chiark / gitweb /
Add mounting of a name=elogind cgroup if no init controller is found.
authorSven Eden <yamakuzure@gmx.net>
Fri, 10 Mar 2017 07:33:22 +0000 (08:33 +0100)
committerSven Eden <yamakuzure@gmx.net>
Tue, 14 Mar 2017 09:23:22 +0000 (10:23 +0100)
This is done for systems, which init systems are no cgroup
controllers. One example is runit on Void Linux.

Makefile.am
configure.ac
src/basic/cgroup-util.c
src/basic/cgroup-util.h
src/core/cgroup.c
src/core/mount-setup.c [new file with mode: 0644]
src/core/mount-setup.h [new file with mode: 0644]
src/login/logind.c

index fbcc25d..c403adc 100644 (file)
@@ -623,7 +623,9 @@ libelogind_core_la_SOURCES = \
        src/login/logind-utmp.c \
        src/login/logind-acl.h \
        src/core/cgroup.h \
-       src/core/cgroup.c
+       src/core/cgroup.c \
+       src/core/mount-setup.h \
+       src/core/mount-setup.c
 
 libelogind_core_la_LIBADD = \
        libshared.la
index d410162..1c5e25f 100644 (file)
@@ -113,10 +113,17 @@ fi
 # ------------------------------------------------------------------------------
 # Find running cgroup controller
 with_cgroupctrl=
-AS_IF(  [test -f /proc/self/cgroup],
-        [with_cgroupctrl=`grep "^1:name=" /proc/self/cgroup | cut -d ':' -f 2`])
+AS_IF(  [test -f /proc/self/cgroup], [
+        # If the init system is a cgroup controler, it will be position 1.
+        # Secondary controllers, like cgmanager, do not work.
+        with_cgroupctrl=`grep "^1:name=" /proc/self/cgroup | cut -d ':' -f 2`
+        AS_IF(  [test -z "$with_cgroupctrl"], [
+                # Try to be our own cgroup controller
+                with_cgroupctrl="name=elogind"
+        ])
+])
 AS_IF(  [test -z "$with_cgroupctrl"],
-        [with_cgroupctrl="name=elogind"])
+        AC_MSG_ERROR([No running cgroup controller found]))
 
 
 # ------------------------------------------------------------------------------
index c83010d..1e54f58 100644 (file)
@@ -520,7 +520,7 @@ int cg_get_path(const char *controller, const char *path, const char *suffix, ch
                  * *below* the controllers, without any prefix. */
 
                 if (!path && !suffix)
-                return -EINVAL;
+                        return -EINVAL;
 
                 if (!suffix)
                         t = strdup(path);
@@ -552,7 +552,7 @@ int cg_get_path(const char *controller, const char *path, const char *suffix, ch
 
         path_kill_slashes(*fs);
         return 0;
-        }
+}
 
 static int controller_is_accessible(const char *controller) {
         int unified;
@@ -2274,6 +2274,10 @@ bool cg_is_unified_wanted(void) {
 bool cg_is_legacy_wanted(void) {
         return !cg_is_unified_wanted();
 }
+#else
+bool cg_is_legacy_wanted(void) {
+        return true;
+}
 #endif // 0
 
 static const char *cgroup_controller_table[_CGROUP_CONTROLLER_MAX] = {
index f14fc37..9743559 100644 (file)
@@ -155,7 +155,7 @@ int cg_unified(void);
 // UNNEEDED void cg_unified_flush(void);
 
 // UNNEEDED bool cg_is_unified_wanted(void);
-// UNNEEDED bool cg_is_legacy_wanted(void);
+bool cg_is_legacy_wanted(void);
 
 const char* cgroup_controller_to_string(CGroupController c) _const_;
 CGroupController cgroup_controller_from_string(const char *s) _pure_;
index c0bfecb..567fe2d 100644 (file)
@@ -1229,6 +1229,7 @@ int manager_setup_cgroup(Manager *m) {
         r = cg_pid_get_path(ELOGIND_CGROUP_CONTROLLER, 0, &m->cgroup_root);
         if (r < 0)
                 return log_error_errno(r, "Cannot determine cgroup we are running in: %m");
+
 /// elogind does not support systemd scopes and slices
 #if 0
         /* Chop off the init scope, if we are already located in it */
@@ -1320,7 +1321,10 @@ int manager_setup_cgroup(Manager *m) {
                 scope_path = strjoina(m->cgroup_root, "/" SPECIAL_INIT_SCOPE);
                 r = cg_create_and_attach(ELOGIND_CGROUP_CONTROLLER, scope_path, 0);
 #else
-                if (streq(m->cgroup_root, "/elogind"))
+                if (streq(ELOGIND_CGROUP_CONTROLLER, "name=elogind"))
+                        // we are our own cgroup controller
+                        scope_path = strjoina("");
+                else if (streq(m->cgroup_root, "/elogind"))
                         // root already is our cgroup
                         scope_path = strjoina(m->cgroup_root);
                 else
diff --git a/src/core/mount-setup.c b/src/core/mount-setup.c
new file mode 100644 (file)
index 0000000..ed95cbb
--- /dev/null
@@ -0,0 +1,425 @@
+/*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
+
+/***
+  This file is part of systemd.
+
+  Copyright 2010 Lennart Poettering
+
+  systemd is free software; you can redistribute it and/or modify it
+  under the terms of the GNU Lesser General Public License as published by
+  the Free Software Foundation; either version 2.1 of the License, or
+  (at your option) any later version.
+
+  systemd is distributed in the hope that it will be useful, but
+  WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+  Lesser General Public License for more details.
+
+  You should have received a copy of the GNU Lesser General Public License
+  along with systemd; If not, see <http://www.gnu.org/licenses/>.
+***/
+
+#include <sys/mount.h>
+#include <errno.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <ftw.h>
+
+#include "mount-setup.h"
+//#include "dev-setup.h"
+//#include "bus-util.h"
+//#include "log.h"
+#include "macro.h"
+//#include "util.h"
+#include "label.h"
+//#include "set.h"
+//#include "strv.h"
+#include "mkdir.h"
+#include "path-util.h"
+//#include "missing.h"
+#include "virt.h"
+//#include "efivars.h"
+//#include "smack-util.h"
+#include "cgroup-util.h"
+
+typedef enum MountMode {
+        MNT_NONE  =        0,
+        MNT_FATAL =        1 <<  0,
+        MNT_IN_CONTAINER = 1 <<  1,
+} MountMode;
+
+typedef struct MountPoint {
+        const char *what;
+        const char *where;
+        const char *type;
+        const char *options;
+        unsigned long flags;
+        bool (*condition_fn)(void);
+        MountMode mode;
+} MountPoint;
+
+/* The first three entries we might need before SELinux is up. The
+ * fourth (securityfs) is needed by IMA to load a custom policy. The
+ * other ones we can delay until SELinux and IMA are loaded. When
+ * SMACK is enabled we need smackfs, too, so it's a fifth one. */
+#ifdef HAVE_SMACK
+#define N_EARLY_MOUNT 5
+#else
+#define N_EARLY_MOUNT 4
+#endif
+
+static const MountPoint mount_table[] = {
+/// UNNEEDED by elogind
+#if 0
+        { "sysfs",       "/sys",                      "sysfs",      NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          NULL,          MNT_FATAL|MNT_IN_CONTAINER },
+        { "proc",        "/proc",                     "proc",       NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          NULL,          MNT_FATAL|MNT_IN_CONTAINER },
+        { "devtmpfs",    "/dev",                      "devtmpfs",   "mode=755",                MS_NOSUID|MS_STRICTATIME,
+          NULL,          MNT_FATAL|MNT_IN_CONTAINER },
+        { "securityfs",  "/sys/kernel/security",      "securityfs", NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          NULL,          MNT_NONE                   },
+#ifdef HAVE_SMACK
+        { "smackfs",     "/sys/fs/smackfs",           "smackfs",    "smackfsdef=*",            MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          mac_smack_use, MNT_FATAL                  },
+        { "tmpfs",       "/dev/shm",                  "tmpfs",      "mode=1777,smackfsroot=*", MS_NOSUID|MS_NODEV|MS_STRICTATIME,
+          mac_smack_use, MNT_FATAL                  },
+#endif
+        { "tmpfs",       "/dev/shm",                  "tmpfs",      "mode=1777",               MS_NOSUID|MS_NODEV|MS_STRICTATIME,
+          NULL,          MNT_FATAL|MNT_IN_CONTAINER },
+        { "devpts",      "/dev/pts",                  "devpts",     "mode=620,gid=" STRINGIFY(TTY_GID), MS_NOSUID|MS_NOEXEC,
+          NULL,          MNT_IN_CONTAINER           },
+#ifdef HAVE_SMACK
+        { "tmpfs",       "/run",                      "tmpfs",      "mode=755,smackfsroot=*",  MS_NOSUID|MS_NODEV|MS_STRICTATIME,
+          mac_smack_use, MNT_FATAL                  },
+#endif
+        { "tmpfs",       "/run",                      "tmpfs",      "mode=755",                MS_NOSUID|MS_NODEV|MS_STRICTATIME,
+          NULL,          MNT_FATAL|MNT_IN_CONTAINER },
+        { "cgroup",      "/sys/fs/cgroup",            "cgroup",     "__DEVEL__sane_behavior",  MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          cg_is_unified_wanted, MNT_FATAL|MNT_IN_CONTAINER },
+#endif // 0
+        { "tmpfs",       "/sys/fs/cgroup",            "tmpfs",      "mode=755",                MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME,
+          cg_is_legacy_wanted, MNT_FATAL|MNT_IN_CONTAINER },
+/// UNNEEDED by elogind
+#if 0
+        { "cgroup",      "/sys/fs/cgroup/systemd",    "cgroup",     "none,name=systemd,xattr", MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          cg_is_legacy_wanted, MNT_IN_CONTAINER           },
+        { "cgroup",      "/sys/fs/cgroup/systemd",    "cgroup",     "none,name=systemd",       MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          cg_is_legacy_wanted, MNT_FATAL|MNT_IN_CONTAINER },
+#else
+        { "cgroup",      "/sys/fs/cgroup/elogind",    "cgroup",     "none,name=elogind,xattr", MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          cg_is_legacy_wanted, MNT_IN_CONTAINER           },
+        { "cgroup",      "/sys/fs/cgroup/elogind",    "cgroup",     "none,name=elogind",       MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          cg_is_legacy_wanted, MNT_FATAL|MNT_IN_CONTAINER },
+#endif // 0
+/// UNNEEDED by elogind
+#if 0
+        { "pstore",      "/sys/fs/pstore",            "pstore",     NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          NULL,          MNT_NONE                   },
+#ifdef ENABLE_EFI
+        { "efivarfs",    "/sys/firmware/efi/efivars", "efivarfs",   NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          is_efi_boot,   MNT_NONE                   },
+#endif
+        { "kdbusfs",    "/sys/fs/kdbus",             "kdbusfs",    NULL, MS_NOSUID|MS_NOEXEC|MS_NODEV,
+          is_kdbus_wanted,       MNT_IN_CONTAINER },
+#endif // 0
+};
+
+/// UNNEEDED by elogind
+#if 0
+/* These are API file systems that might be mounted by other software,
+ * we just list them here so that we know that we should ignore them */
+
+static const char ignore_paths[] =
+        /* SELinux file systems */
+        "/sys/fs/selinux\0"
+        /* Container bind mounts */
+        "/proc/sys\0"
+        "/dev/console\0"
+        "/proc/kmsg\0";
+
+bool mount_point_is_api(const char *path) {
+        unsigned i;
+
+        /* Checks if this mount point is considered "API", and hence
+         * should be ignored */
+
+        for (i = 0; i < ELEMENTSOF(mount_table); i ++)
+                if (path_equal(path, mount_table[i].where))
+                        return true;
+
+        return path_startswith(path, "/sys/fs/cgroup/");
+}
+
+bool mount_point_ignore(const char *path) {
+        const char *i;
+
+        NULSTR_FOREACH(i, ignore_paths)
+                if (path_equal(path, i))
+                        return true;
+
+        return false;
+}
+#endif // 0
+
+static int mount_one(const MountPoint *p, bool relabel) {
+        int r;
+
+        assert(p);
+
+        if (p->condition_fn && !p->condition_fn())
+                return 0;
+
+        /* Relabel first, just in case */
+        if (relabel)
+                label_fix(p->where, true, true);
+
+        r = path_is_mount_point(p->where, AT_SYMLINK_FOLLOW);
+        if (r < 0 && r != -ENOENT)
+                return r;
+        if (r > 0)
+                return 0;
+
+        /* Skip securityfs in a container */
+        if (!(p->mode & MNT_IN_CONTAINER) && detect_container(NULL) > 0)
+                return 0;
+
+        /* The access mode here doesn't really matter too much, since
+         * the mounted file system will take precedence anyway. */
+        if (relabel)
+                mkdir_p_label(p->where, 0755);
+        else
+                mkdir_p(p->where, 0755);
+
+        log_debug("Mounting %s to %s of type %s with options %s.",
+                  p->what,
+                  p->where,
+                  p->type,
+                  strna(p->options));
+
+        if (mount(p->what,
+                  p->where,
+                  p->type,
+                  p->flags,
+                  p->options) < 0) {
+                log_full((p->mode & MNT_FATAL) ? LOG_ERR : LOG_DEBUG, "Failed to mount %s at %s: %m", p->type, p->where);
+                return (p->mode & MNT_FATAL) ? -errno : 0;
+        }
+
+        /* Relabel again, since we now mounted something fresh here */
+        if (relabel)
+                label_fix(p->where, false, false);
+
+        return 1;
+}
+
+/// UNNEEDED by elogind
+#if 0
+int mount_setup_early(void) {
+        unsigned i;
+        int r = 0;
+
+        assert_cc(N_EARLY_MOUNT <= ELEMENTSOF(mount_table));
+
+        /* Do a minimal mount of /proc and friends to enable the most
+         * basic stuff, such as SELinux */
+        for (i = 0; i < N_EARLY_MOUNT; i ++)  {
+                int j;
+
+                j = mount_one(mount_table + i, false);
+                if (r == 0)
+                        r = j;
+        }
+
+        return r;
+}
+
+int mount_cgroup_controllers(char ***join_controllers) {
+        _cleanup_set_free_free_ Set *controllers = NULL;
+        int r;
+
+        if (!cg_is_legacy_wanted())
+                return 0;
+
+        /* Mount all available cgroup controllers that are built into the kernel. */
+
+        controllers = set_new(&string_hash_ops);
+        if (!controllers)
+                return log_oom();
+
+        r = cg_kernel_controllers(controllers);
+        if (r < 0)
+                return log_error_errno(r, "Failed to enumerate cgroup controllers: %m");
+
+        for (;;) {
+                _cleanup_free_ char *options = NULL, *controller = NULL, *where = NULL;
+                MountPoint p = {
+                        .what = "cgroup",
+                        .type = "cgroup",
+                        .flags = MS_NOSUID|MS_NOEXEC|MS_NODEV,
+                        .mode = MNT_IN_CONTAINER,
+                };
+                char ***k = NULL;
+
+                controller = set_steal_first(controllers);
+                if (!controller)
+                        break;
+
+                if (join_controllers)
+                        for (k = join_controllers; *k; k++)
+                                if (strv_find(*k, controller))
+                                        break;
+
+                if (k && *k) {
+                        char **i, **j;
+
+                        for (i = *k, j = *k; *i; i++) {
+
+                                if (!streq(*i, controller)) {
+                                        _cleanup_free_ char *t;
+
+                                        t = set_remove(controllers, *i);
+                                        if (!t) {
+                                                free(*i);
+                                                continue;
+                                        }
+                                }
+
+                                *(j++) = *i;
+                        }
+
+                        *j = NULL;
+
+                        options = strv_join(*k, ",");
+                        if (!options)
+                                return log_oom();
+                } else {
+                        options = controller;
+                        controller = NULL;
+                }
+
+                where = strappend("/sys/fs/cgroup/", options);
+                if (!where)
+                        return log_oom();
+
+                p.where = where;
+                p.options = options;
+
+                r = mount_one(&p, true);
+                if (r < 0)
+                        return r;
+
+                if (r > 0 && k && *k) {
+                        char **i;
+
+                        for (i = *k; *i; i++) {
+                                _cleanup_free_ char *t = NULL;
+
+                                t = strappend("/sys/fs/cgroup/", *i);
+                                if (!t)
+                                        return log_oom();
+
+                                r = symlink(options, t);
+                                if (r < 0 && errno != EEXIST)
+                                        return log_error_errno(errno, "Failed to create symlink %s: %m", t);
+                        }
+                }
+        }
+
+        /* Now that we mounted everything, let's make the tmpfs the
+         * cgroup file systems are mounted into read-only. */
+        (void) mount("tmpfs", "/sys/fs/cgroup", "tmpfs", MS_REMOUNT|MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME|MS_RDONLY, "mode=755");
+
+        return 0;
+}
+
+#if defined(HAVE_SELINUX) || defined(HAVE_SMACK)
+static int nftw_cb(
+                const char *fpath,
+                const struct stat *sb,
+                int tflag,
+                struct FTW *ftwbuf) {
+
+        /* No need to label /dev twice in a row... */
+        if (_unlikely_(ftwbuf->level == 0))
+                return FTW_CONTINUE;
+
+        label_fix(fpath, false, false);
+
+        /* /run/initramfs is static data and big, no need to
+         * dynamically relabel its contents at boot... */
+        if (_unlikely_(ftwbuf->level == 1 &&
+                      tflag == FTW_D &&
+                      streq(fpath, "/run/initramfs")))
+                return FTW_SKIP_SUBTREE;
+
+        return FTW_CONTINUE;
+};
+#endif
+#endif // 0
+
+int mount_setup(bool loaded_policy) {
+        unsigned i;
+        int r = 0;
+
+        for (i = 0; i < ELEMENTSOF(mount_table); i ++) {
+                int j;
+
+                j = mount_one(mount_table + i, loaded_policy);
+                if (r == 0)
+                        r = j;
+        }
+
+        if (r < 0)
+                return r;
+
+/// elogind does not control /, /dev, /run and /run/systemd/* are setup elsewhere.
+#if 0
+#if defined(HAVE_SELINUX) || defined(HAVE_SMACK)
+        /* Nodes in devtmpfs and /run need to be manually updated for
+         * the appropriate labels, after mounting. The other virtual
+         * API file systems like /sys and /proc do not need that, they
+         * use the same label for all their files. */
+        if (loaded_policy) {
+                usec_t before_relabel, after_relabel;
+                char timespan[FORMAT_TIMESPAN_MAX];
+
+                before_relabel = now(CLOCK_MONOTONIC);
+
+                nftw("/dev", nftw_cb, 64, FTW_MOUNT|FTW_PHYS|FTW_ACTIONRETVAL);
+                nftw("/run", nftw_cb, 64, FTW_MOUNT|FTW_PHYS|FTW_ACTIONRETVAL);
+
+                after_relabel = now(CLOCK_MONOTONIC);
+
+                log_info("Relabelled /dev and /run in %s.",
+                         format_timespan(timespan, sizeof(timespan), after_relabel - before_relabel, 0));
+        }
+#endif
+
+        /* Create a few default symlinks, which are normally created
+         * by udevd, but some scripts might need them before we start
+         * udevd. */
+        dev_setup(NULL, UID_INVALID, GID_INVALID);
+
+        /* Mark the root directory as shared in regards to mount
+         * propagation. The kernel defaults to "private", but we think
+         * it makes more sense to have a default of "shared" so that
+         * nspawn and the container tools work out of the box. If
+         * specific setups need other settings they can reset the
+         * propagation mode to private if needed. */
+        if (detect_container(NULL) <= 0)
+                if (mount(NULL, "/", NULL, MS_REC|MS_SHARED, NULL) < 0)
+                        log_warning_errno(errno, "Failed to set up the root directory for shared mount propagation: %m");
+
+        /* Create a few directories we always want around, Note that
+         * sd_booted() checks for /run/systemd/system, so this mkdir
+         * really needs to stay for good, otherwise software that
+         * copied sd-daemon.c into their sources will misdetect
+         * systemd. */
+        mkdir_label("/run/systemd", 0755);
+        mkdir_label("/run/systemd/system", 0755);
+        mkdir_label("/run/systemd/inaccessible", 0000);
+#endif // 0
+
+        return 0;
+}
diff --git a/src/core/mount-setup.h b/src/core/mount-setup.h
new file mode 100644 (file)
index 0000000..395a492
--- /dev/null
@@ -0,0 +1,32 @@
+/*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
+
+#pragma once
+
+/***
+  This file is part of systemd.
+
+  Copyright 2010 Lennart Poettering
+
+  systemd is free software; you can redistribute it and/or modify it
+  under the terms of the GNU Lesser General Public License as published by
+  the Free Software Foundation; either version 2.1 of the License, or
+  (at your option) any later version.
+
+  systemd is distributed in the hope that it will be useful, but
+  WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+  Lesser General Public License for more details.
+
+  You should have received a copy of the GNU Lesser General Public License
+  along with systemd; If not, see <http://www.gnu.org/licenses/>.
+***/
+
+#include <stdbool.h>
+
+// UNNEEDED int mount_setup_early(void);
+int mount_setup(bool loaded_policy);
+
+// UNNEEDED int mount_cgroup_controllers(char ***join_controllers);
+
+// UNNEEDED bool mount_point_is_api(const char *path);
+// UNNEEDED bool mount_point_ignore(const char *path);
index 7288953..23f4059 100644 (file)
@@ -36,6 +36,7 @@
 #include "label.h"
 #include "label.h"
 #include "cgroup.h"
+#include "mount-setup.h"
 #include "virt.h"
 
 static void manager_free(Manager *m);
@@ -95,6 +96,13 @@ static Manager *manager_new(void) {
         if (!m->kill_exclude_users)
                 goto fail;
 
+        /* If elogind should be its own controller, mount its cgroup */
+        if (streq(ELOGIND_CGROUP_CONTROLLER, "name=elogind")) {
+                r = mount_setup(true);
+                if (r < 0)
+                        goto fail;
+        }
+
         /* Make cgroups */
         r = manager_setup_cgroup(m);
         if (r < 0)