xref: /titanic_44/usr/src/lib/libdtrace/common/dt_aggregate.c (revision 48a4016cae8aa2b8b3d8b258eb22e0c781912bed)
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  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
29  * Copyright (c) 2012 by Delphix. All rights reserved.
30  */
31 
32 #include <stdlib.h>
33 #include <strings.h>
34 #include <errno.h>
35 #include <unistd.h>
36 #include <dt_impl.h>
37 #include <assert.h>
38 #include <alloca.h>
39 #include <limits.h>
40 
41 #define	DTRACE_AHASHSIZE	32779		/* big 'ol prime */
42 
43 /*
44  * Because qsort(3C) does not allow an argument to be passed to a comparison
45  * function, the variables that affect comparison must regrettably be global;
46  * they are protected by a global static lock, dt_qsort_lock.
47  */
48 static pthread_mutex_t dt_qsort_lock = PTHREAD_MUTEX_INITIALIZER;
49 
50 static int dt_revsort;
51 static int dt_keysort;
52 static int dt_keypos;
53 
54 #define	DT_LESSTHAN	(dt_revsort == 0 ? -1 : 1)
55 #define	DT_GREATERTHAN	(dt_revsort == 0 ? 1 : -1)
56 
57 static void
58 dt_aggregate_count(int64_t *existing, int64_t *new, size_t size)
59 {
60 	int i;
61 
62 	for (i = 0; i < size / sizeof (int64_t); i++)
63 		existing[i] = existing[i] + new[i];
64 }
65 
66 static int
67 dt_aggregate_countcmp(int64_t *lhs, int64_t *rhs)
68 {
69 	int64_t lvar = *lhs;
70 	int64_t rvar = *rhs;
71 
72 	if (lvar < rvar)
73 		return (DT_LESSTHAN);
74 
75 	if (lvar > rvar)
76 		return (DT_GREATERTHAN);
77 
78 	return (0);
79 }
80 
81 /*ARGSUSED*/
82 static void
83 dt_aggregate_min(int64_t *existing, int64_t *new, size_t size)
84 {
85 	if (*new < *existing)
86 		*existing = *new;
87 }
88 
89 /*ARGSUSED*/
90 static void
91 dt_aggregate_max(int64_t *existing, int64_t *new, size_t size)
92 {
93 	if (*new > *existing)
94 		*existing = *new;
95 }
96 
97 static int
98 dt_aggregate_averagecmp(int64_t *lhs, int64_t *rhs)
99 {
100 	int64_t lavg = lhs[0] ? (lhs[1] / lhs[0]) : 0;
101 	int64_t ravg = rhs[0] ? (rhs[1] / rhs[0]) : 0;
102 
103 	if (lavg < ravg)
104 		return (DT_LESSTHAN);
105 
106 	if (lavg > ravg)
107 		return (DT_GREATERTHAN);
108 
109 	return (0);
110 }
111 
112 static int
113 dt_aggregate_stddevcmp(int64_t *lhs, int64_t *rhs)
114 {
115 	uint64_t lsd = dt_stddev((uint64_t *)lhs, 1);
116 	uint64_t rsd = dt_stddev((uint64_t *)rhs, 1);
117 
118 	if (lsd < rsd)
119 		return (DT_LESSTHAN);
120 
121 	if (lsd > rsd)
122 		return (DT_GREATERTHAN);
123 
124 	return (0);
125 }
126 
127 /*ARGSUSED*/
128 static void
129 dt_aggregate_lquantize(int64_t *existing, int64_t *new, size_t size)
130 {
131 	int64_t arg = *existing++;
132 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
133 	int i;
134 
135 	for (i = 0; i <= levels + 1; i++)
136 		existing[i] = existing[i] + new[i + 1];
137 }
138 
139 static long double
140 dt_aggregate_lquantizedsum(int64_t *lquanta)
141 {
142 	int64_t arg = *lquanta++;
143 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
144 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
145 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg), i;
146 	long double total = (long double)lquanta[0] * (long double)(base - 1);
147 
148 	for (i = 0; i < levels; base += step, i++)
149 		total += (long double)lquanta[i + 1] * (long double)base;
150 
151 	return (total + (long double)lquanta[levels + 1] *
152 	    (long double)(base + 1));
153 }
154 
155 static int64_t
156 dt_aggregate_lquantizedzero(int64_t *lquanta)
157 {
158 	int64_t arg = *lquanta++;
159 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
160 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
161 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg), i;
162 
163 	if (base - 1 == 0)
164 		return (lquanta[0]);
165 
166 	for (i = 0; i < levels; base += step, i++) {
167 		if (base != 0)
168 			continue;
169 
170 		return (lquanta[i + 1]);
171 	}
172 
173 	if (base + 1 == 0)
174 		return (lquanta[levels + 1]);
175 
176 	return (0);
177 }
178 
179 static int
180 dt_aggregate_lquantizedcmp(int64_t *lhs, int64_t *rhs)
181 {
182 	long double lsum = dt_aggregate_lquantizedsum(lhs);
183 	long double rsum = dt_aggregate_lquantizedsum(rhs);
184 	int64_t lzero, rzero;
185 
186 	if (lsum < rsum)
187 		return (DT_LESSTHAN);
188 
189 	if (lsum > rsum)
190 		return (DT_GREATERTHAN);
191 
192 	/*
193 	 * If they're both equal, then we will compare based on the weights at
194 	 * zero.  If the weights at zero are equal (or if zero is not within
195 	 * the range of the linear quantization), then this will be judged a
196 	 * tie and will be resolved based on the key comparison.
197 	 */
198 	lzero = dt_aggregate_lquantizedzero(lhs);
199 	rzero = dt_aggregate_lquantizedzero(rhs);
200 
201 	if (lzero < rzero)
202 		return (DT_LESSTHAN);
203 
204 	if (lzero > rzero)
205 		return (DT_GREATERTHAN);
206 
207 	return (0);
208 }
209 
210 static void
211 dt_aggregate_llquantize(int64_t *existing, int64_t *new, size_t size)
212 {
213 	int i;
214 
215 	for (i = 1; i < size / sizeof (int64_t); i++)
216 		existing[i] = existing[i] + new[i];
217 }
218 
219 static long double
220 dt_aggregate_llquantizedsum(int64_t *llquanta)
221 {
222 	int64_t arg = *llquanta++;
223 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
224 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
225 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
226 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
227 	int bin = 0, order;
228 	int64_t value = 1, next, step;
229 	long double total;
230 
231 	assert(nsteps >= factor);
232 	assert(nsteps % factor == 0);
233 
234 	for (order = 0; order < low; order++)
235 		value *= factor;
236 
237 	total = (long double)llquanta[bin++] * (long double)(value - 1);
238 
239 	next = value * factor;
240 	step = next > nsteps ? next / nsteps : 1;
241 
242 	while (order <= high) {
243 		assert(value < next);
244 		total += (long double)llquanta[bin++] * (long double)(value);
245 
246 		if ((value += step) != next)
247 			continue;
248 
249 		next = value * factor;
250 		step = next > nsteps ? next / nsteps : 1;
251 		order++;
252 	}
253 
254 	return (total + (long double)llquanta[bin] * (long double)value);
255 }
256 
257 static int
258 dt_aggregate_llquantizedcmp(int64_t *lhs, int64_t *rhs)
259 {
260 	long double lsum = dt_aggregate_llquantizedsum(lhs);
261 	long double rsum = dt_aggregate_llquantizedsum(rhs);
262 	int64_t lzero, rzero;
263 
264 	if (lsum < rsum)
265 		return (DT_LESSTHAN);
266 
267 	if (lsum > rsum)
268 		return (DT_GREATERTHAN);
269 
270 	/*
271 	 * If they're both equal, then we will compare based on the weights at
272 	 * zero.  If the weights at zero are equal, then this will be judged a
273 	 * tie and will be resolved based on the key comparison.
274 	 */
275 	lzero = lhs[1];
276 	rzero = rhs[1];
277 
278 	if (lzero < rzero)
279 		return (DT_LESSTHAN);
280 
281 	if (lzero > rzero)
282 		return (DT_GREATERTHAN);
283 
284 	return (0);
285 }
286 
287 static int
288 dt_aggregate_quantizedcmp(int64_t *lhs, int64_t *rhs)
289 {
290 	int nbuckets = DTRACE_QUANTIZE_NBUCKETS, i;
291 	long double ltotal = 0, rtotal = 0;
292 	int64_t lzero, rzero;
293 
294 	for (i = 0; i < nbuckets; i++) {
295 		int64_t bucketval = DTRACE_QUANTIZE_BUCKETVAL(i);
296 
297 		if (bucketval == 0) {
298 			lzero = lhs[i];
299 			rzero = rhs[i];
300 		}
301 
302 		ltotal += (long double)bucketval * (long double)lhs[i];
303 		rtotal += (long double)bucketval * (long double)rhs[i];
304 	}
305 
306 	if (ltotal < rtotal)
307 		return (DT_LESSTHAN);
308 
309 	if (ltotal > rtotal)
310 		return (DT_GREATERTHAN);
311 
312 	/*
313 	 * If they're both equal, then we will compare based on the weights at
314 	 * zero.  If the weights at zero are equal, then this will be judged a
315 	 * tie and will be resolved based on the key comparison.
316 	 */
317 	if (lzero < rzero)
318 		return (DT_LESSTHAN);
319 
320 	if (lzero > rzero)
321 		return (DT_GREATERTHAN);
322 
323 	return (0);
324 }
325 
326 static void
327 dt_aggregate_usym(dtrace_hdl_t *dtp, uint64_t *data)
328 {
329 	uint64_t pid = data[0];
330 	uint64_t *pc = &data[1];
331 	struct ps_prochandle *P;
332 	GElf_Sym sym;
333 
334 	if (dtp->dt_vector != NULL)
335 		return;
336 
337 	if ((P = dt_proc_grab(dtp, pid, PGRAB_RDONLY | PGRAB_FORCE, 0)) == NULL)
338 		return;
339 
340 	dt_proc_lock(dtp, P);
341 
342 	if (Plookup_by_addr(P, *pc, NULL, 0, &sym) == 0)
343 		*pc = sym.st_value;
344 
345 	dt_proc_unlock(dtp, P);
346 	dt_proc_release(dtp, P);
347 }
348 
349 static void
350 dt_aggregate_umod(dtrace_hdl_t *dtp, uint64_t *data)
351 {
352 	uint64_t pid = data[0];
353 	uint64_t *pc = &data[1];
354 	struct ps_prochandle *P;
355 	const prmap_t *map;
356 
357 	if (dtp->dt_vector != NULL)
358 		return;
359 
360 	if ((P = dt_proc_grab(dtp, pid, PGRAB_RDONLY | PGRAB_FORCE, 0)) == NULL)
361 		return;
362 
363 	dt_proc_lock(dtp, P);
364 
365 	if ((map = Paddr_to_map(P, *pc)) != NULL)
366 		*pc = map->pr_vaddr;
367 
368 	dt_proc_unlock(dtp, P);
369 	dt_proc_release(dtp, P);
370 }
371 
372 static void
373 dt_aggregate_sym(dtrace_hdl_t *dtp, uint64_t *data)
374 {
375 	GElf_Sym sym;
376 	uint64_t *pc = data;
377 
378 	if (dtrace_lookup_by_addr(dtp, *pc, &sym, NULL) == 0)
379 		*pc = sym.st_value;
380 }
381 
382 static void
383 dt_aggregate_mod(dtrace_hdl_t *dtp, uint64_t *data)
384 {
385 	uint64_t *pc = data;
386 	dt_module_t *dmp;
387 
388 	if (dtp->dt_vector != NULL) {
389 		/*
390 		 * We don't have a way of just getting the module for a
391 		 * vectored open, and it doesn't seem to be worth defining
392 		 * one.  This means that use of mod() won't get true
393 		 * aggregation in the postmortem case (some modules may
394 		 * appear more than once in aggregation output).  It seems
395 		 * unlikely that anyone will ever notice or care...
396 		 */
397 		return;
398 	}
399 
400 	for (dmp = dt_list_next(&dtp->dt_modlist); dmp != NULL;
401 	    dmp = dt_list_next(dmp)) {
402 		if (*pc - dmp->dm_text_va < dmp->dm_text_size) {
403 			*pc = dmp->dm_text_va;
404 			return;
405 		}
406 	}
407 }
408 
409 static dtrace_aggvarid_t
410 dt_aggregate_aggvarid(dt_ahashent_t *ent)
411 {
412 	dtrace_aggdesc_t *agg = ent->dtahe_data.dtada_desc;
413 	caddr_t data = ent->dtahe_data.dtada_data;
414 	dtrace_recdesc_t *rec = agg->dtagd_rec;
415 
416 	/*
417 	 * First, we'll check the variable ID in the aggdesc.  If it's valid,
418 	 * we'll return it.  If not, we'll use the compiler-generated ID
419 	 * present as the first record.
420 	 */
421 	if (agg->dtagd_varid != DTRACE_AGGVARIDNONE)
422 		return (agg->dtagd_varid);
423 
424 	agg->dtagd_varid = *((dtrace_aggvarid_t *)(uintptr_t)(data +
425 	    rec->dtrd_offset));
426 
427 	return (agg->dtagd_varid);
428 }
429 
430 
431 static int
432 dt_aggregate_snap_cpu(dtrace_hdl_t *dtp, processorid_t cpu)
433 {
434 	dtrace_epid_t id;
435 	uint64_t hashval;
436 	size_t offs, roffs, size, ndx;
437 	int i, j, rval;
438 	caddr_t addr, data;
439 	dtrace_recdesc_t *rec;
440 	dt_aggregate_t *agp = &dtp->dt_aggregate;
441 	dtrace_aggdesc_t *agg;
442 	dt_ahash_t *hash = &agp->dtat_hash;
443 	dt_ahashent_t *h;
444 	dtrace_bufdesc_t b = agp->dtat_buf, *buf = &b;
445 	dtrace_aggdata_t *aggdata;
446 	int flags = agp->dtat_flags;
447 
448 	buf->dtbd_cpu = cpu;
449 
450 	if (dt_ioctl(dtp, DTRACEIOC_AGGSNAP, buf) == -1) {
451 		if (errno == ENOENT) {
452 			/*
453 			 * If that failed with ENOENT, it may be because the
454 			 * CPU was unconfigured.  This is okay; we'll just
455 			 * do nothing but return success.
456 			 */
457 			return (0);
458 		}
459 
460 		return (dt_set_errno(dtp, errno));
461 	}
462 
463 	if (buf->dtbd_drops != 0) {
464 		if (dt_handle_cpudrop(dtp, cpu,
465 		    DTRACEDROP_AGGREGATION, buf->dtbd_drops) == -1)
466 			return (-1);
467 	}
468 
469 	if (buf->dtbd_size == 0)
470 		return (0);
471 
472 	if (hash->dtah_hash == NULL) {
473 		size_t size;
474 
475 		hash->dtah_size = DTRACE_AHASHSIZE;
476 		size = hash->dtah_size * sizeof (dt_ahashent_t *);
477 
478 		if ((hash->dtah_hash = malloc(size)) == NULL)
479 			return (dt_set_errno(dtp, EDT_NOMEM));
480 
481 		bzero(hash->dtah_hash, size);
482 	}
483 
484 	for (offs = 0; offs < buf->dtbd_size; ) {
485 		/*
486 		 * We're guaranteed to have an ID.
487 		 */
488 		id = *((dtrace_epid_t *)((uintptr_t)buf->dtbd_data +
489 		    (uintptr_t)offs));
490 
491 		if (id == DTRACE_AGGIDNONE) {
492 			/*
493 			 * This is filler to assure proper alignment of the
494 			 * next record; we simply ignore it.
495 			 */
496 			offs += sizeof (id);
497 			continue;
498 		}
499 
500 		if ((rval = dt_aggid_lookup(dtp, id, &agg)) != 0)
501 			return (rval);
502 
503 		addr = buf->dtbd_data + offs;
504 		size = agg->dtagd_size;
505 		hashval = 0;
506 
507 		for (j = 0; j < agg->dtagd_nrecs - 1; j++) {
508 			rec = &agg->dtagd_rec[j];
509 			roffs = rec->dtrd_offset;
510 
511 			switch (rec->dtrd_action) {
512 			case DTRACEACT_USYM:
513 				dt_aggregate_usym(dtp,
514 				    /* LINTED - alignment */
515 				    (uint64_t *)&addr[roffs]);
516 				break;
517 
518 			case DTRACEACT_UMOD:
519 				dt_aggregate_umod(dtp,
520 				    /* LINTED - alignment */
521 				    (uint64_t *)&addr[roffs]);
522 				break;
523 
524 			case DTRACEACT_SYM:
525 				/* LINTED - alignment */
526 				dt_aggregate_sym(dtp, (uint64_t *)&addr[roffs]);
527 				break;
528 
529 			case DTRACEACT_MOD:
530 				/* LINTED - alignment */
531 				dt_aggregate_mod(dtp, (uint64_t *)&addr[roffs]);
532 				break;
533 
534 			default:
535 				break;
536 			}
537 
538 			for (i = 0; i < rec->dtrd_size; i++)
539 				hashval += addr[roffs + i];
540 		}
541 
542 		ndx = hashval % hash->dtah_size;
543 
544 		for (h = hash->dtah_hash[ndx]; h != NULL; h = h->dtahe_next) {
545 			if (h->dtahe_hashval != hashval)
546 				continue;
547 
548 			if (h->dtahe_size != size)
549 				continue;
550 
551 			aggdata = &h->dtahe_data;
552 			data = aggdata->dtada_data;
553 
554 			for (j = 0; j < agg->dtagd_nrecs - 1; j++) {
555 				rec = &agg->dtagd_rec[j];
556 				roffs = rec->dtrd_offset;
557 
558 				for (i = 0; i < rec->dtrd_size; i++)
559 					if (addr[roffs + i] != data[roffs + i])
560 						goto hashnext;
561 			}
562 
563 			/*
564 			 * We found it.  Now we need to apply the aggregating
565 			 * action on the data here.
566 			 */
567 			rec = &agg->dtagd_rec[agg->dtagd_nrecs - 1];
568 			roffs = rec->dtrd_offset;
569 			/* LINTED - alignment */
570 			h->dtahe_aggregate((int64_t *)&data[roffs],
571 			    /* LINTED - alignment */
572 			    (int64_t *)&addr[roffs], rec->dtrd_size);
573 
574 			/*
575 			 * If we're keeping per CPU data, apply the aggregating
576 			 * action there as well.
577 			 */
578 			if (aggdata->dtada_percpu != NULL) {
579 				data = aggdata->dtada_percpu[cpu];
580 
581 				/* LINTED - alignment */
582 				h->dtahe_aggregate((int64_t *)data,
583 				    /* LINTED - alignment */
584 				    (int64_t *)&addr[roffs], rec->dtrd_size);
585 			}
586 
587 			goto bufnext;
588 hashnext:
589 			continue;
590 		}
591 
592 		/*
593 		 * If we're here, we couldn't find an entry for this record.
594 		 */
595 		if ((h = malloc(sizeof (dt_ahashent_t))) == NULL)
596 			return (dt_set_errno(dtp, EDT_NOMEM));
597 		bzero(h, sizeof (dt_ahashent_t));
598 		aggdata = &h->dtahe_data;
599 
600 		if ((aggdata->dtada_data = malloc(size)) == NULL) {
601 			free(h);
602 			return (dt_set_errno(dtp, EDT_NOMEM));
603 		}
604 
605 		bcopy(addr, aggdata->dtada_data, size);
606 		aggdata->dtada_size = size;
607 		aggdata->dtada_desc = agg;
608 		aggdata->dtada_handle = dtp;
609 		(void) dt_epid_lookup(dtp, agg->dtagd_epid,
610 		    &aggdata->dtada_edesc, &aggdata->dtada_pdesc);
611 		aggdata->dtada_normal = 1;
612 
613 		h->dtahe_hashval = hashval;
614 		h->dtahe_size = size;
615 		(void) dt_aggregate_aggvarid(h);
616 
617 		rec = &agg->dtagd_rec[agg->dtagd_nrecs - 1];
618 
619 		if (flags & DTRACE_A_PERCPU) {
620 			int max_cpus = agp->dtat_maxcpu;
621 			caddr_t *percpu = malloc(max_cpus * sizeof (caddr_t));
622 
623 			if (percpu == NULL) {
624 				free(aggdata->dtada_data);
625 				free(h);
626 				return (dt_set_errno(dtp, EDT_NOMEM));
627 			}
628 
629 			for (j = 0; j < max_cpus; j++) {
630 				percpu[j] = malloc(rec->dtrd_size);
631 
632 				if (percpu[j] == NULL) {
633 					while (--j >= 0)
634 						free(percpu[j]);
635 
636 					free(aggdata->dtada_data);
637 					free(h);
638 					return (dt_set_errno(dtp, EDT_NOMEM));
639 				}
640 
641 				if (j == cpu) {
642 					bcopy(&addr[rec->dtrd_offset],
643 					    percpu[j], rec->dtrd_size);
644 				} else {
645 					bzero(percpu[j], rec->dtrd_size);
646 				}
647 			}
648 
649 			aggdata->dtada_percpu = percpu;
650 		}
651 
652 		switch (rec->dtrd_action) {
653 		case DTRACEAGG_MIN:
654 			h->dtahe_aggregate = dt_aggregate_min;
655 			break;
656 
657 		case DTRACEAGG_MAX:
658 			h->dtahe_aggregate = dt_aggregate_max;
659 			break;
660 
661 		case DTRACEAGG_LQUANTIZE:
662 			h->dtahe_aggregate = dt_aggregate_lquantize;
663 			break;
664 
665 		case DTRACEAGG_LLQUANTIZE:
666 			h->dtahe_aggregate = dt_aggregate_llquantize;
667 			break;
668 
669 		case DTRACEAGG_COUNT:
670 		case DTRACEAGG_SUM:
671 		case DTRACEAGG_AVG:
672 		case DTRACEAGG_STDDEV:
673 		case DTRACEAGG_QUANTIZE:
674 			h->dtahe_aggregate = dt_aggregate_count;
675 			break;
676 
677 		default:
678 			return (dt_set_errno(dtp, EDT_BADAGG));
679 		}
680 
681 		if (hash->dtah_hash[ndx] != NULL)
682 			hash->dtah_hash[ndx]->dtahe_prev = h;
683 
684 		h->dtahe_next = hash->dtah_hash[ndx];
685 		hash->dtah_hash[ndx] = h;
686 
687 		if (hash->dtah_all != NULL)
688 			hash->dtah_all->dtahe_prevall = h;
689 
690 		h->dtahe_nextall = hash->dtah_all;
691 		hash->dtah_all = h;
692 bufnext:
693 		offs += agg->dtagd_size;
694 	}
695 
696 	return (0);
697 }
698 
699 int
700 dtrace_aggregate_snap(dtrace_hdl_t *dtp)
701 {
702 	int i, rval;
703 	dt_aggregate_t *agp = &dtp->dt_aggregate;
704 	hrtime_t now = gethrtime();
705 	dtrace_optval_t interval = dtp->dt_options[DTRACEOPT_AGGRATE];
706 
707 	if (dtp->dt_lastagg != 0) {
708 		if (now - dtp->dt_lastagg < interval)
709 			return (0);
710 
711 		dtp->dt_lastagg += interval;
712 	} else {
713 		dtp->dt_lastagg = now;
714 	}
715 
716 	if (!dtp->dt_active)
717 		return (dt_set_errno(dtp, EINVAL));
718 
719 	if (agp->dtat_buf.dtbd_size == 0)
720 		return (0);
721 
722 	for (i = 0; i < agp->dtat_ncpus; i++) {
723 		if (rval = dt_aggregate_snap_cpu(dtp, agp->dtat_cpus[i]))
724 			return (rval);
725 	}
726 
727 	return (0);
728 }
729 
730 static int
731 dt_aggregate_hashcmp(const void *lhs, const void *rhs)
732 {
733 	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
734 	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
735 	dtrace_aggdesc_t *lagg = lh->dtahe_data.dtada_desc;
736 	dtrace_aggdesc_t *ragg = rh->dtahe_data.dtada_desc;
737 
738 	if (lagg->dtagd_nrecs < ragg->dtagd_nrecs)
739 		return (DT_LESSTHAN);
740 
741 	if (lagg->dtagd_nrecs > ragg->dtagd_nrecs)
742 		return (DT_GREATERTHAN);
743 
744 	return (0);
745 }
746 
747 static int
748 dt_aggregate_varcmp(const void *lhs, const void *rhs)
749 {
750 	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
751 	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
752 	dtrace_aggvarid_t lid, rid;
753 
754 	lid = dt_aggregate_aggvarid(lh);
755 	rid = dt_aggregate_aggvarid(rh);
756 
757 	if (lid < rid)
758 		return (DT_LESSTHAN);
759 
760 	if (lid > rid)
761 		return (DT_GREATERTHAN);
762 
763 	return (0);
764 }
765 
766 static int
767 dt_aggregate_keycmp(const void *lhs, const void *rhs)
768 {
769 	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
770 	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
771 	dtrace_aggdesc_t *lagg = lh->dtahe_data.dtada_desc;
772 	dtrace_aggdesc_t *ragg = rh->dtahe_data.dtada_desc;
773 	dtrace_recdesc_t *lrec, *rrec;
774 	char *ldata, *rdata;
775 	int rval, i, j, keypos, nrecs;
776 
777 	if ((rval = dt_aggregate_hashcmp(lhs, rhs)) != 0)
778 		return (rval);
779 
780 	nrecs = lagg->dtagd_nrecs - 1;
781 	assert(nrecs == ragg->dtagd_nrecs - 1);
782 
783 	keypos = dt_keypos + 1 >= nrecs ? 0 : dt_keypos;
784 
785 	for (i = 1; i < nrecs; i++) {
786 		uint64_t lval, rval;
787 		int ndx = i + keypos;
788 
789 		if (ndx >= nrecs)
790 			ndx = ndx - nrecs + 1;
791 
792 		lrec = &lagg->dtagd_rec[ndx];
793 		rrec = &ragg->dtagd_rec[ndx];
794 
795 		ldata = lh->dtahe_data.dtada_data + lrec->dtrd_offset;
796 		rdata = rh->dtahe_data.dtada_data + rrec->dtrd_offset;
797 
798 		if (lrec->dtrd_size < rrec->dtrd_size)
799 			return (DT_LESSTHAN);
800 
801 		if (lrec->dtrd_size > rrec->dtrd_size)
802 			return (DT_GREATERTHAN);
803 
804 		switch (lrec->dtrd_size) {
805 		case sizeof (uint64_t):
806 			/* LINTED - alignment */
807 			lval = *((uint64_t *)ldata);
808 			/* LINTED - alignment */
809 			rval = *((uint64_t *)rdata);
810 			break;
811 
812 		case sizeof (uint32_t):
813 			/* LINTED - alignment */
814 			lval = *((uint32_t *)ldata);
815 			/* LINTED - alignment */
816 			rval = *((uint32_t *)rdata);
817 			break;
818 
819 		case sizeof (uint16_t):
820 			/* LINTED - alignment */
821 			lval = *((uint16_t *)ldata);
822 			/* LINTED - alignment */
823 			rval = *((uint16_t *)rdata);
824 			break;
825 
826 		case sizeof (uint8_t):
827 			lval = *((uint8_t *)ldata);
828 			rval = *((uint8_t *)rdata);
829 			break;
830 
831 		default:
832 			switch (lrec->dtrd_action) {
833 			case DTRACEACT_UMOD:
834 			case DTRACEACT_UADDR:
835 			case DTRACEACT_USYM:
836 				for (j = 0; j < 2; j++) {
837 					/* LINTED - alignment */
838 					lval = ((uint64_t *)ldata)[j];
839 					/* LINTED - alignment */
840 					rval = ((uint64_t *)rdata)[j];
841 
842 					if (lval < rval)
843 						return (DT_LESSTHAN);
844 
845 					if (lval > rval)
846 						return (DT_GREATERTHAN);
847 				}
848 
849 				break;
850 
851 			default:
852 				for (j = 0; j < lrec->dtrd_size; j++) {
853 					lval = ((uint8_t *)ldata)[j];
854 					rval = ((uint8_t *)rdata)[j];
855 
856 					if (lval < rval)
857 						return (DT_LESSTHAN);
858 
859 					if (lval > rval)
860 						return (DT_GREATERTHAN);
861 				}
862 			}
863 
864 			continue;
865 		}
866 
867 		if (lval < rval)
868 			return (DT_LESSTHAN);
869 
870 		if (lval > rval)
871 			return (DT_GREATERTHAN);
872 	}
873 
874 	return (0);
875 }
876 
877 static int
878 dt_aggregate_valcmp(const void *lhs, const void *rhs)
879 {
880 	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
881 	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
882 	dtrace_aggdesc_t *lagg = lh->dtahe_data.dtada_desc;
883 	dtrace_aggdesc_t *ragg = rh->dtahe_data.dtada_desc;
884 	caddr_t ldata = lh->dtahe_data.dtada_data;
885 	caddr_t rdata = rh->dtahe_data.dtada_data;
886 	dtrace_recdesc_t *lrec, *rrec;
887 	int64_t *laddr, *raddr;
888 	int rval;
889 
890 	assert(lagg->dtagd_nrecs == ragg->dtagd_nrecs);
891 
892 	lrec = &lagg->dtagd_rec[lagg->dtagd_nrecs - 1];
893 	rrec = &ragg->dtagd_rec[ragg->dtagd_nrecs - 1];
894 
895 	assert(lrec->dtrd_action == rrec->dtrd_action);
896 
897 	laddr = (int64_t *)(uintptr_t)(ldata + lrec->dtrd_offset);
898 	raddr = (int64_t *)(uintptr_t)(rdata + rrec->dtrd_offset);
899 
900 	switch (lrec->dtrd_action) {
901 	case DTRACEAGG_AVG:
902 		rval = dt_aggregate_averagecmp(laddr, raddr);
903 		break;
904 
905 	case DTRACEAGG_STDDEV:
906 		rval = dt_aggregate_stddevcmp(laddr, raddr);
907 		break;
908 
909 	case DTRACEAGG_QUANTIZE:
910 		rval = dt_aggregate_quantizedcmp(laddr, raddr);
911 		break;
912 
913 	case DTRACEAGG_LQUANTIZE:
914 		rval = dt_aggregate_lquantizedcmp(laddr, raddr);
915 		break;
916 
917 	case DTRACEAGG_LLQUANTIZE:
918 		rval = dt_aggregate_llquantizedcmp(laddr, raddr);
919 		break;
920 
921 	case DTRACEAGG_COUNT:
922 	case DTRACEAGG_SUM:
923 	case DTRACEAGG_MIN:
924 	case DTRACEAGG_MAX:
925 		rval = dt_aggregate_countcmp(laddr, raddr);
926 		break;
927 
928 	default:
929 		assert(0);
930 	}
931 
932 	return (rval);
933 }
934 
935 static int
936 dt_aggregate_valkeycmp(const void *lhs, const void *rhs)
937 {
938 	int rval;
939 
940 	if ((rval = dt_aggregate_valcmp(lhs, rhs)) != 0)
941 		return (rval);
942 
943 	/*
944 	 * If we're here, the values for the two aggregation elements are
945 	 * equal.  We already know that the key layout is the same for the two
946 	 * elements; we must now compare the keys themselves as a tie-breaker.
947 	 */
948 	return (dt_aggregate_keycmp(lhs, rhs));
949 }
950 
951 static int
952 dt_aggregate_keyvarcmp(const void *lhs, const void *rhs)
953 {
954 	int rval;
955 
956 	if ((rval = dt_aggregate_keycmp(lhs, rhs)) != 0)
957 		return (rval);
958 
959 	return (dt_aggregate_varcmp(lhs, rhs));
960 }
961 
962 static int
963 dt_aggregate_varkeycmp(const void *lhs, const void *rhs)
964 {
965 	int rval;
966 
967 	if ((rval = dt_aggregate_varcmp(lhs, rhs)) != 0)
968 		return (rval);
969 
970 	return (dt_aggregate_keycmp(lhs, rhs));
971 }
972 
973 static int
974 dt_aggregate_valvarcmp(const void *lhs, const void *rhs)
975 {
976 	int rval;
977 
978 	if ((rval = dt_aggregate_valkeycmp(lhs, rhs)) != 0)
979 		return (rval);
980 
981 	return (dt_aggregate_varcmp(lhs, rhs));
982 }
983 
984 static int
985 dt_aggregate_varvalcmp(const void *lhs, const void *rhs)
986 {
987 	int rval;
988 
989 	if ((rval = dt_aggregate_varcmp(lhs, rhs)) != 0)
990 		return (rval);
991 
992 	return (dt_aggregate_valkeycmp(lhs, rhs));
993 }
994 
995 static int
996 dt_aggregate_keyvarrevcmp(const void *lhs, const void *rhs)
997 {
998 	return (dt_aggregate_keyvarcmp(rhs, lhs));
999 }
1000 
1001 static int
1002 dt_aggregate_varkeyrevcmp(const void *lhs, const void *rhs)
1003 {
1004 	return (dt_aggregate_varkeycmp(rhs, lhs));
1005 }
1006 
1007 static int
1008 dt_aggregate_valvarrevcmp(const void *lhs, const void *rhs)
1009 {
1010 	return (dt_aggregate_valvarcmp(rhs, lhs));
1011 }
1012 
1013 static int
1014 dt_aggregate_varvalrevcmp(const void *lhs, const void *rhs)
1015 {
1016 	return (dt_aggregate_varvalcmp(rhs, lhs));
1017 }
1018 
1019 static int
1020 dt_aggregate_bundlecmp(const void *lhs, const void *rhs)
1021 {
1022 	dt_ahashent_t **lh = *((dt_ahashent_t ***)lhs);
1023 	dt_ahashent_t **rh = *((dt_ahashent_t ***)rhs);
1024 	int i, rval;
1025 
1026 	if (dt_keysort) {
1027 		/*
1028 		 * If we're sorting on keys, we need to scan until we find the
1029 		 * last entry -- that's the representative key.  (The order of
1030 		 * the bundle is values followed by key to accommodate the
1031 		 * default behavior of sorting by value.)  If the keys are
1032 		 * equal, we'll fall into the value comparison loop, below.
1033 		 */
1034 		for (i = 0; lh[i + 1] != NULL; i++)
1035 			continue;
1036 
1037 		assert(i != 0);
1038 		assert(rh[i + 1] == NULL);
1039 
1040 		if ((rval = dt_aggregate_keycmp(&lh[i], &rh[i])) != 0)
1041 			return (rval);
1042 	}
1043 
1044 	for (i = 0; ; i++) {
1045 		if (lh[i + 1] == NULL) {
1046 			/*
1047 			 * All of the values are equal; if we're sorting on
1048 			 * keys, then we're only here because the keys were
1049 			 * found to be equal and these records are therefore
1050 			 * equal.  If we're not sorting on keys, we'll use the
1051 			 * key comparison from the representative key as the
1052 			 * tie-breaker.
1053 			 */
1054 			if (dt_keysort)
1055 				return (0);
1056 
1057 			assert(i != 0);
1058 			assert(rh[i + 1] == NULL);
1059 			return (dt_aggregate_keycmp(&lh[i], &rh[i]));
1060 		} else {
1061 			if ((rval = dt_aggregate_valcmp(&lh[i], &rh[i])) != 0)
1062 				return (rval);
1063 		}
1064 	}
1065 }
1066 
1067 int
1068 dt_aggregate_go(dtrace_hdl_t *dtp)
1069 {
1070 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1071 	dtrace_optval_t size, cpu;
1072 	dtrace_bufdesc_t *buf = &agp->dtat_buf;
1073 	int rval, i;
1074 
1075 	assert(agp->dtat_maxcpu == 0);
1076 	assert(agp->dtat_ncpu == 0);
1077 	assert(agp->dtat_cpus == NULL);
1078 
1079 	agp->dtat_maxcpu = dt_sysconf(dtp, _SC_CPUID_MAX) + 1;
1080 	agp->dtat_ncpu = dt_sysconf(dtp, _SC_NPROCESSORS_MAX);
1081 	agp->dtat_cpus = malloc(agp->dtat_ncpu * sizeof (processorid_t));
1082 
1083 	if (agp->dtat_cpus == NULL)
1084 		return (dt_set_errno(dtp, EDT_NOMEM));
1085 
1086 	/*
1087 	 * Use the aggregation buffer size as reloaded from the kernel.
1088 	 */
1089 	size = dtp->dt_options[DTRACEOPT_AGGSIZE];
1090 
1091 	rval = dtrace_getopt(dtp, "aggsize", &size);
1092 	assert(rval == 0);
1093 
1094 	if (size == 0 || size == DTRACEOPT_UNSET)
1095 		return (0);
1096 
1097 	buf = &agp->dtat_buf;
1098 	buf->dtbd_size = size;
1099 
1100 	if ((buf->dtbd_data = malloc(buf->dtbd_size)) == NULL)
1101 		return (dt_set_errno(dtp, EDT_NOMEM));
1102 
1103 	/*
1104 	 * Now query for the CPUs enabled.
1105 	 */
1106 	rval = dtrace_getopt(dtp, "cpu", &cpu);
1107 	assert(rval == 0 && cpu != DTRACEOPT_UNSET);
1108 
1109 	if (cpu != DTRACE_CPUALL) {
1110 		assert(cpu < agp->dtat_ncpu);
1111 		agp->dtat_cpus[agp->dtat_ncpus++] = (processorid_t)cpu;
1112 
1113 		return (0);
1114 	}
1115 
1116 	agp->dtat_ncpus = 0;
1117 	for (i = 0; i < agp->dtat_maxcpu; i++) {
1118 		if (dt_status(dtp, i) == -1)
1119 			continue;
1120 
1121 		agp->dtat_cpus[agp->dtat_ncpus++] = i;
1122 	}
1123 
1124 	return (0);
1125 }
1126 
1127 static int
1128 dt_aggwalk_rval(dtrace_hdl_t *dtp, dt_ahashent_t *h, int rval)
1129 {
1130 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1131 	dtrace_aggdata_t *data;
1132 	dtrace_aggdesc_t *aggdesc;
1133 	dtrace_recdesc_t *rec;
1134 	int i;
1135 
1136 	switch (rval) {
1137 	case DTRACE_AGGWALK_NEXT:
1138 		break;
1139 
1140 	case DTRACE_AGGWALK_CLEAR: {
1141 		uint32_t size, offs = 0;
1142 
1143 		aggdesc = h->dtahe_data.dtada_desc;
1144 		rec = &aggdesc->dtagd_rec[aggdesc->dtagd_nrecs - 1];
1145 		size = rec->dtrd_size;
1146 		data = &h->dtahe_data;
1147 
1148 		if (rec->dtrd_action == DTRACEAGG_LQUANTIZE) {
1149 			offs = sizeof (uint64_t);
1150 			size -= sizeof (uint64_t);
1151 		}
1152 
1153 		bzero(&data->dtada_data[rec->dtrd_offset] + offs, size);
1154 
1155 		if (data->dtada_percpu == NULL)
1156 			break;
1157 
1158 		for (i = 0; i < dtp->dt_aggregate.dtat_maxcpu; i++)
1159 			bzero(data->dtada_percpu[i] + offs, size);
1160 		break;
1161 	}
1162 
1163 	case DTRACE_AGGWALK_ERROR:
1164 		/*
1165 		 * We assume that errno is already set in this case.
1166 		 */
1167 		return (dt_set_errno(dtp, errno));
1168 
1169 	case DTRACE_AGGWALK_ABORT:
1170 		return (dt_set_errno(dtp, EDT_DIRABORT));
1171 
1172 	case DTRACE_AGGWALK_DENORMALIZE:
1173 		h->dtahe_data.dtada_normal = 1;
1174 		return (0);
1175 
1176 	case DTRACE_AGGWALK_NORMALIZE:
1177 		if (h->dtahe_data.dtada_normal == 0) {
1178 			h->dtahe_data.dtada_normal = 1;
1179 			return (dt_set_errno(dtp, EDT_BADRVAL));
1180 		}
1181 
1182 		return (0);
1183 
1184 	case DTRACE_AGGWALK_REMOVE: {
1185 		dtrace_aggdata_t *aggdata = &h->dtahe_data;
1186 		int i, max_cpus = agp->dtat_maxcpu;
1187 
1188 		/*
1189 		 * First, remove this hash entry from its hash chain.
1190 		 */
1191 		if (h->dtahe_prev != NULL) {
1192 			h->dtahe_prev->dtahe_next = h->dtahe_next;
1193 		} else {
1194 			dt_ahash_t *hash = &agp->dtat_hash;
1195 			size_t ndx = h->dtahe_hashval % hash->dtah_size;
1196 
1197 			assert(hash->dtah_hash[ndx] == h);
1198 			hash->dtah_hash[ndx] = h->dtahe_next;
1199 		}
1200 
1201 		if (h->dtahe_next != NULL)
1202 			h->dtahe_next->dtahe_prev = h->dtahe_prev;
1203 
1204 		/*
1205 		 * Now remove it from the list of all hash entries.
1206 		 */
1207 		if (h->dtahe_prevall != NULL) {
1208 			h->dtahe_prevall->dtahe_nextall = h->dtahe_nextall;
1209 		} else {
1210 			dt_ahash_t *hash = &agp->dtat_hash;
1211 
1212 			assert(hash->dtah_all == h);
1213 			hash->dtah_all = h->dtahe_nextall;
1214 		}
1215 
1216 		if (h->dtahe_nextall != NULL)
1217 			h->dtahe_nextall->dtahe_prevall = h->dtahe_prevall;
1218 
1219 		/*
1220 		 * We're unlinked.  We can safely destroy the data.
1221 		 */
1222 		if (aggdata->dtada_percpu != NULL) {
1223 			for (i = 0; i < max_cpus; i++)
1224 				free(aggdata->dtada_percpu[i]);
1225 			free(aggdata->dtada_percpu);
1226 		}
1227 
1228 		free(aggdata->dtada_data);
1229 		free(h);
1230 
1231 		return (0);
1232 	}
1233 
1234 	default:
1235 		return (dt_set_errno(dtp, EDT_BADRVAL));
1236 	}
1237 
1238 	return (0);
1239 }
1240 
1241 void
1242 dt_aggregate_qsort(dtrace_hdl_t *dtp, void *base, size_t nel, size_t width,
1243     int (*compar)(const void *, const void *))
1244 {
1245 	int rev = dt_revsort, key = dt_keysort, keypos = dt_keypos;
1246 	dtrace_optval_t keyposopt = dtp->dt_options[DTRACEOPT_AGGSORTKEYPOS];
1247 
1248 	dt_revsort = (dtp->dt_options[DTRACEOPT_AGGSORTREV] != DTRACEOPT_UNSET);
1249 	dt_keysort = (dtp->dt_options[DTRACEOPT_AGGSORTKEY] != DTRACEOPT_UNSET);
1250 
1251 	if (keyposopt != DTRACEOPT_UNSET && keyposopt <= INT_MAX) {
1252 		dt_keypos = (int)keyposopt;
1253 	} else {
1254 		dt_keypos = 0;
1255 	}
1256 
1257 	if (compar == NULL) {
1258 		if (!dt_keysort) {
1259 			compar = dt_aggregate_varvalcmp;
1260 		} else {
1261 			compar = dt_aggregate_varkeycmp;
1262 		}
1263 	}
1264 
1265 	qsort(base, nel, width, compar);
1266 
1267 	dt_revsort = rev;
1268 	dt_keysort = key;
1269 	dt_keypos = keypos;
1270 }
1271 
1272 int
1273 dtrace_aggregate_walk(dtrace_hdl_t *dtp, dtrace_aggregate_f *func, void *arg)
1274 {
1275 	dt_ahashent_t *h, *next;
1276 	dt_ahash_t *hash = &dtp->dt_aggregate.dtat_hash;
1277 
1278 	for (h = hash->dtah_all; h != NULL; h = next) {
1279 		/*
1280 		 * dt_aggwalk_rval() can potentially remove the current hash
1281 		 * entry; we need to load the next hash entry before calling
1282 		 * into it.
1283 		 */
1284 		next = h->dtahe_nextall;
1285 
1286 		if (dt_aggwalk_rval(dtp, h, func(&h->dtahe_data, arg)) == -1)
1287 			return (-1);
1288 	}
1289 
1290 	return (0);
1291 }
1292 
1293 static int
1294 dt_aggregate_total(dtrace_hdl_t *dtp, boolean_t clear)
1295 {
1296 	dt_ahashent_t *h;
1297 	dtrace_aggdata_t **total;
1298 	dtrace_aggid_t max = DTRACE_AGGVARIDNONE, id;
1299 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1300 	dt_ahash_t *hash = &agp->dtat_hash;
1301 	uint32_t tflags;
1302 
1303 	tflags = DTRACE_A_TOTAL | DTRACE_A_HASNEGATIVES | DTRACE_A_HASPOSITIVES;
1304 
1305 	/*
1306 	 * If we need to deliver per-aggregation totals, we're going to take
1307 	 * three passes over the aggregate:  one to clear everything out and
1308 	 * determine our maximum aggregation ID, one to actually total
1309 	 * everything up, and a final pass to assign the totals to the
1310 	 * individual elements.
1311 	 */
1312 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1313 		dtrace_aggdata_t *aggdata = &h->dtahe_data;
1314 
1315 		if ((id = dt_aggregate_aggvarid(h)) > max)
1316 			max = id;
1317 
1318 		aggdata->dtada_total = 0;
1319 		aggdata->dtada_flags &= ~tflags;
1320 	}
1321 
1322 	if (clear || max == DTRACE_AGGVARIDNONE)
1323 		return (0);
1324 
1325 	total = dt_zalloc(dtp, (max + 1) * sizeof (dtrace_aggdata_t *));
1326 
1327 	if (total == NULL)
1328 		return (-1);
1329 
1330 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1331 		dtrace_aggdata_t *aggdata = &h->dtahe_data;
1332 		dtrace_aggdesc_t *agg = aggdata->dtada_desc;
1333 		dtrace_recdesc_t *rec;
1334 		caddr_t data;
1335 		int64_t val, *addr;
1336 
1337 		rec = &agg->dtagd_rec[agg->dtagd_nrecs - 1];
1338 		data = aggdata->dtada_data;
1339 		addr = (int64_t *)(uintptr_t)(data + rec->dtrd_offset);
1340 
1341 		switch (rec->dtrd_action) {
1342 		case DTRACEAGG_STDDEV:
1343 			val = dt_stddev((uint64_t *)addr, 1);
1344 			break;
1345 
1346 		case DTRACEAGG_SUM:
1347 		case DTRACEAGG_COUNT:
1348 			val = *addr;
1349 			break;
1350 
1351 		case DTRACEAGG_AVG:
1352 			val = addr[0] ? (addr[1] / addr[0]) : 0;
1353 			break;
1354 
1355 		default:
1356 			continue;
1357 		}
1358 
1359 		if (total[agg->dtagd_varid] == NULL) {
1360 			total[agg->dtagd_varid] = aggdata;
1361 			aggdata->dtada_flags |= DTRACE_A_TOTAL;
1362 		} else {
1363 			aggdata = total[agg->dtagd_varid];
1364 		}
1365 
1366 		if (val > 0)
1367 			aggdata->dtada_flags |= DTRACE_A_HASPOSITIVES;
1368 
1369 		if (val < 0) {
1370 			aggdata->dtada_flags |= DTRACE_A_HASNEGATIVES;
1371 			val = -val;
1372 		}
1373 
1374 		if (dtp->dt_options[DTRACEOPT_AGGZOOM] != DTRACEOPT_UNSET) {
1375 			val = (int64_t)((long double)val *
1376 			    (1 / DTRACE_AGGZOOM_MAX));
1377 
1378 			if (val > aggdata->dtada_total)
1379 				aggdata->dtada_total = val;
1380 		} else {
1381 			aggdata->dtada_total += val;
1382 		}
1383 	}
1384 
1385 	/*
1386 	 * And now one final pass to set everyone's total.
1387 	 */
1388 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1389 		dtrace_aggdata_t *aggdata = &h->dtahe_data, *t;
1390 		dtrace_aggdesc_t *agg = aggdata->dtada_desc;
1391 
1392 		if ((t = total[agg->dtagd_varid]) == NULL || aggdata == t)
1393 			continue;
1394 
1395 		aggdata->dtada_total = t->dtada_total;
1396 		aggdata->dtada_flags |= (t->dtada_flags & tflags);
1397 	}
1398 
1399 	dt_free(dtp, total);
1400 
1401 	return (0);
1402 }
1403 
1404 static int
1405 dt_aggregate_minmaxbin(dtrace_hdl_t *dtp, boolean_t clear)
1406 {
1407 	dt_ahashent_t *h;
1408 	dtrace_aggdata_t **minmax;
1409 	dtrace_aggid_t max = DTRACE_AGGVARIDNONE, id;
1410 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1411 	dt_ahash_t *hash = &agp->dtat_hash;
1412 
1413 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1414 		dtrace_aggdata_t *aggdata = &h->dtahe_data;
1415 
1416 		if ((id = dt_aggregate_aggvarid(h)) > max)
1417 			max = id;
1418 
1419 		aggdata->dtada_minbin = 0;
1420 		aggdata->dtada_maxbin = 0;
1421 		aggdata->dtada_flags &= ~DTRACE_A_MINMAXBIN;
1422 	}
1423 
1424 	if (clear || max == DTRACE_AGGVARIDNONE)
1425 		return (0);
1426 
1427 	minmax = dt_zalloc(dtp, (max + 1) * sizeof (dtrace_aggdata_t *));
1428 
1429 	if (minmax == NULL)
1430 		return (-1);
1431 
1432 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1433 		dtrace_aggdata_t *aggdata = &h->dtahe_data;
1434 		dtrace_aggdesc_t *agg = aggdata->dtada_desc;
1435 		dtrace_recdesc_t *rec;
1436 		caddr_t data;
1437 		int64_t *addr;
1438 		int minbin = -1, maxbin = -1, i;
1439 		int start = 0, size;
1440 
1441 		rec = &agg->dtagd_rec[agg->dtagd_nrecs - 1];
1442 		size = rec->dtrd_size / sizeof (int64_t);
1443 		data = aggdata->dtada_data;
1444 		addr = (int64_t *)(uintptr_t)(data + rec->dtrd_offset);
1445 
1446 		switch (rec->dtrd_action) {
1447 		case DTRACEAGG_LQUANTIZE:
1448 			/*
1449 			 * For lquantize(), we always display the entire range
1450 			 * of the aggregation when aggpack is set.
1451 			 */
1452 			start = 1;
1453 			minbin = start;
1454 			maxbin = size - 1 - start;
1455 			break;
1456 
1457 		case DTRACEAGG_QUANTIZE:
1458 			for (i = start; i < size; i++) {
1459 				if (!addr[i])
1460 					continue;
1461 
1462 				if (minbin == -1)
1463 					minbin = i - start;
1464 
1465 				maxbin = i - start;
1466 			}
1467 
1468 			if (minbin == -1) {
1469 				/*
1470 				 * If we have no data (e.g., due to a clear()
1471 				 * or negative increments), we'll use the
1472 				 * zero bucket as both our min and max.
1473 				 */
1474 				minbin = maxbin = DTRACE_QUANTIZE_ZEROBUCKET;
1475 			}
1476 
1477 			break;
1478 
1479 		default:
1480 			continue;
1481 		}
1482 
1483 		if (minmax[agg->dtagd_varid] == NULL) {
1484 			minmax[agg->dtagd_varid] = aggdata;
1485 			aggdata->dtada_flags |= DTRACE_A_MINMAXBIN;
1486 			aggdata->dtada_minbin = minbin;
1487 			aggdata->dtada_maxbin = maxbin;
1488 			continue;
1489 		}
1490 
1491 		if (minbin < minmax[agg->dtagd_varid]->dtada_minbin)
1492 			minmax[agg->dtagd_varid]->dtada_minbin = minbin;
1493 
1494 		if (maxbin > minmax[agg->dtagd_varid]->dtada_maxbin)
1495 			minmax[agg->dtagd_varid]->dtada_maxbin = maxbin;
1496 	}
1497 
1498 	/*
1499 	 * And now one final pass to set everyone's minbin and maxbin.
1500 	 */
1501 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1502 		dtrace_aggdata_t *aggdata = &h->dtahe_data, *mm;
1503 		dtrace_aggdesc_t *agg = aggdata->dtada_desc;
1504 
1505 		if ((mm = minmax[agg->dtagd_varid]) == NULL || aggdata == mm)
1506 			continue;
1507 
1508 		aggdata->dtada_minbin = mm->dtada_minbin;
1509 		aggdata->dtada_maxbin = mm->dtada_maxbin;
1510 		aggdata->dtada_flags |= DTRACE_A_MINMAXBIN;
1511 	}
1512 
1513 	dt_free(dtp, minmax);
1514 
1515 	return (0);
1516 }
1517 
1518 static int
1519 dt_aggregate_walk_sorted(dtrace_hdl_t *dtp,
1520     dtrace_aggregate_f *func, void *arg,
1521     int (*sfunc)(const void *, const void *))
1522 {
1523 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1524 	dt_ahashent_t *h, **sorted;
1525 	dt_ahash_t *hash = &agp->dtat_hash;
1526 	size_t i, nentries = 0;
1527 	int rval = -1;
1528 
1529 	agp->dtat_flags &= ~(DTRACE_A_TOTAL | DTRACE_A_MINMAXBIN);
1530 
1531 	if (dtp->dt_options[DTRACEOPT_AGGHIST] != DTRACEOPT_UNSET) {
1532 		agp->dtat_flags |= DTRACE_A_TOTAL;
1533 
1534 		if (dt_aggregate_total(dtp, B_FALSE) != 0)
1535 			return (-1);
1536 	}
1537 
1538 	if (dtp->dt_options[DTRACEOPT_AGGPACK] != DTRACEOPT_UNSET) {
1539 		agp->dtat_flags |= DTRACE_A_MINMAXBIN;
1540 
1541 		if (dt_aggregate_minmaxbin(dtp, B_FALSE) != 0)
1542 			return (-1);
1543 	}
1544 
1545 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall)
1546 		nentries++;
1547 
1548 	sorted = dt_alloc(dtp, nentries * sizeof (dt_ahashent_t *));
1549 
1550 	if (sorted == NULL)
1551 		goto out;
1552 
1553 	for (h = hash->dtah_all, i = 0; h != NULL; h = h->dtahe_nextall)
1554 		sorted[i++] = h;
1555 
1556 	(void) pthread_mutex_lock(&dt_qsort_lock);
1557 
1558 	if (sfunc == NULL) {
1559 		dt_aggregate_qsort(dtp, sorted, nentries,
1560 		    sizeof (dt_ahashent_t *), NULL);
1561 	} else {
1562 		/*
1563 		 * If we've been explicitly passed a sorting function,
1564 		 * we'll use that -- ignoring the values of the "aggsortrev",
1565 		 * "aggsortkey" and "aggsortkeypos" options.
1566 		 */
1567 		qsort(sorted, nentries, sizeof (dt_ahashent_t *), sfunc);
1568 	}
1569 
1570 	(void) pthread_mutex_unlock(&dt_qsort_lock);
1571 
1572 	for (i = 0; i < nentries; i++) {
1573 		h = sorted[i];
1574 
1575 		if (dt_aggwalk_rval(dtp, h, func(&h->dtahe_data, arg)) == -1)
1576 			goto out;
1577 	}
1578 
1579 	rval = 0;
1580 out:
1581 	if (agp->dtat_flags & DTRACE_A_TOTAL)
1582 		(void) dt_aggregate_total(dtp, B_TRUE);
1583 
1584 	if (agp->dtat_flags & DTRACE_A_MINMAXBIN)
1585 		(void) dt_aggregate_minmaxbin(dtp, B_TRUE);
1586 
1587 	dt_free(dtp, sorted);
1588 	return (rval);
1589 }
1590 
1591 int
1592 dtrace_aggregate_walk_sorted(dtrace_hdl_t *dtp,
1593     dtrace_aggregate_f *func, void *arg)
1594 {
1595 	return (dt_aggregate_walk_sorted(dtp, func, arg, NULL));
1596 }
1597 
1598 int
1599 dtrace_aggregate_walk_keysorted(dtrace_hdl_t *dtp,
1600     dtrace_aggregate_f *func, void *arg)
1601 {
1602 	return (dt_aggregate_walk_sorted(dtp, func,
1603 	    arg, dt_aggregate_varkeycmp));
1604 }
1605 
1606 int
1607 dtrace_aggregate_walk_valsorted(dtrace_hdl_t *dtp,
1608     dtrace_aggregate_f *func, void *arg)
1609 {
1610 	return (dt_aggregate_walk_sorted(dtp, func,
1611 	    arg, dt_aggregate_varvalcmp));
1612 }
1613 
1614 int
1615 dtrace_aggregate_walk_keyvarsorted(dtrace_hdl_t *dtp,
1616     dtrace_aggregate_f *func, void *arg)
1617 {
1618 	return (dt_aggregate_walk_sorted(dtp, func,
1619 	    arg, dt_aggregate_keyvarcmp));
1620 }
1621 
1622 int
1623 dtrace_aggregate_walk_valvarsorted(dtrace_hdl_t *dtp,
1624     dtrace_aggregate_f *func, void *arg)
1625 {
1626 	return (dt_aggregate_walk_sorted(dtp, func,
1627 	    arg, dt_aggregate_valvarcmp));
1628 }
1629 
1630 int
1631 dtrace_aggregate_walk_keyrevsorted(dtrace_hdl_t *dtp,
1632     dtrace_aggregate_f *func, void *arg)
1633 {
1634 	return (dt_aggregate_walk_sorted(dtp, func,
1635 	    arg, dt_aggregate_varkeyrevcmp));
1636 }
1637 
1638 int
1639 dtrace_aggregate_walk_valrevsorted(dtrace_hdl_t *dtp,
1640     dtrace_aggregate_f *func, void *arg)
1641 {
1642 	return (dt_aggregate_walk_sorted(dtp, func,
1643 	    arg, dt_aggregate_varvalrevcmp));
1644 }
1645 
1646 int
1647 dtrace_aggregate_walk_keyvarrevsorted(dtrace_hdl_t *dtp,
1648     dtrace_aggregate_f *func, void *arg)
1649 {
1650 	return (dt_aggregate_walk_sorted(dtp, func,
1651 	    arg, dt_aggregate_keyvarrevcmp));
1652 }
1653 
1654 int
1655 dtrace_aggregate_walk_valvarrevsorted(dtrace_hdl_t *dtp,
1656     dtrace_aggregate_f *func, void *arg)
1657 {
1658 	return (dt_aggregate_walk_sorted(dtp, func,
1659 	    arg, dt_aggregate_valvarrevcmp));
1660 }
1661 
1662 int
1663 dtrace_aggregate_walk_joined(dtrace_hdl_t *dtp, dtrace_aggvarid_t *aggvars,
1664     int naggvars, dtrace_aggregate_walk_joined_f *func, void *arg)
1665 {
1666 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1667 	dt_ahashent_t *h, **sorted = NULL, ***bundle, **nbundle;
1668 	const dtrace_aggdata_t **data;
1669 	dt_ahashent_t *zaggdata = NULL;
1670 	dt_ahash_t *hash = &agp->dtat_hash;
1671 	size_t nentries = 0, nbundles = 0, start, zsize = 0, bundlesize;
1672 	dtrace_aggvarid_t max = 0, aggvar;
1673 	int rval = -1, *map, *remap = NULL;
1674 	int i, j;
1675 	dtrace_optval_t sortpos = dtp->dt_options[DTRACEOPT_AGGSORTPOS];
1676 
1677 	/*
1678 	 * If the sorting position is greater than the number of aggregation
1679 	 * variable IDs, we silently set it to 0.
1680 	 */
1681 	if (sortpos == DTRACEOPT_UNSET || sortpos >= naggvars)
1682 		sortpos = 0;
1683 
1684 	/*
1685 	 * First we need to translate the specified aggregation variable IDs
1686 	 * into a linear map that will allow us to translate an aggregation
1687 	 * variable ID into its position in the specified aggvars.
1688 	 */
1689 	for (i = 0; i < naggvars; i++) {
1690 		if (aggvars[i] == DTRACE_AGGVARIDNONE || aggvars[i] < 0)
1691 			return (dt_set_errno(dtp, EDT_BADAGGVAR));
1692 
1693 		if (aggvars[i] > max)
1694 			max = aggvars[i];
1695 	}
1696 
1697 	if ((map = dt_zalloc(dtp, (max + 1) * sizeof (int))) == NULL)
1698 		return (-1);
1699 
1700 	zaggdata = dt_zalloc(dtp, naggvars * sizeof (dt_ahashent_t));
1701 
1702 	if (zaggdata == NULL)
1703 		goto out;
1704 
1705 	for (i = 0; i < naggvars; i++) {
1706 		int ndx = i + sortpos;
1707 
1708 		if (ndx >= naggvars)
1709 			ndx -= naggvars;
1710 
1711 		aggvar = aggvars[ndx];
1712 		assert(aggvar <= max);
1713 
1714 		if (map[aggvar]) {
1715 			/*
1716 			 * We have an aggregation variable that is present
1717 			 * more than once in the array of aggregation
1718 			 * variables.  While it's unclear why one might want
1719 			 * to do this, it's legal.  To support this construct,
1720 			 * we will allocate a remap that will indicate the
1721 			 * position from which this aggregation variable
1722 			 * should be pulled.  (That is, where the remap will
1723 			 * map from one position to another.)
1724 			 */
1725 			if (remap == NULL) {
1726 				remap = dt_zalloc(dtp, naggvars * sizeof (int));
1727 
1728 				if (remap == NULL)
1729 					goto out;
1730 			}
1731 
1732 			/*
1733 			 * Given that the variable is already present, assert
1734 			 * that following through the mapping and adjusting
1735 			 * for the sort position yields the same aggregation
1736 			 * variable ID.
1737 			 */
1738 			assert(aggvars[(map[aggvar] - 1 + sortpos) %
1739 			    naggvars] == aggvars[ndx]);
1740 
1741 			remap[i] = map[aggvar];
1742 			continue;
1743 		}
1744 
1745 		map[aggvar] = i + 1;
1746 	}
1747 
1748 	/*
1749 	 * We need to take two passes over the data to size our allocation, so
1750 	 * we'll use the first pass to also fill in the zero-filled data to be
1751 	 * used to properly format a zero-valued aggregation.
1752 	 */
1753 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1754 		dtrace_aggvarid_t id;
1755 		int ndx;
1756 
1757 		if ((id = dt_aggregate_aggvarid(h)) > max || !(ndx = map[id]))
1758 			continue;
1759 
1760 		if (zaggdata[ndx - 1].dtahe_size == 0) {
1761 			zaggdata[ndx - 1].dtahe_size = h->dtahe_size;
1762 			zaggdata[ndx - 1].dtahe_data = h->dtahe_data;
1763 		}
1764 
1765 		nentries++;
1766 	}
1767 
1768 	if (nentries == 0) {
1769 		/*
1770 		 * We couldn't find any entries; there is nothing else to do.
1771 		 */
1772 		rval = 0;
1773 		goto out;
1774 	}
1775 
1776 	/*
1777 	 * Before we sort the data, we're going to look for any holes in our
1778 	 * zero-filled data.  This will occur if an aggregation variable that
1779 	 * we are being asked to print has not yet been assigned the result of
1780 	 * any aggregating action for _any_ tuple.  The issue becomes that we
1781 	 * would like a zero value to be printed for all columns for this
1782 	 * aggregation, but without any record description, we don't know the
1783 	 * aggregating action that corresponds to the aggregation variable.  To
1784 	 * try to find a match, we're simply going to lookup aggregation IDs
1785 	 * (which are guaranteed to be contiguous and to start from 1), looking
1786 	 * for the specified aggregation variable ID.  If we find a match,
1787 	 * we'll use that.  If we iterate over all aggregation IDs and don't
1788 	 * find a match, then we must be an anonymous enabling.  (Anonymous
1789 	 * enablings can't currently derive either aggregation variable IDs or
1790 	 * aggregation variable names given only an aggregation ID.)  In this
1791 	 * obscure case (anonymous enabling, multiple aggregation printa() with
1792 	 * some aggregations not represented for any tuple), our defined
1793 	 * behavior is that the zero will be printed in the format of the first
1794 	 * aggregation variable that contains any non-zero value.
1795 	 */
1796 	for (i = 0; i < naggvars; i++) {
1797 		if (zaggdata[i].dtahe_size == 0) {
1798 			dtrace_aggvarid_t aggvar;
1799 
1800 			aggvar = aggvars[(i - sortpos + naggvars) % naggvars];
1801 			assert(zaggdata[i].dtahe_data.dtada_data == NULL);
1802 
1803 			for (j = DTRACE_AGGIDNONE + 1; ; j++) {
1804 				dtrace_aggdesc_t *agg;
1805 				dtrace_aggdata_t *aggdata;
1806 
1807 				if (dt_aggid_lookup(dtp, j, &agg) != 0)
1808 					break;
1809 
1810 				if (agg->dtagd_varid != aggvar)
1811 					continue;
1812 
1813 				/*
1814 				 * We have our description -- now we need to
1815 				 * cons up the zaggdata entry for it.
1816 				 */
1817 				aggdata = &zaggdata[i].dtahe_data;
1818 				aggdata->dtada_size = agg->dtagd_size;
1819 				aggdata->dtada_desc = agg;
1820 				aggdata->dtada_handle = dtp;
1821 				(void) dt_epid_lookup(dtp, agg->dtagd_epid,
1822 				    &aggdata->dtada_edesc,
1823 				    &aggdata->dtada_pdesc);
1824 				aggdata->dtada_normal = 1;
1825 				zaggdata[i].dtahe_hashval = 0;
1826 				zaggdata[i].dtahe_size = agg->dtagd_size;
1827 				break;
1828 			}
1829 
1830 			if (zaggdata[i].dtahe_size == 0) {
1831 				caddr_t data;
1832 
1833 				/*
1834 				 * We couldn't find this aggregation, meaning
1835 				 * that we have never seen it before for any
1836 				 * tuple _and_ this is an anonymous enabling.
1837 				 * That is, we're in the obscure case outlined
1838 				 * above.  In this case, our defined behavior
1839 				 * is to format the data in the format of the
1840 				 * first non-zero aggregation -- of which, of
1841 				 * course, we know there to be at least one
1842 				 * (or nentries would have been zero).
1843 				 */
1844 				for (j = 0; j < naggvars; j++) {
1845 					if (zaggdata[j].dtahe_size != 0)
1846 						break;
1847 				}
1848 
1849 				assert(j < naggvars);
1850 				zaggdata[i] = zaggdata[j];
1851 
1852 				data = zaggdata[i].dtahe_data.dtada_data;
1853 				assert(data != NULL);
1854 			}
1855 		}
1856 	}
1857 
1858 	/*
1859 	 * Now we need to allocate our zero-filled data for use for
1860 	 * aggregations that don't have a value corresponding to a given key.
1861 	 */
1862 	for (i = 0; i < naggvars; i++) {
1863 		dtrace_aggdata_t *aggdata = &zaggdata[i].dtahe_data;
1864 		dtrace_aggdesc_t *aggdesc = aggdata->dtada_desc;
1865 		dtrace_recdesc_t *rec;
1866 		uint64_t larg;
1867 		caddr_t zdata;
1868 
1869 		zsize = zaggdata[i].dtahe_size;
1870 		assert(zsize != 0);
1871 
1872 		if ((zdata = dt_zalloc(dtp, zsize)) == NULL) {
1873 			/*
1874 			 * If we failed to allocated some zero-filled data, we
1875 			 * need to zero out the remaining dtada_data pointers
1876 			 * to prevent the wrong data from being freed below.
1877 			 */
1878 			for (j = i; j < naggvars; j++)
1879 				zaggdata[j].dtahe_data.dtada_data = NULL;
1880 			goto out;
1881 		}
1882 
1883 		aggvar = aggvars[(i - sortpos + naggvars) % naggvars];
1884 
1885 		/*
1886 		 * First, the easy bit.  To maintain compatibility with
1887 		 * consumers that pull the compiler-generated ID out of the
1888 		 * data, we put that ID at the top of the zero-filled data.
1889 		 */
1890 		rec = &aggdesc->dtagd_rec[0];
1891 		/* LINTED - alignment */
1892 		*((dtrace_aggvarid_t *)(zdata + rec->dtrd_offset)) = aggvar;
1893 
1894 		rec = &aggdesc->dtagd_rec[aggdesc->dtagd_nrecs - 1];
1895 
1896 		/*
1897 		 * Now for the more complicated part.  If (and only if) this
1898 		 * is an lquantize() aggregating action, zero-filled data is
1899 		 * not equivalent to an empty record:  we must also get the
1900 		 * parameters for the lquantize().
1901 		 */
1902 		if (rec->dtrd_action == DTRACEAGG_LQUANTIZE) {
1903 			if (aggdata->dtada_data != NULL) {
1904 				/*
1905 				 * The easier case here is if we actually have
1906 				 * some prototype data -- in which case we
1907 				 * manually dig it out of the aggregation
1908 				 * record.
1909 				 */
1910 				/* LINTED - alignment */
1911 				larg = *((uint64_t *)(aggdata->dtada_data +
1912 				    rec->dtrd_offset));
1913 			} else {
1914 				/*
1915 				 * We don't have any prototype data.  As a
1916 				 * result, we know that we _do_ have the
1917 				 * compiler-generated information.  (If this
1918 				 * were an anonymous enabling, all of our
1919 				 * zero-filled data would have prototype data
1920 				 * -- either directly or indirectly.) So as
1921 				 * gross as it is, we'll grovel around in the
1922 				 * compiler-generated information to find the
1923 				 * lquantize() parameters.
1924 				 */
1925 				dtrace_stmtdesc_t *sdp;
1926 				dt_ident_t *aid;
1927 				dt_idsig_t *isp;
1928 
1929 				sdp = (dtrace_stmtdesc_t *)(uintptr_t)
1930 				    aggdesc->dtagd_rec[0].dtrd_uarg;
1931 				aid = sdp->dtsd_aggdata;
1932 				isp = (dt_idsig_t *)aid->di_data;
1933 				assert(isp->dis_auxinfo != 0);
1934 				larg = isp->dis_auxinfo;
1935 			}
1936 
1937 			/* LINTED - alignment */
1938 			*((uint64_t *)(zdata + rec->dtrd_offset)) = larg;
1939 		}
1940 
1941 		aggdata->dtada_data = zdata;
1942 	}
1943 
1944 	/*
1945 	 * Now that we've dealt with setting up our zero-filled data, we can
1946 	 * allocate our sorted array, and take another pass over the data to
1947 	 * fill it.
1948 	 */
1949 	sorted = dt_alloc(dtp, nentries * sizeof (dt_ahashent_t *));
1950 
1951 	if (sorted == NULL)
1952 		goto out;
1953 
1954 	for (h = hash->dtah_all, i = 0; h != NULL; h = h->dtahe_nextall) {
1955 		dtrace_aggvarid_t id;
1956 
1957 		if ((id = dt_aggregate_aggvarid(h)) > max || !map[id])
1958 			continue;
1959 
1960 		sorted[i++] = h;
1961 	}
1962 
1963 	assert(i == nentries);
1964 
1965 	/*
1966 	 * We've loaded our array; now we need to sort by value to allow us
1967 	 * to create bundles of like value.  We're going to acquire the
1968 	 * dt_qsort_lock here, and hold it across all of our subsequent
1969 	 * comparison and sorting.
1970 	 */
1971 	(void) pthread_mutex_lock(&dt_qsort_lock);
1972 
1973 	qsort(sorted, nentries, sizeof (dt_ahashent_t *),
1974 	    dt_aggregate_keyvarcmp);
1975 
1976 	/*
1977 	 * Now we need to go through and create bundles.  Because the number
1978 	 * of bundles is bounded by the size of the sorted array, we're going
1979 	 * to reuse the underlying storage.  And note that "bundle" is an
1980 	 * array of pointers to arrays of pointers to dt_ahashent_t -- making
1981 	 * its type (regrettably) "dt_ahashent_t ***".  (Regrettable because
1982 	 * '*' -- like '_' and 'X' -- should never appear in triplicate in
1983 	 * an ideal world.)
1984 	 */
1985 	bundle = (dt_ahashent_t ***)sorted;
1986 
1987 	for (i = 1, start = 0; i <= nentries; i++) {
1988 		if (i < nentries &&
1989 		    dt_aggregate_keycmp(&sorted[i], &sorted[i - 1]) == 0)
1990 			continue;
1991 
1992 		/*
1993 		 * We have a bundle boundary.  Everything from start to
1994 		 * (i - 1) belongs in one bundle.
1995 		 */
1996 		assert(i - start <= naggvars);
1997 		bundlesize = (naggvars + 2) * sizeof (dt_ahashent_t *);
1998 
1999 		if ((nbundle = dt_zalloc(dtp, bundlesize)) == NULL) {
2000 			(void) pthread_mutex_unlock(&dt_qsort_lock);
2001 			goto out;
2002 		}
2003 
2004 		for (j = start; j < i; j++) {
2005 			dtrace_aggvarid_t id = dt_aggregate_aggvarid(sorted[j]);
2006 
2007 			assert(id <= max);
2008 			assert(map[id] != 0);
2009 			assert(map[id] - 1 < naggvars);
2010 			assert(nbundle[map[id] - 1] == NULL);
2011 			nbundle[map[id] - 1] = sorted[j];
2012 
2013 			if (nbundle[naggvars] == NULL)
2014 				nbundle[naggvars] = sorted[j];
2015 		}
2016 
2017 		for (j = 0; j < naggvars; j++) {
2018 			if (nbundle[j] != NULL)
2019 				continue;
2020 
2021 			/*
2022 			 * Before we assume that this aggregation variable
2023 			 * isn't present (and fall back to using the
2024 			 * zero-filled data allocated earlier), check the
2025 			 * remap.  If we have a remapping, we'll drop it in
2026 			 * here.  Note that we might be remapping an
2027 			 * aggregation variable that isn't present for this
2028 			 * key; in this case, the aggregation data that we
2029 			 * copy will point to the zeroed data.
2030 			 */
2031 			if (remap != NULL && remap[j]) {
2032 				assert(remap[j] - 1 < j);
2033 				assert(nbundle[remap[j] - 1] != NULL);
2034 				nbundle[j] = nbundle[remap[j] - 1];
2035 			} else {
2036 				nbundle[j] = &zaggdata[j];
2037 			}
2038 		}
2039 
2040 		bundle[nbundles++] = nbundle;
2041 		start = i;
2042 	}
2043 
2044 	/*
2045 	 * Now we need to re-sort based on the first value.
2046 	 */
2047 	dt_aggregate_qsort(dtp, bundle, nbundles, sizeof (dt_ahashent_t **),
2048 	    dt_aggregate_bundlecmp);
2049 
2050 	(void) pthread_mutex_unlock(&dt_qsort_lock);
2051 
2052 	/*
2053 	 * We're done!  Now we just need to go back over the sorted bundles,
2054 	 * calling the function.
2055 	 */
2056 	data = alloca((naggvars + 1) * sizeof (dtrace_aggdata_t *));
2057 
2058 	for (i = 0; i < nbundles; i++) {
2059 		for (j = 0; j < naggvars; j++)
2060 			data[j + 1] = NULL;
2061 
2062 		for (j = 0; j < naggvars; j++) {
2063 			int ndx = j - sortpos;
2064 
2065 			if (ndx < 0)
2066 				ndx += naggvars;
2067 
2068 			assert(bundle[i][ndx] != NULL);
2069 			data[j + 1] = &bundle[i][ndx]->dtahe_data;
2070 		}
2071 
2072 		for (j = 0; j < naggvars; j++)
2073 			assert(data[j + 1] != NULL);
2074 
2075 		/*
2076 		 * The representative key is the last element in the bundle.
2077 		 * Assert that we have one, and then set it to be the first
2078 		 * element of data.
2079 		 */
2080 		assert(bundle[i][j] != NULL);
2081 		data[0] = &bundle[i][j]->dtahe_data;
2082 
2083 		if ((rval = func(data, naggvars + 1, arg)) == -1)
2084 			goto out;
2085 	}
2086 
2087 	rval = 0;
2088 out:
2089 	for (i = 0; i < nbundles; i++)
2090 		dt_free(dtp, bundle[i]);
2091 
2092 	if (zaggdata != NULL) {
2093 		for (i = 0; i < naggvars; i++)
2094 			dt_free(dtp, zaggdata[i].dtahe_data.dtada_data);
2095 	}
2096 
2097 	dt_free(dtp, zaggdata);
2098 	dt_free(dtp, sorted);
2099 	dt_free(dtp, remap);
2100 	dt_free(dtp, map);
2101 
2102 	return (rval);
2103 }
2104 
2105 int
2106 dtrace_aggregate_print(dtrace_hdl_t *dtp, FILE *fp,
2107     dtrace_aggregate_walk_f *func)
2108 {
2109 	dt_print_aggdata_t pd;
2110 
2111 	bzero(&pd, sizeof (pd));
2112 
2113 	pd.dtpa_dtp = dtp;
2114 	pd.dtpa_fp = fp;
2115 	pd.dtpa_allunprint = 1;
2116 
2117 	if (func == NULL)
2118 		func = dtrace_aggregate_walk_sorted;
2119 
2120 	if ((*func)(dtp, dt_print_agg, &pd) == -1)
2121 		return (dt_set_errno(dtp, dtp->dt_errno));
2122 
2123 	return (0);
2124 }
2125 
2126 void
2127 dtrace_aggregate_clear(dtrace_hdl_t *dtp)
2128 {
2129 	dt_aggregate_t *agp = &dtp->dt_aggregate;
2130 	dt_ahash_t *hash = &agp->dtat_hash;
2131 	dt_ahashent_t *h;
2132 	dtrace_aggdata_t *data;
2133 	dtrace_aggdesc_t *aggdesc;
2134 	dtrace_recdesc_t *rec;
2135 	int i, max_cpus = agp->dtat_maxcpu;
2136 
2137 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
2138 		aggdesc = h->dtahe_data.dtada_desc;
2139 		rec = &aggdesc->dtagd_rec[aggdesc->dtagd_nrecs - 1];
2140 		data = &h->dtahe_data;
2141 
2142 		bzero(&data->dtada_data[rec->dtrd_offset], rec->dtrd_size);
2143 
2144 		if (data->dtada_percpu == NULL)
2145 			continue;
2146 
2147 		for (i = 0; i < max_cpus; i++)
2148 			bzero(data->dtada_percpu[i], rec->dtrd_size);
2149 	}
2150 }
2151 
2152 void
2153 dt_aggregate_destroy(dtrace_hdl_t *dtp)
2154 {
2155 	dt_aggregate_t *agp = &dtp->dt_aggregate;
2156 	dt_ahash_t *hash = &agp->dtat_hash;
2157 	dt_ahashent_t *h, *next;
2158 	dtrace_aggdata_t *aggdata;
2159 	int i, max_cpus = agp->dtat_maxcpu;
2160 
2161 	if (hash->dtah_hash == NULL) {
2162 		assert(hash->dtah_all == NULL);
2163 	} else {
2164 		free(hash->dtah_hash);
2165 
2166 		for (h = hash->dtah_all; h != NULL; h = next) {
2167 			next = h->dtahe_nextall;
2168 
2169 			aggdata = &h->dtahe_data;
2170 
2171 			if (aggdata->dtada_percpu != NULL) {
2172 				for (i = 0; i < max_cpus; i++)
2173 					free(aggdata->dtada_percpu[i]);
2174 				free(aggdata->dtada_percpu);
2175 			}
2176 
2177 			free(aggdata->dtada_data);
2178 			free(h);
2179 		}
2180 
2181 		hash->dtah_hash = NULL;
2182 		hash->dtah_all = NULL;
2183 		hash->dtah_size = 0;
2184 	}
2185 
2186 	free(agp->dtat_buf.dtbd_data);
2187 	free(agp->dtat_cpus);
2188 }
2189