untested: add memory.max proxy

TODO: make this optional? probably needs to be if we're going to upstream
it.

Signed-off-by: Tycho Andersen <tycho@tycho.pizza>

src/cgroups/cgfsng.c

@@ -518,6 +518,9 @@ static bool cgfsng_get_hierarchies(struct cgroup_ops *ops, int n, char ***out)
 	return true;
 }
 
+static int cgfsng_get_memory_max(struct cgroup_ops *ops, const char *cgroup,
+				 char **value);
+
 static bool cgfsng_get(struct cgroup_ops *ops, const char *controller,
 		       const char *cgroup, const char *file, char **value)
 {
@@ -530,6 +533,25 @@ static bool cgfsng_get(struct cgroup_ops *ops, const char *controller,
 
 	path = must_make_path_relative(cgroup, file, NULL);
 	*value = readat_file(h->fd, path);
+	/*
+	 * Some applications (aka the JVM on cgroupv2) only check memory.max,
+	 * and then use the sysinfo() syscall to reason about how much memory
+	 * the system has.
+	 *
+	 * In the case of systemd running inside a container, when the unit
+	 * itself doesn't have a limit, memory.max reads "max", so the
+	 * application will then call sysinfo() and decide it has the whole
+	 * system's memory to use.
+	 *
+	 * Let's fix this up by proxying lxcfs' notion of max memory for
+	 * memory.max if the underlying controller does not have a value set.
+	 */
+	if (value && !strcmp("memory.max", file) && !strcmp("max\n", *value)) {
+		free(*value);
+		cgfsng_get_memory_max(ops, cgroup, value);
+		return *value != NULL;
+	}
+
 	return *value != NULL;
 }