summaryrefslogtreecommitdiff
path: root/tools/testing/selftests/kvm/rseq_test.c
blob: 4158da0da2bba8e5e8ea12acd87f6cff6206e3ae (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
// SPDX-License-Identifier: GPL-2.0-only
#define _GNU_SOURCE /* for program_invocation_short_name */
#include <errno.h>
#include <fcntl.h>
#include <pthread.h>
#include <sched.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <signal.h>
#include <syscall.h>
#include <sys/ioctl.h>
#include <sys/sysinfo.h>
#include <asm/barrier.h>
#include <linux/atomic.h>
#include <linux/rseq.h>
#include <linux/unistd.h>

#include "kvm_util.h"
#include "processor.h"
#include "test_util.h"

#define VCPU_ID 0

static __thread volatile struct rseq __rseq = {
	.cpu_id = RSEQ_CPU_ID_UNINITIALIZED,
};

/*
 * Use an arbitrary, bogus signature for configuring rseq, this test does not
 * actually enter an rseq critical section.
 */
#define RSEQ_SIG 0xdeadbeef

/*
 * Any bug related to task migration is likely to be timing-dependent; perform
 * a large number of migrations to reduce the odds of a false negative.
 */
#define NR_TASK_MIGRATIONS 100000

static pthread_t migration_thread;
static cpu_set_t possible_mask;
static int min_cpu, max_cpu;
static bool done;

static atomic_t seq_cnt;

static void guest_code(void)
{
	for (;;)
		GUEST_SYNC(0);
}

static void sys_rseq(int flags)
{
	int r;

	r = syscall(__NR_rseq, &__rseq, sizeof(__rseq), flags, RSEQ_SIG);
	TEST_ASSERT(!r, "rseq failed, errno = %d (%s)", errno, strerror(errno));
}

static int next_cpu(int cpu)
{
	/*
	 * Advance to the next CPU, skipping those that weren't in the original
	 * affinity set.  Sadly, there is no CPU_SET_FOR_EACH, and cpu_set_t's
	 * data storage is considered as opaque.  Note, if this task is pinned
	 * to a small set of discontigous CPUs, e.g. 2 and 1023, this loop will
	 * burn a lot cycles and the test will take longer than normal to
	 * complete.
	 */
	do {
		cpu++;
		if (cpu > max_cpu) {
			cpu = min_cpu;
			TEST_ASSERT(CPU_ISSET(cpu, &possible_mask),
				    "Min CPU = %d must always be usable", cpu);
			break;
		}
	} while (!CPU_ISSET(cpu, &possible_mask));

	return cpu;
}

static void *migration_worker(void *ign)
{
	cpu_set_t allowed_mask;
	int r, i, cpu;

	CPU_ZERO(&allowed_mask);

	for (i = 0, cpu = min_cpu; i < NR_TASK_MIGRATIONS; i++, cpu = next_cpu(cpu)) {
		CPU_SET(cpu, &allowed_mask);

		/*
		 * Bump the sequence count twice to allow the reader to detect
		 * that a migration may have occurred in between rseq and sched
		 * CPU ID reads.  An odd sequence count indicates a migration
		 * is in-progress, while a completely different count indicates
		 * a migration occurred since the count was last read.
		 */
		atomic_inc(&seq_cnt);

		/*
		 * Ensure the odd count is visible while sched_getcpu() isn't
		 * stable, i.e. while changing affinity is in-progress.
		 */
		smp_wmb();
		r = sched_setaffinity(0, sizeof(allowed_mask), &allowed_mask);
		TEST_ASSERT(!r, "sched_setaffinity failed, errno = %d (%s)",
			    errno, strerror(errno));
		smp_wmb();
		atomic_inc(&seq_cnt);

		CPU_CLR(cpu, &allowed_mask);

		/*
		 * Wait 1-10us before proceeding to the next iteration and more
		 * specifically, before bumping seq_cnt again.  A delay is
		 * needed on three fronts:
		 *
		 *  1. To allow sched_setaffinity() to prompt migration before
		 *     ioctl(KVM_RUN) enters the guest so that TIF_NOTIFY_RESUME
		 *     (or TIF_NEED_RESCHED, which indirectly leads to handling
		 *     NOTIFY_RESUME) is handled in KVM context.
		 *
		 *     If NOTIFY_RESUME/NEED_RESCHED is set after KVM enters
		 *     the guest, the guest will trigger a IO/MMIO exit all the
		 *     way to userspace and the TIF flags will be handled by
		 *     the generic "exit to userspace" logic, not by KVM.  The
		 *     exit to userspace is necessary to give the test a chance
		 *     to check the rseq CPU ID (see #2).
		 *
		 *     Alternatively, guest_code() could include an instruction
		 *     to trigger an exit that is handled by KVM, but any such
		 *     exit requires architecture specific code.
		 *
		 *  2. To let ioctl(KVM_RUN) make its way back to the test
		 *     before the next round of migration.  The test's check on
		 *     the rseq CPU ID must wait for migration to complete in
		 *     order to avoid false positive, thus any kernel rseq bug
		 *     will be missed if the next migration starts before the
		 *     check completes.
		 *
		 *  3. To ensure the read-side makes efficient forward progress,
		 *     e.g. if sched_getcpu() involves a syscall.  Stalling the
		 *     read-side means the test will spend more time waiting for
		 *     sched_getcpu() to stabilize and less time trying to hit
		 *     the timing-dependent bug.
		 *
		 * Because any bug in this area is likely to be timing-dependent,
		 * run with a range of delays at 1us intervals from 1us to 10us
		 * as a best effort to avoid tuning the test to the point where
		 * it can hit _only_ the original bug and not detect future
		 * regressions.
		 *
		 * The original bug can reproduce with a delay up to ~500us on
		 * x86-64, but starts to require more iterations to reproduce
		 * as the delay creeps above ~10us, and the average runtime of
		 * each iteration obviously increases as well.  Cap the delay
		 * at 10us to keep test runtime reasonable while minimizing
		 * potential coverage loss.
		 *
		 * The lower bound for reproducing the bug is likely below 1us,
		 * e.g. failures occur on x86-64 with nanosleep(0), but at that
		 * point the overhead of the syscall likely dominates the delay.
		 * Use usleep() for simplicity and to avoid unnecessary kernel
		 * dependencies.
		 */
		usleep((i % 10) + 1);
	}
	done = true;
	return NULL;
}

static int calc_min_max_cpu(void)
{
	int i, cnt, nproc;

	if (CPU_COUNT(&possible_mask) < 2)
		return -EINVAL;

	/*
	 * CPU_SET doesn't provide a FOR_EACH helper, get the min/max CPU that
	 * this task is affined to in order to reduce the time spent querying
	 * unusable CPUs, e.g. if this task is pinned to a small percentage of
	 * total CPUs.
	 */
	nproc = get_nprocs_conf();
	min_cpu = -1;
	max_cpu = -1;
	cnt = 0;

	for (i = 0; i < nproc; i++) {
		if (!CPU_ISSET(i, &possible_mask))
			continue;
		if (min_cpu == -1)
			min_cpu = i;
		max_cpu = i;
		cnt++;
	}

	return (cnt < 2) ? -EINVAL : 0;
}

int main(int argc, char *argv[])
{
	int r, i, snapshot;
	struct kvm_vm *vm;
	u32 cpu, rseq_cpu;

	/* Tell stdout not to buffer its content */
	setbuf(stdout, NULL);

	r = sched_getaffinity(0, sizeof(possible_mask), &possible_mask);
	TEST_ASSERT(!r, "sched_getaffinity failed, errno = %d (%s)", errno,
		    strerror(errno));

	if (calc_min_max_cpu()) {
		print_skip("Only one usable CPU, task migration not possible");
		exit(KSFT_SKIP);
	}

	sys_rseq(0);

	/*
	 * Create and run a dummy VM that immediately exits to userspace via
	 * GUEST_SYNC, while concurrently migrating the process by setting its
	 * CPU affinity.
	 */
	vm = vm_create_default(VCPU_ID, 0, guest_code);
	ucall_init(vm, NULL);

	pthread_create(&migration_thread, NULL, migration_worker, 0);

	for (i = 0; !done; i++) {
		vcpu_run(vm, VCPU_ID);
		TEST_ASSERT(get_ucall(vm, VCPU_ID, NULL) == UCALL_SYNC,
			    "Guest failed?");

		/*
		 * Verify rseq's CPU matches sched's CPU.  Ensure migration
		 * doesn't occur between sched_getcpu() and reading the rseq
		 * cpu_id by rereading both if the sequence count changes, or
		 * if the count is odd (migration in-progress).
		 */
		do {
			/*
			 * Drop bit 0 to force a mismatch if the count is odd,
			 * i.e. if a migration is in-progress.
			 */
			snapshot = atomic_read(&seq_cnt) & ~1;

			/*
			 * Ensure reading sched_getcpu() and rseq.cpu_id
			 * complete in a single "no migration" window, i.e. are
			 * not reordered across the seq_cnt reads.
			 */
			smp_rmb();
			cpu = sched_getcpu();
			rseq_cpu = READ_ONCE(__rseq.cpu_id);
			smp_rmb();
		} while (snapshot != atomic_read(&seq_cnt));

		TEST_ASSERT(rseq_cpu == cpu,
			    "rseq CPU = %d, sched CPU = %d\n", rseq_cpu, cpu);
	}

	/*
	 * Sanity check that the test was able to enter the guest a reasonable
	 * number of times, e.g. didn't get stalled too often/long waiting for
	 * sched_getcpu() to stabilize.  A 2:1 migration:KVM_RUN ratio is a
	 * fairly conservative ratio on x86-64, which can do _more_ KVM_RUNs
	 * than migrations given the 1us+ delay in the migration task.
	 */
	TEST_ASSERT(i > (NR_TASK_MIGRATIONS / 2),
		    "Only performed %d KVM_RUNs, task stalled too much?\n", i);

	pthread_join(migration_thread, NULL);

	kvm_vm_free(vm);

	sys_rseq(RSEQ_FLAG_UNREGISTER);

	return 0;
}