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