diff options
author | Lennart Poettering <lennart@poettering.net> | 2017-02-23 16:23:02 +0100 |
---|---|---|
committer | GitHub <noreply@github.com> | 2017-02-23 16:23:02 +0100 |
commit | ecc0eab247da25a6767ccabd2162a4d03de6ee8c (patch) | |
tree | 4041d203a3611df9705b23a76611f55de019fea3 /src/basic | |
parent | c07afaa5e51d079691c74230377ce6fcc97228e5 (diff) | |
parent | 0a05dcc09dd723dcf017bf42432d33e58ec56a1d (diff) |
Merge pull request #4670 from htejun/systemd-controller-on-unified-v2
Systemd controller on unified v2
Diffstat (limited to 'src/basic')
-rw-r--r-- | src/basic/build.h | 5 | ||||
-rw-r--r-- | src/basic/cgroup-util.c | 305 | ||||
-rw-r--r-- | src/basic/cgroup-util.h | 10 | ||||
-rw-r--r-- | src/basic/def.h | 4 |
4 files changed, 191 insertions, 133 deletions
diff --git a/src/basic/build.h b/src/basic/build.h index 633c2aaccb..91312bd2a3 100644 --- a/src/basic/build.h +++ b/src/basic/build.h @@ -133,6 +133,8 @@ #define _IDN_FEATURE_ "-IDN" #endif +#define _CGROUP_HIEARCHY_ "default-hierarchy=" DEFAULT_HIERARCHY_NAME + #define SYSTEMD_FEATURES \ _PAM_FEATURE_ " " \ _AUDIT_FEATURE_ " " \ @@ -152,4 +154,5 @@ _BLKID_FEATURE_ " " \ _ELFUTILS_FEATURE_ " " \ _KMOD_FEATURE_ " " \ - _IDN_FEATURE_ + _IDN_FEATURE_ " " \ + _CGROUP_HIEARCHY_ diff --git a/src/basic/cgroup-util.c b/src/basic/cgroup-util.c index 6948ed3931..f76b7f47e5 100644 --- a/src/basic/cgroup-util.c +++ b/src/basic/cgroup-util.c @@ -208,6 +208,12 @@ int cg_rmdir(const char *controller, const char *path) { if (r < 0 && errno != ENOENT) return -errno; + if (streq(controller, SYSTEMD_CGROUP_CONTROLLER) && cg_hybrid_unified()) { + r = cg_rmdir(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path); + if (r < 0) + log_warning_errno(r, "Failed to remove compat systemd cgroup %s: %m", path); + } + return 0; } @@ -542,6 +548,13 @@ static const char *controller_to_dirname(const char *controller) { * just cuts off the name= prefixed used for named * hierarchies, if it is specified. */ + if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) { + if (cg_hybrid_unified()) + controller = SYSTEMD_CGROUP_CONTROLLER_HYBRID; + else + controller = SYSTEMD_CGROUP_CONTROLLER_LEGACY; + } + e = startswith(controller, "name="); if (e) return e; @@ -594,7 +607,7 @@ static int join_path_unified(const char *path, const char *suffix, char **fs) { } int cg_get_path(const char *controller, const char *path, const char *suffix, char **fs) { - int unified, r; + int r; assert(fs); @@ -623,11 +636,7 @@ int cg_get_path(const char *controller, const char *path, const char *suffix, ch if (!cg_controller_is_valid(controller)) return -EINVAL; - unified = cg_all_unified(); - if (unified < 0) - return unified; - - if (unified > 0) + if (cg_all_unified()) r = join_path_unified(path, suffix, fs); else r = join_path_legacy(controller, path, suffix, fs); @@ -639,7 +648,6 @@ int cg_get_path(const char *controller, const char *path, const char *suffix, ch } static int controller_is_accessible(const char *controller) { - int unified; assert(controller); @@ -651,10 +659,7 @@ static int controller_is_accessible(const char *controller) { if (!cg_controller_is_valid(controller)) return -EINVAL; - unified = cg_all_unified(); - if (unified < 0) - return unified; - if (unified > 0) { + if (cg_all_unified()) { /* We don't support named hierarchies if we are using * the unified hierarchy. */ @@ -708,7 +713,7 @@ static int trim_cb(const char *path, const struct stat *sb, int typeflag, struct int cg_trim(const char *controller, const char *path, bool delete_root) { _cleanup_free_ char *fs = NULL; - int r = 0; + int r = 0, q; assert(path); @@ -731,6 +736,12 @@ int cg_trim(const char *controller, const char *path, bool delete_root) { return -errno; } + if (streq(controller, SYSTEMD_CGROUP_CONTROLLER) && cg_hybrid_unified()) { + q = cg_trim(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, delete_root); + if (q < 0) + log_warning_errno(q, "Failed to trim compat systemd cgroup %s: %m", path); + } + return r; } @@ -754,6 +765,12 @@ int cg_create(const char *controller, const char *path) { return -errno; } + if (streq(controller, SYSTEMD_CGROUP_CONTROLLER) && cg_hybrid_unified()) { + r = cg_create(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path); + if (r < 0) + log_warning_errno(r, "Failed to create compat systemd cgroup %s: %m", path); + } + return 1; } @@ -791,7 +808,17 @@ int cg_attach(const char *controller, const char *path, pid_t pid) { xsprintf(c, PID_FMT "\n", pid); - return write_string_file(fs, c, 0); + r = write_string_file(fs, c, 0); + if (r < 0) + return r; + + if (streq(controller, SYSTEMD_CGROUP_CONTROLLER) && cg_hybrid_unified()) { + r = cg_attach(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, pid); + if (r < 0) + log_warning_errno(r, "Failed to attach %d to compat systemd cgroup %s: %m", pid, path); + } + + return 0; } int cg_attach_fallback(const char *controller, const char *path, pid_t pid) { @@ -840,7 +867,17 @@ int cg_set_group_access( if (r < 0) return r; - return chmod_and_chown(fs, mode, uid, gid); + r = chmod_and_chown(fs, mode, uid, gid); + if (r < 0) + return r; + + if (streq(controller, SYSTEMD_CGROUP_CONTROLLER) && cg_hybrid_unified()) { + r = cg_set_group_access(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, mode, uid, gid); + if (r < 0) + log_warning_errno(r, "Failed to set group access on compat systemd cgroup %s: %m", path); + } + + return 0; } int cg_set_task_access( @@ -851,7 +888,7 @@ int cg_set_task_access( gid_t gid) { _cleanup_free_ char *fs = NULL, *procs = NULL; - int r, unified; + int r; assert(path); @@ -869,16 +906,18 @@ int cg_set_task_access( if (r < 0) return r; - unified = cg_unified(controller); - if (unified < 0) - return unified; - if (unified) - return 0; + if (!cg_unified(controller)) { + /* Compatibility, Always keep values for "tasks" in sync with + * "cgroup.procs" */ + if (cg_get_path(controller, path, "tasks", &procs) >= 0) + (void) chmod_and_chown(procs, mode, uid, gid); + } - /* Compatibility, Always keep values for "tasks" in sync with - * "cgroup.procs" */ - if (cg_get_path(controller, path, "tasks", &procs) >= 0) - (void) chmod_and_chown(procs, mode, uid, gid); + if (streq(controller, SYSTEMD_CGROUP_CONTROLLER) && cg_hybrid_unified()) { + r = cg_set_task_access(SYSTEMD_CGROUP_CONTROLLER_LEGACY, path, mode, uid, gid); + if (r < 0) + log_warning_errno(r, "Failed to set task access on compat systemd cgroup %s: %m", path); + } return 0; } @@ -923,9 +962,9 @@ int cg_get_xattr(const char *controller, const char *path, const char *name, voi int cg_pid_get_path(const char *controller, pid_t pid, char **path) { _cleanup_fclose_ FILE *f = NULL; char line[LINE_MAX]; - const char *fs; + const char *fs, *controller_str; size_t cs = 0; - int unified; + bool unified; assert(path); assert(pid >= 0); @@ -937,10 +976,14 @@ int cg_pid_get_path(const char *controller, pid_t pid, char **path) { controller = SYSTEMD_CGROUP_CONTROLLER; unified = cg_unified(controller); - if (unified < 0) - return unified; - if (unified == 0) - cs = strlen(controller); + if (!unified) { + if (streq(controller, SYSTEMD_CGROUP_CONTROLLER)) + controller_str = SYSTEMD_CGROUP_CONTROLLER_LEGACY; + else + controller_str = controller; + + cs = strlen(controller_str); + } fs = procfs_file_alloca(pid, "cgroup"); f = fopen(fs, "re"); @@ -977,7 +1020,7 @@ int cg_pid_get_path(const char *controller, pid_t pid, char **path) { *e = 0; FOREACH_WORD_SEPARATOR(word, k, l, ",", state) { - if (k == cs && memcmp(word, controller, cs) == 0) { + if (k == cs && memcmp(word, controller_str, cs) == 0) { found = true; break; } @@ -1001,14 +1044,11 @@ int cg_pid_get_path(const char *controller, pid_t pid, char **path) { int cg_install_release_agent(const char *controller, const char *agent) { _cleanup_free_ char *fs = NULL, *contents = NULL; const char *sc; - int r, unified; + int r; assert(agent); - unified = cg_unified(controller); - if (unified < 0) - return unified; - if (unified) /* doesn't apply to unified hierarchy */ + if (cg_unified(controller)) /* doesn't apply to unified hierarchy */ return -EOPNOTSUPP; r = cg_get_path(controller, NULL, "release_agent", &fs); @@ -1054,12 +1094,9 @@ int cg_install_release_agent(const char *controller, const char *agent) { int cg_uninstall_release_agent(const char *controller) { _cleanup_free_ char *fs = NULL; - int r, unified; + int r; - unified = cg_unified(controller); - if (unified < 0) - return unified; - if (unified) /* Doesn't apply to unified hierarchy */ + if (cg_unified(controller)) /* Doesn't apply to unified hierarchy */ return -EOPNOTSUPP; r = cg_get_path(controller, NULL, "notify_on_release", &fs); @@ -1104,7 +1141,7 @@ int cg_is_empty(const char *controller, const char *path) { } int cg_is_empty_recursive(const char *controller, const char *path) { - int unified, r; + int r; assert(path); @@ -1112,11 +1149,7 @@ int cg_is_empty_recursive(const char *controller, const char *path) { if (controller && (isempty(path) || path_equal(path, "/"))) return false; - unified = cg_unified(controller); - if (unified < 0) - return unified; - - if (unified > 0) { + if (cg_unified(controller)) { _cleanup_free_ char *t = NULL; /* On the unified hierarchy we can check empty state @@ -1833,6 +1866,9 @@ bool cg_controller_is_valid(const char *p) { if (!p) return false; + if (streq(p, SYSTEMD_CGROUP_CONTROLLER)) + return true; + s = startswith(p, "name="); if (s) p = s; @@ -1986,7 +2022,7 @@ int cg_get_keyed_attribute(const char *controller, const char *path, const char int cg_create_everywhere(CGroupMask supported, CGroupMask mask, const char *path) { CGroupController c; - int r, unified; + int r; /* This one will create a cgroup in our private tree, but also * duplicate it in the trees specified in mask, and remove it @@ -1998,10 +2034,7 @@ int cg_create_everywhere(CGroupMask supported, CGroupMask mask, const char *path return r; /* If we are in the unified hierarchy, we are done now */ - unified = cg_all_unified(); - if (unified < 0) - return unified; - if (unified > 0) + if (cg_all_unified()) return 0; /* Otherwise, do the same in the other hierarchies */ @@ -2022,16 +2055,13 @@ int cg_create_everywhere(CGroupMask supported, CGroupMask mask, const char *path int cg_attach_everywhere(CGroupMask supported, const char *path, pid_t pid, cg_migrate_callback_t path_callback, void *userdata) { CGroupController c; - int r, unified; + int r; r = cg_attach(SYSTEMD_CGROUP_CONTROLLER, path, pid); if (r < 0) return r; - unified = cg_all_unified(); - if (unified < 0) - return unified; - if (unified > 0) + if (cg_all_unified()) return 0; for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) { @@ -2072,7 +2102,7 @@ int cg_attach_many_everywhere(CGroupMask supported, const char *path, Set* pids, int cg_migrate_everywhere(CGroupMask supported, const char *from, const char *to, cg_migrate_callback_t to_callback, void *userdata) { CGroupController c; - int r = 0, unified; + int r = 0; if (!path_equal(from, to)) { r = cg_migrate_recursive(SYSTEMD_CGROUP_CONTROLLER, from, SYSTEMD_CGROUP_CONTROLLER, to, CGROUP_REMOVE); @@ -2080,10 +2110,7 @@ int cg_migrate_everywhere(CGroupMask supported, const char *from, const char *to return r; } - unified = cg_all_unified(); - if (unified < 0) - return unified; - if (unified > 0) + if (cg_all_unified()) return r; for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) { @@ -2107,16 +2134,13 @@ int cg_migrate_everywhere(CGroupMask supported, const char *from, const char *to int cg_trim_everywhere(CGroupMask supported, const char *path, bool delete_root) { CGroupController c; - int r, unified; + int r; r = cg_trim(SYSTEMD_CGROUP_CONTROLLER, path, delete_root); if (r < 0) return r; - unified = cg_all_unified(); - if (unified < 0) - return unified; - if (unified > 0) + if (cg_all_unified()) return r; for (c = 0; c < _CGROUP_CONTROLLER_MAX; c++) { @@ -2133,16 +2157,13 @@ int cg_trim_everywhere(CGroupMask supported, const char *path, bool delete_root) int cg_mask_supported(CGroupMask *ret) { CGroupMask mask = 0; - int r, unified; + int r; /* Determines the mask of supported cgroup controllers. Only * includes controllers we can make sense of and that are * actually accessible. */ - unified = cg_all_unified(); - if (unified < 0) - return unified; - if (unified > 0) { + if (cg_all_unified()) { _cleanup_free_ char *root = NULL, *controllers = NULL, *path = NULL; const char *c; @@ -2262,6 +2283,20 @@ int cg_kernel_controllers(Set *controllers) { static thread_local CGroupUnified unified_cache = CGROUP_UNIFIED_UNKNOWN; +/* The hybrid mode was initially implemented in v232 and simply mounted + * cgroup v2 on /sys/fs/cgroup/systemd. This unfortunately broke other + * tools (such as docker) which expected the v1 "name=systemd" hierarchy + * on /sys/fs/cgroup/systemd. From v233 and on, the hybrid mode mountnbs + * v2 on /sys/fs/cgroup/unified and maintains "name=systemd" hierarchy + * on /sys/fs/cgroup/systemd for compatibility with other tools. + * + * To keep live upgrade working, we detect and support v232 layout. When + * v232 layout is detected, to keep cgroup v2 process management but + * disable the compat dual layout, we return %true on + * cg_unified(SYSTEMD_CGROUP_CONTROLLER) and %false on cg_hybrid_unified(). + */ +static thread_local bool unified_systemd_v232; + static int cg_update_unified(void) { struct statfs fs; @@ -2280,24 +2315,30 @@ static int cg_update_unified(void) { if (F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) unified_cache = CGROUP_UNIFIED_ALL; else if (F_TYPE_EQUAL(fs.f_type, TMPFS_MAGIC)) { - if (statfs("/sys/fs/cgroup/systemd/", &fs) < 0) - return -errno; - - unified_cache = F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC) ? - CGROUP_UNIFIED_SYSTEMD : CGROUP_UNIFIED_NONE; + if (statfs("/sys/fs/cgroup/unified/", &fs) == 0 && + F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) { + unified_cache = CGROUP_UNIFIED_SYSTEMD; + unified_systemd_v232 = false; + } else if (statfs("/sys/fs/cgroup/systemd/", &fs) == 0 && + F_TYPE_EQUAL(fs.f_type, CGROUP2_SUPER_MAGIC)) { + unified_cache = CGROUP_UNIFIED_SYSTEMD; + unified_systemd_v232 = true; + } else { + if (statfs("/sys/fs/cgroup/systemd/", &fs) < 0) + return -errno; + if (!F_TYPE_EQUAL(fs.f_type, CGROUP_SUPER_MAGIC)) + return -ENOMEDIUM; + unified_cache = CGROUP_UNIFIED_NONE; + } } else return -ENOMEDIUM; return 0; } -int cg_unified(const char *controller) { +bool cg_unified(const char *controller) { - int r; - - r = cg_update_unified(); - if (r < 0) - return r; + assert(cg_update_unified() >= 0); if (streq_ptr(controller, SYSTEMD_CGROUP_CONTROLLER)) return unified_cache >= CGROUP_UNIFIED_SYSTEMD; @@ -2305,29 +2346,35 @@ int cg_unified(const char *controller) { return unified_cache >= CGROUP_UNIFIED_ALL; } -int cg_all_unified(void) { +bool cg_all_unified(void) { return cg_unified(NULL); } -void cg_unified_flush(void) { +bool cg_hybrid_unified(void) { + + assert(cg_update_unified() >= 0); + + return unified_cache == CGROUP_UNIFIED_SYSTEMD && !unified_systemd_v232; +} + +int cg_unified_flush(void) { unified_cache = CGROUP_UNIFIED_UNKNOWN; + + return cg_update_unified(); } int cg_enable_everywhere(CGroupMask supported, CGroupMask mask, const char *p) { _cleanup_free_ char *fs = NULL; CGroupController c; - int r, unified; + int r; assert(p); if (supported == 0) return 0; - unified = cg_all_unified(); - if (unified < 0) - return unified; - if (!unified) /* on the legacy hiearchy there's no joining of controllers defined */ + if (!cg_all_unified()) /* on the legacy hiearchy there's no joining of controllers defined */ return 0; r = cg_get_path(SYSTEMD_CGROUP_CONTROLLER, p, "cgroup.subtree_control", &fs); @@ -2359,63 +2406,69 @@ int cg_enable_everywhere(CGroupMask supported, CGroupMask mask, const char *p) { bool cg_is_unified_wanted(void) { static thread_local int wanted = -1; - int r, unified; + int r; bool b; + const bool is_default = DEFAULT_HIERARCHY == CGROUP_UNIFIED_ALL; - /* If the hierarchy is already mounted, then follow whatever - * was chosen for it. */ - unified = cg_all_unified(); - if (unified >= 0) - return unified; - - /* Otherwise, let's see what the kernel command line has to - * say. Since checking that is expensive, let's cache the - * result. */ + /* If we have a cached value, return that. */ if (wanted >= 0) return wanted; + /* If the hierarchy is already mounted, then follow whatever + * was chosen for it. */ + if (cg_unified_flush() >= 0) + return (wanted = cg_all_unified()); + + /* Otherwise, let's see what the kernel command line has to say. + * Since checking is expensive, cache a non-error result. */ r = proc_cmdline_get_bool("systemd.unified_cgroup_hierarchy", &b); - if (r < 0) - return false; - return (wanted = r > 0 ? b : false); + return (wanted = r > 0 ? b : is_default); } bool cg_is_legacy_wanted(void) { - return !cg_is_unified_wanted(); + static thread_local int wanted = -1; + + /* If we have a cached value, return that. */ + if (wanted >= 0) + return wanted; + + /* Check if we have cgroups2 already mounted. */ + if (cg_unified_flush() >= 0 && + unified_cache == CGROUP_UNIFIED_ALL) + return (wanted = false); + + /* Otherwise, assume that at least partial legacy is wanted, + * since cgroups2 should already be mounted at this point. */ + return (wanted = true); } -bool cg_is_unified_systemd_controller_wanted(void) { +bool cg_is_hybrid_wanted(void) { static thread_local int wanted = -1; - int r, unified; + int r; bool b; + const bool is_default = DEFAULT_HIERARCHY >= CGROUP_UNIFIED_SYSTEMD; + /* We default to true if the default is "hybrid", obviously, + * but also when the default is "unified", because if we get + * called, it means that unified hierarchy was not mounted. */ - /* If the unified hierarchy is requested in full, no need to - * bother with this. */ - if (cg_is_unified_wanted()) - return 0; + /* If we have a cached value, return that. */ + if (wanted >= 0) + return wanted; /* If the hierarchy is already mounted, then follow whatever * was chosen for it. */ - unified = cg_unified(SYSTEMD_CGROUP_CONTROLLER); - if (unified >= 0) - return unified; - - /* Otherwise, let's see what the kernel command line has to - * say. Since checking that is expensive, let's cache the - * result. */ - if (wanted >= 0) - return wanted; + if (cg_unified_flush() >= 0 && + unified_cache == CGROUP_UNIFIED_ALL) + return (wanted = false); + /* Otherwise, let's see what the kernel command line has to say. + * Since checking is expensive, cache a non-error result. */ r = proc_cmdline_get_bool("systemd.legacy_systemd_cgroup_controller", &b); - if (r < 0) - return false; - - return (wanted = r > 0 ? b : false); -} -bool cg_is_legacy_systemd_controller_wanted(void) { - return cg_is_legacy_wanted() && !cg_is_unified_systemd_controller_wanted(); + /* The meaning of the kernel option is reversed wrt. to the return value + * of this function, hence the negation. */ + return (wanted = r > 0 ? !b : is_default); } int cg_weight_parse(const char *s, uint64_t *ret) { diff --git a/src/basic/cgroup-util.h b/src/basic/cgroup-util.h index 0aa27c4cd7..3d14ec3d9d 100644 --- a/src/basic/cgroup-util.h +++ b/src/basic/cgroup-util.h @@ -240,14 +240,14 @@ int cg_kernel_controllers(Set *controllers); bool cg_ns_supported(void); -int cg_all_unified(void); -int cg_unified(const char *controller); -void cg_unified_flush(void); +bool cg_all_unified(void); +bool cg_hybrid_unified(void); +bool cg_unified(const char *controller); +int cg_unified_flush(void); bool cg_is_unified_wanted(void); bool cg_is_legacy_wanted(void); -bool cg_is_unified_systemd_controller_wanted(void); -bool cg_is_legacy_systemd_controller_wanted(void); +bool cg_is_hybrid_wanted(void); const char* cgroup_controller_to_string(CGroupController c) _const_; CGroupController cgroup_controller_from_string(const char *s) _pure_; diff --git a/src/basic/def.h b/src/basic/def.h index 10d776ec8e..200ea973c1 100644 --- a/src/basic/def.h +++ b/src/basic/def.h @@ -36,7 +36,9 @@ /* The default value for the net.unix.max_dgram_qlen sysctl */ #define DEFAULT_UNIX_MAX_DGRAM_QLEN 512UL -#define SYSTEMD_CGROUP_CONTROLLER "name=systemd" +#define SYSTEMD_CGROUP_CONTROLLER_LEGACY "name=systemd" +#define SYSTEMD_CGROUP_CONTROLLER_HYBRID "name=unified" +#define SYSTEMD_CGROUP_CONTROLLER "_systemd" #define SIGNALS_CRASH_HANDLER SIGSEGV,SIGILL,SIGFPE,SIGBUS,SIGQUIT,SIGABRT #define SIGNALS_IGNORE SIGPIPE |