int
ProcFamily::migrate_to_cgroup(pid_t pid)
{
// Attempt to migrate a given process to a cgroup.
// This can be done without regards to whether the
// process is already in the cgroup
if (!m_cgroup.isValid()) {
return 1;
}
// We want to make sure task migration is turned on for the
// associated memory controller. So, we get to look up the original cgroup.
//
// If there is no memory controller present, we skip all this and just attempt a migrate
int err;
u_int64_t orig_migrate;
bool changed_orig = false;
char * orig_cgroup_string = NULL;
struct cgroup * orig_cgroup;
struct cgroup_controller * memory_controller;
if (m_cm.isMounted(CgroupManager::MEMORY_CONTROLLER) && (err = cgroup_get_current_controller_path(pid, MEMORY_CONTROLLER_STR, &orig_cgroup_string))) {
dprintf(D_PROCFAMILY,
"Unable to determine current memory cgroup for PID %u (ProcFamily %u): %u %s\n",
pid, m_root_pid, err, cgroup_strerror(err));
return 1;
}
// We will migrate the PID to the new cgroup even if it is in the proper memory controller cgroup
// It is possible for the task to be in multiple cgroups.
if (m_cm.isMounted(CgroupManager::MEMORY_CONTROLLER) && (orig_cgroup_string != NULL) && (strcmp(m_cgroup_string.c_str(), orig_cgroup_string))) {
// Yes, there are race conditions here - can't really avoid this.
// Throughout this block, we can assume memory controller exists.
// Get original value of migrate.
orig_cgroup = cgroup_new_cgroup(orig_cgroup_string);
ASSERT (orig_cgroup != NULL);
if ((err = cgroup_get_cgroup(orig_cgroup))) {
dprintf(D_PROCFAMILY,
"Unable to read original cgroup %s (ProcFamily %u): %u %s\n",
orig_cgroup_string, m_root_pid, err, cgroup_strerror(err));
goto after_migrate;
}
if ((memory_controller = cgroup_get_controller(orig_cgroup, MEMORY_CONTROLLER_STR)) == NULL) {
cgroup_free(&orig_cgroup);
goto after_migrate;
}
if ((err = cgroup_get_value_uint64(memory_controller, "memory.move_charge_at_immigrate", &orig_migrate))) {
if (err == ECGROUPVALUENOTEXIST) {
// Older kernels don't have the ability to migrate memory accounting to the new cgroup.
dprintf(D_PROCFAMILY,
"This kernel does not support memory usage migration; cgroup %s memory statistics"
" will be slightly incorrect (ProcFamily %u)\n",
m_cgroup_string.c_str(), m_root_pid);
} else {
dprintf(D_PROCFAMILY,
"Unable to read cgroup %s memory controller settings for "
"migration (ProcFamily %u): %u %s\n",
orig_cgroup_string, m_root_pid, err, cgroup_strerror(err));
}
cgroup_free(&orig_cgroup);
goto after_migrate;
}
if (orig_migrate != 3) {
orig_cgroup = cgroup_new_cgroup(orig_cgroup_string);
memory_controller = cgroup_add_controller(orig_cgroup, MEMORY_CONTROLLER_STR);
ASSERT (memory_controller != NULL); // Memory controller must already exist
cgroup_add_value_uint64(memory_controller, "memory.move_charge_at_immigrate", 3);
if ((err = cgroup_modify_cgroup(orig_cgroup))) {
// Not allowed to change settings
dprintf(D_ALWAYS,
"Unable to change cgroup %s memory controller settings for migration. "
"Some memory accounting will be inaccurate (ProcFamily %u): %u %s\n",
orig_cgroup_string, m_root_pid, err, cgroup_strerror(err));
} else {
changed_orig = true;
}
}
cgroup_free(&orig_cgroup);
}
after_migrate:
orig_cgroup = NULL;
err = cgroup_attach_task_pid(& const_cast<struct cgroup &>(m_cgroup.getCgroup()), pid);
if (err) {
dprintf(D_PROCFAMILY,
"Cannot attach pid %u to cgroup %s for ProcFamily %u: %u %s\n",
pid, m_cgroup_string.c_str(), m_root_pid, err, cgroup_strerror(err));
}
if (changed_orig) {
if ((orig_cgroup = cgroup_new_cgroup(orig_cgroup_string))) {
goto after_restore;
}
if (((memory_controller = cgroup_add_controller(orig_cgroup, MEMORY_CONTROLLER_STR)) != NULL) &&
(!cgroup_add_value_uint64(memory_controller, "memory.move_charge_at_immigrate", orig_migrate))) {
cgroup_modify_cgroup(orig_cgroup);
}
cgroup_free(&orig_cgroup);
}
after_restore:
if (orig_cgroup_string != NULL) {
free(orig_cgroup_string);
}
return err;
}
static int corosync_move_to_root_cgroup(void) {
int res = -1;
#ifdef HAVE_LIBCGROUP
int cg_ret;
struct cgroup *root_cgroup = NULL;
struct cgroup_controller *root_cpu_cgroup_controller = NULL;
char *current_cgroup_path = NULL;
cg_ret = cgroup_init();
if (cg_ret) {
log_printf(LOGSYS_LEVEL_WARNING, "Unable to initialize libcgroup: %s ",
cgroup_strerror(cg_ret));
goto exit_res;
}
cg_ret = cgroup_get_current_controller_path(getpid(), "cpu", ¤t_cgroup_path);
if (cg_ret) {
log_printf(LOGSYS_LEVEL_WARNING, "Unable to get current cpu cgroup path: %s ",
cgroup_strerror(cg_ret));
goto exit_res;
}
if (strcmp(current_cgroup_path, "/") == 0) {
log_printf(LOGSYS_LEVEL_DEBUG, "Corosync is already in root cgroup path");
res = 0;
goto exit_res;
}
root_cgroup = cgroup_new_cgroup("/");
if (root_cgroup == NULL) {
log_printf(LOGSYS_LEVEL_WARNING, "Can't create root cgroup");
goto exit_res;
}
root_cpu_cgroup_controller = cgroup_add_controller(root_cgroup, "cpu");
if (root_cpu_cgroup_controller == NULL) {
log_printf(LOGSYS_LEVEL_WARNING, "Can't create root cgroup cpu controller");
goto exit_res;
}
cg_ret = cgroup_attach_task(root_cgroup);
if (cg_ret) {
log_printf(LOGSYS_LEVEL_WARNING, "Can't attach task to root cgroup: %s ",
cgroup_strerror(cg_ret));
goto exit_res;
}
cg_ret = cgroup_get_current_controller_path(getpid(), "cpu", ¤t_cgroup_path);
if (cg_ret) {
log_printf(LOGSYS_LEVEL_WARNING, "Unable to get current cpu cgroup path: %s ",
cgroup_strerror(cg_ret));
goto exit_res;
}
if (strcmp(current_cgroup_path, "/") == 0) {
log_printf(LOGSYS_LEVEL_NOTICE, "Corosync successfully moved to root cgroup");
res = 0;
} else {
log_printf(LOGSYS_LEVEL_WARNING, "Can't move Corosync to root cgroup");
}
exit_res:
if (root_cgroup != NULL) {
cgroup_free(&root_cgroup);
}
/*
* libcgroup doesn't define something like cgroup_fini so there is no way how to clean
* it's cache. It has to be called when libcgroup authors decide to implement it.
*/
#endif
return (res);
}
