Fix possible bug on threading platforms with running out of fds do to cpuworkers...
[tor.git] / src / or / cpuworker.c
blob155a875fa6b28f03bd3ea108d65189f1f6303261
1 /* Copyright 2003-2004 Roger Dingledine.
2 * Copyright 2004-2005 Roger Dingledine, Nick Mathewson. */
3 /* See LICENSE for licensing information */
4 /* $Id$ */
5 const char cpuworker_c_id[] = "$Id$";
7 /**
8 * \file cpuworker.c
9 * \brief Run computation-intensive tasks (generally for crypto) in
10 * a separate execution context. [OR only.]
12 * Right now, we only use this for processing onionskins.
13 **/
15 #include "or.h"
17 /** The maximum number of cpuworker processes we will keep around. */
18 #define MAX_CPUWORKERS 16
19 /** The minimum number of cpuworker processes we will keep around. */
20 #define MIN_CPUWORKERS 1
22 /** The tag specifies which circuit this onionskin was from. */
23 #define TAG_LEN 8
24 /** How many bytes are sent from tor to the cpuworker? */
25 #define LEN_ONION_QUESTION (1+TAG_LEN+ONIONSKIN_CHALLENGE_LEN)
26 /** How many bytes are sent from the cpuworker back to tor? */
27 #define LEN_ONION_RESPONSE (1+TAG_LEN+ONIONSKIN_REPLY_LEN+40+32)
29 /** How many cpuworkers we have running right now. */
30 static int num_cpuworkers=0;
31 /** How many of the running cpuworkers have an assigned task right now. */
32 static int num_cpuworkers_busy=0;
33 /** We need to spawn new cpuworkers whenever we rotate the onion keys
34 * on platforms where execution contexts==processes. This variable stores
35 * the last time we got a key rotation event. */
36 static time_t last_rotation_time=0;
38 static int cpuworker_main(void *data);
39 static int spawn_cpuworker(void);
40 static void spawn_enough_cpuworkers(void);
41 static void process_pending_task(connection_t *cpuworker);
43 /** Initialize the cpuworker subsystem.
45 void cpu_init(void) {
46 last_rotation_time=time(NULL);
47 spawn_enough_cpuworkers();
50 /** Called when we're done sending a request to a cpuworker. */
51 int connection_cpu_finished_flushing(connection_t *conn) {
52 tor_assert(conn);
53 tor_assert(conn->type == CONN_TYPE_CPUWORKER);
54 connection_stop_writing(conn);
55 return 0;
58 /** Pack addr,port,and circ_id; set *tag to the result. (See note on
59 * cpuworker_main for wire format.) */
60 static void tag_pack(char *tag, uint32_t addr, uint16_t port, uint16_t circ_id) {
61 *(uint32_t *)tag = addr;
62 *(uint16_t *)(tag+4) = port;
63 *(uint16_t *)(tag+6) = circ_id;
66 /** Unpack <b>tag</b> into addr, port, and circ_id.
68 static void tag_unpack(const char *tag, uint32_t *addr, uint16_t *port, uint16_t *circ_id) {
69 struct in_addr in;
70 char addrbuf[INET_NTOA_BUF_LEN];
72 *addr = *(const uint32_t *)tag;
73 *port = *(const uint16_t *)(tag+4);
74 *circ_id = *(const uint16_t *)(tag+6);
76 in.s_addr = htonl(*addr);
77 tor_inet_ntoa(&in, addrbuf, sizeof(addrbuf));
78 log_fn(LOG_DEBUG,"onion was from %s:%d, circ_id %d.", addrbuf, *port, *circ_id);
81 /** Called when the onion key has changed and we need to spawn new
82 * cpuworkers. Close all currently idle cpuworkers, and mark the last
83 * rotation time as now.
85 void cpuworkers_rotate(void)
87 connection_t *cpuworker;
88 while ((cpuworker = connection_get_by_type_state(CONN_TYPE_CPUWORKER,
89 CPUWORKER_STATE_IDLE))) {
90 connection_mark_for_close(cpuworker);
91 --num_cpuworkers;
93 last_rotation_time = time(NULL);
94 spawn_enough_cpuworkers();
97 /** If the cpuworker closes the connection,
98 * mark it as closed and spawn a new one as needed. */
99 int connection_cpu_reached_eof(connection_t *conn) {
100 log_fn(LOG_WARN,"Read eof. Worker died unexpectedly.");
101 if (conn->state != CPUWORKER_STATE_IDLE) {
102 /* the circ associated with this cpuworker will have to wait until
103 * it gets culled in run_connection_housekeeping(), since we have
104 * no way to find out which circ it was. */
105 log_fn(LOG_WARN,"...and it left a circuit queued; abandoning circ.");
106 num_cpuworkers_busy--;
108 num_cpuworkers--;
109 spawn_enough_cpuworkers(); /* try to regrow. hope we don't end up spinning. */
110 connection_mark_for_close(conn);
111 return 0;
114 /** Called when we get data from a cpuworker. If the answer is not complete,
115 * wait for a complete answer. If the answer is complete,
116 * process it as appropriate.
118 int connection_cpu_process_inbuf(connection_t *conn) {
119 char success;
120 unsigned char buf[LEN_ONION_RESPONSE];
121 uint32_t addr;
122 uint16_t port;
123 uint16_t circ_id;
124 connection_t *p_conn;
125 circuit_t *circ;
127 tor_assert(conn);
128 tor_assert(conn->type == CONN_TYPE_CPUWORKER);
130 if (!buf_datalen(conn->inbuf))
131 return 0;
133 if (conn->state == CPUWORKER_STATE_BUSY_ONION) {
134 if (buf_datalen(conn->inbuf) < LEN_ONION_RESPONSE) /* entire answer available? */
135 return 0; /* not yet */
136 tor_assert(buf_datalen(conn->inbuf) == LEN_ONION_RESPONSE);
138 connection_fetch_from_buf(&success,1,conn);
139 connection_fetch_from_buf(buf,LEN_ONION_RESPONSE-1,conn);
141 /* parse out the circ it was talking about */
142 tag_unpack(buf, &addr, &port, &circ_id);
143 circ = NULL;
144 /* (Here we use connection_exact_get_by_addr_port rather than
145 * get_by_identity_digest: we want a specific port here in
146 * case there are multiple connections.) */
147 p_conn = connection_exact_get_by_addr_port(addr,port);
148 if (p_conn)
149 circ = circuit_get_by_circ_id_conn(circ_id, p_conn);
151 if (success == 0) {
152 log_fn(LOG_INFO,"decoding onionskin failed. Closing.");
153 if (circ)
154 circuit_mark_for_close(circ);
155 goto done_processing;
157 if (!circ) {
158 log_fn(LOG_INFO,"processed onion for a circ that's gone. Dropping.");
159 goto done_processing;
161 tor_assert(circ->p_conn);
162 if (onionskin_answer(circ, buf+TAG_LEN, buf+TAG_LEN+ONIONSKIN_REPLY_LEN) < 0) {
163 log_fn(LOG_WARN,"onionskin_answer failed. Closing.");
164 circuit_mark_for_close(circ);
165 goto done_processing;
167 log_fn(LOG_DEBUG,"onionskin_answer succeeded. Yay.");
168 } else {
169 tor_assert(0); /* don't ask me to do handshakes yet */
172 done_processing:
173 conn->state = CPUWORKER_STATE_IDLE;
174 num_cpuworkers_busy--;
175 if (conn->timestamp_created < last_rotation_time) {
176 connection_mark_for_close(conn);
177 num_cpuworkers--;
178 spawn_enough_cpuworkers();
179 } else {
180 process_pending_task(conn);
182 return 0;
185 /** Implement a cpuworker. 'data' is an fdarray as returned by socketpair.
186 * Read and writes from fdarray[1]. Reads requests, writes answers.
188 * Request format:
189 * Task type [1 byte, always CPUWORKER_TASK_ONION]
190 * Opaque tag TAG_LEN
191 * Onionskin challenge ONIONSKIN_CHALLENGE_LEN
192 * Response format:
193 * Success/failure [1 byte, boolean.]
194 * Opaque tag TAG_LEN
195 * Onionskin challenge ONIONSKIN_REPLY_LEN
196 * Negotiated keys KEY_LEN*2+DIGEST_LEN*2
198 * (Note: this _should_ be by addr/port, since we're concerned with specific
199 * connections, not with routers (where we'd use identity).)
201 static int cpuworker_main(void *data) {
202 unsigned char question[ONIONSKIN_CHALLENGE_LEN];
203 unsigned char question_type;
204 int *fdarray = data;
205 int fd;
207 /* variables for onion processing */
208 unsigned char keys[40+32];
209 unsigned char reply_to_proxy[ONIONSKIN_REPLY_LEN];
210 unsigned char buf[LEN_ONION_RESPONSE];
211 char tag[TAG_LEN];
212 crypto_pk_env_t *onion_key = NULL, *last_onion_key = NULL;
214 fd = fdarray[1]; /* this side is ours */
215 #ifndef TOR_IS_MULTITHREADED
216 tor_close_socket(fdarray[0]); /* this is the side of the socketpair the parent uses */
217 connection_free_all(); /* so the child doesn't hold the parent's fd's open */
218 handle_signals(0); /* ignore interrupts from the keyboard, etc */
219 #endif
220 tor_free(data);
222 dup_onion_keys(&onion_key, &last_onion_key);
224 for (;;) {
225 int r;
227 if ((r = recv(fd, &question_type, 1, 0)) != 1) {
228 // log_fn(LOG_ERR,"read type failed. Exiting.");
229 if (r == 0) {
230 log_fn(LOG_INFO,"CPU worker exiting because Tor process closed connection (either rotated keys or died).");
231 } else {
232 log_fn(LOG_INFO,"CPU worker editing because of error on connection to Tor process.");
233 log_fn(LOG_INFO,"(Error on %d was %s)", fd, tor_socket_strerror(tor_socket_errno(fd)));
235 goto end;
237 tor_assert(question_type == CPUWORKER_TASK_ONION);
239 if (read_all(fd, tag, TAG_LEN, 1) != TAG_LEN) {
240 log_fn(LOG_ERR,"read tag failed. Exiting.");
241 goto end;
244 if (read_all(fd, question, ONIONSKIN_CHALLENGE_LEN, 1) != ONIONSKIN_CHALLENGE_LEN) {
245 log_fn(LOG_ERR,"read question failed. Exiting.");
246 goto end;
249 if (question_type == CPUWORKER_TASK_ONION) {
250 if (onion_skin_server_handshake(question, onion_key, last_onion_key,
251 reply_to_proxy, keys, 40+32) < 0) {
252 /* failure */
253 log_fn(LOG_INFO,"onion_skin_server_handshake failed.");
254 memset(buf,0,LEN_ONION_RESPONSE); /* send all zeros for failure */
255 } else {
256 /* success */
257 log_fn(LOG_DEBUG,"onion_skin_server_handshake succeeded.");
258 buf[0] = 1; /* 1 means success */
259 memcpy(buf+1,tag,TAG_LEN);
260 memcpy(buf+1+TAG_LEN,reply_to_proxy,ONIONSKIN_REPLY_LEN);
261 memcpy(buf+1+TAG_LEN+ONIONSKIN_REPLY_LEN,keys,40+32);
263 if (write_all(fd, buf, LEN_ONION_RESPONSE, 1) != LEN_ONION_RESPONSE) {
264 log_fn(LOG_ERR,"writing response buf failed. Exiting.");
265 goto end;
267 log_fn(LOG_DEBUG,"finished writing response.");
270 end:
271 if (onion_key)
272 crypto_free_pk_env(onion_key);
273 if (last_onion_key)
274 crypto_free_pk_env(last_onion_key);
275 close(fd);
276 spawn_exit();
277 return 0; /* windows wants this function to return an int */
280 /** Launch a new cpuworker.
282 static int spawn_cpuworker(void) {
283 int *fdarray;
284 int fd;
285 connection_t *conn;
287 fdarray = tor_malloc(sizeof(int)*2);
288 if (tor_socketpair(AF_UNIX, SOCK_STREAM, 0, fdarray) < 0) {
289 log(LOG_ERR, "Couldn't construct socketpair: %s",
290 tor_socket_strerror(tor_socket_errno(-1)));
291 tor_cleanup();
292 tor_free(fdarray);
293 exit(1);
296 fd = fdarray[0];
297 spawn_func(cpuworker_main, (void*)fdarray);
298 log_fn(LOG_DEBUG,"just spawned a worker.");
299 #ifndef TOR_IS_MULTITHREADED
300 tor_close_socket(fdarray[1]); /* we don't need the worker's side of the pipe */
301 tor_free(fdarray);
302 #endif
304 conn = connection_new(CONN_TYPE_CPUWORKER);
306 set_socket_nonblocking(fd);
308 /* set up conn so it's got all the data we need to remember */
309 conn->s = fd;
310 conn->address = tor_strdup("localhost");
312 if (connection_add(conn) < 0) { /* no space, forget it */
313 log_fn(LOG_WARN,"connection_add failed. Giving up.");
314 connection_free(conn); /* this closes fd */
315 return -1;
318 conn->state = CPUWORKER_STATE_IDLE;
319 connection_start_reading(conn);
321 return 0; /* success */
324 /** If we have too few or too many active cpuworkers, try to spawn new ones
325 * or kill idle ones.
327 static void spawn_enough_cpuworkers(void) {
328 int num_cpuworkers_needed = get_options()->NumCpus;
330 if (num_cpuworkers_needed < MIN_CPUWORKERS)
331 num_cpuworkers_needed = MIN_CPUWORKERS;
332 if (num_cpuworkers_needed > MAX_CPUWORKERS)
333 num_cpuworkers_needed = MAX_CPUWORKERS;
335 while (num_cpuworkers < num_cpuworkers_needed) {
336 if (spawn_cpuworker() < 0) {
337 log_fn(LOG_WARN,"spawn failed!");
338 return;
340 num_cpuworkers++;
344 /** Take a pending task from the queue and assign it to 'cpuworker'. */
345 static void process_pending_task(connection_t *cpuworker) {
346 circuit_t *circ;
348 tor_assert(cpuworker);
350 /* for now only process onion tasks */
352 circ = onion_next_task();
353 if (!circ)
354 return;
355 if (assign_to_cpuworker(cpuworker, CPUWORKER_TASK_ONION, circ) < 0)
356 log_fn(LOG_WARN,"assign_to_cpuworker failed. Ignoring.");
359 /** if cpuworker is defined, assert that he's idle, and use him. else,
360 * look for an idle cpuworker and use him. if none idle, queue task onto
361 * the pending onion list and return.
362 * If question_type is CPUWORKER_TASK_ONION then task is a circ.
363 * No other question_types are allowed.
365 int assign_to_cpuworker(connection_t *cpuworker, unsigned char question_type,
366 void *task) {
367 circuit_t *circ;
368 char tag[TAG_LEN];
370 tor_assert(question_type == CPUWORKER_TASK_ONION);
372 if (question_type == CPUWORKER_TASK_ONION) {
373 circ = task;
375 if (num_cpuworkers_busy == num_cpuworkers) {
376 log_fn(LOG_DEBUG,"No idle cpuworkers. Queuing.");
377 if (onion_pending_add(circ) < 0)
378 return -1;
379 return 0;
382 if (!cpuworker)
383 cpuworker = connection_get_by_type_state(CONN_TYPE_CPUWORKER, CPUWORKER_STATE_IDLE);
385 tor_assert(cpuworker);
387 if (!circ->p_conn) {
388 log_fn(LOG_INFO,"circ->p_conn gone. Failing circ.");
389 return -1;
391 tag_pack(tag, circ->p_conn->addr, circ->p_conn->port, circ->p_circ_id);
393 cpuworker->state = CPUWORKER_STATE_BUSY_ONION;
394 num_cpuworkers_busy++;
396 connection_write_to_buf(&question_type, 1, cpuworker);
397 connection_write_to_buf(tag, sizeof(tag), cpuworker);
398 connection_write_to_buf(circ->onionskin, ONIONSKIN_CHALLENGE_LEN, cpuworker);
400 return 0;