xref: /titanic_51/usr/src/cmd/picl/plugins/sun4v/lib/snmp/snmplib.c (revision c8268b2c32246a4fd927df00921c772baab114e0)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * The snmp library helps to prepare the PDUs and communicate with
29  * the snmp agent on the SP side via the ds_snmp driver.
30  */
31 
32 #include <stdio.h>
33 #include <stdlib.h>
34 #include <string.h>
35 #include <unistd.h>
36 #include <thread.h>
37 #include <synch.h>
38 #include <errno.h>
39 #include <sys/time.h>
40 #include <sys/types.h>
41 #include <sys/stat.h>
42 #include <fcntl.h>
43 #include <libnvpair.h>
44 #include <sys/ds_snmp.h>
45 
46 #include "libpiclsnmp.h"
47 #include "snmplib.h"
48 #include "asn1.h"
49 #include "pdu.h"
50 #include "debug.h"
51 
52 #pragma init(libpiclsnmp_init)		/* need this in .init */
53 
54 /*
55  * Data from the MIB is fetched based on the hints about object
56  * groups received from (possibly many threads in) the application.
57  * However, the fetched data is kept in a common cache for use across
58  * all threads, so even a GETBULK is issued only when absolutely
59  * necessary.
60  *
61  * Note that locking is not fine grained (there's no locking per row)
62  * since we don't expect too many MT consumers right away.
63  *
64  */
65 static mutex_t	mibcache_lock;
66 static nvlist_t	**mibcache = NULL;
67 static uint_t	n_mibcache_rows = 0;
68 
69 static mutex_t snmp_reqid_lock;
70 static int snmp_reqid = 1;
71 
72 #ifdef SNMP_DEBUG
73 uint_t snmp_nsends = 0;
74 uint_t snmp_sentbytes = 0;
75 uint_t snmp_nrecvs = 0;
76 uint_t snmp_rcvdbytes = 0;
77 #endif
78 
79 #ifdef USE_SOCKETS
80 #define	SNMP_DEFAULT_PORT	161
81 #define	SNMP_MAX_RECV_PKTSZ	(64 * 1024)
82 #endif
83 
84 /*
85  * We need a reliably monotonic and stable source of time values to age
86  * entries in the mibcache toward expiration.  The code originally used
87  * gettimeofday(), but since that is subject to time-of-day changes made by
88  * the administrator, the values it returns do not satisfy our needs.
89  * Instead, we use gethrtime(), which is immune to time-of-day changes.
90  * However, since gethrtime() returns a signed 64-bit value in units of
91  * nanoseconds and we are using signed 32-bit timestamps, we always divide
92  * the result by (HRTIME_SCALE * NANOSEC) to scale it down into units of 10
93  * seconds.
94  *
95  * Note that the scaling factor means that the value of MAX_INCACHE_TIME
96  * from snmplib.h should also be in units of 10 seconds.
97  */
98 #define	GET_SCALED_HRTIME()	(int)(gethrtime() / (HRTIME_SCALE * NANOSEC))
99 
100 /*
101  * The mibcache code originally cached values for 300 seconds after fetching
102  * data via SNMP.  Subsequent reads within that 300 second window would come
103  * from the cache - which is quite a bit faster than an SNMP query - but the
104  * first request that came in more than 300 seconds after the previous SNMP
105  * query would trigger a new SNMP query.  This worked well as an
106  * optimization for frequent queries, but when data was only queried less
107  * frequently than every 300 seconds (as proved to be the case at multiple
108  * customer sites), the cache didn't help at all.
109  *
110  * To improve the performance of infrequent queries, code was added to the
111  * library to allow a client (i.e. a thread in the picl plugin) to proactively
112  * refresh cache entries without waiting for them to expire, thereby ensuring
113  * that all volatile entries in the cache at any given time are less than 300
114  * seconds old.  Whenever an SNMP query is generated to retrieve volatile data
115  * that will be cached, an entry is added in a refresh queue that tracks the
116  * parameters of the query and the time that it was made.  A client can query
117  * the age of the oldest item in the refresh queue and - at its discretion - can
118  * then force that query to be repeated in a manner that will update the
119  * mibcache entry even though it hasn't expired.
120  */
121 typedef struct {
122 	struct picl_snmphdl	*smd;
123 	char			*oidstrs;
124 	int			n_oids;
125 	int			row;
126 	int			last_fetch_time;	/* in scaled hrtime */
127 } refreshq_job_t;
128 
129 static mutex_t		refreshq_lock;
130 static refreshq_job_t	*refreshq = NULL;
131 static uint_t		n_refreshq_slots = 0;	/* # of alloc'ed job slots */
132 static uint_t		n_refreshq_jobs = 0;	/* # of unprocessed jobs */
133 static uint_t		refreshq_next_job = 0;	/* oldest unprocessed job */
134 static uint_t		refreshq_next_slot = 0;	/* next available job slot */
135 
136 
137 /*
138  * Static function declarations
139  */
140 static void	libpiclsnmp_init(void);
141 
142 static int	lookup_int(char *, int, int *, int);
143 static int	lookup_str(char *, int, char **, int);
144 static int	lookup_bitstr(char *, int, uchar_t **, uint_t *, int);
145 
146 static oidgroup_t *locate_oid_group(struct picl_snmphdl *, char *);
147 static int	search_oid_in_group(char *, char *, int);
148 
149 static snmp_pdu_t *fetch_single(struct picl_snmphdl *, char *, int, int *);
150 static snmp_pdu_t *fetch_next(struct picl_snmphdl *, char *, int, int *);
151 static void	fetch_bulk(struct picl_snmphdl *, char *, int, int, int, int *);
152 static int	fetch_single_str(struct picl_snmphdl *, char *, int,
153 		    char **, int *);
154 static int	fetch_single_int(struct picl_snmphdl *, char *, int,
155 		    int *, int *);
156 static int	fetch_single_bitstr(struct picl_snmphdl *, char *, int,
157 		    uchar_t **, uint_t *, int *);
158 
159 static int	snmp_send_request(struct picl_snmphdl *, snmp_pdu_t *, int *);
160 static int	snmp_recv_reply(struct picl_snmphdl *, snmp_pdu_t *, int *);
161 
162 static int	mibcache_realloc(int);
163 static void	mibcache_populate(snmp_pdu_t *, int);
164 static char	*oid_to_oidstr(oid *, size_t);
165 
166 static int	refreshq_realloc(int);
167 static int	refreshq_add_job(struct picl_snmphdl *, char *, int, int);
168 
169 
170 static void
171 libpiclsnmp_init(void)
172 {
173 	(void) mutex_init(&mibcache_lock, USYNC_THREAD, NULL);
174 	if (mibcache_realloc(0) < 0)
175 		(void) mutex_destroy(&mibcache_lock);
176 
177 	(void) mutex_init(&refreshq_lock, USYNC_THREAD, NULL);
178 	(void) mutex_init(&snmp_reqid_lock, USYNC_THREAD, NULL);
179 
180 	LOGINIT();
181 }
182 
183 picl_snmphdl_t
184 snmp_init()
185 {
186 	struct picl_snmphdl	*smd;
187 #ifdef USE_SOCKETS
188 	int	sbuf = (1 << 15);	/* 16K */
189 	int	rbuf = (1 << 17);	/* 64K */
190 	char	*snmp_agent_addr;
191 #endif
192 
193 	smd = (struct picl_snmphdl *)calloc(1, sizeof (struct picl_snmphdl));
194 	if (smd == NULL)
195 		return (NULL);
196 
197 #ifdef USE_SOCKETS
198 	if ((snmp_agent_addr = getenv("SNMP_AGENT_IPADDR")) == NULL)
199 		return (NULL);
200 
201 	if ((smd->fd = socket(PF_INET, SOCK_DGRAM, 0)) < 0)
202 		return (NULL);
203 
204 	(void) setsockopt(smd->fd, SOL_SOCKET, SO_SNDBUF, &sbuf, sizeof (int));
205 	(void) setsockopt(smd->fd, SOL_SOCKET, SO_RCVBUF, &rbuf, sizeof (int));
206 
207 	memset(&smd->agent_addr, 0, sizeof (struct sockaddr_in));
208 	smd->agent_addr.sin_family = AF_INET;
209 	smd->agent_addr.sin_port = htons(SNMP_DEFAULT_PORT);
210 	smd->agent_addr.sin_addr.s_addr = inet_addr(snmp_agent_addr);
211 #else
212 	smd->fd = open(DS_SNMP_DRIVER, O_RDWR);
213 	if (smd->fd < 0) {
214 		free(smd);
215 		return (NULL);
216 	}
217 #endif
218 
219 	return ((picl_snmphdl_t)smd);
220 }
221 
222 void
223 snmp_fini(picl_snmphdl_t hdl)
224 {
225 	struct picl_snmphdl	*smd = (struct picl_snmphdl *)hdl;
226 
227 	if (smd) {
228 		if (smd->fd >= 0) {
229 			(void) close(smd->fd);
230 		}
231 		free(smd);
232 	}
233 }
234 
235 int
236 snmp_reinit(picl_snmphdl_t hdl, int clr_linkreset)
237 {
238 	struct picl_snmphdl *smd = (struct picl_snmphdl *)hdl;
239 	nvlist_t *nvl;
240 	int i;
241 
242 	(void) mutex_lock(&mibcache_lock);
243 
244 	for (i = 0; i < n_mibcache_rows; i++) {
245 		if ((nvl = mibcache[i]) != NULL)
246 			nvlist_free(nvl);
247 	}
248 
249 	n_mibcache_rows = 0;
250 	if (mibcache) {
251 		free(mibcache);
252 		mibcache = NULL;
253 	}
254 
255 	(void) mutex_unlock(&mibcache_lock);
256 
257 	if (clr_linkreset) {
258 		if (smd == NULL || smd->fd < 0)
259 			return (-1);
260 		else
261 			return (ioctl(smd->fd, DSSNMP_CLRLNKRESET, NULL));
262 	}
263 
264 	return (0);
265 }
266 
267 void
268 snmp_register_group(picl_snmphdl_t hdl, char *oidstrs, int n_oids, int is_vol)
269 {
270 	struct picl_snmphdl *smd = (struct picl_snmphdl *)hdl;
271 	oidgroup_t	*oidg;
272 	oidgroup_t	*curr, *prev;
273 	char		*p;
274 	int		i, sz;
275 
276 	/*
277 	 * Allocate a new oidgroup_t
278 	 */
279 	oidg = (oidgroup_t *)calloc(1, sizeof (struct oidgroup));
280 	if (oidg == NULL)
281 		return;
282 
283 	/*
284 	 * Determine how much space is required to register this group
285 	 */
286 	sz = 0;
287 	p = oidstrs;
288 	for (i = 0; i < n_oids; i++) {
289 		sz += strlen(p) + 1;
290 		p = oidstrs + sz;
291 	}
292 
293 	/*
294 	 * Create this oid group
295 	 */
296 	if ((p = (char *)malloc(sz)) == NULL) {
297 		free((void *) oidg);
298 		return;
299 	}
300 
301 	(void) memcpy(p, oidstrs, sz);
302 
303 	oidg->next = NULL;
304 	oidg->oidstrs = p;
305 	oidg->n_oids = n_oids;
306 	oidg->is_volatile = is_vol;
307 
308 	/*
309 	 * Link it to the tail of the list of oid groups
310 	 */
311 	for (prev = NULL, curr = smd->group; curr; curr = curr->next)
312 		prev = curr;
313 
314 	if (prev == NULL)
315 		smd->group = oidg;
316 	else
317 		prev->next = oidg;
318 }
319 
320 /*
321  * snmp_get_int() takes in an OID and returns the integer value
322  * of the object referenced in the passed arg. It returns 0 on
323  * success and -1 on failure.
324  */
325 int
326 snmp_get_int(picl_snmphdl_t hdl, char *prefix, int row, int *val,
327     int *snmp_syserr)
328 {
329 	struct picl_snmphdl *smd = (struct picl_snmphdl *)hdl;
330 	oidgroup_t	*grp;
331 	int	ret;
332 	int	err = 0;
333 
334 	if (smd == NULL || prefix == NULL || val == NULL)
335 		return (-1);
336 
337 	/*
338 	 * If this item should not be cached, fetch it directly from
339 	 * the agent using fetch_single_xxx()
340 	 */
341 	if ((grp = locate_oid_group(smd, prefix)) == NULL) {
342 		ret = fetch_single_int(smd, prefix, row, val, &err);
343 
344 		if (snmp_syserr)
345 			*snmp_syserr = err;
346 
347 		return (ret);
348 	}
349 
350 	/*
351 	 * is it in the cache ?
352 	 */
353 	if (lookup_int(prefix, row, val, grp->is_volatile) == 0)
354 		return (0);
355 
356 	/*
357 	 * fetch it from the agent and populate the cache
358 	 */
359 	fetch_bulk(smd, grp->oidstrs, grp->n_oids, row, grp->is_volatile, &err);
360 	if (snmp_syserr)
361 		*snmp_syserr = err;
362 
363 	/*
364 	 * look it up again and return it
365 	 */
366 	if (lookup_int(prefix, row, val, grp->is_volatile) < 0)
367 		return (-1);
368 
369 	return (0);
370 }
371 
372 /*
373  * snmp_get_str() takes in an OID and returns the string value
374  * of the object referenced in the passed arg. Memory for the string
375  * is allocated within snmp_get_str() and is expected to be freed by
376  * the caller when it is no longer needed. The function returns 0
377  * on success and -1 on failure.
378  */
379 int
380 snmp_get_str(picl_snmphdl_t hdl, char *prefix, int row, char **strp,
381     int *snmp_syserr)
382 {
383 	struct picl_snmphdl *smd = (struct picl_snmphdl *)hdl;
384 	oidgroup_t	*grp;
385 	char	*val;
386 	int	ret;
387 	int	err = 0;
388 
389 	if (smd == NULL || prefix == NULL || strp == NULL)
390 		return (-1);
391 
392 	*strp = NULL;
393 	/*
394 	 * Check if this item is cacheable or not. If not, call
395 	 * fetch_single_* to get it directly from the agent
396 	 */
397 	if ((grp = locate_oid_group(smd, prefix)) == NULL) {
398 		ret = fetch_single_str(smd, prefix, row, strp, &err);
399 
400 		if (snmp_syserr)
401 			*snmp_syserr = err;
402 
403 		return (ret);
404 	}
405 
406 	/*
407 	 * See if it's in the cache already
408 	 */
409 	if (lookup_str(prefix, row, &val, grp->is_volatile) == 0) {
410 		if ((*strp = strdup(val)) == NULL)
411 			return (-1);
412 		else
413 			return (0);
414 	}
415 
416 	/*
417 	 * Fetch it from the agent and populate cache
418 	 */
419 	fetch_bulk(smd, grp->oidstrs, grp->n_oids, row, grp->is_volatile, &err);
420 	if (snmp_syserr)
421 		*snmp_syserr = err;
422 
423 	/*
424 	 * Retry lookup
425 	 */
426 	if (lookup_str(prefix, row, &val, grp->is_volatile) < 0)
427 		return (-1);
428 
429 
430 	if ((*strp = strdup(val)) == NULL)
431 		return (-1);
432 	else
433 		return (0);
434 }
435 
436 /*
437  * snmp_get_bitstr() takes in an OID and returns the bit string value
438  * of the object referenced in the passed args. Memory for the bitstring
439  * is allocated within the function and is expected to be freed by
440  * the caller when it is no longer needed. The function returns 0
441  * on success and -1 on failure.
442  */
443 int
444 snmp_get_bitstr(picl_snmphdl_t hdl, char *prefix, int row, uchar_t **bitstrp,
445     uint_t *nbytes, int *snmp_syserr)
446 {
447 	struct picl_snmphdl *smd = (struct picl_snmphdl *)hdl;
448 	oidgroup_t	*grp;
449 	uchar_t	*val;
450 	int	ret;
451 	int	err = 0;
452 
453 	if (smd == NULL || prefix == NULL || bitstrp == NULL || nbytes == NULL)
454 		return (-1);
455 
456 	*bitstrp = NULL;
457 	/*
458 	 * Check if this item is cacheable or not. If not, call
459 	 * fetch_single_* to get it directly from the agent
460 	 */
461 	if ((grp = locate_oid_group(smd, prefix)) == NULL) {
462 		ret = fetch_single_bitstr(smd, prefix, row, bitstrp,
463 		    nbytes, &err);
464 
465 		if (snmp_syserr)
466 			*snmp_syserr = err;
467 
468 		return (ret);
469 	}
470 
471 	/*
472 	 * See if it's in the cache already
473 	 */
474 	if (lookup_bitstr(prefix, row, &val, nbytes, grp->is_volatile) == 0) {
475 		if ((*bitstrp = (uchar_t *)calloc(*nbytes, 1)) == NULL)
476 			return (-1);
477 		(void) memcpy(*bitstrp, (const void *)val, *nbytes);
478 		return (0);
479 	}
480 
481 	/*
482 	 * Fetch it from the agent and populate cache
483 	 */
484 	fetch_bulk(smd, grp->oidstrs, grp->n_oids, row, grp->is_volatile, &err);
485 	if (snmp_syserr)
486 		*snmp_syserr = err;
487 
488 	/*
489 	 * Retry lookup
490 	 */
491 	if (lookup_bitstr(prefix, row, &val, nbytes, grp->is_volatile) < 0)
492 		return (-1);
493 
494 	if ((*bitstrp = (uchar_t *)calloc(*nbytes, 1)) == NULL)
495 		return (-1);
496 	(void) memcpy(*bitstrp, (const void *)val, *nbytes);
497 
498 	return (0);
499 }
500 
501 /*
502  * snmp_get_nextrow() is similar in operation to SNMP_GETNEXT, but
503  * only just. In particular, this is only expected to return the next
504  * valid row number for the same object, not its value. Since we don't
505  * have any other means, we use this to determine the number of rows
506  * in the table (and the valid ones). This function returns 0 on success
507  * and -1 on failure.
508  */
509 int
510 snmp_get_nextrow(picl_snmphdl_t hdl, char *prefix, int row, int *nextrow,
511     int *snmp_syserr)
512 {
513 	struct picl_snmphdl *smd = (struct picl_snmphdl *)hdl;
514 	snmp_pdu_t *reply_pdu;
515 	pdu_varlist_t *vp;
516 	char	*nxt_oidstr;
517 	int	err = 0;
518 
519 	if (smd == NULL || prefix == NULL || nextrow == NULL) {
520 		if (snmp_syserr)
521 			*snmp_syserr = EINVAL;
522 		return (-1);
523 	}
524 
525 	/*
526 	 * The get_nextrow results should *never* go into any cache,
527 	 * since these relationships are dynamically discovered each time.
528 	 */
529 	if ((reply_pdu = fetch_next(smd, prefix, row, &err)) == NULL) {
530 		if (snmp_syserr)
531 			*snmp_syserr = err;
532 		return (-1);
533 	}
534 
535 	/*
536 	 * We are not concerned about the "value" of the lexicographically
537 	 * next object; we only care about the name of that object and
538 	 * its row number (and whether such an object exists or not).
539 	 */
540 	vp = reply_pdu->vars;
541 
542 	/*
543 	 * This indicates that we're at the end of the MIB view.
544 	 */
545 	if (vp == NULL || vp->name == NULL || vp->type == SNMP_NOSUCHOBJECT ||
546 	    vp->type == SNMP_NOSUCHINSTANCE || vp->type == SNMP_ENDOFMIBVIEW) {
547 		snmp_free_pdu(reply_pdu);
548 		if (snmp_syserr)
549 			*snmp_syserr = ENOSPC;
550 		return (-1);
551 	}
552 
553 	/*
554 	 * need to be able to convert the OID
555 	 */
556 	if ((nxt_oidstr = oid_to_oidstr(vp->name, vp->name_len - 1)) == NULL) {
557 		snmp_free_pdu(reply_pdu);
558 		if (snmp_syserr)
559 			*snmp_syserr = ENOMEM;
560 		return (-1);
561 	}
562 
563 	/*
564 	 * We're on to the next table.
565 	 */
566 	if (strcmp(nxt_oidstr, prefix) != 0) {
567 		free(nxt_oidstr);
568 		snmp_free_pdu(reply_pdu);
569 		if (snmp_syserr)
570 			*snmp_syserr = ENOENT;
571 		return (-1);
572 	}
573 
574 	/*
575 	 * Ok, so we've got an oid that's simply the next valid row of the
576 	 * passed on object, return this row number.
577 	 */
578 	*nextrow = (vp->name)[vp->name_len-1];
579 
580 	free(nxt_oidstr);
581 	snmp_free_pdu(reply_pdu);
582 
583 	return (0);
584 }
585 
586 /*
587  * Request ids for snmp messages to the agent are sequenced here.
588  */
589 int
590 snmp_get_reqid(void)
591 {
592 	int	ret;
593 
594 	(void) mutex_lock(&snmp_reqid_lock);
595 
596 	ret = snmp_reqid++;
597 
598 	(void) mutex_unlock(&snmp_reqid_lock);
599 
600 	return (ret);
601 }
602 
603 static int
604 lookup_int(char *prefix, int row, int *valp, int is_vol)
605 {
606 	int32_t	*val_arr;
607 	uint_t	nelem;
608 	int	now;
609 	int	elapsed;
610 
611 	(void) mutex_lock(&mibcache_lock);
612 
613 	if (row >= n_mibcache_rows) {
614 		(void) mutex_unlock(&mibcache_lock);
615 		return (-1);
616 	}
617 
618 	if (mibcache[row] == NULL) {
619 		(void) mutex_unlock(&mibcache_lock);
620 		return (-1);
621 	}
622 
623 	/*
624 	 * If this is a volatile property, we should be searching
625 	 * for an integer-timestamp pair
626 	 */
627 	if (is_vol) {
628 		if (nvlist_lookup_int32_array(mibcache[row], prefix,
629 		    &val_arr, &nelem) != 0) {
630 			(void) mutex_unlock(&mibcache_lock);
631 			return (-1);
632 		}
633 		if (nelem != 2 || val_arr[1] < 0) {
634 			(void) mutex_unlock(&mibcache_lock);
635 			return (-1);
636 		}
637 		now = GET_SCALED_HRTIME();
638 		elapsed = now - val_arr[1];
639 		if (elapsed < 0 || elapsed > MAX_INCACHE_TIME) {
640 			(void) mutex_unlock(&mibcache_lock);
641 			return (-1);
642 		}
643 
644 		*valp = (int)val_arr[0];
645 	} else {
646 		if (nvlist_lookup_int32(mibcache[row], prefix, valp) != 0) {
647 			(void) mutex_unlock(&mibcache_lock);
648 			return (-1);
649 		}
650 	}
651 
652 	(void) mutex_unlock(&mibcache_lock);
653 
654 	return (0);
655 }
656 
657 static int
658 lookup_str(char *prefix, int row, char **valp, int is_vol)
659 {
660 	char	**val_arr;
661 	uint_t	nelem;
662 	int	now;
663 	int	elapsed;
664 
665 	(void) mutex_lock(&mibcache_lock);
666 
667 	if (row >= n_mibcache_rows) {
668 		(void) mutex_unlock(&mibcache_lock);
669 		return (-1);
670 	}
671 
672 	if (mibcache[row] == NULL) {
673 		(void) mutex_unlock(&mibcache_lock);
674 		return (-1);
675 	}
676 
677 	/*
678 	 * If this is a volatile property, we should be searching
679 	 * for a string-timestamp pair
680 	 */
681 	if (is_vol) {
682 		if (nvlist_lookup_string_array(mibcache[row], prefix,
683 		    &val_arr, &nelem) != 0) {
684 			(void) mutex_unlock(&mibcache_lock);
685 			return (-1);
686 		}
687 		if (nelem != 2 || atoi(val_arr[1]) <= 0) {
688 			(void) mutex_unlock(&mibcache_lock);
689 			return (-1);
690 		}
691 		now = GET_SCALED_HRTIME();
692 		elapsed = now - atoi(val_arr[1]);
693 		if (elapsed < 0 || elapsed > MAX_INCACHE_TIME) {
694 			(void) mutex_unlock(&mibcache_lock);
695 			return (-1);
696 		}
697 
698 		*valp = val_arr[0];
699 	} else {
700 		if (nvlist_lookup_string(mibcache[row], prefix, valp) != 0) {
701 			(void) mutex_unlock(&mibcache_lock);
702 			return (-1);
703 		}
704 	}
705 
706 	(void) mutex_unlock(&mibcache_lock);
707 
708 	return (0);
709 }
710 
711 static int
712 lookup_bitstr(char *prefix, int row, uchar_t **valp, uint_t *nelem, int is_vol)
713 {
714 	(void) mutex_lock(&mibcache_lock);
715 
716 	if (row >= n_mibcache_rows) {
717 		(void) mutex_unlock(&mibcache_lock);
718 		return (-1);
719 	}
720 
721 	if (mibcache[row] == NULL) {
722 		(void) mutex_unlock(&mibcache_lock);
723 		return (-1);
724 	}
725 
726 	/*
727 	 * We don't support volatile bit string values yet. The nvlist
728 	 * functions don't support bitstring arrays like they do charstring
729 	 * arrays, so we would need to do things in a convoluted way,
730 	 * probably by attaching the timestamp as part of the byte array
731 	 * itself. However, the need for volatile bitstrings isn't there
732 	 * yet, to justify the effort.
733 	 */
734 	if (is_vol) {
735 		(void) mutex_unlock(&mibcache_lock);
736 		return (-1);
737 	}
738 
739 	if (nvlist_lookup_byte_array(mibcache[row], prefix, valp, nelem) != 0) {
740 		(void) mutex_unlock(&mibcache_lock);
741 		return (-1);
742 	}
743 
744 	(void) mutex_unlock(&mibcache_lock);
745 
746 	return (0);
747 }
748 
749 static int
750 search_oid_in_group(char *prefix, char *oidstrs, int n_oids)
751 {
752 	char	*p;
753 	int	i;
754 
755 	p = oidstrs;
756 	for (i = 0; i < n_oids; i++) {
757 		if (strcmp(p, prefix) == 0)
758 			return (0);
759 
760 		p += strlen(p) + 1;
761 	}
762 
763 	return (-1);
764 }
765 
766 static oidgroup_t *
767 locate_oid_group(struct picl_snmphdl *smd, char *prefix)
768 {
769 	oidgroup_t	*grp;
770 
771 	if (smd == NULL)
772 		return (NULL);
773 
774 	if (smd->group == NULL)
775 		return (NULL);
776 
777 	for (grp = smd->group; grp; grp = grp->next) {
778 		if (search_oid_in_group(prefix, grp->oidstrs,
779 		    grp->n_oids) == 0) {
780 			return (grp);
781 		}
782 	}
783 
784 	return (NULL);
785 }
786 
787 static int
788 fetch_single_int(struct picl_snmphdl *smd, char *prefix, int row, int *ival,
789     int *snmp_syserr)
790 {
791 	snmp_pdu_t *reply_pdu;
792 	pdu_varlist_t *vp;
793 
794 	if ((reply_pdu = fetch_single(smd, prefix, row, snmp_syserr)) == NULL)
795 		return (-1);
796 
797 	/*
798 	 * Note that we don't make any distinction between unsigned int
799 	 * value and signed int value at this point, since we provide
800 	 * only snmp_get_int() at the higher level. While it is possible
801 	 * to provide an entirely separate interface such as snmp_get_uint(),
802 	 * that's quite unnecessary, because we don't do any interpretation
803 	 * of the received value. Besides, the sizes of int and uint are
804 	 * the same and the sizes of all pointers are the same (so val.iptr
805 	 * would be the same as val.uiptr in pdu_varlist_t). If/when we
806 	 * violate any of these assumptions, it will be time to add
807 	 * snmp_get_uint().
808 	 */
809 	vp = reply_pdu->vars;
810 	if (vp == NULL || vp->val.iptr == NULL) {
811 		snmp_free_pdu(reply_pdu);
812 		return (-1);
813 	}
814 
815 	*ival = *(vp->val.iptr);
816 
817 	snmp_free_pdu(reply_pdu);
818 
819 	return (0);
820 }
821 
822 static int
823 fetch_single_str(struct picl_snmphdl *smd, char *prefix, int row, char **valp,
824     int *snmp_syserr)
825 {
826 	snmp_pdu_t *reply_pdu;
827 	pdu_varlist_t *vp;
828 
829 	if ((reply_pdu = fetch_single(smd, prefix, row, snmp_syserr)) == NULL)
830 		return (-1);
831 
832 	vp = reply_pdu->vars;
833 	if (vp == NULL || vp->val.str == NULL) {
834 		snmp_free_pdu(reply_pdu);
835 		return (-1);
836 	}
837 
838 	*valp = strdup((const char *)(vp->val.str));
839 
840 	snmp_free_pdu(reply_pdu);
841 
842 	return (0);
843 }
844 
845 static int
846 fetch_single_bitstr(struct picl_snmphdl *smd, char *prefix, int row,
847     uchar_t **valp, uint_t *nelem, int *snmp_syserr)
848 {
849 	snmp_pdu_t *reply_pdu;
850 	pdu_varlist_t *vp;
851 
852 	if ((reply_pdu = fetch_single(smd, prefix, row, snmp_syserr)) == NULL)
853 		return (-1);
854 
855 	vp = reply_pdu->vars;
856 	if (vp == NULL || vp->val.str == NULL) {
857 		snmp_free_pdu(reply_pdu);
858 		return (-1);
859 	}
860 
861 	if ((*valp = (uchar_t *)calloc(vp->val_len, 1)) == NULL) {
862 		snmp_free_pdu(reply_pdu);
863 		return (-1);
864 	}
865 
866 	*nelem = vp->val_len;
867 	(void) memcpy(*valp, (const void *)(vp->val.str),
868 	    (size_t)(vp->val_len));
869 
870 	snmp_free_pdu(reply_pdu);
871 
872 	return (0);
873 }
874 
875 static snmp_pdu_t *
876 fetch_single(struct picl_snmphdl *smd, char *prefix, int row, int *snmp_syserr)
877 {
878 	snmp_pdu_t	*pdu, *reply_pdu;
879 
880 	LOGGET(TAG_CMD_REQUEST, prefix, row);
881 
882 	if ((pdu = snmp_create_pdu(SNMP_MSG_GET, 0, prefix, 1, row)) == NULL)
883 		return (NULL);
884 
885 	LOGPDU(TAG_REQUEST_PDU, pdu);
886 
887 	if (snmp_make_packet(pdu) < 0) {
888 		snmp_free_pdu(pdu);
889 		return (NULL);
890 	}
891 
892 	LOGPKT(TAG_REQUEST_PKT, pdu->req_pkt, pdu->req_pktsz);
893 
894 	if (snmp_send_request(smd, pdu, snmp_syserr) < 0) {
895 		snmp_free_pdu(pdu);
896 		return (NULL);
897 	}
898 
899 	if (snmp_recv_reply(smd, pdu, snmp_syserr) < 0) {
900 		snmp_free_pdu(pdu);
901 		return (NULL);
902 	}
903 
904 	LOGPKT(TAG_RESPONSE_PKT, pdu->reply_pkt, pdu->reply_pktsz);
905 
906 	reply_pdu = snmp_parse_reply(pdu->reqid, pdu->reply_pkt,
907 	    pdu->reply_pktsz);
908 
909 	LOGPDU(TAG_RESPONSE_PDU, reply_pdu);
910 
911 	snmp_free_pdu(pdu);
912 
913 	return (reply_pdu);
914 }
915 
916 static void
917 fetch_bulk(struct picl_snmphdl *smd, char *oidstrs, int n_oids,
918     int row, int is_vol, int *snmp_syserr)
919 {
920 	snmp_pdu_t	*pdu, *reply_pdu;
921 	int		max_reps;
922 
923 	LOGBULK(TAG_CMD_REQUEST, n_oids, oidstrs, row);
924 
925 	/*
926 	 * If we're fetching volatile properties using BULKGET, don't
927 	 * venture to get multiple rows (passing max_reps=0 will make
928 	 * snmp_create_pdu() fetch SNMP_DEF_MAX_REPETITIONS rows)
929 	 */
930 	max_reps = is_vol ? 1 : 0;
931 
932 	pdu = snmp_create_pdu(SNMP_MSG_GETBULK, max_reps, oidstrs, n_oids, row);
933 	if (pdu == NULL)
934 		return;
935 
936 	LOGPDU(TAG_REQUEST_PDU, pdu);
937 
938 	/*
939 	 * Make an ASN.1 encoded packet from the PDU information
940 	 */
941 	if (snmp_make_packet(pdu) < 0) {
942 		snmp_free_pdu(pdu);
943 		return;
944 	}
945 
946 	LOGPKT(TAG_REQUEST_PKT, pdu->req_pkt, pdu->req_pktsz);
947 
948 	/*
949 	 * Send the request packet to the agent
950 	 */
951 	if (snmp_send_request(smd, pdu, snmp_syserr) < 0) {
952 		snmp_free_pdu(pdu);
953 		return;
954 	}
955 
956 	/*
957 	 * Receive response from the agent into the reply packet buffer
958 	 * in the request PDU
959 	 */
960 	if (snmp_recv_reply(smd, pdu, snmp_syserr) < 0) {
961 		snmp_free_pdu(pdu);
962 		return;
963 	}
964 
965 	LOGPKT(TAG_RESPONSE_PKT, pdu->reply_pkt, pdu->reply_pktsz);
966 
967 	/*
968 	 * Parse the reply, validate the response and create a
969 	 * reply-PDU out of the information. Populate the mibcache
970 	 * with the received values.
971 	 */
972 	reply_pdu = snmp_parse_reply(pdu->reqid, pdu->reply_pkt,
973 	    pdu->reply_pktsz);
974 	if (reply_pdu) {
975 		LOGPDU(TAG_RESPONSE_PDU, reply_pdu);
976 
977 		if (reply_pdu->errstat == SNMP_ERR_NOERROR) {
978 			if (is_vol) {
979 				/* Add a job to the cache refresh work queue */
980 				(void) refreshq_add_job(smd, oidstrs, n_oids,
981 				    row);
982 			}
983 
984 			mibcache_populate(reply_pdu, is_vol);
985 		}
986 
987 		snmp_free_pdu(reply_pdu);
988 	}
989 
990 	snmp_free_pdu(pdu);
991 }
992 
993 static snmp_pdu_t *
994 fetch_next(struct picl_snmphdl *smd, char *prefix, int row, int *snmp_syserr)
995 {
996 	snmp_pdu_t	*pdu, *reply_pdu;
997 
998 	LOGNEXT(TAG_CMD_REQUEST, prefix, row);
999 
1000 	pdu = snmp_create_pdu(SNMP_MSG_GETNEXT, 0, prefix, 1, row);
1001 	if (pdu == NULL)
1002 		return (NULL);
1003 
1004 	LOGPDU(TAG_REQUEST_PDU, pdu);
1005 
1006 	if (snmp_make_packet(pdu) < 0) {
1007 		snmp_free_pdu(pdu);
1008 		return (NULL);
1009 	}
1010 
1011 	LOGPKT(TAG_REQUEST_PKT, pdu->req_pkt, pdu->req_pktsz);
1012 
1013 	if (snmp_send_request(smd, pdu, snmp_syserr) < 0) {
1014 		snmp_free_pdu(pdu);
1015 		return (NULL);
1016 	}
1017 
1018 	if (snmp_recv_reply(smd, pdu, snmp_syserr) < 0) {
1019 		snmp_free_pdu(pdu);
1020 		return (NULL);
1021 	}
1022 
1023 	LOGPKT(TAG_RESPONSE_PKT, pdu->reply_pkt, pdu->reply_pktsz);
1024 
1025 	reply_pdu = snmp_parse_reply(pdu->reqid, pdu->reply_pkt,
1026 	    pdu->reply_pktsz);
1027 
1028 	LOGPDU(TAG_RESPONSE_PDU, reply_pdu);
1029 
1030 	snmp_free_pdu(pdu);
1031 
1032 	return (reply_pdu);
1033 }
1034 
1035 static int
1036 snmp_send_request(struct picl_snmphdl *smd, snmp_pdu_t *pdu, int *snmp_syserr)
1037 {
1038 	extern int	errno;
1039 #ifdef USE_SOCKETS
1040 	int		ret;
1041 #endif
1042 
1043 	if (smd->fd < 0)
1044 		return (-1);
1045 
1046 	if (pdu == NULL || pdu->req_pkt == NULL)
1047 		return (-1);
1048 
1049 #ifdef USE_SOCKETS
1050 	ret = -1;
1051 	while (ret < 0) {
1052 		LOGIO(TAG_SENDTO, smd->fd, pdu->req_pkt, pdu->req_pktsz);
1053 
1054 		ret = sendto(smd->fd, pdu->req_pkt, pdu->req_pktsz, 0,
1055 		    (struct sockaddr *)&smd->agent_addr,
1056 		    sizeof (struct sockaddr));
1057 		if (ret < 0 && errno != EINTR) {
1058 			return (-1);
1059 		}
1060 	}
1061 #else
1062 	LOGIO(TAG_WRITE, smd->fd, pdu->req_pkt, pdu->req_pktsz);
1063 
1064 	if (write(smd->fd, pdu->req_pkt, pdu->req_pktsz) < 0) {
1065 		if (snmp_syserr)
1066 			*snmp_syserr = errno;
1067 		return (-1);
1068 	}
1069 #endif
1070 
1071 #ifdef SNMP_DEBUG
1072 	snmp_nsends++;
1073 	snmp_sentbytes += pdu->req_pktsz;
1074 #endif
1075 
1076 	return (0);
1077 }
1078 
1079 static int
1080 snmp_recv_reply(struct picl_snmphdl *smd, snmp_pdu_t *pdu, int *snmp_syserr)
1081 {
1082 	struct dssnmp_info	snmp_info;
1083 	size_t	pktsz;
1084 	uchar_t	*pkt;
1085 	extern int errno;
1086 #ifdef USE_SOCKETS
1087 	struct sockaddr_in 	from;
1088 	int	fromlen;
1089 	ssize_t	msgsz;
1090 #endif
1091 
1092 	if (smd->fd < 0 || pdu == NULL)
1093 		return (-1);
1094 
1095 #ifdef USE_SOCKETS
1096 	if ((pkt = (uchar_t *)calloc(1, SNMP_MAX_RECV_PKTSZ)) == NULL)
1097 		return (-1);
1098 
1099 	fromlen = sizeof (struct sockaddr_in);
1100 
1101 	LOGIO(TAG_RECVFROM, smd->fd, pkt, SNMP_MAX_RECV_PKTSZ);
1102 
1103 	msgsz = recvfrom(smd->fd, pkt, SNMP_MAX_RECV_PKTSZ, 0,
1104 	    (struct sockaddr *)&from, &fromlen);
1105 	if (msgsz  < 0 || msgsz >= SNMP_MAX_RECV_PKTSZ) {
1106 		free(pkt);
1107 		return (-1);
1108 	}
1109 
1110 	pktsz = (size_t)msgsz;
1111 #else
1112 	LOGIO(TAG_IOCTL, smd->fd, DSSNMP_GETINFO, &snmp_info);
1113 
1114 	/*
1115 	 * The ioctl will block until we have snmp data available
1116 	 */
1117 	if (ioctl(smd->fd, DSSNMP_GETINFO, &snmp_info) < 0) {
1118 		if (snmp_syserr)
1119 			*snmp_syserr = errno;
1120 		return (-1);
1121 	}
1122 
1123 	pktsz = snmp_info.size;
1124 	if ((pkt = (uchar_t *)calloc(1, pktsz)) == NULL)
1125 		return (-1);
1126 
1127 	LOGIO(TAG_READ, smd->fd, pkt, pktsz);
1128 
1129 	if (read(smd->fd, pkt, pktsz) < 0) {
1130 		free(pkt);
1131 		if (snmp_syserr)
1132 			*snmp_syserr = errno;
1133 		return (-1);
1134 	}
1135 #endif
1136 
1137 	pdu->reply_pkt = pkt;
1138 	pdu->reply_pktsz = pktsz;
1139 
1140 #ifdef SNMP_DEBUG
1141 	snmp_nrecvs++;
1142 	snmp_rcvdbytes += pktsz;
1143 #endif
1144 
1145 	return (0);
1146 }
1147 
1148 static int
1149 mibcache_realloc(int hint)
1150 {
1151 	uint_t		count = (uint_t)hint;
1152 	nvlist_t	**p;
1153 
1154 	if (hint < 0)
1155 		return (-1);
1156 
1157 	(void) mutex_lock(&mibcache_lock);
1158 
1159 	if (hint < n_mibcache_rows) {
1160 		(void) mutex_unlock(&mibcache_lock);
1161 		return (0);
1162 	}
1163 
1164 	count =  ((count >> MIBCACHE_BLK_SHIFT) + 1) << MIBCACHE_BLK_SHIFT;
1165 
1166 	p = (nvlist_t **)calloc(count, sizeof (nvlist_t *));
1167 	if (p == NULL) {
1168 		(void) mutex_unlock(&mibcache_lock);
1169 		return (-1);
1170 	}
1171 
1172 	if (mibcache) {
1173 		(void) memcpy((void *) p, (void *) mibcache,
1174 		    n_mibcache_rows * sizeof (nvlist_t *));
1175 		free((void *) mibcache);
1176 	}
1177 
1178 	mibcache = p;
1179 	n_mibcache_rows = count;
1180 
1181 	(void) mutex_unlock(&mibcache_lock);
1182 
1183 	return (0);
1184 }
1185 
1186 
1187 /*
1188  * Scan each variable in the returned PDU's bindings and populate
1189  * the cache appropriately
1190  */
1191 static void
1192 mibcache_populate(snmp_pdu_t *pdu, int is_vol)
1193 {
1194 	pdu_varlist_t	*vp;
1195 	int		row, ret;
1196 	char		*oidstr;
1197 	int		tod;	/* in secs */
1198 	char		tod_str[MAX_INT_LEN];
1199 	int		ival_arr[2];
1200 	char		*sval_arr[2];
1201 
1202 	/*
1203 	 * If we're populating volatile properties, we also store a
1204 	 * timestamp with each property value. When we lookup, we check the
1205 	 * current time against this timestamp to determine if we need to
1206 	 * refetch the value or not (refetch if it has been in for far too
1207 	 * long).
1208 	 */
1209 
1210 	if (is_vol) {
1211 		tod = GET_SCALED_HRTIME();
1212 
1213 		tod_str[0] = 0;
1214 		(void) snprintf(tod_str, MAX_INT_LEN, "%d", tod);
1215 
1216 		ival_arr[1] = tod;
1217 		sval_arr[1] = (char *)tod_str;
1218 	}
1219 
1220 	for (vp = pdu->vars; vp; vp = vp->nextvar) {
1221 		if (vp->type != ASN_INTEGER && vp->type != ASN_OCTET_STR &&
1222 		    vp->type != ASN_BIT_STR) {
1223 			continue;
1224 		}
1225 
1226 		if (vp->name == NULL || vp->val.str == NULL)
1227 			continue;
1228 
1229 		row = (vp->name)[vp->name_len-1];
1230 
1231 		(void) mutex_lock(&mibcache_lock);
1232 
1233 		if (row >= n_mibcache_rows) {
1234 			(void) mutex_unlock(&mibcache_lock);
1235 			if (mibcache_realloc(row) < 0)
1236 				continue;
1237 			(void) mutex_lock(&mibcache_lock);
1238 		}
1239 		ret = 0;
1240 		if (mibcache[row] == NULL)
1241 			ret = nvlist_alloc(&mibcache[row], NV_UNIQUE_NAME, 0);
1242 
1243 		(void) mutex_unlock(&mibcache_lock);
1244 
1245 		if (ret != 0)
1246 			continue;
1247 
1248 		/*
1249 		 * Convert the standard OID form into an oid string that
1250 		 * we can use as the key to lookup. Since we only search
1251 		 * by the prefix (mibcache is really an array of nvlist_t
1252 		 * pointers), ignore the leaf subid.
1253 		 */
1254 		oidstr = oid_to_oidstr(vp->name, vp->name_len - 1);
1255 		if (oidstr == NULL)
1256 			continue;
1257 
1258 		(void) mutex_lock(&mibcache_lock);
1259 
1260 		if (vp->type == ASN_INTEGER) {
1261 			if (is_vol) {
1262 				ival_arr[0] = *(vp->val.iptr);
1263 				(void) nvlist_add_int32_array(mibcache[row],
1264 				    oidstr, ival_arr, 2);
1265 			} else {
1266 				(void) nvlist_add_int32(mibcache[row],
1267 				    oidstr, *(vp->val.iptr));
1268 			}
1269 
1270 		} else if (vp->type == ASN_OCTET_STR) {
1271 			if (is_vol) {
1272 				sval_arr[0] = (char *)vp->val.str;
1273 				(void) nvlist_add_string_array(mibcache[row],
1274 				    oidstr, sval_arr, 2);
1275 			} else {
1276 				(void) nvlist_add_string(mibcache[row],
1277 				    oidstr, (const char *)(vp->val.str));
1278 			}
1279 		} else if (vp->type == ASN_BIT_STR) {
1280 			/*
1281 			 * We don't support yet bit string objects that are
1282 			 * volatile values.
1283 			 */
1284 			if (!is_vol) {
1285 				(void) nvlist_add_byte_array(mibcache[row],
1286 				    oidstr, (uchar_t *)(vp->val.str),
1287 				    (uint_t)vp->val_len);
1288 			}
1289 		}
1290 		(void) mutex_unlock(&mibcache_lock);
1291 
1292 		free(oidstr);
1293 	}
1294 }
1295 
1296 static char *
1297 oid_to_oidstr(oid *objid, size_t n_subids)
1298 {
1299 	char	*oidstr;
1300 	char	subid_str[MAX_INT_LEN];
1301 	int	i, isize;
1302 	size_t	oidstr_sz;
1303 
1304 	/*
1305 	 * ugly, but for now this will have to do.
1306 	 */
1307 	oidstr_sz = sizeof (subid_str) * n_subids;
1308 	oidstr = calloc(1, oidstr_sz);
1309 
1310 	for (i = 0; i < n_subids; i++) {
1311 		(void) memset(subid_str, 0, sizeof (subid_str));
1312 		isize = snprintf(subid_str, sizeof (subid_str), "%d",
1313 		    objid[i]);
1314 		if (isize >= sizeof (subid_str))
1315 			return (NULL);
1316 
1317 		(void) strlcat(oidstr, subid_str, oidstr_sz);
1318 		if (i < (n_subids - 1))
1319 			(void) strlcat(oidstr, ".", oidstr_sz);
1320 	}
1321 
1322 	return (oidstr);
1323 }
1324 
1325 /*
1326  * Expand the refreshq to hold more cache refresh jobs.  Caller must already
1327  * hold refreshq_lock mutex.  Every expansion of the refreshq will add
1328  * REFRESH_BLK_SZ job slots, rather than expanding by one slot every time more
1329  * space is needed.
1330  */
1331 static int
1332 refreshq_realloc(int hint)
1333 {
1334 	uint_t		count = (uint_t)hint;
1335 	refreshq_job_t	*p;
1336 
1337 	if (hint < 0)
1338 		return (-1);
1339 
1340 	if (hint < n_refreshq_slots) {
1341 		return (0);
1342 	}
1343 
1344 	/* Round count up to next multiple of REFRESHQ_BLK_SHIFT */
1345 	count =  ((count >> REFRESHQ_BLK_SHIFT) + 1) << REFRESHQ_BLK_SHIFT;
1346 
1347 	p = (refreshq_job_t *)calloc(count, sizeof (refreshq_job_t));
1348 	if (p == NULL) {
1349 		return (-1);
1350 	}
1351 
1352 	if (refreshq) {
1353 		if (n_refreshq_jobs == 0) {
1354 			/* Simple case, nothing to copy */
1355 			refreshq_next_job = 0;
1356 			refreshq_next_slot = 0;
1357 		} else if (refreshq_next_slot > refreshq_next_job) {
1358 			/* Simple case, single copy preserves everything */
1359 			(void) memcpy((void *) p,
1360 			    (void *) &(refreshq[refreshq_next_job]),
1361 			    n_refreshq_jobs * sizeof (refreshq_job_t));
1362 		} else {
1363 			/*
1364 			 * Complex case.  The jobs in the refresh queue wrap
1365 			 * around the end of the array in which they are stored.
1366 			 * To preserve chronological order in the new allocated
1367 			 * array, we need to copy the jobs at the end of the old
1368 			 * array to the beginning of the new one and place the
1369 			 * jobs from the beginning of the old array after them.
1370 			 */
1371 			uint_t tail_jobs, head_jobs;
1372 
1373 			tail_jobs = n_refreshq_slots - refreshq_next_job;
1374 			head_jobs = n_refreshq_jobs - tail_jobs;
1375 
1376 			/* Copy the jobs from the end of the old array */
1377 			(void) memcpy((void *) p,
1378 			    (void *) &(refreshq[refreshq_next_job]),
1379 			    tail_jobs * sizeof (refreshq_job_t));
1380 
1381 			/* Copy the jobs from the beginning of the old array */
1382 			(void) memcpy((void *) &(p[tail_jobs]),
1383 			    (void *) &(refreshq[0]),
1384 			    head_jobs * sizeof (refreshq_job_t));
1385 
1386 			/* update the job and slot indices to match */
1387 			refreshq_next_job = 0;
1388 			refreshq_next_slot = n_refreshq_jobs;
1389 		}
1390 		free((void *) refreshq);
1391 	} else {
1392 		/* First initialization */
1393 		refreshq_next_job = 0;
1394 		refreshq_next_slot = 0;
1395 		n_refreshq_jobs = 0;
1396 	}
1397 
1398 	refreshq = p;
1399 	n_refreshq_slots = count;
1400 
1401 	return (0);
1402 }
1403 
1404 /*
1405  * Add a new job to the refreshq.  If there aren't any open slots, attempt to
1406  * expand the queue first.  Return -1 if unable to add the job to the work
1407  * queue, or 0 if the job was added OR if an existing job with the same
1408  * parameters is already pending.
1409  */
1410 static int
1411 refreshq_add_job(struct picl_snmphdl *smd, char *oidstrs, int n_oids, int row)
1412 {
1413 	int	i;
1414 	int	job;
1415 
1416 	(void) mutex_lock(&refreshq_lock);
1417 
1418 	/*
1419 	 * Can't do anything without a queue.  Either the client never
1420 	 * initialized the refresh queue or the initial memory allocation
1421 	 * failed.
1422 	 */
1423 	if (refreshq == NULL) {
1424 		(void) mutex_unlock(&refreshq_lock);
1425 		return (-1);
1426 	}
1427 
1428 	/*
1429 	 * If there is already a job pending with the same parameters as the job
1430 	 * we have been asked to add, we apparently let an entry expire and it
1431 	 * is now being reloaded.  Rather than add another job for the same
1432 	 * entry, we skip adding the new job and let the existing job address
1433 	 * it.
1434 	 */
1435 	for (i = 0, job = refreshq_next_job; i < n_refreshq_jobs; i++,
1436 	    job = (job + 1) % n_refreshq_slots) {
1437 		if ((refreshq[job].row == row) &&
1438 		    (refreshq[job].n_oids == n_oids) &&
1439 		    (refreshq[job].oidstrs == oidstrs)) {
1440 			(void) mutex_unlock(&refreshq_lock);
1441 			return (0);
1442 		}
1443 	}
1444 
1445 
1446 	/*
1447 	 * If the queue is full, we need to expand it
1448 	 */
1449 	if (n_refreshq_jobs == n_refreshq_slots) {
1450 		if (refreshq_realloc(n_refreshq_slots + 1) < 0) {
1451 			/*
1452 			 * Can't expand the job queue, so we drop this job on
1453 			 * the floor.  No data is lost... we just allow some
1454 			 * data in the mibcache to expire.
1455 			 */
1456 			(void) mutex_unlock(&refreshq_lock);
1457 			return (-1);
1458 		}
1459 	}
1460 
1461 	/*
1462 	 * There is room in the queue, so add the new job.  We are actually
1463 	 * taking a timestamp for this job that is slightly earlier than when
1464 	 * the mibcache entry will be updated, but since we're trying to update
1465 	 * the mibcache entry before it expires anyway, the earlier timestamp
1466 	 * here is acceptable.
1467 	 */
1468 	refreshq[refreshq_next_slot].smd = smd;
1469 	refreshq[refreshq_next_slot].oidstrs = oidstrs;
1470 	refreshq[refreshq_next_slot].n_oids = n_oids;
1471 	refreshq[refreshq_next_slot].row = row;
1472 	refreshq[refreshq_next_slot].last_fetch_time = GET_SCALED_HRTIME();
1473 
1474 	/*
1475 	 * Update queue management variables
1476 	 */
1477 	n_refreshq_jobs += 1;
1478 	refreshq_next_slot = (refreshq_next_slot + 1) % n_refreshq_slots;
1479 
1480 	(void) mutex_unlock(&refreshq_lock);
1481 
1482 	return (0);
1483 }
1484 
1485 /*
1486  * Almost all of the refresh code remains dormant unless specifically
1487  * initialized by a client (the exception being that fetch_bulk() will still
1488  * call refreshq_add_job(), but the latter will return without doing anything).
1489  */
1490 int
1491 snmp_refresh_init(void)
1492 {
1493 	int ret;
1494 
1495 	(void) mutex_lock(&refreshq_lock);
1496 
1497 	ret = refreshq_realloc(0);
1498 
1499 	(void) mutex_unlock(&refreshq_lock);
1500 
1501 	return (ret);
1502 }
1503 
1504 /*
1505  * If the client is going away, we don't want to keep doing refresh work, so
1506  * clean everything up.
1507  */
1508 void
1509 snmp_refresh_fini(void)
1510 {
1511 	(void) mutex_lock(&refreshq_lock);
1512 
1513 	n_refreshq_jobs = 0;
1514 	n_refreshq_slots = 0;
1515 	refreshq_next_job = 0;
1516 	refreshq_next_slot = 0;
1517 	free(refreshq);
1518 	refreshq = NULL;
1519 
1520 	(void) mutex_unlock(&refreshq_lock);
1521 }
1522 
1523 /*
1524  * Return the number of seconds remaining before the mibcache entry associated
1525  * with the next job in the queue will expire.  Note that this requires
1526  * reversing the scaling normally done on hrtime values.  (The need for scaling
1527  * is purely internal, and should be hidden from clients.)  If there are no jobs
1528  * in the queue, return -1.  If the next job has already expired, return 0.
1529  */
1530 int
1531 snmp_refresh_get_next_expiration(void)
1532 {
1533 	int ret;
1534 	int elapsed;
1535 
1536 	(void) mutex_lock(&refreshq_lock);
1537 
1538 	if (n_refreshq_jobs == 0) {
1539 		ret = -1;
1540 	} else {
1541 		elapsed = GET_SCALED_HRTIME() -
1542 		    refreshq[refreshq_next_job].last_fetch_time;
1543 
1544 		if (elapsed >= MAX_INCACHE_TIME) {
1545 			ret = 0;
1546 		} else {
1547 			ret = (MAX_INCACHE_TIME - elapsed) * HRTIME_SCALE;
1548 		}
1549 	}
1550 
1551 	(void) mutex_unlock(&refreshq_lock);
1552 
1553 	return (ret);
1554 }
1555 
1556 /*
1557  * Given the number of seconds the client wants to spend on each cyle of
1558  * processing jobs and then sleeping, return a suggestion for the number of jobs
1559  * the client should process, calculated by dividing the client's cycle duration
1560  * by MAX_INCACHE_TIME and multiplying the result by the total number of jobs in
1561  * the queue.  (Note that the actual implementation of that calculation is done
1562  * in a different order to avoid losing fractional values during integer
1563  * arithmetic.)
1564  */
1565 int
1566 snmp_refresh_get_cycle_hint(int secs)
1567 {
1568 	int	jobs;
1569 
1570 	(void) mutex_lock(&refreshq_lock);
1571 
1572 	/*
1573 	 * First, we need to scale the client's cycle time to get it into the
1574 	 * same units we use internally (i.e. tens of seconds).  We round up, as
1575 	 * it makes more sense for the client to process extra jobs than
1576 	 * insufficient jobs.  If the client's desired cycle time is greater
1577 	 * than MAX_INCACHE_TIME, we just return the current total number of
1578 	 * jobs.
1579 	 */
1580 	secs = (secs + HRTIME_SCALE - 1) / HRTIME_SCALE;
1581 
1582 	jobs = (n_refreshq_jobs * secs) / MAX_INCACHE_TIME;
1583 	if (jobs > n_refreshq_jobs) {
1584 		jobs = n_refreshq_jobs;
1585 	}
1586 
1587 	(void) mutex_unlock(&refreshq_lock);
1588 
1589 	return (jobs);
1590 }
1591 
1592 /*
1593  * Process the next job on the refresh queue by invoking fetch_bulk() with the
1594  * recorded parameters.  Return -1 if no job was processed (e.g. because there
1595  * aren't any available), or 0 if a job was processed.  We don't actually care
1596  * if fetch_bulk() fails, since we're just working on cache entry refreshing and
1597  * the worst case result of failing here is a longer delay getting that data the
1598  * next time it is requested.
1599  */
1600 int
1601 snmp_refresh_process_job(void)
1602 {
1603 	struct picl_snmphdl	*smd;
1604 	char			*oidstrs;
1605 	int			n_oids;
1606 	int			row;
1607 	int			err;
1608 
1609 	(void) mutex_lock(&refreshq_lock);
1610 
1611 	if (n_refreshq_jobs == 0) {
1612 		(void) mutex_unlock(&refreshq_lock);
1613 
1614 		return (-1);
1615 	}
1616 
1617 	smd = refreshq[refreshq_next_job].smd;
1618 	oidstrs = refreshq[refreshq_next_job].oidstrs;
1619 	n_oids = refreshq[refreshq_next_job].n_oids;
1620 	row = refreshq[refreshq_next_job].row;
1621 
1622 	refreshq_next_job = (refreshq_next_job + 1) % n_refreshq_slots;
1623 	n_refreshq_jobs--;
1624 
1625 	(void) mutex_unlock(&refreshq_lock);
1626 
1627 
1628 	/*
1629 	 * fetch_bulk() is going to come right back into the refresh code to add
1630 	 * a new job for the entry we just loaded, which means we have to make
1631 	 * the call without holding the refreshq_lock mutex.
1632 	 */
1633 	fetch_bulk(smd, oidstrs, n_oids, row, 1, &err);
1634 
1635 	return (0);
1636 }
1637