Whamcloud - gitweb
LU-473 llite: improve error for 2.x cli + 1.8 srv
[fs/lustre-release.git] / lustre / ptlrpc / import.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
31  */
32 /*
33  * Copyright (c) 2011 Whamcloud, Inc.
34  */
35 /*
36  * This file is part of Lustre, http://www.lustre.org/
37  * Lustre is a trademark of Sun Microsystems, Inc.
38  *
39  * lustre/ptlrpc/import.c
40  *
41  * Author: Mike Shaver <shaver@clusterfs.com>
42  */
43
44 #define DEBUG_SUBSYSTEM S_RPC
45 #ifndef __KERNEL__
46 # include <liblustre.h>
47 #endif
48
49 #include <obd_support.h>
50 #include <lustre_ha.h>
51 #include <lustre_net.h>
52 #include <lustre_import.h>
53 #include <lustre_export.h>
54 #include <obd.h>
55 #include <obd_cksum.h>
56 #include <obd_class.h>
57
58 #include "ptlrpc_internal.h"
59
60 struct ptlrpc_connect_async_args {
61          __u64 pcaa_peer_committed;
62         int pcaa_initial_connect;
63 };
64
65 /**
66  * Updates import \a imp current state to provided \a state value
67  * Helper function. Must be called under imp_lock.
68  */
69 static void __import_set_state(struct obd_import *imp,
70                                enum lustre_imp_state state)
71 {
72         imp->imp_state = state;
73         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
74         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
75                 cfs_time_current_sec();
76         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
77                 IMP_STATE_HIST_LEN;
78 }
79
80 /* A CLOSED import should remain so. */
81 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
82 do {                                                                           \
83         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
84                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
85                       imp, obd2cli_tgt(imp->imp_obd),                          \
86                       ptlrpc_import_state_name(imp->imp_state),                \
87                       ptlrpc_import_state_name(state));                        \
88                __import_set_state(imp, state);                                 \
89         }                                                                      \
90 } while(0)
91
92 #define IMPORT_SET_STATE(imp, state)            \
93 do {                                            \
94         cfs_spin_lock(&imp->imp_lock);          \
95         IMPORT_SET_STATE_NOLOCK(imp, state);    \
96         cfs_spin_unlock(&imp->imp_lock);        \
97 } while(0)
98
99
100 static int ptlrpc_connect_interpret(const struct lu_env *env,
101                                     struct ptlrpc_request *request,
102                                     void * data, int rc);
103 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
104
105 /* Only this function is allowed to change the import state when it is
106  * CLOSED. I would rather refcount the import and free it after
107  * disconnection like we do with exports. To do that, the client_obd
108  * will need to save the peer info somewhere other than in the import,
109  * though. */
110 int ptlrpc_init_import(struct obd_import *imp)
111 {
112         cfs_spin_lock(&imp->imp_lock);
113
114         imp->imp_generation++;
115         imp->imp_state =  LUSTRE_IMP_NEW;
116
117         cfs_spin_unlock(&imp->imp_lock);
118
119         return 0;
120 }
121 EXPORT_SYMBOL(ptlrpc_init_import);
122
123 #define UUID_STR "_UUID"
124 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
125                       int *uuid_len)
126 {
127         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
128                 ? uuid : uuid + strlen(prefix);
129
130         *uuid_len = strlen(*uuid_start);
131
132         if (*uuid_len < strlen(UUID_STR))
133                 return;
134
135         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
136                     UUID_STR, strlen(UUID_STR)))
137                 *uuid_len -= strlen(UUID_STR);
138 }
139
140 /**
141  * Returns true if import was FULL, false if import was already not
142  * connected.
143  * @imp - import to be disconnected
144  * @conn_cnt - connection count (epoch) of the request that timed out
145  *             and caused the disconnection.  In some cases, multiple
146  *             inflight requests can fail to a single target (e.g. OST
147  *             bulk requests) and if one has already caused a reconnection
148  *             (increasing the import->conn_cnt) the older failure should
149  *             not also cause a reconnection.  If zero it forces a reconnect.
150  */
151 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
152 {
153         int rc = 0;
154
155         cfs_spin_lock(&imp->imp_lock);
156
157         if (imp->imp_state == LUSTRE_IMP_FULL &&
158             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
159                 char *target_start;
160                 int   target_len;
161
162                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
163                           &target_start, &target_len);
164
165                 if (imp->imp_replayable) {
166                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
167                                "%s was lost; in progress operations using this "
168                                "service will wait for recovery to complete.\n",
169                                imp->imp_obd->obd_name, target_len, target_start,
170                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
171                 } else {
172                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
173                                "%.*s via nid %s was lost; in progress "
174                                "operations using this service will fail.\n",
175                                imp->imp_obd->obd_name,
176                                target_len, target_start,
177                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
178                 }
179                 ptlrpc_deactivate_timeouts(imp);
180                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
181                 cfs_spin_unlock(&imp->imp_lock);
182
183                 if (obd_dump_on_timeout)
184                         libcfs_debug_dumplog();
185
186                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
187                 rc = 1;
188         } else {
189                 cfs_spin_unlock(&imp->imp_lock);
190                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
191                        imp->imp_client->cli_name, imp,
192                        (imp->imp_state == LUSTRE_IMP_FULL &&
193                         imp->imp_conn_cnt > conn_cnt) ?
194                        "reconnected" : "not connected", imp->imp_conn_cnt,
195                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
196         }
197
198         return rc;
199 }
200
201 /* Must be called with imp_lock held! */
202 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
203 {
204         ENTRY;
205         LASSERT_SPIN_LOCKED(&imp->imp_lock);
206
207         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
208         imp->imp_invalid = 1;
209         imp->imp_generation++;
210         cfs_spin_unlock(&imp->imp_lock);
211
212         ptlrpc_abort_inflight(imp);
213         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
214
215         EXIT;
216 }
217
218 /*
219  * This acts as a barrier; all existing requests are rejected, and
220  * no new requests will be accepted until the import is valid again.
221  */
222 void ptlrpc_deactivate_import(struct obd_import *imp)
223 {
224         cfs_spin_lock(&imp->imp_lock);
225         ptlrpc_deactivate_and_unlock_import(imp);
226 }
227
228 static unsigned int
229 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
230 {
231         long dl;
232
233         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
234               (req->rq_phase == RQ_PHASE_BULK) ||
235               (req->rq_phase == RQ_PHASE_NEW)))
236                 return 0;
237
238         if (req->rq_timedout)
239                 return 0;
240
241         if (req->rq_phase == RQ_PHASE_NEW)
242                 dl = req->rq_sent;
243         else
244                 dl = req->rq_deadline;
245
246         if (dl <= now)
247                 return 0;
248
249         return dl - now;
250 }
251
252 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
253 {
254         time_t now = cfs_time_current_sec();
255         cfs_list_t *tmp, *n;
256         struct ptlrpc_request *req;
257         unsigned int timeout = 0;
258
259         cfs_spin_lock(&imp->imp_lock);
260         cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
261                 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
262                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
263         }
264         cfs_spin_unlock(&imp->imp_lock);
265         return timeout;
266 }
267
268 /**
269  * This function will invalidate the import, if necessary, then block
270  * for all the RPC completions, and finally notify the obd to
271  * invalidate its state (ie cancel locks, clear pending requests,
272  * etc).
273  */
274 void ptlrpc_invalidate_import(struct obd_import *imp)
275 {
276         cfs_list_t *tmp, *n;
277         struct ptlrpc_request *req;
278         struct l_wait_info lwi;
279         unsigned int timeout;
280         int rc;
281
282         cfs_atomic_inc(&imp->imp_inval_count);
283
284         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
285                 ptlrpc_deactivate_import(imp);
286
287         LASSERT(imp->imp_invalid);
288
289         /* Wait forever until inflight == 0. We really can't do it another
290          * way because in some cases we need to wait for very long reply
291          * unlink. We can't do anything before that because there is really
292          * no guarantee that some rdma transfer is not in progress right now. */
293         do {
294                 /* Calculate max timeout for waiting on rpcs to error
295                  * out. Use obd_timeout if calculated value is smaller
296                  * than it. */
297                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
298                         timeout = ptlrpc_inflight_timeout(imp);
299                         timeout += timeout / 3;
300
301                         if (timeout == 0)
302                                 timeout = obd_timeout;
303                 } else {
304                         /* decrease the interval to increase race condition */
305                         timeout = 1;
306                 }
307
308                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
309                        timeout);
310
311                 /* Wait for all requests to error out and call completion
312                  * callbacks. Cap it at obd_timeout -- these should all
313                  * have been locally cancelled by ptlrpc_abort_inflight. */
314                 lwi = LWI_TIMEOUT_INTERVAL(
315                         cfs_timeout_cap(cfs_time_seconds(timeout)),
316                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
317                         NULL, NULL);
318                 rc = l_wait_event(imp->imp_recovery_waitq,
319                                   (cfs_atomic_read(&imp->imp_inflight) == 0),
320                                   &lwi);
321                 if (rc) {
322                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
323
324                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
325                                cli_tgt, rc,
326                                cfs_atomic_read(&imp->imp_inflight));
327
328                         cfs_spin_lock(&imp->imp_lock);
329                         if (cfs_atomic_read(&imp->imp_inflight) == 0) {
330                                 int count = cfs_atomic_read(&imp->imp_unregistering);
331
332                                 /* We know that "unregistering" rpcs only can
333                                  * survive in sending or delaying lists (they
334                                  * maybe waiting for long reply unlink in
335                                  * sluggish nets). Let's check this. If there
336                                  * is no inflight and unregistering != 0, this
337                                  * is bug. */
338                                 LASSERTF(count == 0, "Some RPCs are still "
339                                          "unregistering: %d\n", count);
340
341                                 /* Let's save one loop as soon as inflight have
342                                  * dropped to zero. No new inflights possible at
343                                  * this point. */
344                                 rc = 0;
345                         } else {
346                                 cfs_list_for_each_safe(tmp, n,
347                                                        &imp->imp_sending_list) {
348                                         req = cfs_list_entry(tmp,
349                                                              struct ptlrpc_request,
350                                                              rq_list);
351                                         DEBUG_REQ(D_ERROR, req,
352                                                   "still on sending list");
353                                 }
354                                 cfs_list_for_each_safe(tmp, n,
355                                                        &imp->imp_delayed_list) {
356                                         req = cfs_list_entry(tmp,
357                                                              struct ptlrpc_request,
358                                                              rq_list);
359                                         DEBUG_REQ(D_ERROR, req,
360                                                   "still on delayed list");
361                                 }
362
363                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
364                                        "Network is sluggish? Waiting them "
365                                        "to error out.\n", cli_tgt,
366                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
367                                        cfs_atomic_read(&imp->
368                                                        imp_unregistering));
369                         }
370                         cfs_spin_unlock(&imp->imp_lock);
371                   }
372         } while (rc != 0);
373
374         /*
375          * Let's additionally check that no new rpcs added to import in
376          * "invalidate" state.
377          */
378         LASSERT(cfs_atomic_read(&imp->imp_inflight) == 0);
379         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
380         sptlrpc_import_flush_all_ctx(imp);
381
382         cfs_atomic_dec(&imp->imp_inval_count);
383         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
384 }
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         cfs_spin_lock(&imp->imp_lock);
392         imp->imp_invalid = 0;
393         ptlrpc_activate_timeouts(imp);
394         cfs_spin_unlock(&imp->imp_lock);
395         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
396 }
397
398 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
399 {
400         ENTRY;
401
402         LASSERT(!imp->imp_dlm_fake);
403
404         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
405                 if (!imp->imp_replayable) {
406                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
407                                "auto-deactivating\n",
408                                obd2cli_tgt(imp->imp_obd),
409                                imp->imp_connection->c_remote_uuid.uuid,
410                                imp->imp_obd->obd_name);
411                         ptlrpc_deactivate_import(imp);
412                 }
413
414                 CDEBUG(D_HA, "%s: waking up pinger\n",
415                        obd2cli_tgt(imp->imp_obd));
416
417                 cfs_spin_lock(&imp->imp_lock);
418                 imp->imp_force_verify = 1;
419                 cfs_spin_unlock(&imp->imp_lock);
420
421                 ptlrpc_pinger_wake_up();
422         }
423         EXIT;
424 }
425
426 int ptlrpc_reconnect_import(struct obd_import *imp)
427 {
428         ptlrpc_set_import_discon(imp, 0);
429         /* Force a new connect attempt */
430         ptlrpc_invalidate_import(imp);
431         /* Do a fresh connect next time by zeroing the handle */
432         ptlrpc_disconnect_import(imp, 1);
433         /* Wait for all invalidate calls to finish */
434         if (cfs_atomic_read(&imp->imp_inval_count) > 0) {
435                 int rc;
436                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
437                 rc = l_wait_event(imp->imp_recovery_waitq,
438                                   (cfs_atomic_read(&imp->imp_inval_count) == 0),
439                                   &lwi);
440                 if (rc)
441                         CERROR("Interrupted, inval=%d\n",
442                                cfs_atomic_read(&imp->imp_inval_count));
443         }
444
445         /* Allow reconnect attempts */
446         imp->imp_obd->obd_no_recov = 0;
447         /* Remove 'invalid' flag */
448         ptlrpc_activate_import(imp);
449         /* Attempt a new connect */
450         ptlrpc_recover_import(imp, NULL);
451         return 0;
452 }
453
454 EXPORT_SYMBOL(ptlrpc_reconnect_import);
455
456 /**
457  * Connection on import \a imp is changed to another one (if more than one is
458  * present). We typically chose connection that we have not tried to connect to
459  * the longest
460  */
461 static int import_select_connection(struct obd_import *imp)
462 {
463         struct obd_import_conn *imp_conn = NULL, *conn;
464         struct obd_export *dlmexp;
465         int tried_all = 1;
466         ENTRY;
467
468         cfs_spin_lock(&imp->imp_lock);
469
470         if (cfs_list_empty(&imp->imp_conn_list)) {
471                 CERROR("%s: no connections available\n",
472                         imp->imp_obd->obd_name);
473                 cfs_spin_unlock(&imp->imp_lock);
474                 RETURN(-EINVAL);
475         }
476
477         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
478                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
479                        imp->imp_obd->obd_name,
480                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
481                        conn->oic_last_attempt);
482
483                 /* If we have not tried this connection since
484                    the last successful attempt, go with this one */
485                 if ((conn->oic_last_attempt == 0) ||
486                     cfs_time_beforeq_64(conn->oic_last_attempt,
487                                        imp->imp_last_success_conn)) {
488                         imp_conn = conn;
489                         tried_all = 0;
490                         break;
491                 }
492
493                 /* If all of the connections have already been tried
494                    since the last successful connection; just choose the
495                    least recently used */
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
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         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
517                 if (at_get(&imp->imp_at.iat_net_latency) <
518                     CONNECTION_SWITCH_MAX) {
519                         at_measured(&imp->imp_at.iat_net_latency,
520                                     at_get(&imp->imp_at.iat_net_latency) +
521                                     CONNECTION_SWITCH_INC);
522                 }
523                 LASSERT(imp_conn->oic_last_attempt);
524                 CWARN("%s: tried all connections, increasing latency to %ds\n",
525                       imp->imp_obd->obd_name,
526                       at_get(&imp->imp_at.iat_net_latency));
527         }
528
529         imp_conn->oic_last_attempt = cfs_time_current_64();
530
531         /* switch connection, don't mind if it's same as the current one */
532         if (imp->imp_connection)
533                 ptlrpc_connection_put(imp->imp_connection);
534         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
535
536         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
537         LASSERT(dlmexp != NULL);
538         if (dlmexp->exp_connection)
539                 ptlrpc_connection_put(dlmexp->exp_connection);
540         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
541         class_export_put(dlmexp);
542
543         if (imp->imp_conn_current != imp_conn) {
544                 if (imp->imp_conn_current)
545                         CDEBUG(D_HA, "Changing connection for %s to %s/%s\n",
546                                imp->imp_obd->obd_name, imp_conn->oic_uuid.uuid,
547                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
548                 imp->imp_conn_current = imp_conn;
549         }
550
551         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
552                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
553                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
554
555         cfs_spin_unlock(&imp->imp_lock);
556
557         RETURN(0);
558 }
559
560 /*
561  * must be called under imp_lock
562  */
563 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
564 {
565         struct ptlrpc_request *req;
566         cfs_list_t *tmp;
567
568         if (cfs_list_empty(&imp->imp_replay_list))
569                 return 0;
570         tmp = imp->imp_replay_list.next;
571         req = cfs_list_entry(tmp, struct ptlrpc_request, rq_replay_list);
572         *transno = req->rq_transno;
573         if (req->rq_transno == 0) {
574                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
575                 LBUG();
576         }
577
578         return 1;
579 }
580
581 /**
582  * Attempt to (re)connect import \a imp. This includes all preparations,
583  * initializing CONNECT RPC request and passing it to ptlrpcd for
584  * actual sending.
585  * Returns 0 on success or error code.
586  */
587 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
588 {
589         struct obd_device *obd = imp->imp_obd;
590         int initial_connect = 0;
591         int set_transno = 0;
592         __u64 committed_before_reconnect = 0;
593         struct ptlrpc_request *request;
594         char *bufs[] = { NULL,
595                          obd2cli_tgt(imp->imp_obd),
596                          obd->obd_uuid.uuid,
597                          (char *)&imp->imp_dlm_handle,
598                          (char *)&imp->imp_connect_data };
599         struct ptlrpc_connect_async_args *aa;
600         int rc;
601         ENTRY;
602
603         cfs_spin_lock(&imp->imp_lock);
604         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
605                 cfs_spin_unlock(&imp->imp_lock);
606                 CERROR("can't connect to a closed import\n");
607                 RETURN(-EINVAL);
608         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
609                 cfs_spin_unlock(&imp->imp_lock);
610                 CERROR("already connected\n");
611                 RETURN(0);
612         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
613                 cfs_spin_unlock(&imp->imp_lock);
614                 CERROR("already connecting\n");
615                 RETURN(-EALREADY);
616         }
617
618         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
619
620         imp->imp_conn_cnt++;
621         imp->imp_resend_replay = 0;
622
623         if (!lustre_handle_is_used(&imp->imp_remote_handle))
624                 initial_connect = 1;
625         else
626                 committed_before_reconnect = imp->imp_peer_committed_transno;
627
628         set_transno = ptlrpc_first_transno(imp,
629                                            &imp->imp_connect_data.ocd_transno);
630         cfs_spin_unlock(&imp->imp_lock);
631
632         if (new_uuid) {
633                 struct obd_uuid uuid;
634
635                 obd_str2uuid(&uuid, new_uuid);
636                 rc = import_set_conn_priority(imp, &uuid);
637                 if (rc)
638                         GOTO(out, rc);
639         }
640
641         rc = import_select_connection(imp);
642         if (rc)
643                 GOTO(out, rc);
644
645         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
646         if (rc)
647                 GOTO(out, rc);
648
649         /* Reset connect flags to the originally requested flags, in case
650          * the server is updated on-the-fly we will get the new features. */
651         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
652         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
653         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
654
655         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
656                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
657         if (rc)
658                 GOTO(out, rc);
659
660         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
661         if (request == NULL)
662                 GOTO(out, rc = -ENOMEM);
663
664         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
665                                       imp->imp_connect_op, bufs, NULL);
666         if (rc) {
667                 ptlrpc_request_free(request);
668                 GOTO(out, rc);
669         }
670
671         /* Report the rpc service time to the server so that it knows how long
672          * to wait for clients to join recovery */
673         lustre_msg_set_service_time(request->rq_reqmsg,
674                                     at_timeout2est(request->rq_timeout));
675
676         /* The amount of time we give the server to process the connect req.
677          * import_select_connection will increase the net latency on
678          * repeated reconnect attempts to cover slow networks.
679          * We override/ignore the server rpc completion estimate here,
680          * which may be large if this is a reconnect attempt */
681         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
682         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
683
684 #ifndef __KERNEL__
685         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
686 #endif
687         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
688
689         request->rq_no_resend = request->rq_no_delay = 1;
690         request->rq_send_state = LUSTRE_IMP_CONNECTING;
691         /* Allow a slightly larger reply for future growth compatibility */
692         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
693                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
694         ptlrpc_request_set_replen(request);
695         request->rq_interpret_reply = ptlrpc_connect_interpret;
696
697         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
698         aa = ptlrpc_req_async_args(request);
699         memset(aa, 0, sizeof *aa);
700
701         aa->pcaa_peer_committed = committed_before_reconnect;
702         aa->pcaa_initial_connect = initial_connect;
703
704         if (aa->pcaa_initial_connect) {
705                 cfs_spin_lock(&imp->imp_lock);
706                 imp->imp_replayable = 1;
707                 cfs_spin_unlock(&imp->imp_lock);
708                 lustre_msg_add_op_flags(request->rq_reqmsg,
709                                         MSG_CONNECT_INITIAL);
710         }
711
712         if (set_transno)
713                 lustre_msg_add_op_flags(request->rq_reqmsg,
714                                         MSG_CONNECT_TRANSNO);
715
716         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
717                   request->rq_timeout);
718         ptlrpcd_add_req(request, PSCOPE_OTHER);
719         rc = 0;
720 out:
721         if (rc != 0) {
722                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
723         }
724
725         RETURN(rc);
726 }
727 EXPORT_SYMBOL(ptlrpc_connect_import);
728
729 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
730 {
731 #ifdef __KERNEL__
732         /* the pinger takes care of issuing the next reconnect request */
733         return;
734 #else
735         /* liblustre has no pinger thread, so we wakeup pinger anyway */
736         ptlrpc_pinger_wake_up();
737 #endif
738 }
739
740 static int ptlrpc_busy_reconnect(int rc)
741 {
742         return (rc == -EBUSY) || (rc == -EAGAIN);
743 }
744
745 /**
746  * interpret_reply callback for connect RPCs.
747  * Looks into returned status of connect operation and decides
748  * what to do with the import - i.e enter recovery, promote it to
749  * full state for normal operations of disconnect it due to an error.
750  */
751 static int ptlrpc_connect_interpret(const struct lu_env *env,
752                                     struct ptlrpc_request *request,
753                                     void *data, int rc)
754 {
755         struct ptlrpc_connect_async_args *aa = data;
756         struct obd_import *imp = request->rq_import;
757         struct client_obd *cli = &imp->imp_obd->u.cli;
758         struct lustre_handle old_hdl;
759         __u64 old_connect_flags;
760         int msg_flags;
761         ENTRY;
762
763         cfs_spin_lock(&imp->imp_lock);
764         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
765                 cfs_spin_unlock(&imp->imp_lock);
766                 RETURN(0);
767         }
768
769         if (rc) {
770                 /* if this reconnect to busy export - not need select new target
771                  * for connecting*/
772                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
773                 cfs_spin_unlock(&imp->imp_lock);
774                 GOTO(out, rc);
775         }
776
777         LASSERT(imp->imp_conn_current);
778
779         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
780
781         /* All imports are pingable */
782         imp->imp_pingable = 1;
783         imp->imp_force_reconnect = 0;
784
785         if (aa->pcaa_initial_connect) {
786                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
787                         imp->imp_replayable = 1;
788                         cfs_spin_unlock(&imp->imp_lock);
789                         CDEBUG(D_HA, "connected to replayable target: %s\n",
790                                obd2cli_tgt(imp->imp_obd));
791                 } else {
792                         imp->imp_replayable = 0;
793                         cfs_spin_unlock(&imp->imp_lock);
794                 }
795
796                 /* if applies, adjust the imp->imp_msg_magic here
797                  * according to reply flags */
798
799                 imp->imp_remote_handle =
800                                 *lustre_msg_get_handle(request->rq_repmsg);
801
802                 /* Initial connects are allowed for clients with non-random
803                  * uuids when servers are in recovery.  Simply signal the
804                  * servers replay is complete and wait in REPLAY_WAIT. */
805                 if (msg_flags & MSG_CONNECT_RECOVERING) {
806                         CDEBUG(D_HA, "connect to %s during recovery\n",
807                                obd2cli_tgt(imp->imp_obd));
808                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
809                 } else {
810                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
811                         ptlrpc_activate_import(imp);
812                 }
813
814                 GOTO(finish, rc = 0);
815         } else {
816                 cfs_spin_unlock(&imp->imp_lock);
817         }
818
819         /* Determine what recovery state to move the import to. */
820         if (MSG_CONNECT_RECONNECT & msg_flags) {
821                 memset(&old_hdl, 0, sizeof(old_hdl));
822                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
823                             sizeof (old_hdl))) {
824                         CERROR("%s@%s didn't like our handle "LPX64
825                                ", failed\n", obd2cli_tgt(imp->imp_obd),
826                                imp->imp_connection->c_remote_uuid.uuid,
827                                imp->imp_dlm_handle.cookie);
828                         GOTO(out, rc = -ENOTCONN);
829                 }
830
831                 if (memcmp(&imp->imp_remote_handle,
832                            lustre_msg_get_handle(request->rq_repmsg),
833                            sizeof(imp->imp_remote_handle))) {
834                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
835                                 D_HA : D_WARNING;
836
837                         /* Bug 16611/14775: if server handle have changed,
838                          * that means some sort of disconnection happened.
839                          * If the server is not in recovery, that also means it
840                          * already erased all of our state because of previous
841                          * eviction. If it is in recovery - we are safe to
842                          * participate since we can reestablish all of our state
843                          * with server again */
844                         CDEBUG(level,"%s@%s changed server handle from "
845                                      LPX64" to "LPX64"%s\n",
846                                      obd2cli_tgt(imp->imp_obd),
847                                      imp->imp_connection->c_remote_uuid.uuid,
848                                      imp->imp_remote_handle.cookie,
849                                      lustre_msg_get_handle(request->rq_repmsg)->
850                                                                         cookie,
851                                      (MSG_CONNECT_RECOVERING & msg_flags) ?
852                                          " but is still in recovery" : "");
853
854                         imp->imp_remote_handle =
855                                      *lustre_msg_get_handle(request->rq_repmsg);
856
857                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
858                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
859                                 GOTO(finish, rc = 0);
860                         }
861
862                 } else {
863                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
864                                obd2cli_tgt(imp->imp_obd),
865                                imp->imp_connection->c_remote_uuid.uuid);
866                 }
867
868                 if (imp->imp_invalid) {
869                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
870                                "marking evicted\n", imp->imp_obd->obd_name);
871                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
872                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
873                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
874                                imp->imp_obd->obd_name,
875                                obd2cli_tgt(imp->imp_obd));
876
877                         cfs_spin_lock(&imp->imp_lock);
878                         imp->imp_resend_replay = 1;
879                         cfs_spin_unlock(&imp->imp_lock);
880
881                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
882                 } else {
883                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
884                 }
885         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
886                 LASSERT(imp->imp_replayable);
887                 imp->imp_remote_handle =
888                                 *lustre_msg_get_handle(request->rq_repmsg);
889                 imp->imp_last_replay_transno = 0;
890                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
891         } else {
892                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
893                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
894                 imp->imp_remote_handle =
895                                 *lustre_msg_get_handle(request->rq_repmsg);
896                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
897         }
898
899         /* Sanity checks for a reconnected import. */
900         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
901                 CERROR("imp_replayable flag does not match server "
902                        "after reconnect. We should LBUG right here.\n");
903         }
904
905         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
906             lustre_msg_get_last_committed(request->rq_repmsg) <
907             aa->pcaa_peer_committed) {
908                 CERROR("%s went back in time (transno "LPD64
909                        " was previously committed, server now claims "LPD64
910                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
911                        "id=9646\n",
912                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
913                        lustre_msg_get_last_committed(request->rq_repmsg));
914         }
915
916 finish:
917         rc = ptlrpc_import_recovery_state_machine(imp);
918         if (rc != 0) {
919                 if (rc == -ENOTCONN) {
920                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
921                                "invalidating and reconnecting\n",
922                                obd2cli_tgt(imp->imp_obd),
923                                imp->imp_connection->c_remote_uuid.uuid);
924                         ptlrpc_connect_import(imp, NULL);
925                         RETURN(0);
926                 }
927         } else {
928                 struct obd_connect_data *ocd;
929                 struct obd_export *exp;
930                 int ret;
931                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
932                                            RCL_SERVER);
933                 /* server replied obd_connect_data is always bigger */
934                 ocd = req_capsule_server_sized_get(&request->rq_pill,
935                                                    &RMF_CONNECT_DATA, ret);
936
937                 cfs_spin_lock(&imp->imp_lock);
938                 cfs_list_del(&imp->imp_conn_current->oic_item);
939                 cfs_list_add(&imp->imp_conn_current->oic_item,
940                              &imp->imp_conn_list);
941                 imp->imp_last_success_conn =
942                         imp->imp_conn_current->oic_last_attempt;
943
944                 if (ocd == NULL) {
945                         cfs_spin_unlock(&imp->imp_lock);
946                         CERROR("Wrong connect data from server\n");
947                         rc = -EPROTO;
948                         GOTO(out, rc);
949                 }
950
951                 imp->imp_connect_data = *ocd;
952
953                 exp = class_conn2export(&imp->imp_dlm_handle);
954                 cfs_spin_unlock(&imp->imp_lock);
955
956                 /* check that server granted subset of flags we asked for. */
957                 LASSERTF((ocd->ocd_connect_flags &
958                           imp->imp_connect_flags_orig) ==
959                          ocd->ocd_connect_flags, LPX64" != "LPX64,
960                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
961
962                 if (!exp) {
963                         /* This could happen if export is cleaned during the
964                            connect attempt */
965                         CERROR("Missing export for %s\n",
966                                imp->imp_obd->obd_name);
967                         GOTO(out, rc = -ENODEV);
968                 }
969                 old_connect_flags = exp->exp_connect_flags;
970                 exp->exp_connect_flags = ocd->ocd_connect_flags;
971                 imp->imp_obd->obd_self_export->exp_connect_flags =
972                                                         ocd->ocd_connect_flags;
973                 class_export_put(exp);
974
975                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
976
977                 if (!ocd->ocd_ibits_known &&
978                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
979                         CERROR("Inodebits aware server returned zero compatible"
980                                " bits?\n");
981
982                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
983                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
984                                         LUSTRE_VERSION_OFFSET_WARN ||
985                      ocd->ocd_version < LUSTRE_VERSION_CODE -
986                                         LUSTRE_VERSION_OFFSET_WARN)) {
987                         /* Sigh, some compilers do not like #ifdef in the middle
988                            of macro arguments */
989 #ifdef __KERNEL__
990                         const char *older = "older. Consider upgrading server "
991                                             "or downgrading client";
992 #else
993                         const char *older = "older. Consider recompiling this "
994                                             "application";
995 #endif
996                         const char *newer = "newer than client version. "
997                                             "Consider upgrading client";
998
999                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1000                                       "is much %s (%s)\n",
1001                                       obd2cli_tgt(imp->imp_obd),
1002                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1003                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1004                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1005                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1006                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1007                                       newer : older, LUSTRE_VERSION_STRING);
1008                 }
1009
1010                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1011                         /* We sent to the server ocd_cksum_types with bits set
1012                          * for algorithms we understand. The server masked off
1013                          * the checksum types it doesn't support */
1014                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
1015                                 LCONSOLE_WARN("The negotiation of the checksum "
1016                                               "alogrithm to use with server %s "
1017                                               "failed (%x/%x), disabling "
1018                                               "checksums\n",
1019                                               obd2cli_tgt(imp->imp_obd),
1020                                               ocd->ocd_cksum_types,
1021                                               OBD_CKSUM_ALL);
1022                                 cli->cl_checksum = 0;
1023                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1024                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
1025                         } else {
1026                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1027
1028                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
1029                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
1030                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
1031                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
1032                                 else
1033                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
1034                         }
1035                 } else {
1036                         /* The server does not support OBD_CONNECT_CKSUM.
1037                          * Enforce CRC32 for backward compatibility*/
1038                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1039                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
1040                 }
1041
1042                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1043                         cli->cl_max_pages_per_rpc =
1044                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
1045                 else if (imp->imp_connect_op == MDS_CONNECT ||
1046                          imp->imp_connect_op == MGS_CONNECT)
1047                         cli->cl_max_pages_per_rpc = 1;
1048
1049                 /* Reset ns_connect_flags only for initial connect. It might be
1050                  * changed in while using FS and if we reset it in reconnect
1051                  * this leads to losing user settings done before such as
1052                  * disable lru_resize, etc. */
1053                 if (old_connect_flags != exp->exp_connect_flags ||
1054                     aa->pcaa_initial_connect) {
1055                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1056                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1057                               ocd->ocd_connect_flags);
1058                         imp->imp_obd->obd_namespace->ns_connect_flags =
1059                                 ocd->ocd_connect_flags;
1060                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1061                                 ocd->ocd_connect_flags;
1062                 }
1063
1064                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1065                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1066                         /* We need a per-message support flag, because
1067                            a. we don't know if the incoming connect reply
1068                               supports AT or not (in reply_in_callback)
1069                               until we unpack it.
1070                            b. failovered server means export and flags are gone
1071                               (in ptlrpc_send_reply).
1072                            Can only be set when we know AT is supported at
1073                            both ends */
1074                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1075                 else
1076                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1077
1078                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1079                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1080                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1081                 else
1082                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1083
1084                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1085                         (cli->cl_max_pages_per_rpc > 0));
1086         }
1087
1088 out:
1089         if (rc != 0) {
1090                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1091                 if (rc == -EACCES) {
1092                         /*
1093                          * Give up trying to reconnect
1094                          * EACCES means client has no permission for connection
1095                          */
1096                         imp->imp_obd->obd_no_recov = 1;
1097                         ptlrpc_deactivate_import(imp);
1098                 }
1099
1100                 if (rc == -EPROTO) {
1101                         struct obd_connect_data *ocd;
1102
1103                         /* reply message might not be ready */
1104                         if (request->rq_repmsg == NULL)
1105                                 RETURN(-EPROTO);
1106
1107                         ocd = req_capsule_server_get(&request->rq_pill,
1108                                                      &RMF_CONNECT_DATA);
1109                         if (ocd &&
1110                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1111                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1112                            /* Actually servers are only supposed to refuse
1113                               connection from liblustre clients, so we should
1114                               never see this from VFS context */
1115                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1116                                         "(%d.%d.%d.%d)"
1117                                         " refused connection from this client "
1118                                         "with an incompatible version (%s).  "
1119                                         "Client must be recompiled\n",
1120                                         obd2cli_tgt(imp->imp_obd),
1121                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1122                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1123                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1124                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1125                                         LUSTRE_VERSION_STRING);
1126                                 ptlrpc_deactivate_import(imp);
1127                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1128                         }
1129                         RETURN(-EPROTO);
1130                 }
1131
1132                 ptlrpc_maybe_ping_import_soon(imp);
1133
1134                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1135                        obd2cli_tgt(imp->imp_obd),
1136                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1137         }
1138
1139         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1140         RETURN(rc);
1141 }
1142
1143 /**
1144  * interpret callback for "completed replay" RPCs.
1145  * \see signal_completed_replay
1146  */
1147 static int completed_replay_interpret(const struct lu_env *env,
1148                                       struct ptlrpc_request *req,
1149                                       void * data, int rc)
1150 {
1151         ENTRY;
1152         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1153         if (req->rq_status == 0 &&
1154             !req->rq_import->imp_vbr_failed) {
1155                 ptlrpc_import_recovery_state_machine(req->rq_import);
1156         } else {
1157                 if (req->rq_import->imp_vbr_failed) {
1158                         CDEBUG(D_WARNING,
1159                                "%s: version recovery fails, reconnecting\n",
1160                                req->rq_import->imp_obd->obd_name);
1161                 } else {
1162                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1163                                      "reconnecting\n",
1164                                req->rq_import->imp_obd->obd_name,
1165                                req->rq_status);
1166                 }
1167                 ptlrpc_connect_import(req->rq_import, NULL);
1168         }
1169
1170         RETURN(0);
1171 }
1172
1173 /**
1174  * Let server know that we have no requests to replay anymore.
1175  * Achieved by just sending a PING request
1176  */
1177 static int signal_completed_replay(struct obd_import *imp)
1178 {
1179         struct ptlrpc_request *req;
1180         ENTRY;
1181
1182         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1183                 RETURN(0);
1184
1185         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
1186         cfs_atomic_inc(&imp->imp_replay_inflight);
1187
1188         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1189                                         OBD_PING);
1190         if (req == NULL) {
1191                 cfs_atomic_dec(&imp->imp_replay_inflight);
1192                 RETURN(-ENOMEM);
1193         }
1194
1195         ptlrpc_request_set_replen(req);
1196         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1197         lustre_msg_add_flags(req->rq_reqmsg,
1198                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1199         if (AT_OFF)
1200                 req->rq_timeout *= 3;
1201         req->rq_interpret_reply = completed_replay_interpret;
1202
1203         ptlrpcd_add_req(req, PSCOPE_OTHER);
1204         RETURN(0);
1205 }
1206
1207 #ifdef __KERNEL__
1208 /**
1209  * In kernel code all import invalidation happens in its own
1210  * separate thread, so that whatever application happened to encounter
1211  * a problem could still be killed or otherwise continue
1212  */
1213 static int ptlrpc_invalidate_import_thread(void *data)
1214 {
1215         struct obd_import *imp = data;
1216
1217         ENTRY;
1218
1219         cfs_daemonize_ctxt("ll_imp_inval");
1220
1221         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1222                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1223                imp->imp_connection->c_remote_uuid.uuid);
1224
1225         ptlrpc_invalidate_import(imp);
1226
1227         if (obd_dump_on_eviction) {
1228                 CERROR("dump the log upon eviction\n");
1229                 libcfs_debug_dumplog();
1230         }
1231
1232         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1233         ptlrpc_import_recovery_state_machine(imp);
1234
1235         class_import_put(imp);
1236         RETURN(0);
1237 }
1238 #endif
1239
1240 /**
1241  * This is the state machine for client-side recovery on import.
1242  *
1243  * Typicaly we have two possibly paths. If we came to server and it is not
1244  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1245  * state and reconnect from scratch.
1246  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1247  * We go through our list of requests to replay and send them to server one by
1248  * one.
1249  * After sending all request from the list we change import state to
1250  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1251  * and also all the locks we don't yet have and wait for server to grant us.
1252  * After that we send a special "replay completed" request and change import
1253  * state to IMP_REPLAY_WAIT.
1254  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1255  * state and resend all requests from sending list.
1256  * After that we promote import to FULL state and send all delayed requests
1257  * and import is fully operational after that.
1258  *
1259  */
1260 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1261 {
1262         int rc = 0;
1263         int inflight;
1264         char *target_start;
1265         int target_len;
1266
1267         ENTRY;
1268         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1269                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1270                           &target_start, &target_len);
1271                 /* Don't care about MGC eviction */
1272                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1273                            LUSTRE_MGC_NAME) != 0) {
1274                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1275                                            "%.*s; in progress operations using "
1276                                            "this service will fail.\n",
1277                                            target_len, target_start);
1278                 }
1279                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1280                        obd2cli_tgt(imp->imp_obd),
1281                        imp->imp_connection->c_remote_uuid.uuid);
1282                 /* reset vbr_failed flag upon eviction */
1283                 cfs_spin_lock(&imp->imp_lock);
1284                 imp->imp_vbr_failed = 0;
1285                 cfs_spin_unlock(&imp->imp_lock);
1286
1287 #ifdef __KERNEL__
1288                 /* bug 17802:  XXX client_disconnect_export vs connect request
1289                  * race. if client will evicted at this time, we start
1290                  * invalidate thread without reference to import and import can
1291                  * be freed at same time. */
1292                 class_import_get(imp);
1293                 rc = cfs_create_thread(ptlrpc_invalidate_import_thread, imp,
1294                                        CFS_DAEMON_FLAGS);
1295                 if (rc < 0) {
1296                         class_import_put(imp);
1297                         CERROR("error starting invalidate thread: %d\n", rc);
1298                 } else {
1299                         rc = 0;
1300                 }
1301                 RETURN(rc);
1302 #else
1303                 ptlrpc_invalidate_import(imp);
1304
1305                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1306 #endif
1307         }
1308
1309         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1310                 CDEBUG(D_HA, "replay requested by %s\n",
1311                        obd2cli_tgt(imp->imp_obd));
1312                 rc = ptlrpc_replay_next(imp, &inflight);
1313                 if (inflight == 0 &&
1314                     cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1315                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1316                         rc = ldlm_replay_locks(imp);
1317                         if (rc)
1318                                 GOTO(out, rc);
1319                 }
1320                 rc = 0;
1321         }
1322
1323         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1324                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1325                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1326                         rc = signal_completed_replay(imp);
1327                         if (rc)
1328                                 GOTO(out, rc);
1329                 }
1330
1331         }
1332
1333         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1334                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1335                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1336                 }
1337         }
1338
1339         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1340                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1341                        obd2cli_tgt(imp->imp_obd),
1342                        imp->imp_connection->c_remote_uuid.uuid);
1343
1344                 rc = ptlrpc_resend(imp);
1345                 if (rc)
1346                         GOTO(out, rc);
1347                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1348                 ptlrpc_activate_import(imp);
1349
1350                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1351                           &target_start, &target_len);
1352                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1353                               "using nid %s.\n", imp->imp_obd->obd_name,
1354                               target_len, target_start,
1355                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1356         }
1357
1358         if (imp->imp_state == LUSTRE_IMP_FULL) {
1359                 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1360                 ptlrpc_wake_delayed(imp);
1361         }
1362
1363 out:
1364         RETURN(rc);
1365 }
1366
1367 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1368 {
1369         struct ptlrpc_request *req;
1370         int rq_opc, rc = 0;
1371         int nowait = imp->imp_obd->obd_force;
1372         ENTRY;
1373
1374         if (nowait)
1375                 GOTO(set_state, rc);
1376
1377         switch (imp->imp_connect_op) {
1378         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1379         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1380         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1381         default:
1382                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1383                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1384                 RETURN(-EINVAL);
1385         }
1386
1387         if (ptlrpc_import_in_recovery(imp)) {
1388                 struct l_wait_info lwi;
1389                 cfs_duration_t timeout;
1390
1391
1392                 if (AT_OFF) {
1393                         if (imp->imp_server_timeout)
1394                                 timeout = cfs_time_seconds(obd_timeout / 2);
1395                         else
1396                                 timeout = cfs_time_seconds(obd_timeout);
1397                 } else {
1398                         int idx = import_at_get_index(imp,
1399                                 imp->imp_client->cli_request_portal);
1400                         timeout = cfs_time_seconds(
1401                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1402                 }
1403
1404                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1405                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1406                 rc = l_wait_event(imp->imp_recovery_waitq,
1407                                   !ptlrpc_import_in_recovery(imp), &lwi);
1408
1409         }
1410
1411         cfs_spin_lock(&imp->imp_lock);
1412         if (imp->imp_state != LUSTRE_IMP_FULL)
1413                 GOTO(out, 0);
1414
1415         cfs_spin_unlock(&imp->imp_lock);
1416
1417         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1418                                         LUSTRE_OBD_VERSION, rq_opc);
1419         if (req) {
1420                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1421                  * it fails.  We can get through the above with a down server
1422                  * if the client doesn't know the server is gone yet. */
1423                 req->rq_no_resend = 1;
1424
1425 #ifndef CRAY_XT3
1426                 /* We want client umounts to happen quickly, no matter the
1427                    server state... */
1428                 req->rq_timeout = min_t(int, req->rq_timeout,
1429                                         INITIAL_CONNECT_TIMEOUT);
1430 #else
1431                 /* ... but we always want liblustre clients to nicely
1432                    disconnect, so only use the adaptive value. */
1433                 if (AT_OFF)
1434                         req->rq_timeout = obd_timeout / 3;
1435 #endif
1436
1437                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1438                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1439                 ptlrpc_request_set_replen(req);
1440                 rc = ptlrpc_queue_wait(req);
1441                 ptlrpc_req_finished(req);
1442         }
1443
1444 set_state:
1445         cfs_spin_lock(&imp->imp_lock);
1446 out:
1447         if (noclose)
1448                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1449         else
1450                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1451         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1452         cfs_spin_unlock(&imp->imp_lock);
1453
1454         RETURN(rc);
1455 }
1456
1457 void ptlrpc_cleanup_imp(struct obd_import *imp)
1458 {
1459         ENTRY;
1460
1461         cfs_spin_lock(&imp->imp_lock);
1462         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1463         imp->imp_generation++;
1464         cfs_spin_unlock(&imp->imp_lock);
1465         ptlrpc_abort_inflight(imp);
1466
1467         EXIT;
1468 }
1469
1470 /* Adaptive Timeout utils */
1471 extern unsigned int at_min, at_max, at_history;
1472
1473 /* Bin into timeslices using AT_BINS bins.
1474    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1475    but still smoothing out a return to normalcy from a slow response.
1476    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1477 int at_measured(struct adaptive_timeout *at, unsigned int val)
1478 {
1479         unsigned int old = at->at_current;
1480         time_t now = cfs_time_current_sec();
1481         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1482
1483         LASSERT(at);
1484         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1485                val, at, now - at->at_binstart, at->at_current,
1486                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1487
1488         if (val == 0)
1489                 /* 0's don't count, because we never want our timeout to
1490                    drop to 0, and because 0 could mean an error */
1491                 return 0;
1492
1493         cfs_spin_lock(&at->at_lock);
1494
1495         if (unlikely(at->at_binstart == 0)) {
1496                 /* Special case to remove default from history */
1497                 at->at_current = val;
1498                 at->at_worst_ever = val;
1499                 at->at_worst_time = now;
1500                 at->at_hist[0] = val;
1501                 at->at_binstart = now;
1502         } else if (now - at->at_binstart < binlimit ) {
1503                 /* in bin 0 */
1504                 at->at_hist[0] = max(val, at->at_hist[0]);
1505                 at->at_current = max(val, at->at_current);
1506         } else {
1507                 int i, shift;
1508                 unsigned int maxv = val;
1509                 /* move bins over */
1510                 shift = (now - at->at_binstart) / binlimit;
1511                 LASSERT(shift > 0);
1512                 for(i = AT_BINS - 1; i >= 0; i--) {
1513                         if (i >= shift) {
1514                                 at->at_hist[i] = at->at_hist[i - shift];
1515                                 maxv = max(maxv, at->at_hist[i]);
1516                         } else {
1517                                 at->at_hist[i] = 0;
1518                         }
1519                 }
1520                 at->at_hist[0] = val;
1521                 at->at_current = maxv;
1522                 at->at_binstart += shift * binlimit;
1523         }
1524
1525         if (at->at_current > at->at_worst_ever) {
1526                 at->at_worst_ever = at->at_current;
1527                 at->at_worst_time = now;
1528         }
1529
1530         if (at->at_flags & AT_FLG_NOHIST)
1531                 /* Only keep last reported val; keeping the rest of the history
1532                    for proc only */
1533                 at->at_current = val;
1534
1535         if (at_max > 0)
1536                 at->at_current =  min(at->at_current, at_max);
1537         at->at_current =  max(at->at_current, at_min);
1538
1539         if (at->at_current != old)
1540                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1541                        "(val=%u) hist %u %u %u %u\n", at,
1542                        old, at->at_current, at->at_current - old, val,
1543                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1544                        at->at_hist[3]);
1545
1546         /* if we changed, report the old value */
1547         old = (at->at_current != old) ? old : 0;
1548
1549         cfs_spin_unlock(&at->at_lock);
1550         return old;
1551 }
1552
1553 /* Find the imp_at index for a given portal; assign if space available */
1554 int import_at_get_index(struct obd_import *imp, int portal)
1555 {
1556         struct imp_at *at = &imp->imp_at;
1557         int i;
1558
1559         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1560                 if (at->iat_portal[i] == portal)
1561                         return i;
1562                 if (at->iat_portal[i] == 0)
1563                         /* unused */
1564                         break;
1565         }
1566
1567         /* Not found in list, add it under a lock */
1568         cfs_spin_lock(&imp->imp_lock);
1569
1570         /* Check unused under lock */
1571         for (; i < IMP_AT_MAX_PORTALS; i++) {
1572                 if (at->iat_portal[i] == portal)
1573                         goto out;
1574                 if (at->iat_portal[i] == 0)
1575                         /* unused */
1576                         break;
1577         }
1578
1579         /* Not enough portals? */
1580         LASSERT(i < IMP_AT_MAX_PORTALS);
1581
1582         at->iat_portal[i] = portal;
1583 out:
1584         cfs_spin_unlock(&imp->imp_lock);
1585         return i;
1586 }