common - zygaena/hammerhead

root / base / usr / src / cmd / cpc / common
common Plain Text 3041 lines 72.9 KB
/*
 * 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.
 */

#define	 __EXTENSIONS__	/* header bug! strtok_r is overly hidden */
#include <string.h>
#include <strings.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <libintl.h>

#include <libcpc.h>

#include "cpucmds.h"

struct args {
	FILE *fp;
	int colnum;
	int margin;
};

struct evlist {
	char *list;
	int size;
};

#define	MAX_RHS_COLUMN	76
#define	EVENT_MARGIN	17
#define	ATTR_MARGIN	20

/*ARGSUSED*/
static void
list_cap(void *arg, uint_t regno, const char *name)
{
	struct args *args = arg;
	int i;

	if ((args->colnum + strlen(name) + 1) > MAX_RHS_COLUMN) {
		(void) fprintf(args->fp, "\n");
		for (i = 0; i < args->margin; i++)
			(void) fprintf(args->fp, " ");
		args->colnum = args->margin;
	}
	args->colnum += fprintf(args->fp, "%s ", name);
}

static void
list_attr(void *arg, const char *name)
{
	/*
	 * The following attributes are used by the commands but should not be
	 * reported to the user, since they may not be specified directly.
	 */
	if (strncmp(name, "picnum", 7) == 0 ||
	    strncmp(name, "count_sibling_usr", 18) == 0 ||
	    strncmp(name, "count_sibling_sys", 18) == 0)
		return;

	list_cap(arg, 0, name);
}

static void *
emalloc(size_t size)
{
	void *ptr;

	if ((ptr = malloc(size)) == NULL) {
		(void) fprintf(stderr, gettext("no memory available\n"));
		exit(1);
	}

	return (ptr);
}

/*
 * Used by allpics_equal().
 */
/*ARGSUSED*/
static void
cap_walker(void *arg, uint_t regno, const char *name)
{
	struct evlist *list = arg;

	list->size += strlen(name);
	if ((list->list = realloc(list->list, list->size + 1)) == NULL) {
		(void) fprintf(stderr, gettext("no memory available\n"));
		exit(1);
	}

	(void) strcat(list->list, name);
}

/*
 * Returns 1 if all counters on this chip can count all possible events.
 */
static int
allpics_equal(cpc_t *cpc)
{
	int	npics = cpc_npic(cpc);
	int	i;
	struct	evlist **lists;
	int	ret = 1;

	lists = emalloc(npics * sizeof (struct evlist *));

	for (i = 0; i < npics; i++) {
		lists[i] = emalloc(sizeof (struct evlist));
		lists[i]->size = 0;
		lists[i]->list = emalloc(1);
		lists[i]->list[0] = '\0';
		cpc_walk_events_pic(cpc, i, lists[i], cap_walker);
	}

	for (i = 1; i < npics; i++)
		if (lists[i]->size != lists[0]->size ||
		    strncmp(lists[i]->list, lists[0]->list,
		    lists[0]->size) != 0) {
			ret = 0;
			break;
		}

	for (i = 0; i < npics; i++) {
		free(lists[i]->list);
		free(lists[i]);
	}
	free(lists);

	return (ret);
}

int
capabilities(cpc_t *cpc, FILE *fp)
{
	struct args _args, *args = &_args;
	char *text, *tok, *cp;
	const char *ccp;
	int npic = cpc_npic(cpc);
	int i, pics_equal = allpics_equal(cpc);

	args->fp = fp;

	if ((ccp = cpc_cciname(cpc)) == NULL)
		ccp = "No information available";
	(void) fprintf(args->fp, "\t%s: %s\n\n",
	    gettext("CPU performance counter interface"), ccp);

	(void) fprintf(args->fp, gettext("\tevent specification syntax:\n"));

	(void) fprintf(args->fp, "\t[picn=]<eventn>[,attr[n][=<val>]]"
	    "[,[picn=]<eventn>[,attr[n][=<val>]],...]\n");

	(void) fprintf(args->fp, gettext("\n\tGeneric Events:\n"));

	if (pics_equal) {
		args->margin = args->colnum = EVENT_MARGIN;
		(void) fprintf(args->fp, "\n\tevent[0-%d]: ", npic - 1);
		cpc_walk_generic_events_pic(cpc, 0, args, list_cap);
		(void) fprintf(args->fp, "\n");
	} else {
		args->margin = EVENT_MARGIN;
		for (i = 0; i < npic; i++) {
			(void) fprintf(args->fp, "\n\tevent%d: ", i);
			if (i < 10) (void) fprintf(args->fp, " ");
			args->colnum = EVENT_MARGIN;
			cpc_walk_generic_events_pic(cpc, i, args, list_cap);
			(void) fprintf(args->fp, "\n");
		}
	}

	(void) fprintf(args->fp, gettext("\n\tSee generic_events(3CPC) for"
	    " descriptions of these events\n\n"));

	(void) fprintf(args->fp, gettext("\tPlatform Specific Events:\n"));

	if (pics_equal) {
		args->margin = args->colnum = EVENT_MARGIN;
		(void) fprintf(args->fp, "\n\tevent[0-%d]: ", npic - 1);
		cpc_walk_events_pic(cpc, 0, args, list_cap);
		(void) fprintf(args->fp, "\n");
	} else {
		args->margin = EVENT_MARGIN;
		for (i = 0; i < npic; i++) {
			(void) fprintf(args->fp, "\n\tevent%d: ", i);
			if (i < 10) (void) fprintf(args->fp, " ");
			args->colnum = EVENT_MARGIN;
			cpc_walk_events_pic(cpc, i, args, list_cap);
			(void) fprintf(args->fp, "\n");
		}
	}

	(void) fprintf(args->fp, "\n\tattributes: ");
	args->colnum = args->margin = ATTR_MARGIN;
	cpc_walk_attrs(cpc, args, list_attr);
	/*
	 * In addition to the attributes published by the kernel, we allow the
	 * user to specify two additional tokens on all platforms. List them
	 * here.
	 */
	list_cap(args, 0, "nouser");
	list_cap(args, 0, "sys");
	(void) fprintf(args->fp, "\n\n\t");
	args->colnum = 8;

	if ((ccp = cpc_cpuref(cpc)) == NULL)
		ccp = "No information available";
	if ((text = strdup(ccp)) == NULL) {
		(void) fprintf(stderr, gettext("no memory available.\n"));
		exit(1);
	}
	for (cp = strtok_r(text, " ", &tok);
	    cp != NULL; cp = strtok_r(NULL, " ", &tok)) {
		if ((args->colnum + strlen(cp) + 1) > MAX_RHS_COLUMN) {
			(void) fprintf(args->fp, "\n\t");
			args->colnum = 8;
		}
		args->colnum += fprintf(args->fp, "%s ", cp);
	}
	(void) fprintf(args->fp, "\n");
	free(text);

	return (0);
}

/*
 * Returns 1 on SMT processors which do not have full CPC hardware for each
 * logical processor.
 */
int
smt_limited_cpc_hw(cpc_t *cpc)
{
	if (strcmp(cpc_cciname(cpc), "Pentium 4 with HyperThreading") == 0)
		return (1);
	return (0);
}
/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License, Version 1.0 only
 * (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 2004 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#ifndef	_CPUCMDS_H
#define	_CPUCMDS_H

#include <sys/types.h>
#include <stdio.h>
#include <stdarg.h>

#ifdef __cplusplus
extern "C" {
#endif

#include "libcpc_impl.h"

extern cpc_set_t *cpc_strtoset(cpc_t *cpc, const char *spec, int smt);
extern cpc_errhndlr_t *strtoset_errfn;
extern int capabilities(cpc_t *cpc, FILE *);
extern int smt_limited_cpc_hw(cpc_t *cpc);
extern void zerotime(void);
extern float mstimestamp(hrtime_t hrt);

/*
 * Request sets can be manipulated in collections called setgroups.
 */
typedef struct __cpc_setgrp cpc_setgrp_t;

extern cpc_setgrp_t *cpc_setgrp_new(cpc_t *cpc, int smt);
extern cpc_setgrp_t *cpc_setgrp_newset(cpc_setgrp_t *sgrp,
    const char *spec, int *errcnt);
extern int cpc_setgrp_getbufs(cpc_setgrp_t *sgrp, cpc_buf_t ***data1,
    cpc_buf_t ***data2, cpc_buf_t ***scratch);
extern cpc_setgrp_t *cpc_setgrp_clone(cpc_setgrp_t *sgrp);
extern void cpc_setgrp_free(cpc_setgrp_t *sgrp);

extern cpc_set_t *cpc_setgrp_getset(cpc_setgrp_t *sgrp);
extern const char *cpc_setgrp_getname(cpc_setgrp_t *sgrp);
extern const char *cpc_setgrp_gethdr(cpc_setgrp_t *sgrp);
extern int cpc_setgrp_numsets(cpc_setgrp_t *sgrp);
extern cpc_set_t *cpc_setgrp_nextset(cpc_setgrp_t *sgrp);
extern void cpc_setgrp_reset(cpc_setgrp_t *to);
extern void cpc_setgrp_accum(cpc_setgrp_t *accum, cpc_setgrp_t *sgrp);
extern int cpc_setgrp_sysonly(cpc_setgrp_t *sgrp);
extern int cpc_setgrp_has_sysonly(cpc_setgrp_t *sgrp);

#ifdef __cplusplus
}
#endif

#endif	/* _CPUCMDS_H */
/*
 * 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 2009 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

/*
 * Copyright 2019 Joyent, Inc.
 */

#include <sys/types.h>
#include <sys/processor.h>
#include <sys/pset.h>
#include <sys/lwp.h>
#include <sys/priocntl.h>
#include <sys/fxpriocntl.h>
#include <time.h>
#include <stdio.h>
#include <stdlib.h>
#include <inttypes.h>
#include <unistd.h>
#include <limits.h>
#include <string.h>
#include <strings.h>
#include <thread.h>
#include <errno.h>
#include <libintl.h>
#include <locale.h>
#include <kstat.h>
#include <synch.h>
#include <libcpc.h>
#include <sys/resource.h>

#include "cpucmds.h"
#include "statcommon.h"

static struct options {
	int debug;
	int dotitle;
	int dohelp;
	int dotick;
	int dosoaker;
	int doperiod;
	char *pgmname;
	uint_t mseconds;
	uint_t nsamples;
	uint_t nsets;
	uint_t mseconds_rest;
	cpc_setgrp_t *master;
} __options;

/*
 * States for soaker threads.
 */
#define	SOAK_PAUSE	0
#define	SOAK_RUN	1

struct tstate {
	processorid_t	cpuid;
	int		chip_id;
	cpc_setgrp_t	*sgrp;
	int		status;
	thread_t	tid;
	int		soak_state;
	mutex_t		soak_lock;
	cond_t		soak_cv;
};

static const struct options *opts = (const struct options *)&__options;

static cpc_t *cpc;

struct tstate	*gstate;
static int	ncpus;
static int	max_chip_id;
static int	*chip_designees;    /* cpuid of CPU which counts for phs chip */
static int	smt = 0;	    /* If set, cpustat needs to be SMT-aware. */
static pcinfo_t	fxinfo = { 0, "FX", 0 }; /* FX scheduler class info */

static uint_t timestamp_fmt = NODATE;

/*ARGSUSED*/
static void
cpustat_errfn(const char *fn, int subcode, const char *fmt, va_list ap)
{
	(void) fprintf(stderr, "%s: ", opts->pgmname);
	if (opts->debug)
		(void) fprintf(stderr, "%s: ", fn);
	(void) vfprintf(stderr, fmt, ap);
}

static int cpustat(void);
static int get_chipid(kstat_ctl_t *kc, processorid_t cpuid);
static void *soaker(void *arg);


#if !defined(TEXT_DOMAIN)
#define	TEXT_DOMAIN	"SYS_TEST"
#endif

int
main(int argc, char *argv[])
{
	struct options	*opts = &__options;
	int		c, errcnt = 0, ret;
	cpc_setgrp_t	*sgrp;
	char		*errstr;
	double		period;
	char		*endp;
	struct rlimit	rl;

	(void) setlocale(LC_ALL, "");
	(void) textdomain(TEXT_DOMAIN);

	if ((opts->pgmname = strrchr(argv[0], '/')) == NULL)
		opts->pgmname = argv[0];
	else
		opts->pgmname++;

	/* Make sure we can open enough files */
	rl.rlim_max = rl.rlim_cur = RLIM_INFINITY;
	if (setrlimit(RLIMIT_NOFILE, &rl) != 0) {
		errstr = strerror(errno);
		(void) fprintf(stderr,
		    gettext("%s: setrlimit failed - %s\n"),
		    opts->pgmname, errstr);
	}

	if ((cpc = cpc_open(CPC_VER_CURRENT)) == NULL) {
		errstr = strerror(errno);
		(void) fprintf(stderr, gettext("%s: cannot access performance "
		    "counters - %s\n"), opts->pgmname, errstr);
		return (1);
	}

	(void) cpc_seterrhndlr(cpc, cpustat_errfn);
	strtoset_errfn = cpustat_errfn;

	/*
	 * Check to see if cpustat needs to be SMT-aware.
	 */
	smt = smt_limited_cpc_hw(cpc);

	/*
	 * Establish some defaults
	 */
	opts->mseconds = 5000;
	opts->nsamples = UINT_MAX;
	opts->dotitle = 1;
	if ((opts->master = cpc_setgrp_new(cpc, smt)) == NULL) {
		(void) fprintf(stderr, gettext("%s: out of heap\n"),
		    opts->pgmname);
		return (1);
	}

	while ((c = getopt(argc, argv, "Dc:hntT:sp:")) != EOF && errcnt == 0)
		switch (c) {
		case 'D':			/* enable debugging */
			opts->debug++;
			break;
		case 'c':			/* specify statistics */
			if ((sgrp = cpc_setgrp_newset(opts->master,
			    optarg, &errcnt)) != NULL)
				opts->master = sgrp;
			break;
		case 'n':			/* no titles */
			opts->dotitle = 0;
			break;
		case 'p':			/* periodic behavior */
			opts->doperiod = 1;
			period = strtod(optarg, &endp);
			if (*endp != '\0') {
				(void) fprintf(stderr, gettext("%s: invalid "
				    "parameter \"%s\"\n"), opts->pgmname,
				    optarg);
				errcnt++;
			}
			break;
		case 's':			/* run soaker thread */
			opts->dosoaker = 1;
			break;
		case 't':			/* print %tick */
			opts->dotick = 1;
			break;
		case 'T':
			if (optarg) {
				if (*optarg == 'u')
					timestamp_fmt = UDATE;
				else if (*optarg == 'd')
					timestamp_fmt = DDATE;
				else
					errcnt++;
			} else {
				errcnt++;
			}
			break;
		case 'h':			/* help */
			opts->dohelp = 1;
			break;
		case '?':
		default:
			errcnt++;
			break;
		}

	switch (argc - optind) {
	case 0:
		break;
	case 2:
		opts->nsamples = strtol(argv[optind + 1], &endp, 10);
		if (*endp != '\0') {
			(void) fprintf(stderr,
			    gettext("%s: invalid argument \"%s\"\n"),
			    opts->pgmname, argv[optind + 1]);
			errcnt++;
			break;
		}
		/*FALLTHROUGH*/
	case 1:
		opts->mseconds = (uint_t)(strtod(argv[optind], &endp) * 1000.0);
		if (*endp != '\0') {
			(void) fprintf(stderr,
			    gettext("%s: invalid argument \"%s\"\n"),
			    opts->pgmname, argv[optind]);
			errcnt++;
		}
		break;
	default:
		errcnt++;
		break;
	}

	if (opts->nsamples == 0 || opts->mseconds == 0)
		errcnt++;

	if (errcnt != 0 || opts->dohelp ||
	    (opts->nsets = cpc_setgrp_numsets(opts->master)) == 0) {
		(void) fprintf(opts->dohelp ? stdout : stderr, gettext(
		    "Usage:\n\t%s -c spec [-c spec]... [-p period] [-T u|d]\n"
		    "\t\t[-sntD] [interval [count]]\n\n"
		    "\t-c spec\t  specify processor events to be monitored\n"
		    "\t-n\t  suppress titles\n"
		    "\t-p period cycle through event list periodically\n"
		    "\t-s\t  run user soaker thread for system-only events\n"
		    "\t-t\t  include %s register\n"
		    "\t-T d|u\t  Display a timestamp in date (d) or unix "
		    "time_t (u)\n"
		    "\t-D\t  enable debug mode\n"
		    "\t-h\t  print extended usage information\n\n"
		    "\tUse cputrack(1) to monitor per-process statistics.\n"),
		    opts->pgmname, CPC_TICKREG_NAME);
		if (opts->dohelp) {
			(void) putchar('\n');
			(void) capabilities(cpc, stdout);
			exit(0);
		}
		exit(2);
	}

	/*
	 * If the user requested periodic behavior, calculate the rest time
	 * between cycles.
	 */
	if (opts->doperiod) {
		opts->mseconds_rest = (uint_t)((period * 1000.0) -
		    (opts->mseconds * opts->nsets));
		if ((int)opts->mseconds_rest < 0)
			opts->mseconds_rest = 0;
		if (opts->nsamples != UINT_MAX)
			opts->nsamples *= opts->nsets;
	}

	cpc_setgrp_reset(opts->master);
	(void) setvbuf(stdout, NULL, _IOLBF, 0);

	/*
	 * By design, cpustat (regrettably) has multiple threads racing in
	 * write(2) to generate output.  As there are no guarantees made with
	 * respect to the atomicity of concurrent writes on non-O_APPEND file
	 * descriptors, we must set O_APPEND on stdout to assure that no output
	 * is lost.  If cpustat is rearchitected such that only one thread is
	 * generating output (which would also assure that the output is always
	 * in a consistent order), this code should be removed.
	 */
	if (fcntl(1, F_SETFL, fcntl(1, F_GETFL) | O_APPEND) == -1) {
		(void) fprintf(stderr, gettext("%s: cannot set output to be "
		    "append-only - %s\n"), opts->pgmname, strerror(errno));
		return (1);
	}

	/*
	 * If no system-mode only sets were created, no soaker threads will be
	 * needed.
	 */
	if (opts->dosoaker == 1 && cpc_setgrp_has_sysonly(opts->master) == 0)
		opts->dosoaker = 0;

	ret = cpustat();

	(void) cpc_close(cpc);

	return (ret);
}

static void
print_title(cpc_setgrp_t *sgrp)
{
	(void) printf("%7s %3s %5s ", "time", "cpu", "event");
	if (opts->dotick)
		(void) printf("%9s ", CPC_TICKREG_NAME);
	(void) printf("%s\n", cpc_setgrp_gethdr(sgrp));
}

static void
print_sample(processorid_t cpuid, cpc_buf_t *buf, int nreq, const char *setname,
    int sibling)
{
	char		line[1024];
	int		ccnt;
	int		i;
	uint64_t	val;
	uint64_t	tick;
	hrtime_t	hrtime;

	hrtime = cpc_buf_hrtime(cpc, buf);
	tick = cpc_buf_tick(cpc, buf);

	ccnt = snprintf(line, sizeof (line), "%7.3f %3d %5s ",
	    mstimestamp(hrtime), (int)cpuid, "tick");
	if (opts->dotick)
		ccnt += snprintf(line + ccnt, sizeof (line) - ccnt,
		    "%9" PRId64 " ", tick);
	for (i = 0; i < nreq; i++) {
		(void) cpc_buf_get(cpc, buf, i, &val);
		ccnt += snprintf(line + ccnt, sizeof (line) - ccnt,
		    "%9" PRId64 " ", val);
	}
	if (opts->nsets > 1)
		ccnt += snprintf(line + ccnt, sizeof (line) - ccnt,
		    " # %s\n", setname);
	else
		ccnt += snprintf(line + ccnt, sizeof (line) - ccnt, "\n");

	if (sibling) {
		/*
		 * This sample is being printed for a "sibling" CPU -- that is,
		 * a CPU which does not have its own CPC set bound. It is being
		 * measured via a set bound to another CPU sharing its physical
		 * processor.
		 */
		int designee = chip_designees[gstate[cpuid].chip_id];
		char *p;

		if ((p = strrchr(line, '#')) == NULL)
			p = strrchr(line, '\n');

		if (p != NULL) {
			*p = '\0';
			ccnt = strlen(line);
			ccnt += snprintf(line + ccnt, sizeof (line) - ccnt,
			    "# counter shared with CPU %d\n", designee);
		}
	}

	if (timestamp_fmt != NODATE)
		print_timestamp(timestamp_fmt);
	if (ccnt > sizeof (line))
		ccnt = sizeof (line);
	if (ccnt > 0)
		(void) write(1, line, ccnt);

	/*
	 * If this CPU is the chip designee for any other CPUs, print a line for
	 * them here.
	 */
	if (smt && (sibling == 0)) {
		for (i = 0; i < ncpus; i++) {
			if ((i != cpuid) && (gstate[i].cpuid != -1) &&
			    (chip_designees[gstate[i].chip_id] == cpuid))
				print_sample(i, buf, nreq, setname, 1);
		}
	}
}

static void
print_total(int ncpus, cpc_buf_t *buf, int nreq, const char *setname)
{
	int		i;
	uint64_t	val;

	(void) printf("%7.3f %3d %5s ", mstimestamp(cpc_buf_hrtime(cpc, buf)),
	    ncpus, "total");
	if (opts->dotick)
		(void) printf("%9" PRId64 " ", cpc_buf_tick(cpc, buf));
	for (i = 0; i < nreq; i++) {
		(void) cpc_buf_get(cpc, buf, i, &val);
		(void) printf("%9" PRId64 " ", val);
	}
	if (opts->nsets > 1)
		(void) printf(" # %s", setname);
	(void) fputc('\n', stdout);
}

#define	NSECS_PER_MSEC	1000000ll
#define	NSECS_PER_SEC	1000000000ll

static void *
gtick(void *arg)
{
	struct tstate		*state = arg;
	char			*errstr;
	uint_t			nsamples;
	uint_t			sample_cnt = 1;
	hrtime_t		ht, htdelta, restdelta;
	cpc_setgrp_t		*sgrp = state->sgrp;
	cpc_set_t		*this = cpc_setgrp_getset(sgrp);
	const char		*name = cpc_setgrp_getname(sgrp);
	cpc_buf_t		**data1, **data2, **scratch;
	cpc_buf_t		*tmp;
	int			nreqs;
	thread_t		tid;

	htdelta = NSECS_PER_MSEC * opts->mseconds;
	restdelta = NSECS_PER_MSEC * opts->mseconds_rest;
	ht = gethrtime();

	/*
	 * If this CPU is SMT, we run one gtick() thread per _physical_ CPU,
	 * instead of per cpu_t. The following check returns if it detects that
	 * this cpu_t has not been designated to do the counting for this
	 * physical CPU.
	 */
	if (smt && chip_designees[state->chip_id] != state->cpuid)
		return (NULL);

	/*
	 * If we need to run a soaker thread on this CPU, start it here.
	 */
	if (opts->dosoaker) {
		if (cond_init(&state->soak_cv, USYNC_THREAD, NULL) != 0)
			goto bad;
		if (mutex_init(&state->soak_lock, USYNC_THREAD,
		    NULL) != 0)
			goto bad;
		(void) mutex_lock(&state->soak_lock);
		state->soak_state = SOAK_PAUSE;
		if (thr_create(NULL, 0, soaker, state, 0, &tid) != 0)
			goto bad;

		while (state->soak_state == SOAK_PAUSE)
			(void) cond_wait(&state->soak_cv,
			    &state->soak_lock);
		(void) mutex_unlock(&state->soak_lock);

		/*
		 * If the soaker needs to pause for the first set, stop it now.
		 */
		if (cpc_setgrp_sysonly(sgrp) == 0) {
			(void) mutex_lock(&state->soak_lock);
			state->soak_state = SOAK_PAUSE;
			(void) mutex_unlock(&state->soak_lock);
		}
	}
	if (cpc_bind_cpu(cpc, state->cpuid, this, 0) == -1)
		goto bad;

	for (nsamples = opts->nsamples; nsamples; nsamples--, sample_cnt++) {
		hrtime_t htnow;
		struct timespec ts;

		nreqs = cpc_setgrp_getbufs(sgrp, &data1, &data2, &scratch);

		ht += htdelta;
		htnow = gethrtime();
		if (ht <= htnow)
			continue;
		ts.tv_sec = (time_t)((ht - htnow) / NSECS_PER_SEC);
		ts.tv_nsec = (suseconds_t)((ht - htnow) % NSECS_PER_SEC);

		(void) nanosleep(&ts, NULL);

		if (opts->nsets == 1) {
			/*
			 * If we're dealing with one set, buffer usage is:
			 *
			 * data1 = most recent data snapshot
			 * data2 = previous data snapshot
			 * scratch = used for diffing data1 and data2
			 *
			 * Save the snapshot from the previous sample in data2
			 * before putting the current sample in data1.
			 */
			tmp = *data1;
			*data1 = *data2;
			*data2 = tmp;
			if (cpc_set_sample(cpc, this, *data1) != 0)
				goto bad;
			cpc_buf_sub(cpc, *scratch, *data1, *data2);

			print_sample(state->cpuid, *scratch, nreqs, name, 0);
		} else {
			/*
			 * More than one set is in use (multiple -c options
			 * given). Buffer usage in this case is:
			 *
			 * data1 = total counts for this set since program began
			 * data2 = unused
			 * scratch = most recent data snapshot
			 */
			name = cpc_setgrp_getname(sgrp);
			nreqs = cpc_setgrp_getbufs(sgrp, &data1, &data2,
			    &scratch);

			if (cpc_set_sample(cpc, this, *scratch) != 0)
				goto bad;

			cpc_buf_add(cpc, *data1, *data1, *scratch);

			if (cpc_unbind(cpc, this) != 0)
				(void) fprintf(stderr, gettext("%s: error "
				    "unbinding on cpu %d - %s\n"),
				    opts->pgmname, state->cpuid,
				    strerror(errno));

			this = cpc_setgrp_nextset(sgrp);

			print_sample(state->cpuid, *scratch, nreqs, name, 0);

			/*
			 * If periodic behavior was requested, rest here.
			 */
			if (opts->doperiod && opts->mseconds_rest > 0 &&
			    (sample_cnt % opts->nsets) == 0) {
				/*
				 * Stop the soaker while the tool rests.
				 */
				if (opts->dosoaker) {
					(void) mutex_lock(&state->soak_lock);
					if (state->soak_state == SOAK_RUN)
						state->soak_state = SOAK_PAUSE;
					(void) mutex_unlock(&state->soak_lock);
				}

				htnow = gethrtime();
				ht += restdelta;
				ts.tv_sec = (time_t)((ht - htnow) /
				    NSECS_PER_SEC);
				ts.tv_nsec = (suseconds_t)((ht - htnow) %
				    NSECS_PER_SEC);

				(void) nanosleep(&ts, NULL);
			}

			/*
			 * Start or stop the soaker if needed.
			 */
			if (opts->dosoaker) {
				(void) mutex_lock(&state->soak_lock);
				if (cpc_setgrp_sysonly(sgrp) &&
				    state->soak_state == SOAK_PAUSE) {
					/*
					 * Soaker is paused but the next set is
					 * sysonly: start the soaker.
					 */
					state->soak_state = SOAK_RUN;
					(void) cond_signal(&state->soak_cv);
				} else if (cpc_setgrp_sysonly(sgrp) == 0 &&
				    state->soak_state == SOAK_RUN)
					/*
					 * Soaker is running but the next set
					 * counts user events: stop the soaker.
					 */
					state->soak_state = SOAK_PAUSE;
				(void) mutex_unlock(&state->soak_lock);
			}

			if (cpc_bind_cpu(cpc, state->cpuid, this, 0) != 0)
				goto bad;
		}
	}

	if (cpc_unbind(cpc, this) != 0)
		(void) fprintf(stderr, gettext("%s: error unbinding on"
		    " cpu %d - %s\n"), opts->pgmname,
		    state->cpuid, strerror(errno));

	/*
	 * We're done, so stop the soaker if needed.
	 */
	if (opts->dosoaker) {
		(void) mutex_lock(&state->soak_lock);
		if (state->soak_state == SOAK_RUN)
			state->soak_state = SOAK_PAUSE;
		(void) mutex_unlock(&state->soak_lock);
	}

	return (NULL);
bad:
	state->status = 3;
	errstr = strerror(errno);
	(void) fprintf(stderr, gettext("%s: cpu%d - %s\n"),
	    opts->pgmname, state->cpuid, errstr);
	return (NULL);
}

static int
cpustat(void)
{
	cpc_setgrp_t	*accum;
	cpc_set_t	*start;
	int		c, i, retval;
	int		lwps = 0;
	psetid_t	mypset, cpupset;
	char		*errstr;
	cpc_buf_t	**data1, **data2, **scratch;
	int		nreqs;
	kstat_ctl_t	*kc;

	ncpus = (int)sysconf(_SC_NPROCESSORS_CONF);
	if ((gstate = calloc(ncpus, sizeof (*gstate))) == NULL) {
		(void) fprintf(stderr, gettext(
		    "%s: out of heap\n"), opts->pgmname);
		return (1);
	}

	max_chip_id = sysconf(_SC_CPUID_MAX);
	if ((chip_designees = malloc(max_chip_id * sizeof (int))) == NULL) {
		(void) fprintf(stderr, gettext(
		    "%s: out of heap\n"), opts->pgmname);
		return (1);
	}
	for (i = 0; i < max_chip_id; i++)
		chip_designees[i] = -1;

	if (smt) {
		if ((kc = kstat_open()) == NULL) {
			(void) fprintf(stderr, gettext(
			    "%s: kstat_open() failed: %s\n"), opts->pgmname,
			    strerror(errno));
			return (1);
		}
	}

	if (opts->dosoaker)
		if (priocntl(0, 0, PC_GETCID, &fxinfo) == -1) {
			(void) fprintf(stderr, gettext(
			    "%s: couldn't get FX scheduler class: %s\n"),
			    opts->pgmname, strerror(errno));
			return (1);
		}

	/*
	 * Only include processors that are participating in the system
	 */
	for (c = 0, i = 0; i < ncpus; c++) {
		switch (p_online(c, P_STATUS)) {
		case P_ONLINE:
		case P_NOINTR:
			if (smt) {

				gstate[i].chip_id = get_chipid(kc, c);
				if (gstate[i].chip_id != -1 &&
				    chip_designees[gstate[i].chip_id] == -1)
					chip_designees[gstate[i].chip_id] = c;
			}

			gstate[i++].cpuid = c;
			break;
		case P_OFFLINE:
		case P_POWEROFF:
		case P_FAULTED:
		case P_SPARE:
		case P_DISABLED:
			gstate[i++].cpuid = -1;
			break;
		default:
			gstate[i++].cpuid = -1;
			(void) fprintf(stderr,
			    gettext("%s: cpu%d in unknown state\n"),
			    opts->pgmname, c);
			break;
		case -1:
			break;
		}
	}

	/*
	 * Examine the processor sets; if we're in one, only attempt
	 * to report on the set we're in.
	 */
	if (pset_bind(PS_QUERY, P_PID, P_MYID, &mypset) == -1) {
		errstr = strerror(errno);
		(void) fprintf(stderr, gettext("%s: pset_bind - %s\n"),
		    opts->pgmname, errstr);
	} else {
		for (i = 0; i < ncpus; i++) {
			struct tstate *this = &gstate[i];

			if (this->cpuid == -1)
				continue;

			if (pset_assign(PS_QUERY,
			    this->cpuid, &cpupset) == -1) {
				errstr = strerror(errno);
				(void) fprintf(stderr,
				    gettext("%s: pset_assign - %s\n"),
				    opts->pgmname, errstr);
				continue;
			}

			if (mypset != cpupset)
				this->cpuid = -1;
		}
	}

	if (opts->dotitle)
		print_title(opts->master);
	zerotime();

	for (i = 0; i < ncpus; i++) {
		struct tstate *this = &gstate[i];

		if (this->cpuid == -1)
			continue;
		this->sgrp = cpc_setgrp_clone(opts->master);
		if (this->sgrp == NULL) {
			this->cpuid = -1;
			continue;
		}
		if (thr_create(NULL, 0, gtick, this,
		    THR_BOUND|THR_NEW_LWP, &this->tid) == 0)
			lwps++;
		else {
			(void) fprintf(stderr,
			    gettext("%s: cannot create thread for cpu%d\n"),
			    opts->pgmname, this->cpuid);
			this->status = 4;
		}
	}

	if (lwps != 0)
		for (i = 0; i < ncpus; i++)
			(void) thr_join(gstate[i].tid, NULL, NULL);

	if ((accum = cpc_setgrp_clone(opts->master)) == NULL) {
		(void) fprintf(stderr, gettext("%s: out of heap\n"),
		    opts->pgmname);
		return (1);
	}

	retval = 0;
	for (i = 0; i < ncpus; i++) {
		struct tstate *this = &gstate[i];

		if (this->cpuid == -1)
			continue;
		cpc_setgrp_accum(accum, this->sgrp);
		cpc_setgrp_free(this->sgrp);
		this->sgrp = NULL;
		if (this->status != 0)
			retval = 1;
	}

	cpc_setgrp_reset(accum);
	start = cpc_setgrp_getset(accum);
	do {
		nreqs = cpc_setgrp_getbufs(accum, &data1, &data2, &scratch);
		print_total(lwps, *data1, nreqs, cpc_setgrp_getname(accum));
	} while (cpc_setgrp_nextset(accum) != start);

	cpc_setgrp_free(accum);
	accum = NULL;

	free(gstate);
	return (retval);
}

static int
get_chipid(kstat_ctl_t *kc, processorid_t cpuid)
{
	kstat_t		*ksp;
	kstat_named_t	*k;

	if ((ksp = kstat_lookup(kc, "cpu_info", cpuid, NULL)) == NULL)
		return (-1);

	if (kstat_read(kc, ksp, NULL) == -1) {
		(void) fprintf(stderr,
		    gettext("%s: kstat_read() failed for cpu %d: %s\n"),
		    opts->pgmname, cpuid, strerror(errno));
		return (-1);
	}

	if ((k = (kstat_named_t *)kstat_data_lookup(ksp, "chip_id")) == NULL) {
		(void) fprintf(stderr,
		    gettext("%s: chip_id not found for cpu %d: %s\n"),
		    opts->pgmname, cpuid, strerror(errno));
		return (-1);
	}

	return (k->value.i32);
}

static void *
soaker(void *arg)
{
	struct tstate	*state = arg;
	pcparms_t	pcparms;
	fxparms_t	*fx = (fxparms_t *)pcparms.pc_clparms;

	if (processor_bind(P_LWPID, P_MYID, state->cpuid, NULL) != 0)
		(void) fprintf(stderr, gettext("%s: couldn't bind soaker "
		    "thread to cpu%d: %s\n"), opts->pgmname, state->cpuid,
		    strerror(errno));

	/*
	 * Put the soaker thread in the fixed priority (FX) class so it runs
	 * at the lowest possible global priority.
	 */
	pcparms.pc_cid = fxinfo.pc_cid;
	fx->fx_upri = 0;
	fx->fx_uprilim = 0;
	fx->fx_tqsecs = fx->fx_tqnsecs = FX_TQDEF;

	if (priocntl(P_LWPID, P_MYID, PC_SETPARMS, &pcparms) != 0)
		(void) fprintf(stderr, gettext("%s: couldn't put soaker "
		    "thread in FX sched class: %s\n"), opts->pgmname,
		    strerror(errno));

	/*
	 * Let the parent thread know we're ready to roll.
	 */
	(void) mutex_lock(&state->soak_lock);
	state->soak_state = SOAK_RUN;
	(void) cond_signal(&state->soak_cv);
	(void) mutex_unlock(&state->soak_lock);

	for (;;) {
spin:
		(void) mutex_lock(&state->soak_lock);
		if (state->soak_state == SOAK_RUN) {
			(void) mutex_unlock(&state->soak_lock);
			goto spin;
		}

		while (state->soak_state == SOAK_PAUSE)
			(void) cond_wait(&state->soak_cv,
			    &state->soak_lock);
		(void) mutex_unlock(&state->soak_lock);
	}

	/*NOTREACHED*/
	return (NULL);
}
/*
 * 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 2009 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */


#include <sys/time.h>
#include <stdio.h>
#include <stdlib.h>
#include <inttypes.h>
#include <unistd.h>
#include <string.h>
#include <strings.h>
#include <limits.h>
#include <libintl.h>
#include <locale.h>
#include <errno.h>
#include <kstat.h>
#include <libcpc.h>

#include "cpucmds.h"

static struct options {
	int debug;
	int verbose;
	int dotitle;
	int dohelp;
	int dotick;
	int cpuver;
	char *pgmname;
	uint_t mseconds;
	uint_t nsamples;
	uint_t nsets;
	cpc_setgrp_t *master;
	int followfork;
	int followexec;
	pid_t pid;
	FILE *log;
} __options;

static const struct options *opts = (const struct options *)&__options;

static cpc_t *cpc;

/*
 * How many signals caught from terminal
 * We bail out as soon as possible when interrupt is set
 */
static int	interrupt = 0;

/*ARGSUSED*/
static void
cputrack_errfn(const char *fn, int subcode, const char *fmt, va_list ap)
{
	(void) fprintf(stderr, "%s: ", opts->pgmname);
	if (opts->debug)
		(void) fprintf(stderr, "%s: ", fn);
	(void) vfprintf(stderr, fmt, ap);
}

static void
cputrack_pctx_errfn(const char *fn, const char *fmt, va_list ap)
{
	cputrack_errfn(fn, -1, fmt, ap);
}

static int cputrack(int argc, char *argv[], int optind);
static void intr(int);

#if !defined(TEXT_DOMAIN)
#define	TEXT_DOMAIN	"SYS_TEST"
#endif

int
main(int argc, char *argv[])
{
	struct options *opts = &__options;
	int c, errcnt = 0;
	int nsamples;
	cpc_setgrp_t *sgrp;
	char *errstr;
	int ret;

	(void) setlocale(LC_ALL, "");
	(void) textdomain(TEXT_DOMAIN);

	if ((opts->pgmname = strrchr(argv[0], '/')) == NULL)
		opts->pgmname = argv[0];
	else
		opts->pgmname++;

	if ((cpc = cpc_open(CPC_VER_CURRENT)) == NULL) {
		errstr = strerror(errno);
		(void) fprintf(stderr, gettext("%s: cannot access performance "
		    "counter library - %s\n"), opts->pgmname, errstr);
		return (1);
	}

	(void) cpc_seterrhndlr(cpc, cputrack_errfn);
	strtoset_errfn = cputrack_errfn;

	/*
	 * Establish (non-zero) defaults
	 */
	opts->mseconds = 1000;
	opts->dotitle = 1;
	opts->log = stdout;
	if ((opts->master = cpc_setgrp_new(cpc, 0)) == NULL) {
		(void) fprintf(stderr, gettext("%s: no memory available\n"),
		    opts->pgmname);
		exit(1);
	}

	while ((c = getopt(argc, argv, "T:N:Defhntvo:r:c:p:")) != EOF)
		switch (c) {
		case 'T':			/* sample time,	seconds */
			opts->mseconds = (uint_t)(atof(optarg) * 1000.0);
			break;
		case 'N':			/* number of samples */
			nsamples = atoi(optarg);
			if (nsamples < 0)
				errcnt++;
			else
				opts->nsamples = (uint_t)nsamples;
			break;
		case 'D':			/* enable debugging */
			opts->debug++;
			break;
		case 'f':			/* follow fork */
			opts->followfork++;
			break;
		case 'e':			/* follow exec */
			opts->followexec++;
			break;
		case 'n':			/* no titles */
			opts->dotitle = 0;
			break;
		case 't':			/* print %tick */
			opts->dotick = 1;
			break;
		case 'v':
			opts->verbose = 1;	/* more chatty */
			break;
		case 'o':
			if (optarg == NULL) {
				errcnt++;
				break;
			}
			if ((opts->log = fopen(optarg, "w")) == NULL) {
				(void) fprintf(stderr, gettext(
				    "%s: cannot open '%s' for writing\n"),
				    opts->pgmname, optarg);
				return (1);
			}
			break;
		case 'c':			/* specify statistics */
			if ((sgrp = cpc_setgrp_newset(opts->master,
			    optarg, &errcnt)) != NULL)
				opts->master = sgrp;
			break;
		case 'p':			/* grab given pid */
			if ((opts->pid = atoi(optarg)) <= 0)
				errcnt++;
			break;
		case 'h':
			opts->dohelp = 1;
			break;
		case '?':
		default:
			errcnt++;
			break;
		}

	if (opts->nsamples == 0)
		opts->nsamples = UINT_MAX;

	if (errcnt != 0 ||
	    opts->dohelp ||
	    (argc == optind && opts->pid == 0) ||
	    (argc > optind && opts->pid != 0) ||
	    (opts->nsets = cpc_setgrp_numsets(opts->master)) == 0) {
		(void) fprintf(opts->dohelp ? stdout : stderr, gettext(
		    "Usage:\n\t%s [-T secs] [-N count] [-Defhnv] [-o file]\n"
		    "\t\t-c events [command [args] | -p pid]\n\n"
		    "\t-T secs\t  seconds between samples, default 1\n"
		    "\t-N count  number of samples, default unlimited\n"
		    "\t-D\t  enable debug mode\n"
		    "\t-e\t  follow exec(2), and execve(2)\n"
		    "\t-f\t  follow fork(2), fork1(2), and vfork(2)\n"
		    "\t-h\t  print extended usage information\n"
		    "\t-n\t  suppress titles\n"
		    "\t-t\t  include virtualized %s register\n"
		    "\t-v\t  verbose mode\n"
		    "\t-o file\t  write cpu statistics to this file\n"
		    "\t-c events specify processor events to be monitored\n"
		    "\t-p pid\t  pid of existing process to capture\n\n"
		    "\tUse cpustat(8) to monitor system-wide statistics.\n"),
		    opts->pgmname, CPC_TICKREG_NAME);
		if (opts->dohelp) {
			(void) putchar('\n');
			(void) capabilities(cpc, stdout);
			exit(0);
		}
		exit(2);
	}

	/*
	 * Catch signals from terminal, so they can be handled asynchronously
	 * when we're ready instead of when we're not (;-)
	 */
	if (sigset(SIGHUP, SIG_IGN) == SIG_DFL)
		(void) sigset(SIGHUP, intr);
	if (sigset(SIGINT, SIG_IGN) == SIG_DFL)
		(void) sigset(SIGINT, intr);
	if (sigset(SIGQUIT, SIG_IGN) == SIG_DFL)
		(void) sigset(SIGQUIT, intr);
	(void) sigset(SIGPIPE, intr);
	(void) sigset(SIGTERM, intr);

	cpc_setgrp_reset(opts->master);
	(void) setvbuf(opts->log, NULL, _IOLBF, 0);
	ret = cputrack(argc, argv, optind);
	(void) cpc_close(cpc);
	return (ret);
}

static void
print_title(cpc_setgrp_t *sgrp)
{
	(void) fprintf(opts->log, "%7s ", "time");
	if (opts->followfork)
		(void) fprintf(opts->log, "%6s ", "pid");
	(void) fprintf(opts->log, "%3s %10s ", "lwp", "event");
	if (opts->dotick)
		(void) fprintf(opts->log, "%9s ", CPC_TICKREG_NAME);
	(void) fprintf(opts->log, "%s\n", cpc_setgrp_gethdr(sgrp));
	(void) fflush(opts->log);
}

static void
print_exec(float now, pid_t pid, char *name)
{
	if (name == NULL)
		name = "(unknown)";

	(void) fprintf(opts->log, "%7.3f ", now);
	if (opts->followfork)
		(void) fprintf(opts->log, "%6d ", (int)pid);
	(void) fprintf(opts->log, "%3d %10s ", 1, "exec");
	if (opts->dotick)
		(void) fprintf(opts->log, "%9s ", "");
	(void) fprintf(opts->log, "%9s %9s # '%s'\n", "", "", name);
	(void) fflush(opts->log);
}

static void
print_fork(float now, pid_t newpid, id_t lwpid, pid_t oldpid)
{
	(void) fprintf(opts->log, "%7.3f ", now);
	if (opts->followfork)
		(void) fprintf(opts->log, "%6d ", (int)oldpid);
	(void) fprintf(opts->log, "%3d %10s ", (int)lwpid, "fork");
	if (opts->dotick)
		(void) fprintf(opts->log, "%9s ", "");
	(void) fprintf(opts->log, "%9s %9s # %d\n", "", "", (int)newpid);
	(void) fflush(opts->log);
}

static void
print_sample(pid_t pid, id_t lwpid,
    char *pevent, cpc_buf_t *buf, int nreq, const char *evname)
{
	uint64_t	val;
	int		i;

	(void) fprintf(opts->log, "%7.3f ",
	    mstimestamp(cpc_buf_hrtime(cpc, buf)));
	if (opts->followfork)
		(void) fprintf(opts->log, "%6d ", (int)pid);
	(void) fprintf(opts->log, "%3d %10s ", (int)lwpid, pevent);
	if (opts->dotick)
		(void) fprintf(opts->log, "%9" PRId64 " ",
		    cpc_buf_tick(cpc, buf));
	for (i = 0; i < nreq; i++) {
		(void) cpc_buf_get(cpc, buf, i, &val);
		(void) fprintf(opts->log, "%9" PRId64 " ", val);
	}
	if (opts->nsets > 1)
		(void) fprintf(opts->log, " # %s\n", evname);
	else
		(void) fputc('\n', opts->log);
}

struct pstate {
	cpc_setgrp_t *accum;
	cpc_setgrp_t **sgrps;
	int maxlwpid;
};

static int
pinit_lwp(pctx_t *pctx, pid_t pid, id_t lwpid, void *arg)
{
	struct pstate *state = arg;
	cpc_setgrp_t *sgrp;
	cpc_set_t *set;
	cpc_buf_t **data1, **data2, **scratch;
	char *errstr;
	int nreq;

	if (interrupt)
		return (0);

	if (state->maxlwpid < lwpid) {
		state->sgrps = realloc(state->sgrps,
		    lwpid * sizeof (state->sgrps));
		if (state->sgrps == NULL) {
			(void) fprintf(stderr, gettext(
			    "%6d: init_lwp: out of memory\n"), (int)pid);
			return (-1);
		}
		while (state->maxlwpid < lwpid) {
			state->sgrps[state->maxlwpid] = NULL;
			state->maxlwpid++;
		}
	}

	if ((sgrp = state->sgrps[lwpid-1]) == NULL) {
		if ((sgrp = cpc_setgrp_clone(opts->master)) == NULL) {
			(void) fprintf(stderr, gettext(
			    "%6d: init_lwp: out of memory\n"), (int)pid);
			return (-1);
		}
		state->sgrps[lwpid-1] = sgrp;
		set = cpc_setgrp_getset(sgrp);
	} else {
		cpc_setgrp_reset(sgrp);
		set = cpc_setgrp_getset(sgrp);
	}

	nreq = cpc_setgrp_getbufs(sgrp, &data1, &data2, &scratch);

	if (cpc_bind_pctx(cpc, pctx, lwpid, set, 0) != 0 ||
	    cpc_set_sample(cpc, set, *data2) != 0) {
		errstr = strerror(errno);
		if (errno == EAGAIN) {
			(void) cpc_unbind(cpc, set);
		}

		(void) fprintf(stderr, gettext(
		    "%6d: init_lwp: can't bind perf counters "
		    "to lwp%d - %s\n"), (int)pid, (int)lwpid, errstr);
		return (-1);
	}

	if (opts->verbose)
		print_sample(pid, lwpid, "init_lwp",
		    *data2, nreq, cpc_setgrp_getname(sgrp));
	return (0);
}

/*ARGSUSED*/
static int
pfini_lwp(pctx_t *pctx, pid_t pid, id_t lwpid, void *arg)
{
	struct pstate *state = arg;
	cpc_setgrp_t *sgrp = state->sgrps[lwpid-1];
	cpc_set_t *set;
	char *errstr;
	cpc_buf_t **data1, **data2, **scratch;
	int nreq;

	if (interrupt)
		return (0);

	set = cpc_setgrp_getset(sgrp);
	nreq = cpc_setgrp_getbufs(sgrp, &data1, &data2, &scratch);
	if (cpc_set_sample(cpc, set, *scratch) == 0) {
		if (opts->nsets == 1) {
			/*
			 * When we only have one set of counts, the sample
			 * gives us the accumulated count.
			 */
			*data1 = *scratch;
		} else {
			/*
			 * When we have more than one set of counts, the
			 * sample gives us the count for the latest sample
			 * period. *data1 contains the accumulated count but
			 * does not include the count for the latest sample
			 * period for this set of counters.
			 */
			cpc_buf_add(cpc, *data1, *data1, *scratch);
		}
		if (opts->verbose)
			print_sample(pid, lwpid, "fini_lwp",
			    *data1, nreq, cpc_setgrp_getname(sgrp));
		cpc_setgrp_accum(state->accum, sgrp);
		if (cpc_unbind(cpc, set) == 0)
			return (0);
	}

	switch (errno) {
	case EAGAIN:
		(void) fprintf(stderr, gettext("%6d: fini_lwp: "
		    "lwp%d: perf counter contents invalidated\n"),
		    (int)pid, (int)lwpid);
		break;
	default:
		errstr = strerror(errno);
		(void) fprintf(stderr, gettext("%6d: fini_lwp: "
		    "lwp%d: can't access perf counters - %s\n"),
		    (int)pid, (int)lwpid, errstr);
		break;
	}
	return (-1);
}

/*ARGSUSED*/
static int
plwp_create(pctx_t *pctx, pid_t pid, id_t lwpid, void *arg)
{
	cpc_setgrp_t	*sgrp = opts->master;
	cpc_buf_t	**data1, **data2, **scratch;
	int		nreq;

	if (interrupt)
		return (0);

	nreq = cpc_setgrp_getbufs(sgrp, &data1, &data2, &scratch);

	print_sample(pid, lwpid, "lwp_create",
	    *data1, nreq, cpc_setgrp_getname(sgrp));

	return (0);
}

/*ARGSUSED*/
static int
plwp_exit(pctx_t *pctx, pid_t pid, id_t lwpid, void *arg)
{
	struct pstate	*state = arg;
	cpc_setgrp_t	*sgrp = state->sgrps[lwpid-1];
	cpc_set_t	*start;
	int		nreq;
	cpc_buf_t	**data1, **data2, **scratch;

	if (interrupt)
		return (0);

	start = cpc_setgrp_getset(sgrp);
	do {
		nreq = cpc_setgrp_getbufs(sgrp, &data1, &data2, &scratch);
		if (cpc_buf_hrtime(cpc, *data1) == 0)
			continue;
		print_sample(pid, lwpid, "lwp_exit",
		    *data1, nreq, cpc_setgrp_getname(sgrp));
	} while (cpc_setgrp_nextset(sgrp) != start);

	return (0);
}

/*ARGSUSED*/
static int
pexec(pctx_t *pctx, pid_t pid, id_t lwpid, char *name, void *arg)
{
	struct pstate	*state = arg;
	float		now = 0.0;
	cpc_set_t	*start;
	int		nreq;
	cpc_buf_t	**data1, **data2, **scratch;
	hrtime_t	hrt;

	if (interrupt)
		return (0);

	/*
	 * Print the accumulated results from the previous program image
	 */
	cpc_setgrp_reset(state->accum);
	start = cpc_setgrp_getset(state->accum);
	do {
		nreq = cpc_setgrp_getbufs(state->accum, &data1, &data2,
		    &scratch);
		hrt = cpc_buf_hrtime(cpc, *data1);
		if (hrt == 0)
			continue;
		print_sample(pid, lwpid, "exec",
		    *data1, nreq, cpc_setgrp_getname(state->accum));
		if (now < mstimestamp(hrt))
			now = mstimestamp(hrt);
	} while (cpc_setgrp_nextset(state->accum) != start);

	print_exec(now, pid, name);

	if (state->accum != NULL) {
		cpc_setgrp_free(state->accum);
		state->accum = NULL;
	}

	if (opts->followexec) {
		state->accum = cpc_setgrp_clone(opts->master);
		return (0);
	}
	return (-1);
}

/*ARGSUSED*/
static void
pexit(pctx_t *pctx, pid_t pid, id_t lwpid, int status, void *arg)
{
	struct pstate	*state = arg;
	cpc_set_t	*start;
	int		nreq;
	cpc_buf_t	**data1, **data2, **scratch;

	if (interrupt)
		return;

	cpc_setgrp_reset(state->accum);
	start = cpc_setgrp_getset(state->accum);
	do {
		nreq = cpc_setgrp_getbufs(state->accum, &data1, &data2,
		    &scratch);
		if (cpc_buf_hrtime(cpc, *data1) == 0)
			continue;
		print_sample(pid, lwpid, "exit",
		    *data1, nreq, cpc_setgrp_getname(state->accum));
	} while (cpc_setgrp_nextset(state->accum) != start);

	cpc_setgrp_free(state->accum);
	state->accum = NULL;

	for (lwpid = 1; lwpid < state->maxlwpid; lwpid++)
		if (state->sgrps[lwpid-1] != NULL) {
			cpc_setgrp_free(state->sgrps[lwpid-1]);
			state->sgrps[lwpid-1] = NULL;
		}
	free(state->sgrps);
	state->sgrps = NULL;
}

static int
ptick(pctx_t *pctx, pid_t pid, id_t lwpid, void *arg)
{
	struct pstate *state = arg;
	cpc_setgrp_t *sgrp = state->sgrps[lwpid-1];
	cpc_set_t *this = cpc_setgrp_getset(sgrp);
	const char *name = cpc_setgrp_getname(sgrp);
	cpc_buf_t **data1, **data2, **scratch, *tmp;
	char *errstr;
	int nreqs;

	if (interrupt)
		return (0);

	nreqs = cpc_setgrp_getbufs(sgrp, &data1, &data2, &scratch);

	if (opts->nsets == 1) {
		/*
		 * If we're dealing with one set, buffer usage is:
		 *
		 * data1 = most recent data snapshot
		 * data2 = previous data snapshot
		 * scratch = used for diffing data1 and data2
		 *
		 * Save the snapshot from the previous sample in data2
		 * before putting the current sample in data1.
		 */
		tmp = *data1;
		*data1 = *data2;
		*data2 = tmp;
		if (cpc_set_sample(cpc, this, *data1) != 0)
			goto broken;
		cpc_buf_sub(cpc, *scratch, *data1, *data2);
	} else {
		cpc_set_t *next = cpc_setgrp_nextset(sgrp);
		/*
		 * If there is more than set in use, we will need to
		 * unbind and re-bind on each go-around because each
		 * time a counter is bound, it is preset to 0 (as it was
		 * specified when the requests were added to the set).
		 *
		 * Buffer usage in this case is:
		 *
		 * data1 = total counts for this set since program began
		 * data2 = unused
		 * scratch = most recent data snapshot
		 */

		if (cpc_set_sample(cpc, this, *scratch) != 0)
			goto broken;
		cpc_buf_add(cpc, *data1, *data1, *scratch);

		/*
		 * No need to unbind the previous set, as binding another set
		 * automatically unbinds the most recently bound set.
		 */
		if (cpc_bind_pctx(cpc, pctx, lwpid, next, 0) != 0)
			goto broken;
	}
	print_sample(pid, lwpid, "tick", *scratch, nreqs, name);

	return (0);

broken:
	switch (errno) {
	case EAGAIN:
		(void) fprintf(stderr, gettext(
		    "%6d: tick: lwp%d: perf counter contents invalidated\n"),
		    (int)pid, (int)lwpid);
		break;
	default:
		errstr = strerror(errno);
		(void) fprintf(stderr, gettext(
		    "%6d: tick: lwp%d: can't access perf counter - %s\n"),
		    (int)pid, (int)lwpid, errstr);
		break;
	}
	(void) cpc_unbind(cpc, this);
	return (-1);
}

/*
 * The system has just created a new address space that has a new pid.
 * We're running in a child of the controlling process, with a new
 * pctx handle already opened on the child of the original controlled process.
 */
static void
pfork(pctx_t *pctx, pid_t oldpid, pid_t pid, id_t lwpid, void *arg)
{
	struct pstate *state = arg;

	print_fork(mstimestamp(0), pid, lwpid, oldpid);

	if (!opts->followfork)
		return;

	if (pctx_set_events(pctx,
	    PCTX_SYSC_EXEC_EVENT, pexec,
	    PCTX_SYSC_FORK_EVENT, pfork,
	    PCTX_SYSC_EXIT_EVENT, pexit,
	    PCTX_SYSC_LWP_CREATE_EVENT, plwp_create,
	    PCTX_INIT_LWP_EVENT, pinit_lwp,
	    PCTX_FINI_LWP_EVENT, pfini_lwp,
	    PCTX_SYSC_LWP_EXIT_EVENT, plwp_exit,
	    PCTX_NULL_EVENT) == 0) {
		state->accum = cpc_setgrp_clone(opts->master);
		(void) pctx_run(pctx, opts->mseconds, opts->nsamples, ptick);
		if (state->accum) {
			free(state->accum);
			state->accum = NULL;
		}
	}
}

/*
 * Translate the incoming options into actions, and get the
 * tool and the process to control running.
 */
static int
cputrack(int argc, char *argv[], int optind)
{
	struct pstate __state, *state = &__state;
	pctx_t *pctx;
	int err;

	bzero(state, sizeof (*state));

	if (opts->pid == 0) {
		if (argc <= optind) {
			(void) fprintf(stderr, "%s: %s\n",
			    opts->pgmname,
			    gettext("no program to start"));
			return (1);
		}
		pctx = pctx_create(argv[optind],
		    &argv[optind], state, 1, cputrack_pctx_errfn);
		if (pctx == NULL) {
			(void) fprintf(stderr, "%s: %s '%s'\n",
			    opts->pgmname,
			    gettext("failed to start program"),
			    argv[optind]);
			return (1);
		}
	} else {
		pctx = pctx_capture(opts->pid, state, 1, cputrack_pctx_errfn);
		if (pctx == NULL) {
			(void) fprintf(stderr, "%s: %s %d\n",
			    opts->pgmname,
			    gettext("failed to capture pid"),
			    (int)opts->pid);
			return (1);
		}
	}

	err = pctx_set_events(pctx,
	    PCTX_SYSC_EXEC_EVENT, pexec,
	    PCTX_SYSC_FORK_EVENT, pfork,
	    PCTX_SYSC_EXIT_EVENT, pexit,
	    PCTX_SYSC_LWP_CREATE_EVENT, plwp_create,
	    PCTX_INIT_LWP_EVENT, pinit_lwp,
	    PCTX_FINI_LWP_EVENT, pfini_lwp,
	    PCTX_SYSC_LWP_EXIT_EVENT, plwp_exit,
	    PCTX_NULL_EVENT);

	if (err != 0) {
		(void) fprintf(stderr, "%s: %s\n",
		    opts->pgmname,
		    gettext("can't bind process context ops to process"));
	} else {
		if (opts->dotitle)
			print_title(opts->master);
		state->accum = cpc_setgrp_clone(opts->master);
		zerotime();
		err = pctx_run(pctx, opts->mseconds, opts->nsamples, ptick);
		if (state->accum) {
			cpc_setgrp_free(state->accum);
			state->accum = NULL;
		}
	}

	return (err != 0 ? 1 : 0);
}

/*ARGSUSED*/
static void
intr(int sig)
{
	interrupt++;
	if (cpc != NULL)
		cpc_terminate(cpc);
}
/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License, Version 1.0 only
 * (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 2004 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <strings.h>
#include <libintl.h>
#include <sys/types.h>


#include <libcpc.h>
#include "cpucmds.h"

#define	CHARS_PER_REQ 11	/* space required for printing column headers */

/*
 * These routines are solely used to manage a list of request sets.
 */

struct __cpc_setgrp {
	struct setgrp_elem {
		cpc_set_t	*set;
		uint8_t		sysonly;	/* All reqs sys-mode only ? */
		int		nreqs;
		int		*picnums;	/* picnum used per req */
		cpc_buf_t	*data1;
		cpc_buf_t	*data2;
		cpc_buf_t	*scratch;
		char *name;
		char *hdr;
	} *sets;		/* array of events and names */
	int nelem;		/* size of array */
	int current;		/* currently bound event in eventset */
	int smt;		/* Measures physical events on SMT CPU */
	int has_sysonly_set;	/* Does this group have a system-only set? */
	cpc_t *cpc;		/* library handle */
};

static void *emalloc(size_t n);

cpc_setgrp_t *
cpc_setgrp_new(cpc_t *cpc, int smt)
{
	cpc_setgrp_t *sgrp;

	sgrp = emalloc(sizeof (*sgrp));
	sgrp->current = -1;
	sgrp->cpc = cpc;
	sgrp->smt = smt;
	sgrp->has_sysonly_set = 0;
	return (sgrp);
}

/*
 * Walker to count the number of requests in a set, and check if any requests
 * count user-mode events.
 */
/*ARGSUSED*/
static void
cpc_setgrp_walker(void *arg, int index, const char *event, uint64_t preset,
    uint_t flags, int nattrs, const cpc_attr_t *attrs)
{
	struct setgrp_elem *se = arg;

	se->nreqs++;
	if (flags & CPC_COUNT_USER)
		se->sysonly = 0;
}

/*
 * Walker to discover the picnums used by the requests in a set.
 */
/*ARGSUSED*/
static void
cpc_setgrp_picwalker(void *arg, int index, const char *event, uint64_t preset,
    uint_t flags, int nattrs, const cpc_attr_t *attrs)
{
	int *picnums = arg;
	int i;

	for (i = 0; i < nattrs; i++) {
		if (strncmp(attrs[i].ca_name, "picnum", 7) == 0)
			break;
	}
	if (i == nattrs)
		picnums[index] = -1;

	picnums[index] = (int)attrs[i].ca_val;
}

cpc_setgrp_t *
cpc_setgrp_newset(cpc_setgrp_t *sgrp, const char *spec, int *errcnt)
{
	cpc_set_t		*set;
	struct setgrp_elem	*new;
	char			hdr[CHARS_PER_REQ+1];
	int			i;

	if ((set = cpc_strtoset(sgrp->cpc, spec, sgrp->smt)) == NULL) {
		*errcnt += 1;
		return (NULL);
	}

	if ((new = realloc(sgrp->sets, (1 + sgrp->nelem) * sizeof (*new)))
	    == NULL) {
		(void) fprintf(stderr,
		    gettext("cpc_setgrp: no re memory available\n"));
		exit(0);
	}

	sgrp->sets = new;
	sgrp->sets[sgrp->nelem].set = set;
	/*
	 * Count the number of requests in the set we just made. If any requests
	 * in the set have CPC_COUNT_USER in the flags, the sysonly flag will
	 * be cleared.
	 */
	sgrp->sets[sgrp->nelem].nreqs = 0;
	sgrp->sets[sgrp->nelem].sysonly = 1;
	cpc_walk_requests(sgrp->cpc, set, &(sgrp->sets[sgrp->nelem]),
	    cpc_setgrp_walker);

	if (sgrp->sets[sgrp->nelem].sysonly == 1)
		sgrp->has_sysonly_set = 1;

	sgrp->sets[sgrp->nelem].picnums = emalloc(sgrp->sets[sgrp->nelem].nreqs
	    * sizeof (int));

	sgrp->sets[sgrp->nelem].hdr = emalloc((sgrp->sets[sgrp->nelem].nreqs *
	    CHARS_PER_REQ) + 1);

	/*
	 * Find out which picnums the requests are using.
	 */
	cpc_walk_requests(sgrp->cpc, set, sgrp->sets[sgrp->nelem].picnums,
	    cpc_setgrp_picwalker);
	/*
	 * Use the picnums we discovered to build a printable header for this
	 * set.
	 */
	sgrp->sets[sgrp->nelem].hdr[0] = '\0';
	for (i = 0; i < sgrp->sets[sgrp->nelem].nreqs; i++) {
		(void) snprintf(hdr, CHARS_PER_REQ, "%8s%-2d ", "pic",
		    sgrp->sets[sgrp->nelem].picnums[i]);
		(void) strncat(sgrp->sets[sgrp->nelem].hdr, hdr,
		    sgrp->sets[sgrp->nelem].nreqs * CHARS_PER_REQ);
	}
	sgrp->sets[sgrp->nelem].hdr[strlen(sgrp->sets[sgrp->nelem].hdr)] = '\0';

	if ((sgrp->sets[sgrp->nelem].name = strdup(spec)) == NULL) {
		(void) fprintf(stderr,
		    gettext("cpc_setgrp: no memory available\n"));
		exit(0);
	}

	if ((sgrp->sets[sgrp->nelem].data1 = cpc_buf_create(sgrp->cpc, set))
	    == NULL ||
	    (sgrp->sets[sgrp->nelem].data2 = cpc_buf_create(sgrp->cpc, set))
	    == NULL ||
	    (sgrp->sets[sgrp->nelem].scratch = cpc_buf_create(sgrp->cpc, set))
	    == NULL) {
		(void) fprintf(stderr,
		    gettext("cpc_setgrp: no memory available\n"));
		exit(0);
	}

	if (sgrp->current < 0)
		sgrp->current = 0;
	sgrp->nelem++;
	return (sgrp);
}

int
cpc_setgrp_getbufs(cpc_setgrp_t *sgrp, cpc_buf_t ***data1, cpc_buf_t ***data2,
    cpc_buf_t ***scratch)
{
	if ((uint_t)sgrp->current >= sgrp->nelem)
		return (-1);

	*data1   = &(sgrp->sets[sgrp->current].data1);
	*data2   = &(sgrp->sets[sgrp->current].data2);
	*scratch = &(sgrp->sets[sgrp->current].scratch);

	return (sgrp->sets[sgrp->current].nreqs);
}

cpc_setgrp_t *
cpc_setgrp_clone(cpc_setgrp_t *old)
{
	int			i;
	cpc_setgrp_t		*new;
	struct setgrp_elem	*newa;

	new = emalloc(sizeof (*new));
	newa = emalloc(old->nelem * sizeof (*newa));

	new->nelem = old->nelem;
	new->current = old->current;
	new->cpc = old->cpc;
	new->sets = newa;
	new->smt = old->smt;
	new->has_sysonly_set = old->has_sysonly_set;
	for (i = 0; i < old->nelem; i++) {
		if ((newa[i].set = cpc_strtoset(old->cpc, old->sets[i].name,
		    old->smt)) == NULL) {
			(void) fprintf(stderr,
			    gettext("cpc_setgrp: cpc_strtoset() failed\n"));
			exit(0);
		}
		if ((newa[i].name = strdup(old->sets[i].name)) == NULL) {
			(void) fprintf(stderr,
			    gettext("cpc_setgrp: no memory available\n"));
			exit(0);
		}
		newa[i].sysonly = old->sets[i].sysonly;
		newa[i].nreqs = old->sets[i].nreqs;
		newa[i].data1 = cpc_buf_create(old->cpc, newa[i].set);
		newa[i].data2 = cpc_buf_create(old->cpc, newa[i].set);
		newa[i].scratch = cpc_buf_create(old->cpc, newa[i].set);
		if (newa[i].data1 == NULL || newa[i].data2 == NULL ||
		    newa[i].scratch == NULL) {
			(void) fprintf(stderr,
			    gettext("cpc_setgrp: no memory available\n"));
			exit(0);
		}
		cpc_buf_copy(old->cpc, newa[i].data1, old->sets[i].data1);
		cpc_buf_copy(old->cpc, newa[i].data2, old->sets[i].data2);
		cpc_buf_copy(old->cpc, newa[i].scratch, old->sets[i].scratch);
	}
	return (new);
}

static void
cpc_setgrp_delset(cpc_setgrp_t *sgrp)
{
	int l;

	if ((uint_t)sgrp->current >= sgrp->nelem)
		sgrp->current = sgrp->nelem - 1;
	if (sgrp->current < 0)
		return;
	free(sgrp->sets[sgrp->current].name);
	free(sgrp->sets[sgrp->current].hdr);
	free(sgrp->sets[sgrp->current].picnums);
	(void) cpc_buf_destroy(sgrp->cpc, sgrp->sets[sgrp->current].data1);
	(void) cpc_buf_destroy(sgrp->cpc, sgrp->sets[sgrp->current].data2);
	(void) cpc_buf_destroy(sgrp->cpc, sgrp->sets[sgrp->current].scratch);
	for (l = sgrp->current; l < sgrp->nelem - 1; l++)
		sgrp->sets[l] = sgrp->sets[l + 1];
	sgrp->nelem--;
}

void
cpc_setgrp_free(cpc_setgrp_t *sgrp)
{
	if (sgrp->sets) {
		while (sgrp->nelem)
			cpc_setgrp_delset(sgrp);
		free(sgrp->sets);
	}
	free(sgrp);
}

cpc_set_t *
cpc_setgrp_getset(cpc_setgrp_t *sgrp)
{
	if ((uint_t)sgrp->current >= sgrp->nelem)
		return (NULL);
	return (sgrp->sets[sgrp->current].set);
}

const char *
cpc_setgrp_getname(cpc_setgrp_t *sgrp)
{
	if ((uint_t)sgrp->current >= sgrp->nelem)
		return (NULL);
	return (sgrp->sets[sgrp->current].name);
}

const char *
cpc_setgrp_gethdr(cpc_setgrp_t *sgrp)
{
	if ((uint_t)sgrp->current >= sgrp->nelem)
		return (NULL);
	return (sgrp->sets[sgrp->current].hdr);
}

int
cpc_setgrp_numsets(cpc_setgrp_t *sgrp)
{
	return (sgrp->nelem);
}

cpc_set_t *
cpc_setgrp_nextset(cpc_setgrp_t *sgrp)
{
	if (sgrp->current < 0)
		return (NULL);

	if (++sgrp->current >= sgrp->nelem)
		sgrp->current = 0;

	return (cpc_setgrp_getset(sgrp));
}

/*
 * Put the setgrp pointer back to the beginning of the set
 */
void
cpc_setgrp_reset(cpc_setgrp_t *sgrp)
{
	if (sgrp->current > 0)
		sgrp->current = 0;
}

/*
 * Adds the data from the 'data1' buf into the accum setgrp.
 */
void
cpc_setgrp_accum(cpc_setgrp_t *accum, cpc_setgrp_t *sgrp)
{
	int i;

	cpc_setgrp_reset(accum);
	cpc_setgrp_reset(sgrp);
	if (accum->nelem != sgrp->nelem)
		return;

	for (i = 0; i < sgrp->nelem; i++) {
		if (accum->sets[i].nreqs != sgrp->sets[i].nreqs)
			return;
		cpc_buf_add(sgrp->cpc, accum->sets[i].data1,
		    accum->sets[i].data1, sgrp->sets[i].data1);
	}
}

/*
 * Returns 1 if all requests in the current set count only system-mode events.
 */
int
cpc_setgrp_sysonly(cpc_setgrp_t *sgrp)
{
	return ((int)sgrp->sets[sgrp->current].sysonly);
}

/*
 * Returns 1 if any set in the group is a system-mode-only set.
 */
int
cpc_setgrp_has_sysonly(cpc_setgrp_t *sgrp)
{
	return (sgrp->has_sysonly_set);
}

/*
 * If we ever fail to get memory, we print an error message and exit.
 */
static void *
emalloc(size_t n)
{
	/*
	 * Several callers of this routine need zero-filled buffers.
	 */
	void *p = calloc(1, n);

	if (p == NULL) {
		(void) fprintf(stderr,
		    gettext("cpc_setgrp: no memory available\n"));
		exit(0);
	}

	return (p);
}
/*
 * 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 (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
 */

#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <libintl.h>

#include "libcpc.h"

/*
 * Takes a string and converts it to a cpc_set_t.
 *
 * While processing the string using getsubopt(), we will use an array of
 * requests to hold the data, and a proprietary representation of attributes
 * which allow us to avoid a realloc()/bcopy() dance every time we come across
 * a new attribute.
 *
 * Not until after the string has been processed in its entirety do we
 * allocate and specify a request set properly.
 */

/*
 * Leave enough room in token strings for picn, nousern, or sysn where n is
 * picnum.
 */
#define	TOK_SIZE	10

typedef struct __tmp_attr {
	char			*name;
	uint64_t		val;
	struct __tmp_attr	*next;
} tmp_attr_t;

typedef struct __tok_info {
	char			*name;
	int			picnum;
} tok_info_t;

typedef struct __request_t {
	char			cr_event[CPC_MAX_EVENT_LEN];
	uint_t			cr_flags;
	uint_t			cr_nattrs;	/* # CPU-specific attrs */
} request_t;

static void strtoset_cleanup(void);
static void smt_special(int picnum);
static void *emalloc(size_t n);

/*
 * Clients of cpc_strtoset may set this to specify an error handler during
 * string parsing.
 */
cpc_errhndlr_t		*strtoset_errfn = NULL;

static request_t	*reqs;
static int		nreqs;
static int		ncounters;

static tmp_attr_t	**attrs;
static int		ntoks;
static char		**toks;
static tok_info_t	*tok_info;
static int		(*(*tok_funcs))(int, char *);
static char		**attrlist;	/* array of ptrs to toks in attrlistp */
static int		nattrs;
static cpc_t		*cpc;
static int		found;

static void
strtoset_err(const char *fmt, ...)
{
	va_list ap;

	if (strtoset_errfn == NULL)
		return;

	va_start(ap, fmt);
	(*strtoset_errfn)("cpc_strtoset", -1, fmt, ap);
	va_end(ap);
}

/*ARGSUSED*/
static void
event_walker(void *arg, uint_t picno, const char *event)
{
	if (strncmp(arg, event, CPC_MAX_EVENT_LEN) == 0)
		found = 1;
}

static int
event_valid(int picnum, char *event)
{
	char *end_event;
	int err;

	found = 0;

	cpc_walk_events_pic(cpc, picnum, event, event_walker);

	if (found)
		return (1);

	cpc_walk_generic_events_pic(cpc, picnum, event, event_walker);

	if (found)
		return (1);

	/*
	 * Before assuming this is an invalid event, see if we have been given
	 * a raw event code.
	 * Check the second argument of strtol() to ensure invalid events
	 * beginning with number do not go through.
	 */
	err = errno;
	errno = 0;
	(void) strtol(event, &end_event, 0);
	if ((errno == 0) && (*end_event == '\0')) {
		/*
		 * Success - this is a valid raw code in hex, decimal, or octal.
		 */
		errno = err;
		return (1);
	}

	errno = err;
	return (0);
}

/*
 * An unknown token was encountered; check here if it is an implicit event
 * name. We allow users to omit the "picn=" portion of the event spec, and
 * assign such events to available pics in order they are returned from
 * getsubopt(3C). We start our search for an available pic _after_ the highest
 * picnum to be assigned. This ensures that the event spec can never be out of
 * order; i.e. if the event string is "eventa,eventb" we must ensure that the
 * picnum counting eventa is less than the picnum counting eventb.
 */
static int
find_event(char *event)
{
	int i;

	/*
	 * Event names cannot have '=' in them. If present here, it means we
	 * have encountered an unknown token (foo=bar, for example).
	 */
	if (strchr(event, '=') != NULL)
		return (0);

	/*
	 * Find the first unavailable pic, after which we must start our search.
	 */
	for (i = ncounters - 1; i >= 0; i--) {
		if (reqs[i].cr_event[0] != '\0')
			break;
	}
	/*
	 * If the last counter has been assigned, we cannot place this event.
	 */
	if (i == ncounters - 1)
		return (0);

	/*
	 * If none of the counters have been assigned yet, i is -1 and we will
	 * begin our search at 0. Else we begin our search at the counter after
	 * the last one currently assigned.
	 */
	i++;

	for (; i < ncounters; i++) {
		if (event_valid(i, event) == 0)
			continue;

		nreqs++;
		(void) strncpy(reqs[i].cr_event, event, CPC_MAX_EVENT_LEN);
		return (1);
	}

	return (0);
}

static int
pic(int tok, char *val)
{
	int picnum = tok_info[tok].picnum;
	/*
	 * Make sure the each pic only appears in the spec once.
	 */
	if (reqs[picnum].cr_event[0] != '\0') {
		strtoset_err(gettext("repeated 'pic%d' token\n"), picnum);
		return (-1);
	}

	if (val == NULL || val[0] == '\0') {
		strtoset_err(gettext("missing 'pic%d' value\n"), picnum);
		return (-1);
	}

	if (event_valid(picnum, val) == 0) {
		strtoset_err(gettext("pic%d cannot measure event '%s' on this "
		    "cpu\n"), picnum, val);
		return (-1);
	}

	nreqs++;
	(void) strncpy(reqs[picnum].cr_event, val, CPC_MAX_EVENT_LEN);
	return (0);
}

/*
 * We explicitly ignore any value provided for these tokens, as their
 * mere presence signals us to turn on or off the relevant flags.
 */
/*ARGSUSED*/
static int
flag(int tok, char *val)
{
	int i;
	int picnum = tok_info[tok].picnum;

	/*
	 * If picnum is -1, this flag should be applied to all reqs.
	 */
	for (i = (picnum == -1) ? 0 : picnum; i < ncounters; i++) {
		if (strcmp(tok_info[tok].name, "nouser") == 0)
			reqs[i].cr_flags &= ~CPC_COUNT_USER;
		else if (strcmp(tok_info[tok].name, "sys") == 0)
			reqs[i].cr_flags |= CPC_COUNT_SYSTEM;
		else
			return (-1);

		if (picnum != -1)
			break;
	}

	return (0);
}

static int
doattr(int tok, char *val)
{
	int		i;
	int		picnum = tok_info[tok].picnum;
	tmp_attr_t	*tmp;
	char		*endptr;

	/*
	 * If picnum is -1, this attribute should be applied to all reqs.
	 */
	for (i = (picnum == -1) ? 0 : picnum; i < ncounters; i++) {
		tmp = (tmp_attr_t *)emalloc(sizeof (tmp_attr_t));
		tmp->name = tok_info[tok].name;
		if (val != NULL) {
			tmp->val = strtoll(val, &endptr, 0);
			if (endptr == val) {
				strtoset_err(gettext("invalid value '%s' for "
				    "attribute '%s'\n"), val, tmp->name);
				free(tmp);
				return (-1);
			}
		} else
			/*
			 * No value was provided for this attribute,
			 * so specify a default value of 1.
			 */
			tmp->val = 1;

		tmp->next = attrs[i];
		attrs[i] = tmp;
		reqs[i].cr_nattrs++;

		if (picnum != -1)
			break;
	}

	return (0);
}

/*ARGSUSED*/
static void
attr_count_walker(void *arg, const char *attr)
{
	/*
	 * We don't allow picnum to be specified by the user.
	 */
	if (strncmp(attr, "picnum", 7) == 0)
		return;
	(*(int *)arg)++;
}

static int
cpc_count_attrs(cpc_t *cpc)
{
	int nattrs = 0;

	cpc_walk_attrs(cpc, &nattrs, attr_count_walker);

	return (nattrs);
}

static void
attr_walker(void *arg, const char *attr)
{
	int *i = arg;

	if (strncmp(attr, "picnum", 7) == 0)
		return;

	if ((attrlist[(*i)++] = strdup(attr)) == NULL) {
		strtoset_err(gettext("no memory available\n"));
		exit(0);
	}
}

cpc_set_t *
cpc_strtoset(cpc_t *cpcin, const char *spec, int smt)
{
	cpc_set_t		*set;
	cpc_attr_t		*req_attrs;
	tmp_attr_t		*tmp;
	size_t			toklen;
	int			i;
	int			j;
	int			x;
	char			*opts;
	char			*val;

	cpc = cpcin;
	nattrs = 0;

	ncounters = cpc_npic(cpc);

	reqs = (request_t *)emalloc(ncounters * sizeof (request_t));

	attrs = (tmp_attr_t **)emalloc(ncounters * sizeof (tmp_attr_t *));

	for (i = 0; i < ncounters; i++) {
		reqs[i].cr_event[0] = '\0';
		reqs[i].cr_flags = CPC_COUNT_USER;
		/*
		 * Each pic will have at least one attribute: the physical pic
		 * assignment via the "picnum" attribute. Set that up here for
		 * each request.
		 */
		reqs[i].cr_nattrs = 1;
		attrs[i] = emalloc(sizeof (tmp_attr_t));
		attrs[i]->name = "picnum";
		attrs[i]->val = i;
		attrs[i]->next = NULL;
	}

	/*
	 * Build up a list of acceptable tokens.
	 *
	 * Permitted tokens are
	 * picn=event
	 * nousern
	 * sysn
	 * attrn=val
	 * nouser
	 * sys
	 * attr=val
	 *
	 * Where n is a counter number, and attr is any attribute supported by
	 * the current processor.
	 *
	 * If a token appears without a counter number, it applies to all
	 * counters in the request set.
	 *
	 * The number of tokens is:
	 *
	 * picn: ncounters
	 * generic flags: 2 * ncounters (nouser, sys)
	 * attrs: nattrs * ncounters
	 * attrs with no picnum: nattrs
	 * generic flags with no picnum: 2 (nouser, sys)
	 * NULL token to signify end of list to getsubopt(3C).
	 *
	 * Matching each token's index in the token table is a function which
	 * process that token; these are in tok_funcs.
	 */

	/*
	 * Count the number of valid attributes.
	 * Set up the attrlist array to point to the attributes in attrlistp.
	 */
	nattrs = cpc_count_attrs(cpc);
	attrlist = (char **)emalloc(nattrs * sizeof (char *));

	i = 0;
	cpc_walk_attrs(cpc, &i, attr_walker);

	ntoks = ncounters + (2 * ncounters) + (nattrs * ncounters) + nattrs + 3;
	toks = (char **)emalloc(ntoks * sizeof (char *));
	tok_info = (tok_info_t *)emalloc(ntoks * sizeof (tok_info_t));

	tok_funcs = (int (**)(int, char *))emalloc(ntoks *
	    sizeof (int (*)(char *)));

	for (i = 0; i < ntoks; i++) {
		toks[i] = NULL;
		tok_funcs[i] = NULL;
	}

	x = 0;
	for (i = 0; i < ncounters; i++) {
		toks[x] = (char *)emalloc(TOK_SIZE);
		(void) snprintf(toks[x], TOK_SIZE, "pic%d", i);
		tok_info[x].name = "pic";
		tok_info[i].picnum = i;
		tok_funcs[x] = pic;
		x++;
	}

	for (i = 0; i < ncounters; i++) {
		toks[x] = (char *)emalloc(TOK_SIZE);
		(void) snprintf(toks[x], TOK_SIZE, "nouser%d", i);
		tok_info[x].name = "nouser";
		tok_info[x].picnum = i;
		tok_funcs[x] = flag;
		x++;
	}

	for (i = 0; i < ncounters; i++) {
		toks[x] = (char *)emalloc(TOK_SIZE);
		(void) snprintf(toks[x], TOK_SIZE, "sys%d", i);
		tok_info[x].name = "sys";
		tok_info[x].picnum = i;
		tok_funcs[x] = flag;
		x++;
	}
	for (j = 0; j < nattrs; j++) {
		toklen = strlen(attrlist[j]) + 3;
		for (i = 0; i < ncounters; i++) {
			toks[x] = (char *)emalloc(toklen);
			(void) snprintf(toks[x], toklen, "%s%d", attrlist[j],
			    i);
			tok_info[x].name = attrlist[j];
			tok_info[x].picnum = i;
			tok_funcs[x] = doattr;
			x++;
		}

		/*
		 * Now create a token for this attribute with no picnum; if used
		 * it will be applied to all reqs.
		 */
		toks[x] = (char *)emalloc(toklen);
		(void) snprintf(toks[x], toklen, "%s", attrlist[j]);
		tok_info[x].name = attrlist[j];
		tok_info[x].picnum = -1;
		tok_funcs[x] = doattr;
		x++;
	}

	toks[x] = "nouser";
	tok_info[x].name = "nouser";
	tok_info[x].picnum = -1;
	tok_funcs[x] = flag;
	x++;

	toks[x] = "sys";
	tok_info[x].name = "sys";
	tok_info[x].picnum = -1;
	tok_funcs[x] = flag;
	x++;

	toks[x] = NULL;

	opts = strdupa(spec);
	while (*opts != '\0') {
		int idx = getsubopt(&opts, toks, &val);

		if (idx == -1) {
			if (find_event(val) == 0) {
				strtoset_err(gettext("bad token '%s'\n"), val);
				goto inval;
			} else
				continue;
		}

		if (tok_funcs[idx](idx, val) == -1)
			goto inval;
	}

	/*
	 * The string has been processed. Now count how many PICs were used,
	 * create a request set, and specify each request properly.
	 */

	if ((set = cpc_set_create(cpc)) == NULL) {
		strtoset_err(gettext("no memory available\n"));
		exit(0);
	}

	for (i = 0; i < ncounters; i++) {
		if (reqs[i].cr_event[0] == '\0')
			continue;

		/*
		 * If the caller wishes to measure events on the physical CPU,
		 * we need to add SMT attributes to each request.
		 */
		if (smt)
			smt_special(i);

		req_attrs = (cpc_attr_t *)emalloc(reqs[i].cr_nattrs *
		    sizeof (cpc_attr_t));

		j = 0;
		for (tmp = attrs[i]; tmp != NULL; tmp = tmp->next) {
			req_attrs[j].ca_name = tmp->name;
			req_attrs[j].ca_val = tmp->val;
			j++;
		}

		if (cpc_set_add_request(cpc, set, reqs[i].cr_event, 0,
		    reqs[i].cr_flags, reqs[i].cr_nattrs, req_attrs) == -1) {
			free(req_attrs);
			(void) cpc_set_destroy(cpc, set);
			strtoset_err(
			    gettext("cpc_set_add_request() failed: %s\n"),
			    strerror(errno));
			goto inval;
		}

		free(req_attrs);
	}

	strtoset_cleanup();

	return (set);

inval:
	strtoset_cleanup();
	errno = EINVAL;
	return (NULL);
}

static void
strtoset_cleanup(void)
{
	int		i;
	tmp_attr_t	*tmp, *p;

	for (i = 0; i < nattrs; i++)
		free(attrlist[i]);
	free(attrlist);

	for (i = 0; i < ncounters; i++) {
		for (tmp = attrs[i]; tmp != NULL; tmp = p) {
			p = tmp->next;
			free(tmp);
		}
	}
	free(attrs);

	for (i = 0; i < ntoks - 3; i++)
		/*
		 * We free all but the last three tokens: "nouser", "sys", NULL
		 */
		free(toks[i]);
	free(toks);
	free(reqs);
	free(tok_info);
	free(tok_funcs);
}

/*
 * The following is called to modify requests so that they count events on
 * behalf of a physical processor, instead of a logical processor. It duplicates
 * the request flags for the sibling processor (i.e. if the request counts user
 * events, add an attribute to count user events on the sibling processor also).
 */
static void
smt_special(int picnum)
{
	tmp_attr_t *attr;

	if (reqs[picnum].cr_flags & CPC_COUNT_USER) {
		attr = (tmp_attr_t *)emalloc(sizeof (tmp_attr_t));
		attr->name = "count_sibling_usr";
		attr->val = 1;
		attr->next = attrs[picnum];
		attrs[picnum] = attr;
		reqs[picnum].cr_nattrs++;
	}

	if (reqs[picnum].cr_flags & CPC_COUNT_SYSTEM) {
		attr = (tmp_attr_t *)emalloc(sizeof (tmp_attr_t));
		attr->name = "count_sibling_sys";
		attr->val = 1;
		attr->next = attrs[picnum];
		attrs[picnum] = attr;
		reqs[picnum].cr_nattrs++;
	}
}

/*
 * If we ever fail to get memory, we print an error message and exit.
 */
static void *
emalloc(size_t n)
{
	void *p = malloc(n);

	if (p == NULL) {
		strtoset_err(gettext("no memory available\n"));
		exit(0);
	}

	return (p);
}
/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License, Version 1.0 only
 * (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 2004 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#include <sys/types.h>
#include <sys/time.h>
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <libcpc.h>

#include "cpucmds.h"

static hrtime_t timebase;

void
zerotime(void)
{
	timebase = gethrtime();
}

#define	NSECPERSEC	1000000000.0

float
mstimestamp(hrtime_t hrt)
{
	if (hrt == 0)
		hrt = gethrtime();
	return ((float)(hrt - timebase) / NSECPERSEC);
}