Commit | Line | Data |
---|---|---|
d7e09d03 PT |
1 | /* |
2 | * GPL HEADER START | |
3 | * | |
4 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License version 2 only, | |
8 | * as published by the Free Software Foundation. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, but | |
11 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
13 | * General Public License version 2 for more details (a copy is included | |
14 | * in the LICENSE file that accompanied this code). | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * version 2 along with this program; If not, see | |
18 | * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf | |
19 | * | |
20 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
21 | * CA 95054 USA or visit www.sun.com if you need additional information or | |
22 | * have any questions. | |
23 | * | |
24 | * GPL HEADER END | |
25 | */ | |
26 | /* | |
27 | * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
30 | * Copyright (c) 2011, 2012, Intel Corporation. | |
31 | */ | |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | */ | |
36 | ||
37 | #define DEBUG_SUBSYSTEM S_LNET | |
38 | #include <linux/lnet/lib-lnet.h> | |
39 | ||
40 | ||
41 | static int accept_port = 988; | |
42 | static int accept_backlog = 127; | |
43 | static int accept_timeout = 5; | |
44 | ||
45 | struct { | |
46 | int pta_shutdown; | |
47 | socket_t *pta_sock; | |
48 | struct completion pta_signal; | |
49 | } lnet_acceptor_state; | |
50 | ||
51 | int | |
52 | lnet_acceptor_port(void) | |
53 | { | |
54 | return accept_port; | |
55 | } | |
56 | ||
57 | static inline int | |
58 | lnet_accept_magic(__u32 magic, __u32 constant) | |
59 | { | |
60 | return (magic == constant || | |
61 | magic == __swab32(constant)); | |
62 | } | |
63 | ||
64 | ||
65 | EXPORT_SYMBOL(lnet_acceptor_port); | |
66 | ||
67 | static char *accept = "secure"; | |
68 | ||
69 | CFS_MODULE_PARM(accept, "s", charp, 0444, | |
70 | "Accept connections (secure|all|none)"); | |
71 | CFS_MODULE_PARM(accept_port, "i", int, 0444, | |
72 | "Acceptor's port (same on all nodes)"); | |
73 | CFS_MODULE_PARM(accept_backlog, "i", int, 0444, | |
74 | "Acceptor's listen backlog"); | |
75 | CFS_MODULE_PARM(accept_timeout, "i", int, 0644, | |
76 | "Acceptor's timeout (seconds)"); | |
77 | ||
78 | static char *accept_type = NULL; | |
79 | ||
80 | int | |
81 | lnet_acceptor_get_tunables(void) | |
82 | { | |
83 | /* Userland acceptor uses 'accept_type' instead of 'accept', due to | |
84 | * conflict with 'accept(2)', but kernel acceptor still uses 'accept' | |
85 | * for compatibility. Hence the trick. */ | |
86 | accept_type = accept; | |
87 | return 0; | |
88 | } | |
89 | ||
90 | int | |
91 | lnet_acceptor_timeout(void) | |
92 | { | |
93 | return accept_timeout; | |
94 | } | |
95 | EXPORT_SYMBOL(lnet_acceptor_timeout); | |
96 | ||
97 | void | |
98 | lnet_connect_console_error (int rc, lnet_nid_t peer_nid, | |
99 | __u32 peer_ip, int peer_port) | |
100 | { | |
101 | switch (rc) { | |
102 | /* "normal" errors */ | |
103 | case -ECONNREFUSED: | |
104 | CNETERR("Connection to %s at host %u.%u.%u.%u on port %d was " | |
105 | "refused: check that Lustre is running on that node.\n", | |
106 | libcfs_nid2str(peer_nid), | |
107 | HIPQUAD(peer_ip), peer_port); | |
108 | break; | |
109 | case -EHOSTUNREACH: | |
110 | case -ENETUNREACH: | |
111 | CNETERR("Connection to %s at host %u.%u.%u.%u " | |
112 | "was unreachable: the network or that node may " | |
113 | "be down, or Lustre may be misconfigured.\n", | |
114 | libcfs_nid2str(peer_nid), HIPQUAD(peer_ip)); | |
115 | break; | |
116 | case -ETIMEDOUT: | |
117 | CNETERR("Connection to %s at host %u.%u.%u.%u on " | |
118 | "port %d took too long: that node may be hung " | |
119 | "or experiencing high load.\n", | |
120 | libcfs_nid2str(peer_nid), | |
121 | HIPQUAD(peer_ip), peer_port); | |
122 | break; | |
123 | case -ECONNRESET: | |
124 | LCONSOLE_ERROR_MSG(0x11b, "Connection to %s at host %u.%u.%u.%u" | |
125 | " on port %d was reset: " | |
126 | "is it running a compatible version of " | |
127 | "Lustre and is %s one of its NIDs?\n", | |
128 | libcfs_nid2str(peer_nid), | |
129 | HIPQUAD(peer_ip), peer_port, | |
130 | libcfs_nid2str(peer_nid)); | |
131 | break; | |
132 | case -EPROTO: | |
133 | LCONSOLE_ERROR_MSG(0x11c, "Protocol error connecting to %s at " | |
134 | "host %u.%u.%u.%u on port %d: is it running " | |
135 | "a compatible version of Lustre?\n", | |
136 | libcfs_nid2str(peer_nid), | |
137 | HIPQUAD(peer_ip), peer_port); | |
138 | break; | |
139 | case -EADDRINUSE: | |
140 | LCONSOLE_ERROR_MSG(0x11d, "No privileged ports available to " | |
141 | "connect to %s at host %u.%u.%u.%u on port " | |
142 | "%d\n", libcfs_nid2str(peer_nid), | |
143 | HIPQUAD(peer_ip), peer_port); | |
144 | break; | |
145 | default: | |
146 | LCONSOLE_ERROR_MSG(0x11e, "Unexpected error %d connecting to %s" | |
147 | " at host %u.%u.%u.%u on port %d\n", rc, | |
148 | libcfs_nid2str(peer_nid), | |
149 | HIPQUAD(peer_ip), peer_port); | |
150 | break; | |
151 | } | |
152 | } | |
153 | EXPORT_SYMBOL(lnet_connect_console_error); | |
154 | ||
155 | int | |
156 | lnet_connect(socket_t **sockp, lnet_nid_t peer_nid, | |
157 | __u32 local_ip, __u32 peer_ip, int peer_port) | |
158 | { | |
159 | lnet_acceptor_connreq_t cr; | |
160 | socket_t *sock; | |
161 | int rc; | |
162 | int port; | |
163 | int fatal; | |
164 | ||
165 | CLASSERT (sizeof(cr) <= 16); /* not too big to be on the stack */ | |
166 | ||
167 | for (port = LNET_ACCEPTOR_MAX_RESERVED_PORT; | |
168 | port >= LNET_ACCEPTOR_MIN_RESERVED_PORT; | |
169 | --port) { | |
170 | /* Iterate through reserved ports. */ | |
171 | ||
172 | rc = libcfs_sock_connect(&sock, &fatal, | |
173 | local_ip, port, | |
174 | peer_ip, peer_port); | |
175 | if (rc != 0) { | |
176 | if (fatal) | |
177 | goto failed; | |
178 | continue; | |
179 | } | |
180 | ||
181 | CLASSERT (LNET_PROTO_ACCEPTOR_VERSION == 1); | |
182 | ||
183 | cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC; | |
184 | cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION; | |
185 | cr.acr_nid = peer_nid; | |
186 | ||
187 | if (the_lnet.ln_testprotocompat != 0) { | |
188 | /* single-shot proto check */ | |
189 | lnet_net_lock(LNET_LOCK_EX); | |
190 | if ((the_lnet.ln_testprotocompat & 4) != 0) { | |
191 | cr.acr_version++; | |
192 | the_lnet.ln_testprotocompat &= ~4; | |
193 | } | |
194 | if ((the_lnet.ln_testprotocompat & 8) != 0) { | |
195 | cr.acr_magic = LNET_PROTO_MAGIC; | |
196 | the_lnet.ln_testprotocompat &= ~8; | |
197 | } | |
198 | lnet_net_unlock(LNET_LOCK_EX); | |
199 | } | |
200 | ||
201 | rc = libcfs_sock_write(sock, &cr, sizeof(cr), | |
202 | accept_timeout); | |
203 | if (rc != 0) | |
204 | goto failed_sock; | |
205 | ||
206 | *sockp = sock; | |
207 | return 0; | |
208 | } | |
209 | ||
210 | rc = -EADDRINUSE; | |
211 | goto failed; | |
212 | ||
213 | failed_sock: | |
214 | libcfs_sock_release(sock); | |
215 | failed: | |
216 | lnet_connect_console_error(rc, peer_nid, peer_ip, peer_port); | |
217 | return rc; | |
218 | } | |
219 | EXPORT_SYMBOL(lnet_connect); | |
220 | ||
221 | ||
222 | /* Below is the code common for both kernel and MT user-space */ | |
223 | ||
224 | int | |
225 | lnet_accept(socket_t *sock, __u32 magic) | |
226 | { | |
227 | lnet_acceptor_connreq_t cr; | |
228 | __u32 peer_ip; | |
229 | int peer_port; | |
230 | int rc; | |
231 | int flip; | |
232 | lnet_ni_t *ni; | |
233 | char *str; | |
234 | ||
235 | LASSERT (sizeof(cr) <= 16); /* not too big for the stack */ | |
236 | ||
237 | rc = libcfs_sock_getaddr(sock, 1, &peer_ip, &peer_port); | |
238 | LASSERT (rc == 0); /* we succeeded before */ | |
239 | ||
240 | if (!lnet_accept_magic(magic, LNET_PROTO_ACCEPTOR_MAGIC)) { | |
241 | ||
242 | if (lnet_accept_magic(magic, LNET_PROTO_MAGIC)) { | |
243 | /* future version compatibility! | |
244 | * When LNET unifies protocols over all LNDs, the first | |
245 | * thing sent will be a version query. I send back | |
246 | * LNET_PROTO_ACCEPTOR_MAGIC to tell her I'm "old" */ | |
247 | ||
248 | memset (&cr, 0, sizeof(cr)); | |
249 | cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC; | |
250 | cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION; | |
251 | rc = libcfs_sock_write(sock, &cr, sizeof(cr), | |
252 | accept_timeout); | |
253 | ||
254 | if (rc != 0) | |
255 | CERROR("Error sending magic+version in response" | |
256 | "to LNET magic from %u.%u.%u.%u: %d\n", | |
257 | HIPQUAD(peer_ip), rc); | |
258 | return -EPROTO; | |
259 | } | |
260 | ||
261 | if (magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) | |
262 | str = "'old' socknal/tcpnal"; | |
263 | else if (lnet_accept_magic(magic, LNET_PROTO_RA_MAGIC)) | |
264 | str = "'old' ranal"; | |
265 | else | |
266 | str = "unrecognised"; | |
267 | ||
268 | LCONSOLE_ERROR_MSG(0x11f, "Refusing connection from %u.%u.%u.%u" | |
269 | " magic %08x: %s acceptor protocol\n", | |
270 | HIPQUAD(peer_ip), magic, str); | |
271 | return -EPROTO; | |
272 | } | |
273 | ||
274 | flip = (magic != LNET_PROTO_ACCEPTOR_MAGIC); | |
275 | ||
276 | rc = libcfs_sock_read(sock, &cr.acr_version, | |
277 | sizeof(cr.acr_version), | |
278 | accept_timeout); | |
279 | if (rc != 0) { | |
280 | CERROR("Error %d reading connection request version from " | |
281 | "%u.%u.%u.%u\n", rc, HIPQUAD(peer_ip)); | |
282 | return -EIO; | |
283 | } | |
284 | ||
285 | if (flip) | |
286 | __swab32s(&cr.acr_version); | |
287 | ||
288 | if (cr.acr_version != LNET_PROTO_ACCEPTOR_VERSION) { | |
289 | /* future version compatibility! | |
290 | * An acceptor-specific protocol rev will first send a version | |
291 | * query. I send back my current version to tell her I'm | |
292 | * "old". */ | |
293 | int peer_version = cr.acr_version; | |
294 | ||
295 | memset (&cr, 0, sizeof(cr)); | |
296 | cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC; | |
297 | cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION; | |
298 | ||
299 | rc = libcfs_sock_write(sock, &cr, sizeof(cr), | |
300 | accept_timeout); | |
301 | ||
302 | if (rc != 0) | |
303 | CERROR("Error sending magic+version in response" | |
304 | "to version %d from %u.%u.%u.%u: %d\n", | |
305 | peer_version, HIPQUAD(peer_ip), rc); | |
306 | return -EPROTO; | |
307 | } | |
308 | ||
309 | rc = libcfs_sock_read(sock, &cr.acr_nid, | |
310 | sizeof(cr) - | |
311 | offsetof(lnet_acceptor_connreq_t, acr_nid), | |
312 | accept_timeout); | |
313 | if (rc != 0) { | |
314 | CERROR("Error %d reading connection request from " | |
315 | "%u.%u.%u.%u\n", rc, HIPQUAD(peer_ip)); | |
316 | return -EIO; | |
317 | } | |
318 | ||
319 | if (flip) | |
320 | __swab64s(&cr.acr_nid); | |
321 | ||
322 | ni = lnet_net2ni(LNET_NIDNET(cr.acr_nid)); | |
323 | if (ni == NULL || /* no matching net */ | |
324 | ni->ni_nid != cr.acr_nid) { /* right NET, wrong NID! */ | |
325 | if (ni != NULL) | |
326 | lnet_ni_decref(ni); | |
327 | LCONSOLE_ERROR_MSG(0x120, "Refusing connection from %u.%u.%u.%u" | |
328 | " for %s: No matching NI\n", | |
329 | HIPQUAD(peer_ip), libcfs_nid2str(cr.acr_nid)); | |
330 | return -EPERM; | |
331 | } | |
332 | ||
333 | if (ni->ni_lnd->lnd_accept == NULL) { | |
334 | /* This catches a request for the loopback LND */ | |
335 | lnet_ni_decref(ni); | |
336 | LCONSOLE_ERROR_MSG(0x121, "Refusing connection from %u.%u.%u.%u" | |
337 | " for %s: NI doesn not accept IP connections\n", | |
338 | HIPQUAD(peer_ip), libcfs_nid2str(cr.acr_nid)); | |
339 | return -EPERM; | |
340 | } | |
341 | ||
342 | CDEBUG(D_NET, "Accept %s from %u.%u.%u.%u\n", | |
343 | libcfs_nid2str(cr.acr_nid), HIPQUAD(peer_ip)); | |
344 | ||
345 | rc = ni->ni_lnd->lnd_accept(ni, sock); | |
346 | ||
347 | lnet_ni_decref(ni); | |
348 | return rc; | |
349 | } | |
350 | ||
351 | int | |
352 | lnet_acceptor(void *arg) | |
353 | { | |
354 | socket_t *newsock; | |
355 | int rc; | |
356 | __u32 magic; | |
357 | __u32 peer_ip; | |
358 | int peer_port; | |
359 | int secure = (int)((long_ptr_t)arg); | |
360 | ||
361 | LASSERT (lnet_acceptor_state.pta_sock == NULL); | |
362 | ||
363 | cfs_block_allsigs(); | |
364 | ||
365 | rc = libcfs_sock_listen(&lnet_acceptor_state.pta_sock, | |
366 | 0, accept_port, accept_backlog); | |
367 | if (rc != 0) { | |
368 | if (rc == -EADDRINUSE) | |
369 | LCONSOLE_ERROR_MSG(0x122, "Can't start acceptor on port" | |
370 | " %d: port already in use\n", | |
371 | accept_port); | |
372 | else | |
373 | LCONSOLE_ERROR_MSG(0x123, "Can't start acceptor on port " | |
374 | "%d: unexpected error %d\n", | |
375 | accept_port, rc); | |
376 | ||
377 | lnet_acceptor_state.pta_sock = NULL; | |
378 | } else { | |
379 | LCONSOLE(0, "Accept %s, port %d\n", accept_type, accept_port); | |
380 | } | |
381 | ||
382 | /* set init status and unblock parent */ | |
383 | lnet_acceptor_state.pta_shutdown = rc; | |
384 | complete(&lnet_acceptor_state.pta_signal); | |
385 | ||
386 | if (rc != 0) | |
387 | return rc; | |
388 | ||
389 | while (!lnet_acceptor_state.pta_shutdown) { | |
390 | ||
391 | rc = libcfs_sock_accept(&newsock, lnet_acceptor_state.pta_sock); | |
392 | if (rc != 0) { | |
393 | if (rc != -EAGAIN) { | |
394 | CWARN("Accept error %d: pausing...\n", rc); | |
395 | cfs_pause(cfs_time_seconds(1)); | |
396 | } | |
397 | continue; | |
398 | } | |
399 | ||
400 | /* maybe we're waken up with libcfs_sock_abort_accept() */ | |
401 | if (lnet_acceptor_state.pta_shutdown) { | |
402 | libcfs_sock_release(newsock); | |
403 | break; | |
404 | } | |
405 | ||
406 | rc = libcfs_sock_getaddr(newsock, 1, &peer_ip, &peer_port); | |
407 | if (rc != 0) { | |
408 | CERROR("Can't determine new connection's address\n"); | |
409 | goto failed; | |
410 | } | |
411 | ||
412 | if (secure && peer_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) { | |
413 | CERROR("Refusing connection from %u.%u.%u.%u: " | |
414 | "insecure port %d\n", | |
415 | HIPQUAD(peer_ip), peer_port); | |
416 | goto failed; | |
417 | } | |
418 | ||
419 | rc = libcfs_sock_read(newsock, &magic, sizeof(magic), | |
420 | accept_timeout); | |
421 | if (rc != 0) { | |
422 | CERROR("Error %d reading connection request from " | |
423 | "%u.%u.%u.%u\n", rc, HIPQUAD(peer_ip)); | |
424 | goto failed; | |
425 | } | |
426 | ||
427 | rc = lnet_accept(newsock, magic); | |
428 | if (rc != 0) | |
429 | goto failed; | |
430 | ||
431 | continue; | |
432 | ||
433 | failed: | |
434 | libcfs_sock_release(newsock); | |
435 | } | |
436 | ||
437 | libcfs_sock_release(lnet_acceptor_state.pta_sock); | |
438 | lnet_acceptor_state.pta_sock = NULL; | |
439 | ||
440 | CDEBUG(D_NET, "Acceptor stopping\n"); | |
441 | ||
442 | /* unblock lnet_acceptor_stop() */ | |
443 | complete(&lnet_acceptor_state.pta_signal); | |
444 | return 0; | |
445 | } | |
446 | ||
447 | static inline int | |
448 | accept2secure(const char *acc, long *sec) | |
449 | { | |
450 | if (!strcmp(acc, "secure")) { | |
451 | *sec = 1; | |
452 | return 1; | |
453 | } else if (!strcmp(acc, "all")) { | |
454 | *sec = 0; | |
455 | return 1; | |
456 | } else if (!strcmp(acc, "none")) { | |
457 | return 0; | |
458 | } else { | |
459 | LCONSOLE_ERROR_MSG(0x124, "Can't parse 'accept=\"%s\"'\n", | |
460 | acc); | |
461 | return -EINVAL; | |
462 | } | |
463 | } | |
464 | ||
465 | int | |
466 | lnet_acceptor_start(void) | |
467 | { | |
468 | int rc; | |
469 | long rc2; | |
470 | long secure; | |
471 | ||
472 | LASSERT (lnet_acceptor_state.pta_sock == NULL); | |
473 | ||
474 | rc = lnet_acceptor_get_tunables(); | |
475 | if (rc != 0) | |
476 | return rc; | |
477 | ||
478 | ||
479 | init_completion(&lnet_acceptor_state.pta_signal); | |
480 | rc = accept2secure(accept_type, &secure); | |
481 | if (rc <= 0) { | |
482 | fini_completion(&lnet_acceptor_state.pta_signal); | |
483 | return rc; | |
484 | } | |
485 | ||
486 | if (lnet_count_acceptor_nis() == 0) /* not required */ | |
487 | return 0; | |
488 | ||
489 | rc2 = PTR_ERR(kthread_run(lnet_acceptor, | |
490 | (void *)(ulong_ptr_t)secure, | |
491 | "acceptor_%03ld", secure)); | |
492 | if (IS_ERR_VALUE(rc2)) { | |
493 | CERROR("Can't start acceptor thread: %ld\n", rc2); | |
494 | fini_completion(&lnet_acceptor_state.pta_signal); | |
495 | ||
496 | return -ESRCH; | |
497 | } | |
498 | ||
499 | /* wait for acceptor to startup */ | |
500 | wait_for_completion(&lnet_acceptor_state.pta_signal); | |
501 | ||
502 | if (!lnet_acceptor_state.pta_shutdown) { | |
503 | /* started OK */ | |
504 | LASSERT(lnet_acceptor_state.pta_sock != NULL); | |
505 | return 0; | |
506 | } | |
507 | ||
508 | LASSERT(lnet_acceptor_state.pta_sock == NULL); | |
509 | fini_completion(&lnet_acceptor_state.pta_signal); | |
510 | ||
511 | return -ENETDOWN; | |
512 | } | |
513 | ||
514 | void | |
515 | lnet_acceptor_stop(void) | |
516 | { | |
517 | if (lnet_acceptor_state.pta_sock == NULL) /* not running */ | |
518 | return; | |
519 | ||
520 | lnet_acceptor_state.pta_shutdown = 1; | |
521 | libcfs_sock_abort_accept(lnet_acceptor_state.pta_sock); | |
522 | ||
523 | /* block until acceptor signals exit */ | |
524 | wait_for_completion(&lnet_acceptor_state.pta_signal); | |
525 | ||
526 | fini_completion(&lnet_acceptor_state.pta_signal); | |
527 | } |