xref: /freebsd/cddl/contrib/opensolaris/lib/libdtrace/common/dt_aggregate.c (revision 9a30c8d347bf9aaa89277b6e5a275f737be8edce)
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 <dt_oformat.h>
39 #ifdef illumos
40 #include <alloca.h>
41 #else
42 #include <sys/sysctl.h>
43 #include <libproc_compat.h>
44 #endif
45 #include <limits.h>
46 
47 #define	DTRACE_AHASHSIZE	32779		/* big 'ol prime */
48 
49 /*
50  * Because qsort(3C) does not allow an argument to be passed to a comparison
51  * function, the variables that affect comparison must regrettably be global;
52  * they are protected by a global static lock, dt_qsort_lock.
53  */
54 static pthread_mutex_t dt_qsort_lock = PTHREAD_MUTEX_INITIALIZER;
55 
56 static int dt_revsort;
57 static int dt_keysort;
58 static int dt_keypos;
59 
60 #define	DT_LESSTHAN	(dt_revsort == 0 ? -1 : 1)
61 #define	DT_GREATERTHAN	(dt_revsort == 0 ? 1 : -1)
62 
63 static void
dt_aggregate_count(int64_t * existing,int64_t * new,size_t size)64 dt_aggregate_count(int64_t *existing, int64_t *new, size_t size)
65 {
66 	uint_t i;
67 
68 	for (i = 0; i < size / sizeof (int64_t); i++)
69 		existing[i] = existing[i] + new[i];
70 }
71 
72 static int
dt_aggregate_countcmp(int64_t * lhs,int64_t * rhs)73 dt_aggregate_countcmp(int64_t *lhs, int64_t *rhs)
74 {
75 	int64_t lvar = *lhs;
76 	int64_t rvar = *rhs;
77 
78 	if (lvar < rvar)
79 		return (DT_LESSTHAN);
80 
81 	if (lvar > rvar)
82 		return (DT_GREATERTHAN);
83 
84 	return (0);
85 }
86 
87 /*ARGSUSED*/
88 static void
dt_aggregate_min(int64_t * existing,int64_t * new,size_t size)89 dt_aggregate_min(int64_t *existing, int64_t *new, size_t size)
90 {
91 	if (*new < *existing)
92 		*existing = *new;
93 }
94 
95 /*ARGSUSED*/
96 static void
dt_aggregate_max(int64_t * existing,int64_t * new,size_t size)97 dt_aggregate_max(int64_t *existing, int64_t *new, size_t size)
98 {
99 	if (*new > *existing)
100 		*existing = *new;
101 }
102 
103 static int
dt_aggregate_averagecmp(int64_t * lhs,int64_t * rhs)104 dt_aggregate_averagecmp(int64_t *lhs, int64_t *rhs)
105 {
106 	int64_t lavg = lhs[0] ? (lhs[1] / lhs[0]) : 0;
107 	int64_t ravg = rhs[0] ? (rhs[1] / rhs[0]) : 0;
108 
109 	if (lavg < ravg)
110 		return (DT_LESSTHAN);
111 
112 	if (lavg > ravg)
113 		return (DT_GREATERTHAN);
114 
115 	return (0);
116 }
117 
118 static int
dt_aggregate_stddevcmp(int64_t * lhs,int64_t * rhs)119 dt_aggregate_stddevcmp(int64_t *lhs, int64_t *rhs)
120 {
121 	uint64_t lsd = dt_stddev((uint64_t *)lhs, 1);
122 	uint64_t rsd = dt_stddev((uint64_t *)rhs, 1);
123 
124 	if (lsd < rsd)
125 		return (DT_LESSTHAN);
126 
127 	if (lsd > rsd)
128 		return (DT_GREATERTHAN);
129 
130 	return (0);
131 }
132 
133 /*ARGSUSED*/
134 static void
dt_aggregate_lquantize(int64_t * existing,int64_t * new,size_t size)135 dt_aggregate_lquantize(int64_t *existing, int64_t *new, size_t size)
136 {
137 	int64_t arg = *existing++;
138 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
139 	int i;
140 
141 	for (i = 0; i <= levels + 1; i++)
142 		existing[i] = existing[i] + new[i + 1];
143 }
144 
145 static long double
dt_aggregate_lquantizedsum(int64_t * lquanta)146 dt_aggregate_lquantizedsum(int64_t *lquanta)
147 {
148 	int64_t arg = *lquanta++;
149 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
150 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
151 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg), i;
152 	long double total = (long double)lquanta[0] * (long double)(base - 1);
153 
154 	for (i = 0; i < levels; base += step, i++)
155 		total += (long double)lquanta[i + 1] * (long double)base;
156 
157 	return (total + (long double)lquanta[levels + 1] *
158 	    (long double)(base + 1));
159 }
160 
161 static int64_t
dt_aggregate_lquantizedzero(int64_t * lquanta)162 dt_aggregate_lquantizedzero(int64_t *lquanta)
163 {
164 	int64_t arg = *lquanta++;
165 	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
166 	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
167 	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg), i;
168 
169 	if (base - 1 == 0)
170 		return (lquanta[0]);
171 
172 	for (i = 0; i < levels; base += step, i++) {
173 		if (base != 0)
174 			continue;
175 
176 		return (lquanta[i + 1]);
177 	}
178 
179 	if (base + 1 == 0)
180 		return (lquanta[levels + 1]);
181 
182 	return (0);
183 }
184 
185 static int
dt_aggregate_lquantizedcmp(int64_t * lhs,int64_t * rhs)186 dt_aggregate_lquantizedcmp(int64_t *lhs, int64_t *rhs)
187 {
188 	long double lsum = dt_aggregate_lquantizedsum(lhs);
189 	long double rsum = dt_aggregate_lquantizedsum(rhs);
190 	int64_t lzero, rzero;
191 
192 	if (lsum < rsum)
193 		return (DT_LESSTHAN);
194 
195 	if (lsum > rsum)
196 		return (DT_GREATERTHAN);
197 
198 	/*
199 	 * If they're both equal, then we will compare based on the weights at
200 	 * zero.  If the weights at zero are equal (or if zero is not within
201 	 * the range of the linear quantization), then this will be judged a
202 	 * tie and will be resolved based on the key comparison.
203 	 */
204 	lzero = dt_aggregate_lquantizedzero(lhs);
205 	rzero = dt_aggregate_lquantizedzero(rhs);
206 
207 	if (lzero < rzero)
208 		return (DT_LESSTHAN);
209 
210 	if (lzero > rzero)
211 		return (DT_GREATERTHAN);
212 
213 	return (0);
214 }
215 
216 static void
dt_aggregate_llquantize(int64_t * existing,int64_t * new,size_t size)217 dt_aggregate_llquantize(int64_t *existing, int64_t *new, size_t size)
218 {
219 	int i;
220 
221 	for (i = 1; i < size / sizeof (int64_t); i++)
222 		existing[i] = existing[i] + new[i];
223 }
224 
225 static long double
dt_aggregate_llquantizedsum(int64_t * llquanta)226 dt_aggregate_llquantizedsum(int64_t *llquanta)
227 {
228 	int64_t arg = *llquanta++;
229 	uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
230 	uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
231 	uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
232 	uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
233 	int bin = 0, order;
234 	int64_t value = 1, next, step;
235 	long double total;
236 
237 	assert(nsteps >= factor);
238 	assert(nsteps % factor == 0);
239 
240 	for (order = 0; order < low; order++)
241 		value *= factor;
242 
243 	total = (long double)llquanta[bin++] * (long double)(value - 1);
244 
245 	next = value * factor;
246 	step = next > nsteps ? next / nsteps : 1;
247 
248 	while (order <= high) {
249 		assert(value < next);
250 		total += (long double)llquanta[bin++] * (long double)(value);
251 
252 		if ((value += step) != next)
253 			continue;
254 
255 		next = value * factor;
256 		step = next > nsteps ? next / nsteps : 1;
257 		order++;
258 	}
259 
260 	return (total + (long double)llquanta[bin] * (long double)value);
261 }
262 
263 static int
dt_aggregate_llquantizedcmp(int64_t * lhs,int64_t * rhs)264 dt_aggregate_llquantizedcmp(int64_t *lhs, int64_t *rhs)
265 {
266 	long double lsum = dt_aggregate_llquantizedsum(lhs);
267 	long double rsum = dt_aggregate_llquantizedsum(rhs);
268 	int64_t lzero, rzero;
269 
270 	if (lsum < rsum)
271 		return (DT_LESSTHAN);
272 
273 	if (lsum > rsum)
274 		return (DT_GREATERTHAN);
275 
276 	/*
277 	 * If they're both equal, then we will compare based on the weights at
278 	 * zero.  If the weights at zero are equal, then this will be judged a
279 	 * tie and will be resolved based on the key comparison.
280 	 */
281 	lzero = lhs[1];
282 	rzero = rhs[1];
283 
284 	if (lzero < rzero)
285 		return (DT_LESSTHAN);
286 
287 	if (lzero > rzero)
288 		return (DT_GREATERTHAN);
289 
290 	return (0);
291 }
292 
293 static int
dt_aggregate_quantizedcmp(int64_t * lhs,int64_t * rhs)294 dt_aggregate_quantizedcmp(int64_t *lhs, int64_t *rhs)
295 {
296 	int nbuckets = DTRACE_QUANTIZE_NBUCKETS;
297 	long double ltotal = 0, rtotal = 0;
298 	int64_t lzero, rzero;
299 	uint_t i;
300 
301 	for (i = 0; i < nbuckets; i++) {
302 		int64_t bucketval = DTRACE_QUANTIZE_BUCKETVAL(i);
303 
304 		if (bucketval == 0) {
305 			lzero = lhs[i];
306 			rzero = rhs[i];
307 		}
308 
309 		ltotal += (long double)bucketval * (long double)lhs[i];
310 		rtotal += (long double)bucketval * (long double)rhs[i];
311 	}
312 
313 	if (ltotal < rtotal)
314 		return (DT_LESSTHAN);
315 
316 	if (ltotal > rtotal)
317 		return (DT_GREATERTHAN);
318 
319 	/*
320 	 * If they're both equal, then we will compare based on the weights at
321 	 * zero.  If the weights at zero are equal, then this will be judged a
322 	 * tie and will be resolved based on the key comparison.
323 	 */
324 	if (lzero < rzero)
325 		return (DT_LESSTHAN);
326 
327 	if (lzero > rzero)
328 		return (DT_GREATERTHAN);
329 
330 	return (0);
331 }
332 
333 static void
dt_aggregate_usym(dtrace_hdl_t * dtp,uint64_t * data)334 dt_aggregate_usym(dtrace_hdl_t *dtp, uint64_t *data)
335 {
336 	uint64_t pid = data[0];
337 	uint64_t *pc = &data[1];
338 	struct ps_prochandle *P;
339 	GElf_Sym sym;
340 
341 	if (dtp->dt_vector != NULL)
342 		return;
343 
344 	if ((P = dt_proc_grab(dtp, pid, PGRAB_RDONLY | PGRAB_FORCE, 0)) == NULL)
345 		return;
346 
347 	dt_proc_lock(dtp, P);
348 
349 	if (Plookup_by_addr(P, *pc, NULL, 0, &sym) == 0)
350 		*pc = sym.st_value;
351 
352 	dt_proc_unlock(dtp, P);
353 	dt_proc_release(dtp, P);
354 }
355 
356 static void
dt_aggregate_umod(dtrace_hdl_t * dtp,uint64_t * data)357 dt_aggregate_umod(dtrace_hdl_t *dtp, uint64_t *data)
358 {
359 	uint64_t pid = data[0];
360 	uint64_t *pc = &data[1];
361 	struct ps_prochandle *P;
362 	const prmap_t *map;
363 
364 	if (dtp->dt_vector != NULL)
365 		return;
366 
367 	if ((P = dt_proc_grab(dtp, pid, PGRAB_RDONLY | PGRAB_FORCE, 0)) == NULL)
368 		return;
369 
370 	dt_proc_lock(dtp, P);
371 
372 	if ((map = Paddr_to_map(P, *pc)) != NULL)
373 		*pc = map->pr_vaddr;
374 
375 	dt_proc_unlock(dtp, P);
376 	dt_proc_release(dtp, P);
377 }
378 
379 static void
dt_aggregate_sym(dtrace_hdl_t * dtp,uint64_t * data)380 dt_aggregate_sym(dtrace_hdl_t *dtp, uint64_t *data)
381 {
382 	GElf_Sym sym;
383 	uint64_t *pc = data;
384 
385 	if (dtrace_lookup_by_addr(dtp, *pc, &sym, NULL) == 0)
386 		*pc = sym.st_value;
387 }
388 
389 static void
dt_aggregate_mod(dtrace_hdl_t * dtp,uint64_t * data)390 dt_aggregate_mod(dtrace_hdl_t *dtp, uint64_t *data)
391 {
392 	uint64_t *pc = data;
393 	dt_module_t *dmp;
394 
395 	if (dtp->dt_vector != NULL) {
396 		/*
397 		 * We don't have a way of just getting the module for a
398 		 * vectored open, and it doesn't seem to be worth defining
399 		 * one.  This means that use of mod() won't get true
400 		 * aggregation in the postmortem case (some modules may
401 		 * appear more than once in aggregation output).  It seems
402 		 * unlikely that anyone will ever notice or care...
403 		 */
404 		return;
405 	}
406 
407 	for (dmp = dt_list_next(&dtp->dt_modlist); dmp != NULL;
408 	    dmp = dt_list_next(dmp)) {
409 		if (*pc - dmp->dm_text_va < dmp->dm_text_size) {
410 			*pc = dmp->dm_text_va;
411 			return;
412 		}
413 	}
414 }
415 
416 static dtrace_aggvarid_t
dt_aggregate_aggvarid(dt_ahashent_t * ent)417 dt_aggregate_aggvarid(dt_ahashent_t *ent)
418 {
419 	dtrace_aggdesc_t *agg = ent->dtahe_data.dtada_desc;
420 	caddr_t data = ent->dtahe_data.dtada_data;
421 	dtrace_recdesc_t *rec = agg->dtagd_rec;
422 
423 	/*
424 	 * First, we'll check the variable ID in the aggdesc.  If it's valid,
425 	 * we'll return it.  If not, we'll use the compiler-generated ID
426 	 * present as the first record.
427 	 */
428 	if (agg->dtagd_varid != DTRACE_AGGVARIDNONE)
429 		return (agg->dtagd_varid);
430 
431 	agg->dtagd_varid = *((dtrace_aggvarid_t *)(uintptr_t)(data +
432 	    rec->dtrd_offset));
433 
434 	return (agg->dtagd_varid);
435 }
436 
437 
438 static int
dt_aggregate_snap_cpu(dtrace_hdl_t * dtp,processorid_t cpu)439 dt_aggregate_snap_cpu(dtrace_hdl_t *dtp, processorid_t cpu)
440 {
441 	dtrace_epid_t id;
442 	uint64_t hashval;
443 	size_t offs, roffs, size, ndx;
444 	int i, j, rval;
445 	caddr_t addr, data;
446 	dtrace_recdesc_t *rec;
447 	dt_aggregate_t *agp = &dtp->dt_aggregate;
448 	dtrace_aggdesc_t *agg;
449 	dt_ahash_t *hash = &agp->dtat_hash;
450 	dt_ahashent_t *h;
451 	dtrace_bufdesc_t b = agp->dtat_buf, *buf = &b;
452 	dtrace_aggdata_t *aggdata;
453 	int flags = agp->dtat_flags;
454 
455 	buf->dtbd_cpu = cpu;
456 
457 #ifdef illumos
458 	if (dt_ioctl(dtp, DTRACEIOC_AGGSNAP, buf) == -1) {
459 #else
460 	if (dt_ioctl(dtp, DTRACEIOC_AGGSNAP, &buf) == -1) {
461 #endif
462 		if (errno == ENOENT) {
463 			/*
464 			 * If that failed with ENOENT, it may be because the
465 			 * CPU was unconfigured.  This is okay; we'll just
466 			 * do nothing but return success.
467 			 */
468 			return (0);
469 		}
470 
471 		return (dt_set_errno(dtp, errno));
472 	}
473 
474 	if (buf->dtbd_drops != 0) {
475 		xo_open_instance("probes");
476 		dt_oformat_drop(dtp, cpu);
477 		if (dt_handle_cpudrop(dtp, cpu,
478 		    DTRACEDROP_AGGREGATION, buf->dtbd_drops) == -1) {
479 			xo_close_instance("probes");
480 			return (-1);
481 		}
482 		xo_close_instance("probes");
483 	}
484 
485 	if (buf->dtbd_size == 0)
486 		return (0);
487 
488 	if (hash->dtah_hash == NULL) {
489 		size_t size;
490 
491 		hash->dtah_size = DTRACE_AHASHSIZE;
492 		size = hash->dtah_size * sizeof (dt_ahashent_t *);
493 
494 		if ((hash->dtah_hash = malloc(size)) == NULL)
495 			return (dt_set_errno(dtp, EDT_NOMEM));
496 
497 		bzero(hash->dtah_hash, size);
498 	}
499 
500 	for (offs = 0; offs < buf->dtbd_size; ) {
501 		/*
502 		 * We're guaranteed to have an ID.
503 		 */
504 		id = *((dtrace_epid_t *)((uintptr_t)buf->dtbd_data +
505 		    (uintptr_t)offs));
506 
507 		if (id == DTRACE_AGGIDNONE) {
508 			/*
509 			 * This is filler to assure proper alignment of the
510 			 * next record; we simply ignore it.
511 			 */
512 			offs += sizeof (id);
513 			continue;
514 		}
515 
516 		if ((rval = dt_aggid_lookup(dtp, id, &agg)) != 0)
517 			return (rval);
518 
519 		addr = buf->dtbd_data + offs;
520 		size = agg->dtagd_size;
521 		hashval = 0;
522 
523 		for (j = 0; j < agg->dtagd_nrecs - 1; j++) {
524 			rec = &agg->dtagd_rec[j];
525 			roffs = rec->dtrd_offset;
526 
527 			switch (rec->dtrd_action) {
528 			case DTRACEACT_USYM:
529 				dt_aggregate_usym(dtp,
530 				    /* LINTED - alignment */
531 				    (uint64_t *)&addr[roffs]);
532 				break;
533 
534 			case DTRACEACT_UMOD:
535 				dt_aggregate_umod(dtp,
536 				    /* LINTED - alignment */
537 				    (uint64_t *)&addr[roffs]);
538 				break;
539 
540 			case DTRACEACT_SYM:
541 				/* LINTED - alignment */
542 				dt_aggregate_sym(dtp, (uint64_t *)&addr[roffs]);
543 				break;
544 
545 			case DTRACEACT_MOD:
546 				/* LINTED - alignment */
547 				dt_aggregate_mod(dtp, (uint64_t *)&addr[roffs]);
548 				break;
549 
550 			default:
551 				break;
552 			}
553 
554 			for (i = 0; i < rec->dtrd_size; i++)
555 				hashval += addr[roffs + i];
556 		}
557 
558 		ndx = hashval % hash->dtah_size;
559 
560 		for (h = hash->dtah_hash[ndx]; h != NULL; h = h->dtahe_next) {
561 			if (h->dtahe_hashval != hashval)
562 				continue;
563 
564 			if (h->dtahe_size != size)
565 				continue;
566 
567 			aggdata = &h->dtahe_data;
568 			data = aggdata->dtada_data;
569 
570 			for (j = 0; j < agg->dtagd_nrecs - 1; j++) {
571 				rec = &agg->dtagd_rec[j];
572 				roffs = rec->dtrd_offset;
573 
574 				for (i = 0; i < rec->dtrd_size; i++)
575 					if (addr[roffs + i] != data[roffs + i])
576 						goto hashnext;
577 			}
578 
579 			/*
580 			 * We found it.  Now we need to apply the aggregating
581 			 * action on the data here.
582 			 */
583 			rec = &agg->dtagd_rec[agg->dtagd_nrecs - 1];
584 			roffs = rec->dtrd_offset;
585 			/* LINTED - alignment */
586 			h->dtahe_aggregate((int64_t *)&data[roffs],
587 			    /* LINTED - alignment */
588 			    (int64_t *)&addr[roffs], rec->dtrd_size);
589 
590 			/*
591 			 * If we're keeping per CPU data, apply the aggregating
592 			 * action there as well.
593 			 */
594 			if (aggdata->dtada_percpu != NULL) {
595 				data = aggdata->dtada_percpu[cpu];
596 
597 				/* LINTED - alignment */
598 				h->dtahe_aggregate((int64_t *)data,
599 				    /* LINTED - alignment */
600 				    (int64_t *)&addr[roffs], rec->dtrd_size);
601 			}
602 
603 			goto bufnext;
604 hashnext:
605 			continue;
606 		}
607 
608 		/*
609 		 * If we're here, we couldn't find an entry for this record.
610 		 */
611 		if ((h = malloc(sizeof (dt_ahashent_t))) == NULL)
612 			return (dt_set_errno(dtp, EDT_NOMEM));
613 		bzero(h, sizeof (dt_ahashent_t));
614 		aggdata = &h->dtahe_data;
615 
616 		if ((aggdata->dtada_data = malloc(size)) == NULL) {
617 			free(h);
618 			return (dt_set_errno(dtp, EDT_NOMEM));
619 		}
620 
621 		bcopy(addr, aggdata->dtada_data, size);
622 		aggdata->dtada_size = size;
623 		aggdata->dtada_desc = agg;
624 		aggdata->dtada_handle = dtp;
625 		(void) dt_epid_lookup(dtp, agg->dtagd_epid,
626 		    &aggdata->dtada_edesc, &aggdata->dtada_pdesc);
627 		aggdata->dtada_normal = 1;
628 
629 		h->dtahe_hashval = hashval;
630 		h->dtahe_size = size;
631 		(void) dt_aggregate_aggvarid(h);
632 
633 		rec = &agg->dtagd_rec[agg->dtagd_nrecs - 1];
634 
635 		if (flags & DTRACE_A_PERCPU) {
636 			int max_cpus = agp->dtat_maxcpu;
637 			caddr_t *percpu = malloc(max_cpus * sizeof (caddr_t));
638 
639 			if (percpu == NULL) {
640 				free(aggdata->dtada_data);
641 				free(h);
642 				return (dt_set_errno(dtp, EDT_NOMEM));
643 			}
644 
645 			for (j = 0; j < max_cpus; j++) {
646 				percpu[j] = malloc(rec->dtrd_size);
647 
648 				if (percpu[j] == NULL) {
649 					while (--j >= 0)
650 						free(percpu[j]);
651 
652 					free(aggdata->dtada_data);
653 					free(h);
654 					return (dt_set_errno(dtp, EDT_NOMEM));
655 				}
656 
657 				if (j == cpu) {
658 					bcopy(&addr[rec->dtrd_offset],
659 					    percpu[j], rec->dtrd_size);
660 				} else {
661 					bzero(percpu[j], rec->dtrd_size);
662 				}
663 			}
664 
665 			aggdata->dtada_percpu = percpu;
666 		}
667 
668 		switch (rec->dtrd_action) {
669 		case DTRACEAGG_MIN:
670 			h->dtahe_aggregate = dt_aggregate_min;
671 			break;
672 
673 		case DTRACEAGG_MAX:
674 			h->dtahe_aggregate = dt_aggregate_max;
675 			break;
676 
677 		case DTRACEAGG_LQUANTIZE:
678 			h->dtahe_aggregate = dt_aggregate_lquantize;
679 			break;
680 
681 		case DTRACEAGG_LLQUANTIZE:
682 			h->dtahe_aggregate = dt_aggregate_llquantize;
683 			break;
684 
685 		case DTRACEAGG_COUNT:
686 		case DTRACEAGG_SUM:
687 		case DTRACEAGG_AVG:
688 		case DTRACEAGG_STDDEV:
689 		case DTRACEAGG_QUANTIZE:
690 			h->dtahe_aggregate = dt_aggregate_count;
691 			break;
692 
693 		default:
694 			return (dt_set_errno(dtp, EDT_BADAGG));
695 		}
696 
697 		if (hash->dtah_hash[ndx] != NULL)
698 			hash->dtah_hash[ndx]->dtahe_prev = h;
699 
700 		h->dtahe_next = hash->dtah_hash[ndx];
701 		hash->dtah_hash[ndx] = h;
702 
703 		if (hash->dtah_all != NULL)
704 			hash->dtah_all->dtahe_prevall = h;
705 
706 		h->dtahe_nextall = hash->dtah_all;
707 		hash->dtah_all = h;
708 bufnext:
709 		offs += agg->dtagd_size;
710 	}
711 
712 	return (0);
713 }
714 
715 int
716 dtrace_aggregate_snap(dtrace_hdl_t *dtp)
717 {
718 	int i, rval;
719 	dt_aggregate_t *agp = &dtp->dt_aggregate;
720 	hrtime_t now = gethrtime();
721 	dtrace_optval_t interval = dtp->dt_options[DTRACEOPT_AGGRATE];
722 
723 	if (dtp->dt_lastagg != 0) {
724 		if (now - dtp->dt_lastagg < interval)
725 			return (0);
726 
727 		dtp->dt_lastagg += interval;
728 	} else {
729 		dtp->dt_lastagg = now;
730 	}
731 
732 	if (!dtp->dt_active)
733 		return (dt_set_errno(dtp, EINVAL));
734 
735 	if (agp->dtat_buf.dtbd_size == 0)
736 		return (0);
737 
738 	for (i = 0; i < agp->dtat_ncpus; i++) {
739 		if ((rval = dt_aggregate_snap_cpu(dtp, agp->dtat_cpus[i])))
740 			return (rval);
741 	}
742 
743 	return (0);
744 }
745 
746 static int
747 dt_aggregate_hashcmp(const void *lhs, const void *rhs)
748 {
749 	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
750 	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
751 	dtrace_aggdesc_t *lagg = lh->dtahe_data.dtada_desc;
752 	dtrace_aggdesc_t *ragg = rh->dtahe_data.dtada_desc;
753 
754 	if (lagg->dtagd_nrecs < ragg->dtagd_nrecs)
755 		return (DT_LESSTHAN);
756 
757 	if (lagg->dtagd_nrecs > ragg->dtagd_nrecs)
758 		return (DT_GREATERTHAN);
759 
760 	return (0);
761 }
762 
763 static int
764 dt_aggregate_varcmp(const void *lhs, const void *rhs)
765 {
766 	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
767 	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
768 	dtrace_aggvarid_t lid, rid;
769 
770 	lid = dt_aggregate_aggvarid(lh);
771 	rid = dt_aggregate_aggvarid(rh);
772 
773 	if (lid < rid)
774 		return (DT_LESSTHAN);
775 
776 	if (lid > rid)
777 		return (DT_GREATERTHAN);
778 
779 	return (0);
780 }
781 
782 static int
783 dt_aggregate_keycmp(const void *lhs, const void *rhs)
784 {
785 	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
786 	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
787 	dtrace_aggdesc_t *lagg = lh->dtahe_data.dtada_desc;
788 	dtrace_aggdesc_t *ragg = rh->dtahe_data.dtada_desc;
789 	dtrace_recdesc_t *lrec, *rrec;
790 	char *ldata, *rdata;
791 	int rval, i, j, keypos, nrecs;
792 
793 	if ((rval = dt_aggregate_hashcmp(lhs, rhs)) != 0)
794 		return (rval);
795 
796 	nrecs = lagg->dtagd_nrecs - 1;
797 	assert(nrecs == ragg->dtagd_nrecs - 1);
798 
799 	keypos = dt_keypos + 1 >= nrecs ? 0 : dt_keypos;
800 
801 	for (i = 1; i < nrecs; i++) {
802 		uint64_t lval, rval;
803 		int ndx = i + keypos;
804 
805 		if (ndx >= nrecs)
806 			ndx = ndx - nrecs + 1;
807 
808 		lrec = &lagg->dtagd_rec[ndx];
809 		rrec = &ragg->dtagd_rec[ndx];
810 
811 		ldata = lh->dtahe_data.dtada_data + lrec->dtrd_offset;
812 		rdata = rh->dtahe_data.dtada_data + rrec->dtrd_offset;
813 
814 		if (lrec->dtrd_size < rrec->dtrd_size)
815 			return (DT_LESSTHAN);
816 
817 		if (lrec->dtrd_size > rrec->dtrd_size)
818 			return (DT_GREATERTHAN);
819 
820 		switch (lrec->dtrd_size) {
821 		case sizeof (uint64_t):
822 			/* LINTED - alignment */
823 			lval = *((uint64_t *)ldata);
824 			/* LINTED - alignment */
825 			rval = *((uint64_t *)rdata);
826 			break;
827 
828 		case sizeof (uint32_t):
829 			/* LINTED - alignment */
830 			lval = *((uint32_t *)ldata);
831 			/* LINTED - alignment */
832 			rval = *((uint32_t *)rdata);
833 			break;
834 
835 		case sizeof (uint16_t):
836 			/* LINTED - alignment */
837 			lval = *((uint16_t *)ldata);
838 			/* LINTED - alignment */
839 			rval = *((uint16_t *)rdata);
840 			break;
841 
842 		case sizeof (uint8_t):
843 			lval = *((uint8_t *)ldata);
844 			rval = *((uint8_t *)rdata);
845 			break;
846 
847 		default:
848 			switch (lrec->dtrd_action) {
849 			case DTRACEACT_UMOD:
850 			case DTRACEACT_UADDR:
851 			case DTRACEACT_USYM:
852 				for (j = 0; j < 2; j++) {
853 					/* LINTED - alignment */
854 					lval = ((uint64_t *)ldata)[j];
855 					/* LINTED - alignment */
856 					rval = ((uint64_t *)rdata)[j];
857 
858 					if (lval < rval)
859 						return (DT_LESSTHAN);
860 
861 					if (lval > rval)
862 						return (DT_GREATERTHAN);
863 				}
864 
865 				break;
866 
867 			default:
868 				for (j = 0; j < lrec->dtrd_size; j++) {
869 					lval = ((uint8_t *)ldata)[j];
870 					rval = ((uint8_t *)rdata)[j];
871 
872 					if (lval < rval)
873 						return (DT_LESSTHAN);
874 
875 					if (lval > rval)
876 						return (DT_GREATERTHAN);
877 				}
878 			}
879 
880 			continue;
881 		}
882 
883 		if (lval < rval)
884 			return (DT_LESSTHAN);
885 
886 		if (lval > rval)
887 			return (DT_GREATERTHAN);
888 	}
889 
890 	return (0);
891 }
892 
893 static int
894 dt_aggregate_valcmp(const void *lhs, const void *rhs)
895 {
896 	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
897 	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
898 	dtrace_aggdesc_t *lagg = lh->dtahe_data.dtada_desc;
899 	dtrace_aggdesc_t *ragg = rh->dtahe_data.dtada_desc;
900 	caddr_t ldata = lh->dtahe_data.dtada_data;
901 	caddr_t rdata = rh->dtahe_data.dtada_data;
902 	dtrace_recdesc_t *lrec, *rrec;
903 	int64_t *laddr, *raddr;
904 	int rval;
905 
906 	assert(lagg->dtagd_nrecs == ragg->dtagd_nrecs);
907 
908 	lrec = &lagg->dtagd_rec[lagg->dtagd_nrecs - 1];
909 	rrec = &ragg->dtagd_rec[ragg->dtagd_nrecs - 1];
910 
911 	assert(lrec->dtrd_action == rrec->dtrd_action);
912 
913 	laddr = (int64_t *)(uintptr_t)(ldata + lrec->dtrd_offset);
914 	raddr = (int64_t *)(uintptr_t)(rdata + rrec->dtrd_offset);
915 
916 	switch (lrec->dtrd_action) {
917 	case DTRACEAGG_AVG:
918 		rval = dt_aggregate_averagecmp(laddr, raddr);
919 		break;
920 
921 	case DTRACEAGG_STDDEV:
922 		rval = dt_aggregate_stddevcmp(laddr, raddr);
923 		break;
924 
925 	case DTRACEAGG_QUANTIZE:
926 		rval = dt_aggregate_quantizedcmp(laddr, raddr);
927 		break;
928 
929 	case DTRACEAGG_LQUANTIZE:
930 		rval = dt_aggregate_lquantizedcmp(laddr, raddr);
931 		break;
932 
933 	case DTRACEAGG_LLQUANTIZE:
934 		rval = dt_aggregate_llquantizedcmp(laddr, raddr);
935 		break;
936 
937 	case DTRACEAGG_COUNT:
938 	case DTRACEAGG_SUM:
939 	case DTRACEAGG_MIN:
940 	case DTRACEAGG_MAX:
941 		rval = dt_aggregate_countcmp(laddr, raddr);
942 		break;
943 
944 	default:
945 		assert(0);
946 	}
947 
948 	return (rval);
949 }
950 
951 static int
952 dt_aggregate_valkeycmp(const void *lhs, const void *rhs)
953 {
954 	int rval;
955 
956 	if ((rval = dt_aggregate_valcmp(lhs, rhs)) != 0)
957 		return (rval);
958 
959 	/*
960 	 * If we're here, the values for the two aggregation elements are
961 	 * equal.  We already know that the key layout is the same for the two
962 	 * elements; we must now compare the keys themselves as a tie-breaker.
963 	 */
964 	return (dt_aggregate_keycmp(lhs, rhs));
965 }
966 
967 static int
968 dt_aggregate_keyvarcmp(const void *lhs, const void *rhs)
969 {
970 	int rval;
971 
972 	if ((rval = dt_aggregate_keycmp(lhs, rhs)) != 0)
973 		return (rval);
974 
975 	return (dt_aggregate_varcmp(lhs, rhs));
976 }
977 
978 static int
979 dt_aggregate_varkeycmp(const void *lhs, const void *rhs)
980 {
981 	int rval;
982 
983 	if ((rval = dt_aggregate_varcmp(lhs, rhs)) != 0)
984 		return (rval);
985 
986 	return (dt_aggregate_keycmp(lhs, rhs));
987 }
988 
989 static int
990 dt_aggregate_valvarcmp(const void *lhs, const void *rhs)
991 {
992 	int rval;
993 
994 	if ((rval = dt_aggregate_valkeycmp(lhs, rhs)) != 0)
995 		return (rval);
996 
997 	return (dt_aggregate_varcmp(lhs, rhs));
998 }
999 
1000 static int
1001 dt_aggregate_varvalcmp(const void *lhs, const void *rhs)
1002 {
1003 	int rval;
1004 
1005 	if ((rval = dt_aggregate_varcmp(lhs, rhs)) != 0)
1006 		return (rval);
1007 
1008 	return (dt_aggregate_valkeycmp(lhs, rhs));
1009 }
1010 
1011 static int
1012 dt_aggregate_keyvarrevcmp(const void *lhs, const void *rhs)
1013 {
1014 	return (dt_aggregate_keyvarcmp(rhs, lhs));
1015 }
1016 
1017 static int
1018 dt_aggregate_varkeyrevcmp(const void *lhs, const void *rhs)
1019 {
1020 	return (dt_aggregate_varkeycmp(rhs, lhs));
1021 }
1022 
1023 static int
1024 dt_aggregate_valvarrevcmp(const void *lhs, const void *rhs)
1025 {
1026 	return (dt_aggregate_valvarcmp(rhs, lhs));
1027 }
1028 
1029 static int
1030 dt_aggregate_varvalrevcmp(const void *lhs, const void *rhs)
1031 {
1032 	return (dt_aggregate_varvalcmp(rhs, lhs));
1033 }
1034 
1035 static int
1036 dt_aggregate_bundlecmp(const void *lhs, const void *rhs)
1037 {
1038 	dt_ahashent_t **lh = *((dt_ahashent_t ***)lhs);
1039 	dt_ahashent_t **rh = *((dt_ahashent_t ***)rhs);
1040 	int i, rval;
1041 
1042 	if (dt_keysort) {
1043 		/*
1044 		 * If we're sorting on keys, we need to scan until we find the
1045 		 * last entry -- that's the representative key.  (The order of
1046 		 * the bundle is values followed by key to accommodate the
1047 		 * default behavior of sorting by value.)  If the keys are
1048 		 * equal, we'll fall into the value comparison loop, below.
1049 		 */
1050 		for (i = 0; lh[i + 1] != NULL; i++)
1051 			continue;
1052 
1053 		assert(i != 0);
1054 		assert(rh[i + 1] == NULL);
1055 
1056 		if ((rval = dt_aggregate_keycmp(&lh[i], &rh[i])) != 0)
1057 			return (rval);
1058 	}
1059 
1060 	for (i = 0; ; i++) {
1061 		if (lh[i + 1] == NULL) {
1062 			/*
1063 			 * All of the values are equal; if we're sorting on
1064 			 * keys, then we're only here because the keys were
1065 			 * found to be equal and these records are therefore
1066 			 * equal.  If we're not sorting on keys, we'll use the
1067 			 * key comparison from the representative key as the
1068 			 * tie-breaker.
1069 			 */
1070 			if (dt_keysort)
1071 				return (0);
1072 
1073 			assert(i != 0);
1074 			assert(rh[i + 1] == NULL);
1075 			return (dt_aggregate_keycmp(&lh[i], &rh[i]));
1076 		} else {
1077 			if ((rval = dt_aggregate_valcmp(&lh[i], &rh[i])) != 0)
1078 				return (rval);
1079 		}
1080 	}
1081 }
1082 
1083 int
1084 dt_aggregate_go(dtrace_hdl_t *dtp)
1085 {
1086 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1087 	dtrace_optval_t size, cpu;
1088 	dtrace_bufdesc_t *buf = &agp->dtat_buf;
1089 	int rval, i;
1090 
1091 	assert(agp->dtat_maxcpu == 0);
1092 	assert(agp->dtat_ncpu == 0);
1093 	assert(agp->dtat_cpus == NULL);
1094 
1095 	agp->dtat_maxcpu = dt_cpu_maxid(dtp) + 1;
1096 	if (agp->dtat_maxcpu <= 0)
1097 		return (-1);
1098 	agp->dtat_ncpu = dt_sysconf(dtp, _SC_NPROCESSORS_CONF);
1099 	agp->dtat_cpus = malloc(agp->dtat_ncpu * sizeof (processorid_t));
1100 
1101 	if (agp->dtat_cpus == NULL)
1102 		return (dt_set_errno(dtp, EDT_NOMEM));
1103 
1104 	/*
1105 	 * Use the aggregation buffer size as reloaded from the kernel.
1106 	 */
1107 	size = dtp->dt_options[DTRACEOPT_AGGSIZE];
1108 
1109 	rval = dtrace_getopt(dtp, "aggsize", &size);
1110 	assert(rval == 0);
1111 
1112 	if (size == 0 || size == DTRACEOPT_UNSET)
1113 		return (0);
1114 
1115 	buf = &agp->dtat_buf;
1116 	buf->dtbd_size = size;
1117 
1118 	if ((buf->dtbd_data = malloc(buf->dtbd_size)) == NULL)
1119 		return (dt_set_errno(dtp, EDT_NOMEM));
1120 
1121 	/*
1122 	 * Now query for the CPUs enabled.
1123 	 */
1124 	rval = dtrace_getopt(dtp, "cpu", &cpu);
1125 	assert(rval == 0 && cpu != DTRACEOPT_UNSET);
1126 
1127 	if (cpu != DTRACE_CPUALL) {
1128 		assert(cpu < agp->dtat_ncpu);
1129 		agp->dtat_cpus[agp->dtat_ncpus++] = (processorid_t)cpu;
1130 
1131 		return (0);
1132 	}
1133 
1134 	agp->dtat_ncpus = 0;
1135 	for (i = 0; i < agp->dtat_maxcpu; i++) {
1136 		if (dt_status(dtp, i) == -1)
1137 			continue;
1138 
1139 		agp->dtat_cpus[agp->dtat_ncpus++] = i;
1140 	}
1141 
1142 	return (0);
1143 }
1144 
1145 static int
1146 dt_aggwalk_rval(dtrace_hdl_t *dtp, dt_ahashent_t *h, int rval)
1147 {
1148 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1149 	dtrace_aggdata_t *data;
1150 	dtrace_aggdesc_t *aggdesc;
1151 	dtrace_recdesc_t *rec;
1152 	int i;
1153 
1154 	switch (rval) {
1155 	case DTRACE_AGGWALK_NEXT:
1156 		break;
1157 
1158 	case DTRACE_AGGWALK_CLEAR: {
1159 		uint32_t size, offs = 0;
1160 
1161 		aggdesc = h->dtahe_data.dtada_desc;
1162 		rec = &aggdesc->dtagd_rec[aggdesc->dtagd_nrecs - 1];
1163 		size = rec->dtrd_size;
1164 		data = &h->dtahe_data;
1165 
1166 		if (rec->dtrd_action == DTRACEAGG_LQUANTIZE) {
1167 			offs = sizeof (uint64_t);
1168 			size -= sizeof (uint64_t);
1169 		}
1170 
1171 		bzero(&data->dtada_data[rec->dtrd_offset] + offs, size);
1172 
1173 		if (data->dtada_percpu == NULL)
1174 			break;
1175 
1176 		for (i = 0; i < dtp->dt_aggregate.dtat_maxcpu; i++)
1177 			bzero(data->dtada_percpu[i] + offs, size);
1178 		break;
1179 	}
1180 
1181 	case DTRACE_AGGWALK_ERROR:
1182 		/*
1183 		 * We assume that errno is already set in this case.
1184 		 */
1185 		return (dt_set_errno(dtp, errno));
1186 
1187 	case DTRACE_AGGWALK_ABORT:
1188 		return (dt_set_errno(dtp, EDT_DIRABORT));
1189 
1190 	case DTRACE_AGGWALK_DENORMALIZE:
1191 		h->dtahe_data.dtada_normal = 1;
1192 		return (0);
1193 
1194 	case DTRACE_AGGWALK_NORMALIZE:
1195 		if (h->dtahe_data.dtada_normal == 0) {
1196 			h->dtahe_data.dtada_normal = 1;
1197 			return (dt_set_errno(dtp, EDT_BADRVAL));
1198 		}
1199 
1200 		return (0);
1201 
1202 	case DTRACE_AGGWALK_REMOVE: {
1203 		dtrace_aggdata_t *aggdata = &h->dtahe_data;
1204 		int max_cpus = agp->dtat_maxcpu;
1205 
1206 		/*
1207 		 * First, remove this hash entry from its hash chain.
1208 		 */
1209 		if (h->dtahe_prev != NULL) {
1210 			h->dtahe_prev->dtahe_next = h->dtahe_next;
1211 		} else {
1212 			dt_ahash_t *hash = &agp->dtat_hash;
1213 			size_t ndx = h->dtahe_hashval % hash->dtah_size;
1214 
1215 			assert(hash->dtah_hash[ndx] == h);
1216 			hash->dtah_hash[ndx] = h->dtahe_next;
1217 		}
1218 
1219 		if (h->dtahe_next != NULL)
1220 			h->dtahe_next->dtahe_prev = h->dtahe_prev;
1221 
1222 		/*
1223 		 * Now remove it from the list of all hash entries.
1224 		 */
1225 		if (h->dtahe_prevall != NULL) {
1226 			h->dtahe_prevall->dtahe_nextall = h->dtahe_nextall;
1227 		} else {
1228 			dt_ahash_t *hash = &agp->dtat_hash;
1229 
1230 			assert(hash->dtah_all == h);
1231 			hash->dtah_all = h->dtahe_nextall;
1232 		}
1233 
1234 		if (h->dtahe_nextall != NULL)
1235 			h->dtahe_nextall->dtahe_prevall = h->dtahe_prevall;
1236 
1237 		/*
1238 		 * We're unlinked.  We can safely destroy the data.
1239 		 */
1240 		if (aggdata->dtada_percpu != NULL) {
1241 			for (i = 0; i < max_cpus; i++)
1242 				free(aggdata->dtada_percpu[i]);
1243 			free(aggdata->dtada_percpu);
1244 		}
1245 
1246 		free(aggdata->dtada_data);
1247 		free(h);
1248 
1249 		return (0);
1250 	}
1251 
1252 	default:
1253 		return (dt_set_errno(dtp, EDT_BADRVAL));
1254 	}
1255 
1256 	return (0);
1257 }
1258 
1259 void
1260 dt_aggregate_qsort(dtrace_hdl_t *dtp, void *base, size_t nel, size_t width,
1261     int (*compar)(const void *, const void *))
1262 {
1263 	int rev = dt_revsort, key = dt_keysort, keypos = dt_keypos;
1264 	dtrace_optval_t keyposopt = dtp->dt_options[DTRACEOPT_AGGSORTKEYPOS];
1265 
1266 	dt_revsort = (dtp->dt_options[DTRACEOPT_AGGSORTREV] != DTRACEOPT_UNSET);
1267 	dt_keysort = (dtp->dt_options[DTRACEOPT_AGGSORTKEY] != DTRACEOPT_UNSET);
1268 
1269 	if (keyposopt != DTRACEOPT_UNSET && keyposopt <= INT_MAX) {
1270 		dt_keypos = (int)keyposopt;
1271 	} else {
1272 		dt_keypos = 0;
1273 	}
1274 
1275 	if (compar == NULL) {
1276 		if (!dt_keysort) {
1277 			compar = dt_aggregate_varvalcmp;
1278 		} else {
1279 			compar = dt_aggregate_varkeycmp;
1280 		}
1281 	}
1282 
1283 	qsort(base, nel, width, compar);
1284 
1285 	dt_revsort = rev;
1286 	dt_keysort = key;
1287 	dt_keypos = keypos;
1288 }
1289 
1290 int
1291 dtrace_aggregate_walk(dtrace_hdl_t *dtp, dtrace_aggregate_f *func, void *arg)
1292 {
1293 	dt_ahashent_t *h, *next;
1294 	dt_ahash_t *hash = &dtp->dt_aggregate.dtat_hash;
1295 
1296 	for (h = hash->dtah_all; h != NULL; h = next) {
1297 		/*
1298 		 * dt_aggwalk_rval() can potentially remove the current hash
1299 		 * entry; we need to load the next hash entry before calling
1300 		 * into it.
1301 		 */
1302 		next = h->dtahe_nextall;
1303 
1304 		if (dt_aggwalk_rval(dtp, h, func(&h->dtahe_data, arg)) == -1)
1305 			return (-1);
1306 	}
1307 
1308 	return (0);
1309 }
1310 
1311 static int
1312 dt_aggregate_total(dtrace_hdl_t *dtp, boolean_t clear)
1313 {
1314 	dt_ahashent_t *h;
1315 	dtrace_aggdata_t **total;
1316 	dtrace_aggid_t max = DTRACE_AGGVARIDNONE, id;
1317 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1318 	dt_ahash_t *hash = &agp->dtat_hash;
1319 	uint32_t tflags;
1320 
1321 	tflags = DTRACE_A_TOTAL | DTRACE_A_HASNEGATIVES | DTRACE_A_HASPOSITIVES;
1322 
1323 	/*
1324 	 * If we need to deliver per-aggregation totals, we're going to take
1325 	 * three passes over the aggregate:  one to clear everything out and
1326 	 * determine our maximum aggregation ID, one to actually total
1327 	 * everything up, and a final pass to assign the totals to the
1328 	 * individual elements.
1329 	 */
1330 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1331 		dtrace_aggdata_t *aggdata = &h->dtahe_data;
1332 
1333 		if ((id = dt_aggregate_aggvarid(h)) > max)
1334 			max = id;
1335 
1336 		aggdata->dtada_total = 0;
1337 		aggdata->dtada_flags &= ~tflags;
1338 	}
1339 
1340 	if (clear || max == DTRACE_AGGVARIDNONE)
1341 		return (0);
1342 
1343 	total = dt_zalloc(dtp, (max + 1) * sizeof (dtrace_aggdata_t *));
1344 
1345 	if (total == NULL)
1346 		return (-1);
1347 
1348 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1349 		dtrace_aggdata_t *aggdata = &h->dtahe_data;
1350 		dtrace_aggdesc_t *agg = aggdata->dtada_desc;
1351 		dtrace_recdesc_t *rec;
1352 		caddr_t data;
1353 		int64_t val, *addr;
1354 
1355 		rec = &agg->dtagd_rec[agg->dtagd_nrecs - 1];
1356 		data = aggdata->dtada_data;
1357 		addr = (int64_t *)(uintptr_t)(data + rec->dtrd_offset);
1358 
1359 		switch (rec->dtrd_action) {
1360 		case DTRACEAGG_STDDEV:
1361 			val = dt_stddev((uint64_t *)addr, 1);
1362 			break;
1363 
1364 		case DTRACEAGG_SUM:
1365 		case DTRACEAGG_COUNT:
1366 			val = *addr;
1367 			break;
1368 
1369 		case DTRACEAGG_AVG:
1370 			val = addr[0] ? (addr[1] / addr[0]) : 0;
1371 			break;
1372 
1373 		default:
1374 			continue;
1375 		}
1376 
1377 		if (total[agg->dtagd_varid] == NULL) {
1378 			total[agg->dtagd_varid] = aggdata;
1379 			aggdata->dtada_flags |= DTRACE_A_TOTAL;
1380 		} else {
1381 			aggdata = total[agg->dtagd_varid];
1382 		}
1383 
1384 		if (val > 0)
1385 			aggdata->dtada_flags |= DTRACE_A_HASPOSITIVES;
1386 
1387 		if (val < 0) {
1388 			aggdata->dtada_flags |= DTRACE_A_HASNEGATIVES;
1389 			val = -val;
1390 		}
1391 
1392 		if (dtp->dt_options[DTRACEOPT_AGGZOOM] != DTRACEOPT_UNSET) {
1393 			val = (int64_t)((long double)val *
1394 			    (1 / DTRACE_AGGZOOM_MAX));
1395 
1396 			if (val > aggdata->dtada_total)
1397 				aggdata->dtada_total = val;
1398 		} else {
1399 			aggdata->dtada_total += val;
1400 		}
1401 	}
1402 
1403 	/*
1404 	 * And now one final pass to set everyone's total.
1405 	 */
1406 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1407 		dtrace_aggdata_t *aggdata = &h->dtahe_data, *t;
1408 		dtrace_aggdesc_t *agg = aggdata->dtada_desc;
1409 
1410 		if ((t = total[agg->dtagd_varid]) == NULL || aggdata == t)
1411 			continue;
1412 
1413 		aggdata->dtada_total = t->dtada_total;
1414 		aggdata->dtada_flags |= (t->dtada_flags & tflags);
1415 	}
1416 
1417 	dt_free(dtp, total);
1418 
1419 	return (0);
1420 }
1421 
1422 static int
1423 dt_aggregate_minmaxbin(dtrace_hdl_t *dtp, boolean_t clear)
1424 {
1425 	dt_ahashent_t *h;
1426 	dtrace_aggdata_t **minmax;
1427 	dtrace_aggid_t max = DTRACE_AGGVARIDNONE, id;
1428 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1429 	dt_ahash_t *hash = &agp->dtat_hash;
1430 
1431 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1432 		dtrace_aggdata_t *aggdata = &h->dtahe_data;
1433 
1434 		if ((id = dt_aggregate_aggvarid(h)) > max)
1435 			max = id;
1436 
1437 		aggdata->dtada_minbin = 0;
1438 		aggdata->dtada_maxbin = 0;
1439 		aggdata->dtada_flags &= ~DTRACE_A_MINMAXBIN;
1440 	}
1441 
1442 	if (clear || max == DTRACE_AGGVARIDNONE)
1443 		return (0);
1444 
1445 	minmax = dt_zalloc(dtp, (max + 1) * sizeof (dtrace_aggdata_t *));
1446 
1447 	if (minmax == NULL)
1448 		return (-1);
1449 
1450 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1451 		dtrace_aggdata_t *aggdata = &h->dtahe_data;
1452 		dtrace_aggdesc_t *agg = aggdata->dtada_desc;
1453 		dtrace_recdesc_t *rec;
1454 		caddr_t data;
1455 		int64_t *addr;
1456 		int minbin = -1, maxbin = -1, i;
1457 		int start = 0, size;
1458 
1459 		rec = &agg->dtagd_rec[agg->dtagd_nrecs - 1];
1460 		size = rec->dtrd_size / sizeof (int64_t);
1461 		data = aggdata->dtada_data;
1462 		addr = (int64_t *)(uintptr_t)(data + rec->dtrd_offset);
1463 
1464 		switch (rec->dtrd_action) {
1465 		case DTRACEAGG_LQUANTIZE:
1466 			/*
1467 			 * For lquantize(), we always display the entire range
1468 			 * of the aggregation when aggpack is set.
1469 			 */
1470 			start = 1;
1471 			minbin = start;
1472 			maxbin = size - 1 - start;
1473 			break;
1474 
1475 		case DTRACEAGG_QUANTIZE:
1476 			for (i = start; i < size; i++) {
1477 				if (!addr[i])
1478 					continue;
1479 
1480 				if (minbin == -1)
1481 					minbin = i - start;
1482 
1483 				maxbin = i - start;
1484 			}
1485 
1486 			if (minbin == -1) {
1487 				/*
1488 				 * If we have no data (e.g., due to a clear()
1489 				 * or negative increments), we'll use the
1490 				 * zero bucket as both our min and max.
1491 				 */
1492 				minbin = maxbin = DTRACE_QUANTIZE_ZEROBUCKET;
1493 			}
1494 
1495 			break;
1496 
1497 		default:
1498 			continue;
1499 		}
1500 
1501 		if (minmax[agg->dtagd_varid] == NULL) {
1502 			minmax[agg->dtagd_varid] = aggdata;
1503 			aggdata->dtada_flags |= DTRACE_A_MINMAXBIN;
1504 			aggdata->dtada_minbin = minbin;
1505 			aggdata->dtada_maxbin = maxbin;
1506 			continue;
1507 		}
1508 
1509 		if (minbin < minmax[agg->dtagd_varid]->dtada_minbin)
1510 			minmax[agg->dtagd_varid]->dtada_minbin = minbin;
1511 
1512 		if (maxbin > minmax[agg->dtagd_varid]->dtada_maxbin)
1513 			minmax[agg->dtagd_varid]->dtada_maxbin = maxbin;
1514 	}
1515 
1516 	/*
1517 	 * And now one final pass to set everyone's minbin and maxbin.
1518 	 */
1519 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1520 		dtrace_aggdata_t *aggdata = &h->dtahe_data, *mm;
1521 		dtrace_aggdesc_t *agg = aggdata->dtada_desc;
1522 
1523 		if ((mm = minmax[agg->dtagd_varid]) == NULL || aggdata == mm)
1524 			continue;
1525 
1526 		aggdata->dtada_minbin = mm->dtada_minbin;
1527 		aggdata->dtada_maxbin = mm->dtada_maxbin;
1528 		aggdata->dtada_flags |= DTRACE_A_MINMAXBIN;
1529 	}
1530 
1531 	dt_free(dtp, minmax);
1532 
1533 	return (0);
1534 }
1535 
1536 static int
1537 dt_aggregate_walk_sorted(dtrace_hdl_t *dtp,
1538     dtrace_aggregate_f *func, void *arg,
1539     int (*sfunc)(const void *, const void *))
1540 {
1541 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1542 	dt_ahashent_t *h, **sorted;
1543 	dt_ahash_t *hash = &agp->dtat_hash;
1544 	size_t i, nentries = 0;
1545 	int rval = -1;
1546 
1547 	agp->dtat_flags &= ~(DTRACE_A_TOTAL | DTRACE_A_MINMAXBIN);
1548 
1549 	if (dtp->dt_options[DTRACEOPT_AGGHIST] != DTRACEOPT_UNSET) {
1550 		agp->dtat_flags |= DTRACE_A_TOTAL;
1551 
1552 		if (dt_aggregate_total(dtp, B_FALSE) != 0)
1553 			return (-1);
1554 	}
1555 
1556 	if (dtp->dt_options[DTRACEOPT_AGGPACK] != DTRACEOPT_UNSET) {
1557 		agp->dtat_flags |= DTRACE_A_MINMAXBIN;
1558 
1559 		if (dt_aggregate_minmaxbin(dtp, B_FALSE) != 0)
1560 			return (-1);
1561 	}
1562 
1563 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall)
1564 		nentries++;
1565 
1566 	sorted = dt_alloc(dtp, nentries * sizeof (dt_ahashent_t *));
1567 
1568 	if (sorted == NULL)
1569 		goto out;
1570 
1571 	for (h = hash->dtah_all, i = 0; h != NULL; h = h->dtahe_nextall)
1572 		sorted[i++] = h;
1573 
1574 	(void) pthread_mutex_lock(&dt_qsort_lock);
1575 
1576 	if (sfunc == NULL) {
1577 		dt_aggregate_qsort(dtp, sorted, nentries,
1578 		    sizeof (dt_ahashent_t *), NULL);
1579 	} else {
1580 		/*
1581 		 * If we've been explicitly passed a sorting function,
1582 		 * we'll use that -- ignoring the values of the "aggsortrev",
1583 		 * "aggsortkey" and "aggsortkeypos" options.
1584 		 */
1585 		qsort(sorted, nentries, sizeof (dt_ahashent_t *), sfunc);
1586 	}
1587 
1588 	(void) pthread_mutex_unlock(&dt_qsort_lock);
1589 
1590 	for (i = 0; i < nentries; i++) {
1591 		h = sorted[i];
1592 
1593 		if (dt_aggwalk_rval(dtp, h, func(&h->dtahe_data, arg)) == -1)
1594 			goto out;
1595 	}
1596 
1597 	rval = 0;
1598 out:
1599 	if (agp->dtat_flags & DTRACE_A_TOTAL)
1600 		(void) dt_aggregate_total(dtp, B_TRUE);
1601 
1602 	if (agp->dtat_flags & DTRACE_A_MINMAXBIN)
1603 		(void) dt_aggregate_minmaxbin(dtp, B_TRUE);
1604 
1605 	dt_free(dtp, sorted);
1606 	return (rval);
1607 }
1608 
1609 int
1610 dtrace_aggregate_walk_sorted(dtrace_hdl_t *dtp,
1611     dtrace_aggregate_f *func, void *arg)
1612 {
1613 	return (dt_aggregate_walk_sorted(dtp, func, arg, NULL));
1614 }
1615 
1616 int
1617 dtrace_aggregate_walk_keysorted(dtrace_hdl_t *dtp,
1618     dtrace_aggregate_f *func, void *arg)
1619 {
1620 	return (dt_aggregate_walk_sorted(dtp, func,
1621 	    arg, dt_aggregate_varkeycmp));
1622 }
1623 
1624 int
1625 dtrace_aggregate_walk_valsorted(dtrace_hdl_t *dtp,
1626     dtrace_aggregate_f *func, void *arg)
1627 {
1628 	return (dt_aggregate_walk_sorted(dtp, func,
1629 	    arg, dt_aggregate_varvalcmp));
1630 }
1631 
1632 int
1633 dtrace_aggregate_walk_keyvarsorted(dtrace_hdl_t *dtp,
1634     dtrace_aggregate_f *func, void *arg)
1635 {
1636 	return (dt_aggregate_walk_sorted(dtp, func,
1637 	    arg, dt_aggregate_keyvarcmp));
1638 }
1639 
1640 int
1641 dtrace_aggregate_walk_valvarsorted(dtrace_hdl_t *dtp,
1642     dtrace_aggregate_f *func, void *arg)
1643 {
1644 	return (dt_aggregate_walk_sorted(dtp, func,
1645 	    arg, dt_aggregate_valvarcmp));
1646 }
1647 
1648 int
1649 dtrace_aggregate_walk_keyrevsorted(dtrace_hdl_t *dtp,
1650     dtrace_aggregate_f *func, void *arg)
1651 {
1652 	return (dt_aggregate_walk_sorted(dtp, func,
1653 	    arg, dt_aggregate_varkeyrevcmp));
1654 }
1655 
1656 int
1657 dtrace_aggregate_walk_valrevsorted(dtrace_hdl_t *dtp,
1658     dtrace_aggregate_f *func, void *arg)
1659 {
1660 	return (dt_aggregate_walk_sorted(dtp, func,
1661 	    arg, dt_aggregate_varvalrevcmp));
1662 }
1663 
1664 int
1665 dtrace_aggregate_walk_keyvarrevsorted(dtrace_hdl_t *dtp,
1666     dtrace_aggregate_f *func, void *arg)
1667 {
1668 	return (dt_aggregate_walk_sorted(dtp, func,
1669 	    arg, dt_aggregate_keyvarrevcmp));
1670 }
1671 
1672 int
1673 dtrace_aggregate_walk_valvarrevsorted(dtrace_hdl_t *dtp,
1674     dtrace_aggregate_f *func, void *arg)
1675 {
1676 	return (dt_aggregate_walk_sorted(dtp, func,
1677 	    arg, dt_aggregate_valvarrevcmp));
1678 }
1679 
1680 int
1681 dtrace_aggregate_walk_joined(dtrace_hdl_t *dtp, dtrace_aggvarid_t *aggvars,
1682     int naggvars, dtrace_aggregate_walk_joined_f *func, void *arg)
1683 {
1684 	dt_aggregate_t *agp = &dtp->dt_aggregate;
1685 	dt_ahashent_t *h, **sorted = NULL, ***bundle, **nbundle;
1686 	const dtrace_aggdata_t **data;
1687 	dt_ahashent_t *zaggdata = NULL;
1688 	dt_ahash_t *hash = &agp->dtat_hash;
1689 	size_t nentries = 0, nbundles = 0, start, zsize = 0, bundlesize;
1690 	dtrace_aggvarid_t max = 0, aggvar;
1691 	int rval = -1, *map, *remap = NULL;
1692 	int i, j;
1693 	dtrace_optval_t sortpos = dtp->dt_options[DTRACEOPT_AGGSORTPOS];
1694 
1695 	/*
1696 	 * If the sorting position is greater than the number of aggregation
1697 	 * variable IDs, we silently set it to 0.
1698 	 */
1699 	if (sortpos == DTRACEOPT_UNSET || sortpos >= naggvars)
1700 		sortpos = 0;
1701 
1702 	/*
1703 	 * First we need to translate the specified aggregation variable IDs
1704 	 * into a linear map that will allow us to translate an aggregation
1705 	 * variable ID into its position in the specified aggvars.
1706 	 */
1707 	for (i = 0; i < naggvars; i++) {
1708 		if (aggvars[i] == DTRACE_AGGVARIDNONE || aggvars[i] < 0)
1709 			return (dt_set_errno(dtp, EDT_BADAGGVAR));
1710 
1711 		if (aggvars[i] > max)
1712 			max = aggvars[i];
1713 	}
1714 
1715 	if ((map = dt_zalloc(dtp, (max + 1) * sizeof (int))) == NULL)
1716 		return (-1);
1717 
1718 	zaggdata = dt_zalloc(dtp, naggvars * sizeof (dt_ahashent_t));
1719 
1720 	if (zaggdata == NULL)
1721 		goto out;
1722 
1723 	for (i = 0; i < naggvars; i++) {
1724 		int ndx = i + sortpos;
1725 
1726 		if (ndx >= naggvars)
1727 			ndx -= naggvars;
1728 
1729 		aggvar = aggvars[ndx];
1730 		assert(aggvar <= max);
1731 
1732 		if (map[aggvar]) {
1733 			/*
1734 			 * We have an aggregation variable that is present
1735 			 * more than once in the array of aggregation
1736 			 * variables.  While it's unclear why one might want
1737 			 * to do this, it's legal.  To support this construct,
1738 			 * we will allocate a remap that will indicate the
1739 			 * position from which this aggregation variable
1740 			 * should be pulled.  (That is, where the remap will
1741 			 * map from one position to another.)
1742 			 */
1743 			if (remap == NULL) {
1744 				remap = dt_zalloc(dtp, naggvars * sizeof (int));
1745 
1746 				if (remap == NULL)
1747 					goto out;
1748 			}
1749 
1750 			/*
1751 			 * Given that the variable is already present, assert
1752 			 * that following through the mapping and adjusting
1753 			 * for the sort position yields the same aggregation
1754 			 * variable ID.
1755 			 */
1756 			assert(aggvars[(map[aggvar] - 1 + sortpos) %
1757 			    naggvars] == aggvars[ndx]);
1758 
1759 			remap[i] = map[aggvar];
1760 			continue;
1761 		}
1762 
1763 		map[aggvar] = i + 1;
1764 	}
1765 
1766 	/*
1767 	 * We need to take two passes over the data to size our allocation, so
1768 	 * we'll use the first pass to also fill in the zero-filled data to be
1769 	 * used to properly format a zero-valued aggregation.
1770 	 */
1771 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
1772 		dtrace_aggvarid_t id;
1773 		int ndx;
1774 
1775 		if ((id = dt_aggregate_aggvarid(h)) > max || !(ndx = map[id]))
1776 			continue;
1777 
1778 		if (zaggdata[ndx - 1].dtahe_size == 0) {
1779 			zaggdata[ndx - 1].dtahe_size = h->dtahe_size;
1780 			zaggdata[ndx - 1].dtahe_data = h->dtahe_data;
1781 		}
1782 
1783 		nentries++;
1784 	}
1785 
1786 	if (nentries == 0) {
1787 		/*
1788 		 * We couldn't find any entries; there is nothing else to do.
1789 		 */
1790 		rval = 0;
1791 		goto out;
1792 	}
1793 
1794 	/*
1795 	 * Before we sort the data, we're going to look for any holes in our
1796 	 * zero-filled data.  This will occur if an aggregation variable that
1797 	 * we are being asked to print has not yet been assigned the result of
1798 	 * any aggregating action for _any_ tuple.  The issue becomes that we
1799 	 * would like a zero value to be printed for all columns for this
1800 	 * aggregation, but without any record description, we don't know the
1801 	 * aggregating action that corresponds to the aggregation variable.  To
1802 	 * try to find a match, we're simply going to lookup aggregation IDs
1803 	 * (which are guaranteed to be contiguous and to start from 1), looking
1804 	 * for the specified aggregation variable ID.  If we find a match,
1805 	 * we'll use that.  If we iterate over all aggregation IDs and don't
1806 	 * find a match, then we must be an anonymous enabling.  (Anonymous
1807 	 * enablings can't currently derive either aggregation variable IDs or
1808 	 * aggregation variable names given only an aggregation ID.)  In this
1809 	 * obscure case (anonymous enabling, multiple aggregation printa() with
1810 	 * some aggregations not represented for any tuple), our defined
1811 	 * behavior is that the zero will be printed in the format of the first
1812 	 * aggregation variable that contains any non-zero value.
1813 	 */
1814 	for (i = 0; i < naggvars; i++) {
1815 		if (zaggdata[i].dtahe_size == 0) {
1816 			dtrace_aggvarid_t aggvar;
1817 
1818 			aggvar = aggvars[(i - sortpos + naggvars) % naggvars];
1819 			assert(zaggdata[i].dtahe_data.dtada_data == NULL);
1820 
1821 			for (j = DTRACE_AGGIDNONE + 1; ; j++) {
1822 				dtrace_aggdesc_t *agg;
1823 				dtrace_aggdata_t *aggdata;
1824 
1825 				if (dt_aggid_lookup(dtp, j, &agg) != 0)
1826 					break;
1827 
1828 				if (agg->dtagd_varid != aggvar)
1829 					continue;
1830 
1831 				/*
1832 				 * We have our description -- now we need to
1833 				 * cons up the zaggdata entry for it.
1834 				 */
1835 				aggdata = &zaggdata[i].dtahe_data;
1836 				aggdata->dtada_size = agg->dtagd_size;
1837 				aggdata->dtada_desc = agg;
1838 				aggdata->dtada_handle = dtp;
1839 				(void) dt_epid_lookup(dtp, agg->dtagd_epid,
1840 				    &aggdata->dtada_edesc,
1841 				    &aggdata->dtada_pdesc);
1842 				aggdata->dtada_normal = 1;
1843 				zaggdata[i].dtahe_hashval = 0;
1844 				zaggdata[i].dtahe_size = agg->dtagd_size;
1845 				break;
1846 			}
1847 
1848 			if (zaggdata[i].dtahe_size == 0) {
1849 				caddr_t data;
1850 
1851 				/*
1852 				 * We couldn't find this aggregation, meaning
1853 				 * that we have never seen it before for any
1854 				 * tuple _and_ this is an anonymous enabling.
1855 				 * That is, we're in the obscure case outlined
1856 				 * above.  In this case, our defined behavior
1857 				 * is to format the data in the format of the
1858 				 * first non-zero aggregation -- of which, of
1859 				 * course, we know there to be at least one
1860 				 * (or nentries would have been zero).
1861 				 */
1862 				for (j = 0; j < naggvars; j++) {
1863 					if (zaggdata[j].dtahe_size != 0)
1864 						break;
1865 				}
1866 
1867 				assert(j < naggvars);
1868 				zaggdata[i] = zaggdata[j];
1869 
1870 				data = zaggdata[i].dtahe_data.dtada_data;
1871 				assert(data != NULL);
1872 			}
1873 		}
1874 	}
1875 
1876 	/*
1877 	 * Now we need to allocate our zero-filled data for use for
1878 	 * aggregations that don't have a value corresponding to a given key.
1879 	 */
1880 	for (i = 0; i < naggvars; i++) {
1881 		dtrace_aggdata_t *aggdata = &zaggdata[i].dtahe_data;
1882 		dtrace_aggdesc_t *aggdesc = aggdata->dtada_desc;
1883 		dtrace_recdesc_t *rec;
1884 		uint64_t larg;
1885 		caddr_t zdata;
1886 
1887 		zsize = zaggdata[i].dtahe_size;
1888 		assert(zsize != 0);
1889 
1890 		if ((zdata = dt_zalloc(dtp, zsize)) == NULL) {
1891 			/*
1892 			 * If we failed to allocated some zero-filled data, we
1893 			 * need to zero out the remaining dtada_data pointers
1894 			 * to prevent the wrong data from being freed below.
1895 			 */
1896 			for (j = i; j < naggvars; j++)
1897 				zaggdata[j].dtahe_data.dtada_data = NULL;
1898 			goto out;
1899 		}
1900 
1901 		aggvar = aggvars[(i - sortpos + naggvars) % naggvars];
1902 
1903 		/*
1904 		 * First, the easy bit.  To maintain compatibility with
1905 		 * consumers that pull the compiler-generated ID out of the
1906 		 * data, we put that ID at the top of the zero-filled data.
1907 		 */
1908 		rec = &aggdesc->dtagd_rec[0];
1909 		/* LINTED - alignment */
1910 		*((dtrace_aggvarid_t *)(zdata + rec->dtrd_offset)) = aggvar;
1911 
1912 		rec = &aggdesc->dtagd_rec[aggdesc->dtagd_nrecs - 1];
1913 
1914 		/*
1915 		 * Now for the more complicated part.  If (and only if) this
1916 		 * is an lquantize() aggregating action, zero-filled data is
1917 		 * not equivalent to an empty record:  we must also get the
1918 		 * parameters for the lquantize().
1919 		 */
1920 		if (rec->dtrd_action == DTRACEAGG_LQUANTIZE) {
1921 			if (aggdata->dtada_data != NULL) {
1922 				/*
1923 				 * The easier case here is if we actually have
1924 				 * some prototype data -- in which case we
1925 				 * manually dig it out of the aggregation
1926 				 * record.
1927 				 */
1928 				/* LINTED - alignment */
1929 				larg = *((uint64_t *)(aggdata->dtada_data +
1930 				    rec->dtrd_offset));
1931 			} else {
1932 				/*
1933 				 * We don't have any prototype data.  As a
1934 				 * result, we know that we _do_ have the
1935 				 * compiler-generated information.  (If this
1936 				 * were an anonymous enabling, all of our
1937 				 * zero-filled data would have prototype data
1938 				 * -- either directly or indirectly.) So as
1939 				 * gross as it is, we'll grovel around in the
1940 				 * compiler-generated information to find the
1941 				 * lquantize() parameters.
1942 				 */
1943 				dtrace_stmtdesc_t *sdp;
1944 				dt_ident_t *aid;
1945 				dt_idsig_t *isp;
1946 
1947 				sdp = (dtrace_stmtdesc_t *)(uintptr_t)
1948 				    aggdesc->dtagd_rec[0].dtrd_uarg;
1949 				aid = sdp->dtsd_aggdata;
1950 				isp = (dt_idsig_t *)aid->di_data;
1951 				assert(isp->dis_auxinfo != 0);
1952 				larg = isp->dis_auxinfo;
1953 			}
1954 
1955 			/* LINTED - alignment */
1956 			*((uint64_t *)(zdata + rec->dtrd_offset)) = larg;
1957 		}
1958 
1959 		aggdata->dtada_data = zdata;
1960 	}
1961 
1962 	/*
1963 	 * Now that we've dealt with setting up our zero-filled data, we can
1964 	 * allocate our sorted array, and take another pass over the data to
1965 	 * fill it.
1966 	 */
1967 	sorted = dt_alloc(dtp, nentries * sizeof (dt_ahashent_t *));
1968 
1969 	if (sorted == NULL)
1970 		goto out;
1971 
1972 	for (h = hash->dtah_all, i = 0; h != NULL; h = h->dtahe_nextall) {
1973 		dtrace_aggvarid_t id;
1974 
1975 		if ((id = dt_aggregate_aggvarid(h)) > max || !map[id])
1976 			continue;
1977 
1978 		sorted[i++] = h;
1979 	}
1980 
1981 	assert(i == nentries);
1982 
1983 	/*
1984 	 * We've loaded our array; now we need to sort by value to allow us
1985 	 * to create bundles of like value.  We're going to acquire the
1986 	 * dt_qsort_lock here, and hold it across all of our subsequent
1987 	 * comparison and sorting.
1988 	 */
1989 	(void) pthread_mutex_lock(&dt_qsort_lock);
1990 
1991 	qsort(sorted, nentries, sizeof (dt_ahashent_t *),
1992 	    dt_aggregate_keyvarcmp);
1993 
1994 	/*
1995 	 * Now we need to go through and create bundles.  Because the number
1996 	 * of bundles is bounded by the size of the sorted array, we're going
1997 	 * to reuse the underlying storage.  And note that "bundle" is an
1998 	 * array of pointers to arrays of pointers to dt_ahashent_t -- making
1999 	 * its type (regrettably) "dt_ahashent_t ***".  (Regrettable because
2000 	 * '*' -- like '_' and 'X' -- should never appear in triplicate in
2001 	 * an ideal world.)
2002 	 */
2003 	bundle = (dt_ahashent_t ***)sorted;
2004 
2005 	for (i = 1, start = 0; i <= nentries; i++) {
2006 		if (i < nentries &&
2007 		    dt_aggregate_keycmp(&sorted[i], &sorted[i - 1]) == 0)
2008 			continue;
2009 
2010 		/*
2011 		 * We have a bundle boundary.  Everything from start to
2012 		 * (i - 1) belongs in one bundle.
2013 		 */
2014 		assert(i - start <= naggvars);
2015 		bundlesize = (naggvars + 2) * sizeof (dt_ahashent_t *);
2016 
2017 		if ((nbundle = dt_zalloc(dtp, bundlesize)) == NULL) {
2018 			(void) pthread_mutex_unlock(&dt_qsort_lock);
2019 			goto out;
2020 		}
2021 
2022 		for (j = start; j < i; j++) {
2023 			dtrace_aggvarid_t id = dt_aggregate_aggvarid(sorted[j]);
2024 
2025 			assert(id <= max);
2026 			assert(map[id] != 0);
2027 			assert(map[id] - 1 < naggvars);
2028 			assert(nbundle[map[id] - 1] == NULL);
2029 			nbundle[map[id] - 1] = sorted[j];
2030 
2031 			if (nbundle[naggvars] == NULL)
2032 				nbundle[naggvars] = sorted[j];
2033 		}
2034 
2035 		for (j = 0; j < naggvars; j++) {
2036 			if (nbundle[j] != NULL)
2037 				continue;
2038 
2039 			/*
2040 			 * Before we assume that this aggregation variable
2041 			 * isn't present (and fall back to using the
2042 			 * zero-filled data allocated earlier), check the
2043 			 * remap.  If we have a remapping, we'll drop it in
2044 			 * here.  Note that we might be remapping an
2045 			 * aggregation variable that isn't present for this
2046 			 * key; in this case, the aggregation data that we
2047 			 * copy will point to the zeroed data.
2048 			 */
2049 			if (remap != NULL && remap[j]) {
2050 				assert(remap[j] - 1 < j);
2051 				assert(nbundle[remap[j] - 1] != NULL);
2052 				nbundle[j] = nbundle[remap[j] - 1];
2053 			} else {
2054 				nbundle[j] = &zaggdata[j];
2055 			}
2056 		}
2057 
2058 		bundle[nbundles++] = nbundle;
2059 		start = i;
2060 	}
2061 
2062 	/*
2063 	 * Now we need to re-sort based on the first value.
2064 	 */
2065 	dt_aggregate_qsort(dtp, bundle, nbundles, sizeof (dt_ahashent_t **),
2066 	    dt_aggregate_bundlecmp);
2067 
2068 	(void) pthread_mutex_unlock(&dt_qsort_lock);
2069 
2070 	/*
2071 	 * We're done!  Now we just need to go back over the sorted bundles,
2072 	 * calling the function.
2073 	 */
2074 	data = alloca((naggvars + 1) * sizeof (dtrace_aggdata_t *));
2075 
2076 	for (i = 0; i < nbundles; i++) {
2077 		for (j = 0; j < naggvars; j++)
2078 			data[j + 1] = NULL;
2079 
2080 		for (j = 0; j < naggvars; j++) {
2081 			int ndx = j - sortpos;
2082 
2083 			if (ndx < 0)
2084 				ndx += naggvars;
2085 
2086 			assert(bundle[i][ndx] != NULL);
2087 			data[j + 1] = &bundle[i][ndx]->dtahe_data;
2088 		}
2089 
2090 		for (j = 0; j < naggvars; j++)
2091 			assert(data[j + 1] != NULL);
2092 
2093 		/*
2094 		 * The representative key is the last element in the bundle.
2095 		 * Assert that we have one, and then set it to be the first
2096 		 * element of data.
2097 		 */
2098 		assert(bundle[i][j] != NULL);
2099 		data[0] = &bundle[i][j]->dtahe_data;
2100 
2101 		if ((rval = func(data, naggvars + 1, arg)) == -1)
2102 			goto out;
2103 	}
2104 
2105 	rval = 0;
2106 out:
2107 	for (i = 0; i < nbundles; i++)
2108 		dt_free(dtp, bundle[i]);
2109 
2110 	if (zaggdata != NULL) {
2111 		for (i = 0; i < naggvars; i++)
2112 			dt_free(dtp, zaggdata[i].dtahe_data.dtada_data);
2113 	}
2114 
2115 	dt_free(dtp, zaggdata);
2116 	dt_free(dtp, sorted);
2117 	dt_free(dtp, remap);
2118 	dt_free(dtp, map);
2119 
2120 	return (rval);
2121 }
2122 
2123 int
2124 dtrace_aggregate_print(dtrace_hdl_t *dtp, FILE *fp,
2125     dtrace_aggregate_walk_f *func)
2126 {
2127 	dt_print_aggdata_t pd;
2128 
2129 	bzero(&pd, sizeof (pd));
2130 
2131 	pd.dtpa_dtp = dtp;
2132 	pd.dtpa_fp = fp;
2133 	pd.dtpa_allunprint = 1;
2134 
2135 	if (func == NULL)
2136 		func = dtrace_aggregate_walk_sorted;
2137 
2138 	if (dtp->dt_oformat) {
2139 		if ((*func)(dtp, dt_format_agg, &pd) == -1)
2140 			return (dt_set_errno(dtp, dtp->dt_errno));
2141 	} else {
2142 		if ((*func)(dtp, dt_print_agg, &pd) == -1)
2143 			return (dt_set_errno(dtp, dtp->dt_errno));
2144 	}
2145 
2146 	return (0);
2147 }
2148 
2149 void
2150 dtrace_aggregate_clear(dtrace_hdl_t *dtp)
2151 {
2152 	dt_aggregate_t *agp = &dtp->dt_aggregate;
2153 	dt_ahash_t *hash = &agp->dtat_hash;
2154 	dt_ahashent_t *h;
2155 	dtrace_aggdata_t *data;
2156 	dtrace_aggdesc_t *aggdesc;
2157 	dtrace_recdesc_t *rec;
2158 	int i, max_cpus = agp->dtat_maxcpu;
2159 
2160 	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
2161 		aggdesc = h->dtahe_data.dtada_desc;
2162 		rec = &aggdesc->dtagd_rec[aggdesc->dtagd_nrecs - 1];
2163 		data = &h->dtahe_data;
2164 
2165 		bzero(&data->dtada_data[rec->dtrd_offset], rec->dtrd_size);
2166 
2167 		if (data->dtada_percpu == NULL)
2168 			continue;
2169 
2170 		for (i = 0; i < max_cpus; i++)
2171 			bzero(data->dtada_percpu[i], rec->dtrd_size);
2172 	}
2173 }
2174 
2175 void
2176 dt_aggregate_destroy(dtrace_hdl_t *dtp)
2177 {
2178 	dt_aggregate_t *agp = &dtp->dt_aggregate;
2179 	dt_ahash_t *hash = &agp->dtat_hash;
2180 	dt_ahashent_t *h, *next;
2181 	dtrace_aggdata_t *aggdata;
2182 	int i, max_cpus = agp->dtat_maxcpu;
2183 
2184 	if (hash->dtah_hash == NULL) {
2185 		assert(hash->dtah_all == NULL);
2186 	} else {
2187 		free(hash->dtah_hash);
2188 
2189 		for (h = hash->dtah_all; h != NULL; h = next) {
2190 			next = h->dtahe_nextall;
2191 
2192 			aggdata = &h->dtahe_data;
2193 
2194 			if (aggdata->dtada_percpu != NULL) {
2195 				for (i = 0; i < max_cpus; i++)
2196 					free(aggdata->dtada_percpu[i]);
2197 				free(aggdata->dtada_percpu);
2198 			}
2199 
2200 			free(aggdata->dtada_data);
2201 			free(h);
2202 		}
2203 
2204 		hash->dtah_hash = NULL;
2205 		hash->dtah_all = NULL;
2206 		hash->dtah_size = 0;
2207 	}
2208 
2209 	free(agp->dtat_buf.dtbd_data);
2210 	free(agp->dtat_cpus);
2211 }
2212