src/api.c: support /proc mounted with subset=pid

/proc filesystem can be mounted with subset=pid as one of its mount
options.  This option hides all the top-level files and directories,
those are not related to processes. The cgroup v1 filesystem depends
on the /proc/cgroups to populate the cgroups controllers and will fail
during the cgroup_init() phase, when not available, whereas cgroup v2
considers this as a deprecated file and recommends reading the list of
controller from <unified mount point>/cgroup.controllers[1].

Support this valid /proc mount point only when the system is booted with
the unified mode and will fail to initialize in the case we find the
cgroup v1 mounted, i.e, the system booted with legacy or hybrid mode.

[1] https://docs.kernel.org/admin-guide/cgroup-v2.html#deprecated-v1-core-features

Fixes: https://github.com/libcgroup/libcgroup/issues/234
Reported-by: @0n-s (github username)
Signed-off-by: Kamalesh Babulal <kamalesh.babulal@oracle.com>
Signed-off-by: Tom Hromatka <tom.hromatka@oracle.com>

diff --git a/src/api.c b/src/api.c
index 4cda82c81bc22770f5474bb4dbaa27f4a725fbbf..51190ad410386de5d98a68854567fd7d386d6d5b 100644 (file)
--- a/src/api.c
+++ b/src/api.c
@@ -1330,6 +1330,73 @@ static void cgroup_free_cg_mount_table(void)
        memset(&cg_cgroup_v2_empty_mount_paths, 0, sizeof(cg_cgroup_v2_empty_mount_paths));
 }
 
+/*
+ * Parses the mount options of the given mount point and checks for the
+ * option in the list of mount options and sets is_set accordingly.
+ * @mnt: Mount point name to search for mount points.
+ * @mnt_opt: Mount option to be searched.
+ * @is_set: Set to 1, when mount option is found, 0 otherwise.
+ *
+ * Returns 0, in case of success and ECGOTHER on failure.
+ */
+static int check_mount_point_opt(const char *mnt, const char *mnt_opt, int * const is_set)
+{
+       struct mntent *ent, *temp_ent = NULL;
+       char mntent_buffer[4 * FILENAME_MAX];
+       char *mntopt = NULL;
+       char mnt_opt_delim;
+       FILE *proc_mount;
+       int ret = 0;
+
+       if (!mnt || !mnt_opt || !is_set)
+               return ECGINVAL;
+
+       proc_mount = setmntent(mnt, "r");
+       if (!proc_mount) {
+               cgroup_err("cannot open %s: %s\n", mnt, strerror(errno));
+               last_errno = errno;
+               ret = ECGOTHER;
+               goto err;
+       }
+
+       temp_ent = (struct mntent *) malloc(sizeof(struct mntent));
+       if (!temp_ent) {
+               last_errno = errno;
+               ret = ECGOTHER;
+               goto err;
+       }
+
+       ent = getmntent_r(proc_mount, temp_ent, mntent_buffer, sizeof(mntent_buffer));
+       if (!ent) {
+               last_errno = errno;
+               ret = ECGOTHER;
+               goto err;
+       }
+
+       *is_set = 0;
+       while((mntopt = hasmntopt(ent, mnt_opt))) {
+               mnt_opt_delim = mntopt[strlen(mnt_opt)];
+               if (mnt_opt_delim == '\0' || mnt_opt_delim == ',') {
+                       *is_set  = 1;
+                       break;
+               }
+       }
+
+       if (*is_set == 0) {
+               cgroup_dbg("%s, not found in the list of ", mnt_opt);
+               cgroup_cont("mount options of %s\n", mnt);
+       }
+
+err:
+       if (proc_mount)
+               endmntent(proc_mount);
+
+       if (temp_ent)
+               free(temp_ent);
+
+       return ret;
+}
+
 /*
  * Reads /proc/cgroups and populates the controllers/subsys_name. This
  * function should be called with cg_mount_table_lock taken.
@@ -1338,16 +1405,26 @@ static int cgroup_populate_controllers(char *controllers[CG_CONTROLLER_MAX])
 {
        int hierarchy, num_cgroups, enabled;
        char subsys_name[FILENAME_MAX];
+       char mnt_opt[] = "subset=pid";
        FILE *proc_cgroup;
        char *buf = NULL;
+       int mnt_opt_set;
        int ret = 0;
-       int i, err;
+       int err, i = 0;
 
        proc_cgroup = fopen("/proc/cgroups", "re");
        if (!proc_cgroup) {
-               cgroup_err("cannot open /proc/cgroups: %s\n", strerror(errno));
-               last_errno = errno;
-               ret = ECGOTHER;
+               cgroup_warn("cannot open /proc/cgroups: %s\n", strerror(errno));
+               ret = check_mount_point_opt("/proc/self/mounts", mnt_opt, &mnt_opt_set);
+               if (ret)
+                       goto err;
+
+               if (!mnt_opt_set)
+                       ret = ECGINVAL;
+               /*
+                * /proc, mounted with subset=pid is valid. cgroup v2 doesn't
+                * depend on /proc/cgroups to parse the available controllers.
+                */
                goto err;
        }
 
@@ -1369,7 +1446,6 @@ static int cgroup_populate_controllers(char *controllers[CG_CONTROLLER_MAX])
                goto err;
        }
 
-       i = 0;
        while (!feof(proc_cgroup)) {
                /*
                 * check Linux Kernel sources/kernel/cgroup/cgroup.c cgroup_init_early(),
@@ -1438,6 +1514,13 @@ static int cgroup_populate_mount_points(char *controllers[CG_CONTROLLER_MAX])
                                  sizeof(mntent_buffer))) != NULL) {
 
                if (strcmp(ent->mnt_type, "cgroup") == 0) {
+                       if (controllers[0] == NULL) {
+                               cgroup_err("cgroup v1 requires /proc/cgroups, check if /proc ");
+                               cgroup_cont("is mounted with subset=pid option.\n");
+                               ret = ECGINVAL;
+                               goto err;
+                       }
+
                        ret = cgroup_process_v1_mnt(controllers, ent, &found_mnt);
                        if (ret)
                                goto err;