2 * Copyright (C) 2008 Advanced Micro Devices, Inc.
4 * Author: Joerg Roedel <joerg.roedel@amd.com>
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.
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.
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
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>
35 #include <asm/sections.h>
37 #define HASH_SIZE 1024ULL
38 #define HASH_FN_SHIFT 13
39 #define HASH_FN_MASK (HASH_SIZE - 1)
49 MAP_ERR_CHECK_NOT_APPLICABLE
,
54 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
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
70 struct dma_debug_entry
{
71 struct list_head list
;
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
];
88 typedef bool (*match_fn
)(struct dma_debug_entry
*, struct dma_debug_entry
*);
91 struct list_head list
;
93 } ____cacheline_aligned_in_smp
;
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
);
102 /* Global disable flag - will be set in case of an error */
103 static bool global_disable __read_mostly
;
105 /* Early initialization disable flag, set at the end of dma_debug_init */
106 static bool dma_debug_initialized __read_mostly
;
108 static inline bool dma_debug_disabled(void)
110 return global_disable
|| !dma_debug_initialized
;
113 /* Global error count */
114 static u32 error_count
;
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;
121 static u32 num_free_entries
;
122 static u32 min_free_entries
;
123 static u32 nr_total_entries
;
125 /* number of preallocated entries requested by kernel cmdline */
126 static u32 req_entries
;
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
;
138 /* per-driver filter related state */
140 #define NAME_MAX_LEN 64
142 static char current_driver_name
[NAME_MAX_LEN
] __read_mostly
;
143 static struct device_driver
*current_driver __read_mostly
;
145 static DEFINE_RWLOCK(driver_name_lock
);
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",
153 static const char *type2name
[4] = { "single", "page",
154 "scather-gather", "coherent" };
156 static const char *dir2name
[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
157 "DMA_FROM_DEVICE", "DMA_NONE" };
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:
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.
172 static inline void dump_entry_trace(struct dma_debug_entry
*entry
)
174 #ifdef CONFIG_STACKTRACE
176 pr_warning("Mapped at:\n");
177 print_stack_trace(&entry
->stacktrace
, 0);
182 static bool driver_filter(struct device
*dev
)
184 struct device_driver
*drv
;
188 /* driver filter off */
189 if (likely(!current_driver_name
[0]))
192 /* driver filter on and initialized */
193 if (current_driver
&& dev
&& dev
->driver
== current_driver
)
196 /* driver filter on, but we can't filter on a NULL device... */
200 if (current_driver
|| !current_driver_name
[0])
203 /* driver filter on but not yet initialized */
208 /* lock to protect against change of current_driver_name */
209 read_lock_irqsave(&driver_name_lock
, flags
);
213 strncmp(current_driver_name
, drv
->name
, NAME_MAX_LEN
- 1) == 0) {
214 current_driver
= drv
;
218 read_unlock_irqrestore(&driver_name_lock
, flags
);
223 #define err_printk(dev, entry, format, arg...) do { \
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); \
232 if (!show_all_errors && show_num_errors > 0) \
233 show_num_errors -= 1; \
237 * Hash related functions
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.
242 static int hash_fn(struct dma_debug_entry
*entry
)
245 * Hash function is based on the dma address.
246 * We use bits 20-27 here as the index into the hash
248 return (entry
->dev_addr
>> HASH_FN_SHIFT
) & HASH_FN_MASK
;
252 * Request exclusive access to a hash bucket for a given dma_debug_entry.
254 static struct hash_bucket
*get_hash_bucket(struct dma_debug_entry
*entry
,
255 unsigned long *flags
)
257 int idx
= hash_fn(entry
);
258 unsigned long __flags
;
260 spin_lock_irqsave(&dma_entry_hash
[idx
].lock
, __flags
);
262 return &dma_entry_hash
[idx
];
266 * Give up exclusive access to the hash bucket
268 static void put_hash_bucket(struct hash_bucket
*bucket
,
269 unsigned long *flags
)
271 unsigned long __flags
= *flags
;
273 spin_unlock_irqrestore(&bucket
->lock
, __flags
);
276 static bool exact_match(struct dma_debug_entry
*a
, struct dma_debug_entry
*b
)
278 return ((a
->dev_addr
== b
->dev_addr
) &&
279 (a
->dev
== b
->dev
)) ? true : false;
282 static bool containing_match(struct dma_debug_entry
*a
,
283 struct dma_debug_entry
*b
)
285 if (a
->dev
!= b
->dev
)
288 if ((b
->dev_addr
<= a
->dev_addr
) &&
289 ((b
->dev_addr
+ b
->size
) >= (a
->dev_addr
+ a
->size
)))
296 * Search a given entry in the hash bucket list
298 static struct dma_debug_entry
*__hash_bucket_find(struct hash_bucket
*bucket
,
299 struct dma_debug_entry
*ref
,
302 struct dma_debug_entry
*entry
, *ret
= NULL
;
303 int matches
= 0, match_lvl
, last_lvl
= -1;
305 list_for_each_entry(entry
, &bucket
->list
, list
) {
306 if (!match(ref
, entry
))
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.
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;
326 if (match_lvl
== 4) {
327 /* perfect-fit - return the result */
329 } else if (match_lvl
> last_lvl
) {
331 * We found an entry that fits better then the
332 * previous one or it is the 1st match.
334 last_lvl
= match_lvl
;
340 * If we have multiple matches but no perfect-fit, just return
343 ret
= (matches
== 1) ? ret
: NULL
;
348 static struct dma_debug_entry
*bucket_find_exact(struct hash_bucket
*bucket
,
349 struct dma_debug_entry
*ref
)
351 return __hash_bucket_find(bucket
, ref
, exact_match
);
354 static struct dma_debug_entry
*bucket_find_contain(struct hash_bucket
**bucket
,
355 struct dma_debug_entry
*ref
,
356 unsigned long *flags
)
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;
363 while (range
<= max_range
) {
364 entry
= __hash_bucket_find(*bucket
, ref
, containing_match
);
370 * Nothing found, go back a hash bucket
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
);
382 * Add an entry to a hash bucket
384 static void hash_bucket_add(struct hash_bucket
*bucket
,
385 struct dma_debug_entry
*entry
)
387 list_add_tail(&entry
->list
, &bucket
->list
);
391 * Remove entry from a hash bucket list
393 static void hash_bucket_del(struct dma_debug_entry
*entry
)
395 list_del(&entry
->list
);
398 static unsigned long long phys_addr(struct dma_debug_entry
*entry
)
400 return page_to_phys(pfn_to_page(entry
->pfn
)) + entry
->offset
;
404 * Dump mapping entries for debugging purposes
406 void debug_dma_dump_mappings(struct device
*dev
)
410 for (idx
= 0; idx
< HASH_SIZE
; idx
++) {
411 struct hash_bucket
*bucket
= &dma_entry_hash
[idx
];
412 struct dma_debug_entry
*entry
;
415 spin_lock_irqsave(&bucket
->lock
, flags
);
417 list_for_each_entry(entry
, &bucket
->list
, list
) {
418 if (!dev
|| dev
== 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
]);
429 spin_unlock_irqrestore(&bucket
->lock
, flags
);
432 EXPORT_SYMBOL(debug_dma_dump_mappings
);
435 * For each mapping (initial cacheline in the case of
436 * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
437 * scatterlist, or the cacheline specified in dma_map_single) insert
438 * into this tree using the cacheline as the key. At
439 * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry. If
440 * the entry already exists at insertion time add a tag as a reference
441 * count for the overlapping mappings. For now, the overlap tracking
442 * just ensures that 'unmaps' balance 'maps' before marking the
443 * cacheline idle, but we should also be flagging overlaps as an API
446 * Memory usage is mostly constrained by the maximum number of available
447 * dma-debug entries in that we need a free dma_debug_entry before
448 * inserting into the tree. In the case of dma_map_page and
449 * dma_alloc_coherent there is only one dma_debug_entry and one
450 * dma_active_cacheline entry to track per event. dma_map_sg(), on the
451 * other hand, consumes a single dma_debug_entry, but inserts 'nents'
452 * entries into the tree.
454 * At any time debug_dma_assert_idle() can be called to trigger a
455 * warning if any cachelines in the given page are in the active set.
457 static RADIX_TREE(dma_active_cacheline
, GFP_NOWAIT
);
458 static DEFINE_SPINLOCK(radix_lock
);
459 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
460 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
461 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
463 static phys_addr_t
to_cacheline_number(struct dma_debug_entry
*entry
)
465 return (entry
->pfn
<< CACHELINE_PER_PAGE_SHIFT
) +
466 (entry
->offset
>> L1_CACHE_SHIFT
);
469 static int active_cacheline_read_overlap(phys_addr_t cln
)
473 for (i
= RADIX_TREE_MAX_TAGS
- 1; i
>= 0; i
--)
474 if (radix_tree_tag_get(&dma_active_cacheline
, cln
, i
))
479 static int active_cacheline_set_overlap(phys_addr_t cln
, int overlap
)
483 if (overlap
> ACTIVE_CACHELINE_MAX_OVERLAP
|| overlap
< 0)
486 for (i
= RADIX_TREE_MAX_TAGS
- 1; i
>= 0; i
--)
487 if (overlap
& 1 << i
)
488 radix_tree_tag_set(&dma_active_cacheline
, cln
, i
);
490 radix_tree_tag_clear(&dma_active_cacheline
, cln
, i
);
495 static void active_cacheline_inc_overlap(phys_addr_t cln
)
497 int overlap
= active_cacheline_read_overlap(cln
);
499 overlap
= active_cacheline_set_overlap(cln
, ++overlap
);
501 /* If we overflowed the overlap counter then we're potentially
502 * leaking dma-mappings. Otherwise, if maps and unmaps are
503 * balanced then this overflow may cause false negatives in
504 * debug_dma_assert_idle() as the cacheline may be marked idle
507 WARN_ONCE(overlap
> ACTIVE_CACHELINE_MAX_OVERLAP
,
508 "DMA-API: exceeded %d overlapping mappings of cacheline %pa\n",
509 ACTIVE_CACHELINE_MAX_OVERLAP
, &cln
);
512 static int active_cacheline_dec_overlap(phys_addr_t cln
)
514 int overlap
= active_cacheline_read_overlap(cln
);
516 return active_cacheline_set_overlap(cln
, --overlap
);
519 static int active_cacheline_insert(struct dma_debug_entry
*entry
)
521 phys_addr_t cln
= to_cacheline_number(entry
);
525 /* If the device is not writing memory then we don't have any
526 * concerns about the cpu consuming stale data. This mitigates
527 * legitimate usages of overlapping mappings.
529 if (entry
->direction
== DMA_TO_DEVICE
)
532 spin_lock_irqsave(&radix_lock
, flags
);
533 rc
= radix_tree_insert(&dma_active_cacheline
, cln
, entry
);
535 active_cacheline_inc_overlap(cln
);
536 spin_unlock_irqrestore(&radix_lock
, flags
);
541 static void active_cacheline_remove(struct dma_debug_entry
*entry
)
543 phys_addr_t cln
= to_cacheline_number(entry
);
546 /* ...mirror the insert case */
547 if (entry
->direction
== DMA_TO_DEVICE
)
550 spin_lock_irqsave(&radix_lock
, flags
);
551 /* since we are counting overlaps the final put of the
552 * cacheline will occur when the overlap count is 0.
553 * active_cacheline_dec_overlap() returns -1 in that case
555 if (active_cacheline_dec_overlap(cln
) < 0)
556 radix_tree_delete(&dma_active_cacheline
, cln
);
557 spin_unlock_irqrestore(&radix_lock
, flags
);
561 * debug_dma_assert_idle() - assert that a page is not undergoing dma
562 * @page: page to lookup in the dma_active_cacheline tree
564 * Place a call to this routine in cases where the cpu touching the page
565 * before the dma completes (page is dma_unmapped) will lead to data
568 void debug_dma_assert_idle(struct page
*page
)
570 static struct dma_debug_entry
*ents
[CACHELINES_PER_PAGE
];
571 struct dma_debug_entry
*entry
= NULL
;
572 void **results
= (void **) &ents
;
573 unsigned int nents
, i
;
577 if (dma_debug_disabled())
583 cln
= (phys_addr_t
) page_to_pfn(page
) << CACHELINE_PER_PAGE_SHIFT
;
584 spin_lock_irqsave(&radix_lock
, flags
);
585 nents
= radix_tree_gang_lookup(&dma_active_cacheline
, results
, cln
,
586 CACHELINES_PER_PAGE
);
587 for (i
= 0; i
< nents
; i
++) {
588 phys_addr_t ent_cln
= to_cacheline_number(ents
[i
]);
590 if (ent_cln
== cln
) {
593 } else if (ent_cln
>= cln
+ CACHELINES_PER_PAGE
)
596 spin_unlock_irqrestore(&radix_lock
, flags
);
601 cln
= to_cacheline_number(entry
);
602 err_printk(entry
->dev
, entry
,
603 "DMA-API: cpu touching an active dma mapped cacheline [cln=%pa]\n",
608 * Wrapper function for adding an entry to the hash.
609 * This function takes care of locking itself.
611 static void add_dma_entry(struct dma_debug_entry
*entry
)
613 struct hash_bucket
*bucket
;
617 bucket
= get_hash_bucket(entry
, &flags
);
618 hash_bucket_add(bucket
, entry
);
619 put_hash_bucket(bucket
, &flags
);
621 rc
= active_cacheline_insert(entry
);
623 pr_err("DMA-API: cacheline tracking ENOMEM, dma-debug disabled\n");
624 global_disable
= true;
627 /* TODO: report -EEXIST errors here as overlapping mappings are
628 * not supported by the DMA API
632 static struct dma_debug_entry
*__dma_entry_alloc(void)
634 struct dma_debug_entry
*entry
;
636 entry
= list_entry(free_entries
.next
, struct dma_debug_entry
, list
);
637 list_del(&entry
->list
);
638 memset(entry
, 0, sizeof(*entry
));
640 num_free_entries
-= 1;
641 if (num_free_entries
< min_free_entries
)
642 min_free_entries
= num_free_entries
;
647 /* struct dma_entry allocator
649 * The next two functions implement the allocator for
650 * struct dma_debug_entries.
652 static struct dma_debug_entry
*dma_entry_alloc(void)
654 struct dma_debug_entry
*entry
;
657 spin_lock_irqsave(&free_entries_lock
, flags
);
659 if (list_empty(&free_entries
)) {
660 global_disable
= true;
661 spin_unlock_irqrestore(&free_entries_lock
, flags
);
662 pr_err("DMA-API: debugging out of memory - disabling\n");
666 entry
= __dma_entry_alloc();
668 spin_unlock_irqrestore(&free_entries_lock
, flags
);
670 #ifdef CONFIG_STACKTRACE
671 entry
->stacktrace
.max_entries
= DMA_DEBUG_STACKTRACE_ENTRIES
;
672 entry
->stacktrace
.entries
= entry
->st_entries
;
673 entry
->stacktrace
.skip
= 2;
674 save_stack_trace(&entry
->stacktrace
);
680 static void dma_entry_free(struct dma_debug_entry
*entry
)
684 active_cacheline_remove(entry
);
687 * add to beginning of the list - this way the entries are
688 * more likely cache hot when they are reallocated.
690 spin_lock_irqsave(&free_entries_lock
, flags
);
691 list_add(&entry
->list
, &free_entries
);
692 num_free_entries
+= 1;
693 spin_unlock_irqrestore(&free_entries_lock
, flags
);
696 int dma_debug_resize_entries(u32 num_entries
)
698 int i
, delta
, ret
= 0;
700 struct dma_debug_entry
*entry
;
703 spin_lock_irqsave(&free_entries_lock
, flags
);
705 if (nr_total_entries
< num_entries
) {
706 delta
= num_entries
- nr_total_entries
;
708 spin_unlock_irqrestore(&free_entries_lock
, flags
);
710 for (i
= 0; i
< delta
; i
++) {
711 entry
= kzalloc(sizeof(*entry
), GFP_KERNEL
);
715 list_add_tail(&entry
->list
, &tmp
);
718 spin_lock_irqsave(&free_entries_lock
, flags
);
720 list_splice(&tmp
, &free_entries
);
721 nr_total_entries
+= i
;
722 num_free_entries
+= i
;
724 delta
= nr_total_entries
- num_entries
;
726 for (i
= 0; i
< delta
&& !list_empty(&free_entries
); i
++) {
727 entry
= __dma_entry_alloc();
731 nr_total_entries
-= i
;
734 if (nr_total_entries
!= num_entries
)
737 spin_unlock_irqrestore(&free_entries_lock
, flags
);
741 EXPORT_SYMBOL(dma_debug_resize_entries
);
744 * DMA-API debugging init code
746 * The init code does two things:
747 * 1. Initialize core data structures
748 * 2. Preallocate a given number of dma_debug_entry structs
751 static int prealloc_memory(u32 num_entries
)
753 struct dma_debug_entry
*entry
, *next_entry
;
756 for (i
= 0; i
< num_entries
; ++i
) {
757 entry
= kzalloc(sizeof(*entry
), GFP_KERNEL
);
761 list_add_tail(&entry
->list
, &free_entries
);
764 num_free_entries
= num_entries
;
765 min_free_entries
= num_entries
;
767 pr_info("DMA-API: preallocated %d debug entries\n", num_entries
);
773 list_for_each_entry_safe(entry
, next_entry
, &free_entries
, list
) {
774 list_del(&entry
->list
);
781 static ssize_t
filter_read(struct file
*file
, char __user
*user_buf
,
782 size_t count
, loff_t
*ppos
)
784 char buf
[NAME_MAX_LEN
+ 1];
788 if (!current_driver_name
[0])
792 * We can't copy to userspace directly because current_driver_name can
793 * only be read under the driver_name_lock with irqs disabled. So
794 * create a temporary copy first.
796 read_lock_irqsave(&driver_name_lock
, flags
);
797 len
= scnprintf(buf
, NAME_MAX_LEN
+ 1, "%s\n", current_driver_name
);
798 read_unlock_irqrestore(&driver_name_lock
, flags
);
800 return simple_read_from_buffer(user_buf
, count
, ppos
, buf
, len
);
803 static ssize_t
filter_write(struct file
*file
, const char __user
*userbuf
,
804 size_t count
, loff_t
*ppos
)
806 char buf
[NAME_MAX_LEN
];
812 * We can't copy from userspace directly. Access to
813 * current_driver_name is protected with a write_lock with irqs
814 * disabled. Since copy_from_user can fault and may sleep we
815 * need to copy to temporary buffer first
817 len
= min(count
, (size_t)(NAME_MAX_LEN
- 1));
818 if (copy_from_user(buf
, userbuf
, len
))
823 write_lock_irqsave(&driver_name_lock
, flags
);
826 * Now handle the string we got from userspace very carefully.
828 * - only use the first token we got
829 * - token delimiter is everything looking like a space
830 * character (' ', '\n', '\t' ...)
833 if (!isalnum(buf
[0])) {
835 * If the first character userspace gave us is not
836 * alphanumerical then assume the filter should be
839 if (current_driver_name
[0])
840 pr_info("DMA-API: switching off dma-debug driver filter\n");
841 current_driver_name
[0] = 0;
842 current_driver
= NULL
;
847 * Now parse out the first token and use it as the name for the
848 * driver to filter for.
850 for (i
= 0; i
< NAME_MAX_LEN
- 1; ++i
) {
851 current_driver_name
[i
] = buf
[i
];
852 if (isspace(buf
[i
]) || buf
[i
] == ' ' || buf
[i
] == 0)
855 current_driver_name
[i
] = 0;
856 current_driver
= NULL
;
858 pr_info("DMA-API: enable driver filter for driver [%s]\n",
859 current_driver_name
);
862 write_unlock_irqrestore(&driver_name_lock
, flags
);
867 static const struct file_operations filter_fops
= {
869 .write
= filter_write
,
870 .llseek
= default_llseek
,
873 static int dma_debug_fs_init(void)
875 dma_debug_dent
= debugfs_create_dir("dma-api", NULL
);
876 if (!dma_debug_dent
) {
877 pr_err("DMA-API: can not create debugfs directory\n");
881 global_disable_dent
= debugfs_create_bool("disabled", 0444,
884 if (!global_disable_dent
)
887 error_count_dent
= debugfs_create_u32("error_count", 0444,
888 dma_debug_dent
, &error_count
);
889 if (!error_count_dent
)
892 show_all_errors_dent
= debugfs_create_u32("all_errors", 0644,
895 if (!show_all_errors_dent
)
898 show_num_errors_dent
= debugfs_create_u32("num_errors", 0644,
901 if (!show_num_errors_dent
)
904 num_free_entries_dent
= debugfs_create_u32("num_free_entries", 0444,
907 if (!num_free_entries_dent
)
910 min_free_entries_dent
= debugfs_create_u32("min_free_entries", 0444,
913 if (!min_free_entries_dent
)
916 filter_dent
= debugfs_create_file("driver_filter", 0644,
917 dma_debug_dent
, NULL
, &filter_fops
);
924 debugfs_remove_recursive(dma_debug_dent
);
929 static int device_dma_allocations(struct device
*dev
, struct dma_debug_entry
**out_entry
)
931 struct dma_debug_entry
*entry
;
935 local_irq_save(flags
);
937 for (i
= 0; i
< HASH_SIZE
; ++i
) {
938 spin_lock(&dma_entry_hash
[i
].lock
);
939 list_for_each_entry(entry
, &dma_entry_hash
[i
].list
, list
) {
940 if (entry
->dev
== dev
) {
945 spin_unlock(&dma_entry_hash
[i
].lock
);
948 local_irq_restore(flags
);
953 static int dma_debug_device_change(struct notifier_block
*nb
, unsigned long action
, void *data
)
955 struct device
*dev
= data
;
956 struct dma_debug_entry
*uninitialized_var(entry
);
959 if (dma_debug_disabled())
963 case BUS_NOTIFY_UNBOUND_DRIVER
:
964 count
= device_dma_allocations(dev
, &entry
);
967 err_printk(dev
, entry
, "DMA-API: device driver has pending "
968 "DMA allocations while released from device "
970 "One of leaked entries details: "
971 "[device address=0x%016llx] [size=%llu bytes] "
972 "[mapped with %s] [mapped as %s]\n",
973 count
, entry
->dev_addr
, entry
->size
,
974 dir2name
[entry
->direction
], type2name
[entry
->type
]);
983 void dma_debug_add_bus(struct bus_type
*bus
)
985 struct notifier_block
*nb
;
987 if (dma_debug_disabled())
990 nb
= kzalloc(sizeof(struct notifier_block
), GFP_KERNEL
);
992 pr_err("dma_debug_add_bus: out of memory\n");
996 nb
->notifier_call
= dma_debug_device_change
;
998 bus_register_notifier(bus
, nb
);
1002 * Let the architectures decide how many entries should be preallocated.
1004 void dma_debug_init(u32 num_entries
)
1008 /* Do not use dma_debug_initialized here, since we really want to be
1009 * called to set dma_debug_initialized
1014 for (i
= 0; i
< HASH_SIZE
; ++i
) {
1015 INIT_LIST_HEAD(&dma_entry_hash
[i
].list
);
1016 spin_lock_init(&dma_entry_hash
[i
].lock
);
1019 if (dma_debug_fs_init() != 0) {
1020 pr_err("DMA-API: error creating debugfs entries - disabling\n");
1021 global_disable
= true;
1027 num_entries
= req_entries
;
1029 if (prealloc_memory(num_entries
) != 0) {
1030 pr_err("DMA-API: debugging out of memory error - disabled\n");
1031 global_disable
= true;
1036 nr_total_entries
= num_free_entries
;
1038 dma_debug_initialized
= true;
1040 pr_info("DMA-API: debugging enabled by kernel config\n");
1043 static __init
int dma_debug_cmdline(char *str
)
1048 if (strncmp(str
, "off", 3) == 0) {
1049 pr_info("DMA-API: debugging disabled on kernel command line\n");
1050 global_disable
= true;
1056 static __init
int dma_debug_entries_cmdline(char *str
)
1063 res
= get_option(&str
, &req_entries
);
1071 __setup("dma_debug=", dma_debug_cmdline
);
1072 __setup("dma_debug_entries=", dma_debug_entries_cmdline
);
1074 static void check_unmap(struct dma_debug_entry
*ref
)
1076 struct dma_debug_entry
*entry
;
1077 struct hash_bucket
*bucket
;
1078 unsigned long flags
;
1080 bucket
= get_hash_bucket(ref
, &flags
);
1081 entry
= bucket_find_exact(bucket
, ref
);
1084 /* must drop lock before calling dma_mapping_error */
1085 put_hash_bucket(bucket
, &flags
);
1087 if (dma_mapping_error(ref
->dev
, ref
->dev_addr
)) {
1088 err_printk(ref
->dev
, NULL
,
1089 "DMA-API: device driver tries to free an "
1090 "invalid DMA memory address\n");
1092 err_printk(ref
->dev
, NULL
,
1093 "DMA-API: device driver tries to free DMA "
1094 "memory it has not allocated [device "
1095 "address=0x%016llx] [size=%llu bytes]\n",
1096 ref
->dev_addr
, ref
->size
);
1101 if (ref
->size
!= entry
->size
) {
1102 err_printk(ref
->dev
, entry
, "DMA-API: device driver frees "
1103 "DMA memory with different size "
1104 "[device address=0x%016llx] [map size=%llu bytes] "
1105 "[unmap size=%llu bytes]\n",
1106 ref
->dev_addr
, entry
->size
, ref
->size
);
1109 if (ref
->type
!= entry
->type
) {
1110 err_printk(ref
->dev
, entry
, "DMA-API: device driver frees "
1111 "DMA memory with wrong function "
1112 "[device address=0x%016llx] [size=%llu bytes] "
1113 "[mapped as %s] [unmapped as %s]\n",
1114 ref
->dev_addr
, ref
->size
,
1115 type2name
[entry
->type
], type2name
[ref
->type
]);
1116 } else if ((entry
->type
== dma_debug_coherent
) &&
1117 (phys_addr(ref
) != phys_addr(entry
))) {
1118 err_printk(ref
->dev
, entry
, "DMA-API: device driver frees "
1119 "DMA memory with different CPU address "
1120 "[device address=0x%016llx] [size=%llu bytes] "
1121 "[cpu alloc address=0x%016llx] "
1122 "[cpu free address=0x%016llx]",
1123 ref
->dev_addr
, ref
->size
,
1128 if (ref
->sg_call_ents
&& ref
->type
== dma_debug_sg
&&
1129 ref
->sg_call_ents
!= entry
->sg_call_ents
) {
1130 err_printk(ref
->dev
, entry
, "DMA-API: device driver frees "
1131 "DMA sg list with different entry count "
1132 "[map count=%d] [unmap count=%d]\n",
1133 entry
->sg_call_ents
, ref
->sg_call_ents
);
1137 * This may be no bug in reality - but most implementations of the
1138 * DMA API don't handle this properly, so check for it here
1140 if (ref
->direction
!= entry
->direction
) {
1141 err_printk(ref
->dev
, entry
, "DMA-API: device driver frees "
1142 "DMA memory with different direction "
1143 "[device address=0x%016llx] [size=%llu bytes] "
1144 "[mapped with %s] [unmapped with %s]\n",
1145 ref
->dev_addr
, ref
->size
,
1146 dir2name
[entry
->direction
],
1147 dir2name
[ref
->direction
]);
1150 if (entry
->map_err_type
== MAP_ERR_NOT_CHECKED
) {
1151 err_printk(ref
->dev
, entry
,
1152 "DMA-API: device driver failed to check map error"
1153 "[device address=0x%016llx] [size=%llu bytes] "
1155 ref
->dev_addr
, ref
->size
,
1156 type2name
[entry
->type
]);
1159 hash_bucket_del(entry
);
1160 dma_entry_free(entry
);
1162 put_hash_bucket(bucket
, &flags
);
1165 static void check_for_stack(struct device
*dev
, void *addr
)
1167 if (object_is_on_stack(addr
))
1168 err_printk(dev
, NULL
, "DMA-API: device driver maps memory from "
1169 "stack [addr=%p]\n", addr
);
1172 static inline bool overlap(void *addr
, unsigned long len
, void *start
, void *end
)
1174 unsigned long a1
= (unsigned long)addr
;
1175 unsigned long b1
= a1
+ len
;
1176 unsigned long a2
= (unsigned long)start
;
1177 unsigned long b2
= (unsigned long)end
;
1179 return !(b1
<= a2
|| a1
>= b2
);
1182 static void check_for_illegal_area(struct device
*dev
, void *addr
, unsigned long len
)
1184 if (overlap(addr
, len
, _stext
, _etext
) ||
1185 overlap(addr
, len
, __start_rodata
, __end_rodata
))
1186 err_printk(dev
, NULL
, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr
, len
);
1189 static void check_sync(struct device
*dev
,
1190 struct dma_debug_entry
*ref
,
1193 struct dma_debug_entry
*entry
;
1194 struct hash_bucket
*bucket
;
1195 unsigned long flags
;
1197 bucket
= get_hash_bucket(ref
, &flags
);
1199 entry
= bucket_find_contain(&bucket
, ref
, &flags
);
1202 err_printk(dev
, NULL
, "DMA-API: device driver tries "
1203 "to sync DMA memory it has not allocated "
1204 "[device address=0x%016llx] [size=%llu bytes]\n",
1205 (unsigned long long)ref
->dev_addr
, ref
->size
);
1209 if (ref
->size
> entry
->size
) {
1210 err_printk(dev
, entry
, "DMA-API: device driver syncs"
1211 " DMA memory outside allocated range "
1212 "[device address=0x%016llx] "
1213 "[allocation size=%llu bytes] "
1214 "[sync offset+size=%llu]\n",
1215 entry
->dev_addr
, entry
->size
,
1219 if (entry
->direction
== DMA_BIDIRECTIONAL
)
1222 if (ref
->direction
!= entry
->direction
) {
1223 err_printk(dev
, entry
, "DMA-API: device driver syncs "
1224 "DMA memory with different direction "
1225 "[device address=0x%016llx] [size=%llu bytes] "
1226 "[mapped with %s] [synced with %s]\n",
1227 (unsigned long long)ref
->dev_addr
, entry
->size
,
1228 dir2name
[entry
->direction
],
1229 dir2name
[ref
->direction
]);
1232 if (to_cpu
&& !(entry
->direction
== DMA_FROM_DEVICE
) &&
1233 !(ref
->direction
== DMA_TO_DEVICE
))
1234 err_printk(dev
, entry
, "DMA-API: device driver syncs "
1235 "device read-only DMA memory for cpu "
1236 "[device address=0x%016llx] [size=%llu bytes] "
1237 "[mapped with %s] [synced with %s]\n",
1238 (unsigned long long)ref
->dev_addr
, entry
->size
,
1239 dir2name
[entry
->direction
],
1240 dir2name
[ref
->direction
]);
1242 if (!to_cpu
&& !(entry
->direction
== DMA_TO_DEVICE
) &&
1243 !(ref
->direction
== DMA_FROM_DEVICE
))
1244 err_printk(dev
, entry
, "DMA-API: device driver syncs "
1245 "device write-only DMA memory to device "
1246 "[device address=0x%016llx] [size=%llu bytes] "
1247 "[mapped with %s] [synced with %s]\n",
1248 (unsigned long long)ref
->dev_addr
, entry
->size
,
1249 dir2name
[entry
->direction
],
1250 dir2name
[ref
->direction
]);
1252 if (ref
->sg_call_ents
&& ref
->type
== dma_debug_sg
&&
1253 ref
->sg_call_ents
!= entry
->sg_call_ents
) {
1254 err_printk(ref
->dev
, entry
, "DMA-API: device driver syncs "
1255 "DMA sg list with different entry count "
1256 "[map count=%d] [sync count=%d]\n",
1257 entry
->sg_call_ents
, ref
->sg_call_ents
);
1261 put_hash_bucket(bucket
, &flags
);
1264 void debug_dma_map_page(struct device
*dev
, struct page
*page
, size_t offset
,
1265 size_t size
, int direction
, dma_addr_t dma_addr
,
1268 struct dma_debug_entry
*entry
;
1270 if (unlikely(dma_debug_disabled()))
1273 if (dma_mapping_error(dev
, dma_addr
))
1276 entry
= dma_entry_alloc();
1281 entry
->type
= dma_debug_page
;
1282 entry
->pfn
= page_to_pfn(page
);
1283 entry
->offset
= offset
,
1284 entry
->dev_addr
= dma_addr
;
1286 entry
->direction
= direction
;
1287 entry
->map_err_type
= MAP_ERR_NOT_CHECKED
;
1290 entry
->type
= dma_debug_single
;
1292 if (!PageHighMem(page
)) {
1293 void *addr
= page_address(page
) + offset
;
1295 check_for_stack(dev
, addr
);
1296 check_for_illegal_area(dev
, addr
, size
);
1299 add_dma_entry(entry
);
1301 EXPORT_SYMBOL(debug_dma_map_page
);
1303 void debug_dma_mapping_error(struct device
*dev
, dma_addr_t dma_addr
)
1305 struct dma_debug_entry ref
;
1306 struct dma_debug_entry
*entry
;
1307 struct hash_bucket
*bucket
;
1308 unsigned long flags
;
1310 if (unlikely(dma_debug_disabled()))
1314 ref
.dev_addr
= dma_addr
;
1315 bucket
= get_hash_bucket(&ref
, &flags
);
1317 list_for_each_entry(entry
, &bucket
->list
, list
) {
1318 if (!exact_match(&ref
, entry
))
1322 * The same physical address can be mapped multiple
1323 * times. Without a hardware IOMMU this results in the
1324 * same device addresses being put into the dma-debug
1325 * hash multiple times too. This can result in false
1326 * positives being reported. Therefore we implement a
1327 * best-fit algorithm here which updates the first entry
1328 * from the hash which fits the reference value and is
1329 * not currently listed as being checked.
1331 if (entry
->map_err_type
== MAP_ERR_NOT_CHECKED
) {
1332 entry
->map_err_type
= MAP_ERR_CHECKED
;
1337 put_hash_bucket(bucket
, &flags
);
1339 EXPORT_SYMBOL(debug_dma_mapping_error
);
1341 void debug_dma_unmap_page(struct device
*dev
, dma_addr_t addr
,
1342 size_t size
, int direction
, bool map_single
)
1344 struct dma_debug_entry ref
= {
1345 .type
= dma_debug_page
,
1349 .direction
= direction
,
1352 if (unlikely(dma_debug_disabled()))
1356 ref
.type
= dma_debug_single
;
1360 EXPORT_SYMBOL(debug_dma_unmap_page
);
1362 void debug_dma_map_sg(struct device
*dev
, struct scatterlist
*sg
,
1363 int nents
, int mapped_ents
, int direction
)
1365 struct dma_debug_entry
*entry
;
1366 struct scatterlist
*s
;
1369 if (unlikely(dma_debug_disabled()))
1372 for_each_sg(sg
, s
, mapped_ents
, i
) {
1373 entry
= dma_entry_alloc();
1377 entry
->type
= dma_debug_sg
;
1379 entry
->pfn
= page_to_pfn(sg_page(s
));
1380 entry
->offset
= s
->offset
,
1381 entry
->size
= sg_dma_len(s
);
1382 entry
->dev_addr
= sg_dma_address(s
);
1383 entry
->direction
= direction
;
1384 entry
->sg_call_ents
= nents
;
1385 entry
->sg_mapped_ents
= mapped_ents
;
1387 if (!PageHighMem(sg_page(s
))) {
1388 check_for_stack(dev
, sg_virt(s
));
1389 check_for_illegal_area(dev
, sg_virt(s
), sg_dma_len(s
));
1392 add_dma_entry(entry
);
1395 EXPORT_SYMBOL(debug_dma_map_sg
);
1397 static int get_nr_mapped_entries(struct device
*dev
,
1398 struct dma_debug_entry
*ref
)
1400 struct dma_debug_entry
*entry
;
1401 struct hash_bucket
*bucket
;
1402 unsigned long flags
;
1405 bucket
= get_hash_bucket(ref
, &flags
);
1406 entry
= bucket_find_exact(bucket
, ref
);
1410 mapped_ents
= entry
->sg_mapped_ents
;
1411 put_hash_bucket(bucket
, &flags
);
1416 void debug_dma_unmap_sg(struct device
*dev
, struct scatterlist
*sglist
,
1417 int nelems
, int dir
)
1419 struct scatterlist
*s
;
1420 int mapped_ents
= 0, i
;
1422 if (unlikely(dma_debug_disabled()))
1425 for_each_sg(sglist
, s
, nelems
, i
) {
1427 struct dma_debug_entry ref
= {
1428 .type
= dma_debug_sg
,
1430 .pfn
= page_to_pfn(sg_page(s
)),
1431 .offset
= s
->offset
,
1432 .dev_addr
= sg_dma_address(s
),
1433 .size
= sg_dma_len(s
),
1435 .sg_call_ents
= nelems
,
1438 if (mapped_ents
&& i
>= mapped_ents
)
1442 mapped_ents
= get_nr_mapped_entries(dev
, &ref
);
1447 EXPORT_SYMBOL(debug_dma_unmap_sg
);
1449 void debug_dma_alloc_coherent(struct device
*dev
, size_t size
,
1450 dma_addr_t dma_addr
, void *virt
)
1452 struct dma_debug_entry
*entry
;
1454 if (unlikely(dma_debug_disabled()))
1457 if (unlikely(virt
== NULL
))
1460 entry
= dma_entry_alloc();
1464 entry
->type
= dma_debug_coherent
;
1466 entry
->pfn
= page_to_pfn(virt_to_page(virt
));
1467 entry
->offset
= (size_t) virt
& ~PAGE_MASK
;
1469 entry
->dev_addr
= dma_addr
;
1470 entry
->direction
= DMA_BIDIRECTIONAL
;
1472 add_dma_entry(entry
);
1474 EXPORT_SYMBOL(debug_dma_alloc_coherent
);
1476 void debug_dma_free_coherent(struct device
*dev
, size_t size
,
1477 void *virt
, dma_addr_t addr
)
1479 struct dma_debug_entry ref
= {
1480 .type
= dma_debug_coherent
,
1482 .pfn
= page_to_pfn(virt_to_page(virt
)),
1483 .offset
= (size_t) virt
& ~PAGE_MASK
,
1486 .direction
= DMA_BIDIRECTIONAL
,
1489 if (unlikely(dma_debug_disabled()))
1494 EXPORT_SYMBOL(debug_dma_free_coherent
);
1496 void debug_dma_sync_single_for_cpu(struct device
*dev
, dma_addr_t dma_handle
,
1497 size_t size
, int direction
)
1499 struct dma_debug_entry ref
;
1501 if (unlikely(dma_debug_disabled()))
1504 ref
.type
= dma_debug_single
;
1506 ref
.dev_addr
= dma_handle
;
1508 ref
.direction
= direction
;
1509 ref
.sg_call_ents
= 0;
1511 check_sync(dev
, &ref
, true);
1513 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu
);
1515 void debug_dma_sync_single_for_device(struct device
*dev
,
1516 dma_addr_t dma_handle
, size_t size
,
1519 struct dma_debug_entry ref
;
1521 if (unlikely(dma_debug_disabled()))
1524 ref
.type
= dma_debug_single
;
1526 ref
.dev_addr
= dma_handle
;
1528 ref
.direction
= direction
;
1529 ref
.sg_call_ents
= 0;
1531 check_sync(dev
, &ref
, false);
1533 EXPORT_SYMBOL(debug_dma_sync_single_for_device
);
1535 void debug_dma_sync_single_range_for_cpu(struct device
*dev
,
1536 dma_addr_t dma_handle
,
1537 unsigned long offset
, size_t size
,
1540 struct dma_debug_entry ref
;
1542 if (unlikely(dma_debug_disabled()))
1545 ref
.type
= dma_debug_single
;
1547 ref
.dev_addr
= dma_handle
;
1548 ref
.size
= offset
+ size
;
1549 ref
.direction
= direction
;
1550 ref
.sg_call_ents
= 0;
1552 check_sync(dev
, &ref
, true);
1554 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu
);
1556 void debug_dma_sync_single_range_for_device(struct device
*dev
,
1557 dma_addr_t dma_handle
,
1558 unsigned long offset
,
1559 size_t size
, int direction
)
1561 struct dma_debug_entry ref
;
1563 if (unlikely(dma_debug_disabled()))
1566 ref
.type
= dma_debug_single
;
1568 ref
.dev_addr
= dma_handle
;
1569 ref
.size
= offset
+ size
;
1570 ref
.direction
= direction
;
1571 ref
.sg_call_ents
= 0;
1573 check_sync(dev
, &ref
, false);
1575 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device
);
1577 void debug_dma_sync_sg_for_cpu(struct device
*dev
, struct scatterlist
*sg
,
1578 int nelems
, int direction
)
1580 struct scatterlist
*s
;
1581 int mapped_ents
= 0, i
;
1583 if (unlikely(dma_debug_disabled()))
1586 for_each_sg(sg
, s
, nelems
, i
) {
1588 struct dma_debug_entry ref
= {
1589 .type
= dma_debug_sg
,
1591 .pfn
= page_to_pfn(sg_page(s
)),
1592 .offset
= s
->offset
,
1593 .dev_addr
= sg_dma_address(s
),
1594 .size
= sg_dma_len(s
),
1595 .direction
= direction
,
1596 .sg_call_ents
= nelems
,
1600 mapped_ents
= get_nr_mapped_entries(dev
, &ref
);
1602 if (i
>= mapped_ents
)
1605 check_sync(dev
, &ref
, true);
1608 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu
);
1610 void debug_dma_sync_sg_for_device(struct device
*dev
, struct scatterlist
*sg
,
1611 int nelems
, int direction
)
1613 struct scatterlist
*s
;
1614 int mapped_ents
= 0, i
;
1616 if (unlikely(dma_debug_disabled()))
1619 for_each_sg(sg
, s
, nelems
, i
) {
1621 struct dma_debug_entry ref
= {
1622 .type
= dma_debug_sg
,
1624 .pfn
= page_to_pfn(sg_page(s
)),
1625 .offset
= s
->offset
,
1626 .dev_addr
= sg_dma_address(s
),
1627 .size
= sg_dma_len(s
),
1628 .direction
= direction
,
1629 .sg_call_ents
= nelems
,
1632 mapped_ents
= get_nr_mapped_entries(dev
, &ref
);
1634 if (i
>= mapped_ents
)
1637 check_sync(dev
, &ref
, false);
1640 EXPORT_SYMBOL(debug_dma_sync_sg_for_device
);
1642 static int __init
dma_debug_driver_setup(char *str
)
1646 for (i
= 0; i
< NAME_MAX_LEN
- 1; ++i
, ++str
) {
1647 current_driver_name
[i
] = *str
;
1652 if (current_driver_name
[0])
1653 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1654 current_driver_name
);
1659 __setup("dma_debug_driver=", dma_debug_driver_setup
);