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/device.h>
27 #include <linux/types.h>
28 #include <linux/sched.h>
29 #include <linux/list.h>
30 #include <linux/slab.h>
32 #include <asm/sections.h>
34 #define HASH_SIZE 1024ULL
35 #define HASH_FN_SHIFT 13
36 #define HASH_FN_MASK (HASH_SIZE - 1)
45 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
47 struct dma_debug_entry {
48 struct list_head list;
57 #ifdef CONFIG_STACKTRACE
58 struct stack_trace stacktrace;
59 unsigned long st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
64 struct list_head list;
66 } ____cacheline_aligned_in_smp;
68 /* Hash list to save the allocated dma addresses */
69 static struct hash_bucket dma_entry_hash[HASH_SIZE];
70 /* List of pre-allocated dma_debug_entry's */
71 static LIST_HEAD(free_entries);
72 /* Lock for the list above */
73 static DEFINE_SPINLOCK(free_entries_lock);
75 /* Global disable flag - will be set in case of an error */
76 static bool global_disable __read_mostly;
78 /* Global error count */
79 static u32 error_count;
81 /* Global error show enable*/
82 static u32 show_all_errors __read_mostly;
83 /* Number of errors to show */
84 static u32 show_num_errors = 1;
86 static u32 num_free_entries;
87 static u32 min_free_entries;
89 /* number of preallocated entries requested by kernel cmdline */
90 static u32 req_entries;
92 /* debugfs dentry's for the stuff above */
93 static struct dentry *dma_debug_dent __read_mostly;
94 static struct dentry *global_disable_dent __read_mostly;
95 static struct dentry *error_count_dent __read_mostly;
96 static struct dentry *show_all_errors_dent __read_mostly;
97 static struct dentry *show_num_errors_dent __read_mostly;
98 static struct dentry *num_free_entries_dent __read_mostly;
99 static struct dentry *min_free_entries_dent __read_mostly;
101 static const char *type2name[4] = { "single", "page",
102 "scather-gather", "coherent" };
104 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
105 "DMA_FROM_DEVICE", "DMA_NONE" };
108 * The access to some variables in this macro is racy. We can't use atomic_t
109 * here because all these variables are exported to debugfs. Some of them even
110 * writeable. This is also the reason why a lock won't help much. But anyway,
111 * the races are no big deal. Here is why:
113 * error_count: the addition is racy, but the worst thing that can happen is
114 * that we don't count some errors
115 * show_num_errors: the subtraction is racy. Also no big deal because in
116 * worst case this will result in one warning more in the
117 * system log than the user configured. This variable is
118 * writeable via debugfs.
120 static inline void dump_entry_trace(struct dma_debug_entry *entry)
122 #ifdef CONFIG_STACKTRACE
124 printk(KERN_WARNING "Mapped at:\n");
125 print_stack_trace(&entry->stacktrace, 0);
130 #define err_printk(dev, entry, format, arg...) do { \
132 if (show_all_errors || show_num_errors > 0) { \
133 WARN(1, "%s %s: " format, \
134 dev_driver_string(dev), \
135 dev_name(dev) , ## arg); \
136 dump_entry_trace(entry); \
138 if (!show_all_errors && show_num_errors > 0) \
139 show_num_errors -= 1; \
143 * Hash related functions
145 * Every DMA-API request is saved into a struct dma_debug_entry. To
146 * have quick access to these structs they are stored into a hash.
148 static int hash_fn(struct dma_debug_entry *entry)
151 * Hash function is based on the dma address.
152 * We use bits 20-27 here as the index into the hash
154 return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
158 * Request exclusive access to a hash bucket for a given dma_debug_entry.
160 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
161 unsigned long *flags)
163 int idx = hash_fn(entry);
164 unsigned long __flags;
166 spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
168 return &dma_entry_hash[idx];
172 * Give up exclusive access to the hash bucket
174 static void put_hash_bucket(struct hash_bucket *bucket,
175 unsigned long *flags)
177 unsigned long __flags = *flags;
179 spin_unlock_irqrestore(&bucket->lock, __flags);
183 * Search a given entry in the hash bucket list
185 static struct dma_debug_entry *hash_bucket_find(struct hash_bucket *bucket,
186 struct dma_debug_entry *ref)
188 struct dma_debug_entry *entry;
190 list_for_each_entry(entry, &bucket->list, list) {
191 if ((entry->dev_addr == ref->dev_addr) &&
192 (entry->dev == ref->dev))
200 * Add an entry to a hash bucket
202 static void hash_bucket_add(struct hash_bucket *bucket,
203 struct dma_debug_entry *entry)
205 list_add_tail(&entry->list, &bucket->list);
209 * Remove entry from a hash bucket list
211 static void hash_bucket_del(struct dma_debug_entry *entry)
213 list_del(&entry->list);
217 * Dump mapping entries for debugging purposes
219 void debug_dma_dump_mappings(struct device *dev)
223 for (idx = 0; idx < HASH_SIZE; idx++) {
224 struct hash_bucket *bucket = &dma_entry_hash[idx];
225 struct dma_debug_entry *entry;
228 spin_lock_irqsave(&bucket->lock, flags);
230 list_for_each_entry(entry, &bucket->list, list) {
231 if (!dev || dev == entry->dev) {
233 "%s idx %d P=%Lx D=%Lx L=%Lx %s\n",
234 type2name[entry->type], idx,
235 (unsigned long long)entry->paddr,
236 entry->dev_addr, entry->size,
237 dir2name[entry->direction]);
241 spin_unlock_irqrestore(&bucket->lock, flags);
244 EXPORT_SYMBOL(debug_dma_dump_mappings);
247 * Wrapper function for adding an entry to the hash.
248 * This function takes care of locking itself.
250 static void add_dma_entry(struct dma_debug_entry *entry)
252 struct hash_bucket *bucket;
255 bucket = get_hash_bucket(entry, &flags);
256 hash_bucket_add(bucket, entry);
257 put_hash_bucket(bucket, &flags);
260 /* struct dma_entry allocator
262 * The next two functions implement the allocator for
263 * struct dma_debug_entries.
265 static struct dma_debug_entry *dma_entry_alloc(void)
267 struct dma_debug_entry *entry = NULL;
270 spin_lock_irqsave(&free_entries_lock, flags);
272 if (list_empty(&free_entries)) {
273 printk(KERN_ERR "DMA-API: debugging out of memory "
275 global_disable = true;
279 entry = list_entry(free_entries.next, struct dma_debug_entry, list);
280 list_del(&entry->list);
281 memset(entry, 0, sizeof(*entry));
283 #ifdef CONFIG_STACKTRACE
284 entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
285 entry->stacktrace.entries = entry->st_entries;
286 entry->stacktrace.skip = 2;
287 save_stack_trace(&entry->stacktrace);
289 num_free_entries -= 1;
290 if (num_free_entries < min_free_entries)
291 min_free_entries = num_free_entries;
294 spin_unlock_irqrestore(&free_entries_lock, flags);
299 static void dma_entry_free(struct dma_debug_entry *entry)
304 * add to beginning of the list - this way the entries are
305 * more likely cache hot when they are reallocated.
307 spin_lock_irqsave(&free_entries_lock, flags);
308 list_add(&entry->list, &free_entries);
309 num_free_entries += 1;
310 spin_unlock_irqrestore(&free_entries_lock, flags);
314 * DMA-API debugging init code
316 * The init code does two things:
317 * 1. Initialize core data structures
318 * 2. Preallocate a given number of dma_debug_entry structs
321 static int prealloc_memory(u32 num_entries)
323 struct dma_debug_entry *entry, *next_entry;
326 for (i = 0; i < num_entries; ++i) {
327 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
331 list_add_tail(&entry->list, &free_entries);
334 num_free_entries = num_entries;
335 min_free_entries = num_entries;
337 printk(KERN_INFO "DMA-API: preallocated %d debug entries\n",
344 list_for_each_entry_safe(entry, next_entry, &free_entries, list) {
345 list_del(&entry->list);
352 static int dma_debug_fs_init(void)
354 dma_debug_dent = debugfs_create_dir("dma-api", NULL);
355 if (!dma_debug_dent) {
356 printk(KERN_ERR "DMA-API: can not create debugfs directory\n");
360 global_disable_dent = debugfs_create_bool("disabled", 0444,
362 (u32 *)&global_disable);
363 if (!global_disable_dent)
366 error_count_dent = debugfs_create_u32("error_count", 0444,
367 dma_debug_dent, &error_count);
368 if (!error_count_dent)
371 show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
374 if (!show_all_errors_dent)
377 show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
380 if (!show_num_errors_dent)
383 num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
386 if (!num_free_entries_dent)
389 min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
392 if (!min_free_entries_dent)
398 debugfs_remove_recursive(dma_debug_dent);
405 * Let the architectures decide how many entries should be preallocated.
407 void dma_debug_init(u32 num_entries)
414 for (i = 0; i < HASH_SIZE; ++i) {
415 INIT_LIST_HEAD(&dma_entry_hash[i].list);
416 dma_entry_hash[i].lock = SPIN_LOCK_UNLOCKED;
419 if (dma_debug_fs_init() != 0) {
420 printk(KERN_ERR "DMA-API: error creating debugfs entries "
422 global_disable = true;
428 num_entries = req_entries;
430 if (prealloc_memory(num_entries) != 0) {
431 printk(KERN_ERR "DMA-API: debugging out of memory error "
433 global_disable = true;
438 printk(KERN_INFO "DMA-API: debugging enabled by kernel config\n");
441 static __init int dma_debug_cmdline(char *str)
446 if (strncmp(str, "off", 3) == 0) {
447 printk(KERN_INFO "DMA-API: debugging disabled on kernel "
449 global_disable = true;
455 static __init int dma_debug_entries_cmdline(char *str)
462 res = get_option(&str, &req_entries);
470 __setup("dma_debug=", dma_debug_cmdline);
471 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
473 static void check_unmap(struct dma_debug_entry *ref)
475 struct dma_debug_entry *entry;
476 struct hash_bucket *bucket;
479 if (dma_mapping_error(ref->dev, ref->dev_addr))
482 bucket = get_hash_bucket(ref, &flags);
483 entry = hash_bucket_find(bucket, ref);
486 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
487 "to free DMA memory it has not allocated "
488 "[device address=0x%016llx] [size=%llu bytes]\n",
489 ref->dev_addr, ref->size);
493 if (ref->size != entry->size) {
494 err_printk(ref->dev, entry, "DMA-API: device driver frees "
495 "DMA memory with different size "
496 "[device address=0x%016llx] [map size=%llu bytes] "
497 "[unmap size=%llu bytes]\n",
498 ref->dev_addr, entry->size, ref->size);
501 if (ref->type != entry->type) {
502 err_printk(ref->dev, entry, "DMA-API: device driver frees "
503 "DMA memory with wrong function "
504 "[device address=0x%016llx] [size=%llu bytes] "
505 "[mapped as %s] [unmapped as %s]\n",
506 ref->dev_addr, ref->size,
507 type2name[entry->type], type2name[ref->type]);
508 } else if ((entry->type == dma_debug_coherent) &&
509 (ref->paddr != entry->paddr)) {
510 err_printk(ref->dev, entry, "DMA-API: device driver frees "
511 "DMA memory with different CPU address "
512 "[device address=0x%016llx] [size=%llu bytes] "
513 "[cpu alloc address=%p] [cpu free address=%p]",
514 ref->dev_addr, ref->size,
515 (void *)entry->paddr, (void *)ref->paddr);
518 if (ref->sg_call_ents && ref->type == dma_debug_sg &&
519 ref->sg_call_ents != entry->sg_call_ents) {
520 err_printk(ref->dev, entry, "DMA-API: device driver frees "
521 "DMA sg list with different entry count "
522 "[map count=%d] [unmap count=%d]\n",
523 entry->sg_call_ents, ref->sg_call_ents);
527 * This may be no bug in reality - but most implementations of the
528 * DMA API don't handle this properly, so check for it here
530 if (ref->direction != entry->direction) {
531 err_printk(ref->dev, entry, "DMA-API: device driver frees "
532 "DMA memory with different direction "
533 "[device address=0x%016llx] [size=%llu bytes] "
534 "[mapped with %s] [unmapped with %s]\n",
535 ref->dev_addr, ref->size,
536 dir2name[entry->direction],
537 dir2name[ref->direction]);
540 hash_bucket_del(entry);
541 dma_entry_free(entry);
544 put_hash_bucket(bucket, &flags);
547 static void check_for_stack(struct device *dev, void *addr)
549 if (object_is_on_stack(addr))
550 err_printk(dev, NULL, "DMA-API: device driver maps memory from"
551 "stack [addr=%p]\n", addr);
554 static inline bool overlap(void *addr, u64 size, void *start, void *end)
556 void *addr2 = (char *)addr + size;
558 return ((addr >= start && addr < end) ||
559 (addr2 >= start && addr2 < end) ||
560 ((addr < start) && (addr2 >= end)));
563 static void check_for_illegal_area(struct device *dev, void *addr, u64 size)
565 if (overlap(addr, size, _text, _etext) ||
566 overlap(addr, size, __start_rodata, __end_rodata))
567 err_printk(dev, NULL, "DMA-API: device driver maps "
568 "memory from kernel text or rodata "
569 "[addr=%p] [size=%llu]\n", addr, size);
572 static void check_sync(struct device *dev, dma_addr_t addr,
573 u64 size, u64 offset, int direction, bool to_cpu)
575 struct dma_debug_entry ref = {
579 .direction = direction,
581 struct dma_debug_entry *entry;
582 struct hash_bucket *bucket;
585 bucket = get_hash_bucket(&ref, &flags);
587 entry = hash_bucket_find(bucket, &ref);
590 err_printk(dev, NULL, "DMA-API: device driver tries "
591 "to sync DMA memory it has not allocated "
592 "[device address=0x%016llx] [size=%llu bytes]\n",
597 if ((offset + size) > entry->size) {
598 err_printk(dev, entry, "DMA-API: device driver syncs"
599 " DMA memory outside allocated range "
600 "[device address=0x%016llx] "
601 "[allocation size=%llu bytes] [sync offset=%llu] "
602 "[sync size=%llu]\n", entry->dev_addr, entry->size,
606 if (direction != entry->direction) {
607 err_printk(dev, entry, "DMA-API: device driver syncs "
608 "DMA memory with different direction "
609 "[device address=0x%016llx] [size=%llu bytes] "
610 "[mapped with %s] [synced with %s]\n",
612 dir2name[entry->direction],
613 dir2name[direction]);
616 if (entry->direction == DMA_BIDIRECTIONAL)
619 if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
620 !(direction == DMA_TO_DEVICE))
621 err_printk(dev, entry, "DMA-API: device driver syncs "
622 "device read-only DMA memory for cpu "
623 "[device address=0x%016llx] [size=%llu bytes] "
624 "[mapped with %s] [synced with %s]\n",
626 dir2name[entry->direction],
627 dir2name[direction]);
629 if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
630 !(direction == DMA_FROM_DEVICE))
631 err_printk(dev, entry, "DMA-API: device driver syncs "
632 "device write-only DMA memory to device "
633 "[device address=0x%016llx] [size=%llu bytes] "
634 "[mapped with %s] [synced with %s]\n",
636 dir2name[entry->direction],
637 dir2name[direction]);
640 put_hash_bucket(bucket, &flags);
644 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
645 size_t size, int direction, dma_addr_t dma_addr,
648 struct dma_debug_entry *entry;
650 if (unlikely(global_disable))
653 if (unlikely(dma_mapping_error(dev, dma_addr)))
656 entry = dma_entry_alloc();
661 entry->type = dma_debug_page;
662 entry->paddr = page_to_phys(page) + offset;
663 entry->dev_addr = dma_addr;
665 entry->direction = direction;
668 void *addr = ((char *)page_address(page)) + offset;
670 entry->type = dma_debug_single;
671 check_for_stack(dev, addr);
672 check_for_illegal_area(dev, addr, size);
675 add_dma_entry(entry);
677 EXPORT_SYMBOL(debug_dma_map_page);
679 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
680 size_t size, int direction, bool map_single)
682 struct dma_debug_entry ref = {
683 .type = dma_debug_page,
687 .direction = direction,
690 if (unlikely(global_disable))
694 ref.type = dma_debug_single;
698 EXPORT_SYMBOL(debug_dma_unmap_page);
700 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
701 int nents, int mapped_ents, int direction)
703 struct dma_debug_entry *entry;
704 struct scatterlist *s;
707 if (unlikely(global_disable))
710 for_each_sg(sg, s, mapped_ents, i) {
711 entry = dma_entry_alloc();
715 entry->type = dma_debug_sg;
717 entry->paddr = sg_phys(s);
718 entry->size = s->length;
719 entry->dev_addr = s->dma_address;
720 entry->direction = direction;
721 entry->sg_call_ents = nents;
722 entry->sg_mapped_ents = mapped_ents;
724 check_for_stack(dev, sg_virt(s));
725 check_for_illegal_area(dev, sg_virt(s), s->length);
727 add_dma_entry(entry);
730 EXPORT_SYMBOL(debug_dma_map_sg);
732 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
735 struct dma_debug_entry *entry;
736 struct scatterlist *s;
737 int mapped_ents = 0, i;
740 if (unlikely(global_disable))
743 for_each_sg(sglist, s, nelems, i) {
745 struct dma_debug_entry ref = {
746 .type = dma_debug_sg,
749 .dev_addr = s->dma_address,
755 if (mapped_ents && i >= mapped_ents)
758 if (mapped_ents == 0) {
759 struct hash_bucket *bucket;
760 ref.sg_call_ents = nelems;
761 bucket = get_hash_bucket(&ref, &flags);
762 entry = hash_bucket_find(bucket, &ref);
764 mapped_ents = entry->sg_mapped_ents;
765 put_hash_bucket(bucket, &flags);
771 EXPORT_SYMBOL(debug_dma_unmap_sg);
773 void debug_dma_alloc_coherent(struct device *dev, size_t size,
774 dma_addr_t dma_addr, void *virt)
776 struct dma_debug_entry *entry;
778 if (unlikely(global_disable))
781 if (unlikely(virt == NULL))
784 entry = dma_entry_alloc();
788 entry->type = dma_debug_coherent;
790 entry->paddr = virt_to_phys(virt);
792 entry->dev_addr = dma_addr;
793 entry->direction = DMA_BIDIRECTIONAL;
795 add_dma_entry(entry);
797 EXPORT_SYMBOL(debug_dma_alloc_coherent);
799 void debug_dma_free_coherent(struct device *dev, size_t size,
800 void *virt, dma_addr_t addr)
802 struct dma_debug_entry ref = {
803 .type = dma_debug_coherent,
805 .paddr = virt_to_phys(virt),
808 .direction = DMA_BIDIRECTIONAL,
811 if (unlikely(global_disable))
816 EXPORT_SYMBOL(debug_dma_free_coherent);
818 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
819 size_t size, int direction)
821 if (unlikely(global_disable))
824 check_sync(dev, dma_handle, size, 0, direction, true);
826 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
828 void debug_dma_sync_single_for_device(struct device *dev,
829 dma_addr_t dma_handle, size_t size,
832 if (unlikely(global_disable))
835 check_sync(dev, dma_handle, size, 0, direction, false);
837 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
839 void debug_dma_sync_single_range_for_cpu(struct device *dev,
840 dma_addr_t dma_handle,
841 unsigned long offset, size_t size,
844 if (unlikely(global_disable))
847 check_sync(dev, dma_handle, size, offset, direction, true);
849 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
851 void debug_dma_sync_single_range_for_device(struct device *dev,
852 dma_addr_t dma_handle,
853 unsigned long offset,
854 size_t size, int direction)
856 if (unlikely(global_disable))
859 check_sync(dev, dma_handle, size, offset, direction, false);
861 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
863 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
864 int nelems, int direction)
866 struct scatterlist *s;
869 if (unlikely(global_disable))
872 for_each_sg(sg, s, nelems, i) {
873 check_sync(dev, s->dma_address, s->dma_length, 0,
877 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
879 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
880 int nelems, int direction)
882 struct scatterlist *s;
885 if (unlikely(global_disable))
888 for_each_sg(sg, s, nelems, i) {
889 check_sync(dev, s->dma_address, s->dma_length, 0,
893 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);