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 
48 static void guest_code(void)
49 {
50 	for (;;)
51 		GUEST_SYNC(0);
52 }
53 
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 
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 
85 static void *migration_worker(void *__rseq_tid)
86 {
87 	pid_t rseq_tid = (pid_t)(unsigned long)__rseq_tid;
88 	cpu_set_t allowed_mask;
89 	int r, i, cpu;
90 
91 	CPU_ZERO(&allowed_mask);
92 
93 	for (i = 0, cpu = min_cpu; i < NR_TASK_MIGRATIONS; i++, cpu = next_cpu(cpu)) {
94 		CPU_SET(cpu, &allowed_mask);
95 
96 		/*
97 		 * Bump the sequence count twice to allow the reader to detect
98 		 * that a migration may have occurred in between rseq and sched
99 		 * CPU ID reads.  An odd sequence count indicates a migration
100 		 * is in-progress, while a completely different count indicates
101 		 * a migration occurred since the count was last read.
102 		 */
103 		atomic_inc(&seq_cnt);
104 
105 		/*
106 		 * Ensure the odd count is visible while sched_getcpu() isn't
107 		 * stable, i.e. while changing affinity is in-progress.
108 		 */
109 		smp_wmb();
110 		r = sched_setaffinity(rseq_tid, sizeof(allowed_mask), &allowed_mask);
111 		TEST_ASSERT(!r, "sched_setaffinity failed, errno = %d (%s)",
112 			    errno, strerror(errno));
113 		smp_wmb();
114 		atomic_inc(&seq_cnt);
115 
116 		CPU_CLR(cpu, &allowed_mask);
117 
118 		/*
119 		 * Wait 1-10us before proceeding to the next iteration and more
120 		 * specifically, before bumping seq_cnt again.  A delay is
121 		 * needed on three fronts:
122 		 *
123 		 *  1. To allow sched_setaffinity() to prompt migration before
124 		 *     ioctl(KVM_RUN) enters the guest so that TIF_NOTIFY_RESUME
125 		 *     (or TIF_NEED_RESCHED, which indirectly leads to handling
126 		 *     NOTIFY_RESUME) is handled in KVM context.
127 		 *
128 		 *     If NOTIFY_RESUME/NEED_RESCHED is set after KVM enters
129 		 *     the guest, the guest will trigger a IO/MMIO exit all the
130 		 *     way to userspace and the TIF flags will be handled by
131 		 *     the generic "exit to userspace" logic, not by KVM.  The
132 		 *     exit to userspace is necessary to give the test a chance
133 		 *     to check the rseq CPU ID (see #2).
134 		 *
135 		 *     Alternatively, guest_code() could include an instruction
136 		 *     to trigger an exit that is handled by KVM, but any such
137 		 *     exit requires architecture specific code.
138 		 *
139 		 *  2. To let ioctl(KVM_RUN) make its way back to the test
140 		 *     before the next round of migration.  The test's check on
141 		 *     the rseq CPU ID must wait for migration to complete in
142 		 *     order to avoid false positive, thus any kernel rseq bug
143 		 *     will be missed if the next migration starts before the
144 		 *     check completes.
145 		 *
146 		 *  3. To ensure the read-side makes efficient forward progress,
147 		 *     e.g. if sched_getcpu() involves a syscall.  Stalling the
148 		 *     read-side means the test will spend more time waiting for
149 		 *     sched_getcpu() to stabilize and less time trying to hit
150 		 *     the timing-dependent bug.
151 		 *
152 		 * Because any bug in this area is likely to be timing-dependent,
153 		 * run with a range of delays at 1us intervals from 1us to 10us
154 		 * as a best effort to avoid tuning the test to the point where
155 		 * it can hit _only_ the original bug and not detect future
156 		 * regressions.
157 		 *
158 		 * The original bug can reproduce with a delay up to ~500us on
159 		 * x86-64, but starts to require more iterations to reproduce
160 		 * as the delay creeps above ~10us, and the average runtime of
161 		 * each iteration obviously increases as well.  Cap the delay
162 		 * at 10us to keep test runtime reasonable while minimizing
163 		 * potential coverage loss.
164 		 *
165 		 * The lower bound for reproducing the bug is likely below 1us,
166 		 * e.g. failures occur on x86-64 with nanosleep(0), but at that
167 		 * point the overhead of the syscall likely dominates the delay.
168 		 * Use usleep() for simplicity and to avoid unnecessary kernel
169 		 * dependencies.
170 		 */
171 		usleep((i % 10) + 1);
172 	}
173 	done = true;
174 	return NULL;
175 }
176 
177 static int calc_min_max_cpu(void)
178 {
179 	int i, cnt, nproc;
180 
181 	if (CPU_COUNT(&possible_mask) < 2)
182 		return -EINVAL;
183 
184 	/*
185 	 * CPU_SET doesn't provide a FOR_EACH helper, get the min/max CPU that
186 	 * this task is affined to in order to reduce the time spent querying
187 	 * unusable CPUs, e.g. if this task is pinned to a small percentage of
188 	 * total CPUs.
189 	 */
190 	nproc = get_nprocs_conf();
191 	min_cpu = -1;
192 	max_cpu = -1;
193 	cnt = 0;
194 
195 	for (i = 0; i < nproc; i++) {
196 		if (!CPU_ISSET(i, &possible_mask))
197 			continue;
198 		if (min_cpu == -1)
199 			min_cpu = i;
200 		max_cpu = i;
201 		cnt++;
202 	}
203 
204 	return (cnt < 2) ? -EINVAL : 0;
205 }
206 
207 int main(int argc, char *argv[])
208 {
209 	int r, i, snapshot;
210 	struct kvm_vm *vm;
211 	u32 cpu, rseq_cpu;
212 
213 	/* Tell stdout not to buffer its content */
214 	setbuf(stdout, NULL);
215 
216 	r = sched_getaffinity(0, sizeof(possible_mask), &possible_mask);
217 	TEST_ASSERT(!r, "sched_getaffinity failed, errno = %d (%s)", errno,
218 		    strerror(errno));
219 
220 	if (calc_min_max_cpu()) {
221 		print_skip("Only one usable CPU, task migration not possible");
222 		exit(KSFT_SKIP);
223 	}
224 
225 	sys_rseq(0);
226 
227 	/*
228 	 * Create and run a dummy VM that immediately exits to userspace via
229 	 * GUEST_SYNC, while concurrently migrating the process by setting its
230 	 * CPU affinity.
231 	 */
232 	vm = vm_create_default(VCPU_ID, 0, guest_code);
233 	ucall_init(vm, NULL);
234 
235 	pthread_create(&migration_thread, NULL, migration_worker,
236 		       (void *)(unsigned long)gettid());
237 
238 	for (i = 0; !done; i++) {
239 		vcpu_run(vm, VCPU_ID);
240 		TEST_ASSERT(get_ucall(vm, VCPU_ID, NULL) == UCALL_SYNC,
241 			    "Guest failed?");
242 
243 		/*
244 		 * Verify rseq's CPU matches sched's CPU.  Ensure migration
245 		 * doesn't occur between sched_getcpu() and reading the rseq
246 		 * cpu_id by rereading both if the sequence count changes, or
247 		 * if the count is odd (migration in-progress).
248 		 */
249 		do {
250 			/*
251 			 * Drop bit 0 to force a mismatch if the count is odd,
252 			 * i.e. if a migration is in-progress.
253 			 */
254 			snapshot = atomic_read(&seq_cnt) & ~1;
255 
256 			/*
257 			 * Ensure reading sched_getcpu() and rseq.cpu_id
258 			 * complete in a single "no migration" window, i.e. are
259 			 * not reordered across the seq_cnt reads.
260 			 */
261 			smp_rmb();
262 			cpu = sched_getcpu();
263 			rseq_cpu = READ_ONCE(__rseq.cpu_id);
264 			smp_rmb();
265 		} while (snapshot != atomic_read(&seq_cnt));
266 
267 		TEST_ASSERT(rseq_cpu == cpu,
268 			    "rseq CPU = %d, sched CPU = %d\n", rseq_cpu, cpu);
269 	}
270 
271 	/*
272 	 * Sanity check that the test was able to enter the guest a reasonable
273 	 * number of times, e.g. didn't get stalled too often/long waiting for
274 	 * sched_getcpu() to stabilize.  A 2:1 migration:KVM_RUN ratio is a
275 	 * fairly conservative ratio on x86-64, which can do _more_ KVM_RUNs
276 	 * than migrations given the 1us+ delay in the migration task.
277 	 */
278 	TEST_ASSERT(i > (NR_TASK_MIGRATIONS / 2),
279 		    "Only performed %d KVM_RUNs, task stalled too much?\n", i);
280 
281 	pthread_join(migration_thread, NULL);
282 
283 	kvm_vm_free(vm);
284 
285 	sys_rseq(RSEQ_FLAG_UNREGISTER);
286 
287 	return 0;
288 }
289