SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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                 printk(KERN_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         /* driver filter off */
 151         if (likely(!current_driver_name[0]))
 152                 return true;
 153
 154         /* driver filter on and initialized */
 155         if (current_driver && dev->driver == current_driver)
 156                 return true;
 157
 158         /* driver filter on but not yet initialized */
 159         if (!current_driver && current_driver_name[0]) {
 160                 struct device_driver *drv = get_driver(dev->driver);
 161                 unsigned long flags;
 162                 bool ret = false;
 163
 164                 if (!drv)
 165                         return false;
 166
 167                 /* lock to protect against change of current_driver_name */
 168                 read_lock_irqsave(&driver_name_lock, flags);
 169
 170                 if (drv->name &&
 171                     strncmp(current_driver_name, drv->name,
 172                             NAME_MAX_LEN-1) == 0) {
 173                         current_driver = drv;
 174                         ret = true;
 175                 }
 176
 177                 read_unlock_irqrestore(&driver_name_lock, flags);
 178                 put_driver(drv);
 179
 180                 return ret;
 181         }
 182
 183         return false;
 184 }
 185
 186 #define err_printk(dev, entry, format, arg...) do {             \
 187                 error_count += 1;                               \
 188                 if (driver_filter(dev) &&                       \
 189                     (show_all_errors || show_num_errors > 0)) { \
 190                         WARN(1, "%s %s: " format,               \
 191                              dev_driver_string(dev),            \
 192                              dev_name(dev) , ## arg);           \
 193                         dump_entry_trace(entry);                \
 194                 }                                               \
 195                 if (!show_all_errors && show_num_errors > 0)    \
 196                         show_num_errors -= 1;                   \
 197         } while (0);
 198
 199 /*
 200  * Hash related functions
 201  *
 202  * Every DMA-API request is saved into a struct dma_debug_entry. To
 203  * have quick access to these structs they are stored into a hash.
 204  */
 205 static int hash_fn(struct dma_debug_entry *entry)
 206 {
 207         /*
 208          * Hash function is based on the dma address.
 209          * We use bits 20-27 here as the index into the hash
 210          */
 211         return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 212 }
 213
 214 /*
 215  * Request exclusive access to a hash bucket for a given dma_debug_entry.
 216  */
 217 static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 218                                            unsigned long *flags)
 219 {
 220         int idx = hash_fn(entry);
 221         unsigned long __flags;
 222
 223         spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 224         *flags = __flags;
 225         return &dma_entry_hash[idx];
 226 }
 227
 228 /*
 229  * Give up exclusive access to the hash bucket
 230  */
 231 static void put_hash_bucket(struct hash_bucket *bucket,
 232                             unsigned long *flags)
 233 {
 234         unsigned long __flags = *flags;
 235
 236         spin_unlock_irqrestore(&bucket->lock, __flags);
 237 }
 238
 239 /*
 240  * Search a given entry in the hash bucket list
 241  */
 242 static struct dma_debug_entry *hash_bucket_find(struct hash_bucket *bucket,
 243                                                 struct dma_debug_entry *ref)
 244 {
 245         struct dma_debug_entry *entry, *ret = NULL;
 246         int matches = 0, match_lvl, last_lvl = 0;
 247
 248         list_for_each_entry(entry, &bucket->list, list) {
 249                 if ((entry->dev_addr != ref->dev_addr) ||
 250                     (entry->dev != ref->dev))
 251                         continue;
 252
 253                 /*
 254                  * Some drivers map the same physical address multiple
 255                  * times. Without a hardware IOMMU this results in the
 256                  * same device addresses being put into the dma-debug
 257                  * hash multiple times too. This can result in false
 258                  * positives being reported. Therfore we implement a
 259                  * best-fit algorithm here which returns the entry from
 260                  * the hash which fits best to the reference value
 261                  * instead of the first-fit.
 262                  */
 263                 matches += 1;
 264                 match_lvl = 0;
 265                 entry->size      == ref->size      ? ++match_lvl : match_lvl;
 266                 entry->type      == ref->type      ? ++match_lvl : match_lvl;
 267                 entry->direction == ref->direction ? ++match_lvl : match_lvl;
 268
 269                 if (match_lvl == 3) {
 270                         /* perfect-fit - return the result */
 271                         return entry;
 272                 } else if (match_lvl > last_lvl) {
 273                         /*
 274                          * We found an entry that fits better then the
 275                          * previous one
 276                          */
 277                         last_lvl = match_lvl;
 278                         ret      = entry;
 279                 }
 280         }
 281
 282         /*
 283          * If we have multiple matches but no perfect-fit, just return
 284          * NULL.
 285          */
 286         ret = (matches == 1) ? ret : NULL;
 287
 288         return ret;
 289 }
 290
 291 /*
 292  * Add an entry to a hash bucket
 293  */
 294 static void hash_bucket_add(struct hash_bucket *bucket,
 295                             struct dma_debug_entry *entry)
 296 {
 297         list_add_tail(&entry->list, &bucket->list);
 298 }
 299
 300 /*
 301  * Remove entry from a hash bucket list
 302  */
 303 static void hash_bucket_del(struct dma_debug_entry *entry)
 304 {
 305         list_del(&entry->list);
 306 }
 307
 308 /*
 309  * Dump mapping entries for debugging purposes
 310  */
 311 void debug_dma_dump_mappings(struct device *dev)
 312 {
 313         int idx;
 314
 315         for (idx = 0; idx < HASH_SIZE; idx++) {
 316                 struct hash_bucket *bucket = &dma_entry_hash[idx];
 317                 struct dma_debug_entry *entry;
 318                 unsigned long flags;
 319
 320                 spin_lock_irqsave(&bucket->lock, flags);
 321
 322                 list_for_each_entry(entry, &bucket->list, list) {
 323                         if (!dev || dev == entry->dev) {
 324                                 dev_info(entry->dev,
 325                                          "%s idx %d P=%Lx D=%Lx L=%Lx %s\n",
 326                                          type2name[entry->type], idx,
 327                                          (unsigned long long)entry->paddr,
 328                                          entry->dev_addr, entry->size,
 329                                          dir2name[entry->direction]);
 330                         }
 331                 }
 332
 333                 spin_unlock_irqrestore(&bucket->lock, flags);
 334         }
 335 }
 336 EXPORT_SYMBOL(debug_dma_dump_mappings);
 337
 338 /*
 339  * Wrapper function for adding an entry to the hash.
 340  * This function takes care of locking itself.
 341  */
 342 static void add_dma_entry(struct dma_debug_entry *entry)
 343 {
 344         struct hash_bucket *bucket;
 345         unsigned long flags;
 346
 347         bucket = get_hash_bucket(entry, &flags);
 348         hash_bucket_add(bucket, entry);
 349         put_hash_bucket(bucket, &flags);
 350 }
 351
 352 static struct dma_debug_entry *__dma_entry_alloc(void)
 353 {
 354         struct dma_debug_entry *entry;
 355
 356         entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 357         list_del(&entry->list);
 358         memset(entry, 0, sizeof(*entry));
 359
 360         num_free_entries -= 1;
 361         if (num_free_entries < min_free_entries)
 362                 min_free_entries = num_free_entries;
 363
 364         return entry;
 365 }
 366
 367 /* struct dma_entry allocator
 368  *
 369  * The next two functions implement the allocator for
 370  * struct dma_debug_entries.
 371  */
 372 static struct dma_debug_entry *dma_entry_alloc(void)
 373 {
 374         struct dma_debug_entry *entry = NULL;
 375         unsigned long flags;
 376
 377         spin_lock_irqsave(&free_entries_lock, flags);
 378
 379         if (list_empty(&free_entries)) {
 380                 printk(KERN_ERR "DMA-API: debugging out of memory "
 381                                 "- disabling\n");
 382                 global_disable = true;
 383                 goto out;
 384         }
 385
 386         entry = __dma_entry_alloc();
 387
 388 #ifdef CONFIG_STACKTRACE
 389         entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
 390         entry->stacktrace.entries = entry->st_entries;
 391         entry->stacktrace.skip = 2;
 392         save_stack_trace(&entry->stacktrace);
 393 #endif
 394
 395 out:
 396         spin_unlock_irqrestore(&free_entries_lock, flags);
 397
 398         return entry;
 399 }
 400
 401 static void dma_entry_free(struct dma_debug_entry *entry)
 402 {
 403         unsigned long flags;
 404
 405         /*
 406          * add to beginning of the list - this way the entries are
 407          * more likely cache hot when they are reallocated.
 408          */
 409         spin_lock_irqsave(&free_entries_lock, flags);
 410         list_add(&entry->list, &free_entries);
 411         num_free_entries += 1;
 412         spin_unlock_irqrestore(&free_entries_lock, flags);
 413 }
 414
 415 int dma_debug_resize_entries(u32 num_entries)
 416 {
 417         int i, delta, ret = 0;
 418         unsigned long flags;
 419         struct dma_debug_entry *entry;
 420         LIST_HEAD(tmp);
 421
 422         spin_lock_irqsave(&free_entries_lock, flags);
 423
 424         if (nr_total_entries < num_entries) {
 425                 delta = num_entries - nr_total_entries;
 426
 427                 spin_unlock_irqrestore(&free_entries_lock, flags);
 428
 429                 for (i = 0; i < delta; i++) {
 430                         entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 431                         if (!entry)
 432                                 break;
 433
 434                         list_add_tail(&entry->list, &tmp);
 435                 }
 436
 437                 spin_lock_irqsave(&free_entries_lock, flags);
 438
 439                 list_splice(&tmp, &free_entries);
 440                 nr_total_entries += i;
 441                 num_free_entries += i;
 442         } else {
 443                 delta = nr_total_entries - num_entries;
 444
 445                 for (i = 0; i < delta && !list_empty(&free_entries); i++) {
 446                         entry = __dma_entry_alloc();
 447                         kfree(entry);
 448                 }
 449
 450                 nr_total_entries -= i;
 451         }
 452
 453         if (nr_total_entries != num_entries)
 454                 ret = 1;
 455
 456         spin_unlock_irqrestore(&free_entries_lock, flags);
 457
 458         return ret;
 459 }
 460 EXPORT_SYMBOL(dma_debug_resize_entries);
 461
 462 /*
 463  * DMA-API debugging init code
 464  *
 465  * The init code does two things:
 466  *   1. Initialize core data structures
 467  *   2. Preallocate a given number of dma_debug_entry structs
 468  */
 469
 470 static int prealloc_memory(u32 num_entries)
 471 {
 472         struct dma_debug_entry *entry, *next_entry;
 473         int i;
 474
 475         for (i = 0; i < num_entries; ++i) {
 476                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 477                 if (!entry)
 478                         goto out_err;
 479
 480                 list_add_tail(&entry->list, &free_entries);
 481         }
 482
 483         num_free_entries = num_entries;
 484         min_free_entries = num_entries;
 485
 486         printk(KERN_INFO "DMA-API: preallocated %d debug entries\n",
 487                         num_entries);
 488
 489         return 0;
 490
 491 out_err:
 492
 493         list_for_each_entry_safe(entry, next_entry, &free_entries, list) {
 494                 list_del(&entry->list);
 495                 kfree(entry);
 496         }
 497
 498         return -ENOMEM;
 499 }
 500
 501 static ssize_t filter_read(struct file *file, char __user *user_buf,
 502                            size_t count, loff_t *ppos)
 503 {
 504         unsigned long flags;
 505         char buf[NAME_MAX_LEN + 1];
 506         int len;
 507
 508         if (!current_driver_name[0])
 509                 return 0;
 510
 511         /*
 512          * We can't copy to userspace directly because current_driver_name can
 513          * only be read under the driver_name_lock with irqs disabled. So
 514          * create a temporary copy first.
 515          */
 516         read_lock_irqsave(&driver_name_lock, flags);
 517         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 518         read_unlock_irqrestore(&driver_name_lock, flags);
 519
 520         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 521 }
 522
 523 static ssize_t filter_write(struct file *file, const char __user *userbuf,
 524                             size_t count, loff_t *ppos)
 525 {
 526         unsigned long flags;
 527         char buf[NAME_MAX_LEN];
 528         size_t len = NAME_MAX_LEN - 1;
 529         int i;
 530
 531         /*
 532          * We can't copy from userspace directly. Access to
 533          * current_driver_name is protected with a write_lock with irqs
 534          * disabled. Since copy_from_user can fault and may sleep we
 535          * need to copy to temporary buffer first
 536          */
 537         len = min(count, len);
 538         if (copy_from_user(buf, userbuf, len))
 539                 return -EFAULT;
 540
 541         buf[len] = 0;
 542
 543         write_lock_irqsave(&driver_name_lock, flags);
 544
 545         /* Now handle the string we got from userspace very carefully.
 546          * The rules are:
 547          *         - only use the first token we got
 548          *         - token delimiter is everything looking like a space
 549          *           character (' ', '\n', '\t' ...)
 550          *
 551          */
 552         if (!isalnum(buf[0])) {
 553                 /*
 554                    If the first character userspace gave us is not
 555                  * alphanumerical then assume the filter should be
 556                  * switched off.
 557                  */
 558                 if (current_driver_name[0])
 559                         printk(KERN_INFO "DMA-API: switching off dma-debug "
 560                                          "driver filter\n");
 561                 current_driver_name[0] = 0;
 562                 current_driver = NULL;
 563                 goto out_unlock;
 564         }
 565
 566         /*
 567          * Now parse out the first token and use it as the name for the
 568          * driver to filter for.
 569          */
 570         for (i = 0; i < NAME_MAX_LEN; ++i) {
 571                 current_driver_name[i] = buf[i];
 572                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 573                         break;
 574         }
 575         current_driver_name[i] = 0;
 576         current_driver = NULL;
 577
 578         printk(KERN_INFO "DMA-API: enable driver filter for driver [%s]\n",
 579                current_driver_name);
 580
 581 out_unlock:
 582         write_unlock_irqrestore(&driver_name_lock, flags);
 583
 584         return count;
 585 }
 586
 587 const struct file_operations filter_fops = {
 588         .read  = filter_read,
 589         .write = filter_write,
 590 };
 591
 592 static int dma_debug_fs_init(void)
 593 {
 594         dma_debug_dent = debugfs_create_dir("dma-api", NULL);
 595         if (!dma_debug_dent) {
 596                 printk(KERN_ERR "DMA-API: can not create debugfs directory\n");
 597                 return -ENOMEM;
 598         }
 599
 600         global_disable_dent = debugfs_create_bool("disabled", 0444,
 601                         dma_debug_dent,
 602                         (u32 *)&global_disable);
 603         if (!global_disable_dent)
 604                 goto out_err;
 605
 606         error_count_dent = debugfs_create_u32("error_count", 0444,
 607                         dma_debug_dent, &error_count);
 608         if (!error_count_dent)
 609                 goto out_err;
 610
 611         show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
 612                         dma_debug_dent,
 613                         &show_all_errors);
 614         if (!show_all_errors_dent)
 615                 goto out_err;
 616
 617         show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
 618                         dma_debug_dent,
 619                         &show_num_errors);
 620         if (!show_num_errors_dent)
 621                 goto out_err;
 622
 623         num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
 624                         dma_debug_dent,
 625                         &num_free_entries);
 626         if (!num_free_entries_dent)
 627                 goto out_err;
 628
 629         min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
 630                         dma_debug_dent,
 631                         &min_free_entries);
 632         if (!min_free_entries_dent)
 633                 goto out_err;
 634
 635         filter_dent = debugfs_create_file("driver_filter", 0644,
 636                                           dma_debug_dent, NULL, &filter_fops);
 637         if (!filter_dent)
 638                 goto out_err;
 639
 640         return 0;
 641
 642 out_err:
 643         debugfs_remove_recursive(dma_debug_dent);
 644
 645         return -ENOMEM;
 646 }
 647
 648 static int device_dma_allocations(struct device *dev)
 649 {
 650         struct dma_debug_entry *entry;
 651         unsigned long flags;
 652         int count = 0, i;
 653
 654         for (i = 0; i < HASH_SIZE; ++i) {
 655                 spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
 656                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 657                         if (entry->dev == dev)
 658                                 count += 1;
 659                 }
 660                 spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
 661         }
 662
 663         return count;
 664 }
 665
 666 static int dma_debug_device_change(struct notifier_block *nb,
 667                                     unsigned long action, void *data)
 668 {
 669         struct device *dev = data;
 670         int count;
 671
 672
 673         switch (action) {
 674         case BUS_NOTIFY_UNBOUND_DRIVER:
 675                 count = device_dma_allocations(dev);
 676                 if (count == 0)
 677                         break;
 678                 err_printk(dev, NULL, "DMA-API: device driver has pending "
 679                                 "DMA allocations while released from device "
 680                                 "[count=%d]\n", count);
 681                 break;
 682         default:
 683                 break;
 684         }
 685
 686         return 0;
 687 }
 688
 689 void dma_debug_add_bus(struct bus_type *bus)
 690 {
 691         struct notifier_block *nb;
 692
 693         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 694         if (nb == NULL) {
 695                 printk(KERN_ERR "dma_debug_add_bus: out of memory\n");
 696                 return;
 697         }
 698
 699         nb->notifier_call = dma_debug_device_change;
 700
 701         bus_register_notifier(bus, nb);
 702 }
 703
 704 /*
 705  * Let the architectures decide how many entries should be preallocated.
 706  */
 707 void dma_debug_init(u32 num_entries)
 708 {
 709         int i;
 710
 711         if (global_disable)
 712                 return;
 713
 714         for (i = 0; i < HASH_SIZE; ++i) {
 715                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
 716                 dma_entry_hash[i].lock = SPIN_LOCK_UNLOCKED;
 717         }
 718
 719         if (dma_debug_fs_init() != 0) {
 720                 printk(KERN_ERR "DMA-API: error creating debugfs entries "
 721                                 "- disabling\n");
 722                 global_disable = true;
 723
 724                 return;
 725         }
 726
 727         if (req_entries)
 728                 num_entries = req_entries;
 729
 730         if (prealloc_memory(num_entries) != 0) {
 731                 printk(KERN_ERR "DMA-API: debugging out of memory error "
 732                                 "- disabled\n");
 733                 global_disable = true;
 734
 735                 return;
 736         }
 737
 738         nr_total_entries = num_free_entries;
 739
 740         printk(KERN_INFO "DMA-API: debugging enabled by kernel config\n");
 741 }
 742
 743 static __init int dma_debug_cmdline(char *str)
 744 {
 745         if (!str)
 746                 return -EINVAL;
 747
 748         if (strncmp(str, "off", 3) == 0) {
 749                 printk(KERN_INFO "DMA-API: debugging disabled on kernel "
 750                                  "command line\n");
 751                 global_disable = true;
 752         }
 753
 754         return 0;
 755 }
 756
 757 static __init int dma_debug_entries_cmdline(char *str)
 758 {
 759         int res;
 760
 761         if (!str)
 762                 return -EINVAL;
 763
 764         res = get_option(&str, &req_entries);
 765
 766         if (!res)
 767                 req_entries = 0;
 768
 769         return 0;
 770 }
 771
 772 __setup("dma_debug=", dma_debug_cmdline);
 773 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
 774
 775 static void check_unmap(struct dma_debug_entry *ref)
 776 {
 777         struct dma_debug_entry *entry;
 778         struct hash_bucket *bucket;
 779         unsigned long flags;
 780
 781         if (dma_mapping_error(ref->dev, ref->dev_addr)) {
 782                 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
 783                            "to free an invalid DMA memory address\n");
 784                 return;
 785         }
 786
 787         bucket = get_hash_bucket(ref, &flags);
 788         entry = hash_bucket_find(bucket, ref);
 789
 790         if (!entry) {
 791                 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
 792                            "to free DMA memory it has not allocated "
 793                            "[device address=0x%016llx] [size=%llu bytes]\n",
 794                            ref->dev_addr, ref->size);
 795                 goto out;
 796         }
 797
 798         if (ref->size != entry->size) {
 799                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 800                            "DMA memory with different size "
 801                            "[device address=0x%016llx] [map size=%llu bytes] "
 802                            "[unmap size=%llu bytes]\n",
 803                            ref->dev_addr, entry->size, ref->size);
 804         }
 805
 806         if (ref->type != entry->type) {
 807                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 808                            "DMA memory with wrong function "
 809                            "[device address=0x%016llx] [size=%llu bytes] "
 810                            "[mapped as %s] [unmapped as %s]\n",
 811                            ref->dev_addr, ref->size,
 812                            type2name[entry->type], type2name[ref->type]);
 813         } else if ((entry->type == dma_debug_coherent) &&
 814                    (ref->paddr != entry->paddr)) {
 815                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 816                            "DMA memory with different CPU address "
 817                            "[device address=0x%016llx] [size=%llu bytes] "
 818                            "[cpu alloc address=%p] [cpu free address=%p]",
 819                            ref->dev_addr, ref->size,
 820                            (void *)entry->paddr, (void *)ref->paddr);
 821         }
 822
 823         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
 824             ref->sg_call_ents != entry->sg_call_ents) {
 825                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 826                            "DMA sg list with different entry count "
 827                            "[map count=%d] [unmap count=%d]\n",
 828                            entry->sg_call_ents, ref->sg_call_ents);
 829         }
 830
 831         /*
 832          * This may be no bug in reality - but most implementations of the
 833          * DMA API don't handle this properly, so check for it here
 834          */
 835         if (ref->direction != entry->direction) {
 836                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 837                            "DMA memory with different direction "
 838                            "[device address=0x%016llx] [size=%llu bytes] "
 839                            "[mapped with %s] [unmapped with %s]\n",
 840                            ref->dev_addr, ref->size,
 841                            dir2name[entry->direction],
 842                            dir2name[ref->direction]);
 843         }
 844
 845         hash_bucket_del(entry);
 846         dma_entry_free(entry);
 847
 848 out:
 849         put_hash_bucket(bucket, &flags);
 850 }
 851
 852 static void check_for_stack(struct device *dev, void *addr)
 853 {
 854         if (object_is_on_stack(addr))
 855                 err_printk(dev, NULL, "DMA-API: device driver maps memory from"
 856                                 "stack [addr=%p]\n", addr);
 857 }
 858
 859 static inline bool overlap(void *addr, u64 size, void *start, void *end)
 860 {
 861         void *addr2 = (char *)addr + size;
 862
 863         return ((addr >= start && addr < end) ||
 864                 (addr2 >= start && addr2 < end) ||
 865                 ((addr < start) && (addr2 >= end)));
 866 }
 867
 868 static void check_for_illegal_area(struct device *dev, void *addr, u64 size)
 869 {
 870         if (overlap(addr, size, _text, _etext) ||
 871             overlap(addr, size, __start_rodata, __end_rodata))
 872                 err_printk(dev, NULL, "DMA-API: device driver maps "
 873                                 "memory from kernel text or rodata "
 874                                 "[addr=%p] [size=%llu]\n", addr, size);
 875 }
 876
 877 static void check_sync(struct device *dev, dma_addr_t addr,
 878                        u64 size, u64 offset, int direction, bool to_cpu)
 879 {
 880         struct dma_debug_entry ref = {
 881                 .dev            = dev,
 882                 .dev_addr       = addr,
 883                 .size           = size,
 884                 .direction      = direction,
 885         };
 886         struct dma_debug_entry *entry;
 887         struct hash_bucket *bucket;
 888         unsigned long flags;
 889
 890         bucket = get_hash_bucket(&ref, &flags);
 891
 892         entry = hash_bucket_find(bucket, &ref);
 893
 894         if (!entry) {
 895                 err_printk(dev, NULL, "DMA-API: device driver tries "
 896                                 "to sync DMA memory it has not allocated "
 897                                 "[device address=0x%016llx] [size=%llu bytes]\n",
 898                                 (unsigned long long)addr, size);
 899                 goto out;
 900         }
 901
 902         if ((offset + size) > entry->size) {
 903                 err_printk(dev, entry, "DMA-API: device driver syncs"
 904                                 " DMA memory outside allocated range "
 905                                 "[device address=0x%016llx] "
 906                                 "[allocation size=%llu bytes] [sync offset=%llu] "
 907                                 "[sync size=%llu]\n", entry->dev_addr, entry->size,
 908                                 offset, size);
 909         }
 910
 911         if (direction != entry->direction) {
 912                 err_printk(dev, entry, "DMA-API: device driver syncs "
 913                                 "DMA memory with different direction "
 914                                 "[device address=0x%016llx] [size=%llu bytes] "
 915                                 "[mapped with %s] [synced with %s]\n",
 916                                 (unsigned long long)addr, entry->size,
 917                                 dir2name[entry->direction],
 918                                 dir2name[direction]);
 919         }
 920
 921         if (entry->direction == DMA_BIDIRECTIONAL)
 922                 goto out;
 923
 924         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
 925                       !(direction == DMA_TO_DEVICE))
 926                 err_printk(dev, entry, "DMA-API: device driver syncs "
 927                                 "device read-only DMA memory for cpu "
 928                                 "[device address=0x%016llx] [size=%llu bytes] "
 929                                 "[mapped with %s] [synced with %s]\n",
 930                                 (unsigned long long)addr, entry->size,
 931                                 dir2name[entry->direction],
 932                                 dir2name[direction]);
 933
 934         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
 935                        !(direction == DMA_FROM_DEVICE))
 936                 err_printk(dev, entry, "DMA-API: device driver syncs "
 937                                 "device write-only DMA memory to device "
 938                                 "[device address=0x%016llx] [size=%llu bytes] "
 939                                 "[mapped with %s] [synced with %s]\n",
 940                                 (unsigned long long)addr, entry->size,
 941                                 dir2name[entry->direction],
 942                                 dir2name[direction]);
 943
 944 out:
 945         put_hash_bucket(bucket, &flags);
 946
 947 }
 948
 949 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
 950                         size_t size, int direction, dma_addr_t dma_addr,
 951                         bool map_single)
 952 {
 953         struct dma_debug_entry *entry;
 954
 955         if (unlikely(global_disable))
 956                 return;
 957
 958         if (unlikely(dma_mapping_error(dev, dma_addr)))
 959                 return;
 960
 961         entry = dma_entry_alloc();
 962         if (!entry)
 963                 return;
 964
 965         entry->dev       = dev;
 966         entry->type      = dma_debug_page;
 967         entry->paddr     = page_to_phys(page) + offset;
 968         entry->dev_addr  = dma_addr;
 969         entry->size      = size;
 970         entry->direction = direction;
 971
 972         if (map_single)
 973                 entry->type = dma_debug_single;
 974
 975         if (!PageHighMem(page)) {
 976                 void *addr = ((char *)page_address(page)) + offset;
 977                 check_for_stack(dev, addr);
 978                 check_for_illegal_area(dev, addr, size);
 979         }
 980
 981         add_dma_entry(entry);
 982 }
 983 EXPORT_SYMBOL(debug_dma_map_page);
 984
 985 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
 986                           size_t size, int direction, bool map_single)
 987 {
 988         struct dma_debug_entry ref = {
 989                 .type           = dma_debug_page,
 990                 .dev            = dev,
 991                 .dev_addr       = addr,
 992                 .size           = size,
 993                 .direction      = direction,
 994         };
 995
 996         if (unlikely(global_disable))
 997                 return;
 998
 999         if (map_single)
1000                 ref.type = dma_debug_single;
1001
1002         check_unmap(&ref);
1003 }
1004 EXPORT_SYMBOL(debug_dma_unmap_page);
1005
1006 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1007                       int nents, int mapped_ents, int direction)
1008 {
1009         struct dma_debug_entry *entry;
1010         struct scatterlist *s;
1011         int i;
1012
1013         if (unlikely(global_disable))
1014                 return;
1015
1016         for_each_sg(sg, s, mapped_ents, i) {
1017                 entry = dma_entry_alloc();
1018                 if (!entry)
1019                         return;
1020
1021                 entry->type           = dma_debug_sg;
1022                 entry->dev            = dev;
1023                 entry->paddr          = sg_phys(s);
1024                 entry->size           = sg_dma_len(s);
1025                 entry->dev_addr       = sg_dma_address(s);
1026                 entry->direction      = direction;
1027                 entry->sg_call_ents   = nents;
1028                 entry->sg_mapped_ents = mapped_ents;
1029
1030                 if (!PageHighMem(sg_page(s))) {
1031                         check_for_stack(dev, sg_virt(s));
1032                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1033                 }
1034
1035                 add_dma_entry(entry);
1036         }
1037 }
1038 EXPORT_SYMBOL(debug_dma_map_sg);
1039
1040 static int get_nr_mapped_entries(struct device *dev, struct scatterlist *s)
1041 {
1042         struct dma_debug_entry *entry;
1043         struct hash_bucket *bucket;
1044         unsigned long flags;
1045         int mapped_ents = 0;
1046         struct dma_debug_entry ref;
1047
1048         ref.dev = dev;
1049         ref.dev_addr = sg_dma_address(s);
1050         ref.size = sg_dma_len(s),
1051
1052         bucket = get_hash_bucket(&ref, &flags);
1053         entry = hash_bucket_find(bucket, &ref);
1054         if (entry)
1055                 mapped_ents = entry->sg_mapped_ents;
1056         put_hash_bucket(bucket, &flags);
1057
1058         return mapped_ents;
1059 }
1060
1061 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1062                         int nelems, int dir)
1063 {
1064         struct scatterlist *s;
1065         int mapped_ents = 0, i;
1066
1067         if (unlikely(global_disable))
1068                 return;
1069
1070         for_each_sg(sglist, s, nelems, i) {
1071
1072                 struct dma_debug_entry ref = {
1073                         .type           = dma_debug_sg,
1074                         .dev            = dev,
1075                         .paddr          = sg_phys(s),
1076                         .dev_addr       = sg_dma_address(s),
1077                         .size           = sg_dma_len(s),
1078                         .direction      = dir,
1079                         .sg_call_ents   = 0,
1080                 };
1081
1082                 if (mapped_ents && i >= mapped_ents)
1083                         break;
1084
1085                 if (!i) {
1086                         ref.sg_call_ents = nelems;
1087                         mapped_ents = get_nr_mapped_entries(dev, s);
1088                 }
1089
1090                 check_unmap(&ref);
1091         }
1092 }
1093 EXPORT_SYMBOL(debug_dma_unmap_sg);
1094
1095 void debug_dma_alloc_coherent(struct device *dev, size_t size,
1096                               dma_addr_t dma_addr, void *virt)
1097 {
1098         struct dma_debug_entry *entry;
1099
1100         if (unlikely(global_disable))
1101                 return;
1102
1103         if (unlikely(virt == NULL))
1104                 return;
1105
1106         entry = dma_entry_alloc();
1107         if (!entry)
1108                 return;
1109
1110         entry->type      = dma_debug_coherent;
1111         entry->dev       = dev;
1112         entry->paddr     = virt_to_phys(virt);
1113         entry->size      = size;
1114         entry->dev_addr  = dma_addr;
1115         entry->direction = DMA_BIDIRECTIONAL;
1116
1117         add_dma_entry(entry);
1118 }
1119 EXPORT_SYMBOL(debug_dma_alloc_coherent);
1120
1121 void debug_dma_free_coherent(struct device *dev, size_t size,
1122                          void *virt, dma_addr_t addr)
1123 {
1124         struct dma_debug_entry ref = {
1125                 .type           = dma_debug_coherent,
1126                 .dev            = dev,
1127                 .paddr          = virt_to_phys(virt),
1128                 .dev_addr       = addr,
1129                 .size           = size,
1130                 .direction      = DMA_BIDIRECTIONAL,
1131         };
1132
1133         if (unlikely(global_disable))
1134                 return;
1135
1136         check_unmap(&ref);
1137 }
1138 EXPORT_SYMBOL(debug_dma_free_coherent);
1139
1140 void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1141                                    size_t size, int direction)
1142 {
1143         if (unlikely(global_disable))
1144                 return;
1145
1146         check_sync(dev, dma_handle, size, 0, direction, true);
1147 }
1148 EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1149
1150 void debug_dma_sync_single_for_device(struct device *dev,
1151                                       dma_addr_t dma_handle, size_t size,
1152                                       int direction)
1153 {
1154         if (unlikely(global_disable))
1155                 return;
1156
1157         check_sync(dev, dma_handle, size, 0, direction, false);
1158 }
1159 EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1160
1161 void debug_dma_sync_single_range_for_cpu(struct device *dev,
1162                                          dma_addr_t dma_handle,
1163                                          unsigned long offset, size_t size,
1164                                          int direction)
1165 {
1166         if (unlikely(global_disable))
1167                 return;
1168
1169         check_sync(dev, dma_handle, size, offset, direction, true);
1170 }
1171 EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
1172
1173 void debug_dma_sync_single_range_for_device(struct device *dev,
1174                                             dma_addr_t dma_handle,
1175                                             unsigned long offset,
1176                                             size_t size, int direction)
1177 {
1178         if (unlikely(global_disable))
1179                 return;
1180
1181         check_sync(dev, dma_handle, size, offset, direction, false);
1182 }
1183 EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
1184
1185 void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1186                                int nelems, int direction)
1187 {
1188         struct scatterlist *s;
1189         int mapped_ents = 0, i;
1190
1191         if (unlikely(global_disable))
1192                 return;
1193
1194         for_each_sg(sg, s, nelems, i) {
1195                 if (!i)
1196                         mapped_ents = get_nr_mapped_entries(dev, s);
1197
1198                 if (i >= mapped_ents)
1199                         break;
1200
1201                 check_sync(dev, sg_dma_address(s), sg_dma_len(s), 0,
1202                            direction, true);
1203         }
1204 }
1205 EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1206
1207 void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1208                                   int nelems, int direction)
1209 {
1210         struct scatterlist *s;
1211         int mapped_ents = 0, i;
1212
1213         if (unlikely(global_disable))
1214                 return;
1215
1216         for_each_sg(sg, s, nelems, i) {
1217                 if (!i)
1218                         mapped_ents = get_nr_mapped_entries(dev, s);
1219
1220                 if (i >= mapped_ents)
1221                         break;
1222
1223                 check_sync(dev, sg_dma_address(s), sg_dma_len(s), 0,
1224                            direction, false);
1225         }
1226 }
1227 EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1228
1229 static int __init dma_debug_driver_setup(char *str)
1230 {
1231         int i;
1232
1233         for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1234                 current_driver_name[i] = *str;
1235                 if (*str == 0)
1236                         break;
1237         }
1238
1239         if (current_driver_name[0])
1240                 printk(KERN_INFO "DMA-API: enable driver filter for "
1241                                  "driver [%s]\n", current_driver_name);
1242
1243
1244         return 1;
1245 }
1246 __setup("dma_debug_driver=", dma_debug_driver_setup);