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 inline void seq_print_ip_sym(struct seq_file *s, unsigned long ip)
869 seq_printf(s, "[<%p>] %pS\n", (void *)ip, (void *)ip);
872 void seq_print_trace(struct seq_file *s, struct stack_trace *trace)
876 if (WARN_ON(!trace->entries))
879 for (i = trace->skip; i < trace->nr_entries; i++)
880 seq_print_ip_sym(s, trace->entries[i]);
884 * Print all map entries just in the order they are stored. We assume that the
885 * user will be able to parse this later anyway. Detailed output includes stack
886 * traces of allocations.
888 void seq_print_dma_mappings(struct seq_file *s, int detail)
892 for (idx = 0; idx < HASH_SIZE; idx++) {
893 struct hash_bucket *bucket = &dma_entry_hash[idx];
894 struct dma_debug_entry *entry;
897 spin_lock_irqsave(&bucket->lock, flags);
899 list_for_each_entry(entry, &bucket->list, list) {
901 " %s %s idx %d P=%Lx D=%llx L=%llx %s A=%s\n",
902 dev_name(entry->dev),
903 type2name[entry->type], idx,
905 entry->dev_addr, entry->size,
906 dir2name[entry->direction],
907 debug_dma_platformdata(entry->dev));
910 seq_print_trace(s, &entry->stacktrace);
913 spin_unlock_irqrestore(&bucket->lock, flags);
917 void __weak dma_debugfs_platform_info(struct dentry *dent)
921 static int _dump_allocs(struct seq_file *s, void *data)
923 int detail = (int)s->private;
925 seq_print_dma_mappings(s, detail);
929 #define DEFINE_DEBUGFS(__name, __func, __data) \
930 static int __name ## _open(struct inode *inode, struct file *file) \
932 return single_open(file, __func, __data); \
934 static const struct file_operations __name ## _fops = { \
935 .open = __name ## _open, \
937 .llseek = seq_lseek, \
938 .release = single_release, \
941 DEFINE_DEBUGFS(_dump_allocs, _dump_allocs, NULL);
942 DEFINE_DEBUGFS(_dump_allocs_detail, _dump_allocs, (void *)1);
943 #undef DEFINE_DEBUGFS
945 static int map_dump_debug_fs_init(void)
947 #define CREATE_FILE(name) \
948 debugfs_create_file(#name, S_IRUGO, \
949 dma_debug_dent, NULL, \
952 if (!CREATE_FILE(dump_allocs))
955 if (!CREATE_FILE(dump_allocs_detail))
960 dma_debugfs_platform_info(dma_debug_dent);
964 static const struct file_operations filter_fops = {
966 .write = filter_write,
967 .llseek = default_llseek,
970 static int dma_debug_fs_init(void)
972 dma_debug_dent = debugfs_create_dir("dma-api", NULL);
973 if (!dma_debug_dent) {
974 pr_err("DMA-API: can not create debugfs directory\n");
978 global_disable_dent = debugfs_create_bool("disabled", 0444,
981 if (!global_disable_dent)
984 error_count_dent = debugfs_create_u32("error_count", 0444,
985 dma_debug_dent, &error_count);
986 if (!error_count_dent)
989 show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
992 if (!show_all_errors_dent)
995 show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
998 if (!show_num_errors_dent)
1001 num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
1004 if (!num_free_entries_dent)
1007 min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
1010 if (!min_free_entries_dent)
1013 filter_dent = debugfs_create_file("driver_filter", 0644,
1014 dma_debug_dent, NULL, &filter_fops);
1018 if (map_dump_debug_fs_init())
1024 debugfs_remove_recursive(dma_debug_dent);
1029 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
1031 struct dma_debug_entry *entry;
1032 unsigned long flags;
1035 local_irq_save(flags);
1037 for (i = 0; i < HASH_SIZE; ++i) {
1038 spin_lock(&dma_entry_hash[i].lock);
1039 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
1040 if (entry->dev == dev) {
1045 spin_unlock(&dma_entry_hash[i].lock);
1048 local_irq_restore(flags);
1053 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
1055 struct device *dev = data;
1056 struct dma_debug_entry *uninitialized_var(entry);
1059 if (dma_debug_disabled())
1063 case BUS_NOTIFY_UNBOUND_DRIVER:
1064 count = device_dma_allocations(dev, &entry);
1067 err_printk(dev, entry, "DMA-API: device driver has pending "
1068 "DMA allocations while released from device "
1070 "One of leaked entries details: "
1071 "[device address=0x%016llx] [size=%llu bytes] "
1072 "[mapped with %s] [mapped as %s]\n",
1073 count, entry->dev_addr, entry->size,
1074 dir2name[entry->direction], type2name[entry->type]);
1083 void dma_debug_add_bus(struct bus_type *bus)
1085 struct notifier_block *nb;
1087 if (dma_debug_disabled())
1090 nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
1092 pr_err("dma_debug_add_bus: out of memory\n");
1096 nb->notifier_call = dma_debug_device_change;
1098 bus_register_notifier(bus, nb);
1102 * Let the architectures decide how many entries should be preallocated.
1104 void dma_debug_init(u32 num_entries)
1108 /* Do not use dma_debug_initialized here, since we really want to be
1109 * called to set dma_debug_initialized
1114 for (i = 0; i < HASH_SIZE; ++i) {
1115 INIT_LIST_HEAD(&dma_entry_hash[i].list);
1116 spin_lock_init(&dma_entry_hash[i].lock);
1119 if (dma_debug_fs_init() != 0) {
1120 pr_err("DMA-API: error creating debugfs entries - disabling\n");
1121 global_disable = true;
1127 num_entries = req_entries;
1129 if (prealloc_memory(num_entries) != 0) {
1130 pr_err("DMA-API: debugging out of memory error - disabled\n");
1131 global_disable = true;
1136 nr_total_entries = num_free_entries;
1138 dma_debug_initialized = true;
1140 pr_info("DMA-API: debugging enabled by kernel config\n");
1143 static __init int dma_debug_cmdline(char *str)
1148 if (strncmp(str, "off", 3) == 0) {
1149 pr_info("DMA-API: debugging disabled on kernel command line\n");
1150 global_disable = true;
1156 static __init int dma_debug_entries_cmdline(char *str)
1163 res = get_option(&str, &req_entries);
1171 __setup("dma_debug=", dma_debug_cmdline);
1172 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
1174 static void check_unmap(struct dma_debug_entry *ref)
1176 struct dma_debug_entry *entry;
1177 struct hash_bucket *bucket;
1178 unsigned long flags;
1180 bucket = get_hash_bucket(ref, &flags);
1181 entry = bucket_find_exact(bucket, ref);
1184 /* must drop lock before calling dma_mapping_error */
1185 put_hash_bucket(bucket, &flags);
1187 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1188 err_printk(ref->dev, NULL,
1189 "DMA-API: device driver tries to free an "
1190 "invalid DMA memory address\n");
1192 err_printk(ref->dev, NULL,
1193 "DMA-API: device driver tries to free DMA "
1194 "memory it has not allocated [device "
1195 "address=0x%016llx] [size=%llu bytes]\n",
1196 ref->dev_addr, ref->size);
1201 if (ref->size != entry->size) {
1202 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1203 "DMA memory with different size "
1204 "[device address=0x%016llx] [map size=%llu bytes] "
1205 "[unmap size=%llu bytes]\n",
1206 ref->dev_addr, entry->size, ref->size);
1209 if (ref->type != entry->type) {
1210 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1211 "DMA memory with wrong function "
1212 "[device address=0x%016llx] [size=%llu bytes] "
1213 "[mapped as %s] [unmapped as %s]\n",
1214 ref->dev_addr, ref->size,
1215 type2name[entry->type], type2name[ref->type]);
1216 } else if ((entry->type == dma_debug_coherent) &&
1217 (phys_addr(ref) != phys_addr(entry))) {
1218 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1219 "DMA memory with different CPU address "
1220 "[device address=0x%016llx] [size=%llu bytes] "
1221 "[cpu alloc address=0x%016llx] "
1222 "[cpu free address=0x%016llx]",
1223 ref->dev_addr, ref->size,
1228 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1229 ref->sg_call_ents != entry->sg_call_ents) {
1230 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1231 "DMA sg list with different entry count "
1232 "[map count=%d] [unmap count=%d]\n",
1233 entry->sg_call_ents, ref->sg_call_ents);
1237 * This may be no bug in reality - but most implementations of the
1238 * DMA API don't handle this properly, so check for it here
1240 if (ref->direction != entry->direction) {
1241 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1242 "DMA memory with different direction "
1243 "[device address=0x%016llx] [size=%llu bytes] "
1244 "[mapped with %s] [unmapped with %s]\n",
1245 ref->dev_addr, ref->size,
1246 dir2name[entry->direction],
1247 dir2name[ref->direction]);
1250 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1251 err_printk(ref->dev, entry,
1252 "DMA-API: device driver failed to check map error"
1253 "[device address=0x%016llx] [size=%llu bytes] "
1255 ref->dev_addr, ref->size,
1256 type2name[entry->type]);
1259 hash_bucket_del(entry);
1260 dma_entry_free(entry);
1262 put_hash_bucket(bucket, &flags);
1265 static void check_for_stack(struct device *dev, void *addr)
1267 if (object_is_on_stack(addr))
1268 err_printk(dev, NULL, "DMA-API: device driver maps memory from "
1269 "stack [addr=%p]\n", addr);
1272 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1274 unsigned long a1 = (unsigned long)addr;
1275 unsigned long b1 = a1 + len;
1276 unsigned long a2 = (unsigned long)start;
1277 unsigned long b2 = (unsigned long)end;
1279 return !(b1 <= a2 || a1 >= b2);
1282 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1284 if (overlap(addr, len, _stext, _etext) ||
1285 overlap(addr, len, __start_rodata, __end_rodata))
1286 err_printk(dev, NULL, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1289 static void check_sync(struct device *dev,
1290 struct dma_debug_entry *ref,
1293 struct dma_debug_entry *entry;
1294 struct hash_bucket *bucket;
1295 unsigned long flags;
1297 bucket = get_hash_bucket(ref, &flags);
1299 entry = bucket_find_contain(&bucket, ref, &flags);
1302 err_printk(dev, NULL, "DMA-API: device driver tries "
1303 "to sync DMA memory it has not allocated "
1304 "[device address=0x%016llx] [size=%llu bytes]\n",
1305 (unsigned long long)ref->dev_addr, ref->size);
1309 if (ref->size > entry->size) {
1310 err_printk(dev, entry, "DMA-API: device driver syncs"
1311 " DMA memory outside allocated range "
1312 "[device address=0x%016llx] "
1313 "[allocation size=%llu bytes] "
1314 "[sync offset+size=%llu]\n",
1315 entry->dev_addr, entry->size,
1319 if (entry->direction == DMA_BIDIRECTIONAL)
1322 if (ref->direction != entry->direction) {
1323 err_printk(dev, entry, "DMA-API: device driver syncs "
1324 "DMA memory with different direction "
1325 "[device address=0x%016llx] [size=%llu bytes] "
1326 "[mapped with %s] [synced with %s]\n",
1327 (unsigned long long)ref->dev_addr, entry->size,
1328 dir2name[entry->direction],
1329 dir2name[ref->direction]);
1332 if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1333 !(ref->direction == DMA_TO_DEVICE))
1334 err_printk(dev, entry, "DMA-API: device driver syncs "
1335 "device read-only DMA memory for cpu "
1336 "[device address=0x%016llx] [size=%llu bytes] "
1337 "[mapped with %s] [synced with %s]\n",
1338 (unsigned long long)ref->dev_addr, entry->size,
1339 dir2name[entry->direction],
1340 dir2name[ref->direction]);
1342 if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1343 !(ref->direction == DMA_FROM_DEVICE))
1344 err_printk(dev, entry, "DMA-API: device driver syncs "
1345 "device write-only DMA memory to device "
1346 "[device address=0x%016llx] [size=%llu bytes] "
1347 "[mapped with %s] [synced with %s]\n",
1348 (unsigned long long)ref->dev_addr, entry->size,
1349 dir2name[entry->direction],
1350 dir2name[ref->direction]);
1352 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1353 ref->sg_call_ents != entry->sg_call_ents) {
1354 err_printk(ref->dev, entry, "DMA-API: device driver syncs "
1355 "DMA sg list with different entry count "
1356 "[map count=%d] [sync count=%d]\n",
1357 entry->sg_call_ents, ref->sg_call_ents);
1361 put_hash_bucket(bucket, &flags);
1364 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1365 size_t size, int direction, dma_addr_t dma_addr,
1368 struct dma_debug_entry *entry;
1370 if (unlikely(dma_debug_disabled()))
1373 if (dma_mapping_error(dev, dma_addr))
1376 entry = dma_entry_alloc();
1381 entry->type = dma_debug_page;
1382 entry->pfn = page_to_pfn(page);
1383 entry->offset = offset,
1384 entry->dev_addr = dma_addr;
1386 entry->direction = direction;
1387 entry->map_err_type = MAP_ERR_NOT_CHECKED;
1390 entry->type = dma_debug_single;
1392 if (!PageHighMem(page)) {
1393 void *addr = page_address(page) + offset;
1395 check_for_stack(dev, addr);
1396 check_for_illegal_area(dev, addr, size);
1399 add_dma_entry(entry);
1401 EXPORT_SYMBOL(debug_dma_map_page);
1403 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1405 struct dma_debug_entry ref;
1406 struct dma_debug_entry *entry;
1407 struct hash_bucket *bucket;
1408 unsigned long flags;
1410 if (unlikely(dma_debug_disabled()))
1414 ref.dev_addr = dma_addr;
1415 bucket = get_hash_bucket(&ref, &flags);
1417 list_for_each_entry(entry, &bucket->list, list) {
1418 if (!exact_match(&ref, entry))
1422 * The same physical address can be mapped multiple
1423 * times. Without a hardware IOMMU this results in the
1424 * same device addresses being put into the dma-debug
1425 * hash multiple times too. This can result in false
1426 * positives being reported. Therefore we implement a
1427 * best-fit algorithm here which updates the first entry
1428 * from the hash which fits the reference value and is
1429 * not currently listed as being checked.
1431 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1432 entry->map_err_type = MAP_ERR_CHECKED;
1437 put_hash_bucket(bucket, &flags);
1439 EXPORT_SYMBOL(debug_dma_mapping_error);
1441 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1442 size_t size, int direction, bool map_single)
1444 struct dma_debug_entry ref = {
1445 .type = dma_debug_page,
1449 .direction = direction,
1452 if (unlikely(dma_debug_disabled()))
1456 ref.type = dma_debug_single;
1460 EXPORT_SYMBOL(debug_dma_unmap_page);
1462 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1463 int nents, int mapped_ents, int direction)
1465 struct dma_debug_entry *entry;
1466 struct scatterlist *s;
1469 if (unlikely(dma_debug_disabled()))
1472 for_each_sg(sg, s, mapped_ents, i) {
1473 entry = dma_entry_alloc();
1477 entry->type = dma_debug_sg;
1479 entry->pfn = page_to_pfn(sg_page(s));
1480 entry->offset = s->offset,
1481 entry->size = sg_dma_len(s);
1482 entry->dev_addr = sg_dma_address(s);
1483 entry->direction = direction;
1484 entry->sg_call_ents = nents;
1485 entry->sg_mapped_ents = mapped_ents;
1487 if (!PageHighMem(sg_page(s))) {
1488 check_for_stack(dev, sg_virt(s));
1489 check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1492 add_dma_entry(entry);
1495 EXPORT_SYMBOL(debug_dma_map_sg);
1497 static int get_nr_mapped_entries(struct device *dev,
1498 struct dma_debug_entry *ref)
1500 struct dma_debug_entry *entry;
1501 struct hash_bucket *bucket;
1502 unsigned long flags;
1505 bucket = get_hash_bucket(ref, &flags);
1506 entry = bucket_find_exact(bucket, ref);
1510 mapped_ents = entry->sg_mapped_ents;
1511 put_hash_bucket(bucket, &flags);
1516 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1517 int nelems, int dir)
1519 struct scatterlist *s;
1520 int mapped_ents = 0, i;
1522 if (unlikely(dma_debug_disabled()))
1525 for_each_sg(sglist, s, nelems, i) {
1527 struct dma_debug_entry ref = {
1528 .type = dma_debug_sg,
1530 .pfn = page_to_pfn(sg_page(s)),
1531 .offset = s->offset,
1532 .dev_addr = sg_dma_address(s),
1533 .size = sg_dma_len(s),
1535 .sg_call_ents = nelems,
1538 if (mapped_ents && i >= mapped_ents)
1542 mapped_ents = get_nr_mapped_entries(dev, &ref);
1547 EXPORT_SYMBOL(debug_dma_unmap_sg);
1549 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1550 dma_addr_t dma_addr, void *virt)
1552 struct dma_debug_entry *entry;
1554 if (unlikely(dma_debug_disabled()))
1557 if (unlikely(virt == NULL))
1560 entry = dma_entry_alloc();
1564 entry->type = dma_debug_coherent;
1566 entry->pfn = page_to_pfn(virt_to_page(virt));
1567 entry->offset = (size_t) virt & ~PAGE_MASK;
1569 entry->dev_addr = dma_addr;
1570 entry->direction = DMA_BIDIRECTIONAL;
1572 add_dma_entry(entry);
1574 EXPORT_SYMBOL(debug_dma_alloc_coherent);
1576 void debug_dma_free_coherent(struct device *dev, size_t size,
1577 void *virt, dma_addr_t addr)
1579 struct dma_debug_entry ref = {
1580 .type = dma_debug_coherent,
1582 .pfn = page_to_pfn(virt_to_page(virt)),
1583 .offset = (size_t) virt & ~PAGE_MASK,
1586 .direction = DMA_BIDIRECTIONAL,
1589 if (unlikely(dma_debug_disabled()))
1594 EXPORT_SYMBOL(debug_dma_free_coherent);
1596 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1597 size_t size, int direction)
1599 struct dma_debug_entry ref;
1601 if (unlikely(dma_debug_disabled()))
1604 ref.type = dma_debug_single;
1606 ref.dev_addr = dma_handle;
1608 ref.direction = direction;
1609 ref.sg_call_ents = 0;
1611 check_sync(dev, &ref, true);
1613 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1615 void debug_dma_sync_single_for_device(struct device *dev,
1616 dma_addr_t dma_handle, size_t size,
1619 struct dma_debug_entry ref;
1621 if (unlikely(dma_debug_disabled()))
1624 ref.type = dma_debug_single;
1626 ref.dev_addr = dma_handle;
1628 ref.direction = direction;
1629 ref.sg_call_ents = 0;
1631 check_sync(dev, &ref, false);
1633 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1635 void debug_dma_sync_single_range_for_cpu(struct device *dev,
1636 dma_addr_t dma_handle,
1637 unsigned long offset, size_t size,
1640 struct dma_debug_entry ref;
1642 if (unlikely(dma_debug_disabled()))
1645 ref.type = dma_debug_single;
1647 ref.dev_addr = dma_handle;
1648 ref.size = offset + size;
1649 ref.direction = direction;
1650 ref.sg_call_ents = 0;
1652 check_sync(dev, &ref, true);
1654 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
1656 void debug_dma_sync_single_range_for_device(struct device *dev,
1657 dma_addr_t dma_handle,
1658 unsigned long offset,
1659 size_t size, int direction)
1661 struct dma_debug_entry ref;
1663 if (unlikely(dma_debug_disabled()))
1666 ref.type = dma_debug_single;
1668 ref.dev_addr = dma_handle;
1669 ref.size = offset + size;
1670 ref.direction = direction;
1671 ref.sg_call_ents = 0;
1673 check_sync(dev, &ref, false);
1675 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
1677 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1678 int nelems, int direction)
1680 struct scatterlist *s;
1681 int mapped_ents = 0, i;
1683 if (unlikely(dma_debug_disabled()))
1686 for_each_sg(sg, s, nelems, i) {
1688 struct dma_debug_entry ref = {
1689 .type = dma_debug_sg,
1691 .pfn = page_to_pfn(sg_page(s)),
1692 .offset = s->offset,
1693 .dev_addr = sg_dma_address(s),
1694 .size = sg_dma_len(s),
1695 .direction = direction,
1696 .sg_call_ents = nelems,
1700 mapped_ents = get_nr_mapped_entries(dev, &ref);
1702 if (i >= mapped_ents)
1705 check_sync(dev, &ref, true);
1708 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1710 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1711 int nelems, int direction)
1713 struct scatterlist *s;
1714 int mapped_ents = 0, i;
1716 if (unlikely(dma_debug_disabled()))
1719 for_each_sg(sg, s, nelems, i) {
1721 struct dma_debug_entry ref = {
1722 .type = dma_debug_sg,
1724 .pfn = page_to_pfn(sg_page(s)),
1725 .offset = s->offset,
1726 .dev_addr = sg_dma_address(s),
1727 .size = sg_dma_len(s),
1728 .direction = direction,
1729 .sg_call_ents = nelems,
1732 mapped_ents = get_nr_mapped_entries(dev, &ref);
1734 if (i >= mapped_ents)
1737 check_sync(dev, &ref, false);
1740 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1742 static int __init dma_debug_driver_setup(char *str)
1746 for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1747 current_driver_name[i] = *str;
1752 if (current_driver_name[0])
1753 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1754 current_driver_name);
1759 __setup("dma_debug_driver=", dma_debug_driver_setup);