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 <joerg.roedel@amd.com>
   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[] = {
 141         [dma_debug_single] = "single",
 142         [dma_debug_sg] = "scather-gather",
 143         [dma_debug_coherent] = "coherent",
 144         [dma_debug_resource] = "resource",
 145 };
 146
 147 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
 148                                    "DMA_FROM_DEVICE", "DMA_NONE" };
 149
 150 /*
 151  * The access to some variables in this macro is racy. We can't use atomic_t
 152  * here because all these variables are exported to debugfs. Some of them even
 153  * writeable. This is also the reason why a lock won't help much. But anyway,
 154  * the races are no big deal. Here is why:
 155  *
 156  *   error_count: the addition is racy, but the worst thing that can happen is
 157  *                that we don't count some errors
 158  *   show_num_errors: the subtraction is racy. Also no big deal because in
 159  *                    worst case this will result in one warning more in the
 160  *                    system log than the user configured. This variable is
 161  *                    writeable via debugfs.
 162  */
 163 static inline void dump_entry_trace(struct dma_debug_entry *entry)
 164 {
 165 #ifdef CONFIG_STACKTRACE
 166         if (entry) {
 167                 pr_warn("Mapped at:\n");
 168                 stack_trace_print(entry->stack_entries, entry->stack_len, 0);
 169         }
 170 #endif
 171 }
 172
 173 static bool driver_filter(struct device *dev)
 174 {
 175         struct device_driver *drv;
 176         unsigned long flags;
 177         bool ret;
 178
 179         /* driver filter off */
 180         if (likely(!current_driver_name[0]))
 181                 return true;
 182
 183         /* driver filter on and initialized */
 184         if (current_driver && dev && dev->driver == current_driver)
 185                 return true;
 186
 187         /* driver filter on, but we can't filter on a NULL device... */
 188         if (!dev)
 189                 return false;
 190
 191         if (current_driver || !current_driver_name[0])
 192                 return false;
 193
 194         /* driver filter on but not yet initialized */
 195         drv = dev->driver;
 196         if (!drv)
 197                 return false;
 198
 199         /* lock to protect against change of current_driver_name */
 200         read_lock_irqsave(&driver_name_lock, flags);
 201
 202         ret = false;
 203         if (drv->name &&
 204             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 205                 current_driver = drv;
 206                 ret = true;
 207         }
 208
 209         read_unlock_irqrestore(&driver_name_lock, flags);
 210
 211         return ret;
 212 }
 213
 214 #define err_printk(dev, entry, format, arg...) do {                     \
 215                 error_count += 1;                                       \
 216                 if (driver_filter(dev) &&                               \
 217                     (show_all_errors || show_num_errors > 0)) {         \
 218                         WARN(1, pr_fmt("%s %s: ") format,               \
 219                              dev ? dev_driver_string(dev) : "NULL",     \
 220                              dev ? dev_name(dev) : "NULL", ## arg);     \
 221                         dump_entry_trace(entry);                        \
 222                 }                                                       \
 223                 if (!show_all_errors && show_num_errors > 0)            \
 224                         show_num_errors -= 1;                           \
 225         } while (0);
 226
 227 /*
 228  * Hash related functions
 229  *
 230  * Every DMA-API request is saved into a struct dma_debug_entry. To
 231  * have quick access to these structs they are stored into a hash.
 232  */
 233 static int hash_fn(struct dma_debug_entry *entry)
 234 {
 235         /*
 236          * Hash function is based on the dma address.
 237          * We use bits 20-27 here as the index into the hash
 238          */
 239         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 240 }
 241
 242 /*
 243  * Request exclusive access to a hash bucket for a given dma_debug_entry.
 244  */
 245 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 246                                            unsigned long *flags)
 247         __acquires(&dma_entry_hash[idx].lock)
 248 {
 249         int idx = hash_fn(entry);
 250         unsigned long __flags;
 251
 252         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 253         *flags = __flags;
 254         return &dma_entry_hash[idx];
 255 }
 256
 257 /*
 258  * Give up exclusive access to the hash bucket
 259  */
 260 static void put_hash_bucket(struct hash_bucket *bucket,
 261                             unsigned long flags)
 262         __releases(&bucket->lock)
 263 {
 264         spin_unlock_irqrestore(&bucket->lock, flags);
 265 }
 266
 267 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
 268 {
 269         return ((a->dev_addr == b->dev_addr) &&
 270                 (a->dev == b->dev)) ? true : false;
 271 }
 272
 273 static bool containing_match(struct dma_debug_entry *a,
 274                              struct dma_debug_entry *b)
 275 {
 276         if (a->dev != b->dev)
 277                 return false;
 278
 279         if ((b->dev_addr <= a->dev_addr) &&
 280             ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
 281                 return true;
 282
 283         return false;
 284 }
 285
 286 /*
 287  * Search a given entry in the hash bucket list
 288  */
 289 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
 290                                                   struct dma_debug_entry *ref,
 291                                                   match_fn match)
 292 {
 293         struct dma_debug_entry *entry, *ret = NULL;
 294         int matches = 0, match_lvl, last_lvl = -1;
 295
 296         list_for_each_entry(entry, &bucket->list, list) {
 297                 if (!match(ref, entry))
 298                         continue;
 299
 300                 /*
 301                  * Some drivers map the same physical address multiple
 302                  * times. Without a hardware IOMMU this results in the
 303                  * same device addresses being put into the dma-debug
 304                  * hash multiple times too. This can result in false
 305                  * positives being reported. Therefore we implement a
 306                  * best-fit algorithm here which returns the entry from
 307                  * the hash which fits best to the reference value
 308                  * instead of the first-fit.
 309                  */
 310                 matches += 1;
 311                 match_lvl = 0;
 312                 entry->size         == ref->size         ? ++match_lvl : 0;
 313                 entry->type         == ref->type         ? ++match_lvl : 0;
 314                 entry->direction    == ref->direction    ? ++match_lvl : 0;
 315                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
 316
 317                 if (match_lvl == 4) {
 318                         /* perfect-fit - return the result */
 319                         return entry;
 320                 } else if (match_lvl > last_lvl) {
 321                         /*
 322                          * We found an entry that fits better then the
 323                          * previous one or it is the 1st match.
 324                          */
 325                         last_lvl = match_lvl;
 326                         ret      = entry;
 327                 }
 328         }
 329
 330         /*
 331          * If we have multiple matches but no perfect-fit, just return
 332          * NULL.
 333          */
 334         ret = (matches == 1) ? ret : NULL;
 335
 336         return ret;
 337 }
 338
 339 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
 340                                                  struct dma_debug_entry *ref)
 341 {
 342         return __hash_bucket_find(bucket, ref, exact_match);
 343 }
 344
 345 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 346                                                    struct dma_debug_entry *ref,
 347                                                    unsigned long *flags)
 348 {
 349
 350         unsigned int max_range = dma_get_max_seg_size(ref->dev);
 351         struct dma_debug_entry *entry, index = *ref;
 352         unsigned int range = 0;
 353
 354         while (range <= max_range) {
 355                 entry = __hash_bucket_find(*bucket, ref, containing_match);
 356
 357                 if (entry)
 358                         return entry;
 359
 360                 /*
 361                  * Nothing found, go back a hash bucket
 362                  */
 363                 put_hash_bucket(*bucket, *flags);
 364                 range          += (1 << HASH_FN_SHIFT);
 365                 index.dev_addr -= (1 << HASH_FN_SHIFT);
 366                 *bucket = get_hash_bucket(&index, flags);
 367         }
 368
 369         return NULL;
 370 }
 371
 372 /*
 373  * Add an entry to a hash bucket
 374  */
 375 static void hash_bucket_add(struct hash_bucket *bucket,
 376                             struct dma_debug_entry *entry)
 377 {
 378         list_add_tail(&entry->list, &bucket->list);
 379 }
 380
 381 /*
 382  * Remove entry from a hash bucket list
 383  */
 384 static void hash_bucket_del(struct dma_debug_entry *entry)
 385 {
 386         list_del(&entry->list);
 387 }
 388
 389 static unsigned long long phys_addr(struct dma_debug_entry *entry)
 390 {
 391         if (entry->type == dma_debug_resource)
 392                 return __pfn_to_phys(entry->pfn) + entry->offset;
 393
 394         return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
 395 }
 396
 397 /*
 398  * Dump mapping entries for debugging purposes
 399  */
 400 void debug_dma_dump_mappings(struct device *dev)
 401 {
 402         int idx;
 403
 404         for (idx = 0; idx < HASH_SIZE; idx++) {
 405                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 406                 struct dma_debug_entry *entry;
 407                 unsigned long flags;
 408
 409                 spin_lock_irqsave(&bucket->lock, flags);
 410
 411                 list_for_each_entry(entry, &bucket->list, list) {
 412                         if (!dev || dev == entry->dev) {
 413                                 dev_info(entry->dev,
 414                                          "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
 415                                          type2name[entry->type], idx,
 416                                          phys_addr(entry), entry->pfn,
 417                                          entry->dev_addr, entry->size,
 418                                          dir2name[entry->direction],
 419                                          maperr2str[entry->map_err_type]);
 420                         }
 421                 }
 422
 423                 spin_unlock_irqrestore(&bucket->lock, flags);
 424                 cond_resched();
 425         }
 426 }
 427
 428 /*
 429  * For each mapping (initial cacheline in the case of
 430  * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
 431  * scatterlist, or the cacheline specified in dma_map_single) insert
 432  * into this tree using the cacheline as the key. At
 433  * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
 434  * the entry already exists at insertion time add a tag as a reference
 435  * count for the overlapping mappings.  For now, the overlap tracking
 436  * just ensures that 'unmaps' balance 'maps' before marking the
 437  * cacheline idle, but we should also be flagging overlaps as an API
 438  * violation.
 439  *
 440  * Memory usage is mostly constrained by the maximum number of available
 441  * dma-debug entries in that we need a free dma_debug_entry before
 442  * inserting into the tree.  In the case of dma_map_page and
 443  * dma_alloc_coherent there is only one dma_debug_entry and one
 444  * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
 445  * other hand, consumes a single dma_debug_entry, but inserts 'nents'
 446  * entries into the tree.
 447  *
 448  * At any time debug_dma_assert_idle() can be called to trigger a
 449  * warning if any cachelines in the given page are in the active set.
 450  */
 451 static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
 452 static DEFINE_SPINLOCK(radix_lock);
 453 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
 454 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
 455 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
 456
 457 static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
 458 {
 459         return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
 460                 (entry->offset >> L1_CACHE_SHIFT);
 461 }
 462
 463 static int active_cacheline_read_overlap(phys_addr_t cln)
 464 {
 465         int overlap = 0, i;
 466
 467         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 468                 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
 469                         overlap |= 1 << i;
 470         return overlap;
 471 }
 472
 473 static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
 474 {
 475         int i;
 476
 477         if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
 478                 return overlap;
 479
 480         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 481                 if (overlap & 1 << i)
 482                         radix_tree_tag_set(&dma_active_cacheline, cln, i);
 483                 else
 484                         radix_tree_tag_clear(&dma_active_cacheline, cln, i);
 485
 486         return overlap;
 487 }
 488
 489 static void active_cacheline_inc_overlap(phys_addr_t cln)
 490 {
 491         int overlap = active_cacheline_read_overlap(cln);
 492
 493         overlap = active_cacheline_set_overlap(cln, ++overlap);
 494
 495         /* If we overflowed the overlap counter then we're potentially
 496          * leaking dma-mappings.  Otherwise, if maps and unmaps are
 497          * balanced then this overflow may cause false negatives in
 498          * debug_dma_assert_idle() as the cacheline may be marked idle
 499          * prematurely.
 500          */
 501         WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
 502                   pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
 503                   ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 504 }
 505
 506 static int active_cacheline_dec_overlap(phys_addr_t cln)
 507 {
 508         int overlap = active_cacheline_read_overlap(cln);
 509
 510         return active_cacheline_set_overlap(cln, --overlap);
 511 }
 512
 513 static int active_cacheline_insert(struct dma_debug_entry *entry)
 514 {
 515         phys_addr_t cln = to_cacheline_number(entry);
 516         unsigned long flags;
 517         int rc;
 518
 519         /* If the device is not writing memory then we don't have any
 520          * concerns about the cpu consuming stale data.  This mitigates
 521          * legitimate usages of overlapping mappings.
 522          */
 523         if (entry->direction == DMA_TO_DEVICE)
 524                 return 0;
 525
 526         spin_lock_irqsave(&radix_lock, flags);
 527         rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
 528         if (rc == -EEXIST)
 529                 active_cacheline_inc_overlap(cln);
 530         spin_unlock_irqrestore(&radix_lock, flags);
 531
 532         return rc;
 533 }
 534
 535 static void active_cacheline_remove(struct dma_debug_entry *entry)
 536 {
 537         phys_addr_t cln = to_cacheline_number(entry);
 538         unsigned long flags;
 539
 540         /* ...mirror the insert case */
 541         if (entry->direction == DMA_TO_DEVICE)
 542                 return;
 543
 544         spin_lock_irqsave(&radix_lock, flags);
 545         /* since we are counting overlaps the final put of the
 546          * cacheline will occur when the overlap count is 0.
 547          * active_cacheline_dec_overlap() returns -1 in that case
 548          */
 549         if (active_cacheline_dec_overlap(cln) < 0)
 550                 radix_tree_delete(&dma_active_cacheline, cln);
 551         spin_unlock_irqrestore(&radix_lock, flags);
 552 }
 553
 554 /**
 555  * debug_dma_assert_idle() - assert that a page is not undergoing dma
 556  * @page: page to lookup in the dma_active_cacheline tree
 557  *
 558  * Place a call to this routine in cases where the cpu touching the page
 559  * before the dma completes (page is dma_unmapped) will lead to data
 560  * corruption.
 561  */
 562 void debug_dma_assert_idle(struct page *page)
 563 {
 564         static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
 565         struct dma_debug_entry *entry = NULL;
 566         void **results = (void **) &ents;
 567         unsigned int nents, i;
 568         unsigned long flags;
 569         phys_addr_t cln;
 570
 571         if (dma_debug_disabled())
 572                 return;
 573
 574         if (!page)
 575                 return;
 576
 577         cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
 578         spin_lock_irqsave(&radix_lock, flags);
 579         nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
 580                                        CACHELINES_PER_PAGE);
 581         for (i = 0; i < nents; i++) {
 582                 phys_addr_t ent_cln = to_cacheline_number(ents[i]);
 583
 584                 if (ent_cln == cln) {
 585                         entry = ents[i];
 586                         break;
 587                 } else if (ent_cln >= cln + CACHELINES_PER_PAGE)
 588                         break;
 589         }
 590         spin_unlock_irqrestore(&radix_lock, flags);
 591
 592         if (!entry)
 593                 return;
 594
 595         cln = to_cacheline_number(entry);
 596         err_printk(entry->dev, entry,
 597                    "cpu touching an active dma mapped cacheline [cln=%pa]\n",
 598                    &cln);
 599 }
 600
 601 /*
 602  * Wrapper function for adding an entry to the hash.
 603  * This function takes care of locking itself.
 604  */
 605 static void add_dma_entry(struct dma_debug_entry *entry)
 606 {
 607         struct hash_bucket *bucket;
 608         unsigned long flags;
 609         int rc;
 610
 611         bucket = get_hash_bucket(entry, &flags);
 612         hash_bucket_add(bucket, entry);
 613         put_hash_bucket(bucket, flags);
 614
 615         rc = active_cacheline_insert(entry);
 616         if (rc == -ENOMEM) {
 617                 pr_err("cacheline tracking ENOMEM, dma-debug disabled\n");
 618                 global_disable = true;
 619         }
 620
 621         /* TODO: report -EEXIST errors here as overlapping mappings are
 622          * not supported by the DMA API
 623          */
 624 }
 625
 626 static int dma_debug_create_entries(gfp_t gfp)
 627 {
 628         struct dma_debug_entry *entry;
 629         int i;
 630
 631         entry = (void *)get_zeroed_page(gfp);
 632         if (!entry)
 633                 return -ENOMEM;
 634
 635         for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
 636                 list_add_tail(&entry[i].list, &free_entries);
 637
 638         num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 639         nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 640
 641         return 0;
 642 }
 643
 644 static struct dma_debug_entry *__dma_entry_alloc(void)
 645 {
 646         struct dma_debug_entry *entry;
 647
 648         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 649         list_del(&entry->list);
 650         memset(entry, 0, sizeof(*entry));
 651
 652         num_free_entries -= 1;
 653         if (num_free_entries < min_free_entries)
 654                 min_free_entries = num_free_entries;
 655
 656         return entry;
 657 }
 658
 659 void __dma_entry_alloc_check_leak(void)
 660 {
 661         u32 tmp = nr_total_entries % nr_prealloc_entries;
 662
 663         /* Shout each time we tick over some multiple of the initial pool */
 664         if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
 665                 pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
 666                         nr_total_entries,
 667                         (nr_total_entries / nr_prealloc_entries));
 668         }
 669 }
 670
 671 /* struct dma_entry allocator
 672  *
 673  * The next two functions implement the allocator for
 674  * struct dma_debug_entries.
 675  */
 676 static struct dma_debug_entry *dma_entry_alloc(void)
 677 {
 678         struct dma_debug_entry *entry;
 679         unsigned long flags;
 680
 681         spin_lock_irqsave(&free_entries_lock, flags);
 682         if (num_free_entries == 0) {
 683                 if (dma_debug_create_entries(GFP_ATOMIC)) {
 684                         global_disable = true;
 685                         spin_unlock_irqrestore(&free_entries_lock, flags);
 686                         pr_err("debugging out of memory - disabling\n");
 687                         return NULL;
 688                 }
 689                 __dma_entry_alloc_check_leak();
 690         }
 691
 692         entry = __dma_entry_alloc();
 693
 694         spin_unlock_irqrestore(&free_entries_lock, flags);
 695
 696 #ifdef CONFIG_STACKTRACE
 697         entry->stack_len = stack_trace_save(entry->stack_entries,
 698                                             ARRAY_SIZE(entry->stack_entries),
 699                                             1);
 700 #endif
 701         return entry;
 702 }
 703
 704 static void dma_entry_free(struct dma_debug_entry *entry)
 705 {
 706         unsigned long flags;
 707
 708         active_cacheline_remove(entry);
 709
 710         /*
 711          * add to beginning of the list - this way the entries are
 712          * more likely cache hot when they are reallocated.
 713          */
 714         spin_lock_irqsave(&free_entries_lock, flags);
 715         list_add(&entry->list, &free_entries);
 716         num_free_entries += 1;
 717         spin_unlock_irqrestore(&free_entries_lock, flags);
 718 }
 719
 720 /*
 721  * DMA-API debugging init code
 722  *
 723  * The init code does two things:
 724  *   1. Initialize core data structures
 725  *   2. Preallocate a given number of dma_debug_entry structs
 726  */
 727
 728 static ssize_t filter_read(struct file *file, char __user *user_buf,
 729                            size_t count, loff_t *ppos)
 730 {
 731         char buf[NAME_MAX_LEN + 1];
 732         unsigned long flags;
 733         int len;
 734
 735         if (!current_driver_name[0])
 736                 return 0;
 737
 738         /*
 739          * We can't copy to userspace directly because current_driver_name can
 740          * only be read under the driver_name_lock with irqs disabled. So
 741          * create a temporary copy first.
 742          */
 743         read_lock_irqsave(&driver_name_lock, flags);
 744         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 745         read_unlock_irqrestore(&driver_name_lock, flags);
 746
 747         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 748 }
 749
 750 static ssize_t filter_write(struct file *file, const char __user *userbuf,
 751                             size_t count, loff_t *ppos)
 752 {
 753         char buf[NAME_MAX_LEN];
 754         unsigned long flags;
 755         size_t len;
 756         int i;
 757
 758         /*
 759          * We can't copy from userspace directly. Access to
 760          * current_driver_name is protected with a write_lock with irqs
 761          * disabled. Since copy_from_user can fault and may sleep we
 762          * need to copy to temporary buffer first
 763          */
 764         len = min(count, (size_t)(NAME_MAX_LEN - 1));
 765         if (copy_from_user(buf, userbuf, len))
 766                 return -EFAULT;
 767
 768         buf[len] = 0;
 769
 770         write_lock_irqsave(&driver_name_lock, flags);
 771
 772         /*
 773          * Now handle the string we got from userspace very carefully.
 774          * The rules are:
 775          *         - only use the first token we got
 776          *         - token delimiter is everything looking like a space
 777          *           character (' ', '\n', '\t' ...)
 778          *
 779          */
 780         if (!isalnum(buf[0])) {
 781                 /*
 782                  * If the first character userspace gave us is not
 783                  * alphanumerical then assume the filter should be
 784                  * switched off.
 785                  */
 786                 if (current_driver_name[0])
 787                         pr_info("switching off dma-debug driver filter\n");
 788                 current_driver_name[0] = 0;
 789                 current_driver = NULL;
 790                 goto out_unlock;
 791         }
 792
 793         /*
 794          * Now parse out the first token and use it as the name for the
 795          * driver to filter for.
 796          */
 797         for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 798                 current_driver_name[i] = buf[i];
 799                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 800                         break;
 801         }
 802         current_driver_name[i] = 0;
 803         current_driver = NULL;
 804
 805         pr_info("enable driver filter for driver [%s]\n",
 806                 current_driver_name);
 807
 808 out_unlock:
 809         write_unlock_irqrestore(&driver_name_lock, flags);
 810
 811         return count;
 812 }
 813
 814 static const struct file_operations filter_fops = {
 815         .read  = filter_read,
 816         .write = filter_write,
 817         .llseek = default_llseek,
 818 };
 819
 820 static int dump_show(struct seq_file *seq, void *v)
 821 {
 822         int idx;
 823
 824         for (idx = 0; idx < HASH_SIZE; idx++) {
 825                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 826                 struct dma_debug_entry *entry;
 827                 unsigned long flags;
 828
 829                 spin_lock_irqsave(&bucket->lock, flags);
 830                 list_for_each_entry(entry, &bucket->list, list) {
 831                         seq_printf(seq,
 832                                    "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx %s %s\n",
 833                                    dev_name(entry->dev),
 834                                    dev_driver_string(entry->dev),
 835                                    type2name[entry->type], idx,
 836                                    phys_addr(entry), entry->pfn,
 837                                    entry->dev_addr, entry->size,
 838                                    dir2name[entry->direction],
 839                                    maperr2str[entry->map_err_type]);
 840                 }
 841                 spin_unlock_irqrestore(&bucket->lock, flags);
 842         }
 843         return 0;
 844 }
 845 DEFINE_SHOW_ATTRIBUTE(dump);
 846
 847 static void dma_debug_fs_init(void)
 848 {
 849         struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
 850
 851         debugfs_create_bool("disabled", 0444, dentry, &global_disable);
 852         debugfs_create_u32("error_count", 0444, dentry, &error_count);
 853         debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
 854         debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
 855         debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
 856         debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
 857         debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
 858         debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
 859         debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
 860 }
 861
 862 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 863 {
 864         struct dma_debug_entry *entry;
 865         unsigned long flags;
 866         int count = 0, i;
 867
 868         for (i = 0; i < HASH_SIZE; ++i) {
 869                 spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
 870                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 871                         if (entry->dev == dev) {
 872                                 count += 1;
 873                                 *out_entry = entry;
 874                         }
 875                 }
 876                 spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
 877         }
 878
 879         return count;
 880 }
 881
 882 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 883 {
 884         struct device *dev = data;
 885         struct dma_debug_entry *uninitialized_var(entry);
 886         int count;
 887
 888         if (dma_debug_disabled())
 889                 return 0;
 890
 891         switch (action) {
 892         case BUS_NOTIFY_UNBOUND_DRIVER:
 893                 count = device_dma_allocations(dev, &entry);
 894                 if (count == 0)
 895                         break;
 896                 err_printk(dev, entry, "device driver has pending "
 897                                 "DMA allocations while released from device "
 898                                 "[count=%d]\n"
 899                                 "One of leaked entries details: "
 900                                 "[device address=0x%016llx] [size=%llu bytes] "
 901                                 "[mapped with %s] [mapped as %s]\n",
 902                         count, entry->dev_addr, entry->size,
 903                         dir2name[entry->direction], type2name[entry->type]);
 904                 break;
 905         default:
 906                 break;
 907         }
 908
 909         return 0;
 910 }
 911
 912 void dma_debug_add_bus(struct bus_type *bus)
 913 {
 914         struct notifier_block *nb;
 915
 916         if (dma_debug_disabled())
 917                 return;
 918
 919         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 920         if (nb == NULL) {
 921                 pr_err("dma_debug_add_bus: out of memory\n");
 922                 return;
 923         }
 924
 925         nb->notifier_call = dma_debug_device_change;
 926
 927         bus_register_notifier(bus, nb);
 928 }
 929
 930 static int dma_debug_init(void)
 931 {
 932         int i, nr_pages;
 933
 934         /* Do not use dma_debug_initialized here, since we really want to be
 935          * called to set dma_debug_initialized
 936          */
 937         if (global_disable)
 938                 return 0;
 939
 940         for (i = 0; i < HASH_SIZE; ++i) {
 941                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
 942                 spin_lock_init(&dma_entry_hash[i].lock);
 943         }
 944
 945         dma_debug_fs_init();
 946
 947         nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
 948         for (i = 0; i < nr_pages; ++i)
 949                 dma_debug_create_entries(GFP_KERNEL);
 950         if (num_free_entries >= nr_prealloc_entries) {
 951                 pr_info("preallocated %d debug entries\n", nr_total_entries);
 952         } else if (num_free_entries > 0) {
 953                 pr_warn("%d debug entries requested but only %d allocated\n",
 954                         nr_prealloc_entries, nr_total_entries);
 955         } else {
 956                 pr_err("debugging out of memory error - disabled\n");
 957                 global_disable = true;
 958
 959                 return 0;
 960         }
 961         min_free_entries = num_free_entries;
 962
 963         dma_debug_initialized = true;
 964
 965         pr_info("debugging enabled by kernel config\n");
 966         return 0;
 967 }
 968 core_initcall(dma_debug_init);
 969
 970 static __init int dma_debug_cmdline(char *str)
 971 {
 972         if (!str)
 973                 return -EINVAL;
 974
 975         if (strncmp(str, "off", 3) == 0) {
 976                 pr_info("debugging disabled on kernel command line\n");
 977                 global_disable = true;
 978         }
 979
 980         return 0;
 981 }
 982
 983 static __init int dma_debug_entries_cmdline(char *str)
 984 {
 985         if (!str)
 986                 return -EINVAL;
 987         if (!get_option(&str, &nr_prealloc_entries))
 988                 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 989         return 0;
 990 }
 991
 992 __setup("dma_debug=", dma_debug_cmdline);
 993 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
 994
 995 static void check_unmap(struct dma_debug_entry *ref)
 996 {
 997         struct dma_debug_entry *entry;
 998         struct hash_bucket *bucket;
 999         unsigned long flags;
1000
1001         bucket = get_hash_bucket(ref, &flags);
1002         entry = bucket_find_exact(bucket, ref);
1003
1004         if (!entry) {
1005                 /* must drop lock before calling dma_mapping_error */
1006                 put_hash_bucket(bucket, flags);
1007
1008                 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1009                         err_printk(ref->dev, NULL,
1010                                    "device driver tries to free an "
1011                                    "invalid DMA memory address\n");
1012                 } else {
1013                         err_printk(ref->dev, NULL,
1014                                    "device driver tries to free DMA "
1015                                    "memory it has not allocated [device "
1016                                    "address=0x%016llx] [size=%llu bytes]\n",
1017                                    ref->dev_addr, ref->size);
1018                 }
1019                 return;
1020         }
1021
1022         if (ref->size != entry->size) {
1023                 err_printk(ref->dev, entry, "device driver frees "
1024                            "DMA memory with different size "
1025                            "[device address=0x%016llx] [map size=%llu bytes] "
1026                            "[unmap size=%llu bytes]\n",
1027                            ref->dev_addr, entry->size, ref->size);
1028         }
1029
1030         if (ref->type != entry->type) {
1031                 err_printk(ref->dev, entry, "device driver frees "
1032                            "DMA memory with wrong function "
1033                            "[device address=0x%016llx] [size=%llu bytes] "
1034                            "[mapped as %s] [unmapped as %s]\n",
1035                            ref->dev_addr, ref->size,
1036                            type2name[entry->type], type2name[ref->type]);
1037         } else if ((entry->type == dma_debug_coherent) &&
1038                    (phys_addr(ref) != phys_addr(entry))) {
1039                 err_printk(ref->dev, entry, "device driver frees "
1040                            "DMA memory with different CPU address "
1041                            "[device address=0x%016llx] [size=%llu bytes] "
1042                            "[cpu alloc address=0x%016llx] "
1043                            "[cpu free address=0x%016llx]",
1044                            ref->dev_addr, ref->size,
1045                            phys_addr(entry),
1046                            phys_addr(ref));
1047         }
1048
1049         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1050             ref->sg_call_ents != entry->sg_call_ents) {
1051                 err_printk(ref->dev, entry, "device driver frees "
1052                            "DMA sg list with different entry count "
1053                            "[map count=%d] [unmap count=%d]\n",
1054                            entry->sg_call_ents, ref->sg_call_ents);
1055         }
1056
1057         /*
1058          * This may be no bug in reality - but most implementations of the
1059          * DMA API don't handle this properly, so check for it here
1060          */
1061         if (ref->direction != entry->direction) {
1062                 err_printk(ref->dev, entry, "device driver frees "
1063                            "DMA memory with different direction "
1064                            "[device address=0x%016llx] [size=%llu bytes] "
1065                            "[mapped with %s] [unmapped with %s]\n",
1066                            ref->dev_addr, ref->size,
1067                            dir2name[entry->direction],
1068                            dir2name[ref->direction]);
1069         }
1070
1071         /*
1072          * Drivers should use dma_mapping_error() to check the returned
1073          * addresses of dma_map_single() and dma_map_page().
1074          * If not, print this warning message. See Documentation/DMA-API.txt.
1075          */
1076         if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1077                 err_printk(ref->dev, entry,
1078                            "device driver failed to check map error"
1079                            "[device address=0x%016llx] [size=%llu bytes] "
1080                            "[mapped as %s]",
1081                            ref->dev_addr, ref->size,
1082                            type2name[entry->type]);
1083         }
1084
1085         hash_bucket_del(entry);
1086         dma_entry_free(entry);
1087
1088         put_hash_bucket(bucket, flags);
1089 }
1090
1091 static void check_for_stack(struct device *dev,
1092                             struct page *page, size_t offset)
1093 {
1094         void *addr;
1095         struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1096
1097         if (!stack_vm_area) {
1098                 /* Stack is direct-mapped. */
1099                 if (PageHighMem(page))
1100                         return;
1101                 addr = page_address(page) + offset;
1102                 if (object_is_on_stack(addr))
1103                         err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1104         } else {
1105                 /* Stack is vmalloced. */
1106                 int i;
1107
1108                 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1109                         if (page != stack_vm_area->pages[i])
1110                                 continue;
1111
1112                         addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1113                         err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1114                         break;
1115                 }
1116         }
1117 }
1118
1119 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1120 {
1121         unsigned long a1 = (unsigned long)addr;
1122         unsigned long b1 = a1 + len;
1123         unsigned long a2 = (unsigned long)start;
1124         unsigned long b2 = (unsigned long)end;
1125
1126         return !(b1 <= a2 || a1 >= b2);
1127 }
1128
1129 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1130 {
1131         if (overlap(addr, len, _stext, _etext) ||
1132             overlap(addr, len, __start_rodata, __end_rodata))
1133                 err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1134 }
1135
1136 static void check_sync(struct device *dev,
1137                        struct dma_debug_entry *ref,
1138                        bool to_cpu)
1139 {
1140         struct dma_debug_entry *entry;
1141         struct hash_bucket *bucket;
1142         unsigned long flags;
1143
1144         bucket = get_hash_bucket(ref, &flags);
1145
1146         entry = bucket_find_contain(&bucket, ref, &flags);
1147
1148         if (!entry) {
1149                 err_printk(dev, NULL, "device driver tries "
1150                                 "to sync DMA memory it has not allocated "
1151                                 "[device address=0x%016llx] [size=%llu bytes]\n",
1152                                 (unsigned long long)ref->dev_addr, ref->size);
1153                 goto out;
1154         }
1155
1156         if (ref->size > entry->size) {
1157                 err_printk(dev, entry, "device driver syncs"
1158                                 " DMA memory outside allocated range "
1159                                 "[device address=0x%016llx] "
1160                                 "[allocation size=%llu bytes] "
1161                                 "[sync offset+size=%llu]\n",
1162                                 entry->dev_addr, entry->size,
1163                                 ref->size);
1164         }
1165
1166         if (entry->direction == DMA_BIDIRECTIONAL)
1167                 goto out;
1168
1169         if (ref->direction != entry->direction) {
1170                 err_printk(dev, entry, "device driver syncs "
1171                                 "DMA memory with different direction "
1172                                 "[device address=0x%016llx] [size=%llu bytes] "
1173                                 "[mapped with %s] [synced with %s]\n",
1174                                 (unsigned long long)ref->dev_addr, entry->size,
1175                                 dir2name[entry->direction],
1176                                 dir2name[ref->direction]);
1177         }
1178
1179         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1180                       !(ref->direction == DMA_TO_DEVICE))
1181                 err_printk(dev, entry, "device driver syncs "
1182                                 "device read-only DMA memory for cpu "
1183                                 "[device address=0x%016llx] [size=%llu bytes] "
1184                                 "[mapped with %s] [synced with %s]\n",
1185                                 (unsigned long long)ref->dev_addr, entry->size,
1186                                 dir2name[entry->direction],
1187                                 dir2name[ref->direction]);
1188
1189         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1190                        !(ref->direction == DMA_FROM_DEVICE))
1191                 err_printk(dev, entry, "device driver syncs "
1192                                 "device write-only DMA memory to device "
1193                                 "[device address=0x%016llx] [size=%llu bytes] "
1194                                 "[mapped with %s] [synced with %s]\n",
1195                                 (unsigned long long)ref->dev_addr, entry->size,
1196                                 dir2name[entry->direction],
1197                                 dir2name[ref->direction]);
1198
1199         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1200             ref->sg_call_ents != entry->sg_call_ents) {
1201                 err_printk(ref->dev, entry, "device driver syncs "
1202                            "DMA sg list with different entry count "
1203                            "[map count=%d] [sync count=%d]\n",
1204                            entry->sg_call_ents, ref->sg_call_ents);
1205         }
1206
1207 out:
1208         put_hash_bucket(bucket, flags);
1209 }
1210
1211 static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1212 {
1213 #ifdef CONFIG_DMA_API_DEBUG_SG
1214         unsigned int max_seg = dma_get_max_seg_size(dev);
1215         u64 start, end, boundary = dma_get_seg_boundary(dev);
1216
1217         /*
1218          * Either the driver forgot to set dma_parms appropriately, or
1219          * whoever generated the list forgot to check them.
1220          */
1221         if (sg->length > max_seg)
1222                 err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1223                            sg->length, max_seg);
1224         /*
1225          * In some cases this could potentially be the DMA API
1226          * implementation's fault, but it would usually imply that
1227          * the scatterlist was built inappropriately to begin with.
1228          */
1229         start = sg_dma_address(sg);
1230         end = start + sg_dma_len(sg) - 1;
1231         if ((start ^ end) & ~boundary)
1232                 err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1233                            start, end, boundary);
1234 #endif
1235 }
1236
1237 void debug_dma_map_single(struct device *dev, const void *addr,
1238                             unsigned long len)
1239 {
1240         if (unlikely(dma_debug_disabled()))
1241                 return;
1242
1243         if (!virt_addr_valid(addr))
1244                 err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1245                            addr, len);
1246
1247         if (is_vmalloc_addr(addr))
1248                 err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1249                            addr, len);
1250 }
1251 EXPORT_SYMBOL(debug_dma_map_single);
1252
1253 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1254                         size_t size, int direction, dma_addr_t dma_addr)
1255 {
1256         struct dma_debug_entry *entry;
1257
1258         if (unlikely(dma_debug_disabled()))
1259                 return;
1260
1261         if (dma_mapping_error(dev, dma_addr))
1262                 return;
1263
1264         entry = dma_entry_alloc();
1265         if (!entry)
1266                 return;
1267
1268         entry->dev       = dev;
1269         entry->type      = dma_debug_single;
1270         entry->pfn       = page_to_pfn(page);
1271         entry->offset    = offset,
1272         entry->dev_addr  = dma_addr;
1273         entry->size      = size;
1274         entry->direction = direction;
1275         entry->map_err_type = MAP_ERR_NOT_CHECKED;
1276
1277         check_for_stack(dev, page, offset);
1278
1279         if (!PageHighMem(page)) {
1280                 void *addr = page_address(page) + offset;
1281
1282                 check_for_illegal_area(dev, addr, size);
1283         }
1284
1285         add_dma_entry(entry);
1286 }
1287 EXPORT_SYMBOL(debug_dma_map_page);
1288
1289 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1290 {
1291         struct dma_debug_entry ref;
1292         struct dma_debug_entry *entry;
1293         struct hash_bucket *bucket;
1294         unsigned long flags;
1295
1296         if (unlikely(dma_debug_disabled()))
1297                 return;
1298
1299         ref.dev = dev;
1300         ref.dev_addr = dma_addr;
1301         bucket = get_hash_bucket(&ref, &flags);
1302
1303         list_for_each_entry(entry, &bucket->list, list) {
1304                 if (!exact_match(&ref, entry))
1305                         continue;
1306
1307                 /*
1308                  * The same physical address can be mapped multiple
1309                  * times. Without a hardware IOMMU this results in the
1310                  * same device addresses being put into the dma-debug
1311                  * hash multiple times too. This can result in false
1312                  * positives being reported. Therefore we implement a
1313                  * best-fit algorithm here which updates the first entry
1314                  * from the hash which fits the reference value and is
1315                  * not currently listed as being checked.
1316                  */
1317                 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1318                         entry->map_err_type = MAP_ERR_CHECKED;
1319                         break;
1320                 }
1321         }
1322
1323         put_hash_bucket(bucket, flags);
1324 }
1325 EXPORT_SYMBOL(debug_dma_mapping_error);
1326
1327 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1328                           size_t size, int direction)
1329 {
1330         struct dma_debug_entry ref = {
1331                 .type           = dma_debug_single,
1332                 .dev            = dev,
1333                 .dev_addr       = addr,
1334                 .size           = size,
1335                 .direction      = direction,
1336         };
1337
1338         if (unlikely(dma_debug_disabled()))
1339                 return;
1340         check_unmap(&ref);
1341 }
1342 EXPORT_SYMBOL(debug_dma_unmap_page);
1343
1344 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1345                       int nents, int mapped_ents, int direction)
1346 {
1347         struct dma_debug_entry *entry;
1348         struct scatterlist *s;
1349         int i;
1350
1351         if (unlikely(dma_debug_disabled()))
1352                 return;
1353
1354         for_each_sg(sg, s, mapped_ents, i) {
1355                 entry = dma_entry_alloc();
1356                 if (!entry)
1357                         return;
1358
1359                 entry->type           = dma_debug_sg;
1360                 entry->dev            = dev;
1361                 entry->pfn            = page_to_pfn(sg_page(s));
1362                 entry->offset         = s->offset,
1363                 entry->size           = sg_dma_len(s);
1364                 entry->dev_addr       = sg_dma_address(s);
1365                 entry->direction      = direction;
1366                 entry->sg_call_ents   = nents;
1367                 entry->sg_mapped_ents = mapped_ents;
1368
1369                 check_for_stack(dev, sg_page(s), s->offset);
1370
1371                 if (!PageHighMem(sg_page(s))) {
1372                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1373                 }
1374
1375                 check_sg_segment(dev, s);
1376
1377                 add_dma_entry(entry);
1378         }
1379 }
1380 EXPORT_SYMBOL(debug_dma_map_sg);
1381
1382 static int get_nr_mapped_entries(struct device *dev,
1383                                  struct dma_debug_entry *ref)
1384 {
1385         struct dma_debug_entry *entry;
1386         struct hash_bucket *bucket;
1387         unsigned long flags;
1388         int mapped_ents;
1389
1390         bucket       = get_hash_bucket(ref, &flags);
1391         entry        = bucket_find_exact(bucket, ref);
1392         mapped_ents  = 0;
1393
1394         if (entry)
1395                 mapped_ents = entry->sg_mapped_ents;
1396         put_hash_bucket(bucket, flags);
1397
1398         return mapped_ents;
1399 }
1400
1401 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1402                         int nelems, int dir)
1403 {
1404         struct scatterlist *s;
1405         int mapped_ents = 0, i;
1406
1407         if (unlikely(dma_debug_disabled()))
1408                 return;
1409
1410         for_each_sg(sglist, s, nelems, i) {
1411
1412                 struct dma_debug_entry ref = {
1413                         .type           = dma_debug_sg,
1414                         .dev            = dev,
1415                         .pfn            = page_to_pfn(sg_page(s)),
1416                         .offset         = s->offset,
1417                         .dev_addr       = sg_dma_address(s),
1418                         .size           = sg_dma_len(s),
1419                         .direction      = dir,
1420                         .sg_call_ents   = nelems,
1421                 };
1422
1423                 if (mapped_ents && i >= mapped_ents)
1424                         break;
1425
1426                 if (!i)
1427                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1428
1429                 check_unmap(&ref);
1430         }
1431 }
1432 EXPORT_SYMBOL(debug_dma_unmap_sg);
1433
1434 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1435                               dma_addr_t dma_addr, void *virt)
1436 {
1437         struct dma_debug_entry *entry;
1438
1439         if (unlikely(dma_debug_disabled()))
1440                 return;
1441
1442         if (unlikely(virt == NULL))
1443                 return;
1444
1445         /* handle vmalloc and linear addresses */
1446         if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1447                 return;
1448
1449         entry = dma_entry_alloc();
1450         if (!entry)
1451                 return;
1452
1453         entry->type      = dma_debug_coherent;
1454         entry->dev       = dev;
1455         entry->offset    = offset_in_page(virt);
1456         entry->size      = size;
1457         entry->dev_addr  = dma_addr;
1458         entry->direction = DMA_BIDIRECTIONAL;
1459
1460         if (is_vmalloc_addr(virt))
1461                 entry->pfn = vmalloc_to_pfn(virt);
1462         else
1463                 entry->pfn = page_to_pfn(virt_to_page(virt));
1464
1465         add_dma_entry(entry);
1466 }
1467
1468 void debug_dma_free_coherent(struct device *dev, size_t size,
1469                          void *virt, dma_addr_t addr)
1470 {
1471         struct dma_debug_entry ref = {
1472                 .type           = dma_debug_coherent,
1473                 .dev            = dev,
1474                 .offset         = offset_in_page(virt),
1475                 .dev_addr       = addr,
1476                 .size           = size,
1477                 .direction      = DMA_BIDIRECTIONAL,
1478         };
1479
1480         /* handle vmalloc and linear addresses */
1481         if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1482                 return;
1483
1484         if (is_vmalloc_addr(virt))
1485                 ref.pfn = vmalloc_to_pfn(virt);
1486         else
1487                 ref.pfn = page_to_pfn(virt_to_page(virt));
1488
1489         if (unlikely(dma_debug_disabled()))
1490                 return;
1491
1492         check_unmap(&ref);
1493 }
1494
1495 void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1496                             int direction, dma_addr_t dma_addr)
1497 {
1498         struct dma_debug_entry *entry;
1499
1500         if (unlikely(dma_debug_disabled()))
1501                 return;
1502
1503         entry = dma_entry_alloc();
1504         if (!entry)
1505                 return;
1506
1507         entry->type             = dma_debug_resource;
1508         entry->dev              = dev;
1509         entry->pfn              = PHYS_PFN(addr);
1510         entry->offset           = offset_in_page(addr);
1511         entry->size             = size;
1512         entry->dev_addr         = dma_addr;
1513         entry->direction        = direction;
1514         entry->map_err_type     = MAP_ERR_NOT_CHECKED;
1515
1516         add_dma_entry(entry);
1517 }
1518 EXPORT_SYMBOL(debug_dma_map_resource);
1519
1520 void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1521                               size_t size, int direction)
1522 {
1523         struct dma_debug_entry ref = {
1524                 .type           = dma_debug_resource,
1525                 .dev            = dev,
1526                 .dev_addr       = dma_addr,
1527                 .size           = size,
1528                 .direction      = direction,
1529         };
1530
1531         if (unlikely(dma_debug_disabled()))
1532                 return;
1533
1534         check_unmap(&ref);
1535 }
1536 EXPORT_SYMBOL(debug_dma_unmap_resource);
1537
1538 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1539                                    size_t size, int direction)
1540 {
1541         struct dma_debug_entry ref;
1542
1543         if (unlikely(dma_debug_disabled()))
1544                 return;
1545
1546         ref.type         = dma_debug_single;
1547         ref.dev          = dev;
1548         ref.dev_addr     = dma_handle;
1549         ref.size         = size;
1550         ref.direction    = direction;
1551         ref.sg_call_ents = 0;
1552
1553         check_sync(dev, &ref, true);
1554 }
1555 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1556
1557 void debug_dma_sync_single_for_device(struct device *dev,
1558                                       dma_addr_t dma_handle, size_t size,
1559                                       int direction)
1560 {
1561         struct dma_debug_entry ref;
1562
1563         if (unlikely(dma_debug_disabled()))
1564                 return;
1565
1566         ref.type         = dma_debug_single;
1567         ref.dev          = dev;
1568         ref.dev_addr     = dma_handle;
1569         ref.size         = size;
1570         ref.direction    = direction;
1571         ref.sg_call_ents = 0;
1572
1573         check_sync(dev, &ref, false);
1574 }
1575 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1576
1577 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1578                                int nelems, int direction)
1579 {
1580         struct scatterlist *s;
1581         int mapped_ents = 0, i;
1582
1583         if (unlikely(dma_debug_disabled()))
1584                 return;
1585
1586         for_each_sg(sg, s, nelems, i) {
1587
1588                 struct dma_debug_entry ref = {
1589                         .type           = dma_debug_sg,
1590                         .dev            = dev,
1591                         .pfn            = page_to_pfn(sg_page(s)),
1592                         .offset         = s->offset,
1593                         .dev_addr       = sg_dma_address(s),
1594                         .size           = sg_dma_len(s),
1595                         .direction      = direction,
1596                         .sg_call_ents   = nelems,
1597                 };
1598
1599                 if (!i)
1600                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1601
1602                 if (i >= mapped_ents)
1603                         break;
1604
1605                 check_sync(dev, &ref, true);
1606         }
1607 }
1608 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1609
1610 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1611                                   int nelems, int direction)
1612 {
1613         struct scatterlist *s;
1614         int mapped_ents = 0, i;
1615
1616         if (unlikely(dma_debug_disabled()))
1617                 return;
1618
1619         for_each_sg(sg, s, nelems, i) {
1620
1621                 struct dma_debug_entry ref = {
1622                         .type           = dma_debug_sg,
1623                         .dev            = dev,
1624                         .pfn            = page_to_pfn(sg_page(s)),
1625                         .offset         = s->offset,
1626                         .dev_addr       = sg_dma_address(s),
1627                         .size           = sg_dma_len(s),
1628                         .direction      = direction,
1629                         .sg_call_ents   = nelems,
1630                 };
1631                 if (!i)
1632                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1633
1634                 if (i >= mapped_ents)
1635                         break;
1636
1637                 check_sync(dev, &ref, false);
1638         }
1639 }
1640 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1641
1642 static int __init dma_debug_driver_setup(char *str)
1643 {
1644         int i;
1645
1646         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1647                 current_driver_name[i] = *str;
1648                 if (*str == 0)
1649                         break;
1650         }
1651
1652         if (current_driver_name[0])
1653                 pr_info("enable driver filter for driver [%s]\n",
1654                         current_driver_name);
1655
1656
1657         return 1;
1658 }
1659 __setup("dma_debug_driver=", dma_debug_driver_setup);