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 #define pr_fmt(fmt) "DMA-API: " fmt
22 #include <linux/sched/task_stack.h>
23 #include <linux/scatterlist.h>
24 #include <linux/dma-mapping.h>
25 #include <linux/sched/task.h>
26 #include <linux/stacktrace.h>
27 #include <linux/dma-debug.h>
28 #include <linux/spinlock.h>
29 #include <linux/vmalloc.h>
30 #include <linux/debugfs.h>
31 #include <linux/uaccess.h>
32 #include <linux/export.h>
33 #include <linux/device.h>
34 #include <linux/types.h>
35 #include <linux/sched.h>
36 #include <linux/ctype.h>
37 #include <linux/list.h>
38 #include <linux/slab.h>
40 #include <asm/sections.h>
42 #define HASH_SIZE 1024ULL
43 #define HASH_FN_SHIFT 13
44 #define HASH_FN_MASK (HASH_SIZE - 1)
46 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
47 /* If the pool runs out, add this many new entries at once */
48 #define DMA_DEBUG_DYNAMIC_ENTRIES (PAGE_SIZE / sizeof(struct dma_debug_entry))
58 MAP_ERR_CHECK_NOT_APPLICABLE,
63 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
66 * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
67 * @list: node on pre-allocated free_entries list
68 * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
69 * @type: single, page, sg, coherent
70 * @pfn: page frame of the start address
71 * @offset: offset of mapping relative to pfn
72 * @size: length of the mapping
73 * @direction: enum dma_data_direction
74 * @sg_call_ents: 'nents' from dma_map_sg
75 * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
76 * @map_err_type: track whether dma_mapping_error() was checked
77 * @stacktrace: support backtraces when a violation is detected
79 struct dma_debug_entry {
80 struct list_head list;
90 enum map_err_types map_err_type;
91 #ifdef CONFIG_STACKTRACE
92 struct stack_trace stacktrace;
93 unsigned long st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
97 typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
100 struct list_head list;
102 } ____cacheline_aligned_in_smp;
104 /* Hash list to save the allocated dma addresses */
105 static struct hash_bucket dma_entry_hash[HASH_SIZE];
106 /* List of pre-allocated dma_debug_entry's */
107 static LIST_HEAD(free_entries);
108 /* Lock for the list above */
109 static DEFINE_SPINLOCK(free_entries_lock);
111 /* Global disable flag - will be set in case of an error */
112 static bool global_disable __read_mostly;
114 /* Early initialization disable flag, set at the end of dma_debug_init */
115 static bool dma_debug_initialized __read_mostly;
117 static inline bool dma_debug_disabled(void)
119 return global_disable || !dma_debug_initialized;
122 /* Global error count */
123 static u32 error_count;
125 /* Global error show enable*/
126 static u32 show_all_errors __read_mostly;
127 /* Number of errors to show */
128 static u32 show_num_errors = 1;
130 static u32 num_free_entries;
131 static u32 min_free_entries;
132 static u32 nr_total_entries;
134 /* number of preallocated entries requested by kernel cmdline */
135 static u32 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
137 /* per-driver filter related state */
139 #define NAME_MAX_LEN 64
141 static char current_driver_name[NAME_MAX_LEN] __read_mostly;
142 static struct device_driver *current_driver __read_mostly;
144 static DEFINE_RWLOCK(driver_name_lock);
146 static const char *const maperr2str[] = {
147 [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
148 [MAP_ERR_NOT_CHECKED] = "dma map error not checked",
149 [MAP_ERR_CHECKED] = "dma map error checked",
152 static const char *type2name[5] = { "single", "page",
153 "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, pr_fmt("%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)
256 __acquires(&dma_entry_hash[idx].lock)
258 int idx = hash_fn(entry);
259 unsigned long __flags;
261 spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
263 return &dma_entry_hash[idx];
267 * Give up exclusive access to the hash bucket
269 static void put_hash_bucket(struct hash_bucket *bucket,
270 unsigned long *flags)
271 __releases(&bucket->lock)
273 unsigned long __flags = *flags;
275 spin_unlock_irqrestore(&bucket->lock, __flags);
278 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
280 return ((a->dev_addr == b->dev_addr) &&
281 (a->dev == b->dev)) ? true : false;
284 static bool containing_match(struct dma_debug_entry *a,
285 struct dma_debug_entry *b)
287 if (a->dev != b->dev)
290 if ((b->dev_addr <= a->dev_addr) &&
291 ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
298 * Search a given entry in the hash bucket list
300 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
301 struct dma_debug_entry *ref,
304 struct dma_debug_entry *entry, *ret = NULL;
305 int matches = 0, match_lvl, last_lvl = -1;
307 list_for_each_entry(entry, &bucket->list, list) {
308 if (!match(ref, entry))
312 * Some drivers map the same physical address multiple
313 * times. Without a hardware IOMMU this results in the
314 * same device addresses being put into the dma-debug
315 * hash multiple times too. This can result in false
316 * positives being reported. Therefore we implement a
317 * best-fit algorithm here which returns the entry from
318 * the hash which fits best to the reference value
319 * instead of the first-fit.
323 entry->size == ref->size ? ++match_lvl : 0;
324 entry->type == ref->type ? ++match_lvl : 0;
325 entry->direction == ref->direction ? ++match_lvl : 0;
326 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
328 if (match_lvl == 4) {
329 /* perfect-fit - return the result */
331 } else if (match_lvl > last_lvl) {
333 * We found an entry that fits better then the
334 * previous one or it is the 1st match.
336 last_lvl = match_lvl;
342 * If we have multiple matches but no perfect-fit, just return
345 ret = (matches == 1) ? ret : NULL;
350 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
351 struct dma_debug_entry *ref)
353 return __hash_bucket_find(bucket, ref, exact_match);
356 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
357 struct dma_debug_entry *ref,
358 unsigned long *flags)
361 unsigned int max_range = dma_get_max_seg_size(ref->dev);
362 struct dma_debug_entry *entry, index = *ref;
363 unsigned int range = 0;
365 while (range <= max_range) {
366 entry = __hash_bucket_find(*bucket, ref, containing_match);
372 * Nothing found, go back a hash bucket
374 put_hash_bucket(*bucket, flags);
375 range += (1 << HASH_FN_SHIFT);
376 index.dev_addr -= (1 << HASH_FN_SHIFT);
377 *bucket = get_hash_bucket(&index, flags);
384 * Add an entry to a hash bucket
386 static void hash_bucket_add(struct hash_bucket *bucket,
387 struct dma_debug_entry *entry)
389 list_add_tail(&entry->list, &bucket->list);
393 * Remove entry from a hash bucket list
395 static void hash_bucket_del(struct dma_debug_entry *entry)
397 list_del(&entry->list);
400 static unsigned long long phys_addr(struct dma_debug_entry *entry)
402 if (entry->type == dma_debug_resource)
403 return __pfn_to_phys(entry->pfn) + entry->offset;
405 return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
409 * Dump mapping entries for debugging purposes
411 void debug_dma_dump_mappings(struct device *dev)
415 for (idx = 0; idx < HASH_SIZE; idx++) {
416 struct hash_bucket *bucket = &dma_entry_hash[idx];
417 struct dma_debug_entry *entry;
420 spin_lock_irqsave(&bucket->lock, flags);
422 list_for_each_entry(entry, &bucket->list, list) {
423 if (!dev || dev == entry->dev) {
425 "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
426 type2name[entry->type], idx,
427 phys_addr(entry), entry->pfn,
428 entry->dev_addr, entry->size,
429 dir2name[entry->direction],
430 maperr2str[entry->map_err_type]);
434 spin_unlock_irqrestore(&bucket->lock, flags);
439 * For each mapping (initial cacheline in the case of
440 * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
441 * scatterlist, or the cacheline specified in dma_map_single) insert
442 * into this tree using the cacheline as the key. At
443 * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry. If
444 * the entry already exists at insertion time add a tag as a reference
445 * count for the overlapping mappings. For now, the overlap tracking
446 * just ensures that 'unmaps' balance 'maps' before marking the
447 * cacheline idle, but we should also be flagging overlaps as an API
450 * Memory usage is mostly constrained by the maximum number of available
451 * dma-debug entries in that we need a free dma_debug_entry before
452 * inserting into the tree. In the case of dma_map_page and
453 * dma_alloc_coherent there is only one dma_debug_entry and one
454 * dma_active_cacheline entry to track per event. dma_map_sg(), on the
455 * other hand, consumes a single dma_debug_entry, but inserts 'nents'
456 * entries into the tree.
458 * At any time debug_dma_assert_idle() can be called to trigger a
459 * warning if any cachelines in the given page are in the active set.
461 static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
462 static DEFINE_SPINLOCK(radix_lock);
463 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
464 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
465 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
467 static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
469 return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
470 (entry->offset >> L1_CACHE_SHIFT);
473 static int active_cacheline_read_overlap(phys_addr_t cln)
477 for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
478 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
483 static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
487 if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
490 for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
491 if (overlap & 1 << i)
492 radix_tree_tag_set(&dma_active_cacheline, cln, i);
494 radix_tree_tag_clear(&dma_active_cacheline, cln, i);
499 static void active_cacheline_inc_overlap(phys_addr_t cln)
501 int overlap = active_cacheline_read_overlap(cln);
503 overlap = active_cacheline_set_overlap(cln, ++overlap);
505 /* If we overflowed the overlap counter then we're potentially
506 * leaking dma-mappings. Otherwise, if maps and unmaps are
507 * balanced then this overflow may cause false negatives in
508 * debug_dma_assert_idle() as the cacheline may be marked idle
511 WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
512 pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
513 ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
516 static int active_cacheline_dec_overlap(phys_addr_t cln)
518 int overlap = active_cacheline_read_overlap(cln);
520 return active_cacheline_set_overlap(cln, --overlap);
523 static int active_cacheline_insert(struct dma_debug_entry *entry)
525 phys_addr_t cln = to_cacheline_number(entry);
529 /* If the device is not writing memory then we don't have any
530 * concerns about the cpu consuming stale data. This mitigates
531 * legitimate usages of overlapping mappings.
533 if (entry->direction == DMA_TO_DEVICE)
536 spin_lock_irqsave(&radix_lock, flags);
537 rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
539 active_cacheline_inc_overlap(cln);
540 spin_unlock_irqrestore(&radix_lock, flags);
545 static void active_cacheline_remove(struct dma_debug_entry *entry)
547 phys_addr_t cln = to_cacheline_number(entry);
550 /* ...mirror the insert case */
551 if (entry->direction == DMA_TO_DEVICE)
554 spin_lock_irqsave(&radix_lock, flags);
555 /* since we are counting overlaps the final put of the
556 * cacheline will occur when the overlap count is 0.
557 * active_cacheline_dec_overlap() returns -1 in that case
559 if (active_cacheline_dec_overlap(cln) < 0)
560 radix_tree_delete(&dma_active_cacheline, cln);
561 spin_unlock_irqrestore(&radix_lock, flags);
565 * debug_dma_assert_idle() - assert that a page is not undergoing dma
566 * @page: page to lookup in the dma_active_cacheline tree
568 * Place a call to this routine in cases where the cpu touching the page
569 * before the dma completes (page is dma_unmapped) will lead to data
572 void debug_dma_assert_idle(struct page *page)
574 static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
575 struct dma_debug_entry *entry = NULL;
576 void **results = (void **) &ents;
577 unsigned int nents, i;
581 if (dma_debug_disabled())
587 cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
588 spin_lock_irqsave(&radix_lock, flags);
589 nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
590 CACHELINES_PER_PAGE);
591 for (i = 0; i < nents; i++) {
592 phys_addr_t ent_cln = to_cacheline_number(ents[i]);
594 if (ent_cln == cln) {
597 } else if (ent_cln >= cln + CACHELINES_PER_PAGE)
600 spin_unlock_irqrestore(&radix_lock, flags);
605 cln = to_cacheline_number(entry);
606 err_printk(entry->dev, entry,
607 "cpu touching an active dma mapped cacheline [cln=%pa]\n",
612 * Wrapper function for adding an entry to the hash.
613 * This function takes care of locking itself.
615 static void add_dma_entry(struct dma_debug_entry *entry)
617 struct hash_bucket *bucket;
621 bucket = get_hash_bucket(entry, &flags);
622 hash_bucket_add(bucket, entry);
623 put_hash_bucket(bucket, &flags);
625 rc = active_cacheline_insert(entry);
627 pr_err("cacheline tracking ENOMEM, dma-debug disabled\n");
628 global_disable = true;
631 /* TODO: report -EEXIST errors here as overlapping mappings are
632 * not supported by the DMA API
636 static int dma_debug_create_entries(gfp_t gfp)
638 struct dma_debug_entry *entry;
641 entry = (void *)get_zeroed_page(gfp);
645 for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
646 list_add_tail(&entry[i].list, &free_entries);
648 num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
649 nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
654 static struct dma_debug_entry *__dma_entry_alloc(void)
656 struct dma_debug_entry *entry;
658 entry = list_entry(free_entries.next, struct dma_debug_entry, list);
659 list_del(&entry->list);
660 memset(entry, 0, sizeof(*entry));
662 num_free_entries -= 1;
663 if (num_free_entries < min_free_entries)
664 min_free_entries = num_free_entries;
669 void __dma_entry_alloc_check_leak(void)
671 u32 tmp = nr_total_entries % nr_prealloc_entries;
673 /* Shout each time we tick over some multiple of the initial pool */
674 if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
675 pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
677 (nr_total_entries / nr_prealloc_entries));
681 /* struct dma_entry allocator
683 * The next two functions implement the allocator for
684 * struct dma_debug_entries.
686 static struct dma_debug_entry *dma_entry_alloc(void)
688 struct dma_debug_entry *entry;
691 spin_lock_irqsave(&free_entries_lock, flags);
692 if (num_free_entries == 0) {
693 if (dma_debug_create_entries(GFP_ATOMIC)) {
694 global_disable = true;
695 spin_unlock_irqrestore(&free_entries_lock, flags);
696 pr_err("debugging out of memory - disabling\n");
699 __dma_entry_alloc_check_leak();
702 entry = __dma_entry_alloc();
704 spin_unlock_irqrestore(&free_entries_lock, flags);
706 #ifdef CONFIG_STACKTRACE
707 entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
708 entry->stacktrace.entries = entry->st_entries;
709 entry->stacktrace.skip = 2;
710 save_stack_trace(&entry->stacktrace);
716 static void dma_entry_free(struct dma_debug_entry *entry)
720 active_cacheline_remove(entry);
723 * add to beginning of the list - this way the entries are
724 * more likely cache hot when they are reallocated.
726 spin_lock_irqsave(&free_entries_lock, flags);
727 list_add(&entry->list, &free_entries);
728 num_free_entries += 1;
729 spin_unlock_irqrestore(&free_entries_lock, flags);
733 * DMA-API debugging init code
735 * The init code does two things:
736 * 1. Initialize core data structures
737 * 2. Preallocate a given number of dma_debug_entry structs
740 static ssize_t filter_read(struct file *file, char __user *user_buf,
741 size_t count, loff_t *ppos)
743 char buf[NAME_MAX_LEN + 1];
747 if (!current_driver_name[0])
751 * We can't copy to userspace directly because current_driver_name can
752 * only be read under the driver_name_lock with irqs disabled. So
753 * create a temporary copy first.
755 read_lock_irqsave(&driver_name_lock, flags);
756 len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
757 read_unlock_irqrestore(&driver_name_lock, flags);
759 return simple_read_from_buffer(user_buf, count, ppos, buf, len);
762 static ssize_t filter_write(struct file *file, const char __user *userbuf,
763 size_t count, loff_t *ppos)
765 char buf[NAME_MAX_LEN];
771 * We can't copy from userspace directly. Access to
772 * current_driver_name is protected with a write_lock with irqs
773 * disabled. Since copy_from_user can fault and may sleep we
774 * need to copy to temporary buffer first
776 len = min(count, (size_t)(NAME_MAX_LEN - 1));
777 if (copy_from_user(buf, userbuf, len))
782 write_lock_irqsave(&driver_name_lock, flags);
785 * Now handle the string we got from userspace very carefully.
787 * - only use the first token we got
788 * - token delimiter is everything looking like a space
789 * character (' ', '\n', '\t' ...)
792 if (!isalnum(buf[0])) {
794 * If the first character userspace gave us is not
795 * alphanumerical then assume the filter should be
798 if (current_driver_name[0])
799 pr_info("switching off dma-debug driver filter\n");
800 current_driver_name[0] = 0;
801 current_driver = NULL;
806 * Now parse out the first token and use it as the name for the
807 * driver to filter for.
809 for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
810 current_driver_name[i] = buf[i];
811 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
814 current_driver_name[i] = 0;
815 current_driver = NULL;
817 pr_info("enable driver filter for driver [%s]\n",
818 current_driver_name);
821 write_unlock_irqrestore(&driver_name_lock, flags);
826 static const struct file_operations filter_fops = {
828 .write = filter_write,
829 .llseek = default_llseek,
832 static int dump_show(struct seq_file *seq, void *v)
836 for (idx = 0; idx < HASH_SIZE; idx++) {
837 struct hash_bucket *bucket = &dma_entry_hash[idx];
838 struct dma_debug_entry *entry;
841 spin_lock_irqsave(&bucket->lock, flags);
842 list_for_each_entry(entry, &bucket->list, list) {
844 "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx %s %s\n",
845 dev_name(entry->dev),
846 dev_driver_string(entry->dev),
847 type2name[entry->type], idx,
848 phys_addr(entry), entry->pfn,
849 entry->dev_addr, entry->size,
850 dir2name[entry->direction],
851 maperr2str[entry->map_err_type]);
853 spin_unlock_irqrestore(&bucket->lock, flags);
857 DEFINE_SHOW_ATTRIBUTE(dump);
859 static void dma_debug_fs_init(void)
861 struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
863 debugfs_create_bool("disabled", 0444, dentry, &global_disable);
864 debugfs_create_u32("error_count", 0444, dentry, &error_count);
865 debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
866 debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
867 debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
868 debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
869 debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
870 debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
871 debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
874 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
876 struct dma_debug_entry *entry;
880 for (i = 0; i < HASH_SIZE; ++i) {
881 spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
882 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
883 if (entry->dev == dev) {
888 spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
894 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
896 struct device *dev = data;
897 struct dma_debug_entry *uninitialized_var(entry);
900 if (dma_debug_disabled())
904 case BUS_NOTIFY_UNBOUND_DRIVER:
905 count = device_dma_allocations(dev, &entry);
908 err_printk(dev, entry, "device driver has pending "
909 "DMA allocations while released from device "
911 "One of leaked entries details: "
912 "[device address=0x%016llx] [size=%llu bytes] "
913 "[mapped with %s] [mapped as %s]\n",
914 count, entry->dev_addr, entry->size,
915 dir2name[entry->direction], type2name[entry->type]);
924 void dma_debug_add_bus(struct bus_type *bus)
926 struct notifier_block *nb;
928 if (dma_debug_disabled())
931 nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
933 pr_err("dma_debug_add_bus: out of memory\n");
937 nb->notifier_call = dma_debug_device_change;
939 bus_register_notifier(bus, nb);
942 static int dma_debug_init(void)
946 /* Do not use dma_debug_initialized here, since we really want to be
947 * called to set dma_debug_initialized
952 for (i = 0; i < HASH_SIZE; ++i) {
953 INIT_LIST_HEAD(&dma_entry_hash[i].list);
954 spin_lock_init(&dma_entry_hash[i].lock);
959 nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
960 for (i = 0; i < nr_pages; ++i)
961 dma_debug_create_entries(GFP_KERNEL);
962 if (num_free_entries >= nr_prealloc_entries) {
963 pr_info("preallocated %d debug entries\n", nr_total_entries);
964 } else if (num_free_entries > 0) {
965 pr_warn("%d debug entries requested but only %d allocated\n",
966 nr_prealloc_entries, nr_total_entries);
968 pr_err("debugging out of memory error - disabled\n");
969 global_disable = true;
973 min_free_entries = num_free_entries;
975 dma_debug_initialized = true;
977 pr_info("debugging enabled by kernel config\n");
980 core_initcall(dma_debug_init);
982 static __init int dma_debug_cmdline(char *str)
987 if (strncmp(str, "off", 3) == 0) {
988 pr_info("debugging disabled on kernel command line\n");
989 global_disable = true;
995 static __init int dma_debug_entries_cmdline(char *str)
999 if (!get_option(&str, &nr_prealloc_entries))
1000 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
1004 __setup("dma_debug=", dma_debug_cmdline);
1005 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
1007 static void check_unmap(struct dma_debug_entry *ref)
1009 struct dma_debug_entry *entry;
1010 struct hash_bucket *bucket;
1011 unsigned long flags;
1013 bucket = get_hash_bucket(ref, &flags);
1014 entry = bucket_find_exact(bucket, ref);
1017 /* must drop lock before calling dma_mapping_error */
1018 put_hash_bucket(bucket, &flags);
1020 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1021 err_printk(ref->dev, NULL,
1022 "device driver tries to free an "
1023 "invalid DMA memory address\n");
1025 err_printk(ref->dev, NULL,
1026 "device driver tries to free DMA "
1027 "memory it has not allocated [device "
1028 "address=0x%016llx] [size=%llu bytes]\n",
1029 ref->dev_addr, ref->size);
1034 if (ref->size != entry->size) {
1035 err_printk(ref->dev, entry, "device driver frees "
1036 "DMA memory with different size "
1037 "[device address=0x%016llx] [map size=%llu bytes] "
1038 "[unmap size=%llu bytes]\n",
1039 ref->dev_addr, entry->size, ref->size);
1042 if (ref->type != entry->type) {
1043 err_printk(ref->dev, entry, "device driver frees "
1044 "DMA memory with wrong function "
1045 "[device address=0x%016llx] [size=%llu bytes] "
1046 "[mapped as %s] [unmapped as %s]\n",
1047 ref->dev_addr, ref->size,
1048 type2name[entry->type], type2name[ref->type]);
1049 } else if ((entry->type == dma_debug_coherent) &&
1050 (phys_addr(ref) != phys_addr(entry))) {
1051 err_printk(ref->dev, entry, "device driver frees "
1052 "DMA memory with different CPU address "
1053 "[device address=0x%016llx] [size=%llu bytes] "
1054 "[cpu alloc address=0x%016llx] "
1055 "[cpu free address=0x%016llx]",
1056 ref->dev_addr, ref->size,
1061 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1062 ref->sg_call_ents != entry->sg_call_ents) {
1063 err_printk(ref->dev, entry, "device driver frees "
1064 "DMA sg list with different entry count "
1065 "[map count=%d] [unmap count=%d]\n",
1066 entry->sg_call_ents, ref->sg_call_ents);
1070 * This may be no bug in reality - but most implementations of the
1071 * DMA API don't handle this properly, so check for it here
1073 if (ref->direction != entry->direction) {
1074 err_printk(ref->dev, entry, "device driver frees "
1075 "DMA memory with different direction "
1076 "[device address=0x%016llx] [size=%llu bytes] "
1077 "[mapped with %s] [unmapped with %s]\n",
1078 ref->dev_addr, ref->size,
1079 dir2name[entry->direction],
1080 dir2name[ref->direction]);
1084 * Drivers should use dma_mapping_error() to check the returned
1085 * addresses of dma_map_single() and dma_map_page().
1086 * If not, print this warning message. See Documentation/DMA-API.txt.
1088 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1089 err_printk(ref->dev, entry,
1090 "device driver failed to check map error"
1091 "[device address=0x%016llx] [size=%llu bytes] "
1093 ref->dev_addr, ref->size,
1094 type2name[entry->type]);
1097 hash_bucket_del(entry);
1098 dma_entry_free(entry);
1100 put_hash_bucket(bucket, &flags);
1103 static void check_for_stack(struct device *dev,
1104 struct page *page, size_t offset)
1107 struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1109 if (!stack_vm_area) {
1110 /* Stack is direct-mapped. */
1111 if (PageHighMem(page))
1113 addr = page_address(page) + offset;
1114 if (object_is_on_stack(addr))
1115 err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1117 /* Stack is vmalloced. */
1120 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1121 if (page != stack_vm_area->pages[i])
1124 addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1125 err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1131 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1133 unsigned long a1 = (unsigned long)addr;
1134 unsigned long b1 = a1 + len;
1135 unsigned long a2 = (unsigned long)start;
1136 unsigned long b2 = (unsigned long)end;
1138 return !(b1 <= a2 || a1 >= b2);
1141 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1143 if (overlap(addr, len, _stext, _etext) ||
1144 overlap(addr, len, __start_rodata, __end_rodata))
1145 err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1148 static void check_sync(struct device *dev,
1149 struct dma_debug_entry *ref,
1152 struct dma_debug_entry *entry;
1153 struct hash_bucket *bucket;
1154 unsigned long flags;
1156 bucket = get_hash_bucket(ref, &flags);
1158 entry = bucket_find_contain(&bucket, ref, &flags);
1161 err_printk(dev, NULL, "device driver tries "
1162 "to sync DMA memory it has not allocated "
1163 "[device address=0x%016llx] [size=%llu bytes]\n",
1164 (unsigned long long)ref->dev_addr, ref->size);
1168 if (ref->size > entry->size) {
1169 err_printk(dev, entry, "device driver syncs"
1170 " DMA memory outside allocated range "
1171 "[device address=0x%016llx] "
1172 "[allocation size=%llu bytes] "
1173 "[sync offset+size=%llu]\n",
1174 entry->dev_addr, entry->size,
1178 if (entry->direction == DMA_BIDIRECTIONAL)
1181 if (ref->direction != entry->direction) {
1182 err_printk(dev, entry, "device driver syncs "
1183 "DMA memory with different direction "
1184 "[device address=0x%016llx] [size=%llu bytes] "
1185 "[mapped with %s] [synced with %s]\n",
1186 (unsigned long long)ref->dev_addr, entry->size,
1187 dir2name[entry->direction],
1188 dir2name[ref->direction]);
1191 if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1192 !(ref->direction == DMA_TO_DEVICE))
1193 err_printk(dev, entry, "device driver syncs "
1194 "device read-only DMA memory for cpu "
1195 "[device address=0x%016llx] [size=%llu bytes] "
1196 "[mapped with %s] [synced with %s]\n",
1197 (unsigned long long)ref->dev_addr, entry->size,
1198 dir2name[entry->direction],
1199 dir2name[ref->direction]);
1201 if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1202 !(ref->direction == DMA_FROM_DEVICE))
1203 err_printk(dev, entry, "device driver syncs "
1204 "device write-only DMA memory to device "
1205 "[device address=0x%016llx] [size=%llu bytes] "
1206 "[mapped with %s] [synced with %s]\n",
1207 (unsigned long long)ref->dev_addr, entry->size,
1208 dir2name[entry->direction],
1209 dir2name[ref->direction]);
1211 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1212 ref->sg_call_ents != entry->sg_call_ents) {
1213 err_printk(ref->dev, entry, "device driver syncs "
1214 "DMA sg list with different entry count "
1215 "[map count=%d] [sync count=%d]\n",
1216 entry->sg_call_ents, ref->sg_call_ents);
1220 put_hash_bucket(bucket, &flags);
1223 static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1225 #ifdef CONFIG_DMA_API_DEBUG_SG
1226 unsigned int max_seg = dma_get_max_seg_size(dev);
1227 u64 start, end, boundary = dma_get_seg_boundary(dev);
1230 * Either the driver forgot to set dma_parms appropriately, or
1231 * whoever generated the list forgot to check them.
1233 if (sg->length > max_seg)
1234 err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1235 sg->length, max_seg);
1237 * In some cases this could potentially be the DMA API
1238 * implementation's fault, but it would usually imply that
1239 * the scatterlist was built inappropriately to begin with.
1241 start = sg_dma_address(sg);
1242 end = start + sg_dma_len(sg) - 1;
1243 if ((start ^ end) & ~boundary)
1244 err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1245 start, end, boundary);
1249 void debug_dma_map_single(struct device *dev, const void *addr,
1252 if (unlikely(dma_debug_disabled()))
1255 if (!virt_addr_valid(addr))
1256 err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1259 if (is_vmalloc_addr(addr))
1260 err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1263 EXPORT_SYMBOL(debug_dma_map_single);
1265 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1266 size_t size, int direction, dma_addr_t dma_addr)
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_single;
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;
1289 check_for_stack(dev, page, offset);
1291 if (!PageHighMem(page)) {
1292 void *addr = page_address(page) + offset;
1294 check_for_illegal_area(dev, addr, size);
1297 add_dma_entry(entry);
1299 EXPORT_SYMBOL(debug_dma_map_page);
1301 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1303 struct dma_debug_entry ref;
1304 struct dma_debug_entry *entry;
1305 struct hash_bucket *bucket;
1306 unsigned long flags;
1308 if (unlikely(dma_debug_disabled()))
1312 ref.dev_addr = dma_addr;
1313 bucket = get_hash_bucket(&ref, &flags);
1315 list_for_each_entry(entry, &bucket->list, list) {
1316 if (!exact_match(&ref, entry))
1320 * The same physical address can be mapped multiple
1321 * times. Without a hardware IOMMU this results in the
1322 * same device addresses being put into the dma-debug
1323 * hash multiple times too. This can result in false
1324 * positives being reported. Therefore we implement a
1325 * best-fit algorithm here which updates the first entry
1326 * from the hash which fits the reference value and is
1327 * not currently listed as being checked.
1329 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1330 entry->map_err_type = MAP_ERR_CHECKED;
1335 put_hash_bucket(bucket, &flags);
1337 EXPORT_SYMBOL(debug_dma_mapping_error);
1339 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1340 size_t size, int direction)
1342 struct dma_debug_entry ref = {
1343 .type = dma_debug_single,
1347 .direction = direction,
1350 if (unlikely(dma_debug_disabled()))
1354 EXPORT_SYMBOL(debug_dma_unmap_page);
1356 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1357 int nents, int mapped_ents, int direction)
1359 struct dma_debug_entry *entry;
1360 struct scatterlist *s;
1363 if (unlikely(dma_debug_disabled()))
1366 for_each_sg(sg, s, mapped_ents, i) {
1367 entry = dma_entry_alloc();
1371 entry->type = dma_debug_sg;
1373 entry->pfn = page_to_pfn(sg_page(s));
1374 entry->offset = s->offset,
1375 entry->size = sg_dma_len(s);
1376 entry->dev_addr = sg_dma_address(s);
1377 entry->direction = direction;
1378 entry->sg_call_ents = nents;
1379 entry->sg_mapped_ents = mapped_ents;
1381 check_for_stack(dev, sg_page(s), s->offset);
1383 if (!PageHighMem(sg_page(s))) {
1384 check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1387 check_sg_segment(dev, s);
1389 add_dma_entry(entry);
1392 EXPORT_SYMBOL(debug_dma_map_sg);
1394 static int get_nr_mapped_entries(struct device *dev,
1395 struct dma_debug_entry *ref)
1397 struct dma_debug_entry *entry;
1398 struct hash_bucket *bucket;
1399 unsigned long flags;
1402 bucket = get_hash_bucket(ref, &flags);
1403 entry = bucket_find_exact(bucket, ref);
1407 mapped_ents = entry->sg_mapped_ents;
1408 put_hash_bucket(bucket, &flags);
1413 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1414 int nelems, int dir)
1416 struct scatterlist *s;
1417 int mapped_ents = 0, i;
1419 if (unlikely(dma_debug_disabled()))
1422 for_each_sg(sglist, s, nelems, i) {
1424 struct dma_debug_entry ref = {
1425 .type = dma_debug_sg,
1427 .pfn = page_to_pfn(sg_page(s)),
1428 .offset = s->offset,
1429 .dev_addr = sg_dma_address(s),
1430 .size = sg_dma_len(s),
1432 .sg_call_ents = nelems,
1435 if (mapped_ents && i >= mapped_ents)
1439 mapped_ents = get_nr_mapped_entries(dev, &ref);
1444 EXPORT_SYMBOL(debug_dma_unmap_sg);
1446 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1447 dma_addr_t dma_addr, void *virt)
1449 struct dma_debug_entry *entry;
1451 if (unlikely(dma_debug_disabled()))
1454 if (unlikely(virt == NULL))
1457 /* handle vmalloc and linear addresses */
1458 if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1461 entry = dma_entry_alloc();
1465 entry->type = dma_debug_coherent;
1467 entry->offset = offset_in_page(virt);
1469 entry->dev_addr = dma_addr;
1470 entry->direction = DMA_BIDIRECTIONAL;
1472 if (is_vmalloc_addr(virt))
1473 entry->pfn = vmalloc_to_pfn(virt);
1475 entry->pfn = page_to_pfn(virt_to_page(virt));
1477 add_dma_entry(entry);
1480 void debug_dma_free_coherent(struct device *dev, size_t size,
1481 void *virt, dma_addr_t addr)
1483 struct dma_debug_entry ref = {
1484 .type = dma_debug_coherent,
1486 .offset = offset_in_page(virt),
1489 .direction = DMA_BIDIRECTIONAL,
1492 /* handle vmalloc and linear addresses */
1493 if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1496 if (is_vmalloc_addr(virt))
1497 ref.pfn = vmalloc_to_pfn(virt);
1499 ref.pfn = page_to_pfn(virt_to_page(virt));
1501 if (unlikely(dma_debug_disabled()))
1507 void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1508 int direction, dma_addr_t dma_addr)
1510 struct dma_debug_entry *entry;
1512 if (unlikely(dma_debug_disabled()))
1515 entry = dma_entry_alloc();
1519 entry->type = dma_debug_resource;
1521 entry->pfn = PHYS_PFN(addr);
1522 entry->offset = offset_in_page(addr);
1524 entry->dev_addr = dma_addr;
1525 entry->direction = direction;
1526 entry->map_err_type = MAP_ERR_NOT_CHECKED;
1528 add_dma_entry(entry);
1530 EXPORT_SYMBOL(debug_dma_map_resource);
1532 void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1533 size_t size, int direction)
1535 struct dma_debug_entry ref = {
1536 .type = dma_debug_resource,
1538 .dev_addr = dma_addr,
1540 .direction = direction,
1543 if (unlikely(dma_debug_disabled()))
1548 EXPORT_SYMBOL(debug_dma_unmap_resource);
1550 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1551 size_t size, int direction)
1553 struct dma_debug_entry ref;
1555 if (unlikely(dma_debug_disabled()))
1558 ref.type = dma_debug_single;
1560 ref.dev_addr = dma_handle;
1562 ref.direction = direction;
1563 ref.sg_call_ents = 0;
1565 check_sync(dev, &ref, true);
1567 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1569 void debug_dma_sync_single_for_device(struct device *dev,
1570 dma_addr_t dma_handle, size_t size,
1573 struct dma_debug_entry ref;
1575 if (unlikely(dma_debug_disabled()))
1578 ref.type = dma_debug_single;
1580 ref.dev_addr = dma_handle;
1582 ref.direction = direction;
1583 ref.sg_call_ents = 0;
1585 check_sync(dev, &ref, false);
1587 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1589 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1590 int nelems, int direction)
1592 struct scatterlist *s;
1593 int mapped_ents = 0, i;
1595 if (unlikely(dma_debug_disabled()))
1598 for_each_sg(sg, s, nelems, i) {
1600 struct dma_debug_entry ref = {
1601 .type = dma_debug_sg,
1603 .pfn = page_to_pfn(sg_page(s)),
1604 .offset = s->offset,
1605 .dev_addr = sg_dma_address(s),
1606 .size = sg_dma_len(s),
1607 .direction = direction,
1608 .sg_call_ents = nelems,
1612 mapped_ents = get_nr_mapped_entries(dev, &ref);
1614 if (i >= mapped_ents)
1617 check_sync(dev, &ref, true);
1620 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1622 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1623 int nelems, int direction)
1625 struct scatterlist *s;
1626 int mapped_ents = 0, i;
1628 if (unlikely(dma_debug_disabled()))
1631 for_each_sg(sg, s, nelems, i) {
1633 struct dma_debug_entry ref = {
1634 .type = dma_debug_sg,
1636 .pfn = page_to_pfn(sg_page(s)),
1637 .offset = s->offset,
1638 .dev_addr = sg_dma_address(s),
1639 .size = sg_dma_len(s),
1640 .direction = direction,
1641 .sg_call_ents = nelems,
1644 mapped_ents = get_nr_mapped_entries(dev, &ref);
1646 if (i >= mapped_ents)
1649 check_sync(dev, &ref, false);
1652 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1654 static int __init dma_debug_driver_setup(char *str)
1658 for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1659 current_driver_name[i] = *str;
1664 if (current_driver_name[0])
1665 pr_info("enable driver filter for driver [%s]\n",
1666 current_driver_name);
1671 __setup("dma_debug_driver=", dma_debug_driver_setup);