bool
make_sub_cgroup(cgroup_t group, char *parent, char *child)
{
char child_dir[SGE_PATH_MAX];
DENTER(TOP_LAYER, "make_sub_cgroup");
build_path(child_dir, parent, child);
if (!sge_is_directory(child_dir)) {
errno = 0;
if (mkdir(child_dir, 0755) != 0) {
WARNING((SGE_EVENT, MSG_FILE_CREATEDIRFAILED_SS, child_dir,
strerror(errno)));
DRETURN(false);
}
}
/* You need to populate the mems and cpus before you can use the
cpuset -- they're not inherited.
Checkme: it looks as if clone_children can deal with this
(present in Ubuntu's Linux 3.2, at least). */
if (cg_cpuset == group) {
DRETURN(copy_from_parent(child_dir, "mems") &&
copy_from_parent(child_dir, "cpus"));
}
DRETURN(true);
}
/* Remove the job task cpuset directory after first removing shepherd
sub-directories. */
bool
remove_job_cpuset(u_long32 job, u_long32 task)
{
#if ! __linux__
return true;
#else /* using non-portable dirent-isms */
char dirpath[SGE_PATH_MAX];
DIR *dir;
struct dirent *dent;
DENTER(TOP_LAYER, "remove_job_cpuset");
snprintf(dirpath, sizeof dirpath, "%s/"sge_u32"."sge_u32,
cgroup_dir(cg_cpuset), job, task);
INFO((SGE_EVENT, "removing task cpuset "SFN, dirpath));
if (!sge_is_directory(dirpath)) return true;
errno = 0;
/* Maybe this should be made reentrant, though there's existing code
which does the same sort of thing in the same context. */
if ((dir = opendir(dirpath)) == NULL) {
ERROR((SGE_EVENT, MSG_FILE_CANTOPENDIRECTORYX_SS, dirpath,
strerror(errno)));
DRETURN(false);
}
while ((dent = readdir(dir)))
if ((DT_DIR == dent->d_type) && sge_strisint(dent->d_name))
remove_shepherd_cpuset(job, task, atoi(dent->d_name));
closedir(dir);
if (rmdir(dirpath) != 0) {
ERROR((SGE_EVENT, MSG_FILE_RMDIRFAILED_SS, dirpath, strerror(errno)));
DRETURN(false);
}
DRETURN(true);
#endif /* !__linux__ */
}
/* look for an "sge" directory under a mount point of the cgroup */
static bool
find_cgroup_dir(cgroup_t group, char *dir, size_t ldir)
{
FILE *fp;
char path[2048], fstype[64], options[256];
bool ret = false, is_cpuset = (cg_cpuset == group);
*dir = '\0';
if ((fp = fopen("/proc/self/mounts", "r")) == NULL)
return ret;
while (fscanf(fp, "%*s %2047s %63s %255s %*d %*d\n", path, fstype, options)
== 3) {
if (is_cpuset && strcmp(fstype, "cpuset") == 0) {
ret = true;
break;
} else if ((strcmp(fstype, "cgroup") == 0)) {
char *s = options, *tok, *save = NULL;
bool rw = false, got_type = is_cpuset;
while ((tok = strtok_r(s, ",", &save))) {
if (strcmp(tok, "rw") == 0) rw = true;
if (strcmp(tok, group_name[group]) == 0) got_type = true;
if (rw && got_type) break;
s = NULL;
}
sge_strlcat(path, "/sge", sizeof path);
ret = rw && got_type && sge_is_directory(path);
if (ret) break;
}
}
fclose(fp);
if (ret) sge_strlcpy(dir, path, ldir);
return ret;
}
static int elect_path(dstring *aBuffer)
{
const char *d;
d = getenv("TMPDIR");
if ((d != NULL) && sge_is_directory(d)) {
sge_dstring_append(aBuffer, d);
return 0;
} else if (sge_is_directory(P_tmpdir)) {
sge_dstring_append(aBuffer, P_tmpdir);
return 0;
} else if (sge_is_directory("/tmp")) {
sge_dstring_append(aBuffer, "/tmp/");
return 0;
}
return -1;
}
/* Does the controller directory for the job exist? */
bool
have_cgroup_job_dir(cgroup_t group, u_long32 job, u_long32 task)
{
char dir[SGE_PATH_MAX];
if (!get_cgroup_job_dir(group, dir, sizeof dir, job, task))
return false;
return *dir && sge_is_directory(dir);
}
/* Put PROC (string version of pid) into the controller DIR */
static bool
reparent_proc(const char *proc, char *dir)
{
char path[SGE_PATH_MAX];
pid_t pid = atoi(proc);
if (!pid) return false;
/* Don't fail if the process no longer exists. */
build_path(path, "/proc", proc);
if (!sge_is_directory(path)) return true;
return set_pid_cgroup(pid, dir);
}
/* Return the directory for the given controller GROUP of JOB and TASK
in buffer DIR of length LDIR (e.g. /dev/cpuset/sge/123.1). DIR is
zero-length on failure. */
bool
get_cgroup_job_dir(cgroup_t group, char *dir, size_t ldir, u_long32 job, u_long32 task)
{
const char *cdir;
DENTER(TOP_LAYER, "get_cgroup_job_dir");
dir[0] = '\0';
cdir = cgroup_dir(group);
if (*cdir == '\0') DRETURN(false);
if (snprintf(dir, ldir, "%s/"sge_u32"."sge_u32, cdir, job, task) >= ldir) {
WARNING((SGE_EVENT, "Can't build cgroup_job_dir value"));
DRETURN(false);
}
if (!sge_is_directory(dir)) {
dir[0] = '\0';
DRETURN(false);
}
DRETURN(true);
}
/* Move the shepherd tasks to the job cpuset 0 and remove the shepherd
cpuset. */
bool
empty_shepherd_cpuset(u_long32 job, u_long32 task, pid_t pid)
{
char dir[SGE_PATH_MAX], taskfile[SGE_PATH_MAX], taskfile0[SGE_PATH_MAX];
bool ret;
snprintf(dir, sizeof dir, "%s/"sge_u32"."sge_u32"/"pid_t_fmt,
cgroup_dir(cg_cpuset), job, task, pid);
if (!sge_is_directory(dir)) return true;
build_path(taskfile, dir, "tasks");
snprintf(taskfile0, sizeof taskfile0, "%s/"sge_u32"."sge_u32"/0/tasks",
cgroup_dir(cg_cpuset), job, task);
errno = 0;
sge_seteuid(SGE_SUPERUSER_UID);
/* This could fail at least if a task is respawning.
It needs to be done task-wise (line-wise) according to cpuset(7). */
ret = copy_linewise(taskfile, taskfile0);
if (sge_switch2admin_user() != 0)
abort();
if (!ret) return false;
/* If rmdir fails, execd will try to kill it. */
return rmdir(dir) ? false : true;
}