Merge pull request #3891 from brauner/2021-07-01.fixes

[mirror_lxc.git] / src / lxc / cgroups / cgfsng.c

/* SPDX-License-Identifier: LGPL-2.1+ */

/*
 * cgfs-ng.c: this is a new, simplified implementation of a filesystem
 * cgroup backend.  The original cgfs.c was designed to be as flexible
 * as possible.  It would try to find cgroup filesystems no matter where
 * or how you had them mounted, and deduce the most usable mount for
 * each controller.
 *
 * This new implementation assumes that cgroup filesystems are mounted
 * under /sys/fs/cgroup/clist where clist is either the controller, or
 * a comma-separated list of controllers.
 */

#ifndef _GNU_SOURCE
#define _GNU_SOURCE 1
#endif
#include <ctype.h>
#include <dirent.h>
#include <errno.h>
#include <grp.h>
#include <linux/kdev_t.h>
#include <linux/types.h>
#include <poll.h>
#include <signal.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/epoll.h>
#include <sys/types.h>
#include <unistd.h>

#include "af_unix.h"
#include "caps.h"
#include "cgroup.h"
#include "cgroup2_devices.h"
#include "cgroup_utils.h"
#include "commands.h"
#include "commands_utils.h"
#include "conf.h"
#include "config.h"
#include "error_utils.h"
#include "log.h"
#include "macro.h"
#include "mainloop.h"
#include "memory_utils.h"
#include "mount_utils.h"
#include "storage/storage.h"
#include "string_utils.h"
#include "syscall_wrappers.h"
#include "utils.h"

#ifndef HAVE_STRLCPY
#include "include/strlcpy.h"
#endif

#ifndef HAVE_STRLCAT
#include "include/strlcat.h"
#endif

lxc_log_define(cgfsng, cgroup);

/*
 * Given a pointer to a null-terminated array of pointers, realloc to add one
 * entry, and point the new entry to NULL. Do not fail. Return the index to the
 * second-to-last entry - that is, the one which is now available for use
 * (keeping the list null-terminated).
 */
static int list_add(void ***list)
{
        int idx = 0;
        void **p;

        if (*list)
                for (; (*list)[idx]; idx++)
                        ;

        p = realloc(*list, (idx + 2) * sizeof(void **));
        if (!p)
                return ret_errno(ENOMEM);

        p[idx + 1] = NULL;
        *list = p;

        return idx;
}

/* Given a null-terminated array of strings, check whether @entry is one of the
 * strings.
 */
static bool string_in_list(char **list, const char *entry)
{
        if (!list)
                return false;

        for (int i = 0; list[i]; i++)
                if (strequal(list[i], entry))
                        return true;

        return false;
}

/* Given a handler's cgroup data, return the struct hierarchy for the controller
 * @c, or NULL if there is none.
 */
static struct hierarchy *get_hierarchy(const struct cgroup_ops *ops, const char *controller)
{
        if (!ops->hierarchies)
                return log_trace_errno(NULL, errno, "There are no useable cgroup controllers");

        for (int i = 0; ops->hierarchies[i]; i++) {
                if (!controller) {
                        /* This is the empty unified hierarchy. */
                        if (ops->hierarchies[i]->controllers && !ops->hierarchies[i]->controllers[0])
                                return ops->hierarchies[i];

                        continue;
                }

                /*
                 * Handle controllers with significant implementation changes
                 * from cgroup to cgroup2.
                 */
                if (pure_unified_layout(ops)) {
                        if (strequal(controller, "devices")) {
                                if (device_utility_controller(ops->unified))
                                        return ops->unified;

                                break;
                        } else if (strequal(controller, "freezer")) {
                                if (freezer_utility_controller(ops->unified))
                                        return ops->unified;

                                break;
                        }
                }

                if (string_in_list(ops->hierarchies[i]->controllers, controller))
                        return ops->hierarchies[i];
        }

        if (controller)
                WARN("There is no useable %s controller", controller);
        else
                WARN("There is no empty unified cgroup hierarchy");

        return ret_set_errno(NULL, ENOENT);
}

int prepare_cgroup_fd(const struct cgroup_ops *ops, struct cgroup_fd *fd, bool limit)
{
        int dfd;
        const struct hierarchy *h;

        h = get_hierarchy(ops, fd->controller);
        if (!h)
                return ret_errno(ENOENT);

        /*
         * The client requested that the controller must be in a specific
         * cgroup version.
         */
        if (fd->type != 0 && fd->type != h->fs_type)
                return ret_errno(EINVAL);

        if (limit)
                dfd = h->dfd_con;
        else
                dfd = h->dfd_lim;
        if (dfd < 0)
                return ret_errno(EBADF);

        fd->layout = ops->cgroup_layout;
        fd->type = h->fs_type;
        if (fd->type == UNIFIED_HIERARCHY)
                fd->utilities = h->utilities;
        fd->fd = dfd;

        return 0;
}

/* Taken over modified from the kernel sources. */
#define NBITS 32 /* bits in uint32_t */
#define DIV_ROUND_UP(n, d) (((n) + (d)-1) / (d))
#define BITS_TO_LONGS(nr) DIV_ROUND_UP(nr, NBITS)

static void set_bit(unsigned bit, uint32_t *bitarr)
{
        bitarr[bit / NBITS] |= (1 << (bit % NBITS));
}

static void clear_bit(unsigned bit, uint32_t *bitarr)
{
        bitarr[bit / NBITS] &= ~(1 << (bit % NBITS));
}

static bool is_set(unsigned bit, uint32_t *bitarr)
{
        return (bitarr[bit / NBITS] & (1 << (bit % NBITS))) != 0;
}

/* Create cpumask from cpulist aka turn:
 *
 *      0,2-3
 *
 * into bit array
 *
 *      1 0 1 1
 */
static uint32_t *lxc_cpumask(char *buf, size_t nbits)
{
        __do_free uint32_t *bitarr = NULL;
        char *token;
        size_t arrlen;

        arrlen = BITS_TO_LONGS(nbits);
        bitarr = calloc(arrlen, sizeof(uint32_t));
        if (!bitarr)
                return ret_set_errno(NULL, ENOMEM);

        lxc_iterate_parts(token, buf, ",") {
                errno = 0;
                unsigned end, start;
                char *range;

                start = strtoul(token, NULL, 0);
                end = start;
                range = strchr(token, '-');
                if (range)
                        end = strtoul(range + 1, NULL, 0);

                if (!(start <= end))
                        return ret_set_errno(NULL, EINVAL);

                if (end >= nbits)
                        return ret_set_errno(NULL, EINVAL);

                while (start <= end)
                        set_bit(start++, bitarr);
        }

        return move_ptr(bitarr);
}

/* Turn cpumask into simple, comma-separated cpulist. */
static char *lxc_cpumask_to_cpulist(uint32_t *bitarr, size_t nbits)
{
        __do_free_string_list char **cpulist = NULL;
        char numstr[INTTYPE_TO_STRLEN(size_t)] = {0};
        int ret;

        for (size_t i = 0; i <= nbits; i++) {
                if (!is_set(i, bitarr))
                        continue;

                ret = strnprintf(numstr, sizeof(numstr), "%zu", i);
                if (ret < 0)
                        return NULL;

                ret = lxc_append_string(&cpulist, numstr);
                if (ret < 0)
                        return ret_set_errno(NULL, ENOMEM);
        }

        if (!cpulist)
                return ret_set_errno(NULL, ENOMEM);

        return lxc_string_join(",", (const char **)cpulist, false);
}

static ssize_t get_max_cpus(char *cpulist)
{
        char *c1, *c2;
        char *maxcpus = cpulist;
        size_t cpus = 0;

        c1 = strrchr(maxcpus, ',');
        if (c1)
                c1++;

        c2 = strrchr(maxcpus, '-');
        if (c2)
                c2++;

        if (!c1 && !c2)
                c1 = maxcpus;
        else if (c1 > c2)
                c2 = c1;
        else if (c1 < c2)
                c1 = c2;
        else if (!c1 && c2)
                c1 = c2;

        errno = 0;
        cpus = strtoul(c1, NULL, 0);
        if (errno != 0)
                return -1;

        return cpus;
}

static inline bool is_unified_hierarchy(const struct hierarchy *h)
{
        return h->fs_type == UNIFIED_HIERARCHY;
}

/* Return true if the controller @entry is found in the null-terminated list of
 * hierarchies @hlist.
 */
static bool controller_available(struct hierarchy **hlist, char *entry)
{
        if (!hlist)
                return false;

        for (int i = 0; hlist[i]; i++)
                if (string_in_list(hlist[i]->controllers, entry))
                        return true;

        return false;
}

static bool controllers_available(struct cgroup_ops *ops)
{
        struct hierarchy **hlist;

        if (!ops->cgroup_use)
                return true;

        hlist = ops->hierarchies;
        for (char **cur = ops->cgroup_use; cur && *cur; cur++)
                if (!controller_available(hlist, *cur))
                        return log_error(false, "The %s controller found", *cur);

        return true;
}

static char **list_new(void)
{
        __do_free_string_list char **list = NULL;
        int idx;

        idx = list_add((void ***)&list);
        if (idx < 0)
                return NULL;

        list[idx] = NULL;
        return move_ptr(list);
}

static int list_add_string(char ***list, char *entry)
{
        __do_free char *dup = NULL;
        int idx;

        dup = strdup(entry);
        if (!dup)
                return ret_errno(ENOMEM);

        idx = list_add((void ***)list);
        if (idx < 0)
                return idx;

        (*list)[idx] = move_ptr(dup);
        return 0;
}

static char **list_add_controllers(char *controllers)
{
        __do_free_string_list char **list = NULL;
        char *it;

        lxc_iterate_parts(it, controllers, ", \t\n") {
                int ret;

                ret = list_add_string(&list, it);
                if (ret < 0)
                        return NULL;
        }

        return move_ptr(list);
}

static char **unified_controllers(int dfd, const char *file)
{
        __do_free char *buf = NULL;

        buf = read_file_at(dfd, file, PROTECT_OPEN, 0);
        if (!buf)
                return NULL;

        return list_add_controllers(buf);
}

static bool skip_hierarchy(const struct cgroup_ops *ops, char **controllers)
{
        if (!ops->cgroup_use)
                return false;

        for (char **cur_ctrl = controllers; cur_ctrl && *cur_ctrl; cur_ctrl++) {
                bool found = false;

                for (char **cur_use = ops->cgroup_use; cur_use && *cur_use; cur_use++) {
                        if (!strequal(*cur_use, *cur_ctrl))
                                continue;

                        found = true;
                        break;
                }

                if (found)
                        continue;

                return true;
        }

        return false;
}

static int cgroup_hierarchy_add(struct cgroup_ops *ops, int dfd_mnt, char *mnt,
                                int dfd_base, char *base_cgroup,
                                char **controllers, cgroupfs_type_magic_t fs_type)
{
        __do_free struct hierarchy *new = NULL;
        int idx;

        if (abspath(base_cgroup))
                return syserror_set(-EINVAL, "Container base path must be relative to controller mount");

        new = zalloc(sizeof(*new));
        if (!new)
                return ret_errno(ENOMEM);

        new->dfd_con            = -EBADF;
        new->dfd_lim            = -EBADF;
        new->dfd_mon            = -EBADF;

        new->fs_type            = fs_type;
        new->controllers        = controllers;
        new->at_mnt             = mnt;
        new->at_base            = base_cgroup;

        new->dfd_mnt            = dfd_mnt;
        new->dfd_base           = dfd_base;

        TRACE("Adding cgroup hierarchy mounted at %s and base cgroup %s",
              mnt, maybe_empty(base_cgroup));
        for (char *const *it = new->controllers; it && *it; it++)
                TRACE("The hierarchy contains the %s controller", *it);

        idx = list_add((void ***)&ops->hierarchies);
        if (idx < 0)
                return ret_errno(idx);

        if (fs_type == UNIFIED_HIERARCHY)
                ops->unified = new;
        (ops->hierarchies)[idx] = move_ptr(new);

        return 0;
}

static int cgroup_tree_remove(struct hierarchy **hierarchies, const char *path_prune)
{
        if (!path_prune || !hierarchies)
                return 0;

        for (int i = 0; hierarchies[i]; i++) {
                struct hierarchy *h = hierarchies[i];
                int ret;

                ret = cgroup_tree_prune(h->dfd_base, path_prune);
                if (ret < 0)
                        SYSWARN("Failed to destroy %d(%s)", h->dfd_base, path_prune);
                else
                        TRACE("Removed cgroup tree %d(%s)", h->dfd_base, path_prune);

                free_equal(h->path_lim, h->path_con);
        }

        return 0;
}

struct generic_userns_exec_data {
        struct hierarchy **hierarchies;
        const char *path_prune;
        struct lxc_conf *conf;
        uid_t origuid; /* target uid in parent namespace */
        char *path;
};

static int cgroup_tree_remove_wrapper(void *data)
{
        struct generic_userns_exec_data *arg = data;
        uid_t nsuid = (arg->conf->root_nsuid_map != NULL) ? 0 : arg->conf->init_uid;
        gid_t nsgid = (arg->conf->root_nsgid_map != NULL) ? 0 : arg->conf->init_gid;
        int ret;

        if (!lxc_drop_groups() && errno != EPERM)
                return log_error_errno(-1, errno, "Failed to setgroups(0, NULL)");

        ret = setresgid(nsgid, nsgid, nsgid);
        if (ret < 0)
                return log_error_errno(-1, errno, "Failed to setresgid(%d, %d, %d)",
                                       (int)nsgid, (int)nsgid, (int)nsgid);

        ret = setresuid(nsuid, nsuid, nsuid);
        if (ret < 0)
                return log_error_errno(-1, errno, "Failed to setresuid(%d, %d, %d)",
                                       (int)nsuid, (int)nsuid, (int)nsuid);

        return cgroup_tree_remove(arg->hierarchies, arg->path_prune);
}

__cgfsng_ops static void cgfsng_payload_destroy(struct cgroup_ops *ops,
                                                struct lxc_handler *handler)
{
        int ret;

        if (!ops) {
                ERROR("Called with uninitialized cgroup operations");
                return;
        }

        if (!ops->hierarchies)
                return;

        if (!handler) {
                ERROR("Called with uninitialized handler");
                return;
        }

        if (!handler->conf) {
                ERROR("Called with uninitialized conf");
                return;
        }

        if (!ops->container_limit_cgroup) {
                WARN("Uninitialized limit cgroup");
                return;
        }

        ret = bpf_program_cgroup_detach(handler->cgroup_ops->cgroup2_devices);
        if (ret < 0)
                WARN("Failed to detach bpf program from cgroup");

        if (!lxc_list_empty(&handler->conf->id_map)) {
                struct generic_userns_exec_data wrap = {
                        .conf                   = handler->conf,
                        .path_prune             = ops->container_limit_cgroup,
                        .hierarchies            = ops->hierarchies,
                        .origuid                = 0,
                };
                ret = userns_exec_1(handler->conf, cgroup_tree_remove_wrapper,
                                    &wrap, "cgroup_tree_remove_wrapper");
        } else {
                ret = cgroup_tree_remove(ops->hierarchies, ops->container_limit_cgroup);
        }
        if (ret < 0)
                SYSWARN("Failed to destroy cgroups");
}

#define __ISOL_CPUS "/sys/devices/system/cpu/isolated"
#define __OFFLINE_CPUS "/sys/devices/system/cpu/offline"
static bool cpuset1_cpus_initialize(int dfd_parent, int dfd_child,
                                    bool am_initialized)
{
        __do_free char *cpulist = NULL, *fpath = NULL, *isolcpus = NULL,
                       *offlinecpus = NULL, *posscpus = NULL;
        __do_free uint32_t *isolmask = NULL, *offlinemask = NULL,
                           *possmask = NULL;
        int ret;
        ssize_t i;
        ssize_t maxisol = 0, maxoffline = 0, maxposs = 0;
        bool flipped_bit = false;

        posscpus = read_file_at(dfd_parent, "cpuset.cpus", PROTECT_OPEN, 0);
        if (!posscpus)
                return log_error_errno(false, errno, "Failed to read file \"%s\"", fpath);

        /* Get maximum number of cpus found in possible cpuset. */
        maxposs = get_max_cpus(posscpus);
        if (maxposs < 0 || maxposs >= INT_MAX - 1)
                return false;

        if (file_exists(__ISOL_CPUS)) {
                isolcpus = read_file_at(-EBADF, __ISOL_CPUS, PROTECT_OPEN, 0);
                if (!isolcpus)
                        return log_error_errno(false, errno, "Failed to read file \"%s\"", __ISOL_CPUS);

                if (isdigit(isolcpus[0])) {
                        /* Get maximum number of cpus found in isolated cpuset. */
                        maxisol = get_max_cpus(isolcpus);
                        if (maxisol < 0 || maxisol >= INT_MAX - 1)
                                return false;
                }

                if (maxposs < maxisol)
                        maxposs = maxisol;
                maxposs++;
        } else {
                TRACE("The path \""__ISOL_CPUS"\" to read isolated cpus from does not exist");
        }

        if (file_exists(__OFFLINE_CPUS)) {
                offlinecpus = read_file_at(-EBADF, __OFFLINE_CPUS, PROTECT_OPEN, 0);
                if (!offlinecpus)
                        return log_error_errno(false, errno, "Failed to read file \"%s\"", __OFFLINE_CPUS);

                if (isdigit(offlinecpus[0])) {
                        /* Get maximum number of cpus found in offline cpuset. */
                        maxoffline = get_max_cpus(offlinecpus);
                        if (maxoffline < 0 || maxoffline >= INT_MAX - 1)
                                return false;
                }

                if (maxposs < maxoffline)
                        maxposs = maxoffline;
                maxposs++;
        } else {
                TRACE("The path \""__OFFLINE_CPUS"\" to read offline cpus from does not exist");
        }

        if ((maxisol == 0) && (maxoffline == 0)) {
                cpulist = move_ptr(posscpus);
                goto copy_parent;
        }

        possmask = lxc_cpumask(posscpus, maxposs);
        if (!possmask)
                return log_error_errno(false, errno, "Failed to create cpumask for possible cpus");

        if (maxisol > 0) {
                isolmask = lxc_cpumask(isolcpus, maxposs);
                if (!isolmask)
                        return log_error_errno(false, errno, "Failed to create cpumask for isolated cpus");
        }

        if (maxoffline > 0) {
                offlinemask = lxc_cpumask(offlinecpus, maxposs);
                if (!offlinemask)
                        return log_error_errno(false, errno, "Failed to create cpumask for offline cpus");
        }

        for (i = 0; i <= maxposs; i++) {
                if ((isolmask && !is_set(i, isolmask)) ||
                    (offlinemask && !is_set(i, offlinemask)) ||
                    !is_set(i, possmask))
                        continue;

                flipped_bit = true;
                clear_bit(i, possmask);
        }

        if (!flipped_bit) {
                cpulist = lxc_cpumask_to_cpulist(possmask, maxposs);
                TRACE("No isolated or offline cpus present in cpuset");
        } else {
                cpulist = move_ptr(posscpus);
                TRACE("Removed isolated or offline cpus from cpuset");
        }
        if (!cpulist)
                return log_error_errno(false, errno, "Failed to create cpu list");

copy_parent:
        if (!am_initialized) {
                ret = lxc_writeat(dfd_child, "cpuset.cpus", cpulist, strlen(cpulist));
                if (ret < 0)
                        return log_error_errno(false, errno, "Failed to write cpu list to \"%d/cpuset.cpus\"", dfd_child);

                TRACE("Copied cpu settings of parent cgroup");
        }

        return true;
}

static bool cpuset1_initialize(int dfd_base, int dfd_next)
{
        char mems[PATH_MAX];
        ssize_t bytes;
        char v;

        /*
        * Determine whether the base cgroup has cpuset
        * inheritance turned on.
         */
        bytes = lxc_readat(dfd_base, "cgroup.clone_children", &v, 1);
        if (bytes < 0)
                return syserror_ret(false, "Failed to read file %d(cgroup.clone_children)", dfd_base);

        /*
        * Initialize cpuset.cpus and make remove any isolated
        * and offline cpus.
         */
        if (!cpuset1_cpus_initialize(dfd_base, dfd_next, v == '1'))
                return syserror_ret(false, "Failed to initialize cpuset.cpus");

        /* Read cpuset.mems from parent... */
        bytes = lxc_readat(dfd_base, "cpuset.mems", mems, sizeof(mems));
        if (bytes < 0)
                return syserror_ret(false, "Failed to read file %d(cpuset.mems)", dfd_base);

        /* ... and copy to first cgroup in the tree... */
        bytes = lxc_writeat(dfd_next, "cpuset.mems", mems, bytes);
        if (bytes < 0)
                return syserror_ret(false, "Failed to write %d(cpuset.mems)", dfd_next);

        /* ... and finally turn on cpuset inheritance. */
        bytes = lxc_writeat(dfd_next, "cgroup.clone_children", "1", 1);
        if (bytes < 0)
                return syserror_ret(false, "Failed to write %d(cgroup.clone_children)", dfd_next);

        return log_trace(true, "Initialized cpuset in the legacy hierarchy");
}

static int __cgroup_tree_create(int dfd_base, const char *path, mode_t mode,
                                bool cpuset_v1, bool eexist_ignore)
{
        __do_close int dfd_final = -EBADF;
        int dfd_cur = dfd_base;
        int ret = 0;
        size_t len;
        char *cur;
        char buf[PATH_MAX];

        if (is_empty_string(path))
                return ret_errno(EINVAL);

        len = strlcpy(buf, path, sizeof(buf));
        if (len >= sizeof(buf))
                return ret_errno(E2BIG);

        lxc_iterate_parts(cur, buf, "/") {
                /*
                 * Even though we vetted the paths when we parsed the config
                 * we're paranoid here and check that the path is neither
                 * absolute nor walks upwards.
                 */
                if (abspath(cur))
                        return syserror_set(-EINVAL, "No absolute paths allowed");

                if (strnequal(cur, "..", STRLITERALLEN("..")))
                        return syserror_set(-EINVAL, "No upward walking paths allowed");

                ret = mkdirat(dfd_cur, cur, mode);
                if (ret < 0) {
                        if (errno != EEXIST)
                                return syserror("Failed to create %d(%s)", dfd_cur, cur);

                        ret = -EEXIST;
                }
                TRACE("%s %d(%s) cgroup", !ret ? "Created" : "Reusing", dfd_cur, cur);

                dfd_final = open_at(dfd_cur, cur, PROTECT_OPATH_DIRECTORY, PROTECT_LOOKUP_BENEATH, 0);
                if (dfd_final < 0)
                        return syserror("Fail to open%s directory %d(%s)",
                                        !ret ? " newly created" : "", dfd_base, cur);
                if (dfd_cur != dfd_base)
                        close(dfd_cur);
                else if (cpuset_v1 && !cpuset1_initialize(dfd_base, dfd_final))
                        return syserror_set(-EINVAL, "Failed to initialize cpuset controller in the legacy hierarchy");
                /*
                 * Leave dfd_final pointing to the last fd we opened so
                 * it will be automatically zapped if we return early.
                 */
                dfd_cur = dfd_final;
        }

        /* The final cgroup must be succesfully creatd by us. */
        if (ret) {
                if (ret != -EEXIST || !eexist_ignore)
                        return syserror_set(ret, "Creating the final cgroup %d(%s) failed", dfd_base, path);
        }

        return move_fd(dfd_final);
}

static bool cgroup_tree_create(struct cgroup_ops *ops, struct lxc_conf *conf,
                               struct hierarchy *h, const char *cgroup_limit_dir,
                               const char *cgroup_leaf, bool payload)
{
        __do_close int fd_limit = -EBADF, fd_final = -EBADF;
        __do_free char *path = NULL, *limit_path = NULL;
        bool cpuset_v1 = false;

        /*
         * The legacy cpuset controller needs massaging in case inheriting
         * settings from its immediate ancestor cgroup hasn't been turned on.
         */
        cpuset_v1 = !is_unified_hierarchy(h) && string_in_list(h->controllers, "cpuset");

        if (payload && cgroup_leaf) {
                /* With isolation both parts need to not already exist. */
                fd_limit = __cgroup_tree_create(h->dfd_base, cgroup_limit_dir, 0755, cpuset_v1, false);
                if (fd_limit < 0)
                        return syserror_ret(false, "Failed to create limiting cgroup %d(%s)", h->dfd_base, cgroup_limit_dir);

                TRACE("Created limit cgroup %d->%d(%s)",
                      fd_limit, h->dfd_base, cgroup_limit_dir);

                /*
                 * With isolation the devices legacy cgroup needs to be
                 * iinitialized early, as it typically contains an 'a' (all)
                 * line, which is not possible once a subdirectory has been
                 * created.
                 */
                if (string_in_list(h->controllers, "devices") &&
                    !ops->setup_limits_legacy(ops, conf, true))
                        return log_error(false, "Failed to setup legacy device limits");

                limit_path = make_cgroup_path(h, h->at_base, cgroup_limit_dir, NULL);
                path = must_make_path(limit_path, cgroup_leaf, NULL);

                /*
                 * If we use a separate limit cgroup, the leaf cgroup, i.e. the
                 * cgroup the container actually resides in, is below fd_limit.
                 */
                fd_final = __cgroup_tree_create(fd_limit, cgroup_leaf, 0755, cpuset_v1, false);
                if (fd_final < 0) {
                        /* Ensure we don't leave any garbage behind. */
                        if (cgroup_tree_prune(h->dfd_base, cgroup_limit_dir))
                                SYSWARN("Failed to destroy %d(%s)", h->dfd_base, cgroup_limit_dir);
                        else
                                TRACE("Removed cgroup tree %d(%s)", h->dfd_base, cgroup_limit_dir);
                }
        } else {
                path = make_cgroup_path(h, h->at_base, cgroup_limit_dir, NULL);

                fd_final = __cgroup_tree_create(h->dfd_base, cgroup_limit_dir, 0755, cpuset_v1, false);
        }
        if (fd_final < 0)
                return syserror_ret(false, "Failed to create %s cgroup %d(%s)", payload ? "payload" : "monitor", h->dfd_base, cgroup_limit_dir);

        if (payload) {
                h->dfd_con = move_fd(fd_final);
                h->path_con = move_ptr(path);

                if (fd_limit < 0)
                        h->dfd_lim = h->dfd_con;
                else
                        h->dfd_lim = move_fd(fd_limit);

                if (limit_path)
                        h->path_lim = move_ptr(limit_path);
                else
                        h->path_lim = h->path_con;
        } else {
                h->dfd_mon = move_fd(fd_final);
        }

        return true;
}

static void cgroup_tree_prune_leaf(struct hierarchy *h, const char *path_prune,
                                   bool payload)
{
        bool prune = true;

        if (payload) {
                /* Check whether we actually created the cgroup to prune. */
                if (h->dfd_lim < 0)
                        prune = false;

                free_equal(h->path_con, h->path_lim);
                close_equal(h->dfd_con, h->dfd_lim);
        } else {
                /* Check whether we actually created the cgroup to prune. */
                if (h->dfd_mon < 0)
                        prune = false;

                close_prot_errno_disarm(h->dfd_mon);
        }

        /* We didn't create this cgroup. */
        if (!prune)
                return;

        if (cgroup_tree_prune(h->dfd_base, path_prune))
                SYSWARN("Failed to destroy %d(%s)", h->dfd_base, path_prune);
        else
                TRACE("Removed cgroup tree %d(%s)", h->dfd_base, path_prune);
}

__cgfsng_ops static void cgfsng_monitor_destroy(struct cgroup_ops *ops,
                                                struct lxc_handler *handler)
{
        int len;
        char pidstr[INTTYPE_TO_STRLEN(pid_t)];
        const struct lxc_conf *conf;

        if (!ops) {
                ERROR("Called with uninitialized cgroup operations");
                return;
        }

        if (!ops->hierarchies)
                return;

        if (!handler) {
                ERROR("Called with uninitialized handler");
                return;
        }

        if (!handler->conf) {
                ERROR("Called with uninitialized conf");
                return;
        }
        conf = handler->conf;

        if (!ops->monitor_cgroup) {
                WARN("Uninitialized monitor cgroup");
                return;
        }

        len = strnprintf(pidstr, sizeof(pidstr), "%d", handler->monitor_pid);
        if (len < 0)
                return;

        for (int i = 0; ops->hierarchies[i]; i++) {
                __do_close int fd_pivot = -EBADF;
                __do_free char *pivot_path = NULL;
                struct hierarchy *h = ops->hierarchies[i];
                bool cpuset_v1 = false;
                int ret;

                /* Monitor might have died before we entered the cgroup. */
                if (handler->monitor_pid <= 0) {
                        WARN("No valid monitor process found while destroying cgroups");
                        goto cgroup_prune_tree;
                }

                if (conf->cgroup_meta.monitor_pivot_dir)
                        pivot_path = must_make_path(conf->cgroup_meta.monitor_pivot_dir, CGROUP_PIVOT, NULL);
                else if (conf->cgroup_meta.dir)
                        pivot_path = must_make_path(conf->cgroup_meta.dir, CGROUP_PIVOT, NULL);
                else
                        pivot_path = must_make_path(CGROUP_PIVOT, NULL);

                cpuset_v1 = !is_unified_hierarchy(h) && string_in_list(h->controllers, "cpuset");

                fd_pivot = __cgroup_tree_create(h->dfd_base, pivot_path, 0755, cpuset_v1, true);
                if (fd_pivot < 0) {
                        SYSWARN("Failed to create pivot cgroup %d(%s)", h->dfd_base, pivot_path);
                        continue;
                }

                ret = lxc_writeat(fd_pivot, "cgroup.procs", pidstr, len);
                if (ret != 0) {
                        SYSWARN("Failed to move monitor %s to \"%s\"", pidstr, pivot_path);
                        continue;
                }

cgroup_prune_tree:
                ret = cgroup_tree_prune(h->dfd_base, ops->monitor_cgroup);
                if (ret < 0)
                        SYSWARN("Failed to destroy %d(%s)", h->dfd_base, ops->monitor_cgroup);
                else
                        TRACE("Removed cgroup tree %d(%s)", h->dfd_base, ops->monitor_cgroup);
        }
}

/*
 * Check we have no lxc.cgroup.dir, and that lxc.cgroup.dir.limit_prefix is a
 * proper prefix directory of lxc.cgroup.dir.payload.
 *
 * Returns the prefix length if it is set, otherwise zero on success.
 */
static bool check_cgroup_dir_config(struct lxc_conf *conf)
{
        const char *monitor_dir = conf->cgroup_meta.monitor_dir,
                   *container_dir = conf->cgroup_meta.container_dir,
                   *namespace_dir = conf->cgroup_meta.namespace_dir;

        /* none of the new options are set, all is fine */
        if (!monitor_dir && !container_dir && !namespace_dir)
                return true;

        /* some are set, make sure lxc.cgroup.dir is not also set*/
        if (conf->cgroup_meta.dir)
                return log_error_errno(false, EINVAL,
                        "lxc.cgroup.dir conflicts with lxc.cgroup.dir.payload/monitor");

        /* make sure both monitor and payload are set */
        if (!monitor_dir || !container_dir)
                return log_error_errno(false, EINVAL,
                        "lxc.cgroup.dir.payload and lxc.cgroup.dir.monitor must both be set");

        /* namespace_dir may be empty */
        return true;
}

__cgfsng_ops static bool cgfsng_monitor_create(struct cgroup_ops *ops, struct lxc_handler *handler)
{
        __do_free char *monitor_cgroup = NULL;
        int idx = 0;
        int i;
        size_t len;
        char *suffix = NULL;
        struct lxc_conf *conf;

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return true;

        if (ops->monitor_cgroup)
                return ret_set_errno(false, EEXIST);

        if (!handler || !handler->conf)
                return ret_set_errno(false, EINVAL);

        conf = handler->conf;

        if (!check_cgroup_dir_config(conf))
                return false;

        if (conf->cgroup_meta.monitor_dir) {
                monitor_cgroup = strdup(conf->cgroup_meta.monitor_dir);
        } else if (conf->cgroup_meta.dir) {
                monitor_cgroup = must_concat(&len, conf->cgroup_meta.dir, "/",
                                             DEFAULT_MONITOR_CGROUP_PREFIX,
                                             handler->name,
                                             CGROUP_CREATE_RETRY, NULL);
        } else if (ops->cgroup_pattern) {
                __do_free char *cgroup_tree = NULL;

                cgroup_tree = lxc_string_replace("%n", handler->name, ops->cgroup_pattern);
                if (!cgroup_tree)
                        return ret_set_errno(false, ENOMEM);

                monitor_cgroup = must_concat(&len, cgroup_tree, "/",
                                             DEFAULT_MONITOR_CGROUP,
                                             CGROUP_CREATE_RETRY, NULL);
        } else {
                monitor_cgroup = must_concat(&len, DEFAULT_MONITOR_CGROUP_PREFIX,
                                             handler->name,
                                             CGROUP_CREATE_RETRY, NULL);
        }
        if (!monitor_cgroup)
                return ret_set_errno(false, ENOMEM);

        if (!conf->cgroup_meta.monitor_dir) {
                suffix = monitor_cgroup + len - CGROUP_CREATE_RETRY_LEN;
                *suffix = '\0';
        }
        do {
                if (idx && suffix)
                        sprintf(suffix, "-%d", idx);

                for (i = 0; ops->hierarchies[i]; i++) {
                        if (cgroup_tree_create(ops, handler->conf,
                                               ops->hierarchies[i],
                                               monitor_cgroup, NULL, false))
                                continue;

                        DEBUG("Failed to create cgroup %s)", monitor_cgroup);
                        for (int j = 0; j <= i; j++)
                                cgroup_tree_prune_leaf(ops->hierarchies[j],
                                                       monitor_cgroup, false);

                        idx++;
                        break;
                }
        } while (ops->hierarchies[i] && idx > 0 && idx < 1000 && suffix);

        if (idx == 1000 || (!suffix && idx != 0))
                return log_error_errno(false, ERANGE, "Failed to create monitor cgroup");

        ops->monitor_cgroup = move_ptr(monitor_cgroup);
        return log_info(true, "The monitor process uses \"%s\" as cgroup", ops->monitor_cgroup);
}

/*
 * Try to create the same cgroup in all hierarchies. Start with cgroup_pattern;
 * next cgroup_pattern-1, -2, ..., -999.
 */
__cgfsng_ops static bool cgfsng_payload_create(struct cgroup_ops *ops, struct lxc_handler *handler)
{
        __do_free char *container_cgroup = NULL, *__limit_cgroup = NULL;
        char *limit_cgroup;
        int idx = 0;
        int i;
        size_t len;
        char *suffix = NULL;
        struct lxc_conf *conf;

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return true;

        if (ops->container_cgroup || ops->container_limit_cgroup)
                return ret_set_errno(false, EEXIST);

        if (!handler || !handler->conf)
                return ret_set_errno(false, EINVAL);

        conf = handler->conf;

        if (!check_cgroup_dir_config(conf))
                return false;

        if (conf->cgroup_meta.container_dir) {
                __limit_cgroup = strdup(conf->cgroup_meta.container_dir);
                if (!__limit_cgroup)
                        return ret_set_errno(false, ENOMEM);

                if (conf->cgroup_meta.namespace_dir) {
                        container_cgroup = must_make_path(__limit_cgroup,
                                                          conf->cgroup_meta.namespace_dir,
                                                          NULL);
                        limit_cgroup = __limit_cgroup;
                } else {
                        /* explicit paths but without isolation */
                        limit_cgroup = move_ptr(__limit_cgroup);
                        container_cgroup = limit_cgroup;
                }
        } else if (conf->cgroup_meta.dir) {
                limit_cgroup = must_concat(&len, conf->cgroup_meta.dir, "/",
                                           DEFAULT_PAYLOAD_CGROUP_PREFIX,
                                           handler->name,
                                           CGROUP_CREATE_RETRY, NULL);
                container_cgroup = limit_cgroup;
        } else if (ops->cgroup_pattern) {
                __do_free char *cgroup_tree = NULL;

                cgroup_tree = lxc_string_replace("%n", handler->name, ops->cgroup_pattern);
                if (!cgroup_tree)
                        return ret_set_errno(false, ENOMEM);

                limit_cgroup = must_concat(&len, cgroup_tree, "/",
                                           DEFAULT_PAYLOAD_CGROUP,
                                           CGROUP_CREATE_RETRY, NULL);
                container_cgroup = limit_cgroup;
        } else {
                limit_cgroup = must_concat(&len, DEFAULT_PAYLOAD_CGROUP_PREFIX,
                                           handler->name,
                                           CGROUP_CREATE_RETRY, NULL);
                container_cgroup = limit_cgroup;
        }
        if (!limit_cgroup)
                return ret_set_errno(false, ENOMEM);

        if (!conf->cgroup_meta.container_dir) {
                suffix = container_cgroup + len - CGROUP_CREATE_RETRY_LEN;
                *suffix = '\0';
        }
        do {
                if (idx && suffix)
                        sprintf(suffix, "-%d", idx);

                for (i = 0; ops->hierarchies[i]; i++) {
                        if (cgroup_tree_create(ops, handler->conf,
                                               ops->hierarchies[i], limit_cgroup,
                                               conf->cgroup_meta.namespace_dir,
                                               true))
                                continue;

                        DEBUG("Failed to create cgroup \"%s\"", ops->hierarchies[i]->path_con ?: "(null)");
                        for (int j = 0; j <= i; j++)
                                cgroup_tree_prune_leaf(ops->hierarchies[j],
                                                       limit_cgroup, true);

                        idx++;
                        break;
                }
        } while (ops->hierarchies[i] && idx > 0 && idx < 1000 && suffix);

        if (idx == 1000 || (!suffix && idx != 0))
                return log_error_errno(false, ERANGE, "Failed to create container cgroup");

        ops->container_cgroup = move_ptr(container_cgroup);
        if (__limit_cgroup)
                ops->container_limit_cgroup = move_ptr(__limit_cgroup);
        else
                ops->container_limit_cgroup = ops->container_cgroup;
        INFO("The container process uses \"%s\" as inner and \"%s\" as limit cgroup",
             ops->container_cgroup, ops->container_limit_cgroup);
        return true;
}

__cgfsng_ops static bool cgfsng_monitor_enter(struct cgroup_ops *ops,
                                              struct lxc_handler *handler)
{
        int monitor_len, transient_len = 0;
        char monitor[INTTYPE_TO_STRLEN(pid_t)],
            transient[INTTYPE_TO_STRLEN(pid_t)];

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return true;

        if (!ops->monitor_cgroup)
                return ret_set_errno(false, ENOENT);

        if (!handler || !handler->conf)
                return ret_set_errno(false, EINVAL);

        monitor_len = strnprintf(monitor, sizeof(monitor), "%d", handler->monitor_pid);
        if (monitor_len < 0)
                return false;

        if (handler->transient_pid > 0) {
                transient_len = strnprintf(transient, sizeof(transient), "%d", handler->transient_pid);
                if (transient_len < 0)
                        return false;
        }

        for (int i = 0; ops->hierarchies[i]; i++) {
                struct hierarchy *h = ops->hierarchies[i];
                int ret;

                ret = lxc_writeat(h->dfd_mon, "cgroup.procs", monitor, monitor_len);
                if (ret)
                        return log_error_errno(false, errno, "Failed to enter cgroup %d", h->dfd_mon);

                TRACE("Moved monitor into cgroup %d", h->dfd_mon);

                if (handler->transient_pid <= 0)
                        continue;

                ret = lxc_writeat(h->dfd_mon, "cgroup.procs", transient, transient_len);
                if (ret)
                        return log_error_errno(false, errno, "Failed to enter cgroup %d", h->dfd_mon);

                TRACE("Moved transient process into cgroup %d", h->dfd_mon);

                /*
                 * we don't keep the fds for non-unified hierarchies around
                 * mainly because we don't make use of them anymore after the
                 * core cgroup setup is done but also because there are quite a
                 * lot of them.
                 */
                if (!is_unified_hierarchy(h))
                        close_prot_errno_disarm(h->dfd_mon);
        }
        handler->transient_pid = -1;

        return true;
}

__cgfsng_ops static bool cgfsng_payload_enter(struct cgroup_ops *ops,
                                              struct lxc_handler *handler)
{
        int len;
        char pidstr[INTTYPE_TO_STRLEN(pid_t)];

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return true;

        if (!ops->container_cgroup)
                return ret_set_errno(false, ENOENT);

        if (!handler || !handler->conf)
                return ret_set_errno(false, EINVAL);

        len = strnprintf(pidstr, sizeof(pidstr), "%d", handler->pid);
        if (len < 0)
                return false;

        for (int i = 0; ops->hierarchies[i]; i++) {
                struct hierarchy *h = ops->hierarchies[i];
                int ret;

                if (is_unified_hierarchy(h) &&
                    (handler->clone_flags & CLONE_INTO_CGROUP))
                        continue;

                ret = lxc_writeat(h->dfd_con, "cgroup.procs", pidstr, len);
                if (ret != 0)
                        return log_error_errno(false, errno, "Failed to enter cgroup \"%s\"", h->path_con);

                TRACE("Moved container into %s cgroup via %d", h->path_con, h->dfd_con);
        }

        return true;
}

static int fchowmodat(int dirfd, const char *path, uid_t chown_uid,
                      gid_t chown_gid, mode_t chmod_mode)
{
        int ret;

        ret = fchownat(dirfd, path, chown_uid, chown_gid,
                       AT_EMPTY_PATH | AT_SYMLINK_NOFOLLOW);
        if (ret < 0)
                return log_warn_errno(-1,
                                      errno, "Failed to fchownat(%d, %s, %d, %d, AT_EMPTY_PATH | AT_SYMLINK_NOFOLLOW )",
                                      dirfd, path, (int)chown_uid,
                                      (int)chown_gid);

        ret = fchmodat(dirfd, (*path != '\0') ? path : ".", chmod_mode, 0);
        if (ret < 0)
                return log_warn_errno(-1, errno, "Failed to fchmodat(%d, %s, %d, AT_SYMLINK_NOFOLLOW)",
                                      dirfd, path, (int)chmod_mode);

        return 0;
}

/* chgrp the container cgroups to container group.  We leave
 * the container owner as cgroup owner.  So we must make the
 * directories 775 so that the container can create sub-cgroups.
 *
 * Also chown the tasks and cgroup.procs files.  Those may not
 * exist depending on kernel version.
 */
static int chown_cgroup_wrapper(void *data)
{
        int ret;
        uid_t destuid;
        struct generic_userns_exec_data *arg = data;
        uid_t nsuid = (arg->conf->root_nsuid_map != NULL) ? 0 : arg->conf->init_uid;
        gid_t nsgid = (arg->conf->root_nsgid_map != NULL) ? 0 : arg->conf->init_gid;

        if (!lxc_drop_groups() && errno != EPERM)
                return log_error_errno(-1, errno, "Failed to setgroups(0, NULL)");

        ret = setresgid(nsgid, nsgid, nsgid);
        if (ret < 0)
                return log_error_errno(-1, errno, "Failed to setresgid(%d, %d, %d)",
                                       (int)nsgid, (int)nsgid, (int)nsgid);

        ret = setresuid(nsuid, nsuid, nsuid);
        if (ret < 0)
                return log_error_errno(-1, errno, "Failed to setresuid(%d, %d, %d)",
                                       (int)nsuid, (int)nsuid, (int)nsuid);

        destuid = get_ns_uid(arg->origuid);
        if (destuid == LXC_INVALID_UID)
                destuid = 0;

        for (int i = 0; arg->hierarchies[i]; i++) {
                int dirfd = arg->hierarchies[i]->dfd_con;

                if (dirfd < 0)
                        return syserror_set(-EBADF, "Invalid cgroup file descriptor");

                (void)fchowmodat(dirfd, "", destuid, nsgid, 0775);

                /*
                 * Failures to chown() these are inconvenient but not
                 * detrimental We leave these owned by the container launcher,
                 * so that container root can write to the files to attach.  We
                 * chmod() them 664 so that container systemd can write to the
                 * files (which systemd in wily insists on doing).
                 */

                if (arg->hierarchies[i]->fs_type == LEGACY_HIERARCHY)
                        (void)fchowmodat(dirfd, "tasks", destuid, nsgid, 0664);

                (void)fchowmodat(dirfd, "cgroup.procs", destuid, nsgid, 0664);

                if (arg->hierarchies[i]->fs_type != UNIFIED_HIERARCHY)
                        continue;

                for (char **p = arg->hierarchies[i]->delegate; p && *p; p++)
                        (void)fchowmodat(dirfd, *p, destuid, nsgid, 0664);
        }

        return 0;
}

__cgfsng_ops static bool cgfsng_chown(struct cgroup_ops *ops,
                                      struct lxc_conf *conf)
{
        struct generic_userns_exec_data wrap;

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return true;

        if (!ops->container_cgroup)
                return ret_set_errno(false, ENOENT);

        if (!conf)
                return ret_set_errno(false, EINVAL);

        if (lxc_list_empty(&conf->id_map))
                return true;

        wrap.origuid = geteuid();
        wrap.path = NULL;
        wrap.hierarchies = ops->hierarchies;
        wrap.conf = conf;

        if (userns_exec_1(conf, chown_cgroup_wrapper, &wrap, "chown_cgroup_wrapper") < 0)
                return log_error_errno(false, errno, "Error requesting cgroup chown in new user namespace");

        return true;
}

__cgfsng_ops static void cgfsng_finalize(struct cgroup_ops *ops)
{
        if (!ops)
                return;

        if (!ops->hierarchies)
                return;

        for (int i = 0; ops->hierarchies[i]; i++) {
                struct hierarchy *h = ops->hierarchies[i];

                /* Close all monitor cgroup file descriptors. */
                close_prot_errno_disarm(h->dfd_mon);
        }
        /* Close the cgroup root file descriptor. */
        close_prot_errno_disarm(ops->dfd_mnt);

        /*
         * The checking for freezer support should obviously be done at cgroup
         * initialization time but that doesn't work reliable. The freezer
         * controller has been demoted (rightly so) to a simple file located in
         * each non-root cgroup. At the time when the container is created we
         * might still be located in /sys/fs/cgroup and so checking for
         * cgroup.freeze won't tell us anything because this file doesn't exist
         * in the root cgroup. We could then iterate through /sys/fs/cgroup and
         * find an already existing cgroup and then check within that cgroup
         * for the existence of cgroup.freeze but that will only work on
         * systemd based hosts. Other init systems might not manage cgroups and
         * so no cgroup will exist. So we defer until we have created cgroups
         * for our container which means we check here.
         */
        if (pure_unified_layout(ops) &&
            !faccessat(ops->unified->dfd_con, "cgroup.freeze", F_OK,
                       AT_SYMLINK_NOFOLLOW)) {
                TRACE("Unified hierarchy supports freezer");
                ops->unified->utilities |= FREEZER_CONTROLLER;
        }
}

/* cgroup-full:* is done, no need to create subdirs */
static inline bool cg_mount_needs_subdirs(int cgroup_automount_type)
{
        switch (cgroup_automount_type) {
        case LXC_AUTO_CGROUP_RO:
                return true;
        case LXC_AUTO_CGROUP_RW:
                return true;
        case LXC_AUTO_CGROUP_MIXED:
                return true;
        }

        return false;
}

/* After $rootfs/sys/fs/container/controller/the/cg/path has been created,
 * remount controller ro if needed and bindmount the cgroupfs onto
 * control/the/cg/path.
 */
static int cg_legacy_mount_controllers(int cgroup_automount_type, struct hierarchy *h,
                                       char *hierarchy_mnt, char *cgpath,
                                       const char *container_cgroup)
{
        __do_free char *sourcepath = NULL;
        int ret, remount_flags;
        int flags = MS_BIND;

        if ((cgroup_automount_type == LXC_AUTO_CGROUP_RO) ||
            (cgroup_automount_type == LXC_AUTO_CGROUP_MIXED)) {
                ret = mount(hierarchy_mnt, hierarchy_mnt, "cgroup", MS_BIND, NULL);
                if (ret < 0)
                        return log_error_errno(-1, errno, "Failed to bind mount \"%s\" onto \"%s\"",
                                               hierarchy_mnt, hierarchy_mnt);

                remount_flags = add_required_remount_flags(hierarchy_mnt,
                                                           hierarchy_mnt,
                                                           flags | MS_REMOUNT);
                ret = mount(hierarchy_mnt, hierarchy_mnt, "cgroup",
                            remount_flags | MS_REMOUNT | MS_BIND | MS_RDONLY,
                            NULL);
                if (ret < 0)
                        return log_error_errno(-1, errno, "Failed to remount \"%s\" ro", hierarchy_mnt);

                INFO("Remounted %s read-only", hierarchy_mnt);
        }

        sourcepath = make_cgroup_path(h, h->at_base, container_cgroup, NULL);
        if (cgroup_automount_type == LXC_AUTO_CGROUP_RO)
                flags |= MS_RDONLY;

        ret = mount(sourcepath, cgpath, "cgroup", flags, NULL);
        if (ret < 0)
                return log_error_errno(-1, errno, "Failed to mount \"%s\" onto \"%s\"",
                                       h->controllers[0], cgpath);
        INFO("Mounted \"%s\" onto \"%s\"", h->controllers[0], cgpath);

        if (flags & MS_RDONLY) {
                remount_flags = add_required_remount_flags(sourcepath, cgpath,
                                                           flags | MS_REMOUNT);
                ret = mount(sourcepath, cgpath, "cgroup", remount_flags, NULL);
                if (ret < 0)
                        return log_error_errno(-1, errno, "Failed to remount \"%s\" ro", cgpath);
                INFO("Remounted %s read-only", cgpath);
        }

        INFO("Completed second stage cgroup automounts for \"%s\"", cgpath);
        return 0;
}

/* __cgroupfs_mount
 *
 * Mount cgroup hierarchies directly without using bind-mounts. The main
 * uses-cases are mounting cgroup hierarchies in cgroup namespaces and mounting
 * cgroups for the LXC_AUTO_CGROUP_FULL option.
 */
static int __cgroupfs_mount(int cgroup_automount_type, struct hierarchy *h,
                            struct lxc_rootfs *rootfs, int dfd_mnt_cgroupfs,
                            const char *hierarchy_mnt)
{
        __do_close int fd_fs = -EBADF;
        unsigned int flags = 0;
        char *fstype;
        int ret;

        if (dfd_mnt_cgroupfs < 0)
                return ret_errno(EINVAL);

        flags |= MOUNT_ATTR_NOSUID;
        flags |= MOUNT_ATTR_NOEXEC;
        flags |= MOUNT_ATTR_NODEV;
        flags |= MOUNT_ATTR_RELATIME;

        if ((cgroup_automount_type == LXC_AUTO_CGROUP_RO) ||
            (cgroup_automount_type == LXC_AUTO_CGROUP_FULL_RO))
                flags |= MOUNT_ATTR_RDONLY;

        if (is_unified_hierarchy(h))
                fstype = "cgroup2";
        else
                fstype = "cgroup";

        if (can_use_mount_api()) {
                fd_fs = fs_prepare(fstype, -EBADF, "", 0, 0);
                if (fd_fs < 0)
                        return log_error_errno(-errno, errno, "Failed to prepare filesystem context for %s", fstype);

                if (!is_unified_hierarchy(h)) {
                        for (const char **it = (const char **)h->controllers; it && *it; it++) {
                                if (strnequal(*it, "name=", STRLITERALLEN("name=")))
                                        ret = fs_set_property(fd_fs, "name", *it + STRLITERALLEN("name="));
                                else
                                        ret = fs_set_property(fd_fs, *it, "");
                                if (ret < 0)
                                        return log_error_errno(-errno, errno, "Failed to add %s controller to cgroup filesystem context %d(dev)", *it, fd_fs);
                        }
                }

                ret = fs_attach(fd_fs, dfd_mnt_cgroupfs, hierarchy_mnt,
                                PROTECT_OPATH_DIRECTORY, PROTECT_LOOKUP_BENEATH,
                                flags);
        } else {
                __do_free char *controllers = NULL, *target = NULL;
                unsigned int old_flags = 0;
                const char *rootfs_mnt;

                if (!is_unified_hierarchy(h)) {
                        controllers = lxc_string_join(",", (const char **)h->controllers, false);
                        if (!controllers)
                                return ret_errno(ENOMEM);
                }

                rootfs_mnt = get_rootfs_mnt(rootfs);
                ret = mnt_attributes_old(flags, &old_flags);
                if (ret)
                        return log_error_errno(-EINVAL, EINVAL, "Unsupported mount properties specified");

                target = must_make_path(rootfs_mnt, DEFAULT_CGROUP_MOUNTPOINT, hierarchy_mnt, NULL);
                ret = safe_mount(NULL, target, fstype, old_flags, controllers, rootfs_mnt);
        }
        if (ret < 0)
                return log_error_errno(ret, errno, "Failed to mount %s filesystem onto %d(%s)",
                                       fstype, dfd_mnt_cgroupfs, maybe_empty(hierarchy_mnt));

        DEBUG("Mounted cgroup filesystem %s onto %d(%s)",
              fstype, dfd_mnt_cgroupfs, maybe_empty(hierarchy_mnt));
        return 0;
}

static inline int cgroupfs_mount(int cgroup_automount_type, struct hierarchy *h,
                                 struct lxc_rootfs *rootfs,
                                 int dfd_mnt_cgroupfs, const char *hierarchy_mnt)
{
        return __cgroupfs_mount(cgroup_automount_type, h, rootfs,
                                dfd_mnt_cgroupfs, hierarchy_mnt);
}

static inline int cgroupfs_bind_mount(int cgroup_automount_type, struct hierarchy *h,
                                      struct lxc_rootfs *rootfs,
                                      int dfd_mnt_cgroupfs,
                                      const char *hierarchy_mnt)
{
        switch (cgroup_automount_type) {
        case LXC_AUTO_CGROUP_FULL_RO:
                break;
        case LXC_AUTO_CGROUP_FULL_RW:
                break;
        case LXC_AUTO_CGROUP_FULL_MIXED:
                break;
        default:
                return 0;
        }

        return __cgroupfs_mount(cgroup_automount_type, h, rootfs,
                                dfd_mnt_cgroupfs, hierarchy_mnt);
}

__cgfsng_ops static bool cgfsng_mount(struct cgroup_ops *ops,
                                      struct lxc_handler *handler, int cg_flags)
{
        __do_close int dfd_mnt_tmpfs = -EBADF, fd_fs = -EBADF;
        __do_free char *cgroup_root = NULL;
        int cgroup_automount_type;
        bool in_cgroup_ns = false, wants_force_mount = false;
        struct lxc_conf *conf = handler->conf;
        struct lxc_rootfs *rootfs = &conf->rootfs;
        const char *rootfs_mnt = get_rootfs_mnt(rootfs);
        int ret;

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return true;

        if (!conf)
                return ret_set_errno(false, EINVAL);

        if ((cg_flags & LXC_AUTO_CGROUP_MASK) == 0)
                return log_trace(true, "No cgroup mounts requested");

        if (cg_flags & LXC_AUTO_CGROUP_FORCE) {
                cg_flags &= ~LXC_AUTO_CGROUP_FORCE;
                wants_force_mount = true;
        }

        switch (cg_flags) {
        case LXC_AUTO_CGROUP_RO:
                TRACE("Read-only cgroup mounts requested");
                break;
        case LXC_AUTO_CGROUP_RW:
                TRACE("Read-write cgroup mounts requested");
                break;
        case LXC_AUTO_CGROUP_MIXED:
                TRACE("Mixed cgroup mounts requested");
                break;
        case LXC_AUTO_CGROUP_FULL_RO:
                TRACE("Full read-only cgroup mounts requested");
                break;
        case LXC_AUTO_CGROUP_FULL_RW:
                TRACE("Full read-write cgroup mounts requested");
                break;
        case LXC_AUTO_CGROUP_FULL_MIXED:
                TRACE("Full mixed cgroup mounts requested");
                break;
        default:
                return log_error_errno(false, EINVAL, "Invalid cgroup mount options specified");
        }
        cgroup_automount_type = cg_flags;

        if (!wants_force_mount) {
                wants_force_mount = !lxc_wants_cap(CAP_SYS_ADMIN, conf);

                /*
                 * Most recent distro versions currently have init system that
                 * do support cgroup2 but do not mount it by default unless
                 * explicitly told so even if the host is cgroup2 only. That
                 * means they often will fail to boot. Fix this by pre-mounting
                 * cgroup2 by default. We will likely need to be doing this a
                 * few years until all distros have switched over to cgroup2 at
                 * which point we can safely assume that their init systems
                 * will mount it themselves.
                 */
                if (pure_unified_layout(ops))
                        wants_force_mount = true;
        }

        if (cgns_supported() && container_uses_namespace(handler, CLONE_NEWCGROUP))
                in_cgroup_ns = true;

        if (in_cgroup_ns && !wants_force_mount)
                return log_trace(true, "Mounting cgroups not requested or needed");

        /* This is really the codepath that we want. */
        if (pure_unified_layout(ops)) {
                __do_close int dfd_mnt_unified = -EBADF;

                dfd_mnt_unified = open_at(rootfs->dfd_mnt, DEFAULT_CGROUP_MOUNTPOINT_RELATIVE,
                                          PROTECT_OPATH_DIRECTORY, PROTECT_LOOKUP_BENEATH_XDEV, 0);
                if (dfd_mnt_unified < 0)
                        return syserror_ret(false, "Failed to open %d(%s)",
                                            rootfs->dfd_mnt, DEFAULT_CGROUP_MOUNTPOINT_RELATIVE);
                /*
                 * If cgroup namespaces are supported but the container will
                 * not have CAP_SYS_ADMIN after it has started we need to mount
                 * the cgroups manually.
                 *
                 * Note that here we know that wants_force_mount is true.
                 * Otherwise we would've returned early above.
                 */
                if (in_cgroup_ns) {
                        /*
                         *  1. cgroup:rw:force    -> Mount the cgroup2 filesystem.
                         *  2. cgroup:ro:force    -> Mount the cgroup2 filesystem read-only.
                         *  3. cgroup:mixed:force -> See comment above how this
                         *                           does not apply so
                         *                           cgroup:mixed is equal to
                         *                           cgroup:rw when cgroup
                         *                           namespaces are supported.

                         *  4. cgroup:rw    -> No-op; init system responsible for mounting.
                         *  5. cgroup:ro    -> No-op; init system responsible for mounting.
                         *  6. cgroup:mixed -> No-op; init system responsible for mounting.
                         *
                         *  7. cgroup-full:rw    -> Not supported.
                         *  8. cgroup-full:ro    -> Not supported.
                         *  9. cgroup-full:mixed -> Not supported.

                         * 10. cgroup-full:rw:force    -> Not supported.
                         * 11. cgroup-full:ro:force    -> Not supported.
                         * 12. cgroup-full:mixed:force -> Not supported.
                         */
                        ret = cgroupfs_mount(cgroup_automount_type, ops->unified, rootfs, dfd_mnt_unified, "");
                        if (ret < 0)
                                return syserror_ret(false, "Failed to force mount cgroup filesystem in cgroup namespace");

                        return log_trace(true, "Force mounted cgroup filesystem in new cgroup namespace");
                } else {
                        /*
                         * Either no cgroup namespace supported (highly
                         * unlikely unless we're dealing with a Frankenkernel.
                         * Or the user requested to keep the cgroup namespace
                         * of the host or another container.
                         */
                        if (wants_force_mount) {
                                /*
                                 * 1. cgroup:rw:force    -> Bind-mount the cgroup2 filesystem writable.
                                 * 2. cgroup:ro:force    -> Bind-mount the cgroup2 filesystem read-only.
                                 * 3. cgroup:mixed:force -> bind-mount the cgroup2 filesystem and
                                 *                          and make the parent directory of the
                                 *                          container's cgroup read-only but the
                                 *                          container's cgroup writable.
                                 *
                                 * 10. cgroup-full:rw:force    ->
                                 * 11. cgroup-full:ro:force    ->
                                 * 12. cgroup-full:mixed:force ->
                                 */
                                errno = EOPNOTSUPP;
                                SYSWARN("Force-mounting the unified cgroup hierarchy without cgroup namespace support is currently not supported");
                        } else {
                                errno = EOPNOTSUPP;
                                SYSWARN("Mounting the unified cgroup hierarchy without cgroup namespace support is currently not supported");
                        }
                }

                return syserror_ret(false, "Failed to mount cgroups");
        }

        /*
         * Mount a tmpfs over DEFAULT_CGROUP_MOUNTPOINT. Note that we're
         * relying on RESOLVE_BENEATH so we need to skip the leading "/" in the
         * DEFAULT_CGROUP_MOUNTPOINT define.
         */
        if (can_use_mount_api()) {
                fd_fs = fs_prepare("tmpfs", -EBADF, "", 0, 0);
                if (fd_fs < 0)
                        return log_error_errno(-errno, errno, "Failed to create new filesystem context for tmpfs");

                ret = fs_set_property(fd_fs, "mode", "0755");
                if (ret < 0)
                        return log_error_errno(-errno, errno, "Failed to mount tmpfs onto %d(dev)", fd_fs);

                ret = fs_set_property(fd_fs, "size", "10240k");
                if (ret < 0)
                        return log_error_errno(-errno, errno, "Failed to mount tmpfs onto %d(dev)", fd_fs);

                ret = fs_attach(fd_fs, rootfs->dfd_mnt, DEFAULT_CGROUP_MOUNTPOINT_RELATIVE,
                                PROTECT_OPATH_DIRECTORY, PROTECT_LOOKUP_BENEATH_XDEV,
                                MOUNT_ATTR_NOSUID | MOUNT_ATTR_NODEV |
                                MOUNT_ATTR_NOEXEC | MOUNT_ATTR_RELATIME);
        } else {
                cgroup_root = must_make_path(rootfs_mnt, DEFAULT_CGROUP_MOUNTPOINT, NULL);
                ret = safe_mount(NULL, cgroup_root, "tmpfs",
                                 MS_NOSUID | MS_NODEV | MS_NOEXEC | MS_RELATIME,
                                 "size=10240k,mode=755", rootfs_mnt);
        }
        if (ret < 0)
                return log_error_errno(false, errno, "Failed to mount tmpfs on %s",
                                       DEFAULT_CGROUP_MOUNTPOINT_RELATIVE);

        dfd_mnt_tmpfs = open_at(rootfs->dfd_mnt, DEFAULT_CGROUP_MOUNTPOINT_RELATIVE,
                                PROTECT_OPATH_DIRECTORY, PROTECT_LOOKUP_BENEATH_XDEV, 0);
        if (dfd_mnt_tmpfs < 0)
                return syserror_ret(false, "Failed to open %d(%s)",
                                    rootfs->dfd_mnt, DEFAULT_CGROUP_MOUNTPOINT_RELATIVE);

        for (int i = 0; ops->hierarchies[i]; i++) {
                __do_free char *hierarchy_mnt = NULL, *path2 = NULL;
                struct hierarchy *h = ops->hierarchies[i];

                ret = mkdirat(dfd_mnt_tmpfs, h->at_mnt, 0000);
                if (ret < 0)
                        return syserror_ret(false, "Failed to create cgroup at_mnt %d(%s)", dfd_mnt_tmpfs, h->at_mnt);

                if (in_cgroup_ns && wants_force_mount) {
                        /*
                         * If cgroup namespaces are supported but the container
                         * will not have CAP_SYS_ADMIN after it has started we
                         * need to mount the cgroups manually.
                         */
                        ret = cgroupfs_mount(cgroup_automount_type, h, rootfs,
                                             dfd_mnt_tmpfs, h->at_mnt);
                        if (ret < 0)
                                return false;

                        continue;
                }

                /* Here is where the ancient kernel section begins. */
                ret = cgroupfs_bind_mount(cgroup_automount_type, h, rootfs,
                                          dfd_mnt_tmpfs, h->at_mnt);
                if (ret < 0)
                        return false;

                if (!cg_mount_needs_subdirs(cgroup_automount_type))
                        continue;

                if (!cgroup_root)
                        cgroup_root = must_make_path(rootfs_mnt, DEFAULT_CGROUP_MOUNTPOINT, NULL);

                hierarchy_mnt = must_make_path(cgroup_root, h->at_mnt, NULL);
                path2 = must_make_path(hierarchy_mnt, h->at_base,
                                       ops->container_cgroup, NULL);
                ret = mkdir_p(path2, 0755);
                if (ret < 0 && (errno != EEXIST))
                        return false;

                ret = cg_legacy_mount_controllers(cgroup_automount_type, h,
                                                  hierarchy_mnt, path2,
                                                  ops->container_cgroup);
                if (ret < 0)
                        return false;
        }

        return true;
}

/* Only root needs to escape to the cgroup of its init. */
__cgfsng_ops static bool cgfsng_criu_escape(const struct cgroup_ops *ops,
                                            struct lxc_conf *conf)
{
        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return true;

        if (!conf)
                return ret_set_errno(false, EINVAL);

        if (conf->cgroup_meta.relative || geteuid())
                return true;

        for (int i = 0; ops->hierarchies[i]; i++) {
                __do_free char *fullpath = NULL;
                int ret;

                fullpath = make_cgroup_path(ops->hierarchies[i],
                                            ops->hierarchies[i]->at_base,
                                            "cgroup.procs", NULL);
                ret = lxc_write_to_file(fullpath, "0", 2, false, 0666);
                if (ret != 0)
                        return log_error_errno(false, errno, "Failed to escape to cgroup \"%s\"", fullpath);
        }

        return true;
}

__cgfsng_ops static int cgfsng_criu_num_hierarchies(struct cgroup_ops *ops)
{
        int i = 0;

        if (!ops)
                return ret_set_errno(-1, ENOENT);

        if (!ops->hierarchies)
                return 0;

        for (; ops->hierarchies[i]; i++)
                ;

        return i;
}

__cgfsng_ops static bool cgfsng_criu_get_hierarchies(struct cgroup_ops *ops,
                                                     int n, char ***out)
{
        int i;

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return ret_set_errno(false, ENOENT);

        /* consistency check n */
        for (i = 0; i < n; i++)
                if (!ops->hierarchies[i])
                        return ret_set_errno(false, ENOENT);

        *out = ops->hierarchies[i]->controllers;

        return true;
}

static int cg_legacy_freeze(struct cgroup_ops *ops)
{
        struct hierarchy *h;

        h = get_hierarchy(ops, "freezer");
        if (!h)
                return ret_set_errno(-1, ENOENT);

        return lxc_write_openat(h->path_con, "freezer.state",
                                "FROZEN", STRLITERALLEN("FROZEN"));
}

static int freezer_cgroup_events_cb(int fd, uint32_t events, void *cbdata,
                                    struct lxc_epoll_descr *descr)
{
        __do_free char *line = NULL;
        __do_fclose FILE *f = NULL;
        int state = PTR_TO_INT(cbdata);
        size_t len;
        const char *state_string;

        f = fdopen_at(fd, "", "re", PROTECT_OPEN, PROTECT_LOOKUP_BENEATH);
        if (!f)
                return LXC_MAINLOOP_ERROR;

        if (state == 1)
                state_string = "frozen 1";
        else
                state_string = "frozen 0";

        while (getline(&line, &len, f) != -1)
                if (strnequal(line, state_string, STRLITERALLEN("frozen") + 2))
                        return LXC_MAINLOOP_CLOSE;

        rewind(f);

        return LXC_MAINLOOP_CONTINUE;
}

static int cg_unified_freeze_do(struct cgroup_ops *ops, int timeout,
                                const char *state_string,
                                int state_num,
                                const char *epoll_error,
                                const char *wait_error)
{
        __do_close int fd = -EBADF;
        call_cleaner(lxc_mainloop_close) struct lxc_epoll_descr *descr_ptr = NULL;
        int ret;
        struct lxc_epoll_descr descr;
        struct hierarchy *h;

        h = ops->unified;
        if (!h)
                return ret_set_errno(-1, ENOENT);

        if (!h->path_con)
                return ret_set_errno(-1, EEXIST);

        if (timeout != 0) {
                __do_free char *events_file = NULL;

                events_file = must_make_path(h->path_con, "cgroup.events", NULL);
                fd = open(events_file, O_RDONLY | O_CLOEXEC);
                if (fd < 0)
                        return log_error_errno(-1, errno, "Failed to open cgroup.events file");

                ret = lxc_mainloop_open(&descr);
                if (ret)
                        return log_error_errno(-1, errno, "%s", epoll_error);

                /* automatically cleaned up now */
                descr_ptr = &descr;

                ret = lxc_mainloop_add_handler_events(&descr, fd, EPOLLPRI, freezer_cgroup_events_cb, INT_TO_PTR(state_num));
                if (ret < 0)
                        return log_error_errno(-1, errno, "Failed to add cgroup.events fd handler to mainloop");
        }

        ret = lxc_write_openat(h->path_con, "cgroup.freeze", state_string, 1);
        if (ret < 0)
                return log_error_errno(-1, errno, "Failed to open cgroup.freeze file");

        if (timeout != 0 && lxc_mainloop(&descr, timeout))
                return log_error_errno(-1, errno, "%s", wait_error);

        return 0;
}

static int cg_unified_freeze(struct cgroup_ops *ops, int timeout)
{
        return cg_unified_freeze_do(ops, timeout, "1", 1,
                "Failed to create epoll instance to wait for container freeze",
                "Failed to wait for container to be frozen");
}

__cgfsng_ops static int cgfsng_freeze(struct cgroup_ops *ops, int timeout)
{
        if (!ops->hierarchies)
                return ret_set_errno(-1, ENOENT);

        if (ops->cgroup_layout != CGROUP_LAYOUT_UNIFIED)
                return cg_legacy_freeze(ops);

        return cg_unified_freeze(ops, timeout);
}

static int cg_legacy_unfreeze(struct cgroup_ops *ops)
{
        struct hierarchy *h;

        h = get_hierarchy(ops, "freezer");
        if (!h)
                return ret_set_errno(-1, ENOENT);

        return lxc_write_openat(h->path_con, "freezer.state",
                                "THAWED", STRLITERALLEN("THAWED"));
}

static int cg_unified_unfreeze(struct cgroup_ops *ops, int timeout)
{
        return cg_unified_freeze_do(ops, timeout, "0", 0,
                "Failed to create epoll instance to wait for container unfreeze",
                "Failed to wait for container to be unfrozen");
}

__cgfsng_ops static int cgfsng_unfreeze(struct cgroup_ops *ops, int timeout)
{
        if (!ops->hierarchies)
                return ret_set_errno(-1, ENOENT);

        if (ops->cgroup_layout != CGROUP_LAYOUT_UNIFIED)
                return cg_legacy_unfreeze(ops);

        return cg_unified_unfreeze(ops, timeout);
}

static const char *cgfsng_get_cgroup_do(struct cgroup_ops *ops,
                                        const char *controller, bool limiting)
{
        struct hierarchy *h;
        size_t len;
        const char *path;

        h = get_hierarchy(ops, controller);
        if (!h)
                return log_warn_errno(NULL, ENOENT,
                                      "Failed to find hierarchy for controller \"%s\"", maybe_empty(controller));

        if (limiting)
                path = h->path_lim;
        else
                path = h->path_con;
        if (!path)
                return NULL;

        len = strlen(h->at_mnt);
        if (!strnequal(h->at_mnt, DEFAULT_CGROUP_MOUNTPOINT,
                       STRLITERALLEN(DEFAULT_CGROUP_MOUNTPOINT))) {
                path += STRLITERALLEN(DEFAULT_CGROUP_MOUNTPOINT);
                path += strspn(path, "/");
        }
        return path += len;
}

__cgfsng_ops static const char *cgfsng_get_cgroup(struct cgroup_ops *ops,
                                                  const char *controller)
{
    return cgfsng_get_cgroup_do(ops, controller, false);
}

__cgfsng_ops static const char *cgfsng_get_limit_cgroup(struct cgroup_ops *ops,
                                                        const char *controller)
{
    return cgfsng_get_cgroup_do(ops, controller, true);
}

/* Given a cgroup path returned from lxc_cmd_get_cgroup_path, build a full path,
 * which must be freed by the caller.
 */
static inline char *build_full_cgpath_from_monitorpath(struct hierarchy *h,
                                                       const char *inpath,
                                                       const char *filename)
{
        return make_cgroup_path(h, inpath, filename, NULL);
}

static int cgroup_attach_leaf(const struct lxc_conf *conf, int unified_fd, pid_t pid)
{
        int idx = 1;
        int ret;
        char pidstr[INTTYPE_TO_STRLEN(int64_t) + 1];
        ssize_t pidstr_len;

        /* Create leaf cgroup. */
        ret = mkdirat(unified_fd, ".lxc", 0755);
        if (ret < 0 && errno != EEXIST)
                return log_error_errno(-errno, errno, "Failed to create leaf cgroup \".lxc\"");

        pidstr_len = strnprintf(pidstr, sizeof(pidstr), INT64_FMT, (int64_t)pid);
        if (pidstr_len < 0)
                return pidstr_len;

        ret = lxc_writeat(unified_fd, ".lxc/cgroup.procs", pidstr, pidstr_len);
        if (ret < 0)
                ret = lxc_writeat(unified_fd, "cgroup.procs", pidstr, pidstr_len);
        if (ret == 0)
                return log_trace(0, "Moved process %s into cgroup %d(.lxc)", pidstr, unified_fd);

        /* this is a non-leaf node */
        if (errno != EBUSY)
                return log_error_errno(-errno, errno, "Failed to attach to unified cgroup");

        do {
                bool rm = false;
                char attach_cgroup[STRLITERALLEN(".lxc-/cgroup.procs") + INTTYPE_TO_STRLEN(int) + 1];
                char *slash = attach_cgroup;

                ret = strnprintf(attach_cgroup, sizeof(attach_cgroup), ".lxc-%d/cgroup.procs", idx);
                if (ret < 0)
                        return ret;

                /*
                 * This shouldn't really happen but the compiler might complain
                 * that a short write would cause a buffer overrun. So be on
                 * the safe side.
                 */
                if (ret < STRLITERALLEN(".lxc-/cgroup.procs"))
                        return log_error_errno(-EINVAL, EINVAL, "Unexpected short write would cause buffer-overrun");

                slash += (ret - STRLITERALLEN("/cgroup.procs"));
                *slash = '\0';

                ret = mkdirat(unified_fd, attach_cgroup, 0755);
                if (ret < 0 && errno != EEXIST)
                        return log_error_errno(-1, errno, "Failed to create cgroup %s", attach_cgroup);
                if (ret == 0)
                        rm = true;

                *slash = '/';

                ret = lxc_writeat(unified_fd, attach_cgroup, pidstr, pidstr_len);
                if (ret == 0)
                        return log_trace(0, "Moved process %s into cgroup %d(%s)", pidstr, unified_fd, attach_cgroup);

                if (rm && unlinkat(unified_fd, attach_cgroup, AT_REMOVEDIR))
                        SYSERROR("Failed to remove cgroup \"%d(%s)\"", unified_fd, attach_cgroup);

                /* this is a non-leaf node */
                if (errno != EBUSY)
                        return log_error_errno(-1, errno, "Failed to attach to unified cgroup");

                idx++;
        } while (idx < 1000);

        return log_error_errno(-1, errno, "Failed to attach to unified cgroup");
}

static int cgroup_attach_create_leaf(const struct lxc_conf *conf,
                                     int unified_fd, int *sk_fd)
{
        __do_close int sk = *sk_fd, target_fd0 = -EBADF, target_fd1 = -EBADF;
        int target_fds[2];
        ssize_t ret;

        /* Create leaf cgroup. */
        ret = mkdirat(unified_fd, ".lxc", 0755);
        if (ret < 0 && errno != EEXIST)
                return log_error_errno(-1, errno, "Failed to create leaf cgroup \".lxc\"");

        target_fd0 = open_at(unified_fd, ".lxc/cgroup.procs", PROTECT_OPEN_W, PROTECT_LOOKUP_BENEATH, 0);
        if (target_fd0 < 0)
                return log_error_errno(-errno, errno, "Failed to open \".lxc/cgroup.procs\"");
        target_fds[0] = target_fd0;

        target_fd1 = open_at(unified_fd, "cgroup.procs", PROTECT_OPEN_W, PROTECT_LOOKUP_BENEATH, 0);
        if (target_fd1 < 0)
                return log_error_errno(-errno, errno, "Failed to open \".lxc/cgroup.procs\"");
        target_fds[1] = target_fd1;

        ret = lxc_abstract_unix_send_fds(sk, target_fds, 2, NULL, 0);
        if (ret <= 0)
                return log_error_errno(-errno, errno, "Failed to send \".lxc/cgroup.procs\" fds %d and %d",
                                       target_fd0, target_fd1);

        return log_debug(0, "Sent target cgroup fds %d and %d", target_fd0, target_fd1);
}

static int cgroup_attach_move_into_leaf(const struct lxc_conf *conf,
                                        int *sk_fd, pid_t pid)
{
        __do_close int sk = *sk_fd, target_fd0 = -EBADF, target_fd1 = -EBADF;
        char pidstr[INTTYPE_TO_STRLEN(int64_t) + 1];
        size_t pidstr_len;
        ssize_t ret;

        ret = lxc_abstract_unix_recv_two_fds(sk, &target_fd0, &target_fd1);
        if (ret < 0)
                return log_error_errno(-1, errno, "Failed to receive target cgroup fd");

        pidstr_len = sprintf(pidstr, INT64_FMT, (int64_t)pid);

        ret = lxc_write_nointr(target_fd0, pidstr, pidstr_len);
        if (ret > 0 && ret == pidstr_len)
                return log_debug(0, "Moved process into target cgroup via fd %d", target_fd0);

        ret = lxc_write_nointr(target_fd1, pidstr, pidstr_len);
        if (ret > 0 && ret == pidstr_len)
                return log_debug(0, "Moved process into target cgroup via fd %d", target_fd1);

        return log_debug_errno(-1, errno, "Failed to move process into target cgroup via fd %d and %d",
                               target_fd0, target_fd1);
}

struct userns_exec_unified_attach_data {
        const struct lxc_conf *conf;
        int unified_fd;
        int sk_pair[2];
        pid_t pid;
};

static int cgroup_unified_attach_child_wrapper(void *data)
{
        struct userns_exec_unified_attach_data *args = data;

        if (!args->conf || args->unified_fd < 0 || args->pid <= 0 ||
            args->sk_pair[0] < 0 || args->sk_pair[1] < 0)
                return ret_errno(EINVAL);

        close_prot_errno_disarm(args->sk_pair[0]);
        return cgroup_attach_create_leaf(args->conf, args->unified_fd,
                                         &args->sk_pair[1]);
}

static int cgroup_unified_attach_parent_wrapper(void *data)
{
        struct userns_exec_unified_attach_data *args = data;

        if (!args->conf || args->unified_fd < 0 || args->pid <= 0 ||
            args->sk_pair[0] < 0 || args->sk_pair[1] < 0)
                return ret_errno(EINVAL);

        close_prot_errno_disarm(args->sk_pair[1]);
        return cgroup_attach_move_into_leaf(args->conf, &args->sk_pair[0],
                                            args->pid);
}

/* Technically, we're always at a delegation boundary here (This is especially
 * true when cgroup namespaces are available.). The reasoning is that in order
 * for us to have been able to start a container in the first place the root
 * cgroup must have been a leaf node. Now, either the container's init system
 * has populated the cgroup and kept it as a leaf node or it has created
 * subtrees. In the former case we will simply attach to the leaf node we
 * created when we started the container in the latter case we create our own
 * cgroup for the attaching process.
 */
static int __cg_unified_attach(const struct hierarchy *h,
                               const struct lxc_conf *conf, const char *name,
                               const char *lxcpath, pid_t pid,
                               const char *controller)
{
        __do_close int unified_fd = -EBADF;
        __do_free char *path = NULL, *cgroup = NULL;
        int ret;

        if (!conf || !name || !lxcpath || pid <= 0)
                return ret_errno(EINVAL);

        ret = cgroup_attach(conf, name, lxcpath, pid);
        if (ret == 0)
                return log_trace(0, "Attached to unified cgroup via command handler");
        if (!ERRNO_IS_NOT_SUPPORTED(ret) && ret != -ENOCGROUP2)
                return log_error_errno(ret, errno, "Failed to attach to unified cgroup");

        /* Fall back to retrieving the path for the unified cgroup. */
        cgroup = lxc_cmd_get_cgroup_path(name, lxcpath, controller);
        /* not running */
        if (!cgroup)
                return 0;

        path = make_cgroup_path(h, cgroup, NULL);

        unified_fd = open(path, O_PATH | O_DIRECTORY | O_CLOEXEC);
        if (unified_fd < 0)
                return ret_errno(EBADF);

        if (!lxc_list_empty(&conf->id_map)) {
                struct userns_exec_unified_attach_data args = {
                        .conf           = conf,
                        .unified_fd     = unified_fd,
                        .pid            = pid,
                };

                ret = socketpair(PF_LOCAL, SOCK_STREAM | SOCK_CLOEXEC, 0, args.sk_pair);
                if (ret < 0)
                        return -errno;

                ret = userns_exec_minimal(conf,
                                          cgroup_unified_attach_parent_wrapper,
                                          &args,
                                          cgroup_unified_attach_child_wrapper,
                                          &args);
        } else {
                ret = cgroup_attach_leaf(conf, unified_fd, pid);
        }

        return ret;
}

__cgfsng_ops static bool cgfsng_attach(struct cgroup_ops *ops,
                                       const struct lxc_conf *conf,
                                       const char *name, const char *lxcpath,
                                       pid_t pid)
{
        int len, ret;
        char pidstr[INTTYPE_TO_STRLEN(pid_t)];

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return true;

        len = strnprintf(pidstr, sizeof(pidstr), "%d", pid);
        if (len < 0)
                return false;

        for (int i = 0; ops->hierarchies[i]; i++) {
                __do_free char *fullpath = NULL, *path = NULL;
                struct hierarchy *h = ops->hierarchies[i];

                if (h->fs_type == UNIFIED_HIERARCHY) {
                        ret = __cg_unified_attach(h, conf, name, lxcpath, pid,
                                                  h->controllers[0]);
                        if (ret < 0)
                                return false;

                        continue;
                }

                path = lxc_cmd_get_cgroup_path(name, lxcpath, h->controllers[0]);
                if (!path) {
                        /*
                         * Someone might have created a name=<controller>
                         * controller after the container has started and so
                         * the container doesn't make use of this controller.
                         *
                         * Link: https://github.com/lxc/lxd/issues/8577
                         */
                        TRACE("Skipping unused %s controller", maybe_empty(h->controllers[0]));
                        continue;
                }

                fullpath = build_full_cgpath_from_monitorpath(h, path, "cgroup.procs");
                ret = lxc_write_to_file(fullpath, pidstr, len, false, 0666);
                if (ret < 0)
                        return log_error_errno(false, errno, "Failed to attach %d to %s",
                                               (int)pid, fullpath);
        }

        return true;
}

/* Called externally (i.e. from 'lxc-cgroup') to query cgroup limits.  Here we
 * don't have a cgroup_data set up, so we ask the running container through the
 * commands API for the cgroup path.
 */
__cgfsng_ops static int cgfsng_get(struct cgroup_ops *ops, const char *filename,
                                     char *value, size_t len, const char *name,
                                     const char *lxcpath)
{
        __do_free char *path = NULL;
        __do_free char *controller = NULL;
        char *p;
        struct hierarchy *h;
        int ret = -1;

        if (!ops)
                return ret_set_errno(-1, ENOENT);

        controller = strdup(filename);
        if (!controller)
                return ret_errno(ENOMEM);

        p = strchr(controller, '.');
        if (p)
                *p = '\0';

        path = lxc_cmd_get_limit_cgroup_path(name, lxcpath, controller);
        /* not running */
        if (!path)
                return -1;

        h = get_hierarchy(ops, controller);
        if (h) {
                __do_free char *fullpath = NULL;

                fullpath = build_full_cgpath_from_monitorpath(h, path, filename);
                ret = lxc_read_from_file(fullpath, value, len);
        }

        return ret;
}

static int device_cgroup_parse_access(struct device_item *device, const char *val)
{
        for (int count = 0; count < 3; count++, val++) {
                switch (*val) {
                case 'r':
                        device->access[count] = *val;
                        break;
                case 'w':
                        device->access[count] = *val;
                        break;
                case 'm':
                        device->access[count] = *val;
                        break;
                case '\n':
                case '\0':
                        count = 3;
                        break;
                default:
                        return ret_errno(EINVAL);
                }
        }

        return 0;
}

static int device_cgroup_rule_parse(struct device_item *device, const char *key,
                                    const char *val)
{
        int count, ret;
        char temp[50];

        if (strequal("devices.allow", key))
                device->allow = 1; /* allow the device */
        else
                device->allow = 0; /* deny the device */

        if (strequal(val, "a")) {
                /* global rule */
                device->type = 'a';
                device->major = -1;
                device->minor = -1;
                return 0;
        }

        switch (*val) {
        case 'a':
                __fallthrough;
        case 'b':
                __fallthrough;
        case 'c':
                device->type = *val;
                break;
        default:
                return -1;
        }

        val++;
        if (!isspace(*val))
                return -1;
        val++;
        if (*val == '*') {
                device->major = -1;
                val++;
        } else if (isdigit(*val)) {
                memset(temp, 0, sizeof(temp));
                for (count = 0; count < sizeof(temp) - 1; count++) {
                        temp[count] = *val;
                        val++;
                        if (!isdigit(*val))
                                break;
                }
                ret = lxc_safe_int(temp, &device->major);
                if (ret)
                        return -1;
        } else {
                return -1;
        }
        if (*val != ':')
                return -1;
        val++;

        /* read minor */
        if (*val == '*') {
                device->minor = -1;
                val++;
        } else if (isdigit(*val)) {
                memset(temp, 0, sizeof(temp));
                for (count = 0; count < sizeof(temp) - 1; count++) {
                        temp[count] = *val;
                        val++;
                        if (!isdigit(*val))
                                break;
                }
                ret = lxc_safe_int(temp, &device->minor);
                if (ret)
                        return -1;
        } else {
                return -1;
        }
        if (!isspace(*val))
                return -1;

        return device_cgroup_parse_access(device, ++val);
}

/* Called externally (i.e. from 'lxc-cgroup') to set new cgroup limits.  Here we
 * don't have a cgroup_data set up, so we ask the running container through the
 * commands API for the cgroup path.
 */
__cgfsng_ops static int cgfsng_set(struct cgroup_ops *ops,
                                     const char *key, const char *value,
                                     const char *name, const char *lxcpath)
{
        __do_free char *path = NULL;
        __do_free char *controller = NULL;
        char *p;
        struct hierarchy *h;
        int ret = -1;

        if (!ops || is_empty_string(key) || is_empty_string(value) ||
            is_empty_string(name) || is_empty_string(lxcpath))
                return ret_errno(EINVAL);

        controller = strdup(key);
        if (!controller)
                return ret_errno(ENOMEM);

        p = strchr(controller, '.');
        if (p)
                *p = '\0';

        if (pure_unified_layout(ops) && strequal(controller, "devices")) {
                struct device_item device = {};

                ret = device_cgroup_rule_parse(&device, key, value);
                if (ret < 0)
                        return log_error_errno(-1, EINVAL, "Failed to parse device string %s=%s",
                                               key, value);

                ret = lxc_cmd_add_bpf_device_cgroup(name, lxcpath, &device);
                if (ret < 0)
                        return -1;

                return 0;
        }

        path = lxc_cmd_get_limit_cgroup_path(name, lxcpath, controller);
        /* not running */
        if (!path)
                return -1;

        h = get_hierarchy(ops, controller);
        if (h) {
                __do_free char *fullpath = NULL;

                fullpath = build_full_cgpath_from_monitorpath(h, path, key);
                ret = lxc_write_to_file(fullpath, value, strlen(value), false, 0666);
        }

        return ret;
}

/* take devices cgroup line
 *    /dev/foo rwx
 * and convert it to a valid
 *    type major:minor mode
 * line. Return <0 on error. Dest is a preallocated buffer long enough to hold
 * the output.
 */
static int device_cgroup_rule_parse_devpath(struct device_item *device,
                                            const char *devpath)
{
        __do_free char *path = NULL;
        char *mode = NULL;
        int n_parts, ret;
        char *p;
        struct stat sb;

        path = strdup(devpath);
        if (!path)
                return ret_errno(ENOMEM);

        /*
         * Read path followed by mode. Ignore any trailing text.
         * A '    # comment' would be legal. Technically other text is not
         * legal, we could check for that if we cared to.
         */
        for (n_parts = 1, p = path; *p; p++) {
                if (*p != ' ')
                        continue;
                *p = '\0';

                if (n_parts != 1)
                        break;
                p++;
                n_parts++;

                while (*p == ' ')
                        p++;

                mode = p;

                if (*p == '\0')
                        return ret_set_errno(-1, EINVAL);
        }

        if (!mode)
                return ret_errno(EINVAL);

        if (device_cgroup_parse_access(device, mode) < 0)
                return -1;

        ret = stat(path, &sb);
        if (ret < 0)
                return ret_set_errno(-1, errno);

        mode_t m = sb.st_mode & S_IFMT;
        switch (m) {
        case S_IFBLK:
                device->type = 'b';
                break;
        case S_IFCHR:
                device->type = 'c';
                break;
        default:
                return log_error_errno(-1, EINVAL, "Unsupported device type %i for \"%s\"", m, path);
        }

        device->major = MAJOR(sb.st_rdev);
        device->minor = MINOR(sb.st_rdev);
        device->allow = 1;

        return 0;
}

static int convert_devpath(const char *invalue, char *dest)
{
        struct device_item device = {};
        int ret;

        ret = device_cgroup_rule_parse_devpath(&device, invalue);
        if (ret < 0)
                return -1;

        ret = strnprintf(dest, 50, "%c %d:%d %s", device.type, device.major,
                         device.minor, device.access);
        if (ret < 0)
                return log_error_errno(ret, -ret,
                                       "Error on configuration value \"%c %d:%d %s\" (max 50 chars)",
                                       device.type, device.major, device.minor,
                                       device.access);

        return 0;
}

/* Called from setup_limits - here we have the container's cgroup_data because
 * we created the cgroups.
 */
static int cg_legacy_set_data(struct cgroup_ops *ops, const char *filename,
                              const char *value, bool is_cpuset)
{
        __do_free char *controller = NULL;
        char *p;
        /* "b|c <2^64-1>:<2^64-1> r|w|m" = 47 chars max */
        char converted_value[50];
        struct hierarchy *h;

        controller = strdup(filename);
        if (!controller)
                return ret_errno(ENOMEM);

        p = strchr(controller, '.');
        if (p)
                *p = '\0';

        if (strequal("devices.allow", filename) && value[0] == '/') {
                int ret;

                ret = convert_devpath(value, converted_value);
                if (ret < 0)
                        return ret;
                value = converted_value;
        }

        h = get_hierarchy(ops, controller);
        if (!h)
                return log_error_errno(-ENOENT, ENOENT, "Failed to setup limits for the \"%s\" controller. The controller seems to be unused by \"cgfsng\" cgroup driver or not enabled on the cgroup hierarchy", controller);

        if (is_cpuset) {
                int ret = lxc_write_openat(h->path_con, filename, value, strlen(value));
                if (ret)
                        return ret;
        }
        return lxc_write_openat(h->path_lim, filename, value, strlen(value));
}

__cgfsng_ops static bool cgfsng_setup_limits_legacy(struct cgroup_ops *ops,
                                                    struct lxc_conf *conf,
                                                    bool do_devices)
{
        __do_free struct lxc_list *sorted_cgroup_settings = NULL;
        struct lxc_list *cgroup_settings = &conf->cgroup;
        struct lxc_list *iterator, *next;
        struct lxc_cgroup *cg;
        bool ret = false;

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!conf)
                return ret_set_errno(false, EINVAL);

        cgroup_settings = &conf->cgroup;
        if (lxc_list_empty(cgroup_settings))
                return true;

        if (!ops->hierarchies)
                return ret_set_errno(false, EINVAL);

        if (pure_unified_layout(ops))
                return log_warn_errno(true, EINVAL, "Ignoring legacy cgroup limits on pure cgroup2 system");

        sorted_cgroup_settings = sort_cgroup_settings(cgroup_settings);
        if (!sorted_cgroup_settings)
                return false;

        lxc_list_for_each(iterator, sorted_cgroup_settings) {
                cg = iterator->elem;

                if (do_devices == strnequal("devices", cg->subsystem, 7)) {
                        if (cg_legacy_set_data(ops, cg->subsystem, cg->value, strnequal("cpuset", cg->subsystem, 6))) {
                                if (do_devices && (errno == EACCES || errno == EPERM)) {
                                        SYSWARN("Failed to set \"%s\" to \"%s\"", cg->subsystem, cg->value);
                                        continue;
                                }
                                SYSERROR("Failed to set \"%s\" to \"%s\"", cg->subsystem, cg->value);
                                goto out;
                        }
                        DEBUG("Set controller \"%s\" set to \"%s\"", cg->subsystem, cg->value);
                }
        }

        ret = true;
        INFO("Limits for the legacy cgroup hierarchies have been setup");
out:
        lxc_list_for_each_safe(iterator, sorted_cgroup_settings, next) {
                lxc_list_del(iterator);
                free(iterator);
        }

        return ret;
}

/*
 * Some of the parsing logic comes from the original cgroup device v1
 * implementation in the kernel.
 */
static int bpf_device_cgroup_prepare(struct cgroup_ops *ops,
                                     struct lxc_conf *conf, const char *key,
                                     const char *val)
{
        struct device_item device_item = {};
        int ret;

        if (strequal("devices.allow", key) && abspath(val))
                ret = device_cgroup_rule_parse_devpath(&device_item, val);
        else
                ret = device_cgroup_rule_parse(&device_item, key, val);
        if (ret < 0)
                return syserror_set(EINVAL, "Failed to parse device rule %s=%s", key, val);

        /*
         * Note that bpf_list_add_device() returns 1 if it altered the device
         * list and 0 if it didn't; both return values indicate success.
         * Only a negative return value indicates an error.
         */
        ret = bpf_list_add_device(&conf->bpf_devices, &device_item);
        if (ret < 0)
                return -1;

        return 0;
}

__cgfsng_ops static bool cgfsng_setup_limits(struct cgroup_ops *ops,
                                             struct lxc_handler *handler)
{
        struct lxc_list *cgroup_settings, *iterator;
        struct hierarchy *h;
        struct lxc_conf *conf;

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return true;

        if (!ops->container_cgroup)
                return ret_set_errno(false, EINVAL);

        if (!handler || !handler->conf)
                return ret_set_errno(false, EINVAL);
        conf = handler->conf;

        cgroup_settings = &conf->cgroup2;
        if (lxc_list_empty(cgroup_settings))
                return true;

        if (!pure_unified_layout(ops))
                return log_warn_errno(true, EINVAL, "Ignoring cgroup2 limits on legacy cgroup system");

        if (!ops->unified)
                return false;
        h = ops->unified;

        lxc_list_for_each (iterator, cgroup_settings) {
                struct lxc_cgroup *cg = iterator->elem;
                int ret;

                if (strnequal("devices", cg->subsystem, 7))
                        ret = bpf_device_cgroup_prepare(ops, conf, cg->subsystem, cg->value);
                else
                        ret = lxc_write_openat(h->path_lim, cg->subsystem, cg->value, strlen(cg->value));
                if (ret < 0)
                        return log_error_errno(false, errno, "Failed to set \"%s\" to \"%s\"", cg->subsystem, cg->value);

                TRACE("Set \"%s\" to \"%s\"", cg->subsystem, cg->value);
        }

        return log_info(true, "Limits for the unified cgroup hierarchy have been setup");
}

__cgfsng_ops static bool cgfsng_devices_activate(struct cgroup_ops *ops, struct lxc_handler *handler)
{
        struct lxc_conf *conf;
        struct hierarchy *unified;

        if (!ops)
                return ret_set_errno(false, ENOENT);

        if (!ops->hierarchies)
                return true;

        if (!ops->container_cgroup)
                return ret_set_errno(false, EEXIST);

        if (!handler || !handler->conf)
                return ret_set_errno(false, EINVAL);
        conf = handler->conf;

        unified = ops->unified;
        if (!unified || !device_utility_controller(unified) ||
            !unified->path_con ||
            lxc_list_empty(&(conf->bpf_devices).device_item))
                return true;

        return bpf_cgroup_devices_attach(ops, &conf->bpf_devices);
}

static bool __cgfsng_delegate_controllers(struct cgroup_ops *ops, const char *cgroup)
{
        __do_close int dfd_final = -EBADF;
        __do_free char *add_controllers = NULL, *copy = NULL;
        size_t full_len = 0;
        struct hierarchy *unified;
        int dfd_cur, ret;
        char *cur;
        char **it;

        if (!ops->hierarchies || !pure_unified_layout(ops))
                return true;

        unified = ops->unified;
        if (!unified->controllers[0])
                return true;

        /* For now we simply enable all controllers that we have detected by
         * creating a string like "+memory +pids +cpu +io".
         * TODO: In the near future we might want to support "-<controller>"
         * etc. but whether supporting semantics like this make sense will need
         * some thinking.
         */
        for (it = unified->controllers; it && *it; it++) {
                full_len += strlen(*it) + 2;
                add_controllers = must_realloc(add_controllers, full_len + 1);

                if (unified->controllers[0] == *it)
                        add_controllers[0] = '\0';

                (void)strlcat(add_controllers, "+", full_len + 1);
                (void)strlcat(add_controllers, *it, full_len + 1);

                if ((it + 1) && *(it + 1))
                        (void)strlcat(add_controllers, " ", full_len + 1);
        }

        copy = strdup(cgroup);
        if (!copy)
                return false;

        /*
         * Placing the write to cgroup.subtree_control before the open() is
         * intentional because of the cgroup2 delegation model. It enforces
         * that leaf cgroups don't have any controllers enabled for delegation.
         */
        dfd_cur = unified->dfd_base;
        lxc_iterate_parts(cur, copy, "/") {
                /*
                 * Even though we vetted the paths when we parsed the config
                 * we're paranoid here and check that the path is neither
                 * absolute nor walks upwards.
                 */
                if (abspath(cur))
                        return syserror_set(-EINVAL, "No absolute paths allowed");

                if (strnequal(cur, "..", STRLITERALLEN("..")))
                        return syserror_set(-EINVAL, "No upward walking paths allowed");

                ret = lxc_writeat(dfd_cur, "cgroup.subtree_control", add_controllers, full_len);
                if (ret < 0)
                        return syserror("Could not enable \"%s\" controllers in the unified cgroup %d", add_controllers, dfd_cur);

                TRACE("Enabled \"%s\" controllers in the unified cgroup %d", add_controllers, dfd_cur);

                dfd_final = open_at(dfd_cur, cur, PROTECT_OPATH_DIRECTORY, PROTECT_LOOKUP_BENEATH, 0);
                if (dfd_final < 0)
                        return syserror("Fail to open directory %d(%s)", dfd_cur, cur);
                if (dfd_cur != unified->dfd_base)
                        close(dfd_cur);
                /*
                 * Leave dfd_final pointing to the last fd we opened so
                 * it will be automatically zapped if we return early.
                 */
                dfd_cur = dfd_final;
        }

        return true;
}

__cgfsng_ops static bool cgfsng_monitor_delegate_controllers(struct cgroup_ops *ops)
{
        if (!ops)
                return ret_set_errno(false, ENOENT);

        return __cgfsng_delegate_controllers(ops, ops->monitor_cgroup);
}

__cgfsng_ops static bool cgfsng_payload_delegate_controllers(struct cgroup_ops *ops)
{
        if (!ops)
                return ret_set_errno(false, ENOENT);

        return __cgfsng_delegate_controllers(ops, ops->container_cgroup);
}

static inline bool unified_cgroup(const char *line)
{
        return *line == '0';
}

static inline char *current_unified_cgroup(bool relative, char *line)
{
        char *current_cgroup;

        line += STRLITERALLEN("0::");

        if (!abspath(line))
                return ERR_PTR(-EINVAL);

        /* remove init.scope */
        if (!relative)
                line = prune_init_scope(line);

        /* create a relative path */
        line = deabs(line);

        current_cgroup = strdup(line);
        if (!current_cgroup)
                return ERR_PTR(-ENOMEM);

        return current_cgroup;
}

static inline const char *unprefix(const char *controllers)
{
        if (strnequal(controllers, "name=", STRLITERALLEN("name=")))
                return controllers + STRLITERALLEN("name=");
        return controllers;
}

static int __list_cgroup_delegate(char ***delegate)
{
        __do_free char **list = NULL;
        __do_free char *buf = NULL;
        char *standard[] = {
                "cgroup.procs",
                "cgroup.threads",
                "cgroup.subtree_control",
                "memory.oom.group",
                NULL,
        };
        char *token;
        int ret;

        buf = read_file_at(-EBADF, "/sys/kernel/cgroup/delegate", PROTECT_OPEN, 0);
        if (!buf) {
                for (char **p = standard; p && *p; p++) {
                        ret = list_add_string(&list, *p);
                        if (ret < 0)
                                return ret;
                }

                *delegate = move_ptr(list);
                return syswarn_ret(0, "Failed to read /sys/kernel/cgroup/delegate");
        }

        lxc_iterate_parts(token, buf, " \t\n") {
                /*
                 * We always need to chown this for both cgroup and
                 * cgroup2.
                 */
                if (strequal(token, "cgroup.procs"))
                        continue;

                ret = list_add_string(&list, token);
                if (ret < 0)
                        return ret;
        }

        *delegate = move_ptr(list);
        return 0;
}

static bool unified_hierarchy_delegated(int dfd_base, char ***ret_files)
{
        __do_free_string_list char **list = NULL;
        int ret;

        ret = __list_cgroup_delegate(&list);
        if (ret < 0)
                return syserror_ret(ret, "Failed to determine unified cgroup delegation requirements");

        for (char *const *s = list; s && *s; s++) {
                if (!faccessat(dfd_base, *s, W_OK, 0) || errno == ENOENT)
                        continue;

                return sysinfo_ret(false, "The %s file is not writable, skipping unified hierarchy", *s);
        }

        *ret_files = move_ptr(list);
        return true;
}

static bool legacy_hierarchy_delegated(int dfd_base)
{
        int ret;

        ret = faccessat(dfd_base, ".", W_OK, 0);
        if (ret < 0 && errno != ENOENT)
                return sysinfo_ret(false, "Legacy hierarchy not writable, skipping");

        return true;
}

/**
 * systemd guarantees that the order of co-mounted controllers is stable. On
 * some systems the order of the controllers might be reversed though.
 *
 * For example, this is how the order is mismatched on CentOS 7:
 *
 *      [root@localhost ~]# cat /proc/self/cgroup
 *      11:perf_event:/
 *      10:pids:/
 *      9:freezer:/
 * >>>> 8:cpuacct,cpu:/
 *      7:memory:/
 *      6:blkio:/
 *      5:devices:/
 *      4:hugetlb:/
 * >>>> 3:net_prio,net_cls:/
 *      2:cpuset:/
 *      1:name=systemd:/user.slice/user-0.slice/session-c1.scope
 *
 * whereas the mountpoint:
 *
 *      | |-/sys/fs/cgroup                    tmpfs         tmpfs      ro,nosuid,nodev,noexec,mode=755
 *      | | |-/sys/fs/cgroup/systemd          cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,xattr,release_agent=/usr/lib/systemd/systemd-cgroups-agent,name=systemd
 *      | | |-/sys/fs/cgroup/cpuset           cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,cpuset
 * >>>> | | |-/sys/fs/cgroup/net_cls,net_prio cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,net_prio,net_cls
 *      | | |-/sys/fs/cgroup/hugetlb          cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,hugetlb
 *      | | |-/sys/fs/cgroup/devices          cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,devices
 *      | | |-/sys/fs/cgroup/blkio            cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,blkio
 *      | | |-/sys/fs/cgroup/memory           cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,memory
 * >>>> | | |-/sys/fs/cgroup/cpu,cpuacct      cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,cpuacct,cpu
 *      | | |-/sys/fs/cgroup/freezer          cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,freezer
 *      | | |-/sys/fs/cgroup/pids             cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,pids
 *      | | `-/sys/fs/cgroup/perf_event       cgroup        cgroup     rw,nosuid,nodev,noexec,relatime,perf_event
 *
 * Ensure that we always use the systemd-guaranteed stable order when checking
 * for the mountpoint.
 */
__attribute__((returns_nonnull)) __attribute__((nonnull))
static const char *stable_order(const char *controllers)
{
        if (strequal(controllers, "cpuacct,cpu"))
                return "cpu,cpuacct";

        if (strequal(controllers, "net_prio,net_cls"))
                return "net_cls,net_prio";

        return unprefix(controllers);
}

static int __initialize_cgroups(struct cgroup_ops *ops, bool relative,
                                bool unprivileged)
{
        __do_free char *cgroup_info = NULL;
        char *it;

        /*
         * Root spawned containers escape the current cgroup, so use init's
         * cgroups as our base in that case.
         */
        if (!relative && (geteuid() == 0))
                cgroup_info = read_file_at(-EBADF, "/proc/1/cgroup", PROTECT_OPEN, 0);
        else
                cgroup_info = read_file_at(-EBADF, "/proc/self/cgroup", PROTECT_OPEN, 0);
        if (!cgroup_info)
                return ret_errno(ENOMEM);

        lxc_iterate_parts(it, cgroup_info, "\n") {
                __do_close int dfd_base = -EBADF, dfd_mnt = -EBADF;
                __do_free char *controllers = NULL, *current_cgroup = NULL;
                __do_free_string_list char **controller_list = NULL,
                                           **delegate = NULL;
                char *line;
                int dfd, ret, type;

                /* Handle the unified cgroup hierarchy. */
                line = it;
                if (unified_cgroup(line)) {
                        char *unified_mnt;

                        type = UNIFIED_HIERARCHY;

                        current_cgroup = current_unified_cgroup(relative, line);
                        if (IS_ERR(current_cgroup))
                                return PTR_ERR(current_cgroup);

                        if (unified_cgroup_fd(ops->dfd_mnt)) {
                                dfd_mnt = dup_cloexec(ops->dfd_mnt);
                                unified_mnt = "";
                        } else {
                                dfd_mnt = open_at(ops->dfd_mnt,
                                                  "unified",
                                                  PROTECT_OPATH_DIRECTORY,
                                                  PROTECT_LOOKUP_ABSOLUTE_XDEV, 0);
                                unified_mnt = "unified";
                        }
                        if (dfd_mnt < 0) {
                                if (errno != ENOENT)
                                        return syserror("Failed to open %d/unified", ops->dfd_mnt);

                                SYSTRACE("Unified cgroup not mounted");
                                continue;
                        }
                        dfd = dfd_mnt;

                        if (!is_empty_string(current_cgroup)) {
                                dfd_base = open_at(dfd_mnt, current_cgroup,
                                                   PROTECT_OPATH_DIRECTORY,
                                                   PROTECT_LOOKUP_BENEATH_XDEV, 0);
                                if (dfd_base < 0) {
                                        if (errno != ENOENT)
                                                return syserror("Failed to open %d/%s",
                                                                dfd_mnt, current_cgroup);

                                        SYSTRACE("Current cgroup %d/%s does not exist (funky cgroup layout?)",
                                                 dfd_mnt, current_cgroup);
                                        continue;
                                }
                                dfd = dfd_base;
                        }

                        if (!unified_hierarchy_delegated(dfd, &delegate))
                                continue;

                        controller_list = unified_controllers(dfd, "cgroup.controllers");
                        if (!controller_list) {
                                TRACE("No controllers are enabled for delegation in the unified hierarchy");
                                controller_list = list_new();
                                if (!controller_list)
                                        return syserror_set(-ENOMEM, "Failed to create empty controller list");
                        }

                        controllers = strdup(unified_mnt);
                        if (!controllers)
                                return ret_errno(ENOMEM);
                } else {
                        char *__controllers, *__current_cgroup;

                        type = LEGACY_HIERARCHY;

                        __controllers = strchr(line, ':');
                        if (!__controllers)
                                return ret_errno(EINVAL);
                        __controllers++;

                        __current_cgroup = strchr(__controllers, ':');
                        if (!__current_cgroup)
                                return ret_errno(EINVAL);
                        *__current_cgroup = '\0';
                        __current_cgroup++;

                        controllers = strdup(stable_order(__controllers));
                        if (!controllers)
                                return ret_errno(ENOMEM);

                        dfd_mnt = open_at(ops->dfd_mnt,
                                          controllers,
                                          PROTECT_OPATH_DIRECTORY,
                                          PROTECT_LOOKUP_ABSOLUTE_XDEV, 0);
                        if (dfd_mnt < 0) {
                                if (errno != ENOENT)
                                        return syserror("Failed to open %d/%s",
                                                        ops->dfd_mnt, controllers);

                                SYSTRACE("%s not mounted", controllers);
                                continue;
                        }
                        dfd = dfd_mnt;

                        if (!abspath(__current_cgroup))
                                return ret_errno(EINVAL);

                        /* remove init.scope */
                        if (!relative)
                                __current_cgroup = prune_init_scope(__current_cgroup);

                        /* create a relative path */
                        __current_cgroup = deabs(__current_cgroup);

                        current_cgroup = strdup(__current_cgroup);
                        if (!current_cgroup)
                                return ret_errno(ENOMEM);

                        if (!is_empty_string(current_cgroup)) {
                                dfd_base = open_at(dfd_mnt, current_cgroup,
                                                   PROTECT_OPATH_DIRECTORY,
                                                   PROTECT_LOOKUP_BENEATH_XDEV, 0);
                                if (dfd_base < 0) {
                                        if (errno != ENOENT)
                                                return syserror("Failed to open %d/%s",
                                                                dfd_mnt, current_cgroup);

                                        SYSTRACE("Current cgroup %d/%s does not exist (funky cgroup layout?)",
                                                 dfd_mnt, current_cgroup);
                                        continue;
                                }
                                dfd = dfd_base;
                        }

                        if (!legacy_hierarchy_delegated(dfd))
                                continue;

                        /*
                         * We intentionally pass __current_cgroup here and not
                         * controllers because we would otherwise chop the
                         * mountpoint.
                         */
                        controller_list = list_add_controllers(__controllers);
                        if (!controller_list)
                                return syserror_set(-ENOMEM, "Failed to create controller list from %s", __controllers);

                        if (skip_hierarchy(ops, controller_list))
                                continue;

                        ops->cgroup_layout = CGROUP_LAYOUT_LEGACY;
                }

                ret = cgroup_hierarchy_add(ops, dfd_mnt, controllers, dfd,
                                           current_cgroup, controller_list, type);
                if (ret < 0)
                        return syserror_ret(ret, "Failed to add %s hierarchy", controllers);

                /* Transfer ownership. */
                move_fd(dfd_mnt);
                move_fd(dfd_base);
                move_ptr(current_cgroup);
                move_ptr(controllers);
                move_ptr(controller_list);
                if (type == UNIFIED_HIERARCHY)
                        ops->unified->delegate = move_ptr(delegate);
        }

        /* determine cgroup layout */
        if (ops->unified) {
                if (ops->cgroup_layout == CGROUP_LAYOUT_LEGACY) {
                        ops->cgroup_layout = CGROUP_LAYOUT_HYBRID;
                } else {
                        if (bpf_devices_cgroup_supported())
                                ops->unified->utilities |= DEVICES_CONTROLLER;
                        ops->cgroup_layout = CGROUP_LAYOUT_UNIFIED;
                }
        }

        if (!controllers_available(ops))
                return syserror_set(-ENOENT, "One or more requested controllers unavailable or not delegated");

        return 0;
}

static int initialize_cgroups(struct cgroup_ops *ops, struct lxc_conf *conf)
{
        __do_close int dfd = -EBADF;
        int ret;
        const char *controllers_use;

        if (ops->dfd_mnt >= 0)
                return ret_errno(EBUSY);

        /*
         * I don't see the need for allowing symlinks here. If users want to
         * have their hierarchy available in different locations I strongly
         * suggest bind-mounts.
         */
        dfd = open_at(-EBADF, DEFAULT_CGROUP_MOUNTPOINT,
                        PROTECT_OPATH_DIRECTORY, PROTECT_LOOKUP_ABSOLUTE_XDEV, 0);
        if (dfd < 0)
                return syserror("Failed to open " DEFAULT_CGROUP_MOUNTPOINT);

        controllers_use = lxc_global_config_value("lxc.cgroup.use");
        if (controllers_use) {
                __do_free char *dup = NULL;
                char *it;

                dup = strdup(controllers_use);
                if (!dup)
                        return -errno;

                lxc_iterate_parts(it, dup, ",") {
                        ret = list_add_string(&ops->cgroup_use, it);
                        if (ret < 0)
                                return ret;
                }
        }

        /*
         * Keep dfd referenced by the cleanup function and actually move the fd
         * once we know the initialization succeeded. So if we fail we clean up
         * the dfd.
         */
        ops->dfd_mnt = dfd;

        ret = __initialize_cgroups(ops, conf->cgroup_meta.relative, !lxc_list_empty(&conf->id_map));
        if (ret < 0)
                return syserror_ret(ret, "Failed to initialize cgroups");

        /* Transfer ownership to cgroup_ops. */
        move_fd(dfd);
        return 0;
}

__cgfsng_ops static int cgfsng_data_init(struct cgroup_ops *ops)
{
        const char *cgroup_pattern;

        if (!ops)
                return ret_set_errno(-1, ENOENT);

        /* copy system-wide cgroup information */
        cgroup_pattern = lxc_global_config_value("lxc.cgroup.pattern");
        if (cgroup_pattern && !strequal(cgroup_pattern, "")) {
                ops->cgroup_pattern = strdup(cgroup_pattern);
                if (!ops->cgroup_pattern)
                        return ret_errno(ENOMEM);
        }

        return 0;
}

struct cgroup_ops *cgroup_ops_init(struct lxc_conf *conf)
{
        __cleanup_cgroup_ops struct cgroup_ops *cgfsng_ops = NULL;

        cgfsng_ops = zalloc(sizeof(struct cgroup_ops));
        if (!cgfsng_ops)
                return ret_set_errno(NULL, ENOMEM);

        cgfsng_ops->cgroup_layout       = CGROUP_LAYOUT_UNKNOWN;
        cgfsng_ops->dfd_mnt             = -EBADF;

        if (initialize_cgroups(cgfsng_ops, conf))
                return NULL;

        cgfsng_ops->data_init                           = cgfsng_data_init;
        cgfsng_ops->payload_destroy                     = cgfsng_payload_destroy;
        cgfsng_ops->monitor_destroy                     = cgfsng_monitor_destroy;
        cgfsng_ops->monitor_create                      = cgfsng_monitor_create;
        cgfsng_ops->monitor_enter                       = cgfsng_monitor_enter;
        cgfsng_ops->monitor_delegate_controllers        = cgfsng_monitor_delegate_controllers;
        cgfsng_ops->payload_delegate_controllers        = cgfsng_payload_delegate_controllers;
        cgfsng_ops->payload_create                      = cgfsng_payload_create;
        cgfsng_ops->payload_enter                       = cgfsng_payload_enter;
        cgfsng_ops->finalize                            = cgfsng_finalize;
        cgfsng_ops->get_cgroup                          = cgfsng_get_cgroup;
        cgfsng_ops->get                                 = cgfsng_get;
        cgfsng_ops->set                                 = cgfsng_set;
        cgfsng_ops->freeze                              = cgfsng_freeze;
        cgfsng_ops->unfreeze                            = cgfsng_unfreeze;
        cgfsng_ops->setup_limits_legacy                 = cgfsng_setup_limits_legacy;
        cgfsng_ops->setup_limits                        = cgfsng_setup_limits;
        cgfsng_ops->driver                              = "cgfsng";
        cgfsng_ops->version                             = "1.0.0";
        cgfsng_ops->attach                              = cgfsng_attach;
        cgfsng_ops->chown                               = cgfsng_chown;
        cgfsng_ops->mount                               = cgfsng_mount;
        cgfsng_ops->devices_activate                    = cgfsng_devices_activate;
        cgfsng_ops->get_limit_cgroup                    = cgfsng_get_limit_cgroup;

        cgfsng_ops->criu_escape                         = cgfsng_criu_escape;
        cgfsng_ops->criu_num_hierarchies                = cgfsng_criu_num_hierarchies;
        cgfsng_ops->criu_get_hierarchies                = cgfsng_criu_get_hierarchies;

        return move_ptr(cgfsng_ops);
}

static int __unified_attach_fd(const struct lxc_conf *conf, int fd_unified, pid_t pid)
{
        int ret;

        if (!lxc_list_empty(&conf->id_map)) {
                struct userns_exec_unified_attach_data args = {
                        .conf           = conf,
                        .unified_fd     = fd_unified,
                        .pid            = pid,
                };

                ret = socketpair(PF_LOCAL, SOCK_STREAM | SOCK_CLOEXEC, 0, args.sk_pair);
                if (ret < 0)
                        return -errno;

                ret = userns_exec_minimal(conf,
                                          cgroup_unified_attach_parent_wrapper,
                                          &args,
                                          cgroup_unified_attach_child_wrapper,
                                          &args);
        } else {
                ret = cgroup_attach_leaf(conf, fd_unified, pid);
        }

        return ret;
}

static int __cgroup_attach_many(const struct lxc_conf *conf, const char *name,
                                const char *lxcpath, pid_t pid)
{
        call_cleaner(put_cgroup_ctx) struct cgroup_ctx *ctx = &(struct cgroup_ctx){};
        int ret;
        size_t idx;
        ssize_t pidstr_len;
        char pidstr[INTTYPE_TO_STRLEN(pid_t)];

        ret = lxc_cmd_get_cgroup_ctx(name, lxcpath, sizeof(struct cgroup_ctx), ctx);
        if (ret < 0)
                return ret_errno(ENOSYS);

        pidstr_len = strnprintf(pidstr, sizeof(pidstr), "%d", pid);
        if (pidstr_len < 0)
                return pidstr_len;

        for (idx = 0; idx < ctx->fd_len; idx++) {
                int dfd_con = ctx->fd[idx];

                if (unified_cgroup_fd(dfd_con))
                        ret = __unified_attach_fd(conf, dfd_con, pid);
                else
                        ret = lxc_writeat(dfd_con, "cgroup.procs", pidstr, pidstr_len);
                if (ret)
                        return syserror_ret(ret, "Failed to attach to cgroup fd %d", dfd_con);
                else
                        TRACE("Attached to cgroup fd %d", dfd_con);
        }

        if (idx == 0)
                return syserror_set(-ENOENT, "Failed to attach to cgroups");

        TRACE("Attached to %s cgroup layout", cgroup_layout_name(ctx->layout));
        return 0;
}

static int __cgroup_attach_unified(const struct lxc_conf *conf, const char *name,
                                   const char *lxcpath, pid_t pid)
{
        __do_close int dfd_unified = -EBADF;

        if (!conf || is_empty_string(name) || is_empty_string(lxcpath) || pid <= 0)
                return ret_errno(EINVAL);

        dfd_unified = lxc_cmd_get_cgroup2_fd(name, lxcpath);
        if (dfd_unified < 0)
                return ret_errno(ENOSYS);

        return __unified_attach_fd(conf, dfd_unified, pid);
}

int cgroup_attach(const struct lxc_conf *conf, const char *name,
                  const char *lxcpath, pid_t pid)
{
        int ret;

        ret = __cgroup_attach_many(conf, name, lxcpath, pid);
        if (ret < 0) {
                if (!ERRNO_IS_NOT_SUPPORTED(ret))
                        return ret;

                ret = __cgroup_attach_unified(conf, name, lxcpath, pid);
                if (ret < 0 && ERRNO_IS_NOT_SUPPORTED(ret))
                        return ret_errno(ENOSYS);
        }

        return ret;
}

/* Connects to command socket therefore isn't callable from command handler. */
int cgroup_get(const char *name, const char *lxcpath, const char *key, char *buf, size_t len)
{
        __do_close int dfd = -EBADF;
        struct cgroup_fd fd = {
                .fd = -EBADF,
        };
        size_t len_controller;
        int ret;

        if (is_empty_string(name) || is_empty_string(lxcpath) ||
            is_empty_string(key))
                return ret_errno(EINVAL);

        if ((buf && !len) || (len && !buf))
                return ret_errno(EINVAL);

        len_controller = strcspn(key, ".");
        len_controller++; /* Don't forget the \0 byte. */
        if (len_controller >= MAX_CGROUP_ROOT_NAMELEN)
                return ret_errno(EINVAL);
        (void)strlcpy(fd.controller, key, len_controller);

        ret = lxc_cmd_get_limit_cgroup_fd(name, lxcpath, sizeof(struct cgroup_fd), &fd);
        if (ret < 0) {
                if (!ERRNO_IS_NOT_SUPPORTED(ret))
                        return ret;

                dfd = lxc_cmd_get_limit_cgroup2_fd(name, lxcpath);
                if (dfd < 0) {
                        if (!ERRNO_IS_NOT_SUPPORTED(ret))
                                return ret;

                        return ret_errno(ENOSYS);
                }
                fd.type = UNIFIED_HIERARCHY;
                fd.fd = move_fd(dfd);
        }
        dfd = move_fd(fd.fd);

        TRACE("Reading %s from %s cgroup hierarchy", key, cgroup_hierarchy_name(fd.type));

        if (fd.type == UNIFIED_HIERARCHY && strequal(fd.controller, "devices"))
                return ret_errno(EOPNOTSUPP);
        else
                ret = lxc_read_try_buf_at(dfd, key, buf, len);

        return ret;
}

/* Connects to command socket therefore isn't callable from command handler. */
int cgroup_set(const char *name, const char *lxcpath, const char *key, const char *value)
{
        __do_close int dfd = -EBADF;
        struct cgroup_fd fd = {
                .fd = -EBADF,
        };
        size_t len_controller;
        int ret;

        if (is_empty_string(name) || is_empty_string(lxcpath) ||
            is_empty_string(key) || is_empty_string(value))
                return ret_errno(EINVAL);

        len_controller = strcspn(key, ".");
        len_controller++; /* Don't forget the \0 byte. */
        if (len_controller >= MAX_CGROUP_ROOT_NAMELEN)
                return ret_errno(EINVAL);
        (void)strlcpy(fd.controller, key, len_controller);

        ret = lxc_cmd_get_limit_cgroup_fd(name, lxcpath, sizeof(struct cgroup_fd), &fd);
        if (ret < 0) {
                if (!ERRNO_IS_NOT_SUPPORTED(ret))
                        return ret;

                dfd = lxc_cmd_get_limit_cgroup2_fd(name, lxcpath);
                if (dfd < 0) {
                        if (!ERRNO_IS_NOT_SUPPORTED(ret))
                                return ret;

                        return ret_errno(ENOSYS);
                }
                fd.type = UNIFIED_HIERARCHY;
                fd.fd = move_fd(dfd);
        }
        dfd = move_fd(fd.fd);

        TRACE("Setting %s to %s in %s cgroup hierarchy", key, value, cgroup_hierarchy_name(fd.type));

        if (fd.type == UNIFIED_HIERARCHY && strequal(fd.controller, "devices")) {
                struct device_item device = {};

                ret = device_cgroup_rule_parse(&device, key, value);
                if (ret < 0)
                        return log_error_errno(-1, EINVAL, "Failed to parse device string %s=%s",
                                               key, value);

                ret = lxc_cmd_add_bpf_device_cgroup(name, lxcpath, &device);
        } else {
                ret = lxc_writeat(dfd, key, value, strlen(value));
        }

        return ret;
}

static int do_cgroup_freeze(int unified_fd,
                            const char *state_string,
                            int state_num,
                            int timeout,
                            const char *epoll_error,
                            const char *wait_error)
{
        __do_close int events_fd = -EBADF;
        call_cleaner(lxc_mainloop_close) struct lxc_epoll_descr *descr_ptr = NULL;
        int ret;
        struct lxc_epoll_descr descr = {};

        if (timeout != 0) {
                ret = lxc_mainloop_open(&descr);
                if (ret)
                        return log_error_errno(-1, errno, "%s", epoll_error);

                /* automatically cleaned up now */
                descr_ptr = &descr;

                events_fd = open_at(unified_fd, "cgroup.events", PROTECT_OPEN, PROTECT_LOOKUP_BENEATH, 0);
                if (events_fd < 0)
                        return log_error_errno(-errno, errno, "Failed to open cgroup.events file");

                ret = lxc_mainloop_add_handler_events(&descr, events_fd, EPOLLPRI, freezer_cgroup_events_cb, INT_TO_PTR(state_num));
                if (ret < 0)
                        return log_error_errno(-1, errno, "Failed to add cgroup.events fd handler to mainloop");
        }

        ret = lxc_writeat(unified_fd, "cgroup.freeze", state_string, 1);
        if (ret < 0)
                return log_error_errno(-1, errno, "Failed to open cgroup.freeze file");

        if (timeout != 0) {
                ret = lxc_mainloop(&descr, timeout);
                if (ret)
                        return log_error_errno(-1, errno, "%s", wait_error);
        }

        return log_trace(0, "Container now %s", (state_num == 1) ? "frozen" : "unfrozen");
}

static inline int __cgroup_freeze(int unified_fd, int timeout)
{
        return do_cgroup_freeze(unified_fd, "1", 1, timeout,
                                "Failed to create epoll instance to wait for container freeze",
                                "Failed to wait for container to be frozen");
}

int cgroup_freeze(const char *name, const char *lxcpath, int timeout)
{
        __do_close int unified_fd = -EBADF;
        int ret;

        if (is_empty_string(name) || is_empty_string(lxcpath))
                return ret_errno(EINVAL);

        unified_fd = lxc_cmd_get_limit_cgroup2_fd(name, lxcpath);
        if (unified_fd < 0)
                return ret_errno(ENOCGROUP2);

        lxc_cmd_notify_state_listeners(name, lxcpath, FREEZING);
        ret = __cgroup_freeze(unified_fd, timeout);
        lxc_cmd_notify_state_listeners(name, lxcpath, !ret ? FROZEN : RUNNING);
        return ret;
}

int __cgroup_unfreeze(int unified_fd, int timeout)
{
        return do_cgroup_freeze(unified_fd, "0", 0, timeout,
                                "Failed to create epoll instance to wait for container freeze",
                                "Failed to wait for container to be frozen");
}

int cgroup_unfreeze(const char *name, const char *lxcpath, int timeout)
{
        __do_close int unified_fd = -EBADF;
        int ret;

        if (is_empty_string(name) || is_empty_string(lxcpath))
                return ret_errno(EINVAL);

        unified_fd = lxc_cmd_get_limit_cgroup2_fd(name, lxcpath);
        if (unified_fd < 0)
                return ret_errno(ENOCGROUP2);

        lxc_cmd_notify_state_listeners(name, lxcpath, THAWED);
        ret = __cgroup_unfreeze(unified_fd, timeout);
        lxc_cmd_notify_state_listeners(name, lxcpath, !ret ? RUNNING : FROZEN);
        return ret;
}