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