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-map-ops.h>
  13 #include <linux/sched/task.h>
  14 #include <linux/stacktrace.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/vmalloc.h>
  17 #include <linux/debugfs.h>
  18 #include <linux/uaccess.h>
  19 #include <linux/export.h>
  20 #include <linux/device.h>
  21 #include <linux/types.h>
  22 #include <linux/sched.h>
  23 #include <linux/ctype.h>
  24 #include <linux/list.h>
  25 #include <linux/slab.h>
  26 #include <asm/sections.h>
  27 #include "debug.h"
  28
  29 #define HASH_SIZE       16384ULL
  30 #define HASH_FN_SHIFT   13
  31 #define HASH_FN_MASK    (HASH_SIZE - 1)
  32
  33 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
  34 /* If the pool runs out, add this many new entries at once */
  35 #define DMA_DEBUG_DYNAMIC_ENTRIES (PAGE_SIZE / sizeof(struct dma_debug_entry))
  36
  37 enum {
  38         dma_debug_single,
  39         dma_debug_sg,
  40         dma_debug_coherent,
  41         dma_debug_resource,
  42 };
  43
  44 enum map_err_types {
  45         MAP_ERR_CHECK_NOT_APPLICABLE,
  46         MAP_ERR_NOT_CHECKED,
  47         MAP_ERR_CHECKED,
  48 };
  49
  50 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
  51
  52 /**
  53  * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
  54  * @list: node on pre-allocated free_entries list
  55  * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
  56  * @size: length of the mapping
  57  * @type: single, page, sg, coherent
  58  * @direction: enum dma_data_direction
  59  * @sg_call_ents: 'nents' from dma_map_sg
  60  * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
  61  * @pfn: page frame of the start address
  62  * @offset: offset of mapping relative to pfn
  63  * @map_err_type: track whether dma_mapping_error() was checked
  64  * @stacktrace: support backtraces when a violation is detected
  65  */
  66 struct dma_debug_entry {
  67         struct list_head list;
  68         struct device    *dev;
  69         u64              dev_addr;
  70         u64              size;
  71         int              type;
  72         int              direction;
  73         int              sg_call_ents;
  74         int              sg_mapped_ents;
  75         unsigned long    pfn;
  76         size_t           offset;
  77         enum map_err_types  map_err_type;
  78 #ifdef CONFIG_STACKTRACE
  79         unsigned int    stack_len;
  80         unsigned long   stack_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
  81 #endif
  82 } ____cacheline_aligned_in_smp;
  83
  84 typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
  85
  86 struct hash_bucket {
  87         struct list_head list;
  88         spinlock_t lock;
  89 };
  90
  91 /* Hash list to save the allocated dma addresses */
  92 static struct hash_bucket dma_entry_hash[HASH_SIZE];
  93 /* List of pre-allocated dma_debug_entry's */
  94 static LIST_HEAD(free_entries);
  95 /* Lock for the list above */
  96 static DEFINE_SPINLOCK(free_entries_lock);
  97
  98 /* Global disable flag - will be set in case of an error */
  99 static bool global_disable __read_mostly;
 100
 101 /* Early initialization disable flag, set at the end of dma_debug_init */
 102 static bool dma_debug_initialized __read_mostly;
 103
 104 static inline bool dma_debug_disabled(void)
 105 {
 106         return global_disable || !dma_debug_initialized;
 107 }
 108
 109 /* Global error count */
 110 static u32 error_count;
 111
 112 /* Global error show enable*/
 113 static u32 show_all_errors __read_mostly;
 114 /* Number of errors to show */
 115 static u32 show_num_errors = 1;
 116
 117 static u32 num_free_entries;
 118 static u32 min_free_entries;
 119 static u32 nr_total_entries;
 120
 121 /* number of preallocated entries requested by kernel cmdline */
 122 static u32 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 123
 124 /* per-driver filter related state */
 125
 126 #define NAME_MAX_LEN    64
 127
 128 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
 129 static struct device_driver *current_driver                    __read_mostly;
 130
 131 static DEFINE_RWLOCK(driver_name_lock);
 132
 133 static const char *const maperr2str[] = {
 134         [MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
 135         [MAP_ERR_NOT_CHECKED] = "dma map error not checked",
 136         [MAP_ERR_CHECKED] = "dma map error checked",
 137 };
 138
 139 static const char *type2name[] = {
 140         [dma_debug_single] = "single",
 141         [dma_debug_sg] = "scather-gather",
 142         [dma_debug_coherent] = "coherent",
 143         [dma_debug_resource] = "resource",
 144 };
 145
 146 static const char *dir2name[] = {
 147         [DMA_BIDIRECTIONAL]     = "DMA_BIDIRECTIONAL",
 148         [DMA_TO_DEVICE]         = "DMA_TO_DEVICE",
 149         [DMA_FROM_DEVICE]       = "DMA_FROM_DEVICE",
 150         [DMA_NONE]              = "DMA_NONE",
 151 };
 152
 153 /*
 154  * The access to some variables in this macro is racy. We can't use atomic_t
 155  * here because all these variables are exported to debugfs. Some of them even
 156  * writeable. This is also the reason why a lock won't help much. But anyway,
 157  * the races are no big deal. Here is why:
 158  *
 159  *   error_count: the addition is racy, but the worst thing that can happen is
 160  *                that we don't count some errors
 161  *   show_num_errors: the subtraction is racy. Also no big deal because in
 162  *                    worst case this will result in one warning more in the
 163  *                    system log than the user configured. This variable is
 164  *                    writeable via debugfs.
 165  */
 166 static inline void dump_entry_trace(struct dma_debug_entry *entry)
 167 {
 168 #ifdef CONFIG_STACKTRACE
 169         if (entry) {
 170                 pr_warn("Mapped at:\n");
 171                 stack_trace_print(entry->stack_entries, entry->stack_len, 0);
 172         }
 173 #endif
 174 }
 175
 176 static bool driver_filter(struct device *dev)
 177 {
 178         struct device_driver *drv;
 179         unsigned long flags;
 180         bool ret;
 181
 182         /* driver filter off */
 183         if (likely(!current_driver_name[0]))
 184                 return true;
 185
 186         /* driver filter on and initialized */
 187         if (current_driver && dev && dev->driver == current_driver)
 188                 return true;
 189
 190         /* driver filter on, but we can't filter on a NULL device... */
 191         if (!dev)
 192                 return false;
 193
 194         if (current_driver || !current_driver_name[0])
 195                 return false;
 196
 197         /* driver filter on but not yet initialized */
 198         drv = dev->driver;
 199         if (!drv)
 200                 return false;
 201
 202         /* lock to protect against change of current_driver_name */
 203         read_lock_irqsave(&driver_name_lock, flags);
 204
 205         ret = false;
 206         if (drv->name &&
 207             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 208                 current_driver = drv;
 209                 ret = true;
 210         }
 211
 212         read_unlock_irqrestore(&driver_name_lock, flags);
 213
 214         return ret;
 215 }
 216
 217 #define err_printk(dev, entry, format, arg...) do {                     \
 218                 error_count += 1;                                       \
 219                 if (driver_filter(dev) &&                               \
 220                     (show_all_errors || show_num_errors > 0)) {         \
 221                         WARN(1, pr_fmt("%s %s: ") format,               \
 222                              dev ? dev_driver_string(dev) : "NULL",     \
 223                              dev ? dev_name(dev) : "NULL", ## arg);     \
 224                         dump_entry_trace(entry);                        \
 225                 }                                                       \
 226                 if (!show_all_errors && show_num_errors > 0)            \
 227                         show_num_errors -= 1;                           \
 228         } while (0);
 229
 230 /*
 231  * Hash related functions
 232  *
 233  * Every DMA-API request is saved into a struct dma_debug_entry. To
 234  * have quick access to these structs they are stored into a hash.
 235  */
 236 static int hash_fn(struct dma_debug_entry *entry)
 237 {
 238         /*
 239          * Hash function is based on the dma address.
 240          * We use bits 20-27 here as the index into the hash
 241          */
 242         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 243 }
 244
 245 /*
 246  * Request exclusive access to a hash bucket for a given dma_debug_entry.
 247  */
 248 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 249                                            unsigned long *flags)
 250         __acquires(&dma_entry_hash[idx].lock)
 251 {
 252         int idx = hash_fn(entry);
 253         unsigned long __flags;
 254
 255         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 256         *flags = __flags;
 257         return &dma_entry_hash[idx];
 258 }
 259
 260 /*
 261  * Give up exclusive access to the hash bucket
 262  */
 263 static void put_hash_bucket(struct hash_bucket *bucket,
 264                             unsigned long flags)
 265         __releases(&bucket->lock)
 266 {
 267         spin_unlock_irqrestore(&bucket->lock, flags);
 268 }
 269
 270 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
 271 {
 272         return ((a->dev_addr == b->dev_addr) &&
 273                 (a->dev == b->dev)) ? true : false;
 274 }
 275
 276 static bool containing_match(struct dma_debug_entry *a,
 277                              struct dma_debug_entry *b)
 278 {
 279         if (a->dev != b->dev)
 280                 return false;
 281
 282         if ((b->dev_addr <= a->dev_addr) &&
 283             ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
 284                 return true;
 285
 286         return false;
 287 }
 288
 289 /*
 290  * Search a given entry in the hash bucket list
 291  */
 292 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
 293                                                   struct dma_debug_entry *ref,
 294                                                   match_fn match)
 295 {
 296         struct dma_debug_entry *entry, *ret = NULL;
 297         int matches = 0, match_lvl, last_lvl = -1;
 298
 299         list_for_each_entry(entry, &bucket->list, list) {
 300                 if (!match(ref, entry))
 301                         continue;
 302
 303                 /*
 304                  * Some drivers map the same physical address multiple
 305                  * times. Without a hardware IOMMU this results in the
 306                  * same device addresses being put into the dma-debug
 307                  * hash multiple times too. This can result in false
 308                  * positives being reported. Therefore we implement a
 309                  * best-fit algorithm here which returns the entry from
 310                  * the hash which fits best to the reference value
 311                  * instead of the first-fit.
 312                  */
 313                 matches += 1;
 314                 match_lvl = 0;
 315                 entry->size         == ref->size         ? ++match_lvl : 0;
 316                 entry->type         == ref->type         ? ++match_lvl : 0;
 317                 entry->direction    == ref->direction    ? ++match_lvl : 0;
 318                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
 319
 320                 if (match_lvl == 4) {
 321                         /* perfect-fit - return the result */
 322                         return entry;
 323                 } else if (match_lvl > last_lvl) {
 324                         /*
 325                          * We found an entry that fits better then the
 326                          * previous one or it is the 1st match.
 327                          */
 328                         last_lvl = match_lvl;
 329                         ret      = entry;
 330                 }
 331         }
 332
 333         /*
 334          * If we have multiple matches but no perfect-fit, just return
 335          * NULL.
 336          */
 337         ret = (matches == 1) ? ret : NULL;
 338
 339         return ret;
 340 }
 341
 342 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
 343                                                  struct dma_debug_entry *ref)
 344 {
 345         return __hash_bucket_find(bucket, ref, exact_match);
 346 }
 347
 348 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 349                                                    struct dma_debug_entry *ref,
 350                                                    unsigned long *flags)
 351 {
 352
 353         unsigned int max_range = dma_get_max_seg_size(ref->dev);
 354         struct dma_debug_entry *entry, index = *ref;
 355         unsigned int range = 0;
 356
 357         while (range <= max_range) {
 358                 entry = __hash_bucket_find(*bucket, ref, containing_match);
 359
 360                 if (entry)
 361                         return entry;
 362
 363                 /*
 364                  * Nothing found, go back a hash bucket
 365                  */
 366                 put_hash_bucket(*bucket, *flags);
 367                 range          += (1 << HASH_FN_SHIFT);
 368                 index.dev_addr -= (1 << HASH_FN_SHIFT);
 369                 *bucket = get_hash_bucket(&index, flags);
 370         }
 371
 372         return NULL;
 373 }
 374
 375 /*
 376  * Add an entry to a hash bucket
 377  */
 378 static void hash_bucket_add(struct hash_bucket *bucket,
 379                             struct dma_debug_entry *entry)
 380 {
 381         list_add_tail(&entry->list, &bucket->list);
 382 }
 383
 384 /*
 385  * Remove entry from a hash bucket list
 386  */
 387 static void hash_bucket_del(struct dma_debug_entry *entry)
 388 {
 389         list_del(&entry->list);
 390 }
 391
 392 static unsigned long long phys_addr(struct dma_debug_entry *entry)
 393 {
 394         if (entry->type == dma_debug_resource)
 395                 return __pfn_to_phys(entry->pfn) + entry->offset;
 396
 397         return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
 398 }
 399
 400 /*
 401  * Dump mapping entries for debugging purposes
 402  */
 403 void debug_dma_dump_mappings(struct device *dev)
 404 {
 405         int idx;
 406
 407         for (idx = 0; idx < HASH_SIZE; idx++) {
 408                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 409                 struct dma_debug_entry *entry;
 410                 unsigned long flags;
 411
 412                 spin_lock_irqsave(&bucket->lock, flags);
 413
 414                 list_for_each_entry(entry, &bucket->list, list) {
 415                         if (!dev || dev == entry->dev) {
 416                                 dev_info(entry->dev,
 417                                          "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
 418                                          type2name[entry->type], idx,
 419                                          phys_addr(entry), entry->pfn,
 420                                          entry->dev_addr, entry->size,
 421                                          dir2name[entry->direction],
 422                                          maperr2str[entry->map_err_type]);
 423                         }
 424                 }
 425
 426                 spin_unlock_irqrestore(&bucket->lock, flags);
 427                 cond_resched();
 428         }
 429 }
 430
 431 /*
 432  * For each mapping (initial cacheline in the case of
 433  * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
 434  * scatterlist, or the cacheline specified in dma_map_single) insert
 435  * into this tree using the cacheline as the key. At
 436  * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
 437  * the entry already exists at insertion time add a tag as a reference
 438  * count for the overlapping mappings.  For now, the overlap tracking
 439  * just ensures that 'unmaps' balance 'maps' before marking the
 440  * cacheline idle, but we should also be flagging overlaps as an API
 441  * violation.
 442  *
 443  * Memory usage is mostly constrained by the maximum number of available
 444  * dma-debug entries in that we need a free dma_debug_entry before
 445  * inserting into the tree.  In the case of dma_map_page and
 446  * dma_alloc_coherent there is only one dma_debug_entry and one
 447  * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
 448  * other hand, consumes a single dma_debug_entry, but inserts 'nents'
 449  * entries into the tree.
 450  */
 451 static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
 452 static DEFINE_SPINLOCK(radix_lock);
 453 #define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
 454 #define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
 455 #define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
 456
 457 static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
 458 {
 459         return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
 460                 (entry->offset >> L1_CACHE_SHIFT);
 461 }
 462
 463 static int active_cacheline_read_overlap(phys_addr_t cln)
 464 {
 465         int overlap = 0, i;
 466
 467         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 468                 if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
 469                         overlap |= 1 << i;
 470         return overlap;
 471 }
 472
 473 static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
 474 {
 475         int i;
 476
 477         if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
 478                 return overlap;
 479
 480         for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 481                 if (overlap & 1 << i)
 482                         radix_tree_tag_set(&dma_active_cacheline, cln, i);
 483                 else
 484                         radix_tree_tag_clear(&dma_active_cacheline, cln, i);
 485
 486         return overlap;
 487 }
 488
 489 static void active_cacheline_inc_overlap(phys_addr_t cln)
 490 {
 491         int overlap = active_cacheline_read_overlap(cln);
 492
 493         overlap = active_cacheline_set_overlap(cln, ++overlap);
 494
 495         /* If we overflowed the overlap counter then we're potentially
 496          * leaking dma-mappings.
 497          */
 498         WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
 499                   pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
 500                   ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 501 }
 502
 503 static int active_cacheline_dec_overlap(phys_addr_t cln)
 504 {
 505         int overlap = active_cacheline_read_overlap(cln);
 506
 507         return active_cacheline_set_overlap(cln, --overlap);
 508 }
 509
 510 static int active_cacheline_insert(struct dma_debug_entry *entry)
 511 {
 512         phys_addr_t cln = to_cacheline_number(entry);
 513         unsigned long flags;
 514         int rc;
 515
 516         /* If the device is not writing memory then we don't have any
 517          * concerns about the cpu consuming stale data.  This mitigates
 518          * legitimate usages of overlapping mappings.
 519          */
 520         if (entry->direction == DMA_TO_DEVICE)
 521                 return 0;
 522
 523         spin_lock_irqsave(&radix_lock, flags);
 524         rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
 525         if (rc == -EEXIST)
 526                 active_cacheline_inc_overlap(cln);
 527         spin_unlock_irqrestore(&radix_lock, flags);
 528
 529         return rc;
 530 }
 531
 532 static void active_cacheline_remove(struct dma_debug_entry *entry)
 533 {
 534         phys_addr_t cln = to_cacheline_number(entry);
 535         unsigned long flags;
 536
 537         /* ...mirror the insert case */
 538         if (entry->direction == DMA_TO_DEVICE)
 539                 return;
 540
 541         spin_lock_irqsave(&radix_lock, flags);
 542         /* since we are counting overlaps the final put of the
 543          * cacheline will occur when the overlap count is 0.
 544          * active_cacheline_dec_overlap() returns -1 in that case
 545          */
 546         if (active_cacheline_dec_overlap(cln) < 0)
 547                 radix_tree_delete(&dma_active_cacheline, cln);
 548         spin_unlock_irqrestore(&radix_lock, flags);
 549 }
 550
 551 /*
 552  * Wrapper function for adding an entry to the hash.
 553  * This function takes care of locking itself.
 554  */
 555 static void add_dma_entry(struct dma_debug_entry *entry)
 556 {
 557         struct hash_bucket *bucket;
 558         unsigned long flags;
 559         int rc;
 560
 561         bucket = get_hash_bucket(entry, &flags);
 562         hash_bucket_add(bucket, entry);
 563         put_hash_bucket(bucket, flags);
 564
 565         rc = active_cacheline_insert(entry);
 566         if (rc == -ENOMEM) {
 567                 pr_err("cacheline tracking ENOMEM, dma-debug disabled\n");
 568                 global_disable = true;
 569         }
 570
 571         /* TODO: report -EEXIST errors here as overlapping mappings are
 572          * not supported by the DMA API
 573          */
 574 }
 575
 576 static int dma_debug_create_entries(gfp_t gfp)
 577 {
 578         struct dma_debug_entry *entry;
 579         int i;
 580
 581         entry = (void *)get_zeroed_page(gfp);
 582         if (!entry)
 583                 return -ENOMEM;
 584
 585         for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
 586                 list_add_tail(&entry[i].list, &free_entries);
 587
 588         num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 589         nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 590
 591         return 0;
 592 }
 593
 594 static struct dma_debug_entry *__dma_entry_alloc(void)
 595 {
 596         struct dma_debug_entry *entry;
 597
 598         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 599         list_del(&entry->list);
 600         memset(entry, 0, sizeof(*entry));
 601
 602         num_free_entries -= 1;
 603         if (num_free_entries < min_free_entries)
 604                 min_free_entries = num_free_entries;
 605
 606         return entry;
 607 }
 608
 609 static void __dma_entry_alloc_check_leak(void)
 610 {
 611         u32 tmp = nr_total_entries % nr_prealloc_entries;
 612
 613         /* Shout each time we tick over some multiple of the initial pool */
 614         if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
 615                 pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
 616                         nr_total_entries,
 617                         (nr_total_entries / nr_prealloc_entries));
 618         }
 619 }
 620
 621 /* struct dma_entry allocator
 622  *
 623  * The next two functions implement the allocator for
 624  * struct dma_debug_entries.
 625  */
 626 static struct dma_debug_entry *dma_entry_alloc(void)
 627 {
 628         struct dma_debug_entry *entry;
 629         unsigned long flags;
 630
 631         spin_lock_irqsave(&free_entries_lock, flags);
 632         if (num_free_entries == 0) {
 633                 if (dma_debug_create_entries(GFP_ATOMIC)) {
 634                         global_disable = true;
 635                         spin_unlock_irqrestore(&free_entries_lock, flags);
 636                         pr_err("debugging out of memory - disabling\n");
 637                         return NULL;
 638                 }
 639                 __dma_entry_alloc_check_leak();
 640         }
 641
 642         entry = __dma_entry_alloc();
 643
 644         spin_unlock_irqrestore(&free_entries_lock, flags);
 645
 646 #ifdef CONFIG_STACKTRACE
 647         entry->stack_len = stack_trace_save(entry->stack_entries,
 648                                             ARRAY_SIZE(entry->stack_entries),
 649                                             1);
 650 #endif
 651         return entry;
 652 }
 653
 654 static void dma_entry_free(struct dma_debug_entry *entry)
 655 {
 656         unsigned long flags;
 657
 658         active_cacheline_remove(entry);
 659
 660         /*
 661          * add to beginning of the list - this way the entries are
 662          * more likely cache hot when they are reallocated.
 663          */
 664         spin_lock_irqsave(&free_entries_lock, flags);
 665         list_add(&entry->list, &free_entries);
 666         num_free_entries += 1;
 667         spin_unlock_irqrestore(&free_entries_lock, flags);
 668 }
 669
 670 /*
 671  * DMA-API debugging init code
 672  *
 673  * The init code does two things:
 674  *   1. Initialize core data structures
 675  *   2. Preallocate a given number of dma_debug_entry structs
 676  */
 677
 678 static ssize_t filter_read(struct file *file, char __user *user_buf,
 679                            size_t count, loff_t *ppos)
 680 {
 681         char buf[NAME_MAX_LEN + 1];
 682         unsigned long flags;
 683         int len;
 684
 685         if (!current_driver_name[0])
 686                 return 0;
 687
 688         /*
 689          * We can't copy to userspace directly because current_driver_name can
 690          * only be read under the driver_name_lock with irqs disabled. So
 691          * create a temporary copy first.
 692          */
 693         read_lock_irqsave(&driver_name_lock, flags);
 694         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 695         read_unlock_irqrestore(&driver_name_lock, flags);
 696
 697         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 698 }
 699
 700 static ssize_t filter_write(struct file *file, const char __user *userbuf,
 701                             size_t count, loff_t *ppos)
 702 {
 703         char buf[NAME_MAX_LEN];
 704         unsigned long flags;
 705         size_t len;
 706         int i;
 707
 708         /*
 709          * We can't copy from userspace directly. Access to
 710          * current_driver_name is protected with a write_lock with irqs
 711          * disabled. Since copy_from_user can fault and may sleep we
 712          * need to copy to temporary buffer first
 713          */
 714         len = min(count, (size_t)(NAME_MAX_LEN - 1));
 715         if (copy_from_user(buf, userbuf, len))
 716                 return -EFAULT;
 717
 718         buf[len] = 0;
 719
 720         write_lock_irqsave(&driver_name_lock, flags);
 721
 722         /*
 723          * Now handle the string we got from userspace very carefully.
 724          * The rules are:
 725          *         - only use the first token we got
 726          *         - token delimiter is everything looking like a space
 727          *           character (' ', '\n', '\t' ...)
 728          *
 729          */
 730         if (!isalnum(buf[0])) {
 731                 /*
 732                  * If the first character userspace gave us is not
 733                  * alphanumerical then assume the filter should be
 734                  * switched off.
 735                  */
 736                 if (current_driver_name[0])
 737                         pr_info("switching off dma-debug driver filter\n");
 738                 current_driver_name[0] = 0;
 739                 current_driver = NULL;
 740                 goto out_unlock;
 741         }
 742
 743         /*
 744          * Now parse out the first token and use it as the name for the
 745          * driver to filter for.
 746          */
 747         for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 748                 current_driver_name[i] = buf[i];
 749                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 750                         break;
 751         }
 752         current_driver_name[i] = 0;
 753         current_driver = NULL;
 754
 755         pr_info("enable driver filter for driver [%s]\n",
 756                 current_driver_name);
 757
 758 out_unlock:
 759         write_unlock_irqrestore(&driver_name_lock, flags);
 760
 761         return count;
 762 }
 763
 764 static const struct file_operations filter_fops = {
 765         .read  = filter_read,
 766         .write = filter_write,
 767         .llseek = default_llseek,
 768 };
 769
 770 static int dump_show(struct seq_file *seq, void *v)
 771 {
 772         int idx;
 773
 774         for (idx = 0; idx < HASH_SIZE; idx++) {
 775                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 776                 struct dma_debug_entry *entry;
 777                 unsigned long flags;
 778
 779                 spin_lock_irqsave(&bucket->lock, flags);
 780                 list_for_each_entry(entry, &bucket->list, list) {
 781                         seq_printf(seq,
 782                                    "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx %s %s\n",
 783                                    dev_name(entry->dev),
 784                                    dev_driver_string(entry->dev),
 785                                    type2name[entry->type], idx,
 786                                    phys_addr(entry), entry->pfn,
 787                                    entry->dev_addr, entry->size,
 788                                    dir2name[entry->direction],
 789                                    maperr2str[entry->map_err_type]);
 790                 }
 791                 spin_unlock_irqrestore(&bucket->lock, flags);
 792         }
 793         return 0;
 794 }
 795 DEFINE_SHOW_ATTRIBUTE(dump);
 796
 797 static void dma_debug_fs_init(void)
 798 {
 799         struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
 800
 801         debugfs_create_bool("disabled", 0444, dentry, &global_disable);
 802         debugfs_create_u32("error_count", 0444, dentry, &error_count);
 803         debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
 804         debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
 805         debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
 806         debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
 807         debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
 808         debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
 809         debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
 810 }
 811
 812 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 813 {
 814         struct dma_debug_entry *entry;
 815         unsigned long flags;
 816         int count = 0, i;
 817
 818         for (i = 0; i < HASH_SIZE; ++i) {
 819                 spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
 820                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 821                         if (entry->dev == dev) {
 822                                 count += 1;
 823                                 *out_entry = entry;
 824                         }
 825                 }
 826                 spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
 827         }
 828
 829         return count;
 830 }
 831
 832 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 833 {
 834         struct device *dev = data;
 835         struct dma_debug_entry *entry;
 836         int count;
 837
 838         if (dma_debug_disabled())
 839                 return 0;
 840
 841         switch (action) {
 842         case BUS_NOTIFY_UNBOUND_DRIVER:
 843                 count = device_dma_allocations(dev, &entry);
 844                 if (count == 0)
 845                         break;
 846                 err_printk(dev, entry, "device driver has pending "
 847                                 "DMA allocations while released from device "
 848                                 "[count=%d]\n"
 849                                 "One of leaked entries details: "
 850                                 "[device address=0x%016llx] [size=%llu bytes] "
 851                                 "[mapped with %s] [mapped as %s]\n",
 852                         count, entry->dev_addr, entry->size,
 853                         dir2name[entry->direction], type2name[entry->type]);
 854                 break;
 855         default:
 856                 break;
 857         }
 858
 859         return 0;
 860 }
 861
 862 void dma_debug_add_bus(struct bus_type *bus)
 863 {
 864         struct notifier_block *nb;
 865
 866         if (dma_debug_disabled())
 867                 return;
 868
 869         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 870         if (nb == NULL) {
 871                 pr_err("dma_debug_add_bus: out of memory\n");
 872                 return;
 873         }
 874
 875         nb->notifier_call = dma_debug_device_change;
 876
 877         bus_register_notifier(bus, nb);
 878 }
 879
 880 static int dma_debug_init(void)
 881 {
 882         int i, nr_pages;
 883
 884         /* Do not use dma_debug_initialized here, since we really want to be
 885          * called to set dma_debug_initialized
 886          */
 887         if (global_disable)
 888                 return 0;
 889
 890         for (i = 0; i < HASH_SIZE; ++i) {
 891                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
 892                 spin_lock_init(&dma_entry_hash[i].lock);
 893         }
 894
 895         dma_debug_fs_init();
 896
 897         nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
 898         for (i = 0; i < nr_pages; ++i)
 899                 dma_debug_create_entries(GFP_KERNEL);
 900         if (num_free_entries >= nr_prealloc_entries) {
 901                 pr_info("preallocated %d debug entries\n", nr_total_entries);
 902         } else if (num_free_entries > 0) {
 903                 pr_warn("%d debug entries requested but only %d allocated\n",
 904                         nr_prealloc_entries, nr_total_entries);
 905         } else {
 906                 pr_err("debugging out of memory error - disabled\n");
 907                 global_disable = true;
 908
 909                 return 0;
 910         }
 911         min_free_entries = num_free_entries;
 912
 913         dma_debug_initialized = true;
 914
 915         pr_info("debugging enabled by kernel config\n");
 916         return 0;
 917 }
 918 core_initcall(dma_debug_init);
 919
 920 static __init int dma_debug_cmdline(char *str)
 921 {
 922         if (!str)
 923                 return -EINVAL;
 924
 925         if (strncmp(str, "off", 3) == 0) {
 926                 pr_info("debugging disabled on kernel command line\n");
 927                 global_disable = true;
 928         }
 929
 930         return 0;
 931 }
 932
 933 static __init int dma_debug_entries_cmdline(char *str)
 934 {
 935         if (!str)
 936                 return -EINVAL;
 937         if (!get_option(&str, &nr_prealloc_entries))
 938                 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 939         return 0;
 940 }
 941
 942 __setup("dma_debug=", dma_debug_cmdline);
 943 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
 944
 945 static void check_unmap(struct dma_debug_entry *ref)
 946 {
 947         struct dma_debug_entry *entry;
 948         struct hash_bucket *bucket;
 949         unsigned long flags;
 950
 951         bucket = get_hash_bucket(ref, &flags);
 952         entry = bucket_find_exact(bucket, ref);
 953
 954         if (!entry) {
 955                 /* must drop lock before calling dma_mapping_error */
 956                 put_hash_bucket(bucket, flags);
 957
 958                 if (dma_mapping_error(ref->dev, ref->dev_addr)) {
 959                         err_printk(ref->dev, NULL,
 960                                    "device driver tries to free an "
 961                                    "invalid DMA memory address\n");
 962                 } else {
 963                         err_printk(ref->dev, NULL,
 964                                    "device driver tries to free DMA "
 965                                    "memory it has not allocated [device "
 966                                    "address=0x%016llx] [size=%llu bytes]\n",
 967                                    ref->dev_addr, ref->size);
 968                 }
 969                 return;
 970         }
 971
 972         if (ref->size != entry->size) {
 973                 err_printk(ref->dev, entry, "device driver frees "
 974                            "DMA memory with different size "
 975                            "[device address=0x%016llx] [map size=%llu bytes] "
 976                            "[unmap size=%llu bytes]\n",
 977                            ref->dev_addr, entry->size, ref->size);
 978         }
 979
 980         if (ref->type != entry->type) {
 981                 err_printk(ref->dev, entry, "device driver frees "
 982                            "DMA memory with wrong function "
 983                            "[device address=0x%016llx] [size=%llu bytes] "
 984                            "[mapped as %s] [unmapped as %s]\n",
 985                            ref->dev_addr, ref->size,
 986                            type2name[entry->type], type2name[ref->type]);
 987         } else if ((entry->type == dma_debug_coherent) &&
 988                    (phys_addr(ref) != phys_addr(entry))) {
 989                 err_printk(ref->dev, entry, "device driver frees "
 990                            "DMA memory with different CPU address "
 991                            "[device address=0x%016llx] [size=%llu bytes] "
 992                            "[cpu alloc address=0x%016llx] "
 993                            "[cpu free address=0x%016llx]",
 994                            ref->dev_addr, ref->size,
 995                            phys_addr(entry),
 996                            phys_addr(ref));
 997         }
 998
 999         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1000             ref->sg_call_ents != entry->sg_call_ents) {
1001                 err_printk(ref->dev, entry, "device driver frees "
1002                            "DMA sg list with different entry count "
1003                            "[map count=%d] [unmap count=%d]\n",
1004                            entry->sg_call_ents, ref->sg_call_ents);
1005         }
1006
1007         /*
1008          * This may be no bug in reality - but most implementations of the
1009          * DMA API don't handle this properly, so check for it here
1010          */
1011         if (ref->direction != entry->direction) {
1012                 err_printk(ref->dev, entry, "device driver frees "
1013                            "DMA memory with different direction "
1014                            "[device address=0x%016llx] [size=%llu bytes] "
1015                            "[mapped with %s] [unmapped with %s]\n",
1016                            ref->dev_addr, ref->size,
1017                            dir2name[entry->direction],
1018                            dir2name[ref->direction]);
1019         }
1020
1021         /*
1022          * Drivers should use dma_mapping_error() to check the returned
1023          * addresses of dma_map_single() and dma_map_page().
1024          * If not, print this warning message. See Documentation/core-api/dma-api.rst.
1025          */
1026         if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1027                 err_printk(ref->dev, entry,
1028                            "device driver failed to check map error"
1029                            "[device address=0x%016llx] [size=%llu bytes] "
1030                            "[mapped as %s]",
1031                            ref->dev_addr, ref->size,
1032                            type2name[entry->type]);
1033         }
1034
1035         hash_bucket_del(entry);
1036         dma_entry_free(entry);
1037
1038         put_hash_bucket(bucket, flags);
1039 }
1040
1041 static void check_for_stack(struct device *dev,
1042                             struct page *page, size_t offset)
1043 {
1044         void *addr;
1045         struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1046
1047         if (!stack_vm_area) {
1048                 /* Stack is direct-mapped. */
1049                 if (PageHighMem(page))
1050                         return;
1051                 addr = page_address(page) + offset;
1052                 if (object_is_on_stack(addr))
1053                         err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1054         } else {
1055                 /* Stack is vmalloced. */
1056                 int i;
1057
1058                 for (i = 0; i < stack_vm_area->nr_pages; i++) {
1059                         if (page != stack_vm_area->pages[i])
1060                                 continue;
1061
1062                         addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1063                         err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1064                         break;
1065                 }
1066         }
1067 }
1068
1069 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1070 {
1071         unsigned long a1 = (unsigned long)addr;
1072         unsigned long b1 = a1 + len;
1073         unsigned long a2 = (unsigned long)start;
1074         unsigned long b2 = (unsigned long)end;
1075
1076         return !(b1 <= a2 || a1 >= b2);
1077 }
1078
1079 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1080 {
1081         if (overlap(addr, len, _stext, _etext) ||
1082             overlap(addr, len, __start_rodata, __end_rodata))
1083                 err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1084 }
1085
1086 static void check_sync(struct device *dev,
1087                        struct dma_debug_entry *ref,
1088                        bool to_cpu)
1089 {
1090         struct dma_debug_entry *entry;
1091         struct hash_bucket *bucket;
1092         unsigned long flags;
1093
1094         bucket = get_hash_bucket(ref, &flags);
1095
1096         entry = bucket_find_contain(&bucket, ref, &flags);
1097
1098         if (!entry) {
1099                 err_printk(dev, NULL, "device driver tries "
1100                                 "to sync DMA memory it has not allocated "
1101                                 "[device address=0x%016llx] [size=%llu bytes]\n",
1102                                 (unsigned long long)ref->dev_addr, ref->size);
1103                 goto out;
1104         }
1105
1106         if (ref->size > entry->size) {
1107                 err_printk(dev, entry, "device driver syncs"
1108                                 " DMA memory outside allocated range "
1109                                 "[device address=0x%016llx] "
1110                                 "[allocation size=%llu bytes] "
1111                                 "[sync offset+size=%llu]\n",
1112                                 entry->dev_addr, entry->size,
1113                                 ref->size);
1114         }
1115
1116         if (entry->direction == DMA_BIDIRECTIONAL)
1117                 goto out;
1118
1119         if (ref->direction != entry->direction) {
1120                 err_printk(dev, entry, "device driver syncs "
1121                                 "DMA memory with different direction "
1122                                 "[device address=0x%016llx] [size=%llu bytes] "
1123                                 "[mapped with %s] [synced with %s]\n",
1124                                 (unsigned long long)ref->dev_addr, entry->size,
1125                                 dir2name[entry->direction],
1126                                 dir2name[ref->direction]);
1127         }
1128
1129         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1130                       !(ref->direction == DMA_TO_DEVICE))
1131                 err_printk(dev, entry, "device driver syncs "
1132                                 "device read-only DMA memory for cpu "
1133                                 "[device address=0x%016llx] [size=%llu bytes] "
1134                                 "[mapped with %s] [synced with %s]\n",
1135                                 (unsigned long long)ref->dev_addr, entry->size,
1136                                 dir2name[entry->direction],
1137                                 dir2name[ref->direction]);
1138
1139         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1140                        !(ref->direction == DMA_FROM_DEVICE))
1141                 err_printk(dev, entry, "device driver syncs "
1142                                 "device write-only DMA memory to device "
1143                                 "[device address=0x%016llx] [size=%llu bytes] "
1144                                 "[mapped with %s] [synced with %s]\n",
1145                                 (unsigned long long)ref->dev_addr, entry->size,
1146                                 dir2name[entry->direction],
1147                                 dir2name[ref->direction]);
1148
1149         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1150             ref->sg_call_ents != entry->sg_call_ents) {
1151                 err_printk(ref->dev, entry, "device driver syncs "
1152                            "DMA sg list with different entry count "
1153                            "[map count=%d] [sync count=%d]\n",
1154                            entry->sg_call_ents, ref->sg_call_ents);
1155         }
1156
1157 out:
1158         put_hash_bucket(bucket, flags);
1159 }
1160
1161 static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1162 {
1163 #ifdef CONFIG_DMA_API_DEBUG_SG
1164         unsigned int max_seg = dma_get_max_seg_size(dev);
1165         u64 start, end, boundary = dma_get_seg_boundary(dev);
1166
1167         /*
1168          * Either the driver forgot to set dma_parms appropriately, or
1169          * whoever generated the list forgot to check them.
1170          */
1171         if (sg->length > max_seg)
1172                 err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1173                            sg->length, max_seg);
1174         /*
1175          * In some cases this could potentially be the DMA API
1176          * implementation's fault, but it would usually imply that
1177          * the scatterlist was built inappropriately to begin with.
1178          */
1179         start = sg_dma_address(sg);
1180         end = start + sg_dma_len(sg) - 1;
1181         if ((start ^ end) & ~boundary)
1182                 err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1183                            start, end, boundary);
1184 #endif
1185 }
1186
1187 void debug_dma_map_single(struct device *dev, const void *addr,
1188                             unsigned long len)
1189 {
1190         if (unlikely(dma_debug_disabled()))
1191                 return;
1192
1193         if (!virt_addr_valid(addr))
1194                 err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1195                            addr, len);
1196
1197         if (is_vmalloc_addr(addr))
1198                 err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1199                            addr, len);
1200 }
1201 EXPORT_SYMBOL(debug_dma_map_single);
1202
1203 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1204                         size_t size, int direction, dma_addr_t dma_addr)
1205 {
1206         struct dma_debug_entry *entry;
1207
1208         if (unlikely(dma_debug_disabled()))
1209                 return;
1210
1211         if (dma_mapping_error(dev, dma_addr))
1212                 return;
1213
1214         entry = dma_entry_alloc();
1215         if (!entry)
1216                 return;
1217
1218         entry->dev       = dev;
1219         entry->type      = dma_debug_single;
1220         entry->pfn       = page_to_pfn(page);
1221         entry->offset    = offset;
1222         entry->dev_addr  = dma_addr;
1223         entry->size      = size;
1224         entry->direction = direction;
1225         entry->map_err_type = MAP_ERR_NOT_CHECKED;
1226
1227         check_for_stack(dev, page, offset);
1228
1229         if (!PageHighMem(page)) {
1230                 void *addr = page_address(page) + offset;
1231
1232                 check_for_illegal_area(dev, addr, size);
1233         }
1234
1235         add_dma_entry(entry);
1236 }
1237
1238 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1239 {
1240         struct dma_debug_entry ref;
1241         struct dma_debug_entry *entry;
1242         struct hash_bucket *bucket;
1243         unsigned long flags;
1244
1245         if (unlikely(dma_debug_disabled()))
1246                 return;
1247
1248         ref.dev = dev;
1249         ref.dev_addr = dma_addr;
1250         bucket = get_hash_bucket(&ref, &flags);
1251
1252         list_for_each_entry(entry, &bucket->list, list) {
1253                 if (!exact_match(&ref, entry))
1254                         continue;
1255
1256                 /*
1257                  * The same physical address can be mapped multiple
1258                  * times. Without a hardware IOMMU this results in the
1259                  * same device addresses being put into the dma-debug
1260                  * hash multiple times too. This can result in false
1261                  * positives being reported. Therefore we implement a
1262                  * best-fit algorithm here which updates the first entry
1263                  * from the hash which fits the reference value and is
1264                  * not currently listed as being checked.
1265                  */
1266                 if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1267                         entry->map_err_type = MAP_ERR_CHECKED;
1268                         break;
1269                 }
1270         }
1271
1272         put_hash_bucket(bucket, flags);
1273 }
1274 EXPORT_SYMBOL(debug_dma_mapping_error);
1275
1276 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1277                           size_t size, int direction)
1278 {
1279         struct dma_debug_entry ref = {
1280                 .type           = dma_debug_single,
1281                 .dev            = dev,
1282                 .dev_addr       = addr,
1283                 .size           = size,
1284                 .direction      = direction,
1285         };
1286
1287         if (unlikely(dma_debug_disabled()))
1288                 return;
1289         check_unmap(&ref);
1290 }
1291
1292 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1293                       int nents, int mapped_ents, int direction)
1294 {
1295         struct dma_debug_entry *entry;
1296         struct scatterlist *s;
1297         int i;
1298
1299         if (unlikely(dma_debug_disabled()))
1300                 return;
1301
1302         for_each_sg(sg, s, mapped_ents, i) {
1303                 entry = dma_entry_alloc();
1304                 if (!entry)
1305                         return;
1306
1307                 entry->type           = dma_debug_sg;
1308                 entry->dev            = dev;
1309                 entry->pfn            = page_to_pfn(sg_page(s));
1310                 entry->offset         = s->offset;
1311                 entry->size           = sg_dma_len(s);
1312                 entry->dev_addr       = sg_dma_address(s);
1313                 entry->direction      = direction;
1314                 entry->sg_call_ents   = nents;
1315                 entry->sg_mapped_ents = mapped_ents;
1316
1317                 check_for_stack(dev, sg_page(s), s->offset);
1318
1319                 if (!PageHighMem(sg_page(s))) {
1320                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1321                 }
1322
1323                 check_sg_segment(dev, s);
1324
1325                 add_dma_entry(entry);
1326         }
1327 }
1328
1329 static int get_nr_mapped_entries(struct device *dev,
1330                                  struct dma_debug_entry *ref)
1331 {
1332         struct dma_debug_entry *entry;
1333         struct hash_bucket *bucket;
1334         unsigned long flags;
1335         int mapped_ents;
1336
1337         bucket       = get_hash_bucket(ref, &flags);
1338         entry        = bucket_find_exact(bucket, ref);
1339         mapped_ents  = 0;
1340
1341         if (entry)
1342                 mapped_ents = entry->sg_mapped_ents;
1343         put_hash_bucket(bucket, flags);
1344
1345         return mapped_ents;
1346 }
1347
1348 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1349                         int nelems, int dir)
1350 {
1351         struct scatterlist *s;
1352         int mapped_ents = 0, i;
1353
1354         if (unlikely(dma_debug_disabled()))
1355                 return;
1356
1357         for_each_sg(sglist, s, nelems, i) {
1358
1359                 struct dma_debug_entry ref = {
1360                         .type           = dma_debug_sg,
1361                         .dev            = dev,
1362                         .pfn            = page_to_pfn(sg_page(s)),
1363                         .offset         = s->offset,
1364                         .dev_addr       = sg_dma_address(s),
1365                         .size           = sg_dma_len(s),
1366                         .direction      = dir,
1367                         .sg_call_ents   = nelems,
1368                 };
1369
1370                 if (mapped_ents && i >= mapped_ents)
1371                         break;
1372
1373                 if (!i)
1374                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1375
1376                 check_unmap(&ref);
1377         }
1378 }
1379
1380 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1381                               dma_addr_t dma_addr, void *virt)
1382 {
1383         struct dma_debug_entry *entry;
1384
1385         if (unlikely(dma_debug_disabled()))
1386                 return;
1387
1388         if (unlikely(virt == NULL))
1389                 return;
1390
1391         /* handle vmalloc and linear addresses */
1392         if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1393                 return;
1394
1395         entry = dma_entry_alloc();
1396         if (!entry)
1397                 return;
1398
1399         entry->type      = dma_debug_coherent;
1400         entry->dev       = dev;
1401         entry->offset    = offset_in_page(virt);
1402         entry->size      = size;
1403         entry->dev_addr  = dma_addr;
1404         entry->direction = DMA_BIDIRECTIONAL;
1405
1406         if (is_vmalloc_addr(virt))
1407                 entry->pfn = vmalloc_to_pfn(virt);
1408         else
1409                 entry->pfn = page_to_pfn(virt_to_page(virt));
1410
1411         add_dma_entry(entry);
1412 }
1413
1414 void debug_dma_free_coherent(struct device *dev, size_t size,
1415                          void *virt, dma_addr_t addr)
1416 {
1417         struct dma_debug_entry ref = {
1418                 .type           = dma_debug_coherent,
1419                 .dev            = dev,
1420                 .offset         = offset_in_page(virt),
1421                 .dev_addr       = addr,
1422                 .size           = size,
1423                 .direction      = DMA_BIDIRECTIONAL,
1424         };
1425
1426         /* handle vmalloc and linear addresses */
1427         if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1428                 return;
1429
1430         if (is_vmalloc_addr(virt))
1431                 ref.pfn = vmalloc_to_pfn(virt);
1432         else
1433                 ref.pfn = page_to_pfn(virt_to_page(virt));
1434
1435         if (unlikely(dma_debug_disabled()))
1436                 return;
1437
1438         check_unmap(&ref);
1439 }
1440
1441 void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1442                             int direction, dma_addr_t dma_addr)
1443 {
1444         struct dma_debug_entry *entry;
1445
1446         if (unlikely(dma_debug_disabled()))
1447                 return;
1448
1449         entry = dma_entry_alloc();
1450         if (!entry)
1451                 return;
1452
1453         entry->type             = dma_debug_resource;
1454         entry->dev              = dev;
1455         entry->pfn              = PHYS_PFN(addr);
1456         entry->offset           = offset_in_page(addr);
1457         entry->size             = size;
1458         entry->dev_addr         = dma_addr;
1459         entry->direction        = direction;
1460         entry->map_err_type     = MAP_ERR_NOT_CHECKED;
1461
1462         add_dma_entry(entry);
1463 }
1464
1465 void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1466                               size_t size, int direction)
1467 {
1468         struct dma_debug_entry ref = {
1469                 .type           = dma_debug_resource,
1470                 .dev            = dev,
1471                 .dev_addr       = dma_addr,
1472                 .size           = size,
1473                 .direction      = direction,
1474         };
1475
1476         if (unlikely(dma_debug_disabled()))
1477                 return;
1478
1479         check_unmap(&ref);
1480 }
1481
1482 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1483                                    size_t size, int direction)
1484 {
1485         struct dma_debug_entry ref;
1486
1487         if (unlikely(dma_debug_disabled()))
1488                 return;
1489
1490         ref.type         = dma_debug_single;
1491         ref.dev          = dev;
1492         ref.dev_addr     = dma_handle;
1493         ref.size         = size;
1494         ref.direction    = direction;
1495         ref.sg_call_ents = 0;
1496
1497         check_sync(dev, &ref, true);
1498 }
1499
1500 void debug_dma_sync_single_for_device(struct device *dev,
1501                                       dma_addr_t dma_handle, size_t size,
1502                                       int direction)
1503 {
1504         struct dma_debug_entry ref;
1505
1506         if (unlikely(dma_debug_disabled()))
1507                 return;
1508
1509         ref.type         = dma_debug_single;
1510         ref.dev          = dev;
1511         ref.dev_addr     = dma_handle;
1512         ref.size         = size;
1513         ref.direction    = direction;
1514         ref.sg_call_ents = 0;
1515
1516         check_sync(dev, &ref, false);
1517 }
1518
1519 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1520                                int nelems, int direction)
1521 {
1522         struct scatterlist *s;
1523         int mapped_ents = 0, i;
1524
1525         if (unlikely(dma_debug_disabled()))
1526                 return;
1527
1528         for_each_sg(sg, s, nelems, i) {
1529
1530                 struct dma_debug_entry ref = {
1531                         .type           = dma_debug_sg,
1532                         .dev            = dev,
1533                         .pfn            = page_to_pfn(sg_page(s)),
1534                         .offset         = s->offset,
1535                         .dev_addr       = sg_dma_address(s),
1536                         .size           = sg_dma_len(s),
1537                         .direction      = direction,
1538                         .sg_call_ents   = nelems,
1539                 };
1540
1541                 if (!i)
1542                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1543
1544                 if (i >= mapped_ents)
1545                         break;
1546
1547                 check_sync(dev, &ref, true);
1548         }
1549 }
1550
1551 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1552                                   int nelems, int direction)
1553 {
1554         struct scatterlist *s;
1555         int mapped_ents = 0, i;
1556
1557         if (unlikely(dma_debug_disabled()))
1558                 return;
1559
1560         for_each_sg(sg, s, nelems, i) {
1561
1562                 struct dma_debug_entry ref = {
1563                         .type           = dma_debug_sg,
1564                         .dev            = dev,
1565                         .pfn            = page_to_pfn(sg_page(s)),
1566                         .offset         = s->offset,
1567                         .dev_addr       = sg_dma_address(s),
1568                         .size           = sg_dma_len(s),
1569                         .direction      = direction,
1570                         .sg_call_ents   = nelems,
1571                 };
1572                 if (!i)
1573                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1574
1575                 if (i >= mapped_ents)
1576                         break;
1577
1578                 check_sync(dev, &ref, false);
1579         }
1580 }
1581
1582 static int __init dma_debug_driver_setup(char *str)
1583 {
1584         int i;
1585
1586         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1587                 current_driver_name[i] = *str;
1588                 if (*str == 0)
1589                         break;
1590         }
1591
1592         if (current_driver_name[0])
1593                 pr_info("enable driver filter for driver [%s]\n",
1594                         current_driver_name);
1595
1596
1597         return 1;
1598 }
1599 __setup("dma_debug_driver=", dma_debug_driver_setup);