1210 }
1211 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
1212
1213 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1214                                int nelems, int direction)
1215 {
1216         struct scatterlist *s;
1217         int mapped_ents = 0, i;
1218
1219         if (unlikely(global_disable))
1220                 return;
1221
1222         for_each_sg(sg, s, nelems, i) {
1223
1224                 struct dma_debug_entry ref = {
1225                         .type           = dma_debug_sg,
1226                         .dev            = dev,
1227                         .paddr          = sg_phys(s),
1228                         .dev_addr       = sg_dma_address(s),
1229                         .size           = sg_dma_len(s),
1230                         .direction      = direction,
1231                         .sg_call_ents   = nelems,
1232                 };
1233
1234                 if (!i)
1235                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1236
1237                 if (i >= mapped_ents)
1238                         break;
1239
1240                 check_sync(dev, &ref, true);
1241         }
1242 }
1243 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1244
1245 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1246                                   int nelems, int direction)
1247 {
1248         struct scatterlist *s;
1249         int mapped_ents = 0, i;
1250
1251         if (unlikely(global_disable))
1252                 return;
1253
1254         for_each_sg(sg, s, nelems, i) {
1255
1256                 struct dma_debug_entry ref = {
1257                         .type           = dma_debug_sg,
1258                         .dev            = dev,
1259                         .paddr          = sg_phys(s),
1260                         .dev_addr       = sg_dma_address(s),
1261                         .size           = sg_dma_len(s),
1262                         .direction      = direction,
1263                         .sg_call_ents   = nelems,
1264                 };
1265                 if (!i)
1266                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1267
1268                 if (i >= mapped_ents)
1269                         break;
1270
1271                 check_sync(dev, &ref, false);
1272         }
1273 }
1274 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1275
1276 static int __init dma_debug_driver_setup(char *str)
1277 {
1278         int i;
1279
1280         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1281                 current_driver_name[i] = *str;
1282                 if (*str == 0)
1283                         break;
1284         }
1285
1286         if (current_driver_name[0])
1287                 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1288                         current_driver_name);
1289
1290
1291         return 1;
1292 }
1293 __setup("dma_debug_driver=", dma_debug_driver_setup);