tools/testing/selftests/mm/uffd-stress.c

// SPDX-License-Identifier: GPL-2.0-only
/*
 * Stress userfaultfd syscall.
 *
 *  Copyright (C) 2015  Red Hat, Inc.
 *
 * This test allocates two virtual areas and bounces the physical
 * memory across the two virtual areas (from area_src to area_dst)
 * using userfaultfd.
 *
 * There are three threads running per CPU:
 *
 * 1) one per-CPU thread takes a per-page pthread_mutex in a random
 *    page of the area_dst (while the physical page may still be in
 *    area_src), and increments a per-page counter in the same page,
 *    and checks its value against a verification region.
 *
 * 2) another per-CPU thread handles the userfaults generated by
 *    thread 1 above. userfaultfd blocking reads or poll() modes are
 *    exercised interleaved.
 *
 * 3) one last per-CPU thread transfers the memory in the background
 *    at maximum bandwidth (if not already transferred by thread
 *    2). Each cpu thread takes cares of transferring a portion of the
 *    area.
 *
 * When all threads of type 3 completed the transfer, one bounce is
 * complete. area_src and area_dst are then swapped. All threads are
 * respawned and so the bounce is immediately restarted in the
 * opposite direction.
 *
 * per-CPU threads 1 by triggering userfaults inside
 * pthread_mutex_lock will also verify the atomicity of the memory
 * transfer (UFFDIO_COPY).
 */

#include "uffd-common.h"

uint64_t features;
#ifdef __NR_userfaultfd

#define BOUNCE_RANDOM		(1<<0)
#define BOUNCE_RACINGFAULTS	(1<<1)
#define BOUNCE_VERIFY		(1<<2)
#define BOUNCE_POLL		(1<<3)
static int bounces;
/* defined globally for this particular test as the sigalrm handler
 * depends on test_uffdio_*_eexist.
 * XXX: define gopts in main() when we figure out a way to deal with
 * test_uffdio_*_eexist.
 */
static uffd_global_test_opts_t *gopts;

/* exercise the test_uffdio_*_eexist every ALARM_INTERVAL_SECS */
#define ALARM_INTERVAL_SECS 10
static char *zeropage;
pthread_attr_t attr;

#define swap(a, b) \
	do { __auto_type __tmp = (a); (a) = (b); (b) = __tmp; } while (0)

const char *examples =
	"# Run anonymous memory test on 100MiB region with 99999 bounces:\n"
	"./uffd-stress anon 100 99999\n\n"
	"# Run share memory test on 1GiB region with 99 bounces:\n"
	"./uffd-stress shmem 1000 99\n\n"
	"# Run hugetlb memory test on 256MiB region with 50 bounces:\n"
	"./uffd-stress hugetlb 256 50\n\n"
	"# Run the same hugetlb test but using private file:\n"
	"./uffd-stress hugetlb-private 256 50\n\n"
	"# 10MiB-~6GiB 999 bounces anonymous test, "
	"continue forever unless an error triggers\n"
	"while ./uffd-stress anon $[RANDOM % 6000 + 10] 999; do true; done\n\n";

static void usage(void)
{
	fprintf(stderr, "\nUsage: ./uffd-stress <test type> <MiB> <bounces>\n\n");
	fprintf(stderr, "Supported <test type>: anon, hugetlb, "
		"hugetlb-private, shmem, shmem-private\n\n");
	fprintf(stderr, "Examples:\n\n");
	fprintf(stderr, "%s", examples);
	exit(1);
}

static void uffd_stats_reset(uffd_global_test_opts_t *gopts, struct uffd_args *args,
			     unsigned long n_cpus)
{
	int i;

	for (i = 0; i < n_cpus; i++) {
		args[i].cpu = i;
		args[i].apply_wp = gopts->test_uffdio_wp;
		args[i].missing_faults = 0;
		args[i].wp_faults = 0;
		args[i].minor_faults = 0;
		args[i].gopts = gopts;
	}
}

static void *locking_thread(void *arg)
{
	struct uffd_args *args = (struct uffd_args *) arg;
	uffd_global_test_opts_t *gopts = args->gopts;
	unsigned long cpu = (unsigned long) args->cpu;
	unsigned long page_nr;
	unsigned long long count;

	if (!(bounces & BOUNCE_RANDOM)) {
		page_nr = -bounces;
		if (!(bounces & BOUNCE_RACINGFAULTS))
			page_nr += cpu * gopts->nr_pages_per_cpu;
	}

	while (!gopts->finished) {
		if (bounces & BOUNCE_RANDOM) {
			if (getrandom(&page_nr, sizeof(page_nr), 0) != sizeof(page_nr))
				err("getrandom failed");
		} else
			page_nr += 1;
		page_nr %= gopts->nr_pages;
		pthread_mutex_lock(area_mutex(gopts->area_dst, page_nr, gopts));
		count = *area_count(gopts->area_dst, page_nr, gopts);
		if (count != gopts->count_verify[page_nr])
			err("page_nr %lu memory corruption %llu %llu",
			    page_nr, count, gopts->count_verify[page_nr]);
		count++;
		*area_count(gopts->area_dst, page_nr, gopts) = gopts->count_verify[page_nr] = count;
		pthread_mutex_unlock(area_mutex(gopts->area_dst, page_nr, gopts));
	}

	return NULL;
}

static int copy_page_retry(uffd_global_test_opts_t *gopts, unsigned long offset)
{
	return __copy_page(gopts, offset, true, gopts->test_uffdio_wp);
}

pthread_mutex_t uffd_read_mutex = PTHREAD_MUTEX_INITIALIZER;

static void *uffd_read_thread(void *arg)
{
	struct uffd_args *args = (struct uffd_args *)arg;
	uffd_global_test_opts_t *gopts = args->gopts;
	struct uffd_msg msg;

	pthread_mutex_unlock(&uffd_read_mutex);
	/* from here cancellation is ok */

	for (;;) {
		if (uffd_read_msg(gopts, &msg))
			continue;
		uffd_handle_page_fault(gopts, &msg, args);
	}

	return NULL;
}

static void *background_thread(void *arg)
{
	struct uffd_args *args = (struct uffd_args *) arg;
	uffd_global_test_opts_t *gopts = args->gopts;
	unsigned long cpu = (unsigned long) args->cpu;
	unsigned long page_nr, start_nr, mid_nr, end_nr;

	start_nr = cpu * gopts->nr_pages_per_cpu;
	end_nr = (cpu+1) * gopts->nr_pages_per_cpu;
	mid_nr = (start_nr + end_nr) / 2;

	/* Copy the first half of the pages */
	for (page_nr = start_nr; page_nr < mid_nr; page_nr++)
		copy_page_retry(gopts, page_nr * gopts->page_size);

	/*
	 * If we need to test uffd-wp, set it up now.  Then we'll have
	 * at least the first half of the pages mapped already which
	 * can be write-protected for testing
	 */
	if (gopts->test_uffdio_wp)
		wp_range(gopts->uffd, (unsigned long)gopts->area_dst + start_nr * gopts->page_size,
			gopts->nr_pages_per_cpu * gopts->page_size, true);

	/*
	 * Continue the 2nd half of the page copying, handling write
	 * protection faults if any
	 */
	for (page_nr = mid_nr; page_nr < end_nr; page_nr++)
		copy_page_retry(gopts, page_nr * gopts->page_size);

	return NULL;
}

static int stress(struct uffd_args *args)
{
	unsigned long cpu;
	uffd_global_test_opts_t *gopts = args->gopts;
	pthread_t locking_threads[gopts->nr_parallel];
	pthread_t uffd_threads[gopts->nr_parallel];
	pthread_t background_threads[gopts->nr_parallel];

	gopts->finished = 0;
	for (cpu = 0; cpu < gopts->nr_parallel; cpu++) {
		if (pthread_create(&locking_threads[cpu], &attr,
				   locking_thread, (void *)&args[cpu]))
			return 1;
		if (bounces & BOUNCE_POLL) {
			if (pthread_create(&uffd_threads[cpu],
					   &attr,
					   uffd_poll_thread,
					   (void *) &args[cpu]))
				err("uffd_poll_thread create");
		} else {
			if (pthread_create(&uffd_threads[cpu], &attr,
					   uffd_read_thread,
					   (void *)&args[cpu]))
				return 1;
			pthread_mutex_lock(&uffd_read_mutex);
		}
		if (pthread_create(&background_threads[cpu], &attr,
				   background_thread, (void *)&args[cpu]))
			return 1;
	}
	for (cpu = 0; cpu < gopts->nr_parallel; cpu++)
		if (pthread_join(background_threads[cpu], NULL))
			return 1;

	/*
	 * Be strict and immediately zap area_src, the whole area has
	 * been transferred already by the background treads. The
	 * area_src could then be faulted in a racy way by still
	 * running uffdio_threads reading zeropages after we zapped
	 * area_src (but they're guaranteed to get -EEXIST from
	 * UFFDIO_COPY without writing zero pages into area_dst
	 * because the background threads already completed).
	 */
	uffd_test_ops->release_pages(gopts, gopts->area_src);

	gopts->finished = 1;
	for (cpu = 0; cpu < gopts->nr_parallel; cpu++)
		if (pthread_join(locking_threads[cpu], NULL))
			return 1;

	for (cpu = 0; cpu < gopts->nr_parallel; cpu++) {
		char c = '\0';
		if (bounces & BOUNCE_POLL) {
			if (write(gopts->pipefd[cpu*2+1], &c, 1) != 1)
				err("pipefd write error");
			if (pthread_join(uffd_threads[cpu],
					 (void *)&args[cpu]))
				return 1;
		} else {
			if (pthread_cancel(uffd_threads[cpu]))
				return 1;
			if (pthread_join(uffd_threads[cpu], NULL))
				return 1;
		}
	}

	return 0;
}

static int userfaultfd_stress(uffd_global_test_opts_t *gopts)
{
	void *area;
	unsigned long nr;
	struct uffd_args args[gopts->nr_parallel];
	uint64_t mem_size = gopts->nr_pages * gopts->page_size;
	int flags = 0;

	memset(args, 0, sizeof(struct uffd_args) * gopts->nr_parallel);

	if (features & UFFD_FEATURE_WP_UNPOPULATED && gopts->test_type == TEST_ANON)
		flags = UFFD_FEATURE_WP_UNPOPULATED;

	if (uffd_test_ctx_init(gopts, flags, NULL))
		err("context init failed");

	if (posix_memalign(&area, gopts->page_size, gopts->page_size))
		err("out of memory");
	zeropage = area;
	bzero(zeropage, gopts->page_size);

	pthread_mutex_lock(&uffd_read_mutex);

	pthread_attr_init(&attr);
	pthread_attr_setstacksize(&attr, 16*1024*1024);

	while (bounces--) {
		ksft_print_msg("bounces: %d, mode:%s%s%s%s, ",
			       bounces,
			       bounces & BOUNCE_RANDOM ? " rnd" : "",
			       bounces & BOUNCE_RACINGFAULTS ? " racing" : "",
			       bounces & BOUNCE_VERIFY ? " ver" : "",
			       bounces & BOUNCE_POLL ? " poll" : " read");
		fflush(stdout);

		if (bounces & BOUNCE_POLL)
			fcntl(gopts->uffd, F_SETFL, gopts->uffd_flags | O_NONBLOCK);
		else
			fcntl(gopts->uffd, F_SETFL, gopts->uffd_flags & ~O_NONBLOCK);

		/* register */
		if (uffd_register(gopts->uffd, gopts->area_dst, mem_size,
				  true, gopts->test_uffdio_wp, false))
			err("register failure");

		if (gopts->area_dst_alias) {
			if (uffd_register(gopts->uffd, gopts->area_dst_alias, mem_size,
					  true, gopts->test_uffdio_wp, false))
				err("register failure alias");
		}

		/*
		 * The madvise done previously isn't enough: some
		 * uffd_thread could have read userfaults (one of
		 * those already resolved by the background thread)
		 * and it may be in the process of calling
		 * UFFDIO_COPY. UFFDIO_COPY will read the zapped
		 * area_src and it would map a zero page in it (of
		 * course such a UFFDIO_COPY is perfectly safe as it'd
		 * return -EEXIST). The problem comes at the next
		 * bounce though: that racing UFFDIO_COPY would
		 * generate zeropages in the area_src, so invalidating
		 * the previous MADV_DONTNEED. Without this additional
		 * MADV_DONTNEED those zeropages leftovers in the
		 * area_src would lead to -EEXIST failure during the
		 * next bounce, effectively leaving a zeropage in the
		 * area_dst.
		 *
		 * Try to comment this out madvise to see the memory
		 * corruption being caught pretty quick.
		 *
		 * khugepaged is also inhibited to collapse THP after
		 * MADV_DONTNEED only after the UFFDIO_REGISTER, so it's
		 * required to MADV_DONTNEED here.
		 */
		uffd_test_ops->release_pages(gopts, gopts->area_dst);

		uffd_stats_reset(gopts, args, gopts->nr_parallel);

		/* bounce pass */
		if (stress(args)) {
			uffd_test_ctx_clear(gopts);
			return 1;
		}

		/* Clear all the write protections if there is any */
		if (gopts->test_uffdio_wp)
			wp_range(gopts->uffd, (unsigned long)gopts->area_dst,
				 gopts->nr_pages * gopts->page_size, false);

		/* unregister */
		if (uffd_unregister(gopts->uffd, gopts->area_dst, mem_size))
			err("unregister failure");
		if (gopts->area_dst_alias) {
			if (uffd_unregister(gopts->uffd, gopts->area_dst_alias, mem_size))
				err("unregister failure alias");
		}

		/* verification */
		if (bounces & BOUNCE_VERIFY)
			for (nr = 0; nr < gopts->nr_pages; nr++)
				if (*area_count(gopts->area_dst, nr, gopts) !=
						gopts->count_verify[nr])
					err("error area_count %llu %llu %lu\n",
					    *area_count(gopts->area_src, nr, gopts),
					    gopts->count_verify[nr], nr);

		/* prepare next bounce */
		swap(gopts->area_src, gopts->area_dst);

		swap(gopts->area_src_alias, gopts->area_dst_alias);

		uffd_stats_report(args, gopts->nr_parallel);
	}
	uffd_test_ctx_clear(gopts);

	return 0;
}

static void set_test_type(uffd_global_test_opts_t *gopts, const char *type)
{
	if (!strcmp(type, "anon")) {
		gopts->test_type = TEST_ANON;
		uffd_test_ops = &anon_uffd_test_ops;
	} else if (!strcmp(type, "hugetlb")) {
		gopts->test_type = TEST_HUGETLB;
		uffd_test_ops = &hugetlb_uffd_test_ops;
		gopts->map_shared = true;
	} else if (!strcmp(type, "hugetlb-private")) {
		gopts->test_type = TEST_HUGETLB;
		uffd_test_ops = &hugetlb_uffd_test_ops;
	} else if (!strcmp(type, "shmem")) {
		gopts->map_shared = true;
		gopts->test_type = TEST_SHMEM;
		uffd_test_ops = &shmem_uffd_test_ops;
	} else if (!strcmp(type, "shmem-private")) {
		gopts->test_type = TEST_SHMEM;
		uffd_test_ops = &shmem_uffd_test_ops;
	}
}

static void parse_test_type_arg(uffd_global_test_opts_t *gopts, const char *raw_type)
{
	set_test_type(gopts, raw_type);

	if (!gopts->test_type)
		err("failed to parse test type argument: '%s'", raw_type);

	if (gopts->test_type == TEST_HUGETLB)
		gopts->page_size = default_huge_page_size();
	else
		gopts->page_size = sysconf(_SC_PAGE_SIZE);

	if (!gopts->page_size)
		err("Unable to determine page size");
	if ((unsigned long) area_count(NULL, 0, gopts) + sizeof(unsigned long long) * 2
	    > gopts->page_size)
		err("Impossible to run this test");

	/*
	 * Whether we can test certain features depends not just on test type,
	 * but also on whether or not this particular kernel supports the
	 * feature.
	 */

	if (uffd_get_features(&features) && errno == ENOENT)
		ksft_exit_skip("failed to get available features (%d)\n", errno);

	gopts->test_uffdio_wp = gopts->test_uffdio_wp &&
		(features & UFFD_FEATURE_PAGEFAULT_FLAG_WP);

	if (gopts->test_type != TEST_ANON && !(features & UFFD_FEATURE_WP_HUGETLBFS_SHMEM))
		gopts->test_uffdio_wp = false;

	close(gopts->uffd);
	gopts->uffd = -1;
}

static void sigalrm(int sig)
{
	if (sig != SIGALRM)
		abort();
	gopts->test_uffdio_copy_eexist = true;
	alarm(ALARM_INTERVAL_SECS);
}

int main(int argc, char **argv)
{
	unsigned long nr_cpus;
	size_t bytes;

	gopts = (uffd_global_test_opts_t *) malloc(sizeof(uffd_global_test_opts_t));

	if (argc < 4)
		usage();

	ksft_print_header();
	ksft_set_plan(1);

	if (signal(SIGALRM, sigalrm) == SIG_ERR)
		err("failed to arm SIGALRM");
	alarm(ALARM_INTERVAL_SECS);

	parse_test_type_arg(gopts, argv[1]);
	bytes = atol(argv[2]) * 1024 * 1024;

	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
	if (nr_cpus > 32) {
		/* Don't let calculation below go to zero. */
		ksft_print_msg("_SC_NPROCESSORS_ONLN (%lu) too large, capping nr_threads to 32\n",
			       nr_cpus);
		gopts->nr_parallel = 32;
	} else {
		gopts->nr_parallel = nr_cpus;
	}

	/*
	 * src and dst each require bytes / page_size number of hugepages.
	 * Ensure nr_parallel - 1 hugepages on top of that to account
	 * for racy extra reservation of hugepages.
	 */
	if (gopts->test_type == TEST_HUGETLB) {
		unsigned long nr = 2 * (bytes / gopts->page_size) + gopts->nr_parallel - 1;

		if (!hugetlb_setup_default(nr))
			ksft_exit_skip("Skipping userfaultfd... not enough hugepages\n");
	}

	gopts->nr_pages_per_cpu = bytes / gopts->page_size / gopts->nr_parallel;
	if (!gopts->nr_pages_per_cpu) {
		ksft_exit_skip("pages_per_cpu = 0, cannot test (%zu / %lu / %lu)\n",
			       bytes, gopts->page_size, gopts->nr_parallel);
	}

	bounces = atoi(argv[3]);
	if (bounces <= 0) {
		_err("invalid bounces");
		usage();
	}
	gopts->nr_pages = gopts->nr_pages_per_cpu * gopts->nr_parallel;

	ksft_print_msg("nr_pages: %lu, nr_pages_per_cpu: %lu\n",
		       gopts->nr_pages, gopts->nr_pages_per_cpu);

	ksft_test_result(!userfaultfd_stress(gopts),
			 "uffd-stress %s\n", argv[1]);
	ksft_finished();
}

#else /* __NR_userfaultfd */

#warning "missing __NR_userfaultfd definition"

int main(void)
{
	ksft_print_header();
	ksft_exit_skip("missing __NR_userfaultfd definition\n");
}

#endif /* __NR_userfaultfd */