1 #include <AvailabilityMacros.h>
2 #ifdef AVAILABLE_MAC_OS_X_VERSION_10_5_AND_LATER
3 #include </System/Library/Frameworks/System.framework/PrivateHeaders/mach/thread_policy.h>
13 #include <mach/mach.h>
14 #include <mach/mach_error.h>
15 #include <mach/notify.h>
16 #include <servers/bootstrap.h>
17 #include <sys/event.h>
18 #include <sys/select.h>
19 #include <sys/types.h>
21 #include <sys/signal.h>
23 #define MAX(A, B) ((A) < (B) ? (B) : (A))
27 mach_msg_header_t header
;
28 mach_msg_trailer_t trailer
; // subtract this when sending
29 } ipc_trivial_message
;
32 mach_msg_header_t header
;
34 mach_msg_trailer_t trailer
; // subtract this when sending
38 mach_msg_header_t header
;
40 mach_msg_ool_descriptor_t descriptor
;
41 mach_msg_trailer_t trailer
; // subtract this when sending
42 } ipc_complex_message
;
53 mach_msg_header_t
*req_msg
;
55 mach_msg_header_t
*reply_msg
;
66 static boolean_t verbose
= FALSE
;
67 static boolean_t affinity
= FALSE
;
68 static boolean_t timeshare
= FALSE
;
69 static boolean_t threaded
= FALSE
;
70 static boolean_t oneway
= FALSE
;
71 static boolean_t do_select
= FALSE
;
80 char **server_port_name
;
82 void signal_handler(int sig
) {
85 void usage(const char *progname
) {
86 fprintf(stderr
, "usage: %s [options]\n", progname
);
87 fprintf(stderr
, "where options are:\n");
88 fprintf(stderr
, " -affinity\t\tthreads use affinity\n");
89 fprintf(stderr
, " -timeshare\t\tthreads use timeshare\n");
90 fprintf(stderr
, " -threaded\t\tuse (p)threads\n");
91 fprintf(stderr
, " -verbose\t\tbe verbose\n");
92 fprintf(stderr
, " -oneway\t\tdo not request return reply\n");
93 fprintf(stderr
, " -count num\t\tnumber of messages to send\n");
94 fprintf(stderr
, " -type trivial|inline|complex\ttype of messages to send\n");
95 fprintf(stderr
, " -numints num\tnumber of 32-bit ints to send in messages\n");
96 fprintf(stderr
, " -servers num\tnumber of servers threads to run\n");
97 fprintf(stderr
, " -clients num\tnumber of clients per server\n");
98 fprintf(stderr
, " -delay num\t\tmicroseconds to sleep clients between messages\n");
99 fprintf(stderr
, " -work num\t\tmicroseconds of client work\n");
100 fprintf(stderr
, " -pages num\t\tpages of memory touched by client work\n");
101 fprintf(stderr
, " -select \t\tselect prior to calling kevent().\n");
102 fprintf(stderr
, "default values are:\n");
103 fprintf(stderr
, " . no affinity\n");
104 fprintf(stderr
, " . not timeshare\n");
105 fprintf(stderr
, " . not verbose\n");
106 fprintf(stderr
, " . not oneway\n");
107 fprintf(stderr
, " . client sends 100000 messages\n");
108 fprintf(stderr
, " . inline message type\n");
109 fprintf(stderr
, " . 64 32-bit integers in inline/complex messages\n");
110 fprintf(stderr
, " . (num_available_processors+1)%%2 servers\n");
111 fprintf(stderr
, " . 4 clients per server\n");
112 fprintf(stderr
, " . no delay\n");
116 void parse_args(int argc
, char *argv
[]) {
117 host_basic_info_data_t info
;
118 mach_msg_type_number_t count
;
119 kern_return_t result
;
121 /* Initialize defaults */
122 msg_type
= msg_type_trivial
;
128 count
= HOST_BASIC_INFO_COUNT
;
129 result
= host_info(mach_host_self(), HOST_BASIC_INFO
,
130 (host_info_t
)&info
, &count
);
131 if (result
== KERN_SUCCESS
&& info
.avail_cpus
> 1)
132 num_servers
= info
.avail_cpus
/ 2;
136 const char *progname
= argv
[0];
139 if (0 == strcmp("-verbose", argv
[0])) {
142 } else if (0 == strcmp("-affinity", argv
[0])) {
145 } else if (0 == strcmp("-timeshare", argv
[0])) {
148 } else if (0 == strcmp("-threaded", argv
[0])) {
151 } else if (0 == strcmp("-oneway", argv
[0])) {
154 } else if (0 == strcmp("-type", argv
[0])) {
157 if (0 == strcmp("trivial", argv
[1])) {
158 msg_type
= msg_type_trivial
;
159 } else if (0 == strcmp("inline", argv
[1])) {
160 msg_type
= msg_type_inline
;
161 } else if (0 == strcmp("complex", argv
[1])) {
162 msg_type
= msg_type_complex
;
165 argc
-= 2; argv
+= 2;
166 } else if (0 == strcmp("-numints", argv
[0])) {
169 num_ints
= strtoul(argv
[1], NULL
, 0);
170 argc
-= 2; argv
+= 2;
171 } else if (0 == strcmp("-count", argv
[0])) {
174 num_msgs
= strtoul(argv
[1], NULL
, 0);
175 argc
-= 2; argv
+= 2;
176 } else if (0 == strcmp("-clients", argv
[0])) {
179 num_clients
= strtoul(argv
[1], NULL
, 0);
180 argc
-= 2; argv
+= 2;
181 } else if (0 == strcmp("-servers", argv
[0])) {
184 num_servers
= strtoul(argv
[1], NULL
, 0);
185 argc
-= 2; argv
+= 2;
186 } else if (0 == strcmp("-delay", argv
[0])) {
189 client_delay
= strtoul(argv
[1], NULL
, 0);
190 argc
-= 2; argv
+= 2;
191 } else if (0 == strcmp("-spin", argv
[0])) {
194 client_spin
= strtoul(argv
[1], NULL
, 0);
195 argc
-= 2; argv
+= 2;
196 } else if (0 == strcmp("-pages", argv
[0])) {
199 client_pages
= strtoul(argv
[1], NULL
, 0);
200 argc
-= 2; argv
+= 2;
201 } else if (0 == strcmp("-select", argv
[0])) {
209 void setup_server_ports(struct port_args
*ports
)
211 kern_return_t ret
= 0;
214 ports
->req_size
= MAX(sizeof(ipc_inline_message
) +
215 sizeof(u_int32_t
) * num_ints
,
216 sizeof(ipc_complex_message
));
217 ports
->reply_size
= sizeof(ipc_trivial_message
) -
218 sizeof(mach_msg_trailer_t
);
219 ports
->req_msg
= malloc(ports
->req_size
);
220 ports
->reply_msg
= malloc(ports
->reply_size
);
222 ret
= mach_port_allocate(mach_task_self(),
223 MACH_PORT_RIGHT_RECEIVE
,
225 if (KERN_SUCCESS
!= ret
) {
226 mach_error("mach_port_allocate(): ", ret
);
230 ret
= mach_port_allocate(mach_task_self(),
231 MACH_PORT_RIGHT_PORT_SET
,
233 if (KERN_SUCCESS
!= ret
) {
234 mach_error("mach_port_allocate(): ", ret
);
238 ret
= mach_port_insert_member(mach_task_self(),
241 if (KERN_SUCCESS
!= ret
) {
242 mach_error("mach_port_insert_member(): ", ret
);
246 ret
= mach_port_insert_right(mach_task_self(),
249 MACH_MSG_TYPE_MAKE_SEND
);
250 if (KERN_SUCCESS
!= ret
) {
251 mach_error("mach_port_insert_right(): ", ret
);
255 ret
= task_get_bootstrap_port(mach_task_self(), &bsport
);
256 if (KERN_SUCCESS
!= ret
) {
257 mach_error("task_get_bootstrap_port(): ", ret
);
262 printf("server waiting for IPC messages from client on port '%s'.\n",
263 server_port_name
[ports
->server_num
]);
265 ret
= bootstrap_register(bsport
,
266 server_port_name
[ports
->server_num
],
268 if (KERN_SUCCESS
!= ret
) {
269 mach_error("bootstrap_register(): ", ret
);
274 void setup_client_ports(struct port_args
*ports
)
276 kern_return_t ret
= 0;
278 case msg_type_trivial
:
279 ports
->req_size
= sizeof(ipc_trivial_message
);
281 case msg_type_inline
:
282 ports
->req_size
= sizeof(ipc_inline_message
) +
283 sizeof(u_int32_t
) * num_ints
;
285 case msg_type_complex
:
286 ports
->req_size
= sizeof(ipc_complex_message
);
289 ports
->req_size
-= sizeof(mach_msg_trailer_t
);
290 ports
->reply_size
= sizeof(ipc_trivial_message
);
291 ports
->req_msg
= malloc(ports
->req_size
);
292 ports
->reply_msg
= malloc(ports
->reply_size
);
294 ret
= mach_port_allocate(mach_task_self(),
295 MACH_PORT_RIGHT_RECEIVE
,
297 if (KERN_SUCCESS
!= ret
) {
298 mach_error("mach_port_allocate(): ", ret
);
302 printf("Client sending %d %s IPC messages to port '%s' in %s mode.\n",
303 num_msgs
, (msg_type
== msg_type_inline
) ?
304 "inline" : ((msg_type
== msg_type_complex
) ?
305 "complex" : "trivial"),
306 server_port_name
[ports
->server_num
],
307 (oneway
? "oneway" : "rpc"));
314 thread_setup(int tag
) {
315 #ifdef AVAILABLE_MAC_OS_X_VERSION_10_5_AND_LATER
317 thread_extended_policy_data_t epolicy
;
318 thread_affinity_policy_data_t policy
;
321 epolicy
.timeshare
= FALSE
;
322 ret
= thread_policy_set(
323 mach_thread_self(), THREAD_EXTENDED_POLICY
,
324 (thread_policy_t
) &epolicy
,
325 THREAD_EXTENDED_POLICY_COUNT
);
326 if (ret
!= KERN_SUCCESS
)
327 printf("thread_policy_set(THREAD_EXTENDED_POLICY) returned %d\n", ret
);
331 policy
.affinity_tag
= tag
;
332 ret
= thread_policy_set(
333 mach_thread_self(), THREAD_AFFINITY_POLICY
,
334 (thread_policy_t
) &policy
,
335 THREAD_AFFINITY_POLICY_COUNT
);
336 if (ret
!= KERN_SUCCESS
)
337 printf("thread_policy_set(THREAD_AFFINITY_POLICY) returned %d\n", ret
);
343 server(void *serverarg
)
346 struct kevent64_s kev
[1];
349 struct port_args args
;
352 int totalmsg
= num_msgs
* num_clients
;
355 args
.server_num
= (int) (long) serverarg
;
356 setup_server_ports(&args
);
358 thread_setup(args
.server_num
+ 1);
365 EV_SET64(&kev
[0], args
.pset
, EVFILT_MACHPORT
, (EV_ADD
| EV_CLEAR
| EV_DISPATCH
),
367 MACH_RCV_MSG
|MACH_RCV_LARGE
, 0, 0, (mach_vm_address_t
)args
.req_msg
, args
.req_size
);
371 err
= kevent64(kq
, kev
, 1, NULL
, 0, 0, NULL
);
377 for (idx
= 0; idx
< totalmsg
; idx
++) {
380 printf("server awaiting message %d\n", idx
);
384 FD_SET(kq
, &readfds
);
387 printf("Calling select() prior to kevent64().\n");
389 count
= select(kq
+ 1, &readfds
, NULL
, NULL
, NULL
);
396 EV_SET64(&kev
[0], args
.pset
, EVFILT_MACHPORT
, EV_ENABLE
,
398 MACH_RCV_MSG
|MACH_RCV_LARGE
, 0, 0, (mach_vm_address_t
)args
.req_msg
, args
.req_size
);
402 err
= kevent64(kq
, kev
, 1, kev
, 1, 0, NULL
);
408 // printf("kevent64: returned zero\n");
414 if (MACH_MSG_SUCCESS
!= ret
) {
416 printf("kevent64() mach_msg_return=%d", ret
);
417 mach_error("kevent64 (msg receive): ", ret
);
421 if (kev
[0].data
!= args
.port
)
422 printf("kevent64(MACH_PORT_NULL) port name (0x%x) != expected (0x%x)\n", kev
[0].data
, args
.port
);
424 args
.req_msg
->msgh_bits
= 0;
425 args
.req_msg
->msgh_size
= args
.req_size
;
426 args
.req_msg
->msgh_local_port
= args
.port
;
427 ret
= mach_msg(args
.req_msg
,
428 MACH_RCV_MSG
|MACH_RCV_INTERRUPT
|MACH_RCV_LARGE
,
432 MACH_MSG_TIMEOUT_NONE
,
434 if (MACH_RCV_INTERRUPTED
== ret
)
436 if (MACH_MSG_SUCCESS
!= ret
) {
438 printf("mach_msg() ret=%d", ret
);
439 mach_error("mach_msg (receive): ", ret
);
444 printf("server received message %d\n", idx
);
445 if (args
.req_msg
->msgh_bits
& MACH_MSGH_BITS_COMPLEX
) {
446 ret
= vm_deallocate(mach_task_self(),
447 (vm_address_t
)((ipc_complex_message
*)args
.req_msg
)->descriptor
.address
,
448 ((ipc_complex_message
*)args
.req_msg
)->descriptor
.size
);
451 if (1 == args
.req_msg
->msgh_id
) {
453 printf("server sending reply %d\n", idx
);
454 args
.reply_msg
->msgh_bits
= MACH_MSGH_BITS(MACH_MSG_TYPE_COPY_SEND
,
455 MACH_MSG_TYPE_MAKE_SEND
);
456 args
.reply_msg
->msgh_size
= args
.reply_size
;
457 args
.reply_msg
->msgh_remote_port
= args
.req_msg
->msgh_remote_port
;
458 args
.reply_msg
->msgh_local_port
= args
.req_msg
->msgh_local_port
;
459 args
.reply_msg
->msgh_id
= 2;
460 ret
= mach_msg(args
.reply_msg
,
465 MACH_MSG_TIMEOUT_NONE
,
467 if (MACH_MSG_SUCCESS
!= ret
) {
468 mach_error("mach_msg (send): ", ret
);
476 client_spin_loop(unsigned count
, void (fn
)(void))
482 static long dummy_memory
;
483 static long *client_memory
= &dummy_memory
;
485 client_work_atom(void)
489 if (++i
> client_pages
* PAGE_SIZE
/ sizeof(long))
491 client_memory
[i
] = 0;
494 static int calibration_count
= 10000;
495 static int calibration_usec
;
497 calibrate_client_work(void)
500 struct timeval nowtv
;
501 struct timeval warmuptv
= { 0, 100 * 1000 }; /* 100ms */
502 struct timeval starttv
;
503 struct timeval endtv
;
506 /* Warm-up the stepper first... */
507 gettimeofday(&nowtv
, NULL
);
508 timeradd(&nowtv
, &warmuptv
, &endtv
);
510 client_spin_loop(calibration_count
, client_work_atom
);
511 gettimeofday(&nowtv
, NULL
);
512 } while (timercmp(&nowtv
, &endtv
, < ));
514 /* Now do the calibration */
516 gettimeofday(&starttv
, NULL
);
517 client_spin_loop(calibration_count
, client_work_atom
);
518 gettimeofday(&endtv
, NULL
);
519 if (endtv
.tv_sec
- starttv
.tv_sec
> 1) {
520 calibration_count
/= 10;
523 calibration_usec
= endtv
.tv_usec
- starttv
.tv_usec
;
524 if (endtv
.tv_usec
< starttv
.tv_usec
) {
525 calibration_usec
+= 1000000;
527 if (calibration_usec
< 1000) {
528 calibration_count
*= 10;
531 calibration_count
/= calibration_usec
;
535 printf("calibration_count=%d calibration_usec=%d\n",
536 calibration_count
, calibration_usec
);
545 client_spin_loop(calibration_count
*client_spin
,
550 usleep(client_delay
);
554 void *client(void *threadarg
)
556 struct port_args args
;
558 mach_msg_header_t
*req
, *reply
;
559 mach_port_t bsport
, servport
;
561 long server_num
= (long) threadarg
;
562 void *ints
= malloc(sizeof(u_int32_t
) * num_ints
);
565 printf("client(%d) started, server port name %s\n",
566 server_num
, server_port_name
[server_num
]);
568 args
.server_num
= server_num
;
569 thread_setup(server_num
+ 1);
571 /* find server port */
572 ret
= task_get_bootstrap_port(mach_task_self(), &bsport
);
573 if (KERN_SUCCESS
!= ret
) {
574 mach_error("task_get_bootstrap_port(): ", ret
);
577 ret
= bootstrap_look_up(bsport
,
578 server_port_name
[server_num
],
580 if (KERN_SUCCESS
!= ret
) {
581 mach_error("bootstrap_look_up(): ", ret
);
585 setup_client_ports(&args
);
587 /* Allocate and touch memory */
590 client_memory
= (long *) malloc(client_pages
* PAGE_SIZE
);
591 for (i
= 0; i
< client_pages
; i
++)
592 client_memory
[i
* PAGE_SIZE
/ sizeof(long)] = 0;
595 /* start message loop */
596 for (idx
= 0; idx
< num_msgs
; idx
++) {
598 reply
= args
.reply_msg
;
600 req
->msgh_bits
= MACH_MSGH_BITS(MACH_MSG_TYPE_COPY_SEND
,
601 MACH_MSG_TYPE_MAKE_SEND
);
602 req
->msgh_size
= args
.req_size
;
603 req
->msgh_remote_port
= servport
;
604 req
->msgh_local_port
= args
.port
;
605 req
->msgh_id
= oneway
? 0 : 1;
606 if (msg_type
== msg_type_complex
) {
607 (req
)->msgh_bits
|= MACH_MSGH_BITS_COMPLEX
;
608 ((ipc_complex_message
*)req
)->body
.msgh_descriptor_count
= 1;
609 ((ipc_complex_message
*)req
)->descriptor
.address
= ints
;
610 ((ipc_complex_message
*)req
)->descriptor
.size
=
611 num_ints
* sizeof(u_int32_t
);
612 ((ipc_complex_message
*)req
)->descriptor
.deallocate
= FALSE
;
613 ((ipc_complex_message
*)req
)->descriptor
.copy
= MACH_MSG_VIRTUAL_COPY
;
614 ((ipc_complex_message
*)req
)->descriptor
.type
= MACH_MSG_OOL_DESCRIPTOR
;
617 printf("client sending message %d\n", idx
);
623 MACH_MSG_TIMEOUT_NONE
,
625 if (MACH_MSG_SUCCESS
!= ret
) {
626 mach_error("mach_msg (send): ", ret
);
627 fprintf(stderr
, "bailing after %u iterations\n", idx
);
633 printf("client awaiting reply %d\n", idx
);
634 reply
->msgh_bits
= 0;
635 reply
->msgh_size
= args
.reply_size
;
636 reply
->msgh_local_port
= args
.port
;
637 ret
= mach_msg(args
.reply_msg
,
638 MACH_RCV_MSG
|MACH_RCV_INTERRUPT
,
642 MACH_MSG_TIMEOUT_NONE
,
644 if (MACH_MSG_SUCCESS
!= ret
) {
645 mach_error("mach_msg (receive): ", ret
);
646 fprintf(stderr
, "bailing after %u iterations\n",
651 printf("client received reply %d\n", idx
);
662 thread_spawn(thread_id_t
*thread
, void *(fn
)(void *), void *arg
) {
665 ret
= pthread_create(
671 err(1, "pthread_create()");
673 printf("created pthread 0x%x\n", thread
->tid
);
675 thread
->pid
= fork();
676 if (thread
->pid
== 0) {
678 printf("calling 0x%x(0x%x)\n", fn
, arg
);
683 printf("forked pid %d\n", thread
->pid
);
688 thread_join(thread_id_t
*thread
) {
692 printf("joining thread 0x%x\n", thread
->tid
);
693 ret
= pthread_join(thread
->tid
, NULL
);
694 if (ret
!= KERN_SUCCESS
)
695 err(1, "pthread_join(0x%x)", thread
->tid
);
699 printf("waiting for pid %d\n", thread
->pid
);
700 waitpid(thread
->pid
, &stat
, 0);
705 wait_for_servers(void)
708 int retry_count
= 10;
709 mach_port_t bsport
, servport
;
712 /* find server port */
713 ret
= task_get_bootstrap_port(mach_task_self(), &bsport
);
714 if (KERN_SUCCESS
!= ret
) {
715 mach_error("task_get_bootstrap_port(): ", ret
);
719 while (retry_count
-- > 0) {
720 for (i
= 0; i
< num_servers
; i
++) {
721 ret
= bootstrap_look_up(bsport
,
724 if (ret
!= KERN_SUCCESS
) {
728 if (ret
== KERN_SUCCESS
)
730 usleep(100 * 1000); /* 100ms */
732 fprintf(stderr
, "Server(s) failed to register\n");
736 int main(int argc
, char *argv
[])
740 thread_id_t
*client_id
;
741 thread_id_t
*server_id
;
743 signal(SIGINT
, signal_handler
);
744 parse_args(argc
, argv
);
746 calibrate_client_work();
749 * If we're using affinity create an empty namespace now
750 * so this is shared by all our offspring.
755 server_id
= (thread_id_t
*) malloc(num_servers
* sizeof(thread_id_t
));
756 server_port_name
= (char **) malloc(num_servers
* sizeof(char *));
758 printf("creating %d servers\n", num_servers
);
759 for (i
= 0; i
< num_servers
; i
++) {
760 server_port_name
[i
] = (char *) malloc(sizeof("PORT.pppppp.xx"));
761 /* PORT names include pid of main process for disambiguation */
762 sprintf(server_port_name
[i
], "PORT.%06d.%02d", getpid(), i
);
763 thread_spawn(&server_id
[i
], server
, (void *) (long) i
);
766 int totalclients
= num_servers
* num_clients
;
767 int totalmsg
= num_msgs
* totalclients
;
768 struct timeval starttv
, endtv
, deltatv
;
771 * Wait for all servers to have registered all ports before starting
772 * the clients and the clock.
776 printf("%d server%s, %d client%s per server (%d total) %u messages...",
777 num_servers
, (num_servers
> 1)? "s" : "",
778 num_clients
, (num_clients
> 1)? "s" : "",
783 /* Call gettimeofday() once and throw away result; some implementations
784 * (like Mach's) cache some time zone info on first call.
786 gettimeofday(&starttv
, NULL
);
787 gettimeofday(&starttv
, NULL
);
789 client_id
= (thread_id_t
*) malloc(totalclients
* sizeof(thread_id_t
));
791 printf("creating %d clients\n", totalclients
);
792 for (i
= 0; i
< num_servers
; i
++) {
793 for (j
= 0; j
< num_clients
; j
++) {
795 &client_id
[(i
*num_clients
) + j
],
801 /* Wait for servers to complete */
802 for (i
= 0; i
< num_servers
; i
++) {
803 thread_join(&server_id
[i
]);
806 gettimeofday(&endtv
, NULL
);
808 for (i
= 0; i
< totalclients
; i
++) {
809 thread_join(&client_id
[i
]);
813 deltatv
.tv_sec
= endtv
.tv_sec
- starttv
.tv_sec
;
814 deltatv
.tv_usec
= endtv
.tv_usec
- starttv
.tv_usec
;
815 if (endtv
.tv_usec
< starttv
.tv_usec
) {
817 deltatv
.tv_usec
+= 1000000;
820 double dsecs
= (double) deltatv
.tv_sec
+
821 1.0E-6 * (double) deltatv
.tv_usec
;
823 printf(" in %u.%03u seconds\n",
824 deltatv
.tv_sec
, deltatv
.tv_usec
/1000);
825 printf(" throughput in messages/sec: %g\n",
826 (double)totalmsg
/ dsecs
);
827 printf(" average message latency (usec): %2.3g\n",
828 dsecs
* 1.0E6
/ (double) totalmsg
);