lib/dma-debug.c

   1 /*
   2  * Copyright (C) 2008 Advanced Micro Devices, Inc.
   3  *
   4  * Author: Joerg Roedel <joerg.roedel@amd.com>
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 as published
   8  * by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  18  */
  19
  20 #include <linux/scatterlist.h>
  21 #include <linux/dma-mapping.h>
  22 #include <linux/stacktrace.h>
  23 #include <linux/dma-debug.h>
  24 #include <linux/spinlock.h>
  25 #include <linux/debugfs.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/export.h>
  28 #include <linux/device.h>
  29 #include <linux/types.h>
  30 #include <linux/sched.h>
  31 #include <linux/ctype.h>
  32 #include <linux/list.h>
  33 #include <linux/slab.h>
  34
  35 #include <asm/sections.h>
  36
  37 #define HASH_SIZE       1024ULL
  38 #define HASH_FN_SHIFT   13
  39 #define HASH_FN_MASK    (HASH_SIZE - 1)
  40
  41 enum {
  42         dma_debug_single,
  43         dma_debug_page,
  44         dma_debug_sg,
  45         dma_debug_coherent,
  46 };
  47
  48 enum map_err_types {
  49         MAP_ERR_CHECK_NOT_APPLICABLE,
  50         MAP_ERR_NOT_CHECKED,
  51         MAP_ERR_CHECKED,
  52 };
  53
  54 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
  55
  56 /**
  57  * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
  58  * @list: node on pre-allocated free_entries list
  59  * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
  60  * @type: single, page, sg, coherent
  61  * @pfn: page frame of the start address
  62  * @offset: offset of mapping relative to pfn
  63  * @size: length of the mapping
  64  * @direction: enum dma_data_direction
  65  * @sg_call_ents: 'nents' from dma_map_sg
  66  * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
  67  * @map_err_type: track whether dma_mapping_error() was checked
  68  * @stacktrace: support backtraces when a violation is detected
  69  */
  70 struct dma_debug_entry {
  71         struct list_head list;
  72         struct device    *dev;
  73         int              type;
  74         unsigned long    pfn;
  75         size_t           offset;
  76         u64              dev_addr;
  77         u64              size;
  78         int              direction;
  79         int              sg_call_ents;
  80         int              sg_mapped_ents;
  81         enum map_err_types  map_err_type;
  82 #ifdef CONFIG_STACKTRACE
  83         struct           stack_trace stacktrace;
  84         unsigned long    st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
  85 #endif
  86 };
  87
  88 typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
  89
  90 struct hash_bucket {
  91         struct list_head list;
  92         spinlock_t lock;
  93 } ____cacheline_aligned_in_smp;
  94
  95 /* Hash list to save the allocated dma addresses */
  96 static struct hash_bucket dma_entry_hash[HASH_SIZE];
  97 /* List of pre-allocated dma_debug_entry's */
  98 static LIST_HEAD(free_entries);
  99 /* Lock for the list above */
 100 static DEFINE_SPINLOCK(free_entries_lock);
 101
 102 /* Global disable flag - will be set in case of an error */
 103 static u32 global_disable __read_mostly;
 104
 105 /* Early initialization disable flag, set at the end of dma_debug_init */
 106 static bool dma_debug_initialized __read_mostly;
 107
 108 static inline bool dma_debug_disabled(void)
 109 {
 110         return global_disable || !dma_debug_initialized;
 111 }
 112
 113 /* Global error count */
 114 static u32 error_count;
 115
 116 /* Global error show enable*/
 117 static u32 show_all_errors __read_mostly;
 118 /* Number of errors to show */
 119 static u32 show_num_errors = 1;
 120
 121 static u32 num_free_entries;
 122 static u32 min_free_entries;
 123 static u32 nr_total_entries;
 124
 125 /* number of preallocated entries requested by kernel cmdline */
 126 static u32 req_entries;
 127
 128 /* debugfs dentry's for the stuff above */
 129 static struct dentry *dma_debug_dent        __read_mostly;
 130 static struct dentry *global_disable_dent   __read_mostly;
 131 static struct dentry *error_count_dent      __read_mostly;
 132 static struct dentry *show_all_errors_dent  __read_mostly;
 133 static struct dentry *show_num_errors_dent  __read_mostly;
 134 static struct dentry *num_free_entries_dent __read_mostly;
 135 static struct dentry *min_free_entries_dent __read_mostly;
 136 static struct dentry *filter_dent           __read_mostly;
 137
 138 /* per-driver filter related state */
 139
 140 #define NAME_MAX_LEN    64
 141
 142 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
 143 static struct device_driver *current_driver                    __read_mostly;
 144
 145 static DEFINE_RWLOCK(driver_name_lock);
 146
 147 static const char *const maperr2str[] = {
 148         [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
 149         [MAP_ERR_NOT_CHECKED] = "dma map error not checked",
 150         [MAP_ERR_CHECKED] = "dma map error checked",
 151 };
 152
 153 static const char *type2name[4] = { "single", "page",
 154                                     "scather-gather", "coherent" };
 155
 156 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
 157                                    "DMA_FROM_DEVICE", "DMA_NONE" };
 158
 159 /*
 160  * The access to some variables in this macro is racy. We can't use atomic_t
 161  * here because all these variables are exported to debugfs. Some of them even
 162  * writeable. This is also the reason why a lock won't help much. But anyway,
 163  * the races are no big deal. Here is why:
 164  *
 165  *   error_count: the addition is racy, but the worst thing that can happen is
 166  *                that we don't count some errors
 167  *   show_num_errors: the subtraction is racy. Also no big deal because in
 168  *                    worst case this will result in one warning more in the
 169  *                    system log than the user configured. This variable is
 170  *                    writeable via debugfs.
 171  */
 172 static inline void dump_entry_trace(struct dma_debug_entry *entry)
 173 {
 174 #ifdef CONFIG_STACKTRACE
 175         if (entry) {
 176                 pr_warning("Mapped at:\n");
 177                 print_stack_trace(&entry->stacktrace, 0);
 178         }
 179 #endif
 180 }
 181
 182 static bool driver_filter(struct device *dev)
 183 {
 184         struct device_driver *drv;
 185         unsigned long flags;
 186         bool ret;
 187
 188         /* driver filter off */
 189         if (likely(!current_driver_name[0]))
 190                 return true;
 191
 192         /* driver filter on and initialized */
 193         if (current_driver && dev && dev->driver == current_driver)
 194                 return true;
 195
 196         /* driver filter on, but we can't filter on a NULL device... */
 197         if (!dev)
 198                 return false;
 199
 200         if (current_driver || !current_driver_name[0])
 201                 return false;
 202
 203         /* driver filter on but not yet initialized */
 204         drv = dev->driver;
 205         if (!drv)
 206                 return false;
 207
 208         /* lock to protect against change of current_driver_name */
 209         read_lock_irqsave(&driver_name_lock, flags);
 210
 211         ret = false;
 212         if (drv->name &&
 213             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 214                 current_driver = drv;
 215                 ret = true;
 216         }
 217
 218         read_unlock_irqrestore(&driver_name_lock, flags);
 219
 220         return ret;
 221 }
 222
 223 #define err_printk(dev, entry, format, arg...) do {                     \
 224                 error_count += 1;                                       \
 225                 if (driver_filter(dev) &&                               \
 226                     (show_all_errors || show_num_errors > 0)) {         \
 227                         WARN(1, "%s %s: " format,                       \
 228                              dev ? dev_driver_string(dev) : "NULL",     \
 229                              dev ? dev_name(dev) : "NULL", ## arg);     \
 230                         dump_entry_trace(entry);                        \
 231                 }                                                       \
 232                 if (!show_all_errors && show_num_errors > 0)            \
 233                         show_num_errors -= 1;                           \
 234         } while (0);
 235
 236 /*
 237  * Hash related functions
 238  *
 239  * Every DMA-API request is saved into a struct dma_debug_entry. To
 240  * have quick access to these structs they are stored into a hash.
 241  */
 242 static int hash_fn(struct dma_debug_entry *entry)
 243 {
 244         /*
 245          * Hash function is based on the dma address.
 246          * We use bits 20-27 here as the index into the hash
 247          */
 248         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 249 }
 250
 251 /*
 252  * Request exclusive access to a hash bucket for a given dma_debug_entry.
 253  */
 254 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 255                                            unsigned long *flags)
 256 {
 257         int idx = hash_fn(entry);
 258         unsigned long __flags;
 259
 260         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 261         *flags = __flags;
 262         return &dma_entry_hash[idx];
 263 }
 264
 265 /*
 266  * Give up exclusive access to the hash bucket
 267  */
 268 static void put_hash_bucket(struct hash_bucket *bucket,
 269                             unsigned long *flags)
 270 {
 271         unsigned long __flags = *flags;
 272
 273         spin_unlock_irqrestore(&bucket->lock, __flags);
 274 }
 275
 276 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
 277 {
 278         return ((a->dev_addr == b->dev_addr) &&
 279                 (a->dev == b->dev)) ? true : false;
 280 }
 281
 282 static bool containing_match(struct dma_debug_entry *a,
 283                              struct dma_debug_entry *b)
 284 {
 285         if (a->dev != b->dev)
 286                 return false;
 287
 288         if ((b->dev_addr <= a->dev_addr) &&
 289             ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
 290                 return true;
 291
 292         return false;
 293 }
 294
 295 /*
 296  * Search a given entry in the hash bucket list
 297  */
 298 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
 299                                                   struct dma_debug_entry *ref,
 300                                                   match_fn match)
 301 {
 302         struct dma_debug_entry *entry, *ret = NULL;
 303         int matches = 0, match_lvl, last_lvl = -1;
 304
 305         list_for_each_entry(entry, &bucket->list, list) {
 306                 if (!match(ref, entry))
 307                         continue;
 308
 309                 /*
 310                  * Some drivers map the same physical address multiple
 311                  * times. Without a hardware IOMMU this results in the
 312                  * same device addresses being put into the dma-debug
 313                  * hash multiple times too. This can result in false
 314                  * positives being reported. Therefore we implement a
 315                  * best-fit algorithm here which returns the entry from
 316                  * the hash which fits best to the reference value
 317                  * instead of the first-fit.
 318                  */
 319                 matches += 1;
 320                 match_lvl = 0;
 321                 entry->size         == ref->size         ? ++match_lvl : 0;
 322                 entry->type         == ref->type         ? ++match_lvl : 0;
 323                 entry->direction    == ref->direction    ? ++match_lvl : 0;
 324                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
 325
 326                 if (match_lvl == 4) {
 327                         /* perfect-fit - return the result */
 328                         return entry;
 329                 } else if (match_lvl > last_lvl) {
 330                         /*
 331                          * We found an entry that fits better then the
 332                          * previous one or it is the 1st match.
 333                          */
 334                         last_lvl = match_lvl;
 335                         ret      = entry;
 336                 }
 337         }
 338
 339         /*
 340          * If we have multiple matches but no perfect-fit, just return
 341          * NULL.
 342          */
 343         ret = (matches == 1) ? ret : NULL;
 344
 345         return ret;
 346 }
 347
 348 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
 349                                                  struct dma_debug_entry *ref)
 350 {
 351         return __hash_bucket_find(bucket, ref, exact_match);
 352 }
 353
 354 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 355                                                    struct dma_debug_entry *ref,
 356                                                    unsigned long *flags)
 357 {
 358
 359         unsigned int max_range = dma_get_max_seg_size(ref->dev);
 360         struct dma_debug_entry *entry, index = *ref;
 361         unsigned int range = 0;
 362
 363         while (range <= max_range) {
 364                 entry = __hash_bucket_find(*bucket, &index, containing_match);
 365
 366                 if (entry)
 367                         return entry;
 368
 369                 /*
 370                  * Nothing found, go back a hash bucket
 371                  */
 372                 put_hash_bucket(*bucket, flags);
 373                 range          += (1 << HASH_FN_SHIFT);
 374                 index.dev_addr -= (1 << HASH_FN_SHIFT);
 375                 *bucket = get_hash_bucket(&index, flags);
 376         }
 377
 378         return NULL;
 379 }
 380
 381 /*
 382  * Add an entry to a hash bucket
 383  */
 384 static void hash_bucket_add(struct hash_bucket *bucket,
 385                             struct dma_debug_entry *entry)
 386 {
 387         list_add_tail(&entry->list, &bucket->list);
 388 }
 389
 390 /*
 391  * Remove entry from a hash bucket list
 392  */
 393 static void hash_bucket_del(struct dma_debug_entry *entry)
 394 {
 395         list_del(&entry->list);
 396 }
 397
 398 static unsigned long long phys_addr(struct dma_debug_entry *entry)
 399 {
 400         return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
 401 }
 402
 403 /*
 404  * Dump mapping entries for debugging purposes
 405  */
 406 void debug_dma_dump_mappings(struct device *dev)
 407 {
 408         int idx;
 409
 410         for (idx = 0; idx < HASH_SIZE; idx++) {
 411                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 412                 struct dma_debug_entry *entry;
 413                 unsigned long flags;
 414
 415                 spin_lock_irqsave(&bucket->lock, flags);
 416
 417                 list_for_each_entry(entry, &bucket->list, list) {
 418                         if (!dev || dev == entry->dev) {
 419                                 dev_info(entry->dev,
 420                                          "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
 421                                          type2name[entry->type], idx,
 422                                          phys_addr(entry), entry->pfn,
 423                                          entry->dev_addr, entry->size,
 424                                          dir2name[entry->direction],
 425                                          maperr2str[entry->map_err_type]);
 426                         }
 427                 }
 428
 429                 spin_unlock_irqrestore(&bucket->lock, flags);
 430         }
 431 }
 432 EXPORT_SYMBOL(debug_dma_dump_mappings);
 433
 434 /*
 435  * For each mapping (initial cacheline in the case of
 436  * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
 437  * scatterlist, or the cacheline specified in dma_map_single) insert
 438  * into this tree using the cacheline as the key. At
 439  * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
 440  * the entry already exists at insertion time add a tag as a reference
 441  * count for the overlapping mappings.  For now, the overlap tracking
 442  * just ensures that 'unmaps' balance 'maps' before marking the
 443  * cacheline idle, but we should also be flagging overlaps as an API
 444  * violation.
 445  *
 446  * Memory usage is mostly constrained by the maximum number of available
 447  * dma-debug entries in that we need a free dma_debug_entry before
 448  * inserting into the tree.  In the case of dma_map_page and
 449  * dma_alloc_coherent there is only one dma_debug_entry and one
 450  * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
 451  * other hand, consumes a single dma_debug_entry, but inserts 'nents'
 452  * entries into the tree.
 453  *
 454  * At any time debug_dma_assert_idle() can be called to trigger a
 455  * warning if any cachelines in the given page are in the active set.
 456  */
 457 static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
 458 static DEFINE_SPINLOCK(radix_lock);
 459 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
 460 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
 461 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
 462
 463 static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
 464 {
 465         return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
 466                 (entry->offset >> L1_CACHE_SHIFT);
 467 }
 468
 469 static int active_cacheline_read_overlap(phys_addr_t cln)
 470 {
 471         int overlap = 0, i;
 472
 473         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 474                 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
 475                         overlap |= 1 << i;
 476         return overlap;
 477 }
 478
 479 static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
 480 {
 481         int i;
 482
 483         if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
 484                 return overlap;
 485
 486         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 487                 if (overlap & 1 << i)
 488                         radix_tree_tag_set(&dma_active_cacheline, cln, i);
 489                 else
 490                         radix_tree_tag_clear(&dma_active_cacheline, cln, i);
 491
 492         return overlap;
 493 }
 494
 495 static void active_cacheline_inc_overlap(phys_addr_t cln)
 496 {
 497         int overlap = active_cacheline_read_overlap(cln);
 498
 499         overlap = active_cacheline_set_overlap(cln, ++overlap);
 500
 501         /* If we overflowed the overlap counter then we're potentially
 502          * leaking dma-mappings.  Otherwise, if maps and unmaps are
 503          * balanced then this overflow may cause false negatives in
 504          * debug_dma_assert_idle() as the cacheline may be marked idle
 505          * prematurely.
 506          */
 507         WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
 508                   "DMA-API: exceeded %d overlapping mappings of cacheline %pa\n",
 509                   ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 510 }
 511
 512 static int active_cacheline_dec_overlap(phys_addr_t cln)
 513 {
 514         int overlap = active_cacheline_read_overlap(cln);
 515
 516         return active_cacheline_set_overlap(cln, --overlap);
 517 }
 518
 519 static int active_cacheline_insert(struct dma_debug_entry *entry)
 520 {
 521         phys_addr_t cln = to_cacheline_number(entry);
 522         unsigned long flags;
 523         int rc;
 524
 525         /* If the device is not writing memory then we don't have any
 526          * concerns about the cpu consuming stale data.  This mitigates
 527          * legitimate usages of overlapping mappings.
 528          */
 529         if (entry->direction == DMA_TO_DEVICE)
 530                 return 0;
 531
 532         spin_lock_irqsave(&radix_lock, flags);
 533         rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
 534         if (rc == -EEXIST)
 535                 active_cacheline_inc_overlap(cln);
 536         spin_unlock_irqrestore(&radix_lock, flags);
 537
 538         return rc;
 539 }
 540
 541 static void active_cacheline_remove(struct dma_debug_entry *entry)
 542 {
 543         phys_addr_t cln = to_cacheline_number(entry);
 544         unsigned long flags;
 545
 546         /* ...mirror the insert case */
 547         if (entry->direction == DMA_TO_DEVICE)
 548                 return;
 549
 550         spin_lock_irqsave(&radix_lock, flags);
 551         /* since we are counting overlaps the final put of the
 552          * cacheline will occur when the overlap count is 0.
 553          * active_cacheline_dec_overlap() returns -1 in that case
 554          */
 555         if (active_cacheline_dec_overlap(cln) < 0)
 556                 radix_tree_delete(&dma_active_cacheline, cln);
 557         spin_unlock_irqrestore(&radix_lock, flags);
 558 }
 559
 560 /**
 561  * debug_dma_assert_idle() - assert that a page is not undergoing dma
 562  * @page: page to lookup in the dma_active_cacheline tree
 563  *
 564  * Place a call to this routine in cases where the cpu touching the page
 565  * before the dma completes (page is dma_unmapped) will lead to data
 566  * corruption.
 567  */
 568 void debug_dma_assert_idle(struct page *page)
 569 {
 570         static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
 571         struct dma_debug_entry *entry = NULL;
 572         void **results = (void **) &ents;
 573         unsigned int nents, i;
 574         unsigned long flags;
 575         phys_addr_t cln;
 576
 577         if (!page)
 578                 return;
 579
 580         cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
 581         spin_lock_irqsave(&radix_lock, flags);
 582         nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
 583                                        CACHELINES_PER_PAGE);
 584         for (i = 0; i < nents; i++) {
 585                 phys_addr_t ent_cln = to_cacheline_number(ents[i]);
 586
 587                 if (ent_cln == cln) {
 588                         entry = ents[i];
 589                         break;
 590                 } else if (ent_cln >= cln + CACHELINES_PER_PAGE)
 591                         break;
 592         }
 593         spin_unlock_irqrestore(&radix_lock, flags);
 594
 595         if (!entry)
 596                 return;
 597
 598         cln = to_cacheline_number(entry);
 599         err_printk(entry->dev, entry,
 600                    "DMA-API: cpu touching an active dma mapped cacheline [cln=%pa]\n",
 601                    &cln);
 602 }
 603
 604 /*
 605  * Wrapper function for adding an entry to the hash.
 606  * This function takes care of locking itself.
 607  */
 608 static void add_dma_entry(struct dma_debug_entry *entry)
 609 {
 610         struct hash_bucket *bucket;
 611         unsigned long flags;
 612         int rc;
 613
 614         bucket = get_hash_bucket(entry, &flags);
 615         hash_bucket_add(bucket, entry);
 616         put_hash_bucket(bucket, &flags);
 617
 618         rc = active_cacheline_insert(entry);
 619         if (rc == -ENOMEM) {
 620                 pr_err("DMA-API: cacheline tracking ENOMEM, dma-debug disabled\n");
 621                 global_disable = true;
 622         }
 623
 624         /* TODO: report -EEXIST errors here as overlapping mappings are
 625          * not supported by the DMA API
 626          */
 627 }
 628
 629 static struct dma_debug_entry *__dma_entry_alloc(void)
 630 {
 631         struct dma_debug_entry *entry;
 632
 633         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 634         list_del(&entry->list);
 635         memset(entry, 0, sizeof(*entry));
 636
 637         num_free_entries -= 1;
 638         if (num_free_entries < min_free_entries)
 639                 min_free_entries = num_free_entries;
 640
 641         return entry;
 642 }
 643
 644 /* struct dma_entry allocator
 645  *
 646  * The next two functions implement the allocator for
 647  * struct dma_debug_entries.
 648  */
 649 static struct dma_debug_entry *dma_entry_alloc(void)
 650 {
 651         struct dma_debug_entry *entry;
 652         unsigned long flags;
 653
 654         spin_lock_irqsave(&free_entries_lock, flags);
 655
 656         if (list_empty(&free_entries)) {
 657                 pr_err("DMA-API: debugging out of memory - disabling\n");
 658                 global_disable = true;
 659                 spin_unlock_irqrestore(&free_entries_lock, flags);
 660                 return NULL;
 661         }
 662
 663         entry = __dma_entry_alloc();
 664
 665         spin_unlock_irqrestore(&free_entries_lock, flags);
 666
 667 #ifdef CONFIG_STACKTRACE
 668         entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
 669         entry->stacktrace.entries = entry->st_entries;
 670         entry->stacktrace.skip = 2;
 671         save_stack_trace(&entry->stacktrace);
 672 #endif
 673
 674         return entry;
 675 }
 676
 677 static void dma_entry_free(struct dma_debug_entry *entry)
 678 {
 679         unsigned long flags;
 680
 681         active_cacheline_remove(entry);
 682
 683         /*
 684          * add to beginning of the list - this way the entries are
 685          * more likely cache hot when they are reallocated.
 686          */
 687         spin_lock_irqsave(&free_entries_lock, flags);
 688         list_add(&entry->list, &free_entries);
 689         num_free_entries += 1;
 690         spin_unlock_irqrestore(&free_entries_lock, flags);
 691 }
 692
 693 int dma_debug_resize_entries(u32 num_entries)
 694 {
 695         int i, delta, ret = 0;
 696         unsigned long flags;
 697         struct dma_debug_entry *entry;
 698         LIST_HEAD(tmp);
 699
 700         spin_lock_irqsave(&free_entries_lock, flags);
 701
 702         if (nr_total_entries < num_entries) {
 703                 delta = num_entries - nr_total_entries;
 704
 705                 spin_unlock_irqrestore(&free_entries_lock, flags);
 706
 707                 for (i = 0; i < delta; i++) {
 708                         entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 709                         if (!entry)
 710                                 break;
 711
 712                         list_add_tail(&entry->list, &tmp);
 713                 }
 714
 715                 spin_lock_irqsave(&free_entries_lock, flags);
 716
 717                 list_splice(&tmp, &free_entries);
 718                 nr_total_entries += i;
 719                 num_free_entries += i;
 720         } else {
 721                 delta = nr_total_entries - num_entries;
 722
 723                 for (i = 0; i < delta && !list_empty(&free_entries); i++) {
 724                         entry = __dma_entry_alloc();
 725                         kfree(entry);
 726                 }
 727
 728                 nr_total_entries -= i;
 729         }
 730
 731         if (nr_total_entries != num_entries)
 732                 ret = 1;
 733
 734         spin_unlock_irqrestore(&free_entries_lock, flags);
 735
 736         return ret;
 737 }
 738 EXPORT_SYMBOL(dma_debug_resize_entries);
 739
 740 /*
 741  * DMA-API debugging init code
 742  *
 743  * The init code does two things:
 744  *   1. Initialize core data structures
 745  *   2. Preallocate a given number of dma_debug_entry structs
 746  */
 747
 748 static int prealloc_memory(u32 num_entries)
 749 {
 750         struct dma_debug_entry *entry, *next_entry;
 751         int i;
 752
 753         for (i = 0; i < num_entries; ++i) {
 754                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 755                 if (!entry)
 756                         goto out_err;
 757
 758                 list_add_tail(&entry->list, &free_entries);
 759         }
 760
 761         num_free_entries = num_entries;
 762         min_free_entries = num_entries;
 763
 764         pr_info("DMA-API: preallocated %d debug entries\n", num_entries);
 765
 766         return 0;
 767
 768 out_err:
 769
 770         list_for_each_entry_safe(entry, next_entry, &free_entries, list) {
 771                 list_del(&entry->list);
 772                 kfree(entry);
 773         }
 774
 775         return -ENOMEM;
 776 }
 777
 778 static ssize_t filter_read(struct file *file, char __user *user_buf,
 779                            size_t count, loff_t *ppos)
 780 {
 781         char buf[NAME_MAX_LEN + 1];
 782         unsigned long flags;
 783         int len;
 784
 785         if (!current_driver_name[0])
 786                 return 0;
 787
 788         /*
 789          * We can't copy to userspace directly because current_driver_name can
 790          * only be read under the driver_name_lock with irqs disabled. So
 791          * create a temporary copy first.
 792          */
 793         read_lock_irqsave(&driver_name_lock, flags);
 794         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 795         read_unlock_irqrestore(&driver_name_lock, flags);
 796
 797         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 798 }
 799
 800 static ssize_t filter_write(struct file *file, const char __user *userbuf,
 801                             size_t count, loff_t *ppos)
 802 {
 803         char buf[NAME_MAX_LEN];
 804         unsigned long flags;
 805         size_t len;
 806         int i;
 807
 808         /*
 809          * We can't copy from userspace directly. Access to
 810          * current_driver_name is protected with a write_lock with irqs
 811          * disabled. Since copy_from_user can fault and may sleep we
 812          * need to copy to temporary buffer first
 813          */
 814         len = min(count, (size_t)(NAME_MAX_LEN - 1));
 815         if (copy_from_user(buf, userbuf, len))
 816                 return -EFAULT;
 817
 818         buf[len] = 0;
 819
 820         write_lock_irqsave(&driver_name_lock, flags);
 821
 822         /*
 823          * Now handle the string we got from userspace very carefully.
 824          * The rules are:
 825          *         - only use the first token we got
 826          *         - token delimiter is everything looking like a space
 827          *           character (' ', '\n', '\t' ...)
 828          *
 829          */
 830         if (!isalnum(buf[0])) {
 831                 /*
 832                  * If the first character userspace gave us is not
 833                  * alphanumerical then assume the filter should be
 834                  * switched off.
 835                  */
 836                 if (current_driver_name[0])
 837                         pr_info("DMA-API: switching off dma-debug driver filter\n");
 838                 current_driver_name[0] = 0;
 839                 current_driver = NULL;
 840                 goto out_unlock;
 841         }
 842
 843         /*
 844          * Now parse out the first token and use it as the name for the
 845          * driver to filter for.
 846          */
 847         for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 848                 current_driver_name[i] = buf[i];
 849                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 850                         break;
 851         }
 852         current_driver_name[i] = 0;
 853         current_driver = NULL;
 854
 855         pr_info("DMA-API: enable driver filter for driver [%s]\n",
 856                 current_driver_name);
 857
 858 out_unlock:
 859         write_unlock_irqrestore(&driver_name_lock, flags);
 860
 861         return count;
 862 }
 863
 864 static const struct file_operations filter_fops = {
 865         .read  = filter_read,
 866         .write = filter_write,
 867         .llseek = default_llseek,
 868 };
 869
 870 static int dma_debug_fs_init(void)
 871 {
 872         dma_debug_dent = debugfs_create_dir("dma-api", NULL);
 873         if (!dma_debug_dent) {
 874                 pr_err("DMA-API: can not create debugfs directory\n");
 875                 return -ENOMEM;
 876         }
 877
 878         global_disable_dent = debugfs_create_bool("disabled", 0444,
 879                         dma_debug_dent,
 880                         &global_disable);
 881         if (!global_disable_dent)
 882                 goto out_err;
 883
 884         error_count_dent = debugfs_create_u32("error_count", 0444,
 885                         dma_debug_dent, &error_count);
 886         if (!error_count_dent)
 887                 goto out_err;
 888
 889         show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
 890                         dma_debug_dent,
 891                         &show_all_errors);
 892         if (!show_all_errors_dent)
 893                 goto out_err;
 894
 895         show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
 896                         dma_debug_dent,
 897                         &show_num_errors);
 898         if (!show_num_errors_dent)
 899                 goto out_err;
 900
 901         num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
 902                         dma_debug_dent,
 903                         &num_free_entries);
 904         if (!num_free_entries_dent)
 905                 goto out_err;
 906
 907         min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
 908                         dma_debug_dent,
 909                         &min_free_entries);
 910         if (!min_free_entries_dent)
 911                 goto out_err;
 912
 913         filter_dent = debugfs_create_file("driver_filter", 0644,
 914                                           dma_debug_dent, NULL, &filter_fops);
 915         if (!filter_dent)
 916                 goto out_err;
 917
 918         return 0;
 919
 920 out_err:
 921         debugfs_remove_recursive(dma_debug_dent);
 922
 923         return -ENOMEM;
 924 }
 925
 926 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 927 {
 928         struct dma_debug_entry *entry;
 929         unsigned long flags;
 930         int count = 0, i;
 931
 932         local_irq_save(flags);
 933
 934         for (i = 0; i < HASH_SIZE; ++i) {
 935                 spin_lock(&dma_entry_hash[i].lock);
 936                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 937                         if (entry->dev == dev) {
 938                                 count += 1;
 939                                 *out_entry = entry;
 940                         }
 941                 }
 942                 spin_unlock(&dma_entry_hash[i].lock);
 943         }
 944
 945         local_irq_restore(flags);
 946
 947         return count;
 948 }
 949
 950 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 951 {
 952         struct device *dev = data;
 953         struct dma_debug_entry *uninitialized_var(entry);
 954         int count;
 955
 956         if (dma_debug_disabled())
 957                 return 0;
 958
 959         switch (action) {
 960         case BUS_NOTIFY_UNBOUND_DRIVER:
 961                 count = device_dma_allocations(dev, &entry);
 962                 if (count == 0)
 963                         break;
 964                 err_printk(dev, entry, "DMA-API: device driver has pending "
 965                                 "DMA allocations while released from device "
 966                                 "[count=%d]\n"
 967                                 "One of leaked entries details: "
 968                                 "[device address=0x%016llx] [size=%llu bytes] "
 969                                 "[mapped with %s] [mapped as %s]\n",
 970                         count, entry->dev_addr, entry->size,
 971                         dir2name[entry->direction], type2name[entry->type]);
 972                 break;
 973         default:
 974                 break;
 975         }
 976
 977         return 0;
 978 }
 979
 980 void dma_debug_add_bus(struct bus_type *bus)
 981 {
 982         struct notifier_block *nb;
 983
 984         if (dma_debug_disabled())
 985                 return;
 986
 987         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 988         if (nb == NULL) {
 989                 pr_err("dma_debug_add_bus: out of memory\n");
 990                 return;
 991         }
 992
 993         nb->notifier_call = dma_debug_device_change;
 994
 995         bus_register_notifier(bus, nb);
 996 }
 997
 998 /*
 999  * Let the architectures decide how many entries should be preallocated.
1000  */
1001 void dma_debug_init(u32 num_entries)
1002 {
1003         int i;
1004
1005         /* Do not use dma_debug_initialized here, since we really want to be
1006          * called to set dma_debug_initialized
1007          */
1008         if (global_disable)
1009                 return;
1010
1011         for (i = 0; i < HASH_SIZE; ++i) {
1012                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
1013                 spin_lock_init(&dma_entry_hash[i].lock);
1014         }
1015
1016         if (dma_debug_fs_init() != 0) {
1017                 pr_err("DMA-API: error creating debugfs entries - disabling\n");
1018                 global_disable = true;
1019
1020                 return;
1021         }
1022
1023         if (req_entries)
1024                 num_entries = req_entries;
1025
1026         if (prealloc_memory(num_entries) != 0) {
1027                 pr_err("DMA-API: debugging out of memory error - disabled\n");
1028                 global_disable = true;
1029
1030                 return;
1031         }
1032
1033         nr_total_entries = num_free_entries;
1034
1035         dma_debug_initialized = true;
1036
1037         pr_info("DMA-API: debugging enabled by kernel config\n");
1038 }
1039
1040 static __init int dma_debug_cmdline(char *str)
1041 {
1042         if (!str)
1043                 return -EINVAL;
1044
1045         if (strncmp(str, "off", 3) == 0) {
1046                 pr_info("DMA-API: debugging disabled on kernel command line\n");
1047                 global_disable = true;
1048         }
1049
1050         return 0;
1051 }
1052
1053 static __init int dma_debug_entries_cmdline(char *str)
1054 {
1055         int res;
1056
1057         if (!str)
1058                 return -EINVAL;
1059
1060         res = get_option(&str, &req_entries);
1061
1062         if (!res)
1063                 req_entries = 0;
1064
1065         return 0;
1066 }
1067
1068 __setup("dma_debug=", dma_debug_cmdline);
1069 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
1070
1071 static void check_unmap(struct dma_debug_entry *ref)
1072 {
1073         struct dma_debug_entry *entry;
1074         struct hash_bucket *bucket;
1075         unsigned long flags;
1076
1077         bucket = get_hash_bucket(ref, &flags);
1078         entry = bucket_find_exact(bucket, ref);
1079
1080         if (!entry) {
1081                 /* must drop lock before calling dma_mapping_error */
1082                 put_hash_bucket(bucket, &flags);
1083
1084                 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1085                         err_printk(ref->dev, NULL,
1086                                    "DMA-API: device driver tries to free an "
1087                                    "invalid DMA memory address\n");
1088                 } else {
1089                         err_printk(ref->dev, NULL,
1090                                    "DMA-API: device driver tries to free DMA "
1091                                    "memory it has not allocated [device "
1092                                    "address=0x%016llx] [size=%llu bytes]\n",
1093                                    ref->dev_addr, ref->size);
1094                 }
1095                 return;
1096         }
1097
1098         if (ref->size != entry->size) {
1099                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1100                            "DMA memory with different size "
1101                            "[device address=0x%016llx] [map size=%llu bytes] "
1102                            "[unmap size=%llu bytes]\n",
1103                            ref->dev_addr, entry->size, ref->size);
1104         }
1105
1106         if (ref->type != entry->type) {
1107                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1108                            "DMA memory with wrong function "
1109                            "[device address=0x%016llx] [size=%llu bytes] "
1110                            "[mapped as %s] [unmapped as %s]\n",
1111                            ref->dev_addr, ref->size,
1112                            type2name[entry->type], type2name[ref->type]);
1113         } else if ((entry->type == dma_debug_coherent) &&
1114                    (phys_addr(ref) != phys_addr(entry))) {
1115                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1116                            "DMA memory with different CPU address "
1117                            "[device address=0x%016llx] [size=%llu bytes] "
1118                            "[cpu alloc address=0x%016llx] "
1119                            "[cpu free address=0x%016llx]",
1120                            ref->dev_addr, ref->size,
1121                            phys_addr(entry),
1122                            phys_addr(ref));
1123         }
1124
1125         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1126             ref->sg_call_ents != entry->sg_call_ents) {
1127                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1128                            "DMA sg list with different entry count "
1129                            "[map count=%d] [unmap count=%d]\n",
1130                            entry->sg_call_ents, ref->sg_call_ents);
1131         }
1132
1133         /*
1134          * This may be no bug in reality - but most implementations of the
1135          * DMA API don't handle this properly, so check for it here
1136          */
1137         if (ref->direction != entry->direction) {
1138                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
1139                            "DMA memory with different direction "
1140                            "[device address=0x%016llx] [size=%llu bytes] "
1141                            "[mapped with %s] [unmapped with %s]\n",
1142                            ref->dev_addr, ref->size,
1143                            dir2name[entry->direction],
1144                            dir2name[ref->direction]);
1145         }
1146
1147         if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1148                 err_printk(ref->dev, entry,
1149                            "DMA-API: device driver failed to check map error"
1150                            "[device address=0x%016llx] [size=%llu bytes] "
1151                            "[mapped as %s]",
1152                            ref->dev_addr, ref->size,
1153                            type2name[entry->type]);
1154         }
1155
1156         hash_bucket_del(entry);
1157         dma_entry_free(entry);
1158
1159         put_hash_bucket(bucket, &flags);
1160 }
1161
1162 static void check_for_stack(struct device *dev, void *addr)
1163 {
1164         if (object_is_on_stack(addr))
1165                 err_printk(dev, NULL, "DMA-API: device driver maps memory from "
1166                                 "stack [addr=%p]\n", addr);
1167 }
1168
1169 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1170 {
1171         unsigned long a1 = (unsigned long)addr;
1172         unsigned long b1 = a1 + len;
1173         unsigned long a2 = (unsigned long)start;
1174         unsigned long b2 = (unsigned long)end;
1175
1176         return !(b1 <= a2 || a1 >= b2);
1177 }
1178
1179 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1180 {
1181         if (overlap(addr, len, _text, _etext) ||
1182             overlap(addr, len, __start_rodata, __end_rodata))
1183                 err_printk(dev, NULL, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1184 }
1185
1186 static void check_sync(struct device *dev,
1187                        struct dma_debug_entry *ref,
1188                        bool to_cpu)
1189 {
1190         struct dma_debug_entry *entry;
1191         struct hash_bucket *bucket;
1192         unsigned long flags;
1193
1194         bucket = get_hash_bucket(ref, &flags);
1195
1196         entry = bucket_find_contain(&bucket, ref, &flags);
1197
1198         if (!entry) {
1199                 err_printk(dev, NULL, "DMA-API: device driver tries "
1200                                 "to sync DMA memory it has not allocated "
1201                                 "[device address=0x%016llx] [size=%llu bytes]\n",
1202                                 (unsigned long long)ref->dev_addr, ref->size);
1203                 goto out;
1204         }
1205
1206         if (ref->size > entry->size) {
1207                 err_printk(dev, entry, "DMA-API: device driver syncs"
1208                                 " DMA memory outside allocated range "
1209                                 "[device address=0x%016llx] "
1210                                 "[allocation size=%llu bytes] "
1211                                 "[sync offset+size=%llu]\n",
1212                                 entry->dev_addr, entry->size,
1213                                 ref->size);
1214         }
1215
1216         if (entry->direction == DMA_BIDIRECTIONAL)
1217                 goto out;
1218
1219         if (ref->direction != entry->direction) {
1220                 err_printk(dev, entry, "DMA-API: device driver syncs "
1221                                 "DMA memory with different direction "
1222                                 "[device address=0x%016llx] [size=%llu bytes] "
1223                                 "[mapped with %s] [synced with %s]\n",
1224                                 (unsigned long long)ref->dev_addr, entry->size,
1225                                 dir2name[entry->direction],
1226                                 dir2name[ref->direction]);
1227         }
1228
1229         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1230                       !(ref->direction == DMA_TO_DEVICE))
1231                 err_printk(dev, entry, "DMA-API: device driver syncs "
1232                                 "device read-only DMA memory for cpu "
1233                                 "[device address=0x%016llx] [size=%llu bytes] "
1234                                 "[mapped with %s] [synced with %s]\n",
1235                                 (unsigned long long)ref->dev_addr, entry->size,
1236                                 dir2name[entry->direction],
1237                                 dir2name[ref->direction]);
1238
1239         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1240                        !(ref->direction == DMA_FROM_DEVICE))
1241                 err_printk(dev, entry, "DMA-API: device driver syncs "
1242                                 "device write-only DMA memory to device "
1243                                 "[device address=0x%016llx] [size=%llu bytes] "
1244                                 "[mapped with %s] [synced with %s]\n",
1245                                 (unsigned long long)ref->dev_addr, entry->size,
1246                                 dir2name[entry->direction],
1247                                 dir2name[ref->direction]);
1248
1249 out:
1250         put_hash_bucket(bucket, &flags);
1251 }
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                         bool map_single)
1256 {
1257         struct dma_debug_entry *entry;
1258
1259         if (unlikely(dma_debug_disabled()))
1260                 return;
1261
1262         if (dma_mapping_error(dev, dma_addr))
1263                 return;
1264
1265         entry = dma_entry_alloc();
1266         if (!entry)
1267                 return;
1268
1269         entry->dev       = dev;
1270         entry->type      = dma_debug_page;
1271         entry->pfn       = page_to_pfn(page);
1272         entry->offset    = offset,
1273         entry->dev_addr  = dma_addr;
1274         entry->size      = size;
1275         entry->direction = direction;
1276         entry->map_err_type = MAP_ERR_NOT_CHECKED;
1277
1278         if (map_single)
1279                 entry->type = dma_debug_single;
1280
1281         if (!PageHighMem(page)) {
1282                 void *addr = page_address(page) + offset;
1283
1284                 check_for_stack(dev, addr);
1285                 check_for_illegal_area(dev, addr, size);
1286         }
1287
1288         add_dma_entry(entry);
1289 }
1290 EXPORT_SYMBOL(debug_dma_map_page);
1291
1292 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1293 {
1294         struct dma_debug_entry ref;
1295         struct dma_debug_entry *entry;
1296         struct hash_bucket *bucket;
1297         unsigned long flags;
1298
1299         if (unlikely(dma_debug_disabled()))
1300                 return;
1301
1302         ref.dev = dev;
1303         ref.dev_addr = dma_addr;
1304         bucket = get_hash_bucket(&ref, &flags);
1305
1306         list_for_each_entry(entry, &bucket->list, list) {
1307                 if (!exact_match(&ref, entry))
1308                         continue;
1309
1310                 /*
1311                  * The same physical address can be mapped multiple
1312                  * times. Without a hardware IOMMU this results in the
1313                  * same device addresses being put into the dma-debug
1314                  * hash multiple times too. This can result in false
1315                  * positives being reported. Therefore we implement a
1316                  * best-fit algorithm here which updates the first entry
1317                  * from the hash which fits the reference value and is
1318                  * not currently listed as being checked.
1319                  */
1320                 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1321                         entry->map_err_type = MAP_ERR_CHECKED;
1322                         break;
1323                 }
1324         }
1325
1326         put_hash_bucket(bucket, &flags);
1327 }
1328 EXPORT_SYMBOL(debug_dma_mapping_error);
1329
1330 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1331                           size_t size, int direction, bool map_single)
1332 {
1333         struct dma_debug_entry ref = {
1334                 .type           = dma_debug_page,
1335                 .dev            = dev,
1336                 .dev_addr       = addr,
1337                 .size           = size,
1338                 .direction      = direction,
1339         };
1340
1341         if (unlikely(dma_debug_disabled()))
1342                 return;
1343
1344         if (map_single)
1345                 ref.type = dma_debug_single;
1346
1347         check_unmap(&ref);
1348 }
1349 EXPORT_SYMBOL(debug_dma_unmap_page);
1350
1351 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1352                       int nents, int mapped_ents, int direction)
1353 {
1354         struct dma_debug_entry *entry;
1355         struct scatterlist *s;
1356         int i;
1357
1358         if (unlikely(dma_debug_disabled()))
1359                 return;
1360
1361         for_each_sg(sg, s, mapped_ents, i) {
1362                 entry = dma_entry_alloc();
1363                 if (!entry)
1364                         return;
1365
1366                 entry->type           = dma_debug_sg;
1367                 entry->dev            = dev;
1368                 entry->pfn            = page_to_pfn(sg_page(s));
1369                 entry->offset         = s->offset,
1370                 entry->size           = sg_dma_len(s);
1371                 entry->dev_addr       = sg_dma_address(s);
1372                 entry->direction      = direction;
1373                 entry->sg_call_ents   = nents;
1374                 entry->sg_mapped_ents = mapped_ents;
1375
1376                 if (!PageHighMem(sg_page(s))) {
1377                         check_for_stack(dev, sg_virt(s));
1378                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1379                 }
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         entry = dma_entry_alloc();
1450         if (!entry)
1451                 return;
1452
1453         entry->type      = dma_debug_coherent;
1454         entry->dev       = dev;
1455         entry->pfn       = page_to_pfn(virt_to_page(virt));
1456         entry->offset    = (size_t) virt & PAGE_MASK;
1457         entry->size      = size;
1458         entry->dev_addr  = dma_addr;
1459         entry->direction = DMA_BIDIRECTIONAL;
1460
1461         add_dma_entry(entry);
1462 }
1463 EXPORT_SYMBOL(debug_dma_alloc_coherent);
1464
1465 void debug_dma_free_coherent(struct device *dev, size_t size,
1466                          void *virt, dma_addr_t addr)
1467 {
1468         struct dma_debug_entry ref = {
1469                 .type           = dma_debug_coherent,
1470                 .dev            = dev,
1471                 .pfn            = page_to_pfn(virt_to_page(virt)),
1472                 .offset         = (size_t) virt & PAGE_MASK,
1473                 .dev_addr       = addr,
1474                 .size           = size,
1475                 .direction      = DMA_BIDIRECTIONAL,
1476         };
1477
1478         if (unlikely(dma_debug_disabled()))
1479                 return;
1480
1481         check_unmap(&ref);
1482 }
1483 EXPORT_SYMBOL(debug_dma_free_coherent);
1484
1485 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1486                                    size_t size, int direction)
1487 {
1488         struct dma_debug_entry ref;
1489
1490         if (unlikely(dma_debug_disabled()))
1491                 return;
1492
1493         ref.type         = dma_debug_single;
1494         ref.dev          = dev;
1495         ref.dev_addr     = dma_handle;
1496         ref.size         = size;
1497         ref.direction    = direction;
1498         ref.sg_call_ents = 0;
1499
1500         check_sync(dev, &ref, true);
1501 }
1502 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1503
1504 void debug_dma_sync_single_for_device(struct device *dev,
1505                                       dma_addr_t dma_handle, size_t size,
1506                                       int direction)
1507 {
1508         struct dma_debug_entry ref;
1509
1510         if (unlikely(dma_debug_disabled()))
1511                 return;
1512
1513         ref.type         = dma_debug_single;
1514         ref.dev          = dev;
1515         ref.dev_addr     = dma_handle;
1516         ref.size         = size;
1517         ref.direction    = direction;
1518         ref.sg_call_ents = 0;
1519
1520         check_sync(dev, &ref, false);
1521 }
1522 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1523
1524 void debug_dma_sync_single_range_for_cpu(struct device *dev,
1525                                          dma_addr_t dma_handle,
1526                                          unsigned long offset, size_t size,
1527                                          int direction)
1528 {
1529         struct dma_debug_entry ref;
1530
1531         if (unlikely(dma_debug_disabled()))
1532                 return;
1533
1534         ref.type         = dma_debug_single;
1535         ref.dev          = dev;
1536         ref.dev_addr     = dma_handle;
1537         ref.size         = offset + size;
1538         ref.direction    = direction;
1539         ref.sg_call_ents = 0;
1540
1541         check_sync(dev, &ref, true);
1542 }
1543 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
1544
1545 void debug_dma_sync_single_range_for_device(struct device *dev,
1546                                             dma_addr_t dma_handle,
1547                                             unsigned long offset,
1548                                             size_t size, int direction)
1549 {
1550         struct dma_debug_entry ref;
1551
1552         if (unlikely(dma_debug_disabled()))
1553                 return;
1554
1555         ref.type         = dma_debug_single;
1556         ref.dev          = dev;
1557         ref.dev_addr     = dma_handle;
1558         ref.size         = offset + size;
1559         ref.direction    = direction;
1560         ref.sg_call_ents = 0;
1561
1562         check_sync(dev, &ref, false);
1563 }
1564 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
1565
1566 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1567                                int nelems, int direction)
1568 {
1569         struct scatterlist *s;
1570         int mapped_ents = 0, i;
1571
1572         if (unlikely(dma_debug_disabled()))
1573                 return;
1574
1575         for_each_sg(sg, s, nelems, i) {
1576
1577                 struct dma_debug_entry ref = {
1578                         .type           = dma_debug_sg,
1579                         .dev            = dev,
1580                         .pfn            = page_to_pfn(sg_page(s)),
1581                         .offset         = s->offset,
1582                         .dev_addr       = sg_dma_address(s),
1583                         .size           = sg_dma_len(s),
1584                         .direction      = direction,
1585                         .sg_call_ents   = nelems,
1586                 };
1587
1588                 if (!i)
1589                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1590
1591                 if (i >= mapped_ents)
1592                         break;
1593
1594                 check_sync(dev, &ref, true);
1595         }
1596 }
1597 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1598
1599 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1600                                   int nelems, int direction)
1601 {
1602         struct scatterlist *s;
1603         int mapped_ents = 0, i;
1604
1605         if (unlikely(dma_debug_disabled()))
1606                 return;
1607
1608         for_each_sg(sg, s, nelems, i) {
1609
1610                 struct dma_debug_entry ref = {
1611                         .type           = dma_debug_sg,
1612                         .dev            = dev,
1613                         .pfn            = page_to_pfn(sg_page(s)),
1614                         .offset         = s->offset,
1615                         .dev_addr       = sg_dma_address(s),
1616                         .size           = sg_dma_len(s),
1617                         .direction      = direction,
1618                         .sg_call_ents   = nelems,
1619                 };
1620                 if (!i)
1621                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1622
1623                 if (i >= mapped_ents)
1624                         break;
1625
1626                 check_sync(dev, &ref, false);
1627         }
1628 }
1629 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1630
1631 static int __init dma_debug_driver_setup(char *str)
1632 {
1633         int i;
1634
1635         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1636                 current_driver_name[i] = *str;
1637                 if (*str == 0)
1638                         break;
1639         }
1640
1641         if (current_driver_name[0])
1642                 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1643                         current_driver_name);
1644
1645
1646         return 1;
1647 }
1648 __setup("dma_debug_driver=", dma_debug_driver_setup);