lib/dma-debug.c

   1 /*
   2  * Copyright (C) 2008 Advanced Micro Devices, Inc.
   3  *
   4  * Author: Joerg Roedel <joerg.roedel@amd.com>
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 as published
   8  * by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  18  */
  19
  20 #include <linux/scatterlist.h>
  21 #include <linux/dma-mapping.h>
  22 #include <linux/stacktrace.h>
  23 #include <linux/dma-debug.h>
  24 #include <linux/spinlock.h>
  25 #include <linux/debugfs.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/device.h>
  28 #include <linux/types.h>
  29 #include <linux/sched.h>
  30 #include <linux/ctype.h>
  31 #include <linux/list.h>
  32 #include <linux/slab.h>
  33
  34 #include <asm/sections.h>
  35
  36 #define HASH_SIZE       1024ULL
  37 #define HASH_FN_SHIFT   13
  38 #define HASH_FN_MASK    (HASH_SIZE - 1)
  39
  40 enum {
  41         dma_debug_single,
  42         dma_debug_page,
  43         dma_debug_sg,
  44         dma_debug_coherent,
  45 };
  46
  47 #define DMA_DEBUG_STACKTRACE_ENTRIES 5
  48
  49 struct dma_debug_entry {
  50         struct list_head list;
  51         struct device    *dev;
  52         int              type;
  53         phys_addr_t      paddr;
  54         u64              dev_addr;
  55         u64              size;
  56         int              direction;
  57         int              sg_call_ents;
  58         int              sg_mapped_ents;
  59 #ifdef CONFIG_STACKTRACE
  60         struct           stack_trace stacktrace;
  61         unsigned long    st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
  62 #endif
  63 };
  64
  65 struct hash_bucket {
  66         struct list_head list;
  67         spinlock_t lock;
  68 } ____cacheline_aligned_in_smp;
  69
  70 /* Hash list to save the allocated dma addresses */
  71 static struct hash_bucket dma_entry_hash[HASH_SIZE];
  72 /* List of pre-allocated dma_debug_entry's */
  73 static LIST_HEAD(free_entries);
  74 /* Lock for the list above */
  75 static DEFINE_SPINLOCK(free_entries_lock);
  76
  77 /* Global disable flag - will be set in case of an error */
  78 static bool global_disable __read_mostly;
  79
  80 /* Global error count */
  81 static u32 error_count;
  82
  83 /* Global error show enable*/
  84 static u32 show_all_errors __read_mostly;
  85 /* Number of errors to show */
  86 static u32 show_num_errors = 1;
  87
  88 static u32 num_free_entries;
  89 static u32 min_free_entries;
  90 static u32 nr_total_entries;
  91
  92 /* number of preallocated entries requested by kernel cmdline */
  93 static u32 req_entries;
  94
  95 /* debugfs dentry's for the stuff above */
  96 static struct dentry *dma_debug_dent        __read_mostly;
  97 static struct dentry *global_disable_dent   __read_mostly;
  98 static struct dentry *error_count_dent      __read_mostly;
  99 static struct dentry *show_all_errors_dent  __read_mostly;
 100 static struct dentry *show_num_errors_dent  __read_mostly;
 101 static struct dentry *num_free_entries_dent __read_mostly;
 102 static struct dentry *min_free_entries_dent __read_mostly;
 103 static struct dentry *filter_dent           __read_mostly;
 104
 105 /* per-driver filter related state */
 106
 107 #define NAME_MAX_LEN    64
 108
 109 static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
 110 static struct device_driver *current_driver                    __read_mostly;
 111
 112 static DEFINE_RWLOCK(driver_name_lock);
 113
 114 static const char *type2name[4] = { "single", "page",
 115                                     "scather-gather", "coherent" };
 116
 117 static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
 118                                    "DMA_FROM_DEVICE", "DMA_NONE" };
 119
 120 /* little merge helper - remove it after the merge window */
 121 #ifndef BUS_NOTIFY_UNBOUND_DRIVER
 122 #define BUS_NOTIFY_UNBOUND_DRIVER 0x0005
 123 #endif
 124
 125 /*
 126  * The access to some variables in this macro is racy. We can't use atomic_t
 127  * here because all these variables are exported to debugfs. Some of them even
 128  * writeable. This is also the reason why a lock won't help much. But anyway,
 129  * the races are no big deal. Here is why:
 130  *
 131  *   error_count: the addition is racy, but the worst thing that can happen is
 132  *                that we don't count some errors
 133  *   show_num_errors: the subtraction is racy. Also no big deal because in
 134  *                    worst case this will result in one warning more in the
 135  *                    system log than the user configured. This variable is
 136  *                    writeable via debugfs.
 137  */
 138 static inline void dump_entry_trace(struct dma_debug_entry *entry)
 139 {
 140 #ifdef CONFIG_STACKTRACE
 141         if (entry) {
 142                 pr_warning("Mapped at:\n");
 143                 print_stack_trace(&entry->stacktrace, 0);
 144         }
 145 #endif
 146 }
 147
 148 static bool driver_filter(struct device *dev)
 149 {
 150         struct device_driver *drv;
 151         unsigned long flags;
 152         bool ret;
 153
 154         /* driver filter off */
 155         if (likely(!current_driver_name[0]))
 156                 return true;
 157
 158         /* driver filter on and initialized */
 159         if (current_driver && dev->driver == current_driver)
 160                 return true;
 161
 162         if (current_driver || !current_driver_name[0])
 163                 return false;
 164
 165         /* driver filter on but not yet initialized */
 166         drv = get_driver(dev->driver);
 167         if (!drv)
 168                 return false;
 169
 170         /* lock to protect against change of current_driver_name */
 171         read_lock_irqsave(&driver_name_lock, flags);
 172
 173         ret = false;
 174         if (drv->name &&
 175             strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 176                 current_driver = drv;
 177                 ret = true;
 178         }
 179
 180         read_unlock_irqrestore(&driver_name_lock, flags);
 181         put_driver(drv);
 182
 183         return ret;
 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                 pr_err("DMA-API: debugging out of memory - disabling\n");
 381                 global_disable = true;
 382                 goto out;
 383         }
 384
 385         entry = __dma_entry_alloc();
 386
 387 #ifdef CONFIG_STACKTRACE
 388         entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
 389         entry->stacktrace.entries = entry->st_entries;
 390         entry->stacktrace.skip = 2;
 391         save_stack_trace(&entry->stacktrace);
 392 #endif
 393
 394 out:
 395         spin_unlock_irqrestore(&free_entries_lock, flags);
 396
 397         return entry;
 398 }
 399
 400 static void dma_entry_free(struct dma_debug_entry *entry)
 401 {
 402         unsigned long flags;
 403
 404         /*
 405          * add to beginning of the list - this way the entries are
 406          * more likely cache hot when they are reallocated.
 407          */
 408         spin_lock_irqsave(&free_entries_lock, flags);
 409         list_add(&entry->list, &free_entries);
 410         num_free_entries += 1;
 411         spin_unlock_irqrestore(&free_entries_lock, flags);
 412 }
 413
 414 int dma_debug_resize_entries(u32 num_entries)
 415 {
 416         int i, delta, ret = 0;
 417         unsigned long flags;
 418         struct dma_debug_entry *entry;
 419         LIST_HEAD(tmp);
 420
 421         spin_lock_irqsave(&free_entries_lock, flags);
 422
 423         if (nr_total_entries < num_entries) {
 424                 delta = num_entries - nr_total_entries;
 425
 426                 spin_unlock_irqrestore(&free_entries_lock, flags);
 427
 428                 for (i = 0; i < delta; i++) {
 429                         entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 430                         if (!entry)
 431                                 break;
 432
 433                         list_add_tail(&entry->list, &tmp);
 434                 }
 435
 436                 spin_lock_irqsave(&free_entries_lock, flags);
 437
 438                 list_splice(&tmp, &free_entries);
 439                 nr_total_entries += i;
 440                 num_free_entries += i;
 441         } else {
 442                 delta = nr_total_entries - num_entries;
 443
 444                 for (i = 0; i < delta && !list_empty(&free_entries); i++) {
 445                         entry = __dma_entry_alloc();
 446                         kfree(entry);
 447                 }
 448
 449                 nr_total_entries -= i;
 450         }
 451
 452         if (nr_total_entries != num_entries)
 453                 ret = 1;
 454
 455         spin_unlock_irqrestore(&free_entries_lock, flags);
 456
 457         return ret;
 458 }
 459 EXPORT_SYMBOL(dma_debug_resize_entries);
 460
 461 /*
 462  * DMA-API debugging init code
 463  *
 464  * The init code does two things:
 465  *   1. Initialize core data structures
 466  *   2. Preallocate a given number of dma_debug_entry structs
 467  */
 468
 469 static int prealloc_memory(u32 num_entries)
 470 {
 471         struct dma_debug_entry *entry, *next_entry;
 472         int i;
 473
 474         for (i = 0; i < num_entries; ++i) {
 475                 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 476                 if (!entry)
 477                         goto out_err;
 478
 479                 list_add_tail(&entry->list, &free_entries);
 480         }
 481
 482         num_free_entries = num_entries;
 483         min_free_entries = num_entries;
 484
 485         pr_info("DMA-API: preallocated %d debug entries\n", num_entries);
 486
 487         return 0;
 488
 489 out_err:
 490
 491         list_for_each_entry_safe(entry, next_entry, &free_entries, list) {
 492                 list_del(&entry->list);
 493                 kfree(entry);
 494         }
 495
 496         return -ENOMEM;
 497 }
 498
 499 static ssize_t filter_read(struct file *file, char __user *user_buf,
 500                            size_t count, loff_t *ppos)
 501 {
 502         char buf[NAME_MAX_LEN + 1];
 503         unsigned long flags;
 504         int len;
 505
 506         if (!current_driver_name[0])
 507                 return 0;
 508
 509         /*
 510          * We can't copy to userspace directly because current_driver_name can
 511          * only be read under the driver_name_lock with irqs disabled. So
 512          * create a temporary copy first.
 513          */
 514         read_lock_irqsave(&driver_name_lock, flags);
 515         len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 516         read_unlock_irqrestore(&driver_name_lock, flags);
 517
 518         return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 519 }
 520
 521 static ssize_t filter_write(struct file *file, const char __user *userbuf,
 522                             size_t count, loff_t *ppos)
 523 {
 524         char buf[NAME_MAX_LEN];
 525         unsigned long flags;
 526         size_t len;
 527         int i;
 528
 529         /*
 530          * We can't copy from userspace directly. Access to
 531          * current_driver_name is protected with a write_lock with irqs
 532          * disabled. Since copy_from_user can fault and may sleep we
 533          * need to copy to temporary buffer first
 534          */
 535         len = min(count, (size_t)(NAME_MAX_LEN - 1));
 536         if (copy_from_user(buf, userbuf, len))
 537                 return -EFAULT;
 538
 539         buf[len] = 0;
 540
 541         write_lock_irqsave(&driver_name_lock, flags);
 542
 543         /*
 544          * Now handle the string we got from userspace very carefully.
 545          * The rules are:
 546          *         - only use the first token we got
 547          *         - token delimiter is everything looking like a space
 548          *           character (' ', '\n', '\t' ...)
 549          *
 550          */
 551         if (!isalnum(buf[0])) {
 552                 /*
 553                  * If the first character userspace gave us is not
 554                  * alphanumerical then assume the filter should be
 555                  * switched off.
 556                  */
 557                 if (current_driver_name[0])
 558                         pr_info("DMA-API: switching off dma-debug driver filter\n");
 559                 current_driver_name[0] = 0;
 560                 current_driver = NULL;
 561                 goto out_unlock;
 562         }
 563
 564         /*
 565          * Now parse out the first token and use it as the name for the
 566          * driver to filter for.
 567          */
 568         for (i = 0; i < NAME_MAX_LEN; ++i) {
 569                 current_driver_name[i] = buf[i];
 570                 if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 571                         break;
 572         }
 573         current_driver_name[i] = 0;
 574         current_driver = NULL;
 575
 576         pr_info("DMA-API: enable driver filter for driver [%s]\n",
 577                 current_driver_name);
 578
 579 out_unlock:
 580         write_unlock_irqrestore(&driver_name_lock, flags);
 581
 582         return count;
 583 }
 584
 585 const struct file_operations filter_fops = {
 586         .read  = filter_read,
 587         .write = filter_write,
 588 };
 589
 590 static int dma_debug_fs_init(void)
 591 {
 592         dma_debug_dent = debugfs_create_dir("dma-api", NULL);
 593         if (!dma_debug_dent) {
 594                 pr_err("DMA-API: can not create debugfs directory\n");
 595                 return -ENOMEM;
 596         }
 597
 598         global_disable_dent = debugfs_create_bool("disabled", 0444,
 599                         dma_debug_dent,
 600                         (u32 *)&global_disable);
 601         if (!global_disable_dent)
 602                 goto out_err;
 603
 604         error_count_dent = debugfs_create_u32("error_count", 0444,
 605                         dma_debug_dent, &error_count);
 606         if (!error_count_dent)
 607                 goto out_err;
 608
 609         show_all_errors_dent = debugfs_create_u32("all_errors", 0644,
 610                         dma_debug_dent,
 611                         &show_all_errors);
 612         if (!show_all_errors_dent)
 613                 goto out_err;
 614
 615         show_num_errors_dent = debugfs_create_u32("num_errors", 0644,
 616                         dma_debug_dent,
 617                         &show_num_errors);
 618         if (!show_num_errors_dent)
 619                 goto out_err;
 620
 621         num_free_entries_dent = debugfs_create_u32("num_free_entries", 0444,
 622                         dma_debug_dent,
 623                         &num_free_entries);
 624         if (!num_free_entries_dent)
 625                 goto out_err;
 626
 627         min_free_entries_dent = debugfs_create_u32("min_free_entries", 0444,
 628                         dma_debug_dent,
 629                         &min_free_entries);
 630         if (!min_free_entries_dent)
 631                 goto out_err;
 632
 633         filter_dent = debugfs_create_file("driver_filter", 0644,
 634                                           dma_debug_dent, NULL, &filter_fops);
 635         if (!filter_dent)
 636                 goto out_err;
 637
 638         return 0;
 639
 640 out_err:
 641         debugfs_remove_recursive(dma_debug_dent);
 642
 643         return -ENOMEM;
 644 }
 645
 646 static int device_dma_allocations(struct device *dev)
 647 {
 648         struct dma_debug_entry *entry;
 649         unsigned long flags;
 650         int count = 0, i;
 651
 652         local_irq_save(flags);
 653
 654         for (i = 0; i < HASH_SIZE; ++i) {
 655                 spin_lock(&dma_entry_hash[i].lock);
 656                 list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 657                         if (entry->dev == dev)
 658                                 count += 1;
 659                 }
 660                 spin_unlock(&dma_entry_hash[i].lock);
 661         }
 662
 663         local_irq_restore(flags);
 664
 665         return count;
 666 }
 667
 668 static int dma_debug_device_change(struct notifier_block *nb,
 669                                     unsigned long action, void *data)
 670 {
 671         struct device *dev = data;
 672         int count;
 673
 674
 675         switch (action) {
 676         case BUS_NOTIFY_UNBOUND_DRIVER:
 677                 count = device_dma_allocations(dev);
 678                 if (count == 0)
 679                         break;
 680                 err_printk(dev, NULL, "DMA-API: device driver has pending "
 681                                 "DMA allocations while released from device "
 682                                 "[count=%d]\n", count);
 683                 break;
 684         default:
 685                 break;
 686         }
 687
 688         return 0;
 689 }
 690
 691 void dma_debug_add_bus(struct bus_type *bus)
 692 {
 693         struct notifier_block *nb;
 694
 695         nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 696         if (nb == NULL) {
 697                 pr_err("dma_debug_add_bus: out of memory\n");
 698                 return;
 699         }
 700
 701         nb->notifier_call = dma_debug_device_change;
 702
 703         bus_register_notifier(bus, nb);
 704 }
 705
 706 /*
 707  * Let the architectures decide how many entries should be preallocated.
 708  */
 709 void dma_debug_init(u32 num_entries)
 710 {
 711         int i;
 712
 713         if (global_disable)
 714                 return;
 715
 716         for (i = 0; i < HASH_SIZE; ++i) {
 717                 INIT_LIST_HEAD(&dma_entry_hash[i].list);
 718                 dma_entry_hash[i].lock = SPIN_LOCK_UNLOCKED;
 719         }
 720
 721         if (dma_debug_fs_init() != 0) {
 722                 pr_err("DMA-API: error creating debugfs entries - disabling\n");
 723                 global_disable = true;
 724
 725                 return;
 726         }
 727
 728         if (req_entries)
 729                 num_entries = req_entries;
 730
 731         if (prealloc_memory(num_entries) != 0) {
 732                 pr_err("DMA-API: debugging out of memory error - disabled\n");
 733                 global_disable = true;
 734
 735                 return;
 736         }
 737
 738         nr_total_entries = num_free_entries;
 739
 740         pr_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                 pr_info("DMA-API: debugging disabled on kernel command line\n");
 750                 global_disable = true;
 751         }
 752
 753         return 0;
 754 }
 755
 756 static __init int dma_debug_entries_cmdline(char *str)
 757 {
 758         int res;
 759
 760         if (!str)
 761                 return -EINVAL;
 762
 763         res = get_option(&str, &req_entries);
 764
 765         if (!res)
 766                 req_entries = 0;
 767
 768         return 0;
 769 }
 770
 771 __setup("dma_debug=", dma_debug_cmdline);
 772 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
 773
 774 static void check_unmap(struct dma_debug_entry *ref)
 775 {
 776         struct dma_debug_entry *entry;
 777         struct hash_bucket *bucket;
 778         unsigned long flags;
 779
 780         if (dma_mapping_error(ref->dev, ref->dev_addr)) {
 781                 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
 782                            "to free an invalid DMA memory address\n");
 783                 return;
 784         }
 785
 786         bucket = get_hash_bucket(ref, &flags);
 787         entry = hash_bucket_find(bucket, ref);
 788
 789         if (!entry) {
 790                 err_printk(ref->dev, NULL, "DMA-API: device driver tries "
 791                            "to free DMA memory it has not allocated "
 792                            "[device address=0x%016llx] [size=%llu bytes]\n",
 793                            ref->dev_addr, ref->size);
 794                 goto out;
 795         }
 796
 797         if (ref->size != entry->size) {
 798                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 799                            "DMA memory with different size "
 800                            "[device address=0x%016llx] [map size=%llu bytes] "
 801                            "[unmap size=%llu bytes]\n",
 802                            ref->dev_addr, entry->size, ref->size);
 803         }
 804
 805         if (ref->type != entry->type) {
 806                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 807                            "DMA memory with wrong function "
 808                            "[device address=0x%016llx] [size=%llu bytes] "
 809                            "[mapped as %s] [unmapped as %s]\n",
 810                            ref->dev_addr, ref->size,
 811                            type2name[entry->type], type2name[ref->type]);
 812         } else if ((entry->type == dma_debug_coherent) &&
 813                    (ref->paddr != entry->paddr)) {
 814                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 815                            "DMA memory with different CPU address "
 816                            "[device address=0x%016llx] [size=%llu bytes] "
 817                            "[cpu alloc address=%p] [cpu free address=%p]",
 818                            ref->dev_addr, ref->size,
 819                            (void *)entry->paddr, (void *)ref->paddr);
 820         }
 821
 822         if (ref->sg_call_ents && ref->type == dma_debug_sg &&
 823             ref->sg_call_ents != entry->sg_call_ents) {
 824                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 825                            "DMA sg list with different entry count "
 826                            "[map count=%d] [unmap count=%d]\n",
 827                            entry->sg_call_ents, ref->sg_call_ents);
 828         }
 829
 830         /*
 831          * This may be no bug in reality - but most implementations of the
 832          * DMA API don't handle this properly, so check for it here
 833          */
 834         if (ref->direction != entry->direction) {
 835                 err_printk(ref->dev, entry, "DMA-API: device driver frees "
 836                            "DMA memory with different direction "
 837                            "[device address=0x%016llx] [size=%llu bytes] "
 838                            "[mapped with %s] [unmapped with %s]\n",
 839                            ref->dev_addr, ref->size,
 840                            dir2name[entry->direction],
 841                            dir2name[ref->direction]);
 842         }
 843
 844         hash_bucket_del(entry);
 845         dma_entry_free(entry);
 846
 847 out:
 848         put_hash_bucket(bucket, &flags);
 849 }
 850
 851 static void check_for_stack(struct device *dev, void *addr)
 852 {
 853         if (object_is_on_stack(addr))
 854                 err_printk(dev, NULL, "DMA-API: device driver maps memory from"
 855                                 "stack [addr=%p]\n", addr);
 856 }
 857
 858 static inline bool overlap(void *addr, u64 size, void *start, void *end)
 859 {
 860         void *addr2 = (char *)addr + size;
 861
 862         return ((addr >= start && addr < end) ||
 863                 (addr2 >= start && addr2 < end) ||
 864                 ((addr < start) && (addr2 >= end)));
 865 }
 866
 867 static void check_for_illegal_area(struct device *dev, void *addr, u64 size)
 868 {
 869         if (overlap(addr, size, _text, _etext) ||
 870             overlap(addr, size, __start_rodata, __end_rodata))
 871                 err_printk(dev, NULL, "DMA-API: device driver maps "
 872                                 "memory from kernel text or rodata "
 873                                 "[addr=%p] [size=%llu]\n", addr, size);
 874 }
 875
 876 static void check_sync(struct device *dev, dma_addr_t addr,
 877                        u64 size, u64 offset, int direction, bool to_cpu)
 878 {
 879         struct dma_debug_entry ref = {
 880                 .dev            = dev,
 881                 .dev_addr       = addr,
 882                 .size           = size,
 883                 .direction      = direction,
 884         };
 885         struct dma_debug_entry *entry;
 886         struct hash_bucket *bucket;
 887         unsigned long flags;
 888
 889         bucket = get_hash_bucket(&ref, &flags);
 890
 891         entry = hash_bucket_find(bucket, &ref);
 892
 893         if (!entry) {
 894                 err_printk(dev, NULL, "DMA-API: device driver tries "
 895                                 "to sync DMA memory it has not allocated "
 896                                 "[device address=0x%016llx] [size=%llu bytes]\n",
 897                                 (unsigned long long)addr, size);
 898                 goto out;
 899         }
 900
 901         if ((offset + size) > entry->size) {
 902                 err_printk(dev, entry, "DMA-API: device driver syncs"
 903                                 " DMA memory outside allocated range "
 904                                 "[device address=0x%016llx] "
 905                                 "[allocation size=%llu bytes] [sync offset=%llu] "
 906                                 "[sync size=%llu]\n", entry->dev_addr, entry->size,
 907                                 offset, size);
 908         }
 909
 910         if (direction != entry->direction) {
 911                 err_printk(dev, entry, "DMA-API: device driver syncs "
 912                                 "DMA memory with different direction "
 913                                 "[device address=0x%016llx] [size=%llu bytes] "
 914                                 "[mapped with %s] [synced with %s]\n",
 915                                 (unsigned long long)addr, entry->size,
 916                                 dir2name[entry->direction],
 917                                 dir2name[direction]);
 918         }
 919
 920         if (entry->direction == DMA_BIDIRECTIONAL)
 921                 goto out;
 922
 923         if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
 924                       !(direction == DMA_TO_DEVICE))
 925                 err_printk(dev, entry, "DMA-API: device driver syncs "
 926                                 "device read-only DMA memory for cpu "
 927                                 "[device address=0x%016llx] [size=%llu bytes] "
 928                                 "[mapped with %s] [synced with %s]\n",
 929                                 (unsigned long long)addr, entry->size,
 930                                 dir2name[entry->direction],
 931                                 dir2name[direction]);
 932
 933         if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
 934                        !(direction == DMA_FROM_DEVICE))
 935                 err_printk(dev, entry, "DMA-API: device driver syncs "
 936                                 "device write-only DMA memory to device "
 937                                 "[device address=0x%016llx] [size=%llu bytes] "
 938                                 "[mapped with %s] [synced with %s]\n",
 939                                 (unsigned long long)addr, entry->size,
 940                                 dir2name[entry->direction],
 941                                 dir2name[direction]);
 942
 943 out:
 944         put_hash_bucket(bucket, &flags);
 945
 946 }
 947
 948 void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
 949                         size_t size, int direction, dma_addr_t dma_addr,
 950                         bool map_single)
 951 {
 952         struct dma_debug_entry *entry;
 953
 954         if (unlikely(global_disable))
 955                 return;
 956
 957         if (unlikely(dma_mapping_error(dev, dma_addr)))
 958                 return;
 959
 960         entry = dma_entry_alloc();
 961         if (!entry)
 962                 return;
 963
 964         entry->dev       = dev;
 965         entry->type      = dma_debug_page;
 966         entry->paddr     = page_to_phys(page) + offset;
 967         entry->dev_addr  = dma_addr;
 968         entry->size      = size;
 969         entry->direction = direction;
 970
 971         if (map_single)
 972                 entry->type = dma_debug_single;
 973
 974         if (!PageHighMem(page)) {
 975                 void *addr = ((char *)page_address(page)) + offset;
 976                 check_for_stack(dev, addr);
 977                 check_for_illegal_area(dev, addr, size);
 978         }
 979
 980         add_dma_entry(entry);
 981 }
 982 EXPORT_SYMBOL(debug_dma_map_page);
 983
 984 void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
 985                           size_t size, int direction, bool map_single)
 986 {
 987         struct dma_debug_entry ref = {
 988                 .type           = dma_debug_page,
 989                 .dev            = dev,
 990                 .dev_addr       = addr,
 991                 .size           = size,
 992                 .direction      = direction,
 993         };
 994
 995         if (unlikely(global_disable))
 996                 return;
 997
 998         if (map_single)
 999                 ref.type = dma_debug_single;
1000
1001         check_unmap(&ref);
1002 }
1003 EXPORT_SYMBOL(debug_dma_unmap_page);
1004
1005 void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1006                       int nents, int mapped_ents, int direction)
1007 {
1008         struct dma_debug_entry *entry;
1009         struct scatterlist *s;
1010         int i;
1011
1012         if (unlikely(global_disable))
1013                 return;
1014
1015         for_each_sg(sg, s, mapped_ents, i) {
1016                 entry = dma_entry_alloc();
1017                 if (!entry)
1018                         return;
1019
1020                 entry->type           = dma_debug_sg;
1021                 entry->dev            = dev;
1022                 entry->paddr          = sg_phys(s);
1023                 entry->size           = sg_dma_len(s);
1024                 entry->dev_addr       = sg_dma_address(s);
1025                 entry->direction      = direction;
1026                 entry->sg_call_ents   = nents;
1027                 entry->sg_mapped_ents = mapped_ents;
1028
1029                 if (!PageHighMem(sg_page(s))) {
1030                         check_for_stack(dev, sg_virt(s));
1031                         check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1032                 }
1033
1034                 add_dma_entry(entry);
1035         }
1036 }
1037 EXPORT_SYMBOL(debug_dma_map_sg);
1038
1039 static int get_nr_mapped_entries(struct device *dev, struct scatterlist *s)
1040 {
1041         struct dma_debug_entry *entry, ref;
1042         struct hash_bucket *bucket;
1043         unsigned long flags;
1044         int mapped_ents;
1045
1046         ref.dev      = dev;
1047         ref.dev_addr = sg_dma_address(s);
1048         ref.size     = sg_dma_len(s),
1049
1050         bucket       = get_hash_bucket(&ref, &flags);
1051         entry        = hash_bucket_find(bucket, &ref);
1052         mapped_ents  = 0;
1053
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                 pr_info("DMA-API: enable driver filter for driver [%s]\n",
1241                         current_driver_name);
1242
1243
1244         return 1;
1245 }
1246 __setup("dma_debug_driver=", dma_debug_driver_setup);