1 /* The functions implementing the master-slave protocol of the
2 * distributed engine are grouped here. They are independent
3 * of the gtp protocol. See the comments at the top of distributed.c
4 * for a general introduction to the distributed engine. */
6 /* The receive queue is an array of pointers to binary buffers.
7 * These pointers are invalidated in one of two ways when a buffer
8 * is recycled: (1) the queue age is increased when the queue is
9 * emptied at a new move, (2) the pointer itself is set to NULL
10 * immmediately, and stays so until at least the next queue age
25 #include "distributed/distributed.h"
26 #include "distributed/protocol.h"
28 /* All gtp commands for current game separated by \n */
29 static char gtp_cmds
[CMDS_SIZE
];
31 /* Latest gtp command sent to slaves. */
32 static char *gtp_cmd
= NULL
;
34 /* Slaves send gtp_cmd when cmd_count changes. */
35 static int cmd_count
= 0;
37 /* Remember at most 10 gtp ids per move: kgs-rules, boardsize, clear_board,
38 * time_settings, komi, handicap, genmoves, play pass, play pass, final_status_list */
39 #define MAX_CMDS_PER_MOVE 10
41 /* History of gtp commands sent for current game, indexed by move. */
42 static struct cmd_history
{
45 } history
[MAX_GAMELEN
][MAX_CMDS_PER_MOVE
];
47 /* Number of active slave machines working for this master. */
48 int active_slaves
= 0;
50 /* Number of replies to last gtp command already received. */
53 /* All replies to latest gtp command are in gtp_replies[0..reply_count-1]. */
57 struct buf_state
**receive_queue
;
60 static int queue_max_length
;
62 /* Mutex protecting all variables above. receive_queue may be
63 * read without the lock but is only written with lock held. */
64 static pthread_mutex_t slave_lock
= PTHREAD_MUTEX_INITIALIZER
;
66 /* Condition signaled when a new gtp command is available. */
67 static pthread_cond_t cmd_cond
= PTHREAD_COND_INITIALIZER
;
69 /* Condition signaled when reply_count increases. */
70 static pthread_cond_t reply_cond
= PTHREAD_COND_INITIALIZER
;
72 /* Mutex protecting stderr. Must not be held at same time as slave_lock. */
73 static pthread_mutex_t log_lock
= PTHREAD_MUTEX_INITIALIZER
;
75 /* Absolute time when this program was started.
76 * For debugging only. */
77 static double start_time
;
79 /* Default slave state. */
80 struct slave_state default_sstate
;
83 /* Get exclusive access to the threads and commands state. */
87 pthread_mutex_lock(&slave_lock
);
90 /* Release exclusive access to the threads and commands state. */
94 pthread_mutex_unlock(&slave_lock
);
97 /* Write the time, client address, prefix, and string s to stderr atomically.
98 * s should end with a \n */
100 logline(struct in_addr
*client
, char *prefix
, char *s
)
102 double now
= time_now();
104 char addr
[INET_ADDRSTRLEN
];
106 inet_ntop(AF_INET
, client
, addr
, sizeof(addr
));
110 pthread_mutex_lock(&log_lock
);
111 fprintf(stderr
, "%s%15s %9.3f: %s", prefix
, addr
, now
- start_time
, s
);
112 pthread_mutex_unlock(&log_lock
);
115 /* Thread opening a connection on the given socket and copying input
116 * from there to stderr. */
118 proxy_thread(void *arg
)
120 int proxy_sock
= (long)arg
;
121 assert(proxy_sock
>= 0);
123 struct in_addr client
;
124 int conn
= open_server_connection(proxy_sock
, &client
);
125 FILE *f
= fdopen(conn
, "r");
127 while (fgets(buf
, BSIZE
, f
)) {
128 logline(&client
, "< ", buf
);
134 /* Get a reply to one gtp command. Return the gtp command id,
135 * or -1 if error. reply must have at least CMDS_SIZE bytes.
136 * The ascii reply ends with an empty line; if the first line
137 * contains "@size", a binary reply of size bytes follows the
138 * empty line. @size is not standard gtp, it is only used
139 * internally by Pachi for the genmoves command; it must be the
140 * last parameter on the line.
141 * *bin_size is the maximum size upon entry, actual size on return.
142 * slave_lock is not held on either entry or exit of this function. */
144 get_reply(FILE *f
, struct in_addr client
, char *reply
, void *bin_reply
, int *bin_size
)
148 if (!fgets(reply
, CMDS_SIZE
, f
)) return -1;
150 /* Check for binary reply. */
151 char *s
= strchr(reply
, '@');
153 if (s
) size
= atoi(s
+1);
154 assert(size
<= *bin_size
);
158 logline(&client
, "<<", reply
);
159 if ((*reply
== '=' || *reply
== '?') && isdigit(reply
[1]))
160 reply_id
= atoi(reply
+1);
162 /* Read the rest of the ascii reply */
163 char *line
= reply
+ strlen(reply
);
164 while (fgets(line
, reply
+ CMDS_SIZE
- line
, f
) && *line
!= '\n') {
166 logline(&client
, "<<", line
);
167 line
+= strlen(line
);
169 if (*line
!= '\n') return -1;
171 /* Read the binary reply if any. */
172 double start
= time_now();
174 while (size
&& (len
= fread(bin_reply
, 1, size
, f
)) > 0) {
175 bin_reply
= (char *)bin_reply
+ len
;
178 if (*bin_size
&& DEBUGVV(2)) {
180 snprintf(buf
, sizeof(buf
), "read reply %d bytes in %.4fms\n", *bin_size
,
181 (time_now() - start
)*1000);
182 logline(&client
, "= ", buf
);
184 return size
? -1 : reply_id
;
187 /* Send the gtp command to_send and get a reply from the slave machine.
188 * Write the reply in buf which must have at least CMDS_SIZE bytes.
189 * If *bin_size > 0, send bin_buf after the gtp command.
190 * Return any binary reply in bin_buf and set its size in bin_size.
191 * bin_buf is private to the slave and need not be copied.
192 * Return the gtp command id, or -1 if error.
193 * slave_lock is held on both entry and exit of this function. */
195 send_command(char *to_send
, void *bin_buf
, int *bin_size
,
196 FILE *f
, struct slave_state
*sstate
, char *buf
)
198 assert(to_send
&& gtp_cmd
&& bin_buf
&& bin_size
);
199 strncpy(buf
, to_send
, CMDS_SIZE
);
200 bool resend
= to_send
!= gtp_cmd
;
202 pthread_mutex_unlock(&slave_lock
);
204 if (DEBUGL(1) && resend
)
205 logline(&sstate
->client
, "? ",
206 to_send
== gtp_cmds
? "resend all\n" : "partial resend\n");
209 double start
= time_now();
211 fwrite(bin_buf
, 1, *bin_size
, f
);
214 if (DEBUGV(strchr(buf
, '@'), 2)) {
215 double ms
= (time_now() - start
) * 1000.0;
217 char *s
= strchr(buf
, '\n');
220 logline(&sstate
->client
, ">>", buf
);
223 snprintf(b
, sizeof(b
),
224 "sent args %d bytes in %.4fms\n", *bin_size
, ms
);
225 logline(&sstate
->client
, "= ", b
);
229 /* Reuse the buffers for the reply. */
230 *bin_size
= sstate
->max_buf_size
;
231 int reply_id
= get_reply(f
, sstate
->client
, buf
, bin_buf
, bin_size
);
233 pthread_mutex_lock(&slave_lock
);
237 /* Return the command sent after that with the given gtp id,
238 * or gtp_cmds if the id wasn't used in this game. If a play command
239 * has overwritten a genmoves command, return the play command.
240 * slave_lock is held on both entry and exit of this function. */
242 next_command(int cmd_id
)
244 if (cmd_id
== -1) return gtp_cmds
;
246 int last_id
= atoi(gtp_cmd
);
247 int reply_move
= move_number(cmd_id
);
248 if (reply_move
> move_number(last_id
)) return gtp_cmds
;
251 for (slot
= 0; slot
< MAX_CMDS_PER_MOVE
; slot
++) {
252 if (cmd_id
== history
[reply_move
][slot
].gtp_id
) break;
254 if (slot
== MAX_CMDS_PER_MOVE
) return gtp_cmds
;
256 char *next
= history
[reply_move
][slot
].next_cmd
;
261 /* Allocate buffers for a slave thread. The state should have been
262 * initialized already as a copy of the default slave state.
263 * slave_lock is not held on either entry or exit of this function. */
265 slave_state_alloc(struct slave_state
*sstate
)
267 for (int n
= 0; n
< BUFFERS_PER_SLAVE
; n
++) {
268 sstate
->b
[n
].buf
= malloc2(sstate
->max_buf_size
);
269 sstate
->b
[n
].owner
= sstate
->thread_id
;
271 if (sstate
->alloc_hook
) sstate
->alloc_hook(sstate
);
274 /* Get a free binary buffer, first invalidating it in the receive
275 * queue if necessary. In practice all buffers should be used
276 * before they are invalidated, if BUFFERS_PER_SLAVE is large enough.
277 * slave_lock is held on both entry and exit of this function. */
279 get_free_buf(struct slave_state
*sstate
)
281 int newest
= (sstate
->newest_buf
+ 1) & (BUFFERS_PER_SLAVE
- 1);
282 sstate
->newest_buf
= newest
;
283 void *buf
= sstate
->b
[newest
].buf
;
287 snprintf(b
, sizeof(b
),
288 "get free %d index %d buf=%p age %d qlength %d\n", newest
,
289 sstate
->b
[newest
].queue_index
, buf
, queue_age
, queue_length
);
290 logline(&sstate
->client
, "? ", b
);
293 int index
= sstate
->b
[newest
].queue_index
;
294 if (index
< 0) return buf
;
296 /* Invalidate the buffer if the calling thread still owns its previous
297 * entry in the receive queue. The entry may have been overwritten by
298 * another thread, but only after a new move which invalidates the
299 * entire receive queue. */
300 if (receive_queue
[index
] && receive_queue
[index
]->owner
== sstate
->thread_id
) {
301 receive_queue
[index
] = NULL
;
303 sstate
->b
[newest
].queue_index
= -1;
307 /* Insert a buffer in the receive queue. It should be the most
308 * recent buffer allocated by the calling thread.
309 * slave_lock is held on both entry and exit of this function. */
311 insert_buf(struct slave_state
*sstate
, void *buf
, int size
)
313 assert(queue_length
< queue_max_length
);
315 int newest
= sstate
->newest_buf
;
316 assert(buf
== sstate
->b
[newest
].buf
);
318 /* Update the buffer if necessary before making it
319 * available to other threads. */
320 if (sstate
->insert_hook
) sstate
->insert_hook(buf
, size
);
324 snprintf(b
, sizeof(b
),
325 "insert newest %d age %d rq[%d]->%p owner %d\n",
326 newest
, queue_age
, queue_length
, buf
, sstate
->thread_id
);
327 logline(&sstate
->client
, "? ", b
);
329 receive_queue
[queue_length
] = &sstate
->b
[newest
];
330 receive_queue
[queue_length
]->size
= size
;
331 receive_queue
[queue_length
]->queue_index
= queue_length
;
335 /* Clear the receive queue. The buffer pointers do not have to be cleared
336 * here, this is done as each buffer is recycled.
337 * slave_lock is held on both entry and exit of this function. */
339 clear_receive_queue(void)
343 snprintf(buf
, sizeof(buf
), "clear queue, old length %d age %d\n",
344 queue_length
, queue_age
);
345 logline(NULL
, "? ", buf
);
351 /* Process the reply received from a slave machine.
352 * Copy the ascii part to reply_buf and insert the binary part
353 * (if any) in the receive queue.
354 * Return false if ok, true if the slave is out of sync.
355 * slave_lock is held on both entry and exit of this function. */
357 process_reply(int reply_id
, char *reply
, char *reply_buf
,
358 void *bin_reply
, int bin_size
, int *last_reply_id
,
359 int *reply_slot
, struct slave_state
*sstate
)
361 /* Resend everything if slave returned an error. */
366 /* Make sure we are still in sync. cmd_count may have
367 * changed but the reply is valid as long as cmd_id didn't
368 * change (this only occurs for consecutive genmoves). */
369 int cmd_id
= atoi(gtp_cmd
);
370 if (reply_id
!= cmd_id
) {
371 *last_reply_id
= reply_id
;
375 strncpy(reply_buf
, reply
, CMDS_SIZE
);
376 if (reply_id
!= *last_reply_id
)
377 *reply_slot
= reply_count
++;
378 gtp_replies
[*reply_slot
] = reply_buf
;
380 if (bin_size
) insert_buf(sstate
, bin_reply
, bin_size
);
382 pthread_cond_signal(&reply_cond
);
383 *last_reply_id
= reply_id
;
387 /* Get the binary arg for the given command, and update the command
388 * if necessary. For now, only genmoves has a binary argument, and
389 * we return the best stats increments from all other slaves.
390 * Set *bin_size to 0 if the command doesn't take binary arguments,
391 * but still return a buffer, to be used for the reply.
392 * Return NULL if the binary arg is obsolete by the time we have
393 * finished computing it, because a new command is available.
394 * This version only gets the buffer for the reply, to be completed
396 * slave_lock is held on both entry and exit of this function. */
398 get_binary_arg(struct slave_state
*sstate
, char *cmd
, int cmd_size
, int *bin_size
)
400 int cmd_id
= atoi(gtp_cmd
);
401 void *buf
= get_free_buf(sstate
);
404 char *s
= strchr(cmd
, '@');
405 if (!s
|| !sstate
->args_hook
) return buf
;
407 int size
= sstate
->args_hook(buf
, sstate
, cmd_id
);
409 /* Check that the command is still valid. */
410 if (atoi(gtp_cmd
) != cmd_id
) return NULL
;
412 /* Set the correct binary size for this slave.
413 * cmd may have been overwritten with new parameters. */
415 s
= strchr(cmd
, '@');
417 snprintf(s
, cmd
+ cmd_size
- s
, "@%d\n", size
);
421 /* Main loop of a slave thread.
422 * Send the current command to the slave machine and wait for a reply.
423 * Resend command history if the slave machine is out of sync.
424 * Returns when the connection with the slave machine is cut.
425 * slave_lock is held on both entry and exit of this function. */
427 slave_loop(FILE *f
, char *reply_buf
, struct slave_state
*sstate
, bool resend
)
430 int last_cmd_count
= 0;
431 int last_reply_id
= -1;
435 /* Resend complete or partial history */
436 to_send
= next_command(last_reply_id
);
438 /* Wait for a new command. */
439 while (last_cmd_count
== cmd_count
)
440 pthread_cond_wait(&cmd_cond
, &slave_lock
);
444 /* Command available, send it to slave machine.
445 * If slave was out of sync, send the history.
446 * But first get binary arguments if necessary. */
448 void *bin_buf
= get_binary_arg(sstate
, gtp_cmd
,
449 gtp_cmds
+ CMDS_SIZE
- gtp_cmd
,
451 /* Check that the command is still valid. */
453 if (!bin_buf
) continue;
455 /* Send the command and get the reply, which always ends with \n\n
456 * The slave machine sends "=id reply" or "?id reply"
457 * with id == cmd_id if it is in sync. */
458 last_cmd_count
= cmd_count
;
460 int reply_id
= send_command(to_send
, bin_buf
, &bin_size
, f
,
462 if (reply_id
== -1) return;
464 resend
= process_reply(reply_id
, buf
, reply_buf
, bin_buf
, bin_size
,
465 &last_reply_id
, &reply_slot
, sstate
);
469 /* Thread sending gtp commands to one slave machine, and
470 * reading replies. If a slave machine dies, this thread waits
471 * for a connection from another slave.
472 * The large buffers are allocated only once we get a first
473 * connection, to avoid wasting memory if max_slaves is too large.
474 * We do not invalidate the received buffers if a slave disconnects;
475 * they are still useful for other slaves. */
477 slave_thread(void *arg
)
479 struct slave_state sstate
= default_sstate
;
480 sstate
.thread_id
= (long)arg
;
482 assert(sstate
.slave_sock
>= 0);
483 char reply_buf
[CMDS_SIZE
];
487 /* Wait for a connection from any slave. */
488 struct in_addr client
;
489 int conn
= open_server_connection(sstate
.slave_sock
, &client
);
491 FILE *f
= fdopen(conn
, "r+");
493 snprintf(reply_buf
, sizeof(reply_buf
),
494 "new slave, id %d\n", sstate
.thread_id
);
495 logline(&client
, "= ", reply_buf
);
498 /* Minimal check of the slave identity. */
500 if (!fgets(reply_buf
, sizeof(reply_buf
), f
)
501 || strncasecmp(reply_buf
, "= Pachi", 7)
502 || !fgets(reply_buf
, sizeof(reply_buf
), f
)
503 || strcmp(reply_buf
, "\n")) {
504 logline(&client
, "? ", "bad slave\n");
509 if (!resend
) slave_state_alloc(&sstate
);
510 sstate
.client
= client
;
512 pthread_mutex_lock(&slave_lock
);
514 slave_loop(f
, reply_buf
, &sstate
, resend
);
516 assert(active_slaves
> 0);
518 // Unblock main thread if it was waiting for this slave.
519 pthread_cond_signal(&reply_cond
);
520 pthread_mutex_unlock(&slave_lock
);
524 logline(&client
, "= ", "lost slave\n");
529 /* Create a new gtp command for all slaves. The slave lock is held
530 * upon entry and upon return, so the command will actually be
531 * sent when the lock is released. The last command is overwritten
532 * if gtp_cmd points to a non-empty string. cmd is a single word;
533 * args has all arguments and is empty or has a trailing \n */
535 update_cmd(struct board
*b
, char *cmd
, char *args
, bool new_id
)
538 /* To make sure the slaves are in sync, we ignore the original id
539 * and use the board number plus some random bits as gtp id. */
540 static int gtp_id
= -1;
541 int moves
= is_reset(cmd
) ? 0 : b
->moves
;
543 int prev_id
= gtp_id
;
545 /* fast_random() is 16-bit only so the multiplication can't overflow. */
546 gtp_id
= force_reply(moves
+ fast_random(65535) * DIST_GAMELEN
);
547 } while (gtp_id
== prev_id
);
550 snprintf(gtp_cmd
, gtp_cmds
+ CMDS_SIZE
- gtp_cmd
, "%d %s %s",
551 gtp_id
, cmd
, *args
? args
: "\n");
554 /* Remember history for out-of-sync slaves. */
556 static struct cmd_history
*last
= NULL
;
558 if (last
) last
->next_cmd
= gtp_cmd
;
559 slot
= (slot
+ 1) % MAX_CMDS_PER_MOVE
;
560 last
= &history
[moves
][slot
];
561 last
->gtp_id
= gtp_id
;
562 last
->next_cmd
= NULL
;
564 // Notify the slave threads about the new command.
565 pthread_cond_broadcast(&cmd_cond
);
568 /* Update the command history, then create a new gtp command
569 * for all slaves. The slave lock is held upon entry and
570 * upon return, so the command will actually be sent when the
571 * lock is released. cmd is a single word; args has all
572 * arguments and is empty or has a trailing \n */
574 new_cmd(struct board
*b
, char *cmd
, char *args
)
576 // Clear the history when a new game starts:
577 if (!gtp_cmd
|| is_gamestart(cmd
)) {
579 memset(history
, 0, sizeof(history
));
581 /* Preserve command history for new slaves.
582 * To indicate that the slave should only reply to
583 * the last command we force the id of previous
584 * commands to be just the move number. */
585 int id
= prevent_reply(atoi(gtp_cmd
));
586 int len
= strspn(gtp_cmd
, "0123456789");
588 snprintf(buf
, sizeof(buf
), "%0*d", len
, id
);
589 memcpy(gtp_cmd
, buf
, len
);
591 gtp_cmd
+= strlen(gtp_cmd
);
594 // Let the slave threads send the new gtp command:
595 update_cmd(b
, cmd
, args
, true);
598 /* Wait for at least one new reply. Return when at least
599 * min_replies slaves have already replied, or when the
600 * given absolute time is passed.
601 * The replies are returned in gtp_replies[0..reply_count-1]
602 * slave_lock is held on entry and on return. */
604 get_replies(double time_limit
, int min_replies
)
607 if (reply_count
> 0) {
610 ts
.tv_nsec
= (int)(modf(time_limit
, &sec
)*1000000000.0);
611 ts
.tv_sec
= (int)sec
;
612 pthread_cond_timedwait(&reply_cond
, &slave_lock
, &ts
);
614 pthread_cond_wait(&reply_cond
, &slave_lock
);
616 if (reply_count
== 0) continue;
617 if (reply_count
>= min_replies
|| reply_count
>= active_slaves
) return;
618 if (time_now() >= time_limit
) break;
622 snprintf(buf
, sizeof(buf
),
623 "get_replies timeout %.3f >= %.3f, replies %d < min %d, active %d\n",
624 time_now() - start_time
, time_limit
- start_time
,
625 reply_count
, min_replies
, active_slaves
);
626 logline(NULL
, "? ", buf
);
628 assert(reply_count
> 0);
631 /* In a 30s move with at least 5ms per genmoves we get at most
632 * 6000 genmoves per slave. */
633 #define MAX_GENMOVES_PER_SLAVE 6000
635 /* Allocate the receive queue, and create the slave and proxy threads.
636 * max_buf_size and the merge-related fields of default_sstate must
637 * already be initialized. */
639 protocol_init(char *slave_port
, char *proxy_port
, int max_slaves
)
641 start_time
= time_now();
643 queue_max_length
= max_slaves
* MAX_GENMOVES_PER_SLAVE
;
644 receive_queue
= calloc2(queue_max_length
, sizeof(*receive_queue
));
646 default_sstate
.slave_sock
= port_listen(slave_port
, max_slaves
);
647 default_sstate
.last_processed
= -1;
649 for (int n
= 0; n
< BUFFERS_PER_SLAVE
; n
++) {
650 default_sstate
.b
[n
].queue_index
= -1;
654 for (int id
= 0; id
< max_slaves
; id
++) {
655 pthread_create(&thread
, NULL
, slave_thread
, (void *)(long)id
);
659 int proxy_sock
= port_listen(proxy_port
, max_slaves
);
660 for (int id
= 0; id
< max_slaves
; id
++) {
661 pthread_create(&thread
, NULL
, proxy_thread
, (void *)(long)proxy_sock
);