1 // SPDX-License-Identifier: GPL-2.0-only
2 #define _GNU_SOURCE /* for program_invocation_short_name */
3 #include <errno.h>
4 #include <fcntl.h>
5 #include <pthread.h>
6 #include <sched.h>
7 #include <stdio.h>
8 #include <stdlib.h>
9 #include <string.h>
10 #include <signal.h>
11 #include <syscall.h>
12 #include <sys/ioctl.h>
13 #include <sys/sysinfo.h>
14 #include <asm/barrier.h>
15 #include <linux/atomic.h>
16 #include <linux/rseq.h>
17 #include <linux/unistd.h>
18
19 #include "kvm_util.h"
20 #include "processor.h"
21 #include "test_util.h"
22
23 #define VCPU_ID 0
24
25 static __thread volatile struct rseq __rseq = {
26 .cpu_id = RSEQ_CPU_ID_UNINITIALIZED,
27 };
28
29 /*
30 * Use an arbitrary, bogus signature for configuring rseq, this test does not
31 * actually enter an rseq critical section.
32 */
33 #define RSEQ_SIG 0xdeadbeef
34
35 /*
36 * Any bug related to task migration is likely to be timing-dependent; perform
37 * a large number of migrations to reduce the odds of a false negative.
38 */
39 #define NR_TASK_MIGRATIONS 100000
40
41 static pthread_t migration_thread;
42 static cpu_set_t possible_mask;
43 static int min_cpu, max_cpu;
44 static bool done;
45
46 static atomic_t seq_cnt;
47
guest_code(void)48 static void guest_code(void)
49 {
50 for (;;)
51 GUEST_SYNC(0);
52 }
53
sys_rseq(int flags)54 static void sys_rseq(int flags)
55 {
56 int r;
57
58 r = syscall(__NR_rseq, &__rseq, sizeof(__rseq), flags, RSEQ_SIG);
59 TEST_ASSERT(!r, "rseq failed, errno = %d (%s)", errno, strerror(errno));
60 }
61
next_cpu(int cpu)62 static int next_cpu(int cpu)
63 {
64 /*
65 * Advance to the next CPU, skipping those that weren't in the original
66 * affinity set. Sadly, there is no CPU_SET_FOR_EACH, and cpu_set_t's
67 * data storage is considered as opaque. Note, if this task is pinned
68 * to a small set of discontigous CPUs, e.g. 2 and 1023, this loop will
69 * burn a lot cycles and the test will take longer than normal to
70 * complete.
71 */
72 do {
73 cpu++;
74 if (cpu > max_cpu) {
75 cpu = min_cpu;
76 TEST_ASSERT(CPU_ISSET(cpu, &possible_mask),
77 "Min CPU = %d must always be usable", cpu);
78 break;
79 }
80 } while (!CPU_ISSET(cpu, &possible_mask));
81
82 return cpu;
83 }
84
migration_worker(void * ign)85 static void *migration_worker(void *ign)
86 {
87 cpu_set_t allowed_mask;
88 int r, i, cpu;
89
90 CPU_ZERO(&allowed_mask);
91
92 for (i = 0, cpu = min_cpu; i < NR_TASK_MIGRATIONS; i++, cpu = next_cpu(cpu)) {
93 CPU_SET(cpu, &allowed_mask);
94
95 /*
96 * Bump the sequence count twice to allow the reader to detect
97 * that a migration may have occurred in between rseq and sched
98 * CPU ID reads. An odd sequence count indicates a migration
99 * is in-progress, while a completely different count indicates
100 * a migration occurred since the count was last read.
101 */
102 atomic_inc(&seq_cnt);
103
104 /*
105 * Ensure the odd count is visible while sched_getcpu() isn't
106 * stable, i.e. while changing affinity is in-progress.
107 */
108 smp_wmb();
109 r = sched_setaffinity(0, sizeof(allowed_mask), &allowed_mask);
110 TEST_ASSERT(!r, "sched_setaffinity failed, errno = %d (%s)",
111 errno, strerror(errno));
112 smp_wmb();
113 atomic_inc(&seq_cnt);
114
115 CPU_CLR(cpu, &allowed_mask);
116
117 /*
118 * Wait 1-10us before proceeding to the next iteration and more
119 * specifically, before bumping seq_cnt again. A delay is
120 * needed on three fronts:
121 *
122 * 1. To allow sched_setaffinity() to prompt migration before
123 * ioctl(KVM_RUN) enters the guest so that TIF_NOTIFY_RESUME
124 * (or TIF_NEED_RESCHED, which indirectly leads to handling
125 * NOTIFY_RESUME) is handled in KVM context.
126 *
127 * If NOTIFY_RESUME/NEED_RESCHED is set after KVM enters
128 * the guest, the guest will trigger a IO/MMIO exit all the
129 * way to userspace and the TIF flags will be handled by
130 * the generic "exit to userspace" logic, not by KVM. The
131 * exit to userspace is necessary to give the test a chance
132 * to check the rseq CPU ID (see #2).
133 *
134 * Alternatively, guest_code() could include an instruction
135 * to trigger an exit that is handled by KVM, but any such
136 * exit requires architecture specific code.
137 *
138 * 2. To let ioctl(KVM_RUN) make its way back to the test
139 * before the next round of migration. The test's check on
140 * the rseq CPU ID must wait for migration to complete in
141 * order to avoid false positive, thus any kernel rseq bug
142 * will be missed if the next migration starts before the
143 * check completes.
144 *
145 * 3. To ensure the read-side makes efficient forward progress,
146 * e.g. if sched_getcpu() involves a syscall. Stalling the
147 * read-side means the test will spend more time waiting for
148 * sched_getcpu() to stabilize and less time trying to hit
149 * the timing-dependent bug.
150 *
151 * Because any bug in this area is likely to be timing-dependent,
152 * run with a range of delays at 1us intervals from 1us to 10us
153 * as a best effort to avoid tuning the test to the point where
154 * it can hit _only_ the original bug and not detect future
155 * regressions.
156 *
157 * The original bug can reproduce with a delay up to ~500us on
158 * x86-64, but starts to require more iterations to reproduce
159 * as the delay creeps above ~10us, and the average runtime of
160 * each iteration obviously increases as well. Cap the delay
161 * at 10us to keep test runtime reasonable while minimizing
162 * potential coverage loss.
163 *
164 * The lower bound for reproducing the bug is likely below 1us,
165 * e.g. failures occur on x86-64 with nanosleep(0), but at that
166 * point the overhead of the syscall likely dominates the delay.
167 * Use usleep() for simplicity and to avoid unnecessary kernel
168 * dependencies.
169 */
170 usleep((i % 10) + 1);
171 }
172 done = true;
173 return NULL;
174 }
175
calc_min_max_cpu(void)176 static int calc_min_max_cpu(void)
177 {
178 int i, cnt, nproc;
179
180 if (CPU_COUNT(&possible_mask) < 2)
181 return -EINVAL;
182
183 /*
184 * CPU_SET doesn't provide a FOR_EACH helper, get the min/max CPU that
185 * this task is affined to in order to reduce the time spent querying
186 * unusable CPUs, e.g. if this task is pinned to a small percentage of
187 * total CPUs.
188 */
189 nproc = get_nprocs_conf();
190 min_cpu = -1;
191 max_cpu = -1;
192 cnt = 0;
193
194 for (i = 0; i < nproc; i++) {
195 if (!CPU_ISSET(i, &possible_mask))
196 continue;
197 if (min_cpu == -1)
198 min_cpu = i;
199 max_cpu = i;
200 cnt++;
201 }
202
203 return (cnt < 2) ? -EINVAL : 0;
204 }
205
main(int argc,char * argv[])206 int main(int argc, char *argv[])
207 {
208 int r, i, snapshot;
209 struct kvm_vm *vm;
210 u32 cpu, rseq_cpu;
211
212 /* Tell stdout not to buffer its content */
213 setbuf(stdout, NULL);
214
215 r = sched_getaffinity(0, sizeof(possible_mask), &possible_mask);
216 TEST_ASSERT(!r, "sched_getaffinity failed, errno = %d (%s)", errno,
217 strerror(errno));
218
219 if (calc_min_max_cpu()) {
220 print_skip("Only one usable CPU, task migration not possible");
221 exit(KSFT_SKIP);
222 }
223
224 sys_rseq(0);
225
226 /*
227 * Create and run a dummy VM that immediately exits to userspace via
228 * GUEST_SYNC, while concurrently migrating the process by setting its
229 * CPU affinity.
230 */
231 vm = vm_create_default(VCPU_ID, 0, guest_code);
232 ucall_init(vm, NULL);
233
234 pthread_create(&migration_thread, NULL, migration_worker, 0);
235
236 for (i = 0; !done; i++) {
237 vcpu_run(vm, VCPU_ID);
238 TEST_ASSERT(get_ucall(vm, VCPU_ID, NULL) == UCALL_SYNC,
239 "Guest failed?");
240
241 /*
242 * Verify rseq's CPU matches sched's CPU. Ensure migration
243 * doesn't occur between sched_getcpu() and reading the rseq
244 * cpu_id by rereading both if the sequence count changes, or
245 * if the count is odd (migration in-progress).
246 */
247 do {
248 /*
249 * Drop bit 0 to force a mismatch if the count is odd,
250 * i.e. if a migration is in-progress.
251 */
252 snapshot = atomic_read(&seq_cnt) & ~1;
253
254 /*
255 * Ensure reading sched_getcpu() and rseq.cpu_id
256 * complete in a single "no migration" window, i.e. are
257 * not reordered across the seq_cnt reads.
258 */
259 smp_rmb();
260 cpu = sched_getcpu();
261 rseq_cpu = READ_ONCE(__rseq.cpu_id);
262 smp_rmb();
263 } while (snapshot != atomic_read(&seq_cnt));
264
265 TEST_ASSERT(rseq_cpu == cpu,
266 "rseq CPU = %d, sched CPU = %d\n", rseq_cpu, cpu);
267 }
268
269 /*
270 * Sanity check that the test was able to enter the guest a reasonable
271 * number of times, e.g. didn't get stalled too often/long waiting for
272 * sched_getcpu() to stabilize. A 2:1 migration:KVM_RUN ratio is a
273 * fairly conservative ratio on x86-64, which can do _more_ KVM_RUNs
274 * than migrations given the 1us+ delay in the migration task.
275 */
276 TEST_ASSERT(i > (NR_TASK_MIGRATIONS / 2),
277 "Only performed %d KVM_RUNs, task stalled too much?\n", i);
278
279 pthread_join(migration_thread, NULL);
280
281 kvm_vm_free(vm);
282
283 sys_rseq(RSEQ_FLAG_UNREGISTER);
284
285 return 0;
286 }
287