xref: /freebsd/sys/kern/kern_kcov.c (revision 4e99f45480598189d49d45a825533a6c9e12f02c)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (C) 2018 The FreeBSD Foundation. All rights reserved.
5  * Copyright (C) 2018, 2019 Andrew Turner
6  *
7  * This software was developed by Mitchell Horne under sponsorship of
8  * the FreeBSD Foundation.
9  *
10  * This software was developed by SRI International and the University of
11  * Cambridge Computer Laboratory under DARPA/AFRL contract FA8750-10-C-0237
12  * ("CTSRD"), as part of the DARPA CRASH research programme.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  * $FreeBSD$
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/conf.h>
44 #include <sys/eventhandler.h>
45 #include <sys/kcov.h>
46 #include <sys/kernel.h>
47 #include <sys/limits.h>
48 #include <sys/lock.h>
49 #include <sys/malloc.h>
50 #include <sys/mman.h>
51 #include <sys/mutex.h>
52 #include <sys/proc.h>
53 #include <sys/rwlock.h>
54 #include <sys/sysctl.h>
55 
56 #include <vm/vm.h>
57 #include <vm/pmap.h>
58 #include <vm/vm_extern.h>
59 #include <vm/vm_object.h>
60 #include <vm/vm_page.h>
61 #include <vm/vm_pager.h>
62 #include <vm/vm_param.h>
63 
64 MALLOC_DEFINE(M_KCOV_INFO, "kcovinfo", "KCOV info type");
65 
66 #define	KCOV_ELEMENT_SIZE	sizeof(uint64_t)
67 
68 /*
69  * To know what the code can safely perform at any point in time we use a
70  * state machine. In the normal case the state transitions are:
71  *
72  * OPEN -> READY -> RUNNING -> DYING
73  *  |       | ^        |        ^ ^
74  *  |       | +--------+        | |
75  *  |       +-------------------+ |
76  *  +-----------------------------+
77  *
78  * The states are:
79  *  OPEN:   The kcov fd has been opened, but no buffer is available to store
80  *          coverage data.
81  *  READY:  The buffer to store coverage data has been allocated. Userspace
82  *          can set this by using ioctl(fd, KIOSETBUFSIZE, entries);. When
83  *          this has been set the buffer can be written to by the kernel,
84  *          and mmaped by userspace.
85  * RUNNING: The coverage probes are able to store coverage data in the buffer.
86  *          This is entered with ioctl(fd, KIOENABLE, mode);. The READY state
87  *          can be exited by ioctl(fd, KIODISABLE); or exiting the thread to
88  *          return to the READY state to allow tracing to be reused, or by
89  *          closing the kcov fd to enter the DYING state.
90  * DYING:   The fd has been closed. All states can enter into this state when
91  *          userspace closes the kcov fd.
92  *
93  * We need to be careful when moving into and out of the RUNNING state. As
94  * an interrupt may happen while this is happening the ordering of memory
95  * operations is important so struct kcov_info is valid for the tracing
96  * functions.
97  *
98  * When moving into the RUNNING state prior stores to struct kcov_info need
99  * to be observed before the state is set. This allows for interrupts that
100  * may call into one of the coverage functions to fire at any point while
101  * being enabled and see a consistent struct kcov_info.
102  *
103  * When moving out of the RUNNING state any later stores to struct kcov_info
104  * need to be observed after the state is set. As with entering this is to
105  * present a consistent struct kcov_info to interrupts.
106  */
107 typedef enum {
108 	KCOV_STATE_INVALID,
109 	KCOV_STATE_OPEN,	/* The device is open, but with no buffer */
110 	KCOV_STATE_READY,	/* The buffer has been allocated */
111 	KCOV_STATE_RUNNING,	/* Recording trace data */
112 	KCOV_STATE_DYING,	/* The fd was closed */
113 } kcov_state_t;
114 
115 /*
116  * (l) Set while holding the kcov_lock mutex and not in the RUNNING state.
117  * (o) Only set once while in the OPEN state. Cleaned up while in the DYING
118  *     state, and with no thread associated with the struct kcov_info.
119  * (s) Set atomically to enter or exit the RUNNING state, non-atomically
120  *     otherwise. See above for a description of the other constraints while
121  *     moving into or out of the RUNNING state.
122  */
123 struct kcov_info {
124 	struct thread	*thread;	/* (l) */
125 	vm_object_t	bufobj;		/* (o) */
126 	vm_offset_t	kvaddr;		/* (o) */
127 	size_t		entries;	/* (o) */
128 	size_t		bufsize;	/* (o) */
129 	kcov_state_t	state;		/* (s) */
130 	int		mode;		/* (l) */
131 };
132 
133 /* Prototypes */
134 static d_open_t		kcov_open;
135 static d_close_t	kcov_close;
136 static d_mmap_single_t	kcov_mmap_single;
137 static d_ioctl_t	kcov_ioctl;
138 
139 static int  kcov_alloc(struct kcov_info *info, size_t entries);
140 static void kcov_free(struct kcov_info *info);
141 static void kcov_init(const void *unused);
142 
143 static struct cdevsw kcov_cdevsw = {
144 	.d_version =	D_VERSION,
145 	.d_open =	kcov_open,
146 	.d_close =	kcov_close,
147 	.d_mmap_single = kcov_mmap_single,
148 	.d_ioctl =	kcov_ioctl,
149 	.d_name =	"kcov",
150 };
151 
152 SYSCTL_NODE(_kern, OID_AUTO, kcov, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
153     "Kernel coverage");
154 
155 static u_int kcov_max_entries = KCOV_MAXENTRIES;
156 SYSCTL_UINT(_kern_kcov, OID_AUTO, max_entries, CTLFLAG_RW,
157     &kcov_max_entries, 0,
158     "Maximum number of entries in the kcov buffer");
159 
160 static struct mtx kcov_lock;
161 static int active_count;
162 
163 static struct kcov_info *
164 get_kinfo(struct thread *td)
165 {
166 	struct kcov_info *info;
167 
168 	/* We might have a NULL thread when releasing the secondary CPUs */
169 	if (td == NULL)
170 		return (NULL);
171 
172 	/*
173 	 * We are in an interrupt, stop tracing as it is not explicitly
174 	 * part of a syscall.
175 	 */
176 	if (td->td_intr_nesting_level > 0 || td->td_intr_frame != NULL)
177 		return (NULL);
178 
179 	/*
180 	 * If info is NULL or the state is not running we are not tracing.
181 	 */
182 	info = td->td_kcov_info;
183 	if (info == NULL ||
184 	    atomic_load_acq_int(&info->state) != KCOV_STATE_RUNNING)
185 		return (NULL);
186 
187 	return (info);
188 }
189 
190 static void
191 trace_pc(uintptr_t ret)
192 {
193 	struct thread *td;
194 	struct kcov_info *info;
195 	uint64_t *buf, index;
196 
197 	td = curthread;
198 	info = get_kinfo(td);
199 	if (info == NULL)
200 		return;
201 
202 	/*
203 	 * Check we are in the PC-trace mode.
204 	 */
205 	if (info->mode != KCOV_MODE_TRACE_PC)
206 		return;
207 
208 	KASSERT(info->kvaddr != 0,
209 	    ("__sanitizer_cov_trace_pc: NULL buf while running"));
210 
211 	buf = (uint64_t *)info->kvaddr;
212 
213 	/* The first entry of the buffer holds the index */
214 	index = buf[0];
215 	if (index + 2 > info->entries)
216 		return;
217 
218 	buf[index + 1] = ret;
219 	buf[0] = index + 1;
220 }
221 
222 static bool
223 trace_cmp(uint64_t type, uint64_t arg1, uint64_t arg2, uint64_t ret)
224 {
225 	struct thread *td;
226 	struct kcov_info *info;
227 	uint64_t *buf, index;
228 
229 	td = curthread;
230 	info = get_kinfo(td);
231 	if (info == NULL)
232 		return (false);
233 
234 	/*
235 	 * Check we are in the comparison-trace mode.
236 	 */
237 	if (info->mode != KCOV_MODE_TRACE_CMP)
238 		return (false);
239 
240 	KASSERT(info->kvaddr != 0,
241 	    ("__sanitizer_cov_trace_pc: NULL buf while running"));
242 
243 	buf = (uint64_t *)info->kvaddr;
244 
245 	/* The first entry of the buffer holds the index */
246 	index = buf[0];
247 
248 	/* Check we have space to store all elements */
249 	if (index * 4 + 4 + 1 > info->entries)
250 		return (false);
251 
252 	while (1) {
253 		buf[index * 4 + 1] = type;
254 		buf[index * 4 + 2] = arg1;
255 		buf[index * 4 + 3] = arg2;
256 		buf[index * 4 + 4] = ret;
257 
258 		if (atomic_cmpset_64(&buf[0], index, index + 1))
259 			break;
260 		buf[0] = index;
261 	}
262 
263 	return (true);
264 }
265 
266 /*
267  * The fd is being closed, cleanup everything we can.
268  */
269 static void
270 kcov_mmap_cleanup(void *arg)
271 {
272 	struct kcov_info *info = arg;
273 	struct thread *thread;
274 
275 	mtx_lock_spin(&kcov_lock);
276 	/*
277 	 * Move to KCOV_STATE_DYING to stop adding new entries.
278 	 *
279 	 * If the thread is running we need to wait until thread exit to
280 	 * clean up as it may currently be adding a new entry. If this is
281 	 * the case being in KCOV_STATE_DYING will signal that the buffer
282 	 * needs to be cleaned up.
283 	 */
284 	atomic_store_int(&info->state, KCOV_STATE_DYING);
285 	atomic_thread_fence_seq_cst();
286 	thread = info->thread;
287 	mtx_unlock_spin(&kcov_lock);
288 
289 	if (thread != NULL)
290 		return;
291 
292 	/*
293 	 * We can safely clean up the info struct as it is in the
294 	 * KCOV_STATE_DYING state with no thread associated.
295 	 *
296 	 * The KCOV_STATE_DYING stops new threads from using it.
297 	 * The lack of a thread means nothing is currently using the buffers.
298 	 */
299 	kcov_free(info);
300 }
301 
302 static int
303 kcov_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
304 {
305 	struct kcov_info *info;
306 	int error;
307 
308 	info = malloc(sizeof(struct kcov_info), M_KCOV_INFO, M_ZERO | M_WAITOK);
309 	info->state = KCOV_STATE_OPEN;
310 	info->thread = NULL;
311 	info->mode = -1;
312 
313 	if ((error = devfs_set_cdevpriv(info, kcov_mmap_cleanup)) != 0)
314 		kcov_mmap_cleanup(info);
315 
316 	return (error);
317 }
318 
319 static int
320 kcov_close(struct cdev *dev, int fflag, int devtype, struct thread *td)
321 {
322 	struct kcov_info *info;
323 	int error;
324 
325 	if ((error = devfs_get_cdevpriv((void **)&info)) != 0)
326 		return (error);
327 
328 	KASSERT(info != NULL, ("kcov_close with no kcov_info structure"));
329 
330 	/* Trying to close, but haven't disabled */
331 	if (info->state == KCOV_STATE_RUNNING)
332 		return (EBUSY);
333 
334 	return (0);
335 }
336 
337 static int
338 kcov_mmap_single(struct cdev *dev, vm_ooffset_t *offset, vm_size_t size,
339     struct vm_object **object, int nprot)
340 {
341 	struct kcov_info *info;
342 	int error;
343 
344 	if ((nprot & (PROT_EXEC | PROT_READ | PROT_WRITE)) !=
345 	    (PROT_READ | PROT_WRITE))
346 		return (EINVAL);
347 
348 	if ((error = devfs_get_cdevpriv((void **)&info)) != 0)
349 		return (error);
350 
351 	if (info->kvaddr == 0 || size / KCOV_ELEMENT_SIZE != info->entries)
352 		return (EINVAL);
353 
354 	vm_object_reference(info->bufobj);
355 	*offset = 0;
356 	*object = info->bufobj;
357 	return (0);
358 }
359 
360 static int
361 kcov_alloc(struct kcov_info *info, size_t entries)
362 {
363 	size_t n, pages;
364 	vm_page_t m;
365 
366 	KASSERT(info->kvaddr == 0, ("kcov_alloc: Already have a buffer"));
367 	KASSERT(info->state == KCOV_STATE_OPEN,
368 	    ("kcov_alloc: Not in open state (%x)", info->state));
369 
370 	if (entries < 2 || entries > kcov_max_entries)
371 		return (EINVAL);
372 
373 	/* Align to page size so mmap can't access other kernel memory */
374 	info->bufsize = roundup2(entries * KCOV_ELEMENT_SIZE, PAGE_SIZE);
375 	pages = info->bufsize / PAGE_SIZE;
376 
377 	if ((info->kvaddr = kva_alloc(info->bufsize)) == 0)
378 		return (ENOMEM);
379 
380 	info->bufobj = vm_pager_allocate(OBJT_PHYS, 0, info->bufsize,
381 	    PROT_READ | PROT_WRITE, 0, curthread->td_ucred);
382 
383 	VM_OBJECT_WLOCK(info->bufobj);
384 	for (n = 0; n < pages; n++) {
385 		m = vm_page_grab(info->bufobj, n,
386 		    VM_ALLOC_ZERO | VM_ALLOC_WIRED);
387 		vm_page_valid(m);
388 		vm_page_xunbusy(m);
389 		pmap_qenter(info->kvaddr + n * PAGE_SIZE, &m, 1);
390 	}
391 	VM_OBJECT_WUNLOCK(info->bufobj);
392 
393 	info->entries = entries;
394 
395 	return (0);
396 }
397 
398 static void
399 kcov_free(struct kcov_info *info)
400 {
401 	vm_page_t m;
402 	size_t i;
403 
404 	if (info->kvaddr != 0) {
405 		pmap_qremove(info->kvaddr, info->bufsize / PAGE_SIZE);
406 		kva_free(info->kvaddr, info->bufsize);
407 	}
408 	if (info->bufobj != NULL) {
409 		VM_OBJECT_WLOCK(info->bufobj);
410 		m = vm_page_lookup(info->bufobj, 0);
411 		for (i = 0; i < info->bufsize / PAGE_SIZE; i++) {
412 			vm_page_unwire_noq(m);
413 			m = vm_page_next(m);
414 		}
415 		VM_OBJECT_WUNLOCK(info->bufobj);
416 		vm_object_deallocate(info->bufobj);
417 	}
418 	free(info, M_KCOV_INFO);
419 }
420 
421 static int
422 kcov_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag __unused,
423     struct thread *td)
424 {
425 	struct kcov_info *info;
426 	int mode, error;
427 
428 	if ((error = devfs_get_cdevpriv((void **)&info)) != 0)
429 		return (error);
430 
431 	if (cmd == KIOSETBUFSIZE) {
432 		/*
433 		 * Set the size of the coverage buffer. Should be called
434 		 * before enabling coverage collection for that thread.
435 		 */
436 		if (info->state != KCOV_STATE_OPEN) {
437 			return (EBUSY);
438 		}
439 		error = kcov_alloc(info, *(u_int *)data);
440 		if (error == 0)
441 			info->state = KCOV_STATE_READY;
442 		return (error);
443 	}
444 
445 	mtx_lock_spin(&kcov_lock);
446 	switch (cmd) {
447 	case KIOENABLE:
448 		if (info->state != KCOV_STATE_READY) {
449 			error = EBUSY;
450 			break;
451 		}
452 		if (td->td_kcov_info != NULL) {
453 			error = EINVAL;
454 			break;
455 		}
456 		mode = *(int *)data;
457 		if (mode != KCOV_MODE_TRACE_PC && mode != KCOV_MODE_TRACE_CMP) {
458 			error = EINVAL;
459 			break;
460 		}
461 
462 		/* Lets hope nobody opens this 2 billion times */
463 		KASSERT(active_count < INT_MAX,
464 		    ("%s: Open too many times", __func__));
465 		active_count++;
466 		if (active_count == 1) {
467 			cov_register_pc(&trace_pc);
468 			cov_register_cmp(&trace_cmp);
469 		}
470 
471 		KASSERT(info->thread == NULL,
472 		    ("Enabling kcov when already enabled"));
473 		info->thread = td;
474 		info->mode = mode;
475 		/*
476 		 * Ensure the mode has been set before starting coverage
477 		 * tracing.
478 		 */
479 		atomic_store_rel_int(&info->state, KCOV_STATE_RUNNING);
480 		td->td_kcov_info = info;
481 		break;
482 	case KIODISABLE:
483 		/* Only the currently enabled thread may disable itself */
484 		if (info->state != KCOV_STATE_RUNNING ||
485 		    info != td->td_kcov_info) {
486 			error = EINVAL;
487 			break;
488 		}
489 		KASSERT(active_count > 0, ("%s: Open count is zero", __func__));
490 		active_count--;
491 		if (active_count == 0) {
492 			cov_unregister_pc();
493 			cov_unregister_cmp();
494 		}
495 
496 		td->td_kcov_info = NULL;
497 		atomic_store_int(&info->state, KCOV_STATE_READY);
498 		/*
499 		 * Ensure we have exited the READY state before clearing the
500 		 * rest of the info struct.
501 		 */
502 		atomic_thread_fence_rel();
503 		info->mode = -1;
504 		info->thread = NULL;
505 		break;
506 	default:
507 		error = EINVAL;
508 		break;
509 	}
510 	mtx_unlock_spin(&kcov_lock);
511 
512 	return (error);
513 }
514 
515 static void
516 kcov_thread_dtor(void *arg __unused, struct thread *td)
517 {
518 	struct kcov_info *info;
519 
520 	info = td->td_kcov_info;
521 	if (info == NULL)
522 		return;
523 
524 	mtx_lock_spin(&kcov_lock);
525 	KASSERT(active_count > 0, ("%s: Open count is zero", __func__));
526 	active_count--;
527 	if (active_count == 0) {
528 		cov_unregister_pc();
529 		cov_unregister_cmp();
530 	}
531 	td->td_kcov_info = NULL;
532 	if (info->state != KCOV_STATE_DYING) {
533 		/*
534 		 * The kcov file is still open. Mark it as unused and
535 		 * wait for it to be closed before cleaning up.
536 		 */
537 		atomic_store_int(&info->state, KCOV_STATE_READY);
538 		atomic_thread_fence_seq_cst();
539 		/* This info struct is unused */
540 		info->thread = NULL;
541 		mtx_unlock_spin(&kcov_lock);
542 		return;
543 	}
544 	mtx_unlock_spin(&kcov_lock);
545 
546 	/*
547 	 * We can safely clean up the info struct as it is in the
548 	 * KCOV_STATE_DYING state where the info struct is associated with
549 	 * the current thread that's about to exit.
550 	 *
551 	 * The KCOV_STATE_DYING stops new threads from using it.
552 	 * It also stops the current thread from trying to use the info struct.
553 	 */
554 	kcov_free(info);
555 }
556 
557 static void
558 kcov_init(const void *unused)
559 {
560 	struct make_dev_args args;
561 	struct cdev *dev;
562 
563 	mtx_init(&kcov_lock, "kcov lock", NULL, MTX_SPIN);
564 
565 	make_dev_args_init(&args);
566 	args.mda_devsw = &kcov_cdevsw;
567 	args.mda_uid = UID_ROOT;
568 	args.mda_gid = GID_WHEEL;
569 	args.mda_mode = 0600;
570 	if (make_dev_s(&args, &dev, "kcov") != 0) {
571 		printf("%s", "Failed to create kcov device");
572 		return;
573 	}
574 
575 	EVENTHANDLER_REGISTER(thread_dtor, kcov_thread_dtor, NULL,
576 	    EVENTHANDLER_PRI_ANY);
577 }
578 
579 SYSINIT(kcovdev, SI_SUB_LAST, SI_ORDER_ANY, kcov_init, NULL);
580