libdtrace/common/dt_aggregate.c

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */

/*
 * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#pragma ident	"%Z%%M%	%I%	%E% SMI"

#include <stdlib.h>
#include <strings.h>
#include <errno.h>
#include <unistd.h>
#include <dt_impl.h>
#include <assert.h>
#include <alloca.h>
#include <limits.h>

#define	DTRACE_AHASHSIZE	32779		/* big 'ol prime */

/*
 * Because qsort(3C) does not allow an argument to be passed to a comparison
 * function, the variables that affect comparison must regrettably be global;
 * they are protected by a global static lock, dt_qsort_lock.
 */
static pthread_mutex_t dt_qsort_lock = PTHREAD_MUTEX_INITIALIZER;

static int dt_revsort;
static int dt_keysort;
static int dt_keypos;

#define	DT_LESSTHAN	(dt_revsort == 0 ? -1 : 1)
#define	DT_GREATERTHAN	(dt_revsort == 0 ? 1 : -1)

static void
dt_aggregate_count(int64_t *existing, int64_t *new, size_t size)
{
	int i;

	for (i = 0; i < size / sizeof (int64_t); i++)
		existing[i] = existing[i] + new[i];
}

static int
dt_aggregate_countcmp(int64_t *lhs, int64_t *rhs)
{
	int64_t lvar = *lhs;
	int64_t rvar = *rhs;

	if (lvar < rvar)
		return (DT_LESSTHAN);

	if (lvar > rvar)
		return (DT_GREATERTHAN);

	return (0);
}

/*ARGSUSED*/
static void
dt_aggregate_min(int64_t *existing, int64_t *new, size_t size)
{
	if (*new < *existing)
		*existing = *new;
}

/*ARGSUSED*/
static void
dt_aggregate_max(int64_t *existing, int64_t *new, size_t size)
{
	if (*new > *existing)
		*existing = *new;
}

static int
dt_aggregate_averagecmp(int64_t *lhs, int64_t *rhs)
{
	int64_t lavg = lhs[0] ? (lhs[1] / lhs[0]) : 0;
	int64_t ravg = rhs[0] ? (rhs[1] / rhs[0]) : 0;

	if (lavg < ravg)
		return (DT_LESSTHAN);

	if (lavg > ravg)
		return (DT_GREATERTHAN);

	return (0);
}

static int
dt_aggregate_stddevcmp(int64_t *lhs, int64_t *rhs)
{
	uint64_t lsd = dt_stddev((uint64_t *)lhs, 1);
	uint64_t rsd = dt_stddev((uint64_t *)rhs, 1);

	if (lsd < rsd)
		return (DT_LESSTHAN);

	if (lsd > rsd)
		return (DT_GREATERTHAN);

	return (0);
}

/*ARGSUSED*/
static void
dt_aggregate_lquantize(int64_t *existing, int64_t *new, size_t size)
{
	int64_t arg = *existing++;
	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
	int i;

	for (i = 0; i <= levels + 1; i++)
		existing[i] = existing[i] + new[i + 1];
}

static long double
dt_aggregate_lquantizedsum(int64_t *lquanta)
{
	int64_t arg = *lquanta++;
	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg), i;
	long double total = (long double)lquanta[0] * (long double)(base - 1);

	for (i = 0; i < levels; base += step, i++)
		total += (long double)lquanta[i + 1] * (long double)base;

	return (total + (long double)lquanta[levels + 1] *
	    (long double)(base + 1));
}

static int64_t
dt_aggregate_lquantizedzero(int64_t *lquanta)
{
	int64_t arg = *lquanta++;
	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg), i;

	if (base - 1 == 0)
		return (lquanta[0]);

	for (i = 0; i < levels; base += step, i++) {
		if (base != 0)
			continue;

		return (lquanta[i + 1]);
	}

	if (base + 1 == 0)
		return (lquanta[levels + 1]);

	return (0);
}

static int
dt_aggregate_lquantizedcmp(int64_t *lhs, int64_t *rhs)
{
	long double lsum = dt_aggregate_lquantizedsum(lhs);
	long double rsum = dt_aggregate_lquantizedsum(rhs);
	int64_t lzero, rzero;

	if (lsum < rsum)
		return (DT_LESSTHAN);

	if (lsum > rsum)
		return (DT_GREATERTHAN);

	/*
	 * If they're both equal, then we will compare based on the weights at
	 * zero.  If the weights at zero are equal (or if zero is not within
	 * the range of the linear quantization), then this will be judged a
	 * tie and will be resolved based on the key comparison.
	 */
	lzero = dt_aggregate_lquantizedzero(lhs);
	rzero = dt_aggregate_lquantizedzero(rhs);

	if (lzero < rzero)
		return (DT_LESSTHAN);

	if (lzero > rzero)
		return (DT_GREATERTHAN);

	return (0);
}

static int
dt_aggregate_quantizedcmp(int64_t *lhs, int64_t *rhs)
{
	int nbuckets = DTRACE_QUANTIZE_NBUCKETS, i;
	long double ltotal = 0, rtotal = 0;
	int64_t lzero, rzero;

	for (i = 0; i < nbuckets; i++) {
		int64_t bucketval = DTRACE_QUANTIZE_BUCKETVAL(i);

		if (bucketval == 0) {
			lzero = lhs[i];
			rzero = rhs[i];
		}

		ltotal += (long double)bucketval * (long double)lhs[i];
		rtotal += (long double)bucketval * (long double)rhs[i];
	}

	if (ltotal < rtotal)
		return (DT_LESSTHAN);

	if (ltotal > rtotal)
		return (DT_GREATERTHAN);

	/*
	 * If they're both equal, then we will compare based on the weights at
	 * zero.  If the weights at zero are equal, then this will be judged a
	 * tie and will be resolved based on the key comparison.
	 */
	if (lzero < rzero)
		return (DT_LESSTHAN);

	if (lzero > rzero)
		return (DT_GREATERTHAN);

	return (0);
}

static void
dt_aggregate_usym(dtrace_hdl_t *dtp, uint64_t *data)
{
	uint64_t pid = data[0];
	uint64_t *pc = &data[1];
	struct ps_prochandle *P;
	GElf_Sym sym;

	if (dtp->dt_vector != NULL)
		return;

	if ((P = dt_proc_grab(dtp, pid, PGRAB_RDONLY | PGRAB_FORCE, 0)) == NULL)
		return;

	dt_proc_lock(dtp, P);

	if (Plookup_by_addr(P, *pc, NULL, 0, &sym) == 0)
		*pc = sym.st_value;

	dt_proc_unlock(dtp, P);
	dt_proc_release(dtp, P);
}

static void
dt_aggregate_umod(dtrace_hdl_t *dtp, uint64_t *data)
{
	uint64_t pid = data[0];
	uint64_t *pc = &data[1];
	struct ps_prochandle *P;
	const prmap_t *map;

	if (dtp->dt_vector != NULL)
		return;

	if ((P = dt_proc_grab(dtp, pid, PGRAB_RDONLY | PGRAB_FORCE, 0)) == NULL)
		return;

	dt_proc_lock(dtp, P);

	if ((map = Paddr_to_map(P, *pc)) != NULL)
		*pc = map->pr_vaddr;

	dt_proc_unlock(dtp, P);
	dt_proc_release(dtp, P);
}

static void
dt_aggregate_sym(dtrace_hdl_t *dtp, uint64_t *data)
{
	GElf_Sym sym;
	uint64_t *pc = data;

	if (dtrace_lookup_by_addr(dtp, *pc, &sym, NULL) == 0)
		*pc = sym.st_value;
}

static void
dt_aggregate_mod(dtrace_hdl_t *dtp, uint64_t *data)
{
	uint64_t *pc = data;
	dt_module_t *dmp;

	if (dtp->dt_vector != NULL) {
		/*
		 * We don't have a way of just getting the module for a
		 * vectored open, and it doesn't seem to be worth defining
		 * one.  This means that use of mod() won't get true
		 * aggregation in the postmortem case (some modules may
		 * appear more than once in aggregation output).  It seems
		 * unlikely that anyone will ever notice or care...
		 */
		return;
	}

	for (dmp = dt_list_next(&dtp->dt_modlist); dmp != NULL;
	    dmp = dt_list_next(dmp)) {
		if (*pc - dmp->dm_text_va < dmp->dm_text_size) {
			*pc = dmp->dm_text_va;
			return;
		}
	}
}

static dtrace_aggvarid_t
dt_aggregate_aggvarid(dt_ahashent_t *ent)
{
	dtrace_aggdesc_t *agg = ent->dtahe_data.dtada_desc;
	caddr_t data = ent->dtahe_data.dtada_data;
	dtrace_recdesc_t *rec = agg->dtagd_rec;

	/*
	 * First, we'll check the variable ID in the aggdesc.  If it's valid,
	 * we'll return it.  If not, we'll use the compiler-generated ID
	 * present as the first record.
	 */
	if (agg->dtagd_varid != DTRACE_AGGVARIDNONE)
		return (agg->dtagd_varid);

	agg->dtagd_varid = *((dtrace_aggvarid_t *)(uintptr_t)(data +
	    rec->dtrd_offset));

	return (agg->dtagd_varid);
}


static int
dt_aggregate_snap_cpu(dtrace_hdl_t *dtp, processorid_t cpu)
{
	dtrace_epid_t id;
	uint64_t hashval;
	size_t offs, roffs, size, ndx;
	int i, j, rval;
	caddr_t addr, data;
	dtrace_recdesc_t *rec;
	dt_aggregate_t *agp = &dtp->dt_aggregate;
	dtrace_aggdesc_t *agg;
	dt_ahash_t *hash = &agp->dtat_hash;
	dt_ahashent_t *h;
	dtrace_bufdesc_t b = agp->dtat_buf, *buf = &b;
	dtrace_aggdata_t *aggdata;
	int flags = agp->dtat_flags;

	buf->dtbd_cpu = cpu;

	if (dt_ioctl(dtp, DTRACEIOC_AGGSNAP, buf) == -1) {
		if (errno == ENOENT) {
			/*
			 * If that failed with ENOENT, it may be because the
			 * CPU was unconfigured.  This is okay; we'll just
			 * do nothing but return success.
			 */
			return (0);
		}

		return (dt_set_errno(dtp, errno));
	}

	if (buf->dtbd_drops != 0) {
		if (dt_handle_cpudrop(dtp, cpu,
		    DTRACEDROP_AGGREGATION, buf->dtbd_drops) == -1)
			return (-1);
	}

	if (buf->dtbd_size == 0)
		return (0);

	if (hash->dtah_hash == NULL) {
		size_t size;

		hash->dtah_size = DTRACE_AHASHSIZE;
		size = hash->dtah_size * sizeof (dt_ahashent_t *);

		if ((hash->dtah_hash = malloc(size)) == NULL)
			return (dt_set_errno(dtp, EDT_NOMEM));

		bzero(hash->dtah_hash, size);
	}

	for (offs = 0; offs < buf->dtbd_size; ) {
		/*
		 * We're guaranteed to have an ID.
		 */
		id = *((dtrace_epid_t *)((uintptr_t)buf->dtbd_data +
		    (uintptr_t)offs));

		if (id == DTRACE_AGGIDNONE) {
			/*
			 * This is filler to assure proper alignment of the
			 * next record; we simply ignore it.
			 */
			offs += sizeof (id);
			continue;
		}

		if ((rval = dt_aggid_lookup(dtp, id, &agg)) != 0)
			return (rval);

		addr = buf->dtbd_data + offs;
		size = agg->dtagd_size;
		hashval = 0;

		for (j = 0; j < agg->dtagd_nrecs - 1; j++) {
			rec = &agg->dtagd_rec[j];
			roffs = rec->dtrd_offset;

			switch (rec->dtrd_action) {
			case DTRACEACT_USYM:
				dt_aggregate_usym(dtp,
				    /* LINTED - alignment */
				    (uint64_t *)&addr[roffs]);
				break;

			case DTRACEACT_UMOD:
				dt_aggregate_umod(dtp,
				    /* LINTED - alignment */
				    (uint64_t *)&addr[roffs]);
				break;

			case DTRACEACT_SYM:
				/* LINTED - alignment */
				dt_aggregate_sym(dtp, (uint64_t *)&addr[roffs]);
				break;

			case DTRACEACT_MOD:
				/* LINTED - alignment */
				dt_aggregate_mod(dtp, (uint64_t *)&addr[roffs]);
				break;

			default:
				break;
			}

			for (i = 0; i < rec->dtrd_size; i++)
				hashval += addr[roffs + i];
		}

		ndx = hashval % hash->dtah_size;

		for (h = hash->dtah_hash[ndx]; h != NULL; h = h->dtahe_next) {
			if (h->dtahe_hashval != hashval)
				continue;

			if (h->dtahe_size != size)
				continue;

			aggdata = &h->dtahe_data;
			data = aggdata->dtada_data;

			for (j = 0; j < agg->dtagd_nrecs - 1; j++) {
				rec = &agg->dtagd_rec[j];
				roffs = rec->dtrd_offset;

				for (i = 0; i < rec->dtrd_size; i++)
					if (addr[roffs + i] != data[roffs + i])
						goto hashnext;
			}

			/*
			 * We found it.  Now we need to apply the aggregating
			 * action on the data here.
			 */
			rec = &agg->dtagd_rec[agg->dtagd_nrecs - 1];
			roffs = rec->dtrd_offset;
			/* LINTED - alignment */
			h->dtahe_aggregate((int64_t *)&data[roffs],
			    /* LINTED - alignment */
			    (int64_t *)&addr[roffs], rec->dtrd_size);

			/*
			 * If we're keeping per CPU data, apply the aggregating
			 * action there as well.
			 */
			if (aggdata->dtada_percpu != NULL) {
				data = aggdata->dtada_percpu[cpu];

				/* LINTED - alignment */
				h->dtahe_aggregate((int64_t *)data,
				    /* LINTED - alignment */
				    (int64_t *)&addr[roffs], rec->dtrd_size);
			}

			goto bufnext;
hashnext:
			continue;
		}

		/*
		 * If we're here, we couldn't find an entry for this record.
		 */
		if ((h = malloc(sizeof (dt_ahashent_t))) == NULL)
			return (dt_set_errno(dtp, EDT_NOMEM));
		bzero(h, sizeof (dt_ahashent_t));
		aggdata = &h->dtahe_data;

		if ((aggdata->dtada_data = malloc(size)) == NULL) {
			free(h);
			return (dt_set_errno(dtp, EDT_NOMEM));
		}

		bcopy(addr, aggdata->dtada_data, size);
		aggdata->dtada_size = size;
		aggdata->dtada_desc = agg;
		aggdata->dtada_handle = dtp;
		(void) dt_epid_lookup(dtp, agg->dtagd_epid,
		    &aggdata->dtada_edesc, &aggdata->dtada_pdesc);
		aggdata->dtada_normal = 1;

		h->dtahe_hashval = hashval;
		h->dtahe_size = size;
		(void) dt_aggregate_aggvarid(h);

		rec = &agg->dtagd_rec[agg->dtagd_nrecs - 1];

		if (flags & DTRACE_A_PERCPU) {
			int max_cpus = agp->dtat_maxcpu;
			caddr_t *percpu = malloc(max_cpus * sizeof (caddr_t));

			if (percpu == NULL) {
				free(aggdata->dtada_data);
				free(h);
				return (dt_set_errno(dtp, EDT_NOMEM));
			}

			for (j = 0; j < max_cpus; j++) {
				percpu[j] = malloc(rec->dtrd_size);

				if (percpu[j] == NULL) {
					while (--j >= 0)
						free(percpu[j]);

					free(aggdata->dtada_data);
					free(h);
					return (dt_set_errno(dtp, EDT_NOMEM));
				}

				if (j == cpu) {
					bcopy(&addr[rec->dtrd_offset],
					    percpu[j], rec->dtrd_size);
				} else {
					bzero(percpu[j], rec->dtrd_size);
				}
			}

			aggdata->dtada_percpu = percpu;
		}

		switch (rec->dtrd_action) {
		case DTRACEAGG_MIN:
			h->dtahe_aggregate = dt_aggregate_min;
			break;

		case DTRACEAGG_MAX:
			h->dtahe_aggregate = dt_aggregate_max;
			break;

		case DTRACEAGG_LQUANTIZE:
			h->dtahe_aggregate = dt_aggregate_lquantize;
			break;

		case DTRACEAGG_COUNT:
		case DTRACEAGG_SUM:
		case DTRACEAGG_AVG:
		case DTRACEAGG_STDDEV:
		case DTRACEAGG_QUANTIZE:
			h->dtahe_aggregate = dt_aggregate_count;
			break;

		default:
			return (dt_set_errno(dtp, EDT_BADAGG));
		}

		if (hash->dtah_hash[ndx] != NULL)
			hash->dtah_hash[ndx]->dtahe_prev = h;

		h->dtahe_next = hash->dtah_hash[ndx];
		hash->dtah_hash[ndx] = h;

		if (hash->dtah_all != NULL)
			hash->dtah_all->dtahe_prevall = h;

		h->dtahe_nextall = hash->dtah_all;
		hash->dtah_all = h;
bufnext:
		offs += agg->dtagd_size;
	}

	return (0);
}

int
dtrace_aggregate_snap(dtrace_hdl_t *dtp)
{
	int i, rval;
	dt_aggregate_t *agp = &dtp->dt_aggregate;
	hrtime_t now = gethrtime();
	dtrace_optval_t interval = dtp->dt_options[DTRACEOPT_AGGRATE];

	if (dtp->dt_lastagg != 0) {
		if (now - dtp->dt_lastagg < interval)
			return (0);

		dtp->dt_lastagg += interval;
	} else {
		dtp->dt_lastagg = now;
	}

	if (!dtp->dt_active)
		return (dt_set_errno(dtp, EINVAL));

	if (agp->dtat_buf.dtbd_size == 0)
		return (0);

	for (i = 0; i < agp->dtat_ncpus; i++) {
		if (rval = dt_aggregate_snap_cpu(dtp, agp->dtat_cpus[i]))
			return (rval);
	}

	return (0);
}

static int
dt_aggregate_hashcmp(const void *lhs, const void *rhs)
{
	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
	dtrace_aggdesc_t *lagg = lh->dtahe_data.dtada_desc;
	dtrace_aggdesc_t *ragg = rh->dtahe_data.dtada_desc;

	if (lagg->dtagd_nrecs < ragg->dtagd_nrecs)
		return (DT_LESSTHAN);

	if (lagg->dtagd_nrecs > ragg->dtagd_nrecs)
		return (DT_GREATERTHAN);

	return (0);
}

static int
dt_aggregate_varcmp(const void *lhs, const void *rhs)
{
	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
	dtrace_aggvarid_t lid, rid;

	lid = dt_aggregate_aggvarid(lh);
	rid = dt_aggregate_aggvarid(rh);

	if (lid < rid)
		return (DT_LESSTHAN);

	if (lid > rid)
		return (DT_GREATERTHAN);

	return (0);
}

static int
dt_aggregate_keycmp(const void *lhs, const void *rhs)
{
	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
	dtrace_aggdesc_t *lagg = lh->dtahe_data.dtada_desc;
	dtrace_aggdesc_t *ragg = rh->dtahe_data.dtada_desc;
	dtrace_recdesc_t *lrec, *rrec;
	char *ldata, *rdata;
	int rval, i, j, keypos, nrecs;

	if ((rval = dt_aggregate_hashcmp(lhs, rhs)) != 0)
		return (rval);

	nrecs = lagg->dtagd_nrecs - 1;
	assert(nrecs == ragg->dtagd_nrecs - 1);

	keypos = dt_keypos + 1 >= nrecs ? 0 : dt_keypos;

	for (i = 1; i < nrecs; i++) {
		uint64_t lval, rval;
		int ndx = i + keypos;

		if (ndx >= nrecs)
			ndx = ndx - nrecs + 1;

		lrec = &lagg->dtagd_rec[ndx];
		rrec = &ragg->dtagd_rec[ndx];

		ldata = lh->dtahe_data.dtada_data + lrec->dtrd_offset;
		rdata = rh->dtahe_data.dtada_data + rrec->dtrd_offset;

		if (lrec->dtrd_size < rrec->dtrd_size)
			return (DT_LESSTHAN);

		if (lrec->dtrd_size > rrec->dtrd_size)
			return (DT_GREATERTHAN);

		switch (lrec->dtrd_size) {
		case sizeof (uint64_t):
			/* LINTED - alignment */
			lval = *((uint64_t *)ldata);
			/* LINTED - alignment */
			rval = *((uint64_t *)rdata);
			break;

		case sizeof (uint32_t):
			/* LINTED - alignment */
			lval = *((uint32_t *)ldata);
			/* LINTED - alignment */
			rval = *((uint32_t *)rdata);
			break;

		case sizeof (uint16_t):
			/* LINTED - alignment */
			lval = *((uint16_t *)ldata);
			/* LINTED - alignment */
			rval = *((uint16_t *)rdata);
			break;

		case sizeof (uint8_t):
			lval = *((uint8_t *)ldata);
			rval = *((uint8_t *)rdata);
			break;

		default:
			switch (lrec->dtrd_action) {
			case DTRACEACT_UMOD:
			case DTRACEACT_UADDR:
			case DTRACEACT_USYM:
				for (j = 0; j < 2; j++) {
					/* LINTED - alignment */
					lval = ((uint64_t *)ldata)[j];
					/* LINTED - alignment */
					rval = ((uint64_t *)rdata)[j];

					if (lval < rval)
						return (DT_LESSTHAN);

					if (lval > rval)
						return (DT_GREATERTHAN);
				}

				break;

			default:
				for (j = 0; j < lrec->dtrd_size; j++) {
					lval = ((uint8_t *)ldata)[j];
					rval = ((uint8_t *)rdata)[j];

					if (lval < rval)
						return (DT_LESSTHAN);

					if (lval > rval)
						return (DT_GREATERTHAN);
				}
			}

			continue;
		}

		if (lval < rval)
			return (DT_LESSTHAN);

		if (lval > rval)
			return (DT_GREATERTHAN);
	}

	return (0);
}

static int
dt_aggregate_valcmp(const void *lhs, const void *rhs)
{
	dt_ahashent_t *lh = *((dt_ahashent_t **)lhs);
	dt_ahashent_t *rh = *((dt_ahashent_t **)rhs);
	dtrace_aggdesc_t *lagg = lh->dtahe_data.dtada_desc;
	dtrace_aggdesc_t *ragg = rh->dtahe_data.dtada_desc;
	caddr_t ldata = lh->dtahe_data.dtada_data;
	caddr_t rdata = rh->dtahe_data.dtada_data;
	dtrace_recdesc_t *lrec, *rrec;
	int64_t *laddr, *raddr;
	int rval, i;

	if ((rval = dt_aggregate_hashcmp(lhs, rhs)) != 0)
		return (rval);

	if (lagg->dtagd_nrecs > ragg->dtagd_nrecs)
		return (DT_GREATERTHAN);

	if (lagg->dtagd_nrecs < ragg->dtagd_nrecs)
		return (DT_LESSTHAN);

	for (i = 0; i < lagg->dtagd_nrecs; i++) {
		lrec = &lagg->dtagd_rec[i];
		rrec = &ragg->dtagd_rec[i];

		if (lrec->dtrd_offset < rrec->dtrd_offset)
			return (DT_LESSTHAN);

		if (lrec->dtrd_offset > rrec->dtrd_offset)
			return (DT_GREATERTHAN);

		if (lrec->dtrd_action < rrec->dtrd_action)
			return (DT_LESSTHAN);

		if (lrec->dtrd_action > rrec->dtrd_action)
			return (DT_GREATERTHAN);
	}

	laddr = (int64_t *)(uintptr_t)(ldata + lrec->dtrd_offset);
	raddr = (int64_t *)(uintptr_t)(rdata + rrec->dtrd_offset);

	switch (lrec->dtrd_action) {
	case DTRACEAGG_AVG:
		rval = dt_aggregate_averagecmp(laddr, raddr);
		break;

	case DTRACEAGG_STDDEV:
		rval = dt_aggregate_stddevcmp(laddr, raddr);
		break;

	case DTRACEAGG_QUANTIZE:
		rval = dt_aggregate_quantizedcmp(laddr, raddr);
		break;

	case DTRACEAGG_LQUANTIZE:
		rval = dt_aggregate_lquantizedcmp(laddr, raddr);
		break;

	case DTRACEAGG_COUNT:
	case DTRACEAGG_SUM:
	case DTRACEAGG_MIN:
	case DTRACEAGG_MAX:
		rval = dt_aggregate_countcmp(laddr, raddr);
		break;

	default:
		assert(0);
	}

	return (rval);
}

static int
dt_aggregate_valkeycmp(const void *lhs, const void *rhs)
{
	int rval;

	if ((rval = dt_aggregate_valcmp(lhs, rhs)) != 0)
		return (rval);

	/*
	 * If we're here, the values for the two aggregation elements are
	 * equal.  We already know that the key layout is the same for the two
	 * elements; we must now compare the keys themselves as a tie-breaker.
	 */
	return (dt_aggregate_keycmp(lhs, rhs));
}

static int
dt_aggregate_keyvarcmp(const void *lhs, const void *rhs)
{
	int rval;

	if ((rval = dt_aggregate_keycmp(lhs, rhs)) != 0)
		return (rval);

	return (dt_aggregate_varcmp(lhs, rhs));
}

static int
dt_aggregate_varkeycmp(const void *lhs, const void *rhs)
{
	int rval;

	if ((rval = dt_aggregate_varcmp(lhs, rhs)) != 0)
		return (rval);

	return (dt_aggregate_keycmp(lhs, rhs));
}

static int
dt_aggregate_valvarcmp(const void *lhs, const void *rhs)
{
	int rval;

	if ((rval = dt_aggregate_valkeycmp(lhs, rhs)) != 0)
		return (rval);

	return (dt_aggregate_varcmp(lhs, rhs));
}

static int
dt_aggregate_varvalcmp(const void *lhs, const void *rhs)
{
	int rval;

	if ((rval = dt_aggregate_varcmp(lhs, rhs)) != 0)
		return (rval);

	return (dt_aggregate_valkeycmp(lhs, rhs));
}

static int
dt_aggregate_keyvarrevcmp(const void *lhs, const void *rhs)
{
	return (dt_aggregate_keyvarcmp(rhs, lhs));
}

static int
dt_aggregate_varkeyrevcmp(const void *lhs, const void *rhs)
{
	return (dt_aggregate_varkeycmp(rhs, lhs));
}

static int
dt_aggregate_valvarrevcmp(const void *lhs, const void *rhs)
{
	return (dt_aggregate_valvarcmp(rhs, lhs));
}

static int
dt_aggregate_varvalrevcmp(const void *lhs, const void *rhs)
{
	return (dt_aggregate_varvalcmp(rhs, lhs));
}

static int
dt_aggregate_bundlecmp(const void *lhs, const void *rhs)
{
	dt_ahashent_t **lh = *((dt_ahashent_t ***)lhs);
	dt_ahashent_t **rh = *((dt_ahashent_t ***)rhs);
	int i, rval;

	if (dt_keysort) {
		/*
		 * If we're sorting on keys, we need to scan until we find the
		 * last entry -- that's the representative key.  (The order of
		 * the bundle is values followed by key to accommodate the
		 * default behavior of sorting by value.)  If the keys are
		 * equal, we'll fall into the value comparison loop, below.
		 */
		for (i = 0; lh[i + 1] != NULL; i++)
			continue;

		assert(i != 0);
		assert(rh[i + 1] == NULL);

		if ((rval = dt_aggregate_keycmp(&lh[i], &rh[i])) != 0)
			return (rval);
	}

	for (i = 0; ; i++) {
		if (lh[i + 1] == NULL) {
			/*
			 * All of the values are equal; if we're sorting on
			 * keys, then we're only here because the keys were
			 * found to be equal and these records are therefore
			 * equal.  If we're not sorting on keys, we'll use the
			 * key comparison from the representative key as the
			 * tie-breaker.
			 */
			if (dt_keysort)
				return (0);

			assert(i != 0);
			assert(rh[i + 1] == NULL);
			return (dt_aggregate_keycmp(&lh[i], &rh[i]));
		} else {
			if ((rval = dt_aggregate_valcmp(&lh[i], &rh[i])) != 0)
				return (rval);
		}
	}
}

int
dt_aggregate_go(dtrace_hdl_t *dtp)
{
	dt_aggregate_t *agp = &dtp->dt_aggregate;
	dtrace_optval_t size, cpu;
	dtrace_bufdesc_t *buf = &agp->dtat_buf;
	int rval, i;

	assert(agp->dtat_maxcpu == 0);
	assert(agp->dtat_ncpu == 0);
	assert(agp->dtat_cpus == NULL);

	agp->dtat_maxcpu = dt_sysconf(dtp, _SC_CPUID_MAX) + 1;
	agp->dtat_ncpu = dt_sysconf(dtp, _SC_NPROCESSORS_MAX);
	agp->dtat_cpus = malloc(agp->dtat_ncpu * sizeof (processorid_t));

	if (agp->dtat_cpus == NULL)
		return (dt_set_errno(dtp, EDT_NOMEM));

	/*
	 * Use the aggregation buffer size as reloaded from the kernel.
	 */
	size = dtp->dt_options[DTRACEOPT_AGGSIZE];

	rval = dtrace_getopt(dtp, "aggsize", &size);
	assert(rval == 0);

	if (size == 0 || size == DTRACEOPT_UNSET)
		return (0);

	buf = &agp->dtat_buf;
	buf->dtbd_size = size;

	if ((buf->dtbd_data = malloc(buf->dtbd_size)) == NULL)
		return (dt_set_errno(dtp, EDT_NOMEM));

	/*
	 * Now query for the CPUs enabled.
	 */
	rval = dtrace_getopt(dtp, "cpu", &cpu);
	assert(rval == 0 && cpu != DTRACEOPT_UNSET);

	if (cpu != DTRACE_CPUALL) {
		assert(cpu < agp->dtat_ncpu);
		agp->dtat_cpus[agp->dtat_ncpus++] = (processorid_t)cpu;

		return (0);
	}

	agp->dtat_ncpus = 0;
	for (i = 0; i < agp->dtat_maxcpu; i++) {
		if (dt_status(dtp, i) == -1)
			continue;

		agp->dtat_cpus[agp->dtat_ncpus++] = i;
	}

	return (0);
}

static int
dt_aggwalk_rval(dtrace_hdl_t *dtp, dt_ahashent_t *h, int rval)
{
	dt_aggregate_t *agp = &dtp->dt_aggregate;
	dtrace_aggdata_t *data;
	dtrace_aggdesc_t *aggdesc;
	dtrace_recdesc_t *rec;
	int i;

	switch (rval) {
	case DTRACE_AGGWALK_NEXT:
		break;

	case DTRACE_AGGWALK_CLEAR: {
		uint32_t size, offs = 0;

		aggdesc = h->dtahe_data.dtada_desc;
		rec = &aggdesc->dtagd_rec[aggdesc->dtagd_nrecs - 1];
		size = rec->dtrd_size;
		data = &h->dtahe_data;

		if (rec->dtrd_action == DTRACEAGG_LQUANTIZE) {
			offs = sizeof (uint64_t);
			size -= sizeof (uint64_t);
		}

		bzero(&data->dtada_data[rec->dtrd_offset] + offs, size);

		if (data->dtada_percpu == NULL)
			break;

		for (i = 0; i < dtp->dt_aggregate.dtat_maxcpu; i++)
			bzero(data->dtada_percpu[i] + offs, size);
		break;
	}

	case DTRACE_AGGWALK_ERROR:
		/*
		 * We assume that errno is already set in this case.
		 */
		return (dt_set_errno(dtp, errno));

	case DTRACE_AGGWALK_ABORT:
		return (dt_set_errno(dtp, EDT_DIRABORT));

	case DTRACE_AGGWALK_DENORMALIZE:
		h->dtahe_data.dtada_normal = 1;
		return (0);

	case DTRACE_AGGWALK_NORMALIZE:
		if (h->dtahe_data.dtada_normal == 0) {
			h->dtahe_data.dtada_normal = 1;
			return (dt_set_errno(dtp, EDT_BADRVAL));
		}

		return (0);

	case DTRACE_AGGWALK_REMOVE: {
		dtrace_aggdata_t *aggdata = &h->dtahe_data;
		int i, max_cpus = agp->dtat_maxcpu;

		/*
		 * First, remove this hash entry from its hash chain.
		 */
		if (h->dtahe_prev != NULL) {
			h->dtahe_prev->dtahe_next = h->dtahe_next;
		} else {
			dt_ahash_t *hash = &agp->dtat_hash;
			size_t ndx = h->dtahe_hashval % hash->dtah_size;

			assert(hash->dtah_hash[ndx] == h);
			hash->dtah_hash[ndx] = h->dtahe_next;
		}

		if (h->dtahe_next != NULL)
			h->dtahe_next->dtahe_prev = h->dtahe_prev;

		/*
		 * Now remove it from the list of all hash entries.
		 */
		if (h->dtahe_prevall != NULL) {
			h->dtahe_prevall->dtahe_nextall = h->dtahe_nextall;
		} else {
			dt_ahash_t *hash = &agp->dtat_hash;

			assert(hash->dtah_all == h);
			hash->dtah_all = h->dtahe_nextall;
		}

		if (h->dtahe_nextall != NULL)
			h->dtahe_nextall->dtahe_prevall = h->dtahe_prevall;

		/*
		 * We're unlinked.  We can safely destroy the data.
		 */
		if (aggdata->dtada_percpu != NULL) {
			for (i = 0; i < max_cpus; i++)
				free(aggdata->dtada_percpu[i]);
			free(aggdata->dtada_percpu);
		}

		free(aggdata->dtada_data);
		free(h);

		return (0);
	}

	default:
		return (dt_set_errno(dtp, EDT_BADRVAL));
	}

	return (0);
}

void
dt_aggregate_qsort(dtrace_hdl_t *dtp, void *base, size_t nel, size_t width,
    int (*compar)(const void *, const void *))
{
	int rev = dt_revsort, key = dt_keysort, keypos = dt_keypos;
	dtrace_optval_t keyposopt = dtp->dt_options[DTRACEOPT_AGGSORTKEYPOS];

	dt_revsort = (dtp->dt_options[DTRACEOPT_AGGSORTREV] != DTRACEOPT_UNSET);
	dt_keysort = (dtp->dt_options[DTRACEOPT_AGGSORTKEY] != DTRACEOPT_UNSET);

	if (keyposopt != DTRACEOPT_UNSET && keyposopt <= INT_MAX) {
		dt_keypos = (int)keyposopt;
	} else {
		dt_keypos = 0;
	}

	if (compar == NULL) {
		if (!dt_keysort) {
			compar = dt_aggregate_varvalcmp;
		} else {
			compar = dt_aggregate_varkeycmp;
		}
	}

	qsort(base, nel, width, compar);

	dt_revsort = rev;
	dt_keysort = key;
	dt_keypos = keypos;
}

int
dtrace_aggregate_walk(dtrace_hdl_t *dtp, dtrace_aggregate_f *func, void *arg)
{
	dt_ahashent_t *h, *next;
	dt_ahash_t *hash = &dtp->dt_aggregate.dtat_hash;

	for (h = hash->dtah_all; h != NULL; h = next) {
		/*
		 * dt_aggwalk_rval() can potentially remove the current hash
		 * entry; we need to load the next hash entry before calling
		 * into it.
		 */
		next = h->dtahe_nextall;

		if (dt_aggwalk_rval(dtp, h, func(&h->dtahe_data, arg)) == -1)
			return (-1);
	}

	return (0);
}

static int
dt_aggregate_walk_sorted(dtrace_hdl_t *dtp,
    dtrace_aggregate_f *func, void *arg,
    int (*sfunc)(const void *, const void *))
{
	dt_aggregate_t *agp = &dtp->dt_aggregate;
	dt_ahashent_t *h, **sorted;
	dt_ahash_t *hash = &agp->dtat_hash;
	size_t i, nentries = 0;

	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall)
		nentries++;

	sorted = dt_alloc(dtp, nentries * sizeof (dt_ahashent_t *));

	if (sorted == NULL)
		return (-1);

	for (h = hash->dtah_all, i = 0; h != NULL; h = h->dtahe_nextall)
		sorted[i++] = h;

	(void) pthread_mutex_lock(&dt_qsort_lock);

	if (sfunc == NULL) {
		dt_aggregate_qsort(dtp, sorted, nentries,
		    sizeof (dt_ahashent_t *), NULL);
	} else {
		/*
		 * If we've been explicitly passed a sorting function,
		 * we'll use that -- ignoring the values of the "aggsortrev",
		 * "aggsortkey" and "aggsortkeypos" options.
		 */
		qsort(sorted, nentries, sizeof (dt_ahashent_t *), sfunc);
	}

	(void) pthread_mutex_unlock(&dt_qsort_lock);

	for (i = 0; i < nentries; i++) {
		h = sorted[i];

		if (dt_aggwalk_rval(dtp, h, func(&h->dtahe_data, arg)) == -1) {
			dt_free(dtp, sorted);
			return (-1);
		}
	}

	dt_free(dtp, sorted);
	return (0);
}

int
dtrace_aggregate_walk_sorted(dtrace_hdl_t *dtp,
    dtrace_aggregate_f *func, void *arg)
{
	return (dt_aggregate_walk_sorted(dtp, func, arg, NULL));
}

int
dtrace_aggregate_walk_keysorted(dtrace_hdl_t *dtp,
    dtrace_aggregate_f *func, void *arg)
{
	return (dt_aggregate_walk_sorted(dtp, func,
	    arg, dt_aggregate_varkeycmp));
}

int
dtrace_aggregate_walk_valsorted(dtrace_hdl_t *dtp,
    dtrace_aggregate_f *func, void *arg)
{
	return (dt_aggregate_walk_sorted(dtp, func,
	    arg, dt_aggregate_varvalcmp));
}

int
dtrace_aggregate_walk_keyvarsorted(dtrace_hdl_t *dtp,
    dtrace_aggregate_f *func, void *arg)
{
	return (dt_aggregate_walk_sorted(dtp, func,
	    arg, dt_aggregate_keyvarcmp));
}

int
dtrace_aggregate_walk_valvarsorted(dtrace_hdl_t *dtp,
    dtrace_aggregate_f *func, void *arg)
{
	return (dt_aggregate_walk_sorted(dtp, func,
	    arg, dt_aggregate_valvarcmp));
}

int
dtrace_aggregate_walk_keyrevsorted(dtrace_hdl_t *dtp,
    dtrace_aggregate_f *func, void *arg)
{
	return (dt_aggregate_walk_sorted(dtp, func,
	    arg, dt_aggregate_varkeyrevcmp));
}

int
dtrace_aggregate_walk_valrevsorted(dtrace_hdl_t *dtp,
    dtrace_aggregate_f *func, void *arg)
{
	return (dt_aggregate_walk_sorted(dtp, func,
	    arg, dt_aggregate_varvalrevcmp));
}

int
dtrace_aggregate_walk_keyvarrevsorted(dtrace_hdl_t *dtp,
    dtrace_aggregate_f *func, void *arg)
{
	return (dt_aggregate_walk_sorted(dtp, func,
	    arg, dt_aggregate_keyvarrevcmp));
}

int
dtrace_aggregate_walk_valvarrevsorted(dtrace_hdl_t *dtp,
    dtrace_aggregate_f *func, void *arg)
{
	return (dt_aggregate_walk_sorted(dtp, func,
	    arg, dt_aggregate_valvarrevcmp));
}

int
dtrace_aggregate_walk_joined(dtrace_hdl_t *dtp, dtrace_aggvarid_t *aggvars,
    int naggvars, dtrace_aggregate_walk_joined_f *func, void *arg)
{
	dt_aggregate_t *agp = &dtp->dt_aggregate;
	dt_ahashent_t *h, **sorted = NULL, ***bundle, **nbundle;
	const dtrace_aggdata_t **data;
	dt_ahashent_t *zaggdata = NULL;
	dt_ahash_t *hash = &agp->dtat_hash;
	size_t nentries = 0, nbundles = 0, start, zsize = 0, bundlesize;
	dtrace_aggvarid_t max = 0, aggvar;
	int rval = -1, *map, *remap = NULL;
	int i, j;
	dtrace_optval_t sortpos = dtp->dt_options[DTRACEOPT_AGGSORTPOS];

	/*
	 * If the sorting position is greater than the number of aggregation
	 * variable IDs, we silently set it to 0.
	 */
	if (sortpos == DTRACEOPT_UNSET || sortpos >= naggvars)
		sortpos = 0;

	/*
	 * First we need to translate the specified aggregation variable IDs
	 * into a linear map that will allow us to translate an aggregation
	 * variable ID into its position in the specified aggvars.
	 */
	for (i = 0; i < naggvars; i++) {
		if (aggvars[i] == DTRACE_AGGVARIDNONE || aggvars[i] < 0)
			return (dt_set_errno(dtp, EDT_BADAGGVAR));

		if (aggvars[i] > max)
			max = aggvars[i];
	}

	if ((map = dt_zalloc(dtp, (max + 1) * sizeof (int))) == NULL)
		return (-1);

	zaggdata = dt_zalloc(dtp, naggvars * sizeof (dt_ahashent_t));

	if (zaggdata == NULL)
		goto out;

	for (i = 0; i < naggvars; i++) {
		int ndx = i + sortpos;

		if (ndx >= naggvars)
			ndx -= naggvars;

		aggvar = aggvars[ndx];
		assert(aggvar <= max);

		if (map[aggvar]) {
			/*
			 * We have an aggregation variable that is present
			 * more than once in the array of aggregation
			 * variables.  While it's unclear why one might want
			 * to do this, it's legal.  To support this construct,
			 * we will allocate a remap that will indicate the
			 * position from which this aggregation variable
			 * should be pulled.  (That is, where the remap will
			 * map from one position to another.)
			 */
			if (remap == NULL) {
				remap = dt_zalloc(dtp, naggvars * sizeof (int));

				if (remap == NULL)
					goto out;
			}

			/*
			 * Given that the variable is already present, assert
			 * that following through the mapping and adjusting
			 * for the sort position yields the same aggregation
			 * variable ID.
			 */
			assert(aggvars[(map[aggvar] - 1 + sortpos) %
			    naggvars] == aggvars[ndx]);

			remap[i] = map[aggvar];
			continue;
		}

		map[aggvar] = i + 1;
	}

	/*
	 * We need to take two passes over the data to size our allocation, so
	 * we'll use the first pass to also fill in the zero-filled data to be
	 * used to properly format a zero-valued aggregation.
	 */
	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
		dtrace_aggvarid_t id;
		int ndx;

		if ((id = dt_aggregate_aggvarid(h)) > max || !(ndx = map[id]))
			continue;

		if (zaggdata[ndx - 1].dtahe_size == 0) {
			zaggdata[ndx - 1].dtahe_size = h->dtahe_size;
			zaggdata[ndx - 1].dtahe_data = h->dtahe_data;
		}

		nentries++;
	}

	if (nentries == 0) {
		/*
		 * We couldn't find any entries; there is nothing else to do.
		 */
		rval = 0;
		goto out;
	}

	/*
	 * Before we sort the data, we're going to look for any holes in our
	 * zero-filled data.  This will occur if an aggregation variable that
	 * we are being asked to print has not yet been assigned the result of
	 * any aggregating action for _any_ tuple.  The issue becomes that we
	 * would like a zero value to be printed for all columns for this
	 * aggregation, but without any record description, we don't know the
	 * aggregating action that corresponds to the aggregation variable.  To
	 * try to find a match, we're simply going to lookup aggregation IDs
	 * (which are guaranteed to be contiguous and to start from 1), looking
	 * for the specified aggregation variable ID.  If we find a match,
	 * we'll use that.  If we iterate over all aggregation IDs and don't
	 * find a match, then we must be an anonymous enabling.  (Anonymous
	 * enablings can't currently derive either aggregation variable IDs or
	 * aggregation variable names given only an aggregation ID.)  In this
	 * obscure case (anonymous enabling, multiple aggregation printa() with
	 * some aggregations not represented for any tuple), our defined
	 * behavior is that the zero will be printed in the format of the first
	 * aggregation variable that contains any non-zero value.
	 */
	for (i = 0; i < naggvars; i++) {
		if (zaggdata[i].dtahe_size == 0) {
			dtrace_aggvarid_t aggvar;

			aggvar = aggvars[(i - sortpos + naggvars) % naggvars];
			assert(zaggdata[i].dtahe_data.dtada_data == NULL);

			for (j = DTRACE_AGGIDNONE + 1; ; j++) {
				dtrace_aggdesc_t *agg;
				dtrace_aggdata_t *aggdata;

				if (dt_aggid_lookup(dtp, j, &agg) != 0)
					break;

				if (agg->dtagd_varid != aggvar)
					continue;

				/*
				 * We have our description -- now we need to
				 * cons up the zaggdata entry for it.
				 */
				aggdata = &zaggdata[i].dtahe_data;
				aggdata->dtada_size = agg->dtagd_size;
				aggdata->dtada_desc = agg;
				aggdata->dtada_handle = dtp;
				(void) dt_epid_lookup(dtp, agg->dtagd_epid,
				    &aggdata->dtada_edesc,
				    &aggdata->dtada_pdesc);
				aggdata->dtada_normal = 1;
				zaggdata[i].dtahe_hashval = 0;
				zaggdata[i].dtahe_size = agg->dtagd_size;
				break;
			}

			if (zaggdata[i].dtahe_size == 0) {
				caddr_t data;

				/*
				 * We couldn't find this aggregation, meaning
				 * that we have never seen it before for any
				 * tuple _and_ this is an anonymous enabling.
				 * That is, we're in the obscure case outlined
				 * above.  In this case, our defined behavior
				 * is to format the data in the format of the
				 * first non-zero aggregation -- of which, of
				 * course, we know there to be at least one
				 * (or nentries would have been zero).
				 */
				for (j = 0; j < naggvars; j++) {
					if (zaggdata[j].dtahe_size != 0)
						break;
				}

				assert(j < naggvars);
				zaggdata[i] = zaggdata[j];

				data = zaggdata[i].dtahe_data.dtada_data;
				assert(data != NULL);
			}
		}
	}

	/*
	 * Now we need to allocate our zero-filled data for use for
	 * aggregations that don't have a value corresponding to a given key.
	 */
	for (i = 0; i < naggvars; i++) {
		dtrace_aggdata_t *aggdata = &zaggdata[i].dtahe_data;
		dtrace_aggdesc_t *aggdesc = aggdata->dtada_desc;
		dtrace_recdesc_t *rec;
		uint64_t larg;
		caddr_t zdata;

		zsize = zaggdata[i].dtahe_size;
		assert(zsize != 0);

		if ((zdata = dt_zalloc(dtp, zsize)) == NULL) {
			/*
			 * If we failed to allocated some zero-filled data, we
			 * need to zero out the remaining dtada_data pointers
			 * to prevent the wrong data from being freed below.
			 */
			for (j = i; j < naggvars; j++)
				zaggdata[j].dtahe_data.dtada_data = NULL;
			goto out;
		}

		aggvar = aggvars[(i - sortpos + naggvars) % naggvars];

		/*
		 * First, the easy bit.  To maintain compatibility with
		 * consumers that pull the compiler-generated ID out of the
		 * data, we put that ID at the top of the zero-filled data.
		 */
		rec = &aggdesc->dtagd_rec[0];
		/* LINTED - alignment */
		*((dtrace_aggvarid_t *)(zdata + rec->dtrd_offset)) = aggvar;

		rec = &aggdesc->dtagd_rec[aggdesc->dtagd_nrecs - 1];

		/*
		 * Now for the more complicated part.  If (and only if) this
		 * is an lquantize() aggregating action, zero-filled data is
		 * not equivalent to an empty record:  we must also get the
		 * parameters for the lquantize().
		 */
		if (rec->dtrd_action == DTRACEAGG_LQUANTIZE) {
			if (aggdata->dtada_data != NULL) {
				/*
				 * The easier case here is if we actually have
				 * some prototype data -- in which case we
				 * manually dig it out of the aggregation
				 * record.
				 */
				/* LINTED - alignment */
				larg = *((uint64_t *)(aggdata->dtada_data +
				    rec->dtrd_offset));
			} else {
				/*
				 * We don't have any prototype data.  As a
				 * result, we know that we _do_ have the
				 * compiler-generated information.  (If this
				 * were an anonymous enabling, all of our
				 * zero-filled data would have prototype data
				 * -- either directly or indirectly.) So as
				 * gross as it is, we'll grovel around in the
				 * compiler-generated information to find the
				 * lquantize() parameters.
				 */
				dtrace_stmtdesc_t *sdp;
				dt_ident_t *aid;
				dt_idsig_t *isp;

				sdp = (dtrace_stmtdesc_t *)(uintptr_t)
				    aggdesc->dtagd_rec[0].dtrd_uarg;
				aid = sdp->dtsd_aggdata;
				isp = (dt_idsig_t *)aid->di_data;
				assert(isp->dis_auxinfo != 0);
				larg = isp->dis_auxinfo;
			}

			/* LINTED - alignment */
			*((uint64_t *)(zdata + rec->dtrd_offset)) = larg;
		}

		aggdata->dtada_data = zdata;
	}

	/*
	 * Now that we've dealt with setting up our zero-filled data, we can
	 * allocate our sorted array, and take another pass over the data to
	 * fill it.
	 */
	sorted = dt_alloc(dtp, nentries * sizeof (dt_ahashent_t *));

	if (sorted == NULL)
		goto out;

	for (h = hash->dtah_all, i = 0; h != NULL; h = h->dtahe_nextall) {
		dtrace_aggvarid_t id;

		if ((id = dt_aggregate_aggvarid(h)) > max || !map[id])
			continue;

		sorted[i++] = h;
	}

	assert(i == nentries);

	/*
	 * We've loaded our array; now we need to sort by value to allow us
	 * to create bundles of like value.  We're going to acquire the
	 * dt_qsort_lock here, and hold it across all of our subsequent
	 * comparison and sorting.
	 */
	(void) pthread_mutex_lock(&dt_qsort_lock);

	qsort(sorted, nentries, sizeof (dt_ahashent_t *),
	    dt_aggregate_keyvarcmp);

	/*
	 * Now we need to go through and create bundles.  Because the number
	 * of bundles is bounded by the size of the sorted array, we're going
	 * to reuse the underlying storage.  And note that "bundle" is an
	 * array of pointers to arrays of pointers to dt_ahashent_t -- making
	 * its type (regrettably) "dt_ahashent_t ***".  (Regrettable because
	 * '*' -- like '_' and 'X' -- should never appear in triplicate in
	 * an ideal world.)
	 */
	bundle = (dt_ahashent_t ***)sorted;

	for (i = 1, start = 0; i <= nentries; i++) {
		if (i < nentries &&
		    dt_aggregate_keycmp(&sorted[i], &sorted[i - 1]) == 0)
			continue;

		/*
		 * We have a bundle boundary.  Everything from start to
		 * (i - 1) belongs in one bundle.
		 */
		assert(i - start <= naggvars);
		bundlesize = (naggvars + 2) * sizeof (dt_ahashent_t *);

		if ((nbundle = dt_zalloc(dtp, bundlesize)) == NULL) {
			(void) pthread_mutex_unlock(&dt_qsort_lock);
			goto out;
		}

		for (j = start; j < i; j++) {
			dtrace_aggvarid_t id = dt_aggregate_aggvarid(sorted[j]);

			assert(id <= max);
			assert(map[id] != 0);
			assert(map[id] - 1 < naggvars);
			assert(nbundle[map[id] - 1] == NULL);
			nbundle[map[id] - 1] = sorted[j];

			if (nbundle[naggvars] == NULL)
				nbundle[naggvars] = sorted[j];
		}

		for (j = 0; j < naggvars; j++) {
			if (nbundle[j] != NULL)
				continue;

			/*
			 * Before we assume that this aggregation variable
			 * isn't present (and fall back to using the
			 * zero-filled data allocated earlier), check the
			 * remap.  If we have a remapping, we'll drop it in
			 * here.  Note that we might be remapping an
			 * aggregation variable that isn't present for this
			 * key; in this case, the aggregation data that we
			 * copy will point to the zeroed data.
			 */
			if (remap != NULL && remap[j]) {
				assert(remap[j] - 1 < j);
				assert(nbundle[remap[j] - 1] != NULL);
				nbundle[j] = nbundle[remap[j] - 1];
			} else {
				nbundle[j] = &zaggdata[j];
			}
		}

		bundle[nbundles++] = nbundle;
		start = i;
	}

	/*
	 * Now we need to re-sort based on the first value.
	 */
	dt_aggregate_qsort(dtp, bundle, nbundles, sizeof (dt_ahashent_t **),
	    dt_aggregate_bundlecmp);

	(void) pthread_mutex_unlock(&dt_qsort_lock);

	/*
	 * We're done!  Now we just need to go back over the sorted bundles,
	 * calling the function.
	 */
	data = alloca((naggvars + 1) * sizeof (dtrace_aggdata_t *));

	for (i = 0; i < nbundles; i++) {
		for (j = 0; j < naggvars; j++)
			data[j + 1] = NULL;

		for (j = 0; j < naggvars; j++) {
			int ndx = j - sortpos;

			if (ndx < 0)
				ndx += naggvars;

			assert(bundle[i][ndx] != NULL);
			data[j + 1] = &bundle[i][ndx]->dtahe_data;
		}

		for (j = 0; j < naggvars; j++)
			assert(data[j + 1] != NULL);

		/*
		 * The representative key is the last element in the bundle.
		 * Assert that we have one, and then set it to be the first
		 * element of data.
		 */
		assert(bundle[i][j] != NULL);
		data[0] = &bundle[i][j]->dtahe_data;

		if ((rval = func(data, naggvars + 1, arg)) == -1)
			goto out;
	}

	rval = 0;
out:
	for (i = 0; i < nbundles; i++)
		dt_free(dtp, bundle[i]);

	if (zaggdata != NULL) {
		for (i = 0; i < naggvars; i++)
			dt_free(dtp, zaggdata[i].dtahe_data.dtada_data);
	}

	dt_free(dtp, zaggdata);
	dt_free(dtp, sorted);
	dt_free(dtp, remap);
	dt_free(dtp, map);

	return (rval);
}

int
dtrace_aggregate_print(dtrace_hdl_t *dtp, FILE *fp,
    dtrace_aggregate_walk_f *func)
{
	dt_print_aggdata_t pd;

	pd.dtpa_dtp = dtp;
	pd.dtpa_fp = fp;
	pd.dtpa_allunprint = 1;

	if (func == NULL)
		func = dtrace_aggregate_walk_sorted;

	if ((*func)(dtp, dt_print_agg, &pd) == -1)
		return (dt_set_errno(dtp, dtp->dt_errno));

	return (0);
}

void
dtrace_aggregate_clear(dtrace_hdl_t *dtp)
{
	dt_aggregate_t *agp = &dtp->dt_aggregate;
	dt_ahash_t *hash = &agp->dtat_hash;
	dt_ahashent_t *h;
	dtrace_aggdata_t *data;
	dtrace_aggdesc_t *aggdesc;
	dtrace_recdesc_t *rec;
	int i, max_cpus = agp->dtat_maxcpu;

	for (h = hash->dtah_all; h != NULL; h = h->dtahe_nextall) {
		aggdesc = h->dtahe_data.dtada_desc;
		rec = &aggdesc->dtagd_rec[aggdesc->dtagd_nrecs - 1];
		data = &h->dtahe_data;

		bzero(&data->dtada_data[rec->dtrd_offset], rec->dtrd_size);

		if (data->dtada_percpu == NULL)
			continue;

		for (i = 0; i < max_cpus; i++)
			bzero(data->dtada_percpu[i], rec->dtrd_size);
	}
}

void
dt_aggregate_destroy(dtrace_hdl_t *dtp)
{
	dt_aggregate_t *agp = &dtp->dt_aggregate;
	dt_ahash_t *hash = &agp->dtat_hash;
	dt_ahashent_t *h, *next;
	dtrace_aggdata_t *aggdata;
	int i, max_cpus = agp->dtat_maxcpu;

	if (hash->dtah_hash == NULL) {
		assert(hash->dtah_all == NULL);
	} else {
		free(hash->dtah_hash);

		for (h = hash->dtah_all; h != NULL; h = next) {
			next = h->dtahe_nextall;

			aggdata = &h->dtahe_data;

			if (aggdata->dtada_percpu != NULL) {
				for (i = 0; i < max_cpus; i++)
					free(aggdata->dtada_percpu[i]);
				free(aggdata->dtada_percpu);
			}

			free(aggdata->dtada_data);
			free(h);
		}

		hash->dtah_hash = NULL;
		hash->dtah_all = NULL;
		hash->dtah_size = 0;
	}

	free(agp->dtat_buf.dtbd_data);
	free(agp->dtat_cpus);
}