[PATCH] selftests/cgroup: Don't count cgroup.threads in test_cgcore_proc_migration

[PATCH] selftests/cgroup: Don't count cgroup.threads in test_cgcore_proc_migration

[Sasha Levin]

test_cgcore_proc_migration() moves a process with 13 extra threads to
cg_dst and checks that they all moved by counting the lines of
cg_dst/cgroup.threads.

test_stress.sh runs test_core while enabling and disabling cpuset in the
root cgroup.subtree_control in a loop. Each change moves every task in
the hierarchy to a new css_set. Reading cgroup.threads walks the
cgroup's css_sets without holding cgroup_mutex, so when such a change
lands in the middle of the read, threads that were already listed from
the old css_set are listed again from the new one. The test then fails
with "not ok 8 test_cgcore_proc_migration" although the migration
worked.

Running the test's steps 40000 times in a 2 CPU KVM guest with the
cpuset changes going on, 5 reads listed the 14 threads two or three
times over (up to 36 lines), and a second read right after gave 14.
Without the cpuset changes it did not happen in 40000 runs. With
test_stress.sh, the test failed in 2 of 150 runs.

Check the cgroup of each thread of the process in /proc instead, which
other migrations can't affect.

Fixes: 1a99fcc035fb ("selftests: cgroup: Run test_core under interfering stress")
Signed-off-by: Sasha Levin <sashal@kernel.org>
---
 tools/testing/selftests/cgroup/test_core.c | 36 +++++++++++++++++++++-
 1 file changed, 35 insertions(+), 1 deletion(-)

diff --git a/tools/testing/selftests/cgroup/test_core.c b/tools/testing/selftests/cgroup/test_core.c
index 20d2b63774c37..53592bf309fac 100644
--- a/tools/testing/selftests/cgroup/test_core.c
+++ b/tools/testing/selftests/cgroup/test_core.c
@@ -9,6 +9,7 @@
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <unistd.h>
+#include <dirent.h>
 #include <fcntl.h>
 #include <sched.h>
 #include <stdio.h>
@@ -567,6 +568,39 @@ static void *dummy_thread_fn(void *arg)
 	return (void *)(size_t)pause();
 }
 
+/*
+ * Count the threads of the current process, or return -1 if any of them
+ * is not in @cgroup. Unlike reading cgroup.threads, this is not affected
+ * by other migrations happening at the same time.
+ */
+static int count_threads_in_cgroup(const char *root, const char *cgroup)
+{
+	char line[PATH_MAX];
+	struct dirent *de;
+	int cnt = 0;
+	DIR *dir;
+
+	snprintf(line, sizeof(line), CG_PATH_FORMAT, cgroup + strlen(root));
+
+	dir = opendir("/proc/self/task");
+	if (!dir)
+		return -1;
+
+	while ((de = readdir(dir))) {
+		if (de->d_name[0] == '.')
+			continue;
+
+		if (proc_read_strstr(atoi(de->d_name), false, "cgroup", line)) {
+			cnt = -1;
+			break;
+		}
+		cnt++;
+	}
+
+	closedir(dir);
+	return cnt;
+}
+
 /*
  * Test threadgroup migration.
  * All threads of a process are migrated together.
@@ -597,7 +631,7 @@ static int test_cgcore_proc_migration(const char *root)
 	}
 
 	cg_enter_current(dst);
-	if (cg_read_lc(dst, CG_THREADS_FILE) != n_threads + 1)
+	if (count_threads_in_cgroup(root, dst) != n_threads + 1)
 		goto cleanup;
 
 	ret = KSFT_PASS;
-- 
2.53.0