xref: /linux/kernel/kcsan/core.c (revision 216b3f432a36549767ca750e64badd71340b8c0f)
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 
find_watchpoint(unsigned long addr,size_t size,bool expect_write,long * encoded_watchpoint)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 *
insert_watchpoint(unsigned long addr,size_t size,bool is_write)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
try_consume_watchpoint(atomic_long_t * watchpoint,long encoded_watchpoint)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. */
consume_watchpoint(atomic_long_t * watchpoint)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. */
remove_watchpoint(atomic_long_t * watchpoint)195 static inline void remove_watchpoint(atomic_long_t *watchpoint)
196 {
197 	atomic_long_set(watchpoint, INVALID_WATCHPOINT);
198 }
199 
get_ctx(void)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! */
kcsan_check_scoped_accesses(void)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
is_atomic(struct kcsan_ctx * ctx,const volatile void * ptr,size_t size,int type)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
should_watch(struct kcsan_ctx * ctx,const volatile void * ptr,size_t size,int type)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  */
kcsan_prandom_u32_max(u32 ep_ro)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 
reset_kcsan_skip(void)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 
kcsan_is_enabled(struct kcsan_ctx * ctx)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. */
delay_access(int type)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  */
read_instrumented_memory(const volatile void * ptr,size_t size)338 static __always_inline u64 read_instrumented_memory(const volatile void *ptr, size_t size)
339 {
340 	/*
341 	 * In the below we don't necessarily need the read of the location to
342 	 * be atomic, and we don't use READ_ONCE(), since all we need for race
343 	 * detection is to observe 2 different values.
344 	 *
345 	 * Furthermore, on certain architectures (such as arm64), READ_ONCE()
346 	 * may turn into more complex instructions than a plain load that cannot
347 	 * do unaligned accesses.
348 	 */
349 	switch (size) {
350 	case 1:  return *(const volatile u8 *)ptr;
351 	case 2:  return *(const volatile u16 *)ptr;
352 	case 4:  return *(const volatile u32 *)ptr;
353 	case 8:  return *(const volatile u64 *)ptr;
354 	default: return 0; /* Ignore; we do not diff the values. */
355 	}
356 }
357 
kcsan_save_irqtrace(struct task_struct * task)358 void kcsan_save_irqtrace(struct task_struct *task)
359 {
360 #ifdef CONFIG_TRACE_IRQFLAGS
361 	task->kcsan_save_irqtrace = task->irqtrace;
362 #endif
363 }
364 
kcsan_restore_irqtrace(struct task_struct * task)365 void kcsan_restore_irqtrace(struct task_struct *task)
366 {
367 #ifdef CONFIG_TRACE_IRQFLAGS
368 	task->irqtrace = task->kcsan_save_irqtrace;
369 #endif
370 }
371 
get_kcsan_stack_depth(void)372 static __always_inline int get_kcsan_stack_depth(void)
373 {
374 #ifdef CONFIG_KCSAN_WEAK_MEMORY
375 	return current->kcsan_stack_depth;
376 #else
377 	BUILD_BUG();
378 	return 0;
379 #endif
380 }
381 
add_kcsan_stack_depth(int val)382 static __always_inline void add_kcsan_stack_depth(int val)
383 {
384 #ifdef CONFIG_KCSAN_WEAK_MEMORY
385 	current->kcsan_stack_depth += val;
386 #else
387 	BUILD_BUG();
388 #endif
389 }
390 
get_reorder_access(struct kcsan_ctx * ctx)391 static __always_inline struct kcsan_scoped_access *get_reorder_access(struct kcsan_ctx *ctx)
392 {
393 #ifdef CONFIG_KCSAN_WEAK_MEMORY
394 	return ctx->disable_scoped ? NULL : &ctx->reorder_access;
395 #else
396 	return NULL;
397 #endif
398 }
399 
400 static __always_inline bool
find_reorder_access(struct kcsan_ctx * ctx,const volatile void * ptr,size_t size,int type,unsigned long ip)401 find_reorder_access(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size,
402 		    int type, unsigned long ip)
403 {
404 	struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);
405 
406 	if (!reorder_access)
407 		return false;
408 
409 	/*
410 	 * Note: If accesses are repeated while reorder_access is identical,
411 	 * never matches the new access, because !(type & KCSAN_ACCESS_SCOPED).
412 	 */
413 	return reorder_access->ptr == ptr && reorder_access->size == size &&
414 	       reorder_access->type == type && reorder_access->ip == ip;
415 }
416 
417 static inline void
set_reorder_access(struct kcsan_ctx * ctx,const volatile void * ptr,size_t size,int type,unsigned long ip)418 set_reorder_access(struct kcsan_ctx *ctx, const volatile void *ptr, size_t size,
419 		   int type, unsigned long ip)
420 {
421 	struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);
422 
423 	if (!reorder_access || !kcsan_weak_memory)
424 		return;
425 
426 	/*
427 	 * To avoid nested interrupts or scheduler (which share kcsan_ctx)
428 	 * reading an inconsistent reorder_access, ensure that the below has
429 	 * exclusive access to reorder_access by disallowing concurrent use.
430 	 */
431 	ctx->disable_scoped++;
432 	barrier();
433 	reorder_access->ptr		= ptr;
434 	reorder_access->size		= size;
435 	reorder_access->type		= type | KCSAN_ACCESS_SCOPED;
436 	reorder_access->ip		= ip;
437 	reorder_access->stack_depth	= get_kcsan_stack_depth();
438 	barrier();
439 	ctx->disable_scoped--;
440 }
441 
442 /*
443  * Pull everything together: check_access() below contains the performance
444  * critical operations; the fast-path (including check_access) functions should
445  * all be inlinable by the instrumentation functions.
446  *
447  * The slow-path (kcsan_found_watchpoint, kcsan_setup_watchpoint) are
448  * non-inlinable -- note that, we prefix these with "kcsan_" to ensure they can
449  * be filtered from the stacktrace, as well as give them unique names for the
450  * UACCESS whitelist of objtool. Each function uses user_access_save/restore(),
451  * since they do not access any user memory, but instrumentation is still
452  * emitted in UACCESS regions.
453  */
454 
kcsan_found_watchpoint(const volatile void * ptr,size_t size,int type,unsigned long ip,atomic_long_t * watchpoint,long encoded_watchpoint)455 static noinline void kcsan_found_watchpoint(const volatile void *ptr,
456 					    size_t size,
457 					    int type,
458 					    unsigned long ip,
459 					    atomic_long_t *watchpoint,
460 					    long encoded_watchpoint)
461 {
462 	const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
463 	struct kcsan_ctx *ctx = get_ctx();
464 	unsigned long flags;
465 	bool consumed;
466 
467 	/*
468 	 * We know a watchpoint exists. Let's try to keep the race-window
469 	 * between here and finally consuming the watchpoint below as small as
470 	 * possible -- avoid unneccessarily complex code until consumed.
471 	 */
472 
473 	if (!kcsan_is_enabled(ctx))
474 		return;
475 
476 	/*
477 	 * The access_mask check relies on value-change comparison. To avoid
478 	 * reporting a race where e.g. the writer set up the watchpoint, but the
479 	 * reader has access_mask!=0, we have to ignore the found watchpoint.
480 	 *
481 	 * reorder_access is never created from an access with access_mask set.
482 	 */
483 	if (ctx->access_mask && !find_reorder_access(ctx, ptr, size, type, ip))
484 		return;
485 
486 	/*
487 	 * If the other thread does not want to ignore the access, and there was
488 	 * a value change as a result of this thread's operation, we will still
489 	 * generate a report of unknown origin.
490 	 *
491 	 * Use CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN=n to filter.
492 	 */
493 	if (!is_assert && kcsan_ignore_address(ptr))
494 		return;
495 
496 	/*
497 	 * Consuming the watchpoint must be guarded by kcsan_is_enabled() to
498 	 * avoid erroneously triggering reports if the context is disabled.
499 	 */
500 	consumed = try_consume_watchpoint(watchpoint, encoded_watchpoint);
501 
502 	/* keep this after try_consume_watchpoint */
503 	flags = user_access_save();
504 
505 	if (consumed) {
506 		kcsan_save_irqtrace(current);
507 		kcsan_report_set_info(ptr, size, type, ip, watchpoint - watchpoints);
508 		kcsan_restore_irqtrace(current);
509 	} else {
510 		/*
511 		 * The other thread may not print any diagnostics, as it has
512 		 * already removed the watchpoint, or another thread consumed
513 		 * the watchpoint before this thread.
514 		 */
515 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_REPORT_RACES]);
516 	}
517 
518 	if (is_assert)
519 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
520 	else
521 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_DATA_RACES]);
522 
523 	user_access_restore(flags);
524 }
525 
526 static noinline void
kcsan_setup_watchpoint(const volatile void * ptr,size_t size,int type,unsigned long ip)527 kcsan_setup_watchpoint(const volatile void *ptr, size_t size, int type, unsigned long ip)
528 {
529 	const bool is_write = (type & KCSAN_ACCESS_WRITE) != 0;
530 	const bool is_assert = (type & KCSAN_ACCESS_ASSERT) != 0;
531 	atomic_long_t *watchpoint;
532 	u64 old, new, diff;
533 	enum kcsan_value_change value_change = KCSAN_VALUE_CHANGE_MAYBE;
534 	bool interrupt_watcher = kcsan_interrupt_watcher;
535 	unsigned long ua_flags = user_access_save();
536 	struct kcsan_ctx *ctx = get_ctx();
537 	unsigned long access_mask = ctx->access_mask;
538 	unsigned long irq_flags = 0;
539 	bool is_reorder_access;
540 
541 	/*
542 	 * Always reset kcsan_skip counter in slow-path to avoid underflow; see
543 	 * should_watch().
544 	 */
545 	reset_kcsan_skip();
546 
547 	if (!kcsan_is_enabled(ctx))
548 		goto out;
549 
550 	/*
551 	 * Check to-ignore addresses after kcsan_is_enabled(), as we may access
552 	 * memory that is not yet initialized during early boot.
553 	 */
554 	if (!is_assert && kcsan_ignore_address(ptr))
555 		goto out;
556 
557 	if (!check_encodable((unsigned long)ptr, size)) {
558 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_UNENCODABLE_ACCESSES]);
559 		goto out;
560 	}
561 
562 	/*
563 	 * The local CPU cannot observe reordering of its own accesses, and
564 	 * therefore we need to take care of 2 cases to avoid false positives:
565 	 *
566 	 *	1. Races of the reordered access with interrupts. To avoid, if
567 	 *	   the current access is reorder_access, disable interrupts.
568 	 *	2. Avoid races of scoped accesses from nested interrupts (below).
569 	 */
570 	is_reorder_access = find_reorder_access(ctx, ptr, size, type, ip);
571 	if (is_reorder_access)
572 		interrupt_watcher = false;
573 	/*
574 	 * Avoid races of scoped accesses from nested interrupts (or scheduler).
575 	 * Assume setting up a watchpoint for a non-scoped (normal) access that
576 	 * also conflicts with a current scoped access. In a nested interrupt,
577 	 * which shares the context, it would check a conflicting scoped access.
578 	 * To avoid, disable scoped access checking.
579 	 */
580 	ctx->disable_scoped++;
581 
582 	/*
583 	 * Save and restore the IRQ state trace touched by KCSAN, since KCSAN's
584 	 * runtime is entered for every memory access, and potentially useful
585 	 * information is lost if dirtied by KCSAN.
586 	 */
587 	kcsan_save_irqtrace(current);
588 	if (!interrupt_watcher) {
589 		local_irq_save(irq_flags);
590 		/*
591 		 * NMIs can still fire, disable checking for all interrupt
592 		 * contexts.
593 		 */
594 		raw_cpu_ptr(&kcsan_cpu_ctx)->disable_count++;
595 	}
596 
597 	watchpoint = insert_watchpoint((unsigned long)ptr, size, is_write);
598 	if (watchpoint == NULL) {
599 		/*
600 		 * Out of capacity: the size of 'watchpoints', and the frequency
601 		 * with which should_watch() returns true should be tweaked so
602 		 * that this case happens very rarely.
603 		 */
604 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_NO_CAPACITY]);
605 		goto out_unlock;
606 	}
607 
608 	atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_SETUP_WATCHPOINTS]);
609 	atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);
610 
611 	/*
612 	 * Read the current value, to later check and infer a race if the data
613 	 * was modified via a non-instrumented access, e.g. from a device.
614 	 */
615 	old = is_reorder_access ? 0 : read_instrumented_memory(ptr, size);
616 
617 	/*
618 	 * Delay this thread, to increase probability of observing a racy
619 	 * conflicting access.
620 	 */
621 	delay_access(type);
622 
623 	/*
624 	 * Re-read value, and check if it is as expected; if not, we infer a
625 	 * racy access.
626 	 */
627 	if (!is_reorder_access) {
628 		new = read_instrumented_memory(ptr, size);
629 	} else {
630 		/*
631 		 * Reordered accesses cannot be used for value change detection,
632 		 * because the memory location may no longer be accessible and
633 		 * could result in a fault.
634 		 */
635 		new = 0;
636 		access_mask = 0;
637 	}
638 
639 	diff = old ^ new;
640 	if (access_mask)
641 		diff &= access_mask;
642 
643 	/*
644 	 * Check if we observed a value change.
645 	 *
646 	 * Also check if the data race should be ignored (the rules depend on
647 	 * non-zero diff); if it is to be ignored, the below rules for
648 	 * KCSAN_VALUE_CHANGE_MAYBE apply.
649 	 */
650 	if (diff && !kcsan_ignore_data_race(size, type, old, new, diff))
651 		value_change = KCSAN_VALUE_CHANGE_TRUE;
652 
653 	/* Check if this access raced with another. */
654 	if (!consume_watchpoint(watchpoint)) {
655 		/*
656 		 * Depending on the access type, map a value_change of MAYBE to
657 		 * TRUE (always report) or FALSE (never report).
658 		 */
659 		if (value_change == KCSAN_VALUE_CHANGE_MAYBE) {
660 			if (access_mask != 0) {
661 				/*
662 				 * For access with access_mask, we require a
663 				 * value-change, as it is likely that races on
664 				 * ~access_mask bits are expected.
665 				 */
666 				value_change = KCSAN_VALUE_CHANGE_FALSE;
667 			} else if (size > 8 || is_assert) {
668 				/* Always assume a value-change. */
669 				value_change = KCSAN_VALUE_CHANGE_TRUE;
670 			}
671 		}
672 
673 		/*
674 		 * No need to increment 'data_races' counter, as the racing
675 		 * thread already did.
676 		 *
677 		 * Count 'assert_failures' for each failed ASSERT access,
678 		 * therefore both this thread and the racing thread may
679 		 * increment this counter.
680 		 */
681 		if (is_assert && value_change == KCSAN_VALUE_CHANGE_TRUE)
682 			atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
683 
684 		kcsan_report_known_origin(ptr, size, type, ip,
685 					  value_change, watchpoint - watchpoints,
686 					  old, new, access_mask);
687 	} else if (value_change == KCSAN_VALUE_CHANGE_TRUE) {
688 		/* Inferring a race, since the value should not have changed. */
689 
690 		atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_RACES_UNKNOWN_ORIGIN]);
691 		if (is_assert)
692 			atomic_long_inc(&kcsan_counters[KCSAN_COUNTER_ASSERT_FAILURES]);
693 
694 		if (IS_ENABLED(CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN) || is_assert) {
695 			kcsan_report_unknown_origin(ptr, size, type, ip,
696 						    old, new, access_mask);
697 		}
698 	}
699 
700 	/*
701 	 * Remove watchpoint; must be after reporting, since the slot may be
702 	 * reused after this point.
703 	 */
704 	remove_watchpoint(watchpoint);
705 	atomic_long_dec(&kcsan_counters[KCSAN_COUNTER_USED_WATCHPOINTS]);
706 
707 out_unlock:
708 	if (!interrupt_watcher) {
709 		raw_cpu_ptr(&kcsan_cpu_ctx)->disable_count--;
710 		local_irq_restore(irq_flags);
711 	}
712 	kcsan_restore_irqtrace(current);
713 	ctx->disable_scoped--;
714 
715 	/*
716 	 * Reordered accesses cannot be used for value change detection,
717 	 * therefore never consider for reordering if access_mask is set.
718 	 * ASSERT_EXCLUSIVE are not real accesses, ignore them as well.
719 	 */
720 	if (!access_mask && !is_assert)
721 		set_reorder_access(ctx, ptr, size, type, ip);
722 out:
723 	user_access_restore(ua_flags);
724 }
725 
726 static __always_inline void
check_access(const volatile void * ptr,size_t size,int type,unsigned long ip)727 check_access(const volatile void *ptr, size_t size, int type, unsigned long ip)
728 {
729 	atomic_long_t *watchpoint;
730 	long encoded_watchpoint;
731 
732 	/*
733 	 * Do nothing for 0 sized check; this comparison will be optimized out
734 	 * for constant sized instrumentation (__tsan_{read,write}N).
735 	 */
736 	if (unlikely(size == 0))
737 		return;
738 
739 again:
740 	/*
741 	 * Avoid user_access_save in fast-path: find_watchpoint is safe without
742 	 * user_access_save, as the address that ptr points to is only used to
743 	 * check if a watchpoint exists; ptr is never dereferenced.
744 	 */
745 	watchpoint = find_watchpoint((unsigned long)ptr, size,
746 				     !(type & KCSAN_ACCESS_WRITE),
747 				     &encoded_watchpoint);
748 	/*
749 	 * It is safe to check kcsan_is_enabled() after find_watchpoint in the
750 	 * slow-path, as long as no state changes that cause a race to be
751 	 * detected and reported have occurred until kcsan_is_enabled() is
752 	 * checked.
753 	 */
754 
755 	if (unlikely(watchpoint != NULL))
756 		kcsan_found_watchpoint(ptr, size, type, ip, watchpoint, encoded_watchpoint);
757 	else {
758 		struct kcsan_ctx *ctx = get_ctx(); /* Call only once in fast-path. */
759 
760 		if (unlikely(should_watch(ctx, ptr, size, type))) {
761 			kcsan_setup_watchpoint(ptr, size, type, ip);
762 			return;
763 		}
764 
765 		if (!(type & KCSAN_ACCESS_SCOPED)) {
766 			struct kcsan_scoped_access *reorder_access = get_reorder_access(ctx);
767 
768 			if (reorder_access) {
769 				/*
770 				 * reorder_access check: simulates reordering of
771 				 * the access after subsequent operations.
772 				 */
773 				ptr = reorder_access->ptr;
774 				type = reorder_access->type;
775 				ip = reorder_access->ip;
776 				/*
777 				 * Upon a nested interrupt, this context's
778 				 * reorder_access can be modified (shared ctx).
779 				 * We know that upon return, reorder_access is
780 				 * always invalidated by setting size to 0 via
781 				 * __tsan_func_exit(). Therefore we must read
782 				 * and check size after the other fields.
783 				 */
784 				barrier();
785 				size = READ_ONCE(reorder_access->size);
786 				if (size)
787 					goto again;
788 			}
789 		}
790 
791 		/*
792 		 * Always checked last, right before returning from runtime;
793 		 * if reorder_access is valid, checked after it was checked.
794 		 */
795 		if (unlikely(ctx->scoped_accesses.prev))
796 			kcsan_check_scoped_accesses();
797 	}
798 }
799 
800 /* === Public interface ===================================================== */
801 
kcsan_init(void)802 void __init kcsan_init(void)
803 {
804 	int cpu;
805 
806 	BUG_ON(!in_task());
807 
808 	for_each_possible_cpu(cpu)
809 		per_cpu(kcsan_rand_state, cpu) = (u32)get_cycles();
810 
811 	/*
812 	 * We are in the init task, and no other tasks should be running;
813 	 * WRITE_ONCE without memory barrier is sufficient.
814 	 */
815 	if (kcsan_early_enable) {
816 		pr_info("enabled early\n");
817 		WRITE_ONCE(kcsan_enabled, true);
818 	}
819 
820 	if (IS_ENABLED(CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY) ||
821 	    IS_ENABLED(CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC) ||
822 	    IS_ENABLED(CONFIG_KCSAN_PERMISSIVE) ||
823 	    IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {
824 		pr_warn("non-strict mode configured - use CONFIG_KCSAN_STRICT=y to see all data races\n");
825 	} else {
826 		pr_info("strict mode configured\n");
827 	}
828 }
829 
830 /* === Exported interface =================================================== */
831 
kcsan_disable_current(void)832 void kcsan_disable_current(void)
833 {
834 	++get_ctx()->disable_count;
835 }
836 EXPORT_SYMBOL(kcsan_disable_current);
837 
kcsan_enable_current(void)838 void kcsan_enable_current(void)
839 {
840 	if (get_ctx()->disable_count-- == 0) {
841 		/*
842 		 * Warn if kcsan_enable_current() calls are unbalanced with
843 		 * kcsan_disable_current() calls, which causes disable_count to
844 		 * become negative and should not happen.
845 		 */
846 		kcsan_disable_current(); /* restore to 0, KCSAN still enabled */
847 		kcsan_disable_current(); /* disable to generate warning */
848 		WARN(1, "Unbalanced %s()", __func__);
849 		kcsan_enable_current();
850 	}
851 }
852 EXPORT_SYMBOL(kcsan_enable_current);
853 
kcsan_enable_current_nowarn(void)854 void kcsan_enable_current_nowarn(void)
855 {
856 	if (get_ctx()->disable_count-- == 0)
857 		kcsan_disable_current();
858 }
859 EXPORT_SYMBOL(kcsan_enable_current_nowarn);
860 
kcsan_nestable_atomic_begin(void)861 void kcsan_nestable_atomic_begin(void)
862 {
863 	/*
864 	 * Do *not* check and warn if we are in a flat atomic region: nestable
865 	 * and flat atomic regions are independent from each other.
866 	 * See include/linux/kcsan.h: struct kcsan_ctx comments for more
867 	 * comments.
868 	 */
869 
870 	++get_ctx()->atomic_nest_count;
871 }
872 EXPORT_SYMBOL(kcsan_nestable_atomic_begin);
873 
kcsan_nestable_atomic_end(void)874 void kcsan_nestable_atomic_end(void)
875 {
876 	if (get_ctx()->atomic_nest_count-- == 0) {
877 		/*
878 		 * Warn if kcsan_nestable_atomic_end() calls are unbalanced with
879 		 * kcsan_nestable_atomic_begin() calls, which causes
880 		 * atomic_nest_count to become negative and should not happen.
881 		 */
882 		kcsan_nestable_atomic_begin(); /* restore to 0 */
883 		kcsan_disable_current(); /* disable to generate warning */
884 		WARN(1, "Unbalanced %s()", __func__);
885 		kcsan_enable_current();
886 	}
887 }
888 EXPORT_SYMBOL(kcsan_nestable_atomic_end);
889 
kcsan_flat_atomic_begin(void)890 void kcsan_flat_atomic_begin(void)
891 {
892 	get_ctx()->in_flat_atomic = true;
893 }
894 EXPORT_SYMBOL(kcsan_flat_atomic_begin);
895 
kcsan_flat_atomic_end(void)896 void kcsan_flat_atomic_end(void)
897 {
898 	get_ctx()->in_flat_atomic = false;
899 }
900 EXPORT_SYMBOL(kcsan_flat_atomic_end);
901 
kcsan_atomic_next(int n)902 void kcsan_atomic_next(int n)
903 {
904 	get_ctx()->atomic_next = n;
905 }
906 EXPORT_SYMBOL(kcsan_atomic_next);
907 
kcsan_set_access_mask(unsigned long mask)908 void kcsan_set_access_mask(unsigned long mask)
909 {
910 	get_ctx()->access_mask = mask;
911 }
912 EXPORT_SYMBOL(kcsan_set_access_mask);
913 
914 struct kcsan_scoped_access *
kcsan_begin_scoped_access(const volatile void * ptr,size_t size,int type,struct kcsan_scoped_access * sa)915 kcsan_begin_scoped_access(const volatile void *ptr, size_t size, int type,
916 			  struct kcsan_scoped_access *sa)
917 {
918 	struct kcsan_ctx *ctx = get_ctx();
919 
920 	check_access(ptr, size, type, _RET_IP_);
921 
922 	ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */
923 
924 	INIT_LIST_HEAD(&sa->list);
925 	sa->ptr = ptr;
926 	sa->size = size;
927 	sa->type = type;
928 	sa->ip = _RET_IP_;
929 
930 	if (!ctx->scoped_accesses.prev) /* Lazy initialize list head. */
931 		INIT_LIST_HEAD(&ctx->scoped_accesses);
932 	list_add(&sa->list, &ctx->scoped_accesses);
933 
934 	ctx->disable_count--;
935 	return sa;
936 }
937 EXPORT_SYMBOL(kcsan_begin_scoped_access);
938 
kcsan_end_scoped_access(struct kcsan_scoped_access * sa)939 void kcsan_end_scoped_access(struct kcsan_scoped_access *sa)
940 {
941 	struct kcsan_ctx *ctx = get_ctx();
942 
943 	if (WARN(!ctx->scoped_accesses.prev, "Unbalanced %s()?", __func__))
944 		return;
945 
946 	ctx->disable_count++; /* Disable KCSAN, in case list debugging is on. */
947 
948 	list_del(&sa->list);
949 	if (list_empty(&ctx->scoped_accesses))
950 		/*
951 		 * Ensure we do not enter kcsan_check_scoped_accesses()
952 		 * slow-path if unnecessary, and avoids requiring list_empty()
953 		 * in the fast-path (to avoid a READ_ONCE() and potential
954 		 * uaccess warning).
955 		 */
956 		ctx->scoped_accesses.prev = NULL;
957 
958 	ctx->disable_count--;
959 
960 	check_access(sa->ptr, sa->size, sa->type, sa->ip);
961 }
962 EXPORT_SYMBOL(kcsan_end_scoped_access);
963 
__kcsan_check_access(const volatile void * ptr,size_t size,int type)964 void __kcsan_check_access(const volatile void *ptr, size_t size, int type)
965 {
966 	check_access(ptr, size, type, _RET_IP_);
967 }
968 EXPORT_SYMBOL(__kcsan_check_access);
969 
970 #define DEFINE_MEMORY_BARRIER(name, order_before_cond)				\
971 	void __kcsan_##name(void)						\
972 	{									\
973 		struct kcsan_scoped_access *sa = get_reorder_access(get_ctx());	\
974 		if (!sa)							\
975 			return;							\
976 		if (order_before_cond)						\
977 			sa->size = 0;						\
978 	}									\
979 	EXPORT_SYMBOL(__kcsan_##name)
980 
981 DEFINE_MEMORY_BARRIER(mb, true);
982 DEFINE_MEMORY_BARRIER(wmb, sa->type & (KCSAN_ACCESS_WRITE | KCSAN_ACCESS_COMPOUND));
983 DEFINE_MEMORY_BARRIER(rmb, !(sa->type & KCSAN_ACCESS_WRITE) || (sa->type & KCSAN_ACCESS_COMPOUND));
984 DEFINE_MEMORY_BARRIER(release, true);
985 
986 /*
987  * KCSAN uses the same instrumentation that is emitted by supported compilers
988  * for ThreadSanitizer (TSAN).
989  *
990  * When enabled, the compiler emits instrumentation calls (the functions
991  * prefixed with "__tsan" below) for all loads and stores that it generated;
992  * inline asm is not instrumented.
993  *
994  * Note that, not all supported compiler versions distinguish aligned/unaligned
995  * accesses, but e.g. recent versions of Clang do. We simply alias the unaligned
996  * version to the generic version, which can handle both.
997  */
998 
999 #define DEFINE_TSAN_READ_WRITE(size)                                           \
1000 	void __tsan_read##size(void *ptr);                                     \
1001 	void __tsan_read##size(void *ptr)                                      \
1002 	{                                                                      \
1003 		check_access(ptr, size, 0, _RET_IP_);                          \
1004 	}                                                                      \
1005 	EXPORT_SYMBOL(__tsan_read##size);                                      \
1006 	void __tsan_unaligned_read##size(void *ptr)                            \
1007 		__alias(__tsan_read##size);                                    \
1008 	EXPORT_SYMBOL(__tsan_unaligned_read##size);                            \
1009 	void __tsan_write##size(void *ptr);                                    \
1010 	void __tsan_write##size(void *ptr)                                     \
1011 	{                                                                      \
1012 		check_access(ptr, size, KCSAN_ACCESS_WRITE, _RET_IP_);         \
1013 	}                                                                      \
1014 	EXPORT_SYMBOL(__tsan_write##size);                                     \
1015 	void __tsan_unaligned_write##size(void *ptr)                           \
1016 		__alias(__tsan_write##size);                                   \
1017 	EXPORT_SYMBOL(__tsan_unaligned_write##size);                           \
1018 	void __tsan_read_write##size(void *ptr);                               \
1019 	void __tsan_read_write##size(void *ptr)                                \
1020 	{                                                                      \
1021 		check_access(ptr, size,                                        \
1022 			     KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE,       \
1023 			     _RET_IP_);                                        \
1024 	}                                                                      \
1025 	EXPORT_SYMBOL(__tsan_read_write##size);                                \
1026 	void __tsan_unaligned_read_write##size(void *ptr)                      \
1027 		__alias(__tsan_read_write##size);                              \
1028 	EXPORT_SYMBOL(__tsan_unaligned_read_write##size)
1029 
1030 DEFINE_TSAN_READ_WRITE(1);
1031 DEFINE_TSAN_READ_WRITE(2);
1032 DEFINE_TSAN_READ_WRITE(4);
1033 DEFINE_TSAN_READ_WRITE(8);
1034 DEFINE_TSAN_READ_WRITE(16);
1035 
1036 void __tsan_read_range(void *ptr, size_t size);
__tsan_read_range(void * ptr,size_t size)1037 void __tsan_read_range(void *ptr, size_t size)
1038 {
1039 	check_access(ptr, size, 0, _RET_IP_);
1040 }
1041 EXPORT_SYMBOL(__tsan_read_range);
1042 
1043 void __tsan_write_range(void *ptr, size_t size);
__tsan_write_range(void * ptr,size_t size)1044 void __tsan_write_range(void *ptr, size_t size)
1045 {
1046 	check_access(ptr, size, KCSAN_ACCESS_WRITE, _RET_IP_);
1047 }
1048 EXPORT_SYMBOL(__tsan_write_range);
1049 
1050 /*
1051  * Use of explicit volatile is generally disallowed [1], however, volatile is
1052  * still used in various concurrent context, whether in low-level
1053  * synchronization primitives or for legacy reasons.
1054  * [1] https://lwn.net/Articles/233479/
1055  *
1056  * We only consider volatile accesses atomic if they are aligned and would pass
1057  * the size-check of compiletime_assert_rwonce_type().
1058  */
1059 #define DEFINE_TSAN_VOLATILE_READ_WRITE(size)                                  \
1060 	void __tsan_volatile_read##size(void *ptr);                            \
1061 	void __tsan_volatile_read##size(void *ptr)                             \
1062 	{                                                                      \
1063 		const bool is_atomic = size <= sizeof(long long) &&            \
1064 				       IS_ALIGNED((unsigned long)ptr, size);   \
1065 		if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
1066 			return;                                                \
1067 		check_access(ptr, size, is_atomic ? KCSAN_ACCESS_ATOMIC : 0,   \
1068 			     _RET_IP_);                                        \
1069 	}                                                                      \
1070 	EXPORT_SYMBOL(__tsan_volatile_read##size);                             \
1071 	void __tsan_unaligned_volatile_read##size(void *ptr)                   \
1072 		__alias(__tsan_volatile_read##size);                           \
1073 	EXPORT_SYMBOL(__tsan_unaligned_volatile_read##size);                   \
1074 	void __tsan_volatile_write##size(void *ptr);                           \
1075 	void __tsan_volatile_write##size(void *ptr)                            \
1076 	{                                                                      \
1077 		const bool is_atomic = size <= sizeof(long long) &&            \
1078 				       IS_ALIGNED((unsigned long)ptr, size);   \
1079 		if (IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS) && is_atomic)      \
1080 			return;                                                \
1081 		check_access(ptr, size,                                        \
1082 			     KCSAN_ACCESS_WRITE |                              \
1083 				     (is_atomic ? KCSAN_ACCESS_ATOMIC : 0),    \
1084 			     _RET_IP_);                                        \
1085 	}                                                                      \
1086 	EXPORT_SYMBOL(__tsan_volatile_write##size);                            \
1087 	void __tsan_unaligned_volatile_write##size(void *ptr)                  \
1088 		__alias(__tsan_volatile_write##size);                          \
1089 	EXPORT_SYMBOL(__tsan_unaligned_volatile_write##size)
1090 
1091 DEFINE_TSAN_VOLATILE_READ_WRITE(1);
1092 DEFINE_TSAN_VOLATILE_READ_WRITE(2);
1093 DEFINE_TSAN_VOLATILE_READ_WRITE(4);
1094 DEFINE_TSAN_VOLATILE_READ_WRITE(8);
1095 DEFINE_TSAN_VOLATILE_READ_WRITE(16);
1096 
1097 /*
1098  * Function entry and exit are used to determine the validty of reorder_access.
1099  * Reordering of the access ends at the end of the function scope where the
1100  * access happened. This is done for two reasons:
1101  *
1102  *	1. Artificially limits the scope where missing barriers are detected.
1103  *	   This minimizes false positives due to uninstrumented functions that
1104  *	   contain the required barriers but were missed.
1105  *
1106  *	2. Simplifies generating the stack trace of the access.
1107  */
1108 void __tsan_func_entry(void *call_pc);
__tsan_func_entry(void * call_pc)1109 noinline void __tsan_func_entry(void *call_pc)
1110 {
1111 	if (!IS_ENABLED(CONFIG_KCSAN_WEAK_MEMORY))
1112 		return;
1113 
1114 	add_kcsan_stack_depth(1);
1115 }
1116 EXPORT_SYMBOL(__tsan_func_entry);
1117 
1118 void __tsan_func_exit(void);
__tsan_func_exit(void)1119 noinline void __tsan_func_exit(void)
1120 {
1121 	struct kcsan_scoped_access *reorder_access;
1122 
1123 	if (!IS_ENABLED(CONFIG_KCSAN_WEAK_MEMORY))
1124 		return;
1125 
1126 	reorder_access = get_reorder_access(get_ctx());
1127 	if (!reorder_access)
1128 		goto out;
1129 
1130 	if (get_kcsan_stack_depth() <= reorder_access->stack_depth) {
1131 		/*
1132 		 * Access check to catch cases where write without a barrier
1133 		 * (supposed release) was last access in function: because
1134 		 * instrumentation is inserted before the real access, a data
1135 		 * race due to the write giving up a c-s would only be caught if
1136 		 * we do the conflicting access after.
1137 		 */
1138 		check_access(reorder_access->ptr, reorder_access->size,
1139 			     reorder_access->type, reorder_access->ip);
1140 		reorder_access->size = 0;
1141 		reorder_access->stack_depth = INT_MIN;
1142 	}
1143 out:
1144 	add_kcsan_stack_depth(-1);
1145 }
1146 EXPORT_SYMBOL(__tsan_func_exit);
1147 
1148 void __tsan_init(void);
__tsan_init(void)1149 void __tsan_init(void)
1150 {
1151 }
1152 EXPORT_SYMBOL(__tsan_init);
1153 
1154 /*
1155  * Instrumentation for atomic builtins (__atomic_*, __sync_*).
1156  *
1157  * Normal kernel code _should not_ be using them directly, but some
1158  * architectures may implement some or all atomics using the compilers'
1159  * builtins.
1160  *
1161  * Note: If an architecture decides to fully implement atomics using the
1162  * builtins, because they are implicitly instrumented by KCSAN (and KASAN,
1163  * etc.), implementing the ARCH_ATOMIC interface (to get instrumentation via
1164  * atomic-instrumented) is no longer necessary.
1165  *
1166  * TSAN instrumentation replaces atomic accesses with calls to any of the below
1167  * functions, whose job is to also execute the operation itself.
1168  */
1169 
kcsan_atomic_builtin_memorder(int memorder)1170 static __always_inline void kcsan_atomic_builtin_memorder(int memorder)
1171 {
1172 	if (memorder == __ATOMIC_RELEASE ||
1173 	    memorder == __ATOMIC_SEQ_CST ||
1174 	    memorder == __ATOMIC_ACQ_REL)
1175 		__kcsan_release();
1176 }
1177 
1178 #define DEFINE_TSAN_ATOMIC_LOAD_STORE(bits)                                                        \
1179 	u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder);                      \
1180 	u##bits __tsan_atomic##bits##_load(const u##bits *ptr, int memorder)                       \
1181 	{                                                                                          \
1182 		kcsan_atomic_builtin_memorder(memorder);                                           \
1183 		if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1184 			check_access(ptr, bits / BITS_PER_BYTE, KCSAN_ACCESS_ATOMIC, _RET_IP_);    \
1185 		}                                                                                  \
1186 		return __atomic_load_n(ptr, memorder);                                             \
1187 	}                                                                                          \
1188 	EXPORT_SYMBOL(__tsan_atomic##bits##_load);                                                 \
1189 	void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder);                   \
1190 	void __tsan_atomic##bits##_store(u##bits *ptr, u##bits v, int memorder)                    \
1191 	{                                                                                          \
1192 		kcsan_atomic_builtin_memorder(memorder);                                           \
1193 		if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1194 			check_access(ptr, bits / BITS_PER_BYTE,                                    \
1195 				     KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ATOMIC, _RET_IP_);          \
1196 		}                                                                                  \
1197 		__atomic_store_n(ptr, v, memorder);                                                \
1198 	}                                                                                          \
1199 	EXPORT_SYMBOL(__tsan_atomic##bits##_store)
1200 
1201 #define DEFINE_TSAN_ATOMIC_RMW(op, bits, suffix)                                                   \
1202 	u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder);                 \
1203 	u##bits __tsan_atomic##bits##_##op(u##bits *ptr, u##bits v, int memorder)                  \
1204 	{                                                                                          \
1205 		kcsan_atomic_builtin_memorder(memorder);                                           \
1206 		if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1207 			check_access(ptr, bits / BITS_PER_BYTE,                                    \
1208 				     KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1209 					     KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
1210 		}                                                                                  \
1211 		return __atomic_##op##suffix(ptr, v, memorder);                                    \
1212 	}                                                                                          \
1213 	EXPORT_SYMBOL(__tsan_atomic##bits##_##op)
1214 
1215 /*
1216  * Note: CAS operations are always classified as write, even in case they
1217  * fail. We cannot perform check_access() after a write, as it might lead to
1218  * false positives, in cases such as:
1219  *
1220  *	T0: __atomic_compare_exchange_n(&p->flag, &old, 1, ...)
1221  *
1222  *	T1: if (__atomic_load_n(&p->flag, ...)) {
1223  *		modify *p;
1224  *		p->flag = 0;
1225  *	    }
1226  *
1227  * The only downside is that, if there are 3 threads, with one CAS that
1228  * succeeds, another CAS that fails, and an unmarked racing operation, we may
1229  * point at the wrong CAS as the source of the race. However, if we assume that
1230  * all CAS can succeed in some other execution, the data race is still valid.
1231  */
1232 #define DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strength, weak)                                           \
1233 	int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
1234 							      u##bits val, int mo, int fail_mo);   \
1235 	int __tsan_atomic##bits##_compare_exchange_##strength(u##bits *ptr, u##bits *exp,          \
1236 							      u##bits val, int mo, int fail_mo)    \
1237 	{                                                                                          \
1238 		kcsan_atomic_builtin_memorder(mo);                                                 \
1239 		if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1240 			check_access(ptr, bits / BITS_PER_BYTE,                                    \
1241 				     KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1242 					     KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
1243 		}                                                                                  \
1244 		return __atomic_compare_exchange_n(ptr, exp, val, weak, mo, fail_mo);              \
1245 	}                                                                                          \
1246 	EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_##strength)
1247 
1248 #define DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)                                                       \
1249 	u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
1250 							   int mo, int fail_mo);                   \
1251 	u##bits __tsan_atomic##bits##_compare_exchange_val(u##bits *ptr, u##bits exp, u##bits val, \
1252 							   int mo, int fail_mo)                    \
1253 	{                                                                                          \
1254 		kcsan_atomic_builtin_memorder(mo);                                                 \
1255 		if (!IS_ENABLED(CONFIG_KCSAN_IGNORE_ATOMICS)) {                                    \
1256 			check_access(ptr, bits / BITS_PER_BYTE,                                    \
1257 				     KCSAN_ACCESS_COMPOUND | KCSAN_ACCESS_WRITE |                  \
1258 					     KCSAN_ACCESS_ATOMIC, _RET_IP_);                       \
1259 		}                                                                                  \
1260 		__atomic_compare_exchange_n(ptr, &exp, val, 0, mo, fail_mo);                       \
1261 		return exp;                                                                        \
1262 	}                                                                                          \
1263 	EXPORT_SYMBOL(__tsan_atomic##bits##_compare_exchange_val)
1264 
1265 #define DEFINE_TSAN_ATOMIC_OPS(bits)                                                               \
1266 	DEFINE_TSAN_ATOMIC_LOAD_STORE(bits);                                                       \
1267 	DEFINE_TSAN_ATOMIC_RMW(exchange, bits, _n);                                                \
1268 	DEFINE_TSAN_ATOMIC_RMW(fetch_add, bits, );                                                 \
1269 	DEFINE_TSAN_ATOMIC_RMW(fetch_sub, bits, );                                                 \
1270 	DEFINE_TSAN_ATOMIC_RMW(fetch_and, bits, );                                                 \
1271 	DEFINE_TSAN_ATOMIC_RMW(fetch_or, bits, );                                                  \
1272 	DEFINE_TSAN_ATOMIC_RMW(fetch_xor, bits, );                                                 \
1273 	DEFINE_TSAN_ATOMIC_RMW(fetch_nand, bits, );                                                \
1274 	DEFINE_TSAN_ATOMIC_CMPXCHG(bits, strong, 0);                                               \
1275 	DEFINE_TSAN_ATOMIC_CMPXCHG(bits, weak, 1);                                                 \
1276 	DEFINE_TSAN_ATOMIC_CMPXCHG_VAL(bits)
1277 
1278 DEFINE_TSAN_ATOMIC_OPS(8);
1279 DEFINE_TSAN_ATOMIC_OPS(16);
1280 DEFINE_TSAN_ATOMIC_OPS(32);
1281 #ifdef CONFIG_64BIT
1282 DEFINE_TSAN_ATOMIC_OPS(64);
1283 #endif
1284 
1285 void __tsan_atomic_thread_fence(int memorder);
__tsan_atomic_thread_fence(int memorder)1286 void __tsan_atomic_thread_fence(int memorder)
1287 {
1288 	kcsan_atomic_builtin_memorder(memorder);
1289 	__atomic_thread_fence(memorder);
1290 }
1291 EXPORT_SYMBOL(__tsan_atomic_thread_fence);
1292 
1293 /*
1294  * In instrumented files, we emit instrumentation for barriers by mapping the
1295  * kernel barriers to an __atomic_signal_fence(), which is interpreted specially
1296  * and otherwise has no relation to a real __atomic_signal_fence(). No known
1297  * kernel code uses __atomic_signal_fence().
1298  *
1299  * Since fsanitize=thread instrumentation handles __atomic_signal_fence(), which
1300  * are turned into calls to __tsan_atomic_signal_fence(), such instrumentation
1301  * can be disabled via the __no_kcsan function attribute (vs. an explicit call
1302  * which could not). When __no_kcsan is requested, __atomic_signal_fence()
1303  * generates no code.
1304  *
1305  * Note: The result of using __atomic_signal_fence() with KCSAN enabled is
1306  * potentially limiting the compiler's ability to reorder operations; however,
1307  * if barriers were instrumented with explicit calls (without LTO), the compiler
1308  * couldn't optimize much anyway. The result of a hypothetical architecture
1309  * using __atomic_signal_fence() in normal code would be KCSAN false negatives.
1310  */
1311 void __tsan_atomic_signal_fence(int memorder);
__tsan_atomic_signal_fence(int memorder)1312 noinline void __tsan_atomic_signal_fence(int memorder)
1313 {
1314 	switch (memorder) {
1315 	case __KCSAN_BARRIER_TO_SIGNAL_FENCE_mb:
1316 		__kcsan_mb();
1317 		break;
1318 	case __KCSAN_BARRIER_TO_SIGNAL_FENCE_wmb:
1319 		__kcsan_wmb();
1320 		break;
1321 	case __KCSAN_BARRIER_TO_SIGNAL_FENCE_rmb:
1322 		__kcsan_rmb();
1323 		break;
1324 	case __KCSAN_BARRIER_TO_SIGNAL_FENCE_release:
1325 		__kcsan_release();
1326 		break;
1327 	default:
1328 		break;
1329 	}
1330 }
1331 EXPORT_SYMBOL(__tsan_atomic_signal_fence);
1332 
1333 #ifdef __HAVE_ARCH_MEMSET
1334 void *__tsan_memset(void *s, int c, size_t count);
__tsan_memset(void * s,int c,size_t count)1335 noinline void *__tsan_memset(void *s, int c, size_t count)
1336 {
1337 	/*
1338 	 * Instead of not setting up watchpoints where accessed size is greater
1339 	 * than MAX_ENCODABLE_SIZE, truncate checked size to MAX_ENCODABLE_SIZE.
1340 	 */
1341 	size_t check_len = min_t(size_t, count, MAX_ENCODABLE_SIZE);
1342 
1343 	check_access(s, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
1344 	return memset(s, c, count);
1345 }
1346 #else
1347 void *__tsan_memset(void *s, int c, size_t count) __alias(memset);
1348 #endif
1349 EXPORT_SYMBOL(__tsan_memset);
1350 
1351 #ifdef __HAVE_ARCH_MEMMOVE
1352 void *__tsan_memmove(void *dst, const void *src, size_t len);
__tsan_memmove(void * dst,const void * src,size_t len)1353 noinline void *__tsan_memmove(void *dst, const void *src, size_t len)
1354 {
1355 	size_t check_len = min_t(size_t, len, MAX_ENCODABLE_SIZE);
1356 
1357 	check_access(dst, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
1358 	check_access(src, check_len, 0, _RET_IP_);
1359 	return memmove(dst, src, len);
1360 }
1361 #else
1362 void *__tsan_memmove(void *dst, const void *src, size_t len) __alias(memmove);
1363 #endif
1364 EXPORT_SYMBOL(__tsan_memmove);
1365 
1366 #ifdef __HAVE_ARCH_MEMCPY
1367 void *__tsan_memcpy(void *dst, const void *src, size_t len);
__tsan_memcpy(void * dst,const void * src,size_t len)1368 noinline void *__tsan_memcpy(void *dst, const void *src, size_t len)
1369 {
1370 	size_t check_len = min_t(size_t, len, MAX_ENCODABLE_SIZE);
1371 
1372 	check_access(dst, check_len, KCSAN_ACCESS_WRITE, _RET_IP_);
1373 	check_access(src, check_len, 0, _RET_IP_);
1374 	return memcpy(dst, src, len);
1375 }
1376 #else
1377 void *__tsan_memcpy(void *dst, const void *src, size_t len) __alias(memcpy);
1378 #endif
1379 EXPORT_SYMBOL(__tsan_memcpy);
1380