lib/dma-debug.c

   1 /*
   2  * Copyright (C) 2008 Advanced Micro Devices, Inc.
   3  *
   4  * Author: Joerg Roedel <joerg.roedel@amd.com>
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 as published
   8  * by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  18  */
  19
  20 #include <linux/scatterlist.h>
  21 #include <linux/dma-mapping.h>
  22 #include <linux/stacktrace.h>
  23 #include <linux/dma-debug.h>
  24 #include <linux/spinlock.h>
  25 #include <linux/debugfs.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/device.h>
  28 #include <linux/types.h>
  29 #include <linux/sched.h>
  30 #include <linux/ctype.h>
  31 #include <linux/list.h>
  32 #include <linux/slab.h>
  33
  34 #include <asm/sections.h>
  35
  36 #define HASH_SIZE       1024ULL
  37 #define HASH_FN_SHIFT   13
  38 #define HASH_FN_MASK    (HASH_SIZE - 1)
  39
  40 enum {
  41         dma_debug_single,
  42         dma_debug_page,
  43         dma_debug_sg,
  44         dma_debug_coherent,
  45 };
  46
  47 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
  48
  49 struct dma_debug_entry {
  50         struct list_head list;
  51         struct device    *dev;
  52         int              type;
  53         phys_addr_t      paddr;
  54         u64              dev_addr;
  55         u64              size;
  56         int              direction;
  57         int              sg_call_ents;
  58         int              sg_mapped_ents;
  59 #ifdef CONFIG_STACKTRACE
  60         struct           stack_trace stacktrace;
  61         unsigned long    st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
  62 #endif
  63 };
  64
  65 struct hash_bucket {
  66         struct list_head list;
  67         spinlock_t lock;
  68 } ____cacheline_aligned_in_smp;
  69
  70 /* Hash list to save the allocated dma addresses */
  71 static struct hash_bucket dma_entry_hash[HASH_SIZE];
  72 /* List of pre-allocated dma_debug_entry's */
  73 static LIST_HEAD(free_entries);
  74 /* Lock for the list above */
  75 static DEFINE_SPINLOCK(free_entries_lock);
  76
  77 /* Global disable flag - will be set in case of an error */
  78 static bool global_disable __read_mostly;
  79
  80 /* Global error count */
  81 static u32 error_count;
  82
  83 /* Global error show enable*/
  84 static u32 show_all_errors __read_mostly;
  85 /* Number of errors to show */
  86 static u32 show_num_errors = 1;
  87
  88 static u32 num_free_entries;
  89 static u32 min_free_entries;
  90 static u32 nr_total_entries;
  91
  92 /* number of preallocated entries requested by kernel cmdline */
  93 static u32 req_entries;
  94
  95 /* debugfs dentry's for the stuff above */
  96 static struct dentry *dma_debug_dent        __read_mostly;
  97 static struct dentry *global_disable_dent   __read_mostly;
  98 static struct dentry *error_count_dent      __read_mostly;
  99 static struct dentry *show_all_errors_dent  __read_mostly;
 100 static struct dentry *show_num_errors_dent  __read_mostly;
 101 static struct dentry *num_free_entries_dent __read_mostly;
 102 static struct dentry *min_free_entries_dent __read_mostly;
 103 static struct dentry *filter_dent           __read_mostly;
 104
 105 /* per-driver filter related state */
 106
 107 #define NAME_MAX_LEN    64
 108
 109 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
 110 static struct device_driver *current_driver                    __read_mostly;
 111
 112 static DEFINE_RWLOCK(driver_name_lock);
 113
 114 static const char *type2name[4] = { "single", "page",
 115                                     "scather-gather", "coherent" };
 116
 117 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
 118                                    "DMA_FROM_DEVICE", "DMA_NONE" };
 119
 120 /* little merge helper - remove it after the merge window */
 121 #ifndef BUS_NOTIFY_UNBOUND_DRIVER
 122 #define BUS_NOTIFY_UNBOUND_DRIVER 0x0005
 123 #endif
 124
 125 /*
 126  * The access to some variables in this macro is racy. We can't use atomic_t
 127  * here because all these variables are exported to debugfs. Some of them even
 128  * writeable. This is also the reason why a lock won't help much. But anyway,
 129  * the races are no big deal. Here is why:
 130  *
 131  *   error_count: the addition is racy, but the worst thing that can happen is
 132  *                that we don't count some errors
 133  *   show_num_errors: the subtraction is racy. Also no big deal because in
 134  *                    worst case this will result in one warning more in the
 135  *                    system log than the user configured. This variable is
 136  *                    writeable via debugfs.
 137  */
 138 static inline void dump_entry_trace(struct dma_debug_entry *entry)
 139 {
 140 #ifdef CONFIG_STACKTRACE
 141         if (entry) {
 142                 pr_warning("Mapped at:\n");
 143                 print_stack_trace(&entry->stacktrace, 0);
 144         }
 145 #endif
 146 }
 147
 148 static bool driver_filter(struct device *dev)
 149 {
 150         struct device_driver *drv;
 151         unsigned long flags;
 152         bool ret;
 153
 154         /* driver filter off */
 155         if (likely(!current_driver_name[0]))
 156                 return true;
 157
 158         /* driver filter on and initialized */
 159         if (current_driver && dev && dev->driver == current_driver)
 160                 return true;
 161
 162         /* driver filter on, but we can't filter on a NULL device... */
 163         if (!dev)
 164                 return false;
 165
 166         if (current_driver || !current_driver_name[0])
 167                 return false;
 168
 169         /* driver filter on but not yet initialized */
 170         drv = get_driver(dev->driver);
 171         if (!drv)
 172                 return false;
 173
 174         /* lock to protect against change of current_driver_name */
 175         read_lock_irqsave(&driver_name_lock, flags);
 176
 177         ret = false;
 178         if (drv->name &&
 179             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 180                 current_driver = drv;
 181                 ret = true;
 182         }
 183
 184         read_unlock_irqrestore(&driver_name_lock, flags);
 185         put_driver(drv);
 186
 187         return ret;
 188 }
 189
 190 #define err_printk(dev, entry, format, arg...) do {                     \
 191                 error_count += 1;                                       \
 192                 if (driver_filter(dev) &&                               \
 193                     (show_all_errors || show_num_errors > 0)) {         \
 194                         WARN(1, "%s %s: " format,                       \
 195                              dev ? dev_driver_string(dev) : "NULL",     \
 196                              dev ? dev_name(dev) : "NULL", ## arg);     \
 197                         dump_entry_trace(entry);                        \
 198                 }                                                       \
 199                 if (!show_all_errors && show_num_errors > 0)            \
 200                         show_num_errors -= 1;                           \
 201         } while (0);
 202
 203 /*
 204  * Hash related functions
 205  *
 206  * Every DMA-API request is saved into a struct dma_debug_entry. To
 207  * have quick access to these structs they are stored into a hash.
 208  */
 209 static int hash_fn(struct dma_debug_entry *entry)
 210 {
 211         /*
 212          * Hash function is based on the dma address.
 213          * We use bits 20-27 here as the index into the hash
 214          */
 215         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 216 }
 217
 218 /*
 219  * Request exclusive access to a hash bucket for a given dma_debug_entry.
 220  */
 221 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 222                                            unsigned long *flags)
 223 {
 224         int idx = hash_fn(entry);
 225         unsigned long __flags;
 226
 227         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 228         *flags = __flags;
 229         return &dma_entry_hash[idx];
 230 }
 231
 232 /*
 233  * Give up exclusive access to the hash bucket
 234  */
 235 static void put_hash_bucket(struct hash_bucket *bucket,
 236                             unsigned long *flags)
 237 {
 238         unsigned long __flags = *flags;
 239
 240         spin_unlock_irqrestore(&bucket->lock, __flags);
 241 }
 242
 243 /*
 244  * Search a given entry in the hash bucket list
 245  */
 246 static struct dma_debug_entry *hash_bucket_find(struct hash_bucket *bucket,
 247                                                 struct dma_debug_entry *ref)
 248 {
 249         struct dma_debug_entry *entry, *ret = NULL;
 250         int matches = 0, match_lvl, last_lvl = 0;
 251
 252         list_for_each_entry(entry, &bucket->list, list) {
 253                 if ((entry->dev_addr != ref->dev_addr) ||
 254                     (entry->dev != ref->dev))
 255                         continue;
 256
 257                 /*
 258                  * Some drivers map the same physical address multiple
 259                  * times. Without a hardware IOMMU this results in the
 260                  * same device addresses being put into the dma-debug
 261                  * hash multiple times too. This can result in false
 262                  * positives being reported. Therefore we implement a
 263                  * best-fit algorithm here which returns the entry from
 264                  * the hash which fits best to the reference value
 265                  * instead of the first-fit.
 266                  */
 267                 matches += 1;
 268                 match_lvl = 0;
 269                 entry->size         == ref->size         ? ++match_lvl : 0;
 270                 entry->type         == ref->type         ? ++match_lvl : 0;
 271                 entry->direction    == ref->direction    ? ++match_lvl : 0;
 272                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
 273
 274                 if (match_lvl == 4) {
 275                         /* perfect-fit - return the result */
 276                         return entry;
 277                 } else if (match_lvl > last_lvl) {
 278                         /*
 279                          * We found an entry that fits better then the
 280                          * previous one
 281                          */
 282                         last_lvl = match_lvl;
 283                         ret      = entry;
 284                 }
 285         }
 286
 287         /*
 288          * If we have multiple matches but no perfect-fit, just return
 289          * NULL.
 290          */
 291         ret = (matches == 1) ? ret : NULL;
 292
 293         return ret;
 294 }
 295
 296 /*
 297  * Add an entry to a hash bucket
 298  */
 299 static void hash_bucket_add(struct hash_bucket *bucket,
 300                             struct dma_debug_entry *entry)
 301 {
 302         list_add_tail(&entry->list, &bucket->list);
 303 }
 304
 305 /*
 306  * Remove entry from a hash bucket list
 307  */
 308 static void hash_bucket_del(struct dma_debug_entry *entry)
 309 {
 310         list_del(&entry->list);
 311 }
 312
 313 /*
 314  * Dump mapping entries for debugging purposes
 315  */
 316 void debug_dma_dump_mappings(struct device *dev)
 317 {
 318         int idx;
 319
 320         for (idx = 0; idx < HASH_SIZE; idx++) {
 321                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 322                 struct dma_debug_entry *entry;
 323                 unsigned long flags;
 324
 325                 spin_lock_irqsave(&bucket->lock, flags);
 326
 327                 list_for_each_entry(entry, &bucket->list, list) {
 328                         if (!dev || dev == entry->dev) {
 329                                 dev_info(entry->dev,
 330                                          "%s idx %d P=%Lx D=%Lx L=%Lx %s\n",
 331                                          type2name[entry->type], idx,
 332                                          (unsigned long long)entry->paddr,
 333                                          entry->dev_addr, entry->size,
 334                                          dir2name[entry->direction]);
 335                         }
 336                 }
 337
 338                 spin_unlock_irqrestore(&bucket->lock, flags);
 339         }
 340 }
 341 EXPORT_SYMBOL(debug_dma_dump_mappings);
 342
 343 /*
 344  * Wrapper function for adding an entry to the hash.
 345  * This function takes care of locking itself.
 346  */
 347 static void add_dma_entry(struct dma_debug_entry *entry)
 348 {
 349         struct hash_bucket *bucket;
 350         unsigned long flags;
 351
 352         bucket = get_hash_bucket(entry, &flags);
 353         hash_bucket_add(bucket, entry);
 354         put_hash_bucket(bucket, &flags);
 355 }
 356
 357 static struct dma_debug_entry *__dma_entry_alloc(void)
 358 {
 359         struct dma_debug_entry *entry;
 360
 361         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 362         list_del(&entry->list);
 363         memset(entry, 0, sizeof(*entry));
 364
 365         num_free_entries -= 1;
 366         if (num_free_entries < min_free_entries)
 367                 min_free_entries = num_free_entries;
 368
 369         return entry;
 370 }
 371
 372 /* struct dma_entry allocator
 373  *
 374  * The next two functions implement the allocator for
 375  * struct dma_debug_entries.
 376  */
 377 static struct dma_debug_entry *dma_entry_alloc(void)
 378 {
 379         struct dma_debug_entry *entry = NULL;
 380         unsigned long flags;
 381
 382         spin_lock_irqsave(&free_entries_lock, flags);
 383
 384         if (list_empty(&free_entries)) {
 385                 pr_err("DMA-API: debugging out of memory - disabling\n");
 386                 global_disable = true;
 387                 goto out;
 388         }
 389
 390         entry = __dma_entry_alloc();
 391
 392 #ifdef CONFIG_STACKTRACE
 393         entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
 394         entry->stacktrace.entries = entry->st_entries;
 395         entry->stacktrace.skip = 2;
 396         save_stack_trace(&entry->stacktrace);
 397 #endif
 398
 399 out:
 400         spin_unlock_irqrestore(&free_entries_lock, flags);
 401
 402         return entry;
 403 }
 404
 405 static void dma_entry_free(struct dma_debug_entry *entry)
 406 {
 407         unsigned long flags;
 408
 409         /*
 410          * add to beginning of the list - this way the entries are
 411          * more likely cache hot when they are reallocated.
 412          */
 413         spin_lock_irqsave(&free_entries_lock, flags);
 414         list_add(&entry->list, &free_entries);
 415         num_free_entries += 1;
 416         spin_unlock_irqrestore(&free_entries_lock, flags);
 417 }
 418
 419 int dma_debug_resize_entries(u32 num_entries)
 420 {
 421         int i, delta, ret = 0;
 422         unsigned long flags;
 423         struct dma_debug_entry *entry;
 424         LIST_HEAD(tmp);
 425
 426         spin_lock_irqsave(&free_entries_lock, flags);
 427
 428         if (nr_total_entries < num_entries) {
 429                 delta = num_entries - nr_total_entries;
 430
 431                 spin_unlock_irqrestore(&free_entries_lock, flags);
 432
 433                 for (i = 0; i < delta; i++) {
 434                         entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 435                         if (!entry)
 436                                 break;
 437
 438                         list_add_tail(&entry->list, &tmp);
 439                 }
 440
 441                 spin_lock_irqsave(&free_entries_lock, flags);
 442
 443                 list_splice(&tmp, &free_entries);
 444                 nr_total_entries += i;
 445                 num_free_entries += i;
 446         } else {
 447                 delta = nr_total_entries - num_entries;
 448
 449                 for (i = 0; i < delta && !list_empty(&free_entries); i++) {
 450                         entry = __dma_entry_alloc();
 451                         kfree(entry);
 452                 }
 453
 454                 nr_total_entries -= i;
 455         }
 456
 457         if (nr_total_entries != num_entries)
 458                 ret = 1;
 459
 460         spin_unlock_irqrestore(&free_entries_lock, flags);
 461
 462         return ret;
 463 }
 464 EXPORT_SYMBOL(dma_debug_resize_entries);
 465
 466 /*
 467  * DMA-API debugging init code
 468  *
 469  * The init code does two things:
 470  *   1. Initialize core data structures
 471  *   2. Preallocate a given number of dma_debug_entry structs
 472  */
 473
 474 static int prealloc_memory(u32 num_entries)
 475 {
 476         struct dma_debug_entry *entry, *next_entry;
 477         int i;
 478
 479         for (i = 0; i < num_entries; ++i) {
 480                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 481                 if (!entry)
 482                         goto out_err;
 483
 484                 list_add_tail(&entry->list, &free_entries);
 485         }
 486
 487         num_free_entries = num_entries;
 488         min_free_entries = num_entries;
 489
 490         pr_info("DMA-API: preallocated %d debug entries\n", num_entries);
 491
 492         return 0;
 493
 494 out_err:
 495
 496         list_for_each_entry_safe(entry, next_entry, &free_entries, list) {
 497                 list_del(&entry->list);
 498                 kfree(entry);
 499         }
 500
 501         return -ENOMEM;
 502 }
 503
 504 static ssize_t filter_read(struct file *file, char __user *user_buf,
 505                            size_t count, loff_t *ppos)
 506 {
 507         char buf[NAME_MAX_LEN + 1];
 508         unsigned long flags;
 509         int len;
 510
 511         if (!current_driver_name[0])
 512                 return 0;
 513
 514         /*
 515          * We can't copy to userspace directly because current_driver_name can
 516          * only be read under the driver_name_lock with irqs disabled. So
 517          * create a temporary copy first.
 518          */
 519         read_lock_irqsave(&driver_name_lock, flags);
 520         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 521         read_unlock_irqrestore(&driver_name_lock, flags);
 522
 523         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 524 }
 525
 526 static ssize_t filter_write(struct file *file, const char __user *userbuf,
 527                             size_t count, loff_t *ppos)
 528 {
 529         char buf[NAME_MAX_LEN];
 530         unsigned long flags;
 531         size_t len;
 532         int i;
 533
 534         /*
 535          * We can't copy from userspace directly. Access to
 536          * current_driver_name is protected with a write_lock with irqs
 537          * disabled. Since copy_from_user can fault and may sleep we
 538          * need to copy to temporary buffer first
 539          */
 540         len = min(count, (size_t)(NAME_MAX_LEN - 1));
 541         if (copy_from_user(buf, userbuf, len))
 542                 return -EFAULT;
 543
 544         buf[len] = 0;
 545
 546         write_lock_irqsave(&driver_name_lock, flags);
 547
 548         /*
 549          * Now handle the string we got from userspace very carefully.
 550          * The rules are:
 551          *         - only use the first token we got
 552          *         - token delimiter is everything looking like a space
 553          *           character (' ', '\n', '\t' ...)
 554          *
 555          */
 556         if (!isalnum(buf[0])) {
 557                 /*
 558                  * If the first character userspace gave us is not
 559                  * alphanumerical then assume the filter should be
 560                  * switched off.
 561                  */
 562                 if (current_driver_name[0])
 563                         pr_info("DMA-API: switching off dma-debug driver filter\n");
 564                 current_driver_name[0] = 0;
 565                 current_driver = NULL;
 566                 goto out_unlock;
 567         }
 568
 569         /*
 570          * Now parse out the first token and use it as the name for the
 571          * driver to filter for.
 572          */
 573         for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 574                 current_driver_name[i] = buf[i];
 575                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 576                         break;
 577         }
 578         current_driver_name[i] = 0;
 579         current_driver = NULL;
 580
 581         pr_info("DMA-API: enable driver filter for driver [%s]\n",
 582                 current_driver_name);
 583
 584 out_unlock:
 585         write_unlock_irqrestore(&driver_name_lock, flags);
 586
 587         return count;
 588 }
 589
 590 static const struct file_operations filter_fops = {
 591         .read  = filter_read,
 592         .write = filter_write,
 593         .llseek = default_llseek,
 594 };
 595
 596 static int dma_debug_fs_init(void)
 597 {
 598         dma_debug_dent = debugfs_create_dir("dma-api", NULL);
 599         if (!dma_debug_dent) {
 600                 pr_err("DMA-API: can not create debugfs directory\n");
 601                 return -ENOMEM;
 602         }
 603
 604         global_disable_dent = debugfs_create_bool("disabled", 0444,
 605                         dma_debug_dent,
 606                         (u32 *)&global_disable);
 607         if (!global_disable_dent)
 608                 goto out_err;
 609
 610         error_count_dent = debugfs_create_u32("error_count", 0444,
 611                         dma_debug_dent, &error_count);
 612         if (!error_count_dent)
 613                 goto out_err;
 614
 615         show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
 616                         dma_debug_dent,
 617                         &show_all_errors);
 618         if (!show_all_errors_dent)
 619                 goto out_err;
 620
 621         show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
 622                         dma_debug_dent,
 623                         &show_num_errors);
 624         if (!show_num_errors_dent)
 625                 goto out_err;
 626
 627         num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
 628                         dma_debug_dent,
 629                         &num_free_entries);
 630         if (!num_free_entries_dent)
 631                 goto out_err;
 632
 633         min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
 634                         dma_debug_dent,
 635                         &min_free_entries);
 636         if (!min_free_entries_dent)
 637                 goto out_err;
 638
 639         filter_dent = debugfs_create_file("driver_filter", 0644,
 640                                           dma_debug_dent, NULL, &filter_fops);
 641         if (!filter_dent)
 642                 goto out_err;
 643
 644         return 0;
 645
 646 out_err:
 647         debugfs_remove_recursive(dma_debug_dent);
 648
 649         return -ENOMEM;
 650 }
 651
 652 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 653 {
 654         struct dma_debug_entry *entry;
 655         unsigned long flags;
 656         int count = 0, i;
 657
 658         local_irq_save(flags);
 659
 660         for (i = 0; i < HASH_SIZE; ++i) {
 661                 spin_lock(&dma_entry_hash[i].lock);
 662                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 663                         if (entry->dev == dev) {
 664                                 count += 1;
 665                                 *out_entry = entry;
 666                         }
 667                 }
 668                 spin_unlock(&dma_entry_hash[i].lock);
 669         }
 670
 671         local_irq_restore(flags);
 672
 673         return count;
 674 }
 675
 676 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 677 {
 678         struct device *dev = data;
 679         struct dma_debug_entry *uninitialized_var(entry);
 680         int count;
 681
 682         if (global_disable)
 683                 return 0;
 684
 685         switch (action) {
 686         case BUS_NOTIFY_UNBOUND_DRIVER:
 687                 count = device_dma_allocations(dev, &entry);
 688                 if (count == 0)
 689                         break;
 690                 err_printk(dev, entry, "DMA-API: device driver has pending "
 691                                 "DMA allocations while released from device "
 692                                 "[count=%d]\n"
 693                                 "One of leaked entries details: "
 694                                 "[device address=0x%016llx] [size=%llu bytes] "
 695                                 "[mapped with %s] [mapped as %s]\n",
 696                         count, entry->dev_addr, entry->size,
 697                         dir2name[entry->direction], type2name[entry->type]);
 698                 break;
 699         default:
 700                 break;
 701         }
 702
 703         return 0;
 704 }
 705
 706 void dma_debug_add_bus(struct bus_type *bus)
 707 {
 708         struct notifier_block *nb;
 709
 710         if (global_disable)
 711                 return;
 712
 713         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 714         if (nb == NULL) {
 715                 pr_err("dma_debug_add_bus: out of memory\n");
 716                 return;
 717         }
 718
 719         nb->notifier_call = dma_debug_device_change;
 720
 721         bus_register_notifier(bus, nb);
 722 }
 723
 724 /*
 725  * Let the architectures decide how many entries should be preallocated.
 726  */
 727 void dma_debug_init(u32 num_entries)
 728 {
 729         int i;
 730
 731         if (global_disable)
 732                 return;
 733
 734         for (i = 0; i < HASH_SIZE; ++i) {
 735                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
 736                 spin_lock_init(&dma_entry_hash[i].lock);
 737         }
 738
 739         if (dma_debug_fs_init() != 0) {
 740                 pr_err("DMA-API: error creating debugfs entries - disabling\n");
 741                 global_disable = true;
 742
 743                 return;
 744         }
 745
 746         if (req_entries)
 747                 num_entries = req_entries;
 748
 749         if (prealloc_memory(num_entries) != 0) {
 750                 pr_err("DMA-API: debugging out of memory error - disabled\n");
 751                 global_disable = true;
 752
 753                 return;
 754         }
 755
 756         nr_total_entries = num_free_entries;
 757
 758         pr_info("DMA-API: debugging enabled by kernel config\n");
 759 }
 760
 761 static __init int dma_debug_cmdline(char *str)
 762 {
 763         if (!str)
 764                 return -EINVAL;
 765
 766         if (strncmp(str, "off", 3) == 0) {
 767                 pr_info("DMA-API: debugging disabled on kernel command line\n");
 768                 global_disable = true;
 769         }
 770
 771         return 0;
 772 }
 773
 774 static __init int dma_debug_entries_cmdline(char *str)
 775 {
 776         int res;
 777
 778         if (!str)
 779                 return -EINVAL;
 780
 781         res = get_option(&str, &req_entries);
 782
 783         if (!res)
 784                 req_entries = 0;
 785
 786         return 0;
 787 }
 788
 789 __setup("dma_debug=", dma_debug_cmdline);
 790 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
 791
 792 static void check_unmap(struct dma_debug_entry *ref)
 793 {
 794         struct dma_debug_entry *entry;
 795         struct hash_bucket *bucket;
 796         unsigned long flags;
 797
 798         if (dma_mapping_error(ref->dev, ref->dev_addr)) {
 799                 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
 800                            "to free an invalid DMA memory address\n");
 801                 return;
 802         }
 803
 804         bucket = get_hash_bucket(ref, &flags);
 805         entry = hash_bucket_find(bucket, ref);
 806
 807         if (!entry) {
 808                 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
 809                            "to free DMA memory it has not allocated "
 810                            "[device address=0x%016llx] [size=%llu bytes]\n",
 811                            ref->dev_addr, ref->size);
 812                 goto out;
 813         }
 814
 815         if (ref->size != entry->size) {
 816                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 817                            "DMA memory with different size "
 818                            "[device address=0x%016llx] [map size=%llu bytes] "
 819                            "[unmap size=%llu bytes]\n",
 820                            ref->dev_addr, entry->size, ref->size);
 821         }
 822
 823         if (ref->type != entry->type) {
 824                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 825                            "DMA memory with wrong function "
 826                            "[device address=0x%016llx] [size=%llu bytes] "
 827                            "[mapped as %s] [unmapped as %s]\n",
 828                            ref->dev_addr, ref->size,
 829                            type2name[entry->type], type2name[ref->type]);
 830         } else if ((entry->type == dma_debug_coherent) &&
 831                    (ref->paddr != entry->paddr)) {
 832                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 833                            "DMA memory with different CPU address "
 834                            "[device address=0x%016llx] [size=%llu bytes] "
 835                            "[cpu alloc address=0x%016llx] "
 836                            "[cpu free address=0x%016llx]",
 837                            ref->dev_addr, ref->size,
 838                            (unsigned long long)entry->paddr,
 839                            (unsigned long long)ref->paddr);
 840         }
 841
 842         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
 843             ref->sg_call_ents != entry->sg_call_ents) {
 844                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 845                            "DMA sg list with different entry count "
 846                            "[map count=%d] [unmap count=%d]\n",
 847                            entry->sg_call_ents, ref->sg_call_ents);
 848         }
 849
 850         /*
 851          * This may be no bug in reality - but most implementations of the
 852          * DMA API don't handle this properly, so check for it here
 853          */
 854         if (ref->direction != entry->direction) {
 855                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 856                            "DMA memory with different direction "
 857                            "[device address=0x%016llx] [size=%llu bytes] "
 858                            "[mapped with %s] [unmapped with %s]\n",
 859                            ref->dev_addr, ref->size,
 860                            dir2name[entry->direction],
 861                            dir2name[ref->direction]);
 862         }
 863
 864         hash_bucket_del(entry);
 865         dma_entry_free(entry);
 866
 867 out:
 868         put_hash_bucket(bucket, &flags);
 869 }
 870
 871 static void check_for_stack(struct device *dev, void *addr)
 872 {
 873         if (object_is_on_stack(addr))
 874                 err_printk(dev, NULL, "DMA-API: device driver maps memory from"
 875                                 "stack [addr=%p]\n", addr);
 876 }
 877
 878 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
 879 {
 880         unsigned long a1 = (unsigned long)addr;
 881         unsigned long b1 = a1 + len;
 882         unsigned long a2 = (unsigned long)start;
 883         unsigned long b2 = (unsigned long)end;
 884
 885         return !(b1 <= a2 || a1 >= b2);
 886 }
 887
 888 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
 889 {
 890         if (overlap(addr, len, _text, _etext) ||
 891             overlap(addr, len, __start_rodata, __end_rodata))
 892                 err_printk(dev, NULL, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
 893 }
 894
 895 static void check_sync(struct device *dev,
 896                        struct dma_debug_entry *ref,
 897                        bool to_cpu)
 898 {
 899         struct dma_debug_entry *entry;
 900         struct hash_bucket *bucket;
 901         unsigned long flags;
 902
 903         bucket = get_hash_bucket(ref, &flags);
 904
 905         entry = hash_bucket_find(bucket, ref);
 906
 907         if (!entry) {
 908                 err_printk(dev, NULL, "DMA-API: device driver tries "
 909                                 "to sync DMA memory it has not allocated "
 910                                 "[device address=0x%016llx] [size=%llu bytes]\n",
 911                                 (unsigned long long)ref->dev_addr, ref->size);
 912                 goto out;
 913         }
 914
 915         if (ref->size > entry->size) {
 916                 err_printk(dev, entry, "DMA-API: device driver syncs"
 917                                 " DMA memory outside allocated range "
 918                                 "[device address=0x%016llx] "
 919                                 "[allocation size=%llu bytes] "
 920                                 "[sync offset+size=%llu]\n",
 921                                 entry->dev_addr, entry->size,
 922                                 ref->size);
 923         }
 924
 925         if (entry->direction == DMA_BIDIRECTIONAL)
 926                 goto out;
 927
 928         if (ref->direction != entry->direction) {
 929                 err_printk(dev, entry, "DMA-API: device driver syncs "
 930                                 "DMA memory with different direction "
 931                                 "[device address=0x%016llx] [size=%llu bytes] "
 932                                 "[mapped with %s] [synced with %s]\n",
 933                                 (unsigned long long)ref->dev_addr, entry->size,
 934                                 dir2name[entry->direction],
 935                                 dir2name[ref->direction]);
 936         }
 937
 938         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
 939                       !(ref->direction == DMA_TO_DEVICE))
 940                 err_printk(dev, entry, "DMA-API: device driver syncs "
 941                                 "device read-only DMA memory for cpu "
 942                                 "[device address=0x%016llx] [size=%llu bytes] "
 943                                 "[mapped with %s] [synced with %s]\n",
 944                                 (unsigned long long)ref->dev_addr, entry->size,
 945                                 dir2name[entry->direction],
 946                                 dir2name[ref->direction]);
 947
 948         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
 949                        !(ref->direction == DMA_FROM_DEVICE))
 950                 err_printk(dev, entry, "DMA-API: device driver syncs "
 951                                 "device write-only DMA memory to device "
 952                                 "[device address=0x%016llx] [size=%llu bytes] "
 953                                 "[mapped with %s] [synced with %s]\n",
 954                                 (unsigned long long)ref->dev_addr, entry->size,
 955                                 dir2name[entry->direction],
 956                                 dir2name[ref->direction]);
 957
 958 out:
 959         put_hash_bucket(bucket, &flags);
 960 }
 961
 962 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
 963                         size_t size, int direction, dma_addr_t dma_addr,
 964                         bool map_single)
 965 {
 966         struct dma_debug_entry *entry;
 967
 968         if (unlikely(global_disable))
 969                 return;
 970
 971         if (unlikely(dma_mapping_error(dev, dma_addr)))
 972                 return;
 973
 974         entry = dma_entry_alloc();
 975         if (!entry)
 976                 return;
 977
 978         entry->dev       = dev;
 979         entry->type      = dma_debug_page;
 980         entry->paddr     = page_to_phys(page) + offset;
 981         entry->dev_addr  = dma_addr;
 982         entry->size      = size;
 983         entry->direction = direction;
 984
 985         if (map_single)
 986                 entry->type = dma_debug_single;
 987
 988         if (!PageHighMem(page)) {
 989                 void *addr = page_address(page) + offset;
 990
 991                 check_for_stack(dev, addr);
 992                 check_for_illegal_area(dev, addr, size);
 993         }
 994
 995         add_dma_entry(entry);
 996 }
 997 EXPORT_SYMBOL(debug_dma_map_page);
 998
 999 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1000                           size_t size, int direction, bool map_single)
1001 {
1002         struct dma_debug_entry ref = {
1003                 .type           = dma_debug_page,
1004                 .dev            = dev,
1005                 .dev_addr       = addr,
1006                 .size           = size,
1007                 .direction      = direction,
1008         };
1009
1010         if (unlikely(global_disable))
1011                 return;
1012
1013         if (map_single)
1014                 ref.type = dma_debug_single;
1015
1016         check_unmap(&ref);
1017 }
1018 EXPORT_SYMBOL(debug_dma_unmap_page);
1019
1020 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1021                       int nents, int mapped_ents, int direction)
1022 {
1023         struct dma_debug_entry *entry;
1024         struct scatterlist *s;
1025         int i;
1026
1027         if (unlikely(global_disable))
1028                 return;
1029
1030         for_each_sg(sg, s, mapped_ents, i) {
1031                 entry = dma_entry_alloc();
1032                 if (!entry)
1033                         return;
1034
1035                 entry->type           = dma_debug_sg;
1036                 entry->dev            = dev;
1037                 entry->paddr          = sg_phys(s);
1038                 entry->size           = sg_dma_len(s);
1039                 entry->dev_addr       = sg_dma_address(s);
1040                 entry->direction      = direction;
1041                 entry->sg_call_ents   = nents;
1042                 entry->sg_mapped_ents = mapped_ents;
1043
1044                 if (!PageHighMem(sg_page(s))) {
1045                         check_for_stack(dev, sg_virt(s));
1046                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1047                 }
1048
1049                 add_dma_entry(entry);
1050         }
1051 }
1052 EXPORT_SYMBOL(debug_dma_map_sg);
1053
1054 static int get_nr_mapped_entries(struct device *dev,
1055                                  struct dma_debug_entry *ref)
1056 {
1057         struct dma_debug_entry *entry;
1058         struct hash_bucket *bucket;
1059         unsigned long flags;
1060         int mapped_ents;
1061
1062         bucket       = get_hash_bucket(ref, &flags);
1063         entry        = hash_bucket_find(bucket, ref);
1064         mapped_ents  = 0;
1065
1066         if (entry)
1067                 mapped_ents = entry->sg_mapped_ents;
1068         put_hash_bucket(bucket, &flags);
1069
1070         return mapped_ents;
1071 }
1072
1073 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1074                         int nelems, int dir)
1075 {
1076         struct scatterlist *s;
1077         int mapped_ents = 0, i;
1078
1079         if (unlikely(global_disable))
1080                 return;
1081
1082         for_each_sg(sglist, s, nelems, i) {
1083
1084                 struct dma_debug_entry ref = {
1085                         .type           = dma_debug_sg,
1086                         .dev            = dev,
1087                         .paddr          = sg_phys(s),
1088                         .dev_addr       = sg_dma_address(s),
1089                         .size           = sg_dma_len(s),
1090                         .direction      = dir,
1091                         .sg_call_ents   = nelems,
1092                 };
1093
1094                 if (mapped_ents && i >= mapped_ents)
1095                         break;
1096
1097                 if (!i)
1098                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1099
1100                 check_unmap(&ref);
1101         }
1102 }
1103 EXPORT_SYMBOL(debug_dma_unmap_sg);
1104
1105 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1106                               dma_addr_t dma_addr, void *virt)
1107 {
1108         struct dma_debug_entry *entry;
1109
1110         if (unlikely(global_disable))
1111                 return;
1112
1113         if (unlikely(virt == NULL))
1114                 return;
1115
1116         entry = dma_entry_alloc();
1117         if (!entry)
1118                 return;
1119
1120         entry->type      = dma_debug_coherent;
1121         entry->dev       = dev;
1122         entry->paddr     = virt_to_phys(virt);
1123         entry->size      = size;
1124         entry->dev_addr  = dma_addr;
1125         entry->direction = DMA_BIDIRECTIONAL;
1126
1127         add_dma_entry(entry);
1128 }
1129 EXPORT_SYMBOL(debug_dma_alloc_coherent);
1130
1131 void debug_dma_free_coherent(struct device *dev, size_t size,
1132                          void *virt, dma_addr_t addr)
1133 {
1134         struct dma_debug_entry ref = {
1135                 .type           = dma_debug_coherent,
1136                 .dev            = dev,
1137                 .paddr          = virt_to_phys(virt),
1138                 .dev_addr       = addr,
1139                 .size           = size,
1140                 .direction      = DMA_BIDIRECTIONAL,
1141         };
1142
1143         if (unlikely(global_disable))
1144                 return;
1145
1146         check_unmap(&ref);
1147 }
1148 EXPORT_SYMBOL(debug_dma_free_coherent);
1149
1150 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1151                                    size_t size, int direction)
1152 {
1153         struct dma_debug_entry ref;
1154
1155         if (unlikely(global_disable))
1156                 return;
1157
1158         ref.type         = dma_debug_single;
1159         ref.dev          = dev;
1160         ref.dev_addr     = dma_handle;
1161         ref.size         = size;
1162         ref.direction    = direction;
1163         ref.sg_call_ents = 0;
1164
1165         check_sync(dev, &ref, true);
1166 }
1167 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1168
1169 void debug_dma_sync_single_for_device(struct device *dev,
1170                                       dma_addr_t dma_handle, size_t size,
1171                                       int direction)
1172 {
1173         struct dma_debug_entry ref;
1174
1175         if (unlikely(global_disable))
1176                 return;
1177
1178         ref.type         = dma_debug_single;
1179         ref.dev          = dev;
1180         ref.dev_addr     = dma_handle;
1181         ref.size         = size;
1182         ref.direction    = direction;
1183         ref.sg_call_ents = 0;
1184
1185         check_sync(dev, &ref, false);
1186 }
1187 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1188
1189 void debug_dma_sync_single_range_for_cpu(struct device *dev,
1190                                          dma_addr_t dma_handle,
1191                                          unsigned long offset, size_t size,
1192                                          int direction)
1193 {
1194         struct dma_debug_entry ref;
1195
1196         if (unlikely(global_disable))
1197                 return;
1198
1199         ref.type         = dma_debug_single;
1200         ref.dev          = dev;
1201         ref.dev_addr     = dma_handle;
1202         ref.size         = offset + size;
1203         ref.direction    = direction;
1204         ref.sg_call_ents = 0;
1205
1206         check_sync(dev, &ref, true);
1207 }
1208 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
1209
1210 void debug_dma_sync_single_range_for_device(struct device *dev,
1211                                             dma_addr_t dma_handle,
1212                                             unsigned long offset,
1213                                             size_t size, int direction)
1214 {
1215         struct dma_debug_entry ref;
1216
1217         if (unlikely(global_disable))
1218                 return;
1219
1220         ref.type         = dma_debug_single;
1221         ref.dev          = dev;
1222         ref.dev_addr     = dma_handle;
1223         ref.size         = offset + size;
1224         ref.direction    = direction;
1225         ref.sg_call_ents = 0;
1226
1227         check_sync(dev, &ref, false);
1228 }
1229 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
1230
1231 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1232                                int nelems, int direction)
1233 {
1234         struct scatterlist *s;
1235         int mapped_ents = 0, i;
1236
1237         if (unlikely(global_disable))
1238                 return;
1239
1240         for_each_sg(sg, s, nelems, i) {
1241
1242                 struct dma_debug_entry ref = {
1243                         .type           = dma_debug_sg,
1244                         .dev            = dev,
1245                         .paddr          = sg_phys(s),
1246                         .dev_addr       = sg_dma_address(s),
1247                         .size           = sg_dma_len(s),
1248                         .direction      = direction,
1249                         .sg_call_ents   = nelems,
1250                 };
1251
1252                 if (!i)
1253                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1254
1255                 if (i >= mapped_ents)
1256                         break;
1257
1258                 check_sync(dev, &ref, true);
1259         }
1260 }
1261 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1262
1263 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1264                                   int nelems, int direction)
1265 {
1266         struct scatterlist *s;
1267         int mapped_ents = 0, i;
1268
1269         if (unlikely(global_disable))
1270                 return;
1271
1272         for_each_sg(sg, s, nelems, i) {
1273
1274                 struct dma_debug_entry ref = {
1275                         .type           = dma_debug_sg,
1276                         .dev            = dev,
1277                         .paddr          = sg_phys(s),
1278                         .dev_addr       = sg_dma_address(s),
1279                         .size           = sg_dma_len(s),
1280                         .direction      = direction,
1281                         .sg_call_ents   = nelems,
1282                 };
1283                 if (!i)
1284                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1285
1286                 if (i >= mapped_ents)
1287                         break;
1288
1289                 check_sync(dev, &ref, false);
1290         }
1291 }
1292 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1293
1294 static int __init dma_debug_driver_setup(char *str)
1295 {
1296         int i;
1297
1298         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1299                 current_driver_name[i] = *str;
1300                 if (*str == 0)
1301                         break;
1302         }
1303
1304         if (current_driver_name[0])
1305                 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1306                         current_driver_name);
1307
1308
1309         return 1;
1310 }
1311 __setup("dma_debug_driver=", dma_debug_driver_setup);