int CgroupLimits::set_cpu_shares(uint64_t shares)
{
if (!m_cgroup.isValid() || !CgroupManager::getInstance().isMounted(CgroupManager::CPU_CONTROLLER)) {
dprintf(D_ALWAYS, "Unable to set CPU shares because cgroup is invalid.\n");
return 1;
}
int err;
struct cgroup *cpucg = &m_cgroup.getCgroup();
struct cgroup_controller *cpu_controller;
if ((cpu_controller = cgroup_get_controller(cpucg, CPU_CONTROLLER_STR)) == NULL) {
dprintf(D_ALWAYS,
"Unable to add cgroup CPU controller for %s.\n",
m_cgroup_string.c_str());
return 1;
} else if ((err = cgroup_set_value_uint64(cpu_controller, "cpu.shares", shares))) {
dprintf(D_ALWAYS,
"Unable to set CPU shares for %s: %u %s\n",
m_cgroup_string.c_str(), err, cgroup_strerror(err));
return 1;
} else {
TemporaryPrivSentry sentry(PRIV_ROOT);
if ((err = cgroup_modify_cgroup(cpucg))) {
dprintf(D_ALWAYS,
"Unable to commit CPU shares for %s"
": %u %s\n",
m_cgroup_string.c_str(), err, cgroup_strerror(err));
return 1;
}
}
return 0;
}
int CgroupLimits::set_blockio_weight(uint64_t weight)
{
if (!m_cgroup.isValid() || !CgroupManager::getInstance().isMounted(CgroupManager::BLOCK_CONTROLLER)) {
dprintf(D_ALWAYS, "Unable to set blockio weight because cgroup is invalid.\n");
return 1;
}
int err;
struct cgroup *blkiocg = &m_cgroup.getCgroup();
struct cgroup_controller *blkio_controller;
if ((blkio_controller = cgroup_get_controller(blkiocg, BLOCK_CONTROLLER_STR)) == NULL) {
dprintf(D_ALWAYS,
"Unable to get cgroup block IO controller for %s.\n",
m_cgroup_string.c_str());
return 1;
} else if ((err = cgroup_set_value_uint64(blkio_controller, "blkio.weight", weight))) {
dprintf(D_ALWAYS,
"Unable to set block IO weight for %s: %u %s\n",
m_cgroup_string.c_str(), err, cgroup_strerror(err));
return 1;
} else {
TemporaryPrivSentry sentry(PRIV_ROOT);
if ((err = cgroup_modify_cgroup(blkiocg))) {
dprintf(D_ALWAYS,
"Unable to commit block IO weight for %s"
": %u %s\n",
m_cgroup_string.c_str(), err, cgroup_strerror(err));
return 1;
}
}
return 0;
}
int ProcFamily::get_cpu_usage_cgroup(long &user_time, long &sys_time) {
if (!m_cm.isMounted(CgroupManager::CPUACCT_CONTROLLER)) {
return 1;
}
void * handle = NULL;
u_int64_t tmp = 0;
struct cgroup_stat stats;
int err = cgroup_read_stats_begin(CPUACCT_CONTROLLER_STR, m_cgroup_string.c_str(), &handle, &stats);
while (err != ECGEOF) {
if (err > 0) {
dprintf(D_PROCFAMILY,
"Unable to read cgroup %s cpuacct stats (ProcFamily %u): %s.\n",
m_cgroup_string.c_str(), m_root_pid, cgroup_strerror(err));
break;
}
if (_check_stat_uint64(stats, "user", &tmp)) {
user_time = tmp/clock_tick-m_initial_user_cpu;
} else if (_check_stat_uint64(stats, "system", &tmp)) {
sys_time = tmp/clock_tick-m_initial_sys_cpu;
}
err = cgroup_read_stats_next(&handle, &stats);
}
if (handle != NULL) {
cgroup_read_stats_end(&handle);
}
if (err != ECGEOF) {
dprintf(D_ALWAYS, "Internal cgroup error when retrieving CPU statistics: %s\n", cgroup_strerror(err));
return 1;
}
return 0;
}
int
ProcFamily::freezer_cgroup(const char * state)
{
// According to kernel docs, freezer will either succeed
// or return EBUSY in the errno.
//
// This function either returns 0 (success), a positive value (fatal error)
// or -EBUSY.
int err = 0;
struct cgroup_controller* freezer;
struct cgroup *cgroup = cgroup_new_cgroup(m_cgroup_string.c_str());
ASSERT (cgroup != NULL);
if (!m_cm.isMounted(CgroupManager::FREEZE_CONTROLLER)) {
err = 1;
goto ret;
}
freezer = cgroup_add_controller(cgroup, FREEZE_CONTROLLER_STR);
if (NULL == freezer) {
dprintf(D_ALWAYS,
"Unable to access the freezer subsystem for ProcFamily %u "
"for cgroup %s\n",
m_root_pid, m_cgroup_string.c_str());
err = 2;
goto ret;
}
if ((err = cgroup_add_value_string(freezer, "freezer.state", state))) {
dprintf(D_ALWAYS,
"Unable to write %s to freezer for cgroup %s (ProcFamily %u). %u %s\n",
state, m_cgroup_string.c_str(), m_root_pid, err, cgroup_strerror(err));
err = 3;
goto ret;
}
if ((err = cgroup_modify_cgroup(cgroup))) {
if (ECGROUPVALUENOTEXIST == err) {
dprintf(D_ALWAYS,
"Does not appear condor_procd is allowed to freeze"
" cgroup %s (ProcFamily %u).\n",
m_cgroup_string.c_str(), m_root_pid);
} else if ((ECGOTHER == err) && (EBUSY == cgroup_get_last_errno())) {
dprintf(D_ALWAYS, "Kernel was unable to freeze cgroup %s "
"(ProcFamily %u) due to process state; signal delivery "
"won't be atomic\n", m_cgroup_string.c_str(), m_root_pid);
err = -EBUSY;
} else {
dprintf(D_ALWAYS,
"Unable to commit freezer change %s for cgroup %s (ProcFamily %u). %u %s\n",
state, m_cgroup_string.c_str(), m_root_pid, err, cgroup_strerror(err));
}
err = 4;
goto ret;
}
ret:
cgroup_free(&cgroup);
return err;
}
int
ProcFamily::set_cgroup(const std::string &cgroup_string)
{
if (cgroup_string == "/") {
dprintf(D_ALWAYS,
"Cowardly refusing to monitor the root cgroup out "
"of security concerns.\n");
return 1;
}
// Ignore this command if we've done this before.
if (m_cgroup.isValid()) {
if (cgroup_string == m_cgroup.getCgroupString()) {
return 0;
} else {
m_cgroup.destroy();
}
}
dprintf(D_PROCFAMILY, "Setting cgroup to %s for ProcFamily %u.\n",
cgroup_string.c_str(), m_root_pid);
m_cm.create(cgroup_string, m_cgroup, CgroupManager::ALL_CONTROLLERS, CgroupManager::NO_CONTROLLERS);
m_cgroup_string = m_cgroup.getCgroupString();
if (!m_cgroup.isValid()) {
return 1;
}
// Now that we have a cgroup, let's move all the existing processes to it
ProcFamilyMember* member = m_member_list;
while (member != NULL) {
migrate_to_cgroup(member->get_proc_info()->pid);
member = member->m_next;
}
// Record the amount of pre-existing CPU usage here.
m_initial_user_cpu = 0;
m_initial_sys_cpu = 0;
get_cpu_usage_cgroup(m_initial_user_cpu, m_initial_sys_cpu);
// Reset block IO controller
if (m_cm.isMounted(CgroupManager::BLOCK_CONTROLLER)) {
struct cgroup *tmp_cgroup = cgroup_new_cgroup(m_cgroup_string.c_str());
struct cgroup_controller *blkio_controller = cgroup_add_controller(tmp_cgroup, BLOCK_CONTROLLER_STR);
ASSERT (blkio_controller != NULL); // Block IO controller should already exist.
cgroup_add_value_uint64(blkio_controller, "blkio.reset_stats", 0);
int err;
if ((err = cgroup_modify_cgroup(tmp_cgroup))) {
// Not allowed to reset stats?
dprintf(D_ALWAYS,
"Unable to reset cgroup %s block IO statistics. "
"Some block IO accounting will be inaccurate (ProcFamily %u): %u %s\n",
m_cgroup_string.c_str(), m_root_pid, err, cgroup_strerror(err));
}
cgroup_free(&tmp_cgroup);
}
return 0;
}
/* display all controllers attached to the given hierarchy */
static int print_all_controllers_in_hierarchy(const char *tname,
int hierarchy, int flags)
{
int ret = 0;
void *handle;
struct controller_data info;
int first = 1;
cont_name_t cont_names;
cont_name_t cont_name;
/*
* Initialize libcgroup and intentionally ignore its result,
* no mounted controller is valid use case.
*/
(void) cgroup_init();
ret = cgroup_get_all_controller_begin(&handle, &info);
if ((ret != 0) && (ret != ECGEOF)) {
fprintf(stderr, "cannot read controller data: %s\n",
cgroup_strerror(ret));
return ret;
}
while (ret != ECGEOF) {
/* controller is in the hierrachy */
if (info.hierarchy != hierarchy)
goto next;
if (first) {
/* the first controller in the hierarchy */
memset(cont_name, 0, FILENAME_MAX);
strncpy(cont_name, info.name, FILENAME_MAX-1);
memset(cont_names, 0, FILENAME_MAX);
strncpy(cont_names, info.name, FILENAME_MAX-1);
first = 0;
} else {
/* the next controller in the hierarchy */
strncat(cont_names, ",", FILENAME_MAX-1);
strncat(cont_names, info.name, FILENAME_MAX-1);
}
next:
ret = cgroup_get_all_controller_next(&handle, &info);
if (ret && ret != ECGEOF)
goto end;
}
ret = print_controller_mount(cont_name,
flags, cont_names, hierarchy);
end:
cgroup_get_all_controller_end(&handle);
if (ret == ECGEOF)
ret = 0;
return ret;
}
/*
* Skip adding controller which points to the same cgroup when delete
* cgroup with specifying multi controllers. Just skip controller which
* cgroup and hierarchy number is same
*/
static int skip_add_controller(int counter, int *skip,
struct ext_cgroup_record *ecg_list)
{
int k;
struct controller_data info;
void *handle;
int ret = 0;
/* find out hierarchy number of added cgroup */
ecg_list[counter].h_number = 0;
ret = cgroup_get_all_controller_begin(&handle, &info);
while (ret == 0) {
if (!strcmp(info.name, ecg_list[counter].name)) {
/* hierarchy number found out, set it */
ecg_list[counter].h_number = info.hierarchy;
break;
}
ret = cgroup_get_all_controller_next(&handle, &info);
}
cgroup_get_all_controller_end(&handle);
/* deal with cgroup_get_controller_begin/next ret values */
if (ret == ECGEOF)
ret = 0;
if (ret) {
fprintf(stderr, "cgroup_get_controller_begin/next failed(%s)\n",
cgroup_strerror(ret));
return ret;
}
/* found out whether the hierarchy should be skipped */
*skip = 0;
for (k = 0; k < counter; k++) {
if ((!strcmp(ecg_list[k].name, ecg_list[counter].name)) &&
(ecg_list[k].h_number == ecg_list[counter].h_number)) {
/* we found a control group in the same hierarchy */
if (strcmp(ecg_list[k].controller,
ecg_list[counter].controller)) {
/*
* it is a different controller ->
* if there is not one cgroup for the same
* controller, skip it
*/
*skip = 1;
} else {
/*
* there is the identical group,controller pair
* don't skip it
*/
*skip = 0;
return ret;
}
}
}
return ret;
}
int
ProcFamily::spree_cgroup(int sig)
{
// The general idea here is we freeze the cgroup, give the signal,
// then thaw everything out. This way, signals are given in an atomic manner.
//
// Note that if the FREEZE call could be attempted, but not 100% completed, we
// proceed anyway.
bool use_freezer = !m_last_signal_was_sigstop;
m_last_signal_was_sigstop = sig == SIGSTOP ? true : false;
if (!use_freezer) {
dprintf(D_ALWAYS,
"Not using freezer controller to send signal; last "
"signal was SIGSTOP.\n");
} else {
dprintf(D_FULLDEBUG,
"Using freezer controller to send signal to process family.\n");
}
int err = use_freezer ? freezer_cgroup(FROZEN) : 0;
if ((err != 0) && (err != -EBUSY)) {
return err;
}
ASSERT (m_cgroup.isValid());
cgroup_get_cgroup(&const_cast<struct cgroup&>(m_cgroup.getCgroup()));
void **handle = (void **)malloc(sizeof(void*));
ASSERT (handle != NULL);
pid_t pid;
err = cgroup_get_task_begin(m_cgroup_string.c_str(), FREEZE_CONTROLLER_STR, handle, &pid);
if ((err > 0) && (err != ECGEOF))
handle = NULL;
while (err != ECGEOF) {
if (err > 0) {
dprintf(D_ALWAYS,
"Unable to iterate through cgroup %s (ProcFamily %u): %u %s\n",
m_cgroup_string.c_str(), m_root_pid, err, cgroup_strerror(err));
goto release;
}
send_signal(pid, sig);
err = cgroup_get_task_next(handle, &pid);
}
err = 0;
release:
if (handle != NULL) {
cgroup_get_task_end(handle);
free(handle);
}
if (use_freezer) freezer_cgroup(THAWED);
return err;
}
int
ProcFamily::aggregate_usage_cgroup_blockio_io_serviced(ProcFamilyUsage* usage)
{
if (!m_cm.isMounted(CgroupManager::BLOCK_CONTROLLER) || !m_cgroup.isValid())
return 1;
int ret;
void *handle;
char line_contents[BLOCK_STATS_LINE_MAX], sep[]=" ", *tok_handle, *word, *info[3];
char blkio_stats_name[] = "blkio.io_serviced";
short ctr;
int64_t reads=0, writes=0;
ret = cgroup_read_value_begin(BLOCK_CONTROLLER_STR, m_cgroup_string.c_str(),
blkio_stats_name, &handle, line_contents, BLOCK_STATS_LINE_MAX);
while (ret == 0) {
ctr = 0;
word = strtok_r(line_contents, sep, &tok_handle);
while (word && ctr < 3) {
info[ctr++] = word;
word = strtok_r(NULL, sep, &tok_handle);
}
if (ctr == 3) {
errno = 0;
int64_t ctrval = strtoll(info[2], NULL, 10);
if (errno) {
dprintf(D_FULLDEBUG, "Error parsing kernel value to a long: %s; %s\n",
info[2], strerror(errno));
break;
}
if (strcmp(info[1], "Read") == 0) {
reads += ctrval;
} else if (strcmp(info[1], "Write") == 0) {
writes += ctrval;
}
}
ret = cgroup_read_value_next(&handle, line_contents, BLOCK_STATS_LINE_MAX);
}
if (handle != NULL) {
cgroup_read_value_end(&handle);
}
if (ret != ECGEOF) {
dprintf(D_ALWAYS, "Internal cgroup error when retrieving block statistics: %s\n", cgroup_strerror(ret));
return 1;
}
usage->block_reads = reads;
usage->block_writes = writes;
return 0;
}
static mrb_value mrb_cgroup_create(mrb_state *mrb, mrb_value self)
{
int code;
mrb_cgroup_context *mrb_cg_cxt = mrb_cgroup_get_context(mrb, self, "mrb_cgroup_context");
// BUG1 : cgroup_create_cgroup returns an error(Invalid argument:50016:ECGOTHER), despite actually succeeding
// BUG2 : cgroup_delete_cgroup returns an error(This kernel does not support this feature:50029:ECGCANTSETVALUE), despite actually succeeding
// REFS : libcgroup/src/api.c 1620 - 1630 comments
//
// error = cg_set_control_value(path,
// cgroup->controller[k]->values[j]->value);
// /*
// * Should we undo, what we've done in the loops above?
// * An error should not be treated as fatal, since we
// * have several read-only files and several files that
// * are only conditionally created in the child.
// *
// * A middle ground would be to track that there
// * was an error and return a diagnostic value--
// * callers don't get context for the error, but can
// * ignore it specifically if they wish.
// */
// if (error) {
// cgroup_dbg("failed to set %s: %s (%d)\n",
// path,
// cgroup_strerror(error), error);
// retval = ECGCANTSETVALUE;
// continue;
// }
//
if ((code = cgroup_create_cgroup(mrb_cg_cxt->cg, 1)) && code != ECGOTHER && code != ECGCANTSETVALUE) {
mrb_raisef(mrb
, E_RUNTIME_ERROR
, "cgroup_create failed: %S(%S)"
, mrb_str_new_cstr(mrb, cgroup_strerror(code))
, mrb_fixnum_value(code)
);
}
mrb_cg_cxt->already_exist = 1;
mrb_iv_set(mrb
, self
, mrb_intern_cstr(mrb, "mrb_cgroup_context")
, mrb_obj_value(Data_Wrap_Struct(mrb
, mrb->object_class
, &mrb_cgroup_context_type
, (void *)mrb_cg_cxt)
)
);
return self;
}
// error handling
static int l_cgroup_strerror (lua_State *L) {
int eint = 0;
const char *out = NULL;
if(lua_isnumber(L, 1))
eint = lua_tointeger(L, 1);
else
eint = cgroup_get_last_errno();
out = cgroup_strerror(eint);
if(out) {
lua_pushstring(L, out);
return 1;
}
return 0;
}
int CgroupLimits::set_memory_limit_bytes(uint64_t mem_bytes, bool soft)
{
if (!m_cgroup.isValid() || !CgroupManager::getInstance().isMounted(CgroupManager::MEMORY_CONTROLLER)) {
dprintf(D_ALWAYS, "Unable to set memory limit because cgroup is invalid.\n");
return 1;
}
int err;
struct cgroup_controller * mem_controller;
const char * limit = soft ? mem_soft_limit : mem_hard_limit;
dprintf(D_ALWAYS, "Limitting memory usage to %ld bytes\n", mem_bytes);
struct cgroup *memcg = &m_cgroup.getCgroup();
if ((mem_controller = cgroup_get_controller(memcg, MEMORY_CONTROLLER_STR)) == NULL) {
dprintf(D_ALWAYS,
"Unable to get cgroup memory controller for %s.\n",
m_cgroup_string.c_str());
return 1;
} else if ((err = cgroup_set_value_uint64(mem_controller, limit, mem_bytes))) {
dprintf(D_ALWAYS,
"Unable to set memory soft limit for %s: %u %s\n",
m_cgroup_string.c_str(), err, cgroup_strerror(err));
return 1;
} else {
TemporaryPrivSentry sentry(PRIV_ROOT);
if ((err = cgroup_modify_cgroup(memcg))) {
dprintf(D_ALWAYS,
"Unable to commit memory soft limit for %s "
": %u %s\n",
m_cgroup_string.c_str(), err, cgroup_strerror(err));
return 1;
}
}
return 0;
}
/*
* Assumes the cgroup is already mounted at /cgroup/memory/a
*
* Assumes some processes are already in the cgroup
*
* Assumes it is the memory controller is mounted in at that
* point
*/
int main()
{
int size;
pid_t *pids;
int ret;
int i;
ret = cgroup_init();
if (ret) {
printf("FAIL: cgroup_init failed with %s\n", cgroup_strerror(ret));
exit(3);
}
ret = cgroup_get_procs("a", "memory", &pids, &size);
if (ret) {
printf("FAIL: cgroup_get_procs failed with %s\n", cgroup_strerror(ret));
exit(3);
}
for (i = 0; i < size; i++)
printf("%u\n", pids[i]);
return 0;
}
static mrb_value mrb_cgroup_delete(mrb_state *mrb, mrb_value self)
{
int code;
mrb_cgroup_context *mrb_cg_cxt = mrb_cgroup_get_context(mrb, self, "mrb_cgroup_context");
// BUG1 : cgroup_delete_cgroup returns an error(No such file or directory:50016:ECGOTHER), despite actually succeeding
if ((code = cgroup_delete_cgroup(mrb_cg_cxt->cg, 1)) && code != ECGOTHER) {
mrb_raisef(mrb
, E_RUNTIME_ERROR
, "cgroup_delete faild: %S(%S)"
, mrb_str_new_cstr(mrb, cgroup_strerror(code))
, mrb_fixnum_value(code)
);
}
return self;
}
/*
* We send a kill signal to all processes. This is racy in theory, since they
* could spawn new processes faster than we kill. But since one of them is the
* init process, (we don't really know which), then eventually the init process
* will die taking away all the others, so this is fine.
*
* This is a big hack, and only exists because we have no way to enter a PID
* namespace from the outside (yet). From there, we could just issue a normal
* reboot.
*/
int hackish_empty_container(envid_t veid)
{
char cgrp[CT_MAX_STR_SIZE];
struct cgroup *ct;
int ret = 0;
void *task_handle;
pid_t pid;
int i;
veid_to_name(cgrp, veid);
ct = cgroup_new_cgroup(cgrp);
ret = cgroup_get_cgroup(ct);
if (ret == ECGROUPNOTEXIST) {
ret = 0;
goto out;
}
/* Any controller will do */
ret = cgroup_get_task_begin(cgrp, "cpu", &task_handle, &pid);
while (!ret) {
kill(pid, SIGKILL);
ret = cgroup_get_task_next(&task_handle, &pid);
}
cgroup_get_task_end(&task_handle);
if (ret != ECGEOF) {
logger(-1, 0, "Could not finish all tasks: %s",
cgroup_strerror(ret));
goto out;
}
ret = 0;
for (i = 0; i < DEF_STOP_TIMEOUT; i++) {
if (!container_is_running(veid))
goto out;
usleep(500000);
}
logger(-1, 0, "Failed to wait for CT tasks to die");
ret = VZ_STOP_ERROR;
out:
cgroup_free(&ct);
return ret;
}
void
ProcFamily::update_max_image_size_cgroup()
{
if (!m_cm.isMounted(CgroupManager::MEMORY_CONTROLLER) || !m_cgroup.isValid()) {
return;
}
int err;
u_int64_t max_image;
struct cgroup_controller *memct;
Cgroup memcg;
if (m_cm.create(m_cgroup_string, memcg, CgroupManager::MEMORY_CONTROLLER, CgroupManager::MEMORY_CONTROLLER) ||
!memcg.isValid()) {
dprintf(D_PROCFAMILY,
"Unable to create cgroup %s (ProcFamily %u).\n",
m_cgroup_string.c_str(), m_root_pid);
return;
}
if ((memct = cgroup_get_controller(&const_cast<struct cgroup &>(memcg.getCgroup()), MEMORY_CONTROLLER_STR)) == NULL) {
dprintf(D_PROCFAMILY,
"Unable to load memory controller for cgroup %s (ProcFamily %u).\n",
m_cgroup_string.c_str(), m_root_pid);
return;
}
if ((err = cgroup_get_value_uint64(memct, "memory.memsw.max_usage_in_bytes", &max_image))) {
// On newer nodes, swap accounting is disabled by default.
// In some cases, swap accounting causes a kernel oops at the time of writing.
// So, we check memory.max_usage_in_bytes instead.
int err2 = cgroup_get_value_uint64(memct, "memory.max_usage_in_bytes", &max_image);
if (err2) {
dprintf(D_PROCFAMILY,
"Unable to load max memory usage for cgroup %s (ProcFamily %u): %u %s\n",
m_cgroup_string.c_str(), m_root_pid, err, cgroup_strerror(err));
return;
} else if (!have_warned_about_memsw) {
have_warned_about_memsw = true;
dprintf(D_ALWAYS, "Swap acounting is not available; only doing RAM accounting.\n");
}
}
m_max_image_size = max_image/1024;
}
int
ProcFamily::count_tasks_cgroup()
{
if (!m_cm.isMounted(CgroupManager::CPUACCT_CONTROLLER) || !m_cgroup.isValid()) {
return -1;
}
int tasks = 0, err = 0;
pid_t pid;
void **handle = (void **)malloc(sizeof(void*));
ASSERT (handle != NULL)
*handle = NULL;
err = cgroup_get_task_begin(m_cgroup_string.c_str(), CPUACCT_CONTROLLER_STR, handle, &pid);
while (err != ECGEOF) {
if (err > 0) {
dprintf(D_PROCFAMILY,
"Unable to read cgroup %s memory stats (ProcFamily %u): %u %s.\n",
m_cgroup_string.c_str(), m_root_pid, err, cgroup_strerror(err));
break;
}
tasks ++;
err = cgroup_get_task_next(handle, &pid);
}
// Reset err to 0
if (err == ECGEOF) {
err = 0;
}
if (*handle) {
cgroup_get_task_end(handle);
}
if (handle) {
free(handle);
}
if (err) {
return -err;
}
return tasks;
}
int main(int argc, char *argv[])
{
int ret = 0;
int i, j;
int c;
static struct option long_opts[] = {
{"help", no_argument, NULL, 'h'},
{"task", required_argument, NULL, 't'},
{"admin", required_argument, NULL, 'a'},
{"", required_argument, NULL, 'g'},
{"dperm", required_argument, NULL, 'd'},
{"fperm", required_argument, NULL, 'f' },
{"tperm", required_argument, NULL, 's' },
{0, 0, 0, 0},
};
uid_t tuid = CGRULE_INVALID, auid = CGRULE_INVALID;
gid_t tgid = CGRULE_INVALID, agid = CGRULE_INVALID;
struct cgroup_group_spec **cgroup_list;
struct cgroup *cgroup;
struct cgroup_controller *cgc;
/* approximation of max. numbers of groups that will be created */
int capacity = argc;
/* permission variables */
mode_t dir_mode = NO_PERMS;
mode_t file_mode = NO_PERMS;
mode_t tasks_mode = NO_PERMS;
int dirm_change = 0;
int filem_change = 0;
/* no parametr on input */
if (argc < 2) {
usage(1, argv[0]);
return -1;
}
cgroup_list = calloc(capacity, sizeof(struct cgroup_group_spec *));
if (cgroup_list == NULL) {
fprintf(stderr, "%s: out of memory\n", argv[0]);
ret = -1;
goto err;
}
/* parse arguments */
while ((c = getopt_long(argc, argv, "a:t:g:hd:f:s:", long_opts, NULL))
> 0) {
switch (c) {
case 'h':
usage(0, argv[0]);
ret = 0;
goto err;
case 'a':
/* set admin uid/gid */
if (parse_uid_gid(optarg, &auid, &agid, argv[0]))
goto err;
break;
case 't':
/* set task uid/gid */
if (parse_uid_gid(optarg, &tuid, &tgid, argv[0]))
goto err;
break;
case 'g':
ret = parse_cgroup_spec(cgroup_list, optarg, capacity);
if (ret) {
fprintf(stderr, "%s: "
"cgroup controller and path"
"parsing failed (%s)\n",
argv[0], argv[optind]);
ret = -1;
goto err;
}
break;
case 'd':
dirm_change = 1;
ret = parse_mode(optarg, &dir_mode, argv[0]);
break;
case 'f':
filem_change = 1;
ret = parse_mode(optarg, &file_mode, argv[0]);
break;
case 's':
filem_change = 1;
ret = parse_mode(optarg, &tasks_mode, argv[0]);
break;
default:
usage(1, argv[0]);
ret = -1;
goto err;
}
}
/* no cgroup name */
if (argv[optind]) {
fprintf(stderr, "%s: "
"wrong arguments (%s)\n",
argv[0], argv[optind]);
ret = -1;
goto err;
}
/* initialize libcg */
ret = cgroup_init();
if (ret) {
fprintf(stderr, "%s: "
"libcgroup initialization failed: %s\n",
argv[0], cgroup_strerror(ret));
goto err;
}
/* for each new cgroup */
for (i = 0; i < capacity; i++) {
if (!cgroup_list[i])
break;
/* create the new cgroup structure */
cgroup = cgroup_new_cgroup(cgroup_list[i]->path);
if (!cgroup) {
ret = ECGFAIL;
fprintf(stderr, "%s: can't add new cgroup: %s\n",
argv[0], cgroup_strerror(ret));
goto err;
}
/* set uid and gid for the new cgroup based on input options */
ret = cgroup_set_uid_gid(cgroup, tuid, tgid, auid, agid);
if (ret)
goto err;
/* add controllers to the new cgroup */
j = 0;
while (cgroup_list[i]->controllers[j]) {
cgc = cgroup_add_controller(cgroup,
cgroup_list[i]->controllers[j]);
if (!cgc) {
ret = ECGINVAL;
fprintf(stderr, "%s: "
"controller %s can't be add\n",
argv[0],
cgroup_list[i]->controllers[j]);
cgroup_free(&cgroup);
goto err;
}
j++;
}
/* all variables set so create cgroup */
if (dirm_change | filem_change)
cgroup_set_permissions(cgroup, dir_mode, file_mode,
tasks_mode);
ret = cgroup_create_cgroup(cgroup, 0);
if (ret) {
fprintf(stderr, "%s: "
"can't create cgroup %s: %s\n",
argv[0], cgroup->name, cgroup_strerror(ret));
cgroup_free(&cgroup);
goto err;
}
cgroup_free(&cgroup);
}
err:
if (cgroup_list) {
for (i = 0; i < capacity; i++) {
if (cgroup_list[i])
cgroup_free_group_spec(cgroup_list[i]);
}
free(cgroup_list);
}
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
}
int main(int argc, char *argv[])
{
int ret = 0;
int i, j;
int c;
int flags = 0;
int final_ret = 0;
int counter = 0;
int max = 0;
struct ext_cgroup_record *ecg_list = NULL;
int skip;
struct cgroup_group_spec **cgroup_list = NULL;
struct cgroup *cgroup;
struct cgroup_controller *cgc;
/* initialize libcg */
ret = cgroup_init();
if (ret) {
fprintf(stderr, "%s: "
"libcgroup initialization failed: %s\n",
argv[0], cgroup_strerror(ret));
goto err;
}
cgroup_list = calloc(argc, sizeof(struct cgroup_group_spec *));
if (cgroup_list == NULL) {
fprintf(stderr, "%s: out of memory\n", argv[0]);
ret = -1;
goto err;
}
ecg_list = calloc(argc, sizeof(struct ext_cgroup_record *));
if (cgroup_list == NULL) {
fprintf(stderr, "%s: out of memory\n", argv[0]);
ret = -1;
goto err;
}
/*
* Parse arguments
*/
while ((c = getopt_long(argc, argv, "rhg:",
long_options, NULL)) > 0) {
switch (c) {
case 'r':
flags |= CGFLAG_DELETE_RECURSIVE;
break;
case 'g':
ret = parse_cgroup_spec(cgroup_list, optarg, argc);
if (ret != 0) {
fprintf(stderr,
"%s: error parsing cgroup '%s'\n",
argv[0], optarg);
ret = -1;
goto err;
}
break;
case 'h':
usage(0, argv[0]);
ret = 0;
goto err;
default:
usage(1, argv[0]);
ret = -1;
goto err;
}
}
/* parse groups on command line */
for (i = optind; i < argc; i++) {
ret = parse_cgroup_spec(cgroup_list, argv[i], argc);
if (ret != 0) {
fprintf(stderr, "%s: error parsing cgroup '%s'\n",
argv[0], argv[i]);
ret = -1;
goto err;
}
}
/* for each cgroup to be deleted */
for (i = 0; i < argc; i++) {
if (!cgroup_list[i])
break;
/* create the new cgroup structure */
cgroup = cgroup_new_cgroup(cgroup_list[i]->path);
if (!cgroup) {
ret = ECGFAIL;
fprintf(stderr, "%s: can't create new cgroup: %s\n",
argv[0], cgroup_strerror(ret));
goto err;
}
/* add controllers to the cgroup */
j = 0;
while (cgroup_list[i]->controllers[j]) {
skip = 0;
/*
* save controller name, cg name and hierarchy number
* to determine whether we should skip adding controller
*/
if (counter == max) {
/*
* there is not enough space to store them,
* create it
*/
max = max + argc;
ecg_list = (struct ext_cgroup_record *)
realloc(ecg_list,
max * sizeof(struct ext_cgroup_record));
if (!ecg_list) {
fprintf(stderr, "%s: ", argv[0]);
fprintf(stderr, "not enough memory\n");
final_ret = -1;
goto err;
}
}
strncpy(ecg_list[counter].controller,
cgroup_list[i]->controllers[j], FILENAME_MAX);
ecg_list[counter].controller[FILENAME_MAX - 1] = '\0';
strncpy(ecg_list[counter].name,
cgroup_list[i]->path, FILENAME_MAX);
ecg_list[counter].name[FILENAME_MAX - 1] = '\0';
ret = skip_add_controller(counter, &skip, ecg_list);
if (ret)
goto err;
if (skip) {
/* don't add the controller, goto next one */
goto next;
}
cgc = cgroup_add_controller(cgroup,
cgroup_list[i]->controllers[j]);
if (!cgc) {
ret = ECGFAIL;
fprintf(stderr, "%s: "
"controller %s can't be added\n",
argv[0],
cgroup_list[i]->controllers[j]);
cgroup_free(&cgroup);
goto err;
}
next:
counter++;
j++;
}
ret = cgroup_delete_cgroup_ext(cgroup, flags);
/*
* Remember the errors and continue, try to remove all groups.
*/
if (ret != 0) {
fprintf(stderr, "%s: cannot remove group '%s': %s\n",
argv[0], cgroup->name,
cgroup_strerror(ret));
final_ret = ret;
}
cgroup_free(&cgroup);
}
ret = final_ret;
err:
if (ecg_list)
free(ecg_list);
if (cgroup_list) {
for (i = 0; i < argc; i++) {
if (cgroup_list[i])
cgroup_free_group_spec(cgroup_list[i]);
}
free(cgroup_list);
}
return ret;
}
static void cgroup_child_init(apr_pool_t *pool, server_rec *server)
{
cgroup *mygroup;
int ret;
cgroup_config *cgconf = ap_get_module_config(server->module_config, &cgroup_module);
if ((ret = cgroup_init()) > 0) {
ap_log_error(APLOG_MARK, APLOG_ERR, errno, server, "Could not initialize CGroups: %s", cgroup_strerror(ret));
}
else if ((mygroup = cgroup_new_cgroup(cgconf->default_cgroup)) == NULL) {
ap_log_error(APLOG_MARK, APLOG_ERR, errno, server, "Cannot allocate CGroup %s resources: %s", cgconf->default_cgroup, cgroup_strerror(ret));
}
else if ((ret = cgroup_get_cgroup(mygroup)) > 0) {
ap_log_error(APLOG_MARK, APLOG_ERR, errno, server, "Cannot get CGroup %s: %s", cgconf->default_cgroup, cgroup_strerror(ret));
}
else if ((ret = cgroup_attach_task(mygroup)) > 0) {
ap_log_error(APLOG_MARK, APLOG_ERR, errno, server, "Cannot assign to CGroup %s: %s", cgconf->default_cgroup, cgroup_strerror(ret));
}
else {
cg_enabled = 1;
cgroup_free(&mygroup);
}
}
static int cgroup_handler(request_rec *r)
{
cgroup *mygroup;
int ret =0;
if (!cg_enabled) {
return DECLINED;
}
cgroup_config *cgconf = ap_get_module_config(r->server->module_config, &cgroup_module);
if ((mygroup = cgroup_new_cgroup(cgconf->cgroup)) == NULL) {
ap_log_rerror(APLOG_MARK, APLOG_ERR, errno, r, "Cannot allocate CGroup %s resources: %s", cgconf->cgroup, cgroup_strerror(ret));
}
else if ((ret = cgroup_get_cgroup(mygroup)) > 0) {
ap_log_rerror(APLOG_MARK, APLOG_ERR, errno, r, "Cannot get CGroup %s: %s", cgconf->cgroup, cgroup_strerror(ret));
}
else if ((ret = cgroup_attach_task(mygroup)) > 0) {
ap_log_rerror(APLOG_MARK, APLOG_ERR, errno, r, "Cannot assign to CGroup %s: %s", cgconf->cgroup, cgroup_strerror(ret));
}
return DECLINED;
}
/* go through the list of all controllers gather them based on hierarchy number
and print them */
static int cgroup_list_all_controllers(const char *tname,
cont_name_t cont_name[CG_CONTROLLER_MAX], int c_number, int flags)
{
int ret = 0;
void *handle;
struct controller_data info;
int h_list[CG_CONTROLLER_MAX]; /* list of hierarchies */
int counter = 0;
int j;
int is_on_list = 0;
ret = cgroup_get_all_controller_begin(&handle, &info);
while (ret == 0) {
if (info.hierarchy == 0) {
/* the controller is not attached to any hierrachy */
if (flags & FL_ALL)
/* display only if -a flag is set */
printf("%s\n", info.name);
}
is_on_list = 0;
j = 0;
while ((is_on_list == 0) && (j < c_number)) {
if (strcmp(info.name, cont_name[j]) == 0) {
is_on_list = 1;
break;
}
j++;
}
if ((info.hierarchy != 0) &&
((flags & FL_ALL) ||
(!(flags & FL_LIST) || (is_on_list == 1)))) {
/* the controller is attached to some hierarchy
and either should be output all controllers,
or the controller is on the output list */
h_list[counter] = info.hierarchy;
counter++;
for (j = 0; j < counter-1; j++) {
/*
* the hierarchy already was on the list
* so remove the new record
*/
if (h_list[j] == info.hierarchy) {
counter--;
break;
}
}
}
ret = cgroup_get_all_controller_next(&handle, &info);
}
cgroup_get_all_controller_end(&handle);
if (ret == ECGEOF)
ret = 0;
if (ret) {
fprintf(stderr,
"cgroup_get_controller_begin/next failed (%s)\n",
cgroup_strerror(ret));
return ret;
}
for (j = 0; j < counter; j++)
ret = print_all_controllers_in_hierarchy(tname,
h_list[j], flags);
return ret;
}
int main(int argc, char *argv[])
{
int ret = 0, i;
int cg_specified = 0;
int flag_child = 0;
uid_t uid;
gid_t gid;
pid_t pid;
int c;
struct cgroup_group_spec *cgroup_list[CG_HIER_MAX];
memset(cgroup_list, 0, sizeof(cgroup_list));
while ((c = getopt_long(argc, argv, "+g:sh", longopts, NULL)) > 0) {
switch (c) {
case 'g':
ret = parse_cgroup_spec(cgroup_list, optarg,
CG_HIER_MAX);
if (ret) {
fprintf(stderr, "cgroup controller and path"
"parsing failed\n");
return -1;
}
cg_specified = 1;
break;
case 's':
flag_child |= CGROUP_DAEMON_UNCHANGE_CHILDREN;
break;
case 'h':
usage(0, argv[0]);
exit(0);
default:
usage(1, argv[0]);
exit(1);
}
}
/* Executable name */
if (!argv[optind]) {
usage(1, argv[0]);
exit(1);
}
/* Initialize libcg */
ret = cgroup_init();
if (ret) {
fprintf(stderr, "libcgroup initialization failed: %s\n",
cgroup_strerror(ret));
return ret;
}
/* Just for debugging purposes. */
uid = geteuid();
gid = getegid();
cgroup_dbg("My euid and eguid is: %d,%d\n", (int) uid, (int) gid);
uid = getuid();
gid = getgid();
pid = getpid();
ret = cgroup_register_unchanged_process(pid, flag_child);
if (ret) {
fprintf(stderr, "registration of process failed\n");
return ret;
}
/*
* 'cgexec' command file needs the root privilege for executing
* a cgroup_register_unchanged_process() by using unix domain
* socket, and an euid/egid should be changed to the executing user
* from a root user.
*/
if (setresuid(uid, uid, uid)) {
fprintf(stderr, "%s", strerror(errno));
return -1;
}
if (setresgid(gid, gid, gid)) {
fprintf(stderr, "%s", strerror(errno));
return -1;
}
if (cg_specified) {
/*
* User has specified the list of control group and
* controllers
* */
for (i = 0; i < CG_HIER_MAX; i++) {
if (!cgroup_list[i])
break;
ret = cgroup_change_cgroup_path(cgroup_list[i]->path,
pid,
(const char*const*) cgroup_list[i]->controllers);
if (ret) {
fprintf(stderr,
"cgroup change of group failed\n");
return ret;
}
}
} else {
/* Change the cgroup by determining the rules based on uid */
ret = cgroup_change_cgroup_flags(uid, gid,
argv[optind], pid, 0);
if (ret) {
fprintf(stderr, "cgroup change of group failed\n");
return ret;
}
}
/* Now exec the new process */
ret = execvp(argv[optind], &argv[optind]);
if (ret == -1) {
fprintf(stderr, "%s", strerror(errno));
return -1;
}
return 0;
}
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;
}
int
ProcFamily::aggregate_usage_cgroup(ProcFamilyUsage* usage)
{
if (!m_cm.isMounted(CgroupManager::MEMORY_CONTROLLER) || !m_cm.isMounted(CgroupManager::CPUACCT_CONTROLLER)
|| !m_cgroup.isValid()) {
return -1;
}
int err;
struct cgroup_stat stats;
void *handle = NULL;
u_int64_t tmp = 0, image = 0;
bool found_rss = false;
// Update memory
err = cgroup_read_stats_begin(MEMORY_CONTROLLER_STR, m_cgroup_string.c_str(), &handle, &stats);
while (err != ECGEOF) {
if (err > 0) {
dprintf(D_PROCFAMILY,
"Unable to read cgroup %s memory stats (ProcFamily %u): %u %s.\n",
m_cgroup_string.c_str(), m_root_pid, err, cgroup_strerror(err));
break;
}
if (_check_stat_uint64(stats, "total_rss", &tmp)) {
image += tmp;
usage->total_resident_set_size = tmp/1024;
found_rss = true;
} else if (_check_stat_uint64(stats, "total_mapped_file", &tmp)) {
image += tmp;
} else if (_check_stat_uint64(stats, "total_swap", &tmp)) {
image += tmp;
}
err = cgroup_read_stats_next(&handle, &stats);
}
if (handle != NULL) {
cgroup_read_stats_end(&handle);
}
if (found_rss) {
usage->total_image_size = image/1024;
} else {
dprintf(D_PROCFAMILY,
"Unable to find all necesary memory structures for cgroup %s"
" (ProcFamily %u)\n",
m_cgroup_string.c_str(), m_root_pid);
}
// The poor man's way of updating the max image size.
if (image > m_max_image_size) {
m_max_image_size = image/1024;
}
// XXX: Try again at using this at a later date.
// Currently, it reports the max size *including* the page cache.
// Doh!
//
// Try updating the max size using cgroups
//update_max_image_size_cgroup();
// Update CPU
get_cpu_usage_cgroup(usage->user_cpu_time, usage->sys_cpu_time);
aggregate_usage_cgroup_blockio(usage);
aggregate_usage_cgroup_blockio_io_serviced(usage);
// Finally, update the list of tasks
if ((err = count_tasks_cgroup()) < 0) {
return -err;
} else {
usage->num_procs = err;
}
return 0;
}
int main(int argc, char *argv[])
{
/* Patch to the log file */
const char *logp = NULL;
/* Syslog facility */
int facility = 0;
/* Verbose level */
int verbosity = 2;
/* For catching signals */
struct sigaction sa;
/* Should we daemonize? */
unsigned char daemon = 1;
/* Return codes */
int ret = 0;
struct passwd *pw;
struct group *gr;
/* Command line arguments */
const char *short_options = "hvqf:s::ndQu:g:";
struct option long_options[] = {
{"help", no_argument, NULL, 'h'},
{"verbose", no_argument, NULL, 'v'},
{"quiet", no_argument, NULL, 'q'},
{"logfile", required_argument, NULL, 'f'},
{"syslog", optional_argument, NULL, 's'},
{"nodaemon", no_argument, NULL, 'n'},
{"debug", no_argument, NULL, 'd'},
{"nolog", no_argument, NULL, 'Q'},
{"socket-user", required_argument, NULL, 'u'},
{"socket-group", required_argument, NULL, 'g'},
{NULL, 0, NULL, 0}
};
/* Make sure the user is root. */
if (getuid() != 0) {
fprintf(stderr, "Error: Only root can start/stop the control"
" group rules engine daemon\n");
ret = 1;
goto finished;
}
while (1) {
int c;
c = getopt_long(argc, argv, short_options, long_options, NULL);
if (c == -1)
break;
switch (c) {
case 'h': /* --help */
usage(stdout, "Help:\n");
ret = 0;
goto finished;
case 'v': /* --verbose */
verbosity++;
break;
case 'q': /* --quiet */
verbosity--;
break;
case 'Q': /* --nolog */
verbosity = 0;
break;
case 'f': /* --logfile=<filename> */
logp = optarg;
break;
case 's': /* --syslog=[facility] */
if (optarg) {
facility = cgre_parse_syslog_facility(optarg);
if (facility == 0) {
fprintf(stderr,
"Unknown syslog facility: %s\n",
optarg);
ret = 2;
goto finished;
}
} else {
facility = LOG_DAEMON;
}
break;
case 'n': /* --no-fork */
daemon = 0;
break;
case 'd': /* --debug */
/* same as -vvn */
daemon = 0;
verbosity = 4;
logp = "-";
break;
case 'u': /* --socket-user */
pw = getpwnam(optarg);
if (pw == NULL) {
usage(stderr, "Cannot find user %s", optarg);
ret = 3;
goto finished;
}
socket_user = pw->pw_uid;
cgroup_dbg("Using socket user %s id %d\n",
optarg, (int)socket_user);
break;
case 'g': /* --socket-group */
gr = getgrnam(optarg);
if (gr == NULL) {
usage(stderr, "Cannot find group %s", optarg);
ret = 3;
goto finished;
}
socket_group = gr->gr_gid;
cgroup_dbg("Using socket group %s id %d\n",
optarg, (int)socket_group);
break;
default:
usage(stderr, "");
ret = 2;
goto finished;
}
}
/* Initialize libcgroup. */
if ((ret = cgroup_init()) != 0) {
fprintf(stderr, "Error: libcgroup initialization failed, %s\n",
cgroup_strerror(ret));
goto finished;
}
/* Ask libcgroup to load the configuration rules. */
if ((ret = cgroup_init_rules_cache()) != 0) {
fprintf(stderr, "Error: libcgroup failed to initialize rules"
"cache from %s. %s\n", CGRULES_CONF_FILE,
cgroup_strerror(ret));
goto finished;
}
/* Now, start the daemon. */
ret = cgre_start_daemon(logp, facility, daemon, verbosity);
if (ret < 0) {
fprintf(stderr, "Error: Failed to launch the daemon, %s\n",
cgroup_strerror(ret));
goto finished;
}
/*
* Set up the signal handler to reload the cached rules upon reception
* of a SIGUSR2 signal.
*/
memset(&sa, 0, sizeof(sa));
sa.sa_handler = &cgre_flash_rules;
sigemptyset(&sa.sa_mask);
if ((ret = sigaction(SIGUSR2, &sa, NULL))) {
flog(LOG_ERR, "Failed to set up signal handler for SIGUSR2."
" Error: %s", strerror(errno));
goto finished;
}
/*
* Set up the signal handler to catch SIGINT and SIGTERM so that we
* can exit gracefully.
*/
sa.sa_handler = &cgre_catch_term;
ret = sigaction(SIGINT, &sa, NULL);
ret |= sigaction(SIGTERM, &sa, NULL);
if (ret) {
flog(LOG_ERR, "Failed to set up the signal handler. Error:"
" %s", strerror(errno));
goto finished;
}
/* Print the configuration to the log file, or stdout. */
if (logfile && loglevel >= LOG_INFO)
cgroup_print_rules_config(logfile);
flog(LOG_NOTICE, "Started the CGroup Rules Engine Daemon.");
/* We loop endlesly in this function, unless we encounter an error. */
ret = cgre_create_netlink_socket_process_msg();
finished:
if (logfile && logfile != stdout)
fclose(logfile);
return ret;
}
int main(int argc, char *argv[])
{
int ret = 0;
int i, j;
int c;
int flags = 0;
int final_ret = 0;
struct cgroup_group_spec **cgroup_list = NULL;
struct cgroup *cgroup;
struct cgroup_controller *cgc;
/* initialize libcg */
ret = cgroup_init();
if (ret) {
fprintf(stderr, "%s: "
"libcgroup initialization failed: %s\n",
argv[0], cgroup_strerror(ret));
goto err;
}
cgroup_list = calloc(argc, sizeof(struct cgroup_group_spec *));
if (cgroup_list == NULL) {
fprintf(stderr, "%s: out of memory\n", argv[0]);
ret = -1;
goto err;
}
/*
* Parse arguments
*/
while ((c = getopt_long(argc, argv, "rhg:",
long_options, NULL)) > 0) {
switch (c) {
case 'r':
flags |= CGFLAG_DELETE_RECURSIVE;
break;
case 'g':
ret = parse_cgroup_spec(cgroup_list, optarg, argc);
if (ret != 0) {
fprintf(stderr,
"%s: error parsing cgroup '%s'\n",
argv[0], optarg);
ret = -1;
goto err;
}
break;
case 'h':
usage(0, argv[0]);
ret = 0;
goto err;
default:
usage(1, argv[0]);
ret = -1;
goto err;
}
}
/* parse groups on command line */
for (i = optind; i < argc; i++) {
ret = parse_cgroup_spec(cgroup_list, argv[i], argc);
if (ret != 0) {
fprintf(stderr, "%s: error parsing cgroup '%s'\n",
argv[0], argv[i]);
ret = -1;
goto err;
}
}
/* for each cgroup to be deleted */
for (i = 0; i < argc; i++) {
if (!cgroup_list[i])
break;
/* create the new cgroup structure */
cgroup = cgroup_new_cgroup(cgroup_list[i]->path);
if (!cgroup) {
ret = ECGFAIL;
fprintf(stderr, "%s: can't create new cgroup: %s\n",
argv[0], cgroup_strerror(ret));
goto err;
}
/* add controllers to the cgroup */
j = 0;
while (cgroup_list[i]->controllers[j]) {
cgc = cgroup_add_controller(cgroup,
cgroup_list[i]->controllers[j]);
if (!cgc) {
ret = ECGFAIL;
fprintf(stderr, "%s: "
"controller %s can't be added\n",
argv[0],
cgroup_list[i]->controllers[j]);
cgroup_free(&cgroup);
goto err;
}
j++;
}
ret = cgroup_delete_cgroup_ext(cgroup, flags);
/*
* Remember the errors and continue, try to remove all groups.
*/
if (ret != 0) {
fprintf(stderr, "%s: cannot remove group '%s': %s\n",
argv[0], cgroup->name,
cgroup_strerror(ret));
final_ret = ret;
}
cgroup_free(&cgroup);
}
ret = final_ret;
err:
if (cgroup_list) {
for (i = 0; i < argc; i++) {
if (cgroup_list[i])
cgroup_free_group_spec(cgroup_list[i]);
}
free(cgroup_list);
}
return ret;
}
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);
}