lib/dma-debug.c

   1 /*
   2  * Copyright (C) 2008 Advanced Micro Devices, Inc.
   3  *
   4  * Author: Joerg Roedel <joerg.roedel@amd.com>
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 as published
   8  * by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  18  */
  19
  20 #include <linux/scatterlist.h>
  21 #include <linux/dma-mapping.h>
  22 #include <linux/stacktrace.h>
  23 #include <linux/dma-debug.h>
  24 #include <linux/spinlock.h>
  25 #include <linux/debugfs.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/export.h>
  28 #include <linux/device.h>
  29 #include <linux/types.h>
  30 #include <linux/sched.h>
  31 #include <linux/ctype.h>
  32 #include <linux/list.h>
  33 #include <linux/slab.h>
  34
  35 #include <asm/sections.h>
  36
  37 #define HASH_SIZE       1024ULL
  38 #define HASH_FN_SHIFT   13
  39 #define HASH_FN_MASK    (HASH_SIZE - 1)
  40
  41 enum {
  42         dma_debug_single,
  43         dma_debug_page,
  44         dma_debug_sg,
  45         dma_debug_coherent,
  46 };
  47
  48 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
  49
  50 struct dma_debug_entry {
  51         struct list_head list;
  52         struct device    *dev;
  53         int              type;
  54         phys_addr_t      paddr;
  55         u64              dev_addr;
  56         u64              size;
  57         int              direction;
  58         int              sg_call_ents;
  59         int              sg_mapped_ents;
  60 #ifdef CONFIG_STACKTRACE
  61         struct           stack_trace stacktrace;
  62         unsigned long    st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
  63 #endif
  64 };
  65
  66 typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
  67
  68 struct hash_bucket {
  69         struct list_head list;
  70         spinlock_t lock;
  71 } ____cacheline_aligned_in_smp;
  72
  73 /* Hash list to save the allocated dma addresses */
  74 static struct hash_bucket dma_entry_hash[HASH_SIZE];
  75 /* List of pre-allocated dma_debug_entry's */
  76 static LIST_HEAD(free_entries);
  77 /* Lock for the list above */
  78 static DEFINE_SPINLOCK(free_entries_lock);
  79
  80 /* Global disable flag - will be set in case of an error */
  81 static bool global_disable __read_mostly;
  82
  83 /* Global error count */
  84 static u32 error_count;
  85
  86 /* Global error show enable*/
  87 static u32 show_all_errors __read_mostly;
  88 /* Number of errors to show */
  89 static u32 show_num_errors = 1;
  90
  91 static u32 num_free_entries;
  92 static u32 min_free_entries;
  93 static u32 nr_total_entries;
  94
  95 /* number of preallocated entries requested by kernel cmdline */
  96 static u32 req_entries;
  97
  98 /* debugfs dentry's for the stuff above */
  99 static struct dentry *dma_debug_dent        __read_mostly;
 100 static struct dentry *global_disable_dent   __read_mostly;
 101 static struct dentry *error_count_dent      __read_mostly;
 102 static struct dentry *show_all_errors_dent  __read_mostly;
 103 static struct dentry *show_num_errors_dent  __read_mostly;
 104 static struct dentry *num_free_entries_dent __read_mostly;
 105 static struct dentry *min_free_entries_dent __read_mostly;
 106 static struct dentry *filter_dent           __read_mostly;
 107
 108 /* per-driver filter related state */
 109
 110 #define NAME_MAX_LEN    64
 111
 112 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
 113 static struct device_driver *current_driver                    __read_mostly;
 114
 115 static DEFINE_RWLOCK(driver_name_lock);
 116
 117 static const char *type2name[4] = { "single", "page",
 118                                     "scather-gather", "coherent" };
 119
 120 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
 121                                    "DMA_FROM_DEVICE", "DMA_NONE" };
 122
 123 /* little merge helper - remove it after the merge window */
 124 #ifndef BUS_NOTIFY_UNBOUND_DRIVER
 125 #define BUS_NOTIFY_UNBOUND_DRIVER 0x0005
 126 #endif
 127
 128 /*
 129  * The access to some variables in this macro is racy. We can't use atomic_t
 130  * here because all these variables are exported to debugfs. Some of them even
 131  * writeable. This is also the reason why a lock won't help much. But anyway,
 132  * the races are no big deal. Here is why:
 133  *
 134  *   error_count: the addition is racy, but the worst thing that can happen is
 135  *                that we don't count some errors
 136  *   show_num_errors: the subtraction is racy. Also no big deal because in
 137  *                    worst case this will result in one warning more in the
 138  *                    system log than the user configured. This variable is
 139  *                    writeable via debugfs.
 140  */
 141 static inline void dump_entry_trace(struct dma_debug_entry *entry)
 142 {
 143 #ifdef CONFIG_STACKTRACE
 144         if (entry) {
 145                 pr_warning("Mapped at:\n");
 146                 print_stack_trace(&entry->stacktrace, 0);
 147         }
 148 #endif
 149 }
 150
 151 static bool driver_filter(struct device *dev)
 152 {
 153         struct device_driver *drv;
 154         unsigned long flags;
 155         bool ret;
 156
 157         /* driver filter off */
 158         if (likely(!current_driver_name[0]))
 159                 return true;
 160
 161         /* driver filter on and initialized */
 162         if (current_driver && dev && dev->driver == current_driver)
 163                 return true;
 164
 165         /* driver filter on, but we can't filter on a NULL device... */
 166         if (!dev)
 167                 return false;
 168
 169         if (current_driver || !current_driver_name[0])
 170                 return false;
 171
 172         /* driver filter on but not yet initialized */
 173         drv = dev->driver;
 174         if (!drv)
 175                 return false;
 176
 177         /* lock to protect against change of current_driver_name */
 178         read_lock_irqsave(&driver_name_lock, flags);
 179
 180         ret = false;
 181         if (drv->name &&
 182             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 183                 current_driver = drv;
 184                 ret = true;
 185         }
 186
 187         read_unlock_irqrestore(&driver_name_lock, flags);
 188
 189         return ret;
 190 }
 191
 192 #define err_printk(dev, entry, format, arg...) do {                     \
 193                 error_count += 1;                                       \
 194                 if (driver_filter(dev) &&                               \
 195                     (show_all_errors || show_num_errors > 0)) {         \
 196                         WARN(1, "%s %s: " format,                       \
 197                              dev ? dev_driver_string(dev) : "NULL",     \
 198                              dev ? dev_name(dev) : "NULL", ## arg);     \
 199                         dump_entry_trace(entry);                        \
 200                 }                                                       \
 201                 if (!show_all_errors && show_num_errors > 0)            \
 202                         show_num_errors -= 1;                           \
 203         } while (0);
 204
 205 /*
 206  * Hash related functions
 207  *
 208  * Every DMA-API request is saved into a struct dma_debug_entry. To
 209  * have quick access to these structs they are stored into a hash.
 210  */
 211 static int hash_fn(struct dma_debug_entry *entry)
 212 {
 213         /*
 214          * Hash function is based on the dma address.
 215          * We use bits 20-27 here as the index into the hash
 216          */
 217         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 218 }
 219
 220 /*
 221  * Request exclusive access to a hash bucket for a given dma_debug_entry.
 222  */
 223 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 224                                            unsigned long *flags)
 225 {
 226         int idx = hash_fn(entry);
 227         unsigned long __flags;
 228
 229         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 230         *flags = __flags;
 231         return &dma_entry_hash[idx];
 232 }
 233
 234 /*
 235  * Give up exclusive access to the hash bucket
 236  */
 237 static void put_hash_bucket(struct hash_bucket *bucket,
 238                             unsigned long *flags)
 239 {
 240         unsigned long __flags = *flags;
 241
 242         spin_unlock_irqrestore(&bucket->lock, __flags);
 243 }
 244
 245 static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
 246 {
 247         return ((a->dev_addr == b->dev_addr) &&
 248                 (a->dev == b->dev)) ? true : false;
 249 }
 250
 251 static bool containing_match(struct dma_debug_entry *a,
 252                              struct dma_debug_entry *b)
 253 {
 254         if (a->dev != b->dev)
 255                 return false;
 256
 257         if ((b->dev_addr <= a->dev_addr) &&
 258             ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
 259                 return true;
 260
 261         return false;
 262 }
 263
 264 /*
 265  * Search a given entry in the hash bucket list
 266  */
 267 static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
 268                                                   struct dma_debug_entry *ref,
 269                                                   match_fn match)
 270 {
 271         struct dma_debug_entry *entry, *ret = NULL;
 272         int matches = 0, match_lvl, last_lvl = 0;
 273
 274         list_for_each_entry(entry, &bucket->list, list) {
 275                 if (!match(ref, entry))
 276                         continue;
 277
 278                 /*
 279                  * Some drivers map the same physical address multiple
 280                  * times. Without a hardware IOMMU this results in the
 281                  * same device addresses being put into the dma-debug
 282                  * hash multiple times too. This can result in false
 283                  * positives being reported. Therefore we implement a
 284                  * best-fit algorithm here which returns the entry from
 285                  * the hash which fits best to the reference value
 286                  * instead of the first-fit.
 287                  */
 288                 matches += 1;
 289                 match_lvl = 0;
 290                 entry->size         == ref->size         ? ++match_lvl : 0;
 291                 entry->type         == ref->type         ? ++match_lvl : 0;
 292                 entry->direction    == ref->direction    ? ++match_lvl : 0;
 293                 entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
 294
 295                 if (match_lvl == 4) {
 296                         /* perfect-fit - return the result */
 297                         return entry;
 298                 } else if (match_lvl > last_lvl) {
 299                         /*
 300                          * We found an entry that fits better then the
 301                          * previous one
 302                          */
 303                         last_lvl = match_lvl;
 304                         ret      = entry;
 305                 }
 306         }
 307
 308         /*
 309          * If we have multiple matches but no perfect-fit, just return
 310          * NULL.
 311          */
 312         ret = (matches == 1) ? ret : NULL;
 313
 314         return ret;
 315 }
 316
 317 static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
 318                                                  struct dma_debug_entry *ref)
 319 {
 320         return __hash_bucket_find(bucket, ref, exact_match);
 321 }
 322
 323 static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 324                                                    struct dma_debug_entry *ref,
 325                                                    unsigned long *flags)
 326 {
 327
 328         unsigned int max_range = dma_get_max_seg_size(ref->dev);
 329         struct dma_debug_entry *entry, index = *ref;
 330         unsigned int range = 0;
 331
 332         while (range <= max_range) {
 333                 entry = __hash_bucket_find(*bucket, &index, containing_match);
 334
 335                 if (entry)
 336                         return entry;
 337
 338                 /*
 339                  * Nothing found, go back a hash bucket
 340                  */
 341                 put_hash_bucket(*bucket, flags);
 342                 range          += (1 << HASH_FN_SHIFT);
 343                 index.dev_addr -= (1 << HASH_FN_SHIFT);
 344                 *bucket = get_hash_bucket(&index, flags);
 345         }
 346
 347         return NULL;
 348 }
 349
 350 /*
 351  * Add an entry to a hash bucket
 352  */
 353 static void hash_bucket_add(struct hash_bucket *bucket,
 354                             struct dma_debug_entry *entry)
 355 {
 356         list_add_tail(&entry->list, &bucket->list);
 357 }
 358
 359 /*
 360  * Remove entry from a hash bucket list
 361  */
 362 static void hash_bucket_del(struct dma_debug_entry *entry)
 363 {
 364         list_del(&entry->list);
 365 }
 366
 367 /*
 368  * Dump mapping entries for debugging purposes
 369  */
 370 void debug_dma_dump_mappings(struct device *dev)
 371 {
 372         int idx;
 373
 374         for (idx = 0; idx < HASH_SIZE; idx++) {
 375                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 376                 struct dma_debug_entry *entry;
 377                 unsigned long flags;
 378
 379                 spin_lock_irqsave(&bucket->lock, flags);
 380
 381                 list_for_each_entry(entry, &bucket->list, list) {
 382                         if (!dev || dev == entry->dev) {
 383                                 dev_info(entry->dev,
 384                                          "%s idx %d P=%Lx D=%Lx L=%Lx %s\n",
 385                                          type2name[entry->type], idx,
 386                                          (unsigned long long)entry->paddr,
 387                                          entry->dev_addr, entry->size,
 388                                          dir2name[entry->direction]);
 389                         }
 390                 }
 391
 392                 spin_unlock_irqrestore(&bucket->lock, flags);
 393         }
 394 }
 395 EXPORT_SYMBOL(debug_dma_dump_mappings);
 396
 397 /*
 398  * Wrapper function for adding an entry to the hash.
 399  * This function takes care of locking itself.
 400  */
 401 static void add_dma_entry(struct dma_debug_entry *entry)
 402 {
 403         struct hash_bucket *bucket;
 404         unsigned long flags;
 405
 406         bucket = get_hash_bucket(entry, &flags);
 407         hash_bucket_add(bucket, entry);
 408         put_hash_bucket(bucket, &flags);
 409 }
 410
 411 static struct dma_debug_entry *__dma_entry_alloc(void)
 412 {
 413         struct dma_debug_entry *entry;
 414
 415         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 416         list_del(&entry->list);
 417         memset(entry, 0, sizeof(*entry));
 418
 419         num_free_entries -= 1;
 420         if (num_free_entries < min_free_entries)
 421                 min_free_entries = num_free_entries;
 422
 423         return entry;
 424 }
 425
 426 /* struct dma_entry allocator
 427  *
 428  * The next two functions implement the allocator for
 429  * struct dma_debug_entries.
 430  */
 431 static struct dma_debug_entry *dma_entry_alloc(void)
 432 {
 433         struct dma_debug_entry *entry;
 434         unsigned long flags;
 435
 436         spin_lock_irqsave(&free_entries_lock, flags);
 437
 438         if (list_empty(&free_entries)) {
 439                 pr_err("DMA-API: debugging out of memory - disabling\n");
 440                 global_disable = true;
 441                 spin_unlock_irqrestore(&free_entries_lock, flags);
 442                 return NULL;
 443         }
 444
 445         entry = __dma_entry_alloc();
 446
 447         spin_unlock_irqrestore(&free_entries_lock, flags);
 448
 449 #ifdef CONFIG_STACKTRACE
 450         entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
 451         entry->stacktrace.entries = entry->st_entries;
 452         entry->stacktrace.skip = 2;
 453         save_stack_trace(&entry->stacktrace);
 454 #endif
 455
 456         return entry;
 457 }
 458
 459 static void dma_entry_free(struct dma_debug_entry *entry)
 460 {
 461         unsigned long flags;
 462
 463         /*
 464          * add to beginning of the list - this way the entries are
 465          * more likely cache hot when they are reallocated.
 466          */
 467         spin_lock_irqsave(&free_entries_lock, flags);
 468         list_add(&entry->list, &free_entries);
 469         num_free_entries += 1;
 470         spin_unlock_irqrestore(&free_entries_lock, flags);
 471 }
 472
 473 int dma_debug_resize_entries(u32 num_entries)
 474 {
 475         int i, delta, ret = 0;
 476         unsigned long flags;
 477         struct dma_debug_entry *entry;
 478         LIST_HEAD(tmp);
 479
 480         spin_lock_irqsave(&free_entries_lock, flags);
 481
 482         if (nr_total_entries < num_entries) {
 483                 delta = num_entries - nr_total_entries;
 484
 485                 spin_unlock_irqrestore(&free_entries_lock, flags);
 486
 487                 for (i = 0; i < delta; i++) {
 488                         entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 489                         if (!entry)
 490                                 break;
 491
 492                         list_add_tail(&entry->list, &tmp);
 493                 }
 494
 495                 spin_lock_irqsave(&free_entries_lock, flags);
 496
 497                 list_splice(&tmp, &free_entries);
 498                 nr_total_entries += i;
 499                 num_free_entries += i;
 500         } else {
 501                 delta = nr_total_entries - num_entries;
 502
 503                 for (i = 0; i < delta && !list_empty(&free_entries); i++) {
 504                         entry = __dma_entry_alloc();
 505                         kfree(entry);
 506                 }
 507
 508                 nr_total_entries -= i;
 509         }
 510
 511         if (nr_total_entries != num_entries)
 512                 ret = 1;
 513
 514         spin_unlock_irqrestore(&free_entries_lock, flags);
 515
 516         return ret;
 517 }
 518 EXPORT_SYMBOL(dma_debug_resize_entries);
 519
 520 /*
 521  * DMA-API debugging init code
 522  *
 523  * The init code does two things:
 524  *   1. Initialize core data structures
 525  *   2. Preallocate a given number of dma_debug_entry structs
 526  */
 527
 528 static int prealloc_memory(u32 num_entries)
 529 {
 530         struct dma_debug_entry *entry, *next_entry;
 531         int i;
 532
 533         for (i = 0; i < num_entries; ++i) {
 534                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 535                 if (!entry)
 536                         goto out_err;
 537
 538                 list_add_tail(&entry->list, &free_entries);
 539         }
 540
 541         num_free_entries = num_entries;
 542         min_free_entries = num_entries;
 543
 544         pr_info("DMA-API: preallocated %d debug entries\n", num_entries);
 545
 546         return 0;
 547
 548 out_err:
 549
 550         list_for_each_entry_safe(entry, next_entry, &free_entries, list) {
 551                 list_del(&entry->list);
 552                 kfree(entry);
 553         }
 554
 555         return -ENOMEM;
 556 }
 557
 558 static ssize_t filter_read(struct file *file, char __user *user_buf,
 559                            size_t count, loff_t *ppos)
 560 {
 561         char buf[NAME_MAX_LEN + 1];
 562         unsigned long flags;
 563         int len;
 564
 565         if (!current_driver_name[0])
 566                 return 0;
 567
 568         /*
 569          * We can't copy to userspace directly because current_driver_name can
 570          * only be read under the driver_name_lock with irqs disabled. So
 571          * create a temporary copy first.
 572          */
 573         read_lock_irqsave(&driver_name_lock, flags);
 574         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 575         read_unlock_irqrestore(&driver_name_lock, flags);
 576
 577         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 578 }
 579
 580 static ssize_t filter_write(struct file *file, const char __user *userbuf,
 581                             size_t count, loff_t *ppos)
 582 {
 583         char buf[NAME_MAX_LEN];
 584         unsigned long flags;
 585         size_t len;
 586         int i;
 587
 588         /*
 589          * We can't copy from userspace directly. Access to
 590          * current_driver_name is protected with a write_lock with irqs
 591          * disabled. Since copy_from_user can fault and may sleep we
 592          * need to copy to temporary buffer first
 593          */
 594         len = min(count, (size_t)(NAME_MAX_LEN - 1));
 595         if (copy_from_user(buf, userbuf, len))
 596                 return -EFAULT;
 597
 598         buf[len] = 0;
 599
 600         write_lock_irqsave(&driver_name_lock, flags);
 601
 602         /*
 603          * Now handle the string we got from userspace very carefully.
 604          * The rules are:
 605          *         - only use the first token we got
 606          *         - token delimiter is everything looking like a space
 607          *           character (' ', '\n', '\t' ...)
 608          *
 609          */
 610         if (!isalnum(buf[0])) {
 611                 /*
 612                  * If the first character userspace gave us is not
 613                  * alphanumerical then assume the filter should be
 614                  * switched off.
 615                  */
 616                 if (current_driver_name[0])
 617                         pr_info("DMA-API: switching off dma-debug driver filter\n");
 618                 current_driver_name[0] = 0;
 619                 current_driver = NULL;
 620                 goto out_unlock;
 621         }
 622
 623         /*
 624          * Now parse out the first token and use it as the name for the
 625          * driver to filter for.
 626          */
 627         for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 628                 current_driver_name[i] = buf[i];
 629                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 630                         break;
 631         }
 632         current_driver_name[i] = 0;
 633         current_driver = NULL;
 634
 635         pr_info("DMA-API: enable driver filter for driver [%s]\n",
 636                 current_driver_name);
 637
 638 out_unlock:
 639         write_unlock_irqrestore(&driver_name_lock, flags);
 640
 641         return count;
 642 }
 643
 644 static const struct file_operations filter_fops = {
 645         .read  = filter_read,
 646         .write = filter_write,
 647         .llseek = default_llseek,
 648 };
 649
 650 static int dma_debug_fs_init(void)
 651 {
 652         dma_debug_dent = debugfs_create_dir("dma-api", NULL);
 653         if (!dma_debug_dent) {
 654                 pr_err("DMA-API: can not create debugfs directory\n");
 655                 return -ENOMEM;
 656         }
 657
 658         global_disable_dent = debugfs_create_bool("disabled", 0444,
 659                         dma_debug_dent,
 660                         (u32 *)&global_disable);
 661         if (!global_disable_dent)
 662                 goto out_err;
 663
 664         error_count_dent = debugfs_create_u32("error_count", 0444,
 665                         dma_debug_dent, &error_count);
 666         if (!error_count_dent)
 667                 goto out_err;
 668
 669         show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
 670                         dma_debug_dent,
 671                         &show_all_errors);
 672         if (!show_all_errors_dent)
 673                 goto out_err;
 674
 675         show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
 676                         dma_debug_dent,
 677                         &show_num_errors);
 678         if (!show_num_errors_dent)
 679                 goto out_err;
 680
 681         num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
 682                         dma_debug_dent,
 683                         &num_free_entries);
 684         if (!num_free_entries_dent)
 685                 goto out_err;
 686
 687         min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
 688                         dma_debug_dent,
 689                         &min_free_entries);
 690         if (!min_free_entries_dent)
 691                 goto out_err;
 692
 693         filter_dent = debugfs_create_file("driver_filter", 0644,
 694                                           dma_debug_dent, NULL, &filter_fops);
 695         if (!filter_dent)
 696                 goto out_err;
 697
 698         return 0;
 699
 700 out_err:
 701         debugfs_remove_recursive(dma_debug_dent);
 702
 703         return -ENOMEM;
 704 }
 705
 706 static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 707 {
 708         struct dma_debug_entry *entry;
 709         unsigned long flags;
 710         int count = 0, i;
 711
 712         local_irq_save(flags);
 713
 714         for (i = 0; i < HASH_SIZE; ++i) {
 715                 spin_lock(&dma_entry_hash[i].lock);
 716                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 717                         if (entry->dev == dev) {
 718                                 count += 1;
 719                                 *out_entry = entry;
 720                         }
 721                 }
 722                 spin_unlock(&dma_entry_hash[i].lock);
 723         }
 724
 725         local_irq_restore(flags);
 726
 727         return count;
 728 }
 729
 730 static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 731 {
 732         struct device *dev = data;
 733         struct dma_debug_entry *uninitialized_var(entry);
 734         int count;
 735
 736         if (global_disable)
 737                 return 0;
 738
 739         switch (action) {
 740         case BUS_NOTIFY_UNBOUND_DRIVER:
 741                 count = device_dma_allocations(dev, &entry);
 742                 if (count == 0)
 743                         break;
 744                 err_printk(dev, entry, "DMA-API: device driver has pending "
 745                                 "DMA allocations while released from device "
 746                                 "[count=%d]\n"
 747                                 "One of leaked entries details: "
 748                                 "[device address=0x%016llx] [size=%llu bytes] "
 749                                 "[mapped with %s] [mapped as %s]\n",
 750                         count, entry->dev_addr, entry->size,
 751                         dir2name[entry->direction], type2name[entry->type]);
 752                 break;
 753         default:
 754                 break;
 755         }
 756
 757         return 0;
 758 }
 759
 760 void dma_debug_add_bus(struct bus_type *bus)
 761 {
 762         struct notifier_block *nb;
 763
 764         if (global_disable)
 765                 return;
 766
 767         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 768         if (nb == NULL) {
 769                 pr_err("dma_debug_add_bus: out of memory\n");
 770                 return;
 771         }
 772
 773         nb->notifier_call = dma_debug_device_change;
 774
 775         bus_register_notifier(bus, nb);
 776 }
 777
 778 /*
 779  * Let the architectures decide how many entries should be preallocated.
 780  */
 781 void dma_debug_init(u32 num_entries)
 782 {
 783         int i;
 784
 785         if (global_disable)
 786                 return;
 787
 788         for (i = 0; i < HASH_SIZE; ++i) {
 789                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
 790                 spin_lock_init(&dma_entry_hash[i].lock);
 791         }
 792
 793         if (dma_debug_fs_init() != 0) {
 794                 pr_err("DMA-API: error creating debugfs entries - disabling\n");
 795                 global_disable = true;
 796
 797                 return;
 798         }
 799
 800         if (req_entries)
 801                 num_entries = req_entries;
 802
 803         if (prealloc_memory(num_entries) != 0) {
 804                 pr_err("DMA-API: debugging out of memory error - disabled\n");
 805                 global_disable = true;
 806
 807                 return;
 808         }
 809
 810         nr_total_entries = num_free_entries;
 811
 812         pr_info("DMA-API: debugging enabled by kernel config\n");
 813 }
 814
 815 static __init int dma_debug_cmdline(char *str)
 816 {
 817         if (!str)
 818                 return -EINVAL;
 819
 820         if (strncmp(str, "off", 3) == 0) {
 821                 pr_info("DMA-API: debugging disabled on kernel command line\n");
 822                 global_disable = true;
 823         }
 824
 825         return 0;
 826 }
 827
 828 static __init int dma_debug_entries_cmdline(char *str)
 829 {
 830         int res;
 831
 832         if (!str)
 833                 return -EINVAL;
 834
 835         res = get_option(&str, &req_entries);
 836
 837         if (!res)
 838                 req_entries = 0;
 839
 840         return 0;
 841 }
 842
 843 __setup("dma_debug=", dma_debug_cmdline);
 844 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
 845
 846 static void check_unmap(struct dma_debug_entry *ref)
 847 {
 848         struct dma_debug_entry *entry;
 849         struct hash_bucket *bucket;
 850         unsigned long flags;
 851
 852         if (dma_mapping_error(ref->dev, ref->dev_addr)) {
 853                 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
 854                            "to free an invalid DMA memory address\n");
 855                 return;
 856         }
 857
 858         bucket = get_hash_bucket(ref, &flags);
 859         entry = bucket_find_exact(bucket, ref);
 860
 861         if (!entry) {
 862                 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
 863                            "to free DMA memory it has not allocated "
 864                            "[device address=0x%016llx] [size=%llu bytes]\n",
 865                            ref->dev_addr, ref->size);
 866                 goto out;
 867         }
 868
 869         if (ref->size != entry->size) {
 870                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 871                            "DMA memory with different size "
 872                            "[device address=0x%016llx] [map size=%llu bytes] "
 873                            "[unmap size=%llu bytes]\n",
 874                            ref->dev_addr, entry->size, ref->size);
 875         }
 876
 877         if (ref->type != entry->type) {
 878                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 879                            "DMA memory with wrong function "
 880                            "[device address=0x%016llx] [size=%llu bytes] "
 881                            "[mapped as %s] [unmapped as %s]\n",
 882                            ref->dev_addr, ref->size,
 883                            type2name[entry->type], type2name[ref->type]);
 884         } else if ((entry->type == dma_debug_coherent) &&
 885                    (ref->paddr != entry->paddr)) {
 886                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 887                            "DMA memory with different CPU address "
 888                            "[device address=0x%016llx] [size=%llu bytes] "
 889                            "[cpu alloc address=0x%016llx] "
 890                            "[cpu free address=0x%016llx]",
 891                            ref->dev_addr, ref->size,
 892                            (unsigned long long)entry->paddr,
 893                            (unsigned long long)ref->paddr);
 894         }
 895
 896         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
 897             ref->sg_call_ents != entry->sg_call_ents) {
 898                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 899                            "DMA sg list with different entry count "
 900                            "[map count=%d] [unmap count=%d]\n",
 901                            entry->sg_call_ents, ref->sg_call_ents);
 902         }
 903
 904         /*
 905          * This may be no bug in reality - but most implementations of the
 906          * DMA API don't handle this properly, so check for it here
 907          */
 908         if (ref->direction != entry->direction) {
 909                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 910                            "DMA memory with different direction "
 911                            "[device address=0x%016llx] [size=%llu bytes] "
 912                            "[mapped with %s] [unmapped with %s]\n",
 913                            ref->dev_addr, ref->size,
 914                            dir2name[entry->direction],
 915                            dir2name[ref->direction]);
 916         }
 917
 918         hash_bucket_del(entry);
 919         dma_entry_free(entry);
 920
 921 out:
 922         put_hash_bucket(bucket, &flags);
 923 }
 924
 925 static void check_for_stack(struct device *dev, void *addr)
 926 {
 927         if (object_is_on_stack(addr))
 928                 err_printk(dev, NULL, "DMA-API: device driver maps memory from"
 929                                 "stack [addr=%p]\n", addr);
 930 }
 931
 932 static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
 933 {
 934         unsigned long a1 = (unsigned long)addr;
 935         unsigned long b1 = a1 + len;
 936         unsigned long a2 = (unsigned long)start;
 937         unsigned long b2 = (unsigned long)end;
 938
 939         return !(b1 <= a2 || a1 >= b2);
 940 }
 941
 942 static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
 943 {
 944         if (overlap(addr, len, _text, _etext) ||
 945             overlap(addr, len, __start_rodata, __end_rodata))
 946                 err_printk(dev, NULL, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
 947 }
 948
 949 static void check_sync(struct device *dev,
 950                        struct dma_debug_entry *ref,
 951                        bool to_cpu)
 952 {
 953         struct dma_debug_entry *entry;
 954         struct hash_bucket *bucket;
 955         unsigned long flags;
 956
 957         bucket = get_hash_bucket(ref, &flags);
 958
 959         entry = bucket_find_contain(&bucket, ref, &flags);
 960
 961         if (!entry) {
 962                 err_printk(dev, NULL, "DMA-API: device driver tries "
 963                                 "to sync DMA memory it has not allocated "
 964                                 "[device address=0x%016llx] [size=%llu bytes]\n",
 965                                 (unsigned long long)ref->dev_addr, ref->size);
 966                 goto out;
 967         }
 968
 969         if (ref->size > entry->size) {
 970                 err_printk(dev, entry, "DMA-API: device driver syncs"
 971                                 " DMA memory outside allocated range "
 972                                 "[device address=0x%016llx] "
 973                                 "[allocation size=%llu bytes] "
 974                                 "[sync offset+size=%llu]\n",
 975                                 entry->dev_addr, entry->size,
 976                                 ref->size);
 977         }
 978
 979         if (entry->direction == DMA_BIDIRECTIONAL)
 980                 goto out;
 981
 982         if (ref->direction != entry->direction) {
 983                 err_printk(dev, entry, "DMA-API: device driver syncs "
 984                                 "DMA memory with different direction "
 985                                 "[device address=0x%016llx] [size=%llu bytes] "
 986                                 "[mapped with %s] [synced with %s]\n",
 987                                 (unsigned long long)ref->dev_addr, entry->size,
 988                                 dir2name[entry->direction],
 989                                 dir2name[ref->direction]);
 990         }
 991
 992         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
 993                       !(ref->direction == DMA_TO_DEVICE))
 994                 err_printk(dev, entry, "DMA-API: device driver syncs "
 995                                 "device read-only DMA memory for cpu "
 996                                 "[device address=0x%016llx] [size=%llu bytes] "
 997                                 "[mapped with %s] [synced with %s]\n",
 998                                 (unsigned long long)ref->dev_addr, entry->size,
 999                                 dir2name[entry->direction],
1000                                 dir2name[ref->direction]);
1001
1002         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1003                        !(ref->direction == DMA_FROM_DEVICE))
1004                 err_printk(dev, entry, "DMA-API: device driver syncs "
1005                                 "device write-only DMA memory to device "
1006                                 "[device address=0x%016llx] [size=%llu bytes] "
1007                                 "[mapped with %s] [synced with %s]\n",
1008                                 (unsigned long long)ref->dev_addr, entry->size,
1009                                 dir2name[entry->direction],
1010                                 dir2name[ref->direction]);
1011
1012 out:
1013         put_hash_bucket(bucket, &flags);
1014 }
1015
1016 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1017                         size_t size, int direction, dma_addr_t dma_addr,
1018                         bool map_single)
1019 {
1020         struct dma_debug_entry *entry;
1021
1022         if (unlikely(global_disable))
1023                 return;
1024
1025         if (unlikely(dma_mapping_error(dev, dma_addr)))
1026                 return;
1027
1028         entry = dma_entry_alloc();
1029         if (!entry)
1030                 return;
1031
1032         entry->dev       = dev;
1033         entry->type      = dma_debug_page;
1034         entry->paddr     = page_to_phys(page) + offset;
1035         entry->dev_addr  = dma_addr;
1036         entry->size      = size;
1037         entry->direction = direction;
1038
1039         if (map_single)
1040                 entry->type = dma_debug_single;
1041
1042         if (!PageHighMem(page)) {
1043                 void *addr = page_address(page) + offset;
1044
1045                 check_for_stack(dev, addr);
1046                 check_for_illegal_area(dev, addr, size);
1047         }
1048
1049         add_dma_entry(entry);
1050 }
1051 EXPORT_SYMBOL(debug_dma_map_page);
1052
1053 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1054                           size_t size, int direction, bool map_single)
1055 {
1056         struct dma_debug_entry ref = {
1057                 .type           = dma_debug_page,
1058                 .dev            = dev,
1059                 .dev_addr       = addr,
1060                 .size           = size,
1061                 .direction      = direction,
1062         };
1063
1064         if (unlikely(global_disable))
1065                 return;
1066
1067         if (map_single)
1068                 ref.type = dma_debug_single;
1069
1070         check_unmap(&ref);
1071 }
1072 EXPORT_SYMBOL(debug_dma_unmap_page);
1073
1074 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1075                       int nents, int mapped_ents, int direction)
1076 {
1077         struct dma_debug_entry *entry;
1078         struct scatterlist *s;
1079         int i;
1080
1081         if (unlikely(global_disable))
1082                 return;
1083
1084         for_each_sg(sg, s, mapped_ents, i) {
1085                 entry = dma_entry_alloc();
1086                 if (!entry)
1087                         return;
1088
1089                 entry->type           = dma_debug_sg;
1090                 entry->dev            = dev;
1091                 entry->paddr          = sg_phys(s);
1092                 entry->size           = sg_dma_len(s);
1093                 entry->dev_addr       = sg_dma_address(s);
1094                 entry->direction      = direction;
1095                 entry->sg_call_ents   = nents;
1096                 entry->sg_mapped_ents = mapped_ents;
1097
1098                 if (!PageHighMem(sg_page(s))) {
1099                         check_for_stack(dev, sg_virt(s));
1100                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1101                 }
1102
1103                 add_dma_entry(entry);
1104         }
1105 }
1106 EXPORT_SYMBOL(debug_dma_map_sg);
1107
1108 static int get_nr_mapped_entries(struct device *dev,
1109                                  struct dma_debug_entry *ref)
1110 {
1111         struct dma_debug_entry *entry;
1112         struct hash_bucket *bucket;
1113         unsigned long flags;
1114         int mapped_ents;
1115
1116         bucket       = get_hash_bucket(ref, &flags);
1117         entry        = bucket_find_exact(bucket, ref);
1118         mapped_ents  = 0;
1119
1120         if (entry)
1121                 mapped_ents = entry->sg_mapped_ents;
1122         put_hash_bucket(bucket, &flags);
1123
1124         return mapped_ents;
1125 }
1126
1127 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1128                         int nelems, int dir)
1129 {
1130         struct scatterlist *s;
1131         int mapped_ents = 0, i;
1132
1133         if (unlikely(global_disable))
1134                 return;
1135
1136         for_each_sg(sglist, s, nelems, i) {
1137
1138                 struct dma_debug_entry ref = {
1139                         .type           = dma_debug_sg,
1140                         .dev            = dev,
1141                         .paddr          = sg_phys(s),
1142                         .dev_addr       = sg_dma_address(s),
1143                         .size           = sg_dma_len(s),
1144                         .direction      = dir,
1145                         .sg_call_ents   = nelems,
1146                 };
1147
1148                 if (mapped_ents && i >= mapped_ents)
1149                         break;
1150
1151                 if (!i)
1152                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1153
1154                 check_unmap(&ref);
1155         }
1156 }
1157 EXPORT_SYMBOL(debug_dma_unmap_sg);
1158
1159 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1160                               dma_addr_t dma_addr, void *virt)
1161 {
1162         struct dma_debug_entry *entry;
1163
1164         if (unlikely(global_disable))
1165                 return;
1166
1167         if (unlikely(virt == NULL))
1168                 return;
1169
1170         entry = dma_entry_alloc();
1171         if (!entry)
1172                 return;
1173
1174         entry->type      = dma_debug_coherent;
1175         entry->dev       = dev;
1176         entry->paddr     = virt_to_phys(virt);
1177         entry->size      = size;
1178         entry->dev_addr  = dma_addr;
1179         entry->direction = DMA_BIDIRECTIONAL;
1180
1181         add_dma_entry(entry);
1182 }
1183 EXPORT_SYMBOL(debug_dma_alloc_coherent);
1184
1185 void debug_dma_free_coherent(struct device *dev, size_t size,
1186                          void *virt, dma_addr_t addr)
1187 {
1188         struct dma_debug_entry ref = {
1189                 .type           = dma_debug_coherent,
1190                 .dev            = dev,
1191                 .paddr          = virt_to_phys(virt),
1192                 .dev_addr       = addr,
1193                 .size           = size,
1194                 .direction      = DMA_BIDIRECTIONAL,
1195         };
1196
1197         if (unlikely(global_disable))
1198                 return;
1199
1200         check_unmap(&ref);
1201 }
1202 EXPORT_SYMBOL(debug_dma_free_coherent);
1203
1204 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1205                                    size_t size, int direction)
1206 {
1207         struct dma_debug_entry ref;
1208
1209         if (unlikely(global_disable))
1210                 return;
1211
1212         ref.type         = dma_debug_single;
1213         ref.dev          = dev;
1214         ref.dev_addr     = dma_handle;
1215         ref.size         = size;
1216         ref.direction    = direction;
1217         ref.sg_call_ents = 0;
1218
1219         check_sync(dev, &ref, true);
1220 }
1221 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1222
1223 void debug_dma_sync_single_for_device(struct device *dev,
1224                                       dma_addr_t dma_handle, size_t size,
1225                                       int direction)
1226 {
1227         struct dma_debug_entry ref;
1228
1229         if (unlikely(global_disable))
1230                 return;
1231
1232         ref.type         = dma_debug_single;
1233         ref.dev          = dev;
1234         ref.dev_addr     = dma_handle;
1235         ref.size         = size;
1236         ref.direction    = direction;
1237         ref.sg_call_ents = 0;
1238
1239         check_sync(dev, &ref, false);
1240 }
1241 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1242
1243 void debug_dma_sync_single_range_for_cpu(struct device *dev,
1244                                          dma_addr_t dma_handle,
1245                                          unsigned long offset, size_t size,
1246                                          int direction)
1247 {
1248         struct dma_debug_entry ref;
1249
1250         if (unlikely(global_disable))
1251                 return;
1252
1253         ref.type         = dma_debug_single;
1254         ref.dev          = dev;
1255         ref.dev_addr     = dma_handle;
1256         ref.size         = offset + size;
1257         ref.direction    = direction;
1258         ref.sg_call_ents = 0;
1259
1260         check_sync(dev, &ref, true);
1261 }
1262 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
1263
1264 void debug_dma_sync_single_range_for_device(struct device *dev,
1265                                             dma_addr_t dma_handle,
1266                                             unsigned long offset,
1267                                             size_t size, int direction)
1268 {
1269         struct dma_debug_entry ref;
1270
1271         if (unlikely(global_disable))
1272                 return;
1273
1274         ref.type         = dma_debug_single;
1275         ref.dev          = dev;
1276         ref.dev_addr     = dma_handle;
1277         ref.size         = offset + size;
1278         ref.direction    = direction;
1279         ref.sg_call_ents = 0;
1280
1281         check_sync(dev, &ref, false);
1282 }
1283 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
1284
1285 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1286                                int nelems, int direction)
1287 {
1288         struct scatterlist *s;
1289         int mapped_ents = 0, i;
1290
1291         if (unlikely(global_disable))
1292                 return;
1293
1294         for_each_sg(sg, s, nelems, i) {
1295
1296                 struct dma_debug_entry ref = {
1297                         .type           = dma_debug_sg,
1298                         .dev            = dev,
1299                         .paddr          = sg_phys(s),
1300                         .dev_addr       = sg_dma_address(s),
1301                         .size           = sg_dma_len(s),
1302                         .direction      = direction,
1303                         .sg_call_ents   = nelems,
1304                 };
1305
1306                 if (!i)
1307                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1308
1309                 if (i >= mapped_ents)
1310                         break;
1311
1312                 check_sync(dev, &ref, true);
1313         }
1314 }
1315 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1316
1317 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1318                                   int nelems, int direction)
1319 {
1320         struct scatterlist *s;
1321         int mapped_ents = 0, i;
1322
1323         if (unlikely(global_disable))
1324                 return;
1325
1326         for_each_sg(sg, s, nelems, i) {
1327
1328                 struct dma_debug_entry ref = {
1329                         .type           = dma_debug_sg,
1330                         .dev            = dev,
1331                         .paddr          = sg_phys(s),
1332                         .dev_addr       = sg_dma_address(s),
1333                         .size           = sg_dma_len(s),
1334                         .direction      = direction,
1335                         .sg_call_ents   = nelems,
1336                 };
1337                 if (!i)
1338                         mapped_ents = get_nr_mapped_entries(dev, &ref);
1339
1340                 if (i >= mapped_ents)
1341                         break;
1342
1343                 check_sync(dev, &ref, false);
1344         }
1345 }
1346 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1347
1348 static int __init dma_debug_driver_setup(char *str)
1349 {
1350         int i;
1351
1352         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1353                 current_driver_name[i] = *str;
1354                 if (*str == 0)
1355                         break;
1356         }
1357
1358         if (current_driver_name[0])
1359                 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1360                         current_driver_name);
1361
1362
1363         return 1;
1364 }
1365 __setup("dma_debug_driver=", dma_debug_driver_setup);