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) 2003, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
1dc563a6 | 30 | * Copyright (c) 2011, 2015, Intel Corporation. |
d7e09d03 PT |
31 | */ |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | * | |
36 | * lustre/ptlrpc/import.c | |
37 | * | |
38 | * Author: Mike Shaver <shaver@clusterfs.com> | |
39 | */ | |
40 | ||
41 | #define DEBUG_SUBSYSTEM S_RPC | |
42 | ||
e27db149 GKH |
43 | #include "../include/obd_support.h" |
44 | #include "../include/lustre_ha.h" | |
45 | #include "../include/lustre_net.h" | |
46 | #include "../include/lustre_import.h" | |
47 | #include "../include/lustre_export.h" | |
48 | #include "../include/obd.h" | |
49 | #include "../include/obd_cksum.h" | |
50 | #include "../include/obd_class.h" | |
d7e09d03 PT |
51 | |
52 | #include "ptlrpc_internal.h" | |
53 | ||
54 | struct ptlrpc_connect_async_args { | |
55 | __u64 pcaa_peer_committed; | |
56 | int pcaa_initial_connect; | |
57 | }; | |
58 | ||
59 | /** | |
60 | * Updates import \a imp current state to provided \a state value | |
61 | * Helper function. Must be called under imp_lock. | |
62 | */ | |
63 | static void __import_set_state(struct obd_import *imp, | |
64 | enum lustre_imp_state state) | |
65 | { | |
502cb58e AS |
66 | switch (state) { |
67 | case LUSTRE_IMP_CLOSED: | |
68 | case LUSTRE_IMP_NEW: | |
69 | case LUSTRE_IMP_DISCON: | |
70 | case LUSTRE_IMP_CONNECTING: | |
71 | break; | |
72 | case LUSTRE_IMP_REPLAY_WAIT: | |
73 | imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS; | |
74 | break; | |
75 | default: | |
76 | imp->imp_replay_state = LUSTRE_IMP_REPLAY; | |
77 | } | |
78 | ||
d7e09d03 PT |
79 | imp->imp_state = state; |
80 | imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state; | |
81 | imp->imp_state_hist[imp->imp_state_hist_idx].ish_time = | |
74e489aa | 82 | ktime_get_real_seconds(); |
d7e09d03 PT |
83 | imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) % |
84 | IMP_STATE_HIST_LEN; | |
85 | } | |
86 | ||
87 | /* A CLOSED import should remain so. */ | |
532118c0 KM |
88 | #define IMPORT_SET_STATE_NOLOCK(imp, state) \ |
89 | do { \ | |
90 | if (imp->imp_state != LUSTRE_IMP_CLOSED) { \ | |
91 | CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n", \ | |
92 | imp, obd2cli_tgt(imp->imp_obd), \ | |
93 | ptlrpc_import_state_name(imp->imp_state), \ | |
94 | ptlrpc_import_state_name(state)); \ | |
95 | __import_set_state(imp, state); \ | |
96 | } \ | |
3949015e | 97 | } while (0) |
d7e09d03 PT |
98 | |
99 | #define IMPORT_SET_STATE(imp, state) \ | |
100 | do { \ | |
101 | spin_lock(&imp->imp_lock); \ | |
102 | IMPORT_SET_STATE_NOLOCK(imp, state); \ | |
103 | spin_unlock(&imp->imp_lock); \ | |
3949015e | 104 | } while (0) |
d7e09d03 | 105 | |
d7e09d03 PT |
106 | static int ptlrpc_connect_interpret(const struct lu_env *env, |
107 | struct ptlrpc_request *request, | |
aff9d8e8 | 108 | void *data, int rc); |
d7e09d03 PT |
109 | int ptlrpc_import_recovery_state_machine(struct obd_import *imp); |
110 | ||
111 | /* Only this function is allowed to change the import state when it is | |
112 | * CLOSED. I would rather refcount the import and free it after | |
113 | * disconnection like we do with exports. To do that, the client_obd | |
114 | * will need to save the peer info somewhere other than in the import, | |
dadfcdab OD |
115 | * though. |
116 | */ | |
d7e09d03 PT |
117 | int ptlrpc_init_import(struct obd_import *imp) |
118 | { | |
119 | spin_lock(&imp->imp_lock); | |
120 | ||
121 | imp->imp_generation++; | |
d0bfef31 | 122 | imp->imp_state = LUSTRE_IMP_NEW; |
d7e09d03 PT |
123 | |
124 | spin_unlock(&imp->imp_lock); | |
125 | ||
126 | return 0; | |
127 | } | |
128 | EXPORT_SYMBOL(ptlrpc_init_import); | |
129 | ||
130 | #define UUID_STR "_UUID" | |
9dd7d427 SB |
131 | static void deuuidify(char *uuid, const char *prefix, char **uuid_start, |
132 | int *uuid_len) | |
d7e09d03 PT |
133 | { |
134 | *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix)) | |
135 | ? uuid : uuid + strlen(prefix); | |
136 | ||
137 | *uuid_len = strlen(*uuid_start); | |
138 | ||
139 | if (*uuid_len < strlen(UUID_STR)) | |
140 | return; | |
141 | ||
142 | if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR), | |
30c0aa39 | 143 | UUID_STR, strlen(UUID_STR))) |
d7e09d03 PT |
144 | *uuid_len -= strlen(UUID_STR); |
145 | } | |
d7e09d03 PT |
146 | |
147 | /** | |
148 | * Returns true if import was FULL, false if import was already not | |
149 | * connected. | |
150 | * @imp - import to be disconnected | |
151 | * @conn_cnt - connection count (epoch) of the request that timed out | |
152 | * and caused the disconnection. In some cases, multiple | |
153 | * inflight requests can fail to a single target (e.g. OST | |
154 | * bulk requests) and if one has already caused a reconnection | |
155 | * (increasing the import->conn_cnt) the older failure should | |
156 | * not also cause a reconnection. If zero it forces a reconnect. | |
157 | */ | |
158 | int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt) | |
159 | { | |
160 | int rc = 0; | |
161 | ||
162 | spin_lock(&imp->imp_lock); | |
163 | ||
164 | if (imp->imp_state == LUSTRE_IMP_FULL && | |
165 | (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) { | |
166 | char *target_start; | |
167 | int target_len; | |
168 | ||
169 | deuuidify(obd2cli_tgt(imp->imp_obd), NULL, | |
170 | &target_start, &target_len); | |
171 | ||
172 | if (imp->imp_replayable) { | |
2d00bd17 JP |
173 | LCONSOLE_WARN("%s: Connection to %.*s (at %s) was lost; in progress operations using this service will wait for recovery to complete\n", |
174 | imp->imp_obd->obd_name, target_len, target_start, | |
175 | libcfs_nid2str(imp->imp_connection->c_peer.nid)); | |
d7e09d03 | 176 | } else { |
2d00bd17 JP |
177 | LCONSOLE_ERROR_MSG(0x166, "%s: Connection to %.*s (at %s) was lost; in progress operations using this service will fail\n", |
178 | imp->imp_obd->obd_name, | |
179 | target_len, target_start, | |
180 | libcfs_nid2str(imp->imp_connection->c_peer.nid)); | |
d7e09d03 | 181 | } |
d7e09d03 PT |
182 | IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON); |
183 | spin_unlock(&imp->imp_lock); | |
184 | ||
185 | if (obd_dump_on_timeout) | |
186 | libcfs_debug_dumplog(); | |
187 | ||
188 | obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON); | |
189 | rc = 1; | |
190 | } else { | |
191 | spin_unlock(&imp->imp_lock); | |
192 | CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n", | |
193 | imp->imp_client->cli_name, imp, | |
194 | (imp->imp_state == LUSTRE_IMP_FULL && | |
195 | imp->imp_conn_cnt > conn_cnt) ? | |
196 | "reconnected" : "not connected", imp->imp_conn_cnt, | |
197 | conn_cnt, ptlrpc_import_state_name(imp->imp_state)); | |
198 | } | |
199 | ||
200 | return rc; | |
201 | } | |
202 | ||
441fda84 ML |
203 | /* |
204 | * This acts as a barrier; all existing requests are rejected, and | |
205 | * no new requests will be accepted until the import is valid again. | |
206 | */ | |
207 | void ptlrpc_deactivate_import(struct obd_import *imp) | |
d7e09d03 | 208 | { |
d7e09d03 | 209 | CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd)); |
441fda84 ML |
210 | |
211 | spin_lock(&imp->imp_lock); | |
d7e09d03 PT |
212 | imp->imp_invalid = 1; |
213 | imp->imp_generation++; | |
214 | spin_unlock(&imp->imp_lock); | |
215 | ||
216 | ptlrpc_abort_inflight(imp); | |
217 | obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE); | |
d7e09d03 | 218 | } |
d7e09d03 PT |
219 | EXPORT_SYMBOL(ptlrpc_deactivate_import); |
220 | ||
221 | static unsigned int | |
219e6de6 | 222 | ptlrpc_inflight_deadline(struct ptlrpc_request *req, time64_t now) |
d7e09d03 PT |
223 | { |
224 | long dl; | |
225 | ||
226 | if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) || | |
227 | (req->rq_phase == RQ_PHASE_BULK) || | |
228 | (req->rq_phase == RQ_PHASE_NEW))) | |
229 | return 0; | |
230 | ||
231 | if (req->rq_timedout) | |
232 | return 0; | |
233 | ||
234 | if (req->rq_phase == RQ_PHASE_NEW) | |
235 | dl = req->rq_sent; | |
236 | else | |
237 | dl = req->rq_deadline; | |
238 | ||
239 | if (dl <= now) | |
240 | return 0; | |
241 | ||
242 | return dl - now; | |
243 | } | |
244 | ||
245 | static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp) | |
246 | { | |
219e6de6 | 247 | time64_t now = ktime_get_real_seconds(); |
d7e09d03 PT |
248 | struct list_head *tmp, *n; |
249 | struct ptlrpc_request *req; | |
250 | unsigned int timeout = 0; | |
251 | ||
252 | spin_lock(&imp->imp_lock); | |
253 | list_for_each_safe(tmp, n, &imp->imp_sending_list) { | |
254 | req = list_entry(tmp, struct ptlrpc_request, rq_list); | |
255 | timeout = max(ptlrpc_inflight_deadline(req, now), timeout); | |
256 | } | |
257 | spin_unlock(&imp->imp_lock); | |
258 | return timeout; | |
259 | } | |
260 | ||
261 | /** | |
262 | * This function will invalidate the import, if necessary, then block | |
263 | * for all the RPC completions, and finally notify the obd to | |
264 | * invalidate its state (ie cancel locks, clear pending requests, | |
265 | * etc). | |
266 | */ | |
267 | void ptlrpc_invalidate_import(struct obd_import *imp) | |
268 | { | |
269 | struct list_head *tmp, *n; | |
270 | struct ptlrpc_request *req; | |
271 | struct l_wait_info lwi; | |
272 | unsigned int timeout; | |
273 | int rc; | |
274 | ||
275 | atomic_inc(&imp->imp_inval_count); | |
276 | ||
277 | if (!imp->imp_invalid || imp->imp_obd->obd_no_recov) | |
278 | ptlrpc_deactivate_import(imp); | |
279 | ||
cca8fca1 | 280 | CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2); |
d7e09d03 PT |
281 | LASSERT(imp->imp_invalid); |
282 | ||
283 | /* Wait forever until inflight == 0. We really can't do it another | |
284 | * way because in some cases we need to wait for very long reply | |
285 | * unlink. We can't do anything before that because there is really | |
dadfcdab OD |
286 | * no guarantee that some rdma transfer is not in progress right now. |
287 | */ | |
d7e09d03 PT |
288 | do { |
289 | /* Calculate max timeout for waiting on rpcs to error | |
290 | * out. Use obd_timeout if calculated value is smaller | |
dadfcdab OD |
291 | * than it. |
292 | */ | |
d7e09d03 PT |
293 | if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) { |
294 | timeout = ptlrpc_inflight_timeout(imp); | |
295 | timeout += timeout / 3; | |
296 | ||
297 | if (timeout == 0) | |
298 | timeout = obd_timeout; | |
299 | } else { | |
300 | /* decrease the interval to increase race condition */ | |
301 | timeout = 1; | |
302 | } | |
303 | ||
1d8cb70c GD |
304 | CDEBUG(D_RPCTRACE, |
305 | "Sleeping %d sec for inflight to error out\n", | |
d7e09d03 PT |
306 | timeout); |
307 | ||
308 | /* Wait for all requests to error out and call completion | |
309 | * callbacks. Cap it at obd_timeout -- these should all | |
dadfcdab OD |
310 | * have been locally cancelled by ptlrpc_abort_inflight. |
311 | */ | |
d7e09d03 PT |
312 | lwi = LWI_TIMEOUT_INTERVAL( |
313 | cfs_timeout_cap(cfs_time_seconds(timeout)), | |
314 | (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2, | |
315 | NULL, NULL); | |
316 | rc = l_wait_event(imp->imp_recovery_waitq, | |
317 | (atomic_read(&imp->imp_inflight) == 0), | |
318 | &lwi); | |
319 | if (rc) { | |
320 | const char *cli_tgt = obd2cli_tgt(imp->imp_obd); | |
321 | ||
322 | CERROR("%s: rc = %d waiting for callback (%d != 0)\n", | |
323 | cli_tgt, rc, | |
324 | atomic_read(&imp->imp_inflight)); | |
325 | ||
326 | spin_lock(&imp->imp_lock); | |
327 | if (atomic_read(&imp->imp_inflight) == 0) { | |
328 | int count = atomic_read(&imp->imp_unregistering); | |
329 | ||
330 | /* We know that "unregistering" rpcs only can | |
331 | * survive in sending or delaying lists (they | |
332 | * maybe waiting for long reply unlink in | |
333 | * sluggish nets). Let's check this. If there | |
334 | * is no inflight and unregistering != 0, this | |
dadfcdab OD |
335 | * is bug. |
336 | */ | |
2d00bd17 JP |
337 | LASSERTF(count == 0, "Some RPCs are still unregistering: %d\n", |
338 | count); | |
d7e09d03 PT |
339 | |
340 | /* Let's save one loop as soon as inflight have | |
341 | * dropped to zero. No new inflights possible at | |
dadfcdab OD |
342 | * this point. |
343 | */ | |
d7e09d03 PT |
344 | rc = 0; |
345 | } else { | |
346 | list_for_each_safe(tmp, n, | |
30c0aa39 | 347 | &imp->imp_sending_list) { |
d7e09d03 | 348 | req = list_entry(tmp, |
30c0aa39 OD |
349 | struct ptlrpc_request, |
350 | rq_list); | |
d7e09d03 PT |
351 | DEBUG_REQ(D_ERROR, req, |
352 | "still on sending list"); | |
353 | } | |
354 | list_for_each_safe(tmp, n, | |
30c0aa39 | 355 | &imp->imp_delayed_list) { |
d7e09d03 | 356 | req = list_entry(tmp, |
30c0aa39 OD |
357 | struct ptlrpc_request, |
358 | rq_list); | |
d7e09d03 PT |
359 | DEBUG_REQ(D_ERROR, req, |
360 | "still on delayed list"); | |
361 | } | |
362 | ||
2d00bd17 JP |
363 | CERROR("%s: RPCs in \"%s\" phase found (%d). Network is sluggish? Waiting them to error out.\n", |
364 | cli_tgt, | |
d7e09d03 PT |
365 | ptlrpc_phase2str(RQ_PHASE_UNREGISTERING), |
366 | atomic_read(&imp-> | |
2d00bd17 | 367 | imp_unregistering)); |
d7e09d03 PT |
368 | } |
369 | spin_unlock(&imp->imp_lock); | |
d0bfef31 | 370 | } |
d7e09d03 PT |
371 | } while (rc != 0); |
372 | ||
373 | /* | |
374 | * Let's additionally check that no new rpcs added to import in | |
375 | * "invalidate" state. | |
376 | */ | |
377 | LASSERT(atomic_read(&imp->imp_inflight) == 0); | |
378 | obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE); | |
379 | sptlrpc_import_flush_all_ctx(imp); | |
380 | ||
381 | atomic_dec(&imp->imp_inval_count); | |
382 | wake_up_all(&imp->imp_recovery_waitq); | |
383 | } | |
384 | EXPORT_SYMBOL(ptlrpc_invalidate_import); | |
385 | ||
386 | /* unset imp_invalid */ | |
387 | void ptlrpc_activate_import(struct obd_import *imp) | |
388 | { | |
389 | struct obd_device *obd = imp->imp_obd; | |
390 | ||
391 | spin_lock(&imp->imp_lock); | |
0b291b9a HZ |
392 | if (imp->imp_deactive != 0) { |
393 | spin_unlock(&imp->imp_lock); | |
394 | return; | |
395 | } | |
396 | ||
d7e09d03 | 397 | imp->imp_invalid = 0; |
d7e09d03 PT |
398 | spin_unlock(&imp->imp_lock); |
399 | obd_import_event(obd, imp, IMP_EVENT_ACTIVE); | |
400 | } | |
401 | EXPORT_SYMBOL(ptlrpc_activate_import); | |
402 | ||
cca8fca1 AS |
403 | static void ptlrpc_pinger_force(struct obd_import *imp) |
404 | { | |
405 | CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd), | |
406 | ptlrpc_import_state_name(imp->imp_state)); | |
407 | ||
408 | spin_lock(&imp->imp_lock); | |
409 | imp->imp_force_verify = 1; | |
410 | spin_unlock(&imp->imp_lock); | |
411 | ||
412 | if (imp->imp_state != LUSTRE_IMP_CONNECTING) | |
413 | ptlrpc_pinger_wake_up(); | |
414 | } | |
415 | ||
d7e09d03 PT |
416 | void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt) |
417 | { | |
d7e09d03 PT |
418 | LASSERT(!imp->imp_dlm_fake); |
419 | ||
420 | if (ptlrpc_set_import_discon(imp, conn_cnt)) { | |
421 | if (!imp->imp_replayable) { | |
2d00bd17 | 422 | CDEBUG(D_HA, "import %s@%s for %s not replayable, auto-deactivating\n", |
d7e09d03 PT |
423 | obd2cli_tgt(imp->imp_obd), |
424 | imp->imp_connection->c_remote_uuid.uuid, | |
425 | imp->imp_obd->obd_name); | |
426 | ptlrpc_deactivate_import(imp); | |
427 | } | |
428 | ||
cca8fca1 | 429 | ptlrpc_pinger_force(imp); |
d7e09d03 | 430 | } |
d7e09d03 PT |
431 | } |
432 | EXPORT_SYMBOL(ptlrpc_fail_import); | |
433 | ||
434 | int ptlrpc_reconnect_import(struct obd_import *imp) | |
435 | { | |
cca8fca1 AS |
436 | struct l_wait_info lwi; |
437 | int secs = cfs_time_seconds(obd_timeout); | |
438 | int rc; | |
439 | ||
440 | ptlrpc_pinger_force(imp); | |
441 | ||
442 | CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n", | |
443 | obd2cli_tgt(imp->imp_obd), secs); | |
444 | ||
445 | lwi = LWI_TIMEOUT(secs, NULL, NULL); | |
446 | rc = l_wait_event(imp->imp_recovery_waitq, | |
447 | !ptlrpc_import_in_recovery(imp), &lwi); | |
448 | CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd), | |
449 | ptlrpc_import_state_name(imp->imp_state)); | |
450 | return rc; | |
d7e09d03 PT |
451 | } |
452 | EXPORT_SYMBOL(ptlrpc_reconnect_import); | |
453 | ||
454 | /** | |
455 | * Connection on import \a imp is changed to another one (if more than one is | |
456 | * present). We typically chose connection that we have not tried to connect to | |
457 | * the longest | |
458 | */ | |
459 | static int import_select_connection(struct obd_import *imp) | |
460 | { | |
461 | struct obd_import_conn *imp_conn = NULL, *conn; | |
462 | struct obd_export *dlmexp; | |
463 | char *target_start; | |
464 | int target_len, tried_all = 1; | |
d7e09d03 PT |
465 | |
466 | spin_lock(&imp->imp_lock); | |
467 | ||
468 | if (list_empty(&imp->imp_conn_list)) { | |
469 | CERROR("%s: no connections available\n", | |
470 | imp->imp_obd->obd_name); | |
471 | spin_unlock(&imp->imp_lock); | |
0a3bdb00 | 472 | return -EINVAL; |
d7e09d03 PT |
473 | } |
474 | ||
475 | list_for_each_entry(conn, &imp->imp_conn_list, oic_item) { | |
b0f5aad5 | 476 | CDEBUG(D_HA, "%s: connect to NID %s last attempt %llu\n", |
d7e09d03 PT |
477 | imp->imp_obd->obd_name, |
478 | libcfs_nid2str(conn->oic_conn->c_peer.nid), | |
479 | conn->oic_last_attempt); | |
480 | ||
481 | /* If we have not tried this connection since | |
dadfcdab OD |
482 | * the last successful attempt, go with this one |
483 | */ | |
d7e09d03 PT |
484 | if ((conn->oic_last_attempt == 0) || |
485 | cfs_time_beforeq_64(conn->oic_last_attempt, | |
30c0aa39 | 486 | imp->imp_last_success_conn)) { |
d7e09d03 PT |
487 | imp_conn = conn; |
488 | tried_all = 0; | |
489 | break; | |
490 | } | |
491 | ||
492 | /* If all of the connections have already been tried | |
dadfcdab OD |
493 | * since the last successful connection; just choose the |
494 | * least recently used | |
495 | */ | |
d7e09d03 PT |
496 | if (!imp_conn) |
497 | imp_conn = conn; | |
498 | else if (cfs_time_before_64(conn->oic_last_attempt, | |
499 | imp_conn->oic_last_attempt)) | |
500 | imp_conn = conn; | |
501 | } | |
502 | ||
503 | /* if not found, simply choose the current one */ | |
504 | if (!imp_conn || imp->imp_force_reconnect) { | |
505 | LASSERT(imp->imp_conn_current); | |
506 | imp_conn = imp->imp_conn_current; | |
507 | tried_all = 0; | |
508 | } | |
509 | LASSERT(imp_conn->oic_conn); | |
510 | ||
511 | /* If we've tried everything, and we're back to the beginning of the | |
dadfcdab OD |
512 | * list, increase our timeout and try again. It will be reset when |
513 | * we do finally connect. (FIXME: really we should wait for all network | |
514 | * state associated with the last connection attempt to drain before | |
515 | * trying to reconnect on it.) | |
516 | */ | |
d7e09d03 PT |
517 | if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) { |
518 | struct adaptive_timeout *at = &imp->imp_at.iat_net_latency; | |
50ffcb7e | 519 | |
d7e09d03 PT |
520 | if (at_get(at) < CONNECTION_SWITCH_MAX) { |
521 | at_measured(at, at_get(at) + CONNECTION_SWITCH_INC); | |
522 | if (at_get(at) > CONNECTION_SWITCH_MAX) | |
523 | at_reset(at, CONNECTION_SWITCH_MAX); | |
524 | } | |
525 | LASSERT(imp_conn->oic_last_attempt); | |
2d00bd17 JP |
526 | CDEBUG(D_HA, "%s: tried all connections, increasing latency to %ds\n", |
527 | imp->imp_obd->obd_name, at_get(at)); | |
d7e09d03 PT |
528 | } |
529 | ||
530 | imp_conn->oic_last_attempt = cfs_time_current_64(); | |
531 | ||
532 | /* switch connection, don't mind if it's same as the current one */ | |
a5cb8880 | 533 | ptlrpc_connection_put(imp->imp_connection); |
d7e09d03 PT |
534 | imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn); |
535 | ||
d0bfef31 | 536 | dlmexp = class_conn2export(&imp->imp_dlm_handle); |
a5cb8880 | 537 | ptlrpc_connection_put(dlmexp->exp_connection); |
d7e09d03 PT |
538 | dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn); |
539 | class_export_put(dlmexp); | |
540 | ||
541 | if (imp->imp_conn_current != imp_conn) { | |
542 | if (imp->imp_conn_current) { | |
543 | deuuidify(obd2cli_tgt(imp->imp_obd), NULL, | |
544 | &target_start, &target_len); | |
545 | ||
2d00bd17 | 546 | CDEBUG(D_HA, "%s: Connection changing to %.*s (at %s)\n", |
d7e09d03 PT |
547 | imp->imp_obd->obd_name, |
548 | target_len, target_start, | |
549 | libcfs_nid2str(imp_conn->oic_conn->c_peer.nid)); | |
550 | } | |
551 | ||
552 | imp->imp_conn_current = imp_conn; | |
553 | } | |
554 | ||
555 | CDEBUG(D_HA, "%s: import %p using connection %s/%s\n", | |
556 | imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid, | |
557 | libcfs_nid2str(imp_conn->oic_conn->c_peer.nid)); | |
558 | ||
559 | spin_unlock(&imp->imp_lock); | |
560 | ||
0a3bdb00 | 561 | return 0; |
d7e09d03 PT |
562 | } |
563 | ||
564 | /* | |
565 | * must be called under imp_lock | |
566 | */ | |
567 | static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno) | |
568 | { | |
569 | struct ptlrpc_request *req; | |
570 | struct list_head *tmp; | |
571 | ||
63d42578 | 572 | /* The requests in committed_list always have smaller transnos than |
dadfcdab OD |
573 | * the requests in replay_list |
574 | */ | |
63d42578 HZ |
575 | if (!list_empty(&imp->imp_committed_list)) { |
576 | tmp = imp->imp_committed_list.next; | |
577 | req = list_entry(tmp, struct ptlrpc_request, rq_replay_list); | |
578 | *transno = req->rq_transno; | |
579 | if (req->rq_transno == 0) { | |
580 | DEBUG_REQ(D_ERROR, req, | |
581 | "zero transno in committed_list"); | |
582 | LBUG(); | |
583 | } | |
584 | return 1; | |
d7e09d03 | 585 | } |
63d42578 HZ |
586 | if (!list_empty(&imp->imp_replay_list)) { |
587 | tmp = imp->imp_replay_list.next; | |
588 | req = list_entry(tmp, struct ptlrpc_request, rq_replay_list); | |
589 | *transno = req->rq_transno; | |
590 | if (req->rq_transno == 0) { | |
591 | DEBUG_REQ(D_ERROR, req, "zero transno in replay_list"); | |
592 | LBUG(); | |
593 | } | |
594 | return 1; | |
595 | } | |
596 | return 0; | |
d7e09d03 PT |
597 | } |
598 | ||
599 | /** | |
600 | * Attempt to (re)connect import \a imp. This includes all preparations, | |
601 | * initializing CONNECT RPC request and passing it to ptlrpcd for | |
602 | * actual sending. | |
603 | * Returns 0 on success or error code. | |
604 | */ | |
605 | int ptlrpc_connect_import(struct obd_import *imp) | |
606 | { | |
607 | struct obd_device *obd = imp->imp_obd; | |
608 | int initial_connect = 0; | |
609 | int set_transno = 0; | |
610 | __u64 committed_before_reconnect = 0; | |
611 | struct ptlrpc_request *request; | |
612 | char *bufs[] = { NULL, | |
613 | obd2cli_tgt(imp->imp_obd), | |
614 | obd->obd_uuid.uuid, | |
615 | (char *)&imp->imp_dlm_handle, | |
616 | (char *)&imp->imp_connect_data }; | |
617 | struct ptlrpc_connect_async_args *aa; | |
618 | int rc; | |
d7e09d03 PT |
619 | |
620 | spin_lock(&imp->imp_lock); | |
621 | if (imp->imp_state == LUSTRE_IMP_CLOSED) { | |
622 | spin_unlock(&imp->imp_lock); | |
623 | CERROR("can't connect to a closed import\n"); | |
0a3bdb00 | 624 | return -EINVAL; |
d7e09d03 PT |
625 | } else if (imp->imp_state == LUSTRE_IMP_FULL) { |
626 | spin_unlock(&imp->imp_lock); | |
627 | CERROR("already connected\n"); | |
0a3bdb00 | 628 | return 0; |
d7e09d03 PT |
629 | } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) { |
630 | spin_unlock(&imp->imp_lock); | |
631 | CERROR("already connecting\n"); | |
0a3bdb00 | 632 | return -EALREADY; |
d7e09d03 PT |
633 | } |
634 | ||
635 | IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING); | |
636 | ||
637 | imp->imp_conn_cnt++; | |
638 | imp->imp_resend_replay = 0; | |
639 | ||
640 | if (!lustre_handle_is_used(&imp->imp_remote_handle)) | |
641 | initial_connect = 1; | |
642 | else | |
643 | committed_before_reconnect = imp->imp_peer_committed_transno; | |
644 | ||
645 | set_transno = ptlrpc_first_transno(imp, | |
646 | &imp->imp_connect_data.ocd_transno); | |
647 | spin_unlock(&imp->imp_lock); | |
648 | ||
649 | rc = import_select_connection(imp); | |
650 | if (rc) | |
a9b3e8f3 | 651 | goto out; |
d7e09d03 | 652 | |
5bcfab13 | 653 | rc = sptlrpc_import_sec_adapt(imp, NULL, NULL); |
d7e09d03 | 654 | if (rc) |
a9b3e8f3 | 655 | goto out; |
d7e09d03 PT |
656 | |
657 | /* Reset connect flags to the originally requested flags, in case | |
dadfcdab OD |
658 | * the server is updated on-the-fly we will get the new features. |
659 | */ | |
d7e09d03 PT |
660 | imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig; |
661 | /* Reset ocd_version each time so the server knows the exact versions */ | |
662 | imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE; | |
663 | imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; | |
664 | imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18; | |
665 | ||
666 | rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd, | |
667 | &obd->obd_uuid, &imp->imp_connect_data, NULL); | |
668 | if (rc) | |
a9b3e8f3 | 669 | goto out; |
d7e09d03 PT |
670 | |
671 | request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT); | |
8b382089 | 672 | if (!request) { |
a9b3e8f3 JL |
673 | rc = -ENOMEM; |
674 | goto out; | |
675 | } | |
d7e09d03 PT |
676 | |
677 | rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION, | |
678 | imp->imp_connect_op, bufs, NULL); | |
679 | if (rc) { | |
680 | ptlrpc_request_free(request); | |
a9b3e8f3 | 681 | goto out; |
d7e09d03 PT |
682 | } |
683 | ||
684 | /* Report the rpc service time to the server so that it knows how long | |
dadfcdab OD |
685 | * to wait for clients to join recovery |
686 | */ | |
d7e09d03 PT |
687 | lustre_msg_set_service_time(request->rq_reqmsg, |
688 | at_timeout2est(request->rq_timeout)); | |
689 | ||
690 | /* The amount of time we give the server to process the connect req. | |
691 | * import_select_connection will increase the net latency on | |
692 | * repeated reconnect attempts to cover slow networks. | |
693 | * We override/ignore the server rpc completion estimate here, | |
dadfcdab OD |
694 | * which may be large if this is a reconnect attempt |
695 | */ | |
d7e09d03 PT |
696 | request->rq_timeout = INITIAL_CONNECT_TIMEOUT; |
697 | lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout); | |
698 | ||
699 | lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER); | |
700 | ||
701 | request->rq_no_resend = request->rq_no_delay = 1; | |
702 | request->rq_send_state = LUSTRE_IMP_CONNECTING; | |
703 | /* Allow a slightly larger reply for future growth compatibility */ | |
704 | req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER, | |
705 | sizeof(struct obd_connect_data)+16*sizeof(__u64)); | |
706 | ptlrpc_request_set_replen(request); | |
707 | request->rq_interpret_reply = ptlrpc_connect_interpret; | |
708 | ||
3949015e | 709 | CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args)); |
d7e09d03 | 710 | aa = ptlrpc_req_async_args(request); |
ec83e611 | 711 | memset(aa, 0, sizeof(*aa)); |
d7e09d03 PT |
712 | |
713 | aa->pcaa_peer_committed = committed_before_reconnect; | |
714 | aa->pcaa_initial_connect = initial_connect; | |
715 | ||
716 | if (aa->pcaa_initial_connect) { | |
717 | spin_lock(&imp->imp_lock); | |
718 | imp->imp_replayable = 1; | |
719 | spin_unlock(&imp->imp_lock); | |
720 | lustre_msg_add_op_flags(request->rq_reqmsg, | |
721 | MSG_CONNECT_INITIAL); | |
722 | } | |
723 | ||
724 | if (set_transno) | |
725 | lustre_msg_add_op_flags(request->rq_reqmsg, | |
726 | MSG_CONNECT_TRANSNO); | |
727 | ||
728 | DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)", | |
729 | request->rq_timeout); | |
c5c4c6fa | 730 | ptlrpcd_add_req(request); |
d7e09d03 PT |
731 | rc = 0; |
732 | out: | |
c5c4c6fa | 733 | if (rc != 0) |
d7e09d03 | 734 | IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON); |
d7e09d03 | 735 | |
0a3bdb00 | 736 | return rc; |
d7e09d03 PT |
737 | } |
738 | EXPORT_SYMBOL(ptlrpc_connect_import); | |
739 | ||
740 | static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp) | |
741 | { | |
742 | int force_verify; | |
743 | ||
744 | spin_lock(&imp->imp_lock); | |
745 | force_verify = imp->imp_force_verify != 0; | |
746 | spin_unlock(&imp->imp_lock); | |
747 | ||
748 | if (force_verify) | |
749 | ptlrpc_pinger_wake_up(); | |
750 | } | |
751 | ||
752 | static int ptlrpc_busy_reconnect(int rc) | |
753 | { | |
754 | return (rc == -EBUSY) || (rc == -EAGAIN); | |
755 | } | |
756 | ||
757 | /** | |
758 | * interpret_reply callback for connect RPCs. | |
759 | * Looks into returned status of connect operation and decides | |
760 | * what to do with the import - i.e enter recovery, promote it to | |
761 | * full state for normal operations of disconnect it due to an error. | |
762 | */ | |
763 | static int ptlrpc_connect_interpret(const struct lu_env *env, | |
764 | struct ptlrpc_request *request, | |
765 | void *data, int rc) | |
766 | { | |
767 | struct ptlrpc_connect_async_args *aa = data; | |
768 | struct obd_import *imp = request->rq_import; | |
769 | struct client_obd *cli = &imp->imp_obd->u.cli; | |
770 | struct lustre_handle old_hdl; | |
771 | __u64 old_connect_flags; | |
772 | int msg_flags; | |
773 | struct obd_connect_data *ocd; | |
774 | struct obd_export *exp; | |
775 | int ret; | |
d7e09d03 PT |
776 | |
777 | spin_lock(&imp->imp_lock); | |
778 | if (imp->imp_state == LUSTRE_IMP_CLOSED) { | |
779 | imp->imp_connect_tried = 1; | |
780 | spin_unlock(&imp->imp_lock); | |
0a3bdb00 | 781 | return 0; |
d7e09d03 PT |
782 | } |
783 | ||
784 | if (rc) { | |
785 | /* if this reconnect to busy export - not need select new target | |
dadfcdab OD |
786 | * for connecting |
787 | */ | |
d7e09d03 PT |
788 | imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc); |
789 | spin_unlock(&imp->imp_lock); | |
790 | ptlrpc_maybe_ping_import_soon(imp); | |
a9b3e8f3 | 791 | goto out; |
d7e09d03 PT |
792 | } |
793 | spin_unlock(&imp->imp_lock); | |
794 | ||
795 | LASSERT(imp->imp_conn_current); | |
796 | ||
797 | msg_flags = lustre_msg_get_op_flags(request->rq_repmsg); | |
798 | ||
799 | ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA, | |
800 | RCL_SERVER); | |
801 | /* server replied obd_connect_data is always bigger */ | |
802 | ocd = req_capsule_server_sized_get(&request->rq_pill, | |
803 | &RMF_CONNECT_DATA, ret); | |
804 | ||
8b382089 | 805 | if (!ocd) { |
d7e09d03 PT |
806 | CERROR("%s: no connect data from server\n", |
807 | imp->imp_obd->obd_name); | |
808 | rc = -EPROTO; | |
a9b3e8f3 | 809 | goto out; |
d7e09d03 PT |
810 | } |
811 | ||
812 | spin_lock(&imp->imp_lock); | |
813 | ||
814 | /* All imports are pingable */ | |
815 | imp->imp_pingable = 1; | |
816 | imp->imp_force_reconnect = 0; | |
817 | imp->imp_force_verify = 0; | |
818 | ||
819 | imp->imp_connect_data = *ocd; | |
820 | ||
821 | CDEBUG(D_HA, "%s: connect to target with instance %u\n", | |
822 | imp->imp_obd->obd_name, ocd->ocd_instance); | |
823 | exp = class_conn2export(&imp->imp_dlm_handle); | |
824 | ||
825 | spin_unlock(&imp->imp_lock); | |
826 | ||
827 | /* check that server granted subset of flags we asked for. */ | |
828 | if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) != | |
829 | ocd->ocd_connect_flags) { | |
55f5a824 | 830 | CERROR("%s: Server didn't granted asked subset of flags: asked=%#llx grranted=%#llx\n", |
1d8cb70c | 831 | imp->imp_obd->obd_name, imp->imp_connect_flags_orig, |
d7e09d03 | 832 | ocd->ocd_connect_flags); |
a9b3e8f3 JL |
833 | rc = -EPROTO; |
834 | goto out; | |
d7e09d03 PT |
835 | } |
836 | ||
837 | if (!exp) { | |
838 | /* This could happen if export is cleaned during the | |
dadfcdab OD |
839 | * connect attempt |
840 | */ | |
d7e09d03 PT |
841 | CERROR("%s: missing export after connect\n", |
842 | imp->imp_obd->obd_name); | |
a9b3e8f3 JL |
843 | rc = -ENODEV; |
844 | goto out; | |
d7e09d03 PT |
845 | } |
846 | old_connect_flags = exp_connect_flags(exp); | |
847 | exp->exp_connect_data = *ocd; | |
848 | imp->imp_obd->obd_self_export->exp_connect_data = *ocd; | |
849 | class_export_put(exp); | |
850 | ||
851 | obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD); | |
852 | ||
853 | if (aa->pcaa_initial_connect) { | |
854 | spin_lock(&imp->imp_lock); | |
855 | if (msg_flags & MSG_CONNECT_REPLAYABLE) { | |
856 | imp->imp_replayable = 1; | |
857 | spin_unlock(&imp->imp_lock); | |
858 | CDEBUG(D_HA, "connected to replayable target: %s\n", | |
859 | obd2cli_tgt(imp->imp_obd)); | |
860 | } else { | |
861 | imp->imp_replayable = 0; | |
862 | spin_unlock(&imp->imp_lock); | |
863 | } | |
864 | ||
865 | /* if applies, adjust the imp->imp_msg_magic here | |
dadfcdab OD |
866 | * according to reply flags |
867 | */ | |
d7e09d03 PT |
868 | |
869 | imp->imp_remote_handle = | |
870 | *lustre_msg_get_handle(request->rq_repmsg); | |
871 | ||
872 | /* Initial connects are allowed for clients with non-random | |
873 | * uuids when servers are in recovery. Simply signal the | |
dadfcdab OD |
874 | * servers replay is complete and wait in REPLAY_WAIT. |
875 | */ | |
d7e09d03 PT |
876 | if (msg_flags & MSG_CONNECT_RECOVERING) { |
877 | CDEBUG(D_HA, "connect to %s during recovery\n", | |
878 | obd2cli_tgt(imp->imp_obd)); | |
879 | IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS); | |
880 | } else { | |
881 | IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL); | |
882 | ptlrpc_activate_import(imp); | |
883 | } | |
884 | ||
a9b3e8f3 JL |
885 | rc = 0; |
886 | goto finish; | |
d7e09d03 PT |
887 | } |
888 | ||
889 | /* Determine what recovery state to move the import to. */ | |
2b241d31 | 890 | if (msg_flags & MSG_CONNECT_RECONNECT) { |
d7e09d03 PT |
891 | memset(&old_hdl, 0, sizeof(old_hdl)); |
892 | if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg), | |
3949015e | 893 | sizeof(old_hdl))) { |
55f5a824 | 894 | LCONSOLE_WARN("Reconnect to %s (at @%s) failed due bad handle %#llx\n", |
d7e09d03 PT |
895 | obd2cli_tgt(imp->imp_obd), |
896 | imp->imp_connection->c_remote_uuid.uuid, | |
897 | imp->imp_dlm_handle.cookie); | |
a9b3e8f3 JL |
898 | rc = -ENOTCONN; |
899 | goto out; | |
d7e09d03 PT |
900 | } |
901 | ||
902 | if (memcmp(&imp->imp_remote_handle, | |
903 | lustre_msg_get_handle(request->rq_repmsg), | |
904 | sizeof(imp->imp_remote_handle))) { | |
905 | int level = msg_flags & MSG_CONNECT_RECOVERING ? | |
906 | D_HA : D_WARNING; | |
907 | ||
908 | /* Bug 16611/14775: if server handle have changed, | |
909 | * that means some sort of disconnection happened. | |
910 | * If the server is not in recovery, that also means it | |
911 | * already erased all of our state because of previous | |
912 | * eviction. If it is in recovery - we are safe to | |
913 | * participate since we can reestablish all of our state | |
dadfcdab OD |
914 | * with server again |
915 | */ | |
2b241d31 | 916 | if ((msg_flags & MSG_CONNECT_RECOVERING)) { |
b533ff4b | 917 | CDEBUG(level, "%s@%s changed server handle from %#llx to %#llx but is still in recovery\n", |
d7e09d03 PT |
918 | obd2cli_tgt(imp->imp_obd), |
919 | imp->imp_connection->c_remote_uuid.uuid, | |
920 | imp->imp_remote_handle.cookie, | |
921 | lustre_msg_get_handle( | |
922 | request->rq_repmsg)->cookie); | |
923 | } else { | |
2d00bd17 | 924 | LCONSOLE_WARN("Evicted from %s (at %s) after server handle changed from %#llx to %#llx\n", |
d7e09d03 PT |
925 | obd2cli_tgt(imp->imp_obd), |
926 | imp->imp_connection-> \ | |
927 | c_remote_uuid.uuid, | |
928 | imp->imp_remote_handle.cookie, | |
929 | lustre_msg_get_handle( | |
2d00bd17 | 930 | request->rq_repmsg)->cookie); |
d7e09d03 PT |
931 | } |
932 | ||
d7e09d03 PT |
933 | imp->imp_remote_handle = |
934 | *lustre_msg_get_handle(request->rq_repmsg); | |
935 | ||
2b241d31 | 936 | if (!(msg_flags & MSG_CONNECT_RECOVERING)) { |
d7e09d03 | 937 | IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED); |
a9b3e8f3 JL |
938 | rc = 0; |
939 | goto finish; | |
d7e09d03 PT |
940 | } |
941 | ||
942 | } else { | |
943 | CDEBUG(D_HA, "reconnected to %s@%s after partition\n", | |
944 | obd2cli_tgt(imp->imp_obd), | |
945 | imp->imp_connection->c_remote_uuid.uuid); | |
946 | } | |
947 | ||
948 | if (imp->imp_invalid) { | |
2d00bd17 JP |
949 | CDEBUG(D_HA, "%s: reconnected but import is invalid; marking evicted\n", |
950 | imp->imp_obd->obd_name); | |
d7e09d03 | 951 | IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED); |
2b241d31 | 952 | } else if (msg_flags & MSG_CONNECT_RECOVERING) { |
d7e09d03 PT |
953 | CDEBUG(D_HA, "%s: reconnected to %s during replay\n", |
954 | imp->imp_obd->obd_name, | |
955 | obd2cli_tgt(imp->imp_obd)); | |
956 | ||
957 | spin_lock(&imp->imp_lock); | |
958 | imp->imp_resend_replay = 1; | |
959 | spin_unlock(&imp->imp_lock); | |
960 | ||
502cb58e | 961 | IMPORT_SET_STATE(imp, imp->imp_replay_state); |
d7e09d03 PT |
962 | } else { |
963 | IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); | |
964 | } | |
2b241d31 | 965 | } else if ((msg_flags & MSG_CONNECT_RECOVERING) && !imp->imp_invalid) { |
d7e09d03 PT |
966 | LASSERT(imp->imp_replayable); |
967 | imp->imp_remote_handle = | |
968 | *lustre_msg_get_handle(request->rq_repmsg); | |
969 | imp->imp_last_replay_transno = 0; | |
970 | IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY); | |
971 | } else { | |
2d00bd17 JP |
972 | DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags not set: %x)", |
973 | imp->imp_obd->obd_name, msg_flags); | |
d7e09d03 PT |
974 | imp->imp_remote_handle = |
975 | *lustre_msg_get_handle(request->rq_repmsg); | |
976 | IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED); | |
977 | } | |
978 | ||
979 | /* Sanity checks for a reconnected import. */ | |
980 | if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) { | |
2d00bd17 | 981 | CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n"); |
d7e09d03 PT |
982 | } |
983 | ||
984 | if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 && | |
985 | lustre_msg_get_last_committed(request->rq_repmsg) < | |
986 | aa->pcaa_peer_committed) { | |
2d00bd17 | 987 | CERROR("%s went back in time (transno %lld was previously committed, server now claims %lld)! See https://bugzilla.lustre.org/show_bug.cgi?id=9646\n", |
d7e09d03 PT |
988 | obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed, |
989 | lustre_msg_get_last_committed(request->rq_repmsg)); | |
990 | } | |
991 | ||
992 | finish: | |
993 | rc = ptlrpc_import_recovery_state_machine(imp); | |
994 | if (rc != 0) { | |
995 | if (rc == -ENOTCONN) { | |
2d00bd17 | 996 | CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery; invalidating and reconnecting\n", |
d7e09d03 PT |
997 | obd2cli_tgt(imp->imp_obd), |
998 | imp->imp_connection->c_remote_uuid.uuid); | |
999 | ptlrpc_connect_import(imp); | |
1000 | imp->imp_connect_tried = 1; | |
0a3bdb00 | 1001 | return 0; |
d7e09d03 PT |
1002 | } |
1003 | } else { | |
1004 | ||
1005 | spin_lock(&imp->imp_lock); | |
1006 | list_del(&imp->imp_conn_current->oic_item); | |
30c0aa39 | 1007 | list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list); |
d7e09d03 PT |
1008 | imp->imp_last_success_conn = |
1009 | imp->imp_conn_current->oic_last_attempt; | |
1010 | ||
1011 | spin_unlock(&imp->imp_lock); | |
1012 | ||
f261f48a FY |
1013 | if ((imp->imp_connect_flags_orig & OBD_CONNECT_IBITS) && |
1014 | !(ocd->ocd_connect_flags & OBD_CONNECT_IBITS)) { | |
2d00bd17 | 1015 | LCONSOLE_WARN("%s: MDS %s does not support ibits lock, either very old or invalid: requested %llx, replied %llx\n", |
f261f48a FY |
1016 | imp->imp_obd->obd_name, |
1017 | imp->imp_connection->c_remote_uuid.uuid, | |
1018 | imp->imp_connect_flags_orig, | |
1019 | ocd->ocd_connect_flags); | |
a9b3e8f3 JL |
1020 | rc = -EPROTO; |
1021 | goto out; | |
f261f48a | 1022 | } |
d7e09d03 PT |
1023 | |
1024 | if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) && | |
1025 | (ocd->ocd_version > LUSTRE_VERSION_CODE + | |
1026 | LUSTRE_VERSION_OFFSET_WARN || | |
1027 | ocd->ocd_version < LUSTRE_VERSION_CODE - | |
1028 | LUSTRE_VERSION_OFFSET_WARN)) { | |
1029 | /* Sigh, some compilers do not like #ifdef in the middle | |
dadfcdab OD |
1030 | * of macro arguments |
1031 | */ | |
2d00bd17 JP |
1032 | const char *older = "older. Consider upgrading server or downgrading client" |
1033 | ; | |
1034 | const char *newer = "newer than client version. Consider upgrading client" | |
1035 | ; | |
d7e09d03 | 1036 | |
2d00bd17 | 1037 | LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) is much %s (%s)\n", |
d7e09d03 PT |
1038 | obd2cli_tgt(imp->imp_obd), |
1039 | OBD_OCD_VERSION_MAJOR(ocd->ocd_version), | |
1040 | OBD_OCD_VERSION_MINOR(ocd->ocd_version), | |
1041 | OBD_OCD_VERSION_PATCH(ocd->ocd_version), | |
1042 | OBD_OCD_VERSION_FIX(ocd->ocd_version), | |
1043 | ocd->ocd_version > LUSTRE_VERSION_CODE ? | |
1044 | newer : older, LUSTRE_VERSION_STRING); | |
1045 | } | |
1046 | ||
1047 | #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0) | |
1048 | /* Check if server has LU-1252 fix applied to not always swab | |
1049 | * the IR MNE entries. Do this only once per connection. This | |
1050 | * fixup is version-limited, because we don't want to carry the | |
1051 | * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we | |
1052 | * need interop with unpatched 2.2 servers. For newer servers, | |
dadfcdab OD |
1053 | * the client will do MNE swabbing only as needed. LU-1644 |
1054 | */ | |
d7e09d03 PT |
1055 | if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) && |
1056 | !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) && | |
1057 | OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 && | |
1058 | OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 && | |
1059 | OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 && | |
1060 | strcmp(imp->imp_obd->obd_type->typ_name, | |
1061 | LUSTRE_MGC_NAME) == 0)) | |
1062 | imp->imp_need_mne_swab = 1; | |
1063 | else /* clear if server was upgraded since last connect */ | |
1064 | imp->imp_need_mne_swab = 0; | |
1065 | #else | |
1066 | #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab" | |
1067 | #endif | |
1068 | ||
1069 | if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) { | |
1070 | /* We sent to the server ocd_cksum_types with bits set | |
1071 | * for algorithms we understand. The server masked off | |
dadfcdab OD |
1072 | * the checksum types it doesn't support |
1073 | */ | |
d7e09d03 PT |
1074 | if ((ocd->ocd_cksum_types & |
1075 | cksum_types_supported_client()) == 0) { | |
2d00bd17 | 1076 | LCONSOLE_WARN("The negotiation of the checksum algorithm to use with server %s failed (%x/%x), disabling checksums\n", |
d7e09d03 PT |
1077 | obd2cli_tgt(imp->imp_obd), |
1078 | ocd->ocd_cksum_types, | |
1079 | cksum_types_supported_client()); | |
1080 | cli->cl_checksum = 0; | |
1081 | cli->cl_supp_cksum_types = OBD_CKSUM_ADLER; | |
1082 | } else { | |
1083 | cli->cl_supp_cksum_types = ocd->ocd_cksum_types; | |
1084 | } | |
1085 | } else { | |
1086 | /* The server does not support OBD_CONNECT_CKSUM. | |
dadfcdab OD |
1087 | * Enforce ADLER for backward compatibility |
1088 | */ | |
d7e09d03 PT |
1089 | cli->cl_supp_cksum_types = OBD_CKSUM_ADLER; |
1090 | } | |
b533ff4b | 1091 | cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types); |
d7e09d03 PT |
1092 | |
1093 | if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) | |
1094 | cli->cl_max_pages_per_rpc = | |
09cbfeaf | 1095 | min(ocd->ocd_brw_size >> PAGE_SHIFT, |
d7e09d03 PT |
1096 | cli->cl_max_pages_per_rpc); |
1097 | else if (imp->imp_connect_op == MDS_CONNECT || | |
1098 | imp->imp_connect_op == MGS_CONNECT) | |
1099 | cli->cl_max_pages_per_rpc = 1; | |
1100 | ||
1101 | /* Reset ns_connect_flags only for initial connect. It might be | |
1102 | * changed in while using FS and if we reset it in reconnect | |
1103 | * this leads to losing user settings done before such as | |
dadfcdab OD |
1104 | * disable lru_resize, etc. |
1105 | */ | |
d7e09d03 PT |
1106 | if (old_connect_flags != exp_connect_flags(exp) || |
1107 | aa->pcaa_initial_connect) { | |
55f5a824 GKH |
1108 | CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server flags: %#llx\n", |
1109 | imp->imp_obd->obd_name, ocd->ocd_connect_flags); | |
d7e09d03 PT |
1110 | imp->imp_obd->obd_namespace->ns_connect_flags = |
1111 | ocd->ocd_connect_flags; | |
1112 | imp->imp_obd->obd_namespace->ns_orig_connect_flags = | |
1113 | ocd->ocd_connect_flags; | |
1114 | } | |
1115 | ||
1116 | if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) && | |
1117 | (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2)) | |
1118 | /* We need a per-message support flag, because | |
dadfcdab OD |
1119 | * a. we don't know if the incoming connect reply |
1120 | * supports AT or not (in reply_in_callback) | |
1121 | * until we unpack it. | |
1122 | * b. failovered server means export and flags are gone | |
1123 | * (in ptlrpc_send_reply). | |
1124 | * Can only be set when we know AT is supported at | |
1125 | * both ends | |
1126 | */ | |
d7e09d03 PT |
1127 | imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT; |
1128 | else | |
1129 | imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; | |
1130 | ||
1131 | if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) && | |
1132 | (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2)) | |
1133 | imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; | |
1134 | else | |
1135 | imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18; | |
1136 | ||
1137 | LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) && | |
1138 | (cli->cl_max_pages_per_rpc > 0)); | |
1139 | } | |
1140 | ||
1141 | out: | |
1142 | imp->imp_connect_tried = 1; | |
1143 | ||
1144 | if (rc != 0) { | |
1145 | IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON); | |
1146 | if (rc == -EACCES) { | |
1147 | /* | |
1148 | * Give up trying to reconnect | |
1149 | * EACCES means client has no permission for connection | |
1150 | */ | |
1151 | imp->imp_obd->obd_no_recov = 1; | |
1152 | ptlrpc_deactivate_import(imp); | |
1153 | } | |
1154 | ||
1155 | if (rc == -EPROTO) { | |
1156 | struct obd_connect_data *ocd; | |
1157 | ||
1158 | /* reply message might not be ready */ | |
8b382089 | 1159 | if (!request->rq_repmsg) |
0a3bdb00 | 1160 | return -EPROTO; |
d7e09d03 PT |
1161 | |
1162 | ocd = req_capsule_server_get(&request->rq_pill, | |
1163 | &RMF_CONNECT_DATA); | |
1164 | if (ocd && | |
1165 | (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) && | |
1166 | (ocd->ocd_version != LUSTRE_VERSION_CODE)) { | |
532118c0 KM |
1167 | /* |
1168 | * Actually servers are only supposed to refuse | |
1169 | * connection from liblustre clients, so we | |
1170 | * should never see this from VFS context | |
1171 | */ | |
2d00bd17 JP |
1172 | LCONSOLE_ERROR_MSG(0x16a, "Server %s version (%d.%d.%d.%d) refused connection from this client with an incompatible version (%s). Client must be recompiled\n", |
1173 | obd2cli_tgt(imp->imp_obd), | |
1174 | OBD_OCD_VERSION_MAJOR(ocd->ocd_version), | |
1175 | OBD_OCD_VERSION_MINOR(ocd->ocd_version), | |
1176 | OBD_OCD_VERSION_PATCH(ocd->ocd_version), | |
1177 | OBD_OCD_VERSION_FIX(ocd->ocd_version), | |
1178 | LUSTRE_VERSION_STRING); | |
d7e09d03 PT |
1179 | ptlrpc_deactivate_import(imp); |
1180 | IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED); | |
1181 | } | |
0a3bdb00 | 1182 | return -EPROTO; |
d7e09d03 PT |
1183 | } |
1184 | ||
1185 | ptlrpc_maybe_ping_import_soon(imp); | |
1186 | ||
1187 | CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n", | |
1188 | obd2cli_tgt(imp->imp_obd), | |
1189 | (char *)imp->imp_connection->c_remote_uuid.uuid, rc); | |
1190 | } | |
1191 | ||
1192 | wake_up_all(&imp->imp_recovery_waitq); | |
0a3bdb00 | 1193 | return rc; |
d7e09d03 PT |
1194 | } |
1195 | ||
1196 | /** | |
1197 | * interpret callback for "completed replay" RPCs. | |
1198 | * \see signal_completed_replay | |
1199 | */ | |
1200 | static int completed_replay_interpret(const struct lu_env *env, | |
1201 | struct ptlrpc_request *req, | |
aff9d8e8 | 1202 | void *data, int rc) |
d7e09d03 | 1203 | { |
d7e09d03 PT |
1204 | atomic_dec(&req->rq_import->imp_replay_inflight); |
1205 | if (req->rq_status == 0 && | |
1206 | !req->rq_import->imp_vbr_failed) { | |
1207 | ptlrpc_import_recovery_state_machine(req->rq_import); | |
1208 | } else { | |
1209 | if (req->rq_import->imp_vbr_failed) { | |
1210 | CDEBUG(D_WARNING, | |
1211 | "%s: version recovery fails, reconnecting\n", | |
1212 | req->rq_import->imp_obd->obd_name); | |
1213 | } else { | |
2d00bd17 | 1214 | CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, reconnecting\n", |
d7e09d03 PT |
1215 | req->rq_import->imp_obd->obd_name, |
1216 | req->rq_status); | |
1217 | } | |
1218 | ptlrpc_connect_import(req->rq_import); | |
1219 | } | |
1220 | ||
0a3bdb00 | 1221 | return 0; |
d7e09d03 PT |
1222 | } |
1223 | ||
1224 | /** | |
1225 | * Let server know that we have no requests to replay anymore. | |
1226 | * Achieved by just sending a PING request | |
1227 | */ | |
1228 | static int signal_completed_replay(struct obd_import *imp) | |
1229 | { | |
1230 | struct ptlrpc_request *req; | |
d7e09d03 PT |
1231 | |
1232 | if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY))) | |
0a3bdb00 | 1233 | return 0; |
d7e09d03 PT |
1234 | |
1235 | LASSERT(atomic_read(&imp->imp_replay_inflight) == 0); | |
1236 | atomic_inc(&imp->imp_replay_inflight); | |
1237 | ||
1238 | req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION, | |
1239 | OBD_PING); | |
8b382089 | 1240 | if (!req) { |
d7e09d03 | 1241 | atomic_dec(&imp->imp_replay_inflight); |
0a3bdb00 | 1242 | return -ENOMEM; |
d7e09d03 PT |
1243 | } |
1244 | ||
1245 | ptlrpc_request_set_replen(req); | |
1246 | req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT; | |
1247 | lustre_msg_add_flags(req->rq_reqmsg, | |
1248 | MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE); | |
1249 | if (AT_OFF) | |
1250 | req->rq_timeout *= 3; | |
1251 | req->rq_interpret_reply = completed_replay_interpret; | |
1252 | ||
c5c4c6fa | 1253 | ptlrpcd_add_req(req); |
0a3bdb00 | 1254 | return 0; |
d7e09d03 PT |
1255 | } |
1256 | ||
1257 | /** | |
1258 | * In kernel code all import invalidation happens in its own | |
1259 | * separate thread, so that whatever application happened to encounter | |
1260 | * a problem could still be killed or otherwise continue | |
1261 | */ | |
1262 | static int ptlrpc_invalidate_import_thread(void *data) | |
1263 | { | |
1264 | struct obd_import *imp = data; | |
1265 | ||
d7e09d03 PT |
1266 | unshare_fs_struct(); |
1267 | ||
1268 | CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n", | |
1269 | imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd), | |
1270 | imp->imp_connection->c_remote_uuid.uuid); | |
1271 | ||
1272 | ptlrpc_invalidate_import(imp); | |
1273 | ||
1274 | if (obd_dump_on_eviction) { | |
1275 | CERROR("dump the log upon eviction\n"); | |
1276 | libcfs_debug_dumplog(); | |
1277 | } | |
1278 | ||
1279 | IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); | |
1280 | ptlrpc_import_recovery_state_machine(imp); | |
1281 | ||
1282 | class_import_put(imp); | |
0a3bdb00 | 1283 | return 0; |
d7e09d03 PT |
1284 | } |
1285 | ||
1286 | /** | |
1287 | * This is the state machine for client-side recovery on import. | |
1288 | * | |
b6da17f3 | 1289 | * Typically we have two possibly paths. If we came to server and it is not |
d7e09d03 PT |
1290 | * in recovery, we just enter IMP_EVICTED state, invalidate our import |
1291 | * state and reconnect from scratch. | |
1292 | * If we came to server that is in recovery, we enter IMP_REPLAY import state. | |
1293 | * We go through our list of requests to replay and send them to server one by | |
1294 | * one. | |
1295 | * After sending all request from the list we change import state to | |
1296 | * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server | |
1297 | * and also all the locks we don't yet have and wait for server to grant us. | |
1298 | * After that we send a special "replay completed" request and change import | |
1299 | * state to IMP_REPLAY_WAIT. | |
1300 | * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER | |
1301 | * state and resend all requests from sending list. | |
1302 | * After that we promote import to FULL state and send all delayed requests | |
1303 | * and import is fully operational after that. | |
1304 | * | |
1305 | */ | |
1306 | int ptlrpc_import_recovery_state_machine(struct obd_import *imp) | |
1307 | { | |
1308 | int rc = 0; | |
1309 | int inflight; | |
1310 | char *target_start; | |
1311 | int target_len; | |
1312 | ||
d7e09d03 PT |
1313 | if (imp->imp_state == LUSTRE_IMP_EVICTED) { |
1314 | deuuidify(obd2cli_tgt(imp->imp_obd), NULL, | |
1315 | &target_start, &target_len); | |
1316 | /* Don't care about MGC eviction */ | |
1317 | if (strcmp(imp->imp_obd->obd_type->typ_name, | |
1318 | LUSTRE_MGC_NAME) != 0) { | |
2d00bd17 | 1319 | LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted by %.*s; in progress operations using this service will fail.\n", |
d7e09d03 PT |
1320 | imp->imp_obd->obd_name, target_len, |
1321 | target_start); | |
1322 | } | |
1323 | CDEBUG(D_HA, "evicted from %s@%s; invalidating\n", | |
1324 | obd2cli_tgt(imp->imp_obd), | |
1325 | imp->imp_connection->c_remote_uuid.uuid); | |
1326 | /* reset vbr_failed flag upon eviction */ | |
1327 | spin_lock(&imp->imp_lock); | |
1328 | imp->imp_vbr_failed = 0; | |
1329 | spin_unlock(&imp->imp_lock); | |
1330 | ||
1331 | { | |
68b636b6 | 1332 | struct task_struct *task; |
d7e09d03 | 1333 | /* bug 17802: XXX client_disconnect_export vs connect request |
9c379663 | 1334 | * race. if client is evicted at this time, we start |
d7e09d03 | 1335 | * invalidate thread without reference to import and import can |
dadfcdab OD |
1336 | * be freed at same time. |
1337 | */ | |
d7e09d03 PT |
1338 | class_import_get(imp); |
1339 | task = kthread_run(ptlrpc_invalidate_import_thread, imp, | |
30c0aa39 | 1340 | "ll_imp_inval"); |
d7e09d03 PT |
1341 | if (IS_ERR(task)) { |
1342 | class_import_put(imp); | |
1343 | CERROR("error starting invalidate thread: %d\n", rc); | |
1344 | rc = PTR_ERR(task); | |
1345 | } else { | |
1346 | rc = 0; | |
1347 | } | |
0a3bdb00 | 1348 | return rc; |
d7e09d03 PT |
1349 | } |
1350 | } | |
1351 | ||
1352 | if (imp->imp_state == LUSTRE_IMP_REPLAY) { | |
1353 | CDEBUG(D_HA, "replay requested by %s\n", | |
1354 | obd2cli_tgt(imp->imp_obd)); | |
1355 | rc = ptlrpc_replay_next(imp, &inflight); | |
1356 | if (inflight == 0 && | |
1357 | atomic_read(&imp->imp_replay_inflight) == 0) { | |
1358 | IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS); | |
1359 | rc = ldlm_replay_locks(imp); | |
1360 | if (rc) | |
a9b3e8f3 | 1361 | goto out; |
d7e09d03 PT |
1362 | } |
1363 | rc = 0; | |
1364 | } | |
1365 | ||
1366 | if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) { | |
1367 | if (atomic_read(&imp->imp_replay_inflight) == 0) { | |
1368 | IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT); | |
1369 | rc = signal_completed_replay(imp); | |
1370 | if (rc) | |
a9b3e8f3 | 1371 | goto out; |
d7e09d03 PT |
1372 | } |
1373 | ||
1374 | } | |
1375 | ||
1376 | if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) { | |
1377 | if (atomic_read(&imp->imp_replay_inflight) == 0) { | |
1378 | IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER); | |
1379 | } | |
1380 | } | |
1381 | ||
1382 | if (imp->imp_state == LUSTRE_IMP_RECOVER) { | |
1383 | CDEBUG(D_HA, "reconnected to %s@%s\n", | |
1384 | obd2cli_tgt(imp->imp_obd), | |
1385 | imp->imp_connection->c_remote_uuid.uuid); | |
1386 | ||
1387 | rc = ptlrpc_resend(imp); | |
1388 | if (rc) | |
a9b3e8f3 | 1389 | goto out; |
d7e09d03 PT |
1390 | IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL); |
1391 | ptlrpc_activate_import(imp); | |
1392 | ||
1393 | deuuidify(obd2cli_tgt(imp->imp_obd), NULL, | |
1394 | &target_start, &target_len); | |
1395 | LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n", | |
1396 | imp->imp_obd->obd_name, | |
1397 | target_len, target_start, | |
1398 | libcfs_nid2str(imp->imp_connection->c_peer.nid)); | |
1399 | } | |
1400 | ||
1401 | if (imp->imp_state == LUSTRE_IMP_FULL) { | |
1402 | wake_up_all(&imp->imp_recovery_waitq); | |
1403 | ptlrpc_wake_delayed(imp); | |
1404 | } | |
1405 | ||
1406 | out: | |
0a3bdb00 | 1407 | return rc; |
d7e09d03 PT |
1408 | } |
1409 | ||
1410 | int ptlrpc_disconnect_import(struct obd_import *imp, int noclose) | |
1411 | { | |
1412 | struct ptlrpc_request *req; | |
1413 | int rq_opc, rc = 0; | |
d7e09d03 | 1414 | |
88291a7a | 1415 | if (imp->imp_obd->obd_force) |
a9b3e8f3 | 1416 | goto set_state; |
d7e09d03 PT |
1417 | |
1418 | switch (imp->imp_connect_op) { | |
88291a7a AD |
1419 | case OST_CONNECT: |
1420 | rq_opc = OST_DISCONNECT; | |
1421 | break; | |
1422 | case MDS_CONNECT: | |
1423 | rq_opc = MDS_DISCONNECT; | |
1424 | break; | |
1425 | case MGS_CONNECT: | |
1426 | rq_opc = MGS_DISCONNECT; | |
1427 | break; | |
d7e09d03 | 1428 | default: |
88291a7a | 1429 | rc = -EINVAL; |
2d00bd17 | 1430 | CERROR("%s: don't know how to disconnect from %s (connect_op %d): rc = %d\n", |
88291a7a AD |
1431 | imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd), |
1432 | imp->imp_connect_op, rc); | |
1433 | return rc; | |
d7e09d03 PT |
1434 | } |
1435 | ||
1436 | if (ptlrpc_import_in_recovery(imp)) { | |
1437 | struct l_wait_info lwi; | |
b2d201bd | 1438 | long timeout; |
d7e09d03 | 1439 | |
d7e09d03 PT |
1440 | if (AT_OFF) { |
1441 | if (imp->imp_server_timeout) | |
1442 | timeout = cfs_time_seconds(obd_timeout / 2); | |
1443 | else | |
1444 | timeout = cfs_time_seconds(obd_timeout); | |
1445 | } else { | |
1446 | int idx = import_at_get_index(imp, | |
1447 | imp->imp_client->cli_request_portal); | |
1448 | timeout = cfs_time_seconds( | |
1449 | at_get(&imp->imp_at.iat_service_estimate[idx])); | |
1450 | } | |
1451 | ||
1452 | lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout), | |
1453 | back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL); | |
1454 | rc = l_wait_event(imp->imp_recovery_waitq, | |
1455 | !ptlrpc_import_in_recovery(imp), &lwi); | |
1456 | ||
1457 | } | |
1458 | ||
1459 | spin_lock(&imp->imp_lock); | |
1460 | if (imp->imp_state != LUSTRE_IMP_FULL) | |
a9b3e8f3 | 1461 | goto out; |
d7e09d03 PT |
1462 | spin_unlock(&imp->imp_lock); |
1463 | ||
1464 | req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT, | |
1465 | LUSTRE_OBD_VERSION, rq_opc); | |
1466 | if (req) { | |
1467 | /* We are disconnecting, do not retry a failed DISCONNECT rpc if | |
1468 | * it fails. We can get through the above with a down server | |
dadfcdab OD |
1469 | * if the client doesn't know the server is gone yet. |
1470 | */ | |
d7e09d03 PT |
1471 | req->rq_no_resend = 1; |
1472 | ||
1473 | /* We want client umounts to happen quickly, no matter the | |
dadfcdab OD |
1474 | * server state... |
1475 | */ | |
d7e09d03 PT |
1476 | req->rq_timeout = min_t(int, req->rq_timeout, |
1477 | INITIAL_CONNECT_TIMEOUT); | |
1478 | ||
1479 | IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING); | |
d0bfef31 | 1480 | req->rq_send_state = LUSTRE_IMP_CONNECTING; |
d7e09d03 PT |
1481 | ptlrpc_request_set_replen(req); |
1482 | rc = ptlrpc_queue_wait(req); | |
1483 | ptlrpc_req_finished(req); | |
1484 | } | |
1485 | ||
1486 | set_state: | |
1487 | spin_lock(&imp->imp_lock); | |
1488 | out: | |
1489 | if (noclose) | |
1490 | IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON); | |
1491 | else | |
1492 | IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED); | |
1493 | memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle)); | |
1494 | spin_unlock(&imp->imp_lock); | |
1495 | ||
88291a7a AD |
1496 | if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN) |
1497 | rc = 0; | |
1498 | ||
0a3bdb00 | 1499 | return rc; |
d7e09d03 PT |
1500 | } |
1501 | EXPORT_SYMBOL(ptlrpc_disconnect_import); | |
1502 | ||
d7e09d03 PT |
1503 | /* Adaptive Timeout utils */ |
1504 | extern unsigned int at_min, at_max, at_history; | |
1505 | ||
1506 | /* Bin into timeslices using AT_BINS bins. | |
dadfcdab OD |
1507 | * This gives us a max of the last binlimit*AT_BINS secs without the storage, |
1508 | * but still smoothing out a return to normalcy from a slow response. | |
1509 | * (E.g. remember the maximum latency in each minute of the last 4 minutes.) | |
1510 | */ | |
d7e09d03 PT |
1511 | int at_measured(struct adaptive_timeout *at, unsigned int val) |
1512 | { | |
1513 | unsigned int old = at->at_current; | |
0ac0478b AB |
1514 | time64_t now = ktime_get_real_seconds(); |
1515 | long binlimit = max_t(long, at_history / AT_BINS, 1); | |
d7e09d03 PT |
1516 | |
1517 | LASSERT(at); | |
1518 | CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n", | |
0ac0478b | 1519 | val, at, (long)(now - at->at_binstart), at->at_current, |
d7e09d03 PT |
1520 | at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]); |
1521 | ||
1522 | if (val == 0) | |
1523 | /* 0's don't count, because we never want our timeout to | |
dadfcdab OD |
1524 | * drop to 0, and because 0 could mean an error |
1525 | */ | |
d7e09d03 PT |
1526 | return 0; |
1527 | ||
1528 | spin_lock(&at->at_lock); | |
1529 | ||
1530 | if (unlikely(at->at_binstart == 0)) { | |
1531 | /* Special case to remove default from history */ | |
1532 | at->at_current = val; | |
1533 | at->at_worst_ever = val; | |
1534 | at->at_worst_time = now; | |
1535 | at->at_hist[0] = val; | |
1536 | at->at_binstart = now; | |
3949015e | 1537 | } else if (now - at->at_binstart < binlimit) { |
d7e09d03 PT |
1538 | /* in bin 0 */ |
1539 | at->at_hist[0] = max(val, at->at_hist[0]); | |
1540 | at->at_current = max(val, at->at_current); | |
1541 | } else { | |
1542 | int i, shift; | |
1543 | unsigned int maxv = val; | |
1544 | /* move bins over */ | |
0ac0478b | 1545 | shift = (u32)(now - at->at_binstart) / binlimit; |
d7e09d03 | 1546 | LASSERT(shift > 0); |
3949015e | 1547 | for (i = AT_BINS - 1; i >= 0; i--) { |
d7e09d03 PT |
1548 | if (i >= shift) { |
1549 | at->at_hist[i] = at->at_hist[i - shift]; | |
1550 | maxv = max(maxv, at->at_hist[i]); | |
1551 | } else { | |
1552 | at->at_hist[i] = 0; | |
1553 | } | |
1554 | } | |
1555 | at->at_hist[0] = val; | |
1556 | at->at_current = maxv; | |
1557 | at->at_binstart += shift * binlimit; | |
1558 | } | |
1559 | ||
1560 | if (at->at_current > at->at_worst_ever) { | |
1561 | at->at_worst_ever = at->at_current; | |
1562 | at->at_worst_time = now; | |
1563 | } | |
1564 | ||
1565 | if (at->at_flags & AT_FLG_NOHIST) | |
1566 | /* Only keep last reported val; keeping the rest of the history | |
dadfcdab OD |
1567 | * for debugfs only |
1568 | */ | |
d7e09d03 PT |
1569 | at->at_current = val; |
1570 | ||
1571 | if (at_max > 0) | |
1572 | at->at_current = min(at->at_current, at_max); | |
1573 | at->at_current = max(at->at_current, at_min); | |
1574 | ||
1575 | if (at->at_current != old) | |
2d00bd17 JP |
1576 | CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d (val=%u) hist %u %u %u %u\n", |
1577 | at, | |
d7e09d03 PT |
1578 | old, at->at_current, at->at_current - old, val, |
1579 | at->at_hist[0], at->at_hist[1], at->at_hist[2], | |
1580 | at->at_hist[3]); | |
1581 | ||
1582 | /* if we changed, report the old value */ | |
1583 | old = (at->at_current != old) ? old : 0; | |
1584 | ||
1585 | spin_unlock(&at->at_lock); | |
1586 | return old; | |
1587 | } | |
1588 | ||
1589 | /* Find the imp_at index for a given portal; assign if space available */ | |
1590 | int import_at_get_index(struct obd_import *imp, int portal) | |
1591 | { | |
1592 | struct imp_at *at = &imp->imp_at; | |
1593 | int i; | |
1594 | ||
1595 | for (i = 0; i < IMP_AT_MAX_PORTALS; i++) { | |
1596 | if (at->iat_portal[i] == portal) | |
1597 | return i; | |
1598 | if (at->iat_portal[i] == 0) | |
1599 | /* unused */ | |
1600 | break; | |
1601 | } | |
1602 | ||
1603 | /* Not found in list, add it under a lock */ | |
1604 | spin_lock(&imp->imp_lock); | |
1605 | ||
1606 | /* Check unused under lock */ | |
1607 | for (; i < IMP_AT_MAX_PORTALS; i++) { | |
1608 | if (at->iat_portal[i] == portal) | |
1609 | goto out; | |
1610 | if (at->iat_portal[i] == 0) | |
1611 | /* unused */ | |
1612 | break; | |
1613 | } | |
1614 | ||
1615 | /* Not enough portals? */ | |
1616 | LASSERT(i < IMP_AT_MAX_PORTALS); | |
1617 | ||
1618 | at->iat_portal[i] = portal; | |
1619 | out: | |
1620 | spin_unlock(&imp->imp_lock); | |
1621 | return i; | |
1622 | } |