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