1/*
2 * Stress userfaultfd syscall.
3 *
4 *  Copyright (C) 2015  Red Hat, Inc.
5 *
6 *  This work is licensed under the terms of the GNU GPL, version 2. See
7 *  the COPYING file in the top-level directory.
8 *
9 * This test allocates two virtual areas and bounces the physical
10 * memory across the two virtual areas (from area_src to area_dst)
11 * using userfaultfd.
12 *
13 * There are three threads running per CPU:
14 *
15 * 1) one per-CPU thread takes a per-page pthread_mutex in a random
16 *    page of the area_dst (while the physical page may still be in
17 *    area_src), and increments a per-page counter in the same page,
18 *    and checks its value against a verification region.
19 *
20 * 2) another per-CPU thread handles the userfaults generated by
21 *    thread 1 above. userfaultfd blocking reads or poll() modes are
22 *    exercised interleaved.
23 *
24 * 3) one last per-CPU thread transfers the memory in the background
25 *    at maximum bandwidth (if not already transferred by thread
26 *    2). Each cpu thread takes cares of transferring a portion of the
27 *    area.
28 *
29 * When all threads of type 3 completed the transfer, one bounce is
30 * complete. area_src and area_dst are then swapped. All threads are
31 * respawned and so the bounce is immediately restarted in the
32 * opposite direction.
33 *
34 * per-CPU threads 1 by triggering userfaults inside
35 * pthread_mutex_lock will also verify the atomicity of the memory
36 * transfer (UFFDIO_COPY).
37 *
38 * The program takes two parameters: the amounts of physical memory in
39 * megabytes (MiB) of the area and the number of bounces to execute.
40 *
41 * # 100MiB 99999 bounces
42 * ./userfaultfd 100 99999
43 *
44 * # 1GiB 99 bounces
45 * ./userfaultfd 1000 99
46 *
47 * # 10MiB-~6GiB 999 bounces, continue forever unless an error triggers
48 * while ./userfaultfd $[RANDOM % 6000 + 10] 999; do true; done
49 */
50
51#define _GNU_SOURCE
52#include <stdio.h>
53#include <errno.h>
54#include <unistd.h>
55#include <stdlib.h>
56#include <sys/types.h>
57#include <sys/stat.h>
58#include <fcntl.h>
59#include <time.h>
60#include <signal.h>
61#include <poll.h>
62#include <string.h>
63#include <sys/mman.h>
64#include <sys/syscall.h>
65#include <sys/ioctl.h>
66#include <pthread.h>
67#include <linux/userfaultfd.h>
68
69#ifdef __NR_userfaultfd
70
71static unsigned long nr_cpus, nr_pages, nr_pages_per_cpu, page_size;
72
73#define BOUNCE_RANDOM		(1<<0)
74#define BOUNCE_RACINGFAULTS	(1<<1)
75#define BOUNCE_VERIFY		(1<<2)
76#define BOUNCE_POLL		(1<<3)
77static int bounces;
78
79static unsigned long long *count_verify;
80static int uffd, finished, *pipefd;
81static char *area_src, *area_dst;
82static char *zeropage;
83pthread_attr_t attr;
84
85/* pthread_mutex_t starts at page offset 0 */
86#define area_mutex(___area, ___nr)					\
87	((pthread_mutex_t *) ((___area) + (___nr)*page_size))
88/*
89 * count is placed in the page after pthread_mutex_t naturally aligned
90 * to avoid non alignment faults on non-x86 archs.
91 */
92#define area_count(___area, ___nr)					\
93	((volatile unsigned long long *) ((unsigned long)		\
94				 ((___area) + (___nr)*page_size +	\
95				  sizeof(pthread_mutex_t) +		\
96				  sizeof(unsigned long long) - 1) &	\
97				 ~(unsigned long)(sizeof(unsigned long long) \
98						  -  1)))
99
100static int my_bcmp(char *str1, char *str2, size_t n)
101{
102	unsigned long i;
103	for (i = 0; i < n; i++)
104		if (str1[i] != str2[i])
105			return 1;
106	return 0;
107}
108
109static void *locking_thread(void *arg)
110{
111	unsigned long cpu = (unsigned long) arg;
112	struct random_data rand;
113	unsigned long page_nr = *(&(page_nr)); /* uninitialized warning */
114	int32_t rand_nr;
115	unsigned long long count;
116	char randstate[64];
117	unsigned int seed;
118	time_t start;
119
120	if (bounces & BOUNCE_RANDOM) {
121		seed = (unsigned int) time(NULL) - bounces;
122		if (!(bounces & BOUNCE_RACINGFAULTS))
123			seed += cpu;
124		bzero(&rand, sizeof(rand));
125		bzero(&randstate, sizeof(randstate));
126		if (initstate_r(seed, randstate, sizeof(randstate), &rand))
127			fprintf(stderr, "srandom_r error\n"), exit(1);
128	} else {
129		page_nr = -bounces;
130		if (!(bounces & BOUNCE_RACINGFAULTS))
131			page_nr += cpu * nr_pages_per_cpu;
132	}
133
134	while (!finished) {
135		if (bounces & BOUNCE_RANDOM) {
136			if (random_r(&rand, &rand_nr))
137				fprintf(stderr, "random_r 1 error\n"), exit(1);
138			page_nr = rand_nr;
139			if (sizeof(page_nr) > sizeof(rand_nr)) {
140				if (random_r(&rand, &rand_nr))
141					fprintf(stderr, "random_r 2 error\n"), exit(1);
142				page_nr |= (((unsigned long) rand_nr) << 16) <<
143					   16;
144			}
145		} else
146			page_nr += 1;
147		page_nr %= nr_pages;
148
149		start = time(NULL);
150		if (bounces & BOUNCE_VERIFY) {
151			count = *area_count(area_dst, page_nr);
152			if (!count)
153				fprintf(stderr,
154					"page_nr %lu wrong count %Lu %Lu\n",
155					page_nr, count,
156					count_verify[page_nr]), exit(1);
157
158
159			/*
160			 * We can't use bcmp (or memcmp) because that
161			 * returns 0 erroneously if the memory is
162			 * changing under it (even if the end of the
163			 * page is never changing and always
164			 * different).
165			 */
166#if 1
167			if (!my_bcmp(area_dst + page_nr * page_size, zeropage,
168				     page_size))
169				fprintf(stderr,
170					"my_bcmp page_nr %lu wrong count %Lu %Lu\n",
171					page_nr, count,
172					count_verify[page_nr]), exit(1);
173#else
174			unsigned long loops;
175
176			loops = 0;
177			/* uncomment the below line to test with mutex */
178			/* pthread_mutex_lock(area_mutex(area_dst, page_nr)); */
179			while (!bcmp(area_dst + page_nr * page_size, zeropage,
180				     page_size)) {
181				loops += 1;
182				if (loops > 10)
183					break;
184			}
185			/* uncomment below line to test with mutex */
186			/* pthread_mutex_unlock(area_mutex(area_dst, page_nr)); */
187			if (loops) {
188				fprintf(stderr,
189					"page_nr %lu all zero thread %lu %p %lu\n",
190					page_nr, cpu, area_dst + page_nr * page_size,
191					loops);
192				if (loops > 10)
193					exit(1);
194			}
195#endif
196		}
197
198		pthread_mutex_lock(area_mutex(area_dst, page_nr));
199		count = *area_count(area_dst, page_nr);
200		if (count != count_verify[page_nr]) {
201			fprintf(stderr,
202				"page_nr %lu memory corruption %Lu %Lu\n",
203				page_nr, count,
204				count_verify[page_nr]), exit(1);
205		}
206		count++;
207		*area_count(area_dst, page_nr) = count_verify[page_nr] = count;
208		pthread_mutex_unlock(area_mutex(area_dst, page_nr));
209
210		if (time(NULL) - start > 1)
211			fprintf(stderr,
212				"userfault too slow %ld "
213				"possible false positive with overcommit\n",
214				time(NULL) - start);
215	}
216
217	return NULL;
218}
219
220static int copy_page(unsigned long offset)
221{
222	struct uffdio_copy uffdio_copy;
223
224	if (offset >= nr_pages * page_size)
225		fprintf(stderr, "unexpected offset %lu\n",
226			offset), exit(1);
227	uffdio_copy.dst = (unsigned long) area_dst + offset;
228	uffdio_copy.src = (unsigned long) area_src + offset;
229	uffdio_copy.len = page_size;
230	uffdio_copy.mode = 0;
231	uffdio_copy.copy = 0;
232	if (ioctl(uffd, UFFDIO_COPY, &uffdio_copy)) {
233		/* real retval in ufdio_copy.copy */
234		if (uffdio_copy.copy != -EEXIST)
235			fprintf(stderr, "UFFDIO_COPY error %Ld\n",
236				uffdio_copy.copy), exit(1);
237	} else if (uffdio_copy.copy != page_size) {
238		fprintf(stderr, "UFFDIO_COPY unexpected copy %Ld\n",
239			uffdio_copy.copy), exit(1);
240	} else
241		return 1;
242	return 0;
243}
244
245static void *uffd_poll_thread(void *arg)
246{
247	unsigned long cpu = (unsigned long) arg;
248	struct pollfd pollfd[2];
249	struct uffd_msg msg;
250	int ret;
251	unsigned long offset;
252	char tmp_chr;
253	unsigned long userfaults = 0;
254
255	pollfd[0].fd = uffd;
256	pollfd[0].events = POLLIN;
257	pollfd[1].fd = pipefd[cpu*2];
258	pollfd[1].events = POLLIN;
259
260	for (;;) {
261		ret = poll(pollfd, 2, -1);
262		if (!ret)
263			fprintf(stderr, "poll error %d\n", ret), exit(1);
264		if (ret < 0)
265			perror("poll"), exit(1);
266		if (pollfd[1].revents & POLLIN) {
267			if (read(pollfd[1].fd, &tmp_chr, 1) != 1)
268				fprintf(stderr, "read pipefd error\n"),
269					exit(1);
270			break;
271		}
272		if (!(pollfd[0].revents & POLLIN))
273			fprintf(stderr, "pollfd[0].revents %d\n",
274				pollfd[0].revents), exit(1);
275		ret = read(uffd, &msg, sizeof(msg));
276		if (ret < 0) {
277			if (errno == EAGAIN)
278				continue;
279			perror("nonblocking read error"), exit(1);
280		}
281		if (msg.event != UFFD_EVENT_PAGEFAULT)
282			fprintf(stderr, "unexpected msg event %u\n",
283				msg.event), exit(1);
284		if (msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
285			fprintf(stderr, "unexpected write fault\n"), exit(1);
286		offset = (char *)(unsigned long)msg.arg.pagefault.address -
287			 area_dst;
288		offset &= ~(page_size-1);
289		if (copy_page(offset))
290			userfaults++;
291	}
292	return (void *)userfaults;
293}
294
295pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;
296
297static void *uffd_read_thread(void *arg)
298{
299	unsigned long *this_cpu_userfaults;
300	struct uffd_msg msg;
301	unsigned long offset;
302	int ret;
303
304	this_cpu_userfaults = (unsigned long *) arg;
305	*this_cpu_userfaults = 0;
306
307	pthread_mutex_unlock(&uffd_read_mutex);
308	/* from here cancellation is ok */
309
310	for (;;) {
311		ret = read(uffd, &msg, sizeof(msg));
312		if (ret != sizeof(msg)) {
313			if (ret < 0)
314				perror("blocking read error"), exit(1);
315			else
316				fprintf(stderr, "short read\n"), exit(1);
317		}
318		if (msg.event != UFFD_EVENT_PAGEFAULT)
319			fprintf(stderr, "unexpected msg event %u\n",
320				msg.event), exit(1);
321		if (bounces & BOUNCE_VERIFY &&
322		    msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WRITE)
323			fprintf(stderr, "unexpected write fault\n"), exit(1);
324		offset = (char *)(unsigned long)msg.arg.pagefault.address -
325			 area_dst;
326		offset &= ~(page_size-1);
327		if (copy_page(offset))
328			(*this_cpu_userfaults)++;
329	}
330	return (void *)NULL;
331}
332
333static void *background_thread(void *arg)
334{
335	unsigned long cpu = (unsigned long) arg;
336	unsigned long page_nr;
337
338	for (page_nr = cpu * nr_pages_per_cpu;
339	     page_nr < (cpu+1) * nr_pages_per_cpu;
340	     page_nr++)
341		copy_page(page_nr * page_size);
342
343	return NULL;
344}
345
346static int stress(unsigned long *userfaults)
347{
348	unsigned long cpu;
349	pthread_t locking_threads[nr_cpus];
350	pthread_t uffd_threads[nr_cpus];
351	pthread_t background_threads[nr_cpus];
352	void **_userfaults = (void **) userfaults;
353
354	finished = 0;
355	for (cpu = 0; cpu < nr_cpus; cpu++) {
356		if (pthread_create(&locking_threads[cpu], &attr,
357				   locking_thread, (void *)cpu))
358			return 1;
359		if (bounces & BOUNCE_POLL) {
360			if (pthread_create(&uffd_threads[cpu], &attr,
361					   uffd_poll_thread, (void *)cpu))
362				return 1;
363		} else {
364			if (pthread_create(&uffd_threads[cpu], &attr,
365					   uffd_read_thread,
366					   &_userfaults[cpu]))
367				return 1;
368			pthread_mutex_lock(&uffd_read_mutex);
369		}
370		if (pthread_create(&background_threads[cpu], &attr,
371				   background_thread, (void *)cpu))
372			return 1;
373	}
374	for (cpu = 0; cpu < nr_cpus; cpu++)
375		if (pthread_join(background_threads[cpu], NULL))
376			return 1;
377
378	/*
379	 * Be strict and immediately zap area_src, the whole area has
380	 * been transferred already by the background treads. The
381	 * area_src could then be faulted in in a racy way by still
382	 * running uffdio_threads reading zeropages after we zapped
383	 * area_src (but they're guaranteed to get -EEXIST from
384	 * UFFDIO_COPY without writing zero pages into area_dst
385	 * because the background threads already completed).
386	 */
387	if (madvise(area_src, nr_pages * page_size, MADV_DONTNEED)) {
388		perror("madvise");
389		return 1;
390	}
391
392	for (cpu = 0; cpu < nr_cpus; cpu++) {
393		char c;
394		if (bounces & BOUNCE_POLL) {
395			if (write(pipefd[cpu*2+1], &c, 1) != 1) {
396				fprintf(stderr, "pipefd write error\n");
397				return 1;
398			}
399			if (pthread_join(uffd_threads[cpu], &_userfaults[cpu]))
400				return 1;
401		} else {
402			if (pthread_cancel(uffd_threads[cpu]))
403				return 1;
404			if (pthread_join(uffd_threads[cpu], NULL))
405				return 1;
406		}
407	}
408
409	finished = 1;
410	for (cpu = 0; cpu < nr_cpus; cpu++)
411		if (pthread_join(locking_threads[cpu], NULL))
412			return 1;
413
414	return 0;
415}
416
417static int userfaultfd_stress(void)
418{
419	void *area;
420	char *tmp_area;
421	unsigned long nr;
422	struct uffdio_register uffdio_register;
423	struct uffdio_api uffdio_api;
424	unsigned long cpu;
425	int uffd_flags, err;
426	unsigned long userfaults[nr_cpus];
427
428	if (posix_memalign(&area, page_size, nr_pages * page_size)) {
429		fprintf(stderr, "out of memory\n");
430		return 1;
431	}
432	area_src = area;
433	if (posix_memalign(&area, page_size, nr_pages * page_size)) {
434		fprintf(stderr, "out of memory\n");
435		return 1;
436	}
437	area_dst = area;
438
439	uffd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
440	if (uffd < 0) {
441		fprintf(stderr,
442			"userfaultfd syscall not available in this kernel\n");
443		return 1;
444	}
445	uffd_flags = fcntl(uffd, F_GETFD, NULL);
446
447	uffdio_api.api = UFFD_API;
448	uffdio_api.features = 0;
449	if (ioctl(uffd, UFFDIO_API, &uffdio_api)) {
450		fprintf(stderr, "UFFDIO_API\n");
451		return 1;
452	}
453	if (uffdio_api.api != UFFD_API) {
454		fprintf(stderr, "UFFDIO_API error %Lu\n", uffdio_api.api);
455		return 1;
456	}
457
458	count_verify = malloc(nr_pages * sizeof(unsigned long long));
459	if (!count_verify) {
460		perror("count_verify");
461		return 1;
462	}
463
464	for (nr = 0; nr < nr_pages; nr++) {
465		*area_mutex(area_src, nr) = (pthread_mutex_t)
466			PTHREAD_MUTEX_INITIALIZER;
467		count_verify[nr] = *area_count(area_src, nr) = 1;
468		/*
469		 * In the transition between 255 to 256, powerpc will
470		 * read out of order in my_bcmp and see both bytes as
471		 * zero, so leave a placeholder below always non-zero
472		 * after the count, to avoid my_bcmp to trigger false
473		 * positives.
474		 */
475		*(area_count(area_src, nr) + 1) = 1;
476	}
477
478	pipefd = malloc(sizeof(int) * nr_cpus * 2);
479	if (!pipefd) {
480		perror("pipefd");
481		return 1;
482	}
483	for (cpu = 0; cpu < nr_cpus; cpu++) {
484		if (pipe2(&pipefd[cpu*2], O_CLOEXEC | O_NONBLOCK)) {
485			perror("pipe");
486			return 1;
487		}
488	}
489
490	if (posix_memalign(&area, page_size, page_size)) {
491		fprintf(stderr, "out of memory\n");
492		return 1;
493	}
494	zeropage = area;
495	bzero(zeropage, page_size);
496
497	pthread_mutex_lock(&uffd_read_mutex);
498
499	pthread_attr_init(&attr);
500	pthread_attr_setstacksize(&attr, 16*1024*1024);
501
502	err = 0;
503	while (bounces--) {
504		unsigned long expected_ioctls;
505
506		printf("bounces: %d, mode:", bounces);
507		if (bounces & BOUNCE_RANDOM)
508			printf(" rnd");
509		if (bounces & BOUNCE_RACINGFAULTS)
510			printf(" racing");
511		if (bounces & BOUNCE_VERIFY)
512			printf(" ver");
513		if (bounces & BOUNCE_POLL)
514			printf(" poll");
515		printf(", ");
516		fflush(stdout);
517
518		if (bounces & BOUNCE_POLL)
519			fcntl(uffd, F_SETFL, uffd_flags | O_NONBLOCK);
520		else
521			fcntl(uffd, F_SETFL, uffd_flags & ~O_NONBLOCK);
522
523		/* register */
524		uffdio_register.range.start = (unsigned long) area_dst;
525		uffdio_register.range.len = nr_pages * page_size;
526		uffdio_register.mode = UFFDIO_REGISTER_MODE_MISSING;
527		if (ioctl(uffd, UFFDIO_REGISTER, &uffdio_register)) {
528			fprintf(stderr, "register failure\n");
529			return 1;
530		}
531		expected_ioctls = (1 << _UFFDIO_WAKE) |
532				  (1 << _UFFDIO_COPY) |
533				  (1 << _UFFDIO_ZEROPAGE);
534		if ((uffdio_register.ioctls & expected_ioctls) !=
535		    expected_ioctls) {
536			fprintf(stderr,
537				"unexpected missing ioctl for anon memory\n");
538			return 1;
539		}
540
541		/*
542		 * The madvise done previously isn't enough: some
543		 * uffd_thread could have read userfaults (one of
544		 * those already resolved by the background thread)
545		 * and it may be in the process of calling
546		 * UFFDIO_COPY. UFFDIO_COPY will read the zapped
547		 * area_src and it would map a zero page in it (of
548		 * course such a UFFDIO_COPY is perfectly safe as it'd
549		 * return -EEXIST). The problem comes at the next
550		 * bounce though: that racing UFFDIO_COPY would
551		 * generate zeropages in the area_src, so invalidating
552		 * the previous MADV_DONTNEED. Without this additional
553		 * MADV_DONTNEED those zeropages leftovers in the
554		 * area_src would lead to -EEXIST failure during the
555		 * next bounce, effectively leaving a zeropage in the
556		 * area_dst.
557		 *
558		 * Try to comment this out madvise to see the memory
559		 * corruption being caught pretty quick.
560		 *
561		 * khugepaged is also inhibited to collapse THP after
562		 * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
563		 * required to MADV_DONTNEED here.
564		 */
565		if (madvise(area_dst, nr_pages * page_size, MADV_DONTNEED)) {
566			perror("madvise 2");
567			return 1;
568		}
569
570		/* bounce pass */
571		if (stress(userfaults))
572			return 1;
573
574		/* unregister */
575		if (ioctl(uffd, UFFDIO_UNREGISTER, &uffdio_register.range)) {
576			fprintf(stderr, "register failure\n");
577			return 1;
578		}
579
580		/* verification */
581		if (bounces & BOUNCE_VERIFY) {
582			for (nr = 0; nr < nr_pages; nr++) {
583				if (*area_count(area_dst, nr) != count_verify[nr]) {
584					fprintf(stderr,
585						"error area_count %Lu %Lu %lu\n",
586						*area_count(area_src, nr),
587						count_verify[nr],
588						nr);
589					err = 1;
590					bounces = 0;
591				}
592			}
593		}
594
595		/* prepare next bounce */
596		tmp_area = area_src;
597		area_src = area_dst;
598		area_dst = tmp_area;
599
600		printf("userfaults:");
601		for (cpu = 0; cpu < nr_cpus; cpu++)
602			printf(" %lu", userfaults[cpu]);
603		printf("\n");
604	}
605
606	return err;
607}
608
609int main(int argc, char **argv)
610{
611	if (argc < 3)
612		fprintf(stderr, "Usage: <MiB> <bounces>\n"), exit(1);
613	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
614	page_size = sysconf(_SC_PAGE_SIZE);
615	if ((unsigned long) area_count(NULL, 0) + sizeof(unsigned long long) * 2
616	    > page_size)
617		fprintf(stderr, "Impossible to run this test\n"), exit(2);
618	nr_pages_per_cpu = atol(argv[1]) * 1024*1024 / page_size /
619		nr_cpus;
620	if (!nr_pages_per_cpu) {
621		fprintf(stderr, "invalid MiB\n");
622		fprintf(stderr, "Usage: <MiB> <bounces>\n"), exit(1);
623	}
624	bounces = atoi(argv[2]);
625	if (bounces <= 0) {
626		fprintf(stderr, "invalid bounces\n");
627		fprintf(stderr, "Usage: <MiB> <bounces>\n"), exit(1);
628	}
629	nr_pages = nr_pages_per_cpu * nr_cpus;
630	printf("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
631	       nr_pages, nr_pages_per_cpu);
632	return userfaultfd_stress();
633}
634
635#else /* __NR_userfaultfd */
636
637#warning "missing __NR_userfaultfd definition"
638
639int main(void)
640{
641	printf("skip: Skipping userfaultfd test (missing __NR_userfaultfd)\n");
642	return 0;
643}
644
645#endif /* __NR_userfaultfd */
646