Commit | Line | Data |
---|---|---|
f2f45e5f JR |
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 | ||
972aa45c | 20 | #include <linux/scatterlist.h> |
2d62ece1 | 21 | #include <linux/dma-mapping.h> |
6c132d1b | 22 | #include <linux/stacktrace.h> |
f2f45e5f | 23 | #include <linux/dma-debug.h> |
30dfa90c | 24 | #include <linux/spinlock.h> |
788dcfa6 | 25 | #include <linux/debugfs.h> |
8a6fc708 | 26 | #include <linux/uaccess.h> |
23a7bfae | 27 | #include <linux/export.h> |
2d62ece1 | 28 | #include <linux/device.h> |
f2f45e5f | 29 | #include <linux/types.h> |
2d62ece1 | 30 | #include <linux/sched.h> |
8a6fc708 | 31 | #include <linux/ctype.h> |
f2f45e5f | 32 | #include <linux/list.h> |
6bf07871 | 33 | #include <linux/slab.h> |
f2f45e5f | 34 | |
2e34bde1 JR |
35 | #include <asm/sections.h> |
36 | ||
30dfa90c JR |
37 | #define HASH_SIZE 1024ULL |
38 | #define HASH_FN_SHIFT 13 | |
39 | #define HASH_FN_MASK (HASH_SIZE - 1) | |
40 | ||
f2f45e5f JR |
41 | enum { |
42 | dma_debug_single, | |
43 | dma_debug_page, | |
44 | dma_debug_sg, | |
45 | dma_debug_coherent, | |
46 | }; | |
47 | ||
6c9c6d63 SK |
48 | enum map_err_types { |
49 | MAP_ERR_CHECK_NOT_APPLICABLE, | |
50 | MAP_ERR_NOT_CHECKED, | |
51 | MAP_ERR_CHECKED, | |
52 | }; | |
53 | ||
6c132d1b DW |
54 | #define DMA_DEBUG_STACKTRACE_ENTRIES 5 |
55 | ||
0abdd7a8 DW |
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 | */ | |
f2f45e5f JR |
70 | struct dma_debug_entry { |
71 | struct list_head list; | |
72 | struct device *dev; | |
73 | int type; | |
0abdd7a8 DW |
74 | unsigned long pfn; |
75 | size_t offset; | |
f2f45e5f JR |
76 | u64 dev_addr; |
77 | u64 size; | |
78 | int direction; | |
79 | int sg_call_ents; | |
80 | int sg_mapped_ents; | |
6c9c6d63 | 81 | enum map_err_types map_err_type; |
6c132d1b DW |
82 | #ifdef CONFIG_STACKTRACE |
83 | struct stack_trace stacktrace; | |
84 | unsigned long st_entries[DMA_DEBUG_STACKTRACE_ENTRIES]; | |
85 | #endif | |
f2f45e5f JR |
86 | }; |
87 | ||
c6a21d0b NH |
88 | typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *); |
89 | ||
30dfa90c JR |
90 | struct hash_bucket { |
91 | struct list_head list; | |
92 | spinlock_t lock; | |
2d62ece1 | 93 | } ____cacheline_aligned_in_smp; |
30dfa90c JR |
94 | |
95 | /* Hash list to save the allocated dma addresses */ | |
96 | static struct hash_bucket dma_entry_hash[HASH_SIZE]; | |
3b1e79ed JR |
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 */ | |
68ee6d22 | 103 | static u32 global_disable __read_mostly; |
3b1e79ed | 104 | |
788dcfa6 JR |
105 | /* Global error count */ |
106 | static u32 error_count; | |
107 | ||
108 | /* Global error show enable*/ | |
109 | static u32 show_all_errors __read_mostly; | |
110 | /* Number of errors to show */ | |
111 | static u32 show_num_errors = 1; | |
112 | ||
3b1e79ed JR |
113 | static u32 num_free_entries; |
114 | static u32 min_free_entries; | |
e6a1a89d | 115 | static u32 nr_total_entries; |
30dfa90c | 116 | |
59d3daaf JR |
117 | /* number of preallocated entries requested by kernel cmdline */ |
118 | static u32 req_entries; | |
119 | ||
788dcfa6 JR |
120 | /* debugfs dentry's for the stuff above */ |
121 | static struct dentry *dma_debug_dent __read_mostly; | |
122 | static struct dentry *global_disable_dent __read_mostly; | |
123 | static struct dentry *error_count_dent __read_mostly; | |
124 | static struct dentry *show_all_errors_dent __read_mostly; | |
125 | static struct dentry *show_num_errors_dent __read_mostly; | |
126 | static struct dentry *num_free_entries_dent __read_mostly; | |
127 | static struct dentry *min_free_entries_dent __read_mostly; | |
8a6fc708 | 128 | static struct dentry *filter_dent __read_mostly; |
788dcfa6 | 129 | |
2e507d84 JR |
130 | /* per-driver filter related state */ |
131 | ||
132 | #define NAME_MAX_LEN 64 | |
133 | ||
134 | static char current_driver_name[NAME_MAX_LEN] __read_mostly; | |
135 | static struct device_driver *current_driver __read_mostly; | |
136 | ||
137 | static DEFINE_RWLOCK(driver_name_lock); | |
788dcfa6 | 138 | |
6c9c6d63 SK |
139 | static const char *const maperr2str[] = { |
140 | [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable", | |
141 | [MAP_ERR_NOT_CHECKED] = "dma map error not checked", | |
142 | [MAP_ERR_CHECKED] = "dma map error checked", | |
143 | }; | |
144 | ||
2d62ece1 JR |
145 | static const char *type2name[4] = { "single", "page", |
146 | "scather-gather", "coherent" }; | |
147 | ||
148 | static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE", | |
149 | "DMA_FROM_DEVICE", "DMA_NONE" }; | |
150 | ||
151 | /* | |
152 | * The access to some variables in this macro is racy. We can't use atomic_t | |
153 | * here because all these variables are exported to debugfs. Some of them even | |
154 | * writeable. This is also the reason why a lock won't help much. But anyway, | |
155 | * the races are no big deal. Here is why: | |
156 | * | |
157 | * error_count: the addition is racy, but the worst thing that can happen is | |
158 | * that we don't count some errors | |
159 | * show_num_errors: the subtraction is racy. Also no big deal because in | |
160 | * worst case this will result in one warning more in the | |
161 | * system log than the user configured. This variable is | |
162 | * writeable via debugfs. | |
163 | */ | |
6c132d1b DW |
164 | static inline void dump_entry_trace(struct dma_debug_entry *entry) |
165 | { | |
166 | #ifdef CONFIG_STACKTRACE | |
167 | if (entry) { | |
e7ed70ee | 168 | pr_warning("Mapped at:\n"); |
6c132d1b DW |
169 | print_stack_trace(&entry->stacktrace, 0); |
170 | } | |
171 | #endif | |
172 | } | |
173 | ||
2e507d84 JR |
174 | static bool driver_filter(struct device *dev) |
175 | { | |
0bf84128 JR |
176 | struct device_driver *drv; |
177 | unsigned long flags; | |
178 | bool ret; | |
179 | ||
2e507d84 JR |
180 | /* driver filter off */ |
181 | if (likely(!current_driver_name[0])) | |
182 | return true; | |
183 | ||
184 | /* driver filter on and initialized */ | |
ec9c96ef | 185 | if (current_driver && dev && dev->driver == current_driver) |
2e507d84 JR |
186 | return true; |
187 | ||
ec9c96ef KM |
188 | /* driver filter on, but we can't filter on a NULL device... */ |
189 | if (!dev) | |
190 | return false; | |
191 | ||
0bf84128 JR |
192 | if (current_driver || !current_driver_name[0]) |
193 | return false; | |
2e507d84 | 194 | |
0bf84128 | 195 | /* driver filter on but not yet initialized */ |
f3ff9247 | 196 | drv = dev->driver; |
0bf84128 JR |
197 | if (!drv) |
198 | return false; | |
199 | ||
200 | /* lock to protect against change of current_driver_name */ | |
201 | read_lock_irqsave(&driver_name_lock, flags); | |
202 | ||
203 | ret = false; | |
204 | if (drv->name && | |
205 | strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) { | |
206 | current_driver = drv; | |
207 | ret = true; | |
2e507d84 JR |
208 | } |
209 | ||
0bf84128 | 210 | read_unlock_irqrestore(&driver_name_lock, flags); |
0bf84128 JR |
211 | |
212 | return ret; | |
2e507d84 JR |
213 | } |
214 | ||
ec9c96ef KM |
215 | #define err_printk(dev, entry, format, arg...) do { \ |
216 | error_count += 1; \ | |
217 | if (driver_filter(dev) && \ | |
218 | (show_all_errors || show_num_errors > 0)) { \ | |
219 | WARN(1, "%s %s: " format, \ | |
220 | dev ? dev_driver_string(dev) : "NULL", \ | |
221 | dev ? dev_name(dev) : "NULL", ## arg); \ | |
222 | dump_entry_trace(entry); \ | |
223 | } \ | |
224 | if (!show_all_errors && show_num_errors > 0) \ | |
225 | show_num_errors -= 1; \ | |
2d62ece1 JR |
226 | } while (0); |
227 | ||
30dfa90c JR |
228 | /* |
229 | * Hash related functions | |
230 | * | |
231 | * Every DMA-API request is saved into a struct dma_debug_entry. To | |
232 | * have quick access to these structs they are stored into a hash. | |
233 | */ | |
234 | static int hash_fn(struct dma_debug_entry *entry) | |
235 | { | |
236 | /* | |
237 | * Hash function is based on the dma address. | |
238 | * We use bits 20-27 here as the index into the hash | |
239 | */ | |
240 | return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK; | |
241 | } | |
242 | ||
243 | /* | |
244 | * Request exclusive access to a hash bucket for a given dma_debug_entry. | |
245 | */ | |
246 | static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry, | |
247 | unsigned long *flags) | |
248 | { | |
249 | int idx = hash_fn(entry); | |
250 | unsigned long __flags; | |
251 | ||
252 | spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags); | |
253 | *flags = __flags; | |
254 | return &dma_entry_hash[idx]; | |
255 | } | |
256 | ||
257 | /* | |
258 | * Give up exclusive access to the hash bucket | |
259 | */ | |
260 | static void put_hash_bucket(struct hash_bucket *bucket, | |
261 | unsigned long *flags) | |
262 | { | |
263 | unsigned long __flags = *flags; | |
264 | ||
265 | spin_unlock_irqrestore(&bucket->lock, __flags); | |
266 | } | |
267 | ||
c6a21d0b NH |
268 | static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b) |
269 | { | |
91ec37cc | 270 | return ((a->dev_addr == b->dev_addr) && |
c6a21d0b NH |
271 | (a->dev == b->dev)) ? true : false; |
272 | } | |
273 | ||
274 | static bool containing_match(struct dma_debug_entry *a, | |
275 | struct dma_debug_entry *b) | |
276 | { | |
277 | if (a->dev != b->dev) | |
278 | return false; | |
279 | ||
280 | if ((b->dev_addr <= a->dev_addr) && | |
281 | ((b->dev_addr + b->size) >= (a->dev_addr + a->size))) | |
282 | return true; | |
283 | ||
284 | return false; | |
285 | } | |
286 | ||
30dfa90c JR |
287 | /* |
288 | * Search a given entry in the hash bucket list | |
289 | */ | |
c6a21d0b NH |
290 | static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket, |
291 | struct dma_debug_entry *ref, | |
292 | match_fn match) | |
30dfa90c | 293 | { |
7caf6a49 | 294 | struct dma_debug_entry *entry, *ret = NULL; |
fe73fbe1 | 295 | int matches = 0, match_lvl, last_lvl = -1; |
30dfa90c JR |
296 | |
297 | list_for_each_entry(entry, &bucket->list, list) { | |
c6a21d0b | 298 | if (!match(ref, entry)) |
7caf6a49 JR |
299 | continue; |
300 | ||
301 | /* | |
302 | * Some drivers map the same physical address multiple | |
303 | * times. Without a hardware IOMMU this results in the | |
304 | * same device addresses being put into the dma-debug | |
305 | * hash multiple times too. This can result in false | |
af901ca1 | 306 | * positives being reported. Therefore we implement a |
7caf6a49 JR |
307 | * best-fit algorithm here which returns the entry from |
308 | * the hash which fits best to the reference value | |
309 | * instead of the first-fit. | |
310 | */ | |
311 | matches += 1; | |
312 | match_lvl = 0; | |
e5e8c5b9 JR |
313 | entry->size == ref->size ? ++match_lvl : 0; |
314 | entry->type == ref->type ? ++match_lvl : 0; | |
315 | entry->direction == ref->direction ? ++match_lvl : 0; | |
316 | entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0; | |
7caf6a49 | 317 | |
e5e8c5b9 | 318 | if (match_lvl == 4) { |
7caf6a49 | 319 | /* perfect-fit - return the result */ |
30dfa90c | 320 | return entry; |
7caf6a49 JR |
321 | } else if (match_lvl > last_lvl) { |
322 | /* | |
323 | * We found an entry that fits better then the | |
fe73fbe1 | 324 | * previous one or it is the 1st match. |
7caf6a49 JR |
325 | */ |
326 | last_lvl = match_lvl; | |
327 | ret = entry; | |
328 | } | |
30dfa90c JR |
329 | } |
330 | ||
7caf6a49 JR |
331 | /* |
332 | * If we have multiple matches but no perfect-fit, just return | |
333 | * NULL. | |
334 | */ | |
335 | ret = (matches == 1) ? ret : NULL; | |
336 | ||
337 | return ret; | |
30dfa90c JR |
338 | } |
339 | ||
c6a21d0b NH |
340 | static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket, |
341 | struct dma_debug_entry *ref) | |
342 | { | |
343 | return __hash_bucket_find(bucket, ref, exact_match); | |
344 | } | |
345 | ||
346 | static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket, | |
347 | struct dma_debug_entry *ref, | |
348 | unsigned long *flags) | |
349 | { | |
350 | ||
351 | unsigned int max_range = dma_get_max_seg_size(ref->dev); | |
352 | struct dma_debug_entry *entry, index = *ref; | |
353 | unsigned int range = 0; | |
354 | ||
355 | while (range <= max_range) { | |
356 | entry = __hash_bucket_find(*bucket, &index, containing_match); | |
357 | ||
358 | if (entry) | |
359 | return entry; | |
360 | ||
361 | /* | |
362 | * Nothing found, go back a hash bucket | |
363 | */ | |
364 | put_hash_bucket(*bucket, flags); | |
365 | range += (1 << HASH_FN_SHIFT); | |
366 | index.dev_addr -= (1 << HASH_FN_SHIFT); | |
367 | *bucket = get_hash_bucket(&index, flags); | |
368 | } | |
369 | ||
370 | return NULL; | |
371 | } | |
372 | ||
30dfa90c JR |
373 | /* |
374 | * Add an entry to a hash bucket | |
375 | */ | |
376 | static void hash_bucket_add(struct hash_bucket *bucket, | |
377 | struct dma_debug_entry *entry) | |
378 | { | |
379 | list_add_tail(&entry->list, &bucket->list); | |
380 | } | |
381 | ||
382 | /* | |
383 | * Remove entry from a hash bucket list | |
384 | */ | |
385 | static void hash_bucket_del(struct dma_debug_entry *entry) | |
386 | { | |
387 | list_del(&entry->list); | |
388 | } | |
389 | ||
0abdd7a8 DW |
390 | static unsigned long long phys_addr(struct dma_debug_entry *entry) |
391 | { | |
392 | return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset; | |
393 | } | |
394 | ||
ac26c18b DW |
395 | /* |
396 | * Dump mapping entries for debugging purposes | |
397 | */ | |
398 | void debug_dma_dump_mappings(struct device *dev) | |
399 | { | |
400 | int idx; | |
401 | ||
402 | for (idx = 0; idx < HASH_SIZE; idx++) { | |
403 | struct hash_bucket *bucket = &dma_entry_hash[idx]; | |
404 | struct dma_debug_entry *entry; | |
405 | unsigned long flags; | |
406 | ||
407 | spin_lock_irqsave(&bucket->lock, flags); | |
408 | ||
409 | list_for_each_entry(entry, &bucket->list, list) { | |
410 | if (!dev || dev == entry->dev) { | |
411 | dev_info(entry->dev, | |
0abdd7a8 | 412 | "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n", |
ac26c18b | 413 | type2name[entry->type], idx, |
0abdd7a8 | 414 | phys_addr(entry), entry->pfn, |
ac26c18b | 415 | entry->dev_addr, entry->size, |
6c9c6d63 SK |
416 | dir2name[entry->direction], |
417 | maperr2str[entry->map_err_type]); | |
ac26c18b DW |
418 | } |
419 | } | |
420 | ||
421 | spin_unlock_irqrestore(&bucket->lock, flags); | |
422 | } | |
423 | } | |
424 | EXPORT_SYMBOL(debug_dma_dump_mappings); | |
425 | ||
0abdd7a8 DW |
426 | /* |
427 | * For each page mapped (initial page in the case of | |
428 | * dma_alloc_coherent/dma_map_{single|page}, or each page in a | |
429 | * scatterlist) insert into this tree using the pfn as the key. At | |
430 | * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry. If | |
431 | * the pfn already exists at insertion time add a tag as a reference | |
432 | * count for the overlapping mappings. For now, the overlap tracking | |
433 | * just ensures that 'unmaps' balance 'maps' before marking the pfn | |
434 | * idle, but we should also be flagging overlaps as an API violation. | |
435 | * | |
436 | * Memory usage is mostly constrained by the maximum number of available | |
437 | * dma-debug entries in that we need a free dma_debug_entry before | |
438 | * inserting into the tree. In the case of dma_map_{single|page} and | |
439 | * dma_alloc_coherent there is only one dma_debug_entry and one pfn to | |
440 | * track per event. dma_map_sg(), on the other hand, | |
441 | * consumes a single dma_debug_entry, but inserts 'nents' entries into | |
442 | * the tree. | |
443 | * | |
444 | * At any time debug_dma_assert_idle() can be called to trigger a | |
445 | * warning if the given page is in the active set. | |
446 | */ | |
447 | static RADIX_TREE(dma_active_pfn, GFP_NOWAIT); | |
448 | static DEFINE_SPINLOCK(radix_lock); | |
449 | #define ACTIVE_PFN_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1) | |
450 | ||
451 | static int active_pfn_read_overlap(unsigned long pfn) | |
452 | { | |
453 | int overlap = 0, i; | |
454 | ||
455 | for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--) | |
456 | if (radix_tree_tag_get(&dma_active_pfn, pfn, i)) | |
457 | overlap |= 1 << i; | |
458 | return overlap; | |
459 | } | |
460 | ||
461 | static int active_pfn_set_overlap(unsigned long pfn, int overlap) | |
462 | { | |
463 | int i; | |
464 | ||
465 | if (overlap > ACTIVE_PFN_MAX_OVERLAP || overlap < 0) | |
59f2e7df | 466 | return overlap; |
0abdd7a8 DW |
467 | |
468 | for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--) | |
469 | if (overlap & 1 << i) | |
470 | radix_tree_tag_set(&dma_active_pfn, pfn, i); | |
471 | else | |
472 | radix_tree_tag_clear(&dma_active_pfn, pfn, i); | |
473 | ||
474 | return overlap; | |
475 | } | |
476 | ||
477 | static void active_pfn_inc_overlap(unsigned long pfn) | |
478 | { | |
479 | int overlap = active_pfn_read_overlap(pfn); | |
480 | ||
481 | overlap = active_pfn_set_overlap(pfn, ++overlap); | |
482 | ||
483 | /* If we overflowed the overlap counter then we're potentially | |
484 | * leaking dma-mappings. Otherwise, if maps and unmaps are | |
485 | * balanced then this overflow may cause false negatives in | |
486 | * debug_dma_assert_idle() as the pfn may be marked idle | |
487 | * prematurely. | |
488 | */ | |
59f2e7df | 489 | WARN_ONCE(overlap > ACTIVE_PFN_MAX_OVERLAP, |
0abdd7a8 DW |
490 | "DMA-API: exceeded %d overlapping mappings of pfn %lx\n", |
491 | ACTIVE_PFN_MAX_OVERLAP, pfn); | |
492 | } | |
493 | ||
494 | static int active_pfn_dec_overlap(unsigned long pfn) | |
495 | { | |
496 | int overlap = active_pfn_read_overlap(pfn); | |
497 | ||
498 | return active_pfn_set_overlap(pfn, --overlap); | |
499 | } | |
500 | ||
501 | static int active_pfn_insert(struct dma_debug_entry *entry) | |
502 | { | |
503 | unsigned long flags; | |
504 | int rc; | |
505 | ||
506 | spin_lock_irqsave(&radix_lock, flags); | |
507 | rc = radix_tree_insert(&dma_active_pfn, entry->pfn, entry); | |
508 | if (rc == -EEXIST) | |
509 | active_pfn_inc_overlap(entry->pfn); | |
510 | spin_unlock_irqrestore(&radix_lock, flags); | |
511 | ||
512 | return rc; | |
513 | } | |
514 | ||
515 | static void active_pfn_remove(struct dma_debug_entry *entry) | |
516 | { | |
517 | unsigned long flags; | |
518 | ||
519 | spin_lock_irqsave(&radix_lock, flags); | |
59f2e7df DW |
520 | /* since we are counting overlaps the final put of the |
521 | * entry->pfn will occur when the overlap count is 0. | |
522 | * active_pfn_dec_overlap() returns -1 in that case | |
523 | */ | |
524 | if (active_pfn_dec_overlap(entry->pfn) < 0) | |
0abdd7a8 DW |
525 | radix_tree_delete(&dma_active_pfn, entry->pfn); |
526 | spin_unlock_irqrestore(&radix_lock, flags); | |
527 | } | |
528 | ||
529 | /** | |
530 | * debug_dma_assert_idle() - assert that a page is not undergoing dma | |
531 | * @page: page to lookup in the dma_active_pfn tree | |
532 | * | |
533 | * Place a call to this routine in cases where the cpu touching the page | |
534 | * before the dma completes (page is dma_unmapped) will lead to data | |
535 | * corruption. | |
536 | */ | |
537 | void debug_dma_assert_idle(struct page *page) | |
538 | { | |
539 | unsigned long flags; | |
540 | struct dma_debug_entry *entry; | |
541 | ||
542 | if (!page) | |
543 | return; | |
544 | ||
545 | spin_lock_irqsave(&radix_lock, flags); | |
546 | entry = radix_tree_lookup(&dma_active_pfn, page_to_pfn(page)); | |
547 | spin_unlock_irqrestore(&radix_lock, flags); | |
548 | ||
549 | if (!entry) | |
550 | return; | |
551 | ||
552 | err_printk(entry->dev, entry, | |
553 | "DMA-API: cpu touching an active dma mapped page " | |
554 | "[pfn=0x%lx]\n", entry->pfn); | |
555 | } | |
556 | ||
30dfa90c JR |
557 | /* |
558 | * Wrapper function for adding an entry to the hash. | |
559 | * This function takes care of locking itself. | |
560 | */ | |
561 | static void add_dma_entry(struct dma_debug_entry *entry) | |
562 | { | |
563 | struct hash_bucket *bucket; | |
564 | unsigned long flags; | |
0abdd7a8 | 565 | int rc; |
30dfa90c JR |
566 | |
567 | bucket = get_hash_bucket(entry, &flags); | |
568 | hash_bucket_add(bucket, entry); | |
569 | put_hash_bucket(bucket, &flags); | |
0abdd7a8 DW |
570 | |
571 | rc = active_pfn_insert(entry); | |
572 | if (rc == -ENOMEM) { | |
573 | pr_err("DMA-API: pfn tracking ENOMEM, dma-debug disabled\n"); | |
574 | global_disable = true; | |
575 | } | |
576 | ||
577 | /* TODO: report -EEXIST errors here as overlapping mappings are | |
578 | * not supported by the DMA API | |
579 | */ | |
30dfa90c JR |
580 | } |
581 | ||
e6a1a89d FT |
582 | static struct dma_debug_entry *__dma_entry_alloc(void) |
583 | { | |
584 | struct dma_debug_entry *entry; | |
585 | ||
586 | entry = list_entry(free_entries.next, struct dma_debug_entry, list); | |
587 | list_del(&entry->list); | |
588 | memset(entry, 0, sizeof(*entry)); | |
589 | ||
590 | num_free_entries -= 1; | |
591 | if (num_free_entries < min_free_entries) | |
592 | min_free_entries = num_free_entries; | |
593 | ||
594 | return entry; | |
595 | } | |
596 | ||
3b1e79ed JR |
597 | /* struct dma_entry allocator |
598 | * | |
599 | * The next two functions implement the allocator for | |
600 | * struct dma_debug_entries. | |
601 | */ | |
602 | static struct dma_debug_entry *dma_entry_alloc(void) | |
603 | { | |
29cdd4e4 | 604 | struct dma_debug_entry *entry; |
3b1e79ed JR |
605 | unsigned long flags; |
606 | ||
607 | spin_lock_irqsave(&free_entries_lock, flags); | |
608 | ||
609 | if (list_empty(&free_entries)) { | |
e7ed70ee | 610 | pr_err("DMA-API: debugging out of memory - disabling\n"); |
3b1e79ed | 611 | global_disable = true; |
29cdd4e4 JK |
612 | spin_unlock_irqrestore(&free_entries_lock, flags); |
613 | return NULL; | |
3b1e79ed JR |
614 | } |
615 | ||
e6a1a89d | 616 | entry = __dma_entry_alloc(); |
3b1e79ed | 617 | |
29cdd4e4 JK |
618 | spin_unlock_irqrestore(&free_entries_lock, flags); |
619 | ||
6c132d1b DW |
620 | #ifdef CONFIG_STACKTRACE |
621 | entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES; | |
622 | entry->stacktrace.entries = entry->st_entries; | |
623 | entry->stacktrace.skip = 2; | |
624 | save_stack_trace(&entry->stacktrace); | |
625 | #endif | |
3b1e79ed | 626 | |
3b1e79ed JR |
627 | return entry; |
628 | } | |
629 | ||
630 | static void dma_entry_free(struct dma_debug_entry *entry) | |
631 | { | |
632 | unsigned long flags; | |
633 | ||
0abdd7a8 DW |
634 | active_pfn_remove(entry); |
635 | ||
3b1e79ed JR |
636 | /* |
637 | * add to beginning of the list - this way the entries are | |
638 | * more likely cache hot when they are reallocated. | |
639 | */ | |
640 | spin_lock_irqsave(&free_entries_lock, flags); | |
641 | list_add(&entry->list, &free_entries); | |
642 | num_free_entries += 1; | |
643 | spin_unlock_irqrestore(&free_entries_lock, flags); | |
644 | } | |
645 | ||
e6a1a89d FT |
646 | int dma_debug_resize_entries(u32 num_entries) |
647 | { | |
648 | int i, delta, ret = 0; | |
649 | unsigned long flags; | |
650 | struct dma_debug_entry *entry; | |
651 | LIST_HEAD(tmp); | |
652 | ||
653 | spin_lock_irqsave(&free_entries_lock, flags); | |
654 | ||
655 | if (nr_total_entries < num_entries) { | |
656 | delta = num_entries - nr_total_entries; | |
657 | ||
658 | spin_unlock_irqrestore(&free_entries_lock, flags); | |
659 | ||
660 | for (i = 0; i < delta; i++) { | |
661 | entry = kzalloc(sizeof(*entry), GFP_KERNEL); | |
662 | if (!entry) | |
663 | break; | |
664 | ||
665 | list_add_tail(&entry->list, &tmp); | |
666 | } | |
667 | ||
668 | spin_lock_irqsave(&free_entries_lock, flags); | |
669 | ||
670 | list_splice(&tmp, &free_entries); | |
671 | nr_total_entries += i; | |
672 | num_free_entries += i; | |
673 | } else { | |
674 | delta = nr_total_entries - num_entries; | |
675 | ||
676 | for (i = 0; i < delta && !list_empty(&free_entries); i++) { | |
677 | entry = __dma_entry_alloc(); | |
678 | kfree(entry); | |
679 | } | |
680 | ||
681 | nr_total_entries -= i; | |
682 | } | |
683 | ||
684 | if (nr_total_entries != num_entries) | |
685 | ret = 1; | |
686 | ||
687 | spin_unlock_irqrestore(&free_entries_lock, flags); | |
688 | ||
689 | return ret; | |
690 | } | |
691 | EXPORT_SYMBOL(dma_debug_resize_entries); | |
692 | ||
6bf07871 JR |
693 | /* |
694 | * DMA-API debugging init code | |
695 | * | |
696 | * The init code does two things: | |
697 | * 1. Initialize core data structures | |
698 | * 2. Preallocate a given number of dma_debug_entry structs | |
699 | */ | |
700 | ||
701 | static int prealloc_memory(u32 num_entries) | |
702 | { | |
703 | struct dma_debug_entry *entry, *next_entry; | |
704 | int i; | |
705 | ||
706 | for (i = 0; i < num_entries; ++i) { | |
707 | entry = kzalloc(sizeof(*entry), GFP_KERNEL); | |
708 | if (!entry) | |
709 | goto out_err; | |
710 | ||
711 | list_add_tail(&entry->list, &free_entries); | |
712 | } | |
713 | ||
714 | num_free_entries = num_entries; | |
715 | min_free_entries = num_entries; | |
716 | ||
e7ed70ee | 717 | pr_info("DMA-API: preallocated %d debug entries\n", num_entries); |
6bf07871 JR |
718 | |
719 | return 0; | |
720 | ||
721 | out_err: | |
722 | ||
723 | list_for_each_entry_safe(entry, next_entry, &free_entries, list) { | |
724 | list_del(&entry->list); | |
725 | kfree(entry); | |
726 | } | |
727 | ||
728 | return -ENOMEM; | |
729 | } | |
730 | ||
8a6fc708 JR |
731 | static ssize_t filter_read(struct file *file, char __user *user_buf, |
732 | size_t count, loff_t *ppos) | |
733 | { | |
8a6fc708 | 734 | char buf[NAME_MAX_LEN + 1]; |
c17e2cf7 | 735 | unsigned long flags; |
8a6fc708 JR |
736 | int len; |
737 | ||
738 | if (!current_driver_name[0]) | |
739 | return 0; | |
740 | ||
741 | /* | |
742 | * We can't copy to userspace directly because current_driver_name can | |
743 | * only be read under the driver_name_lock with irqs disabled. So | |
744 | * create a temporary copy first. | |
745 | */ | |
746 | read_lock_irqsave(&driver_name_lock, flags); | |
747 | len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name); | |
748 | read_unlock_irqrestore(&driver_name_lock, flags); | |
749 | ||
750 | return simple_read_from_buffer(user_buf, count, ppos, buf, len); | |
751 | } | |
752 | ||
753 | static ssize_t filter_write(struct file *file, const char __user *userbuf, | |
754 | size_t count, loff_t *ppos) | |
755 | { | |
8a6fc708 | 756 | char buf[NAME_MAX_LEN]; |
c17e2cf7 JR |
757 | unsigned long flags; |
758 | size_t len; | |
8a6fc708 JR |
759 | int i; |
760 | ||
761 | /* | |
762 | * We can't copy from userspace directly. Access to | |
763 | * current_driver_name is protected with a write_lock with irqs | |
764 | * disabled. Since copy_from_user can fault and may sleep we | |
765 | * need to copy to temporary buffer first | |
766 | */ | |
e7ed70ee | 767 | len = min(count, (size_t)(NAME_MAX_LEN - 1)); |
8a6fc708 JR |
768 | if (copy_from_user(buf, userbuf, len)) |
769 | return -EFAULT; | |
770 | ||
771 | buf[len] = 0; | |
772 | ||
773 | write_lock_irqsave(&driver_name_lock, flags); | |
774 | ||
31232509 JR |
775 | /* |
776 | * Now handle the string we got from userspace very carefully. | |
8a6fc708 JR |
777 | * The rules are: |
778 | * - only use the first token we got | |
779 | * - token delimiter is everything looking like a space | |
780 | * character (' ', '\n', '\t' ...) | |
781 | * | |
782 | */ | |
783 | if (!isalnum(buf[0])) { | |
784 | /* | |
31232509 | 785 | * If the first character userspace gave us is not |
8a6fc708 JR |
786 | * alphanumerical then assume the filter should be |
787 | * switched off. | |
788 | */ | |
789 | if (current_driver_name[0]) | |
e7ed70ee | 790 | pr_info("DMA-API: switching off dma-debug driver filter\n"); |
8a6fc708 JR |
791 | current_driver_name[0] = 0; |
792 | current_driver = NULL; | |
793 | goto out_unlock; | |
794 | } | |
795 | ||
796 | /* | |
797 | * Now parse out the first token and use it as the name for the | |
798 | * driver to filter for. | |
799 | */ | |
39a37ce1 | 800 | for (i = 0; i < NAME_MAX_LEN - 1; ++i) { |
8a6fc708 JR |
801 | current_driver_name[i] = buf[i]; |
802 | if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0) | |
803 | break; | |
804 | } | |
805 | current_driver_name[i] = 0; | |
806 | current_driver = NULL; | |
807 | ||
e7ed70ee JR |
808 | pr_info("DMA-API: enable driver filter for driver [%s]\n", |
809 | current_driver_name); | |
8a6fc708 JR |
810 | |
811 | out_unlock: | |
812 | write_unlock_irqrestore(&driver_name_lock, flags); | |
813 | ||
814 | return count; | |
815 | } | |
816 | ||
aeb583d0 | 817 | static const struct file_operations filter_fops = { |
8a6fc708 JR |
818 | .read = filter_read, |
819 | .write = filter_write, | |
6038f373 | 820 | .llseek = default_llseek, |
8a6fc708 JR |
821 | }; |
822 | ||
788dcfa6 JR |
823 | static int dma_debug_fs_init(void) |
824 | { | |
825 | dma_debug_dent = debugfs_create_dir("dma-api", NULL); | |
826 | if (!dma_debug_dent) { | |
e7ed70ee | 827 | pr_err("DMA-API: can not create debugfs directory\n"); |
788dcfa6 JR |
828 | return -ENOMEM; |
829 | } | |
830 | ||
831 | global_disable_dent = debugfs_create_bool("disabled", 0444, | |
832 | dma_debug_dent, | |
68ee6d22 | 833 | &global_disable); |
788dcfa6 JR |
834 | if (!global_disable_dent) |
835 | goto out_err; | |
836 | ||
837 | error_count_dent = debugfs_create_u32("error_count", 0444, | |
838 | dma_debug_dent, &error_count); | |
839 | if (!error_count_dent) | |
840 | goto out_err; | |
841 | ||
842 | show_all_errors_dent = debugfs_create_u32("all_errors", 0644, | |
843 | dma_debug_dent, | |
844 | &show_all_errors); | |
845 | if (!show_all_errors_dent) | |
846 | goto out_err; | |
847 | ||
848 | show_num_errors_dent = debugfs_create_u32("num_errors", 0644, | |
849 | dma_debug_dent, | |
850 | &show_num_errors); | |
851 | if (!show_num_errors_dent) | |
852 | goto out_err; | |
853 | ||
854 | num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444, | |
855 | dma_debug_dent, | |
856 | &num_free_entries); | |
857 | if (!num_free_entries_dent) | |
858 | goto out_err; | |
859 | ||
860 | min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444, | |
861 | dma_debug_dent, | |
862 | &min_free_entries); | |
863 | if (!min_free_entries_dent) | |
864 | goto out_err; | |
865 | ||
8a6fc708 JR |
866 | filter_dent = debugfs_create_file("driver_filter", 0644, |
867 | dma_debug_dent, NULL, &filter_fops); | |
868 | if (!filter_dent) | |
869 | goto out_err; | |
870 | ||
788dcfa6 JR |
871 | return 0; |
872 | ||
873 | out_err: | |
874 | debugfs_remove_recursive(dma_debug_dent); | |
875 | ||
876 | return -ENOMEM; | |
877 | } | |
878 | ||
ba4b87ad | 879 | static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry) |
ed888aef JR |
880 | { |
881 | struct dma_debug_entry *entry; | |
882 | unsigned long flags; | |
883 | int count = 0, i; | |
884 | ||
be81c6ea JR |
885 | local_irq_save(flags); |
886 | ||
ed888aef | 887 | for (i = 0; i < HASH_SIZE; ++i) { |
be81c6ea | 888 | spin_lock(&dma_entry_hash[i].lock); |
ed888aef | 889 | list_for_each_entry(entry, &dma_entry_hash[i].list, list) { |
ba4b87ad | 890 | if (entry->dev == dev) { |
ed888aef | 891 | count += 1; |
ba4b87ad SG |
892 | *out_entry = entry; |
893 | } | |
ed888aef | 894 | } |
be81c6ea | 895 | spin_unlock(&dma_entry_hash[i].lock); |
ed888aef JR |
896 | } |
897 | ||
be81c6ea JR |
898 | local_irq_restore(flags); |
899 | ||
ed888aef JR |
900 | return count; |
901 | } | |
902 | ||
a8fe9ea2 | 903 | static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data) |
ed888aef JR |
904 | { |
905 | struct device *dev = data; | |
ba4b87ad | 906 | struct dma_debug_entry *uninitialized_var(entry); |
ed888aef JR |
907 | int count; |
908 | ||
f797d988 | 909 | if (global_disable) |
a8fe9ea2 | 910 | return 0; |
ed888aef JR |
911 | |
912 | switch (action) { | |
913 | case BUS_NOTIFY_UNBOUND_DRIVER: | |
ba4b87ad | 914 | count = device_dma_allocations(dev, &entry); |
ed888aef JR |
915 | if (count == 0) |
916 | break; | |
ba4b87ad | 917 | err_printk(dev, entry, "DMA-API: device driver has pending " |
ed888aef | 918 | "DMA allocations while released from device " |
ba4b87ad SG |
919 | "[count=%d]\n" |
920 | "One of leaked entries details: " | |
921 | "[device address=0x%016llx] [size=%llu bytes] " | |
922 | "[mapped with %s] [mapped as %s]\n", | |
923 | count, entry->dev_addr, entry->size, | |
924 | dir2name[entry->direction], type2name[entry->type]); | |
ed888aef JR |
925 | break; |
926 | default: | |
927 | break; | |
928 | } | |
929 | ||
930 | return 0; | |
931 | } | |
932 | ||
41531c8f JR |
933 | void dma_debug_add_bus(struct bus_type *bus) |
934 | { | |
ed888aef JR |
935 | struct notifier_block *nb; |
936 | ||
f797d988 SR |
937 | if (global_disable) |
938 | return; | |
939 | ||
ed888aef JR |
940 | nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL); |
941 | if (nb == NULL) { | |
e7ed70ee | 942 | pr_err("dma_debug_add_bus: out of memory\n"); |
ed888aef JR |
943 | return; |
944 | } | |
945 | ||
946 | nb->notifier_call = dma_debug_device_change; | |
947 | ||
948 | bus_register_notifier(bus, nb); | |
41531c8f | 949 | } |
788dcfa6 | 950 | |
6bf07871 JR |
951 | /* |
952 | * Let the architectures decide how many entries should be preallocated. | |
953 | */ | |
954 | void dma_debug_init(u32 num_entries) | |
955 | { | |
956 | int i; | |
957 | ||
958 | if (global_disable) | |
959 | return; | |
960 | ||
961 | for (i = 0; i < HASH_SIZE; ++i) { | |
962 | INIT_LIST_HEAD(&dma_entry_hash[i].list); | |
b0a5b83e | 963 | spin_lock_init(&dma_entry_hash[i].lock); |
6bf07871 JR |
964 | } |
965 | ||
788dcfa6 | 966 | if (dma_debug_fs_init() != 0) { |
e7ed70ee | 967 | pr_err("DMA-API: error creating debugfs entries - disabling\n"); |
788dcfa6 JR |
968 | global_disable = true; |
969 | ||
970 | return; | |
971 | } | |
972 | ||
59d3daaf JR |
973 | if (req_entries) |
974 | num_entries = req_entries; | |
975 | ||
6bf07871 | 976 | if (prealloc_memory(num_entries) != 0) { |
e7ed70ee | 977 | pr_err("DMA-API: debugging out of memory error - disabled\n"); |
6bf07871 JR |
978 | global_disable = true; |
979 | ||
980 | return; | |
981 | } | |
982 | ||
e6a1a89d FT |
983 | nr_total_entries = num_free_entries; |
984 | ||
e7ed70ee | 985 | pr_info("DMA-API: debugging enabled by kernel config\n"); |
6bf07871 JR |
986 | } |
987 | ||
59d3daaf JR |
988 | static __init int dma_debug_cmdline(char *str) |
989 | { | |
990 | if (!str) | |
991 | return -EINVAL; | |
992 | ||
993 | if (strncmp(str, "off", 3) == 0) { | |
e7ed70ee | 994 | pr_info("DMA-API: debugging disabled on kernel command line\n"); |
59d3daaf JR |
995 | global_disable = true; |
996 | } | |
997 | ||
998 | return 0; | |
999 | } | |
1000 | ||
1001 | static __init int dma_debug_entries_cmdline(char *str) | |
1002 | { | |
1003 | int res; | |
1004 | ||
1005 | if (!str) | |
1006 | return -EINVAL; | |
1007 | ||
1008 | res = get_option(&str, &req_entries); | |
1009 | ||
1010 | if (!res) | |
1011 | req_entries = 0; | |
1012 | ||
1013 | return 0; | |
1014 | } | |
1015 | ||
1016 | __setup("dma_debug=", dma_debug_cmdline); | |
1017 | __setup("dma_debug_entries=", dma_debug_entries_cmdline); | |
1018 | ||
2d62ece1 JR |
1019 | static void check_unmap(struct dma_debug_entry *ref) |
1020 | { | |
1021 | struct dma_debug_entry *entry; | |
1022 | struct hash_bucket *bucket; | |
1023 | unsigned long flags; | |
1024 | ||
2d62ece1 | 1025 | bucket = get_hash_bucket(ref, &flags); |
c6a21d0b | 1026 | entry = bucket_find_exact(bucket, ref); |
2d62ece1 JR |
1027 | |
1028 | if (!entry) { | |
8d640a51 AD |
1029 | /* must drop lock before calling dma_mapping_error */ |
1030 | put_hash_bucket(bucket, &flags); | |
1031 | ||
bfe0fb0f SK |
1032 | if (dma_mapping_error(ref->dev, ref->dev_addr)) { |
1033 | err_printk(ref->dev, NULL, | |
8d640a51 AD |
1034 | "DMA-API: device driver tries to free an " |
1035 | "invalid DMA memory address\n"); | |
1036 | } else { | |
1037 | err_printk(ref->dev, NULL, | |
1038 | "DMA-API: device driver tries to free DMA " | |
1039 | "memory it has not allocated [device " | |
1040 | "address=0x%016llx] [size=%llu bytes]\n", | |
1041 | ref->dev_addr, ref->size); | |
bfe0fb0f | 1042 | } |
8d640a51 | 1043 | return; |
2d62ece1 JR |
1044 | } |
1045 | ||
1046 | if (ref->size != entry->size) { | |
6c132d1b | 1047 | err_printk(ref->dev, entry, "DMA-API: device driver frees " |
2d62ece1 JR |
1048 | "DMA memory with different size " |
1049 | "[device address=0x%016llx] [map size=%llu bytes] " | |
1050 | "[unmap size=%llu bytes]\n", | |
1051 | ref->dev_addr, entry->size, ref->size); | |
1052 | } | |
1053 | ||
1054 | if (ref->type != entry->type) { | |
6c132d1b | 1055 | err_printk(ref->dev, entry, "DMA-API: device driver frees " |
2d62ece1 JR |
1056 | "DMA memory with wrong function " |
1057 | "[device address=0x%016llx] [size=%llu bytes] " | |
1058 | "[mapped as %s] [unmapped as %s]\n", | |
1059 | ref->dev_addr, ref->size, | |
1060 | type2name[entry->type], type2name[ref->type]); | |
1061 | } else if ((entry->type == dma_debug_coherent) && | |
0abdd7a8 | 1062 | (phys_addr(ref) != phys_addr(entry))) { |
6c132d1b | 1063 | err_printk(ref->dev, entry, "DMA-API: device driver frees " |
2d62ece1 JR |
1064 | "DMA memory with different CPU address " |
1065 | "[device address=0x%016llx] [size=%llu bytes] " | |
59a40e70 JR |
1066 | "[cpu alloc address=0x%016llx] " |
1067 | "[cpu free address=0x%016llx]", | |
2d62ece1 | 1068 | ref->dev_addr, ref->size, |
0abdd7a8 DW |
1069 | phys_addr(entry), |
1070 | phys_addr(ref)); | |
2d62ece1 JR |
1071 | } |
1072 | ||
1073 | if (ref->sg_call_ents && ref->type == dma_debug_sg && | |
1074 | ref->sg_call_ents != entry->sg_call_ents) { | |
6c132d1b | 1075 | err_printk(ref->dev, entry, "DMA-API: device driver frees " |
2d62ece1 JR |
1076 | "DMA sg list with different entry count " |
1077 | "[map count=%d] [unmap count=%d]\n", | |
1078 | entry->sg_call_ents, ref->sg_call_ents); | |
1079 | } | |
1080 | ||
1081 | /* | |
1082 | * This may be no bug in reality - but most implementations of the | |
1083 | * DMA API don't handle this properly, so check for it here | |
1084 | */ | |
1085 | if (ref->direction != entry->direction) { | |
6c132d1b | 1086 | err_printk(ref->dev, entry, "DMA-API: device driver frees " |
2d62ece1 JR |
1087 | "DMA memory with different direction " |
1088 | "[device address=0x%016llx] [size=%llu bytes] " | |
1089 | "[mapped with %s] [unmapped with %s]\n", | |
1090 | ref->dev_addr, ref->size, | |
1091 | dir2name[entry->direction], | |
1092 | dir2name[ref->direction]); | |
1093 | } | |
1094 | ||
6c9c6d63 SK |
1095 | if (entry->map_err_type == MAP_ERR_NOT_CHECKED) { |
1096 | err_printk(ref->dev, entry, | |
1097 | "DMA-API: device driver failed to check map error" | |
1098 | "[device address=0x%016llx] [size=%llu bytes] " | |
1099 | "[mapped as %s]", | |
1100 | ref->dev_addr, ref->size, | |
1101 | type2name[entry->type]); | |
1102 | } | |
1103 | ||
2d62ece1 JR |
1104 | hash_bucket_del(entry); |
1105 | dma_entry_free(entry); | |
1106 | ||
2d62ece1 JR |
1107 | put_hash_bucket(bucket, &flags); |
1108 | } | |
1109 | ||
1110 | static void check_for_stack(struct device *dev, void *addr) | |
1111 | { | |
1112 | if (object_is_on_stack(addr)) | |
6c132d1b DW |
1113 | err_printk(dev, NULL, "DMA-API: device driver maps memory from" |
1114 | "stack [addr=%p]\n", addr); | |
2d62ece1 JR |
1115 | } |
1116 | ||
f39d1b97 | 1117 | static inline bool overlap(void *addr, unsigned long len, void *start, void *end) |
2e34bde1 | 1118 | { |
f39d1b97 IM |
1119 | unsigned long a1 = (unsigned long)addr; |
1120 | unsigned long b1 = a1 + len; | |
1121 | unsigned long a2 = (unsigned long)start; | |
1122 | unsigned long b2 = (unsigned long)end; | |
2e34bde1 | 1123 | |
f39d1b97 | 1124 | return !(b1 <= a2 || a1 >= b2); |
2e34bde1 JR |
1125 | } |
1126 | ||
f39d1b97 | 1127 | static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len) |
2e34bde1 | 1128 | { |
f39d1b97 IM |
1129 | if (overlap(addr, len, _text, _etext) || |
1130 | overlap(addr, len, __start_rodata, __end_rodata)) | |
1131 | err_printk(dev, NULL, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len); | |
2e34bde1 JR |
1132 | } |
1133 | ||
aa010efb JR |
1134 | static void check_sync(struct device *dev, |
1135 | struct dma_debug_entry *ref, | |
1136 | bool to_cpu) | |
2d62ece1 | 1137 | { |
2d62ece1 JR |
1138 | struct dma_debug_entry *entry; |
1139 | struct hash_bucket *bucket; | |
1140 | unsigned long flags; | |
1141 | ||
aa010efb | 1142 | bucket = get_hash_bucket(ref, &flags); |
2d62ece1 | 1143 | |
c6a21d0b | 1144 | entry = bucket_find_contain(&bucket, ref, &flags); |
2d62ece1 JR |
1145 | |
1146 | if (!entry) { | |
6c132d1b | 1147 | err_printk(dev, NULL, "DMA-API: device driver tries " |
2d62ece1 JR |
1148 | "to sync DMA memory it has not allocated " |
1149 | "[device address=0x%016llx] [size=%llu bytes]\n", | |
aa010efb | 1150 | (unsigned long long)ref->dev_addr, ref->size); |
2d62ece1 JR |
1151 | goto out; |
1152 | } | |
1153 | ||
aa010efb | 1154 | if (ref->size > entry->size) { |
6c132d1b | 1155 | err_printk(dev, entry, "DMA-API: device driver syncs" |
2d62ece1 JR |
1156 | " DMA memory outside allocated range " |
1157 | "[device address=0x%016llx] " | |
aa010efb JR |
1158 | "[allocation size=%llu bytes] " |
1159 | "[sync offset+size=%llu]\n", | |
1160 | entry->dev_addr, entry->size, | |
1161 | ref->size); | |
2d62ece1 JR |
1162 | } |
1163 | ||
42d53b4f KH |
1164 | if (entry->direction == DMA_BIDIRECTIONAL) |
1165 | goto out; | |
1166 | ||
aa010efb | 1167 | if (ref->direction != entry->direction) { |
6c132d1b | 1168 | err_printk(dev, entry, "DMA-API: device driver syncs " |
2d62ece1 JR |
1169 | "DMA memory with different direction " |
1170 | "[device address=0x%016llx] [size=%llu bytes] " | |
1171 | "[mapped with %s] [synced with %s]\n", | |
aa010efb | 1172 | (unsigned long long)ref->dev_addr, entry->size, |
2d62ece1 | 1173 | dir2name[entry->direction], |
aa010efb | 1174 | dir2name[ref->direction]); |
2d62ece1 JR |
1175 | } |
1176 | ||
2d62ece1 | 1177 | if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) && |
aa010efb | 1178 | !(ref->direction == DMA_TO_DEVICE)) |
6c132d1b | 1179 | err_printk(dev, entry, "DMA-API: device driver syncs " |
2d62ece1 JR |
1180 | "device read-only DMA memory for cpu " |
1181 | "[device address=0x%016llx] [size=%llu bytes] " | |
1182 | "[mapped with %s] [synced with %s]\n", | |
aa010efb | 1183 | (unsigned long long)ref->dev_addr, entry->size, |
2d62ece1 | 1184 | dir2name[entry->direction], |
aa010efb | 1185 | dir2name[ref->direction]); |
2d62ece1 JR |
1186 | |
1187 | if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) && | |
aa010efb | 1188 | !(ref->direction == DMA_FROM_DEVICE)) |
6c132d1b | 1189 | err_printk(dev, entry, "DMA-API: device driver syncs " |
2d62ece1 JR |
1190 | "device write-only DMA memory to device " |
1191 | "[device address=0x%016llx] [size=%llu bytes] " | |
1192 | "[mapped with %s] [synced with %s]\n", | |
aa010efb | 1193 | (unsigned long long)ref->dev_addr, entry->size, |
2d62ece1 | 1194 | dir2name[entry->direction], |
aa010efb | 1195 | dir2name[ref->direction]); |
2d62ece1 JR |
1196 | |
1197 | out: | |
1198 | put_hash_bucket(bucket, &flags); | |
2d62ece1 JR |
1199 | } |
1200 | ||
f62bc980 JR |
1201 | void debug_dma_map_page(struct device *dev, struct page *page, size_t offset, |
1202 | size_t size, int direction, dma_addr_t dma_addr, | |
1203 | bool map_single) | |
1204 | { | |
1205 | struct dma_debug_entry *entry; | |
1206 | ||
1207 | if (unlikely(global_disable)) | |
1208 | return; | |
1209 | ||
bfe0fb0f | 1210 | if (dma_mapping_error(dev, dma_addr)) |
f62bc980 JR |
1211 | return; |
1212 | ||
1213 | entry = dma_entry_alloc(); | |
1214 | if (!entry) | |
1215 | return; | |
1216 | ||
1217 | entry->dev = dev; | |
1218 | entry->type = dma_debug_page; | |
0abdd7a8 DW |
1219 | entry->pfn = page_to_pfn(page); |
1220 | entry->offset = offset, | |
f62bc980 JR |
1221 | entry->dev_addr = dma_addr; |
1222 | entry->size = size; | |
1223 | entry->direction = direction; | |
6c9c6d63 | 1224 | entry->map_err_type = MAP_ERR_NOT_CHECKED; |
f62bc980 | 1225 | |
9537a48e | 1226 | if (map_single) |
f62bc980 | 1227 | entry->type = dma_debug_single; |
9537a48e JR |
1228 | |
1229 | if (!PageHighMem(page)) { | |
f39d1b97 IM |
1230 | void *addr = page_address(page) + offset; |
1231 | ||
2e34bde1 JR |
1232 | check_for_stack(dev, addr); |
1233 | check_for_illegal_area(dev, addr, size); | |
f62bc980 JR |
1234 | } |
1235 | ||
1236 | add_dma_entry(entry); | |
1237 | } | |
1238 | EXPORT_SYMBOL(debug_dma_map_page); | |
1239 | ||
6c9c6d63 SK |
1240 | void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr) |
1241 | { | |
1242 | struct dma_debug_entry ref; | |
1243 | struct dma_debug_entry *entry; | |
1244 | struct hash_bucket *bucket; | |
1245 | unsigned long flags; | |
1246 | ||
1247 | if (unlikely(global_disable)) | |
1248 | return; | |
1249 | ||
1250 | ref.dev = dev; | |
1251 | ref.dev_addr = dma_addr; | |
1252 | bucket = get_hash_bucket(&ref, &flags); | |
6c9c6d63 | 1253 | |
96e7d7a1 AD |
1254 | list_for_each_entry(entry, &bucket->list, list) { |
1255 | if (!exact_match(&ref, entry)) | |
1256 | continue; | |
1257 | ||
1258 | /* | |
1259 | * The same physical address can be mapped multiple | |
1260 | * times. Without a hardware IOMMU this results in the | |
1261 | * same device addresses being put into the dma-debug | |
1262 | * hash multiple times too. This can result in false | |
1263 | * positives being reported. Therefore we implement a | |
1264 | * best-fit algorithm here which updates the first entry | |
1265 | * from the hash which fits the reference value and is | |
1266 | * not currently listed as being checked. | |
1267 | */ | |
1268 | if (entry->map_err_type == MAP_ERR_NOT_CHECKED) { | |
1269 | entry->map_err_type = MAP_ERR_CHECKED; | |
1270 | break; | |
1271 | } | |
1272 | } | |
6c9c6d63 | 1273 | |
6c9c6d63 SK |
1274 | put_hash_bucket(bucket, &flags); |
1275 | } | |
1276 | EXPORT_SYMBOL(debug_dma_mapping_error); | |
1277 | ||
f62bc980 JR |
1278 | void debug_dma_unmap_page(struct device *dev, dma_addr_t addr, |
1279 | size_t size, int direction, bool map_single) | |
1280 | { | |
1281 | struct dma_debug_entry ref = { | |
1282 | .type = dma_debug_page, | |
1283 | .dev = dev, | |
1284 | .dev_addr = addr, | |
1285 | .size = size, | |
1286 | .direction = direction, | |
1287 | }; | |
1288 | ||
1289 | if (unlikely(global_disable)) | |
1290 | return; | |
1291 | ||
1292 | if (map_single) | |
1293 | ref.type = dma_debug_single; | |
1294 | ||
1295 | check_unmap(&ref); | |
1296 | } | |
1297 | EXPORT_SYMBOL(debug_dma_unmap_page); | |
1298 | ||
972aa45c JR |
1299 | void debug_dma_map_sg(struct device *dev, struct scatterlist *sg, |
1300 | int nents, int mapped_ents, int direction) | |
1301 | { | |
1302 | struct dma_debug_entry *entry; | |
1303 | struct scatterlist *s; | |
1304 | int i; | |
1305 | ||
1306 | if (unlikely(global_disable)) | |
1307 | return; | |
1308 | ||
1309 | for_each_sg(sg, s, mapped_ents, i) { | |
1310 | entry = dma_entry_alloc(); | |
1311 | if (!entry) | |
1312 | return; | |
1313 | ||
1314 | entry->type = dma_debug_sg; | |
1315 | entry->dev = dev; | |
0abdd7a8 DW |
1316 | entry->pfn = page_to_pfn(sg_page(s)); |
1317 | entry->offset = s->offset, | |
884d0597 | 1318 | entry->size = sg_dma_len(s); |
15aedea4 | 1319 | entry->dev_addr = sg_dma_address(s); |
972aa45c JR |
1320 | entry->direction = direction; |
1321 | entry->sg_call_ents = nents; | |
1322 | entry->sg_mapped_ents = mapped_ents; | |
1323 | ||
9537a48e JR |
1324 | if (!PageHighMem(sg_page(s))) { |
1325 | check_for_stack(dev, sg_virt(s)); | |
884d0597 | 1326 | check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s)); |
9537a48e | 1327 | } |
972aa45c JR |
1328 | |
1329 | add_dma_entry(entry); | |
1330 | } | |
1331 | } | |
1332 | EXPORT_SYMBOL(debug_dma_map_sg); | |
1333 | ||
aa010efb JR |
1334 | static int get_nr_mapped_entries(struct device *dev, |
1335 | struct dma_debug_entry *ref) | |
88f3907f | 1336 | { |
aa010efb | 1337 | struct dma_debug_entry *entry; |
88f3907f FT |
1338 | struct hash_bucket *bucket; |
1339 | unsigned long flags; | |
c17e2cf7 | 1340 | int mapped_ents; |
88f3907f | 1341 | |
aa010efb | 1342 | bucket = get_hash_bucket(ref, &flags); |
c6a21d0b | 1343 | entry = bucket_find_exact(bucket, ref); |
c17e2cf7 | 1344 | mapped_ents = 0; |
88f3907f | 1345 | |
88f3907f FT |
1346 | if (entry) |
1347 | mapped_ents = entry->sg_mapped_ents; | |
1348 | put_hash_bucket(bucket, &flags); | |
1349 | ||
1350 | return mapped_ents; | |
1351 | } | |
1352 | ||
972aa45c JR |
1353 | void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist, |
1354 | int nelems, int dir) | |
1355 | { | |
972aa45c JR |
1356 | struct scatterlist *s; |
1357 | int mapped_ents = 0, i; | |
972aa45c JR |
1358 | |
1359 | if (unlikely(global_disable)) | |
1360 | return; | |
1361 | ||
1362 | for_each_sg(sglist, s, nelems, i) { | |
1363 | ||
1364 | struct dma_debug_entry ref = { | |
1365 | .type = dma_debug_sg, | |
1366 | .dev = dev, | |
0abdd7a8 DW |
1367 | .pfn = page_to_pfn(sg_page(s)), |
1368 | .offset = s->offset, | |
15aedea4 | 1369 | .dev_addr = sg_dma_address(s), |
884d0597 | 1370 | .size = sg_dma_len(s), |
972aa45c | 1371 | .direction = dir, |
e5e8c5b9 | 1372 | .sg_call_ents = nelems, |
972aa45c JR |
1373 | }; |
1374 | ||
1375 | if (mapped_ents && i >= mapped_ents) | |
1376 | break; | |
1377 | ||
e5e8c5b9 | 1378 | if (!i) |
aa010efb | 1379 | mapped_ents = get_nr_mapped_entries(dev, &ref); |
972aa45c JR |
1380 | |
1381 | check_unmap(&ref); | |
1382 | } | |
1383 | } | |
1384 | EXPORT_SYMBOL(debug_dma_unmap_sg); | |
1385 | ||
6bfd4498 JR |
1386 | void debug_dma_alloc_coherent(struct device *dev, size_t size, |
1387 | dma_addr_t dma_addr, void *virt) | |
1388 | { | |
1389 | struct dma_debug_entry *entry; | |
1390 | ||
1391 | if (unlikely(global_disable)) | |
1392 | return; | |
1393 | ||
1394 | if (unlikely(virt == NULL)) | |
1395 | return; | |
1396 | ||
1397 | entry = dma_entry_alloc(); | |
1398 | if (!entry) | |
1399 | return; | |
1400 | ||
1401 | entry->type = dma_debug_coherent; | |
1402 | entry->dev = dev; | |
0abdd7a8 DW |
1403 | entry->pfn = page_to_pfn(virt_to_page(virt)); |
1404 | entry->offset = (size_t) virt & PAGE_MASK; | |
6bfd4498 JR |
1405 | entry->size = size; |
1406 | entry->dev_addr = dma_addr; | |
1407 | entry->direction = DMA_BIDIRECTIONAL; | |
1408 | ||
1409 | add_dma_entry(entry); | |
1410 | } | |
1411 | EXPORT_SYMBOL(debug_dma_alloc_coherent); | |
1412 | ||
1413 | void debug_dma_free_coherent(struct device *dev, size_t size, | |
1414 | void *virt, dma_addr_t addr) | |
1415 | { | |
1416 | struct dma_debug_entry ref = { | |
1417 | .type = dma_debug_coherent, | |
1418 | .dev = dev, | |
0abdd7a8 DW |
1419 | .pfn = page_to_pfn(virt_to_page(virt)), |
1420 | .offset = (size_t) virt & PAGE_MASK, | |
6bfd4498 JR |
1421 | .dev_addr = addr, |
1422 | .size = size, | |
1423 | .direction = DMA_BIDIRECTIONAL, | |
1424 | }; | |
1425 | ||
1426 | if (unlikely(global_disable)) | |
1427 | return; | |
1428 | ||
1429 | check_unmap(&ref); | |
1430 | } | |
1431 | EXPORT_SYMBOL(debug_dma_free_coherent); | |
1432 | ||
b9d2317e JR |
1433 | void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle, |
1434 | size_t size, int direction) | |
1435 | { | |
aa010efb JR |
1436 | struct dma_debug_entry ref; |
1437 | ||
b9d2317e JR |
1438 | if (unlikely(global_disable)) |
1439 | return; | |
1440 | ||
aa010efb JR |
1441 | ref.type = dma_debug_single; |
1442 | ref.dev = dev; | |
1443 | ref.dev_addr = dma_handle; | |
1444 | ref.size = size; | |
1445 | ref.direction = direction; | |
1446 | ref.sg_call_ents = 0; | |
1447 | ||
1448 | check_sync(dev, &ref, true); | |
b9d2317e JR |
1449 | } |
1450 | EXPORT_SYMBOL(debug_dma_sync_single_for_cpu); | |
1451 | ||
1452 | void debug_dma_sync_single_for_device(struct device *dev, | |
1453 | dma_addr_t dma_handle, size_t size, | |
1454 | int direction) | |
1455 | { | |
aa010efb JR |
1456 | struct dma_debug_entry ref; |
1457 | ||
b9d2317e JR |
1458 | if (unlikely(global_disable)) |
1459 | return; | |
1460 | ||
aa010efb JR |
1461 | ref.type = dma_debug_single; |
1462 | ref.dev = dev; | |
1463 | ref.dev_addr = dma_handle; | |
1464 | ref.size = size; | |
1465 | ref.direction = direction; | |
1466 | ref.sg_call_ents = 0; | |
1467 | ||
1468 | check_sync(dev, &ref, false); | |
b9d2317e JR |
1469 | } |
1470 | EXPORT_SYMBOL(debug_dma_sync_single_for_device); | |
1471 | ||
948408ba JR |
1472 | void debug_dma_sync_single_range_for_cpu(struct device *dev, |
1473 | dma_addr_t dma_handle, | |
1474 | unsigned long offset, size_t size, | |
1475 | int direction) | |
1476 | { | |
aa010efb JR |
1477 | struct dma_debug_entry ref; |
1478 | ||
948408ba JR |
1479 | if (unlikely(global_disable)) |
1480 | return; | |
1481 | ||
aa010efb JR |
1482 | ref.type = dma_debug_single; |
1483 | ref.dev = dev; | |
1484 | ref.dev_addr = dma_handle; | |
1485 | ref.size = offset + size; | |
1486 | ref.direction = direction; | |
1487 | ref.sg_call_ents = 0; | |
1488 | ||
1489 | check_sync(dev, &ref, true); | |
948408ba JR |
1490 | } |
1491 | EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu); | |
1492 | ||
1493 | void debug_dma_sync_single_range_for_device(struct device *dev, | |
1494 | dma_addr_t dma_handle, | |
1495 | unsigned long offset, | |
1496 | size_t size, int direction) | |
1497 | { | |
aa010efb JR |
1498 | struct dma_debug_entry ref; |
1499 | ||
948408ba JR |
1500 | if (unlikely(global_disable)) |
1501 | return; | |
1502 | ||
aa010efb JR |
1503 | ref.type = dma_debug_single; |
1504 | ref.dev = dev; | |
1505 | ref.dev_addr = dma_handle; | |
1506 | ref.size = offset + size; | |
1507 | ref.direction = direction; | |
1508 | ref.sg_call_ents = 0; | |
1509 | ||
1510 | check_sync(dev, &ref, false); | |
948408ba JR |
1511 | } |
1512 | EXPORT_SYMBOL(debug_dma_sync_single_range_for_device); | |
1513 | ||
a31fba5d JR |
1514 | void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg, |
1515 | int nelems, int direction) | |
1516 | { | |
1517 | struct scatterlist *s; | |
88f3907f | 1518 | int mapped_ents = 0, i; |
a31fba5d JR |
1519 | |
1520 | if (unlikely(global_disable)) | |
1521 | return; | |
1522 | ||
1523 | for_each_sg(sg, s, nelems, i) { | |
aa010efb JR |
1524 | |
1525 | struct dma_debug_entry ref = { | |
1526 | .type = dma_debug_sg, | |
1527 | .dev = dev, | |
0abdd7a8 DW |
1528 | .pfn = page_to_pfn(sg_page(s)), |
1529 | .offset = s->offset, | |
aa010efb JR |
1530 | .dev_addr = sg_dma_address(s), |
1531 | .size = sg_dma_len(s), | |
1532 | .direction = direction, | |
1533 | .sg_call_ents = nelems, | |
1534 | }; | |
1535 | ||
88f3907f | 1536 | if (!i) |
aa010efb | 1537 | mapped_ents = get_nr_mapped_entries(dev, &ref); |
88f3907f FT |
1538 | |
1539 | if (i >= mapped_ents) | |
1540 | break; | |
1541 | ||
aa010efb | 1542 | check_sync(dev, &ref, true); |
a31fba5d JR |
1543 | } |
1544 | } | |
1545 | EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu); | |
1546 | ||
1547 | void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg, | |
1548 | int nelems, int direction) | |
1549 | { | |
1550 | struct scatterlist *s; | |
88f3907f | 1551 | int mapped_ents = 0, i; |
a31fba5d JR |
1552 | |
1553 | if (unlikely(global_disable)) | |
1554 | return; | |
1555 | ||
1556 | for_each_sg(sg, s, nelems, i) { | |
aa010efb JR |
1557 | |
1558 | struct dma_debug_entry ref = { | |
1559 | .type = dma_debug_sg, | |
1560 | .dev = dev, | |
0abdd7a8 DW |
1561 | .pfn = page_to_pfn(sg_page(s)), |
1562 | .offset = s->offset, | |
aa010efb JR |
1563 | .dev_addr = sg_dma_address(s), |
1564 | .size = sg_dma_len(s), | |
1565 | .direction = direction, | |
1566 | .sg_call_ents = nelems, | |
1567 | }; | |
88f3907f | 1568 | if (!i) |
aa010efb | 1569 | mapped_ents = get_nr_mapped_entries(dev, &ref); |
88f3907f FT |
1570 | |
1571 | if (i >= mapped_ents) | |
1572 | break; | |
1573 | ||
aa010efb | 1574 | check_sync(dev, &ref, false); |
a31fba5d JR |
1575 | } |
1576 | } | |
1577 | EXPORT_SYMBOL(debug_dma_sync_sg_for_device); | |
1578 | ||
1745de5e JR |
1579 | static int __init dma_debug_driver_setup(char *str) |
1580 | { | |
1581 | int i; | |
1582 | ||
1583 | for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) { | |
1584 | current_driver_name[i] = *str; | |
1585 | if (*str == 0) | |
1586 | break; | |
1587 | } | |
1588 | ||
1589 | if (current_driver_name[0]) | |
e7ed70ee JR |
1590 | pr_info("DMA-API: enable driver filter for driver [%s]\n", |
1591 | current_driver_name); | |
1745de5e JR |
1592 | |
1593 | ||
1594 | return 1; | |
1595 | } | |
1596 | __setup("dma_debug_driver=", dma_debug_driver_setup); |