xnu-7195.101.1.tar.gz

[apple/xnu.git] / tools / tests / MPMMTest / MPMMtest.c

#include <AvailabilityMacros.h>
#include <mach/thread_policy.h>

#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <libgen.h>
#include <string.h>
#include <err.h>
#include <unistd.h>

#include <pthread.h>
#include <mach/mach.h>
#include <mach/mach_error.h>
#include <mach/mach_time.h>
#include <mach/notify.h>
#include <servers/bootstrap.h>
#include <sys/types.h>
#include <sys/time.h>
#include <sys/signal.h>
#include <errno.h>
#include "../unit_tests/tests_common.h" /* for record_perf_data() */

#include <libkern/OSAtomic.h>

#define MAX(A, B) ((A) < (B) ? (B) : (A))


typedef struct {
	mach_msg_header_t       header;
	mach_msg_trailer_t      trailer;                // subtract this when sending
} ipc_trivial_message;

typedef struct {
	mach_msg_header_t       header;
	u_int32_t               numbers[0];
	mach_msg_trailer_t      trailer;                // subtract this when sending
} ipc_inline_message;

typedef struct {
	mach_msg_header_t               header;
	mach_msg_body_t                 body;
	mach_msg_ool_descriptor_t       descriptor;
	mach_msg_trailer_t              trailer;        // subtract this when sending
} ipc_complex_message;

enum {
	msg_type_trivial = 0,
	msg_type_inline = 1,
	msg_type_complex = 2
};

struct port_args {
	int server_num;
	int req_size;
	mach_msg_header_t *req_msg;
	int reply_size;
	mach_msg_header_t *reply_msg;
	mach_port_t port;
	mach_port_t rcv_set;

	mach_port_t *set;
	mach_port_t *port_list;
};

typedef union {
	pid_t           pid;
	pthread_t       tid;
} thread_id_t;

/* Global options */
static int              verbose = 0;
static boolean_t        affinity = FALSE;
static boolean_t        timeshare = FALSE;
static boolean_t        threaded = FALSE;
static boolean_t        oneway = FALSE;
static boolean_t        useset = FALSE;
static boolean_t        save_perfdata = FALSE;
int                     msg_type;
int                     num_ints;
int                     num_msgs;
int                     num_clients;
int                     num_servers;
int                     client_delay;
int                     client_spin;
int                     client_pages;
int                     portcount = 1;
int                     setcount = 0;
boolean_t               stress_prepost = FALSE;
char                    **server_port_name;

struct port_args        *server_port_args;

/* global data */
mach_timebase_info_data_t g_timebase;
int64_t g_client_send_time = 0;

static inline uint64_t
ns_to_abs(uint64_t ns)
{
	return ns * g_timebase.denom / g_timebase.numer;
}

static inline uint64_t
abs_to_ns(uint64_t abs)
{
	return abs * g_timebase.numer / g_timebase.denom;
}


void
signal_handler(int sig)
{
}

void
usage(const char *progname)
{
	fprintf(stderr, "usage: %s [options]\n", progname);
	fprintf(stderr, "where options are:\n");
	fprintf(stderr, "    -affinity\t\tthreads use affinity\n");
	fprintf(stderr, "    -timeshare\t\tthreads use timeshare\n");
	fprintf(stderr, "    -threaded\t\tuse (p)threads\n");
	fprintf(stderr, "    -verbose\t\tbe verbose (use multiple times to increase verbosity)\n");
	fprintf(stderr, "    -oneway\t\tdo not request return reply\n");
	fprintf(stderr, "    -count num\t\tnumber of messages to send\n");
	fprintf(stderr, "    -perf   \t\tCreate perfdata files for metrics.\n");
	fprintf(stderr, "    -type trivial|inline|complex\ttype of messages to send\n");
	fprintf(stderr, "    -numints num\tnumber of 32-bit ints to send in messages\n");
	fprintf(stderr, "    -servers num\tnumber of server threads to run\n");
	fprintf(stderr, "    -clients num\tnumber of clients per server\n");
	fprintf(stderr, "    -delay num\t\tmicroseconds to sleep clients between messages\n");
	fprintf(stderr, "    -work num\t\tmicroseconds of client work\n");
	fprintf(stderr, "    -pages num\t\tpages of memory touched by client work\n");
	fprintf(stderr, "    -set nset num\tcreate [nset] portsets and [num] ports in each server.\n");
	fprintf(stderr, "                 \tEach port is connected to each set.\n");
	fprintf(stderr, "    -prepost\t\tstress the prepost system (implies -threaded, requires -set X Y)\n");
	fprintf(stderr, "default values are:\n");
	fprintf(stderr, "    . no affinity\n");
	fprintf(stderr, "    . not timeshare\n");
	fprintf(stderr, "    . not threaded\n");
	fprintf(stderr, "    . not verbose\n");
	fprintf(stderr, "    . not oneway\n");
	fprintf(stderr, "    . client sends 100000 messages\n");
	fprintf(stderr, "    . inline message type\n");
	fprintf(stderr, "    . 64 32-bit integers in inline/complex messages\n");
	fprintf(stderr, "    . (num_available_processors+1)%%2 servers\n");
	fprintf(stderr, "    . 4 clients per server\n");
	fprintf(stderr, "    . no delay\n");
	fprintf(stderr, "    . no sets / extra ports\n");
	fprintf(stderr, "    . no prepost stress\n");
	exit(1);
}

void
parse_args(int argc, char *argv[])
{
	host_basic_info_data_t          info;
	mach_msg_type_number_t          count;
	kern_return_t                   result;

	/* Initialize defaults */
	msg_type = msg_type_trivial;
	num_ints = 64;
	num_msgs = 100000;
	client_delay = 0;
	num_clients = 4;

	count = HOST_BASIC_INFO_COUNT;
	result = host_info(mach_host_self(), HOST_BASIC_INFO,
	    (host_info_t)&info, &count);
	if (result == KERN_SUCCESS && info.avail_cpus > 1) {
		num_servers = info.avail_cpus / 2;
	} else {
		num_servers = 1;
	}

	const char *progname = argv[0];
	argc--; argv++;
	while (0 < argc) {
		if (0 == strcmp("-verbose", argv[0])) {
			verbose++;
			argc--; argv++;
		} else if (0 == strcmp("-affinity", argv[0])) {
			affinity = TRUE;
			argc--; argv++;
		} else if (0 == strcmp("-timeshare", argv[0])) {
			timeshare = TRUE;
			argc--; argv++;
		} else if (0 == strcmp("-threaded", argv[0])) {
			threaded = TRUE;
			argc--; argv++;
		} else if (0 == strcmp("-oneway", argv[0])) {
			oneway = TRUE;
			argc--; argv++;
		} else if (0 == strcmp("-perf", argv[0])) {
			save_perfdata = TRUE;
			argc--; argv++;
		} else if (0 == strcmp("-type", argv[0])) {
			if (argc < 2) {
				usage(progname);
			}
			if (0 == strcmp("trivial", argv[1])) {
				msg_type = msg_type_trivial;
			} else if (0 == strcmp("inline", argv[1])) {
				msg_type = msg_type_inline;
			} else if (0 == strcmp("complex", argv[1])) {
				msg_type = msg_type_complex;
			} else {
				usage(progname);
			}
			argc -= 2; argv += 2;
		} else if (0 == strcmp("-numints", argv[0])) {
			if (argc < 2) {
				usage(progname);
			}
			num_ints = strtoul(argv[1], NULL, 0);
			argc -= 2; argv += 2;
		} else if (0 == strcmp("-count", argv[0])) {
			if (argc < 2) {
				usage(progname);
			}
			num_msgs = strtoul(argv[1], NULL, 0);
			argc -= 2; argv += 2;
		} else if (0 == strcmp("-clients", argv[0])) {
			if (argc < 2) {
				usage(progname);
			}
			num_clients = strtoul(argv[1], NULL, 0);
			argc -= 2; argv += 2;
		} else if (0 == strcmp("-servers", argv[0])) {
			if (argc < 2) {
				usage(progname);
			}
			num_servers = strtoul(argv[1], NULL, 0);
			argc -= 2; argv += 2;
		} else if (0 == strcmp("-delay", argv[0])) {
			if (argc < 2) {
				usage(progname);
			}
			client_delay = strtoul(argv[1], NULL, 0);
			argc -= 2; argv += 2;
		} else if (0 == strcmp("-spin", argv[0])) {
			if (argc < 2) {
				usage(progname);
			}
			client_spin = strtoul(argv[1], NULL, 0);
			argc -= 2; argv += 2;
		} else if (0 == strcmp("-pages", argv[0])) {
			if (argc < 2) {
				usage(progname);
			}
			client_pages = strtoul(argv[1], NULL, 0);
			argc -= 2; argv += 2;
		} else if (0 == strcmp("-set", argv[0])) {
			if (argc < 3) {
				usage(progname);
			}
			setcount = strtoul(argv[1], NULL, 0);
			portcount = strtoul(argv[2], NULL, 0);
			if (setcount <= 0 || portcount <= 0) {
				usage(progname);
			}
			useset = TRUE;
			argc -= 3; argv += 3;
		} else if (0 == strcmp("-prepost", argv[0])) {
			stress_prepost = TRUE;
			threaded = TRUE;
			argc--; argv++;
		} else {
			fprintf(stderr, "unknown option '%s'\n", argv[0]);
			usage(progname);
		}
	}

	if (stress_prepost) {
		if (!threaded) {
			fprintf(stderr, "Prepost stress test _must_ be threaded\n");
			exit(1);
		}
		if (portcount < 1 || setcount < 1) {
			fprintf(stderr, "Prepost stress test requires >= 1 port in >= 1 set.\n");
			exit(1);
		}
	}
}

void
setup_server_ports(struct port_args *ports)
{
	kern_return_t ret = 0;
	mach_port_t bsport;
	mach_port_t port;

	ports->req_size = MAX(sizeof(ipc_inline_message) +
	    sizeof(u_int32_t) * num_ints,
	    sizeof(ipc_complex_message));
	ports->reply_size = sizeof(ipc_trivial_message) -
	    sizeof(mach_msg_trailer_t);
	ports->req_msg = malloc(ports->req_size);
	ports->reply_msg = malloc(ports->reply_size);
	if (setcount > 0) {
		ports->set = (mach_port_t *)calloc(sizeof(mach_port_t), setcount);
		if (!ports->set) {
			fprintf(stderr, "calloc(%lu, %d) failed!\n", sizeof(mach_port_t), setcount);
			exit(1);
		}
	}
	if (stress_prepost) {
		ports->port_list = (mach_port_t *)calloc(sizeof(mach_port_t), portcount);
		if (!ports->port_list) {
			fprintf(stderr, "calloc(%lu, %d) failed!\n", sizeof(mach_port_t), portcount);
			exit(1);
		}
	}

	if (useset) {
		mach_port_t set;
		if (setcount < 1) {
			fprintf(stderr, "Can't use sets with a setcount of %d\n", setcount);
			exit(1);
		}

		for (int ns = 0; ns < setcount; ns++) {
			ret = mach_port_allocate(mach_task_self(),
			    MACH_PORT_RIGHT_PORT_SET,
			    &ports->set[ns]);
			if (KERN_SUCCESS != ret) {
				mach_error("mach_port_allocate(SET): ", ret);
				exit(1);
			}
			if (verbose > 1) {
				printf("SVR[%d] allocated set[%d] %#x\n",
				    ports->server_num, ns, ports->set[ns]);
			}

			set = ports->set[ns];
		}

		/* receive on a port set (always use the first in the chain) */
		ports->rcv_set = ports->set[0];
	}

	/* stuff the portset(s) with ports */
	for (int i = 0; i < portcount; i++) {
		ret = mach_port_allocate(mach_task_self(),
		    MACH_PORT_RIGHT_RECEIVE,
		    &port);
		if (KERN_SUCCESS != ret) {
			mach_error("mach_port_allocate(PORT): ", ret);
			exit(1);
		}

		if (stress_prepost) {
			ports->port_list[i] = port;
		}

		if (useset) {
			/* insert the port into _all_ allocated lowest-level sets */
			for (int ns = 0; ns < setcount; ns++) {
				if (verbose > 1) {
					printf("SVR[%d] moving port %#x into set %#x...\n",
					    ports->server_num, port, ports->set[ns]);
				}
				ret = mach_port_insert_member(mach_task_self(),
				    port, ports->set[ns]);
				if (KERN_SUCCESS != ret) {
					mach_error("mach_port_insert_member(): ", ret);
					exit(1);
				}
			}
		}
	}

	/* use the last one as the server's bootstrap port */
	ports->port = port;

	if (stress_prepost) {
		/* insert a send right for _each_ port */
		for (int i = 0; i < portcount; i++) {
			ret = mach_port_insert_right(mach_task_self(),
			    ports->port_list[i],
			    ports->port_list[i],
			    MACH_MSG_TYPE_MAKE_SEND);
			if (KERN_SUCCESS != ret) {
				mach_error("mach_port_insert_right(): ", ret);
				exit(1);
			}
		}
	} else {
		ret = mach_port_insert_right(mach_task_self(),
		    ports->port,
		    ports->port,
		    MACH_MSG_TYPE_MAKE_SEND);
		if (KERN_SUCCESS != ret) {
			mach_error("mach_port_insert_right(): ", ret);
			exit(1);
		}
	}

	ret = task_get_bootstrap_port(mach_task_self(), &bsport);
	if (KERN_SUCCESS != ret) {
		mach_error("task_get_bootstrap_port(): ", ret);
		exit(1);
	}

	if (verbose) {
		printf("server waiting for IPC messages from client on port '%s' (%#x).\n",
		    server_port_name[ports->server_num], ports->port);
	}
	ret = bootstrap_register(bsport,
	    server_port_name[ports->server_num],
	    ports->port);
	if (KERN_SUCCESS != ret) {
		mach_error("bootstrap_register(): ", ret);
		exit(1);
	}
}

void
setup_client_ports(struct port_args *ports)
{
	kern_return_t ret = 0;
	switch (msg_type) {
	case msg_type_trivial:
		ports->req_size = sizeof(ipc_trivial_message);
		break;
	case msg_type_inline:
		ports->req_size = sizeof(ipc_inline_message) +
		    sizeof(u_int32_t) * num_ints;
		break;
	case msg_type_complex:
		ports->req_size = sizeof(ipc_complex_message);
		break;
	}
	ports->req_size -= sizeof(mach_msg_trailer_t);
	ports->reply_size = sizeof(ipc_trivial_message);
	ports->req_msg = malloc(ports->req_size);
	ports->reply_msg = malloc(ports->reply_size);

	ret = mach_port_allocate(mach_task_self(),
	    MACH_PORT_RIGHT_RECEIVE,
	    &(ports->port));
	if (KERN_SUCCESS != ret) {
		mach_error("mach_port_allocate(): ", ret);
		exit(1);
	}
	if (verbose) {
		printf("Client sending %d %s IPC messages to port '%s' in %s mode\n",
		    num_msgs, (msg_type == msg_type_inline) ?
		    "inline" :  ((msg_type == msg_type_complex) ?
		    "complex" : "trivial"),
		    server_port_name[ports->server_num],
		    (oneway ? "oneway" : "rpc"));
	}
}


static void
thread_setup(int tag)
{
	kern_return_t                   ret;
	thread_extended_policy_data_t   epolicy;
	thread_affinity_policy_data_t   policy;

	if (!timeshare) {
		epolicy.timeshare = FALSE;
		ret = thread_policy_set(
			mach_thread_self(), THREAD_EXTENDED_POLICY,
			(thread_policy_t) &epolicy,
			THREAD_EXTENDED_POLICY_COUNT);
		if (ret != KERN_SUCCESS) {
			printf("thread_policy_set(THREAD_EXTENDED_POLICY) returned %d\n", ret);
		}
	}

	if (affinity) {
		policy.affinity_tag = tag;
		ret = thread_policy_set(
			mach_thread_self(), THREAD_AFFINITY_POLICY,
			(thread_policy_t) &policy,
			THREAD_AFFINITY_POLICY_COUNT);
		if (ret != KERN_SUCCESS) {
			printf("thread_policy_set(THREAD_AFFINITY_POLICY) returned %d\n", ret);
		}
	}
}

void *
server(void *serverarg)
{
	int idx;
	kern_return_t ret;
	int totalmsg = num_msgs * num_clients;
	mach_port_t recv_port;
	uint64_t starttm, endtm;

	int svr_num = (int)(uintptr_t)serverarg;
	struct port_args *args = &server_port_args[svr_num];

	args->server_num = svr_num;
	setup_server_ports(args);

	thread_setup(args->server_num + 1);

	recv_port = (useset) ? args->rcv_set : args->port;

	for (idx = 0; idx < totalmsg; idx++) {
		if (verbose > 2) {
			printf("server awaiting message %d\n", idx);
		}
		ret = mach_msg(args->req_msg,
		    MACH_RCV_MSG | MACH_RCV_INTERRUPT | MACH_RCV_LARGE,
		    0,
		    args->req_size,
		    recv_port,
		    MACH_MSG_TIMEOUT_NONE,
		    MACH_PORT_NULL);
		if (MACH_RCV_INTERRUPTED == ret) {
			break;
		}
		if (MACH_MSG_SUCCESS != ret) {
			if (verbose) {
				printf("mach_msg() ret=%d", ret);
			}
			mach_error("mach_msg (receive): ", ret);
			exit(1);
		}
		if (verbose > 2) {
			printf("server received message %d\n", idx);
		}
		if (args->req_msg->msgh_bits & MACH_MSGH_BITS_COMPLEX) {
			ret = vm_deallocate(mach_task_self(),
			    (vm_address_t)((ipc_complex_message *)args->req_msg)->descriptor.address,
			    ((ipc_complex_message *)args->req_msg)->descriptor.size);
		}

		if (1 == args->req_msg->msgh_id) {
			if (verbose > 2) {
				printf("server sending reply %d\n", idx);
			}
			args->reply_msg->msgh_bits = MACH_MSGH_BITS(MACH_MSG_TYPE_MOVE_SEND_ONCE, 0);
			args->reply_msg->msgh_size = args->reply_size;
			args->reply_msg->msgh_remote_port = args->req_msg->msgh_remote_port;
			args->reply_msg->msgh_local_port = MACH_PORT_NULL;
			args->reply_msg->msgh_id = 2;
			ret = mach_msg(args->reply_msg,
			    MACH_SEND_MSG,
			    args->reply_size,
			    0,
			    MACH_PORT_NULL,
			    MACH_MSG_TIMEOUT_NONE,
			    MACH_PORT_NULL);
			if (MACH_MSG_SUCCESS != ret) {
				mach_error("mach_msg (send): ", ret);
				exit(1);
			}
		}
	}

	if (!useset) {
		return NULL;
	}

	if (verbose < 1) {
		return NULL;
	}

	uint64_t deltans = 0;
	/*
	 * If we're using multiple sets, explicitly tear them all down
	 * and measure the time.
	 */
	for (int ns = 0; ns < setcount; ns++) {
		if (verbose > 1) {
			printf("\tTearing down set[%d] %#x...\n", ns, args->set[ns]);
		}
		starttm = mach_absolute_time();
		ret = mach_port_mod_refs(mach_task_self(), args->set[ns], MACH_PORT_RIGHT_PORT_SET, -1);
		endtm = mach_absolute_time();
		deltans += abs_to_ns(endtm - starttm);
		if (ret != KERN_SUCCESS) {
			mach_error("mach_port_mod_refs(): ", ret);
			exit(1);
		}
	}

	uint64_t nlinks = (uint64_t)setcount * (uint64_t)portcount;

	printf("\tteardown of %llu links took %llu ns\n", nlinks, deltans);
	printf("\t%lluns per set\n", deltans / (uint64_t)setcount);

	return NULL;
}

static inline void
client_spin_loop(unsigned count, void(fn)(void))
{
	while (count--) {
		fn();
	}
}

static long     dummy_memory;
static long     *client_memory = &dummy_memory;
static void
client_work_atom(void)
{
	static int      i;

	if (++i > client_pages * PAGE_SIZE / sizeof(long)) {
		i = 0;
	}
	client_memory[i] = 0;
}

static  int     calibration_count = 10000;
static  int     calibration_usec;
static void *
calibrate_client_work(void)
{
	long            dummy;
	struct timeval  nowtv;
	struct timeval  warmuptv = { 0, 100 * 1000 }; /* 100ms */
	struct timeval  starttv;
	struct timeval  endtv;

	if (client_spin) {
		/* Warm-up the stepper first... */
		gettimeofday(&nowtv, NULL);
		timeradd(&nowtv, &warmuptv, &endtv);
		do {
			client_spin_loop(calibration_count, client_work_atom);
			gettimeofday(&nowtv, NULL);
		} while (timercmp(&nowtv, &endtv, < ));

		/* Now do the calibration */
		while (TRUE) {
			gettimeofday(&starttv, NULL);
			client_spin_loop(calibration_count, client_work_atom);
			gettimeofday(&endtv, NULL);
			if (endtv.tv_sec - starttv.tv_sec > 1) {
				calibration_count /= 10;
				continue;
			}
			calibration_usec = endtv.tv_usec - starttv.tv_usec;
			if (endtv.tv_usec < starttv.tv_usec) {
				calibration_usec += 1000000;
			}
			if (calibration_usec < 1000) {
				calibration_count *= 10;
				continue;
			}
			calibration_count /= calibration_usec;
			break;
		}
		if (verbose > 1) {
			printf("calibration_count=%d calibration_usec=%d\n",
			    calibration_count, calibration_usec);
		}
	}
	return NULL;
}

static void *
client_work(void)
{
	if (client_spin) {
		client_spin_loop(calibration_count * client_spin,
		    client_work_atom);
	}

	if (client_delay) {
		usleep(client_delay);
	}
	return NULL;
}

void *
client(void *threadarg)
{
	struct port_args args;
	struct port_args *svr_args = NULL;
	int idx;
	mach_msg_header_t *req, *reply;
	mach_port_t bsport, servport;
	kern_return_t ret;
	int server_num = (int)(uintptr_t)threadarg;
	void *ints = malloc(sizeof(u_int32_t) * num_ints);

	if (verbose) {
		printf("client(%d) started, server port name %s\n",
		    server_num, server_port_name[server_num]);
	}

	args.server_num = server_num;
	thread_setup(server_num + 1);

	if (stress_prepost) {
		svr_args = &server_port_args[server_num];
	}

	/* find server port */
	ret = task_get_bootstrap_port(mach_task_self(), &bsport);
	if (KERN_SUCCESS != ret) {
		mach_error("task_get_bootstrap_port(): ", ret);
		exit(1);
	}
	ret = bootstrap_look_up(bsport,
	    server_port_name[server_num],
	    &servport);
	if (KERN_SUCCESS != ret) {
		mach_error("bootstrap_look_up(): ", ret);
		exit(1);
	}

	setup_client_ports(&args);

	/* Allocate and touch memory */
	if (client_pages) {
		unsigned        i;
		client_memory = (long *) malloc(client_pages * PAGE_SIZE);
		for (i = 0; i < client_pages; i++) {
			client_memory[i * PAGE_SIZE / sizeof(long)] = 0;
		}
	}

	uint64_t starttm, endtm;

	/* start message loop */
	for (idx = 0; idx < num_msgs; idx++) {
		req = args.req_msg;
		reply = args.reply_msg;

		req->msgh_size = args.req_size;
		if (stress_prepost) {
			req->msgh_remote_port = svr_args->port_list[idx % portcount];
		} else {
			req->msgh_remote_port = servport;
		}
		if (oneway) {
			req->msgh_bits = MACH_MSGH_BITS(MACH_MSG_TYPE_COPY_SEND, 0);
			req->msgh_local_port = MACH_PORT_NULL;
		} else {
			req->msgh_bits = MACH_MSGH_BITS(MACH_MSG_TYPE_COPY_SEND,
			    MACH_MSG_TYPE_MAKE_SEND_ONCE);
			req->msgh_local_port = args.port;
		}
		req->msgh_id = oneway ? 0 : 1;
		if (msg_type == msg_type_complex) {
			(req)->msgh_bits |=  MACH_MSGH_BITS_COMPLEX;
			((ipc_complex_message *)req)->body.msgh_descriptor_count = 1;
			((ipc_complex_message *)req)->descriptor.address = ints;
			((ipc_complex_message *)req)->descriptor.size =
			    num_ints * sizeof(u_int32_t);
			((ipc_complex_message *)req)->descriptor.deallocate = FALSE;
			((ipc_complex_message *)req)->descriptor.copy = MACH_MSG_VIRTUAL_COPY;
			((ipc_complex_message *)req)->descriptor.type = MACH_MSG_OOL_DESCRIPTOR;
		}
		if (verbose > 2) {
			printf("client sending message %d to port %#x\n",
			    idx, req->msgh_remote_port);
		}
		starttm = mach_absolute_time();
		ret = mach_msg(req,
		    MACH_SEND_MSG,
		    args.req_size,
		    0,
		    MACH_PORT_NULL,
		    MACH_MSG_TIMEOUT_NONE,
		    MACH_PORT_NULL);
		endtm = mach_absolute_time();
		if (MACH_MSG_SUCCESS != ret) {
			mach_error("mach_msg (send): ", ret);
			fprintf(stderr, "bailing after %u iterations\n", idx);
			exit(1);
			break;
		}
		if (stress_prepost) {
			OSAtomicAdd64(endtm - starttm, &g_client_send_time);
		}

		if (!oneway) {
			if (verbose > 2) {
				printf("client awaiting reply %d\n", idx);
			}
			reply->msgh_bits = 0;
			reply->msgh_size = args.reply_size;
			reply->msgh_local_port = args.port;
			ret = mach_msg(args.reply_msg,
			    MACH_RCV_MSG | MACH_RCV_INTERRUPT,
			    0,
			    args.reply_size,
			    args.port,
			    MACH_MSG_TIMEOUT_NONE,
			    MACH_PORT_NULL);
			if (MACH_MSG_SUCCESS != ret) {
				mach_error("mach_msg (receive): ", ret);
				fprintf(stderr, "bailing after %u iterations\n",
				    idx);
				exit(1);
			}
			if (verbose > 2) {
				printf("client received reply %d\n", idx);
			}
		}

		client_work();
	}

	free(ints);
	return NULL;
}

static void
thread_spawn(thread_id_t *thread, void *(fn)(void *), void *arg)
{
	if (threaded) {
		kern_return_t   ret;
		ret = pthread_create(
			&thread->tid,
			NULL,
			fn,
			arg);
		if (ret != 0) {
			err(1, "pthread_create()");
		}
		if (verbose > 1) {
			printf("created pthread %p\n", thread->tid);
		}
	} else {
		thread->pid = fork();
		if (thread->pid == 0) {
			if (verbose > 1) {
				printf("calling %p(%p)\n", fn, arg);
			}
			fn(arg);
			exit(0);
		}
		if (verbose > 1) {
			printf("forked pid %d\n", thread->pid);
		}
	}
}

static void
thread_join(thread_id_t *thread)
{
	if (threaded) {
		kern_return_t   ret;
		if (verbose > 1) {
			printf("joining thread %p\n", thread->tid);
		}
		ret = pthread_join(thread->tid, NULL);
		if (ret != KERN_SUCCESS) {
			err(1, "pthread_join(%p)", thread->tid);
		}
	} else {
		int     stat;
		if (verbose > 1) {
			printf("waiting for pid %d\n", thread->pid);
		}
		waitpid(thread->pid, &stat, 0);
	}
}

static void
wait_for_servers(void)
{
	int             i;
	int             retry_count = 10;
	mach_port_t     bsport, servport;
	kern_return_t   ret;

	/* find server port */
	ret = task_get_bootstrap_port(mach_task_self(), &bsport);
	if (KERN_SUCCESS != ret) {
		mach_error("task_get_bootstrap_port(): ", ret);
		exit(1);
	}

	while (retry_count-- > 0) {
		for (i = 0; i < num_servers; i++) {
			ret = bootstrap_look_up(bsport,
			    server_port_name[i],
			    &servport);
			if (ret != KERN_SUCCESS) {
				break;
			}
		}
		if (ret == KERN_SUCCESS) {
			return;
		}
		usleep(100 * 1000);     /* 100ms */
	}
	fprintf(stderr, "Server(s) failed to register\n");
	exit(1);
}

int
main(int argc, char *argv[])
{
	int             i;
	int             j;
	thread_id_t     *client_id;
	thread_id_t     *server_id;

	signal(SIGINT, signal_handler);
	parse_args(argc, argv);

	if (mach_timebase_info(&g_timebase) != KERN_SUCCESS) {
		fprintf(stderr, "Can't get mach_timebase_info!\n");
		exit(1);
	}

	calibrate_client_work();

	/*
	 * If we're using affinity create an empty namespace now
	 * so this is shared by all our offspring.
	 */
	if (affinity) {
		thread_setup(0);
	}

	server_id = (thread_id_t *) malloc(num_servers * sizeof(thread_id_t));
	server_port_name = (char **) malloc(num_servers * sizeof(char *));
	server_port_args = (struct port_args *)calloc(sizeof(struct port_args), num_servers);
	if (!server_id || !server_port_name || !server_port_args) {
		fprintf(stderr, "malloc/calloc of %d server book keeping structs failed\n", num_servers);
		exit(1);
	}

	if (verbose) {
		printf("creating %d servers\n", num_servers);
	}
	for (i = 0; i < num_servers; i++) {
		server_port_name[i] = (char *) malloc(sizeof("PORT.pppppp.xx"));
		/* PORT names include pid of main process for disambiguation */
		sprintf(server_port_name[i], "PORT.%06d.%02d", getpid(), i);
		thread_spawn(&server_id[i], server, (void *) (long) i);
	}

	int totalclients = num_servers * num_clients;
	int totalmsg = num_msgs * totalclients;
	struct timeval starttv, endtv, deltatv;

	/*
	 * Wait for all servers to have registered all ports before starting
	 * the clients and the clock.
	 */
	wait_for_servers();

	printf("%d server%s, %d client%s per server (%d total) %u messages...",
	    num_servers, (num_servers > 1)? "s" : "",
	    num_clients, (num_clients > 1)? "s" : "",
	    totalclients,
	    totalmsg);
	fflush(stdout);

	/* Call gettimeofday() once and throw away result; some implementations
	 * (like Mach's) cache some time zone info on first call.
	 */
	gettimeofday(&starttv, NULL);
	gettimeofday(&starttv, NULL);

	client_id = (thread_id_t *) malloc(totalclients * sizeof(thread_id_t));
	if (verbose) {
		printf("creating %d clients\n", totalclients);
	}
	for (i = 0; i < num_servers; i++) {
		for (j = 0; j < num_clients; j++) {
			thread_spawn(
				&client_id[(i * num_clients) + j],
				client,
				(void *) (long) i);
		}
	}

	/* Wait for servers to complete */
	for (i = 0; i < num_servers; i++) {
		thread_join(&server_id[i]);
	}

	gettimeofday(&endtv, NULL);
	if (verbose) {
		printf("all servers complete: waiting for clients...\n");
	}

	for (i = 0; i < totalclients; i++) {
		thread_join(&client_id[i]);
	}

	/* report results */
	deltatv.tv_sec = endtv.tv_sec - starttv.tv_sec;
	deltatv.tv_usec = endtv.tv_usec - starttv.tv_usec;
	if (endtv.tv_usec < starttv.tv_usec) {
		deltatv.tv_sec--;
		deltatv.tv_usec += 1000000;
	}

	double dsecs = (double) deltatv.tv_sec +
	    1.0E-6 * (double) deltatv.tv_usec;

	printf(" in %lu.%03u seconds\n",
	    deltatv.tv_sec, deltatv.tv_usec / 1000);
	printf("  throughput in messages/sec:     %g\n",
	    (double)totalmsg / dsecs);
	printf("  average message latency (usec): %2.3g\n",
	    dsecs * 1.0E6 / (double) totalmsg);

	double time_in_sec = (double)deltatv.tv_sec + (double)deltatv.tv_usec / 1000.0;
	double throughput_msg_p_sec = (double) totalmsg / dsecs;
	double avg_msg_latency = dsecs * 1.0E6 / (double)totalmsg;

	if (save_perfdata == TRUE) {
		char name[256];
		snprintf(name, sizeof(name), "%s_avg_msg_latency", basename(argv[0]));
		record_perf_data(name, "usec", avg_msg_latency, "Message latency measured in microseconds. Lower is better", stderr);
	}

	if (stress_prepost) {
		int64_t sendns = abs_to_ns(g_client_send_time);
		dsecs = (double)sendns / (double)NSEC_PER_SEC;
		printf("  total send time: %2.3gs\n", dsecs);
		printf("  average send time (usec): %2.3g\n",
		    dsecs * 1.0E6 / (double)totalmsg);
	}

	return 0;
}