freeze_processes: implement kludges for cgroup v1

Cgroup v1 freezer has always been problematic, failing to freeze a
cgroup.

In runc, we have implemented a few kludges to increase the chance of
succeeding, but those are used when runc freezes a cgroup for its own
purposes (for "runc pause" and to modify device properties for cgroup
v1).

When criu is used, it fails to freeze a cgroup from time to time
(see [1], [2]). Let's try adding kludges similar to ones in runc.

Alas, I have absolutely no way to test this, so please review carefully.

[1]: opencontainers/runc#4273
[2]: opencontainers/runc#4457

Signed-off-by: Kir Kolyshkin <[email protected]>

criu/seize.c

@@ -542,6 +542,7 @@ static int freeze_processes(void)
 	enum freezer_state state = THAWED;
 
 	static const unsigned long step_ms = 100;
+	/* Since opts.timeout is in seconds, multiply it by 1000 to convert to milliseconds. */
 	unsigned long nr_attempts = (opts.timeout * 1000) / step_ms;
 	unsigned long i = 0;
 
@@ -599,6 +600,35 @@ static int freeze_processes(void)
 				goto err;
 			}
 			nanosleep(&req, NULL);
+
+			if (cgroup_v2)
+				continue;
+
+			/* As per older kernel docs (freezer-subsystem.txt before
+			 * the kernel commit ef9fe980c6fcc1821), if FREEZING is seen,
+			 * userspace should either retry or thaw. While current
+			 * kernel cgroup v1 docs no longer mention a need to retry,
+			 * even recent kernels can't reliably freeze a cgroup v1.
+			 *
+			 * Let's keep asking the kernel to freeze from time to time.
+			 * In addition, do occasional thaw/sleep/freeze.
+			 *
+			 * This is still a game of chances (the real fix belongs to the kernel)
+			 * but these kludges might improve the probability of success.
+			 *
+			 * Cgroup v2 does not have this problem.
+			 */
+			switch (i % 32) {
+			case 9:
+			case 20:
+				freezer_write_state(fd, FROZEN);
+				break;
+			case 31:
+				freezer_write_state(fd, THAWED);
+				nanosleep(&req, NULL);
+				freezer_write_state(fd, FROZEN);
+				break;
+			}
 		}
 
 		if (i > nr_attempts) {