selftests/perf_events: Add kselftest for remove_on_exec
authorMarco Elver <elver@google.com>
Thu, 8 Apr 2021 10:36:03 +0000 (12:36 +0200)
committerPeter Zijlstra <peterz@infradead.org>
Fri, 16 Apr 2021 14:32:42 +0000 (16:32 +0200)
Add kselftest to test that remove_on_exec removes inherited events from
child tasks.

Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20210408103605.1676875-9-elver@google.com
tools/testing/selftests/perf_events/.gitignore
tools/testing/selftests/perf_events/Makefile
tools/testing/selftests/perf_events/remove_on_exec.c [new file with mode: 0644]

index 973a2c3..fcafa5f 100644 (file)
@@ -2,5 +2,5 @@
 CFLAGS += -Wl,-no-as-needed -Wall -I../../../../usr/include
 LDFLAGS += -lpthread
 
-TEST_GEN_PROGS := sigtrap_threads
+TEST_GEN_PROGS := sigtrap_threads remove_on_exec
 include ../lib.mk
diff --git a/tools/testing/selftests/perf_events/remove_on_exec.c b/tools/testing/selftests/perf_events/remove_on_exec.c
new file mode 100644 (file)
index 0000000..5814611
--- /dev/null
@@ -0,0 +1,260 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Test for remove_on_exec.
+ *
+ * Copyright (C) 2021, Google LLC.
+ */
+
+#define _GNU_SOURCE
+
+/* We need the latest siginfo from the kernel repo. */
+#include <sys/types.h>
+#include <asm/siginfo.h>
+#define __have_siginfo_t 1
+#define __have_sigval_t 1
+#define __have_sigevent_t 1
+#define __siginfo_t_defined
+#define __sigval_t_defined
+#define __sigevent_t_defined
+#define _BITS_SIGINFO_CONSTS_H 1
+#define _BITS_SIGEVENT_CONSTS_H 1
+
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <linux/perf_event.h>
+#include <pthread.h>
+#include <signal.h>
+#include <sys/ioctl.h>
+#include <sys/syscall.h>
+#include <unistd.h>
+
+#include "../kselftest_harness.h"
+
+static volatile int signal_count;
+
+static struct perf_event_attr make_event_attr(void)
+{
+       struct perf_event_attr attr = {
+               .type           = PERF_TYPE_HARDWARE,
+               .size           = sizeof(attr),
+               .config         = PERF_COUNT_HW_INSTRUCTIONS,
+               .sample_period  = 1000,
+               .exclude_kernel = 1,
+               .exclude_hv     = 1,
+               .disabled       = 1,
+               .inherit        = 1,
+               /*
+                * Children normally retain their inherited event on exec; with
+                * remove_on_exec, we'll remove their event, but the parent and
+                * any other non-exec'd children will keep their events.
+                */
+               .remove_on_exec = 1,
+               .sigtrap        = 1,
+       };
+       return attr;
+}
+
+static void sigtrap_handler(int signum, siginfo_t *info, void *ucontext)
+{
+       if (info->si_code != TRAP_PERF) {
+               fprintf(stderr, "%s: unexpected si_code %d\n", __func__, info->si_code);
+               return;
+       }
+
+       signal_count++;
+}
+
+FIXTURE(remove_on_exec)
+{
+       struct sigaction oldact;
+       int fd;
+};
+
+FIXTURE_SETUP(remove_on_exec)
+{
+       struct perf_event_attr attr = make_event_attr();
+       struct sigaction action = {};
+
+       signal_count = 0;
+
+       /* Initialize sigtrap handler. */
+       action.sa_flags = SA_SIGINFO | SA_NODEFER;
+       action.sa_sigaction = sigtrap_handler;
+       sigemptyset(&action.sa_mask);
+       ASSERT_EQ(sigaction(SIGTRAP, &action, &self->oldact), 0);
+
+       /* Initialize perf event. */
+       self->fd = syscall(__NR_perf_event_open, &attr, 0, -1, -1, PERF_FLAG_FD_CLOEXEC);
+       ASSERT_NE(self->fd, -1);
+}
+
+FIXTURE_TEARDOWN(remove_on_exec)
+{
+       close(self->fd);
+       sigaction(SIGTRAP, &self->oldact, NULL);
+}
+
+/* Verify event propagates to fork'd child. */
+TEST_F(remove_on_exec, fork_only)
+{
+       int status;
+       pid_t pid = fork();
+
+       if (pid == 0) {
+               ASSERT_EQ(signal_count, 0);
+               ASSERT_EQ(ioctl(self->fd, PERF_EVENT_IOC_ENABLE, 0), 0);
+               while (!signal_count);
+               _exit(42);
+       }
+
+       while (!signal_count); /* Child enables event. */
+       EXPECT_EQ(waitpid(pid, &status, 0), pid);
+       EXPECT_EQ(WEXITSTATUS(status), 42);
+}
+
+/*
+ * Verify that event does _not_ propagate to fork+exec'd child; event enabled
+ * after fork+exec.
+ */
+TEST_F(remove_on_exec, fork_exec_then_enable)
+{
+       pid_t pid_exec, pid_only_fork;
+       int pipefd[2];
+       int tmp;
+
+       /*
+        * Non-exec child, to ensure exec does not affect inherited events of
+        * other children.
+        */
+       pid_only_fork = fork();
+       if (pid_only_fork == 0) {
+               /* Block until parent enables event. */
+               while (!signal_count);
+               _exit(42);
+       }
+
+       ASSERT_NE(pipe(pipefd), -1);
+       pid_exec = fork();
+       if (pid_exec == 0) {
+               ASSERT_NE(dup2(pipefd[1], STDOUT_FILENO), -1);
+               close(pipefd[0]);
+               execl("/proc/self/exe", "exec_child", NULL);
+               _exit((perror("exec failed"), 1));
+       }
+       close(pipefd[1]);
+
+       ASSERT_EQ(waitpid(pid_exec, &tmp, WNOHANG), 0); /* Child is running. */
+       /* Wait for exec'd child to start spinning. */
+       EXPECT_EQ(read(pipefd[0], &tmp, sizeof(int)), sizeof(int));
+       EXPECT_EQ(tmp, 42);
+       close(pipefd[0]);
+       /* Now we can enable the event, knowing the child is doing work. */
+       EXPECT_EQ(ioctl(self->fd, PERF_EVENT_IOC_ENABLE, 0), 0);
+       /* If the event propagated to the exec'd child, it will exit normally... */
+       usleep(100000); /* ... give time for event to trigger (in case of bug). */
+       EXPECT_EQ(waitpid(pid_exec, &tmp, WNOHANG), 0); /* Should still be running. */
+       EXPECT_EQ(kill(pid_exec, SIGKILL), 0);
+
+       /* Verify removal from child did not affect this task's event. */
+       tmp = signal_count;
+       while (signal_count == tmp); /* Should not hang! */
+       /* Nor should it have affected the first child. */
+       EXPECT_EQ(waitpid(pid_only_fork, &tmp, 0), pid_only_fork);
+       EXPECT_EQ(WEXITSTATUS(tmp), 42);
+}
+
+/*
+ * Verify that event does _not_ propagate to fork+exec'd child; event enabled
+ * before fork+exec.
+ */
+TEST_F(remove_on_exec, enable_then_fork_exec)
+{
+       pid_t pid_exec;
+       int tmp;
+
+       EXPECT_EQ(ioctl(self->fd, PERF_EVENT_IOC_ENABLE, 0), 0);
+
+       pid_exec = fork();
+       if (pid_exec == 0) {
+               execl("/proc/self/exe", "exec_child", NULL);
+               _exit((perror("exec failed"), 1));
+       }
+
+       /*
+        * The child may exit abnormally at any time if the event propagated and
+        * a SIGTRAP is sent before the handler was set up.
+        */
+       usleep(100000); /* ... give time for event to trigger (in case of bug). */
+       EXPECT_EQ(waitpid(pid_exec, &tmp, WNOHANG), 0); /* Should still be running. */
+       EXPECT_EQ(kill(pid_exec, SIGKILL), 0);
+
+       /* Verify removal from child did not affect this task's event. */
+       tmp = signal_count;
+       while (signal_count == tmp); /* Should not hang! */
+}
+
+TEST_F(remove_on_exec, exec_stress)
+{
+       pid_t pids[30];
+       int i, tmp;
+
+       for (i = 0; i < sizeof(pids) / sizeof(pids[0]); i++) {
+               pids[i] = fork();
+               if (pids[i] == 0) {
+                       execl("/proc/self/exe", "exec_child", NULL);
+                       _exit((perror("exec failed"), 1));
+               }
+
+               /* Some forked with event disabled, rest with enabled. */
+               if (i > 10)
+                       EXPECT_EQ(ioctl(self->fd, PERF_EVENT_IOC_ENABLE, 0), 0);
+       }
+
+       usleep(100000); /* ... give time for event to trigger (in case of bug). */
+
+       for (i = 0; i < sizeof(pids) / sizeof(pids[0]); i++) {
+               /* All children should still be running. */
+               EXPECT_EQ(waitpid(pids[i], &tmp, WNOHANG), 0);
+               EXPECT_EQ(kill(pids[i], SIGKILL), 0);
+       }
+
+       /* Verify event is still alive. */
+       tmp = signal_count;
+       while (signal_count == tmp);
+}
+
+/* For exec'd child. */
+static void exec_child(void)
+{
+       struct sigaction action = {};
+       const int val = 42;
+
+       /* Set up sigtrap handler in case we erroneously receive a trap. */
+       action.sa_flags = SA_SIGINFO | SA_NODEFER;
+       action.sa_sigaction = sigtrap_handler;
+       sigemptyset(&action.sa_mask);
+       if (sigaction(SIGTRAP, &action, NULL))
+               _exit((perror("sigaction failed"), 1));
+
+       /* Signal parent that we're starting to spin. */
+       if (write(STDOUT_FILENO, &val, sizeof(int)) == -1)
+               _exit((perror("write failed"), 1));
+
+       /* Should hang here until killed. */
+       while (!signal_count);
+}
+
+#define main test_main
+TEST_HARNESS_MAIN
+#undef main
+int main(int argc, char *argv[])
+{
+       if (!strcmp(argv[0], "exec_child")) {
+               exec_child();
+               return 1;
+       }
+
+       return test_main(argc, argv);
+}