xref: /illumos-gate/usr/src/cmd/sa/sadc.c (revision 05ede3db5e3b7d540afbccdc872c735770e82ef3)
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  * Copyright (c) 1989, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
26 /*	  All Rights Reserved  	*/
27 
28 
29 /*
30  * sadc.c writes system activity binary data to a file or stdout.
31  *
32  * Usage: sadc [t n] [file]
33  *
34  * if t and n are not specified, it writes a dummy record to data file. This
35  * usage is particularly used at system booting.  If t and n are specified, it
36  * writes system data n times to file every t seconds.  In both cases, if file
37  * is not specified, it writes data to stdout.
38  */
39 
40 #include <sys/fcntl.h>
41 #include <sys/flock.h>
42 #include <sys/proc.h>
43 #include <sys/stat.h>
44 #include <sys/sysinfo.h>
45 #include <sys/time.h>
46 #include <sys/types.h>
47 #include <sys/var.h>
48 
49 #include <ctype.h>
50 #include <errno.h>
51 #include <fcntl.h>
52 #include <kstat.h>
53 #include <memory.h>
54 #include <nlist.h>
55 #include <signal.h>
56 #include <stdarg.h>
57 #include <stdio.h>
58 #include <stdlib.h>
59 #include <string.h>
60 #include <time.h>
61 #include <unistd.h>
62 #include <strings.h>
63 
64 #include "sa.h"
65 
66 #define	MAX(x1, x2)	((x1) >= (x2) ? (x1) : (x2))
67 
68 static	kstat_ctl_t	*kc;		/* libkstat cookie */
69 static	int	ncpus;
70 static	int	oncpus;
71 static	kstat_t	**cpu_stat_list = NULL;
72 static	kstat_t	**ocpu_stat_list = NULL;
73 static	int	ncaches;
74 static	kstat_t	**kmem_cache_list = NULL;
75 
76 static	kstat_t	*sysinfo_ksp, *vminfo_ksp, *var_ksp;
77 static	kstat_t *system_misc_ksp, *ufs_inode_ksp, *kmem_oversize_ksp;
78 static	kstat_t *file_cache_ksp;
79 static	kstat_named_t *ufs_inode_size_knp, *nproc_knp;
80 static	kstat_named_t *file_total_knp, *file_avail_knp;
81 static	kstat_named_t *oversize_alloc_knp, *oversize_fail_knp;
82 static	int slab_create_index, slab_destroy_index, slab_size_index;
83 static	int buf_size_index, buf_avail_index, alloc_fail_index;
84 
85 static	struct	iodevinfo zeroiodev = { NULL, NULL };
86 static	struct	iodevinfo *firstiodev = NULL;
87 static	struct	iodevinfo *lastiodev = NULL;
88 static	struct	iodevinfo *snip = NULL;
89 static	ulong_t	niodevs;
90 
91 static	void	all_stat_init(void);
92 static	int	all_stat_load(void);
93 static	void	fail(int, char *, ...);
94 static	void	safe_zalloc(void **, int, int);
95 static	kid_t	safe_kstat_read(kstat_ctl_t *, kstat_t *, void *);
96 static	kstat_t	*safe_kstat_lookup(kstat_ctl_t *, char *, int, char *);
97 static	void	*safe_kstat_data_lookup(kstat_t *, char *);
98 static	int	safe_kstat_data_index(kstat_t *, char *);
99 static	void	init_iodevs(void);
100 static	int	iodevinfo_load(void);
101 static	int	kstat_copy(const kstat_t *, kstat_t *);
102 static	void	diff_two_arrays(kstat_t ** const [], size_t, size_t,
103     kstat_t ** const []);
104 static	void	compute_cpu_stat_adj(void);
105 
106 static	char	*cmdname = "sadc";
107 
108 static	struct var var;
109 
110 static	struct sa d;
111 static	int64_t	cpu_stat_adj[CPU_STATES] = {0};
112 
113 static	long	ninode;
114 
115 int caught_cont = 0;
116 
117 /*
118  * Sleep until *wakeup + interval, keeping cadence where desired
119  *
120  * *wakeup -	The time we last wanted to wake up. Updated.
121  * interval -	We want to sleep until *wakeup + interval
122  * *caught_cont - Global set by signal handler if we got a SIGCONT
123  */
124 void
125 sleep_until(hrtime_t *wakeup, hrtime_t interval, int *caught_cont)
126 {
127 	hrtime_t now, pause, pause_left;
128 	struct timespec pause_tv;
129 	int status;
130 	now = gethrtime();
131 	pause = *wakeup + interval - now;
132 
133 	if (pause <= 0 || pause < (interval / 4))
134 		if (*caught_cont) {
135 			/* Reset our cadence (see comment below) */
136 			*wakeup = now + interval;
137 			pause = interval;
138 		} else {
139 			/*
140 			 * If we got here, then the time between the
141 			 * output we just did, and the scheduled time
142 			 * for the next output is < 1/4 of our requested
143 			 * interval AND the number of intervals has been
144 			 * requested AND we have never caught a SIGCONT
145 			 * (so we have never been suspended).  In this
146 			 * case, we'll try to stay to the desired
147 			 * cadence, and we will pause for 1/2 the normal
148 			 * interval this time.
149 			 */
150 			pause = interval / 2;
151 			*wakeup += interval;
152 		}
153 	else
154 		*wakeup += interval;
155 	if (pause < 1000)
156 		/* Near enough */
157 		return;
158 
159 	/* Now do the actual sleep */
160 	pause_left = pause;
161 	do {
162 		pause_tv.tv_sec = pause_left / NANOSEC;
163 		pause_tv.tv_nsec = pause_left % NANOSEC;
164 		status = nanosleep(&pause_tv, (struct timespec *)NULL);
165 		if (status < 0)
166 			if (errno == EINTR) {
167 				now = gethrtime();
168 				pause_left = *wakeup - now;
169 				if (pause_left < 1000)
170 					/* Near enough */
171 					return;
172 			} else {
173 				fail(1, "nanosleep failed");
174 			}
175 	} while (status != 0);
176 }
177 
178 /*
179  * Signal handler - so we can be aware of SIGCONT
180  */
181 void
182 cont_handler(int sig_number)
183 {
184 	/* Re-set the signal handler */
185 	(void) signal(sig_number, cont_handler);
186 	caught_cont = 1;
187 }
188 
189 int
190 main(int argc, char *argv[])
191 {
192 	int ct;
193 	unsigned ti;
194 	int fp;
195 	time_t min;
196 	struct stat buf;
197 	char *fname;
198 	struct iodevinfo *iodev;
199 	off_t flength;
200 	hrtime_t start_n;
201 	hrtime_t period_n;
202 
203 
204 	ct = argc >= 3? atoi(argv[2]): 0;
205 	min = time((time_t *)0);
206 	ti = argc >= 3? atoi(argv[1]): 0;
207 
208 	period_n = (hrtime_t)ti * NANOSEC;
209 
210 	if ((kc = kstat_open()) == NULL)
211 		fail(1, "kstat_open(): can't open /dev/kstat");
212 
213 	/* Set up handler for SIGCONT */
214 	if (signal(SIGCONT, cont_handler) == SIG_ERR)
215 		fail(1, "signal failed");
216 
217 	all_stat_init();
218 	init_iodevs();
219 
220 	if (argc == 3 || argc == 1) {
221 		/*
222 		 * no data file is specified, direct data to stdout.
223 		 */
224 		fp = 1;
225 	} else {
226 		struct flock lock;
227 
228 		fname = (argc == 2) ? argv[1] : argv[3];
229 		/*
230 		 * Open or Create a data file. If the file doesn't exist, then
231 		 * it will be created.
232 		 */
233 		if ((fp = open(fname, O_WRONLY | O_APPEND | O_CREAT, 0644))
234 		    == -1)
235 			fail(1, "can't open data file");
236 		/*
237 		 * Lock the entire data file to prevent data corruption
238 		 */
239 		lock.l_type = F_WRLCK;
240 		lock.l_whence = SEEK_SET;
241 		lock.l_start = 0;
242 		lock.l_len = 0;
243 		if (fcntl(fp, F_SETLK, &lock) == -1)
244 			fail(1, "can't lock data file");
245 		/*
246 		 * Get data file statistics for use in determining whether
247 		 * truncation required and where rollback recovery should
248 		 * be applied.
249 		 */
250 		if (fstat(fp, &buf) == -1)
251 			fail(1, "can't get data file information");
252 		/*
253 		 * If the data file was opened and is too old, truncate it
254 		 */
255 		if (min - buf.st_mtime > 86400)
256 			if (ftruncate(fp, 0) == -1)
257 				fail(1, "can't truncate data file");
258 		/*
259 		 * Remember filesize for rollback on error (bug #1223549)
260 		 */
261 		flength = buf.st_size;
262 	}
263 
264 	memset(&d, 0, sizeof (d));
265 
266 	/*
267 	 * If n == 0, write the additional dummy record.
268 	 */
269 	if (ct == 0) {
270 		d.valid = 0;
271 		d.ts = min;
272 		d.niodevs = niodevs;
273 
274 		if (write(fp, &d, sizeof (struct sa)) != sizeof (struct sa))
275 			ftruncate(fp, flength), fail(1, "write failed");
276 
277 		for (iodev = firstiodev; iodev; iodev = iodev->next) {
278 			if (write(fp, iodev, sizeof (struct iodevinfo)) !=
279 			    sizeof (struct iodevinfo))
280 				ftruncate(fp, flength), fail(1, "write failed");
281 		}
282 	}
283 
284 	start_n = gethrtime();
285 
286 	for (;;) {
287 		do {
288 			(void) kstat_chain_update(kc);
289 			all_stat_init();
290 			init_iodevs();
291 		} while (all_stat_load() || iodevinfo_load());
292 
293 		d.ts = time((time_t *)0);
294 		d.valid = 1;
295 		d.niodevs = niodevs;
296 
297 		if (write(fp, &d, sizeof (struct sa)) != sizeof (struct sa))
298 			ftruncate(fp, flength), fail(1, "write failed");
299 
300 		for (iodev = firstiodev; iodev; iodev = iodev->next) {
301 			if (write(fp, iodev, sizeof (struct iodevinfo)) !=
302 			    sizeof (struct iodevinfo))
303 				ftruncate(fp, flength), fail(1, "write failed");
304 		}
305 		if (--ct > 0) {
306 			sleep_until(&start_n, period_n, &caught_cont);
307 		} else {
308 			close(fp);
309 			return (0);
310 		}
311 	}
312 
313 	/*NOTREACHED*/
314 }
315 
316 /*
317  * Get various KIDs for subsequent all_stat_load operations.
318  */
319 
320 static void
321 all_stat_init(void)
322 {
323 	kstat_t *ksp;
324 
325 	/*
326 	 * Initialize global statistics
327 	 */
328 
329 	sysinfo_ksp	= safe_kstat_lookup(kc, "unix", 0, "sysinfo");
330 	vminfo_ksp	= safe_kstat_lookup(kc, "unix", 0, "vminfo");
331 	kmem_oversize_ksp = safe_kstat_lookup(kc, "vmem", -1, "kmem_oversize");
332 	var_ksp		= safe_kstat_lookup(kc, "unix", 0, "var");
333 	system_misc_ksp	= safe_kstat_lookup(kc, "unix", 0, "system_misc");
334 	file_cache_ksp	= safe_kstat_lookup(kc, "unix", 0, "file_cache");
335 	ufs_inode_ksp	= kstat_lookup(kc, "ufs", 0, "inode_cache");
336 
337 	safe_kstat_read(kc, system_misc_ksp, NULL);
338 	nproc_knp	= safe_kstat_data_lookup(system_misc_ksp, "nproc");
339 
340 	safe_kstat_read(kc, file_cache_ksp, NULL);
341 	file_avail_knp = safe_kstat_data_lookup(file_cache_ksp, "buf_avail");
342 	file_total_knp = safe_kstat_data_lookup(file_cache_ksp, "buf_total");
343 
344 	safe_kstat_read(kc, kmem_oversize_ksp, NULL);
345 	oversize_alloc_knp = safe_kstat_data_lookup(kmem_oversize_ksp,
346 	    "mem_total");
347 	oversize_fail_knp = safe_kstat_data_lookup(kmem_oversize_ksp, "fail");
348 
349 	if (ufs_inode_ksp != NULL) {
350 		safe_kstat_read(kc, ufs_inode_ksp, NULL);
351 		ufs_inode_size_knp = safe_kstat_data_lookup(ufs_inode_ksp,
352 		    "size");
353 		ninode = ((kstat_named_t *)
354 		    safe_kstat_data_lookup(ufs_inode_ksp,
355 		    "maxsize"))->value.l;
356 	}
357 
358 	/*
359 	 * Load constant values now -- no need to reread each time
360 	 */
361 
362 	safe_kstat_read(kc, var_ksp, (void *) &var);
363 
364 	/*
365 	 * Initialize per-CPU and per-kmem-cache statistics
366 	 */
367 
368 	ncpus = ncaches = 0;
369 	for (ksp = kc->kc_chain; ksp; ksp = ksp->ks_next) {
370 		if (strncmp(ksp->ks_name, "cpu_stat", 8) == 0)
371 			ncpus++;
372 		if (strcmp(ksp->ks_class, "kmem_cache") == 0)
373 			ncaches++;
374 	}
375 
376 	safe_zalloc((void **)&cpu_stat_list, ncpus * sizeof (kstat_t *), 1);
377 	safe_zalloc((void **)&kmem_cache_list, ncaches * sizeof (kstat_t *), 1);
378 
379 	ncpus = ncaches = 0;
380 	for (ksp = kc->kc_chain; ksp; ksp = ksp->ks_next) {
381 		if (strncmp(ksp->ks_name, "cpu_stat", 8) == 0 &&
382 		    kstat_read(kc, ksp, NULL) != -1)
383 			cpu_stat_list[ncpus++] = ksp;
384 		if (strcmp(ksp->ks_class, "kmem_cache") == 0 &&
385 		    kstat_read(kc, ksp, NULL) != -1)
386 			kmem_cache_list[ncaches++] = ksp;
387 	}
388 
389 	if (ncpus == 0)
390 		fail(1, "can't find any cpu statistics");
391 
392 	if (ncaches == 0)
393 		fail(1, "can't find any kmem_cache statistics");
394 
395 	ksp = kmem_cache_list[0];
396 	safe_kstat_read(kc, ksp, NULL);
397 	buf_size_index = safe_kstat_data_index(ksp, "buf_size");
398 	slab_create_index = safe_kstat_data_index(ksp, "slab_create");
399 	slab_destroy_index = safe_kstat_data_index(ksp, "slab_destroy");
400 	slab_size_index = safe_kstat_data_index(ksp, "slab_size");
401 	buf_avail_index = safe_kstat_data_index(ksp, "buf_avail");
402 	alloc_fail_index = safe_kstat_data_index(ksp, "alloc_fail");
403 }
404 
405 /*
406  * load statistics, summing across CPUs where needed
407  */
408 
409 static int
410 all_stat_load(void)
411 {
412 	int i, j;
413 	cpu_stat_t cs;
414 	ulong_t *np, *tp;
415 	uint64_t cpu_tick[4] = {0, 0, 0, 0};
416 
417 	memset(&d, 0, sizeof (d));
418 
419 	/*
420 	 * Global statistics
421 	 */
422 
423 	safe_kstat_read(kc, sysinfo_ksp, (void *) &d.si);
424 	safe_kstat_read(kc, vminfo_ksp, (void *) &d.vmi);
425 	safe_kstat_read(kc, system_misc_ksp, NULL);
426 	safe_kstat_read(kc, file_cache_ksp, NULL);
427 
428 	if (ufs_inode_ksp != NULL) {
429 		safe_kstat_read(kc, ufs_inode_ksp, NULL);
430 		d.szinode = ufs_inode_size_knp->value.ul;
431 	}
432 
433 	d.szfile = file_total_knp->value.ui64 - file_avail_knp->value.ui64;
434 	d.szproc = nproc_knp->value.ul;
435 
436 	d.mszinode = (ninode > d.szinode) ? ninode : d.szinode;
437 	d.mszfile = d.szfile;
438 	d.mszproc = var.v_proc;
439 
440 	/*
441 	 * Per-CPU statistics.
442 	 */
443 
444 	for (i = 0; i < ncpus; i++) {
445 		if (kstat_read(kc, cpu_stat_list[i], (void *) &cs) == -1)
446 			return (1);
447 
448 		np = (ulong_t *)&d.csi;
449 		tp = (ulong_t *)&cs.cpu_sysinfo;
450 
451 		/*
452 		 * Accumulate cpu ticks for CPU_IDLE, CPU_USER, CPU_KERNEL and
453 		 * CPU_WAIT with respect to each of the cpus.
454 		 */
455 		for (j = 0; j < CPU_STATES; j++)
456 			cpu_tick[j] += tp[j];
457 
458 		for (j = 0; j < sizeof (cpu_sysinfo_t); j += sizeof (ulong_t))
459 			*np++ += *tp++;
460 		np = (ulong_t *)&d.cvmi;
461 		tp = (ulong_t *)&cs.cpu_vminfo;
462 		for (j = 0; j < sizeof (cpu_vminfo_t); j += sizeof (ulong_t))
463 			*np++ += *tp++;
464 	}
465 
466 	/*
467 	 * Per-cache kmem statistics.
468 	 */
469 
470 	for (i = 0; i < ncaches; i++) {
471 		kstat_named_t *knp;
472 		u_longlong_t slab_create, slab_destroy, slab_size, mem_total;
473 		u_longlong_t buf_size, buf_avail, alloc_fail;
474 		int kmi_index;
475 
476 		if (kstat_read(kc, kmem_cache_list[i], NULL) == -1)
477 			return (1);
478 		knp = kmem_cache_list[i]->ks_data;
479 		slab_create	= knp[slab_create_index].value.ui64;
480 		slab_destroy	= knp[slab_destroy_index].value.ui64;
481 		slab_size	= knp[slab_size_index].value.ui64;
482 		buf_size	= knp[buf_size_index].value.ui64;
483 		buf_avail	= knp[buf_avail_index].value.ui64;
484 		alloc_fail	= knp[alloc_fail_index].value.ui64;
485 		if (buf_size <= 256)
486 			kmi_index = KMEM_SMALL;
487 		else
488 			kmi_index = KMEM_LARGE;
489 		mem_total = (slab_create - slab_destroy) * slab_size;
490 
491 		d.kmi.km_mem[kmi_index] += (ulong_t)mem_total;
492 		d.kmi.km_alloc[kmi_index] +=
493 		    (ulong_t)mem_total - buf_size * buf_avail;
494 		d.kmi.km_fail[kmi_index] += (ulong_t)alloc_fail;
495 	}
496 
497 	safe_kstat_read(kc, kmem_oversize_ksp, NULL);
498 
499 	d.kmi.km_alloc[KMEM_OSIZE] = d.kmi.km_mem[KMEM_OSIZE] =
500 	    oversize_alloc_knp->value.ui64;
501 	d.kmi.km_fail[KMEM_OSIZE] = oversize_fail_knp->value.ui64;
502 
503 	/*
504 	 * Adjust CPU statistics so the delta calculations in sar will
505 	 * be correct when facing changes to the set of online CPUs.
506 	 */
507 	compute_cpu_stat_adj();
508 	for (i = 0; i < CPU_STATES; i++)
509 		d.csi.cpu[i] = (cpu_tick[i] + cpu_stat_adj[i]) / ncpus;
510 
511 	return (0);
512 }
513 
514 static void
515 fail(int do_perror, char *message, ...)
516 {
517 	va_list args;
518 
519 	va_start(args, message);
520 	fprintf(stderr, "%s: ", cmdname);
521 	vfprintf(stderr, message, args);
522 	va_end(args);
523 	if (do_perror)
524 		fprintf(stderr, ": %s", strerror(errno));
525 	fprintf(stderr, "\n");
526 	exit(2);
527 }
528 
529 static void
530 safe_zalloc(void **ptr, int size, int free_first)
531 {
532 	if (free_first && *ptr != NULL)
533 		free(*ptr);
534 	if ((*ptr = malloc(size)) == NULL)
535 		fail(1, "malloc failed");
536 	memset(*ptr, 0, size);
537 }
538 
539 static kid_t
540 safe_kstat_read(kstat_ctl_t *kc, kstat_t *ksp, void *data)
541 {
542 	kid_t kstat_chain_id = kstat_read(kc, ksp, data);
543 
544 	if (kstat_chain_id == -1)
545 		fail(1, "kstat_read(%x, '%s') failed", kc, ksp->ks_name);
546 	return (kstat_chain_id);
547 }
548 
549 static kstat_t *
550 safe_kstat_lookup(kstat_ctl_t *kc, char *ks_module, int ks_instance,
551 	char *ks_name)
552 {
553 	kstat_t *ksp = kstat_lookup(kc, ks_module, ks_instance, ks_name);
554 
555 	if (ksp == NULL)
556 		fail(0, "kstat_lookup('%s', %d, '%s') failed",
557 		    ks_module == NULL ? "" : ks_module,
558 		    ks_instance,
559 		    ks_name == NULL ? "" : ks_name);
560 	return (ksp);
561 }
562 
563 static void *
564 safe_kstat_data_lookup(kstat_t *ksp, char *name)
565 {
566 	void *fp = kstat_data_lookup(ksp, name);
567 
568 	if (fp == NULL)
569 		fail(0, "kstat_data_lookup('%s', '%s') failed",
570 		    ksp->ks_name, name);
571 	return (fp);
572 }
573 
574 static int
575 safe_kstat_data_index(kstat_t *ksp, char *name)
576 {
577 	return ((int)((char *)safe_kstat_data_lookup(ksp, name) -
578 	    (char *)ksp->ks_data) / (ksp->ks_data_size / ksp->ks_ndata));
579 }
580 
581 static int
582 kscmp(kstat_t *ks1, kstat_t *ks2)
583 {
584 	int cmp;
585 
586 	cmp = strcmp(ks1->ks_module, ks2->ks_module);
587 	if (cmp != 0)
588 		return (cmp);
589 	cmp = ks1->ks_instance - ks2->ks_instance;
590 	if (cmp != 0)
591 		return (cmp);
592 	return (strcmp(ks1->ks_name, ks2->ks_name));
593 }
594 
595 static void
596 init_iodevs(void)
597 {
598 	struct iodevinfo *iodev, *previodev, *comp;
599 	kstat_t *ksp;
600 
601 	iodev = &zeroiodev;
602 	niodevs = 0;
603 
604 	/*
605 	 * Patch the snip in the iodevinfo list (see below)
606 	 */
607 	if (snip)
608 		lastiodev->next = snip;
609 
610 	for (ksp = kc->kc_chain; ksp; ksp = ksp->ks_next) {
611 
612 		if (ksp->ks_type != KSTAT_TYPE_IO)
613 			continue;
614 		previodev = iodev;
615 		if (iodev->next)
616 			iodev = iodev->next;
617 		else {
618 			safe_zalloc((void **) &iodev->next,
619 			    sizeof (struct iodevinfo), 0);
620 			iodev = iodev->next;
621 			iodev->next = NULL;
622 		}
623 		iodev->ksp = ksp;
624 		iodev->ks = *ksp;
625 		memset((void *)&iodev->kios, 0, sizeof (kstat_io_t));
626 		iodev->kios.wlastupdate = iodev->ks.ks_crtime;
627 		iodev->kios.rlastupdate = iodev->ks.ks_crtime;
628 
629 		/*
630 		 * Insertion sort on (ks_module, ks_instance, ks_name)
631 		 */
632 		comp = &zeroiodev;
633 		while (kscmp(&iodev->ks, &comp->next->ks) > 0)
634 			comp = comp->next;
635 		if (previodev != comp) {
636 			previodev->next = iodev->next;
637 			iodev->next = comp->next;
638 			comp->next = iodev;
639 			iodev = previodev;
640 		}
641 		niodevs++;
642 	}
643 	/*
644 	 * Put a snip in the linked list of iodevinfos.  The idea:
645 	 * If there was a state change such that now there are fewer
646 	 * iodevs, we snip the list and retain the tail, rather than
647 	 * freeing it.  At the next state change, we clip the tail back on.
648 	 * This prevents a lot of malloc/free activity, and it's simpler.
649 	 */
650 	lastiodev = iodev;
651 	snip = iodev->next;
652 	iodev->next = NULL;
653 
654 	firstiodev = zeroiodev.next;
655 }
656 
657 static int
658 iodevinfo_load(void)
659 {
660 	struct iodevinfo *iodev;
661 
662 	for (iodev = firstiodev; iodev; iodev = iodev->next) {
663 		if (kstat_read(kc, iodev->ksp, (void *) &iodev->kios) == -1)
664 			return (1);
665 	}
666 	return (0);
667 }
668 
669 static int
670 kstat_copy(const kstat_t *src, kstat_t *dst)
671 {
672 	*dst = *src;
673 
674 	if (src->ks_data != NULL) {
675 		if ((dst->ks_data = malloc(src->ks_data_size)) == NULL)
676 			return (-1);
677 		bcopy(src->ks_data, dst->ks_data, src->ks_data_size);
678 	} else {
679 		dst->ks_data = NULL;
680 		dst->ks_data_size = 0;
681 	}
682 	return (0);
683 }
684 
685 /*
686  * Determine what is different between two sets of kstats; s[0] and s[1]
687  * are arrays of kstats of size ns0 and ns1, respectively, and sorted by
688  * instance number.  u[0] and u[1] are two arrays which must be
689  * caller-zallocated; each must be of size MAX(ns0, ns1).  When the
690  * function terminates, u[0] contains all s[0]-unique items and u[1]
691  * contains all s[1]-unique items.  Any unused entries in u[0] and u[1]
692  * are left NULL.
693  */
694 static void
695 diff_two_arrays(kstat_t ** const s[], size_t ns0, size_t ns1,
696     kstat_t ** const u[])
697 {
698 	kstat_t **s0p = s[0], **s1p = s[1];
699 	kstat_t **u0p = u[0], **u1p = u[1];
700 	int i = 0, j = 0;
701 
702 	while (i < ns0 && j < ns1) {
703 		if ((*s0p)->ks_instance == (*s1p)->ks_instance) {
704 			if ((*s0p)->ks_kid != (*s1p)->ks_kid) {
705 				/*
706 				 * The instance is the same, but this
707 				 * CPU has been offline during the
708 				 * interval, so we consider *u0p to
709 				 * be s0p-unique, and similarly for
710 				 * *u1p.
711 				 */
712 				*(u0p++) = *s0p;
713 				*(u1p++) = *s1p;
714 			}
715 			s0p++;
716 			i++;
717 			s1p++;
718 			j++;
719 		} else if ((*s0p)->ks_instance < (*s1p)->ks_instance) {
720 			*(u0p++) = *(s0p++);
721 			i++;
722 		} else {
723 			*(u1p++) = *(s1p++);
724 			j++;
725 		}
726 	}
727 
728 	while (i < ns0) {
729 		*(u0p++) = *(s0p++);
730 		i++;
731 	}
732 	while (j < ns1) {
733 		*(u1p++) = *(s1p++);
734 		j++;
735 	}
736 }
737 
738 static int
739 cpuid_compare(const void *p1, const void *p2)
740 {
741 	return ((*(kstat_t **)p1)->ks_instance -
742 	    (*(kstat_t **)p2)->ks_instance);
743 }
744 
745 /*
746  * Identify those CPUs which were not present for the whole interval so
747  * their statistics can be removed from the aggregate.
748  */
749 static void
750 compute_cpu_stat_adj(void)
751 {
752 	int i, j;
753 
754 	if (ocpu_stat_list) {
755 		kstat_t **s[2];
756 		kstat_t **inarray[2];
757 		int max_cpus = MAX(ncpus, oncpus);
758 
759 		qsort(cpu_stat_list, ncpus, sizeof (*cpu_stat_list),
760 		    cpuid_compare);
761 		qsort(ocpu_stat_list, oncpus, sizeof (*ocpu_stat_list),
762 		    cpuid_compare);
763 
764 		s[0] = ocpu_stat_list;
765 		s[1] = cpu_stat_list;
766 
767 		safe_zalloc((void *)&inarray[0], sizeof (**inarray) * max_cpus,
768 		    0);
769 		safe_zalloc((void *)&inarray[1], sizeof (**inarray) * max_cpus,
770 		    0);
771 		diff_two_arrays(s, oncpus, ncpus, inarray);
772 
773 		for (i = 0; i < max_cpus; i++) {
774 			if (inarray[0][i])
775 				for (j = 0; j < CPU_STATES; j++)
776 					cpu_stat_adj[j] +=
777 					    ((cpu_stat_t *)inarray[0][i]
778 					    ->ks_data)->cpu_sysinfo.cpu[j];
779 			if (inarray[1][i])
780 				for (j = 0; j < CPU_STATES; j++)
781 					cpu_stat_adj[j] -=
782 					    ((cpu_stat_t *)inarray[1][i]
783 					    ->ks_data)->cpu_sysinfo.cpu[j];
784 		}
785 
786 		free(inarray[0]);
787 		free(inarray[1]);
788 	}
789 
790 	/*
791 	 * Preserve the last interval's CPU stats.
792 	 */
793 	if (cpu_stat_list) {
794 		for (i = 0; i < oncpus; i++)
795 			free(ocpu_stat_list[i]->ks_data);
796 
797 		oncpus = ncpus;
798 		safe_zalloc((void **)&ocpu_stat_list, oncpus *
799 		    sizeof (*ocpu_stat_list), 1);
800 		for (i = 0; i < ncpus; i++) {
801 			safe_zalloc((void *)&ocpu_stat_list[i],
802 			    sizeof (*ocpu_stat_list[0]), 0);
803 			if (kstat_copy(cpu_stat_list[i], ocpu_stat_list[i]))
804 				fail(1, "kstat_copy() failed");
805 		}
806 	}
807 }
808