/* We don't check the uid in this function, anyone may list the task info. */
static int
_handle_list_pids(int fd, slurmd_job_t *job)
{
int i;
pid_t *pids = NULL;
int npids = 0;
uint32_t pid;
debug("_handle_list_pids for job %u.%u", job->jobid, job->stepid);
slurm_container_get_pids(job->cont_id, &pids, &npids);
safe_write(fd, &npids, sizeof(uint32_t));
for (i = 0; i < npids; i++) {
pid = (uint32_t)pids[i];
safe_write(fd, &pid, sizeof(uint32_t));
}
if (npids > 0)
xfree(pids);
return SLURM_SUCCESS;
rwfail:
if (npids > 0)
xfree(pids);
return SLURM_FAILURE;
}
/*
* _get_process_data() - Build a table of all current processes
*
* IN: pid.
*
* OUT: none
*
* THREADSAFE! Only one thread ever gets here.
*
* Assumption:
* Any file with a name of the form "/proc/[0-9]+/stat"
* is a Linux-style stat entry. We disregard the data if they look
* wrong.
*/
static void _get_process_data(void)
{
struct procsinfo proc;
pid_t *pids = NULL;
int npids = 0;
int i;
uint32_t total_job_mem = 0, total_job_vsize = 0;
int pid = 0;
static int processing = 0;
prec_t *prec = NULL;
struct jobacctinfo *jobacct = NULL;
List prec_list = NULL;
ListIterator itr;
ListIterator itr2;
if (!pgid_plugin && (cont_id == (uint64_t)NO_VAL)) {
debug("cont_id hasn't been set yet not running poll");
return;
}
if(processing) {
debug("already running, returning");
return;
}
processing = 1;
prec_list = list_create(_destroy_prec);
if(!pgid_plugin) {
/* get only the processes in the proctrack container */
slurm_container_get_pids(cont_id, &pids, &npids);
if (!npids) {
debug4("no pids in this container %"PRIu64"", cont_id);
goto finished;
}
for (i = 0; i < npids; i++) {
pid = pids[i];
if(!getprocs(&proc, sizeof(proc), 0, 0, &pid, 1))
continue; /* Assume the process went away */
prec = xmalloc(sizeof(prec_t));
list_append(prec_list, prec);
prec->pid = proc.pi_pid;
prec->ppid = proc.pi_ppid;
prec->usec = proc.pi_ru.ru_utime.tv_sec +
proc.pi_ru.ru_utime.tv_usec * 1e-6;
prec->ssec = proc.pi_ru.ru_stime.tv_sec +
proc.pi_ru.ru_stime.tv_usec * 1e-6;
prec->pages = proc.pi_majflt;
prec->rss = (proc.pi_trss + proc.pi_drss) * pagesize;
//prec->rss *= 1024;
prec->vsize = (proc.pi_tsize / 1024);
prec->vsize += (proc.pi_dvm * pagesize);
//prec->vsize *= 1024;
/* debug("vsize = %f = (%d/1024)+(%d*%d)", */
/* prec->vsize, proc.pi_tsize, proc.pi_dvm, pagesize); */
}
} else {
while(getprocs(&proc, sizeof(proc), 0, 0, &pid, 1) == 1) {
prec = xmalloc(sizeof(prec_t));
list_append(prec_list, prec);
prec->pid = proc.pi_pid;
prec->ppid = proc.pi_ppid;
prec->usec = proc.pi_ru.ru_utime.tv_sec +
proc.pi_ru.ru_utime.tv_usec * 1e-6;
prec->ssec = proc.pi_ru.ru_stime.tv_sec +
proc.pi_ru.ru_stime.tv_usec * 1e-6;
prec->pages = proc.pi_majflt;
prec->rss = (proc.pi_trss + proc.pi_drss) * pagesize;
//prec->rss *= 1024;
prec->vsize = (proc.pi_tsize / 1024);
prec->vsize += (proc.pi_dvm * pagesize);
//prec->vsize *= 1024;
/* debug("vsize = %f = (%d/1024)+(%d*%d)", */
/* prec->vsize, proc.pi_tsize, proc.pi_dvm, pagesize); */
}
}
if(!list_count(prec_list))
goto finished;
slurm_mutex_lock(&jobacct_lock);
if(!task_list || !list_count(task_list)) {
slurm_mutex_unlock(&jobacct_lock);
goto finished;
}
itr = list_iterator_create(task_list);
while((jobacct = list_next(itr))) {
itr2 = list_iterator_create(prec_list);
while((prec = list_next(itr2))) {
//debug2("pid %d ? %d", prec->ppid, jobacct->pid);
if (prec->pid == jobacct->pid) {
/* find all my descendents */
_get_offspring_data(prec_list, prec,
prec->pid);
/* tally their usage */
jobacct->max_rss = jobacct->tot_rss =
MAX(jobacct->max_rss, (int)prec->rss);
total_job_mem += jobacct->max_rss;
jobacct->max_vsize = jobacct->tot_vsize =
MAX(jobacct->max_vsize,
(int)prec->vsize);
total_job_vsize += prec->vsize;
jobacct->max_pages = jobacct->tot_pages =
MAX(jobacct->max_pages, prec->pages);
jobacct->min_cpu = jobacct->tot_cpu =
MAX(jobacct->min_cpu,
(prec->usec + prec->ssec));
debug2("%d size now %d %d time %d",
jobacct->pid, jobacct->max_rss,
jobacct->max_vsize, jobacct->tot_cpu);
break;
}
}
list_iterator_destroy(itr2);
}
list_iterator_destroy(itr);
slurm_mutex_unlock(&jobacct_lock);
if (jobacct_mem_limit) {
debug("Step %u.%u memory used:%u limit:%u KB",
jobacct_job_id, jobacct_step_id,
total_job_mem, jobacct_mem_limit);
}
if (jobacct_job_id && jobacct_mem_limit &&
(total_job_mem > jobacct_mem_limit)) {
if (jobacct_step_id == NO_VAL) {
error("Job %u exceeded %u KB memory limit, being "
"killed", jobacct_job_id, jobacct_mem_limit);
} else {
error("Step %u.%u exceeded %u KB memory limit, being "
"killed", jobacct_job_id, jobacct_step_id,
jobacct_mem_limit);
}
_acct_kill_step();
} else if (jobacct_job_id && jobacct_vmem_limit &&
(total_job_vsize > jobacct_vmem_limit)) {
if (jobacct_step_id == NO_VAL) {
error("Job %u exceeded %u KB virtual memory limit, "
"being killed", jobacct_job_id,
jobacct_vmem_limit);
} else {
error("Step %u.%u exceeded %u KB virtual memory "
"limit, being killed", jobacct_job_id,
jobacct_step_id, jobacct_vmem_limit);
}
_acct_kill_step();
}
finished:
list_destroy(prec_list);
processing = 0;
return;
}
/*
* _get_process_data() - Build a table of all current processes
*
* IN: pid.
*
* OUT: none
*
* THREADSAFE! Only one thread ever gets here.
*
* Assumption:
* Any file with a name of the form "/proc/[0-9]+/stat"
* is a Linux-style stat entry. We disregard the data if they look
* wrong.
*/
static void _get_process_data(void)
{
static int slash_proc_open = 0;
struct dirent *slash_proc_entry;
char *iptr = NULL, *optr = NULL;
FILE *stat_fp = NULL;
char proc_stat_file[256]; /* Allow ~20x extra length */
List prec_list = NULL;
pid_t *pids = NULL;
int npids = 0;
uint32_t total_job_mem = 0, total_job_vsize = 0;
int i, fd;
ListIterator itr;
ListIterator itr2;
prec_t *prec = NULL;
struct jobacctinfo *jobacct = NULL;
static int processing = 0;
long hertz;
if (!pgid_plugin && (cont_id == (uint64_t)NO_VAL)) {
debug("cont_id hasn't been set yet not running poll");
return;
}
if(processing) {
debug("already running, returning");
return;
}
processing = 1;
prec_list = list_create(_destroy_prec);
hertz = sysconf(_SC_CLK_TCK);
if (hertz < 1) {
error ("_get_process_data: unable to get clock rate");
hertz = 100; /* default on many systems */
}
if(!pgid_plugin) {
/* get only the processes in the proctrack container */
slurm_container_get_pids(cont_id, &pids, &npids);
if(!npids) {
debug4("no pids in this container %"PRIu64"", cont_id);
goto finished;
}
for (i = 0; i < npids; i++) {
snprintf(proc_stat_file, 256,
"/proc/%d/stat", pids[i]);
if ((stat_fp = fopen(proc_stat_file, "r"))==NULL)
continue; /* Assume the process went away */
/*
* Close the file on exec() of user tasks.
*
* NOTE: If we fork() slurmstepd after the
* fopen() above and before the fcntl() below,
* then the user task may have this extra file
* open, which can cause problems for
* checkpoint/restart, but this should be a very rare
* problem in practice.
*/
fd = fileno(stat_fp);
fcntl(fd, F_SETFD, FD_CLOEXEC);
prec = xmalloc(sizeof(prec_t));
if (_get_process_data_line(fd, prec))
list_append(prec_list, prec);
else
xfree(prec);
fclose(stat_fp);
}
} else {
slurm_mutex_lock(&reading_mutex);
if (slash_proc_open) {
rewinddir(slash_proc);
} else {
slash_proc=opendir("/proc");
if (slash_proc == NULL) {
perror("opening /proc");
slurm_mutex_unlock(&reading_mutex);
goto finished;
}
slash_proc_open=1;
}
strcpy(proc_stat_file, "/proc/");
while ((slash_proc_entry = readdir(slash_proc))) {
/* Save a few cyles by simulating
strcat(statFileName, slash_proc_entry->d_name);
strcat(statFileName, "/stat");
while checking for a numeric filename (which really
should be a pid).
*/
optr = proc_stat_file + sizeof("/proc");
iptr = slash_proc_entry->d_name;
i = 0;
do {
if((*iptr < '0')
|| ((*optr++ = *iptr++) > '9')) {
i = -1;
break;
}
} while (*iptr);
if(i == -1)
continue;
iptr = (char*)"/stat";
do {
*optr++ = *iptr++;
} while (*iptr);
*optr = 0;
if ((stat_fp = fopen(proc_stat_file,"r"))==NULL)
continue; /* Assume the process went away */
/*
* Close the file on exec() of user tasks.
*
* NOTE: If we fork() slurmstepd after the
* fopen() above and before the fcntl() below,
* then the user task may have this extra file
* open, which can cause problems for
* checkpoint/restart, but this should be a very rare
* problem in practice.
*/
fd = fileno(stat_fp);
fcntl(fd, F_SETFD, FD_CLOEXEC);
prec = xmalloc(sizeof(prec_t));
if (_get_process_data_line(fd, prec))
list_append(prec_list, prec);
else
xfree(prec);
fclose(stat_fp);
}
slurm_mutex_unlock(&reading_mutex);
}
if (!list_count(prec_list)) {
goto finished; /* We have no business being here! */
}
slurm_mutex_lock(&jobacct_lock);
if(!task_list || !list_count(task_list)) {
slurm_mutex_unlock(&jobacct_lock);
goto finished;
}
itr = list_iterator_create(task_list);
while((jobacct = list_next(itr))) {
itr2 = list_iterator_create(prec_list);
while((prec = list_next(itr2))) {
if (prec->pid == jobacct->pid) {
#if _DEBUG
info("pid:%u ppid:%u rss:%d KB",
prec->pid, prec->ppid, prec->rss);
#endif
/* find all my descendents */
_get_offspring_data(prec_list,
prec, prec->pid);
/* tally their usage */
jobacct->max_rss = jobacct->tot_rss =
MAX(jobacct->max_rss, prec->rss);
total_job_mem += prec->rss;
jobacct->max_vsize = jobacct->tot_vsize =
MAX(jobacct->max_vsize, prec->vsize);
total_job_vsize += prec->vsize;
jobacct->max_pages = jobacct->tot_pages =
MAX(jobacct->max_pages, prec->pages);
jobacct->min_cpu = jobacct->tot_cpu =
MAX(jobacct->min_cpu,
(prec->ssec / hertz +
prec->usec / hertz));
debug2("%d mem size %u %u time %u(%u+%u)",
jobacct->pid, jobacct->max_rss,
jobacct->max_vsize, jobacct->tot_cpu,
prec->usec, prec->ssec);
break;
}
}
list_iterator_destroy(itr2);
}
list_iterator_destroy(itr);
slurm_mutex_unlock(&jobacct_lock);
if (jobacct_mem_limit) {
if (jobacct_step_id == NO_VAL) {
debug("Job %u memory used:%u limit:%u KB",
jobacct_job_id, total_job_mem, jobacct_mem_limit);
} else {
debug("Step %u.%u memory used:%u limit:%u KB",
jobacct_job_id, jobacct_step_id,
total_job_mem, jobacct_mem_limit);
}
}
if (jobacct_job_id && jobacct_mem_limit &&
(total_job_mem > jobacct_mem_limit)) {
if (jobacct_step_id == NO_VAL) {
error("Job %u exceeded %u KB memory limit, being "
"killed", jobacct_job_id, jobacct_mem_limit);
} else {
error("Step %u.%u exceeded %u KB memory limit, being "
"killed", jobacct_job_id, jobacct_step_id,
jobacct_mem_limit);
}
_acct_kill_step();
} else if (jobacct_job_id && jobacct_vmem_limit &&
(total_job_vsize > jobacct_vmem_limit)) {
if (jobacct_step_id == NO_VAL) {
error("Job %u exceeded %u KB virtual memory limit, "
"being killed", jobacct_job_id,
jobacct_vmem_limit);
} else {
error("Step %u.%u exceeded %u KB virtual memory "
"limit, being killed", jobacct_job_id,
jobacct_step_id, jobacct_vmem_limit);
}
_acct_kill_step();
}
finished:
list_destroy(prec_list);
processing = 0;
return;
}
