kernel/dma/debug.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2008 Advanced Micro Devices, Inc.
   4  *
   5  * Author: Joerg Roedel <[email protected]>
   6  */
   7
   8 #define pr_fmt(fmt)     "DMA-API: " fmt
   9
  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>
  27
  28 #include <asm/sections.h>
  29
  30 #define HASH_SIZE       16384ULL
  31 #define HASH_FN_SHIFT   13
  32 #define HASH_FN_MASK    (HASH_SIZE - 1)
  33
  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))
  37
  38 enum {
  39         dma_debug_single,
  40         dma_debug_sg,
  41         dma_debug_coherent,
  42         dma_debug_resource,
  43 };
  44
  45 enum map_err_types {
  46         MAP_ERR_CHECK_NOT_APPLICABLE,
  47         MAP_ERR_NOT_CHECKED,
  48         MAP_ERR_CHECKED,
  49 };
  50
  51 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
  52
  53 /**
  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
  66  */
  67 struct dma_debug_entry {
  68         struct list_head list;
  69         struct device    *dev;
  70         u64              dev_addr;
  71         u64              size;
  72         int              type;
  73         int              direction;
  74         int              sg_call_ents;
  75         int              sg_mapped_ents;
  76         unsigned long    pfn;
  77         size_t           offset;
  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];
  82 #endif
  83 } ____cacheline_aligned_in_smp;
  84
  85 typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
  86
  87 struct hash_bucket {
  88         struct list_head list;
  89         spinlock_t lock;
  90 };
  91
  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);
  98
  99 /* Global disable flag - will be set in case of an error */
 100 static bool global_disable __read_mostly;
 101
 102 /* Early initialization disable flag, set at the end of dma_debug_init */
 103 static bool dma_debug_initialized __read_mostly;
 104
 105 static inline bool dma_debug_disabled(void)
 106 {
 107         return global_disable || !dma_debug_initialized;
 108 }
 109
 110 /* Global error count */
 111 static u32 error_count;
 112
 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;
 117
 118 static u32 num_free_entries;
 119 static u32 min_free_entries;
 120 static u32 nr_total_entries;
 121
 122 /* number of preallocated entries requested by kernel cmdline */
 123 static u32 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 124
 125 /* per-driver filter related state */
 126
 127 #define NAME_MAX_LEN    64
 128
 129 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
 130 static struct device_driver *current_driver                    __read_mostly;
 131
 132 static DEFINE_RWLOCK(driver_name_lock);
 133
 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",
 138 };
 139
 140 static const char *type2name[5] = { "single", "page",
 141                                     "scather-gather", "coherent",
 142                                     "resource" };
 143
 144 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
 145                                    "DMA_FROM_DEVICE", "DMA_NONE" };
 146
 147 /*
 148  * The access to some variables in this macro is racy. We can't use atomic_t
 149  * here because all these variables are exported to debugfs. Some of them even
 150  * writeable. This is also the reason why a lock won't help much. But anyway,
 151  * the races are no big deal. Here is why:
 152  *
 153  *   error_count: the addition is racy, but the worst thing that can happen is
 154  *                that we don't count some errors
 155  *   show_num_errors: the subtraction is racy. Also no big deal because in
 156  *                    worst case this will result in one warning more in the
 157  *                    system log than the user configured. This variable is
 158  *                    writeable via debugfs.
 159  */
 160 static inline void dump_entry_trace(struct dma_debug_entry *entry)
 161 {
 162 #ifdef CONFIG_STACKTRACE
 163         if (entry) {
 164                 pr_warn("Mapped at:\n");
 165                 stack_trace_print(entry->stack_entries, entry->stack_len, 0);
 166         }
 167 #endif
 168 }
 169
 170 static bool driver_filter(struct device *dev)
 171 {
 172         struct device_driver *drv;
 173         unsigned long flags;
 174         bool ret;
 175
 176         /* driver filter off */
 177         if (likely(!current_driver_name[0]))
 178                 return true;
 179
 180         /* driver filter on and initialized */
 181         if (current_driver && dev && dev->driver == current_driver)
 182                 return true;
 183
 184         /* driver filter on, but we can't filter on a NULL device... */
 185         if (!dev)
 186                 return false;
 187
 188         if (current_driver || !current_driver_name[0])
 189                 return false;
 190
 191         /* driver filter on but not yet initialized */
 192         drv = dev->driver;
 193         if (!drv)
 194                 return false;
 195
 196         /* lock to protect against change of current_driver_name */
 197         read_lock_irqsave(&driver_name_lock, flags);
 198
 199         ret = false;
 200         if (drv->name &&
 201             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 202                 current_driver = drv;
 203                 ret = true;
 204         }
 205
 206         read_unlock_irqrestore(&driver_name_lock, flags);
 207
 208         return ret;
 209 }
 210
 211 #define err_printk(dev, entry, format, arg...) do {                     \
 212                 error_count += 1;                                       \
 213                 if (driver_filter(dev) &&                               \
 214                     (show_all_errors || show_num_errors > 0)) {         \
 215                         WARN(1, pr_fmt("%s %s: ") format,               \
 216                              dev ? dev_driver_string(dev) : "NULL",     \
 217                              dev ? dev_name(dev) : "NULL", ## arg);     \
 218                         dump_entry_trace(entry);                        \
 219                 }                                                       \
 220                 if (!show_all_errors && show_num_errors > 0)            \
 221                         show_num_errors -= 1;                           \
 222         } while (0);
 223
 224 /*
 225  * Hash related functions
 226  *
 227  * Every DMA-API request is saved into a struct dma_debug_entry. To
 228  * have quick access to these structs they are stored into a hash.
 229  */
 230 static int hash_fn(struct dma_debug_entry *entry)
 231 {
 232         /*
 233          * Hash function is based on the dma address.
 234          * We use bits 20-27 here as the index into the hash
 235          */
 236         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 237 }
 238
 239 /*
 240  * Request exclusive access to a hash bucket for a given dma_debug_entry.
 241  */
 242 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 243                                            unsigned long *flags)
 244         __acquires(&dma_entry_hash[idx].lock)
 245 {
 246         int idx = hash_fn(entry);
 247         unsigned long __flags;
 248
 249         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 250         *flags = __flags;
 251         return &dma_entry_hash[idx];
 252 }
 253
 254 /*
 255  * Give up exclusive access to the hash bucket
 256  */
 257 static void put_hash_bucket(struct hash_bucket *bucket,
 258                             unsigned long flags)
 259         __releases(&bucket->lock)
 260 {
 261         spin_unlock_irqrestore(&bucket->lock, flags);
 262 }
 263
 264 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
 265 {
 266         return ((a->dev_addr == b->dev_addr) &&
 267                 (a->dev == b->dev)) ? true : false;
 268 }
 269
 270 static bool containing_match(struct dma_debug_entry *a,
 271                              struct dma_debug_entry *b)
 272 {
 273         if (a->dev != b->dev)
 274                 return false;
 275
 276         if ((b->dev_addr <= a->dev_addr) &&
 277             ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
 278                 return true;
 279
 280         return false;
 281 }
 282
 283 /*
 284  * Search a given entry in the hash bucket list
 285  */
 286 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
 287                                                   struct dma_debug_entry *ref,
 288                                                   match_fn match)
 289 {
 290         struct dma_debug_entry *entry, *ret = NULL;
 291         int matches = 0, match_lvl, last_lvl = -1;
 292
 293         list_for_each_entry(entry, &bucket->list, list) {
 294                 if (!match(ref, entry))
 295                         continue;
 296
 297                 /*
 298                  * Some drivers map the same physical address multiple
 299                  * times. Without a hardware IOMMU this results in the
 300                  * same device addresses being put into the dma-debug
 301                  * hash multiple times too. This can result in false
 302                  * positives being reported. Therefore we implement a
 303                  * best-fit algorithm here which returns the entry from
 304                  * the hash which fits best to the reference value
 305                  * instead of the first-fit.
 306                  */
 307                 matches += 1;
 308                 match_lvl = 0;
 309                 entry->size         == ref->size         ? ++match_lvl : 0;
 310                 entry->type         == ref->type         ? ++match_lvl : 0;
 311                 entry->direction    == ref->direction    ? ++match_lvl : 0;
 312                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
 313
 314                 if (match_lvl == 4) {
 315                         /* perfect-fit - return the result */
 316                         return entry;
 317                 } else if (match_lvl > last_lvl) {
 318                         /*
 319                          * We found an entry that fits better then the
 320                          * previous one or it is the 1st match.
 321                          */
 322                         last_lvl = match_lvl;
 323                         ret      = entry;
 324                 }
 325         }
 326
 327         /*
 328          * If we have multiple matches but no perfect-fit, just return
 329          * NULL.
 330          */
 331         ret = (matches == 1) ? ret : NULL;
 332
 333         return ret;
 334 }
 335
 336 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
 337                                                  struct dma_debug_entry *ref)
 338 {
 339         return __hash_bucket_find(bucket, ref, exact_match);
 340 }
 341
 342 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 343                                                    struct dma_debug_entry *ref,
 344                                                    unsigned long *flags)
 345 {
 346
 347         unsigned int max_range = dma_get_max_seg_size(ref->dev);
 348         struct dma_debug_entry *entry, index = *ref;
 349         unsigned int range = 0;
 350
 351         while (range <= max_range) {
 352                 entry = __hash_bucket_find(*bucket, ref, containing_match);
 353
 354                 if (entry)
 355                         return entry;
 356
 357                 /*
 358                  * Nothing found, go back a hash bucket
 359                  */
 360                 put_hash_bucket(*bucket, *flags);
 361                 range          += (1 << HASH_FN_SHIFT);
 362                 index.dev_addr -= (1 << HASH_FN_SHIFT);
 363                 *bucket = get_hash_bucket(&index, flags);
 364         }
 365
 366         return NULL;
 367 }
 368
 369 /*
 370  * Add an entry to a hash bucket
 371  */
 372 static void hash_bucket_add(struct hash_bucket *bucket,
 373                             struct dma_debug_entry *entry)
 374 {
 375         list_add_tail(&entry->list, &bucket->list);
 376 }
 377
 378 /*
 379  * Remove entry from a hash bucket list
 380  */
 381 static void hash_bucket_del(struct dma_debug_entry *entry)
 382 {
 383         list_del(&entry->list);
 384 }
 385
 386 static unsigned long long phys_addr(struct dma_debug_entry *entry)
 387 {
 388         if (entry->type == dma_debug_resource)
 389                 return __pfn_to_phys(entry->pfn) + entry->offset;
 390
 391         return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
 392 }
 393
 394 /*
 395  * Dump mapping entries for debugging purposes
 396  */
 397 void debug_dma_dump_mappings(struct device *dev)
 398 {
 399         int idx;
 400
 401         for (idx = 0; idx < HASH_SIZE; idx++) {
 402                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 403                 struct dma_debug_entry *entry;
 404                 unsigned long flags;
 405
 406                 spin_lock_irqsave(&bucket->lock, flags);
 407
 408                 list_for_each_entry(entry, &bucket->list, list) {
 409                         if (!dev || dev == entry->dev) {
 410                                 dev_info(entry->dev,
 411                                          "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
 412                                          type2name[entry->type], idx,
 413                                          phys_addr(entry), entry->pfn,
 414                                          entry->dev_addr, entry->size,
 415                                          dir2name[entry->direction],
 416                                          maperr2str[entry->map_err_type]);
 417                         }
 418                 }
 419
 420                 spin_unlock_irqrestore(&bucket->lock, flags);
 421                 cond_resched();
 422         }
 423 }
 424
 425 /*
 426  * For each mapping (initial cacheline in the case of
 427  * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
 428  * scatterlist, or the cacheline specified in dma_map_single) insert
 429  * into this tree using the cacheline as the key. At
 430  * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
 431  * the entry already exists at insertion time add a tag as a reference
 432  * count for the overlapping mappings.  For now, the overlap tracking
 433  * just ensures that 'unmaps' balance 'maps' before marking the
 434  * cacheline idle, but we should also be flagging overlaps as an API
 435  * violation.
 436  *
 437  * Memory usage is mostly constrained by the maximum number of available
 438  * dma-debug entries in that we need a free dma_debug_entry before
 439  * inserting into the tree.  In the case of dma_map_page and
 440  * dma_alloc_coherent there is only one dma_debug_entry and one
 441  * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
 442  * other hand, consumes a single dma_debug_entry, but inserts 'nents'
 443  * entries into the tree.
 444  *
 445  * At any time debug_dma_assert_idle() can be called to trigger a
 446  * warning if any cachelines in the given page are in the active set.
 447  */
 448 static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
 449 static DEFINE_SPINLOCK(radix_lock);
 450 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
 451 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
 452 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
 453
 454 static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
 455 {
 456         return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
 457                 (entry->offset >> L1_CACHE_SHIFT);
 458 }
 459
 460 static int active_cacheline_read_overlap(phys_addr_t cln)
 461 {
 462         int overlap = 0, i;
 463
 464         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 465                 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
 466                         overlap |= 1 << i;
 467         return overlap;
 468 }
 469
 470 static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
 471 {
 472         int i;
 473
 474         if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
 475                 return overlap;
 476
 477         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 478                 if (overlap & 1 << i)
 479                         radix_tree_tag_set(&dma_active_cacheline, cln, i);
 480                 else
 481                         radix_tree_tag_clear(&dma_active_cacheline, cln, i);
 482
 483         return overlap;
 484 }
 485
 486 static void active_cacheline_inc_overlap(phys_addr_t cln)
 487 {
 488         int overlap = active_cacheline_read_overlap(cln);
 489
 490         overlap = active_cacheline_set_overlap(cln, ++overlap);
 491
 492         /* If we overflowed the overlap counter then we're potentially
 493          * leaking dma-mappings.  Otherwise, if maps and unmaps are
 494          * balanced then this overflow may cause false negatives in
 495          * debug_dma_assert_idle() as the cacheline may be marked idle
 496          * prematurely.
 497          */
 498         WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
 499                   pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
 500                   ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 501 }
 502
 503 static int active_cacheline_dec_overlap(phys_addr_t cln)
 504 {
 505         int overlap = active_cacheline_read_overlap(cln);
 506
 507         return active_cacheline_set_overlap(cln, --overlap);
 508 }
 509
 510 static int active_cacheline_insert(struct dma_debug_entry *entry)
 511 {
 512         phys_addr_t cln = to_cacheline_number(entry);
 513         unsigned long flags;
 514         int rc;
 515
 516         /* If the device is not writing memory then we don't have any
 517          * concerns about the cpu consuming stale data.  This mitigates
 518          * legitimate usages of overlapping mappings.
 519          */
 520         if (entry->direction == DMA_TO_DEVICE)
 521                 return 0;
 522
 523         spin_lock_irqsave(&radix_lock, flags);
 524         rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
 525         if (rc == -EEXIST)
 526                 active_cacheline_inc_overlap(cln);
 527         spin_unlock_irqrestore(&radix_lock, flags);
 528
 529         return rc;
 530 }
 531
 532 static void active_cacheline_remove(struct dma_debug_entry *entry)
 533 {
 534         phys_addr_t cln = to_cacheline_number(entry);
 535         unsigned long flags;
 536
 537         /* ...mirror the insert case */
 538         if (entry->direction == DMA_TO_DEVICE)
 539                 return;
 540
 541         spin_lock_irqsave(&radix_lock, flags);
 542         /* since we are counting overlaps the final put of the
 543          * cacheline will occur when the overlap count is 0.
 544          * active_cacheline_dec_overlap() returns -1 in that case
 545          */
 546         if (active_cacheline_dec_overlap(cln) < 0)
 547                 radix_tree_delete(&dma_active_cacheline, cln);
 548         spin_unlock_irqrestore(&radix_lock, flags);
 549 }
 550
 551 /**
 552  * debug_dma_assert_idle() - assert that a page is not undergoing dma
 553  * @page: page to lookup in the dma_active_cacheline tree
 554  *
 555  * Place a call to this routine in cases where the cpu touching the page
 556  * before the dma completes (page is dma_unmapped) will lead to data
 557  * corruption.
 558  */
 559 void debug_dma_assert_idle(struct page *page)
 560 {
 561         static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
 562         struct dma_debug_entry *entry = NULL;
 563         void **results = (void **) &ents;
 564         unsigned int nents, i;
 565         unsigned long flags;
 566         phys_addr_t cln;
 567
 568         if (dma_debug_disabled())
 569                 return;
 570
 571         if (!page)
 572                 return;
 573
 574         cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
 575         spin_lock_irqsave(&radix_lock, flags);
 576         nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
 577                                        CACHELINES_PER_PAGE);
 578         for (i = 0; i < nents; i++) {
 579                 phys_addr_t ent_cln = to_cacheline_number(ents[i]);
 580
 581                 if (ent_cln == cln) {
 582                         entry = ents[i];
 583                         break;
 584                 } else if (ent_cln >= cln + CACHELINES_PER_PAGE)
 585                         break;
 586         }
 587         spin_unlock_irqrestore(&radix_lock, flags);
 588
 589         if (!entry)
 590                 return;
 591
 592         cln = to_cacheline_number(entry);
 593         err_printk(entry->dev, entry,
 594                    "cpu touching an active dma mapped cacheline [cln=%pa]\n",
 595                    &cln);
 596 }
 597
 598 /*
 599  * Wrapper function for adding an entry to the hash.
 600  * This function takes care of locking itself.
 601  */
 602 static void add_dma_entry(struct dma_debug_entry *entry)
 603 {
 604         struct hash_bucket *bucket;
 605         unsigned long flags;
 606         int rc;
 607
 608         bucket = get_hash_bucket(entry, &flags);
 609         hash_bucket_add(bucket, entry);
 610         put_hash_bucket(bucket, flags);
 611
 612         rc = active_cacheline_insert(entry);
 613         if (rc == -ENOMEM) {
 614                 pr_err("cacheline tracking ENOMEM, dma-debug disabled\n");
 615                 global_disable = true;
 616         }
 617
 618         /* TODO: report -EEXIST errors here as overlapping mappings are
 619          * not supported by the DMA API
 620          */
 621 }
 622
 623 static int dma_debug_create_entries(gfp_t gfp)
 624 {
 625         struct dma_debug_entry *entry;
 626         int i;
 627
 628         entry = (void *)get_zeroed_page(gfp);
 629         if (!entry)
 630                 return -ENOMEM;
 631
 632         for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
 633                 list_add_tail(&entry[i].list, &free_entries);
 634
 635         num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 636         nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 637
 638         return 0;
 639 }
 640
 641 static struct dma_debug_entry *__dma_entry_alloc(void)
 642 {
 643         struct dma_debug_entry *entry;
 644
 645         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 646         list_del(&entry->list);
 647         memset(entry, 0, sizeof(*entry));
 648
 649         num_free_entries -= 1;
 650         if (num_free_entries < min_free_entries)
 651                 min_free_entries = num_free_entries;
 652
 653         return entry;
 654 }
 655
 656 void __dma_entry_alloc_check_leak(void)
 657 {
 658         u32 tmp = nr_total_entries % nr_prealloc_entries;
 659
 660         /* Shout each time we tick over some multiple of the initial pool */
 661         if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
 662                 pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
 663                         nr_total_entries,
 664                         (nr_total_entries / nr_prealloc_entries));
 665         }
 666 }
 667
 668 /* struct dma_entry allocator
 669  *
 670  * The next two functions implement the allocator for
 671  * struct dma_debug_entries.
 672  */
 673 static struct dma_debug_entry *dma_entry_alloc(void)
 674 {
 675         struct dma_debug_entry *entry;
 676         unsigned long flags;
 677
 678         spin_lock_irqsave(&free_entries_lock, flags);
 679         if (num_free_entries == 0) {
 680                 if (dma_debug_create_entries(GFP_ATOMIC)) {
 681                         global_disable = true;
 682                         spin_unlock_irqrestore(&free_entries_lock, flags);
 683                         pr_err("debugging out of memory - disabling\n");
 684                         return NULL;
 685                 }
 686                 __dma_entry_alloc_check_leak();
 687         }
 688
 689         entry = __dma_entry_alloc();
 690
 691         spin_unlock_irqrestore(&free_entries_lock, flags);
 692
 693 #ifdef CONFIG_STACKTRACE
 694         entry->stack_len = stack_trace_save(entry->stack_entries,
 695                                             ARRAY_SIZE(entry->stack_entries),
 696                                             1);
 697 #endif
 698         return entry;
 699 }
 700
 701 static void dma_entry_free(struct dma_debug_entry *entry)
 702 {
 703         unsigned long flags;
 704
 705         active_cacheline_remove(entry);
 706
 707         /*
 708          * add to beginning of the list - this way the entries are
 709          * more likely cache hot when they are reallocated.
 710          */
 711         spin_lock_irqsave(&free_entries_lock, flags);
 712         list_add(&entry->list, &free_entries);
 713         num_free_entries += 1;
 714         spin_unlock_irqrestore(&free_entries_lock, flags);
 715 }
 716
 717 /*
 718  * DMA-API debugging init code
 719  *
 720  * The init code does two things:
 721  *   1. Initialize core data structures
 722  *   2. Preallocate a given number of dma_debug_entry structs
 723  */
 724
 725 static ssize_t filter_read(struct file *file, char __user *user_buf,
 726                            size_t count, loff_t *ppos)
 727 {
 728         char buf[NAME_MAX_LEN + 1];
 729         unsigned long flags;
 730         int len;
 731
 732         if (!current_driver_name[0])
 733                 return 0;
 734
 735         /*
 736          * We can't copy to userspace directly because current_driver_name can
 737          * only be read under the driver_name_lock with irqs disabled. So
 738          * create a temporary copy first.
 739          */
 740         read_lock_irqsave(&driver_name_lock, flags);
 741         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 742         read_unlock_irqrestore(&driver_name_lock, flags);
 743
 744         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 745 }
 746
 747 static ssize_t filter_write(struct file *file, const char __user *userbuf,
 748                             size_t count, loff_t *ppos)
 749 {
 750         char buf[NAME_MAX_LEN];
 751         unsigned long flags;
 752         size_t len;
 753         int i;
 754
 755         /*
 756          * We can't copy from userspace directly. Access to
 757          * current_driver_name is protected with a write_lock with irqs
 758          * disabled. Since copy_from_user can fault and may sleep we
 759          * need to copy to temporary buffer first
 760          */
 761         len = min(count, (size_t)(NAME_MAX_LEN - 1));
 762         if (copy_from_user(buf, userbuf, len))
 763                 return -EFAULT;
 764
 765         buf[len] = 0;
 766
 767         write_lock_irqsave(&driver_name_lock, flags);
 768
 769         /*
 770          * Now handle the string we got from userspace very carefully.
 771          * The rules are:
 772          *         - only use the first token we got
 773          *         - token delimiter is everything looking like a space
 774          *           character (' ', '\n', '\t' ...)
 775          *
 776          */
 777         if (!isalnum(buf[0])) {
 778                 /*
 779                  * If the first character userspace gave us is not
 780                  * alphanumerical then assume the filter should be
 781                  * switched off.
 782                  */
 783                 if (current_driver_name[0])
 784                         pr_info("switching off dma-debug driver filter\n");
 785                 current_driver_name[0] = 0;
 786                 current_driver = NULL;
 787                 goto out_unlock;
 788         }
 789
 790         /*
 791          * Now parse out the first token and use it as the name for the
 792          * driver to filter for.
 793          */
 794         for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 795                 current_driver_name[i] = buf[i];
 796                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 797                         break;
 798         }
 799         current_driver_name[i] = 0;
 800         current_driver = NULL;
 801
 802         pr_info("enable driver filter for driver [%s]\n",
 803                 current_driver_name);
 804
 805 out_unlock:
 806         write_unlock_irqrestore(&driver_name_lock, flags);
 807
 808         return count;
 809 }
 810
 811 static const struct file_operations filter_fops = {
 812         .read  = filter_read,
 813         .write = filter_write,
 814         .llseek = default_llseek,
 815 };
 816
 817 static int dump_show(struct seq_file *seq, void *v)
 818 {
 819         int idx;
 820
 821         for (idx = 0; idx < HASH_SIZE; idx++) {
 822                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 823                 struct dma_debug_entry *entry;
 824                 unsigned long flags;
 825
 826                 spin_lock_irqsave(&bucket->lock, flags);
 827                 list_for_each_entry(entry, &bucket->list, list) {
 828                         seq_printf(seq,
 829                                    "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx %s %s\n",
 830                                    dev_name(entry->dev),
 831                                    dev_driver_string(entry->dev),
 832                                    type2name[entry->type], idx,
 833                                    phys_addr(entry), entry->pfn,
 834                                    entry->dev_addr, entry->size,
 835                                    dir2name[entry->direction],
 836                                    maperr2str[entry->map_err_type]);
 837                 }
 838                 spin_unlock_irqrestore(&bucket->lock, flags);
 839         }
 840         return 0;
 841 }
 842 DEFINE_SHOW_ATTRIBUTE(dump);
 843
 844 static void dma_debug_fs_init(void)
 845 {
 846         struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
 847
 848         debugfs_create_bool("disabled", 0444, dentry, &global_disable);
 849         debugfs_create_u32("error_count", 0444, dentry, &error_count);
 850         debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
 851         debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
 852         debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
 853         debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
 854         debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
 855         debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
 856         debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
 857 }
 858
 859 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 860 {
 861         struct dma_debug_entry *entry;
 862         unsigned long flags;
 863         int count = 0, i;
 864
 865         for (i = 0; i < HASH_SIZE; ++i) {
 866                 spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
 867                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 868                         if (entry->dev == dev) {
 869                                 count += 1;
 870                                 *out_entry = entry;
 871                         }
 872                 }
 873                 spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
 874         }
 875
 876         return count;
 877 }
 878
 879 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 880 {
 881         struct device *dev = data;
 882         struct dma_debug_entry *uninitialized_var(entry);
 883         int count;
 884
 885         if (dma_debug_disabled())
 886                 return 0;
 887
 888         switch (action) {
 889         case BUS_NOTIFY_UNBOUND_DRIVER:
 890                 count = device_dma_allocations(dev, &entry);
 891                 if (count == 0)
 892                         break;
 893                 err_printk(dev, entry, "device driver has pending "
 894                                 "DMA allocations while released from device "
 895                                 "[count=%d]\n"
 896                                 "One of leaked entries details: "
 897                                 "[device address=0x%016llx] [size=%llu bytes] "
 898                                 "[mapped with %s] [mapped as %s]\n",
 899                         count, entry->dev_addr, entry->size,
 900                         dir2name[entry->direction], type2name[entry->type]);
 901                 break;
 902         default:
 903                 break;
 904         }
 905
 906         return 0;
 907 }
 908
 909 void dma_debug_add_bus(struct bus_type *bus)
 910 {
 911         struct notifier_block *nb;
 912
 913         if (dma_debug_disabled())
 914                 return;
 915
 916         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 917         if (nb == NULL) {
 918                 pr_err("dma_debug_add_bus: out of memory\n");
 919                 return;
 920         }
 921
 922         nb->notifier_call = dma_debug_device_change;
 923
 924         bus_register_notifier(bus, nb);
 925 }
 926
 927 static int dma_debug_init(void)
 928 {
 929         int i, nr_pages;
 930
 931         /* Do not use dma_debug_initialized here, since we really want to be
 932          * called to set dma_debug_initialized
 933          */
 934         if (global_disable)
 935                 return 0;
 936
 937         for (i = 0; i < HASH_SIZE; ++i) {
 938                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
 939                 spin_lock_init(&dma_entry_hash[i].lock);
 940         }
 941
 942         dma_debug_fs_init();
 943
 944         nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
 945         for (i = 0; i < nr_pages; ++i)
 946                 dma_debug_create_entries(GFP_KERNEL);
 947         if (num_free_entries >= nr_prealloc_entries) {
 948                 pr_info("preallocated %d debug entries\n", nr_total_entries);
 949         } else if (num_free_entries > 0) {
 950                 pr_warn("%d debug entries requested but only %d allocated\n",
 951                         nr_prealloc_entries, nr_total_entries);
 952         } else {
 953                 pr_err("debugging out of memory error - disabled\n");
 954                 global_disable = true;
 955
 956                 return 0;
 957         }
 958         min_free_entries = num_free_entries;
 959
 960         dma_debug_initialized = true;
 961
 962         pr_info("debugging enabled by kernel config\n");
 963         return 0;
 964 }
 965 core_initcall(dma_debug_init);
 966
 967 static __init int dma_debug_cmdline(char *str)
 968 {
 969         if (!str)
 970                 return -EINVAL;
 971
 972         if (strncmp(str, "off", 3) == 0) {
 973                 pr_info("debugging disabled on kernel command line\n");
 974                 global_disable = true;
 975         }
 976
 977         return 0;
 978 }
 979
 980 static __init int dma_debug_entries_cmdline(char *str)
 981 {
 982         if (!str)
 983                 return -EINVAL;
 984         if (!get_option(&str, &nr_prealloc_entries))
 985                 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 986         return 0;
 987 }
 988
 989 __setup("dma_debug=", dma_debug_cmdline);
 990 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
 991
 992 static void check_unmap(struct dma_debug_entry *ref)
 993 {
 994         struct dma_debug_entry *entry;
 995         struct hash_bucket *bucket;
 996         unsigned long flags;
 997
 998         bucket = get_hash_bucket(ref, &flags);
 999         entry = bucket_find_exact(bucket, ref);
1000
1001         if (!entry) {
1002                 /* must drop lock before calling dma_mapping_error */
1003                 put_hash_bucket(bucket, flags);
1004
1005                 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1006                         err_printk(ref->dev, NULL,
1007                                    "device driver tries to free an "
1008                                    "invalid DMA memory address\n");
1009                 } else {
1010                         err_printk(ref->dev, NULL,
1011                                    "device driver tries to free DMA "
1012                                    "memory it has not allocated [device "
1013                                    "address=0x%016llx] [size=%llu bytes]\n",
1014                                    ref->dev_addr, ref->size);
1015                 }
1016                 return;
1017         }
1018
1019         if (ref->size != entry->size) {
1020                 err_printk(ref->dev, entry, "device driver frees "
1021                            "DMA memory with different size "
1022                            "[device address=0x%016llx] [map size=%llu bytes] "
1023                            "[unmap size=%llu bytes]\n",
1024                            ref->dev_addr, entry->size, ref->size);
1025         }
1026
1027         if (ref->type != entry->type) {
1028                 err_printk(ref->dev, entry, "device driver frees "
1029                            "DMA memory with wrong function "
1030                            "[device address=0x%016llx] [size=%llu bytes] "
1031                            "[mapped as %s] [unmapped as %s]\n",
1032                            ref->dev_addr, ref->size,
1033                            type2name[entry->type], type2name[ref->type]);
1034         } else if ((entry->type == dma_debug_coherent) &&
1035                    (phys_addr(ref) != phys_addr(entry))) {
1036                 err_printk(ref->dev, entry, "device driver frees "
1037                            "DMA memory with different CPU address "
1038                            "[device address=0x%016llx] [size=%llu bytes] "
1039                            "[cpu alloc address=0x%016llx] "
1040                            "[cpu free address=0x%016llx]",
1041                            ref->dev_addr, ref->size,
1042                            phys_addr(entry),
1043                            phys_addr(ref));
1044         }
1045
1046         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1047             ref->sg_call_ents != entry->sg_call_ents) {
1048                 err_printk(ref->dev, entry, "device driver frees "
1049                            "DMA sg list with different entry count "
1050                            "[map count=%d] [unmap count=%d]\n",
1051                            entry->sg_call_ents, ref->sg_call_ents);
1052         }
1053
1054         /*
1055          * This may be no bug in reality - but most implementations of the
1056          * DMA API don't handle this properly, so check for it here
1057          */
1058         if (ref->direction != entry->direction) {
1059                 err_printk(ref->dev, entry, "device driver frees "
1060                            "DMA memory with different direction "
1061                            "[device address=0x%016llx] [size=%llu bytes] "
1062                            "[mapped with %s] [unmapped with %s]\n",
1063                            ref->dev_addr, ref->size,
1064                            dir2name[entry->direction],
1065                            dir2name[ref->direction]);
1066         }
1067
1068         /*
1069          * Drivers should use dma_mapping_error() to check the returned
1070          * addresses of dma_map_single() and dma_map_page().
1071          * If not, print this warning message. See Documentation/DMA-API.txt.
1072          */
1073         if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1074                 err_printk(ref->dev, entry,
1075                            "device driver failed to check map error"
1076                            "[device address=0x%016llx] [size=%llu bytes] "
1077                            "[mapped as %s]",
1078                            ref->dev_addr, ref->size,
1079                            type2name[entry->type]);
1080         }
1081
1082         hash_bucket_del(entry);
1083         dma_entry_free(entry);
1084
1085         put_hash_bucket(bucket, flags);
1086 }
1087
1088 static void check_for_stack(struct device *dev,
1089                             struct page *page, size_t offset)
1090 {
1091         void *addr;
1092         struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1093
1094         if (!stack_vm_area) {
1095                 /* Stack is direct-mapped. */
1096                 if (PageHighMem(page))
1097                         return;
1098                 addr = page_address(page) + offset;
1099                 if (object_is_on_stack(addr))
1100                         err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1101         } else {
1102                 /* Stack is vmalloced. */
1103                 int i;
1104
1105                 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1106                         if (page != stack_vm_area->pages[i])
1107                                 continue;
1108
1109                         addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1110                         err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1111                         break;
1112                 }
1113         }
1114 }
1115
1116 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1117 {
1118         unsigned long a1 = (unsigned long)addr;
1119         unsigned long b1 = a1 + len;
1120         unsigned long a2 = (unsigned long)start;
1121         unsigned long b2 = (unsigned long)end;
1122
1123         return !(b1 <= a2 || a1 >= b2);
1124 }
1125
1126 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1127 {
1128         if (overlap(addr, len, _stext, _etext) ||
1129             overlap(addr, len, __start_rodata, __end_rodata))
1130                 err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1131 }
1132
1133 static void check_sync(struct device *dev,
1134                        struct dma_debug_entry *ref,
1135                        bool to_cpu)
1136 {
1137         struct dma_debug_entry *entry;
1138         struct hash_bucket *bucket;
1139         unsigned long flags;
1140
1141         bucket = get_hash_bucket(ref, &flags);
1142
1143         entry = bucket_find_contain(&bucket, ref, &flags);
1144
1145         if (!entry) {
1146                 err_printk(dev, NULL, "device driver tries "
1147                                 "to sync DMA memory it has not allocated "
1148                                 "[device address=0x%016llx] [size=%llu bytes]\n",
1149                                 (unsigned long long)ref->dev_addr, ref->size);
1150                 goto out;
1151         }
1152
1153         if (ref->size > entry->size) {
1154                 err_printk(dev, entry, "device driver syncs"
1155                                 " DMA memory outside allocated range "
1156                                 "[device address=0x%016llx] "
1157                                 "[allocation size=%llu bytes] "
1158                                 "[sync offset+size=%llu]\n",
1159                                 entry->dev_addr, entry->size,
1160                                 ref->size);
1161         }
1162
1163         if (entry->direction == DMA_BIDIRECTIONAL)
1164                 goto out;
1165
1166         if (ref->direction != entry->direction) {
1167                 err_printk(dev, entry, "device driver syncs "
1168                                 "DMA memory with different direction "
1169                                 "[device address=0x%016llx] [size=%llu bytes] "
1170                                 "[mapped with %s] [synced with %s]\n",
1171                                 (unsigned long long)ref->dev_addr, entry->size,
1172                                 dir2name[entry->direction],
1173                                 dir2name[ref->direction]);
1174         }
1175
1176         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1177                       !(ref->direction == DMA_TO_DEVICE))
1178                 err_printk(dev, entry, "device driver syncs "
1179                                 "device read-only DMA memory for cpu "
1180                                 "[device address=0x%016llx] [size=%llu bytes] "
1181                                 "[mapped with %s] [synced with %s]\n",
1182                                 (unsigned long long)ref->dev_addr, entry->size,
1183                                 dir2name[entry->direction],
1184                                 dir2name[ref->direction]);
1185
1186         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1187                        !(ref->direction == DMA_FROM_DEVICE))
1188                 err_printk(dev, entry, "device driver syncs "
1189                                 "device write-only DMA memory to device "
1190                                 "[device address=0x%016llx] [size=%llu bytes] "
1191                                 "[mapped with %s] [synced with %s]\n",
1192                                 (unsigned long long)ref->dev_addr, entry->size,
1193                                 dir2name[entry->direction],
1194                                 dir2name[ref->direction]);
1195
1196         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1197             ref->sg_call_ents != entry->sg_call_ents) {
1198                 err_printk(ref->dev, entry, "device driver syncs "
1199                            "DMA sg list with different entry count "
1200                            "[map count=%d] [sync count=%d]\n",
1201                            entry->sg_call_ents, ref->sg_call_ents);
1202         }
1203
1204 out:
1205         put_hash_bucket(bucket, flags);
1206 }
1207
1208 static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1209 {
1210 #ifdef CONFIG_DMA_API_DEBUG_SG
1211         unsigned int max_seg = dma_get_max_seg_size(dev);
1212         u64 start, end, boundary = dma_get_seg_boundary(dev);
1213
1214         /*
1215          * Either the driver forgot to set dma_parms appropriately, or
1216          * whoever generated the list forgot to check them.
1217          */
1218         if (sg->length > max_seg)
1219                 err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1220                            sg->length, max_seg);
1221         /*
1222          * In some cases this could potentially be the DMA API
1223          * implementation's fault, but it would usually imply that
1224          * the scatterlist was built inappropriately to begin with.
1225          */
1226         start = sg_dma_address(sg);
1227         end = start + sg_dma_len(sg) - 1;
1228         if ((start ^ end) & ~boundary)
1229                 err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1230                            start, end, boundary);
1231 #endif
1232 }
1233
1234 void debug_dma_map_single(struct device *dev, const void *addr,
1235                             unsigned long len)
1236 {
1237         if (unlikely(dma_debug_disabled()))
1238                 return;
1239
1240         if (!virt_addr_valid(addr))
1241                 err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1242                            addr, len);
1243
1244         if (is_vmalloc_addr(addr))
1245                 err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1246                            addr, len);
1247 }
1248 EXPORT_SYMBOL(debug_dma_map_single);
1249
1250 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1251                         size_t size, int direction, dma_addr_t dma_addr)
1252 {
1253         struct dma_debug_entry *entry;
1254
1255         if (unlikely(dma_debug_disabled()))
1256                 return;
1257
1258         if (dma_mapping_error(dev, dma_addr))
1259                 return;
1260
1261         entry = dma_entry_alloc();
1262         if (!entry)
1263                 return;
1264
1265         entry->dev       = dev;
1266         entry->type      = dma_debug_single;
1267         entry->pfn       = page_to_pfn(page);
1268         entry->offset    = offset,
1269         entry->dev_addr  = dma_addr;
1270         entry->size      = size;
1271         entry->direction = direction;
1272         entry->map_err_type = MAP_ERR_NOT_CHECKED;
1273
1274         check_for_stack(dev, page, offset);
1275
1276         if (!PageHighMem(page)) {
1277                 void *addr = page_address(page) + offset;
1278
1279                 check_for_illegal_area(dev, addr, size);
1280         }
1281
1282         add_dma_entry(entry);
1283 }
1284 EXPORT_SYMBOL(debug_dma_map_page);
1285
1286 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1287 {
1288         struct dma_debug_entry ref;
1289         struct dma_debug_entry *entry;
1290         struct hash_bucket *bucket;
1291         unsigned long flags;
1292
1293         if (unlikely(dma_debug_disabled()))
1294                 return;
1295
1296         ref.dev = dev;
1297         ref.dev_addr = dma_addr;
1298         bucket = get_hash_bucket(&ref, &flags);
1299
1300         list_for_each_entry(entry, &bucket->list, list) {
1301                 if (!exact_match(&ref, entry))
1302                         continue;
1303
1304                 /*
1305                  * The same physical address can be mapped multiple
1306                  * times. Without a hardware IOMMU this results in the
1307                  * same device addresses being put into the dma-debug
1308                  * hash multiple times too. This can result in false
1309                  * positives being reported. Therefore we implement a
1310                  * best-fit algorithm here which updates the first entry
1311                  * from the hash which fits the reference value and is
1312                  * not currently listed as being checked.
1313                  */
1314                 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1315                         entry->map_err_type = MAP_ERR_CHECKED;
1316                         break;
1317                 }
1318         }
1319
1320         put_hash_bucket(bucket, flags);
1321 }
1322 EXPORT_SYMBOL(debug_dma_mapping_error);
1323
1324 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1325                           size_t size, int direction)
1326 {
1327         struct dma_debug_entry ref = {
1328                 .type           = dma_debug_single,
1329                 .dev            = dev,
1330                 .dev_addr       = addr,
1331                 .size           = size,
1332                 .direction      = direction,
1333         };
1334
1335         if (unlikely(dma_debug_disabled()))
1336                 return;
1337         check_unmap(&ref);
1338 }
1339 EXPORT_SYMBOL(debug_dma_unmap_page);
1340
1341 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1342                       int nents, int mapped_ents, int direction)
1343 {
1344         struct dma_debug_entry *entry;
1345         struct scatterlist *s;
1346         int i;
1347
1348         if (unlikely(dma_debug_disabled()))
1349                 return;
1350
1351         for_each_sg(sg, s, mapped_ents, i) {
1352                 entry = dma_entry_alloc();
1353                 if (!entry)
1354                         return;
1355
1356                 entry->type           = dma_debug_sg;
1357                 entry->dev            = dev;
1358                 entry->pfn            = page_to_pfn(sg_page(s));
1359                 entry->offset         = s->offset,
1360                 entry->size           = sg_dma_len(s);
1361                 entry->dev_addr       = sg_dma_address(s);
1362                 entry->direction      = direction;
1363                 entry->sg_call_ents   = nents;
1364                 entry->sg_mapped_ents = mapped_ents;
1365
1366                 check_for_stack(dev, sg_page(s), s->offset);
1367
1368                 if (!PageHighMem(sg_page(s))) {
1369                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1370                 }
1371
1372                 check_sg_segment(dev, s);
1373
1374                 add_dma_entry(entry);
1375         }
1376 }
1377 EXPORT_SYMBOL(debug_dma_map_sg);
1378
1379 static int get_nr_mapped_entries(struct device *dev,
1380                                  struct dma_debug_entry *ref)
1381 {
1382         struct dma_debug_entry *entry;
1383         struct hash_bucket *bucket;
1384         unsigned long flags;
1385         int mapped_ents;
1386
1387         bucket       = get_hash_bucket(ref, &flags);
1388         entry        = bucket_find_exact(bucket, ref);
1389         mapped_ents  = 0;
1390
1391         if (entry)
1392                 mapped_ents = entry->sg_mapped_ents;
1393         put_hash_bucket(bucket, flags);
1394
1395         return mapped_ents;
1396 }
1397
1398 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1399                         int nelems, int dir)
1400 {
1401         struct scatterlist *s;
1402         int mapped_ents = 0, i;
1403
1404         if (unlikely(dma_debug_disabled()))
1405                 return;
1406
1407         for_each_sg(sglist, s, nelems, i) {
1408
1409                 struct dma_debug_entry ref = {
1410                         .type           = dma_debug_sg,
1411                         .dev            = dev,
1412                         .pfn            = page_to_pfn(sg_page(s)),
1413                         .offset         = s->offset,
1414                         .dev_addr       = sg_dma_address(s),
1415                         .size           = sg_dma_len(s),
1416                         .direction      = dir,
1417                         .sg_call_ents   = nelems,
1418                 };
1419
1420                 if (mapped_ents && i >= mapped_ents)
1421                         break;
1422
1423                 if (!i)
1424                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1425
1426                 check_unmap(&ref);
1427         }
1428 }
1429 EXPORT_SYMBOL(debug_dma_unmap_sg);
1430
1431 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1432                               dma_addr_t dma_addr, void *virt)
1433 {
1434         struct dma_debug_entry *entry;
1435
1436         if (unlikely(dma_debug_disabled()))
1437                 return;
1438
1439         if (unlikely(virt == NULL))
1440                 return;
1441
1442         /* handle vmalloc and linear addresses */
1443         if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1444                 return;
1445
1446         entry = dma_entry_alloc();
1447         if (!entry)
1448                 return;
1449
1450         entry->type      = dma_debug_coherent;
1451         entry->dev       = dev;
1452         entry->offset    = offset_in_page(virt);
1453         entry->size      = size;
1454         entry->dev_addr  = dma_addr;
1455         entry->direction = DMA_BIDIRECTIONAL;
1456
1457         if (is_vmalloc_addr(virt))
1458                 entry->pfn = vmalloc_to_pfn(virt);
1459         else
1460                 entry->pfn = page_to_pfn(virt_to_page(virt));
1461
1462         add_dma_entry(entry);
1463 }
1464
1465 void debug_dma_free_coherent(struct device *dev, size_t size,
1466                          void *virt, dma_addr_t addr)
1467 {
1468         struct dma_debug_entry ref = {
1469                 .type           = dma_debug_coherent,
1470                 .dev            = dev,
1471                 .offset         = offset_in_page(virt),
1472                 .dev_addr       = addr,
1473                 .size           = size,
1474                 .direction      = DMA_BIDIRECTIONAL,
1475         };
1476
1477         /* handle vmalloc and linear addresses */
1478         if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1479                 return;
1480
1481         if (is_vmalloc_addr(virt))
1482                 ref.pfn = vmalloc_to_pfn(virt);
1483         else
1484                 ref.pfn = page_to_pfn(virt_to_page(virt));
1485
1486         if (unlikely(dma_debug_disabled()))
1487                 return;
1488
1489         check_unmap(&ref);
1490 }
1491
1492 void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1493                             int direction, dma_addr_t dma_addr)
1494 {
1495         struct dma_debug_entry *entry;
1496
1497         if (unlikely(dma_debug_disabled()))
1498                 return;
1499
1500         entry = dma_entry_alloc();
1501         if (!entry)
1502                 return;
1503
1504         entry->type             = dma_debug_resource;
1505         entry->dev              = dev;
1506         entry->pfn              = PHYS_PFN(addr);
1507         entry->offset           = offset_in_page(addr);
1508         entry->size             = size;
1509         entry->dev_addr         = dma_addr;
1510         entry->direction        = direction;
1511         entry->map_err_type     = MAP_ERR_NOT_CHECKED;
1512
1513         add_dma_entry(entry);
1514 }
1515 EXPORT_SYMBOL(debug_dma_map_resource);
1516
1517 void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1518                               size_t size, int direction)
1519 {
1520         struct dma_debug_entry ref = {
1521                 .type           = dma_debug_resource,
1522                 .dev            = dev,
1523                 .dev_addr       = dma_addr,
1524                 .size           = size,
1525                 .direction      = direction,
1526         };
1527
1528         if (unlikely(dma_debug_disabled()))
1529                 return;
1530
1531         check_unmap(&ref);
1532 }
1533 EXPORT_SYMBOL(debug_dma_unmap_resource);
1534
1535 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1536                                    size_t size, int direction)
1537 {
1538         struct dma_debug_entry ref;
1539
1540         if (unlikely(dma_debug_disabled()))
1541                 return;
1542
1543         ref.type         = dma_debug_single;
1544         ref.dev          = dev;
1545         ref.dev_addr     = dma_handle;
1546         ref.size         = size;
1547         ref.direction    = direction;
1548         ref.sg_call_ents = 0;
1549
1550         check_sync(dev, &ref, true);
1551 }
1552 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1553
1554 void debug_dma_sync_single_for_device(struct device *dev,
1555                                       dma_addr_t dma_handle, size_t size,
1556                                       int direction)
1557 {
1558         struct dma_debug_entry ref;
1559
1560         if (unlikely(dma_debug_disabled()))
1561                 return;
1562
1563         ref.type         = dma_debug_single;
1564         ref.dev          = dev;
1565         ref.dev_addr     = dma_handle;
1566         ref.size         = size;
1567         ref.direction    = direction;
1568         ref.sg_call_ents = 0;
1569
1570         check_sync(dev, &ref, false);
1571 }
1572 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1573
1574 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1575                                int nelems, int direction)
1576 {
1577         struct scatterlist *s;
1578         int mapped_ents = 0, i;
1579
1580         if (unlikely(dma_debug_disabled()))
1581                 return;
1582
1583         for_each_sg(sg, s, nelems, i) {
1584
1585                 struct dma_debug_entry ref = {
1586                         .type           = dma_debug_sg,
1587                         .dev            = dev,
1588                         .pfn            = page_to_pfn(sg_page(s)),
1589                         .offset         = s->offset,
1590                         .dev_addr       = sg_dma_address(s),
1591                         .size           = sg_dma_len(s),
1592                         .direction      = direction,
1593                         .sg_call_ents   = nelems,
1594                 };
1595
1596                 if (!i)
1597                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1598
1599                 if (i >= mapped_ents)
1600                         break;
1601
1602                 check_sync(dev, &ref, true);
1603         }
1604 }
1605 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1606
1607 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1608                                   int nelems, int direction)
1609 {
1610         struct scatterlist *s;
1611         int mapped_ents = 0, i;
1612
1613         if (unlikely(dma_debug_disabled()))
1614                 return;
1615
1616         for_each_sg(sg, s, nelems, i) {
1617
1618                 struct dma_debug_entry ref = {
1619                         .type           = dma_debug_sg,
1620                         .dev            = dev,
1621                         .pfn            = page_to_pfn(sg_page(s)),
1622                         .offset         = s->offset,
1623                         .dev_addr       = sg_dma_address(s),
1624                         .size           = sg_dma_len(s),
1625                         .direction      = direction,
1626                         .sg_call_ents   = nelems,
1627                 };
1628                 if (!i)
1629                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1630
1631                 if (i >= mapped_ents)
1632                         break;
1633
1634                 check_sync(dev, &ref, false);
1635         }
1636 }
1637 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1638
1639 static int __init dma_debug_driver_setup(char *str)
1640 {
1641         int i;
1642
1643         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1644                 current_driver_name[i] = *str;
1645                 if (*str == 0)
1646                         break;
1647         }
1648
1649         if (current_driver_name[0])
1650                 pr_info("enable driver filter for driver [%s]\n",
1651                         current_driver_name);
1652
1653
1654         return 1;
1655 }
1656 __setup("dma_debug_driver=", dma_debug_driver_setup);