xref: /openbmc/linux/kernel/kcov.c (revision 0e6774ec)
1 // SPDX-License-Identifier: GPL-2.0
2 #define pr_fmt(fmt) "kcov: " fmt
3 
4 #define DISABLE_BRANCH_PROFILING
5 #include <linux/atomic.h>
6 #include <linux/compiler.h>
7 #include <linux/errno.h>
8 #include <linux/export.h>
9 #include <linux/types.h>
10 #include <linux/file.h>
11 #include <linux/fs.h>
12 #include <linux/hashtable.h>
13 #include <linux/init.h>
14 #include <linux/kmsan-checks.h>
15 #include <linux/mm.h>
16 #include <linux/preempt.h>
17 #include <linux/printk.h>
18 #include <linux/sched.h>
19 #include <linux/slab.h>
20 #include <linux/spinlock.h>
21 #include <linux/vmalloc.h>
22 #include <linux/debugfs.h>
23 #include <linux/uaccess.h>
24 #include <linux/kcov.h>
25 #include <linux/refcount.h>
26 #include <linux/log2.h>
27 #include <asm/setup.h>
28 
29 #define kcov_debug(fmt, ...) pr_debug("%s: " fmt, __func__, ##__VA_ARGS__)
30 
31 /* Number of 64-bit words written per one comparison: */
32 #define KCOV_WORDS_PER_CMP 4
33 
34 /*
35  * kcov descriptor (one per opened debugfs file).
36  * State transitions of the descriptor:
37  *  - initial state after open()
38  *  - then there must be a single ioctl(KCOV_INIT_TRACE) call
39  *  - then, mmap() call (several calls are allowed but not useful)
40  *  - then, ioctl(KCOV_ENABLE, arg), where arg is
41  *	KCOV_TRACE_PC - to trace only the PCs
42  *	or
43  *	KCOV_TRACE_CMP - to trace only the comparison operands
44  *  - then, ioctl(KCOV_DISABLE) to disable the task.
45  * Enabling/disabling ioctls can be repeated (only one task a time allowed).
46  */
47 struct kcov {
48 	/*
49 	 * Reference counter. We keep one for:
50 	 *  - opened file descriptor
51 	 *  - task with enabled coverage (we can't unwire it from another task)
52 	 *  - each code section for remote coverage collection
53 	 */
54 	refcount_t		refcount;
55 	/* The lock protects mode, size, area and t. */
56 	spinlock_t		lock;
57 	enum kcov_mode		mode;
58 	/* Size of arena (in long's). */
59 	unsigned int		size;
60 	/* Coverage buffer shared with user space. */
61 	void			*area;
62 	/* Task for which we collect coverage, or NULL. */
63 	struct task_struct	*t;
64 	/* Collecting coverage from remote (background) threads. */
65 	bool			remote;
66 	/* Size of remote area (in long's). */
67 	unsigned int		remote_size;
68 	/*
69 	 * Sequence is incremented each time kcov is reenabled, used by
70 	 * kcov_remote_stop(), see the comment there.
71 	 */
72 	int			sequence;
73 };
74 
75 struct kcov_remote_area {
76 	struct list_head	list;
77 	unsigned int		size;
78 };
79 
80 struct kcov_remote {
81 	u64			handle;
82 	struct kcov		*kcov;
83 	struct hlist_node	hnode;
84 };
85 
86 static DEFINE_SPINLOCK(kcov_remote_lock);
87 static DEFINE_HASHTABLE(kcov_remote_map, 4);
88 static struct list_head kcov_remote_areas = LIST_HEAD_INIT(kcov_remote_areas);
89 
90 struct kcov_percpu_data {
91 	void			*irq_area;
92 	local_lock_t		lock;
93 
94 	unsigned int		saved_mode;
95 	unsigned int		saved_size;
96 	void			*saved_area;
97 	struct kcov		*saved_kcov;
98 	int			saved_sequence;
99 };
100 
101 static DEFINE_PER_CPU(struct kcov_percpu_data, kcov_percpu_data) = {
102 	.lock = INIT_LOCAL_LOCK(lock),
103 };
104 
105 /* Must be called with kcov_remote_lock locked. */
106 static struct kcov_remote *kcov_remote_find(u64 handle)
107 {
108 	struct kcov_remote *remote;
109 
110 	hash_for_each_possible(kcov_remote_map, remote, hnode, handle) {
111 		if (remote->handle == handle)
112 			return remote;
113 	}
114 	return NULL;
115 }
116 
117 /* Must be called with kcov_remote_lock locked. */
118 static struct kcov_remote *kcov_remote_add(struct kcov *kcov, u64 handle)
119 {
120 	struct kcov_remote *remote;
121 
122 	if (kcov_remote_find(handle))
123 		return ERR_PTR(-EEXIST);
124 	remote = kmalloc(sizeof(*remote), GFP_ATOMIC);
125 	if (!remote)
126 		return ERR_PTR(-ENOMEM);
127 	remote->handle = handle;
128 	remote->kcov = kcov;
129 	hash_add(kcov_remote_map, &remote->hnode, handle);
130 	return remote;
131 }
132 
133 /* Must be called with kcov_remote_lock locked. */
134 static struct kcov_remote_area *kcov_remote_area_get(unsigned int size)
135 {
136 	struct kcov_remote_area *area;
137 	struct list_head *pos;
138 
139 	list_for_each(pos, &kcov_remote_areas) {
140 		area = list_entry(pos, struct kcov_remote_area, list);
141 		if (area->size == size) {
142 			list_del(&area->list);
143 			return area;
144 		}
145 	}
146 	return NULL;
147 }
148 
149 /* Must be called with kcov_remote_lock locked. */
150 static void kcov_remote_area_put(struct kcov_remote_area *area,
151 					unsigned int size)
152 {
153 	INIT_LIST_HEAD(&area->list);
154 	area->size = size;
155 	list_add(&area->list, &kcov_remote_areas);
156 	/*
157 	 * KMSAN doesn't instrument this file, so it may not know area->list
158 	 * is initialized. Unpoison it explicitly to avoid reports in
159 	 * kcov_remote_area_get().
160 	 */
161 	kmsan_unpoison_memory(&area->list, sizeof(area->list));
162 }
163 
164 /*
165  * Unlike in_serving_softirq(), this function returns false when called during
166  * a hardirq or an NMI that happened in the softirq context.
167  */
168 static inline bool in_softirq_really(void)
169 {
170 	return in_serving_softirq() && !in_hardirq() && !in_nmi();
171 }
172 
173 static notrace bool check_kcov_mode(enum kcov_mode needed_mode, struct task_struct *t)
174 {
175 	unsigned int mode;
176 
177 	/*
178 	 * We are interested in code coverage as a function of a syscall inputs,
179 	 * so we ignore code executed in interrupts, unless we are in a remote
180 	 * coverage collection section in a softirq.
181 	 */
182 	if (!in_task() && !(in_softirq_really() && t->kcov_softirq))
183 		return false;
184 	mode = READ_ONCE(t->kcov_mode);
185 	/*
186 	 * There is some code that runs in interrupts but for which
187 	 * in_interrupt() returns false (e.g. preempt_schedule_irq()).
188 	 * READ_ONCE()/barrier() effectively provides load-acquire wrt
189 	 * interrupts, there are paired barrier()/WRITE_ONCE() in
190 	 * kcov_start().
191 	 */
192 	barrier();
193 	return mode == needed_mode;
194 }
195 
196 static notrace unsigned long canonicalize_ip(unsigned long ip)
197 {
198 #ifdef CONFIG_RANDOMIZE_BASE
199 	ip -= kaslr_offset();
200 #endif
201 	return ip;
202 }
203 
204 /*
205  * Entry point from instrumented code.
206  * This is called once per basic-block/edge.
207  */
208 void notrace __sanitizer_cov_trace_pc(void)
209 {
210 	struct task_struct *t;
211 	unsigned long *area;
212 	unsigned long ip = canonicalize_ip(_RET_IP_);
213 	unsigned long pos;
214 
215 	t = current;
216 	if (!check_kcov_mode(KCOV_MODE_TRACE_PC, t))
217 		return;
218 
219 	area = t->kcov_area;
220 	/* The first 64-bit word is the number of subsequent PCs. */
221 	pos = READ_ONCE(area[0]) + 1;
222 	if (likely(pos < t->kcov_size)) {
223 		/* Previously we write pc before updating pos. However, some
224 		 * early interrupt code could bypass check_kcov_mode() check
225 		 * and invoke __sanitizer_cov_trace_pc(). If such interrupt is
226 		 * raised between writing pc and updating pos, the pc could be
227 		 * overitten by the recursive __sanitizer_cov_trace_pc().
228 		 * Update pos before writing pc to avoid such interleaving.
229 		 */
230 		WRITE_ONCE(area[0], pos);
231 		barrier();
232 		area[pos] = ip;
233 	}
234 }
235 EXPORT_SYMBOL(__sanitizer_cov_trace_pc);
236 
237 #ifdef CONFIG_KCOV_ENABLE_COMPARISONS
238 static void notrace write_comp_data(u64 type, u64 arg1, u64 arg2, u64 ip)
239 {
240 	struct task_struct *t;
241 	u64 *area;
242 	u64 count, start_index, end_pos, max_pos;
243 
244 	t = current;
245 	if (!check_kcov_mode(KCOV_MODE_TRACE_CMP, t))
246 		return;
247 
248 	ip = canonicalize_ip(ip);
249 
250 	/*
251 	 * We write all comparison arguments and types as u64.
252 	 * The buffer was allocated for t->kcov_size unsigned longs.
253 	 */
254 	area = (u64 *)t->kcov_area;
255 	max_pos = t->kcov_size * sizeof(unsigned long);
256 
257 	count = READ_ONCE(area[0]);
258 
259 	/* Every record is KCOV_WORDS_PER_CMP 64-bit words. */
260 	start_index = 1 + count * KCOV_WORDS_PER_CMP;
261 	end_pos = (start_index + KCOV_WORDS_PER_CMP) * sizeof(u64);
262 	if (likely(end_pos <= max_pos)) {
263 		/* See comment in __sanitizer_cov_trace_pc(). */
264 		WRITE_ONCE(area[0], count + 1);
265 		barrier();
266 		area[start_index] = type;
267 		area[start_index + 1] = arg1;
268 		area[start_index + 2] = arg2;
269 		area[start_index + 3] = ip;
270 	}
271 }
272 
273 void notrace __sanitizer_cov_trace_cmp1(u8 arg1, u8 arg2)
274 {
275 	write_comp_data(KCOV_CMP_SIZE(0), arg1, arg2, _RET_IP_);
276 }
277 EXPORT_SYMBOL(__sanitizer_cov_trace_cmp1);
278 
279 void notrace __sanitizer_cov_trace_cmp2(u16 arg1, u16 arg2)
280 {
281 	write_comp_data(KCOV_CMP_SIZE(1), arg1, arg2, _RET_IP_);
282 }
283 EXPORT_SYMBOL(__sanitizer_cov_trace_cmp2);
284 
285 void notrace __sanitizer_cov_trace_cmp4(u32 arg1, u32 arg2)
286 {
287 	write_comp_data(KCOV_CMP_SIZE(2), arg1, arg2, _RET_IP_);
288 }
289 EXPORT_SYMBOL(__sanitizer_cov_trace_cmp4);
290 
291 void notrace __sanitizer_cov_trace_cmp8(kcov_u64 arg1, kcov_u64 arg2)
292 {
293 	write_comp_data(KCOV_CMP_SIZE(3), arg1, arg2, _RET_IP_);
294 }
295 EXPORT_SYMBOL(__sanitizer_cov_trace_cmp8);
296 
297 void notrace __sanitizer_cov_trace_const_cmp1(u8 arg1, u8 arg2)
298 {
299 	write_comp_data(KCOV_CMP_SIZE(0) | KCOV_CMP_CONST, arg1, arg2,
300 			_RET_IP_);
301 }
302 EXPORT_SYMBOL(__sanitizer_cov_trace_const_cmp1);
303 
304 void notrace __sanitizer_cov_trace_const_cmp2(u16 arg1, u16 arg2)
305 {
306 	write_comp_data(KCOV_CMP_SIZE(1) | KCOV_CMP_CONST, arg1, arg2,
307 			_RET_IP_);
308 }
309 EXPORT_SYMBOL(__sanitizer_cov_trace_const_cmp2);
310 
311 void notrace __sanitizer_cov_trace_const_cmp4(u32 arg1, u32 arg2)
312 {
313 	write_comp_data(KCOV_CMP_SIZE(2) | KCOV_CMP_CONST, arg1, arg2,
314 			_RET_IP_);
315 }
316 EXPORT_SYMBOL(__sanitizer_cov_trace_const_cmp4);
317 
318 void notrace __sanitizer_cov_trace_const_cmp8(kcov_u64 arg1, kcov_u64 arg2)
319 {
320 	write_comp_data(KCOV_CMP_SIZE(3) | KCOV_CMP_CONST, arg1, arg2,
321 			_RET_IP_);
322 }
323 EXPORT_SYMBOL(__sanitizer_cov_trace_const_cmp8);
324 
325 void notrace __sanitizer_cov_trace_switch(kcov_u64 val, void *arg)
326 {
327 	u64 i;
328 	u64 *cases = arg;
329 	u64 count = cases[0];
330 	u64 size = cases[1];
331 	u64 type = KCOV_CMP_CONST;
332 
333 	switch (size) {
334 	case 8:
335 		type |= KCOV_CMP_SIZE(0);
336 		break;
337 	case 16:
338 		type |= KCOV_CMP_SIZE(1);
339 		break;
340 	case 32:
341 		type |= KCOV_CMP_SIZE(2);
342 		break;
343 	case 64:
344 		type |= KCOV_CMP_SIZE(3);
345 		break;
346 	default:
347 		return;
348 	}
349 	for (i = 0; i < count; i++)
350 		write_comp_data(type, cases[i + 2], val, _RET_IP_);
351 }
352 EXPORT_SYMBOL(__sanitizer_cov_trace_switch);
353 #endif /* ifdef CONFIG_KCOV_ENABLE_COMPARISONS */
354 
355 static void kcov_start(struct task_struct *t, struct kcov *kcov,
356 			unsigned int size, void *area, enum kcov_mode mode,
357 			int sequence)
358 {
359 	kcov_debug("t = %px, size = %u, area = %px\n", t, size, area);
360 	t->kcov = kcov;
361 	/* Cache in task struct for performance. */
362 	t->kcov_size = size;
363 	t->kcov_area = area;
364 	t->kcov_sequence = sequence;
365 	/* See comment in check_kcov_mode(). */
366 	barrier();
367 	WRITE_ONCE(t->kcov_mode, mode);
368 }
369 
370 static void kcov_stop(struct task_struct *t)
371 {
372 	WRITE_ONCE(t->kcov_mode, KCOV_MODE_DISABLED);
373 	barrier();
374 	t->kcov = NULL;
375 	t->kcov_size = 0;
376 	t->kcov_area = NULL;
377 }
378 
379 static void kcov_task_reset(struct task_struct *t)
380 {
381 	kcov_stop(t);
382 	t->kcov_sequence = 0;
383 	t->kcov_handle = 0;
384 }
385 
386 void kcov_task_init(struct task_struct *t)
387 {
388 	kcov_task_reset(t);
389 	t->kcov_handle = current->kcov_handle;
390 }
391 
392 static void kcov_reset(struct kcov *kcov)
393 {
394 	kcov->t = NULL;
395 	kcov->mode = KCOV_MODE_INIT;
396 	kcov->remote = false;
397 	kcov->remote_size = 0;
398 	kcov->sequence++;
399 }
400 
401 static void kcov_remote_reset(struct kcov *kcov)
402 {
403 	int bkt;
404 	struct kcov_remote *remote;
405 	struct hlist_node *tmp;
406 	unsigned long flags;
407 
408 	spin_lock_irqsave(&kcov_remote_lock, flags);
409 	hash_for_each_safe(kcov_remote_map, bkt, tmp, remote, hnode) {
410 		if (remote->kcov != kcov)
411 			continue;
412 		hash_del(&remote->hnode);
413 		kfree(remote);
414 	}
415 	/* Do reset before unlock to prevent races with kcov_remote_start(). */
416 	kcov_reset(kcov);
417 	spin_unlock_irqrestore(&kcov_remote_lock, flags);
418 }
419 
420 static void kcov_disable(struct task_struct *t, struct kcov *kcov)
421 {
422 	kcov_task_reset(t);
423 	if (kcov->remote)
424 		kcov_remote_reset(kcov);
425 	else
426 		kcov_reset(kcov);
427 }
428 
429 static void kcov_get(struct kcov *kcov)
430 {
431 	refcount_inc(&kcov->refcount);
432 }
433 
434 static void kcov_put(struct kcov *kcov)
435 {
436 	if (refcount_dec_and_test(&kcov->refcount)) {
437 		kcov_remote_reset(kcov);
438 		vfree(kcov->area);
439 		kfree(kcov);
440 	}
441 }
442 
443 void kcov_task_exit(struct task_struct *t)
444 {
445 	struct kcov *kcov;
446 	unsigned long flags;
447 
448 	kcov = t->kcov;
449 	if (kcov == NULL)
450 		return;
451 
452 	spin_lock_irqsave(&kcov->lock, flags);
453 	kcov_debug("t = %px, kcov->t = %px\n", t, kcov->t);
454 	/*
455 	 * For KCOV_ENABLE devices we want to make sure that t->kcov->t == t,
456 	 * which comes down to:
457 	 *        WARN_ON(!kcov->remote && kcov->t != t);
458 	 *
459 	 * For KCOV_REMOTE_ENABLE devices, the exiting task is either:
460 	 *
461 	 * 1. A remote task between kcov_remote_start() and kcov_remote_stop().
462 	 *    In this case we should print a warning right away, since a task
463 	 *    shouldn't be exiting when it's in a kcov coverage collection
464 	 *    section. Here t points to the task that is collecting remote
465 	 *    coverage, and t->kcov->t points to the thread that created the
466 	 *    kcov device. Which means that to detect this case we need to
467 	 *    check that t != t->kcov->t, and this gives us the following:
468 	 *        WARN_ON(kcov->remote && kcov->t != t);
469 	 *
470 	 * 2. The task that created kcov exiting without calling KCOV_DISABLE,
471 	 *    and then again we make sure that t->kcov->t == t:
472 	 *        WARN_ON(kcov->remote && kcov->t != t);
473 	 *
474 	 * By combining all three checks into one we get:
475 	 */
476 	if (WARN_ON(kcov->t != t)) {
477 		spin_unlock_irqrestore(&kcov->lock, flags);
478 		return;
479 	}
480 	/* Just to not leave dangling references behind. */
481 	kcov_disable(t, kcov);
482 	spin_unlock_irqrestore(&kcov->lock, flags);
483 	kcov_put(kcov);
484 }
485 
486 static int kcov_mmap(struct file *filep, struct vm_area_struct *vma)
487 {
488 	int res = 0;
489 	struct kcov *kcov = vma->vm_file->private_data;
490 	unsigned long size, off;
491 	struct page *page;
492 	unsigned long flags;
493 
494 	spin_lock_irqsave(&kcov->lock, flags);
495 	size = kcov->size * sizeof(unsigned long);
496 	if (kcov->area == NULL || vma->vm_pgoff != 0 ||
497 	    vma->vm_end - vma->vm_start != size) {
498 		res = -EINVAL;
499 		goto exit;
500 	}
501 	spin_unlock_irqrestore(&kcov->lock, flags);
502 	vm_flags_set(vma, VM_DONTEXPAND);
503 	for (off = 0; off < size; off += PAGE_SIZE) {
504 		page = vmalloc_to_page(kcov->area + off);
505 		res = vm_insert_page(vma, vma->vm_start + off, page);
506 		if (res) {
507 			pr_warn_once("kcov: vm_insert_page() failed\n");
508 			return res;
509 		}
510 	}
511 	return 0;
512 exit:
513 	spin_unlock_irqrestore(&kcov->lock, flags);
514 	return res;
515 }
516 
517 static int kcov_open(struct inode *inode, struct file *filep)
518 {
519 	struct kcov *kcov;
520 
521 	kcov = kzalloc(sizeof(*kcov), GFP_KERNEL);
522 	if (!kcov)
523 		return -ENOMEM;
524 	kcov->mode = KCOV_MODE_DISABLED;
525 	kcov->sequence = 1;
526 	refcount_set(&kcov->refcount, 1);
527 	spin_lock_init(&kcov->lock);
528 	filep->private_data = kcov;
529 	return nonseekable_open(inode, filep);
530 }
531 
532 static int kcov_close(struct inode *inode, struct file *filep)
533 {
534 	kcov_put(filep->private_data);
535 	return 0;
536 }
537 
538 static int kcov_get_mode(unsigned long arg)
539 {
540 	if (arg == KCOV_TRACE_PC)
541 		return KCOV_MODE_TRACE_PC;
542 	else if (arg == KCOV_TRACE_CMP)
543 #ifdef CONFIG_KCOV_ENABLE_COMPARISONS
544 		return KCOV_MODE_TRACE_CMP;
545 #else
546 		return -ENOTSUPP;
547 #endif
548 	else
549 		return -EINVAL;
550 }
551 
552 /*
553  * Fault in a lazily-faulted vmalloc area before it can be used by
554  * __santizer_cov_trace_pc(), to avoid recursion issues if any code on the
555  * vmalloc fault handling path is instrumented.
556  */
557 static void kcov_fault_in_area(struct kcov *kcov)
558 {
559 	unsigned long stride = PAGE_SIZE / sizeof(unsigned long);
560 	unsigned long *area = kcov->area;
561 	unsigned long offset;
562 
563 	for (offset = 0; offset < kcov->size; offset += stride)
564 		READ_ONCE(area[offset]);
565 }
566 
567 static inline bool kcov_check_handle(u64 handle, bool common_valid,
568 				bool uncommon_valid, bool zero_valid)
569 {
570 	if (handle & ~(KCOV_SUBSYSTEM_MASK | KCOV_INSTANCE_MASK))
571 		return false;
572 	switch (handle & KCOV_SUBSYSTEM_MASK) {
573 	case KCOV_SUBSYSTEM_COMMON:
574 		return (handle & KCOV_INSTANCE_MASK) ?
575 			common_valid : zero_valid;
576 	case KCOV_SUBSYSTEM_USB:
577 		return uncommon_valid;
578 	default:
579 		return false;
580 	}
581 	return false;
582 }
583 
584 static int kcov_ioctl_locked(struct kcov *kcov, unsigned int cmd,
585 			     unsigned long arg)
586 {
587 	struct task_struct *t;
588 	unsigned long flags, unused;
589 	int mode, i;
590 	struct kcov_remote_arg *remote_arg;
591 	struct kcov_remote *remote;
592 
593 	switch (cmd) {
594 	case KCOV_ENABLE:
595 		/*
596 		 * Enable coverage for the current task.
597 		 * At this point user must have been enabled trace mode,
598 		 * and mmapped the file. Coverage collection is disabled only
599 		 * at task exit or voluntary by KCOV_DISABLE. After that it can
600 		 * be enabled for another task.
601 		 */
602 		if (kcov->mode != KCOV_MODE_INIT || !kcov->area)
603 			return -EINVAL;
604 		t = current;
605 		if (kcov->t != NULL || t->kcov != NULL)
606 			return -EBUSY;
607 		mode = kcov_get_mode(arg);
608 		if (mode < 0)
609 			return mode;
610 		kcov_fault_in_area(kcov);
611 		kcov->mode = mode;
612 		kcov_start(t, kcov, kcov->size, kcov->area, kcov->mode,
613 				kcov->sequence);
614 		kcov->t = t;
615 		/* Put either in kcov_task_exit() or in KCOV_DISABLE. */
616 		kcov_get(kcov);
617 		return 0;
618 	case KCOV_DISABLE:
619 		/* Disable coverage for the current task. */
620 		unused = arg;
621 		if (unused != 0 || current->kcov != kcov)
622 			return -EINVAL;
623 		t = current;
624 		if (WARN_ON(kcov->t != t))
625 			return -EINVAL;
626 		kcov_disable(t, kcov);
627 		kcov_put(kcov);
628 		return 0;
629 	case KCOV_REMOTE_ENABLE:
630 		if (kcov->mode != KCOV_MODE_INIT || !kcov->area)
631 			return -EINVAL;
632 		t = current;
633 		if (kcov->t != NULL || t->kcov != NULL)
634 			return -EBUSY;
635 		remote_arg = (struct kcov_remote_arg *)arg;
636 		mode = kcov_get_mode(remote_arg->trace_mode);
637 		if (mode < 0)
638 			return mode;
639 		if (remote_arg->area_size > LONG_MAX / sizeof(unsigned long))
640 			return -EINVAL;
641 		kcov->mode = mode;
642 		t->kcov = kcov;
643 	        t->kcov_mode = KCOV_MODE_REMOTE;
644 		kcov->t = t;
645 		kcov->remote = true;
646 		kcov->remote_size = remote_arg->area_size;
647 		spin_lock_irqsave(&kcov_remote_lock, flags);
648 		for (i = 0; i < remote_arg->num_handles; i++) {
649 			if (!kcov_check_handle(remote_arg->handles[i],
650 						false, true, false)) {
651 				spin_unlock_irqrestore(&kcov_remote_lock,
652 							flags);
653 				kcov_disable(t, kcov);
654 				return -EINVAL;
655 			}
656 			remote = kcov_remote_add(kcov, remote_arg->handles[i]);
657 			if (IS_ERR(remote)) {
658 				spin_unlock_irqrestore(&kcov_remote_lock,
659 							flags);
660 				kcov_disable(t, kcov);
661 				return PTR_ERR(remote);
662 			}
663 		}
664 		if (remote_arg->common_handle) {
665 			if (!kcov_check_handle(remote_arg->common_handle,
666 						true, false, false)) {
667 				spin_unlock_irqrestore(&kcov_remote_lock,
668 							flags);
669 				kcov_disable(t, kcov);
670 				return -EINVAL;
671 			}
672 			remote = kcov_remote_add(kcov,
673 					remote_arg->common_handle);
674 			if (IS_ERR(remote)) {
675 				spin_unlock_irqrestore(&kcov_remote_lock,
676 							flags);
677 				kcov_disable(t, kcov);
678 				return PTR_ERR(remote);
679 			}
680 			t->kcov_handle = remote_arg->common_handle;
681 		}
682 		spin_unlock_irqrestore(&kcov_remote_lock, flags);
683 		/* Put either in kcov_task_exit() or in KCOV_DISABLE. */
684 		kcov_get(kcov);
685 		return 0;
686 	default:
687 		return -ENOTTY;
688 	}
689 }
690 
691 static long kcov_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
692 {
693 	struct kcov *kcov;
694 	int res;
695 	struct kcov_remote_arg *remote_arg = NULL;
696 	unsigned int remote_num_handles;
697 	unsigned long remote_arg_size;
698 	unsigned long size, flags;
699 	void *area;
700 
701 	kcov = filep->private_data;
702 	switch (cmd) {
703 	case KCOV_INIT_TRACE:
704 		/*
705 		 * Enable kcov in trace mode and setup buffer size.
706 		 * Must happen before anything else.
707 		 *
708 		 * First check the size argument - it must be at least 2
709 		 * to hold the current position and one PC.
710 		 */
711 		size = arg;
712 		if (size < 2 || size > INT_MAX / sizeof(unsigned long))
713 			return -EINVAL;
714 		area = vmalloc_user(size * sizeof(unsigned long));
715 		if (area == NULL)
716 			return -ENOMEM;
717 		spin_lock_irqsave(&kcov->lock, flags);
718 		if (kcov->mode != KCOV_MODE_DISABLED) {
719 			spin_unlock_irqrestore(&kcov->lock, flags);
720 			vfree(area);
721 			return -EBUSY;
722 		}
723 		kcov->area = area;
724 		kcov->size = size;
725 		kcov->mode = KCOV_MODE_INIT;
726 		spin_unlock_irqrestore(&kcov->lock, flags);
727 		return 0;
728 	case KCOV_REMOTE_ENABLE:
729 		if (get_user(remote_num_handles, (unsigned __user *)(arg +
730 				offsetof(struct kcov_remote_arg, num_handles))))
731 			return -EFAULT;
732 		if (remote_num_handles > KCOV_REMOTE_MAX_HANDLES)
733 			return -EINVAL;
734 		remote_arg_size = struct_size(remote_arg, handles,
735 					remote_num_handles);
736 		remote_arg = memdup_user((void __user *)arg, remote_arg_size);
737 		if (IS_ERR(remote_arg))
738 			return PTR_ERR(remote_arg);
739 		if (remote_arg->num_handles != remote_num_handles) {
740 			kfree(remote_arg);
741 			return -EINVAL;
742 		}
743 		arg = (unsigned long)remote_arg;
744 		fallthrough;
745 	default:
746 		/*
747 		 * All other commands can be normally executed under a spin lock, so we
748 		 * obtain and release it here in order to simplify kcov_ioctl_locked().
749 		 */
750 		spin_lock_irqsave(&kcov->lock, flags);
751 		res = kcov_ioctl_locked(kcov, cmd, arg);
752 		spin_unlock_irqrestore(&kcov->lock, flags);
753 		kfree(remote_arg);
754 		return res;
755 	}
756 }
757 
758 static const struct file_operations kcov_fops = {
759 	.open		= kcov_open,
760 	.unlocked_ioctl	= kcov_ioctl,
761 	.compat_ioctl	= kcov_ioctl,
762 	.mmap		= kcov_mmap,
763 	.release        = kcov_close,
764 };
765 
766 /*
767  * kcov_remote_start() and kcov_remote_stop() can be used to annotate a section
768  * of code in a kernel background thread or in a softirq to allow kcov to be
769  * used to collect coverage from that part of code.
770  *
771  * The handle argument of kcov_remote_start() identifies a code section that is
772  * used for coverage collection. A userspace process passes this handle to
773  * KCOV_REMOTE_ENABLE ioctl to make the used kcov device start collecting
774  * coverage for the code section identified by this handle.
775  *
776  * The usage of these annotations in the kernel code is different depending on
777  * the type of the kernel thread whose code is being annotated.
778  *
779  * For global kernel threads that are spawned in a limited number of instances
780  * (e.g. one USB hub_event() worker thread is spawned per USB HCD) and for
781  * softirqs, each instance must be assigned a unique 4-byte instance id. The
782  * instance id is then combined with a 1-byte subsystem id to get a handle via
783  * kcov_remote_handle(subsystem_id, instance_id).
784  *
785  * For local kernel threads that are spawned from system calls handler when a
786  * user interacts with some kernel interface (e.g. vhost workers), a handle is
787  * passed from a userspace process as the common_handle field of the
788  * kcov_remote_arg struct (note, that the user must generate a handle by using
789  * kcov_remote_handle() with KCOV_SUBSYSTEM_COMMON as the subsystem id and an
790  * arbitrary 4-byte non-zero number as the instance id). This common handle
791  * then gets saved into the task_struct of the process that issued the
792  * KCOV_REMOTE_ENABLE ioctl. When this process issues system calls that spawn
793  * kernel threads, the common handle must be retrieved via kcov_common_handle()
794  * and passed to the spawned threads via custom annotations. Those kernel
795  * threads must in turn be annotated with kcov_remote_start(common_handle) and
796  * kcov_remote_stop(). All of the threads that are spawned by the same process
797  * obtain the same handle, hence the name "common".
798  *
799  * See Documentation/dev-tools/kcov.rst for more details.
800  *
801  * Internally, kcov_remote_start() looks up the kcov device associated with the
802  * provided handle, allocates an area for coverage collection, and saves the
803  * pointers to kcov and area into the current task_struct to allow coverage to
804  * be collected via __sanitizer_cov_trace_pc().
805  * In turns kcov_remote_stop() clears those pointers from task_struct to stop
806  * collecting coverage and copies all collected coverage into the kcov area.
807  */
808 
809 static inline bool kcov_mode_enabled(unsigned int mode)
810 {
811 	return (mode & ~KCOV_IN_CTXSW) != KCOV_MODE_DISABLED;
812 }
813 
814 static void kcov_remote_softirq_start(struct task_struct *t)
815 {
816 	struct kcov_percpu_data *data = this_cpu_ptr(&kcov_percpu_data);
817 	unsigned int mode;
818 
819 	mode = READ_ONCE(t->kcov_mode);
820 	barrier();
821 	if (kcov_mode_enabled(mode)) {
822 		data->saved_mode = mode;
823 		data->saved_size = t->kcov_size;
824 		data->saved_area = t->kcov_area;
825 		data->saved_sequence = t->kcov_sequence;
826 		data->saved_kcov = t->kcov;
827 		kcov_stop(t);
828 	}
829 }
830 
831 static void kcov_remote_softirq_stop(struct task_struct *t)
832 {
833 	struct kcov_percpu_data *data = this_cpu_ptr(&kcov_percpu_data);
834 
835 	if (data->saved_kcov) {
836 		kcov_start(t, data->saved_kcov, data->saved_size,
837 				data->saved_area, data->saved_mode,
838 				data->saved_sequence);
839 		data->saved_mode = 0;
840 		data->saved_size = 0;
841 		data->saved_area = NULL;
842 		data->saved_sequence = 0;
843 		data->saved_kcov = NULL;
844 	}
845 }
846 
847 void kcov_remote_start(u64 handle)
848 {
849 	struct task_struct *t = current;
850 	struct kcov_remote *remote;
851 	struct kcov *kcov;
852 	unsigned int mode;
853 	void *area;
854 	unsigned int size;
855 	int sequence;
856 	unsigned long flags;
857 
858 	if (WARN_ON(!kcov_check_handle(handle, true, true, true)))
859 		return;
860 	if (!in_task() && !in_softirq_really())
861 		return;
862 
863 	local_lock_irqsave(&kcov_percpu_data.lock, flags);
864 
865 	/*
866 	 * Check that kcov_remote_start() is not called twice in background
867 	 * threads nor called by user tasks (with enabled kcov).
868 	 */
869 	mode = READ_ONCE(t->kcov_mode);
870 	if (WARN_ON(in_task() && kcov_mode_enabled(mode))) {
871 		local_unlock_irqrestore(&kcov_percpu_data.lock, flags);
872 		return;
873 	}
874 	/*
875 	 * Check that kcov_remote_start() is not called twice in softirqs.
876 	 * Note, that kcov_remote_start() can be called from a softirq that
877 	 * happened while collecting coverage from a background thread.
878 	 */
879 	if (WARN_ON(in_serving_softirq() && t->kcov_softirq)) {
880 		local_unlock_irqrestore(&kcov_percpu_data.lock, flags);
881 		return;
882 	}
883 
884 	spin_lock(&kcov_remote_lock);
885 	remote = kcov_remote_find(handle);
886 	if (!remote) {
887 		spin_unlock(&kcov_remote_lock);
888 		local_unlock_irqrestore(&kcov_percpu_data.lock, flags);
889 		return;
890 	}
891 	kcov_debug("handle = %llx, context: %s\n", handle,
892 			in_task() ? "task" : "softirq");
893 	kcov = remote->kcov;
894 	/* Put in kcov_remote_stop(). */
895 	kcov_get(kcov);
896 	/*
897 	 * Read kcov fields before unlock to prevent races with
898 	 * KCOV_DISABLE / kcov_remote_reset().
899 	 */
900 	mode = kcov->mode;
901 	sequence = kcov->sequence;
902 	if (in_task()) {
903 		size = kcov->remote_size;
904 		area = kcov_remote_area_get(size);
905 	} else {
906 		size = CONFIG_KCOV_IRQ_AREA_SIZE;
907 		area = this_cpu_ptr(&kcov_percpu_data)->irq_area;
908 	}
909 	spin_unlock(&kcov_remote_lock);
910 
911 	/* Can only happen when in_task(). */
912 	if (!area) {
913 		local_unlock_irqrestore(&kcov_percpu_data.lock, flags);
914 		area = vmalloc(size * sizeof(unsigned long));
915 		if (!area) {
916 			kcov_put(kcov);
917 			return;
918 		}
919 		local_lock_irqsave(&kcov_percpu_data.lock, flags);
920 	}
921 
922 	/* Reset coverage size. */
923 	*(u64 *)area = 0;
924 
925 	if (in_serving_softirq()) {
926 		kcov_remote_softirq_start(t);
927 		t->kcov_softirq = 1;
928 	}
929 	kcov_start(t, kcov, size, area, mode, sequence);
930 
931 	local_unlock_irqrestore(&kcov_percpu_data.lock, flags);
932 
933 }
934 EXPORT_SYMBOL(kcov_remote_start);
935 
936 static void kcov_move_area(enum kcov_mode mode, void *dst_area,
937 				unsigned int dst_area_size, void *src_area)
938 {
939 	u64 word_size = sizeof(unsigned long);
940 	u64 count_size, entry_size_log;
941 	u64 dst_len, src_len;
942 	void *dst_entries, *src_entries;
943 	u64 dst_occupied, dst_free, bytes_to_move, entries_moved;
944 
945 	kcov_debug("%px %u <= %px %lu\n",
946 		dst_area, dst_area_size, src_area, *(unsigned long *)src_area);
947 
948 	switch (mode) {
949 	case KCOV_MODE_TRACE_PC:
950 		dst_len = READ_ONCE(*(unsigned long *)dst_area);
951 		src_len = *(unsigned long *)src_area;
952 		count_size = sizeof(unsigned long);
953 		entry_size_log = __ilog2_u64(sizeof(unsigned long));
954 		break;
955 	case KCOV_MODE_TRACE_CMP:
956 		dst_len = READ_ONCE(*(u64 *)dst_area);
957 		src_len = *(u64 *)src_area;
958 		count_size = sizeof(u64);
959 		BUILD_BUG_ON(!is_power_of_2(KCOV_WORDS_PER_CMP));
960 		entry_size_log = __ilog2_u64(sizeof(u64) * KCOV_WORDS_PER_CMP);
961 		break;
962 	default:
963 		WARN_ON(1);
964 		return;
965 	}
966 
967 	/* As arm can't divide u64 integers use log of entry size. */
968 	if (dst_len > ((dst_area_size * word_size - count_size) >>
969 				entry_size_log))
970 		return;
971 	dst_occupied = count_size + (dst_len << entry_size_log);
972 	dst_free = dst_area_size * word_size - dst_occupied;
973 	bytes_to_move = min(dst_free, src_len << entry_size_log);
974 	dst_entries = dst_area + dst_occupied;
975 	src_entries = src_area + count_size;
976 	memcpy(dst_entries, src_entries, bytes_to_move);
977 	entries_moved = bytes_to_move >> entry_size_log;
978 
979 	switch (mode) {
980 	case KCOV_MODE_TRACE_PC:
981 		WRITE_ONCE(*(unsigned long *)dst_area, dst_len + entries_moved);
982 		break;
983 	case KCOV_MODE_TRACE_CMP:
984 		WRITE_ONCE(*(u64 *)dst_area, dst_len + entries_moved);
985 		break;
986 	default:
987 		break;
988 	}
989 }
990 
991 /* See the comment before kcov_remote_start() for usage details. */
992 void kcov_remote_stop(void)
993 {
994 	struct task_struct *t = current;
995 	struct kcov *kcov;
996 	unsigned int mode;
997 	void *area;
998 	unsigned int size;
999 	int sequence;
1000 	unsigned long flags;
1001 
1002 	if (!in_task() && !in_softirq_really())
1003 		return;
1004 
1005 	local_lock_irqsave(&kcov_percpu_data.lock, flags);
1006 
1007 	mode = READ_ONCE(t->kcov_mode);
1008 	barrier();
1009 	if (!kcov_mode_enabled(mode)) {
1010 		local_unlock_irqrestore(&kcov_percpu_data.lock, flags);
1011 		return;
1012 	}
1013 	/*
1014 	 * When in softirq, check if the corresponding kcov_remote_start()
1015 	 * actually found the remote handle and started collecting coverage.
1016 	 */
1017 	if (in_serving_softirq() && !t->kcov_softirq) {
1018 		local_unlock_irqrestore(&kcov_percpu_data.lock, flags);
1019 		return;
1020 	}
1021 	/* Make sure that kcov_softirq is only set when in softirq. */
1022 	if (WARN_ON(!in_serving_softirq() && t->kcov_softirq)) {
1023 		local_unlock_irqrestore(&kcov_percpu_data.lock, flags);
1024 		return;
1025 	}
1026 
1027 	kcov = t->kcov;
1028 	area = t->kcov_area;
1029 	size = t->kcov_size;
1030 	sequence = t->kcov_sequence;
1031 
1032 	kcov_stop(t);
1033 	if (in_serving_softirq()) {
1034 		t->kcov_softirq = 0;
1035 		kcov_remote_softirq_stop(t);
1036 	}
1037 
1038 	spin_lock(&kcov->lock);
1039 	/*
1040 	 * KCOV_DISABLE could have been called between kcov_remote_start()
1041 	 * and kcov_remote_stop(), hence the sequence check.
1042 	 */
1043 	if (sequence == kcov->sequence && kcov->remote)
1044 		kcov_move_area(kcov->mode, kcov->area, kcov->size, area);
1045 	spin_unlock(&kcov->lock);
1046 
1047 	if (in_task()) {
1048 		spin_lock(&kcov_remote_lock);
1049 		kcov_remote_area_put(area, size);
1050 		spin_unlock(&kcov_remote_lock);
1051 	}
1052 
1053 	local_unlock_irqrestore(&kcov_percpu_data.lock, flags);
1054 
1055 	/* Get in kcov_remote_start(). */
1056 	kcov_put(kcov);
1057 }
1058 EXPORT_SYMBOL(kcov_remote_stop);
1059 
1060 /* See the comment before kcov_remote_start() for usage details. */
1061 u64 kcov_common_handle(void)
1062 {
1063 	if (!in_task())
1064 		return 0;
1065 	return current->kcov_handle;
1066 }
1067 EXPORT_SYMBOL(kcov_common_handle);
1068 
1069 static int __init kcov_init(void)
1070 {
1071 	int cpu;
1072 
1073 	for_each_possible_cpu(cpu) {
1074 		void *area = vmalloc_node(CONFIG_KCOV_IRQ_AREA_SIZE *
1075 				sizeof(unsigned long), cpu_to_node(cpu));
1076 		if (!area)
1077 			return -ENOMEM;
1078 		per_cpu_ptr(&kcov_percpu_data, cpu)->irq_area = area;
1079 	}
1080 
1081 	/*
1082 	 * The kcov debugfs file won't ever get removed and thus,
1083 	 * there is no need to protect it against removal races. The
1084 	 * use of debugfs_create_file_unsafe() is actually safe here.
1085 	 */
1086 	debugfs_create_file_unsafe("kcov", 0600, NULL, NULL, &kcov_fops);
1087 
1088 	return 0;
1089 }
1090 
1091 device_initcall(kcov_init);
1092