mom_mach.c

openPBS的开放源代码

 *
 *	If set_mode is SET_LIMIT_ALTER, we are beening called to modify
 *	existing limits.  Cannot alter those set by setrlimit (kernel)
 *	because we are the wrong process.  
 */
int mom_set_limits(pjob, set_mode)
    job			*pjob;
    int			 set_mode;	/* SET_LIMIT_SET or SET_LIMIT_ALTER */
{
	char		*id = "mom_set_limits";
	char		*pname;
	int		retval;
	unsigned long	value;	/* place in which to build resource value */
	resource	*pres;
       	struct rlimit	reslim;
	unsigned long	mem_limit  = 0;

	DBPRT(("%s: entered\n", id))
	assert(pjob != NULL);
	assert(pjob->ji_wattr[(int)JOB_ATR_resource].at_type == ATR_TYPE_RESC);
	pres = (resource *)
	    GET_NEXT(pjob->ji_wattr[(int)JOB_ATR_resource].at_val.at_list);

/*
 * Cycle through all the resource specifications,
 * setting limits appropriately (SET_LIMIT_SET).
 */

	while (pres != NULL) {
		assert(pres->rs_defin != NULL);
		pname = pres->rs_defin->rs_name;
		assert(pname != NULL);
		assert(*pname != '\0');

		if (strcmp(pname, "cput") == 0) {
			/* cpu time - check, if less than pcput use it */
			retval = gettime(pres, &value);
			if (retval != PBSE_NONE)
			        return (error(pname, retval));
		} else if (strcmp(pname, "pcput") == 0) {
			/* process cpu time - set */
			retval = gettime(pres, &value);
			if (retval != PBSE_NONE)
			        return (error(pname, retval));
			reslim.rlim_cur = reslim.rlim_max = 
				(unsigned long)((double)value / cputfactor);
			if (setrlimit(RLIMIT_CPU, &reslim) < 0)
	        		return (error("RLIMIT_CPU", PBSE_SYSTEM));
		} else if (strcmp(pname, "file") == 0) {	/* set */
			if (set_mode == SET_LIMIT_SET)  {
			    retval = getsize(pres, &value);
			    if (retval != PBSE_NONE)
			        return (error(pname, retval));
			    if (value > INT_MAX)
			        return (error(pname, PBSE_BADATVAL));
			    reslim.rlim_cur = reslim.rlim_max = value;
			    if (setrlimit(RLIMIT_FSIZE, &reslim) < 0)
			        return (error(pname, PBSE_SYSTEM));
			}
		} else if (strcmp(pname, "vmem") == 0) {	/* check */
			retval = getsize(pres, &value);
			if (retval != PBSE_NONE)
			        return (error(pname, retval));
			if ((mem_limit == 0) || (value < mem_limit))
				mem_limit = value;
		} else if (strcmp(pname, "pvmem") == 0) {	/* set */
			retval = getsize(pres, &value);
			if (retval != PBSE_NONE)
			        return (error(pname, retval));
			if (value > INT_MAX)
			        return (error(pname, PBSE_BADATVAL));
			if ((mem_limit == 0) || (value < mem_limit))
				mem_limit = value;
		} else if (strcmp(pname, "mem") == 0) {		/* ignore */
		} else if (strcmp(pname, "pmem") == 0) {	/* set */
			if (set_mode == SET_LIMIT_SET)  {
			    retval = getsize(pres, &value);
			    if (retval != PBSE_NONE)
			        return (error(pname, retval));
			    reslim.rlim_cur = reslim.rlim_max = value;
			    if (setrlimit(RLIMIT_RSS, &reslim) < 0)
	        		return (error("RLIMIT_RSS", PBSE_SYSTEM));
			}
		} else if (strcmp(pname, "walltime") == 0) {	/* Check */
			retval = gettime(pres, &value);
			if (retval != PBSE_NONE)
			        return (error(pname, retval));
		} else if (strcmp(pname, "nice") == 0) {	/* set nice */
			if (set_mode == SET_LIMIT_SET)  {
			    errno = 0;
			    if ((nice((int)pres->rs_value.at_val.at_long) == -1)
				&& (errno != 0))
				return (error(pname, PBSE_BADATVAL));
			}
		} else if ((pres->rs_defin->rs_flags & ATR_DFLAG_RMOMIG) == 0)
			/* don't recognize and not marked as ignore by mom */
			return (error(pname, PBSE_UNKRESC));
		pres = (resource *)GET_NEXT(pres->rs_link);
	}
	if (set_mode == SET_LIMIT_SET)  {
	    /* if either of vmem or pvmem was given, set sys limit to lesser */
	    if (mem_limit != 0) {
		reslim.rlim_cur = reslim.rlim_max = mem_limit;
		if (setrlimit(RLIMIT_DATA, &reslim) < 0)
	        	return (error("RLIMIT_DATA", PBSE_SYSTEM));
		if (setrlimit(RLIMIT_STACK, &reslim) < 0)
	        	return (error("RLIMIT_STACK", PBSE_SYSTEM));
	    }
	}
	return (PBSE_NONE);
}

/*
 * State whether MOM main loop has to poll this job to determine if some
 * limits are being exceeded.
 *
 *	Sets flag TRUE if polling is necessary, FALSE otherwise.  Actual
 *	polling is done using the mom_over_limit machine-dependent function.
 */

int mom_do_poll(pjob)
    job			*pjob;
{
	char		*id = "mom_do_poll";
	char		*pname;
	resource	*pres;

	DBPRT(("%s: entered\n", id))
	assert(pjob != NULL);
	assert(pjob->ji_wattr[(int)JOB_ATR_resource].at_type == ATR_TYPE_RESC);
	pres = (resource *)
	    GET_NEXT(pjob->ji_wattr[(int)JOB_ATR_resource].at_val.at_list);

	while (pres != NULL) {
		assert(pres->rs_defin != NULL);
		pname = pres->rs_defin->rs_name;
		assert(pname != NULL);
		assert(*pname != '\0');

		if (strcmp(pname, "walltime") == 0 ||
		    strcmp(pname, "cput") == 0 ||
		    strcmp(pname, "pvmem") == 0 ||
		    strcmp(pname, "vmem") == 0)
			return (TRUE);
		pres = (resource *)GET_NEXT(pres->rs_link);
	}

	return (FALSE);
}

/*
 * Setup for polling.
 *
 *	Open kernel device and get namelist info.
 */
int mom_open_poll()
{
	char		*id = "mom_open_poll";

	DBPRT(("%s: entered\n", id))
	if (kd == NULL) {
		kd = kvm_open(NULL, NULL, NULL, O_RDONLY, "mom");
		if (kd == NULL) {
			log_err(errno, id, "kvm_open");
			return (PBSE_SYSTEM);
		}
	}

	return (PBSE_NONE);
}


int
qs_cmp(a, b)
    struct	kinfo_proc	*a;
    struct	kinfo_proc	*b;
{
	return((int)a->kp_eproc.e_paddr - (int)b->kp_eproc.e_paddr);
}

int
bs_cmp(key, member)
    struct	session		*key;
    struct	kinfo_proc	*member;
{
	return((int)key->s_leader - (int)member->kp_eproc.e_paddr);
}

/*
 * Declare start of polling loop.
 *
 *	Until the next call to mom_get_sample, all mom_over_limit calls will
 *	use the same data.  Returns a PBS error code.
 */

int
mom_get_sample()
{
	char			*id = "mom_get_sample";
	int			i;
	struct	session		ss;
	struct	kinfo_proc	*kp;
	struct	kinfo_proc	*leader;
	pid_t			sid;

	DBPRT(("%s: entered\n", id))
	if (sess_tbl)
		free(sess_tbl);

	if (kd == NULL)
		return (PBSE_INTERNAL);

	proc_tbl = kvm_getprocs(kd, KERN_PROC_ALL, 0, &nproc);
	if (proc_tbl == NULL) {
		sprintf(log_buffer,
		    "kvm_getprocs: %s", kvm_geterr(kd));
		log_err(errno, id, log_buffer);
		return (PBSE_SYSTEM);
	}
	sess_tbl = (pid_t *)calloc(nproc, sizeof(pid_t));
	if (sess_tbl == NULL) {
		sprintf(log_buffer,
		    "can't allocate memory for session table");
		log_err(errno, id, log_buffer);
		return (PBSE_SYSTEM);
	}

	qsort(proc_tbl, nproc, sizeof(struct kinfo_proc), qs_cmp);

	for (i=0, kp=proc_tbl; i<nproc; i++, kp++) {
		if (kvm_read(kd, (u_long)kp->kp_eproc.e_sess, &ss, sizeof(ss))
				!= sizeof(ss)) {
			sprintf(log_buffer,
				"kvm_read: %s", kvm_geterr(kd));
			log_err(errno, id, log_buffer);
			return (PBSE_SYSTEM);
		}
		DBPRT(("%s: pid %d ", id, kp->kp_proc.p_pid))
		if (ss.s_leader == kp->kp_eproc.e_paddr ||
				ss.s_leader == NULL) {
			DBPRT(("own leader 0x%x\n", ss.s_leader))
			sid = kp->kp_proc.p_pid;
		}
		else {
			leader = bsearch(&ss, proc_tbl, nproc,
				sizeof(struct kinfo_proc), bs_cmp);
			sid = leader ? leader->kp_proc.p_pid : 0;
			DBPRT(("search got %d\n", sid))
		}
		sess_tbl[i] = sid;
	}

	return (PBSE_NONE);
}

/*
 * Measure job resource usage and compare with its limits.
 *
 *	If it has exceeded any well-formed polled limit return TRUE.
 *	Otherwise, return FALSE.
 */

int mom_over_limit(pjob)
    job			*pjob;
{
	char		*id = "mom_over_limit";
	char		*pname;
	int		retval;
	unsigned long	value, num;
	resource	*pres;

	assert(pjob != NULL);
	assert(pjob->ji_wattr[(int)JOB_ATR_resource].at_type == ATR_TYPE_RESC);
	pres = (resource *)
	    GET_NEXT(pjob->ji_wattr[(int)JOB_ATR_resource].at_val.at_list);

	DBPRT(("%s: entered\n", id))
	for ( ; pres != NULL; pres = (resource *)GET_NEXT(pres->rs_link)) {
		assert(pres->rs_defin != NULL);
		pname = pres->rs_defin->rs_name;
		assert(pname != NULL);
		assert(*pname != '\0');
		if (strcmp(pname, "cput") == 0) {
			retval = gettime(pres, &value);
			if (retval != PBSE_NONE)
				continue;
			if ((num = cput_sum(pjob)) > value) {
				sprintf(log_buffer,
					"cput %lu exceeded limit %lu",
					num, value);
				return (TRUE);
			}
		} else if (strcmp(pname, "vmem") == 0) {
			retval = getsize(pres, &value);
			if (retval != PBSE_NONE)
				continue;
			if ((num = mem_sum(pjob)) > value) {
				sprintf(log_buffer,
					"vmem %lu exceeded limit %lu",
					num, value);
				return (TRUE);
			}
		} else if (strcmp(pname, "pvmem") == 0) {
			retval = getsize(pres, &value);
			if (retval != PBSE_NONE)
				continue;
			if (overmem_proc(pjob, value)) {
				sprintf(log_buffer, "pvmem exceeded limit %lu",
					value);
				return (TRUE);
			}
		} else if (strcmp(pname, "walltime") == 0) {
			retval = gettime(pres, &value);
			if (retval != PBSE_NONE)
				continue;
			num = (unsigned long)(wallfactor * (double)(time_now - pjob->ji_qs.ji_stime));
			if (num > value) {
				sprintf(log_buffer,
					"walltime %lu exceeded limit %lu",
					num, value);
				return (TRUE);
			}
		}
	}

	return (FALSE);
}

/*
 * Update the job attribute for resources used.
 *
 *	The first time this is called for a job, set up resource entries for
 *	each resource that can be reported for this machine.  Fill in the
 *	correct values.  Return an error code.
 *
 *	Assumes that the session ID attribute has already been set.
 */
int mom_set_use(pjob)
    job		*pjob;
{
	char		*id = "mom_set_use";
	resource	*pres;
	attribute	*at;
	resource_def	*rd;
	unsigned long	*lp, lnum;

	assert(pjob != NULL);

	at = &pjob->ji_wattr[(int)JOB_ATR_resc_used];
	assert(at->at_type == ATR_TYPE_RESC);

	at->at_flags |= ATR_VFLAG_MODIFY;
	if ((at->at_flags & ATR_VFLAG_SET) == 0) {
		at->at_flags |= ATR_VFLAG_SET;

		rd = find_resc_def(svr_resc_def, "cput", svr_resc_size);
		assert(rd != NULL);
		pres = add_resource_entry(at, rd);
		assert(pres != NULL);
		pres->rs_value.at_flags |= ATR_VFLAG_SET;
		pres->rs_value.at_type = ATR_TYPE_LONG;
		pres->rs_value.at_val.at_long = 0;

		rd = find_resc_def(svr_resc_def, "vmem", svr_resc_size);
		assert(rd != NULL);
		pres = add_resource_entry(at, rd);
		assert(pres != NULL);
		pres->rs_value.at_flags |= ATR_VFLAG_SET;
		pres->rs_value.at_type = ATR_TYPE_SIZE;
		pres->rs_value.at_val.at_size.atsv_shift = 10; /* in KB */
		pres->rs_value.at_val.at_size.atsv_units = ATR_SV_BYTESZ;
		pres->rs_value.at_val.at_size.atsv_num = 0;

		rd = find_resc_def(svr_resc_def, "walltime", svr_resc_size);
		assert(rd != NULL);
		pres = add_resource_entry(at, rd);
		assert(pres != NULL);
		pres->rs_value.at_flags |= ATR_VFLAG_SET;
		pres->rs_value.at_type = ATR_TYPE_LONG;
		pres->rs_value.at_val.at_long = 0;

		rd = find_resc_def(svr_resc_def, "mem", svr_resc_size);
		assert(rd != NULL);
		pres = add_resource_entry(at, rd);
		assert(pres != NULL);
		pres->rs_value.at_flags |= ATR_VFLAG_SET;
		pres->rs_value.at_type = ATR_TYPE_SIZE;
		pres->rs_value.at_val.at_size.atsv_shift = 0;
		pres->rs_value.at_val.at_size.atsv_units = ATR_SV_BYTESZ;
		pres->rs_value.at_val.at_size.atsv_num = 0;
	}

	rd = find_resc_def(svr_resc_def, "cput", svr_resc_size);
	assert(rd != NULL);
	pres = find_resc_entry(at, rd);
	assert(pres != NULL);
	lp = (unsigned long *)&pres->rs_value.at_val.at_long;
	lnum = cput_sum(pjob);
	*lp = MAX(*lp, lnum);

	rd = find_resc_def(svr_resc_def, "vmem", svr_resc_size);
	assert(rd != NULL);
	pres = find_resc_entry(at, rd);
	assert(pres != NULL);
	lp = &pres->rs_value.at_val.at_size.atsv_num;
	lnum = (mem_sum(pjob) + 1023) >> 10;	/* in KB */
	*lp = MAX(*lp, lnum);

	rd = find_resc_def(svr_resc_def, "walltime", svr_resc_size);
	assert(rd != NULL);
	pres = find_resc_entry(at, rd);
	assert(pres != NULL);
	pres->rs_value.at_val.at_long = (long)((double)(time_now - pjob->ji_qs.ji_stime) * wallfactor);

	rd = find_resc_def(svr_resc_def, "mem", svr_resc_size);
	assert(rd != NULL);
	pres = find_resc_entry(at, rd);
	assert(pres != NULL);
	lp = &pres->rs_value.at_val.at_size.atsv_num;
	lnum = resi_sum(pjob);
	*lp = MAX(*lp, lnum);

	return (PBSE_NONE);
}

/*
 *	Kill a job task.
 *	Call with the job and a signal number.
 */
int kill_task(ptask, sig)
    task	*ptask;
    int		sig;
{
	char	*id = "kill_task";
	int	ct = 0;
	int	i, err;
	int	sesid;

	DBPRT(("%s entered\n", id))

	sesid = ptask->ti_qs.ti_sid;
	if (sesid <= 1)
		return 0;

	if ((err = mom_get_sample()) != PBSE_NONE)
		return 0;

	for (i=0; i<nproc; i++) {
		struct	kinfo_proc	*pp = &proc_tbl[i];

		if (sesid != sess_tbl[i])
			continue;

		DBPRT(("%s: send signal %d to pid %d\n", id,
				sig, pp->kp_proc.p_pid))
		(void)kill(pp->kp_proc.p_pid, sig);
		++ct;
	}
	return ct;
}

/*
 * Clean up everything related to polling.
 *
 *	In the case of the sun, close the kernal if it is open.
 */
int mom_close_poll()
{
	DBPRT(("mom_close_poll entered\n"))
	if (kd) {
		if (kvm_close(kd) != 0) {
			log_err(errno, "mom_close_poll", "kvm_close");
			return (PBSE_SYSTEM);
		}
		kd = NULL;
	}

	return (PBSE_NONE);
}

/*
 * mom_does_chkpnt - return 1 if mom supports checkpoint
 *			    0 if not
 */

int mom_does_chkpnt()
{
	return (0);
}

/*
 * Checkpoint the job.
 *
 *	If abort is true, kill it too.
 */

int mach_checkpoint(ptask, file, abort)
    task	*ptask;
    char	*file;
    int		abort;
{
       	return (-1);
}