Linux Audio

Check our new training course

Loading...
v5.4
   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-mapping.h>
  13#include <linux/sched/task.h>
  14#include <linux/stacktrace.h>
  15#include <linux/dma-debug.h>
  16#include <linux/spinlock.h>
  17#include <linux/vmalloc.h>
  18#include <linux/debugfs.h>
  19#include <linux/uaccess.h>
  20#include <linux/export.h>
  21#include <linux/device.h>
  22#include <linux/types.h>
  23#include <linux/sched.h>
  24#include <linux/ctype.h>
  25#include <linux/list.h>
  26#include <linux/slab.h>
  27
  28#include <asm/sections.h>
 
  29
  30#define HASH_SIZE       1024ULL
  31#define HASH_FN_SHIFT   13
  32#define HASH_FN_MASK    (HASH_SIZE - 1)
  33
  34#define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
  35/* If the pool runs out, add this many new entries at once */
  36#define DMA_DEBUG_DYNAMIC_ENTRIES (PAGE_SIZE / sizeof(struct dma_debug_entry))
  37
  38enum {
  39	dma_debug_single,
  40	dma_debug_sg,
  41	dma_debug_coherent,
  42	dma_debug_resource,
  43};
  44
  45enum map_err_types {
  46	MAP_ERR_CHECK_NOT_APPLICABLE,
  47	MAP_ERR_NOT_CHECKED,
  48	MAP_ERR_CHECKED,
  49};
  50
  51#define DMA_DEBUG_STACKTRACE_ENTRIES 5
  52
  53/**
  54 * struct dma_debug_entry - track a dma_map* or dma_alloc_coherent mapping
  55 * @list: node on pre-allocated free_entries list
  56 * @dev: 'dev' argument to dma_map_{page|single|sg} or dma_alloc_coherent
  57 * @type: single, page, sg, coherent
  58 * @pfn: page frame of the start address
  59 * @offset: offset of mapping relative to pfn
  60 * @size: length of the mapping
 
  61 * @direction: enum dma_data_direction
  62 * @sg_call_ents: 'nents' from dma_map_sg
  63 * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
 
  64 * @map_err_type: track whether dma_mapping_error() was checked
  65 * @stacktrace: support backtraces when a violation is detected
 
  66 */
  67struct dma_debug_entry {
  68	struct list_head list;
  69	struct device    *dev;
  70	int              type;
  71	unsigned long	 pfn;
  72	size_t		 offset;
  73	u64              dev_addr;
  74	u64              size;
 
  75	int              direction;
  76	int		 sg_call_ents;
  77	int		 sg_mapped_ents;
 
  78	enum map_err_types  map_err_type;
  79#ifdef CONFIG_STACKTRACE
  80	unsigned int	stack_len;
  81	unsigned long	stack_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
  82#endif
  83};
  84
  85typedef bool (*match_fn)(struct dma_debug_entry *, struct dma_debug_entry *);
  86
  87struct hash_bucket {
  88	struct list_head list;
  89	spinlock_t lock;
  90} ____cacheline_aligned_in_smp;
  91
  92/* Hash list to save the allocated dma addresses */
  93static struct hash_bucket dma_entry_hash[HASH_SIZE];
  94/* List of pre-allocated dma_debug_entry's */
  95static LIST_HEAD(free_entries);
  96/* Lock for the list above */
  97static DEFINE_SPINLOCK(free_entries_lock);
  98
  99/* Global disable flag - will be set in case of an error */
 100static bool global_disable __read_mostly;
 101
 102/* Early initialization disable flag, set at the end of dma_debug_init */
 103static bool dma_debug_initialized __read_mostly;
 104
 105static inline bool dma_debug_disabled(void)
 106{
 107	return global_disable || !dma_debug_initialized;
 108}
 109
 110/* Global error count */
 111static u32 error_count;
 112
 113/* Global error show enable*/
 114static u32 show_all_errors __read_mostly;
 115/* Number of errors to show */
 116static u32 show_num_errors = 1;
 117
 118static u32 num_free_entries;
 119static u32 min_free_entries;
 120static u32 nr_total_entries;
 121
 122/* number of preallocated entries requested by kernel cmdline */
 123static u32 nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 124
 125/* per-driver filter related state */
 126
 127#define NAME_MAX_LEN	64
 128
 129static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
 130static struct device_driver *current_driver                    __read_mostly;
 131
 132static DEFINE_RWLOCK(driver_name_lock);
 133
 134static const char *const maperr2str[] = {
 135	[MAP_ERR_CHECK_NOT_APPLICABLE] = "dma map error check not applicable",
 136	[MAP_ERR_NOT_CHECKED] = "dma map error not checked",
 137	[MAP_ERR_CHECKED] = "dma map error checked",
 138};
 139
 140static const char *type2name[5] = { "single", "page",
 141				    "scather-gather", "coherent",
 142				    "resource" };
 
 
 
 143
 144static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
 145				   "DMA_FROM_DEVICE", "DMA_NONE" };
 
 
 
 
 146
 147/*
 148 * The access to some variables in this macro is racy. We can't use atomic_t
 149 * here because all these variables are exported to debugfs. Some of them even
 150 * writeable. This is also the reason why a lock won't help much. But anyway,
 151 * the races are no big deal. Here is why:
 152 *
 153 *   error_count: the addition is racy, but the worst thing that can happen is
 154 *                that we don't count some errors
 155 *   show_num_errors: the subtraction is racy. Also no big deal because in
 156 *                    worst case this will result in one warning more in the
 157 *                    system log than the user configured. This variable is
 158 *                    writeable via debugfs.
 159 */
 160static inline void dump_entry_trace(struct dma_debug_entry *entry)
 161{
 162#ifdef CONFIG_STACKTRACE
 163	if (entry) {
 164		pr_warning("Mapped at:\n");
 165		stack_trace_print(entry->stack_entries, entry->stack_len, 0);
 166	}
 167#endif
 168}
 169
 170static bool driver_filter(struct device *dev)
 171{
 172	struct device_driver *drv;
 173	unsigned long flags;
 174	bool ret;
 175
 176	/* driver filter off */
 177	if (likely(!current_driver_name[0]))
 178		return true;
 179
 180	/* driver filter on and initialized */
 181	if (current_driver && dev && dev->driver == current_driver)
 182		return true;
 183
 184	/* driver filter on, but we can't filter on a NULL device... */
 185	if (!dev)
 186		return false;
 187
 188	if (current_driver || !current_driver_name[0])
 189		return false;
 190
 191	/* driver filter on but not yet initialized */
 192	drv = dev->driver;
 193	if (!drv)
 194		return false;
 195
 196	/* lock to protect against change of current_driver_name */
 197	read_lock_irqsave(&driver_name_lock, flags);
 198
 199	ret = false;
 200	if (drv->name &&
 201	    strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
 202		current_driver = drv;
 203		ret = true;
 204	}
 205
 206	read_unlock_irqrestore(&driver_name_lock, flags);
 207
 208	return ret;
 209}
 210
 211#define err_printk(dev, entry, format, arg...) do {			\
 212		error_count += 1;					\
 213		if (driver_filter(dev) &&				\
 214		    (show_all_errors || show_num_errors > 0)) {		\
 215			WARN(1, pr_fmt("%s %s: ") format,		\
 216			     dev ? dev_driver_string(dev) : "NULL",	\
 217			     dev ? dev_name(dev) : "NULL", ## arg);	\
 218			dump_entry_trace(entry);			\
 219		}							\
 220		if (!show_all_errors && show_num_errors > 0)		\
 221			show_num_errors -= 1;				\
 222	} while (0);
 223
 224/*
 225 * Hash related functions
 226 *
 227 * Every DMA-API request is saved into a struct dma_debug_entry. To
 228 * have quick access to these structs they are stored into a hash.
 229 */
 230static int hash_fn(struct dma_debug_entry *entry)
 231{
 232	/*
 233	 * Hash function is based on the dma address.
 234	 * We use bits 20-27 here as the index into the hash
 235	 */
 236	return (entry->dev_addr >> HASH_FN_SHIFT) & HASH_FN_MASK;
 237}
 238
 239/*
 240 * Request exclusive access to a hash bucket for a given dma_debug_entry.
 241 */
 242static struct hash_bucket *get_hash_bucket(struct dma_debug_entry *entry,
 243					   unsigned long *flags)
 244	__acquires(&dma_entry_hash[idx].lock)
 245{
 246	int idx = hash_fn(entry);
 247	unsigned long __flags;
 248
 249	spin_lock_irqsave(&dma_entry_hash[idx].lock, __flags);
 250	*flags = __flags;
 251	return &dma_entry_hash[idx];
 252}
 253
 254/*
 255 * Give up exclusive access to the hash bucket
 256 */
 257static void put_hash_bucket(struct hash_bucket *bucket,
 258			    unsigned long *flags)
 259	__releases(&bucket->lock)
 260{
 261	unsigned long __flags = *flags;
 262
 263	spin_unlock_irqrestore(&bucket->lock, __flags);
 264}
 265
 266static bool exact_match(struct dma_debug_entry *a, struct dma_debug_entry *b)
 267{
 268	return ((a->dev_addr == b->dev_addr) &&
 269		(a->dev == b->dev)) ? true : false;
 270}
 271
 272static bool containing_match(struct dma_debug_entry *a,
 273			     struct dma_debug_entry *b)
 274{
 275	if (a->dev != b->dev)
 276		return false;
 277
 278	if ((b->dev_addr <= a->dev_addr) &&
 279	    ((b->dev_addr + b->size) >= (a->dev_addr + a->size)))
 280		return true;
 281
 282	return false;
 283}
 284
 285/*
 286 * Search a given entry in the hash bucket list
 287 */
 288static struct dma_debug_entry *__hash_bucket_find(struct hash_bucket *bucket,
 289						  struct dma_debug_entry *ref,
 290						  match_fn match)
 291{
 292	struct dma_debug_entry *entry, *ret = NULL;
 293	int matches = 0, match_lvl, last_lvl = -1;
 294
 295	list_for_each_entry(entry, &bucket->list, list) {
 296		if (!match(ref, entry))
 297			continue;
 298
 299		/*
 300		 * Some drivers map the same physical address multiple
 301		 * times. Without a hardware IOMMU this results in the
 302		 * same device addresses being put into the dma-debug
 303		 * hash multiple times too. This can result in false
 304		 * positives being reported. Therefore we implement a
 305		 * best-fit algorithm here which returns the entry from
 306		 * the hash which fits best to the reference value
 307		 * instead of the first-fit.
 308		 */
 309		matches += 1;
 310		match_lvl = 0;
 311		entry->size         == ref->size         ? ++match_lvl : 0;
 312		entry->type         == ref->type         ? ++match_lvl : 0;
 313		entry->direction    == ref->direction    ? ++match_lvl : 0;
 314		entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
 315
 316		if (match_lvl == 4) {
 317			/* perfect-fit - return the result */
 318			return entry;
 319		} else if (match_lvl > last_lvl) {
 320			/*
 321			 * We found an entry that fits better then the
 322			 * previous one or it is the 1st match.
 323			 */
 324			last_lvl = match_lvl;
 325			ret      = entry;
 326		}
 327	}
 328
 329	/*
 330	 * If we have multiple matches but no perfect-fit, just return
 331	 * NULL.
 332	 */
 333	ret = (matches == 1) ? ret : NULL;
 334
 335	return ret;
 336}
 337
 338static struct dma_debug_entry *bucket_find_exact(struct hash_bucket *bucket,
 339						 struct dma_debug_entry *ref)
 340{
 341	return __hash_bucket_find(bucket, ref, exact_match);
 342}
 343
 344static struct dma_debug_entry *bucket_find_contain(struct hash_bucket **bucket,
 345						   struct dma_debug_entry *ref,
 346						   unsigned long *flags)
 347{
 348
 349	unsigned int max_range = dma_get_max_seg_size(ref->dev);
 350	struct dma_debug_entry *entry, index = *ref;
 351	unsigned int range = 0;
 352
 353	while (range <= max_range) {
 354		entry = __hash_bucket_find(*bucket, ref, containing_match);
 355
 356		if (entry)
 357			return entry;
 358
 359		/*
 360		 * Nothing found, go back a hash bucket
 361		 */
 362		put_hash_bucket(*bucket, flags);
 363		range          += (1 << HASH_FN_SHIFT);
 364		index.dev_addr -= (1 << HASH_FN_SHIFT);
 365		*bucket = get_hash_bucket(&index, flags);
 366	}
 367
 368	return NULL;
 369}
 370
 371/*
 372 * Add an entry to a hash bucket
 373 */
 374static void hash_bucket_add(struct hash_bucket *bucket,
 375			    struct dma_debug_entry *entry)
 376{
 377	list_add_tail(&entry->list, &bucket->list);
 378}
 379
 380/*
 381 * Remove entry from a hash bucket list
 382 */
 383static void hash_bucket_del(struct dma_debug_entry *entry)
 384{
 385	list_del(&entry->list);
 386}
 387
 388static unsigned long long phys_addr(struct dma_debug_entry *entry)
 389{
 390	if (entry->type == dma_debug_resource)
 391		return __pfn_to_phys(entry->pfn) + entry->offset;
 392
 393	return page_to_phys(pfn_to_page(entry->pfn)) + entry->offset;
 394}
 395
 396/*
 397 * Dump mapping entries for debugging purposes
 398 */
 399void debug_dma_dump_mappings(struct device *dev)
 400{
 401	int idx;
 402
 403	for (idx = 0; idx < HASH_SIZE; idx++) {
 404		struct hash_bucket *bucket = &dma_entry_hash[idx];
 405		struct dma_debug_entry *entry;
 406		unsigned long flags;
 407
 408		spin_lock_irqsave(&bucket->lock, flags);
 409
 410		list_for_each_entry(entry, &bucket->list, list) {
 411			if (!dev || dev == entry->dev) {
 412				dev_info(entry->dev,
 413					 "%s idx %d P=%Lx N=%lx D=%Lx L=%Lx %s %s\n",
 414					 type2name[entry->type], idx,
 415					 phys_addr(entry), entry->pfn,
 416					 entry->dev_addr, entry->size,
 417					 dir2name[entry->direction],
 418					 maperr2str[entry->map_err_type]);
 419			}
 420		}
 421
 422		spin_unlock_irqrestore(&bucket->lock, flags);
 423	}
 424}
 425
 426/*
 427 * For each mapping (initial cacheline in the case of
 428 * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
 429 * scatterlist, or the cacheline specified in dma_map_single) insert
 430 * into this tree using the cacheline as the key. At
 431 * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
 432 * the entry already exists at insertion time add a tag as a reference
 433 * count for the overlapping mappings.  For now, the overlap tracking
 434 * just ensures that 'unmaps' balance 'maps' before marking the
 435 * cacheline idle, but we should also be flagging overlaps as an API
 436 * violation.
 437 *
 438 * Memory usage is mostly constrained by the maximum number of available
 439 * dma-debug entries in that we need a free dma_debug_entry before
 440 * inserting into the tree.  In the case of dma_map_page and
 441 * dma_alloc_coherent there is only one dma_debug_entry and one
 442 * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
 443 * other hand, consumes a single dma_debug_entry, but inserts 'nents'
 444 * entries into the tree.
 445 *
 446 * At any time debug_dma_assert_idle() can be called to trigger a
 447 * warning if any cachelines in the given page are in the active set.
 448 */
 449static RADIX_TREE(dma_active_cacheline, GFP_NOWAIT);
 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.  Otherwise, if maps and unmaps are
 495	 * balanced then this overflow may cause false negatives in
 496	 * debug_dma_assert_idle() as the cacheline may be marked idle
 497	 * prematurely.
 498	 */
 499	WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
 500		  pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
 501		  ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 502}
 503
 504static int active_cacheline_dec_overlap(phys_addr_t cln)
 505{
 506	int overlap = active_cacheline_read_overlap(cln);
 507
 508	return active_cacheline_set_overlap(cln, --overlap);
 509}
 510
 511static int active_cacheline_insert(struct dma_debug_entry *entry)
 512{
 513	phys_addr_t cln = to_cacheline_number(entry);
 514	unsigned long flags;
 515	int rc;
 516
 517	/* If the device is not writing memory then we don't have any
 518	 * concerns about the cpu consuming stale data.  This mitigates
 519	 * legitimate usages of overlapping mappings.
 520	 */
 521	if (entry->direction == DMA_TO_DEVICE)
 522		return 0;
 523
 524	spin_lock_irqsave(&radix_lock, flags);
 525	rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
 526	if (rc == -EEXIST)
 527		active_cacheline_inc_overlap(cln);
 528	spin_unlock_irqrestore(&radix_lock, flags);
 529
 530	return rc;
 531}
 532
 533static void active_cacheline_remove(struct dma_debug_entry *entry)
 534{
 535	phys_addr_t cln = to_cacheline_number(entry);
 536	unsigned long flags;
 537
 538	/* ...mirror the insert case */
 539	if (entry->direction == DMA_TO_DEVICE)
 540		return;
 541
 542	spin_lock_irqsave(&radix_lock, flags);
 543	/* since we are counting overlaps the final put of the
 544	 * cacheline will occur when the overlap count is 0.
 545	 * active_cacheline_dec_overlap() returns -1 in that case
 546	 */
 547	if (active_cacheline_dec_overlap(cln) < 0)
 548		radix_tree_delete(&dma_active_cacheline, cln);
 549	spin_unlock_irqrestore(&radix_lock, flags);
 550}
 551
 552/**
 553 * debug_dma_assert_idle() - assert that a page is not undergoing dma
 554 * @page: page to lookup in the dma_active_cacheline tree
 555 *
 556 * Place a call to this routine in cases where the cpu touching the page
 557 * before the dma completes (page is dma_unmapped) will lead to data
 558 * corruption.
 559 */
 560void debug_dma_assert_idle(struct page *page)
 561{
 562	static struct dma_debug_entry *ents[CACHELINES_PER_PAGE];
 563	struct dma_debug_entry *entry = NULL;
 564	void **results = (void **) &ents;
 565	unsigned int nents, i;
 566	unsigned long flags;
 567	phys_addr_t cln;
 568
 569	if (dma_debug_disabled())
 570		return;
 571
 572	if (!page)
 573		return;
 574
 575	cln = (phys_addr_t) page_to_pfn(page) << CACHELINE_PER_PAGE_SHIFT;
 576	spin_lock_irqsave(&radix_lock, flags);
 577	nents = radix_tree_gang_lookup(&dma_active_cacheline, results, cln,
 578				       CACHELINES_PER_PAGE);
 579	for (i = 0; i < nents; i++) {
 580		phys_addr_t ent_cln = to_cacheline_number(ents[i]);
 
 
 
 
 
 
 
 
 581
 582		if (ent_cln == cln) {
 583			entry = ents[i];
 584			break;
 585		} else if (ent_cln >= cln + CACHELINES_PER_PAGE)
 586			break;
 587	}
 588	spin_unlock_irqrestore(&radix_lock, flags);
 589
 590	if (!entry)
 591		return;
 
 
 
 
 
 
 
 
 
 
 592
 593	cln = to_cacheline_number(entry);
 594	err_printk(entry->dev, entry,
 595		   "cpu touching an active dma mapped cacheline [cln=%pa]\n",
 596		   &cln);
 
 
 
 
 
 
 
 
 
 
 
 
 597}
 
 598
 599/*
 600 * Wrapper function for adding an entry to the hash.
 601 * This function takes care of locking itself.
 602 */
 603static void add_dma_entry(struct dma_debug_entry *entry)
 604{
 605	struct hash_bucket *bucket;
 606	unsigned long flags;
 607	int rc;
 608
 609	bucket = get_hash_bucket(entry, &flags);
 610	hash_bucket_add(bucket, entry);
 611	put_hash_bucket(bucket, &flags);
 612
 613	rc = active_cacheline_insert(entry);
 614	if (rc == -ENOMEM) {
 615		pr_err("cacheline tracking ENOMEM, dma-debug disabled\n");
 616		global_disable = true;
 
 
 
 617	}
 618
 619	/* TODO: report -EEXIST errors here as overlapping mappings are
 620	 * not supported by the DMA API
 621	 */
 622}
 623
 624static int dma_debug_create_entries(gfp_t gfp)
 625{
 626	struct dma_debug_entry *entry;
 627	int i;
 628
 629	entry = (void *)get_zeroed_page(gfp);
 630	if (!entry)
 631		return -ENOMEM;
 632
 633	for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
 634		list_add_tail(&entry[i].list, &free_entries);
 635
 636	num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 637	nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 638
 639	return 0;
 640}
 641
 642static struct dma_debug_entry *__dma_entry_alloc(void)
 643{
 644	struct dma_debug_entry *entry;
 645
 646	entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 647	list_del(&entry->list);
 648	memset(entry, 0, sizeof(*entry));
 649
 650	num_free_entries -= 1;
 651	if (num_free_entries < min_free_entries)
 652		min_free_entries = num_free_entries;
 653
 654	return entry;
 655}
 656
 657void __dma_entry_alloc_check_leak(void)
 
 
 
 
 658{
 659	u32 tmp = nr_total_entries % nr_prealloc_entries;
 660
 661	/* Shout each time we tick over some multiple of the initial pool */
 662	if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
 663		pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
 664			nr_total_entries,
 665			(nr_total_entries / nr_prealloc_entries));
 666	}
 667}
 668
 669/* struct dma_entry allocator
 670 *
 671 * The next two functions implement the allocator for
 672 * struct dma_debug_entries.
 673 */
 674static struct dma_debug_entry *dma_entry_alloc(void)
 675{
 
 676	struct dma_debug_entry *entry;
 677	unsigned long flags;
 
 678
 679	spin_lock_irqsave(&free_entries_lock, flags);
 680	if (num_free_entries == 0) {
 681		if (dma_debug_create_entries(GFP_ATOMIC)) {
 682			global_disable = true;
 683			spin_unlock_irqrestore(&free_entries_lock, flags);
 684			pr_err("debugging out of memory - disabling\n");
 685			return NULL;
 686		}
 687		__dma_entry_alloc_check_leak();
 
 688	}
 689
 690	entry = __dma_entry_alloc();
 691
 692	spin_unlock_irqrestore(&free_entries_lock, flags);
 693
 
 
 
 694#ifdef CONFIG_STACKTRACE
 695	entry->stack_len = stack_trace_save(entry->stack_entries,
 696					    ARRAY_SIZE(entry->stack_entries),
 697					    1);
 698#endif
 699	return entry;
 700}
 701
 702static void dma_entry_free(struct dma_debug_entry *entry)
 703{
 704	unsigned long flags;
 705
 706	active_cacheline_remove(entry);
 707
 708	/*
 709	 * add to beginning of the list - this way the entries are
 710	 * more likely cache hot when they are reallocated.
 711	 */
 712	spin_lock_irqsave(&free_entries_lock, flags);
 713	list_add(&entry->list, &free_entries);
 714	num_free_entries += 1;
 715	spin_unlock_irqrestore(&free_entries_lock, flags);
 716}
 717
 718/*
 719 * DMA-API debugging init code
 720 *
 721 * The init code does two things:
 722 *   1. Initialize core data structures
 723 *   2. Preallocate a given number of dma_debug_entry structs
 724 */
 725
 726static ssize_t filter_read(struct file *file, char __user *user_buf,
 727			   size_t count, loff_t *ppos)
 728{
 729	char buf[NAME_MAX_LEN + 1];
 730	unsigned long flags;
 731	int len;
 732
 733	if (!current_driver_name[0])
 734		return 0;
 735
 736	/*
 737	 * We can't copy to userspace directly because current_driver_name can
 738	 * only be read under the driver_name_lock with irqs disabled. So
 739	 * create a temporary copy first.
 740	 */
 741	read_lock_irqsave(&driver_name_lock, flags);
 742	len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 743	read_unlock_irqrestore(&driver_name_lock, flags);
 744
 745	return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 746}
 747
 748static ssize_t filter_write(struct file *file, const char __user *userbuf,
 749			    size_t count, loff_t *ppos)
 750{
 751	char buf[NAME_MAX_LEN];
 752	unsigned long flags;
 753	size_t len;
 754	int i;
 755
 756	/*
 757	 * We can't copy from userspace directly. Access to
 758	 * current_driver_name is protected with a write_lock with irqs
 759	 * disabled. Since copy_from_user can fault and may sleep we
 760	 * need to copy to temporary buffer first
 761	 */
 762	len = min(count, (size_t)(NAME_MAX_LEN - 1));
 763	if (copy_from_user(buf, userbuf, len))
 764		return -EFAULT;
 765
 766	buf[len] = 0;
 767
 768	write_lock_irqsave(&driver_name_lock, flags);
 769
 770	/*
 771	 * Now handle the string we got from userspace very carefully.
 772	 * The rules are:
 773	 *         - only use the first token we got
 774	 *         - token delimiter is everything looking like a space
 775	 *           character (' ', '\n', '\t' ...)
 776	 *
 777	 */
 778	if (!isalnum(buf[0])) {
 779		/*
 780		 * If the first character userspace gave us is not
 781		 * alphanumerical then assume the filter should be
 782		 * switched off.
 783		 */
 784		if (current_driver_name[0])
 785			pr_info("switching off dma-debug driver filter\n");
 786		current_driver_name[0] = 0;
 787		current_driver = NULL;
 788		goto out_unlock;
 789	}
 790
 791	/*
 792	 * Now parse out the first token and use it as the name for the
 793	 * driver to filter for.
 794	 */
 795	for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 796		current_driver_name[i] = buf[i];
 797		if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 798			break;
 799	}
 800	current_driver_name[i] = 0;
 801	current_driver = NULL;
 802
 803	pr_info("enable driver filter for driver [%s]\n",
 804		current_driver_name);
 805
 806out_unlock:
 807	write_unlock_irqrestore(&driver_name_lock, flags);
 808
 809	return count;
 810}
 811
 812static const struct file_operations filter_fops = {
 813	.read  = filter_read,
 814	.write = filter_write,
 815	.llseek = default_llseek,
 816};
 817
 818static int dump_show(struct seq_file *seq, void *v)
 819{
 820	int idx;
 821
 822	for (idx = 0; idx < HASH_SIZE; idx++) {
 823		struct hash_bucket *bucket = &dma_entry_hash[idx];
 824		struct dma_debug_entry *entry;
 825		unsigned long flags;
 826
 827		spin_lock_irqsave(&bucket->lock, flags);
 828		list_for_each_entry(entry, &bucket->list, list) {
 829			seq_printf(seq,
 830				   "%s %s %s idx %d P=%llx N=%lx D=%llx L=%llx %s %s\n",
 831				   dev_name(entry->dev),
 832				   dev_driver_string(entry->dev),
 833				   type2name[entry->type], idx,
 834				   phys_addr(entry), entry->pfn,
 835				   entry->dev_addr, entry->size,
 836				   dir2name[entry->direction],
 837				   maperr2str[entry->map_err_type]);
 838		}
 839		spin_unlock_irqrestore(&bucket->lock, flags);
 840	}
 841	return 0;
 842}
 843DEFINE_SHOW_ATTRIBUTE(dump);
 844
 845static void dma_debug_fs_init(void)
 846{
 847	struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
 848
 849	debugfs_create_bool("disabled", 0444, dentry, &global_disable);
 850	debugfs_create_u32("error_count", 0444, dentry, &error_count);
 851	debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
 852	debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
 853	debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
 854	debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
 855	debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
 856	debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
 857	debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
 
 
 858}
 
 859
 860static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 861{
 862	struct dma_debug_entry *entry;
 863	unsigned long flags;
 864	int count = 0, i;
 865
 866	for (i = 0; i < HASH_SIZE; ++i) {
 867		spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
 868		list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 869			if (entry->dev == dev) {
 870				count += 1;
 871				*out_entry = entry;
 872			}
 873		}
 874		spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
 875	}
 876
 877	return count;
 878}
 879
 880static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 881{
 882	struct device *dev = data;
 883	struct dma_debug_entry *uninitialized_var(entry);
 884	int count;
 885
 886	if (dma_debug_disabled())
 887		return 0;
 888
 889	switch (action) {
 890	case BUS_NOTIFY_UNBOUND_DRIVER:
 891		count = device_dma_allocations(dev, &entry);
 892		if (count == 0)
 893			break;
 894		err_printk(dev, entry, "device driver has pending "
 895				"DMA allocations while released from device "
 896				"[count=%d]\n"
 897				"One of leaked entries details: "
 898				"[device address=0x%016llx] [size=%llu bytes] "
 899				"[mapped with %s] [mapped as %s]\n",
 900			count, entry->dev_addr, entry->size,
 901			dir2name[entry->direction], type2name[entry->type]);
 902		break;
 903	default:
 904		break;
 905	}
 906
 907	return 0;
 908}
 909
 910void dma_debug_add_bus(struct bus_type *bus)
 911{
 912	struct notifier_block *nb;
 913
 914	if (dma_debug_disabled())
 915		return;
 916
 917	nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 918	if (nb == NULL) {
 919		pr_err("dma_debug_add_bus: out of memory\n");
 920		return;
 921	}
 922
 923	nb->notifier_call = dma_debug_device_change;
 924
 925	bus_register_notifier(bus, nb);
 926}
 927
 928static int dma_debug_init(void)
 929{
 930	int i, nr_pages;
 931
 932	/* Do not use dma_debug_initialized here, since we really want to be
 933	 * called to set dma_debug_initialized
 934	 */
 935	if (global_disable)
 936		return 0;
 937
 938	for (i = 0; i < HASH_SIZE; ++i) {
 939		INIT_LIST_HEAD(&dma_entry_hash[i].list);
 940		spin_lock_init(&dma_entry_hash[i].lock);
 941	}
 942
 943	dma_debug_fs_init();
 944
 945	nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
 946	for (i = 0; i < nr_pages; ++i)
 947		dma_debug_create_entries(GFP_KERNEL);
 948	if (num_free_entries >= nr_prealloc_entries) {
 949		pr_info("preallocated %d debug entries\n", nr_total_entries);
 950	} else if (num_free_entries > 0) {
 951		pr_warn("%d debug entries requested but only %d allocated\n",
 952			nr_prealloc_entries, nr_total_entries);
 953	} else {
 954		pr_err("debugging out of memory error - disabled\n");
 955		global_disable = true;
 956
 957		return 0;
 958	}
 959	min_free_entries = num_free_entries;
 960
 961	dma_debug_initialized = true;
 962
 963	pr_info("debugging enabled by kernel config\n");
 964	return 0;
 965}
 966core_initcall(dma_debug_init);
 967
 968static __init int dma_debug_cmdline(char *str)
 969{
 970	if (!str)
 971		return -EINVAL;
 972
 973	if (strncmp(str, "off", 3) == 0) {
 974		pr_info("debugging disabled on kernel command line\n");
 975		global_disable = true;
 976	}
 977
 978	return 0;
 979}
 980
 981static __init int dma_debug_entries_cmdline(char *str)
 982{
 983	if (!str)
 984		return -EINVAL;
 985	if (!get_option(&str, &nr_prealloc_entries))
 986		nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 987	return 0;
 988}
 989
 990__setup("dma_debug=", dma_debug_cmdline);
 991__setup("dma_debug_entries=", dma_debug_entries_cmdline);
 992
 993static void check_unmap(struct dma_debug_entry *ref)
 994{
 995	struct dma_debug_entry *entry;
 996	struct hash_bucket *bucket;
 997	unsigned long flags;
 998
 999	bucket = get_hash_bucket(ref, &flags);
1000	entry = bucket_find_exact(bucket, ref);
1001
1002	if (!entry) {
1003		/* must drop lock before calling dma_mapping_error */
1004		put_hash_bucket(bucket, &flags);
1005
1006		if (dma_mapping_error(ref->dev, ref->dev_addr)) {
1007			err_printk(ref->dev, NULL,
1008				   "device driver tries to free an "
1009				   "invalid DMA memory address\n");
1010		} else {
1011			err_printk(ref->dev, NULL,
1012				   "device driver tries to free DMA "
1013				   "memory it has not allocated [device "
1014				   "address=0x%016llx] [size=%llu bytes]\n",
1015				   ref->dev_addr, ref->size);
1016		}
1017		return;
1018	}
1019
1020	if (ref->size != entry->size) {
1021		err_printk(ref->dev, entry, "device driver frees "
1022			   "DMA memory with different size "
1023			   "[device address=0x%016llx] [map size=%llu bytes] "
1024			   "[unmap size=%llu bytes]\n",
1025			   ref->dev_addr, entry->size, ref->size);
1026	}
1027
1028	if (ref->type != entry->type) {
1029		err_printk(ref->dev, entry, "device driver frees "
1030			   "DMA memory with wrong function "
1031			   "[device address=0x%016llx] [size=%llu bytes] "
1032			   "[mapped as %s] [unmapped as %s]\n",
1033			   ref->dev_addr, ref->size,
1034			   type2name[entry->type], type2name[ref->type]);
1035	} else if ((entry->type == dma_debug_coherent) &&
1036		   (phys_addr(ref) != phys_addr(entry))) {
1037		err_printk(ref->dev, entry, "device driver frees "
1038			   "DMA memory with different CPU address "
1039			   "[device address=0x%016llx] [size=%llu bytes] "
1040			   "[cpu alloc address=0x%016llx] "
1041			   "[cpu free address=0x%016llx]",
1042			   ref->dev_addr, ref->size,
1043			   phys_addr(entry),
1044			   phys_addr(ref));
1045	}
1046
1047	if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1048	    ref->sg_call_ents != entry->sg_call_ents) {
1049		err_printk(ref->dev, entry, "device driver frees "
1050			   "DMA sg list with different entry count "
1051			   "[map count=%d] [unmap count=%d]\n",
1052			   entry->sg_call_ents, ref->sg_call_ents);
1053	}
1054
1055	/*
1056	 * This may be no bug in reality - but most implementations of the
1057	 * DMA API don't handle this properly, so check for it here
1058	 */
1059	if (ref->direction != entry->direction) {
1060		err_printk(ref->dev, entry, "device driver frees "
1061			   "DMA memory with different direction "
1062			   "[device address=0x%016llx] [size=%llu bytes] "
1063			   "[mapped with %s] [unmapped with %s]\n",
1064			   ref->dev_addr, ref->size,
1065			   dir2name[entry->direction],
1066			   dir2name[ref->direction]);
1067	}
1068
1069	/*
1070	 * Drivers should use dma_mapping_error() to check the returned
1071	 * addresses of dma_map_single() and dma_map_page().
1072	 * If not, print this warning message. See Documentation/DMA-API.txt.
1073	 */
1074	if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1075		err_printk(ref->dev, entry,
1076			   "device driver failed to check map error"
1077			   "[device address=0x%016llx] [size=%llu bytes] "
1078			   "[mapped as %s]",
1079			   ref->dev_addr, ref->size,
1080			   type2name[entry->type]);
1081	}
1082
1083	hash_bucket_del(entry);
1084	dma_entry_free(entry);
1085
1086	put_hash_bucket(bucket, &flags);
 
 
 
 
1087}
1088
1089static void check_for_stack(struct device *dev,
1090			    struct page *page, size_t offset)
1091{
1092	void *addr;
1093	struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1094
1095	if (!stack_vm_area) {
1096		/* Stack is direct-mapped. */
1097		if (PageHighMem(page))
1098			return;
1099		addr = page_address(page) + offset;
1100		if (object_is_on_stack(addr))
1101			err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1102	} else {
1103		/* Stack is vmalloced. */
1104		int i;
1105
1106		for (i = 0; i < stack_vm_area->nr_pages; i++) {
1107			if (page != stack_vm_area->pages[i])
1108				continue;
1109
1110			addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1111			err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1112			break;
1113		}
1114	}
1115}
1116
1117static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
1118{
1119	unsigned long a1 = (unsigned long)addr;
1120	unsigned long b1 = a1 + len;
1121	unsigned long a2 = (unsigned long)start;
1122	unsigned long b2 = (unsigned long)end;
1123
1124	return !(b1 <= a2 || a1 >= b2);
1125}
1126
1127static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1128{
1129	if (overlap(addr, len, _stext, _etext) ||
1130	    overlap(addr, len, __start_rodata, __end_rodata))
1131		err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1132}
1133
1134static void check_sync(struct device *dev,
1135		       struct dma_debug_entry *ref,
1136		       bool to_cpu)
1137{
1138	struct dma_debug_entry *entry;
1139	struct hash_bucket *bucket;
1140	unsigned long flags;
1141
1142	bucket = get_hash_bucket(ref, &flags);
1143
1144	entry = bucket_find_contain(&bucket, ref, &flags);
1145
1146	if (!entry) {
1147		err_printk(dev, NULL, "device driver tries "
1148				"to sync DMA memory it has not allocated "
1149				"[device address=0x%016llx] [size=%llu bytes]\n",
1150				(unsigned long long)ref->dev_addr, ref->size);
1151		goto out;
1152	}
1153
1154	if (ref->size > entry->size) {
1155		err_printk(dev, entry, "device driver syncs"
1156				" DMA memory outside allocated range "
1157				"[device address=0x%016llx] "
1158				"[allocation size=%llu bytes] "
1159				"[sync offset+size=%llu]\n",
1160				entry->dev_addr, entry->size,
1161				ref->size);
1162	}
1163
1164	if (entry->direction == DMA_BIDIRECTIONAL)
1165		goto out;
1166
1167	if (ref->direction != entry->direction) {
1168		err_printk(dev, entry, "device driver syncs "
1169				"DMA memory with different direction "
1170				"[device address=0x%016llx] [size=%llu bytes] "
1171				"[mapped with %s] [synced with %s]\n",
1172				(unsigned long long)ref->dev_addr, entry->size,
1173				dir2name[entry->direction],
1174				dir2name[ref->direction]);
1175	}
1176
1177	if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1178		      !(ref->direction == DMA_TO_DEVICE))
1179		err_printk(dev, entry, "device driver syncs "
1180				"device read-only DMA memory for cpu "
1181				"[device address=0x%016llx] [size=%llu bytes] "
1182				"[mapped with %s] [synced with %s]\n",
1183				(unsigned long long)ref->dev_addr, entry->size,
1184				dir2name[entry->direction],
1185				dir2name[ref->direction]);
1186
1187	if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1188		       !(ref->direction == DMA_FROM_DEVICE))
1189		err_printk(dev, entry, "device driver syncs "
1190				"device write-only DMA memory to device "
1191				"[device address=0x%016llx] [size=%llu bytes] "
1192				"[mapped with %s] [synced with %s]\n",
1193				(unsigned long long)ref->dev_addr, entry->size,
1194				dir2name[entry->direction],
1195				dir2name[ref->direction]);
1196
1197	if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1198	    ref->sg_call_ents != entry->sg_call_ents) {
1199		err_printk(ref->dev, entry, "device driver syncs "
1200			   "DMA sg list with different entry count "
1201			   "[map count=%d] [sync count=%d]\n",
1202			   entry->sg_call_ents, ref->sg_call_ents);
1203	}
1204
1205out:
1206	put_hash_bucket(bucket, &flags);
1207}
1208
1209static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1210{
1211#ifdef CONFIG_DMA_API_DEBUG_SG
1212	unsigned int max_seg = dma_get_max_seg_size(dev);
1213	u64 start, end, boundary = dma_get_seg_boundary(dev);
1214
1215	/*
1216	 * Either the driver forgot to set dma_parms appropriately, or
1217	 * whoever generated the list forgot to check them.
1218	 */
1219	if (sg->length > max_seg)
1220		err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1221			   sg->length, max_seg);
1222	/*
1223	 * In some cases this could potentially be the DMA API
1224	 * implementation's fault, but it would usually imply that
1225	 * the scatterlist was built inappropriately to begin with.
1226	 */
1227	start = sg_dma_address(sg);
1228	end = start + sg_dma_len(sg) - 1;
1229	if ((start ^ end) & ~boundary)
1230		err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1231			   start, end, boundary);
1232#endif
1233}
1234
1235void debug_dma_map_single(struct device *dev, const void *addr,
1236			    unsigned long len)
1237{
1238	if (unlikely(dma_debug_disabled()))
1239		return;
1240
1241	if (!virt_addr_valid(addr))
1242		err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1243			   addr, len);
1244
1245	if (is_vmalloc_addr(addr))
1246		err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1247			   addr, len);
1248}
1249EXPORT_SYMBOL(debug_dma_map_single);
1250
1251void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1252			size_t size, int direction, dma_addr_t dma_addr)
 
1253{
1254	struct dma_debug_entry *entry;
1255
1256	if (unlikely(dma_debug_disabled()))
1257		return;
1258
1259	if (dma_mapping_error(dev, dma_addr))
1260		return;
1261
1262	entry = dma_entry_alloc();
1263	if (!entry)
1264		return;
1265
1266	entry->dev       = dev;
1267	entry->type      = dma_debug_single;
1268	entry->pfn	 = page_to_pfn(page);
1269	entry->offset	 = offset,
1270	entry->dev_addr  = dma_addr;
1271	entry->size      = size;
1272	entry->direction = direction;
1273	entry->map_err_type = MAP_ERR_NOT_CHECKED;
1274
1275	check_for_stack(dev, page, offset);
1276
1277	if (!PageHighMem(page)) {
1278		void *addr = page_address(page) + offset;
1279
1280		check_for_illegal_area(dev, addr, size);
1281	}
1282
1283	add_dma_entry(entry);
1284}
1285EXPORT_SYMBOL(debug_dma_map_page);
1286
1287void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1288{
1289	struct dma_debug_entry ref;
1290	struct dma_debug_entry *entry;
1291	struct hash_bucket *bucket;
1292	unsigned long flags;
1293
1294	if (unlikely(dma_debug_disabled()))
1295		return;
1296
1297	ref.dev = dev;
1298	ref.dev_addr = dma_addr;
1299	bucket = get_hash_bucket(&ref, &flags);
1300
1301	list_for_each_entry(entry, &bucket->list, list) {
1302		if (!exact_match(&ref, entry))
1303			continue;
1304
1305		/*
1306		 * The same physical address can be mapped multiple
1307		 * times. Without a hardware IOMMU this results in the
1308		 * same device addresses being put into the dma-debug
1309		 * hash multiple times too. This can result in false
1310		 * positives being reported. Therefore we implement a
1311		 * best-fit algorithm here which updates the first entry
1312		 * from the hash which fits the reference value and is
1313		 * not currently listed as being checked.
1314		 */
1315		if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1316			entry->map_err_type = MAP_ERR_CHECKED;
1317			break;
1318		}
1319	}
1320
1321	put_hash_bucket(bucket, &flags);
1322}
1323EXPORT_SYMBOL(debug_dma_mapping_error);
1324
1325void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
1326			  size_t size, int direction)
1327{
1328	struct dma_debug_entry ref = {
1329		.type           = dma_debug_single,
1330		.dev            = dev,
1331		.dev_addr       = addr,
1332		.size           = size,
1333		.direction      = direction,
1334	};
1335
1336	if (unlikely(dma_debug_disabled()))
1337		return;
1338	check_unmap(&ref);
1339}
1340EXPORT_SYMBOL(debug_dma_unmap_page);
1341
1342void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1343		      int nents, int mapped_ents, int direction)
 
1344{
1345	struct dma_debug_entry *entry;
1346	struct scatterlist *s;
1347	int i;
1348
1349	if (unlikely(dma_debug_disabled()))
1350		return;
1351
 
 
 
 
 
 
1352	for_each_sg(sg, s, mapped_ents, i) {
1353		entry = dma_entry_alloc();
1354		if (!entry)
1355			return;
1356
1357		entry->type           = dma_debug_sg;
1358		entry->dev            = dev;
1359		entry->pfn	      = page_to_pfn(sg_page(s));
1360		entry->offset	      = s->offset,
1361		entry->size           = sg_dma_len(s);
1362		entry->dev_addr       = sg_dma_address(s);
1363		entry->direction      = direction;
1364		entry->sg_call_ents   = nents;
1365		entry->sg_mapped_ents = mapped_ents;
1366
1367		check_for_stack(dev, sg_page(s), s->offset);
1368
1369		if (!PageHighMem(sg_page(s))) {
1370			check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
1371		}
1372
1373		check_sg_segment(dev, s);
1374
1375		add_dma_entry(entry);
1376	}
1377}
1378EXPORT_SYMBOL(debug_dma_map_sg);
1379
1380static int get_nr_mapped_entries(struct device *dev,
1381				 struct dma_debug_entry *ref)
1382{
1383	struct dma_debug_entry *entry;
1384	struct hash_bucket *bucket;
1385	unsigned long flags;
1386	int mapped_ents;
1387
1388	bucket       = get_hash_bucket(ref, &flags);
1389	entry        = bucket_find_exact(bucket, ref);
1390	mapped_ents  = 0;
1391
1392	if (entry)
1393		mapped_ents = entry->sg_mapped_ents;
1394	put_hash_bucket(bucket, &flags);
1395
1396	return mapped_ents;
1397}
1398
1399void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1400			int nelems, int dir)
1401{
1402	struct scatterlist *s;
1403	int mapped_ents = 0, i;
1404
1405	if (unlikely(dma_debug_disabled()))
1406		return;
1407
1408	for_each_sg(sglist, s, nelems, i) {
1409
1410		struct dma_debug_entry ref = {
1411			.type           = dma_debug_sg,
1412			.dev            = dev,
1413			.pfn		= page_to_pfn(sg_page(s)),
1414			.offset		= s->offset,
1415			.dev_addr       = sg_dma_address(s),
1416			.size           = sg_dma_len(s),
1417			.direction      = dir,
1418			.sg_call_ents   = nelems,
1419		};
1420
1421		if (mapped_ents && i >= mapped_ents)
1422			break;
1423
1424		if (!i)
1425			mapped_ents = get_nr_mapped_entries(dev, &ref);
1426
1427		check_unmap(&ref);
1428	}
1429}
1430EXPORT_SYMBOL(debug_dma_unmap_sg);
 
 
 
 
 
 
 
 
 
 
 
1431
1432void debug_dma_alloc_coherent(struct device *dev, size_t size,
1433			      dma_addr_t dma_addr, void *virt)
 
1434{
1435	struct dma_debug_entry *entry;
1436
1437	if (unlikely(dma_debug_disabled()))
1438		return;
1439
1440	if (unlikely(virt == NULL))
1441		return;
1442
1443	/* handle vmalloc and linear addresses */
1444	if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1445		return;
1446
1447	entry = dma_entry_alloc();
1448	if (!entry)
1449		return;
1450
1451	entry->type      = dma_debug_coherent;
1452	entry->dev       = dev;
1453	entry->offset	 = offset_in_page(virt);
1454	entry->size      = size;
1455	entry->dev_addr  = dma_addr;
1456	entry->direction = DMA_BIDIRECTIONAL;
1457
1458	if (is_vmalloc_addr(virt))
1459		entry->pfn = vmalloc_to_pfn(virt);
1460	else
1461		entry->pfn = page_to_pfn(virt_to_page(virt));
1462
1463	add_dma_entry(entry);
1464}
1465
1466void debug_dma_free_coherent(struct device *dev, size_t size,
1467			 void *virt, dma_addr_t addr)
1468{
1469	struct dma_debug_entry ref = {
1470		.type           = dma_debug_coherent,
1471		.dev            = dev,
1472		.offset		= offset_in_page(virt),
1473		.dev_addr       = addr,
1474		.size           = size,
1475		.direction      = DMA_BIDIRECTIONAL,
1476	};
1477
1478	/* handle vmalloc and linear addresses */
1479	if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1480		return;
1481
1482	if (is_vmalloc_addr(virt))
1483		ref.pfn = vmalloc_to_pfn(virt);
1484	else
1485		ref.pfn = page_to_pfn(virt_to_page(virt));
1486
1487	if (unlikely(dma_debug_disabled()))
1488		return;
1489
1490	check_unmap(&ref);
1491}
1492
1493void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1494			    int direction, dma_addr_t dma_addr)
 
1495{
1496	struct dma_debug_entry *entry;
1497
1498	if (unlikely(dma_debug_disabled()))
1499		return;
1500
1501	entry = dma_entry_alloc();
1502	if (!entry)
1503		return;
1504
1505	entry->type		= dma_debug_resource;
1506	entry->dev		= dev;
1507	entry->pfn		= PHYS_PFN(addr);
1508	entry->offset		= offset_in_page(addr);
1509	entry->size		= size;
1510	entry->dev_addr		= dma_addr;
1511	entry->direction	= direction;
1512	entry->map_err_type	= MAP_ERR_NOT_CHECKED;
1513
1514	add_dma_entry(entry);
1515}
1516EXPORT_SYMBOL(debug_dma_map_resource);
1517
1518void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1519			      size_t size, int direction)
1520{
1521	struct dma_debug_entry ref = {
1522		.type           = dma_debug_resource,
1523		.dev            = dev,
1524		.dev_addr       = dma_addr,
1525		.size           = size,
1526		.direction      = direction,
1527	};
1528
1529	if (unlikely(dma_debug_disabled()))
1530		return;
1531
1532	check_unmap(&ref);
1533}
1534EXPORT_SYMBOL(debug_dma_unmap_resource);
1535
1536void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1537				   size_t size, int direction)
1538{
1539	struct dma_debug_entry ref;
1540
1541	if (unlikely(dma_debug_disabled()))
1542		return;
1543
1544	ref.type         = dma_debug_single;
1545	ref.dev          = dev;
1546	ref.dev_addr     = dma_handle;
1547	ref.size         = size;
1548	ref.direction    = direction;
1549	ref.sg_call_ents = 0;
1550
1551	check_sync(dev, &ref, true);
1552}
1553EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
1554
1555void debug_dma_sync_single_for_device(struct device *dev,
1556				      dma_addr_t dma_handle, size_t size,
1557				      int direction)
1558{
1559	struct dma_debug_entry ref;
1560
1561	if (unlikely(dma_debug_disabled()))
1562		return;
1563
1564	ref.type         = dma_debug_single;
1565	ref.dev          = dev;
1566	ref.dev_addr     = dma_handle;
1567	ref.size         = size;
1568	ref.direction    = direction;
1569	ref.sg_call_ents = 0;
1570
1571	check_sync(dev, &ref, false);
1572}
1573EXPORT_SYMBOL(debug_dma_sync_single_for_device);
1574
1575void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1576			       int nelems, int direction)
1577{
1578	struct scatterlist *s;
1579	int mapped_ents = 0, i;
1580
1581	if (unlikely(dma_debug_disabled()))
1582		return;
1583
1584	for_each_sg(sg, s, nelems, i) {
1585
1586		struct dma_debug_entry ref = {
1587			.type           = dma_debug_sg,
1588			.dev            = dev,
1589			.pfn		= page_to_pfn(sg_page(s)),
1590			.offset		= s->offset,
1591			.dev_addr       = sg_dma_address(s),
1592			.size           = sg_dma_len(s),
1593			.direction      = direction,
1594			.sg_call_ents   = nelems,
1595		};
1596
1597		if (!i)
1598			mapped_ents = get_nr_mapped_entries(dev, &ref);
1599
1600		if (i >= mapped_ents)
1601			break;
1602
1603		check_sync(dev, &ref, true);
1604	}
1605}
1606EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
1607
1608void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1609				  int nelems, int direction)
1610{
1611	struct scatterlist *s;
1612	int mapped_ents = 0, i;
1613
1614	if (unlikely(dma_debug_disabled()))
1615		return;
1616
1617	for_each_sg(sg, s, nelems, i) {
1618
1619		struct dma_debug_entry ref = {
1620			.type           = dma_debug_sg,
1621			.dev            = dev,
1622			.pfn		= page_to_pfn(sg_page(s)),
1623			.offset		= s->offset,
1624			.dev_addr       = sg_dma_address(s),
1625			.size           = sg_dma_len(s),
1626			.direction      = direction,
1627			.sg_call_ents   = nelems,
1628		};
1629		if (!i)
1630			mapped_ents = get_nr_mapped_entries(dev, &ref);
1631
1632		if (i >= mapped_ents)
1633			break;
1634
1635		check_sync(dev, &ref, false);
1636	}
1637}
1638EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
1639
1640static int __init dma_debug_driver_setup(char *str)
1641{
1642	int i;
1643
1644	for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1645		current_driver_name[i] = *str;
1646		if (*str == 0)
1647			break;
1648	}
1649
1650	if (current_driver_name[0])
1651		pr_info("enable driver filter for driver [%s]\n",
1652			current_driver_name);
1653
1654
1655	return 1;
1656}
1657__setup("dma_debug_driver=", dma_debug_driver_setup);
v6.13.7
   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 * @paddr: physical start address of the mapping
  63 * @map_err_type: track whether dma_mapping_error() was checked
  64 * @stack_len: number of backtrace entries in @stack_entries
  65 * @stack_entries: stack of backtrace history
  66 */
  67struct dma_debug_entry {
  68	struct list_head list;
  69	struct device    *dev;
 
 
 
  70	u64              dev_addr;
  71	u64              size;
  72	int              type;
  73	int              direction;
  74	int		 sg_call_ents;
  75	int		 sg_mapped_ents;
  76	phys_addr_t	 paddr;
  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] = "scatter-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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 390/*
 391 * For each mapping (initial cacheline in the case of
 392 * dma_alloc_coherent/dma_map_page, initial cacheline in each page of a
 393 * scatterlist, or the cacheline specified in dma_map_single) insert
 394 * into this tree using the cacheline as the key. At
 395 * dma_unmap_{single|sg|page} or dma_free_coherent delete the entry.  If
 396 * the entry already exists at insertion time add a tag as a reference
 397 * count for the overlapping mappings.  For now, the overlap tracking
 398 * just ensures that 'unmaps' balance 'maps' before marking the
 399 * cacheline idle, but we should also be flagging overlaps as an API
 400 * violation.
 401 *
 402 * Memory usage is mostly constrained by the maximum number of available
 403 * dma-debug entries in that we need a free dma_debug_entry before
 404 * inserting into the tree.  In the case of dma_map_page and
 405 * dma_alloc_coherent there is only one dma_debug_entry and one
 406 * dma_active_cacheline entry to track per event.  dma_map_sg(), on the
 407 * other hand, consumes a single dma_debug_entry, but inserts 'nents'
 408 * entries into the tree.
 409 *
 410 * Use __GFP_NOWARN because the printk from an OOM, to netconsole, could end
 411 * up right back in the DMA debugging code, leading to a deadlock.
 412 */
 413static RADIX_TREE(dma_active_cacheline, GFP_ATOMIC | __GFP_NOWARN);
 414static DEFINE_SPINLOCK(radix_lock);
 415#define ACTIVE_CACHELINE_MAX_OVERLAP ((1 << RADIX_TREE_MAX_TAGS) - 1)
 416#define CACHELINE_PER_PAGE_SHIFT (PAGE_SHIFT - L1_CACHE_SHIFT)
 417#define CACHELINES_PER_PAGE (1 << CACHELINE_PER_PAGE_SHIFT)
 418
 419static phys_addr_t to_cacheline_number(struct dma_debug_entry *entry)
 420{
 421	return ((entry->paddr >> PAGE_SHIFT) << CACHELINE_PER_PAGE_SHIFT) +
 422		(offset_in_page(entry->paddr) >> L1_CACHE_SHIFT);
 423}
 424
 425static int active_cacheline_read_overlap(phys_addr_t cln)
 426{
 427	int overlap = 0, i;
 428
 429	for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 430		if (radix_tree_tag_get(&dma_active_cacheline, cln, i))
 431			overlap |= 1 << i;
 432	return overlap;
 433}
 434
 435static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
 436{
 437	int i;
 438
 439	if (overlap > ACTIVE_CACHELINE_MAX_OVERLAP || overlap < 0)
 440		return overlap;
 441
 442	for (i = RADIX_TREE_MAX_TAGS - 1; i >= 0; i--)
 443		if (overlap & 1 << i)
 444			radix_tree_tag_set(&dma_active_cacheline, cln, i);
 445		else
 446			radix_tree_tag_clear(&dma_active_cacheline, cln, i);
 447
 448	return overlap;
 449}
 450
 451static void active_cacheline_inc_overlap(phys_addr_t cln)
 452{
 453	int overlap = active_cacheline_read_overlap(cln);
 454
 455	overlap = active_cacheline_set_overlap(cln, ++overlap);
 456
 457	/* If we overflowed the overlap counter then we're potentially
 458	 * leaking dma-mappings.
 
 
 
 459	 */
 460	WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
 461		  pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
 462		  ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 463}
 464
 465static int active_cacheline_dec_overlap(phys_addr_t cln)
 466{
 467	int overlap = active_cacheline_read_overlap(cln);
 468
 469	return active_cacheline_set_overlap(cln, --overlap);
 470}
 471
 472static int active_cacheline_insert(struct dma_debug_entry *entry)
 473{
 474	phys_addr_t cln = to_cacheline_number(entry);
 475	unsigned long flags;
 476	int rc;
 477
 478	/* If the device is not writing memory then we don't have any
 479	 * concerns about the cpu consuming stale data.  This mitigates
 480	 * legitimate usages of overlapping mappings.
 481	 */
 482	if (entry->direction == DMA_TO_DEVICE)
 483		return 0;
 484
 485	spin_lock_irqsave(&radix_lock, flags);
 486	rc = radix_tree_insert(&dma_active_cacheline, cln, entry);
 487	if (rc == -EEXIST)
 488		active_cacheline_inc_overlap(cln);
 489	spin_unlock_irqrestore(&radix_lock, flags);
 490
 491	return rc;
 492}
 493
 494static void active_cacheline_remove(struct dma_debug_entry *entry)
 495{
 496	phys_addr_t cln = to_cacheline_number(entry);
 497	unsigned long flags;
 498
 499	/* ...mirror the insert case */
 500	if (entry->direction == DMA_TO_DEVICE)
 501		return;
 502
 503	spin_lock_irqsave(&radix_lock, flags);
 504	/* since we are counting overlaps the final put of the
 505	 * cacheline will occur when the overlap count is 0.
 506	 * active_cacheline_dec_overlap() returns -1 in that case
 507	 */
 508	if (active_cacheline_dec_overlap(cln) < 0)
 509		radix_tree_delete(&dma_active_cacheline, cln);
 510	spin_unlock_irqrestore(&radix_lock, flags);
 511}
 512
 513/*
 514 * Dump mappings entries on kernel space for debugging purposes
 
 
 
 
 
 515 */
 516void debug_dma_dump_mappings(struct device *dev)
 517{
 518	int idx;
 
 
 
 
 519	phys_addr_t cln;
 520
 521	for (idx = 0; idx < HASH_SIZE; idx++) {
 522		struct hash_bucket *bucket = &dma_entry_hash[idx];
 523		struct dma_debug_entry *entry;
 524		unsigned long flags;
 
 525
 526		spin_lock_irqsave(&bucket->lock, flags);
 527		list_for_each_entry(entry, &bucket->list, list) {
 528			if (!dev || dev == entry->dev) {
 529				cln = to_cacheline_number(entry);
 530				dev_info(entry->dev,
 531					 "%s idx %d P=%pa D=%llx L=%llx cln=%pa %s %s\n",
 532					 type2name[entry->type], idx,
 533					 &entry->paddr, entry->dev_addr,
 534					 entry->size, &cln,
 535					 dir2name[entry->direction],
 536					 maperr2str[entry->map_err_type]);
 537			}
 538		}
 539		spin_unlock_irqrestore(&bucket->lock, flags);
 540
 541		cond_resched();
 
 
 
 
 542	}
 543}
 544
 545/*
 546 * Dump mappings entries on user space via debugfs
 547 */
 548static int dump_show(struct seq_file *seq, void *v)
 549{
 550	int idx;
 551	phys_addr_t cln;
 552
 553	for (idx = 0; idx < HASH_SIZE; idx++) {
 554		struct hash_bucket *bucket = &dma_entry_hash[idx];
 555		struct dma_debug_entry *entry;
 556		unsigned long flags;
 557
 558		spin_lock_irqsave(&bucket->lock, flags);
 559		list_for_each_entry(entry, &bucket->list, list) {
 560			cln = to_cacheline_number(entry);
 561			seq_printf(seq,
 562				   "%s %s %s idx %d P=%pa D=%llx L=%llx cln=%pa %s %s\n",
 563				   dev_driver_string(entry->dev),
 564				   dev_name(entry->dev),
 565				   type2name[entry->type], idx,
 566				   &entry->paddr, entry->dev_addr,
 567				   entry->size, &cln,
 568				   dir2name[entry->direction],
 569				   maperr2str[entry->map_err_type]);
 570		}
 571		spin_unlock_irqrestore(&bucket->lock, flags);
 572	}
 573	return 0;
 574}
 575DEFINE_SHOW_ATTRIBUTE(dump);
 576
 577/*
 578 * Wrapper function for adding an entry to the hash.
 579 * This function takes care of locking itself.
 580 */
 581static void add_dma_entry(struct dma_debug_entry *entry, unsigned long attrs)
 582{
 583	struct hash_bucket *bucket;
 584	unsigned long flags;
 585	int rc;
 586
 587	bucket = get_hash_bucket(entry, &flags);
 588	hash_bucket_add(bucket, entry);
 589	put_hash_bucket(bucket, flags);
 590
 591	rc = active_cacheline_insert(entry);
 592	if (rc == -ENOMEM) {
 593		pr_err_once("cacheline tracking ENOMEM, dma-debug disabled\n");
 594		global_disable = true;
 595	} else if (rc == -EEXIST && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
 596		err_printk(entry->dev, entry,
 597			"cacheline tracking EEXIST, overlapping mappings aren't supported\n");
 598	}
 
 
 
 
 599}
 600
 601static int dma_debug_create_entries(gfp_t gfp)
 602{
 603	struct dma_debug_entry *entry;
 604	int i;
 605
 606	entry = (void *)get_zeroed_page(gfp);
 607	if (!entry)
 608		return -ENOMEM;
 609
 610	for (i = 0; i < DMA_DEBUG_DYNAMIC_ENTRIES; i++)
 611		list_add_tail(&entry[i].list, &free_entries);
 612
 613	num_free_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 614	nr_total_entries += DMA_DEBUG_DYNAMIC_ENTRIES;
 615
 616	return 0;
 617}
 618
 619static struct dma_debug_entry *__dma_entry_alloc(void)
 620{
 621	struct dma_debug_entry *entry;
 622
 623	entry = list_entry(free_entries.next, struct dma_debug_entry, list);
 624	list_del(&entry->list);
 625	memset(entry, 0, sizeof(*entry));
 626
 627	num_free_entries -= 1;
 628	if (num_free_entries < min_free_entries)
 629		min_free_entries = num_free_entries;
 630
 631	return entry;
 632}
 633
 634/*
 635 * This should be called outside of free_entries_lock scope to avoid potential
 636 * deadlocks with serial consoles that use DMA.
 637 */
 638static void __dma_entry_alloc_check_leak(u32 nr_entries)
 639{
 640	u32 tmp = nr_entries % nr_prealloc_entries;
 641
 642	/* Shout each time we tick over some multiple of the initial pool */
 643	if (tmp < DMA_DEBUG_DYNAMIC_ENTRIES) {
 644		pr_info("dma_debug_entry pool grown to %u (%u00%%)\n",
 645			nr_entries,
 646			(nr_entries / nr_prealloc_entries));
 647	}
 648}
 649
 650/* struct dma_entry allocator
 651 *
 652 * The next two functions implement the allocator for
 653 * struct dma_debug_entries.
 654 */
 655static struct dma_debug_entry *dma_entry_alloc(void)
 656{
 657	bool alloc_check_leak = false;
 658	struct dma_debug_entry *entry;
 659	unsigned long flags;
 660	u32 nr_entries;
 661
 662	spin_lock_irqsave(&free_entries_lock, flags);
 663	if (num_free_entries == 0) {
 664		if (dma_debug_create_entries(GFP_ATOMIC)) {
 665			global_disable = true;
 666			spin_unlock_irqrestore(&free_entries_lock, flags);
 667			pr_err("debugging out of memory - disabling\n");
 668			return NULL;
 669		}
 670		alloc_check_leak = true;
 671		nr_entries = nr_total_entries;
 672	}
 673
 674	entry = __dma_entry_alloc();
 675
 676	spin_unlock_irqrestore(&free_entries_lock, flags);
 677
 678	if (alloc_check_leak)
 679		__dma_entry_alloc_check_leak(nr_entries);
 680
 681#ifdef CONFIG_STACKTRACE
 682	entry->stack_len = stack_trace_save(entry->stack_entries,
 683					    ARRAY_SIZE(entry->stack_entries),
 684					    1);
 685#endif
 686	return entry;
 687}
 688
 689static void dma_entry_free(struct dma_debug_entry *entry)
 690{
 691	unsigned long flags;
 692
 693	active_cacheline_remove(entry);
 694
 695	/*
 696	 * add to beginning of the list - this way the entries are
 697	 * more likely cache hot when they are reallocated.
 698	 */
 699	spin_lock_irqsave(&free_entries_lock, flags);
 700	list_add(&entry->list, &free_entries);
 701	num_free_entries += 1;
 702	spin_unlock_irqrestore(&free_entries_lock, flags);
 703}
 704
 705/*
 706 * DMA-API debugging init code
 707 *
 708 * The init code does two things:
 709 *   1. Initialize core data structures
 710 *   2. Preallocate a given number of dma_debug_entry structs
 711 */
 712
 713static ssize_t filter_read(struct file *file, char __user *user_buf,
 714			   size_t count, loff_t *ppos)
 715{
 716	char buf[NAME_MAX_LEN + 1];
 717	unsigned long flags;
 718	int len;
 719
 720	if (!current_driver_name[0])
 721		return 0;
 722
 723	/*
 724	 * We can't copy to userspace directly because current_driver_name can
 725	 * only be read under the driver_name_lock with irqs disabled. So
 726	 * create a temporary copy first.
 727	 */
 728	read_lock_irqsave(&driver_name_lock, flags);
 729	len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
 730	read_unlock_irqrestore(&driver_name_lock, flags);
 731
 732	return simple_read_from_buffer(user_buf, count, ppos, buf, len);
 733}
 734
 735static ssize_t filter_write(struct file *file, const char __user *userbuf,
 736			    size_t count, loff_t *ppos)
 737{
 738	char buf[NAME_MAX_LEN];
 739	unsigned long flags;
 740	size_t len;
 741	int i;
 742
 743	/*
 744	 * We can't copy from userspace directly. Access to
 745	 * current_driver_name is protected with a write_lock with irqs
 746	 * disabled. Since copy_from_user can fault and may sleep we
 747	 * need to copy to temporary buffer first
 748	 */
 749	len = min(count, (size_t)(NAME_MAX_LEN - 1));
 750	if (copy_from_user(buf, userbuf, len))
 751		return -EFAULT;
 752
 753	buf[len] = 0;
 754
 755	write_lock_irqsave(&driver_name_lock, flags);
 756
 757	/*
 758	 * Now handle the string we got from userspace very carefully.
 759	 * The rules are:
 760	 *         - only use the first token we got
 761	 *         - token delimiter is everything looking like a space
 762	 *           character (' ', '\n', '\t' ...)
 763	 *
 764	 */
 765	if (!isalnum(buf[0])) {
 766		/*
 767		 * If the first character userspace gave us is not
 768		 * alphanumerical then assume the filter should be
 769		 * switched off.
 770		 */
 771		if (current_driver_name[0])
 772			pr_info("switching off dma-debug driver filter\n");
 773		current_driver_name[0] = 0;
 774		current_driver = NULL;
 775		goto out_unlock;
 776	}
 777
 778	/*
 779	 * Now parse out the first token and use it as the name for the
 780	 * driver to filter for.
 781	 */
 782	for (i = 0; i < NAME_MAX_LEN - 1; ++i) {
 783		current_driver_name[i] = buf[i];
 784		if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
 785			break;
 786	}
 787	current_driver_name[i] = 0;
 788	current_driver = NULL;
 789
 790	pr_info("enable driver filter for driver [%s]\n",
 791		current_driver_name);
 792
 793out_unlock:
 794	write_unlock_irqrestore(&driver_name_lock, flags);
 795
 796	return count;
 797}
 798
 799static const struct file_operations filter_fops = {
 800	.read  = filter_read,
 801	.write = filter_write,
 802	.llseek = default_llseek,
 803};
 804
 805static int __init dma_debug_fs_init(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 806{
 807	struct dentry *dentry = debugfs_create_dir("dma-api", NULL);
 808
 809	debugfs_create_bool("disabled", 0444, dentry, &global_disable);
 810	debugfs_create_u32("error_count", 0444, dentry, &error_count);
 811	debugfs_create_u32("all_errors", 0644, dentry, &show_all_errors);
 812	debugfs_create_u32("num_errors", 0644, dentry, &show_num_errors);
 813	debugfs_create_u32("num_free_entries", 0444, dentry, &num_free_entries);
 814	debugfs_create_u32("min_free_entries", 0444, dentry, &min_free_entries);
 815	debugfs_create_u32("nr_total_entries", 0444, dentry, &nr_total_entries);
 816	debugfs_create_file("driver_filter", 0644, dentry, NULL, &filter_fops);
 817	debugfs_create_file("dump", 0444, dentry, NULL, &dump_fops);
 818
 819	return 0;
 820}
 821core_initcall_sync(dma_debug_fs_init);
 822
 823static int device_dma_allocations(struct device *dev, struct dma_debug_entry **out_entry)
 824{
 825	struct dma_debug_entry *entry;
 826	unsigned long flags;
 827	int count = 0, i;
 828
 829	for (i = 0; i < HASH_SIZE; ++i) {
 830		spin_lock_irqsave(&dma_entry_hash[i].lock, flags);
 831		list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
 832			if (entry->dev == dev) {
 833				count += 1;
 834				*out_entry = entry;
 835			}
 836		}
 837		spin_unlock_irqrestore(&dma_entry_hash[i].lock, flags);
 838	}
 839
 840	return count;
 841}
 842
 843static int dma_debug_device_change(struct notifier_block *nb, unsigned long action, void *data)
 844{
 845	struct device *dev = data;
 846	struct dma_debug_entry *entry;
 847	int count;
 848
 849	if (dma_debug_disabled())
 850		return 0;
 851
 852	switch (action) {
 853	case BUS_NOTIFY_UNBOUND_DRIVER:
 854		count = device_dma_allocations(dev, &entry);
 855		if (count == 0)
 856			break;
 857		err_printk(dev, entry, "device driver has pending "
 858				"DMA allocations while released from device "
 859				"[count=%d]\n"
 860				"One of leaked entries details: "
 861				"[device address=0x%016llx] [size=%llu bytes] "
 862				"[mapped with %s] [mapped as %s]\n",
 863			count, entry->dev_addr, entry->size,
 864			dir2name[entry->direction], type2name[entry->type]);
 865		break;
 866	default:
 867		break;
 868	}
 869
 870	return 0;
 871}
 872
 873void dma_debug_add_bus(const struct bus_type *bus)
 874{
 875	struct notifier_block *nb;
 876
 877	if (dma_debug_disabled())
 878		return;
 879
 880	nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
 881	if (nb == NULL) {
 882		pr_err("dma_debug_add_bus: out of memory\n");
 883		return;
 884	}
 885
 886	nb->notifier_call = dma_debug_device_change;
 887
 888	bus_register_notifier(bus, nb);
 889}
 890
 891static int dma_debug_init(void)
 892{
 893	int i, nr_pages;
 894
 895	/* Do not use dma_debug_initialized here, since we really want to be
 896	 * called to set dma_debug_initialized
 897	 */
 898	if (global_disable)
 899		return 0;
 900
 901	for (i = 0; i < HASH_SIZE; ++i) {
 902		INIT_LIST_HEAD(&dma_entry_hash[i].list);
 903		spin_lock_init(&dma_entry_hash[i].lock);
 904	}
 905
 
 
 906	nr_pages = DIV_ROUND_UP(nr_prealloc_entries, DMA_DEBUG_DYNAMIC_ENTRIES);
 907	for (i = 0; i < nr_pages; ++i)
 908		dma_debug_create_entries(GFP_KERNEL);
 909	if (num_free_entries >= nr_prealloc_entries) {
 910		pr_info("preallocated %d debug entries\n", nr_total_entries);
 911	} else if (num_free_entries > 0) {
 912		pr_warn("%d debug entries requested but only %d allocated\n",
 913			nr_prealloc_entries, nr_total_entries);
 914	} else {
 915		pr_err("debugging out of memory error - disabled\n");
 916		global_disable = true;
 917
 918		return 0;
 919	}
 920	min_free_entries = num_free_entries;
 921
 922	dma_debug_initialized = true;
 923
 924	pr_info("debugging enabled by kernel config\n");
 925	return 0;
 926}
 927core_initcall(dma_debug_init);
 928
 929static __init int dma_debug_cmdline(char *str)
 930{
 931	if (!str)
 932		return -EINVAL;
 933
 934	if (strncmp(str, "off", 3) == 0) {
 935		pr_info("debugging disabled on kernel command line\n");
 936		global_disable = true;
 937	}
 938
 939	return 1;
 940}
 941
 942static __init int dma_debug_entries_cmdline(char *str)
 943{
 944	if (!str)
 945		return -EINVAL;
 946	if (!get_option(&str, &nr_prealloc_entries))
 947		nr_prealloc_entries = PREALLOC_DMA_DEBUG_ENTRIES;
 948	return 1;
 949}
 950
 951__setup("dma_debug=", dma_debug_cmdline);
 952__setup("dma_debug_entries=", dma_debug_entries_cmdline);
 953
 954static void check_unmap(struct dma_debug_entry *ref)
 955{
 956	struct dma_debug_entry *entry;
 957	struct hash_bucket *bucket;
 958	unsigned long flags;
 959
 960	bucket = get_hash_bucket(ref, &flags);
 961	entry = bucket_find_exact(bucket, ref);
 962
 963	if (!entry) {
 964		/* must drop lock before calling dma_mapping_error */
 965		put_hash_bucket(bucket, flags);
 966
 967		if (dma_mapping_error(ref->dev, ref->dev_addr)) {
 968			err_printk(ref->dev, NULL,
 969				   "device driver tries to free an "
 970				   "invalid DMA memory address\n");
 971		} else {
 972			err_printk(ref->dev, NULL,
 973				   "device driver tries to free DMA "
 974				   "memory it has not allocated [device "
 975				   "address=0x%016llx] [size=%llu bytes]\n",
 976				   ref->dev_addr, ref->size);
 977		}
 978		return;
 979	}
 980
 981	if (ref->size != entry->size) {
 982		err_printk(ref->dev, entry, "device driver frees "
 983			   "DMA memory with different size "
 984			   "[device address=0x%016llx] [map size=%llu bytes] "
 985			   "[unmap size=%llu bytes]\n",
 986			   ref->dev_addr, entry->size, ref->size);
 987	}
 988
 989	if (ref->type != entry->type) {
 990		err_printk(ref->dev, entry, "device driver frees "
 991			   "DMA memory with wrong function "
 992			   "[device address=0x%016llx] [size=%llu bytes] "
 993			   "[mapped as %s] [unmapped as %s]\n",
 994			   ref->dev_addr, ref->size,
 995			   type2name[entry->type], type2name[ref->type]);
 996	} else if (entry->type == dma_debug_coherent &&
 997		   ref->paddr != entry->paddr) {
 998		err_printk(ref->dev, entry, "device driver frees "
 999			   "DMA memory with different CPU address "
1000			   "[device address=0x%016llx] [size=%llu bytes] "
1001			   "[cpu alloc address=0x%pa] "
1002			   "[cpu free address=0x%pa]",
1003			   ref->dev_addr, ref->size,
1004			   &entry->paddr,
1005			   &ref->paddr);
1006	}
1007
1008	if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1009	    ref->sg_call_ents != entry->sg_call_ents) {
1010		err_printk(ref->dev, entry, "device driver frees "
1011			   "DMA sg list with different entry count "
1012			   "[map count=%d] [unmap count=%d]\n",
1013			   entry->sg_call_ents, ref->sg_call_ents);
1014	}
1015
1016	/*
1017	 * This may be no bug in reality - but most implementations of the
1018	 * DMA API don't handle this properly, so check for it here
1019	 */
1020	if (ref->direction != entry->direction) {
1021		err_printk(ref->dev, entry, "device driver frees "
1022			   "DMA memory with different direction "
1023			   "[device address=0x%016llx] [size=%llu bytes] "
1024			   "[mapped with %s] [unmapped with %s]\n",
1025			   ref->dev_addr, ref->size,
1026			   dir2name[entry->direction],
1027			   dir2name[ref->direction]);
1028	}
1029
1030	/*
1031	 * Drivers should use dma_mapping_error() to check the returned
1032	 * addresses of dma_map_single() and dma_map_page().
1033	 * If not, print this warning message. See Documentation/core-api/dma-api.rst.
1034	 */
1035	if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1036		err_printk(ref->dev, entry,
1037			   "device driver failed to check map error"
1038			   "[device address=0x%016llx] [size=%llu bytes] "
1039			   "[mapped as %s]",
1040			   ref->dev_addr, ref->size,
1041			   type2name[entry->type]);
1042	}
1043
1044	hash_bucket_del(entry);
1045	put_hash_bucket(bucket, flags);
1046
1047	/*
1048	 * Free the entry outside of bucket_lock to avoid ABBA deadlocks
1049	 * between that and radix_lock.
1050	 */
1051	dma_entry_free(entry);
1052}
1053
1054static void check_for_stack(struct device *dev,
1055			    struct page *page, size_t offset)
1056{
1057	void *addr;
1058	struct vm_struct *stack_vm_area = task_stack_vm_area(current);
1059
1060	if (!stack_vm_area) {
1061		/* Stack is direct-mapped. */
1062		if (PageHighMem(page))
1063			return;
1064		addr = page_address(page) + offset;
1065		if (object_is_on_stack(addr))
1066			err_printk(dev, NULL, "device driver maps memory from stack [addr=%p]\n", addr);
1067	} else {
1068		/* Stack is vmalloced. */
1069		int i;
1070
1071		for (i = 0; i < stack_vm_area->nr_pages; i++) {
1072			if (page != stack_vm_area->pages[i])
1073				continue;
1074
1075			addr = (u8 *)current->stack + i * PAGE_SIZE + offset;
1076			err_printk(dev, NULL, "device driver maps memory from stack [probable addr=%p]\n", addr);
1077			break;
1078		}
1079	}
1080}
1081
 
 
 
 
 
 
 
 
 
 
1082static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
1083{
1084	if (memory_intersects(_stext, _etext, addr, len) ||
1085	    memory_intersects(__start_rodata, __end_rodata, addr, len))
1086		err_printk(dev, NULL, "device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
1087}
1088
1089static void check_sync(struct device *dev,
1090		       struct dma_debug_entry *ref,
1091		       bool to_cpu)
1092{
1093	struct dma_debug_entry *entry;
1094	struct hash_bucket *bucket;
1095	unsigned long flags;
1096
1097	bucket = get_hash_bucket(ref, &flags);
1098
1099	entry = bucket_find_contain(&bucket, ref, &flags);
1100
1101	if (!entry) {
1102		err_printk(dev, NULL, "device driver tries "
1103				"to sync DMA memory it has not allocated "
1104				"[device address=0x%016llx] [size=%llu bytes]\n",
1105				(unsigned long long)ref->dev_addr, ref->size);
1106		goto out;
1107	}
1108
1109	if (ref->size > entry->size) {
1110		err_printk(dev, entry, "device driver syncs"
1111				" DMA memory outside allocated range "
1112				"[device address=0x%016llx] "
1113				"[allocation size=%llu bytes] "
1114				"[sync offset+size=%llu]\n",
1115				entry->dev_addr, entry->size,
1116				ref->size);
1117	}
1118
1119	if (entry->direction == DMA_BIDIRECTIONAL)
1120		goto out;
1121
1122	if (ref->direction != entry->direction) {
1123		err_printk(dev, entry, "device driver syncs "
1124				"DMA memory with different direction "
1125				"[device address=0x%016llx] [size=%llu bytes] "
1126				"[mapped with %s] [synced with %s]\n",
1127				(unsigned long long)ref->dev_addr, entry->size,
1128				dir2name[entry->direction],
1129				dir2name[ref->direction]);
1130	}
1131
1132	if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
1133		      !(ref->direction == DMA_TO_DEVICE))
1134		err_printk(dev, entry, "device driver syncs "
1135				"device read-only DMA memory for cpu "
1136				"[device address=0x%016llx] [size=%llu bytes] "
1137				"[mapped with %s] [synced with %s]\n",
1138				(unsigned long long)ref->dev_addr, entry->size,
1139				dir2name[entry->direction],
1140				dir2name[ref->direction]);
1141
1142	if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
1143		       !(ref->direction == DMA_FROM_DEVICE))
1144		err_printk(dev, entry, "device driver syncs "
1145				"device write-only DMA memory to device "
1146				"[device address=0x%016llx] [size=%llu bytes] "
1147				"[mapped with %s] [synced with %s]\n",
1148				(unsigned long long)ref->dev_addr, entry->size,
1149				dir2name[entry->direction],
1150				dir2name[ref->direction]);
1151
1152	if (ref->sg_call_ents && ref->type == dma_debug_sg &&
1153	    ref->sg_call_ents != entry->sg_call_ents) {
1154		err_printk(ref->dev, entry, "device driver syncs "
1155			   "DMA sg list with different entry count "
1156			   "[map count=%d] [sync count=%d]\n",
1157			   entry->sg_call_ents, ref->sg_call_ents);
1158	}
1159
1160out:
1161	put_hash_bucket(bucket, flags);
1162}
1163
1164static void check_sg_segment(struct device *dev, struct scatterlist *sg)
1165{
 
1166	unsigned int max_seg = dma_get_max_seg_size(dev);
1167	u64 start, end, boundary = dma_get_seg_boundary(dev);
1168
1169	/*
1170	 * Either the driver forgot to set dma_parms appropriately, or
1171	 * whoever generated the list forgot to check them.
1172	 */
1173	if (sg->length > max_seg)
1174		err_printk(dev, NULL, "mapping sg segment longer than device claims to support [len=%u] [max=%u]\n",
1175			   sg->length, max_seg);
1176	/*
1177	 * In some cases this could potentially be the DMA API
1178	 * implementation's fault, but it would usually imply that
1179	 * the scatterlist was built inappropriately to begin with.
1180	 */
1181	start = sg_dma_address(sg);
1182	end = start + sg_dma_len(sg) - 1;
1183	if ((start ^ end) & ~boundary)
1184		err_printk(dev, NULL, "mapping sg segment across boundary [start=0x%016llx] [end=0x%016llx] [boundary=0x%016llx]\n",
1185			   start, end, boundary);
 
1186}
1187
1188void debug_dma_map_single(struct device *dev, const void *addr,
1189			    unsigned long len)
1190{
1191	if (unlikely(dma_debug_disabled()))
1192		return;
1193
1194	if (!virt_addr_valid(addr))
1195		err_printk(dev, NULL, "device driver maps memory from invalid area [addr=%p] [len=%lu]\n",
1196			   addr, len);
1197
1198	if (is_vmalloc_addr(addr))
1199		err_printk(dev, NULL, "device driver maps memory from vmalloc area [addr=%p] [len=%lu]\n",
1200			   addr, len);
1201}
1202EXPORT_SYMBOL(debug_dma_map_single);
1203
1204void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
1205			size_t size, int direction, dma_addr_t dma_addr,
1206			unsigned long attrs)
1207{
1208	struct dma_debug_entry *entry;
1209
1210	if (unlikely(dma_debug_disabled()))
1211		return;
1212
1213	if (dma_mapping_error(dev, dma_addr))
1214		return;
1215
1216	entry = dma_entry_alloc();
1217	if (!entry)
1218		return;
1219
1220	entry->dev       = dev;
1221	entry->type      = dma_debug_single;
1222	entry->paddr	 = page_to_phys(page) + offset;
 
1223	entry->dev_addr  = dma_addr;
1224	entry->size      = size;
1225	entry->direction = direction;
1226	entry->map_err_type = MAP_ERR_NOT_CHECKED;
1227
1228	check_for_stack(dev, page, offset);
1229
1230	if (!PageHighMem(page)) {
1231		void *addr = page_address(page) + offset;
1232
1233		check_for_illegal_area(dev, addr, size);
1234	}
1235
1236	add_dma_entry(entry, attrs);
1237}
 
1238
1239void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
1240{
1241	struct dma_debug_entry ref;
1242	struct dma_debug_entry *entry;
1243	struct hash_bucket *bucket;
1244	unsigned long flags;
1245
1246	if (unlikely(dma_debug_disabled()))
1247		return;
1248
1249	ref.dev = dev;
1250	ref.dev_addr = dma_addr;
1251	bucket = get_hash_bucket(&ref, &flags);
1252
1253	list_for_each_entry(entry, &bucket->list, list) {
1254		if (!exact_match(&ref, entry))
1255			continue;
1256
1257		/*
1258		 * The same physical address can be mapped multiple
1259		 * times. Without a hardware IOMMU this results in the
1260		 * same device addresses being put into the dma-debug
1261		 * hash multiple times too. This can result in false
1262		 * positives being reported. Therefore we implement a
1263		 * best-fit algorithm here which updates the first entry
1264		 * from the hash which fits the reference value and is
1265		 * not currently listed as being checked.
1266		 */
1267		if (entry->map_err_type == MAP_ERR_NOT_CHECKED) {
1268			entry->map_err_type = MAP_ERR_CHECKED;
1269			break;
1270		}
1271	}
1272
1273	put_hash_bucket(bucket, flags);
1274}
1275EXPORT_SYMBOL(debug_dma_mapping_error);
1276
1277void debug_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
1278			  size_t size, int direction)
1279{
1280	struct dma_debug_entry ref = {
1281		.type           = dma_debug_single,
1282		.dev            = dev,
1283		.dev_addr       = dma_addr,
1284		.size           = size,
1285		.direction      = direction,
1286	};
1287
1288	if (unlikely(dma_debug_disabled()))
1289		return;
1290	check_unmap(&ref);
1291}
 
1292
1293void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
1294		      int nents, int mapped_ents, int direction,
1295		      unsigned long attrs)
1296{
1297	struct dma_debug_entry *entry;
1298	struct scatterlist *s;
1299	int i;
1300
1301	if (unlikely(dma_debug_disabled()))
1302		return;
1303
1304	for_each_sg(sg, s, nents, i) {
1305		check_for_stack(dev, sg_page(s), s->offset);
1306		if (!PageHighMem(sg_page(s)))
1307			check_for_illegal_area(dev, sg_virt(s), s->length);
1308	}
1309
1310	for_each_sg(sg, s, mapped_ents, i) {
1311		entry = dma_entry_alloc();
1312		if (!entry)
1313			return;
1314
1315		entry->type           = dma_debug_sg;
1316		entry->dev            = dev;
1317		entry->paddr	      = sg_phys(s);
 
1318		entry->size           = sg_dma_len(s);
1319		entry->dev_addr       = sg_dma_address(s);
1320		entry->direction      = direction;
1321		entry->sg_call_ents   = nents;
1322		entry->sg_mapped_ents = mapped_ents;
1323
 
 
 
 
 
 
1324		check_sg_segment(dev, s);
1325
1326		add_dma_entry(entry, attrs);
1327	}
1328}
 
1329
1330static int get_nr_mapped_entries(struct device *dev,
1331				 struct dma_debug_entry *ref)
1332{
1333	struct dma_debug_entry *entry;
1334	struct hash_bucket *bucket;
1335	unsigned long flags;
1336	int mapped_ents;
1337
1338	bucket       = get_hash_bucket(ref, &flags);
1339	entry        = bucket_find_exact(bucket, ref);
1340	mapped_ents  = 0;
1341
1342	if (entry)
1343		mapped_ents = entry->sg_mapped_ents;
1344	put_hash_bucket(bucket, flags);
1345
1346	return mapped_ents;
1347}
1348
1349void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
1350			int nelems, int dir)
1351{
1352	struct scatterlist *s;
1353	int mapped_ents = 0, i;
1354
1355	if (unlikely(dma_debug_disabled()))
1356		return;
1357
1358	for_each_sg(sglist, s, nelems, i) {
1359
1360		struct dma_debug_entry ref = {
1361			.type           = dma_debug_sg,
1362			.dev            = dev,
1363			.paddr		= sg_phys(s),
 
1364			.dev_addr       = sg_dma_address(s),
1365			.size           = sg_dma_len(s),
1366			.direction      = dir,
1367			.sg_call_ents   = nelems,
1368		};
1369
1370		if (mapped_ents && i >= mapped_ents)
1371			break;
1372
1373		if (!i)
1374			mapped_ents = get_nr_mapped_entries(dev, &ref);
1375
1376		check_unmap(&ref);
1377	}
1378}
1379
1380static phys_addr_t virt_to_paddr(void *virt)
1381{
1382	struct page *page;
1383
1384	if (is_vmalloc_addr(virt))
1385		page = vmalloc_to_page(virt);
1386	else
1387		page = virt_to_page(virt);
1388
1389	return page_to_phys(page) + offset_in_page(virt);
1390}
1391
1392void debug_dma_alloc_coherent(struct device *dev, size_t size,
1393			      dma_addr_t dma_addr, void *virt,
1394			      unsigned long attrs)
1395{
1396	struct dma_debug_entry *entry;
1397
1398	if (unlikely(dma_debug_disabled()))
1399		return;
1400
1401	if (unlikely(virt == NULL))
1402		return;
1403
1404	/* handle vmalloc and linear addresses */
1405	if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1406		return;
1407
1408	entry = dma_entry_alloc();
1409	if (!entry)
1410		return;
1411
1412	entry->type      = dma_debug_coherent;
1413	entry->dev       = dev;
1414	entry->paddr	 = virt_to_paddr(virt);
1415	entry->size      = size;
1416	entry->dev_addr  = dma_addr;
1417	entry->direction = DMA_BIDIRECTIONAL;
1418
1419	add_dma_entry(entry, attrs);
 
 
 
 
 
1420}
1421
1422void debug_dma_free_coherent(struct device *dev, size_t size,
1423			 void *virt, dma_addr_t dma_addr)
1424{
1425	struct dma_debug_entry ref = {
1426		.type           = dma_debug_coherent,
1427		.dev            = dev,
1428		.dev_addr       = dma_addr,
 
1429		.size           = size,
1430		.direction      = DMA_BIDIRECTIONAL,
1431	};
1432
1433	/* handle vmalloc and linear addresses */
1434	if (!is_vmalloc_addr(virt) && !virt_addr_valid(virt))
1435		return;
1436
1437	ref.paddr = virt_to_paddr(virt);
 
 
 
1438
1439	if (unlikely(dma_debug_disabled()))
1440		return;
1441
1442	check_unmap(&ref);
1443}
1444
1445void debug_dma_map_resource(struct device *dev, phys_addr_t addr, size_t size,
1446			    int direction, dma_addr_t dma_addr,
1447			    unsigned long attrs)
1448{
1449	struct dma_debug_entry *entry;
1450
1451	if (unlikely(dma_debug_disabled()))
1452		return;
1453
1454	entry = dma_entry_alloc();
1455	if (!entry)
1456		return;
1457
1458	entry->type		= dma_debug_resource;
1459	entry->dev		= dev;
1460	entry->paddr		= addr;
 
1461	entry->size		= size;
1462	entry->dev_addr		= dma_addr;
1463	entry->direction	= direction;
1464	entry->map_err_type	= MAP_ERR_NOT_CHECKED;
1465
1466	add_dma_entry(entry, attrs);
1467}
 
1468
1469void debug_dma_unmap_resource(struct device *dev, dma_addr_t dma_addr,
1470			      size_t size, int direction)
1471{
1472	struct dma_debug_entry ref = {
1473		.type           = dma_debug_resource,
1474		.dev            = dev,
1475		.dev_addr       = dma_addr,
1476		.size           = size,
1477		.direction      = direction,
1478	};
1479
1480	if (unlikely(dma_debug_disabled()))
1481		return;
1482
1483	check_unmap(&ref);
1484}
 
1485
1486void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
1487				   size_t size, int direction)
1488{
1489	struct dma_debug_entry ref;
1490
1491	if (unlikely(dma_debug_disabled()))
1492		return;
1493
1494	ref.type         = dma_debug_single;
1495	ref.dev          = dev;
1496	ref.dev_addr     = dma_handle;
1497	ref.size         = size;
1498	ref.direction    = direction;
1499	ref.sg_call_ents = 0;
1500
1501	check_sync(dev, &ref, true);
1502}
 
1503
1504void debug_dma_sync_single_for_device(struct device *dev,
1505				      dma_addr_t dma_handle, size_t size,
1506				      int direction)
1507{
1508	struct dma_debug_entry ref;
1509
1510	if (unlikely(dma_debug_disabled()))
1511		return;
1512
1513	ref.type         = dma_debug_single;
1514	ref.dev          = dev;
1515	ref.dev_addr     = dma_handle;
1516	ref.size         = size;
1517	ref.direction    = direction;
1518	ref.sg_call_ents = 0;
1519
1520	check_sync(dev, &ref, false);
1521}
 
1522
1523void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
1524			       int nelems, int direction)
1525{
1526	struct scatterlist *s;
1527	int mapped_ents = 0, i;
1528
1529	if (unlikely(dma_debug_disabled()))
1530		return;
1531
1532	for_each_sg(sg, s, nelems, i) {
1533
1534		struct dma_debug_entry ref = {
1535			.type           = dma_debug_sg,
1536			.dev            = dev,
1537			.paddr		= sg_phys(s),
 
1538			.dev_addr       = sg_dma_address(s),
1539			.size           = sg_dma_len(s),
1540			.direction      = direction,
1541			.sg_call_ents   = nelems,
1542		};
1543
1544		if (!i)
1545			mapped_ents = get_nr_mapped_entries(dev, &ref);
1546
1547		if (i >= mapped_ents)
1548			break;
1549
1550		check_sync(dev, &ref, true);
1551	}
1552}
 
1553
1554void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
1555				  int nelems, int direction)
1556{
1557	struct scatterlist *s;
1558	int mapped_ents = 0, i;
1559
1560	if (unlikely(dma_debug_disabled()))
1561		return;
1562
1563	for_each_sg(sg, s, nelems, i) {
1564
1565		struct dma_debug_entry ref = {
1566			.type           = dma_debug_sg,
1567			.dev            = dev,
1568			.paddr		= sg_phys(sg),
 
1569			.dev_addr       = sg_dma_address(s),
1570			.size           = sg_dma_len(s),
1571			.direction      = direction,
1572			.sg_call_ents   = nelems,
1573		};
1574		if (!i)
1575			mapped_ents = get_nr_mapped_entries(dev, &ref);
1576
1577		if (i >= mapped_ents)
1578			break;
1579
1580		check_sync(dev, &ref, false);
1581	}
1582}
 
1583
1584static int __init dma_debug_driver_setup(char *str)
1585{
1586	int i;
1587
1588	for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
1589		current_driver_name[i] = *str;
1590		if (*str == 0)
1591			break;
1592	}
1593
1594	if (current_driver_name[0])
1595		pr_info("enable driver filter for driver [%s]\n",
1596			current_driver_name);
1597
1598
1599	return 1;
1600}
1601__setup("dma_debug_driver=", dma_debug_driver_setup);