xref: /linux-6.15/kernel/dma/debug.c (revision 7543c3e3)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2008 Advanced Micro Devices, Inc.
4  *
5  * Author: Joerg Roedel <[email protected]>
6  */
7 
8 #define pr_fmt(fmt)	"DMA-API: " fmt
9 
10 #include <linux/sched/task_stack.h>
11 #include <linux/scatterlist.h>
12 #include <linux/dma-map-ops.h>
13 #include <linux/sched/task.h>
14 #include <linux/stacktrace.h>
15 #include <linux/spinlock.h>
16 #include <linux/vmalloc.h>
17 #include <linux/debugfs.h>
18 #include <linux/uaccess.h>
19 #include <linux/export.h>
20 #include <linux/device.h>
21 #include <linux/types.h>
22 #include <linux/sched.h>
23 #include <linux/ctype.h>
24 #include <linux/list.h>
25 #include <linux/slab.h>
26 #include <asm/sections.h>
27 #include "debug.h"
28 
29 #define HASH_SIZE       16384ULL
30 #define HASH_FN_SHIFT   13
31 #define HASH_FN_MASK    (HASH_SIZE - 1)
32 
33 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
34 /* If the pool runs out, add this many new entries at once */
35 #define DMA_DEBUG_DYNAMIC_ENTRIES (PAGE_SIZE / sizeof(struct dma_debug_entry))
36 
37 enum {
38 	dma_debug_single,
39 	dma_debug_sg,
40 	dma_debug_coherent,
41 	dma_debug_resource,
42 };
43 
44 enum map_err_types {
45 	MAP_ERR_CHECK_NOT_APPLICABLE,
46 	MAP_ERR_NOT_CHECKED,
47 	MAP_ERR_CHECKED,
48 };
49 
50 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
51 
52 /**
53  * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
54  * @list: node on pre-allocated free_entries list
55  * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
56  * @dev_addr: dma address
57  * @size: length of the mapping
58  * @type: single, page, sg, coherent
59  * @direction: enum dma_data_direction
60  * @sg_call_ents: 'nents' from dma_map_sg
61  * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
62  * @pfn: page frame of the start address
63  * @offset: offset of mapping relative to pfn
64  * @map_err_type: track whether dma_mapping_error() was checked
65  * @stack_len: number of backtrace entries in @stack_entries
66  * @stack_entries: stack of backtrace history
67  */
68 struct dma_debug_entry {
69 	struct list_head list;
70 	struct device    *dev;
71 	u64              dev_addr;
72 	u64              size;
73 	int              type;
74 	int              direction;
75 	int		 sg_call_ents;
76 	int		 sg_mapped_ents;
77 	unsigned long	 pfn;
78 	size_t		 offset;
79 	enum map_err_types  map_err_type;
80 #ifdef CONFIG_STACKTRACE
81 	unsigned int	stack_len;
82 	unsigned long	stack_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
83 #endif
84 } ____cacheline_aligned_in_smp;
85 
86 typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
87 
88 struct hash_bucket {
89 	struct list_head list;
90 	spinlock_t lock;
91 };
92 
93 /* Hash list to save the allocated dma addresses */
94 static struct hash_bucket dma_entry_hash[HASH_SIZE];
95 /* List of pre-allocated dma_debug_entry's */
96 static LIST_HEAD(free_entries);
97 /* Lock for the list above */
98 static DEFINE_SPINLOCK(free_entries_lock);
99 
100 /* Global disable flag - will be set in case of an error */
101 static bool global_disable __read_mostly;
102 
103 /* Early initialization disable flag, set at the end of dma_debug_init */
104 static bool dma_debug_initialized __read_mostly;
105 
106 static inline bool dma_debug_disabled(void)
107 {
108 	return global_disable || !dma_debug_initialized;
109 }
110 
111 /* Global error count */
112 static u32 error_count;
113 
114 /* Global error show enable*/
115 static u32 show_all_errors __read_mostly;
116 /* Number of errors to show */
117 static u32 show_num_errors = 1;
118 
119 static u32 num_free_entries;
120 static u32 min_free_entries;
121 static u32 nr_total_entries;
122 
123 /* number of preallocated entries requested by kernel cmdline */
124 static u32 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
125 
126 /* per-driver filter related state */
127 
128 #define NAME_MAX_LEN	64
129 
130 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
131 static struct device_driver *current_driver                    __read_mostly;
132 
133 static DEFINE_RWLOCK(driver_name_lock);
134 
135 static const char *const maperr2str[] = {
136 	[MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
137 	[MAP_ERR_NOT_CHECKED] = "dma map error not checked",
138 	[MAP_ERR_CHECKED] = "dma map error checked",
139 };
140 
141 static const char *type2name[] = {
142 	[dma_debug_single] = "single",
143 	[dma_debug_sg] = "scatter-gather",
144 	[dma_debug_coherent] = "coherent",
145 	[dma_debug_resource] = "resource",
146 };
147 
148 static const char *dir2name[] = {
149 	[DMA_BIDIRECTIONAL]	= "DMA_BIDIRECTIONAL",
150 	[DMA_TO_DEVICE]		= "DMA_TO_DEVICE",
151 	[DMA_FROM_DEVICE]	= "DMA_FROM_DEVICE",
152 	[DMA_NONE]		= "DMA_NONE",
153 };
154 
155 /*
156  * The access to some variables in this macro is racy. We can't use atomic_t
157  * here because all these variables are exported to debugfs. Some of them even
158  * writeable. This is also the reason why a lock won't help much. But anyway,
159  * the races are no big deal. Here is why:
160  *
161  *   error_count: the addition is racy, but the worst thing that can happen is
162  *                that we don't count some errors
163  *   show_num_errors: the subtraction is racy. Also no big deal because in
164  *                    worst case this will result in one warning more in the
165  *                    system log than the user configured. This variable is
166  *                    writeable via debugfs.
167  */
168 static inline void dump_entry_trace(struct dma_debug_entry *entry)
169 {
170 #ifdef CONFIG_STACKTRACE
171 	if (entry) {
172 		pr_warn("Mapped at:\n");
173 		stack_trace_print(entry->stack_entries, entry->stack_len, 0);
174 	}
175 #endif
176 }
177 
178 static bool driver_filter(struct device *dev)
179 {
180 	struct device_driver *drv;
181 	unsigned long flags;
182 	bool ret;
183 
184 	/* driver filter off */
185 	if (likely(!current_driver_name[0]))
186 		return true;
187 
188 	/* driver filter on and initialized */
189 	if (current_driver && dev && dev->driver == current_driver)
190 		return true;
191 
192 	/* driver filter on, but we can't filter on a NULL device... */
193 	if (!dev)
194 		return false;
195 
196 	if (current_driver || !current_driver_name[0])
197 		return false;
198 
199 	/* driver filter on but not yet initialized */
200 	drv = dev->driver;
201 	if (!drv)
202 		return false;
203 
204 	/* lock to protect against change of current_driver_name */
205 	read_lock_irqsave(&driver_name_lock, flags);
206 
207 	ret = false;
208 	if (drv->name &&
209 	    strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
210 		current_driver = drv;
211 		ret = true;
212 	}
213 
214 	read_unlock_irqrestore(&driver_name_lock, flags);
215 
216 	return ret;
217 }
218 
219 #define err_printk(dev, entry, format, arg...) do {			\
220 		error_count += 1;					\
221 		if (driver_filter(dev) &&				\
222 		    (show_all_errors || show_num_errors > 0)) {		\
223 			WARN(1, pr_fmt("%s %s: ") format,		\
224 			     dev ? dev_driver_string(dev) : "NULL",	\
225 			     dev ? dev_name(dev) : "NULL", ## arg);	\
226 			dump_entry_trace(entry);			\
227 		}							\
228 		if (!show_all_errors && show_num_errors > 0)		\
229 			show_num_errors -= 1;				\
230 	} while (0);
231 
232 /*
233  * Hash related functions
234  *
235  * Every DMA-API request is saved into a struct dma_debug_entry. To
236  * have quick access to these structs they are stored into a hash.
237  */
238 static int hash_fn(struct dma_debug_entry *entry)
239 {
240 	/*
241 	 * Hash function is based on the dma address.
242 	 * We use bits 20-27 here as the index into the hash
243 	 */
244 	return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
245 }
246 
247 /*
248  * Request exclusive access to a hash bucket for a given dma_debug_entry.
249  */
250 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
251 					   unsigned long *flags)
252 	__acquires(&dma_entry_hash[idx].lock)
253 {
254 	int idx = hash_fn(entry);
255 	unsigned long __flags;
256 
257 	spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
258 	*flags = __flags;
259 	return &dma_entry_hash[idx];
260 }
261 
262 /*
263  * Give up exclusive access to the hash bucket
264  */
265 static void put_hash_bucket(struct hash_bucket *bucket,
266 			    unsigned long flags)
267 	__releases(&bucket->lock)
268 {
269 	spin_unlock_irqrestore(&bucket->lock, flags);
270 }
271 
272 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
273 {
274 	return ((a->dev_addr == b->dev_addr) &&
275 		(a->dev == b->dev)) ? true : false;
276 }
277 
278 static bool containing_match(struct dma_debug_entry *a,
279 			     struct dma_debug_entry *b)
280 {
281 	if (a->dev != b->dev)
282 		return false;
283 
284 	if ((b->dev_addr <= a->dev_addr) &&
285 	    ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
286 		return true;
287 
288 	return false;
289 }
290 
291 /*
292  * Search a given entry in the hash bucket list
293  */
294 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
295 						  struct dma_debug_entry *ref,
296 						  match_fn match)
297 {
298 	struct dma_debug_entry *entry, *ret = NULL;
299 	int matches = 0, match_lvl, last_lvl = -1;
300 
301 	list_for_each_entry(entry, &bucket->list, list) {
302 		if (!match(ref, entry))
303 			continue;
304 
305 		/*
306 		 * Some drivers map the same physical address multiple
307 		 * times. Without a hardware IOMMU this results in the
308 		 * same device addresses being put into the dma-debug
309 		 * hash multiple times too. This can result in false
310 		 * positives being reported. Therefore we implement a
311 		 * best-fit algorithm here which returns the entry from
312 		 * the hash which fits best to the reference value
313 		 * instead of the first-fit.
314 		 */
315 		matches += 1;
316 		match_lvl = 0;
317 		entry->size         == ref->size         ? ++match_lvl : 0;
318 		entry->type         == ref->type         ? ++match_lvl : 0;
319 		entry->direction    == ref->direction    ? ++match_lvl : 0;
320 		entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
321 
322 		if (match_lvl == 4) {
323 			/* perfect-fit - return the result */
324 			return entry;
325 		} else if (match_lvl > last_lvl) {
326 			/*
327 			 * We found an entry that fits better then the
328 			 * previous one or it is the 1st match.
329 			 */
330 			last_lvl = match_lvl;
331 			ret      = entry;
332 		}
333 	}
334 
335 	/*
336 	 * If we have multiple matches but no perfect-fit, just return
337 	 * NULL.
338 	 */
339 	ret = (matches == 1) ? ret : NULL;
340 
341 	return ret;
342 }
343 
344 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
345 						 struct dma_debug_entry *ref)
346 {
347 	return __hash_bucket_find(bucket, ref, exact_match);
348 }
349 
350 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
351 						   struct dma_debug_entry *ref,
352 						   unsigned long *flags)
353 {
354 
355 	struct dma_debug_entry *entry, index = *ref;
356 	int limit = min(HASH_SIZE, (index.dev_addr >> HASH_FN_SHIFT) + 1);
357 
358 	for (int i = 0; i < limit; i++) {
359 		entry = __hash_bucket_find(*bucket, ref, containing_match);
360 
361 		if (entry)
362 			return entry;
363 
364 		/*
365 		 * Nothing found, go back a hash bucket
366 		 */
367 		put_hash_bucket(*bucket, *flags);
368 		index.dev_addr -= (1 << HASH_FN_SHIFT);
369 		*bucket = get_hash_bucket(&index, flags);
370 	}
371 
372 	return NULL;
373 }
374 
375 /*
376  * Add an entry to a hash bucket
377  */
378 static void hash_bucket_add(struct hash_bucket *bucket,
379 			    struct dma_debug_entry *entry)
380 {
381 	list_add_tail(&entry->list, &bucket->list);
382 }
383 
384 /*
385  * Remove entry from a hash bucket list
386  */
387 static void hash_bucket_del(struct dma_debug_entry *entry)
388 {
389 	list_del(&entry->list);
390 }
391 
392 static unsigned long long phys_addr(struct dma_debug_entry *entry)
393 {
394 	if (entry->type == dma_debug_resource)
395 		return __pfn_to_phys(entry->pfn) + entry->offset;
396 
397 	return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
398 }
399 
400 /*
401  * For each mapping (initial cacheline in the case of
402  * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
403  * scatterlist, or the cacheline specified in dma_map_single) insert
404  * into this tree using the cacheline as the key. At
405  * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
406  * the entry already exists at insertion time add a tag as a reference
407  * count for the overlapping mappings.  For now, the overlap tracking
408  * just ensures that 'unmaps' balance 'maps' before marking the
409  * cacheline idle, but we should also be flagging overlaps as an API
410  * violation.
411  *
412  * Memory usage is mostly constrained by the maximum number of available
413  * dma-debug entries in that we need a free dma_debug_entry before
414  * inserting into the tree.  In the case of dma_map_page and
415  * dma_alloc_coherent there is only one dma_debug_entry and one
416  * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
417  * other hand, consumes a single dma_debug_entry, but inserts 'nents'
418  * entries into the tree.
419  *
420  * Use __GFP_NOWARN because the printk from an OOM, to netconsole, could end
421  * up right back in the DMA debugging code, leading to a deadlock.
422  */
423 static RADIX_TREE(dma_active_cacheline, GFP_ATOMIC | __GFP_NOWARN);
424 static DEFINE_SPINLOCK(radix_lock);
425 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
426 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
427 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
428 
429 static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
430 {
431 	return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
432 		(entry->offset >> L1_CACHE_SHIFT);
433 }
434 
435 static int active_cacheline_read_overlap(phys_addr_t cln)
436 {
437 	int overlap = 0, i;
438 
439 	for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
440 		if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
441 			overlap |= 1 << i;
442 	return overlap;
443 }
444 
445 static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
446 {
447 	int i;
448 
449 	if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
450 		return overlap;
451 
452 	for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
453 		if (overlap & 1 << i)
454 			radix_tree_tag_set(&dma_active_cacheline, cln, i);
455 		else
456 			radix_tree_tag_clear(&dma_active_cacheline, cln, i);
457 
458 	return overlap;
459 }
460 
461 static void active_cacheline_inc_overlap(phys_addr_t cln)
462 {
463 	int overlap = active_cacheline_read_overlap(cln);
464 
465 	overlap = active_cacheline_set_overlap(cln, ++overlap);
466 
467 	/* If we overflowed the overlap counter then we're potentially
468 	 * leaking dma-mappings.
469 	 */
470 	WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
471 		  pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
472 		  ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
473 }
474 
475 static int active_cacheline_dec_overlap(phys_addr_t cln)
476 {
477 	int overlap = active_cacheline_read_overlap(cln);
478 
479 	return active_cacheline_set_overlap(cln, --overlap);
480 }
481 
482 static int active_cacheline_insert(struct dma_debug_entry *entry)
483 {
484 	phys_addr_t cln = to_cacheline_number(entry);
485 	unsigned long flags;
486 	int rc;
487 
488 	/* If the device is not writing memory then we don't have any
489 	 * concerns about the cpu consuming stale data.  This mitigates
490 	 * legitimate usages of overlapping mappings.
491 	 */
492 	if (entry->direction == DMA_TO_DEVICE)
493 		return 0;
494 
495 	spin_lock_irqsave(&radix_lock, flags);
496 	rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
497 	if (rc == -EEXIST)
498 		active_cacheline_inc_overlap(cln);
499 	spin_unlock_irqrestore(&radix_lock, flags);
500 
501 	return rc;
502 }
503 
504 static void active_cacheline_remove(struct dma_debug_entry *entry)
505 {
506 	phys_addr_t cln = to_cacheline_number(entry);
507 	unsigned long flags;
508 
509 	/* ...mirror the insert case */
510 	if (entry->direction == DMA_TO_DEVICE)
511 		return;
512 
513 	spin_lock_irqsave(&radix_lock, flags);
514 	/* since we are counting overlaps the final put of the
515 	 * cacheline will occur when the overlap count is 0.
516 	 * active_cacheline_dec_overlap() returns -1 in that case
517 	 */
518 	if (active_cacheline_dec_overlap(cln) < 0)
519 		radix_tree_delete(&dma_active_cacheline, cln);
520 	spin_unlock_irqrestore(&radix_lock, flags);
521 }
522 
523 /*
524  * Dump mappings entries on kernel space for debugging purposes
525  */
526 void debug_dma_dump_mappings(struct device *dev)
527 {
528 	int idx;
529 	phys_addr_t cln;
530 
531 	for (idx = 0; idx < HASH_SIZE; idx++) {
532 		struct hash_bucket *bucket = &dma_entry_hash[idx];
533 		struct dma_debug_entry *entry;
534 		unsigned long flags;
535 
536 		spin_lock_irqsave(&bucket->lock, flags);
537 		list_for_each_entry(entry, &bucket->list, list) {
538 			if (!dev || dev == entry->dev) {
539 				cln = to_cacheline_number(entry);
540 				dev_info(entry->dev,
541 					 "%s idx %d P=%llx N=%lx D=%llx L=%llx cln=%pa %s %s\n",
542 					 type2name[entry->type], idx,
543 					 phys_addr(entry), entry->pfn,
544 					 entry->dev_addr, entry->size,
545 					 &cln, dir2name[entry->direction],
546 					 maperr2str[entry->map_err_type]);
547 			}
548 		}
549 		spin_unlock_irqrestore(&bucket->lock, flags);
550 
551 		cond_resched();
552 	}
553 }
554 
555 /*
556  * Dump mappings entries on user space via debugfs
557  */
558 static int dump_show(struct seq_file *seq, void *v)
559 {
560 	int idx;
561 	phys_addr_t cln;
562 
563 	for (idx = 0; idx < HASH_SIZE; idx++) {
564 		struct hash_bucket *bucket = &dma_entry_hash[idx];
565 		struct dma_debug_entry *entry;
566 		unsigned long flags;
567 
568 		spin_lock_irqsave(&bucket->lock, flags);
569 		list_for_each_entry(entry, &bucket->list, list) {
570 			cln = to_cacheline_number(entry);
571 			seq_printf(seq,
572 				   "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx cln=%pa %s %s\n",
573 				   dev_driver_string(entry->dev),
574 				   dev_name(entry->dev),
575 				   type2name[entry->type], idx,
576 				   phys_addr(entry), entry->pfn,
577 				   entry->dev_addr, entry->size,
578 				   &cln, dir2name[entry->direction],
579 				   maperr2str[entry->map_err_type]);
580 		}
581 		spin_unlock_irqrestore(&bucket->lock, flags);
582 	}
583 	return 0;
584 }
585 DEFINE_SHOW_ATTRIBUTE(dump);
586 
587 /*
588  * Wrapper function for adding an entry to the hash.
589  * This function takes care of locking itself.
590  */
591 static void add_dma_entry(struct dma_debug_entry *entry, unsigned long attrs)
592 {
593 	struct hash_bucket *bucket;
594 	unsigned long flags;
595 	int rc;
596 
597 	bucket = get_hash_bucket(entry, &flags);
598 	hash_bucket_add(bucket, entry);
599 	put_hash_bucket(bucket, flags);
600 
601 	rc = active_cacheline_insert(entry);
602 	if (rc == -ENOMEM) {
603 		pr_err_once("cacheline tracking ENOMEM, dma-debug disabled\n");
604 		global_disable = true;
605 	} else if (rc == -EEXIST && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
606 		err_printk(entry->dev, entry,
607 			"cacheline tracking EEXIST, overlapping mappings aren't supported\n");
608 	}
609 }
610 
611 static int dma_debug_create_entries(gfp_t gfp)
612 {
613 	struct dma_debug_entry *entry;
614 	int i;
615 
616 	entry = (void *)get_zeroed_page(gfp);
617 	if (!entry)
618 		return -ENOMEM;
619 
620 	for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
621 		list_add_tail(&entry[i].list, &free_entries);
622 
623 	num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
624 	nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
625 
626 	return 0;
627 }
628 
629 static struct dma_debug_entry *__dma_entry_alloc(void)
630 {
631 	struct dma_debug_entry *entry;
632 
633 	entry = list_entry(free_entries.next, struct dma_debug_entry, list);
634 	list_del(&entry->list);
635 	memset(entry, 0, sizeof(*entry));
636 
637 	num_free_entries -= 1;
638 	if (num_free_entries < min_free_entries)
639 		min_free_entries = num_free_entries;
640 
641 	return entry;
642 }
643 
644 /*
645  * This should be called outside of free_entries_lock scope to avoid potential
646  * deadlocks with serial consoles that use DMA.
647  */
648 static void __dma_entry_alloc_check_leak(u32 nr_entries)
649 {
650 	u32 tmp = nr_entries % nr_prealloc_entries;
651 
652 	/* Shout each time we tick over some multiple of the initial pool */
653 	if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
654 		pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
655 			nr_entries,
656 			(nr_entries / nr_prealloc_entries));
657 	}
658 }
659 
660 /* struct dma_entry allocator
661  *
662  * The next two functions implement the allocator for
663  * struct dma_debug_entries.
664  */
665 static struct dma_debug_entry *dma_entry_alloc(void)
666 {
667 	bool alloc_check_leak = false;
668 	struct dma_debug_entry *entry;
669 	unsigned long flags;
670 	u32 nr_entries;
671 
672 	spin_lock_irqsave(&free_entries_lock, flags);
673 	if (num_free_entries == 0) {
674 		if (dma_debug_create_entries(GFP_ATOMIC)) {
675 			global_disable = true;
676 			spin_unlock_irqrestore(&free_entries_lock, flags);
677 			pr_err("debugging out of memory - disabling\n");
678 			return NULL;
679 		}
680 		alloc_check_leak = true;
681 		nr_entries = nr_total_entries;
682 	}
683 
684 	entry = __dma_entry_alloc();
685 
686 	spin_unlock_irqrestore(&free_entries_lock, flags);
687 
688 	if (alloc_check_leak)
689 		__dma_entry_alloc_check_leak(nr_entries);
690 
691 #ifdef CONFIG_STACKTRACE
692 	entry->stack_len = stack_trace_save(entry->stack_entries,
693 					    ARRAY_SIZE(entry->stack_entries),
694 					    1);
695 #endif
696 	return entry;
697 }
698 
699 static void dma_entry_free(struct dma_debug_entry *entry)
700 {
701 	unsigned long flags;
702 
703 	active_cacheline_remove(entry);
704 
705 	/*
706 	 * add to beginning of the list - this way the entries are
707 	 * more likely cache hot when they are reallocated.
708 	 */
709 	spin_lock_irqsave(&free_entries_lock, flags);
710 	list_add(&entry->list, &free_entries);
711 	num_free_entries += 1;
712 	spin_unlock_irqrestore(&free_entries_lock, flags);
713 }
714 
715 /*
716  * DMA-API debugging init code
717  *
718  * The init code does two things:
719  *   1. Initialize core data structures
720  *   2. Preallocate a given number of dma_debug_entry structs
721  */
722 
723 static ssize_t filter_read(struct file *file, char __user *user_buf,
724 			   size_t count, loff_t *ppos)
725 {
726 	char buf[NAME_MAX_LEN + 1];
727 	unsigned long flags;
728 	int len;
729 
730 	if (!current_driver_name[0])
731 		return 0;
732 
733 	/*
734 	 * We can't copy to userspace directly because current_driver_name can
735 	 * only be read under the driver_name_lock with irqs disabled. So
736 	 * create a temporary copy first.
737 	 */
738 	read_lock_irqsave(&driver_name_lock, flags);
739 	len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
740 	read_unlock_irqrestore(&driver_name_lock, flags);
741 
742 	return simple_read_from_buffer(user_buf, count, ppos, buf, len);
743 }
744 
745 static ssize_t filter_write(struct file *file, const char __user *userbuf,
746 			    size_t count, loff_t *ppos)
747 {
748 	char buf[NAME_MAX_LEN];
749 	unsigned long flags;
750 	size_t len;
751 	int i;
752 
753 	/*
754 	 * We can't copy from userspace directly. Access to
755 	 * current_driver_name is protected with a write_lock with irqs
756 	 * disabled. Since copy_from_user can fault and may sleep we
757 	 * need to copy to temporary buffer first
758 	 */
759 	len = min(count, (size_t)(NAME_MAX_LEN - 1));
760 	if (copy_from_user(buf, userbuf, len))
761 		return -EFAULT;
762 
763 	buf[len] = 0;
764 
765 	write_lock_irqsave(&driver_name_lock, flags);
766 
767 	/*
768 	 * Now handle the string we got from userspace very carefully.
769 	 * The rules are:
770 	 *         - only use the first token we got
771 	 *         - token delimiter is everything looking like a space
772 	 *           character (' ', '\n', '\t' ...)
773 	 *
774 	 */
775 	if (!isalnum(buf[0])) {
776 		/*
777 		 * If the first character userspace gave us is not
778 		 * alphanumerical then assume the filter should be
779 		 * switched off.
780 		 */
781 		if (current_driver_name[0])
782 			pr_info("switching off dma-debug driver filter\n");
783 		current_driver_name[0] = 0;
784 		current_driver = NULL;
785 		goto out_unlock;
786 	}
787 
788 	/*
789 	 * Now parse out the first token and use it as the name for the
790 	 * driver to filter for.
791 	 */
792 	for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
793 		current_driver_name[i] = buf[i];
794 		if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
795 			break;
796 	}
797 	current_driver_name[i] = 0;
798 	current_driver = NULL;
799 
800 	pr_info("enable driver filter for driver [%s]\n",
801 		current_driver_name);
802 
803 out_unlock:
804 	write_unlock_irqrestore(&driver_name_lock, flags);
805 
806 	return count;
807 }
808 
809 static const struct file_operations filter_fops = {
810 	.read  = filter_read,
811 	.write = filter_write,
812 	.llseek = default_llseek,
813 };
814 
815 static int __init dma_debug_fs_init(void)
816 {
817 	struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
818 
819 	debugfs_create_bool("disabled", 0444, dentry, &global_disable);
820 	debugfs_create_u32("error_count", 0444, dentry, &error_count);
821 	debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
822 	debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
823 	debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
824 	debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
825 	debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
826 	debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
827 	debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
828 
829 	return 0;
830 }
831 core_initcall_sync(dma_debug_fs_init);
832 
833 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
834 {
835 	struct dma_debug_entry *entry;
836 	unsigned long flags;
837 	int count = 0, i;
838 
839 	for (i = 0; i < HASH_SIZE; ++i) {
840 		spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
841 		list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
842 			if (entry->dev == dev) {
843 				count += 1;
844 				*out_entry = entry;
845 			}
846 		}
847 		spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
848 	}
849 
850 	return count;
851 }
852 
853 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
854 {
855 	struct device *dev = data;
856 	struct dma_debug_entry *entry;
857 	int count;
858 
859 	if (dma_debug_disabled())
860 		return 0;
861 
862 	switch (action) {
863 	case BUS_NOTIFY_UNBOUND_DRIVER:
864 		count = device_dma_allocations(dev, &entry);
865 		if (count == 0)
866 			break;
867 		err_printk(dev, entry, "device driver has pending "
868 				"DMA allocations while released from device "
869 				"[count=%d]\n"
870 				"One of leaked entries details: "
871 				"[device address=0x%016llx] [size=%llu bytes] "
872 				"[mapped with %s] [mapped as %s]\n",
873 			count, entry->dev_addr, entry->size,
874 			dir2name[entry->direction], type2name[entry->type]);
875 		break;
876 	default:
877 		break;
878 	}
879 
880 	return 0;
881 }
882 
883 void dma_debug_add_bus(const struct bus_type *bus)
884 {
885 	struct notifier_block *nb;
886 
887 	if (dma_debug_disabled())
888 		return;
889 
890 	nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
891 	if (nb == NULL) {
892 		pr_err("dma_debug_add_bus: out of memory\n");
893 		return;
894 	}
895 
896 	nb->notifier_call = dma_debug_device_change;
897 
898 	bus_register_notifier(bus, nb);
899 }
900 
901 static int dma_debug_init(void)
902 {
903 	int i, nr_pages;
904 
905 	/* Do not use dma_debug_initialized here, since we really want to be
906 	 * called to set dma_debug_initialized
907 	 */
908 	if (global_disable)
909 		return 0;
910 
911 	for (i = 0; i < HASH_SIZE; ++i) {
912 		INIT_LIST_HEAD(&dma_entry_hash[i].list);
913 		spin_lock_init(&dma_entry_hash[i].lock);
914 	}
915 
916 	nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
917 	for (i = 0; i < nr_pages; ++i)
918 		dma_debug_create_entries(GFP_KERNEL);
919 	if (num_free_entries >= nr_prealloc_entries) {
920 		pr_info("preallocated %d debug entries\n", nr_total_entries);
921 	} else if (num_free_entries > 0) {
922 		pr_warn("%d debug entries requested but only %d allocated\n",
923 			nr_prealloc_entries, nr_total_entries);
924 	} else {
925 		pr_err("debugging out of memory error - disabled\n");
926 		global_disable = true;
927 
928 		return 0;
929 	}
930 	min_free_entries = num_free_entries;
931 
932 	dma_debug_initialized = true;
933 
934 	pr_info("debugging enabled by kernel config\n");
935 	return 0;
936 }
937 core_initcall(dma_debug_init);
938 
939 static __init int dma_debug_cmdline(char *str)
940 {
941 	if (!str)
942 		return -EINVAL;
943 
944 	if (strncmp(str, "off", 3) == 0) {
945 		pr_info("debugging disabled on kernel command line\n");
946 		global_disable = true;
947 	}
948 
949 	return 1;
950 }
951 
952 static __init int dma_debug_entries_cmdline(char *str)
953 {
954 	if (!str)
955 		return -EINVAL;
956 	if (!get_option(&str, &nr_prealloc_entries))
957 		nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
958 	return 1;
959 }
960 
961 __setup("dma_debug=", dma_debug_cmdline);
962 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
963 
964 static void check_unmap(struct dma_debug_entry *ref)
965 {
966 	struct dma_debug_entry *entry;
967 	struct hash_bucket *bucket;
968 	unsigned long flags;
969 
970 	bucket = get_hash_bucket(ref, &flags);
971 	entry = bucket_find_exact(bucket, ref);
972 
973 	if (!entry) {
974 		/* must drop lock before calling dma_mapping_error */
975 		put_hash_bucket(bucket, flags);
976 
977 		if (dma_mapping_error(ref->dev, ref->dev_addr)) {
978 			err_printk(ref->dev, NULL,
979 				   "device driver tries to free an "
980 				   "invalid DMA memory address\n");
981 		} else {
982 			err_printk(ref->dev, NULL,
983 				   "device driver tries to free DMA "
984 				   "memory it has not allocated [device "
985 				   "address=0x%016llx] [size=%llu bytes]\n",
986 				   ref->dev_addr, ref->size);
987 		}
988 		return;
989 	}
990 
991 	if (ref->size != entry->size) {
992 		err_printk(ref->dev, entry, "device driver frees "
993 			   "DMA memory with different size "
994 			   "[device address=0x%016llx] [map size=%llu bytes] "
995 			   "[unmap size=%llu bytes]\n",
996 			   ref->dev_addr, entry->size, ref->size);
997 	}
998 
999 	if (ref->type != entry->type) {
1000 		err_printk(ref->dev, entry, "device driver frees "
1001 			   "DMA memory with wrong function "
1002 			   "[device address=0x%016llx] [size=%llu bytes] "
1003 			   "[mapped as %s] [unmapped as %s]\n",
1004 			   ref->dev_addr, ref->size,
1005 			   type2name[entry->type], type2name[ref->type]);
1006 	} else if ((entry->type == dma_debug_coherent) &&
1007 		   (phys_addr(ref) != phys_addr(entry))) {
1008 		err_printk(ref->dev, entry, "device driver frees "
1009 			   "DMA memory with different CPU address "
1010 			   "[device address=0x%016llx] [size=%llu bytes] "
1011 			   "[cpu alloc address=0x%016llx] "
1012 			   "[cpu free address=0x%016llx]",
1013 			   ref->dev_addr, ref->size,
1014 			   phys_addr(entry),
1015 			   phys_addr(ref));
1016 	}
1017 
1018 	if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1019 	    ref->sg_call_ents != entry->sg_call_ents) {
1020 		err_printk(ref->dev, entry, "device driver frees "
1021 			   "DMA sg list with different entry count "
1022 			   "[map count=%d] [unmap count=%d]\n",
1023 			   entry->sg_call_ents, ref->sg_call_ents);
1024 	}
1025 
1026 	/*
1027 	 * This may be no bug in reality - but most implementations of the
1028 	 * DMA API don't handle this properly, so check for it here
1029 	 */
1030 	if (ref->direction != entry->direction) {
1031 		err_printk(ref->dev, entry, "device driver frees "
1032 			   "DMA memory with different direction "
1033 			   "[device address=0x%016llx] [size=%llu bytes] "
1034 			   "[mapped with %s] [unmapped with %s]\n",
1035 			   ref->dev_addr, ref->size,
1036 			   dir2name[entry->direction],
1037 			   dir2name[ref->direction]);
1038 	}
1039 
1040 	/*
1041 	 * Drivers should use dma_mapping_error() to check the returned
1042 	 * addresses of dma_map_single() and dma_map_page().
1043 	 * If not, print this warning message. See Documentation/core-api/dma-api.rst.
1044 	 */
1045 	if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1046 		err_printk(ref->dev, entry,
1047 			   "device driver failed to check map error"
1048 			   "[device address=0x%016llx] [size=%llu bytes] "
1049 			   "[mapped as %s]",
1050 			   ref->dev_addr, ref->size,
1051 			   type2name[entry->type]);
1052 	}
1053 
1054 	hash_bucket_del(entry);
1055 	put_hash_bucket(bucket, flags);
1056 
1057 	/*
1058 	 * Free the entry outside of bucket_lock to avoid ABBA deadlocks
1059 	 * between that and radix_lock.
1060 	 */
1061 	dma_entry_free(entry);
1062 }
1063 
1064 static void check_for_stack(struct device *dev,
1065 			    struct page *page, size_t offset)
1066 {
1067 	void *addr;
1068 	struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1069 
1070 	if (!stack_vm_area) {
1071 		/* Stack is direct-mapped. */
1072 		if (PageHighMem(page))
1073 			return;
1074 		addr = page_address(page) + offset;
1075 		if (object_is_on_stack(addr))
1076 			err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1077 	} else {
1078 		/* Stack is vmalloced. */
1079 		int i;
1080 
1081 		for (i = 0; i < stack_vm_area->nr_pages; i++) {
1082 			if (page != stack_vm_area->pages[i])
1083 				continue;
1084 
1085 			addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1086 			err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1087 			break;
1088 		}
1089 	}
1090 }
1091 
1092 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1093 {
1094 	if (memory_intersects(_stext, _etext, addr, len) ||
1095 	    memory_intersects(__start_rodata, __end_rodata, addr, len))
1096 		err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1097 }
1098 
1099 static void check_sync(struct device *dev,
1100 		       struct dma_debug_entry *ref,
1101 		       bool to_cpu)
1102 {
1103 	struct dma_debug_entry *entry;
1104 	struct hash_bucket *bucket;
1105 	unsigned long flags;
1106 
1107 	bucket = get_hash_bucket(ref, &flags);
1108 
1109 	entry = bucket_find_contain(&bucket, ref, &flags);
1110 
1111 	if (!entry) {
1112 		err_printk(dev, NULL, "device driver tries "
1113 				"to sync DMA memory it has not allocated "
1114 				"[device address=0x%016llx] [size=%llu bytes]\n",
1115 				(unsigned long long)ref->dev_addr, ref->size);
1116 		goto out;
1117 	}
1118 
1119 	if (ref->size > entry->size) {
1120 		err_printk(dev, entry, "device driver syncs"
1121 				" DMA memory outside allocated range "
1122 				"[device address=0x%016llx] "
1123 				"[allocation size=%llu bytes] "
1124 				"[sync offset+size=%llu]\n",
1125 				entry->dev_addr, entry->size,
1126 				ref->size);
1127 	}
1128 
1129 	if (entry->direction == DMA_BIDIRECTIONAL)
1130 		goto out;
1131 
1132 	if (ref->direction != entry->direction) {
1133 		err_printk(dev, entry, "device driver syncs "
1134 				"DMA memory with different direction "
1135 				"[device address=0x%016llx] [size=%llu bytes] "
1136 				"[mapped with %s] [synced with %s]\n",
1137 				(unsigned long long)ref->dev_addr, entry->size,
1138 				dir2name[entry->direction],
1139 				dir2name[ref->direction]);
1140 	}
1141 
1142 	if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1143 		      !(ref->direction == DMA_TO_DEVICE))
1144 		err_printk(dev, entry, "device driver syncs "
1145 				"device read-only DMA memory for cpu "
1146 				"[device address=0x%016llx] [size=%llu bytes] "
1147 				"[mapped with %s] [synced with %s]\n",
1148 				(unsigned long long)ref->dev_addr, entry->size,
1149 				dir2name[entry->direction],
1150 				dir2name[ref->direction]);
1151 
1152 	if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1153 		       !(ref->direction == DMA_FROM_DEVICE))
1154 		err_printk(dev, entry, "device driver syncs "
1155 				"device write-only DMA memory to device "
1156 				"[device address=0x%016llx] [size=%llu bytes] "
1157 				"[mapped with %s] [synced with %s]\n",
1158 				(unsigned long long)ref->dev_addr, entry->size,
1159 				dir2name[entry->direction],
1160 				dir2name[ref->direction]);
1161 
1162 	if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1163 	    ref->sg_call_ents != entry->sg_call_ents) {
1164 		err_printk(ref->dev, entry, "device driver syncs "
1165 			   "DMA sg list with different entry count "
1166 			   "[map count=%d] [sync count=%d]\n",
1167 			   entry->sg_call_ents, ref->sg_call_ents);
1168 	}
1169 
1170 out:
1171 	put_hash_bucket(bucket, flags);
1172 }
1173 
1174 static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1175 {
1176 #ifdef CONFIG_DMA_API_DEBUG_SG
1177 	unsigned int max_seg = dma_get_max_seg_size(dev);
1178 	u64 start, end, boundary = dma_get_seg_boundary(dev);
1179 
1180 	/*
1181 	 * Either the driver forgot to set dma_parms appropriately, or
1182 	 * whoever generated the list forgot to check them.
1183 	 */
1184 	if (sg->length > max_seg)
1185 		err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1186 			   sg->length, max_seg);
1187 	/*
1188 	 * In some cases this could potentially be the DMA API
1189 	 * implementation's fault, but it would usually imply that
1190 	 * the scatterlist was built inappropriately to begin with.
1191 	 */
1192 	start = sg_dma_address(sg);
1193 	end = start + sg_dma_len(sg) - 1;
1194 	if ((start ^ end) & ~boundary)
1195 		err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1196 			   start, end, boundary);
1197 #endif
1198 }
1199 
1200 void debug_dma_map_single(struct device *dev, const void *addr,
1201 			    unsigned long len)
1202 {
1203 	if (unlikely(dma_debug_disabled()))
1204 		return;
1205 
1206 	if (!virt_addr_valid(addr))
1207 		err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1208 			   addr, len);
1209 
1210 	if (is_vmalloc_addr(addr))
1211 		err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1212 			   addr, len);
1213 }
1214 EXPORT_SYMBOL(debug_dma_map_single);
1215 
1216 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1217 			size_t size, int direction, dma_addr_t dma_addr,
1218 			unsigned long attrs)
1219 {
1220 	struct dma_debug_entry *entry;
1221 
1222 	if (unlikely(dma_debug_disabled()))
1223 		return;
1224 
1225 	if (dma_mapping_error(dev, dma_addr))
1226 		return;
1227 
1228 	entry = dma_entry_alloc();
1229 	if (!entry)
1230 		return;
1231 
1232 	entry->dev       = dev;
1233 	entry->type      = dma_debug_single;
1234 	entry->pfn	 = page_to_pfn(page);
1235 	entry->offset	 = offset;
1236 	entry->dev_addr  = dma_addr;
1237 	entry->size      = size;
1238 	entry->direction = direction;
1239 	entry->map_err_type = MAP_ERR_NOT_CHECKED;
1240 
1241 	check_for_stack(dev, page, offset);
1242 
1243 	if (!PageHighMem(page)) {
1244 		void *addr = page_address(page) + offset;
1245 
1246 		check_for_illegal_area(dev, addr, size);
1247 	}
1248 
1249 	add_dma_entry(entry, attrs);
1250 }
1251 
1252 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1253 {
1254 	struct dma_debug_entry ref;
1255 	struct dma_debug_entry *entry;
1256 	struct hash_bucket *bucket;
1257 	unsigned long flags;
1258 
1259 	if (unlikely(dma_debug_disabled()))
1260 		return;
1261 
1262 	ref.dev = dev;
1263 	ref.dev_addr = dma_addr;
1264 	bucket = get_hash_bucket(&ref, &flags);
1265 
1266 	list_for_each_entry(entry, &bucket->list, list) {
1267 		if (!exact_match(&ref, entry))
1268 			continue;
1269 
1270 		/*
1271 		 * The same physical address can be mapped multiple
1272 		 * times. Without a hardware IOMMU this results in the
1273 		 * same device addresses being put into the dma-debug
1274 		 * hash multiple times too. This can result in false
1275 		 * positives being reported. Therefore we implement a
1276 		 * best-fit algorithm here which updates the first entry
1277 		 * from the hash which fits the reference value and is
1278 		 * not currently listed as being checked.
1279 		 */
1280 		if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1281 			entry->map_err_type = MAP_ERR_CHECKED;
1282 			break;
1283 		}
1284 	}
1285 
1286 	put_hash_bucket(bucket, flags);
1287 }
1288 EXPORT_SYMBOL(debug_dma_mapping_error);
1289 
1290 void debug_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
1291 			  size_t size, int direction)
1292 {
1293 	struct dma_debug_entry ref = {
1294 		.type           = dma_debug_single,
1295 		.dev            = dev,
1296 		.dev_addr       = dma_addr,
1297 		.size           = size,
1298 		.direction      = direction,
1299 	};
1300 
1301 	if (unlikely(dma_debug_disabled()))
1302 		return;
1303 	check_unmap(&ref);
1304 }
1305 
1306 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1307 		      int nents, int mapped_ents, int direction,
1308 		      unsigned long attrs)
1309 {
1310 	struct dma_debug_entry *entry;
1311 	struct scatterlist *s;
1312 	int i;
1313 
1314 	if (unlikely(dma_debug_disabled()))
1315 		return;
1316 
1317 	for_each_sg(sg, s, nents, i) {
1318 		check_for_stack(dev, sg_page(s), s->offset);
1319 		if (!PageHighMem(sg_page(s)))
1320 			check_for_illegal_area(dev, sg_virt(s), s->length);
1321 	}
1322 
1323 	for_each_sg(sg, s, mapped_ents, i) {
1324 		entry = dma_entry_alloc();
1325 		if (!entry)
1326 			return;
1327 
1328 		entry->type           = dma_debug_sg;
1329 		entry->dev            = dev;
1330 		entry->pfn	      = page_to_pfn(sg_page(s));
1331 		entry->offset	      = s->offset;
1332 		entry->size           = sg_dma_len(s);
1333 		entry->dev_addr       = sg_dma_address(s);
1334 		entry->direction      = direction;
1335 		entry->sg_call_ents   = nents;
1336 		entry->sg_mapped_ents = mapped_ents;
1337 
1338 		check_sg_segment(dev, s);
1339 
1340 		add_dma_entry(entry, attrs);
1341 	}
1342 }
1343 
1344 static int get_nr_mapped_entries(struct device *dev,
1345 				 struct dma_debug_entry *ref)
1346 {
1347 	struct dma_debug_entry *entry;
1348 	struct hash_bucket *bucket;
1349 	unsigned long flags;
1350 	int mapped_ents;
1351 
1352 	bucket       = get_hash_bucket(ref, &flags);
1353 	entry        = bucket_find_exact(bucket, ref);
1354 	mapped_ents  = 0;
1355 
1356 	if (entry)
1357 		mapped_ents = entry->sg_mapped_ents;
1358 	put_hash_bucket(bucket, flags);
1359 
1360 	return mapped_ents;
1361 }
1362 
1363 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1364 			int nelems, int dir)
1365 {
1366 	struct scatterlist *s;
1367 	int mapped_ents = 0, i;
1368 
1369 	if (unlikely(dma_debug_disabled()))
1370 		return;
1371 
1372 	for_each_sg(sglist, s, nelems, i) {
1373 
1374 		struct dma_debug_entry ref = {
1375 			.type           = dma_debug_sg,
1376 			.dev            = dev,
1377 			.pfn		= page_to_pfn(sg_page(s)),
1378 			.offset		= s->offset,
1379 			.dev_addr       = sg_dma_address(s),
1380 			.size           = sg_dma_len(s),
1381 			.direction      = dir,
1382 			.sg_call_ents   = nelems,
1383 		};
1384 
1385 		if (mapped_ents && i >= mapped_ents)
1386 			break;
1387 
1388 		if (!i)
1389 			mapped_ents = get_nr_mapped_entries(dev, &ref);
1390 
1391 		check_unmap(&ref);
1392 	}
1393 }
1394 
1395 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1396 			      dma_addr_t dma_addr, void *virt,
1397 			      unsigned long attrs)
1398 {
1399 	struct dma_debug_entry *entry;
1400 
1401 	if (unlikely(dma_debug_disabled()))
1402 		return;
1403 
1404 	if (unlikely(virt == NULL))
1405 		return;
1406 
1407 	/* handle vmalloc and linear addresses */
1408 	if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1409 		return;
1410 
1411 	entry = dma_entry_alloc();
1412 	if (!entry)
1413 		return;
1414 
1415 	entry->type      = dma_debug_coherent;
1416 	entry->dev       = dev;
1417 	entry->offset	 = offset_in_page(virt);
1418 	entry->size      = size;
1419 	entry->dev_addr  = dma_addr;
1420 	entry->direction = DMA_BIDIRECTIONAL;
1421 
1422 	if (is_vmalloc_addr(virt))
1423 		entry->pfn = vmalloc_to_pfn(virt);
1424 	else
1425 		entry->pfn = page_to_pfn(virt_to_page(virt));
1426 
1427 	add_dma_entry(entry, attrs);
1428 }
1429 
1430 void debug_dma_free_coherent(struct device *dev, size_t size,
1431 			 void *virt, dma_addr_t dma_addr)
1432 {
1433 	struct dma_debug_entry ref = {
1434 		.type           = dma_debug_coherent,
1435 		.dev            = dev,
1436 		.offset		= offset_in_page(virt),
1437 		.dev_addr       = dma_addr,
1438 		.size           = size,
1439 		.direction      = DMA_BIDIRECTIONAL,
1440 	};
1441 
1442 	/* handle vmalloc and linear addresses */
1443 	if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1444 		return;
1445 
1446 	if (is_vmalloc_addr(virt))
1447 		ref.pfn = vmalloc_to_pfn(virt);
1448 	else
1449 		ref.pfn = page_to_pfn(virt_to_page(virt));
1450 
1451 	if (unlikely(dma_debug_disabled()))
1452 		return;
1453 
1454 	check_unmap(&ref);
1455 }
1456 
1457 void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1458 			    int direction, dma_addr_t dma_addr,
1459 			    unsigned long attrs)
1460 {
1461 	struct dma_debug_entry *entry;
1462 
1463 	if (unlikely(dma_debug_disabled()))
1464 		return;
1465 
1466 	entry = dma_entry_alloc();
1467 	if (!entry)
1468 		return;
1469 
1470 	entry->type		= dma_debug_resource;
1471 	entry->dev		= dev;
1472 	entry->pfn		= PHYS_PFN(addr);
1473 	entry->offset		= offset_in_page(addr);
1474 	entry->size		= size;
1475 	entry->dev_addr		= dma_addr;
1476 	entry->direction	= direction;
1477 	entry->map_err_type	= MAP_ERR_NOT_CHECKED;
1478 
1479 	add_dma_entry(entry, attrs);
1480 }
1481 
1482 void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1483 			      size_t size, int direction)
1484 {
1485 	struct dma_debug_entry ref = {
1486 		.type           = dma_debug_resource,
1487 		.dev            = dev,
1488 		.dev_addr       = dma_addr,
1489 		.size           = size,
1490 		.direction      = direction,
1491 	};
1492 
1493 	if (unlikely(dma_debug_disabled()))
1494 		return;
1495 
1496 	check_unmap(&ref);
1497 }
1498 
1499 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1500 				   size_t size, int direction)
1501 {
1502 	struct dma_debug_entry ref;
1503 
1504 	if (unlikely(dma_debug_disabled()))
1505 		return;
1506 
1507 	ref.type         = dma_debug_single;
1508 	ref.dev          = dev;
1509 	ref.dev_addr     = dma_handle;
1510 	ref.size         = size;
1511 	ref.direction    = direction;
1512 	ref.sg_call_ents = 0;
1513 
1514 	check_sync(dev, &ref, true);
1515 }
1516 
1517 void debug_dma_sync_single_for_device(struct device *dev,
1518 				      dma_addr_t dma_handle, size_t size,
1519 				      int direction)
1520 {
1521 	struct dma_debug_entry ref;
1522 
1523 	if (unlikely(dma_debug_disabled()))
1524 		return;
1525 
1526 	ref.type         = dma_debug_single;
1527 	ref.dev          = dev;
1528 	ref.dev_addr     = dma_handle;
1529 	ref.size         = size;
1530 	ref.direction    = direction;
1531 	ref.sg_call_ents = 0;
1532 
1533 	check_sync(dev, &ref, false);
1534 }
1535 
1536 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1537 			       int nelems, int direction)
1538 {
1539 	struct scatterlist *s;
1540 	int mapped_ents = 0, i;
1541 
1542 	if (unlikely(dma_debug_disabled()))
1543 		return;
1544 
1545 	for_each_sg(sg, s, nelems, i) {
1546 
1547 		struct dma_debug_entry ref = {
1548 			.type           = dma_debug_sg,
1549 			.dev            = dev,
1550 			.pfn		= page_to_pfn(sg_page(s)),
1551 			.offset		= s->offset,
1552 			.dev_addr       = sg_dma_address(s),
1553 			.size           = sg_dma_len(s),
1554 			.direction      = direction,
1555 			.sg_call_ents   = nelems,
1556 		};
1557 
1558 		if (!i)
1559 			mapped_ents = get_nr_mapped_entries(dev, &ref);
1560 
1561 		if (i >= mapped_ents)
1562 			break;
1563 
1564 		check_sync(dev, &ref, true);
1565 	}
1566 }
1567 
1568 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1569 				  int nelems, int direction)
1570 {
1571 	struct scatterlist *s;
1572 	int mapped_ents = 0, i;
1573 
1574 	if (unlikely(dma_debug_disabled()))
1575 		return;
1576 
1577 	for_each_sg(sg, s, nelems, i) {
1578 
1579 		struct dma_debug_entry ref = {
1580 			.type           = dma_debug_sg,
1581 			.dev            = dev,
1582 			.pfn		= page_to_pfn(sg_page(s)),
1583 			.offset		= s->offset,
1584 			.dev_addr       = sg_dma_address(s),
1585 			.size           = sg_dma_len(s),
1586 			.direction      = direction,
1587 			.sg_call_ents   = nelems,
1588 		};
1589 		if (!i)
1590 			mapped_ents = get_nr_mapped_entries(dev, &ref);
1591 
1592 		if (i >= mapped_ents)
1593 			break;
1594 
1595 		check_sync(dev, &ref, false);
1596 	}
1597 }
1598 
1599 static int __init dma_debug_driver_setup(char *str)
1600 {
1601 	int i;
1602 
1603 	for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1604 		current_driver_name[i] = *str;
1605 		if (*str == 0)
1606 			break;
1607 	}
1608 
1609 	if (current_driver_name[0])
1610 		pr_info("enable driver filter for driver [%s]\n",
1611 			current_driver_name);
1612 
1613 
1614 	return 1;
1615 }
1616 __setup("dma_debug_driver=", dma_debug_driver_setup);
1617