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