1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2008 Advanced Micro Devices, Inc.
5 * Author: Joerg Roedel <joerg.roedel@amd.com>
8 #define pr_fmt(fmt) "DMA-API: " fmt
10 #include <linux/sched/task_stack.h>
11 #include <linux/scatterlist.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/sched/task.h>
14 #include <linux/stacktrace.h>
15 #include <linux/dma-debug.h>
16 #include <linux/spinlock.h>
17 #include <linux/vmalloc.h>
18 #include <linux/debugfs.h>
19 #include <linux/uaccess.h>
20 #include <linux/export.h>
21 #include <linux/device.h>
22 #include <linux/types.h>
23 #include <linux/sched.h>
24 #include <linux/ctype.h>
25 #include <linux/list.h>
26 #include <linux/slab.h>
28 #include <asm/sections.h>
30 #define HASH_SIZE 16384ULL
31 #define HASH_FN_SHIFT 13
32 #define HASH_FN_MASK (HASH_SIZE - 1)
34 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
35 /* If the pool runs out, add this many new entries at once */
36 #define DMA_DEBUG_DYNAMIC_ENTRIES (PAGE_SIZE / sizeof(struct dma_debug_entry))
46 MAP_ERR_CHECK_NOT_APPLICABLE,
51 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
54 * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
55 * @list: node on pre-allocated free_entries list
56 * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
57 * @size: length of the mapping
58 * @type: single, page, sg, coherent
59 * @direction: enum dma_data_direction
60 * @sg_call_ents: 'nents' from dma_map_sg
61 * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
62 * @pfn: page frame of the start address
63 * @offset: offset of mapping relative to pfn
64 * @map_err_type: track whether dma_mapping_error() was checked
65 * @stacktrace: support backtraces when a violation is detected
67 struct dma_debug_entry {
68 struct list_head list;
78 enum map_err_types map_err_type;
79 #ifdef CONFIG_STACKTRACE
80 unsigned int stack_len;
81 unsigned long stack_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
83 } ____cacheline_aligned_in_smp;
85 typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
88 struct list_head list;
92 /* Hash list to save the allocated dma addresses */
93 static struct hash_bucket dma_entry_hash[HASH_SIZE];
94 /* List of pre-allocated dma_debug_entry's */
95 static LIST_HEAD(free_entries);
96 /* Lock for the list above */
97 static DEFINE_SPINLOCK(free_entries_lock);
99 /* Global disable flag - will be set in case of an error */
100 static bool global_disable __read_mostly;
102 /* Early initialization disable flag, set at the end of dma_debug_init */
103 static bool dma_debug_initialized __read_mostly;
105 static inline bool dma_debug_disabled(void)
107 return global_disable || !dma_debug_initialized;
110 /* Global error count */
111 static u32 error_count;
113 /* Global error show enable*/
114 static u32 show_all_errors __read_mostly;
115 /* Number of errors to show */
116 static u32 show_num_errors = 1;
118 static u32 num_free_entries;
119 static u32 min_free_entries;
120 static u32 nr_total_entries;
122 /* number of preallocated entries requested by kernel cmdline */
123 static u32 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
125 /* per-driver filter related state */
127 #define NAME_MAX_LEN 64
129 static char current_driver_name[NAME_MAX_LEN] __read_mostly;
130 static struct device_driver *current_driver __read_mostly;
132 static DEFINE_RWLOCK(driver_name_lock);
134 static const char *const maperr2str[] = {
135 [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
136 [MAP_ERR_NOT_CHECKED] = "dma map error not checked",
137 [MAP_ERR_CHECKED] = "dma map error checked",
140 static const char *type2name[] = {
141 [dma_debug_single] = "single",
142 [dma_debug_sg] = "scather-gather",
143 [dma_debug_coherent] = "coherent",
144 [dma_debug_resource] = "resource",
147 static const char *dir2name[] = {
148 [DMA_BIDIRECTIONAL] = "DMA_BIDIRECTIONAL",
149 [DMA_TO_DEVICE] = "DMA_TO_DEVICE",
150 [DMA_FROM_DEVICE] = "DMA_FROM_DEVICE",
151 [DMA_NONE] = "DMA_NONE",
155 * The access to some variables in this macro is racy. We can't use atomic_t
156 * here because all these variables are exported to debugfs. Some of them even
157 * writeable. This is also the reason why a lock won't help much. But anyway,
158 * the races are no big deal. Here is why:
160 * error_count: the addition is racy, but the worst thing that can happen is
161 * that we don't count some errors
162 * show_num_errors: the subtraction is racy. Also no big deal because in
163 * worst case this will result in one warning more in the
164 * system log than the user configured. This variable is
165 * writeable via debugfs.
167 static inline void dump_entry_trace(struct dma_debug_entry *entry)
169 #ifdef CONFIG_STACKTRACE
171 pr_warn("Mapped at:\n");
172 stack_trace_print(entry->stack_entries, entry->stack_len, 0);
177 static bool driver_filter(struct device *dev)
179 struct device_driver *drv;
183 /* driver filter off */
184 if (likely(!current_driver_name[0]))
187 /* driver filter on and initialized */
188 if (current_driver && dev && dev->driver == current_driver)
191 /* driver filter on, but we can't filter on a NULL device... */
195 if (current_driver || !current_driver_name[0])
198 /* driver filter on but not yet initialized */
203 /* lock to protect against change of current_driver_name */
204 read_lock_irqsave(&driver_name_lock, flags);
208 strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
209 current_driver = drv;
213 read_unlock_irqrestore(&driver_name_lock, flags);
218 #define err_printk(dev, entry, format, arg...) do { \
220 if (driver_filter(dev) && \
221 (show_all_errors || show_num_errors > 0)) { \
222 WARN(1, pr_fmt("%s %s: ") format, \
223 dev ? dev_driver_string(dev) : "NULL", \
224 dev ? dev_name(dev) : "NULL", ## arg); \
225 dump_entry_trace(entry); \
227 if (!show_all_errors && show_num_errors > 0) \
228 show_num_errors -= 1; \
232 * Hash related functions
234 * Every DMA-API request is saved into a struct dma_debug_entry. To
235 * have quick access to these structs they are stored into a hash.
237 static int hash_fn(struct dma_debug_entry *entry)
240 * Hash function is based on the dma address.
241 * We use bits 20-27 here as the index into the hash
243 return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
247 * Request exclusive access to a hash bucket for a given dma_debug_entry.
249 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
250 unsigned long *flags)
251 __acquires(&dma_entry_hash[idx].lock)
253 int idx = hash_fn(entry);
254 unsigned long __flags;
256 spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
258 return &dma_entry_hash[idx];
262 * Give up exclusive access to the hash bucket
264 static void put_hash_bucket(struct hash_bucket *bucket,
266 __releases(&bucket->lock)
268 spin_unlock_irqrestore(&bucket->lock, flags);
271 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
273 return ((a->dev_addr == b->dev_addr) &&
274 (a->dev == b->dev)) ? true : false;
277 static bool containing_match(struct dma_debug_entry *a,
278 struct dma_debug_entry *b)
280 if (a->dev != b->dev)
283 if ((b->dev_addr <= a->dev_addr) &&
284 ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
291 * Search a given entry in the hash bucket list
293 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
294 struct dma_debug_entry *ref,
297 struct dma_debug_entry *entry, *ret = NULL;
298 int matches = 0, match_lvl, last_lvl = -1;
300 list_for_each_entry(entry, &bucket->list, list) {
301 if (!match(ref, entry))
305 * Some drivers map the same physical address multiple
306 * times. Without a hardware IOMMU this results in the
307 * same device addresses being put into the dma-debug
308 * hash multiple times too. This can result in false
309 * positives being reported. Therefore we implement a
310 * best-fit algorithm here which returns the entry from
311 * the hash which fits best to the reference value
312 * instead of the first-fit.
316 entry->size == ref->size ? ++match_lvl : 0;
317 entry->type == ref->type ? ++match_lvl : 0;
318 entry->direction == ref->direction ? ++match_lvl : 0;
319 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
321 if (match_lvl == 4) {
322 /* perfect-fit - return the result */
324 } else if (match_lvl > last_lvl) {
326 * We found an entry that fits better then the
327 * previous one or it is the 1st match.
329 last_lvl = match_lvl;
335 * If we have multiple matches but no perfect-fit, just return
338 ret = (matches == 1) ? ret : NULL;
343 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
344 struct dma_debug_entry *ref)
346 return __hash_bucket_find(bucket, ref, exact_match);
349 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
350 struct dma_debug_entry *ref,
351 unsigned long *flags)
354 unsigned int max_range = dma_get_max_seg_size(ref->dev);
355 struct dma_debug_entry *entry, index = *ref;
356 unsigned int range = 0;
358 while (range <= max_range) {
359 entry = __hash_bucket_find(*bucket, ref, containing_match);
365 * Nothing found, go back a hash bucket
367 put_hash_bucket(*bucket, *flags);
368 range += (1 << HASH_FN_SHIFT);
369 index.dev_addr -= (1 << HASH_FN_SHIFT);
370 *bucket = get_hash_bucket(&index, flags);
377 * Add an entry to a hash bucket
379 static void hash_bucket_add(struct hash_bucket *bucket,
380 struct dma_debug_entry *entry)
382 list_add_tail(&entry->list, &bucket->list);
386 * Remove entry from a hash bucket list
388 static void hash_bucket_del(struct dma_debug_entry *entry)
390 list_del(&entry->list);
393 static unsigned long long phys_addr(struct dma_debug_entry *entry)
395 if (entry->type == dma_debug_resource)
396 return __pfn_to_phys(entry->pfn) + entry->offset;
398 return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
402 * Dump mapping entries for debugging purposes
404 void debug_dma_dump_mappings(struct device *dev)
408 for (idx = 0; idx < HASH_SIZE; idx++) {
409 struct hash_bucket *bucket = &dma_entry_hash[idx];
410 struct dma_debug_entry *entry;
413 spin_lock_irqsave(&bucket->lock, flags);
415 list_for_each_entry(entry, &bucket->list, list) {
416 if (!dev || dev == entry->dev) {
418 "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
419 type2name[entry->type], idx,
420 phys_addr(entry), entry->pfn,
421 entry->dev_addr, entry->size,
422 dir2name[entry->direction],
423 maperr2str[entry->map_err_type]);
427 spin_unlock_irqrestore(&bucket->lock, flags);
433 * For each mapping (initial cacheline in the case of
434 * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
435 * scatterlist, or the cacheline specified in dma_map_single) insert
436 * into this tree using the cacheline as the key. At
437 * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry. If
438 * the entry already exists at insertion time add a tag as a reference
439 * count for the overlapping mappings. For now, the overlap tracking
440 * just ensures that 'unmaps' balance 'maps' before marking the
441 * cacheline idle, but we should also be flagging overlaps as an API
444 * Memory usage is mostly constrained by the maximum number of available
445 * dma-debug entries in that we need a free dma_debug_entry before
446 * inserting into the tree. In the case of dma_map_page and
447 * dma_alloc_coherent there is only one dma_debug_entry and one
448 * dma_active_cacheline entry to track per event. dma_map_sg(), on the
449 * other hand, consumes a single dma_debug_entry, but inserts 'nents'
450 * entries into the tree.
452 static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
453 static DEFINE_SPINLOCK(radix_lock);
454 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
455 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
456 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
458 static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
460 return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
461 (entry->offset >> L1_CACHE_SHIFT);
464 static int active_cacheline_read_overlap(phys_addr_t cln)
468 for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
469 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
474 static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
478 if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
481 for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
482 if (overlap & 1 << i)
483 radix_tree_tag_set(&dma_active_cacheline, cln, i);
485 radix_tree_tag_clear(&dma_active_cacheline, cln, i);
490 static void active_cacheline_inc_overlap(phys_addr_t cln)
492 int overlap = active_cacheline_read_overlap(cln);
494 overlap = active_cacheline_set_overlap(cln, ++overlap);
496 /* If we overflowed the overlap counter then we're potentially
497 * leaking dma-mappings.
499 WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
500 pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
501 ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
504 static int active_cacheline_dec_overlap(phys_addr_t cln)
506 int overlap = active_cacheline_read_overlap(cln);
508 return active_cacheline_set_overlap(cln, --overlap);
511 static int active_cacheline_insert(struct dma_debug_entry *entry)
513 phys_addr_t cln = to_cacheline_number(entry);
517 /* If the device is not writing memory then we don't have any
518 * concerns about the cpu consuming stale data. This mitigates
519 * legitimate usages of overlapping mappings.
521 if (entry->direction == DMA_TO_DEVICE)
524 spin_lock_irqsave(&radix_lock, flags);
525 rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
527 active_cacheline_inc_overlap(cln);
528 spin_unlock_irqrestore(&radix_lock, flags);
533 static void active_cacheline_remove(struct dma_debug_entry *entry)
535 phys_addr_t cln = to_cacheline_number(entry);
538 /* ...mirror the insert case */
539 if (entry->direction == DMA_TO_DEVICE)
542 spin_lock_irqsave(&radix_lock, flags);
543 /* since we are counting overlaps the final put of the
544 * cacheline will occur when the overlap count is 0.
545 * active_cacheline_dec_overlap() returns -1 in that case
547 if (active_cacheline_dec_overlap(cln) < 0)
548 radix_tree_delete(&dma_active_cacheline, cln);
549 spin_unlock_irqrestore(&radix_lock, flags);
553 * Wrapper function for adding an entry to the hash.
554 * This function takes care of locking itself.
556 static void add_dma_entry(struct dma_debug_entry *entry)
558 struct hash_bucket *bucket;
562 bucket = get_hash_bucket(entry, &flags);
563 hash_bucket_add(bucket, entry);
564 put_hash_bucket(bucket, flags);
566 rc = active_cacheline_insert(entry);
568 pr_err("cacheline tracking ENOMEM, dma-debug disabled\n");
569 global_disable = true;
572 /* TODO: report -EEXIST errors here as overlapping mappings are
573 * not supported by the DMA API
577 static int dma_debug_create_entries(gfp_t gfp)
579 struct dma_debug_entry *entry;
582 entry = (void *)get_zeroed_page(gfp);
586 for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
587 list_add_tail(&entry[i].list, &free_entries);
589 num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
590 nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
595 static struct dma_debug_entry *__dma_entry_alloc(void)
597 struct dma_debug_entry *entry;
599 entry = list_entry(free_entries.next, struct dma_debug_entry, list);
600 list_del(&entry->list);
601 memset(entry, 0, sizeof(*entry));
603 num_free_entries -= 1;
604 if (num_free_entries < min_free_entries)
605 min_free_entries = num_free_entries;
610 static void __dma_entry_alloc_check_leak(void)
612 u32 tmp = nr_total_entries % nr_prealloc_entries;
614 /* Shout each time we tick over some multiple of the initial pool */
615 if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
616 pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
618 (nr_total_entries / nr_prealloc_entries));
622 /* struct dma_entry allocator
624 * The next two functions implement the allocator for
625 * struct dma_debug_entries.
627 static struct dma_debug_entry *dma_entry_alloc(void)
629 struct dma_debug_entry *entry;
632 spin_lock_irqsave(&free_entries_lock, flags);
633 if (num_free_entries == 0) {
634 if (dma_debug_create_entries(GFP_ATOMIC)) {
635 global_disable = true;
636 spin_unlock_irqrestore(&free_entries_lock, flags);
637 pr_err("debugging out of memory - disabling\n");
640 __dma_entry_alloc_check_leak();
643 entry = __dma_entry_alloc();
645 spin_unlock_irqrestore(&free_entries_lock, flags);
647 #ifdef CONFIG_STACKTRACE
648 entry->stack_len = stack_trace_save(entry->stack_entries,
649 ARRAY_SIZE(entry->stack_entries),
655 static void dma_entry_free(struct dma_debug_entry *entry)
659 active_cacheline_remove(entry);
662 * add to beginning of the list - this way the entries are
663 * more likely cache hot when they are reallocated.
665 spin_lock_irqsave(&free_entries_lock, flags);
666 list_add(&entry->list, &free_entries);
667 num_free_entries += 1;
668 spin_unlock_irqrestore(&free_entries_lock, flags);
672 * DMA-API debugging init code
674 * The init code does two things:
675 * 1. Initialize core data structures
676 * 2. Preallocate a given number of dma_debug_entry structs
679 static ssize_t filter_read(struct file *file, char __user *user_buf,
680 size_t count, loff_t *ppos)
682 char buf[NAME_MAX_LEN + 1];
686 if (!current_driver_name[0])
690 * We can't copy to userspace directly because current_driver_name can
691 * only be read under the driver_name_lock with irqs disabled. So
692 * create a temporary copy first.
694 read_lock_irqsave(&driver_name_lock, flags);
695 len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
696 read_unlock_irqrestore(&driver_name_lock, flags);
698 return simple_read_from_buffer(user_buf, count, ppos, buf, len);
701 static ssize_t filter_write(struct file *file, const char __user *userbuf,
702 size_t count, loff_t *ppos)
704 char buf[NAME_MAX_LEN];
710 * We can't copy from userspace directly. Access to
711 * current_driver_name is protected with a write_lock with irqs
712 * disabled. Since copy_from_user can fault and may sleep we
713 * need to copy to temporary buffer first
715 len = min(count, (size_t)(NAME_MAX_LEN - 1));
716 if (copy_from_user(buf, userbuf, len))
721 write_lock_irqsave(&driver_name_lock, flags);
724 * Now handle the string we got from userspace very carefully.
726 * - only use the first token we got
727 * - token delimiter is everything looking like a space
728 * character (' ', '\n', '\t' ...)
731 if (!isalnum(buf[0])) {
733 * If the first character userspace gave us is not
734 * alphanumerical then assume the filter should be
737 if (current_driver_name[0])
738 pr_info("switching off dma-debug driver filter\n");
739 current_driver_name[0] = 0;
740 current_driver = NULL;
745 * Now parse out the first token and use it as the name for the
746 * driver to filter for.
748 for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
749 current_driver_name[i] = buf[i];
750 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
753 current_driver_name[i] = 0;
754 current_driver = NULL;
756 pr_info("enable driver filter for driver [%s]\n",
757 current_driver_name);
760 write_unlock_irqrestore(&driver_name_lock, flags);
765 static const struct file_operations filter_fops = {
767 .write = filter_write,
768 .llseek = default_llseek,
771 static int dump_show(struct seq_file *seq, void *v)
775 for (idx = 0; idx < HASH_SIZE; idx++) {
776 struct hash_bucket *bucket = &dma_entry_hash[idx];
777 struct dma_debug_entry *entry;
780 spin_lock_irqsave(&bucket->lock, flags);
781 list_for_each_entry(entry, &bucket->list, list) {
783 "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx %s %s\n",
784 dev_name(entry->dev),
785 dev_driver_string(entry->dev),
786 type2name[entry->type], idx,
787 phys_addr(entry), entry->pfn,
788 entry->dev_addr, entry->size,
789 dir2name[entry->direction],
790 maperr2str[entry->map_err_type]);
792 spin_unlock_irqrestore(&bucket->lock, flags);
796 DEFINE_SHOW_ATTRIBUTE(dump);
798 static void dma_debug_fs_init(void)
800 struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
802 debugfs_create_bool("disabled", 0444, dentry, &global_disable);
803 debugfs_create_u32("error_count", 0444, dentry, &error_count);
804 debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
805 debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
806 debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
807 debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
808 debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
809 debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
810 debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
813 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
815 struct dma_debug_entry *entry;
819 for (i = 0; i < HASH_SIZE; ++i) {
820 spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
821 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
822 if (entry->dev == dev) {
827 spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
833 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
835 struct device *dev = data;
836 struct dma_debug_entry *entry;
839 if (dma_debug_disabled())
843 case BUS_NOTIFY_UNBOUND_DRIVER:
844 count = device_dma_allocations(dev, &entry);
847 err_printk(dev, entry, "device driver has pending "
848 "DMA allocations while released from device "
850 "One of leaked entries details: "
851 "[device address=0x%016llx] [size=%llu bytes] "
852 "[mapped with %s] [mapped as %s]\n",
853 count, entry->dev_addr, entry->size,
854 dir2name[entry->direction], type2name[entry->type]);
863 void dma_debug_add_bus(struct bus_type *bus)
865 struct notifier_block *nb;
867 if (dma_debug_disabled())
870 nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
872 pr_err("dma_debug_add_bus: out of memory\n");
876 nb->notifier_call = dma_debug_device_change;
878 bus_register_notifier(bus, nb);
881 static int dma_debug_init(void)
885 /* Do not use dma_debug_initialized here, since we really want to be
886 * called to set dma_debug_initialized
891 for (i = 0; i < HASH_SIZE; ++i) {
892 INIT_LIST_HEAD(&dma_entry_hash[i].list);
893 spin_lock_init(&dma_entry_hash[i].lock);
898 nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
899 for (i = 0; i < nr_pages; ++i)
900 dma_debug_create_entries(GFP_KERNEL);
901 if (num_free_entries >= nr_prealloc_entries) {
902 pr_info("preallocated %d debug entries\n", nr_total_entries);
903 } else if (num_free_entries > 0) {
904 pr_warn("%d debug entries requested but only %d allocated\n",
905 nr_prealloc_entries, nr_total_entries);
907 pr_err("debugging out of memory error - disabled\n");
908 global_disable = true;
912 min_free_entries = num_free_entries;
914 dma_debug_initialized = true;
916 pr_info("debugging enabled by kernel config\n");
919 core_initcall(dma_debug_init);
921 static __init int dma_debug_cmdline(char *str)
926 if (strncmp(str, "off", 3) == 0) {
927 pr_info("debugging disabled on kernel command line\n");
928 global_disable = true;
934 static __init int dma_debug_entries_cmdline(char *str)
938 if (!get_option(&str, &nr_prealloc_entries))
939 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
943 __setup("dma_debug=", dma_debug_cmdline);
944 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
946 static void check_unmap(struct dma_debug_entry *ref)
948 struct dma_debug_entry *entry;
949 struct hash_bucket *bucket;
952 bucket = get_hash_bucket(ref, &flags);
953 entry = bucket_find_exact(bucket, ref);
956 /* must drop lock before calling dma_mapping_error */
957 put_hash_bucket(bucket, flags);
959 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
960 err_printk(ref->dev, NULL,
961 "device driver tries to free an "
962 "invalid DMA memory address\n");
964 err_printk(ref->dev, NULL,
965 "device driver tries to free DMA "
966 "memory it has not allocated [device "
967 "address=0x%016llx] [size=%llu bytes]\n",
968 ref->dev_addr, ref->size);
973 if (ref->size != entry->size) {
974 err_printk(ref->dev, entry, "device driver frees "
975 "DMA memory with different size "
976 "[device address=0x%016llx] [map size=%llu bytes] "
977 "[unmap size=%llu bytes]\n",
978 ref->dev_addr, entry->size, ref->size);
981 if (ref->type != entry->type) {
982 err_printk(ref->dev, entry, "device driver frees "
983 "DMA memory with wrong function "
984 "[device address=0x%016llx] [size=%llu bytes] "
985 "[mapped as %s] [unmapped as %s]\n",
986 ref->dev_addr, ref->size,
987 type2name[entry->type], type2name[ref->type]);
988 } else if ((entry->type == dma_debug_coherent) &&
989 (phys_addr(ref) != phys_addr(entry))) {
990 err_printk(ref->dev, entry, "device driver frees "
991 "DMA memory with different CPU address "
992 "[device address=0x%016llx] [size=%llu bytes] "
993 "[cpu alloc address=0x%016llx] "
994 "[cpu free address=0x%016llx]",
995 ref->dev_addr, ref->size,
1000 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1001 ref->sg_call_ents != entry->sg_call_ents) {
1002 err_printk(ref->dev, entry, "device driver frees "
1003 "DMA sg list with different entry count "
1004 "[map count=%d] [unmap count=%d]\n",
1005 entry->sg_call_ents, ref->sg_call_ents);
1009 * This may be no bug in reality - but most implementations of the
1010 * DMA API don't handle this properly, so check for it here
1012 if (ref->direction != entry->direction) {
1013 err_printk(ref->dev, entry, "device driver frees "
1014 "DMA memory with different direction "
1015 "[device address=0x%016llx] [size=%llu bytes] "
1016 "[mapped with %s] [unmapped with %s]\n",
1017 ref->dev_addr, ref->size,
1018 dir2name[entry->direction],
1019 dir2name[ref->direction]);
1023 * Drivers should use dma_mapping_error() to check the returned
1024 * addresses of dma_map_single() and dma_map_page().
1025 * If not, print this warning message. See Documentation/core-api/dma-api.rst.
1027 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1028 err_printk(ref->dev, entry,
1029 "device driver failed to check map error"
1030 "[device address=0x%016llx] [size=%llu bytes] "
1032 ref->dev_addr, ref->size,
1033 type2name[entry->type]);
1036 hash_bucket_del(entry);
1037 dma_entry_free(entry);
1039 put_hash_bucket(bucket, flags);
1042 static void check_for_stack(struct device *dev,
1043 struct page *page, size_t offset)
1046 struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1048 if (!stack_vm_area) {
1049 /* Stack is direct-mapped. */
1050 if (PageHighMem(page))
1052 addr = page_address(page) + offset;
1053 if (object_is_on_stack(addr))
1054 err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1056 /* Stack is vmalloced. */
1059 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1060 if (page != stack_vm_area->pages[i])
1063 addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1064 err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1070 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1072 unsigned long a1 = (unsigned long)addr;
1073 unsigned long b1 = a1 + len;
1074 unsigned long a2 = (unsigned long)start;
1075 unsigned long b2 = (unsigned long)end;
1077 return !(b1 <= a2 || a1 >= b2);
1080 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1082 if (overlap(addr, len, _stext, _etext) ||
1083 overlap(addr, len, __start_rodata, __end_rodata))
1084 err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1087 static void check_sync(struct device *dev,
1088 struct dma_debug_entry *ref,
1091 struct dma_debug_entry *entry;
1092 struct hash_bucket *bucket;
1093 unsigned long flags;
1095 bucket = get_hash_bucket(ref, &flags);
1097 entry = bucket_find_contain(&bucket, ref, &flags);
1100 err_printk(dev, NULL, "device driver tries "
1101 "to sync DMA memory it has not allocated "
1102 "[device address=0x%016llx] [size=%llu bytes]\n",
1103 (unsigned long long)ref->dev_addr, ref->size);
1107 if (ref->size > entry->size) {
1108 err_printk(dev, entry, "device driver syncs"
1109 " DMA memory outside allocated range "
1110 "[device address=0x%016llx] "
1111 "[allocation size=%llu bytes] "
1112 "[sync offset+size=%llu]\n",
1113 entry->dev_addr, entry->size,
1117 if (entry->direction == DMA_BIDIRECTIONAL)
1120 if (ref->direction != entry->direction) {
1121 err_printk(dev, entry, "device driver syncs "
1122 "DMA memory with different direction "
1123 "[device address=0x%016llx] [size=%llu bytes] "
1124 "[mapped with %s] [synced with %s]\n",
1125 (unsigned long long)ref->dev_addr, entry->size,
1126 dir2name[entry->direction],
1127 dir2name[ref->direction]);
1130 if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1131 !(ref->direction == DMA_TO_DEVICE))
1132 err_printk(dev, entry, "device driver syncs "
1133 "device read-only DMA memory for cpu "
1134 "[device address=0x%016llx] [size=%llu bytes] "
1135 "[mapped with %s] [synced with %s]\n",
1136 (unsigned long long)ref->dev_addr, entry->size,
1137 dir2name[entry->direction],
1138 dir2name[ref->direction]);
1140 if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1141 !(ref->direction == DMA_FROM_DEVICE))
1142 err_printk(dev, entry, "device driver syncs "
1143 "device write-only DMA memory to device "
1144 "[device address=0x%016llx] [size=%llu bytes] "
1145 "[mapped with %s] [synced with %s]\n",
1146 (unsigned long long)ref->dev_addr, entry->size,
1147 dir2name[entry->direction],
1148 dir2name[ref->direction]);
1150 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1151 ref->sg_call_ents != entry->sg_call_ents) {
1152 err_printk(ref->dev, entry, "device driver syncs "
1153 "DMA sg list with different entry count "
1154 "[map count=%d] [sync count=%d]\n",
1155 entry->sg_call_ents, ref->sg_call_ents);
1159 put_hash_bucket(bucket, flags);
1162 static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1164 #ifdef CONFIG_DMA_API_DEBUG_SG
1165 unsigned int max_seg = dma_get_max_seg_size(dev);
1166 u64 start, end, boundary = dma_get_seg_boundary(dev);
1169 * Either the driver forgot to set dma_parms appropriately, or
1170 * whoever generated the list forgot to check them.
1172 if (sg->length > max_seg)
1173 err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1174 sg->length, max_seg);
1176 * In some cases this could potentially be the DMA API
1177 * implementation's fault, but it would usually imply that
1178 * the scatterlist was built inappropriately to begin with.
1180 start = sg_dma_address(sg);
1181 end = start + sg_dma_len(sg) - 1;
1182 if ((start ^ end) & ~boundary)
1183 err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1184 start, end, boundary);
1188 void debug_dma_map_single(struct device *dev, const void *addr,
1191 if (unlikely(dma_debug_disabled()))
1194 if (!virt_addr_valid(addr))
1195 err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1198 if (is_vmalloc_addr(addr))
1199 err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1202 EXPORT_SYMBOL(debug_dma_map_single);
1204 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1205 size_t size, int direction, dma_addr_t dma_addr)
1207 struct dma_debug_entry *entry;
1209 if (unlikely(dma_debug_disabled()))
1212 if (dma_mapping_error(dev, dma_addr))
1215 entry = dma_entry_alloc();
1220 entry->type = dma_debug_single;
1221 entry->pfn = page_to_pfn(page);
1222 entry->offset = offset,
1223 entry->dev_addr = dma_addr;
1225 entry->direction = direction;
1226 entry->map_err_type = MAP_ERR_NOT_CHECKED;
1228 check_for_stack(dev, page, offset);
1230 if (!PageHighMem(page)) {
1231 void *addr = page_address(page) + offset;
1233 check_for_illegal_area(dev, addr, size);
1236 add_dma_entry(entry);
1238 EXPORT_SYMBOL(debug_dma_map_page);
1240 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1242 struct dma_debug_entry ref;
1243 struct dma_debug_entry *entry;
1244 struct hash_bucket *bucket;
1245 unsigned long flags;
1247 if (unlikely(dma_debug_disabled()))
1251 ref.dev_addr = dma_addr;
1252 bucket = get_hash_bucket(&ref, &flags);
1254 list_for_each_entry(entry, &bucket->list, list) {
1255 if (!exact_match(&ref, entry))
1259 * The same physical address can be mapped multiple
1260 * times. Without a hardware IOMMU this results in the
1261 * same device addresses being put into the dma-debug
1262 * hash multiple times too. This can result in false
1263 * positives being reported. Therefore we implement a
1264 * best-fit algorithm here which updates the first entry
1265 * from the hash which fits the reference value and is
1266 * not currently listed as being checked.
1268 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1269 entry->map_err_type = MAP_ERR_CHECKED;
1274 put_hash_bucket(bucket, flags);
1276 EXPORT_SYMBOL(debug_dma_mapping_error);
1278 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1279 size_t size, int direction)
1281 struct dma_debug_entry ref = {
1282 .type = dma_debug_single,
1286 .direction = direction,
1289 if (unlikely(dma_debug_disabled()))
1293 EXPORT_SYMBOL(debug_dma_unmap_page);
1295 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1296 int nents, int mapped_ents, int direction)
1298 struct dma_debug_entry *entry;
1299 struct scatterlist *s;
1302 if (unlikely(dma_debug_disabled()))
1305 for_each_sg(sg, s, mapped_ents, i) {
1306 entry = dma_entry_alloc();
1310 entry->type = dma_debug_sg;
1312 entry->pfn = page_to_pfn(sg_page(s));
1313 entry->offset = s->offset,
1314 entry->size = sg_dma_len(s);
1315 entry->dev_addr = sg_dma_address(s);
1316 entry->direction = direction;
1317 entry->sg_call_ents = nents;
1318 entry->sg_mapped_ents = mapped_ents;
1320 check_for_stack(dev, sg_page(s), s->offset);
1322 if (!PageHighMem(sg_page(s))) {
1323 check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1326 check_sg_segment(dev, s);
1328 add_dma_entry(entry);
1331 EXPORT_SYMBOL(debug_dma_map_sg);
1333 static int get_nr_mapped_entries(struct device *dev,
1334 struct dma_debug_entry *ref)
1336 struct dma_debug_entry *entry;
1337 struct hash_bucket *bucket;
1338 unsigned long flags;
1341 bucket = get_hash_bucket(ref, &flags);
1342 entry = bucket_find_exact(bucket, ref);
1346 mapped_ents = entry->sg_mapped_ents;
1347 put_hash_bucket(bucket, flags);
1352 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1353 int nelems, int dir)
1355 struct scatterlist *s;
1356 int mapped_ents = 0, i;
1358 if (unlikely(dma_debug_disabled()))
1361 for_each_sg(sglist, s, nelems, i) {
1363 struct dma_debug_entry ref = {
1364 .type = dma_debug_sg,
1366 .pfn = page_to_pfn(sg_page(s)),
1367 .offset = s->offset,
1368 .dev_addr = sg_dma_address(s),
1369 .size = sg_dma_len(s),
1371 .sg_call_ents = nelems,
1374 if (mapped_ents && i >= mapped_ents)
1378 mapped_ents = get_nr_mapped_entries(dev, &ref);
1383 EXPORT_SYMBOL(debug_dma_unmap_sg);
1385 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1386 dma_addr_t dma_addr, void *virt)
1388 struct dma_debug_entry *entry;
1390 if (unlikely(dma_debug_disabled()))
1393 if (unlikely(virt == NULL))
1396 /* handle vmalloc and linear addresses */
1397 if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1400 entry = dma_entry_alloc();
1404 entry->type = dma_debug_coherent;
1406 entry->offset = offset_in_page(virt);
1408 entry->dev_addr = dma_addr;
1409 entry->direction = DMA_BIDIRECTIONAL;
1411 if (is_vmalloc_addr(virt))
1412 entry->pfn = vmalloc_to_pfn(virt);
1414 entry->pfn = page_to_pfn(virt_to_page(virt));
1416 add_dma_entry(entry);
1419 void debug_dma_free_coherent(struct device *dev, size_t size,
1420 void *virt, dma_addr_t addr)
1422 struct dma_debug_entry ref = {
1423 .type = dma_debug_coherent,
1425 .offset = offset_in_page(virt),
1428 .direction = DMA_BIDIRECTIONAL,
1431 /* handle vmalloc and linear addresses */
1432 if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1435 if (is_vmalloc_addr(virt))
1436 ref.pfn = vmalloc_to_pfn(virt);
1438 ref.pfn = page_to_pfn(virt_to_page(virt));
1440 if (unlikely(dma_debug_disabled()))
1446 void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1447 int direction, dma_addr_t dma_addr)
1449 struct dma_debug_entry *entry;
1451 if (unlikely(dma_debug_disabled()))
1454 entry = dma_entry_alloc();
1458 entry->type = dma_debug_resource;
1460 entry->pfn = PHYS_PFN(addr);
1461 entry->offset = offset_in_page(addr);
1463 entry->dev_addr = dma_addr;
1464 entry->direction = direction;
1465 entry->map_err_type = MAP_ERR_NOT_CHECKED;
1467 add_dma_entry(entry);
1469 EXPORT_SYMBOL(debug_dma_map_resource);
1471 void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1472 size_t size, int direction)
1474 struct dma_debug_entry ref = {
1475 .type = dma_debug_resource,
1477 .dev_addr = dma_addr,
1479 .direction = direction,
1482 if (unlikely(dma_debug_disabled()))
1487 EXPORT_SYMBOL(debug_dma_unmap_resource);
1489 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1490 size_t size, int direction)
1492 struct dma_debug_entry ref;
1494 if (unlikely(dma_debug_disabled()))
1497 ref.type = dma_debug_single;
1499 ref.dev_addr = dma_handle;
1501 ref.direction = direction;
1502 ref.sg_call_ents = 0;
1504 check_sync(dev, &ref, true);
1506 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1508 void debug_dma_sync_single_for_device(struct device *dev,
1509 dma_addr_t dma_handle, size_t size,
1512 struct dma_debug_entry ref;
1514 if (unlikely(dma_debug_disabled()))
1517 ref.type = dma_debug_single;
1519 ref.dev_addr = dma_handle;
1521 ref.direction = direction;
1522 ref.sg_call_ents = 0;
1524 check_sync(dev, &ref, false);
1526 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1528 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1529 int nelems, int direction)
1531 struct scatterlist *s;
1532 int mapped_ents = 0, i;
1534 if (unlikely(dma_debug_disabled()))
1537 for_each_sg(sg, s, nelems, i) {
1539 struct dma_debug_entry ref = {
1540 .type = dma_debug_sg,
1542 .pfn = page_to_pfn(sg_page(s)),
1543 .offset = s->offset,
1544 .dev_addr = sg_dma_address(s),
1545 .size = sg_dma_len(s),
1546 .direction = direction,
1547 .sg_call_ents = nelems,
1551 mapped_ents = get_nr_mapped_entries(dev, &ref);
1553 if (i >= mapped_ents)
1556 check_sync(dev, &ref, true);
1559 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1561 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1562 int nelems, int direction)
1564 struct scatterlist *s;
1565 int mapped_ents = 0, i;
1567 if (unlikely(dma_debug_disabled()))
1570 for_each_sg(sg, s, nelems, i) {
1572 struct dma_debug_entry ref = {
1573 .type = dma_debug_sg,
1575 .pfn = page_to_pfn(sg_page(s)),
1576 .offset = s->offset,
1577 .dev_addr = sg_dma_address(s),
1578 .size = sg_dma_len(s),
1579 .direction = direction,
1580 .sg_call_ents = nelems,
1583 mapped_ents = get_nr_mapped_entries(dev, &ref);
1585 if (i >= mapped_ents)
1588 check_sync(dev, &ref, false);
1591 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1593 static int __init dma_debug_driver_setup(char *str)
1597 for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1598 current_driver_name[i] = *str;
1603 if (current_driver_name[0])
1604 pr_info("enable driver filter for driver [%s]\n",
1605 current_driver_name);
1610 __setup("dma_debug_driver=", dma_debug_driver_setup);