Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2008 Advanced Micro Devices, Inc.
   4 *
   5 * Author: Joerg Roedel <joerg.roedel@amd.com>
   6 */
   7
   8#define pr_fmt(fmt)	"DMA-API: " fmt
   9
  10#include <linux/sched/task_stack.h>
  11#include <linux/scatterlist.h>
  12#include <linux/dma-map-ops.h>
  13#include <linux/sched/task.h>
  14#include <linux/stacktrace.h>
  15#include <linux/spinlock.h>
  16#include <linux/vmalloc.h>
  17#include <linux/debugfs.h>
  18#include <linux/uaccess.h>
  19#include <linux/export.h>
  20#include <linux/device.h>
  21#include <linux/types.h>
  22#include <linux/sched.h>
  23#include <linux/ctype.h>
  24#include <linux/list.h>
  25#include <linux/slab.h>
  26#include <asm/sections.h>
  27#include "debug.h"
  28
  29#define HASH_SIZE       16384ULL
  30#define HASH_FN_SHIFT   13
  31#define HASH_FN_MASK    (HASH_SIZE - 1)
  32
  33#define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
  34/* If the pool runs out, add this many new entries at once */
  35#define DMA_DEBUG_DYNAMIC_ENTRIES (PAGE_SIZE / sizeof(struct dma_debug_entry))
  36
  37enum {
  38	dma_debug_single,
  39	dma_debug_sg,
  40	dma_debug_coherent,
  41	dma_debug_resource,
  42};
  43
  44enum map_err_types {
  45	MAP_ERR_CHECK_NOT_APPLICABLE,
  46	MAP_ERR_NOT_CHECKED,
  47	MAP_ERR_CHECKED,
  48};
  49
  50#define DMA_DEBUG_STACKTRACE_ENTRIES 5
  51
  52/**
  53 * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
  54 * @list: node on pre-allocated free_entries list
  55 * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
 
  56 * @size: length of the mapping
  57 * @type: single, page, sg, coherent
  58 * @direction: enum dma_data_direction
  59 * @sg_call_ents: 'nents' from dma_map_sg
  60 * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
  61 * @pfn: page frame of the start address
  62 * @offset: offset of mapping relative to pfn
  63 * @map_err_type: track whether dma_mapping_error() was checked
  64 * @stacktrace: support backtraces when a violation is detected
 
  65 */
  66struct dma_debug_entry {
  67	struct list_head list;
  68	struct device    *dev;
  69	u64              dev_addr;
  70	u64              size;
  71	int              type;
  72	int              direction;
  73	int		 sg_call_ents;
  74	int		 sg_mapped_ents;
  75	unsigned long	 pfn;
  76	size_t		 offset;
  77	enum map_err_types  map_err_type;
  78#ifdef CONFIG_STACKTRACE
  79	unsigned int	stack_len;
  80	unsigned long	stack_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
  81#endif
  82} ____cacheline_aligned_in_smp;
  83
  84typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
  85
  86struct hash_bucket {
  87	struct list_head list;
  88	spinlock_t lock;
  89};
  90
  91/* Hash list to save the allocated dma addresses */
  92static struct hash_bucket dma_entry_hash[HASH_SIZE];
  93/* List of pre-allocated dma_debug_entry's */
  94static LIST_HEAD(free_entries);
  95/* Lock for the list above */
  96static DEFINE_SPINLOCK(free_entries_lock);
  97
  98/* Global disable flag - will be set in case of an error */
  99static bool global_disable __read_mostly;
 100
 101/* Early initialization disable flag, set at the end of dma_debug_init */
 102static bool dma_debug_initialized __read_mostly;
 103
 104static inline bool dma_debug_disabled(void)
 105{
 106	return global_disable || !dma_debug_initialized;
 107}
 108
 109/* Global error count */
 110static u32 error_count;
 111
 112/* Global error show enable*/
 113static u32 show_all_errors __read_mostly;
 114/* Number of errors to show */
 115static u32 show_num_errors = 1;
 116
 117static u32 num_free_entries;
 118static u32 min_free_entries;
 119static u32 nr_total_entries;
 120
 121/* number of preallocated entries requested by kernel cmdline */
 122static u32 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 123
 124/* per-driver filter related state */
 125
 126#define NAME_MAX_LEN	64
 127
 128static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
 129static struct device_driver *current_driver                    __read_mostly;
 130
 131static DEFINE_RWLOCK(driver_name_lock);
 132
 133static const char *const maperr2str[] = {
 134	[MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
 135	[MAP_ERR_NOT_CHECKED] = "dma map error not checked",
 136	[MAP_ERR_CHECKED] = "dma map error checked",
 137};
 138
 139static const char *type2name[] = {
 140	[dma_debug_single] = "single",
 141	[dma_debug_sg] = "scather-gather",
 142	[dma_debug_coherent] = "coherent",
 143	[dma_debug_resource] = "resource",
 144};
 145
 146static const char *dir2name[] = {
 147	[DMA_BIDIRECTIONAL]	= "DMA_BIDIRECTIONAL",
 148	[DMA_TO_DEVICE]		= "DMA_TO_DEVICE",
 149	[DMA_FROM_DEVICE]	= "DMA_FROM_DEVICE",
 150	[DMA_NONE]		= "DMA_NONE",
 151};
 152
 153/*
 154 * The access to some variables in this macro is racy. We can't use atomic_t
 155 * here because all these variables are exported to debugfs. Some of them even
 156 * writeable. This is also the reason why a lock won't help much. But anyway,
 157 * the races are no big deal. Here is why:
 158 *
 159 *   error_count: the addition is racy, but the worst thing that can happen is
 160 *                that we don't count some errors
 161 *   show_num_errors: the subtraction is racy. Also no big deal because in
 162 *                    worst case this will result in one warning more in the
 163 *                    system log than the user configured. This variable is
 164 *                    writeable via debugfs.
 165 */
 166static inline void dump_entry_trace(struct dma_debug_entry *entry)
 167{
 168#ifdef CONFIG_STACKTRACE
 169	if (entry) {
 170		pr_warn("Mapped at:\n");
 171		stack_trace_print(entry->stack_entries, entry->stack_len, 0);
 172	}
 173#endif
 174}
 175
 176static bool driver_filter(struct device *dev)
 177{
 178	struct device_driver *drv;
 179	unsigned long flags;
 180	bool ret;
 181
 182	/* driver filter off */
 183	if (likely(!current_driver_name[0]))
 184		return true;
 185
 186	/* driver filter on and initialized */
 187	if (current_driver && dev && dev->driver == current_driver)
 188		return true;
 189
 190	/* driver filter on, but we can't filter on a NULL device... */
 191	if (!dev)
 192		return false;
 193
 194	if (current_driver || !current_driver_name[0])
 195		return false;
 196
 197	/* driver filter on but not yet initialized */
 198	drv = dev->driver;
 199	if (!drv)
 200		return false;
 201
 202	/* lock to protect against change of current_driver_name */
 203	read_lock_irqsave(&driver_name_lock, flags);
 204
 205	ret = false;
 206	if (drv->name &&
 207	    strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 208		current_driver = drv;
 209		ret = true;
 210	}
 211
 212	read_unlock_irqrestore(&driver_name_lock, flags);
 213
 214	return ret;
 215}
 216
 217#define err_printk(dev, entry, format, arg...) do {			\
 218		error_count += 1;					\
 219		if (driver_filter(dev) &&				\
 220		    (show_all_errors || show_num_errors > 0)) {		\
 221			WARN(1, pr_fmt("%s %s: ") format,		\
 222			     dev ? dev_driver_string(dev) : "NULL",	\
 223			     dev ? dev_name(dev) : "NULL", ## arg);	\
 224			dump_entry_trace(entry);			\
 225		}							\
 226		if (!show_all_errors && show_num_errors > 0)		\
 227			show_num_errors -= 1;				\
 228	} while (0);
 229
 230/*
 231 * Hash related functions
 232 *
 233 * Every DMA-API request is saved into a struct dma_debug_entry. To
 234 * have quick access to these structs they are stored into a hash.
 235 */
 236static int hash_fn(struct dma_debug_entry *entry)
 237{
 238	/*
 239	 * Hash function is based on the dma address.
 240	 * We use bits 20-27 here as the index into the hash
 241	 */
 242	return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 243}
 244
 245/*
 246 * Request exclusive access to a hash bucket for a given dma_debug_entry.
 247 */
 248static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 249					   unsigned long *flags)
 250	__acquires(&dma_entry_hash[idx].lock)
 251{
 252	int idx = hash_fn(entry);
 253	unsigned long __flags;
 254
 255	spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 256	*flags = __flags;
 257	return &dma_entry_hash[idx];
 258}
 259
 260/*
 261 * Give up exclusive access to the hash bucket
 262 */
 263static void put_hash_bucket(struct hash_bucket *bucket,
 264			    unsigned long flags)
 265	__releases(&bucket->lock)
 266{
 267	spin_unlock_irqrestore(&bucket->lock, flags);
 268}
 269
 270static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
 271{
 272	return ((a->dev_addr == b->dev_addr) &&
 273		(a->dev == b->dev)) ? true : false;
 274}
 275
 276static bool containing_match(struct dma_debug_entry *a,
 277			     struct dma_debug_entry *b)
 278{
 279	if (a->dev != b->dev)
 280		return false;
 281
 282	if ((b->dev_addr <= a->dev_addr) &&
 283	    ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
 284		return true;
 285
 286	return false;
 287}
 288
 289/*
 290 * Search a given entry in the hash bucket list
 291 */
 292static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
 293						  struct dma_debug_entry *ref,
 294						  match_fn match)
 295{
 296	struct dma_debug_entry *entry, *ret = NULL;
 297	int matches = 0, match_lvl, last_lvl = -1;
 298
 299	list_for_each_entry(entry, &bucket->list, list) {
 300		if (!match(ref, entry))
 301			continue;
 302
 303		/*
 304		 * Some drivers map the same physical address multiple
 305		 * times. Without a hardware IOMMU this results in the
 306		 * same device addresses being put into the dma-debug
 307		 * hash multiple times too. This can result in false
 308		 * positives being reported. Therefore we implement a
 309		 * best-fit algorithm here which returns the entry from
 310		 * the hash which fits best to the reference value
 311		 * instead of the first-fit.
 312		 */
 313		matches += 1;
 314		match_lvl = 0;
 315		entry->size         == ref->size         ? ++match_lvl : 0;
 316		entry->type         == ref->type         ? ++match_lvl : 0;
 317		entry->direction    == ref->direction    ? ++match_lvl : 0;
 318		entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
 319
 320		if (match_lvl == 4) {
 321			/* perfect-fit - return the result */
 322			return entry;
 323		} else if (match_lvl > last_lvl) {
 324			/*
 325			 * We found an entry that fits better then the
 326			 * previous one or it is the 1st match.
 327			 */
 328			last_lvl = match_lvl;
 329			ret      = entry;
 330		}
 331	}
 332
 333	/*
 334	 * If we have multiple matches but no perfect-fit, just return
 335	 * NULL.
 336	 */
 337	ret = (matches == 1) ? ret : NULL;
 338
 339	return ret;
 340}
 341
 342static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
 343						 struct dma_debug_entry *ref)
 344{
 345	return __hash_bucket_find(bucket, ref, exact_match);
 346}
 347
 348static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 349						   struct dma_debug_entry *ref,
 350						   unsigned long *flags)
 351{
 352
 353	struct dma_debug_entry *entry, index = *ref;
 354	int limit = min(HASH_SIZE, (index.dev_addr >> HASH_FN_SHIFT) + 1);
 355
 356	for (int i = 0; i < limit; i++) {
 357		entry = __hash_bucket_find(*bucket, ref, containing_match);
 358
 359		if (entry)
 360			return entry;
 361
 362		/*
 363		 * Nothing found, go back a hash bucket
 364		 */
 365		put_hash_bucket(*bucket, *flags);
 366		index.dev_addr -= (1 << HASH_FN_SHIFT);
 367		*bucket = get_hash_bucket(&index, flags);
 368	}
 369
 370	return NULL;
 371}
 372
 373/*
 374 * Add an entry to a hash bucket
 375 */
 376static void hash_bucket_add(struct hash_bucket *bucket,
 377			    struct dma_debug_entry *entry)
 378{
 379	list_add_tail(&entry->list, &bucket->list);
 380}
 381
 382/*
 383 * Remove entry from a hash bucket list
 384 */
 385static void hash_bucket_del(struct dma_debug_entry *entry)
 386{
 387	list_del(&entry->list);
 388}
 389
 390static unsigned long long phys_addr(struct dma_debug_entry *entry)
 391{
 392	if (entry->type == dma_debug_resource)
 393		return __pfn_to_phys(entry->pfn) + entry->offset;
 394
 395	return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
 396}
 397
 398/*
 399 * Dump mapping entries for debugging purposes
 400 */
 401void debug_dma_dump_mappings(struct device *dev)
 402{
 403	int idx;
 404
 405	for (idx = 0; idx < HASH_SIZE; idx++) {
 406		struct hash_bucket *bucket = &dma_entry_hash[idx];
 407		struct dma_debug_entry *entry;
 408		unsigned long flags;
 409
 410		spin_lock_irqsave(&bucket->lock, flags);
 411
 412		list_for_each_entry(entry, &bucket->list, list) {
 413			if (!dev || dev == entry->dev) {
 414				dev_info(entry->dev,
 415					 "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
 416					 type2name[entry->type], idx,
 417					 phys_addr(entry), entry->pfn,
 418					 entry->dev_addr, entry->size,
 419					 dir2name[entry->direction],
 420					 maperr2str[entry->map_err_type]);
 421			}
 422		}
 423
 424		spin_unlock_irqrestore(&bucket->lock, flags);
 425		cond_resched();
 426	}
 427}
 428
 429/*
 430 * For each mapping (initial cacheline in the case of
 431 * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
 432 * scatterlist, or the cacheline specified in dma_map_single) insert
 433 * into this tree using the cacheline as the key. At
 434 * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
 435 * the entry already exists at insertion time add a tag as a reference
 436 * count for the overlapping mappings.  For now, the overlap tracking
 437 * just ensures that 'unmaps' balance 'maps' before marking the
 438 * cacheline idle, but we should also be flagging overlaps as an API
 439 * violation.
 440 *
 441 * Memory usage is mostly constrained by the maximum number of available
 442 * dma-debug entries in that we need a free dma_debug_entry before
 443 * inserting into the tree.  In the case of dma_map_page and
 444 * dma_alloc_coherent there is only one dma_debug_entry and one
 445 * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
 446 * other hand, consumes a single dma_debug_entry, but inserts 'nents'
 447 * entries into the tree.
 448 */
 449static RADIX_TREE(dma_active_cacheline, GFP_ATOMIC);
 450static DEFINE_SPINLOCK(radix_lock);
 451#define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
 452#define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
 453#define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
 454
 455static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
 456{
 457	return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
 458		(entry->offset >> L1_CACHE_SHIFT);
 459}
 460
 461static int active_cacheline_read_overlap(phys_addr_t cln)
 462{
 463	int overlap = 0, i;
 464
 465	for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 466		if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
 467			overlap |= 1 << i;
 468	return overlap;
 469}
 470
 471static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
 472{
 473	int i;
 474
 475	if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
 476		return overlap;
 477
 478	for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 479		if (overlap & 1 << i)
 480			radix_tree_tag_set(&dma_active_cacheline, cln, i);
 481		else
 482			radix_tree_tag_clear(&dma_active_cacheline, cln, i);
 483
 484	return overlap;
 485}
 486
 487static void active_cacheline_inc_overlap(phys_addr_t cln)
 488{
 489	int overlap = active_cacheline_read_overlap(cln);
 490
 491	overlap = active_cacheline_set_overlap(cln, ++overlap);
 492
 493	/* If we overflowed the overlap counter then we're potentially
 494	 * leaking dma-mappings.
 495	 */
 496	WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
 497		  pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
 498		  ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 499}
 500
 501static int active_cacheline_dec_overlap(phys_addr_t cln)
 502{
 503	int overlap = active_cacheline_read_overlap(cln);
 504
 505	return active_cacheline_set_overlap(cln, --overlap);
 506}
 507
 508static int active_cacheline_insert(struct dma_debug_entry *entry)
 509{
 510	phys_addr_t cln = to_cacheline_number(entry);
 511	unsigned long flags;
 512	int rc;
 513
 514	/* If the device is not writing memory then we don't have any
 515	 * concerns about the cpu consuming stale data.  This mitigates
 516	 * legitimate usages of overlapping mappings.
 517	 */
 518	if (entry->direction == DMA_TO_DEVICE)
 519		return 0;
 520
 521	spin_lock_irqsave(&radix_lock, flags);
 522	rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
 523	if (rc == -EEXIST)
 524		active_cacheline_inc_overlap(cln);
 525	spin_unlock_irqrestore(&radix_lock, flags);
 526
 527	return rc;
 528}
 529
 530static void active_cacheline_remove(struct dma_debug_entry *entry)
 531{
 532	phys_addr_t cln = to_cacheline_number(entry);
 533	unsigned long flags;
 534
 535	/* ...mirror the insert case */
 536	if (entry->direction == DMA_TO_DEVICE)
 537		return;
 538
 539	spin_lock_irqsave(&radix_lock, flags);
 540	/* since we are counting overlaps the final put of the
 541	 * cacheline will occur when the overlap count is 0.
 542	 * active_cacheline_dec_overlap() returns -1 in that case
 543	 */
 544	if (active_cacheline_dec_overlap(cln) < 0)
 545		radix_tree_delete(&dma_active_cacheline, cln);
 546	spin_unlock_irqrestore(&radix_lock, flags);
 547}
 548
 549/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 550 * Wrapper function for adding an entry to the hash.
 551 * This function takes care of locking itself.
 552 */
 553static void add_dma_entry(struct dma_debug_entry *entry, unsigned long attrs)
 554{
 555	struct hash_bucket *bucket;
 556	unsigned long flags;
 557	int rc;
 558
 559	bucket = get_hash_bucket(entry, &flags);
 560	hash_bucket_add(bucket, entry);
 561	put_hash_bucket(bucket, flags);
 562
 563	rc = active_cacheline_insert(entry);
 564	if (rc == -ENOMEM) {
 565		pr_err_once("cacheline tracking ENOMEM, dma-debug disabled\n");
 566		global_disable = true;
 567	} else if (rc == -EEXIST && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
 568		err_printk(entry->dev, entry,
 569			"cacheline tracking EEXIST, overlapping mappings aren't supported\n");
 570	}
 571}
 572
 573static int dma_debug_create_entries(gfp_t gfp)
 574{
 575	struct dma_debug_entry *entry;
 576	int i;
 577
 578	entry = (void *)get_zeroed_page(gfp);
 579	if (!entry)
 580		return -ENOMEM;
 581
 582	for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
 583		list_add_tail(&entry[i].list, &free_entries);
 584
 585	num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 586	nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 587
 588	return 0;
 589}
 590
 591static struct dma_debug_entry *__dma_entry_alloc(void)
 592{
 593	struct dma_debug_entry *entry;
 594
 595	entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 596	list_del(&entry->list);
 597	memset(entry, 0, sizeof(*entry));
 598
 599	num_free_entries -= 1;
 600	if (num_free_entries < min_free_entries)
 601		min_free_entries = num_free_entries;
 602
 603	return entry;
 604}
 605
 606static void __dma_entry_alloc_check_leak(void)
 
 
 
 
 607{
 608	u32 tmp = nr_total_entries % nr_prealloc_entries;
 609
 610	/* Shout each time we tick over some multiple of the initial pool */
 611	if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
 612		pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
 613			nr_total_entries,
 614			(nr_total_entries / nr_prealloc_entries));
 615	}
 616}
 617
 618/* struct dma_entry allocator
 619 *
 620 * The next two functions implement the allocator for
 621 * struct dma_debug_entries.
 622 */
 623static struct dma_debug_entry *dma_entry_alloc(void)
 624{
 
 625	struct dma_debug_entry *entry;
 626	unsigned long flags;
 
 627
 628	spin_lock_irqsave(&free_entries_lock, flags);
 629	if (num_free_entries == 0) {
 630		if (dma_debug_create_entries(GFP_ATOMIC)) {
 631			global_disable = true;
 632			spin_unlock_irqrestore(&free_entries_lock, flags);
 633			pr_err("debugging out of memory - disabling\n");
 634			return NULL;
 635		}
 636		__dma_entry_alloc_check_leak();
 
 637	}
 638
 639	entry = __dma_entry_alloc();
 640
 641	spin_unlock_irqrestore(&free_entries_lock, flags);
 642
 
 
 
 643#ifdef CONFIG_STACKTRACE
 644	entry->stack_len = stack_trace_save(entry->stack_entries,
 645					    ARRAY_SIZE(entry->stack_entries),
 646					    1);
 647#endif
 648	return entry;
 649}
 650
 651static void dma_entry_free(struct dma_debug_entry *entry)
 652{
 653	unsigned long flags;
 654
 655	active_cacheline_remove(entry);
 656
 657	/*
 658	 * add to beginning of the list - this way the entries are
 659	 * more likely cache hot when they are reallocated.
 660	 */
 661	spin_lock_irqsave(&free_entries_lock, flags);
 662	list_add(&entry->list, &free_entries);
 663	num_free_entries += 1;
 664	spin_unlock_irqrestore(&free_entries_lock, flags);
 665}
 666
 667/*
 668 * DMA-API debugging init code
 669 *
 670 * The init code does two things:
 671 *   1. Initialize core data structures
 672 *   2. Preallocate a given number of dma_debug_entry structs
 673 */
 674
 675static ssize_t filter_read(struct file *file, char __user *user_buf,
 676			   size_t count, loff_t *ppos)
 677{
 678	char buf[NAME_MAX_LEN + 1];
 679	unsigned long flags;
 680	int len;
 681
 682	if (!current_driver_name[0])
 683		return 0;
 684
 685	/*
 686	 * We can't copy to userspace directly because current_driver_name can
 687	 * only be read under the driver_name_lock with irqs disabled. So
 688	 * create a temporary copy first.
 689	 */
 690	read_lock_irqsave(&driver_name_lock, flags);
 691	len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 692	read_unlock_irqrestore(&driver_name_lock, flags);
 693
 694	return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 695}
 696
 697static ssize_t filter_write(struct file *file, const char __user *userbuf,
 698			    size_t count, loff_t *ppos)
 699{
 700	char buf[NAME_MAX_LEN];
 701	unsigned long flags;
 702	size_t len;
 703	int i;
 704
 705	/*
 706	 * We can't copy from userspace directly. Access to
 707	 * current_driver_name is protected with a write_lock with irqs
 708	 * disabled. Since copy_from_user can fault and may sleep we
 709	 * need to copy to temporary buffer first
 710	 */
 711	len = min(count, (size_t)(NAME_MAX_LEN - 1));
 712	if (copy_from_user(buf, userbuf, len))
 713		return -EFAULT;
 714
 715	buf[len] = 0;
 716
 717	write_lock_irqsave(&driver_name_lock, flags);
 718
 719	/*
 720	 * Now handle the string we got from userspace very carefully.
 721	 * The rules are:
 722	 *         - only use the first token we got
 723	 *         - token delimiter is everything looking like a space
 724	 *           character (' ', '\n', '\t' ...)
 725	 *
 726	 */
 727	if (!isalnum(buf[0])) {
 728		/*
 729		 * If the first character userspace gave us is not
 730		 * alphanumerical then assume the filter should be
 731		 * switched off.
 732		 */
 733		if (current_driver_name[0])
 734			pr_info("switching off dma-debug driver filter\n");
 735		current_driver_name[0] = 0;
 736		current_driver = NULL;
 737		goto out_unlock;
 738	}
 739
 740	/*
 741	 * Now parse out the first token and use it as the name for the
 742	 * driver to filter for.
 743	 */
 744	for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 745		current_driver_name[i] = buf[i];
 746		if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 747			break;
 748	}
 749	current_driver_name[i] = 0;
 750	current_driver = NULL;
 751
 752	pr_info("enable driver filter for driver [%s]\n",
 753		current_driver_name);
 754
 755out_unlock:
 756	write_unlock_irqrestore(&driver_name_lock, flags);
 757
 758	return count;
 759}
 760
 761static const struct file_operations filter_fops = {
 762	.read  = filter_read,
 763	.write = filter_write,
 764	.llseek = default_llseek,
 765};
 766
 767static int dump_show(struct seq_file *seq, void *v)
 768{
 769	int idx;
 770
 771	for (idx = 0; idx < HASH_SIZE; idx++) {
 772		struct hash_bucket *bucket = &dma_entry_hash[idx];
 773		struct dma_debug_entry *entry;
 774		unsigned long flags;
 775
 776		spin_lock_irqsave(&bucket->lock, flags);
 777		list_for_each_entry(entry, &bucket->list, list) {
 778			seq_printf(seq,
 779				   "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx %s %s\n",
 780				   dev_name(entry->dev),
 781				   dev_driver_string(entry->dev),
 782				   type2name[entry->type], idx,
 783				   phys_addr(entry), entry->pfn,
 784				   entry->dev_addr, entry->size,
 785				   dir2name[entry->direction],
 786				   maperr2str[entry->map_err_type]);
 787		}
 788		spin_unlock_irqrestore(&bucket->lock, flags);
 789	}
 790	return 0;
 791}
 792DEFINE_SHOW_ATTRIBUTE(dump);
 793
 794static int __init dma_debug_fs_init(void)
 795{
 796	struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
 797
 798	debugfs_create_bool("disabled", 0444, dentry, &global_disable);
 799	debugfs_create_u32("error_count", 0444, dentry, &error_count);
 800	debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
 801	debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
 802	debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
 803	debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
 804	debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
 805	debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
 806	debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
 807
 808	return 0;
 809}
 810core_initcall_sync(dma_debug_fs_init);
 811
 812static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 813{
 814	struct dma_debug_entry *entry;
 815	unsigned long flags;
 816	int count = 0, i;
 817
 818	for (i = 0; i < HASH_SIZE; ++i) {
 819		spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
 820		list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 821			if (entry->dev == dev) {
 822				count += 1;
 823				*out_entry = entry;
 824			}
 825		}
 826		spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
 827	}
 828
 829	return count;
 830}
 831
 832static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 833{
 834	struct device *dev = data;
 835	struct dma_debug_entry *entry;
 836	int count;
 837
 838	if (dma_debug_disabled())
 839		return 0;
 840
 841	switch (action) {
 842	case BUS_NOTIFY_UNBOUND_DRIVER:
 843		count = device_dma_allocations(dev, &entry);
 844		if (count == 0)
 845			break;
 846		err_printk(dev, entry, "device driver has pending "
 847				"DMA allocations while released from device "
 848				"[count=%d]\n"
 849				"One of leaked entries details: "
 850				"[device address=0x%016llx] [size=%llu bytes] "
 851				"[mapped with %s] [mapped as %s]\n",
 852			count, entry->dev_addr, entry->size,
 853			dir2name[entry->direction], type2name[entry->type]);
 854		break;
 855	default:
 856		break;
 857	}
 858
 859	return 0;
 860}
 861
 862void dma_debug_add_bus(struct bus_type *bus)
 863{
 864	struct notifier_block *nb;
 865
 866	if (dma_debug_disabled())
 867		return;
 868
 869	nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 870	if (nb == NULL) {
 871		pr_err("dma_debug_add_bus: out of memory\n");
 872		return;
 873	}
 874
 875	nb->notifier_call = dma_debug_device_change;
 876
 877	bus_register_notifier(bus, nb);
 878}
 879
 880static int dma_debug_init(void)
 881{
 882	int i, nr_pages;
 883
 884	/* Do not use dma_debug_initialized here, since we really want to be
 885	 * called to set dma_debug_initialized
 886	 */
 887	if (global_disable)
 888		return 0;
 889
 890	for (i = 0; i < HASH_SIZE; ++i) {
 891		INIT_LIST_HEAD(&dma_entry_hash[i].list);
 892		spin_lock_init(&dma_entry_hash[i].lock);
 893	}
 894
 895	nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
 896	for (i = 0; i < nr_pages; ++i)
 897		dma_debug_create_entries(GFP_KERNEL);
 898	if (num_free_entries >= nr_prealloc_entries) {
 899		pr_info("preallocated %d debug entries\n", nr_total_entries);
 900	} else if (num_free_entries > 0) {
 901		pr_warn("%d debug entries requested but only %d allocated\n",
 902			nr_prealloc_entries, nr_total_entries);
 903	} else {
 904		pr_err("debugging out of memory error - disabled\n");
 905		global_disable = true;
 906
 907		return 0;
 908	}
 909	min_free_entries = num_free_entries;
 910
 911	dma_debug_initialized = true;
 912
 913	pr_info("debugging enabled by kernel config\n");
 914	return 0;
 915}
 916core_initcall(dma_debug_init);
 917
 918static __init int dma_debug_cmdline(char *str)
 919{
 920	if (!str)
 921		return -EINVAL;
 922
 923	if (strncmp(str, "off", 3) == 0) {
 924		pr_info("debugging disabled on kernel command line\n");
 925		global_disable = true;
 926	}
 927
 928	return 1;
 929}
 930
 931static __init int dma_debug_entries_cmdline(char *str)
 932{
 933	if (!str)
 934		return -EINVAL;
 935	if (!get_option(&str, &nr_prealloc_entries))
 936		nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 937	return 1;
 938}
 939
 940__setup("dma_debug=", dma_debug_cmdline);
 941__setup("dma_debug_entries=", dma_debug_entries_cmdline);
 942
 943static void check_unmap(struct dma_debug_entry *ref)
 944{
 945	struct dma_debug_entry *entry;
 946	struct hash_bucket *bucket;
 947	unsigned long flags;
 948
 949	bucket = get_hash_bucket(ref, &flags);
 950	entry = bucket_find_exact(bucket, ref);
 951
 952	if (!entry) {
 953		/* must drop lock before calling dma_mapping_error */
 954		put_hash_bucket(bucket, flags);
 955
 956		if (dma_mapping_error(ref->dev, ref->dev_addr)) {
 957			err_printk(ref->dev, NULL,
 958				   "device driver tries to free an "
 959				   "invalid DMA memory address\n");
 960		} else {
 961			err_printk(ref->dev, NULL,
 962				   "device driver tries to free DMA "
 963				   "memory it has not allocated [device "
 964				   "address=0x%016llx] [size=%llu bytes]\n",
 965				   ref->dev_addr, ref->size);
 966		}
 967		return;
 968	}
 969
 970	if (ref->size != entry->size) {
 971		err_printk(ref->dev, entry, "device driver frees "
 972			   "DMA memory with different size "
 973			   "[device address=0x%016llx] [map size=%llu bytes] "
 974			   "[unmap size=%llu bytes]\n",
 975			   ref->dev_addr, entry->size, ref->size);
 976	}
 977
 978	if (ref->type != entry->type) {
 979		err_printk(ref->dev, entry, "device driver frees "
 980			   "DMA memory with wrong function "
 981			   "[device address=0x%016llx] [size=%llu bytes] "
 982			   "[mapped as %s] [unmapped as %s]\n",
 983			   ref->dev_addr, ref->size,
 984			   type2name[entry->type], type2name[ref->type]);
 985	} else if ((entry->type == dma_debug_coherent) &&
 986		   (phys_addr(ref) != phys_addr(entry))) {
 987		err_printk(ref->dev, entry, "device driver frees "
 988			   "DMA memory with different CPU address "
 989			   "[device address=0x%016llx] [size=%llu bytes] "
 990			   "[cpu alloc address=0x%016llx] "
 991			   "[cpu free address=0x%016llx]",
 992			   ref->dev_addr, ref->size,
 993			   phys_addr(entry),
 994			   phys_addr(ref));
 995	}
 996
 997	if (ref->sg_call_ents && ref->type == dma_debug_sg &&
 998	    ref->sg_call_ents != entry->sg_call_ents) {
 999		err_printk(ref->dev, entry, "device driver frees "
1000			   "DMA sg list with different entry count "
1001			   "[map count=%d] [unmap count=%d]\n",
1002			   entry->sg_call_ents, ref->sg_call_ents);
1003	}
1004
1005	/*
1006	 * This may be no bug in reality - but most implementations of the
1007	 * DMA API don't handle this properly, so check for it here
1008	 */
1009	if (ref->direction != entry->direction) {
1010		err_printk(ref->dev, entry, "device driver frees "
1011			   "DMA memory with different direction "
1012			   "[device address=0x%016llx] [size=%llu bytes] "
1013			   "[mapped with %s] [unmapped with %s]\n",
1014			   ref->dev_addr, ref->size,
1015			   dir2name[entry->direction],
1016			   dir2name[ref->direction]);
1017	}
1018
1019	/*
1020	 * Drivers should use dma_mapping_error() to check the returned
1021	 * addresses of dma_map_single() and dma_map_page().
1022	 * If not, print this warning message. See Documentation/core-api/dma-api.rst.
1023	 */
1024	if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1025		err_printk(ref->dev, entry,
1026			   "device driver failed to check map error"
1027			   "[device address=0x%016llx] [size=%llu bytes] "
1028			   "[mapped as %s]",
1029			   ref->dev_addr, ref->size,
1030			   type2name[entry->type]);
1031	}
1032
1033	hash_bucket_del(entry);
1034	dma_entry_free(entry);
1035
1036	put_hash_bucket(bucket, flags);
1037}
1038
1039static void check_for_stack(struct device *dev,
1040			    struct page *page, size_t offset)
1041{
1042	void *addr;
1043	struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1044
1045	if (!stack_vm_area) {
1046		/* Stack is direct-mapped. */
1047		if (PageHighMem(page))
1048			return;
1049		addr = page_address(page) + offset;
1050		if (object_is_on_stack(addr))
1051			err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1052	} else {
1053		/* Stack is vmalloced. */
1054		int i;
1055
1056		for (i = 0; i < stack_vm_area->nr_pages; i++) {
1057			if (page != stack_vm_area->pages[i])
1058				continue;
1059
1060			addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1061			err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1062			break;
1063		}
1064	}
1065}
1066
1067static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1068{
1069	if (memory_intersects(_stext, _etext, addr, len) ||
1070	    memory_intersects(__start_rodata, __end_rodata, addr, len))
1071		err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1072}
1073
1074static void check_sync(struct device *dev,
1075		       struct dma_debug_entry *ref,
1076		       bool to_cpu)
1077{
1078	struct dma_debug_entry *entry;
1079	struct hash_bucket *bucket;
1080	unsigned long flags;
1081
1082	bucket = get_hash_bucket(ref, &flags);
1083
1084	entry = bucket_find_contain(&bucket, ref, &flags);
1085
1086	if (!entry) {
1087		err_printk(dev, NULL, "device driver tries "
1088				"to sync DMA memory it has not allocated "
1089				"[device address=0x%016llx] [size=%llu bytes]\n",
1090				(unsigned long long)ref->dev_addr, ref->size);
1091		goto out;
1092	}
1093
1094	if (ref->size > entry->size) {
1095		err_printk(dev, entry, "device driver syncs"
1096				" DMA memory outside allocated range "
1097				"[device address=0x%016llx] "
1098				"[allocation size=%llu bytes] "
1099				"[sync offset+size=%llu]\n",
1100				entry->dev_addr, entry->size,
1101				ref->size);
1102	}
1103
1104	if (entry->direction == DMA_BIDIRECTIONAL)
1105		goto out;
1106
1107	if (ref->direction != entry->direction) {
1108		err_printk(dev, entry, "device driver syncs "
1109				"DMA memory with different direction "
1110				"[device address=0x%016llx] [size=%llu bytes] "
1111				"[mapped with %s] [synced with %s]\n",
1112				(unsigned long long)ref->dev_addr, entry->size,
1113				dir2name[entry->direction],
1114				dir2name[ref->direction]);
1115	}
1116
1117	if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1118		      !(ref->direction == DMA_TO_DEVICE))
1119		err_printk(dev, entry, "device driver syncs "
1120				"device read-only DMA memory for cpu "
1121				"[device address=0x%016llx] [size=%llu bytes] "
1122				"[mapped with %s] [synced with %s]\n",
1123				(unsigned long long)ref->dev_addr, entry->size,
1124				dir2name[entry->direction],
1125				dir2name[ref->direction]);
1126
1127	if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1128		       !(ref->direction == DMA_FROM_DEVICE))
1129		err_printk(dev, entry, "device driver syncs "
1130				"device write-only DMA memory to device "
1131				"[device address=0x%016llx] [size=%llu bytes] "
1132				"[mapped with %s] [synced with %s]\n",
1133				(unsigned long long)ref->dev_addr, entry->size,
1134				dir2name[entry->direction],
1135				dir2name[ref->direction]);
1136
1137	if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1138	    ref->sg_call_ents != entry->sg_call_ents) {
1139		err_printk(ref->dev, entry, "device driver syncs "
1140			   "DMA sg list with different entry count "
1141			   "[map count=%d] [sync count=%d]\n",
1142			   entry->sg_call_ents, ref->sg_call_ents);
1143	}
1144
1145out:
1146	put_hash_bucket(bucket, flags);
1147}
1148
1149static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1150{
1151#ifdef CONFIG_DMA_API_DEBUG_SG
1152	unsigned int max_seg = dma_get_max_seg_size(dev);
1153	u64 start, end, boundary = dma_get_seg_boundary(dev);
1154
1155	/*
1156	 * Either the driver forgot to set dma_parms appropriately, or
1157	 * whoever generated the list forgot to check them.
1158	 */
1159	if (sg->length > max_seg)
1160		err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1161			   sg->length, max_seg);
1162	/*
1163	 * In some cases this could potentially be the DMA API
1164	 * implementation's fault, but it would usually imply that
1165	 * the scatterlist was built inappropriately to begin with.
1166	 */
1167	start = sg_dma_address(sg);
1168	end = start + sg_dma_len(sg) - 1;
1169	if ((start ^ end) & ~boundary)
1170		err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1171			   start, end, boundary);
1172#endif
1173}
1174
1175void debug_dma_map_single(struct device *dev, const void *addr,
1176			    unsigned long len)
1177{
1178	if (unlikely(dma_debug_disabled()))
1179		return;
1180
1181	if (!virt_addr_valid(addr))
1182		err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1183			   addr, len);
1184
1185	if (is_vmalloc_addr(addr))
1186		err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1187			   addr, len);
1188}
1189EXPORT_SYMBOL(debug_dma_map_single);
1190
1191void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1192			size_t size, int direction, dma_addr_t dma_addr,
1193			unsigned long attrs)
1194{
1195	struct dma_debug_entry *entry;
1196
1197	if (unlikely(dma_debug_disabled()))
1198		return;
1199
1200	if (dma_mapping_error(dev, dma_addr))
1201		return;
1202
1203	entry = dma_entry_alloc();
1204	if (!entry)
1205		return;
1206
1207	entry->dev       = dev;
1208	entry->type      = dma_debug_single;
1209	entry->pfn	 = page_to_pfn(page);
1210	entry->offset	 = offset;
1211	entry->dev_addr  = dma_addr;
1212	entry->size      = size;
1213	entry->direction = direction;
1214	entry->map_err_type = MAP_ERR_NOT_CHECKED;
1215
1216	check_for_stack(dev, page, offset);
1217
1218	if (!PageHighMem(page)) {
1219		void *addr = page_address(page) + offset;
1220
1221		check_for_illegal_area(dev, addr, size);
1222	}
1223
1224	add_dma_entry(entry, attrs);
1225}
1226
1227void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1228{
1229	struct dma_debug_entry ref;
1230	struct dma_debug_entry *entry;
1231	struct hash_bucket *bucket;
1232	unsigned long flags;
1233
1234	if (unlikely(dma_debug_disabled()))
1235		return;
1236
1237	ref.dev = dev;
1238	ref.dev_addr = dma_addr;
1239	bucket = get_hash_bucket(&ref, &flags);
1240
1241	list_for_each_entry(entry, &bucket->list, list) {
1242		if (!exact_match(&ref, entry))
1243			continue;
1244
1245		/*
1246		 * The same physical address can be mapped multiple
1247		 * times. Without a hardware IOMMU this results in the
1248		 * same device addresses being put into the dma-debug
1249		 * hash multiple times too. This can result in false
1250		 * positives being reported. Therefore we implement a
1251		 * best-fit algorithm here which updates the first entry
1252		 * from the hash which fits the reference value and is
1253		 * not currently listed as being checked.
1254		 */
1255		if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1256			entry->map_err_type = MAP_ERR_CHECKED;
1257			break;
1258		}
1259	}
1260
1261	put_hash_bucket(bucket, flags);
1262}
1263EXPORT_SYMBOL(debug_dma_mapping_error);
1264
1265void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1266			  size_t size, int direction)
1267{
1268	struct dma_debug_entry ref = {
1269		.type           = dma_debug_single,
1270		.dev            = dev,
1271		.dev_addr       = addr,
1272		.size           = size,
1273		.direction      = direction,
1274	};
1275
1276	if (unlikely(dma_debug_disabled()))
1277		return;
1278	check_unmap(&ref);
1279}
1280
1281void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1282		      int nents, int mapped_ents, int direction,
1283		      unsigned long attrs)
1284{
1285	struct dma_debug_entry *entry;
1286	struct scatterlist *s;
1287	int i;
1288
1289	if (unlikely(dma_debug_disabled()))
1290		return;
1291
1292	for_each_sg(sg, s, nents, i) {
1293		check_for_stack(dev, sg_page(s), s->offset);
1294		if (!PageHighMem(sg_page(s)))
1295			check_for_illegal_area(dev, sg_virt(s), s->length);
1296	}
1297
1298	for_each_sg(sg, s, mapped_ents, i) {
1299		entry = dma_entry_alloc();
1300		if (!entry)
1301			return;
1302
1303		entry->type           = dma_debug_sg;
1304		entry->dev            = dev;
1305		entry->pfn	      = page_to_pfn(sg_page(s));
1306		entry->offset	      = s->offset;
1307		entry->size           = sg_dma_len(s);
1308		entry->dev_addr       = sg_dma_address(s);
1309		entry->direction      = direction;
1310		entry->sg_call_ents   = nents;
1311		entry->sg_mapped_ents = mapped_ents;
1312
1313		check_sg_segment(dev, s);
1314
1315		add_dma_entry(entry, attrs);
1316	}
1317}
1318
1319static int get_nr_mapped_entries(struct device *dev,
1320				 struct dma_debug_entry *ref)
1321{
1322	struct dma_debug_entry *entry;
1323	struct hash_bucket *bucket;
1324	unsigned long flags;
1325	int mapped_ents;
1326
1327	bucket       = get_hash_bucket(ref, &flags);
1328	entry        = bucket_find_exact(bucket, ref);
1329	mapped_ents  = 0;
1330
1331	if (entry)
1332		mapped_ents = entry->sg_mapped_ents;
1333	put_hash_bucket(bucket, flags);
1334
1335	return mapped_ents;
1336}
1337
1338void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1339			int nelems, int dir)
1340{
1341	struct scatterlist *s;
1342	int mapped_ents = 0, i;
1343
1344	if (unlikely(dma_debug_disabled()))
1345		return;
1346
1347	for_each_sg(sglist, s, nelems, i) {
1348
1349		struct dma_debug_entry ref = {
1350			.type           = dma_debug_sg,
1351			.dev            = dev,
1352			.pfn		= page_to_pfn(sg_page(s)),
1353			.offset		= s->offset,
1354			.dev_addr       = sg_dma_address(s),
1355			.size           = sg_dma_len(s),
1356			.direction      = dir,
1357			.sg_call_ents   = nelems,
1358		};
1359
1360		if (mapped_ents && i >= mapped_ents)
1361			break;
1362
1363		if (!i)
1364			mapped_ents = get_nr_mapped_entries(dev, &ref);
1365
1366		check_unmap(&ref);
1367	}
1368}
1369
1370void debug_dma_alloc_coherent(struct device *dev, size_t size,
1371			      dma_addr_t dma_addr, void *virt,
1372			      unsigned long attrs)
1373{
1374	struct dma_debug_entry *entry;
1375
1376	if (unlikely(dma_debug_disabled()))
1377		return;
1378
1379	if (unlikely(virt == NULL))
1380		return;
1381
1382	/* handle vmalloc and linear addresses */
1383	if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1384		return;
1385
1386	entry = dma_entry_alloc();
1387	if (!entry)
1388		return;
1389
1390	entry->type      = dma_debug_coherent;
1391	entry->dev       = dev;
1392	entry->offset	 = offset_in_page(virt);
1393	entry->size      = size;
1394	entry->dev_addr  = dma_addr;
1395	entry->direction = DMA_BIDIRECTIONAL;
1396
1397	if (is_vmalloc_addr(virt))
1398		entry->pfn = vmalloc_to_pfn(virt);
1399	else
1400		entry->pfn = page_to_pfn(virt_to_page(virt));
1401
1402	add_dma_entry(entry, attrs);
1403}
1404
1405void debug_dma_free_coherent(struct device *dev, size_t size,
1406			 void *virt, dma_addr_t addr)
1407{
1408	struct dma_debug_entry ref = {
1409		.type           = dma_debug_coherent,
1410		.dev            = dev,
1411		.offset		= offset_in_page(virt),
1412		.dev_addr       = addr,
1413		.size           = size,
1414		.direction      = DMA_BIDIRECTIONAL,
1415	};
1416
1417	/* handle vmalloc and linear addresses */
1418	if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1419		return;
1420
1421	if (is_vmalloc_addr(virt))
1422		ref.pfn = vmalloc_to_pfn(virt);
1423	else
1424		ref.pfn = page_to_pfn(virt_to_page(virt));
1425
1426	if (unlikely(dma_debug_disabled()))
1427		return;
1428
1429	check_unmap(&ref);
1430}
1431
1432void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1433			    int direction, dma_addr_t dma_addr,
1434			    unsigned long attrs)
1435{
1436	struct dma_debug_entry *entry;
1437
1438	if (unlikely(dma_debug_disabled()))
1439		return;
1440
1441	entry = dma_entry_alloc();
1442	if (!entry)
1443		return;
1444
1445	entry->type		= dma_debug_resource;
1446	entry->dev		= dev;
1447	entry->pfn		= PHYS_PFN(addr);
1448	entry->offset		= offset_in_page(addr);
1449	entry->size		= size;
1450	entry->dev_addr		= dma_addr;
1451	entry->direction	= direction;
1452	entry->map_err_type	= MAP_ERR_NOT_CHECKED;
1453
1454	add_dma_entry(entry, attrs);
1455}
1456
1457void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1458			      size_t size, int direction)
1459{
1460	struct dma_debug_entry ref = {
1461		.type           = dma_debug_resource,
1462		.dev            = dev,
1463		.dev_addr       = dma_addr,
1464		.size           = size,
1465		.direction      = direction,
1466	};
1467
1468	if (unlikely(dma_debug_disabled()))
1469		return;
1470
1471	check_unmap(&ref);
1472}
1473
1474void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1475				   size_t size, int direction)
1476{
1477	struct dma_debug_entry ref;
1478
1479	if (unlikely(dma_debug_disabled()))
1480		return;
1481
1482	ref.type         = dma_debug_single;
1483	ref.dev          = dev;
1484	ref.dev_addr     = dma_handle;
1485	ref.size         = size;
1486	ref.direction    = direction;
1487	ref.sg_call_ents = 0;
1488
1489	check_sync(dev, &ref, true);
1490}
1491
1492void debug_dma_sync_single_for_device(struct device *dev,
1493				      dma_addr_t dma_handle, size_t size,
1494				      int direction)
1495{
1496	struct dma_debug_entry ref;
1497
1498	if (unlikely(dma_debug_disabled()))
1499		return;
1500
1501	ref.type         = dma_debug_single;
1502	ref.dev          = dev;
1503	ref.dev_addr     = dma_handle;
1504	ref.size         = size;
1505	ref.direction    = direction;
1506	ref.sg_call_ents = 0;
1507
1508	check_sync(dev, &ref, false);
1509}
1510
1511void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1512			       int nelems, int direction)
1513{
1514	struct scatterlist *s;
1515	int mapped_ents = 0, i;
1516
1517	if (unlikely(dma_debug_disabled()))
1518		return;
1519
1520	for_each_sg(sg, s, nelems, i) {
1521
1522		struct dma_debug_entry ref = {
1523			.type           = dma_debug_sg,
1524			.dev            = dev,
1525			.pfn		= page_to_pfn(sg_page(s)),
1526			.offset		= s->offset,
1527			.dev_addr       = sg_dma_address(s),
1528			.size           = sg_dma_len(s),
1529			.direction      = direction,
1530			.sg_call_ents   = nelems,
1531		};
1532
1533		if (!i)
1534			mapped_ents = get_nr_mapped_entries(dev, &ref);
1535
1536		if (i >= mapped_ents)
1537			break;
1538
1539		check_sync(dev, &ref, true);
1540	}
1541}
1542
1543void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1544				  int nelems, int direction)
1545{
1546	struct scatterlist *s;
1547	int mapped_ents = 0, i;
1548
1549	if (unlikely(dma_debug_disabled()))
1550		return;
1551
1552	for_each_sg(sg, s, nelems, i) {
1553
1554		struct dma_debug_entry ref = {
1555			.type           = dma_debug_sg,
1556			.dev            = dev,
1557			.pfn		= page_to_pfn(sg_page(s)),
1558			.offset		= s->offset,
1559			.dev_addr       = sg_dma_address(s),
1560			.size           = sg_dma_len(s),
1561			.direction      = direction,
1562			.sg_call_ents   = nelems,
1563		};
1564		if (!i)
1565			mapped_ents = get_nr_mapped_entries(dev, &ref);
1566
1567		if (i >= mapped_ents)
1568			break;
1569
1570		check_sync(dev, &ref, false);
1571	}
1572}
1573
1574static int __init dma_debug_driver_setup(char *str)
1575{
1576	int i;
1577
1578	for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1579		current_driver_name[i] = *str;
1580		if (*str == 0)
1581			break;
1582	}
1583
1584	if (current_driver_name[0])
1585		pr_info("enable driver filter for driver [%s]\n",
1586			current_driver_name);
1587
1588
1589	return 1;
1590}
1591__setup("dma_debug_driver=", dma_debug_driver_setup);
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2008 Advanced Micro Devices, Inc.
   4 *
   5 * Author: Joerg Roedel <joerg.roedel@amd.com>
   6 */
   7
   8#define pr_fmt(fmt)	"DMA-API: " fmt
   9
  10#include <linux/sched/task_stack.h>
  11#include <linux/scatterlist.h>
  12#include <linux/dma-map-ops.h>
  13#include <linux/sched/task.h>
  14#include <linux/stacktrace.h>
  15#include <linux/spinlock.h>
  16#include <linux/vmalloc.h>
  17#include <linux/debugfs.h>
  18#include <linux/uaccess.h>
  19#include <linux/export.h>
  20#include <linux/device.h>
  21#include <linux/types.h>
  22#include <linux/sched.h>
  23#include <linux/ctype.h>
  24#include <linux/list.h>
  25#include <linux/slab.h>
  26#include <asm/sections.h>
  27#include "debug.h"
  28
  29#define HASH_SIZE       16384ULL
  30#define HASH_FN_SHIFT   13
  31#define HASH_FN_MASK    (HASH_SIZE - 1)
  32
  33#define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
  34/* If the pool runs out, add this many new entries at once */
  35#define DMA_DEBUG_DYNAMIC_ENTRIES (PAGE_SIZE / sizeof(struct dma_debug_entry))
  36
  37enum {
  38	dma_debug_single,
  39	dma_debug_sg,
  40	dma_debug_coherent,
  41	dma_debug_resource,
  42};
  43
  44enum map_err_types {
  45	MAP_ERR_CHECK_NOT_APPLICABLE,
  46	MAP_ERR_NOT_CHECKED,
  47	MAP_ERR_CHECKED,
  48};
  49
  50#define DMA_DEBUG_STACKTRACE_ENTRIES 5
  51
  52/**
  53 * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
  54 * @list: node on pre-allocated free_entries list
  55 * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
  56 * @dev_addr: dma address
  57 * @size: length of the mapping
  58 * @type: single, page, sg, coherent
  59 * @direction: enum dma_data_direction
  60 * @sg_call_ents: 'nents' from dma_map_sg
  61 * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
  62 * @pfn: page frame of the start address
  63 * @offset: offset of mapping relative to pfn
  64 * @map_err_type: track whether dma_mapping_error() was checked
  65 * @stack_len: number of backtrace entries in @stack_entries
  66 * @stack_entries: stack of backtrace history
  67 */
  68struct dma_debug_entry {
  69	struct list_head list;
  70	struct device    *dev;
  71	u64              dev_addr;
  72	u64              size;
  73	int              type;
  74	int              direction;
  75	int		 sg_call_ents;
  76	int		 sg_mapped_ents;
  77	unsigned long	 pfn;
  78	size_t		 offset;
  79	enum map_err_types  map_err_type;
  80#ifdef CONFIG_STACKTRACE
  81	unsigned int	stack_len;
  82	unsigned long	stack_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
  83#endif
  84} ____cacheline_aligned_in_smp;
  85
  86typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
  87
  88struct hash_bucket {
  89	struct list_head list;
  90	spinlock_t lock;
  91};
  92
  93/* Hash list to save the allocated dma addresses */
  94static struct hash_bucket dma_entry_hash[HASH_SIZE];
  95/* List of pre-allocated dma_debug_entry's */
  96static LIST_HEAD(free_entries);
  97/* Lock for the list above */
  98static DEFINE_SPINLOCK(free_entries_lock);
  99
 100/* Global disable flag - will be set in case of an error */
 101static bool global_disable __read_mostly;
 102
 103/* Early initialization disable flag, set at the end of dma_debug_init */
 104static bool dma_debug_initialized __read_mostly;
 105
 106static inline bool dma_debug_disabled(void)
 107{
 108	return global_disable || !dma_debug_initialized;
 109}
 110
 111/* Global error count */
 112static u32 error_count;
 113
 114/* Global error show enable*/
 115static u32 show_all_errors __read_mostly;
 116/* Number of errors to show */
 117static u32 show_num_errors = 1;
 118
 119static u32 num_free_entries;
 120static u32 min_free_entries;
 121static u32 nr_total_entries;
 122
 123/* number of preallocated entries requested by kernel cmdline */
 124static u32 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 125
 126/* per-driver filter related state */
 127
 128#define NAME_MAX_LEN	64
 129
 130static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
 131static struct device_driver *current_driver                    __read_mostly;
 132
 133static DEFINE_RWLOCK(driver_name_lock);
 134
 135static const char *const maperr2str[] = {
 136	[MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
 137	[MAP_ERR_NOT_CHECKED] = "dma map error not checked",
 138	[MAP_ERR_CHECKED] = "dma map error checked",
 139};
 140
 141static const char *type2name[] = {
 142	[dma_debug_single] = "single",
 143	[dma_debug_sg] = "scatter-gather",
 144	[dma_debug_coherent] = "coherent",
 145	[dma_debug_resource] = "resource",
 146};
 147
 148static const char *dir2name[] = {
 149	[DMA_BIDIRECTIONAL]	= "DMA_BIDIRECTIONAL",
 150	[DMA_TO_DEVICE]		= "DMA_TO_DEVICE",
 151	[DMA_FROM_DEVICE]	= "DMA_FROM_DEVICE",
 152	[DMA_NONE]		= "DMA_NONE",
 153};
 154
 155/*
 156 * The access to some variables in this macro is racy. We can't use atomic_t
 157 * here because all these variables are exported to debugfs. Some of them even
 158 * writeable. This is also the reason why a lock won't help much. But anyway,
 159 * the races are no big deal. Here is why:
 160 *
 161 *   error_count: the addition is racy, but the worst thing that can happen is
 162 *                that we don't count some errors
 163 *   show_num_errors: the subtraction is racy. Also no big deal because in
 164 *                    worst case this will result in one warning more in the
 165 *                    system log than the user configured. This variable is
 166 *                    writeable via debugfs.
 167 */
 168static inline void dump_entry_trace(struct dma_debug_entry *entry)
 169{
 170#ifdef CONFIG_STACKTRACE
 171	if (entry) {
 172		pr_warn("Mapped at:\n");
 173		stack_trace_print(entry->stack_entries, entry->stack_len, 0);
 174	}
 175#endif
 176}
 177
 178static bool driver_filter(struct device *dev)
 179{
 180	struct device_driver *drv;
 181	unsigned long flags;
 182	bool ret;
 183
 184	/* driver filter off */
 185	if (likely(!current_driver_name[0]))
 186		return true;
 187
 188	/* driver filter on and initialized */
 189	if (current_driver && dev && dev->driver == current_driver)
 190		return true;
 191
 192	/* driver filter on, but we can't filter on a NULL device... */
 193	if (!dev)
 194		return false;
 195
 196	if (current_driver || !current_driver_name[0])
 197		return false;
 198
 199	/* driver filter on but not yet initialized */
 200	drv = dev->driver;
 201	if (!drv)
 202		return false;
 203
 204	/* lock to protect against change of current_driver_name */
 205	read_lock_irqsave(&driver_name_lock, flags);
 206
 207	ret = false;
 208	if (drv->name &&
 209	    strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 210		current_driver = drv;
 211		ret = true;
 212	}
 213
 214	read_unlock_irqrestore(&driver_name_lock, flags);
 215
 216	return ret;
 217}
 218
 219#define err_printk(dev, entry, format, arg...) do {			\
 220		error_count += 1;					\
 221		if (driver_filter(dev) &&				\
 222		    (show_all_errors || show_num_errors > 0)) {		\
 223			WARN(1, pr_fmt("%s %s: ") format,		\
 224			     dev ? dev_driver_string(dev) : "NULL",	\
 225			     dev ? dev_name(dev) : "NULL", ## arg);	\
 226			dump_entry_trace(entry);			\
 227		}							\
 228		if (!show_all_errors && show_num_errors > 0)		\
 229			show_num_errors -= 1;				\
 230	} while (0);
 231
 232/*
 233 * Hash related functions
 234 *
 235 * Every DMA-API request is saved into a struct dma_debug_entry. To
 236 * have quick access to these structs they are stored into a hash.
 237 */
 238static int hash_fn(struct dma_debug_entry *entry)
 239{
 240	/*
 241	 * Hash function is based on the dma address.
 242	 * We use bits 20-27 here as the index into the hash
 243	 */
 244	return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 245}
 246
 247/*
 248 * Request exclusive access to a hash bucket for a given dma_debug_entry.
 249 */
 250static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 251					   unsigned long *flags)
 252	__acquires(&dma_entry_hash[idx].lock)
 253{
 254	int idx = hash_fn(entry);
 255	unsigned long __flags;
 256
 257	spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 258	*flags = __flags;
 259	return &dma_entry_hash[idx];
 260}
 261
 262/*
 263 * Give up exclusive access to the hash bucket
 264 */
 265static void put_hash_bucket(struct hash_bucket *bucket,
 266			    unsigned long flags)
 267	__releases(&bucket->lock)
 268{
 269	spin_unlock_irqrestore(&bucket->lock, flags);
 270}
 271
 272static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
 273{
 274	return ((a->dev_addr == b->dev_addr) &&
 275		(a->dev == b->dev)) ? true : false;
 276}
 277
 278static bool containing_match(struct dma_debug_entry *a,
 279			     struct dma_debug_entry *b)
 280{
 281	if (a->dev != b->dev)
 282		return false;
 283
 284	if ((b->dev_addr <= a->dev_addr) &&
 285	    ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
 286		return true;
 287
 288	return false;
 289}
 290
 291/*
 292 * Search a given entry in the hash bucket list
 293 */
 294static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
 295						  struct dma_debug_entry *ref,
 296						  match_fn match)
 297{
 298	struct dma_debug_entry *entry, *ret = NULL;
 299	int matches = 0, match_lvl, last_lvl = -1;
 300
 301	list_for_each_entry(entry, &bucket->list, list) {
 302		if (!match(ref, entry))
 303			continue;
 304
 305		/*
 306		 * Some drivers map the same physical address multiple
 307		 * times. Without a hardware IOMMU this results in the
 308		 * same device addresses being put into the dma-debug
 309		 * hash multiple times too. This can result in false
 310		 * positives being reported. Therefore we implement a
 311		 * best-fit algorithm here which returns the entry from
 312		 * the hash which fits best to the reference value
 313		 * instead of the first-fit.
 314		 */
 315		matches += 1;
 316		match_lvl = 0;
 317		entry->size         == ref->size         ? ++match_lvl : 0;
 318		entry->type         == ref->type         ? ++match_lvl : 0;
 319		entry->direction    == ref->direction    ? ++match_lvl : 0;
 320		entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
 321
 322		if (match_lvl == 4) {
 323			/* perfect-fit - return the result */
 324			return entry;
 325		} else if (match_lvl > last_lvl) {
 326			/*
 327			 * We found an entry that fits better then the
 328			 * previous one or it is the 1st match.
 329			 */
 330			last_lvl = match_lvl;
 331			ret      = entry;
 332		}
 333	}
 334
 335	/*
 336	 * If we have multiple matches but no perfect-fit, just return
 337	 * NULL.
 338	 */
 339	ret = (matches == 1) ? ret : NULL;
 340
 341	return ret;
 342}
 343
 344static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
 345						 struct dma_debug_entry *ref)
 346{
 347	return __hash_bucket_find(bucket, ref, exact_match);
 348}
 349
 350static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 351						   struct dma_debug_entry *ref,
 352						   unsigned long *flags)
 353{
 354
 355	struct dma_debug_entry *entry, index = *ref;
 356	int limit = min(HASH_SIZE, (index.dev_addr >> HASH_FN_SHIFT) + 1);
 357
 358	for (int i = 0; i < limit; i++) {
 359		entry = __hash_bucket_find(*bucket, ref, containing_match);
 360
 361		if (entry)
 362			return entry;
 363
 364		/*
 365		 * Nothing found, go back a hash bucket
 366		 */
 367		put_hash_bucket(*bucket, *flags);
 368		index.dev_addr -= (1 << HASH_FN_SHIFT);
 369		*bucket = get_hash_bucket(&index, flags);
 370	}
 371
 372	return NULL;
 373}
 374
 375/*
 376 * Add an entry to a hash bucket
 377 */
 378static void hash_bucket_add(struct hash_bucket *bucket,
 379			    struct dma_debug_entry *entry)
 380{
 381	list_add_tail(&entry->list, &bucket->list);
 382}
 383
 384/*
 385 * Remove entry from a hash bucket list
 386 */
 387static void hash_bucket_del(struct dma_debug_entry *entry)
 388{
 389	list_del(&entry->list);
 390}
 391
 392static unsigned long long phys_addr(struct dma_debug_entry *entry)
 393{
 394	if (entry->type == dma_debug_resource)
 395		return __pfn_to_phys(entry->pfn) + entry->offset;
 396
 397	return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
 398}
 399
 400/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 401 * For each mapping (initial cacheline in the case of
 402 * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
 403 * scatterlist, or the cacheline specified in dma_map_single) insert
 404 * into this tree using the cacheline as the key. At
 405 * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
 406 * the entry already exists at insertion time add a tag as a reference
 407 * count for the overlapping mappings.  For now, the overlap tracking
 408 * just ensures that 'unmaps' balance 'maps' before marking the
 409 * cacheline idle, but we should also be flagging overlaps as an API
 410 * violation.
 411 *
 412 * Memory usage is mostly constrained by the maximum number of available
 413 * dma-debug entries in that we need a free dma_debug_entry before
 414 * inserting into the tree.  In the case of dma_map_page and
 415 * dma_alloc_coherent there is only one dma_debug_entry and one
 416 * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
 417 * other hand, consumes a single dma_debug_entry, but inserts 'nents'
 418 * entries into the tree.
 419 */
 420static RADIX_TREE(dma_active_cacheline, GFP_ATOMIC);
 421static DEFINE_SPINLOCK(radix_lock);
 422#define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
 423#define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
 424#define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
 425
 426static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
 427{
 428	return (entry->pfn << CACHELINE_PER_PAGE_SHIFT) +
 429		(entry->offset >> L1_CACHE_SHIFT);
 430}
 431
 432static int active_cacheline_read_overlap(phys_addr_t cln)
 433{
 434	int overlap = 0, i;
 435
 436	for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 437		if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
 438			overlap |= 1 << i;
 439	return overlap;
 440}
 441
 442static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
 443{
 444	int i;
 445
 446	if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
 447		return overlap;
 448
 449	for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 450		if (overlap & 1 << i)
 451			radix_tree_tag_set(&dma_active_cacheline, cln, i);
 452		else
 453			radix_tree_tag_clear(&dma_active_cacheline, cln, i);
 454
 455	return overlap;
 456}
 457
 458static void active_cacheline_inc_overlap(phys_addr_t cln)
 459{
 460	int overlap = active_cacheline_read_overlap(cln);
 461
 462	overlap = active_cacheline_set_overlap(cln, ++overlap);
 463
 464	/* If we overflowed the overlap counter then we're potentially
 465	 * leaking dma-mappings.
 466	 */
 467	WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
 468		  pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
 469		  ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 470}
 471
 472static int active_cacheline_dec_overlap(phys_addr_t cln)
 473{
 474	int overlap = active_cacheline_read_overlap(cln);
 475
 476	return active_cacheline_set_overlap(cln, --overlap);
 477}
 478
 479static int active_cacheline_insert(struct dma_debug_entry *entry)
 480{
 481	phys_addr_t cln = to_cacheline_number(entry);
 482	unsigned long flags;
 483	int rc;
 484
 485	/* If the device is not writing memory then we don't have any
 486	 * concerns about the cpu consuming stale data.  This mitigates
 487	 * legitimate usages of overlapping mappings.
 488	 */
 489	if (entry->direction == DMA_TO_DEVICE)
 490		return 0;
 491
 492	spin_lock_irqsave(&radix_lock, flags);
 493	rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
 494	if (rc == -EEXIST)
 495		active_cacheline_inc_overlap(cln);
 496	spin_unlock_irqrestore(&radix_lock, flags);
 497
 498	return rc;
 499}
 500
 501static void active_cacheline_remove(struct dma_debug_entry *entry)
 502{
 503	phys_addr_t cln = to_cacheline_number(entry);
 504	unsigned long flags;
 505
 506	/* ...mirror the insert case */
 507	if (entry->direction == DMA_TO_DEVICE)
 508		return;
 509
 510	spin_lock_irqsave(&radix_lock, flags);
 511	/* since we are counting overlaps the final put of the
 512	 * cacheline will occur when the overlap count is 0.
 513	 * active_cacheline_dec_overlap() returns -1 in that case
 514	 */
 515	if (active_cacheline_dec_overlap(cln) < 0)
 516		radix_tree_delete(&dma_active_cacheline, cln);
 517	spin_unlock_irqrestore(&radix_lock, flags);
 518}
 519
 520/*
 521 * Dump mappings entries on kernel space for debugging purposes
 522 */
 523void debug_dma_dump_mappings(struct device *dev)
 524{
 525	int idx;
 526	phys_addr_t cln;
 527
 528	for (idx = 0; idx < HASH_SIZE; idx++) {
 529		struct hash_bucket *bucket = &dma_entry_hash[idx];
 530		struct dma_debug_entry *entry;
 531		unsigned long flags;
 532
 533		spin_lock_irqsave(&bucket->lock, flags);
 534		list_for_each_entry(entry, &bucket->list, list) {
 535			if (!dev || dev == entry->dev) {
 536				cln = to_cacheline_number(entry);
 537				dev_info(entry->dev,
 538					 "%s idx %d P=%llx N=%lx D=%llx L=%llx cln=%pa %s %s\n",
 539					 type2name[entry->type], idx,
 540					 phys_addr(entry), entry->pfn,
 541					 entry->dev_addr, entry->size,
 542					 &cln, dir2name[entry->direction],
 543					 maperr2str[entry->map_err_type]);
 544			}
 545		}
 546		spin_unlock_irqrestore(&bucket->lock, flags);
 547
 548		cond_resched();
 549	}
 550}
 551
 552/*
 553 * Dump mappings entries on user space via debugfs
 554 */
 555static int dump_show(struct seq_file *seq, void *v)
 556{
 557	int idx;
 558	phys_addr_t cln;
 559
 560	for (idx = 0; idx < HASH_SIZE; idx++) {
 561		struct hash_bucket *bucket = &dma_entry_hash[idx];
 562		struct dma_debug_entry *entry;
 563		unsigned long flags;
 564
 565		spin_lock_irqsave(&bucket->lock, flags);
 566		list_for_each_entry(entry, &bucket->list, list) {
 567			cln = to_cacheline_number(entry);
 568			seq_printf(seq,
 569				   "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx cln=%pa %s %s\n",
 570				   dev_driver_string(entry->dev),
 571				   dev_name(entry->dev),
 572				   type2name[entry->type], idx,
 573				   phys_addr(entry), entry->pfn,
 574				   entry->dev_addr, entry->size,
 575				   &cln, dir2name[entry->direction],
 576				   maperr2str[entry->map_err_type]);
 577		}
 578		spin_unlock_irqrestore(&bucket->lock, flags);
 579	}
 580	return 0;
 581}
 582DEFINE_SHOW_ATTRIBUTE(dump);
 583
 584/*
 585 * Wrapper function for adding an entry to the hash.
 586 * This function takes care of locking itself.
 587 */
 588static void add_dma_entry(struct dma_debug_entry *entry, unsigned long attrs)
 589{
 590	struct hash_bucket *bucket;
 591	unsigned long flags;
 592	int rc;
 593
 594	bucket = get_hash_bucket(entry, &flags);
 595	hash_bucket_add(bucket, entry);
 596	put_hash_bucket(bucket, flags);
 597
 598	rc = active_cacheline_insert(entry);
 599	if (rc == -ENOMEM) {
 600		pr_err_once("cacheline tracking ENOMEM, dma-debug disabled\n");
 601		global_disable = true;
 602	} else if (rc == -EEXIST && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
 603		err_printk(entry->dev, entry,
 604			"cacheline tracking EEXIST, overlapping mappings aren't supported\n");
 605	}
 606}
 607
 608static int dma_debug_create_entries(gfp_t gfp)
 609{
 610	struct dma_debug_entry *entry;
 611	int i;
 612
 613	entry = (void *)get_zeroed_page(gfp);
 614	if (!entry)
 615		return -ENOMEM;
 616
 617	for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
 618		list_add_tail(&entry[i].list, &free_entries);
 619
 620	num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 621	nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 622
 623	return 0;
 624}
 625
 626static struct dma_debug_entry *__dma_entry_alloc(void)
 627{
 628	struct dma_debug_entry *entry;
 629
 630	entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 631	list_del(&entry->list);
 632	memset(entry, 0, sizeof(*entry));
 633
 634	num_free_entries -= 1;
 635	if (num_free_entries < min_free_entries)
 636		min_free_entries = num_free_entries;
 637
 638	return entry;
 639}
 640
 641/*
 642 * This should be called outside of free_entries_lock scope to avoid potential
 643 * deadlocks with serial consoles that use DMA.
 644 */
 645static void __dma_entry_alloc_check_leak(u32 nr_entries)
 646{
 647	u32 tmp = nr_entries % nr_prealloc_entries;
 648
 649	/* Shout each time we tick over some multiple of the initial pool */
 650	if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
 651		pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
 652			nr_entries,
 653			(nr_entries / nr_prealloc_entries));
 654	}
 655}
 656
 657/* struct dma_entry allocator
 658 *
 659 * The next two functions implement the allocator for
 660 * struct dma_debug_entries.
 661 */
 662static struct dma_debug_entry *dma_entry_alloc(void)
 663{
 664	bool alloc_check_leak = false;
 665	struct dma_debug_entry *entry;
 666	unsigned long flags;
 667	u32 nr_entries;
 668
 669	spin_lock_irqsave(&free_entries_lock, flags);
 670	if (num_free_entries == 0) {
 671		if (dma_debug_create_entries(GFP_ATOMIC)) {
 672			global_disable = true;
 673			spin_unlock_irqrestore(&free_entries_lock, flags);
 674			pr_err("debugging out of memory - disabling\n");
 675			return NULL;
 676		}
 677		alloc_check_leak = true;
 678		nr_entries = nr_total_entries;
 679	}
 680
 681	entry = __dma_entry_alloc();
 682
 683	spin_unlock_irqrestore(&free_entries_lock, flags);
 684
 685	if (alloc_check_leak)
 686		__dma_entry_alloc_check_leak(nr_entries);
 687
 688#ifdef CONFIG_STACKTRACE
 689	entry->stack_len = stack_trace_save(entry->stack_entries,
 690					    ARRAY_SIZE(entry->stack_entries),
 691					    1);
 692#endif
 693	return entry;
 694}
 695
 696static void dma_entry_free(struct dma_debug_entry *entry)
 697{
 698	unsigned long flags;
 699
 700	active_cacheline_remove(entry);
 701
 702	/*
 703	 * add to beginning of the list - this way the entries are
 704	 * more likely cache hot when they are reallocated.
 705	 */
 706	spin_lock_irqsave(&free_entries_lock, flags);
 707	list_add(&entry->list, &free_entries);
 708	num_free_entries += 1;
 709	spin_unlock_irqrestore(&free_entries_lock, flags);
 710}
 711
 712/*
 713 * DMA-API debugging init code
 714 *
 715 * The init code does two things:
 716 *   1. Initialize core data structures
 717 *   2. Preallocate a given number of dma_debug_entry structs
 718 */
 719
 720static ssize_t filter_read(struct file *file, char __user *user_buf,
 721			   size_t count, loff_t *ppos)
 722{
 723	char buf[NAME_MAX_LEN + 1];
 724	unsigned long flags;
 725	int len;
 726
 727	if (!current_driver_name[0])
 728		return 0;
 729
 730	/*
 731	 * We can't copy to userspace directly because current_driver_name can
 732	 * only be read under the driver_name_lock with irqs disabled. So
 733	 * create a temporary copy first.
 734	 */
 735	read_lock_irqsave(&driver_name_lock, flags);
 736	len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 737	read_unlock_irqrestore(&driver_name_lock, flags);
 738
 739	return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 740}
 741
 742static ssize_t filter_write(struct file *file, const char __user *userbuf,
 743			    size_t count, loff_t *ppos)
 744{
 745	char buf[NAME_MAX_LEN];
 746	unsigned long flags;
 747	size_t len;
 748	int i;
 749
 750	/*
 751	 * We can't copy from userspace directly. Access to
 752	 * current_driver_name is protected with a write_lock with irqs
 753	 * disabled. Since copy_from_user can fault and may sleep we
 754	 * need to copy to temporary buffer first
 755	 */
 756	len = min(count, (size_t)(NAME_MAX_LEN - 1));
 757	if (copy_from_user(buf, userbuf, len))
 758		return -EFAULT;
 759
 760	buf[len] = 0;
 761
 762	write_lock_irqsave(&driver_name_lock, flags);
 763
 764	/*
 765	 * Now handle the string we got from userspace very carefully.
 766	 * The rules are:
 767	 *         - only use the first token we got
 768	 *         - token delimiter is everything looking like a space
 769	 *           character (' ', '\n', '\t' ...)
 770	 *
 771	 */
 772	if (!isalnum(buf[0])) {
 773		/*
 774		 * If the first character userspace gave us is not
 775		 * alphanumerical then assume the filter should be
 776		 * switched off.
 777		 */
 778		if (current_driver_name[0])
 779			pr_info("switching off dma-debug driver filter\n");
 780		current_driver_name[0] = 0;
 781		current_driver = NULL;
 782		goto out_unlock;
 783	}
 784
 785	/*
 786	 * Now parse out the first token and use it as the name for the
 787	 * driver to filter for.
 788	 */
 789	for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 790		current_driver_name[i] = buf[i];
 791		if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 792			break;
 793	}
 794	current_driver_name[i] = 0;
 795	current_driver = NULL;
 796
 797	pr_info("enable driver filter for driver [%s]\n",
 798		current_driver_name);
 799
 800out_unlock:
 801	write_unlock_irqrestore(&driver_name_lock, flags);
 802
 803	return count;
 804}
 805
 806static const struct file_operations filter_fops = {
 807	.read  = filter_read,
 808	.write = filter_write,
 809	.llseek = default_llseek,
 810};
 811
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 812static int __init dma_debug_fs_init(void)
 813{
 814	struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
 815
 816	debugfs_create_bool("disabled", 0444, dentry, &global_disable);
 817	debugfs_create_u32("error_count", 0444, dentry, &error_count);
 818	debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
 819	debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
 820	debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
 821	debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
 822	debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
 823	debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
 824	debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
 825
 826	return 0;
 827}
 828core_initcall_sync(dma_debug_fs_init);
 829
 830static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 831{
 832	struct dma_debug_entry *entry;
 833	unsigned long flags;
 834	int count = 0, i;
 835
 836	for (i = 0; i < HASH_SIZE; ++i) {
 837		spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
 838		list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 839			if (entry->dev == dev) {
 840				count += 1;
 841				*out_entry = entry;
 842			}
 843		}
 844		spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
 845	}
 846
 847	return count;
 848}
 849
 850static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 851{
 852	struct device *dev = data;
 853	struct dma_debug_entry *entry;
 854	int count;
 855
 856	if (dma_debug_disabled())
 857		return 0;
 858
 859	switch (action) {
 860	case BUS_NOTIFY_UNBOUND_DRIVER:
 861		count = device_dma_allocations(dev, &entry);
 862		if (count == 0)
 863			break;
 864		err_printk(dev, entry, "device driver has pending "
 865				"DMA allocations while released from device "
 866				"[count=%d]\n"
 867				"One of leaked entries details: "
 868				"[device address=0x%016llx] [size=%llu bytes] "
 869				"[mapped with %s] [mapped as %s]\n",
 870			count, entry->dev_addr, entry->size,
 871			dir2name[entry->direction], type2name[entry->type]);
 872		break;
 873	default:
 874		break;
 875	}
 876
 877	return 0;
 878}
 879
 880void dma_debug_add_bus(const struct bus_type *bus)
 881{
 882	struct notifier_block *nb;
 883
 884	if (dma_debug_disabled())
 885		return;
 886
 887	nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 888	if (nb == NULL) {
 889		pr_err("dma_debug_add_bus: out of memory\n");
 890		return;
 891	}
 892
 893	nb->notifier_call = dma_debug_device_change;
 894
 895	bus_register_notifier(bus, nb);
 896}
 897
 898static int dma_debug_init(void)
 899{
 900	int i, nr_pages;
 901
 902	/* Do not use dma_debug_initialized here, since we really want to be
 903	 * called to set dma_debug_initialized
 904	 */
 905	if (global_disable)
 906		return 0;
 907
 908	for (i = 0; i < HASH_SIZE; ++i) {
 909		INIT_LIST_HEAD(&dma_entry_hash[i].list);
 910		spin_lock_init(&dma_entry_hash[i].lock);
 911	}
 912
 913	nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
 914	for (i = 0; i < nr_pages; ++i)
 915		dma_debug_create_entries(GFP_KERNEL);
 916	if (num_free_entries >= nr_prealloc_entries) {
 917		pr_info("preallocated %d debug entries\n", nr_total_entries);
 918	} else if (num_free_entries > 0) {
 919		pr_warn("%d debug entries requested but only %d allocated\n",
 920			nr_prealloc_entries, nr_total_entries);
 921	} else {
 922		pr_err("debugging out of memory error - disabled\n");
 923		global_disable = true;
 924
 925		return 0;
 926	}
 927	min_free_entries = num_free_entries;
 928
 929	dma_debug_initialized = true;
 930
 931	pr_info("debugging enabled by kernel config\n");
 932	return 0;
 933}
 934core_initcall(dma_debug_init);
 935
 936static __init int dma_debug_cmdline(char *str)
 937{
 938	if (!str)
 939		return -EINVAL;
 940
 941	if (strncmp(str, "off", 3) == 0) {
 942		pr_info("debugging disabled on kernel command line\n");
 943		global_disable = true;
 944	}
 945
 946	return 1;
 947}
 948
 949static __init int dma_debug_entries_cmdline(char *str)
 950{
 951	if (!str)
 952		return -EINVAL;
 953	if (!get_option(&str, &nr_prealloc_entries))
 954		nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 955	return 1;
 956}
 957
 958__setup("dma_debug=", dma_debug_cmdline);
 959__setup("dma_debug_entries=", dma_debug_entries_cmdline);
 960
 961static void check_unmap(struct dma_debug_entry *ref)
 962{
 963	struct dma_debug_entry *entry;
 964	struct hash_bucket *bucket;
 965	unsigned long flags;
 966
 967	bucket = get_hash_bucket(ref, &flags);
 968	entry = bucket_find_exact(bucket, ref);
 969
 970	if (!entry) {
 971		/* must drop lock before calling dma_mapping_error */
 972		put_hash_bucket(bucket, flags);
 973
 974		if (dma_mapping_error(ref->dev, ref->dev_addr)) {
 975			err_printk(ref->dev, NULL,
 976				   "device driver tries to free an "
 977				   "invalid DMA memory address\n");
 978		} else {
 979			err_printk(ref->dev, NULL,
 980				   "device driver tries to free DMA "
 981				   "memory it has not allocated [device "
 982				   "address=0x%016llx] [size=%llu bytes]\n",
 983				   ref->dev_addr, ref->size);
 984		}
 985		return;
 986	}
 987
 988	if (ref->size != entry->size) {
 989		err_printk(ref->dev, entry, "device driver frees "
 990			   "DMA memory with different size "
 991			   "[device address=0x%016llx] [map size=%llu bytes] "
 992			   "[unmap size=%llu bytes]\n",
 993			   ref->dev_addr, entry->size, ref->size);
 994	}
 995
 996	if (ref->type != entry->type) {
 997		err_printk(ref->dev, entry, "device driver frees "
 998			   "DMA memory with wrong function "
 999			   "[device address=0x%016llx] [size=%llu bytes] "
1000			   "[mapped as %s] [unmapped as %s]\n",
1001			   ref->dev_addr, ref->size,
1002			   type2name[entry->type], type2name[ref->type]);
1003	} else if ((entry->type == dma_debug_coherent) &&
1004		   (phys_addr(ref) != phys_addr(entry))) {
1005		err_printk(ref->dev, entry, "device driver frees "
1006			   "DMA memory with different CPU address "
1007			   "[device address=0x%016llx] [size=%llu bytes] "
1008			   "[cpu alloc address=0x%016llx] "
1009			   "[cpu free address=0x%016llx]",
1010			   ref->dev_addr, ref->size,
1011			   phys_addr(entry),
1012			   phys_addr(ref));
1013	}
1014
1015	if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1016	    ref->sg_call_ents != entry->sg_call_ents) {
1017		err_printk(ref->dev, entry, "device driver frees "
1018			   "DMA sg list with different entry count "
1019			   "[map count=%d] [unmap count=%d]\n",
1020			   entry->sg_call_ents, ref->sg_call_ents);
1021	}
1022
1023	/*
1024	 * This may be no bug in reality - but most implementations of the
1025	 * DMA API don't handle this properly, so check for it here
1026	 */
1027	if (ref->direction != entry->direction) {
1028		err_printk(ref->dev, entry, "device driver frees "
1029			   "DMA memory with different direction "
1030			   "[device address=0x%016llx] [size=%llu bytes] "
1031			   "[mapped with %s] [unmapped with %s]\n",
1032			   ref->dev_addr, ref->size,
1033			   dir2name[entry->direction],
1034			   dir2name[ref->direction]);
1035	}
1036
1037	/*
1038	 * Drivers should use dma_mapping_error() to check the returned
1039	 * addresses of dma_map_single() and dma_map_page().
1040	 * If not, print this warning message. See Documentation/core-api/dma-api.rst.
1041	 */
1042	if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1043		err_printk(ref->dev, entry,
1044			   "device driver failed to check map error"
1045			   "[device address=0x%016llx] [size=%llu bytes] "
1046			   "[mapped as %s]",
1047			   ref->dev_addr, ref->size,
1048			   type2name[entry->type]);
1049	}
1050
1051	hash_bucket_del(entry);
1052	dma_entry_free(entry);
1053
1054	put_hash_bucket(bucket, flags);
1055}
1056
1057static void check_for_stack(struct device *dev,
1058			    struct page *page, size_t offset)
1059{
1060	void *addr;
1061	struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1062
1063	if (!stack_vm_area) {
1064		/* Stack is direct-mapped. */
1065		if (PageHighMem(page))
1066			return;
1067		addr = page_address(page) + offset;
1068		if (object_is_on_stack(addr))
1069			err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1070	} else {
1071		/* Stack is vmalloced. */
1072		int i;
1073
1074		for (i = 0; i < stack_vm_area->nr_pages; i++) {
1075			if (page != stack_vm_area->pages[i])
1076				continue;
1077
1078			addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1079			err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1080			break;
1081		}
1082	}
1083}
1084
1085static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1086{
1087	if (memory_intersects(_stext, _etext, addr, len) ||
1088	    memory_intersects(__start_rodata, __end_rodata, addr, len))
1089		err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1090}
1091
1092static void check_sync(struct device *dev,
1093		       struct dma_debug_entry *ref,
1094		       bool to_cpu)
1095{
1096	struct dma_debug_entry *entry;
1097	struct hash_bucket *bucket;
1098	unsigned long flags;
1099
1100	bucket = get_hash_bucket(ref, &flags);
1101
1102	entry = bucket_find_contain(&bucket, ref, &flags);
1103
1104	if (!entry) {
1105		err_printk(dev, NULL, "device driver tries "
1106				"to sync DMA memory it has not allocated "
1107				"[device address=0x%016llx] [size=%llu bytes]\n",
1108				(unsigned long long)ref->dev_addr, ref->size);
1109		goto out;
1110	}
1111
1112	if (ref->size > entry->size) {
1113		err_printk(dev, entry, "device driver syncs"
1114				" DMA memory outside allocated range "
1115				"[device address=0x%016llx] "
1116				"[allocation size=%llu bytes] "
1117				"[sync offset+size=%llu]\n",
1118				entry->dev_addr, entry->size,
1119				ref->size);
1120	}
1121
1122	if (entry->direction == DMA_BIDIRECTIONAL)
1123		goto out;
1124
1125	if (ref->direction != entry->direction) {
1126		err_printk(dev, entry, "device driver syncs "
1127				"DMA memory with different direction "
1128				"[device address=0x%016llx] [size=%llu bytes] "
1129				"[mapped with %s] [synced with %s]\n",
1130				(unsigned long long)ref->dev_addr, entry->size,
1131				dir2name[entry->direction],
1132				dir2name[ref->direction]);
1133	}
1134
1135	if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1136		      !(ref->direction == DMA_TO_DEVICE))
1137		err_printk(dev, entry, "device driver syncs "
1138				"device read-only DMA memory for cpu "
1139				"[device address=0x%016llx] [size=%llu bytes] "
1140				"[mapped with %s] [synced with %s]\n",
1141				(unsigned long long)ref->dev_addr, entry->size,
1142				dir2name[entry->direction],
1143				dir2name[ref->direction]);
1144
1145	if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1146		       !(ref->direction == DMA_FROM_DEVICE))
1147		err_printk(dev, entry, "device driver syncs "
1148				"device write-only DMA memory to device "
1149				"[device address=0x%016llx] [size=%llu bytes] "
1150				"[mapped with %s] [synced with %s]\n",
1151				(unsigned long long)ref->dev_addr, entry->size,
1152				dir2name[entry->direction],
1153				dir2name[ref->direction]);
1154
1155	if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1156	    ref->sg_call_ents != entry->sg_call_ents) {
1157		err_printk(ref->dev, entry, "device driver syncs "
1158			   "DMA sg list with different entry count "
1159			   "[map count=%d] [sync count=%d]\n",
1160			   entry->sg_call_ents, ref->sg_call_ents);
1161	}
1162
1163out:
1164	put_hash_bucket(bucket, flags);
1165}
1166
1167static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1168{
1169#ifdef CONFIG_DMA_API_DEBUG_SG
1170	unsigned int max_seg = dma_get_max_seg_size(dev);
1171	u64 start, end, boundary = dma_get_seg_boundary(dev);
1172
1173	/*
1174	 * Either the driver forgot to set dma_parms appropriately, or
1175	 * whoever generated the list forgot to check them.
1176	 */
1177	if (sg->length > max_seg)
1178		err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1179			   sg->length, max_seg);
1180	/*
1181	 * In some cases this could potentially be the DMA API
1182	 * implementation's fault, but it would usually imply that
1183	 * the scatterlist was built inappropriately to begin with.
1184	 */
1185	start = sg_dma_address(sg);
1186	end = start + sg_dma_len(sg) - 1;
1187	if ((start ^ end) & ~boundary)
1188		err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1189			   start, end, boundary);
1190#endif
1191}
1192
1193void debug_dma_map_single(struct device *dev, const void *addr,
1194			    unsigned long len)
1195{
1196	if (unlikely(dma_debug_disabled()))
1197		return;
1198
1199	if (!virt_addr_valid(addr))
1200		err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1201			   addr, len);
1202
1203	if (is_vmalloc_addr(addr))
1204		err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1205			   addr, len);
1206}
1207EXPORT_SYMBOL(debug_dma_map_single);
1208
1209void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1210			size_t size, int direction, dma_addr_t dma_addr,
1211			unsigned long attrs)
1212{
1213	struct dma_debug_entry *entry;
1214
1215	if (unlikely(dma_debug_disabled()))
1216		return;
1217
1218	if (dma_mapping_error(dev, dma_addr))
1219		return;
1220
1221	entry = dma_entry_alloc();
1222	if (!entry)
1223		return;
1224
1225	entry->dev       = dev;
1226	entry->type      = dma_debug_single;
1227	entry->pfn	 = page_to_pfn(page);
1228	entry->offset	 = offset;
1229	entry->dev_addr  = dma_addr;
1230	entry->size      = size;
1231	entry->direction = direction;
1232	entry->map_err_type = MAP_ERR_NOT_CHECKED;
1233
1234	check_for_stack(dev, page, offset);
1235
1236	if (!PageHighMem(page)) {
1237		void *addr = page_address(page) + offset;
1238
1239		check_for_illegal_area(dev, addr, size);
1240	}
1241
1242	add_dma_entry(entry, attrs);
1243}
1244
1245void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1246{
1247	struct dma_debug_entry ref;
1248	struct dma_debug_entry *entry;
1249	struct hash_bucket *bucket;
1250	unsigned long flags;
1251
1252	if (unlikely(dma_debug_disabled()))
1253		return;
1254
1255	ref.dev = dev;
1256	ref.dev_addr = dma_addr;
1257	bucket = get_hash_bucket(&ref, &flags);
1258
1259	list_for_each_entry(entry, &bucket->list, list) {
1260		if (!exact_match(&ref, entry))
1261			continue;
1262
1263		/*
1264		 * The same physical address can be mapped multiple
1265		 * times. Without a hardware IOMMU this results in the
1266		 * same device addresses being put into the dma-debug
1267		 * hash multiple times too. This can result in false
1268		 * positives being reported. Therefore we implement a
1269		 * best-fit algorithm here which updates the first entry
1270		 * from the hash which fits the reference value and is
1271		 * not currently listed as being checked.
1272		 */
1273		if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1274			entry->map_err_type = MAP_ERR_CHECKED;
1275			break;
1276		}
1277	}
1278
1279	put_hash_bucket(bucket, flags);
1280}
1281EXPORT_SYMBOL(debug_dma_mapping_error);
1282
1283void debug_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
1284			  size_t size, int direction)
1285{
1286	struct dma_debug_entry ref = {
1287		.type           = dma_debug_single,
1288		.dev            = dev,
1289		.dev_addr       = dma_addr,
1290		.size           = size,
1291		.direction      = direction,
1292	};
1293
1294	if (unlikely(dma_debug_disabled()))
1295		return;
1296	check_unmap(&ref);
1297}
1298
1299void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1300		      int nents, int mapped_ents, int direction,
1301		      unsigned long attrs)
1302{
1303	struct dma_debug_entry *entry;
1304	struct scatterlist *s;
1305	int i;
1306
1307	if (unlikely(dma_debug_disabled()))
1308		return;
1309
1310	for_each_sg(sg, s, nents, i) {
1311		check_for_stack(dev, sg_page(s), s->offset);
1312		if (!PageHighMem(sg_page(s)))
1313			check_for_illegal_area(dev, sg_virt(s), s->length);
1314	}
1315
1316	for_each_sg(sg, s, mapped_ents, i) {
1317		entry = dma_entry_alloc();
1318		if (!entry)
1319			return;
1320
1321		entry->type           = dma_debug_sg;
1322		entry->dev            = dev;
1323		entry->pfn	      = page_to_pfn(sg_page(s));
1324		entry->offset	      = s->offset;
1325		entry->size           = sg_dma_len(s);
1326		entry->dev_addr       = sg_dma_address(s);
1327		entry->direction      = direction;
1328		entry->sg_call_ents   = nents;
1329		entry->sg_mapped_ents = mapped_ents;
1330
1331		check_sg_segment(dev, s);
1332
1333		add_dma_entry(entry, attrs);
1334	}
1335}
1336
1337static int get_nr_mapped_entries(struct device *dev,
1338				 struct dma_debug_entry *ref)
1339{
1340	struct dma_debug_entry *entry;
1341	struct hash_bucket *bucket;
1342	unsigned long flags;
1343	int mapped_ents;
1344
1345	bucket       = get_hash_bucket(ref, &flags);
1346	entry        = bucket_find_exact(bucket, ref);
1347	mapped_ents  = 0;
1348
1349	if (entry)
1350		mapped_ents = entry->sg_mapped_ents;
1351	put_hash_bucket(bucket, flags);
1352
1353	return mapped_ents;
1354}
1355
1356void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1357			int nelems, int dir)
1358{
1359	struct scatterlist *s;
1360	int mapped_ents = 0, i;
1361
1362	if (unlikely(dma_debug_disabled()))
1363		return;
1364
1365	for_each_sg(sglist, s, nelems, i) {
1366
1367		struct dma_debug_entry ref = {
1368			.type           = dma_debug_sg,
1369			.dev            = dev,
1370			.pfn		= page_to_pfn(sg_page(s)),
1371			.offset		= s->offset,
1372			.dev_addr       = sg_dma_address(s),
1373			.size           = sg_dma_len(s),
1374			.direction      = dir,
1375			.sg_call_ents   = nelems,
1376		};
1377
1378		if (mapped_ents && i >= mapped_ents)
1379			break;
1380
1381		if (!i)
1382			mapped_ents = get_nr_mapped_entries(dev, &ref);
1383
1384		check_unmap(&ref);
1385	}
1386}
1387
1388void debug_dma_alloc_coherent(struct device *dev, size_t size,
1389			      dma_addr_t dma_addr, void *virt,
1390			      unsigned long attrs)
1391{
1392	struct dma_debug_entry *entry;
1393
1394	if (unlikely(dma_debug_disabled()))
1395		return;
1396
1397	if (unlikely(virt == NULL))
1398		return;
1399
1400	/* handle vmalloc and linear addresses */
1401	if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1402		return;
1403
1404	entry = dma_entry_alloc();
1405	if (!entry)
1406		return;
1407
1408	entry->type      = dma_debug_coherent;
1409	entry->dev       = dev;
1410	entry->offset	 = offset_in_page(virt);
1411	entry->size      = size;
1412	entry->dev_addr  = dma_addr;
1413	entry->direction = DMA_BIDIRECTIONAL;
1414
1415	if (is_vmalloc_addr(virt))
1416		entry->pfn = vmalloc_to_pfn(virt);
1417	else
1418		entry->pfn = page_to_pfn(virt_to_page(virt));
1419
1420	add_dma_entry(entry, attrs);
1421}
1422
1423void debug_dma_free_coherent(struct device *dev, size_t size,
1424			 void *virt, dma_addr_t dma_addr)
1425{
1426	struct dma_debug_entry ref = {
1427		.type           = dma_debug_coherent,
1428		.dev            = dev,
1429		.offset		= offset_in_page(virt),
1430		.dev_addr       = dma_addr,
1431		.size           = size,
1432		.direction      = DMA_BIDIRECTIONAL,
1433	};
1434
1435	/* handle vmalloc and linear addresses */
1436	if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1437		return;
1438
1439	if (is_vmalloc_addr(virt))
1440		ref.pfn = vmalloc_to_pfn(virt);
1441	else
1442		ref.pfn = page_to_pfn(virt_to_page(virt));
1443
1444	if (unlikely(dma_debug_disabled()))
1445		return;
1446
1447	check_unmap(&ref);
1448}
1449
1450void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1451			    int direction, dma_addr_t dma_addr,
1452			    unsigned long attrs)
1453{
1454	struct dma_debug_entry *entry;
1455
1456	if (unlikely(dma_debug_disabled()))
1457		return;
1458
1459	entry = dma_entry_alloc();
1460	if (!entry)
1461		return;
1462
1463	entry->type		= dma_debug_resource;
1464	entry->dev		= dev;
1465	entry->pfn		= PHYS_PFN(addr);
1466	entry->offset		= offset_in_page(addr);
1467	entry->size		= size;
1468	entry->dev_addr		= dma_addr;
1469	entry->direction	= direction;
1470	entry->map_err_type	= MAP_ERR_NOT_CHECKED;
1471
1472	add_dma_entry(entry, attrs);
1473}
1474
1475void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1476			      size_t size, int direction)
1477{
1478	struct dma_debug_entry ref = {
1479		.type           = dma_debug_resource,
1480		.dev            = dev,
1481		.dev_addr       = dma_addr,
1482		.size           = size,
1483		.direction      = direction,
1484	};
1485
1486	if (unlikely(dma_debug_disabled()))
1487		return;
1488
1489	check_unmap(&ref);
1490}
1491
1492void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1493				   size_t size, int direction)
1494{
1495	struct dma_debug_entry ref;
1496
1497	if (unlikely(dma_debug_disabled()))
1498		return;
1499
1500	ref.type         = dma_debug_single;
1501	ref.dev          = dev;
1502	ref.dev_addr     = dma_handle;
1503	ref.size         = size;
1504	ref.direction    = direction;
1505	ref.sg_call_ents = 0;
1506
1507	check_sync(dev, &ref, true);
1508}
1509
1510void debug_dma_sync_single_for_device(struct device *dev,
1511				      dma_addr_t dma_handle, size_t size,
1512				      int direction)
1513{
1514	struct dma_debug_entry ref;
1515
1516	if (unlikely(dma_debug_disabled()))
1517		return;
1518
1519	ref.type         = dma_debug_single;
1520	ref.dev          = dev;
1521	ref.dev_addr     = dma_handle;
1522	ref.size         = size;
1523	ref.direction    = direction;
1524	ref.sg_call_ents = 0;
1525
1526	check_sync(dev, &ref, false);
1527}
1528
1529void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1530			       int nelems, int direction)
1531{
1532	struct scatterlist *s;
1533	int mapped_ents = 0, i;
1534
1535	if (unlikely(dma_debug_disabled()))
1536		return;
1537
1538	for_each_sg(sg, s, nelems, i) {
1539
1540		struct dma_debug_entry ref = {
1541			.type           = dma_debug_sg,
1542			.dev            = dev,
1543			.pfn		= page_to_pfn(sg_page(s)),
1544			.offset		= s->offset,
1545			.dev_addr       = sg_dma_address(s),
1546			.size           = sg_dma_len(s),
1547			.direction      = direction,
1548			.sg_call_ents   = nelems,
1549		};
1550
1551		if (!i)
1552			mapped_ents = get_nr_mapped_entries(dev, &ref);
1553
1554		if (i >= mapped_ents)
1555			break;
1556
1557		check_sync(dev, &ref, true);
1558	}
1559}
1560
1561void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1562				  int nelems, int direction)
1563{
1564	struct scatterlist *s;
1565	int mapped_ents = 0, i;
1566
1567	if (unlikely(dma_debug_disabled()))
1568		return;
1569
1570	for_each_sg(sg, s, nelems, i) {
1571
1572		struct dma_debug_entry ref = {
1573			.type           = dma_debug_sg,
1574			.dev            = dev,
1575			.pfn		= page_to_pfn(sg_page(s)),
1576			.offset		= s->offset,
1577			.dev_addr       = sg_dma_address(s),
1578			.size           = sg_dma_len(s),
1579			.direction      = direction,
1580			.sg_call_ents   = nelems,
1581		};
1582		if (!i)
1583			mapped_ents = get_nr_mapped_entries(dev, &ref);
1584
1585		if (i >= mapped_ents)
1586			break;
1587
1588		check_sync(dev, &ref, false);
1589	}
1590}
1591
1592static int __init dma_debug_driver_setup(char *str)
1593{
1594	int i;
1595
1596	for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1597		current_driver_name[i] = *str;
1598		if (*str == 0)
1599			break;
1600	}
1601
1602	if (current_driver_name[0])
1603		pr_info("enable driver filter for driver [%s]\n",
1604			current_driver_name);
1605
1606
1607	return 1;
1608}
1609__setup("dma_debug_driver=", dma_debug_driver_setup);