xref: /openbmc/linux/kernel/kcsan/core.c (revision 1edd0337)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * KCSAN core runtime.
4  *
5  * Copyright (C) 2019, Google LLC.
6  */
7 
8 #define pr_fmt(fmt) "kcsan: " fmt
9 
10 #include <linux/atomic.h>
11 #include <linux/bug.h>
12 #include <linux/delay.h>
13 #include <linux/export.h>
14 #include <linux/init.h>
15 #include <linux/kernel.h>
16 #include <linux/list.h>
17 #include <linux/minmax.h>
18 #include <linux/moduleparam.h>
19 #include <linux/percpu.h>
20 #include <linux/preempt.h>
21 #include <linux/sched.h>
22 #include <linux/string.h>
23 #include <linux/uaccess.h>
24 
25 #include "encoding.h"
26 #include "kcsan.h"
27 #include "permissive.h"
28 
29 static bool kcsan_early_enable = IS_ENABLED(CONFIG_KCSAN_EARLY_ENABLE);
30 unsigned int kcsan_udelay_task = CONFIG_KCSAN_UDELAY_TASK;
31 unsigned int kcsan_udelay_interrupt = CONFIG_KCSAN_UDELAY_INTERRUPT;
32 static long kcsan_skip_watch = CONFIG_KCSAN_SKIP_WATCH;
33 static bool kcsan_interrupt_watcher = IS_ENABLED(CONFIG_KCSAN_INTERRUPT_WATCHER);
34 
35 #ifdef MODULE_PARAM_PREFIX
36 #undef MODULE_PARAM_PREFIX
37 #endif
38 #define MODULE_PARAM_PREFIX "kcsan."
39 module_param_named(early_enable, kcsan_early_enable, bool, 0);
40 module_param_named(udelay_task, kcsan_udelay_task, uint, 0644);
41 module_param_named(udelay_interrupt, kcsan_udelay_interrupt, uint, 0644);
42 module_param_named(skip_watch, kcsan_skip_watch, long, 0644);
43 module_param_named(interrupt_watcher, kcsan_interrupt_watcher, bool, 0444);
44 
45 #ifdef CONFIG_KCSAN_WEAK_MEMORY
46 static bool kcsan_weak_memory = true;
47 module_param_named(weak_memory, kcsan_weak_memory, bool, 0644);
48 #else
49 #define kcsan_weak_memory false
50 #endif
51 
52 bool kcsan_enabled;
53 
54 /* Per-CPU kcsan_ctx for interrupts */
55 static DEFINE_PER_CPU(struct kcsan_ctx, kcsan_cpu_ctx) = {
56 	.scoped_accesses	= {LIST_POISON1, NULL},
57 };
58 
59 /*
60  * Helper macros to index into adjacent slots, starting from address slot
61  * itself, followed by the right and left slots.
62  *
63  * The purpose is 2-fold:
64  *
65  *	1. if during insertion the address slot is already occupied, check if
66  *	   any adjacent slots are free;
67  *	2. accesses that straddle a slot boundary due to size that exceeds a
68  *	   slot's range may check adjacent slots if any watchpoint matches.
69  *
70  * Note that accesses with very large size may still miss a watchpoint; however,
71  * given this should be rare, this is a reasonable trade-off to make, since this
72  * will avoid:
73  *
74  *	1. excessive contention between watchpoint checks and setup;
75  *	2. larger number of simultaneous watchpoints without sacrificing
76  *	   performance.
77  *
78  * Example: SLOT_IDX values for KCSAN_CHECK_ADJACENT=1, where i is [0, 1, 2]:
79  *
80  *   slot=0:  [ 1,  2,  0]
81  *   slot=9:  [10, 11,  9]
82  *   slot=63: [64, 65, 63]
83  */
84 #define SLOT_IDX(slot, i) (slot + ((i + KCSAN_CHECK_ADJACENT) % NUM_SLOTS))
85 
86 /*
87  * SLOT_IDX_FAST is used in the fast-path. Not first checking the address's primary
88  * slot (middle) is fine if we assume that races occur rarely. The set of
89  * indices {SLOT_IDX(slot, i) | i in [0, NUM_SLOTS)} is equivalent to
90  * {SLOT_IDX_FAST(slot, i) | i in [0, NUM_SLOTS)}.
91  */
92 #define SLOT_IDX_FAST(slot, i) (slot + i)
93 
94 /*
95  * Watchpoints, with each entry encoded as defined in encoding.h: in order to be
96  * able to safely update and access a watchpoint without introducing locking
97  * overhead, we encode each watchpoint as a single atomic long. The initial
98  * zero-initialized state matches INVALID_WATCHPOINT.
99  *
100  * Add NUM_SLOTS-1 entries to account for overflow; this helps avoid having to
101  * use more complicated SLOT_IDX_FAST calculation with modulo in the fast-path.
102  */
103 static atomic_long_t watchpoints[CONFIG_KCSAN_NUM_WATCHPOINTS + NUM_SLOTS-1];
104 
105 /*
106  * Instructions to skip watching counter, used in should_watch(). We use a
107  * per-CPU counter to avoid excessive contention.
108  */
109 static DEFINE_PER_CPU(long, kcsan_skip);
110 
111 /* For kcsan_prandom_u32_max(). */
112 static DEFINE_PER_CPU(u32, kcsan_rand_state);
113 
114 static __always_inline atomic_long_t *find_watchpoint(unsigned long addr,
115 						      size_t size,
116 						      bool expect_write,
117 						      long *encoded_watchpoint)
118 {
119 	const int slot = watchpoint_slot(addr);
120 	const unsigned long addr_masked = addr & WATCHPOINT_ADDR_MASK;
121 	atomic_long_t *watchpoint;
122 	unsigned long wp_addr_masked;
123 	size_t wp_size;
124 	bool is_write;
125 	int i;
126 
127 	BUILD_BUG_ON(CONFIG_KCSAN_NUM_WATCHPOINTS < NUM_SLOTS);
128 
129 	for (i = 0; i < NUM_SLOTS; ++i) {
130 		watchpoint = &watchpoints[SLOT_IDX_FAST(slot, i)];
131 		*encoded_watchpoint = atomic_long_read(watchpoint);
132 		if (!decode_watchpoint(*encoded_watchpoint, &wp_addr_masked,
133 				       &wp_size, &is_write))
134 			continue;
135 
136 		if (expect_write && !is_write)
137 			continue;
138 
139 		/* Check if the watchpoint matches the access. */
140 		if (matching_access(wp_addr_masked, wp_size, addr_masked, size))
141 			return watchpoint;
142 	}
143 
144 	return NULL;
145 }
146 
147 static inline atomic_long_t *
148 insert_watchpoint(unsigned long addr, size_t size, bool is_write)
149 {
150 	const int slot = watchpoint_slot(addr);
151 	const long encoded_watchpoint = encode_watchpoint(addr, size, is_write);
152 	atomic_long_t *watchpoint;
153 	int i;
154 
155 	/* Check slot index logic, ensuring we stay within array bounds. */
156 	BUILD_BUG_ON(SLOT_IDX(0, 0) != KCSAN_CHECK_ADJACENT);
157 	BUILD_BUG_ON(SLOT_IDX(0, KCSAN_CHECK_ADJACENT+1) != 0);
158 	BUILD_BUG_ON(SLOT_IDX(CONFIG_KCSAN_NUM_WATCHPOINTS-1, KCSAN_CHECK_ADJACENT) != ARRAY_SIZE(watchpoints)-1);
159 	BUILD_BUG_ON(SLOT_IDX(CONFIG_KCSAN_NUM_WATCHPOINTS-1, KCSAN_CHECK_ADJACENT+1) != ARRAY_SIZE(watchpoints) - NUM_SLOTS);
160 
161 	for (i = 0; i < NUM_SLOTS; ++i) {
162 		long expect_val = INVALID_WATCHPOINT;
163 
164 		/* Try to acquire this slot. */
165 		watchpoint = &watchpoints[SLOT_IDX(slot, i)];
166 		if (atomic_long_try_cmpxchg_relaxed(watchpoint, &expect_val, encoded_watchpoint))
167 			return watchpoint;
168 	}
169 
170 	return NULL;
171 }
172 
173 /*
174  * Return true if watchpoint was successfully consumed, false otherwise.
175  *
176  * This may return false if:
177  *
178  *	1. another thread already consumed the watchpoint;
179  *	2. the thread that set up the watchpoint already removed it;
180  *	3. the watchpoint was removed and then re-used.
181  */
182 static __always_inline bool
183 try_consume_watchpoint(atomic_long_t *watchpoint, long encoded_watchpoint)
184 {
185 	return atomic_long_try_cmpxchg_relaxed(watchpoint, &encoded_watchpoint, CONSUMED_WATCHPOINT);
186 }
187 
188 /* Return true if watchpoint was not touched, false if already consumed. */
189 static inline bool consume_watchpoint(atomic_long_t *watchpoint)
190 {
191 	return atomic_long_xchg_relaxed(watchpoint, CONSUMED_WATCHPOINT) != CONSUMED_WATCHPOINT;
192 }
193 
194 /* Remove the watchpoint -- its slot may be reused after. */
195 static inline void remove_watchpoint(atomic_long_t *watchpoint)
196 {
197 	atomic_long_set(watchpoint, INVALID_WATCHPOINT);
198 }
199 
200 static __always_inline struct kcsan_ctx *get_ctx(void)
201 {
202 	/*
203 	 * In interrupts, use raw_cpu_ptr to avoid unnecessary checks, that would
204 	 * also result in calls that generate warnings in uaccess regions.
205 	 */
206 	return in_task() ? &current->kcsan_ctx : raw_cpu_ptr(&kcsan_cpu_ctx);
207 }
208 
209 static __always_inline void
210 check_access(const volatile void *ptr, size_t size, int type, unsigned long ip);
211 
212 /* Check scoped accesses; never inline because this is a slow-path! */
213 static noinline void kcsan_check_scoped_accesses(void)
214 {
215 	struct kcsan_ctx *ctx = get_ctx();
216 	struct kcsan_scoped_access *scoped_access;
217 
218 	if (ctx->disable_scoped)
219 		return;
220 
221 	ctx->disable_scoped++;
222 	list_for_each_entry(scoped_access, &ctx->scoped_accesses, list) {
223 		check_access(scoped_access->ptr, scoped_access->size,
224 			     scoped_access->type, scoped_access->ip);
225 	}
226 	ctx->disable_scoped--;
227 }
228 
229 /* Rules for generic atomic accesses. Called from fast-path. */
230 static __always_inline bool
231 is_atomic(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size, int type)
232 {
233 	if (type & KCSAN_ACCESS_ATOMIC)
234 		return true;
235 
236 	/*
237 	 * Unless explicitly declared atomic, never consider an assertion access
238 	 * as atomic. This allows using them also in atomic regions, such as
239 	 * seqlocks, without implicitly changing their semantics.
240 	 */
241 	if (type & KCSAN_ACCESS_ASSERT)
242 		return false;
243 
244 	if (IS_ENABLED(CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC) &&
245 	    (type & KCSAN_ACCESS_WRITE) && size <= sizeof(long) &&
246 	    !(type & KCSAN_ACCESS_COMPOUND) && IS_ALIGNED((unsigned long)ptr, size))
247 		return true; /* Assume aligned writes up to word size are atomic. */
248 
249 	if (ctx->atomic_next > 0) {
250 		/*
251 		 * Because we do not have separate contexts for nested
252 		 * interrupts, in case atomic_next is set, we simply assume that
253 		 * the outer interrupt set atomic_next. In the worst case, we
254 		 * will conservatively consider operations as atomic. This is a
255 		 * reasonable trade-off to make, since this case should be
256 		 * extremely rare; however, even if extremely rare, it could
257 		 * lead to false positives otherwise.
258 		 */
259 		if ((hardirq_count() >> HARDIRQ_SHIFT) < 2)
260 			--ctx->atomic_next; /* in task, or outer interrupt */
261 		return true;
262 	}
263 
264 	return ctx->atomic_nest_count > 0 || ctx->in_flat_atomic;
265 }
266 
267 static __always_inline bool
268 should_watch(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size, int type)
269 {
270 	/*
271 	 * Never set up watchpoints when memory operations are atomic.
272 	 *
273 	 * Need to check this first, before kcsan_skip check below: (1) atomics
274 	 * should not count towards skipped instructions, and (2) to actually
275 	 * decrement kcsan_atomic_next for consecutive instruction stream.
276 	 */
277 	if (is_atomic(ctx, ptr, size, type))
278 		return false;
279 
280 	if (this_cpu_dec_return(kcsan_skip) >= 0)
281 		return false;
282 
283 	/*
284 	 * NOTE: If we get here, kcsan_skip must always be reset in slow path
285 	 * via reset_kcsan_skip() to avoid underflow.
286 	 */
287 
288 	/* this operation should be watched */
289 	return true;
290 }
291 
292 /*
293  * Returns a pseudo-random number in interval [0, ep_ro). Simple linear
294  * congruential generator, using constants from "Numerical Recipes".
295  */
296 static u32 kcsan_prandom_u32_max(u32 ep_ro)
297 {
298 	u32 state = this_cpu_read(kcsan_rand_state);
299 
300 	state = 1664525 * state + 1013904223;
301 	this_cpu_write(kcsan_rand_state, state);
302 
303 	return state % ep_ro;
304 }
305 
306 static inline void reset_kcsan_skip(void)
307 {
308 	long skip_count = kcsan_skip_watch -
309 			  (IS_ENABLED(CONFIG_KCSAN_SKIP_WATCH_RANDOMIZE) ?
310 				   kcsan_prandom_u32_max(kcsan_skip_watch) :
311 				   0);
312 	this_cpu_write(kcsan_skip, skip_count);
313 }
314 
315 static __always_inline bool kcsan_is_enabled(struct kcsan_ctx *ctx)
316 {
317 	return READ_ONCE(kcsan_enabled) && !ctx->disable_count;
318 }
319 
320 /* Introduce delay depending on context and configuration. */
321 static void delay_access(int type)
322 {
323 	unsigned int delay = in_task() ? kcsan_udelay_task : kcsan_udelay_interrupt;
324 	/* For certain access types, skew the random delay to be longer. */
325 	unsigned int skew_delay_order =
326 		(type & (KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_ASSERT)) ? 1 : 0;
327 
328 	delay -= IS_ENABLED(CONFIG_KCSAN_DELAY_RANDOMIZE) ?
329 			       kcsan_prandom_u32_max(delay >> skew_delay_order) :
330 			       0;
331 	udelay(delay);
332 }
333 
334 /*
335  * Reads the instrumented memory for value change detection; value change
336  * detection is currently done for accesses up to a size of 8 bytes.
337  */
338 static __always_inline u64 read_instrumented_memory(const volatile void *ptr, size_t size)
339 {
340 	switch (size) {
341 	case 1:  return READ_ONCE(*(const u8 *)ptr);
342 	case 2:  return READ_ONCE(*(const u16 *)ptr);
343 	case 4:  return READ_ONCE(*(const u32 *)ptr);
344 	case 8:  return READ_ONCE(*(const u64 *)ptr);
345 	default: return 0; /* Ignore; we do not diff the values. */
346 	}
347 }
348 
349 void kcsan_save_irqtrace(struct task_struct *task)
350 {
351 #ifdef CONFIG_TRACE_IRQFLAGS
352 	task->kcsan_save_irqtrace = task->irqtrace;
353 #endif
354 }
355 
356 void kcsan_restore_irqtrace(struct task_struct *task)
357 {
358 #ifdef CONFIG_TRACE_IRQFLAGS
359 	task->irqtrace = task->kcsan_save_irqtrace;
360 #endif
361 }
362 
363 static __always_inline int get_kcsan_stack_depth(void)
364 {
365 #ifdef CONFIG_KCSAN_WEAK_MEMORY
366 	return current->kcsan_stack_depth;
367 #else
368 	BUILD_BUG();
369 	return 0;
370 #endif
371 }
372 
373 static __always_inline void add_kcsan_stack_depth(int val)
374 {
375 #ifdef CONFIG_KCSAN_WEAK_MEMORY
376 	current->kcsan_stack_depth += val;
377 #else
378 	BUILD_BUG();
379 #endif
380 }
381 
382 static __always_inline struct kcsan_scoped_access *get_reorder_access(struct kcsan_ctx *ctx)
383 {
384 #ifdef CONFIG_KCSAN_WEAK_MEMORY
385 	return ctx->disable_scoped ? NULL : &ctx->reorder_access;
386 #else
387 	return NULL;
388 #endif
389 }
390 
391 static __always_inline bool
392 find_reorder_access(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size,
393 		    int type, unsigned long ip)
394 {
395 	struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);
396 
397 	if (!reorder_access)
398 		return false;
399 
400 	/*
401 	 * Note: If accesses are repeated while reorder_access is identical,
402 	 * never matches the new access, because !(type & KCSAN_ACCESS_SCOPED).
403 	 */
404 	return reorder_access->ptr == ptr && reorder_access->size == size &&
405 	       reorder_access->type == type && reorder_access->ip == ip;
406 }
407 
408 static inline void
409 set_reorder_access(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size,
410 		   int type, unsigned long ip)
411 {
412 	struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);
413 
414 	if (!reorder_access || !kcsan_weak_memory)
415 		return;
416 
417 	/*
418 	 * To avoid nested interrupts or scheduler (which share kcsan_ctx)
419 	 * reading an inconsistent reorder_access, ensure that the below has
420 	 * exclusive access to reorder_access by disallowing concurrent use.
421 	 */
422 	ctx->disable_scoped++;
423 	barrier();
424 	reorder_access->ptr		= ptr;
425 	reorder_access->size		= size;
426 	reorder_access->type		= type | KCSAN_ACCESS_SCOPED;
427 	reorder_access->ip		= ip;
428 	reorder_access->stack_depth	= get_kcsan_stack_depth();
429 	barrier();
430 	ctx->disable_scoped--;
431 }
432 
433 /*
434  * Pull everything together: check_access() below contains the performance
435  * critical operations; the fast-path (including check_access) functions should
436  * all be inlinable by the instrumentation functions.
437  *
438  * The slow-path (kcsan_found_watchpoint, kcsan_setup_watchpoint) are
439  * non-inlinable -- note that, we prefix these with "kcsan_" to ensure they can
440  * be filtered from the stacktrace, as well as give them unique names for the
441  * UACCESS whitelist of objtool. Each function uses user_access_save/restore(),
442  * since they do not access any user memory, but instrumentation is still
443  * emitted in UACCESS regions.
444  */
445 
446 static noinline void kcsan_found_watchpoint(const volatile void *ptr,
447 					    size_t size,
448 					    int type,
449 					    unsigned long ip,
450 					    atomic_long_t *watchpoint,
451 					    long encoded_watchpoint)
452 {
453 	const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
454 	struct kcsan_ctx *ctx = get_ctx();
455 	unsigned long flags;
456 	bool consumed;
457 
458 	/*
459 	 * We know a watchpoint exists. Let's try to keep the race-window
460 	 * between here and finally consuming the watchpoint below as small as
461 	 * possible -- avoid unneccessarily complex code until consumed.
462 	 */
463 
464 	if (!kcsan_is_enabled(ctx))
465 		return;
466 
467 	/*
468 	 * The access_mask check relies on value-change comparison. To avoid
469 	 * reporting a race where e.g. the writer set up the watchpoint, but the
470 	 * reader has access_mask!=0, we have to ignore the found watchpoint.
471 	 *
472 	 * reorder_access is never created from an access with access_mask set.
473 	 */
474 	if (ctx->access_mask && !find_reorder_access(ctx, ptr, size, type, ip))
475 		return;
476 
477 	/*
478 	 * If the other thread does not want to ignore the access, and there was
479 	 * a value change as a result of this thread's operation, we will still
480 	 * generate a report of unknown origin.
481 	 *
482 	 * Use CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN=n to filter.
483 	 */
484 	if (!is_assert && kcsan_ignore_address(ptr))
485 		return;
486 
487 	/*
488 	 * Consuming the watchpoint must be guarded by kcsan_is_enabled() to
489 	 * avoid erroneously triggering reports if the context is disabled.
490 	 */
491 	consumed = try_consume_watchpoint(watchpoint, encoded_watchpoint);
492 
493 	/* keep this after try_consume_watchpoint */
494 	flags = user_access_save();
495 
496 	if (consumed) {
497 		kcsan_save_irqtrace(current);
498 		kcsan_report_set_info(ptr, size, type, ip, watchpoint - watchpoints);
499 		kcsan_restore_irqtrace(current);
500 	} else {
501 		/*
502 		 * The other thread may not print any diagnostics, as it has
503 		 * already removed the watchpoint, or another thread consumed
504 		 * the watchpoint before this thread.
505 		 */
506 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_REPORT_RACES]);
507 	}
508 
509 	if (is_assert)
510 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
511 	else
512 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_DATA_RACES]);
513 
514 	user_access_restore(flags);
515 }
516 
517 static noinline void
518 kcsan_setup_watchpoint(const volatile void *ptr, size_t size, int type, unsigned long ip)
519 {
520 	const bool is_write = (type & KCSAN_ACCESS_WRITE) != 0;
521 	const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
522 	atomic_long_t *watchpoint;
523 	u64 old, new, diff;
524 	enum kcsan_value_change value_change = KCSAN_VALUE_CHANGE_MAYBE;
525 	bool interrupt_watcher = kcsan_interrupt_watcher;
526 	unsigned long ua_flags = user_access_save();
527 	struct kcsan_ctx *ctx = get_ctx();
528 	unsigned long access_mask = ctx->access_mask;
529 	unsigned long irq_flags = 0;
530 	bool is_reorder_access;
531 
532 	/*
533 	 * Always reset kcsan_skip counter in slow-path to avoid underflow; see
534 	 * should_watch().
535 	 */
536 	reset_kcsan_skip();
537 
538 	if (!kcsan_is_enabled(ctx))
539 		goto out;
540 
541 	/*
542 	 * Check to-ignore addresses after kcsan_is_enabled(), as we may access
543 	 * memory that is not yet initialized during early boot.
544 	 */
545 	if (!is_assert && kcsan_ignore_address(ptr))
546 		goto out;
547 
548 	if (!check_encodable((unsigned long)ptr, size)) {
549 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_UNENCODABLE_ACCESSES]);
550 		goto out;
551 	}
552 
553 	/*
554 	 * The local CPU cannot observe reordering of its own accesses, and
555 	 * therefore we need to take care of 2 cases to avoid false positives:
556 	 *
557 	 *	1. Races of the reordered access with interrupts. To avoid, if
558 	 *	   the current access is reorder_access, disable interrupts.
559 	 *	2. Avoid races of scoped accesses from nested interrupts (below).
560 	 */
561 	is_reorder_access = find_reorder_access(ctx, ptr, size, type, ip);
562 	if (is_reorder_access)
563 		interrupt_watcher = false;
564 	/*
565 	 * Avoid races of scoped accesses from nested interrupts (or scheduler).
566 	 * Assume setting up a watchpoint for a non-scoped (normal) access that
567 	 * also conflicts with a current scoped access. In a nested interrupt,
568 	 * which shares the context, it would check a conflicting scoped access.
569 	 * To avoid, disable scoped access checking.
570 	 */
571 	ctx->disable_scoped++;
572 
573 	/*
574 	 * Save and restore the IRQ state trace touched by KCSAN, since KCSAN's
575 	 * runtime is entered for every memory access, and potentially useful
576 	 * information is lost if dirtied by KCSAN.
577 	 */
578 	kcsan_save_irqtrace(current);
579 	if (!interrupt_watcher)
580 		local_irq_save(irq_flags);
581 
582 	watchpoint = insert_watchpoint((unsigned long)ptr, size, is_write);
583 	if (watchpoint == NULL) {
584 		/*
585 		 * Out of capacity: the size of 'watchpoints', and the frequency
586 		 * with which should_watch() returns true should be tweaked so
587 		 * that this case happens very rarely.
588 		 */
589 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_NO_CAPACITY]);
590 		goto out_unlock;
591 	}
592 
593 	atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_SETUP_WATCHPOINTS]);
594 	atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);
595 
596 	/*
597 	 * Read the current value, to later check and infer a race if the data
598 	 * was modified via a non-instrumented access, e.g. from a device.
599 	 */
600 	old = is_reorder_access ? 0 : read_instrumented_memory(ptr, size);
601 
602 	/*
603 	 * Delay this thread, to increase probability of observing a racy
604 	 * conflicting access.
605 	 */
606 	delay_access(type);
607 
608 	/*
609 	 * Re-read value, and check if it is as expected; if not, we infer a
610 	 * racy access.
611 	 */
612 	if (!is_reorder_access) {
613 		new = read_instrumented_memory(ptr, size);
614 	} else {
615 		/*
616 		 * Reordered accesses cannot be used for value change detection,
617 		 * because the memory location may no longer be accessible and
618 		 * could result in a fault.
619 		 */
620 		new = 0;
621 		access_mask = 0;
622 	}
623 
624 	diff = old ^ new;
625 	if (access_mask)
626 		diff &= access_mask;
627 
628 	/*
629 	 * Check if we observed a value change.
630 	 *
631 	 * Also check if the data race should be ignored (the rules depend on
632 	 * non-zero diff); if it is to be ignored, the below rules for
633 	 * KCSAN_VALUE_CHANGE_MAYBE apply.
634 	 */
635 	if (diff && !kcsan_ignore_data_race(size, type, old, new, diff))
636 		value_change = KCSAN_VALUE_CHANGE_TRUE;
637 
638 	/* Check if this access raced with another. */
639 	if (!consume_watchpoint(watchpoint)) {
640 		/*
641 		 * Depending on the access type, map a value_change of MAYBE to
642 		 * TRUE (always report) or FALSE (never report).
643 		 */
644 		if (value_change == KCSAN_VALUE_CHANGE_MAYBE) {
645 			if (access_mask != 0) {
646 				/*
647 				 * For access with access_mask, we require a
648 				 * value-change, as it is likely that races on
649 				 * ~access_mask bits are expected.
650 				 */
651 				value_change = KCSAN_VALUE_CHANGE_FALSE;
652 			} else if (size > 8 || is_assert) {
653 				/* Always assume a value-change. */
654 				value_change = KCSAN_VALUE_CHANGE_TRUE;
655 			}
656 		}
657 
658 		/*
659 		 * No need to increment 'data_races' counter, as the racing
660 		 * thread already did.
661 		 *
662 		 * Count 'assert_failures' for each failed ASSERT access,
663 		 * therefore both this thread and the racing thread may
664 		 * increment this counter.
665 		 */
666 		if (is_assert && value_change == KCSAN_VALUE_CHANGE_TRUE)
667 			atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
668 
669 		kcsan_report_known_origin(ptr, size, type, ip,
670 					  value_change, watchpoint - watchpoints,
671 					  old, new, access_mask);
672 	} else if (value_change == KCSAN_VALUE_CHANGE_TRUE) {
673 		/* Inferring a race, since the value should not have changed. */
674 
675 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_RACES_UNKNOWN_ORIGIN]);
676 		if (is_assert)
677 			atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
678 
679 		if (IS_ENABLED(CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN) || is_assert) {
680 			kcsan_report_unknown_origin(ptr, size, type, ip,
681 						    old, new, access_mask);
682 		}
683 	}
684 
685 	/*
686 	 * Remove watchpoint; must be after reporting, since the slot may be
687 	 * reused after this point.
688 	 */
689 	remove_watchpoint(watchpoint);
690 	atomic_long_dec(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);
691 
692 out_unlock:
693 	if (!interrupt_watcher)
694 		local_irq_restore(irq_flags);
695 	kcsan_restore_irqtrace(current);
696 	ctx->disable_scoped--;
697 
698 	/*
699 	 * Reordered accesses cannot be used for value change detection,
700 	 * therefore never consider for reordering if access_mask is set.
701 	 * ASSERT_EXCLUSIVE are not real accesses, ignore them as well.
702 	 */
703 	if (!access_mask && !is_assert)
704 		set_reorder_access(ctx, ptr, size, type, ip);
705 out:
706 	user_access_restore(ua_flags);
707 }
708 
709 static __always_inline void
710 check_access(const volatile void *ptr, size_t size, int type, unsigned long ip)
711 {
712 	atomic_long_t *watchpoint;
713 	long encoded_watchpoint;
714 
715 	/*
716 	 * Do nothing for 0 sized check; this comparison will be optimized out
717 	 * for constant sized instrumentation (__tsan_{read,write}N).
718 	 */
719 	if (unlikely(size == 0))
720 		return;
721 
722 again:
723 	/*
724 	 * Avoid user_access_save in fast-path: find_watchpoint is safe without
725 	 * user_access_save, as the address that ptr points to is only used to
726 	 * check if a watchpoint exists; ptr is never dereferenced.
727 	 */
728 	watchpoint = find_watchpoint((unsigned long)ptr, size,
729 				     !(type & KCSAN_ACCESS_WRITE),
730 				     &encoded_watchpoint);
731 	/*
732 	 * It is safe to check kcsan_is_enabled() after find_watchpoint in the
733 	 * slow-path, as long as no state changes that cause a race to be
734 	 * detected and reported have occurred until kcsan_is_enabled() is
735 	 * checked.
736 	 */
737 
738 	if (unlikely(watchpoint != NULL))
739 		kcsan_found_watchpoint(ptr, size, type, ip, watchpoint, encoded_watchpoint);
740 	else {
741 		struct kcsan_ctx *ctx = get_ctx(); /* Call only once in fast-path. */
742 
743 		if (unlikely(should_watch(ctx, ptr, size, type))) {
744 			kcsan_setup_watchpoint(ptr, size, type, ip);
745 			return;
746 		}
747 
748 		if (!(type & KCSAN_ACCESS_SCOPED)) {
749 			struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);
750 
751 			if (reorder_access) {
752 				/*
753 				 * reorder_access check: simulates reordering of
754 				 * the access after subsequent operations.
755 				 */
756 				ptr = reorder_access->ptr;
757 				type = reorder_access->type;
758 				ip = reorder_access->ip;
759 				/*
760 				 * Upon a nested interrupt, this context's
761 				 * reorder_access can be modified (shared ctx).
762 				 * We know that upon return, reorder_access is
763 				 * always invalidated by setting size to 0 via
764 				 * __tsan_func_exit(). Therefore we must read
765 				 * and check size after the other fields.
766 				 */
767 				barrier();
768 				size = READ_ONCE(reorder_access->size);
769 				if (size)
770 					goto again;
771 			}
772 		}
773 
774 		/*
775 		 * Always checked last, right before returning from runtime;
776 		 * if reorder_access is valid, checked after it was checked.
777 		 */
778 		if (unlikely(ctx->scoped_accesses.prev))
779 			kcsan_check_scoped_accesses();
780 	}
781 }
782 
783 /* === Public interface ===================================================== */
784 
785 void __init kcsan_init(void)
786 {
787 	int cpu;
788 
789 	BUG_ON(!in_task());
790 
791 	for_each_possible_cpu(cpu)
792 		per_cpu(kcsan_rand_state, cpu) = (u32)get_cycles();
793 
794 	/*
795 	 * We are in the init task, and no other tasks should be running;
796 	 * WRITE_ONCE without memory barrier is sufficient.
797 	 */
798 	if (kcsan_early_enable) {
799 		pr_info("enabled early\n");
800 		WRITE_ONCE(kcsan_enabled, true);
801 	}
802 
803 	if (IS_ENABLED(CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY) ||
804 	    IS_ENABLED(CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC) ||
805 	    IS_ENABLED(CONFIG_KCSAN_PERMISSIVE) ||
806 	    IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {
807 		pr_warn("non-strict mode configured - use CONFIG_KCSAN_STRICT=y to see all data races\n");
808 	} else {
809 		pr_info("strict mode configured\n");
810 	}
811 }
812 
813 /* === Exported interface =================================================== */
814 
815 void kcsan_disable_current(void)
816 {
817 	++get_ctx()->disable_count;
818 }
819 EXPORT_SYMBOL(kcsan_disable_current);
820 
821 void kcsan_enable_current(void)
822 {
823 	if (get_ctx()->disable_count-- == 0) {
824 		/*
825 		 * Warn if kcsan_enable_current() calls are unbalanced with
826 		 * kcsan_disable_current() calls, which causes disable_count to
827 		 * become negative and should not happen.
828 		 */
829 		kcsan_disable_current(); /* restore to 0, KCSAN still enabled */
830 		kcsan_disable_current(); /* disable to generate warning */
831 		WARN(1, "Unbalanced %s()", __func__);
832 		kcsan_enable_current();
833 	}
834 }
835 EXPORT_SYMBOL(kcsan_enable_current);
836 
837 void kcsan_enable_current_nowarn(void)
838 {
839 	if (get_ctx()->disable_count-- == 0)
840 		kcsan_disable_current();
841 }
842 EXPORT_SYMBOL(kcsan_enable_current_nowarn);
843 
844 void kcsan_nestable_atomic_begin(void)
845 {
846 	/*
847 	 * Do *not* check and warn if we are in a flat atomic region: nestable
848 	 * and flat atomic regions are independent from each other.
849 	 * See include/linux/kcsan.h: struct kcsan_ctx comments for more
850 	 * comments.
851 	 */
852 
853 	++get_ctx()->atomic_nest_count;
854 }
855 EXPORT_SYMBOL(kcsan_nestable_atomic_begin);
856 
857 void kcsan_nestable_atomic_end(void)
858 {
859 	if (get_ctx()->atomic_nest_count-- == 0) {
860 		/*
861 		 * Warn if kcsan_nestable_atomic_end() calls are unbalanced with
862 		 * kcsan_nestable_atomic_begin() calls, which causes
863 		 * atomic_nest_count to become negative and should not happen.
864 		 */
865 		kcsan_nestable_atomic_begin(); /* restore to 0 */
866 		kcsan_disable_current(); /* disable to generate warning */
867 		WARN(1, "Unbalanced %s()", __func__);
868 		kcsan_enable_current();
869 	}
870 }
871 EXPORT_SYMBOL(kcsan_nestable_atomic_end);
872 
873 void kcsan_flat_atomic_begin(void)
874 {
875 	get_ctx()->in_flat_atomic = true;
876 }
877 EXPORT_SYMBOL(kcsan_flat_atomic_begin);
878 
879 void kcsan_flat_atomic_end(void)
880 {
881 	get_ctx()->in_flat_atomic = false;
882 }
883 EXPORT_SYMBOL(kcsan_flat_atomic_end);
884 
885 void kcsan_atomic_next(int n)
886 {
887 	get_ctx()->atomic_next = n;
888 }
889 EXPORT_SYMBOL(kcsan_atomic_next);
890 
891 void kcsan_set_access_mask(unsigned long mask)
892 {
893 	get_ctx()->access_mask = mask;
894 }
895 EXPORT_SYMBOL(kcsan_set_access_mask);
896 
897 struct kcsan_scoped_access *
898 kcsan_begin_scoped_access(const volatile void *ptr, size_t size, int type,
899 			  struct kcsan_scoped_access *sa)
900 {
901 	struct kcsan_ctx *ctx = get_ctx();
902 
903 	check_access(ptr, size, type, _RET_IP_);
904 
905 	ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */
906 
907 	INIT_LIST_HEAD(&sa->list);
908 	sa->ptr = ptr;
909 	sa->size = size;
910 	sa->type = type;
911 	sa->ip = _RET_IP_;
912 
913 	if (!ctx->scoped_accesses.prev) /* Lazy initialize list head. */
914 		INIT_LIST_HEAD(&ctx->scoped_accesses);
915 	list_add(&sa->list, &ctx->scoped_accesses);
916 
917 	ctx->disable_count--;
918 	return sa;
919 }
920 EXPORT_SYMBOL(kcsan_begin_scoped_access);
921 
922 void kcsan_end_scoped_access(struct kcsan_scoped_access *sa)
923 {
924 	struct kcsan_ctx *ctx = get_ctx();
925 
926 	if (WARN(!ctx->scoped_accesses.prev, "Unbalanced %s()?", __func__))
927 		return;
928 
929 	ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */
930 
931 	list_del(&sa->list);
932 	if (list_empty(&ctx->scoped_accesses))
933 		/*
934 		 * Ensure we do not enter kcsan_check_scoped_accesses()
935 		 * slow-path if unnecessary, and avoids requiring list_empty()
936 		 * in the fast-path (to avoid a READ_ONCE() and potential
937 		 * uaccess warning).
938 		 */
939 		ctx->scoped_accesses.prev = NULL;
940 
941 	ctx->disable_count--;
942 
943 	check_access(sa->ptr, sa->size, sa->type, sa->ip);
944 }
945 EXPORT_SYMBOL(kcsan_end_scoped_access);
946 
947 void __kcsan_check_access(const volatile void *ptr, size_t size, int type)
948 {
949 	check_access(ptr, size, type, _RET_IP_);
950 }
951 EXPORT_SYMBOL(__kcsan_check_access);
952 
953 #define DEFINE_MEMORY_BARRIER(name, order_before_cond)				\
954 	void __kcsan_##name(void)						\
955 	{									\
956 		struct kcsan_scoped_access *sa = get_reorder_access(get_ctx());	\
957 		if (!sa)							\
958 			return;							\
959 		if (order_before_cond)						\
960 			sa->size = 0;						\
961 	}									\
962 	EXPORT_SYMBOL(__kcsan_##name)
963 
964 DEFINE_MEMORY_BARRIER(mb, true);
965 DEFINE_MEMORY_BARRIER(wmb, sa->type & (KCSAN_ACCESS_WRITE | KCSAN_ACCESS_COMPOUND));
966 DEFINE_MEMORY_BARRIER(rmb, !(sa->type & KCSAN_ACCESS_WRITE) || (sa->type & KCSAN_ACCESS_COMPOUND));
967 DEFINE_MEMORY_BARRIER(release, true);
968 
969 /*
970  * KCSAN uses the same instrumentation that is emitted by supported compilers
971  * for ThreadSanitizer (TSAN).
972  *
973  * When enabled, the compiler emits instrumentation calls (the functions
974  * prefixed with "__tsan" below) for all loads and stores that it generated;
975  * inline asm is not instrumented.
976  *
977  * Note that, not all supported compiler versions distinguish aligned/unaligned
978  * accesses, but e.g. recent versions of Clang do. We simply alias the unaligned
979  * version to the generic version, which can handle both.
980  */
981 
982 #define DEFINE_TSAN_READ_WRITE(size)                                           \
983 	void __tsan_read##size(void *ptr);                                     \
984 	void __tsan_read##size(void *ptr)                                      \
985 	{                                                                      \
986 		check_access(ptr, size, 0, _RET_IP_);                          \
987 	}                                                                      \
988 	EXPORT_SYMBOL(__tsan_read##size);                                      \
989 	void __tsan_unaligned_read##size(void *ptr)                            \
990 		__alias(__tsan_read##size);                                    \
991 	EXPORT_SYMBOL(__tsan_unaligned_read##size);                            \
992 	void __tsan_write##size(void *ptr);                                    \
993 	void __tsan_write##size(void *ptr)                                     \
994 	{                                                                      \
995 		check_access(ptr, size, KCSAN_ACCESS_WRITE, _RET_IP_);         \
996 	}                                                                      \
997 	EXPORT_SYMBOL(__tsan_write##size);                                     \
998 	void __tsan_unaligned_write##size(void *ptr)                           \
999 		__alias(__tsan_write##size);                                   \
1000 	EXPORT_SYMBOL(__tsan_unaligned_write##size);                           \
1001 	void __tsan_read_write##size(void *ptr);                               \
1002 	void __tsan_read_write##size(void *ptr)                                \
1003 	{                                                                      \
1004 		check_access(ptr, size,                                        \
1005 			     KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE,       \
1006 			     _RET_IP_);                                        \
1007 	}                                                                      \
1008 	EXPORT_SYMBOL(__tsan_read_write##size);                                \
1009 	void __tsan_unaligned_read_write##size(void *ptr)                      \
1010 		__alias(__tsan_read_write##size);                              \
1011 	EXPORT_SYMBOL(__tsan_unaligned_read_write##size)
1012 
1013 DEFINE_TSAN_READ_WRITE(1);
1014 DEFINE_TSAN_READ_WRITE(2);
1015 DEFINE_TSAN_READ_WRITE(4);
1016 DEFINE_TSAN_READ_WRITE(8);
1017 DEFINE_TSAN_READ_WRITE(16);
1018 
1019 void __tsan_read_range(void *ptr, size_t size);
1020 void __tsan_read_range(void *ptr, size_t size)
1021 {
1022 	check_access(ptr, size, 0, _RET_IP_);
1023 }
1024 EXPORT_SYMBOL(__tsan_read_range);
1025 
1026 void __tsan_write_range(void *ptr, size_t size);
1027 void __tsan_write_range(void *ptr, size_t size)
1028 {
1029 	check_access(ptr, size, KCSAN_ACCESS_WRITE, _RET_IP_);
1030 }
1031 EXPORT_SYMBOL(__tsan_write_range);
1032 
1033 /*
1034  * Use of explicit volatile is generally disallowed [1], however, volatile is
1035  * still used in various concurrent context, whether in low-level
1036  * synchronization primitives or for legacy reasons.
1037  * [1] https://lwn.net/Articles/233479/
1038  *
1039  * We only consider volatile accesses atomic if they are aligned and would pass
1040  * the size-check of compiletime_assert_rwonce_type().
1041  */
1042 #define DEFINE_TSAN_VOLATILE_READ_WRITE(size)                                  \
1043 	void __tsan_volatile_read##size(void *ptr);                            \
1044 	void __tsan_volatile_read##size(void *ptr)                             \
1045 	{                                                                      \
1046 		const bool is_atomic = size <= sizeof(long long) &&            \
1047 				       IS_ALIGNED((unsigned long)ptr, size);   \
1048 		if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
1049 			return;                                                \
1050 		check_access(ptr, size, is_atomic ? KCSAN_ACCESS_ATOMIC : 0,   \
1051 			     _RET_IP_);                                        \
1052 	}                                                                      \
1053 	EXPORT_SYMBOL(__tsan_volatile_read##size);                             \
1054 	void __tsan_unaligned_volatile_read##size(void *ptr)                   \
1055 		__alias(__tsan_volatile_read##size);                           \
1056 	EXPORT_SYMBOL(__tsan_unaligned_volatile_read##size);                   \
1057 	void __tsan_volatile_write##size(void *ptr);                           \
1058 	void __tsan_volatile_write##size(void *ptr)                            \
1059 	{                                                                      \
1060 		const bool is_atomic = size <= sizeof(long long) &&            \
1061 				       IS_ALIGNED((unsigned long)ptr, size);   \
1062 		if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
1063 			return;                                                \
1064 		check_access(ptr, size,                                        \
1065 			     KCSAN_ACCESS_WRITE |                              \
1066 				     (is_atomic ? KCSAN_ACCESS_ATOMIC : 0),    \
1067 			     _RET_IP_);                                        \
1068 	}                                                                      \
1069 	EXPORT_SYMBOL(__tsan_volatile_write##size);                            \
1070 	void __tsan_unaligned_volatile_write##size(void *ptr)                  \
1071 		__alias(__tsan_volatile_write##size);                          \
1072 	EXPORT_SYMBOL(__tsan_unaligned_volatile_write##size)
1073 
1074 DEFINE_TSAN_VOLATILE_READ_WRITE(1);
1075 DEFINE_TSAN_VOLATILE_READ_WRITE(2);
1076 DEFINE_TSAN_VOLATILE_READ_WRITE(4);
1077 DEFINE_TSAN_VOLATILE_READ_WRITE(8);
1078 DEFINE_TSAN_VOLATILE_READ_WRITE(16);
1079 
1080 /*
1081  * Function entry and exit are used to determine the validty of reorder_access.
1082  * Reordering of the access ends at the end of the function scope where the
1083  * access happened. This is done for two reasons:
1084  *
1085  *	1. Artificially limits the scope where missing barriers are detected.
1086  *	   This minimizes false positives due to uninstrumented functions that
1087  *	   contain the required barriers but were missed.
1088  *
1089  *	2. Simplifies generating the stack trace of the access.
1090  */
1091 void __tsan_func_entry(void *call_pc);
1092 noinline void __tsan_func_entry(void *call_pc)
1093 {
1094 	if (!IS_ENABLED(CONFIG_KCSAN_WEAK_MEMORY))
1095 		return;
1096 
1097 	add_kcsan_stack_depth(1);
1098 }
1099 EXPORT_SYMBOL(__tsan_func_entry);
1100 
1101 void __tsan_func_exit(void);
1102 noinline void __tsan_func_exit(void)
1103 {
1104 	struct kcsan_scoped_access *reorder_access;
1105 
1106 	if (!IS_ENABLED(CONFIG_KCSAN_WEAK_MEMORY))
1107 		return;
1108 
1109 	reorder_access = get_reorder_access(get_ctx());
1110 	if (!reorder_access)
1111 		goto out;
1112 
1113 	if (get_kcsan_stack_depth() <= reorder_access->stack_depth) {
1114 		/*
1115 		 * Access check to catch cases where write without a barrier
1116 		 * (supposed release) was last access in function: because
1117 		 * instrumentation is inserted before the real access, a data
1118 		 * race due to the write giving up a c-s would only be caught if
1119 		 * we do the conflicting access after.
1120 		 */
1121 		check_access(reorder_access->ptr, reorder_access->size,
1122 			     reorder_access->type, reorder_access->ip);
1123 		reorder_access->size = 0;
1124 		reorder_access->stack_depth = INT_MIN;
1125 	}
1126 out:
1127 	add_kcsan_stack_depth(-1);
1128 }
1129 EXPORT_SYMBOL(__tsan_func_exit);
1130 
1131 void __tsan_init(void);
1132 void __tsan_init(void)
1133 {
1134 }
1135 EXPORT_SYMBOL(__tsan_init);
1136 
1137 /*
1138  * Instrumentation for atomic builtins (__atomic_*, __sync_*).
1139  *
1140  * Normal kernel code _should not_ be using them directly, but some
1141  * architectures may implement some or all atomics using the compilers'
1142  * builtins.
1143  *
1144  * Note: If an architecture decides to fully implement atomics using the
1145  * builtins, because they are implicitly instrumented by KCSAN (and KASAN,
1146  * etc.), implementing the ARCH_ATOMIC interface (to get instrumentation via
1147  * atomic-instrumented) is no longer necessary.
1148  *
1149  * TSAN instrumentation replaces atomic accesses with calls to any of the below
1150  * functions, whose job is to also execute the operation itself.
1151  */
1152 
1153 static __always_inline void kcsan_atomic_builtin_memorder(int memorder)
1154 {
1155 	if (memorder == __ATOMIC_RELEASE ||
1156 	    memorder == __ATOMIC_SEQ_CST ||
1157 	    memorder == __ATOMIC_ACQ_REL)
1158 		__kcsan_release();
1159 }
1160 
1161 #define DEFINE_TSAN_ATOMIC_LOAD_STORE(bits)                                                        \
1162 	u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder);                      \
1163 	u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder)                       \
1164 	{                                                                                          \
1165 		kcsan_atomic_builtin_memorder(memorder);                                           \
1166 		if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1167 			check_access(ptr, bits / BITS_PER_BYTE, KCSAN_ACCESS_ATOMIC, _RET_IP_);    \
1168 		}                                                                                  \
1169 		return __atomic_load_n(ptr, memorder);                                             \
1170 	}                                                                                          \
1171 	EXPORT_SYMBOL(__tsan_atomic##bits##_load);                                                 \
1172 	void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder);                   \
1173 	void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder)                    \
1174 	{                                                                                          \
1175 		kcsan_atomic_builtin_memorder(memorder);                                           \
1176 		if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1177 			check_access(ptr, bits / BITS_PER_BYTE,                                    \
1178 				     KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ATOMIC, _RET_IP_);          \
1179 		}                                                                                  \
1180 		__atomic_store_n(ptr, v, memorder);                                                \
1181 	}                                                                                          \
1182 	EXPORT_SYMBOL(__tsan_atomic##bits##_store)
1183 
1184 #define DEFINE_TSAN_ATOMIC_RMW(op, bits, suffix)                                                   \
1185 	u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder);                 \
1186 	u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder)                  \
1187 	{                                                                                          \
1188 		kcsan_atomic_builtin_memorder(memorder);                                           \
1189 		if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1190 			check_access(ptr, bits / BITS_PER_BYTE,                                    \
1191 				     KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1192 					     KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
1193 		}                                                                                  \
1194 		return __atomic_##op##suffix(ptr, v, memorder);                                    \
1195 	}                                                                                          \
1196 	EXPORT_SYMBOL(__tsan_atomic##bits##_##op)
1197 
1198 /*
1199  * Note: CAS operations are always classified as write, even in case they
1200  * fail. We cannot perform check_access() after a write, as it might lead to
1201  * false positives, in cases such as:
1202  *
1203  *	T0: __atomic_compare_exchange_n(&p->flag, &old, 1, ...)
1204  *
1205  *	T1: if (__atomic_load_n(&p->flag, ...)) {
1206  *		modify *p;
1207  *		p->flag = 0;
1208  *	    }
1209  *
1210  * The only downside is that, if there are 3 threads, with one CAS that
1211  * succeeds, another CAS that fails, and an unmarked racing operation, we may
1212  * point at the wrong CAS as the source of the race. However, if we assume that
1213  * all CAS can succeed in some other execution, the data race is still valid.
1214  */
1215 #define DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strength, weak)                                           \
1216 	int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
1217 							      u##bits val, int mo, int fail_mo);   \
1218 	int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
1219 							      u##bits val, int mo, int fail_mo)    \
1220 	{                                                                                          \
1221 		kcsan_atomic_builtin_memorder(mo);                                                 \
1222 		if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1223 			check_access(ptr, bits / BITS_PER_BYTE,                                    \
1224 				     KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1225 					     KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
1226 		}                                                                                  \
1227 		return __atomic_compare_exchange_n(ptr, exp, val, weak, mo, fail_mo);              \
1228 	}                                                                                          \
1229 	EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_##strength)
1230 
1231 #define DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)                                                       \
1232 	u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
1233 							   int mo, int fail_mo);                   \
1234 	u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
1235 							   int mo, int fail_mo)                    \
1236 	{                                                                                          \
1237 		kcsan_atomic_builtin_memorder(mo);                                                 \
1238 		if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1239 			check_access(ptr, bits / BITS_PER_BYTE,                                    \
1240 				     KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1241 					     KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
1242 		}                                                                                  \
1243 		__atomic_compare_exchange_n(ptr, &exp, val, 0, mo, fail_mo);                       \
1244 		return exp;                                                                        \
1245 	}                                                                                          \
1246 	EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_val)
1247 
1248 #define DEFINE_TSAN_ATOMIC_OPS(bits)                                                               \
1249 	DEFINE_TSAN_ATOMIC_LOAD_STORE(bits);                                                       \
1250 	DEFINE_TSAN_ATOMIC_RMW(exchange, bits, _n);                                                \
1251 	DEFINE_TSAN_ATOMIC_RMW(fetch_add, bits, );                                                 \
1252 	DEFINE_TSAN_ATOMIC_RMW(fetch_sub, bits, );                                                 \
1253 	DEFINE_TSAN_ATOMIC_RMW(fetch_and, bits, );                                                 \
1254 	DEFINE_TSAN_ATOMIC_RMW(fetch_or, bits, );                                                  \
1255 	DEFINE_TSAN_ATOMIC_RMW(fetch_xor, bits, );                                                 \
1256 	DEFINE_TSAN_ATOMIC_RMW(fetch_nand, bits, );                                                \
1257 	DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strong, 0);                                               \
1258 	DEFINE_TSAN_ATOMIC_CMPXCHG(bits, weak, 1);                                                 \
1259 	DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)
1260 
1261 DEFINE_TSAN_ATOMIC_OPS(8);
1262 DEFINE_TSAN_ATOMIC_OPS(16);
1263 DEFINE_TSAN_ATOMIC_OPS(32);
1264 DEFINE_TSAN_ATOMIC_OPS(64);
1265 
1266 void __tsan_atomic_thread_fence(int memorder);
1267 void __tsan_atomic_thread_fence(int memorder)
1268 {
1269 	kcsan_atomic_builtin_memorder(memorder);
1270 	__atomic_thread_fence(memorder);
1271 }
1272 EXPORT_SYMBOL(__tsan_atomic_thread_fence);
1273 
1274 /*
1275  * In instrumented files, we emit instrumentation for barriers by mapping the
1276  * kernel barriers to an __atomic_signal_fence(), which is interpreted specially
1277  * and otherwise has no relation to a real __atomic_signal_fence(). No known
1278  * kernel code uses __atomic_signal_fence().
1279  *
1280  * Since fsanitize=thread instrumentation handles __atomic_signal_fence(), which
1281  * are turned into calls to __tsan_atomic_signal_fence(), such instrumentation
1282  * can be disabled via the __no_kcsan function attribute (vs. an explicit call
1283  * which could not). When __no_kcsan is requested, __atomic_signal_fence()
1284  * generates no code.
1285  *
1286  * Note: The result of using __atomic_signal_fence() with KCSAN enabled is
1287  * potentially limiting the compiler's ability to reorder operations; however,
1288  * if barriers were instrumented with explicit calls (without LTO), the compiler
1289  * couldn't optimize much anyway. The result of a hypothetical architecture
1290  * using __atomic_signal_fence() in normal code would be KCSAN false negatives.
1291  */
1292 void __tsan_atomic_signal_fence(int memorder);
1293 noinline void __tsan_atomic_signal_fence(int memorder)
1294 {
1295 	switch (memorder) {
1296 	case __KCSAN_BARRIER_TO_SIGNAL_FENCE_mb:
1297 		__kcsan_mb();
1298 		break;
1299 	case __KCSAN_BARRIER_TO_SIGNAL_FENCE_wmb:
1300 		__kcsan_wmb();
1301 		break;
1302 	case __KCSAN_BARRIER_TO_SIGNAL_FENCE_rmb:
1303 		__kcsan_rmb();
1304 		break;
1305 	case __KCSAN_BARRIER_TO_SIGNAL_FENCE_release:
1306 		__kcsan_release();
1307 		break;
1308 	default:
1309 		break;
1310 	}
1311 }
1312 EXPORT_SYMBOL(__tsan_atomic_signal_fence);
1313 
1314 #ifdef __HAVE_ARCH_MEMSET
1315 void *__tsan_memset(void *s, int c, size_t count);
1316 noinline void *__tsan_memset(void *s, int c, size_t count)
1317 {
1318 	/*
1319 	 * Instead of not setting up watchpoints where accessed size is greater
1320 	 * than MAX_ENCODABLE_SIZE, truncate checked size to MAX_ENCODABLE_SIZE.
1321 	 */
1322 	size_t check_len = min_t(size_t, count, MAX_ENCODABLE_SIZE);
1323 
1324 	check_access(s, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
1325 	return memset(s, c, count);
1326 }
1327 #else
1328 void *__tsan_memset(void *s, int c, size_t count) __alias(memset);
1329 #endif
1330 EXPORT_SYMBOL(__tsan_memset);
1331 
1332 #ifdef __HAVE_ARCH_MEMMOVE
1333 void *__tsan_memmove(void *dst, const void *src, size_t len);
1334 noinline void *__tsan_memmove(void *dst, const void *src, size_t len)
1335 {
1336 	size_t check_len = min_t(size_t, len, MAX_ENCODABLE_SIZE);
1337 
1338 	check_access(dst, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
1339 	check_access(src, check_len, 0, _RET_IP_);
1340 	return memmove(dst, src, len);
1341 }
1342 #else
1343 void *__tsan_memmove(void *dst, const void *src, size_t len) __alias(memmove);
1344 #endif
1345 EXPORT_SYMBOL(__tsan_memmove);
1346 
1347 #ifdef __HAVE_ARCH_MEMCPY
1348 void *__tsan_memcpy(void *dst, const void *src, size_t len);
1349 noinline void *__tsan_memcpy(void *dst, const void *src, size_t len)
1350 {
1351 	size_t check_len = min_t(size_t, len, MAX_ENCODABLE_SIZE);
1352 
1353 	check_access(dst, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
1354 	check_access(src, check_len, 0, _RET_IP_);
1355 	return memcpy(dst, src, len);
1356 }
1357 #else
1358 void *__tsan_memcpy(void *dst, const void *src, size_t len) __alias(memcpy);
1359 #endif
1360 EXPORT_SYMBOL(__tsan_memcpy);
1361