/*
* This function assumes that all pids inside a cgroup
* belong to the same namespace, that is the container namespace.
* Therefore, from the host box, any of them will do.
*/
pid_t get_pid_from_container(envid_t veid)
{
char cgrp[CT_MAX_STR_SIZE];
struct cgroup *ct;
void *task_handle;
void *cont_handle;
struct cgroup_mount_point mnt;
pid_t pid = -1;
int ret;
veid_to_name(cgrp, veid);
ct = cgroup_new_cgroup(cgrp);
ret = cgroup_get_cgroup(ct);
if (ret == ECGROUPNOTEXIST)
goto out_free;
ret = cgroup_get_controller_begin(&cont_handle, &mnt);
if (ret != 0) /* no controllers, something is wrong */
goto out_free;
ret = cgroup_get_task_begin(cgrp, mnt.name, &task_handle, &pid);
if (ret != 0) /* no tasks, something is also wrong */
goto out_end_cont;
cgroup_get_task_end(&task_handle);
out_end_cont:
cgroup_get_controller_end(&cont_handle);
out_free:
cgroup_free(&ct);
return pid;
}
static int do_create_container(struct cgroup *ct, struct cgroup *parent)
{
struct cgroup_mount_point mnt;
struct cgroup_controller *controller;
void *handle;
int ret;
ret = cgroup_get_controller_begin(&handle, &mnt);
cgroup_get_cgroup(parent);
do {
controller = cgroup_add_controller(ct, mnt.name);
ret = controller_apply_config(ct, parent, controller, mnt.name);
if (!ret)
ret = cgroup_get_controller_next(&handle, &mnt);
} while (!ret);
cgroup_get_controller_end(&handle);
if (ret == ECGEOF)
ret = cgroup_create_cgroup(ct, 0);
return ret;
}
int container_is_running(envid_t veid)
{
int ret = 0;
void *handle;
struct cgroup_mount_point mnt;
struct cgroup *ct;
char cgrp[CT_MAX_STR_SIZE];
veid_to_name(cgrp, veid);
ct = cgroup_new_cgroup(cgrp);
ret = cgroup_get_cgroup(ct);
if (ret == ECGROUPNOTEXIST) {
ret = 0;
goto out_free;
}
ret = cgroup_get_controller_begin(&handle, &mnt);
do {
struct cgroup_controller *controller;
controller = cgroup_get_controller(ct, mnt.name);
if (!controller) {
logger(0, 0, "Controller %s seems to be missing!", mnt.name);
continue;
}
if ((ret = controller_has_tasks(cgrp, mnt.name)) != 0)
goto out;
} while ((ret = cgroup_get_controller_next(&handle, &mnt)) == 0);
if (ret != ECGEOF)
ret = -ret;
else
ret = 0;
out:
cgroup_get_controller_end(&handle);
out_free:
cgroup_free(&ct);
return ret;
}
int
VanillaProc::setupOOMEvent(const std::string &cgroup_string)
{
#if !(defined(HAVE_EVENTFD) && defined(HAVE_EXT_LIBCGROUP))
// Shut the compiler up.
cgroup_string.size();
return 0;
#else
// Initialize the event descriptor
int tmp_efd = eventfd(0, EFD_CLOEXEC);
if (tmp_efd == -1) {
dprintf(D_ALWAYS,
"Unable to create new event FD for starter: %u %s\n",
errno, strerror(errno));
return 1;
}
// Find the memcg location on disk
void * handle = NULL;
struct cgroup_mount_point mount_info;
int ret = cgroup_get_controller_begin(&handle, &mount_info);
std::stringstream oom_control;
std::stringstream event_control;
bool found_memcg = false;
while (ret == 0) {
if (strcmp(mount_info.name, MEMORY_CONTROLLER_STR) == 0) {
found_memcg = true;
oom_control << mount_info.path << "/";
event_control << mount_info.path << "/";
break;
}
cgroup_get_controller_next(&handle, &mount_info);
}
if (!found_memcg && (ret != ECGEOF)) {
dprintf(D_ALWAYS,
"Error while locating memcg controller for starter: %u %s\n",
ret, cgroup_strerror(ret));
return 1;
}
cgroup_get_controller_end(&handle);
if (found_memcg == false) {
dprintf(D_ALWAYS,
"Memcg is not available; OOM notification disabled for starter.\n");
return 1;
}
// Finish constructing the location of the control files
oom_control << cgroup_string << "/memory.oom_control";
std::string oom_control_str = oom_control.str();
event_control << cgroup_string << "/cgroup.event_control";
std::string event_control_str = event_control.str();
// Open the oom_control and event control files
TemporaryPrivSentry sentry(PRIV_ROOT);
m_oom_fd = open(oom_control_str.c_str(), O_RDONLY | O_CLOEXEC);
if (m_oom_fd == -1) {
dprintf(D_ALWAYS,
"Unable to open the OOM control file for starter: %u %s\n",
errno, strerror(errno));
return 1;
}
int event_ctrl_fd = open(event_control_str.c_str(), O_WRONLY | O_CLOEXEC);
if (event_ctrl_fd == -1) {
dprintf(D_ALWAYS,
"Unable to open event control for starter: %u %s\n",
errno, strerror(errno));
return 1;
}
// Inform Linux we will be handling the OOM events for this container.
int oom_fd2 = open(oom_control_str.c_str(), O_WRONLY | O_CLOEXEC);
if (oom_fd2 == -1) {
dprintf(D_ALWAYS,
"Unable to open the OOM control file for writing for starter: %u %s\n",
errno, strerror(errno));
close(event_ctrl_fd);
return 1;
}
const char limits [] = "1";
ssize_t nwritten = full_write(oom_fd2, &limits, 1);
if (nwritten < 0) {
/* Newer kernels return EINVAL if you attempt to enable OOM management
* on a cgroup where use_hierarchy is set to 1 and it is not the parent
* cgroup.
*
* This is a common setup, so we log and move along.
*
* See also #4435.
*/
if (errno == EINVAL)
{
dprintf(D_FULLDEBUG, "Unable to setup OOM killer management because"
" memory.use_hierarchy is enabled for this cgroup; consider"
" disabling it for this host or set BASE_CGROUP=/. The hold"
" message for an OOM event may not be reliably set.\n");
}
else
{
dprintf(D_ALWAYS, "Failure when attempting to enable OOM killer "
" management for this job (errno=%d, %s).\n", errno, strerror(errno));
close(event_ctrl_fd);
close(oom_fd2);
close(tmp_efd);
return 1;
}
}
close(oom_fd2);
// Create the subscription string:
std::stringstream sub_ss;
sub_ss << tmp_efd << " " << m_oom_fd;
std::string sub_str = sub_ss.str();
if ((nwritten = full_write(event_ctrl_fd, sub_str.c_str(), sub_str.size())) < 0) {
dprintf(D_ALWAYS,
"Unable to write into event control file for starter: %u %s\n",
errno, strerror(errno));
close(event_ctrl_fd);
close(tmp_efd);
return 1;
}
close(event_ctrl_fd);
// Fool DC into talking to the eventfd
int pipes[2]; pipes[0] = -1; pipes[1] = -1;
int fd_to_replace = -1;
if (!daemonCore->Create_Pipe(pipes, true) || pipes[0] == -1) {
dprintf(D_ALWAYS, "Unable to create a DC pipe\n");
close(tmp_efd);
close(m_oom_fd);
m_oom_fd = -1;
return 1;
}
if (!daemonCore->Get_Pipe_FD(pipes[0], &fd_to_replace) || fd_to_replace == -1) {
dprintf(D_ALWAYS, "Unable to lookup pipe's FD\n");
close(tmp_efd);
close(m_oom_fd); m_oom_fd = -1;
daemonCore->Close_Pipe(pipes[0]);
daemonCore->Close_Pipe(pipes[1]);
return 1;
}
dup3(tmp_efd, fd_to_replace, O_CLOEXEC);
close(tmp_efd);
m_oom_efd = pipes[0];
m_oom_efd2 = pipes[1];
// Inform DC we want to recieve notifications from this FD.
if (-1 == daemonCore->Register_Pipe(pipes[0],"OOM event fd", static_cast<PipeHandlercpp>(&VanillaProc::outOfMemoryEvent),"OOM Event Handler",this,HANDLE_READ))
{
dprintf(D_ALWAYS, "Failed to register OOM event FD pipe.\n");
daemonCore->Close_Pipe(pipes[0]);
daemonCore->Close_Pipe(pipes[1]);
m_oom_fd = -1;
m_oom_efd = -1;
m_oom_efd2 = -1;
}
dprintf(D_FULLDEBUG, "Subscribed the starter to OOM notification for this cgroup; jobs triggering an OOM will be put on hold.\n");
return 0;
#endif
}
int
VanillaProc::setupOOMEvent(const std::string &cgroup_string)
{
#if !(defined(HAVE_EVENTFD) && defined(HAVE_EXT_LIBCGROUP))
// Shut the compiler up.
cgroup_string.size();
return 0;
#else
// Initialize the event descriptor
m_oom_efd = eventfd(0, EFD_CLOEXEC);
if (m_oom_efd == -1) {
dprintf(D_ALWAYS,
"Unable to create new event FD for starter: %u %s\n",
errno, strerror(errno));
return 1;
}
// Find the memcg location on disk
void * handle = NULL;
struct cgroup_mount_point mount_info;
int ret = cgroup_get_controller_begin(&handle, &mount_info);
std::stringstream oom_control;
std::stringstream event_control;
bool found_memcg = false;
while (ret == 0) {
if (strcmp(mount_info.name, MEMORY_CONTROLLER_STR) == 0) {
found_memcg = true;
oom_control << mount_info.path << "/";
event_control << mount_info.path << "/";
break;
}
cgroup_get_controller_next(&handle, &mount_info);
}
if (!found_memcg && (ret != ECGEOF)) {
dprintf(D_ALWAYS,
"Error while locating memcg controller for starter: %u %s\n",
ret, cgroup_strerror(ret));
return 1;
}
cgroup_get_controller_end(&handle);
if (found_memcg == false) {
dprintf(D_ALWAYS,
"Memcg is not available; OOM notification disabled for starter.\n");
return 1;
}
// Finish constructing the location of the control files
oom_control << cgroup_string << "/memory.oom_control";
std::string oom_control_str = oom_control.str();
event_control << cgroup_string << "/cgroup.event_control";
std::string event_control_str = event_control.str();
// Open the oom_control and event control files
TemporaryPrivSentry sentry(PRIV_ROOT);
m_oom_fd = open(oom_control_str.c_str(), O_RDONLY | O_CLOEXEC);
if (m_oom_fd == -1) {
dprintf(D_ALWAYS,
"Unable to open the OOM control file for starter: %u %s\n",
errno, strerror(errno));
return 1;
}
int event_ctrl_fd = open(event_control_str.c_str(), O_WRONLY | O_CLOEXEC);
if (event_ctrl_fd == -1) {
dprintf(D_ALWAYS,
"Unable to open event control for starter: %u %s\n",
errno, strerror(errno));
return 1;
}
// Inform Linux we will be handling the OOM events for this container.
int oom_fd2 = open(oom_control_str.c_str(), O_WRONLY | O_CLOEXEC);
if (oom_fd2 == -1) {
dprintf(D_ALWAYS,
"Unable to open the OOM control file for writing for starter: %u %s\n",
errno, strerror(errno));
close(event_ctrl_fd);
return 1;
}
const char limits [] = "1";
ssize_t nwritten = full_write(oom_fd2, &limits, 1);
if (nwritten < 0) {
dprintf(D_ALWAYS,
"Unable to set OOM control to %s for starter: %u %s\n",
limits, errno, strerror(errno));
close(event_ctrl_fd);
close(oom_fd2);
return 1;
}
close(oom_fd2);
// Create the subscription string:
std::stringstream sub_ss;
sub_ss << m_oom_efd << " " << m_oom_fd;
std::string sub_str = sub_ss.str();
if ((nwritten = full_write(event_ctrl_fd, sub_str.c_str(), sub_str.size())) < 0) {
dprintf(D_ALWAYS,
"Unable to write into event control file for starter: %u %s\n",
errno, strerror(errno));
close(event_ctrl_fd);
return 1;
}
close(event_ctrl_fd);
// Fool DC into talking to the eventfd
int pipes[2]; pipes[0] = -1; pipes[1] = -1;
int fd_to_replace = -1;
if (daemonCore->Create_Pipe(pipes, true) == -1 || pipes[0] == -1) {
dprintf(D_ALWAYS, "Unable to create a DC pipe\n");
close(m_oom_efd);
m_oom_efd = -1;
close(m_oom_fd);
m_oom_fd = -1;
return 1;
}
if ( daemonCore->Get_Pipe_FD(pipes[0], &fd_to_replace) == -1 || fd_to_replace == -1) {
dprintf(D_ALWAYS, "Unable to lookup pipe's FD\n");
close(m_oom_efd); m_oom_efd = -1;
close(m_oom_fd); m_oom_fd = -1;
daemonCore->Close_Pipe(pipes[0]);
daemonCore->Close_Pipe(pipes[1]);
return 1;
}
dup3(m_oom_efd, fd_to_replace, O_CLOEXEC);
close(m_oom_efd);
m_oom_efd = pipes[0];
// Inform DC we want to recieve notifications from this FD.
daemonCore->Register_Pipe(pipes[0],"OOM event fd", static_cast<PipeHandlercpp>(&VanillaProc::outOfMemoryEvent),"OOM Event Handler",this,HANDLE_READ);
return 0;
#endif
}