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[] = {
 148         [DMA_BIDIRECTIONAL]     = "DMA_BIDIRECTIONAL",
 149         [DMA_TO_DEVICE]         = "DMA_TO_DEVICE",
 150         [DMA_FROM_DEVICE]       = "DMA_FROM_DEVICE",
 151         [DMA_NONE]              = "DMA_NONE",
 152 };
 153
 154 /*
 155  * The access to some variables in this macro is racy. We can't use atomic_t
 156  * here because all these variables are exported to debugfs. Some of them even
 157  * writeable. This is also the reason why a lock won't help much. But anyway,
 158  * the races are no big deal. Here is why:
 159  *
 160  *   error_count: the addition is racy, but the worst thing that can happen is
 161  *                that we don't count some errors
 162  *   show_num_errors: the subtraction is racy. Also no big deal because in
 163  *                    worst case this will result in one warning more in the
 164  *                    system log than the user configured. This variable is
 165  *                    writeable via debugfs.
 166  */
 167 static inline void dump_entry_trace(struct dma_debug_entry *entry)
 168 {
 169 #ifdef CONFIG_STACKTRACE
 170         if (entry) {
 171                 pr_warn("Mapped at:\n");
 172                 stack_trace_print(entry->stack_entries, entry->stack_len, 0);
 173         }
 174 #endif
 175 }
 176
 177 static bool driver_filter(struct device *dev)
 178 {
 179         struct device_driver *drv;
 180         unsigned long flags;
 181         bool ret;
 182
 183         /* driver filter off */
 184         if (likely(!current_driver_name[0]))
 185                 return true;
 186
 187         /* driver filter on and initialized */
 188         if (current_driver && dev && dev->driver == current_driver)
 189                 return true;
 190
 191         /* driver filter on, but we can't filter on a NULL device... */
 192         if (!dev)
 193                 return false;
 194
 195         if (current_driver || !current_driver_name[0])
 196                 return false;
 197
 198         /* driver filter on but not yet initialized */
 199         drv = dev->driver;
 200         if (!drv)
 201                 return false;
 202
 203         /* lock to protect against change of current_driver_name */
 204         read_lock_irqsave(&driver_name_lock, flags);
 205
 206         ret = false;
 207         if (drv->name &&
 208             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 209                 current_driver = drv;
 210                 ret = true;
 211         }
 212
 213         read_unlock_irqrestore(&driver_name_lock, flags);
 214
 215         return ret;
 216 }
 217
 218 #define err_printk(dev, entry, format, arg...) do {                     \
 219                 error_count += 1;                                       \
 220                 if (driver_filter(dev) &&                               \
 221                     (show_all_errors || show_num_errors > 0)) {         \
 222                         WARN(1, pr_fmt("%s %s: ") format,               \
 223                              dev ? dev_driver_string(dev) : "NULL",     \
 224                              dev ? dev_name(dev) : "NULL", ## arg);     \
 225                         dump_entry_trace(entry);                        \
 226                 }                                                       \
 227                 if (!show_all_errors && show_num_errors > 0)            \
 228                         show_num_errors -= 1;                           \
 229         } while (0);
 230
 231 /*
 232  * Hash related functions
 233  *
 234  * Every DMA-API request is saved into a struct dma_debug_entry. To
 235  * have quick access to these structs they are stored into a hash.
 236  */
 237 static int hash_fn(struct dma_debug_entry *entry)
 238 {
 239         /*
 240          * Hash function is based on the dma address.
 241          * We use bits 20-27 here as the index into the hash
 242          */
 243         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 244 }
 245
 246 /*
 247  * Request exclusive access to a hash bucket for a given dma_debug_entry.
 248  */
 249 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 250                                            unsigned long *flags)
 251         __acquires(&dma_entry_hash[idx].lock)
 252 {
 253         int idx = hash_fn(entry);
 254         unsigned long __flags;
 255
 256         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 257         *flags = __flags;
 258         return &dma_entry_hash[idx];
 259 }
 260
 261 /*
 262  * Give up exclusive access to the hash bucket
 263  */
 264 static void put_hash_bucket(struct hash_bucket *bucket,
 265                             unsigned long flags)
 266         __releases(&bucket->lock)
 267 {
 268         spin_unlock_irqrestore(&bucket->lock, flags);
 269 }
 270
 271 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
 272 {
 273         return ((a->dev_addr == b->dev_addr) &&
 274                 (a->dev == b->dev)) ? true : false;
 275 }
 276
 277 static bool containing_match(struct dma_debug_entry *a,
 278                              struct dma_debug_entry *b)
 279 {
 280         if (a->dev != b->dev)
 281                 return false;
 282
 283         if ((b->dev_addr <= a->dev_addr) &&
 284             ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
 285                 return true;
 286
 287         return false;
 288 }
 289
 290 /*
 291  * Search a given entry in the hash bucket list
 292  */
 293 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
 294                                                   struct dma_debug_entry *ref,
 295                                                   match_fn match)
 296 {
 297         struct dma_debug_entry *entry, *ret = NULL;
 298         int matches = 0, match_lvl, last_lvl = -1;
 299
 300         list_for_each_entry(entry, &bucket->list, list) {
 301                 if (!match(ref, entry))
 302                         continue;
 303
 304                 /*
 305                  * Some drivers map the same physical address multiple
 306                  * times. Without a hardware IOMMU this results in the
 307                  * same device addresses being put into the dma-debug
 308                  * hash multiple times too. This can result in false
 309                  * positives being reported. Therefore we implement a
 310                  * best-fit algorithm here which returns the entry from
 311                  * the hash which fits best to the reference value
 312                  * instead of the first-fit.
 313                  */
 314                 matches += 1;
 315                 match_lvl = 0;
 316                 entry->size         == ref->size         ? ++match_lvl : 0;
 317                 entry->type         == ref->type         ? ++match_lvl : 0;
 318                 entry->direction    == ref->direction    ? ++match_lvl : 0;
 319                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
 320
 321                 if (match_lvl == 4) {
 322                         /* perfect-fit - return the result */
 323                         return entry;
 324                 } else if (match_lvl > last_lvl) {
 325                         /*
 326                          * We found an entry that fits better then the
 327                          * previous one or it is the 1st match.
 328                          */
 329                         last_lvl = match_lvl;
 330                         ret      = entry;
 331                 }
 332         }
 333
 334         /*
 335          * If we have multiple matches but no perfect-fit, just return
 336          * NULL.
 337          */
 338         ret = (matches == 1) ? ret : NULL;
 339
 340         return ret;
 341 }
 342
 343 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
 344                                                  struct dma_debug_entry *ref)
 345 {
 346         return __hash_bucket_find(bucket, ref, exact_match);
 347 }
 348
 349 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 350                                                    struct dma_debug_entry *ref,
 351                                                    unsigned long *flags)
 352 {
 353
 354         unsigned int max_range = dma_get_max_seg_size(ref->dev);
 355         struct dma_debug_entry *entry, index = *ref;
 356         unsigned int range = 0;
 357
 358         while (range <= max_range) {
 359                 entry = __hash_bucket_find(*bucket, ref, containing_match);
 360
 361                 if (entry)
 362                         return entry;
 363
 364                 /*
 365                  * Nothing found, go back a hash bucket
 366                  */
 367                 put_hash_bucket(*bucket, *flags);
 368                 range          += (1 << HASH_FN_SHIFT);
 369                 index.dev_addr -= (1 << HASH_FN_SHIFT);
 370                 *bucket = get_hash_bucket(&index, flags);
 371         }
 372
 373         return NULL;
 374 }
 375
 376 /*
 377  * Add an entry to a hash bucket
 378  */
 379 static void hash_bucket_add(struct hash_bucket *bucket,
 380                             struct dma_debug_entry *entry)
 381 {
 382         list_add_tail(&entry->list, &bucket->list);
 383 }
 384
 385 /*
 386  * Remove entry from a hash bucket list
 387  */
 388 static void hash_bucket_del(struct dma_debug_entry *entry)
 389 {
 390         list_del(&entry->list);
 391 }
 392
 393 static unsigned long long phys_addr(struct dma_debug_entry *entry)
 394 {
 395         if (entry->type == dma_debug_resource)
 396                 return __pfn_to_phys(entry->pfn) + entry->offset;
 397
 398         return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
 399 }
 400
 401 /*
 402  * Dump mapping entries for debugging purposes
 403  */
 404 void debug_dma_dump_mappings(struct device *dev)
 405 {
 406         int idx;
 407
 408         for (idx = 0; idx < HASH_SIZE; idx++) {
 409                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 410                 struct dma_debug_entry *entry;
 411                 unsigned long flags;
 412
 413                 spin_lock_irqsave(&bucket->lock, flags);
 414
 415                 list_for_each_entry(entry, &bucket->list, list) {
 416                         if (!dev || dev == entry->dev) {
 417                                 dev_info(entry->dev,
 418                                          "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
 419                                          type2name[entry->type], idx,
 420                                          phys_addr(entry), entry->pfn,
 421                                          entry->dev_addr, entry->size,
 422                                          dir2name[entry->direction],
 423                                          maperr2str[entry->map_err_type]);
 424                         }
 425                 }
 426
 427                 spin_unlock_irqrestore(&bucket->lock, flags);
 428                 cond_resched();
 429         }
 430 }
 431
 432 /*
 433  * For each mapping (initial cacheline in the case of
 434  * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
 435  * scatterlist, or the cacheline specified in dma_map_single) insert
 436  * into this tree using the cacheline as the key. At
 437  * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
 438  * the entry already exists at insertion time add a tag as a reference
 439  * count for the overlapping mappings.  For now, the overlap tracking
 440  * just ensures that 'unmaps' balance 'maps' before marking the
 441  * cacheline idle, but we should also be flagging overlaps as an API
 442  * violation.
 443  *
 444  * Memory usage is mostly constrained by the maximum number of available
 445  * dma-debug entries in that we need a free dma_debug_entry before
 446  * inserting into the tree.  In the case of dma_map_page and
 447  * dma_alloc_coherent there is only one dma_debug_entry and one
 448  * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
 449  * other hand, consumes a single dma_debug_entry, but inserts 'nents'
 450  * entries into the tree.
 451  *
 452  * At any time debug_dma_assert_idle() can be called to trigger a
 453  * warning if any cachelines in the given page are in the active set.
 454  */
 455 static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
 456 static DEFINE_SPINLOCK(radix_lock);
 457 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
 458 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
 459 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
 460
 461 static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
 462 {
 463         return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
 464                 (entry->offset >> L1_CACHE_SHIFT);
 465 }
 466
 467 static int active_cacheline_read_overlap(phys_addr_t cln)
 468 {
 469         int overlap = 0, i;
 470
 471         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 472                 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
 473                         overlap |= 1 << i;
 474         return overlap;
 475 }
 476
 477 static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
 478 {
 479         int i;
 480
 481         if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
 482                 return overlap;
 483
 484         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 485                 if (overlap & 1 << i)
 486                         radix_tree_tag_set(&dma_active_cacheline, cln, i);
 487                 else
 488                         radix_tree_tag_clear(&dma_active_cacheline, cln, i);
 489
 490         return overlap;
 491 }
 492
 493 static void active_cacheline_inc_overlap(phys_addr_t cln)
 494 {
 495         int overlap = active_cacheline_read_overlap(cln);
 496
 497         overlap = active_cacheline_set_overlap(cln, ++overlap);
 498
 499         /* If we overflowed the overlap counter then we're potentially
 500          * leaking dma-mappings.  Otherwise, if maps and unmaps are
 501          * balanced then this overflow may cause false negatives in
 502          * debug_dma_assert_idle() as the cacheline may be marked idle
 503          * prematurely.
 504          */
 505         WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
 506                   pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
 507                   ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 508 }
 509
 510 static int active_cacheline_dec_overlap(phys_addr_t cln)
 511 {
 512         int overlap = active_cacheline_read_overlap(cln);
 513
 514         return active_cacheline_set_overlap(cln, --overlap);
 515 }
 516
 517 static int active_cacheline_insert(struct dma_debug_entry *entry)
 518 {
 519         phys_addr_t cln = to_cacheline_number(entry);
 520         unsigned long flags;
 521         int rc;
 522
 523         /* If the device is not writing memory then we don't have any
 524          * concerns about the cpu consuming stale data.  This mitigates
 525          * legitimate usages of overlapping mappings.
 526          */
 527         if (entry->direction == DMA_TO_DEVICE)
 528                 return 0;
 529
 530         spin_lock_irqsave(&radix_lock, flags);
 531         rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
 532         if (rc == -EEXIST)
 533                 active_cacheline_inc_overlap(cln);
 534         spin_unlock_irqrestore(&radix_lock, flags);
 535
 536         return rc;
 537 }
 538
 539 static void active_cacheline_remove(struct dma_debug_entry *entry)
 540 {
 541         phys_addr_t cln = to_cacheline_number(entry);
 542         unsigned long flags;
 543
 544         /* ...mirror the insert case */
 545         if (entry->direction == DMA_TO_DEVICE)
 546                 return;
 547
 548         spin_lock_irqsave(&radix_lock, flags);
 549         /* since we are counting overlaps the final put of the
 550          * cacheline will occur when the overlap count is 0.
 551          * active_cacheline_dec_overlap() returns -1 in that case
 552          */
 553         if (active_cacheline_dec_overlap(cln) < 0)
 554                 radix_tree_delete(&dma_active_cacheline, cln);
 555         spin_unlock_irqrestore(&radix_lock, flags);
 556 }
 557
 558 /**
 559  * debug_dma_assert_idle() - assert that a page is not undergoing dma
 560  * @page: page to lookup in the dma_active_cacheline tree
 561  *
 562  * Place a call to this routine in cases where the cpu touching the page
 563  * before the dma completes (page is dma_unmapped) will lead to data
 564  * corruption.
 565  */
 566 void debug_dma_assert_idle(struct page *page)
 567 {
 568         static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
 569         struct dma_debug_entry *entry = NULL;
 570         void **results = (void **) &ents;
 571         unsigned int nents, i;
 572         unsigned long flags;
 573         phys_addr_t cln;
 574
 575         if (dma_debug_disabled())
 576                 return;
 577
 578         if (!page)
 579                 return;
 580
 581         cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
 582         spin_lock_irqsave(&radix_lock, flags);
 583         nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
 584                                        CACHELINES_PER_PAGE);
 585         for (i = 0; i < nents; i++) {
 586                 phys_addr_t ent_cln = to_cacheline_number(ents[i]);
 587
 588                 if (ent_cln == cln) {
 589                         entry = ents[i];
 590                         break;
 591                 } else if (ent_cln >= cln + CACHELINES_PER_PAGE)
 592                         break;
 593         }
 594         spin_unlock_irqrestore(&radix_lock, flags);
 595
 596         if (!entry)
 597                 return;
 598
 599         cln = to_cacheline_number(entry);
 600         err_printk(entry->dev, entry,
 601                    "cpu touching an active dma mapped cacheline [cln=%pa]\n",
 602                    &cln);
 603 }
 604
 605 /*
 606  * Wrapper function for adding an entry to the hash.
 607  * This function takes care of locking itself.
 608  */
 609 static void add_dma_entry(struct dma_debug_entry *entry)
 610 {
 611         struct hash_bucket *bucket;
 612         unsigned long flags;
 613         int rc;
 614
 615         bucket = get_hash_bucket(entry, &flags);
 616         hash_bucket_add(bucket, entry);
 617         put_hash_bucket(bucket, flags);
 618
 619         rc = active_cacheline_insert(entry);
 620         if (rc == -ENOMEM) {
 621                 pr_err("cacheline tracking ENOMEM, dma-debug disabled\n");
 622                 global_disable = true;
 623         }
 624
 625         /* TODO: report -EEXIST errors here as overlapping mappings are
 626          * not supported by the DMA API
 627          */
 628 }
 629
 630 static int dma_debug_create_entries(gfp_t gfp)
 631 {
 632         struct dma_debug_entry *entry;
 633         int i;
 634
 635         entry = (void *)get_zeroed_page(gfp);
 636         if (!entry)
 637                 return -ENOMEM;
 638
 639         for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
 640                 list_add_tail(&entry[i].list, &free_entries);
 641
 642         num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 643         nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 644
 645         return 0;
 646 }
 647
 648 static struct dma_debug_entry *__dma_entry_alloc(void)
 649 {
 650         struct dma_debug_entry *entry;
 651
 652         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 653         list_del(&entry->list);
 654         memset(entry, 0, sizeof(*entry));
 655
 656         num_free_entries -= 1;
 657         if (num_free_entries < min_free_entries)
 658                 min_free_entries = num_free_entries;
 659
 660         return entry;
 661 }
 662
 663 static void __dma_entry_alloc_check_leak(void)
 664 {
 665         u32 tmp = nr_total_entries % nr_prealloc_entries;
 666
 667         /* Shout each time we tick over some multiple of the initial pool */
 668         if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
 669                 pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
 670                         nr_total_entries,
 671                         (nr_total_entries / nr_prealloc_entries));
 672         }
 673 }
 674
 675 /* struct dma_entry allocator
 676  *
 677  * The next two functions implement the allocator for
 678  * struct dma_debug_entries.
 679  */
 680 static struct dma_debug_entry *dma_entry_alloc(void)
 681 {
 682         struct dma_debug_entry *entry;
 683         unsigned long flags;
 684
 685         spin_lock_irqsave(&free_entries_lock, flags);
 686         if (num_free_entries == 0) {
 687                 if (dma_debug_create_entries(GFP_ATOMIC)) {
 688                         global_disable = true;
 689                         spin_unlock_irqrestore(&free_entries_lock, flags);
 690                         pr_err("debugging out of memory - disabling\n");
 691                         return NULL;
 692                 }
 693                 __dma_entry_alloc_check_leak();
 694         }
 695
 696         entry = __dma_entry_alloc();
 697
 698         spin_unlock_irqrestore(&free_entries_lock, flags);
 699
 700 #ifdef CONFIG_STACKTRACE
 701         entry->stack_len = stack_trace_save(entry->stack_entries,
 702                                             ARRAY_SIZE(entry->stack_entries),
 703                                             1);
 704 #endif
 705         return entry;
 706 }
 707
 708 static void dma_entry_free(struct dma_debug_entry *entry)
 709 {
 710         unsigned long flags;
 711
 712         active_cacheline_remove(entry);
 713
 714         /*
 715          * add to beginning of the list - this way the entries are
 716          * more likely cache hot when they are reallocated.
 717          */
 718         spin_lock_irqsave(&free_entries_lock, flags);
 719         list_add(&entry->list, &free_entries);
 720         num_free_entries += 1;
 721         spin_unlock_irqrestore(&free_entries_lock, flags);
 722 }
 723
 724 /*
 725  * DMA-API debugging init code
 726  *
 727  * The init code does two things:
 728  *   1. Initialize core data structures
 729  *   2. Preallocate a given number of dma_debug_entry structs
 730  */
 731
 732 static ssize_t filter_read(struct file *file, char __user *user_buf,
 733                            size_t count, loff_t *ppos)
 734 {
 735         char buf[NAME_MAX_LEN + 1];
 736         unsigned long flags;
 737         int len;
 738
 739         if (!current_driver_name[0])
 740                 return 0;
 741
 742         /*
 743          * We can't copy to userspace directly because current_driver_name can
 744          * only be read under the driver_name_lock with irqs disabled. So
 745          * create a temporary copy first.
 746          */
 747         read_lock_irqsave(&driver_name_lock, flags);
 748         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 749         read_unlock_irqrestore(&driver_name_lock, flags);
 750
 751         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 752 }
 753
 754 static ssize_t filter_write(struct file *file, const char __user *userbuf,
 755                             size_t count, loff_t *ppos)
 756 {
 757         char buf[NAME_MAX_LEN];
 758         unsigned long flags;
 759         size_t len;
 760         int i;
 761
 762         /*
 763          * We can't copy from userspace directly. Access to
 764          * current_driver_name is protected with a write_lock with irqs
 765          * disabled. Since copy_from_user can fault and may sleep we
 766          * need to copy to temporary buffer first
 767          */
 768         len = min(count, (size_t)(NAME_MAX_LEN - 1));
 769         if (copy_from_user(buf, userbuf, len))
 770                 return -EFAULT;
 771
 772         buf[len] = 0;
 773
 774         write_lock_irqsave(&driver_name_lock, flags);
 775
 776         /*
 777          * Now handle the string we got from userspace very carefully.
 778          * The rules are:
 779          *         - only use the first token we got
 780          *         - token delimiter is everything looking like a space
 781          *           character (' ', '\n', '\t' ...)
 782          *
 783          */
 784         if (!isalnum(buf[0])) {
 785                 /*
 786                  * If the first character userspace gave us is not
 787                  * alphanumerical then assume the filter should be
 788                  * switched off.
 789                  */
 790                 if (current_driver_name[0])
 791                         pr_info("switching off dma-debug driver filter\n");
 792                 current_driver_name[0] = 0;
 793                 current_driver = NULL;
 794                 goto out_unlock;
 795         }
 796
 797         /*
 798          * Now parse out the first token and use it as the name for the
 799          * driver to filter for.
 800          */
 801         for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 802                 current_driver_name[i] = buf[i];
 803                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 804                         break;
 805         }
 806         current_driver_name[i] = 0;
 807         current_driver = NULL;
 808
 809         pr_info("enable driver filter for driver [%s]\n",
 810                 current_driver_name);
 811
 812 out_unlock:
 813         write_unlock_irqrestore(&driver_name_lock, flags);
 814
 815         return count;
 816 }
 817
 818 static const struct file_operations filter_fops = {
 819         .read  = filter_read,
 820         .write = filter_write,
 821         .llseek = default_llseek,
 822 };
 823
 824 static int dump_show(struct seq_file *seq, void *v)
 825 {
 826         int idx;
 827
 828         for (idx = 0; idx < HASH_SIZE; idx++) {
 829                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 830                 struct dma_debug_entry *entry;
 831                 unsigned long flags;
 832
 833                 spin_lock_irqsave(&bucket->lock, flags);
 834                 list_for_each_entry(entry, &bucket->list, list) {
 835                         seq_printf(seq,
 836                                    "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx %s %s\n",
 837                                    dev_name(entry->dev),
 838                                    dev_driver_string(entry->dev),
 839                                    type2name[entry->type], idx,
 840                                    phys_addr(entry), entry->pfn,
 841                                    entry->dev_addr, entry->size,
 842                                    dir2name[entry->direction],
 843                                    maperr2str[entry->map_err_type]);
 844                 }
 845                 spin_unlock_irqrestore(&bucket->lock, flags);
 846         }
 847         return 0;
 848 }
 849 DEFINE_SHOW_ATTRIBUTE(dump);
 850
 851 static void dma_debug_fs_init(void)
 852 {
 853         struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
 854
 855         debugfs_create_bool("disabled", 0444, dentry, &global_disable);
 856         debugfs_create_u32("error_count", 0444, dentry, &error_count);
 857         debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
 858         debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
 859         debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
 860         debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
 861         debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
 862         debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
 863         debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
 864 }
 865
 866 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 867 {
 868         struct dma_debug_entry *entry;
 869         unsigned long flags;
 870         int count = 0, i;
 871
 872         for (i = 0; i < HASH_SIZE; ++i) {
 873                 spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
 874                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 875                         if (entry->dev == dev) {
 876                                 count += 1;
 877                                 *out_entry = entry;
 878                         }
 879                 }
 880                 spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
 881         }
 882
 883         return count;
 884 }
 885
 886 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 887 {
 888         struct device *dev = data;
 889         struct dma_debug_entry *entry;
 890         int count;
 891
 892         if (dma_debug_disabled())
 893                 return 0;
 894
 895         switch (action) {
 896         case BUS_NOTIFY_UNBOUND_DRIVER:
 897                 count = device_dma_allocations(dev, &entry);
 898                 if (count == 0)
 899                         break;
 900                 err_printk(dev, entry, "device driver has pending "
 901                                 "DMA allocations while released from device "
 902                                 "[count=%d]\n"
 903                                 "One of leaked entries details: "
 904                                 "[device address=0x%016llx] [size=%llu bytes] "
 905                                 "[mapped with %s] [mapped as %s]\n",
 906                         count, entry->dev_addr, entry->size,
 907                         dir2name[entry->direction], type2name[entry->type]);
 908                 break;
 909         default:
 910                 break;
 911         }
 912
 913         return 0;
 914 }
 915
 916 void dma_debug_add_bus(struct bus_type *bus)
 917 {
 918         struct notifier_block *nb;
 919
 920         if (dma_debug_disabled())
 921                 return;
 922
 923         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 924         if (nb == NULL) {
 925                 pr_err("dma_debug_add_bus: out of memory\n");
 926                 return;
 927         }
 928
 929         nb->notifier_call = dma_debug_device_change;
 930
 931         bus_register_notifier(bus, nb);
 932 }
 933
 934 static int dma_debug_init(void)
 935 {
 936         int i, nr_pages;
 937
 938         /* Do not use dma_debug_initialized here, since we really want to be
 939          * called to set dma_debug_initialized
 940          */
 941         if (global_disable)
 942                 return 0;
 943
 944         for (i = 0; i < HASH_SIZE; ++i) {
 945                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
 946                 spin_lock_init(&dma_entry_hash[i].lock);
 947         }
 948
 949         dma_debug_fs_init();
 950
 951         nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
 952         for (i = 0; i < nr_pages; ++i)
 953                 dma_debug_create_entries(GFP_KERNEL);
 954         if (num_free_entries >= nr_prealloc_entries) {
 955                 pr_info("preallocated %d debug entries\n", nr_total_entries);
 956         } else if (num_free_entries > 0) {
 957                 pr_warn("%d debug entries requested but only %d allocated\n",
 958                         nr_prealloc_entries, nr_total_entries);
 959         } else {
 960                 pr_err("debugging out of memory error - disabled\n");
 961                 global_disable = true;
 962
 963                 return 0;
 964         }
 965         min_free_entries = num_free_entries;
 966
 967         dma_debug_initialized = true;
 968
 969         pr_info("debugging enabled by kernel config\n");
 970         return 0;
 971 }
 972 core_initcall(dma_debug_init);
 973
 974 static __init int dma_debug_cmdline(char *str)
 975 {
 976         if (!str)
 977                 return -EINVAL;
 978
 979         if (strncmp(str, "off", 3) == 0) {
 980                 pr_info("debugging disabled on kernel command line\n");
 981                 global_disable = true;
 982         }
 983
 984         return 0;
 985 }
 986
 987 static __init int dma_debug_entries_cmdline(char *str)
 988 {
 989         if (!str)
 990                 return -EINVAL;
 991         if (!get_option(&str, &nr_prealloc_entries))
 992                 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 993         return 0;
 994 }
 995
 996 __setup("dma_debug=", dma_debug_cmdline);
 997 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
 998
 999 static void check_unmap(struct dma_debug_entry *ref)
1000 {
1001         struct dma_debug_entry *entry;
1002         struct hash_bucket *bucket;
1003         unsigned long flags;
1004
1005         bucket = get_hash_bucket(ref, &flags);
1006         entry = bucket_find_exact(bucket, ref);
1007
1008         if (!entry) {
1009                 /* must drop lock before calling dma_mapping_error */
1010                 put_hash_bucket(bucket, flags);
1011
1012                 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1013                         err_printk(ref->dev, NULL,
1014                                    "device driver tries to free an "
1015                                    "invalid DMA memory address\n");
1016                 } else {
1017                         err_printk(ref->dev, NULL,
1018                                    "device driver tries to free DMA "
1019                                    "memory it has not allocated [device "
1020                                    "address=0x%016llx] [size=%llu bytes]\n",
1021                                    ref->dev_addr, ref->size);
1022                 }
1023                 return;
1024         }
1025
1026         if (ref->size != entry->size) {
1027                 err_printk(ref->dev, entry, "device driver frees "
1028                            "DMA memory with different size "
1029                            "[device address=0x%016llx] [map size=%llu bytes] "
1030                            "[unmap size=%llu bytes]\n",
1031                            ref->dev_addr, entry->size, ref->size);
1032         }
1033
1034         if (ref->type != entry->type) {
1035                 err_printk(ref->dev, entry, "device driver frees "
1036                            "DMA memory with wrong function "
1037                            "[device address=0x%016llx] [size=%llu bytes] "
1038                            "[mapped as %s] [unmapped as %s]\n",
1039                            ref->dev_addr, ref->size,
1040                            type2name[entry->type], type2name[ref->type]);
1041         } else if ((entry->type == dma_debug_coherent) &&
1042                    (phys_addr(ref) != phys_addr(entry))) {
1043                 err_printk(ref->dev, entry, "device driver frees "
1044                            "DMA memory with different CPU address "
1045                            "[device address=0x%016llx] [size=%llu bytes] "
1046                            "[cpu alloc address=0x%016llx] "
1047                            "[cpu free address=0x%016llx]",
1048                            ref->dev_addr, ref->size,
1049                            phys_addr(entry),
1050                            phys_addr(ref));
1051         }
1052
1053         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1054             ref->sg_call_ents != entry->sg_call_ents) {
1055                 err_printk(ref->dev, entry, "device driver frees "
1056                            "DMA sg list with different entry count "
1057                            "[map count=%d] [unmap count=%d]\n",
1058                            entry->sg_call_ents, ref->sg_call_ents);
1059         }
1060
1061         /*
1062          * This may be no bug in reality - but most implementations of the
1063          * DMA API don't handle this properly, so check for it here
1064          */
1065         if (ref->direction != entry->direction) {
1066                 err_printk(ref->dev, entry, "device driver frees "
1067                            "DMA memory with different direction "
1068                            "[device address=0x%016llx] [size=%llu bytes] "
1069                            "[mapped with %s] [unmapped with %s]\n",
1070                            ref->dev_addr, ref->size,
1071                            dir2name[entry->direction],
1072                            dir2name[ref->direction]);
1073         }
1074
1075         /*
1076          * Drivers should use dma_mapping_error() to check the returned
1077          * addresses of dma_map_single() and dma_map_page().
1078          * If not, print this warning message. See Documentation/core-api/dma-api.rst.
1079          */
1080         if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1081                 err_printk(ref->dev, entry,
1082                            "device driver failed to check map error"
1083                            "[device address=0x%016llx] [size=%llu bytes] "
1084                            "[mapped as %s]",
1085                            ref->dev_addr, ref->size,
1086                            type2name[entry->type]);
1087         }
1088
1089         hash_bucket_del(entry);
1090         dma_entry_free(entry);
1091
1092         put_hash_bucket(bucket, flags);
1093 }
1094
1095 static void check_for_stack(struct device *dev,
1096                             struct page *page, size_t offset)
1097 {
1098         void *addr;
1099         struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1100
1101         if (!stack_vm_area) {
1102                 /* Stack is direct-mapped. */
1103                 if (PageHighMem(page))
1104                         return;
1105                 addr = page_address(page) + offset;
1106                 if (object_is_on_stack(addr))
1107                         err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1108         } else {
1109                 /* Stack is vmalloced. */
1110                 int i;
1111
1112                 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1113                         if (page != stack_vm_area->pages[i])
1114                                 continue;
1115
1116                         addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1117                         err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1118                         break;
1119                 }
1120         }
1121 }
1122
1123 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1124 {
1125         unsigned long a1 = (unsigned long)addr;
1126         unsigned long b1 = a1 + len;
1127         unsigned long a2 = (unsigned long)start;
1128         unsigned long b2 = (unsigned long)end;
1129
1130         return !(b1 <= a2 || a1 >= b2);
1131 }
1132
1133 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1134 {
1135         if (overlap(addr, len, _stext, _etext) ||
1136             overlap(addr, len, __start_rodata, __end_rodata))
1137                 err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1138 }
1139
1140 static void check_sync(struct device *dev,
1141                        struct dma_debug_entry *ref,
1142                        bool to_cpu)
1143 {
1144         struct dma_debug_entry *entry;
1145         struct hash_bucket *bucket;
1146         unsigned long flags;
1147
1148         bucket = get_hash_bucket(ref, &flags);
1149
1150         entry = bucket_find_contain(&bucket, ref, &flags);
1151
1152         if (!entry) {
1153                 err_printk(dev, NULL, "device driver tries "
1154                                 "to sync DMA memory it has not allocated "
1155                                 "[device address=0x%016llx] [size=%llu bytes]\n",
1156                                 (unsigned long long)ref->dev_addr, ref->size);
1157                 goto out;
1158         }
1159
1160         if (ref->size > entry->size) {
1161                 err_printk(dev, entry, "device driver syncs"
1162                                 " DMA memory outside allocated range "
1163                                 "[device address=0x%016llx] "
1164                                 "[allocation size=%llu bytes] "
1165                                 "[sync offset+size=%llu]\n",
1166                                 entry->dev_addr, entry->size,
1167                                 ref->size);
1168         }
1169
1170         if (entry->direction == DMA_BIDIRECTIONAL)
1171                 goto out;
1172
1173         if (ref->direction != entry->direction) {
1174                 err_printk(dev, entry, "device driver syncs "
1175                                 "DMA memory with different direction "
1176                                 "[device address=0x%016llx] [size=%llu bytes] "
1177                                 "[mapped with %s] [synced with %s]\n",
1178                                 (unsigned long long)ref->dev_addr, entry->size,
1179                                 dir2name[entry->direction],
1180                                 dir2name[ref->direction]);
1181         }
1182
1183         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1184                       !(ref->direction == DMA_TO_DEVICE))
1185                 err_printk(dev, entry, "device driver syncs "
1186                                 "device read-only DMA memory for cpu "
1187                                 "[device address=0x%016llx] [size=%llu bytes] "
1188                                 "[mapped with %s] [synced with %s]\n",
1189                                 (unsigned long long)ref->dev_addr, entry->size,
1190                                 dir2name[entry->direction],
1191                                 dir2name[ref->direction]);
1192
1193         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1194                        !(ref->direction == DMA_FROM_DEVICE))
1195                 err_printk(dev, entry, "device driver syncs "
1196                                 "device write-only DMA memory to device "
1197                                 "[device address=0x%016llx] [size=%llu bytes] "
1198                                 "[mapped with %s] [synced with %s]\n",
1199                                 (unsigned long long)ref->dev_addr, entry->size,
1200                                 dir2name[entry->direction],
1201                                 dir2name[ref->direction]);
1202
1203         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1204             ref->sg_call_ents != entry->sg_call_ents) {
1205                 err_printk(ref->dev, entry, "device driver syncs "
1206                            "DMA sg list with different entry count "
1207                            "[map count=%d] [sync count=%d]\n",
1208                            entry->sg_call_ents, ref->sg_call_ents);
1209         }
1210
1211 out:
1212         put_hash_bucket(bucket, flags);
1213 }
1214
1215 static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1216 {
1217 #ifdef CONFIG_DMA_API_DEBUG_SG
1218         unsigned int max_seg = dma_get_max_seg_size(dev);
1219         u64 start, end, boundary = dma_get_seg_boundary(dev);
1220
1221         /*
1222          * Either the driver forgot to set dma_parms appropriately, or
1223          * whoever generated the list forgot to check them.
1224          */
1225         if (sg->length > max_seg)
1226                 err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1227                            sg->length, max_seg);
1228         /*
1229          * In some cases this could potentially be the DMA API
1230          * implementation's fault, but it would usually imply that
1231          * the scatterlist was built inappropriately to begin with.
1232          */
1233         start = sg_dma_address(sg);
1234         end = start + sg_dma_len(sg) - 1;
1235         if ((start ^ end) & ~boundary)
1236                 err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1237                            start, end, boundary);
1238 #endif
1239 }
1240
1241 void debug_dma_map_single(struct device *dev, const void *addr,
1242                             unsigned long len)
1243 {
1244         if (unlikely(dma_debug_disabled()))
1245                 return;
1246
1247         if (!virt_addr_valid(addr))
1248                 err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1249                            addr, len);
1250
1251         if (is_vmalloc_addr(addr))
1252                 err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1253                            addr, len);
1254 }
1255 EXPORT_SYMBOL(debug_dma_map_single);
1256
1257 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1258                         size_t size, int direction, dma_addr_t dma_addr)
1259 {
1260         struct dma_debug_entry *entry;
1261
1262         if (unlikely(dma_debug_disabled()))
1263                 return;
1264
1265         if (dma_mapping_error(dev, dma_addr))
1266                 return;
1267
1268         entry = dma_entry_alloc();
1269         if (!entry)
1270                 return;
1271
1272         entry->dev       = dev;
1273         entry->type      = dma_debug_single;
1274         entry->pfn       = page_to_pfn(page);
1275         entry->offset    = offset,
1276         entry->dev_addr  = dma_addr;
1277         entry->size      = size;
1278         entry->direction = direction;
1279         entry->map_err_type = MAP_ERR_NOT_CHECKED;
1280
1281         check_for_stack(dev, page, offset);
1282
1283         if (!PageHighMem(page)) {
1284                 void *addr = page_address(page) + offset;
1285
1286                 check_for_illegal_area(dev, addr, size);
1287         }
1288
1289         add_dma_entry(entry);
1290 }
1291 EXPORT_SYMBOL(debug_dma_map_page);
1292
1293 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1294 {
1295         struct dma_debug_entry ref;
1296         struct dma_debug_entry *entry;
1297         struct hash_bucket *bucket;
1298         unsigned long flags;
1299
1300         if (unlikely(dma_debug_disabled()))
1301                 return;
1302
1303         ref.dev = dev;
1304         ref.dev_addr = dma_addr;
1305         bucket = get_hash_bucket(&ref, &flags);
1306
1307         list_for_each_entry(entry, &bucket->list, list) {
1308                 if (!exact_match(&ref, entry))
1309                         continue;
1310
1311                 /*
1312                  * The same physical address can be mapped multiple
1313                  * times. Without a hardware IOMMU this results in the
1314                  * same device addresses being put into the dma-debug
1315                  * hash multiple times too. This can result in false
1316                  * positives being reported. Therefore we implement a
1317                  * best-fit algorithm here which updates the first entry
1318                  * from the hash which fits the reference value and is
1319                  * not currently listed as being checked.
1320                  */
1321                 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1322                         entry->map_err_type = MAP_ERR_CHECKED;
1323                         break;
1324                 }
1325         }
1326
1327         put_hash_bucket(bucket, flags);
1328 }
1329 EXPORT_SYMBOL(debug_dma_mapping_error);
1330
1331 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1332                           size_t size, int direction)
1333 {
1334         struct dma_debug_entry ref = {
1335                 .type           = dma_debug_single,
1336                 .dev            = dev,
1337                 .dev_addr       = addr,
1338                 .size           = size,
1339                 .direction      = direction,
1340         };
1341
1342         if (unlikely(dma_debug_disabled()))
1343                 return;
1344         check_unmap(&ref);
1345 }
1346 EXPORT_SYMBOL(debug_dma_unmap_page);
1347
1348 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1349                       int nents, int mapped_ents, int direction)
1350 {
1351         struct dma_debug_entry *entry;
1352         struct scatterlist *s;
1353         int i;
1354
1355         if (unlikely(dma_debug_disabled()))
1356                 return;
1357
1358         for_each_sg(sg, s, mapped_ents, i) {
1359                 entry = dma_entry_alloc();
1360                 if (!entry)
1361                         return;
1362
1363                 entry->type           = dma_debug_sg;
1364                 entry->dev            = dev;
1365                 entry->pfn            = page_to_pfn(sg_page(s));
1366                 entry->offset         = s->offset,
1367                 entry->size           = sg_dma_len(s);
1368                 entry->dev_addr       = sg_dma_address(s);
1369                 entry->direction      = direction;
1370                 entry->sg_call_ents   = nents;
1371                 entry->sg_mapped_ents = mapped_ents;
1372
1373                 check_for_stack(dev, sg_page(s), s->offset);
1374
1375                 if (!PageHighMem(sg_page(s))) {
1376                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1377                 }
1378
1379                 check_sg_segment(dev, s);
1380
1381                 add_dma_entry(entry);
1382         }
1383 }
1384 EXPORT_SYMBOL(debug_dma_map_sg);
1385
1386 static int get_nr_mapped_entries(struct device *dev,
1387                                  struct dma_debug_entry *ref)
1388 {
1389         struct dma_debug_entry *entry;
1390         struct hash_bucket *bucket;
1391         unsigned long flags;
1392         int mapped_ents;
1393
1394         bucket       = get_hash_bucket(ref, &flags);
1395         entry        = bucket_find_exact(bucket, ref);
1396         mapped_ents  = 0;
1397
1398         if (entry)
1399                 mapped_ents = entry->sg_mapped_ents;
1400         put_hash_bucket(bucket, flags);
1401
1402         return mapped_ents;
1403 }
1404
1405 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1406                         int nelems, int dir)
1407 {
1408         struct scatterlist *s;
1409         int mapped_ents = 0, i;
1410
1411         if (unlikely(dma_debug_disabled()))
1412                 return;
1413
1414         for_each_sg(sglist, s, nelems, i) {
1415
1416                 struct dma_debug_entry ref = {
1417                         .type           = dma_debug_sg,
1418                         .dev            = dev,
1419                         .pfn            = page_to_pfn(sg_page(s)),
1420                         .offset         = s->offset,
1421                         .dev_addr       = sg_dma_address(s),
1422                         .size           = sg_dma_len(s),
1423                         .direction      = dir,
1424                         .sg_call_ents   = nelems,
1425                 };
1426
1427                 if (mapped_ents && i >= mapped_ents)
1428                         break;
1429
1430                 if (!i)
1431                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1432
1433                 check_unmap(&ref);
1434         }
1435 }
1436 EXPORT_SYMBOL(debug_dma_unmap_sg);
1437
1438 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1439                               dma_addr_t dma_addr, void *virt)
1440 {
1441         struct dma_debug_entry *entry;
1442
1443         if (unlikely(dma_debug_disabled()))
1444                 return;
1445
1446         if (unlikely(virt == NULL))
1447                 return;
1448
1449         /* handle vmalloc and linear addresses */
1450         if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1451                 return;
1452
1453         entry = dma_entry_alloc();
1454         if (!entry)
1455                 return;
1456
1457         entry->type      = dma_debug_coherent;
1458         entry->dev       = dev;
1459         entry->offset    = offset_in_page(virt);
1460         entry->size      = size;
1461         entry->dev_addr  = dma_addr;
1462         entry->direction = DMA_BIDIRECTIONAL;
1463
1464         if (is_vmalloc_addr(virt))
1465                 entry->pfn = vmalloc_to_pfn(virt);
1466         else
1467                 entry->pfn = page_to_pfn(virt_to_page(virt));
1468
1469         add_dma_entry(entry);
1470 }
1471
1472 void debug_dma_free_coherent(struct device *dev, size_t size,
1473                          void *virt, dma_addr_t addr)
1474 {
1475         struct dma_debug_entry ref = {
1476                 .type           = dma_debug_coherent,
1477                 .dev            = dev,
1478                 .offset         = offset_in_page(virt),
1479                 .dev_addr       = addr,
1480                 .size           = size,
1481                 .direction      = DMA_BIDIRECTIONAL,
1482         };
1483
1484         /* handle vmalloc and linear addresses */
1485         if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1486                 return;
1487
1488         if (is_vmalloc_addr(virt))
1489                 ref.pfn = vmalloc_to_pfn(virt);
1490         else
1491                 ref.pfn = page_to_pfn(virt_to_page(virt));
1492
1493         if (unlikely(dma_debug_disabled()))
1494                 return;
1495
1496         check_unmap(&ref);
1497 }
1498
1499 void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1500                             int direction, dma_addr_t dma_addr)
1501 {
1502         struct dma_debug_entry *entry;
1503
1504         if (unlikely(dma_debug_disabled()))
1505                 return;
1506
1507         entry = dma_entry_alloc();
1508         if (!entry)
1509                 return;
1510
1511         entry->type             = dma_debug_resource;
1512         entry->dev              = dev;
1513         entry->pfn              = PHYS_PFN(addr);
1514         entry->offset           = offset_in_page(addr);
1515         entry->size             = size;
1516         entry->dev_addr         = dma_addr;
1517         entry->direction        = direction;
1518         entry->map_err_type     = MAP_ERR_NOT_CHECKED;
1519
1520         add_dma_entry(entry);
1521 }
1522 EXPORT_SYMBOL(debug_dma_map_resource);
1523
1524 void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1525                               size_t size, int direction)
1526 {
1527         struct dma_debug_entry ref = {
1528                 .type           = dma_debug_resource,
1529                 .dev            = dev,
1530                 .dev_addr       = dma_addr,
1531                 .size           = size,
1532                 .direction      = direction,
1533         };
1534
1535         if (unlikely(dma_debug_disabled()))
1536                 return;
1537
1538         check_unmap(&ref);
1539 }
1540 EXPORT_SYMBOL(debug_dma_unmap_resource);
1541
1542 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1543                                    size_t size, int direction)
1544 {
1545         struct dma_debug_entry ref;
1546
1547         if (unlikely(dma_debug_disabled()))
1548                 return;
1549
1550         ref.type         = dma_debug_single;
1551         ref.dev          = dev;
1552         ref.dev_addr     = dma_handle;
1553         ref.size         = size;
1554         ref.direction    = direction;
1555         ref.sg_call_ents = 0;
1556
1557         check_sync(dev, &ref, true);
1558 }
1559 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1560
1561 void debug_dma_sync_single_for_device(struct device *dev,
1562                                       dma_addr_t dma_handle, size_t size,
1563                                       int direction)
1564 {
1565         struct dma_debug_entry ref;
1566
1567         if (unlikely(dma_debug_disabled()))
1568                 return;
1569
1570         ref.type         = dma_debug_single;
1571         ref.dev          = dev;
1572         ref.dev_addr     = dma_handle;
1573         ref.size         = size;
1574         ref.direction    = direction;
1575         ref.sg_call_ents = 0;
1576
1577         check_sync(dev, &ref, false);
1578 }
1579 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1580
1581 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1582                                int nelems, int direction)
1583 {
1584         struct scatterlist *s;
1585         int mapped_ents = 0, i;
1586
1587         if (unlikely(dma_debug_disabled()))
1588                 return;
1589
1590         for_each_sg(sg, s, nelems, i) {
1591
1592                 struct dma_debug_entry ref = {
1593                         .type           = dma_debug_sg,
1594                         .dev            = dev,
1595                         .pfn            = page_to_pfn(sg_page(s)),
1596                         .offset         = s->offset,
1597                         .dev_addr       = sg_dma_address(s),
1598                         .size           = sg_dma_len(s),
1599                         .direction      = direction,
1600                         .sg_call_ents   = nelems,
1601                 };
1602
1603                 if (!i)
1604                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1605
1606                 if (i >= mapped_ents)
1607                         break;
1608
1609                 check_sync(dev, &ref, true);
1610         }
1611 }
1612 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1613
1614 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1615                                   int nelems, int direction)
1616 {
1617         struct scatterlist *s;
1618         int mapped_ents = 0, i;
1619
1620         if (unlikely(dma_debug_disabled()))
1621                 return;
1622
1623         for_each_sg(sg, s, nelems, i) {
1624
1625                 struct dma_debug_entry ref = {
1626                         .type           = dma_debug_sg,
1627                         .dev            = dev,
1628                         .pfn            = page_to_pfn(sg_page(s)),
1629                         .offset         = s->offset,
1630                         .dev_addr       = sg_dma_address(s),
1631                         .size           = sg_dma_len(s),
1632                         .direction      = direction,
1633                         .sg_call_ents   = nelems,
1634                 };
1635                 if (!i)
1636                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1637
1638                 if (i >= mapped_ents)
1639                         break;
1640
1641                 check_sync(dev, &ref, false);
1642         }
1643 }
1644 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1645
1646 static int __init dma_debug_driver_setup(char *str)
1647 {
1648         int i;
1649
1650         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1651                 current_driver_name[i] = *str;
1652                 if (*str == 0)
1653                         break;
1654         }
1655
1656         if (current_driver_name[0])
1657                 pr_info("enable driver filter for driver [%s]\n",
1658                         current_driver_name);
1659
1660
1661         return 1;
1662 }
1663 __setup("dma_debug_driver=", dma_debug_driver_setup);