Whamcloud - gitweb
LU-5569 ptlrpc: change reverse import life cycle
[fs/lustre-release.git] / lustre / ptlrpc / import.c
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  *
30  * Copyright (c) 2011, 2014, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/ptlrpc/import.c
37  *
38  * Author: Mike Shaver <shaver@clusterfs.com>
39  */
40
41 #define DEBUG_SUBSYSTEM S_RPC
42
43 #include <linux/kthread.h>
44 #include <obd_support.h>
45 #include <lustre_ha.h>
46 #include <lustre_net.h>
47 #include <lustre_import.h>
48 #include <lustre_export.h>
49 #include <obd.h>
50 #include <obd_cksum.h>
51 #include <obd_class.h>
52
53 #include "ptlrpc_internal.h"
54
55 struct ptlrpc_connect_async_args {
56          __u64 pcaa_peer_committed;
57         int pcaa_initial_connect;
58 };
59
60 /**
61  * Updates import \a imp current state to provided \a state value
62  * Helper function. Must be called under imp_lock.
63  */
64 static void __import_set_state(struct obd_import *imp,
65                                enum lustre_imp_state state)
66 {
67         switch (state) {
68         case LUSTRE_IMP_CLOSED:
69         case LUSTRE_IMP_NEW:
70         case LUSTRE_IMP_DISCON:
71         case LUSTRE_IMP_CONNECTING:
72                 break;
73         case LUSTRE_IMP_REPLAY_WAIT:
74                 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
75                 break;
76         default:
77                 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
78         }
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 =
82                 cfs_time_current_sec();
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. */
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         }                                                                      \
97 } while(0)
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);                                    \
104 } while(0)
105
106 void ptlrpc_import_enter_resend(struct obd_import *imp)
107 {
108         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
109 }
110 EXPORT_SYMBOL(ptlrpc_import_enter_resend);
111
112
113 static int ptlrpc_connect_interpret(const struct lu_env *env,
114                                     struct ptlrpc_request *request,
115                                     void * data, int rc);
116 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
117
118 /* Only this function is allowed to change the import state when it is
119  * CLOSED. I would rather refcount the import and free it after
120  * disconnection like we do with exports. To do that, the client_obd
121  * will need to save the peer info somewhere other than in the import,
122  * though. */
123 int ptlrpc_init_import(struct obd_import *imp)
124 {
125         spin_lock(&imp->imp_lock);
126
127         imp->imp_generation++;
128         imp->imp_state =  LUSTRE_IMP_NEW;
129
130         spin_unlock(&imp->imp_lock);
131
132         return 0;
133 }
134 EXPORT_SYMBOL(ptlrpc_init_import);
135
136 #define UUID_STR "_UUID"
137 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
138 {
139         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
140                 ? uuid : uuid + strlen(prefix);
141
142         *uuid_len = strlen(*uuid_start);
143
144         if (*uuid_len < strlen(UUID_STR))
145                 return;
146
147         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
148                     UUID_STR, strlen(UUID_STR)))
149                 *uuid_len -= strlen(UUID_STR);
150 }
151
152 /**
153  * Returns true if import was FULL, false if import was already not
154  * connected.
155  * @imp - import to be disconnected
156  * @conn_cnt - connection count (epoch) of the request that timed out
157  *             and caused the disconnection.  In some cases, multiple
158  *             inflight requests can fail to a single target (e.g. OST
159  *             bulk requests) and if one has already caused a reconnection
160  *             (increasing the import->conn_cnt) the older failure should
161  *             not also cause a reconnection.  If zero it forces a reconnect.
162  */
163 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
164 {
165         int rc = 0;
166
167         spin_lock(&imp->imp_lock);
168
169         if (imp->imp_state == LUSTRE_IMP_FULL &&
170             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
171                 char *target_start;
172                 int   target_len;
173
174                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
175                           &target_start, &target_len);
176
177                 if (imp->imp_replayable) {
178                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
179                                "lost; in progress operations using this "
180                                "service will wait for recovery to complete\n",
181                                imp->imp_obd->obd_name, target_len, target_start,
182                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
183                 } else {
184                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
185                                "%.*s (at %s) was lost; in progress "
186                                "operations using this service will fail\n",
187                                imp->imp_obd->obd_name,
188                                target_len, target_start,
189                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
190                 }
191                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
192                 spin_unlock(&imp->imp_lock);
193
194                 if (obd_dump_on_timeout)
195                         libcfs_debug_dumplog();
196
197                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
198                 rc = 1;
199         } else {
200                 spin_unlock(&imp->imp_lock);
201                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
202                        imp->imp_client->cli_name, imp,
203                        (imp->imp_state == LUSTRE_IMP_FULL &&
204                         imp->imp_conn_cnt > conn_cnt) ?
205                        "reconnected" : "not connected", imp->imp_conn_cnt,
206                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
207         }
208
209         return rc;
210 }
211
212 /* Must be called with imp_lock held! */
213 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
214 {
215         ENTRY;
216         assert_spin_locked(&imp->imp_lock);
217
218         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
219         imp->imp_invalid = 1;
220         imp->imp_generation++;
221         spin_unlock(&imp->imp_lock);
222
223         ptlrpc_abort_inflight(imp);
224         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
225
226         EXIT;
227 }
228
229 /*
230  * This acts as a barrier; all existing requests are rejected, and
231  * no new requests will be accepted until the import is valid again.
232  */
233 void ptlrpc_deactivate_import(struct obd_import *imp)
234 {
235         spin_lock(&imp->imp_lock);
236         ptlrpc_deactivate_and_unlock_import(imp);
237 }
238 EXPORT_SYMBOL(ptlrpc_deactivate_import);
239
240 static unsigned int
241 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
242 {
243         long dl;
244
245         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
246               (req->rq_phase == RQ_PHASE_BULK) ||
247               (req->rq_phase == RQ_PHASE_NEW)))
248                 return 0;
249
250         if (req->rq_timedout)
251                 return 0;
252
253         if (req->rq_phase == RQ_PHASE_NEW)
254                 dl = req->rq_sent;
255         else
256                 dl = req->rq_deadline;
257
258         if (dl <= now)
259                 return 0;
260
261         return dl - now;
262 }
263
264 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
265 {
266         time_t now = cfs_time_current_sec();
267         struct list_head *tmp, *n;
268         struct ptlrpc_request *req;
269         unsigned int timeout = 0;
270
271         spin_lock(&imp->imp_lock);
272         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
273                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
274                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
275         }
276         spin_unlock(&imp->imp_lock);
277         return timeout;
278 }
279
280 /**
281  * This function will invalidate the import, if necessary, then block
282  * for all the RPC completions, and finally notify the obd to
283  * invalidate its state (ie cancel locks, clear pending requests,
284  * etc).
285  */
286 void ptlrpc_invalidate_import(struct obd_import *imp)
287 {
288         struct list_head *tmp, *n;
289         struct ptlrpc_request *req;
290         struct l_wait_info lwi;
291         unsigned int timeout;
292         int rc;
293
294         atomic_inc(&imp->imp_inval_count);
295
296         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
297                 ptlrpc_deactivate_import(imp);
298
299         CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
300         LASSERT(imp->imp_invalid);
301
302         /* Wait forever until inflight == 0. We really can't do it another
303          * way because in some cases we need to wait for very long reply
304          * unlink. We can't do anything before that because there is really
305          * no guarantee that some rdma transfer is not in progress right now. */
306         do {
307                 /* Calculate max timeout for waiting on rpcs to error
308                  * out. Use obd_timeout if calculated value is smaller
309                  * than it. */
310                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
311                         timeout = ptlrpc_inflight_timeout(imp);
312                         timeout += timeout / 3;
313
314                         if (timeout == 0)
315                                 timeout = obd_timeout;
316                 } else {
317                         /* decrease the interval to increase race condition */
318                         timeout = 1;
319                 }
320
321                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
322                        timeout);
323
324                 /* Wait for all requests to error out and call completion
325                  * callbacks. Cap it at obd_timeout -- these should all
326                  * have been locally cancelled by ptlrpc_abort_inflight. */
327                 lwi = LWI_TIMEOUT_INTERVAL(
328                         cfs_timeout_cap(cfs_time_seconds(timeout)),
329                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
330                         NULL, NULL);
331                 rc = l_wait_event(imp->imp_recovery_waitq,
332                                   (atomic_read(&imp->imp_inflight) == 0),
333                                   &lwi);
334                 if (rc) {
335                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
336
337                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
338                                cli_tgt, rc, atomic_read(&imp->imp_inflight));
339
340                         spin_lock(&imp->imp_lock);
341                         if (atomic_read(&imp->imp_inflight) == 0) {
342                                 int count = atomic_read(&imp->imp_unregistering);
343
344                                 /* We know that "unregistering" rpcs only can
345                                  * survive in sending or delaying lists (they
346                                  * maybe waiting for long reply unlink in
347                                  * sluggish nets). Let's check this. If there
348                                  * is no inflight and unregistering != 0, this
349                                  * is bug. */
350                                 LASSERTF(count == 0, "Some RPCs are still "
351                                          "unregistering: %d\n", count);
352
353                                 /* Let's save one loop as soon as inflight have
354                                  * dropped to zero. No new inflights possible at
355                                  * this point. */
356                                 rc = 0;
357                         } else {
358                                 list_for_each_safe(tmp, n,
359                                                    &imp->imp_sending_list) {
360                                         req = list_entry(tmp,
361                                                          struct ptlrpc_request,
362                                                          rq_list);
363                                         DEBUG_REQ(D_ERROR, req,
364                                                   "still on sending list");
365                                 }
366                                 list_for_each_safe(tmp, n,
367                                                    &imp->imp_delayed_list) {
368                                         req = list_entry(tmp,
369                                                          struct ptlrpc_request,
370                                                          rq_list);
371                                         DEBUG_REQ(D_ERROR, req,
372                                                   "still on delayed list");
373                                 }
374
375                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
376                                        "Network is sluggish? Waiting them "
377                                        "to error out.\n", cli_tgt,
378                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
379                                        atomic_read(&imp->imp_unregistering));
380                         }
381                         spin_unlock(&imp->imp_lock);
382                 }
383         } while (rc != 0);
384
385         /*
386          * Let's additionally check that no new rpcs added to import in
387          * "invalidate" state.
388          */
389         LASSERT(atomic_read(&imp->imp_inflight) == 0);
390         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
391         sptlrpc_import_flush_all_ctx(imp);
392
393         atomic_dec(&imp->imp_inval_count);
394         wake_up_all(&imp->imp_recovery_waitq);
395 }
396 EXPORT_SYMBOL(ptlrpc_invalidate_import);
397
398 /* unset imp_invalid */
399 void ptlrpc_activate_import(struct obd_import *imp)
400 {
401         struct obd_device *obd = imp->imp_obd;
402
403         spin_lock(&imp->imp_lock);
404         if (imp->imp_deactive != 0) {
405                 spin_unlock(&imp->imp_lock);
406                 return;
407         }
408
409         imp->imp_invalid = 0;
410         spin_unlock(&imp->imp_lock);
411         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
412 }
413 EXPORT_SYMBOL(ptlrpc_activate_import);
414
415 void ptlrpc_pinger_force(struct obd_import *imp)
416 {
417         CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
418                ptlrpc_import_state_name(imp->imp_state));
419
420         spin_lock(&imp->imp_lock);
421         imp->imp_force_verify = 1;
422         spin_unlock(&imp->imp_lock);
423
424         if (imp->imp_state != LUSTRE_IMP_CONNECTING)
425                 ptlrpc_pinger_wake_up();
426 }
427 EXPORT_SYMBOL(ptlrpc_pinger_force);
428
429 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
430 {
431         ENTRY;
432
433         LASSERT(!imp->imp_dlm_fake);
434
435         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
436                 if (!imp->imp_replayable) {
437                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
438                                "auto-deactivating\n",
439                                obd2cli_tgt(imp->imp_obd),
440                                imp->imp_connection->c_remote_uuid.uuid,
441                                imp->imp_obd->obd_name);
442                         ptlrpc_deactivate_import(imp);
443                 }
444
445                 ptlrpc_pinger_force(imp);
446         }
447         EXIT;
448 }
449
450 int ptlrpc_reconnect_import(struct obd_import *imp)
451 {
452 #ifdef ENABLE_PINGER
453         struct l_wait_info lwi;
454         int secs = cfs_time_seconds(obd_timeout);
455         int rc;
456
457         ptlrpc_pinger_force(imp);
458
459         CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
460                obd2cli_tgt(imp->imp_obd), secs);
461
462         lwi = LWI_TIMEOUT(secs, NULL, NULL);
463         rc = l_wait_event(imp->imp_recovery_waitq,
464                           !ptlrpc_import_in_recovery(imp), &lwi);
465         CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
466                ptlrpc_import_state_name(imp->imp_state));
467         return rc;
468 #else
469         ptlrpc_set_import_discon(imp, 0);
470         /* Force a new connect attempt */
471         ptlrpc_invalidate_import(imp);
472         /* Do a fresh connect next time by zeroing the handle */
473         ptlrpc_disconnect_import(imp, 1);
474         /* Wait for all invalidate calls to finish */
475         if (atomic_read(&imp->imp_inval_count) > 0) {
476                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
477                 int rc;
478
479                 rc = l_wait_event(imp->imp_recovery_waitq,
480                                   (atomic_read(&imp->imp_inval_count) == 0),
481                                   &lwi);
482                 if (rc)
483                         CERROR("Interrupted, inval=%d\n",
484                                atomic_read(&imp->imp_inval_count));
485         }
486
487         /* Allow reconnect attempts */
488         imp->imp_obd->obd_no_recov = 0;
489         /* Remove 'invalid' flag */
490         ptlrpc_activate_import(imp);
491         /* Attempt a new connect */
492         ptlrpc_recover_import(imp, NULL, 0);
493         return 0;
494 #endif
495 }
496 EXPORT_SYMBOL(ptlrpc_reconnect_import);
497
498 /**
499  * Connection on import \a imp is changed to another one (if more than one is
500  * present). We typically chose connection that we have not tried to connect to
501  * the longest
502  */
503 static int import_select_connection(struct obd_import *imp)
504 {
505         struct obd_import_conn *imp_conn = NULL, *conn;
506         struct obd_export *dlmexp;
507         char *target_start;
508         int target_len, tried_all = 1;
509         ENTRY;
510
511         spin_lock(&imp->imp_lock);
512
513         if (list_empty(&imp->imp_conn_list)) {
514                 CERROR("%s: no connections available\n",
515                        imp->imp_obd->obd_name);
516                 spin_unlock(&imp->imp_lock);
517                 RETURN(-EINVAL);
518         }
519
520         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
521                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
522                        imp->imp_obd->obd_name,
523                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
524                        conn->oic_last_attempt);
525
526                 /* If we have not tried this connection since
527                    the last successful attempt, go with this one */
528                 if ((conn->oic_last_attempt == 0) ||
529                     cfs_time_beforeq_64(conn->oic_last_attempt,
530                                        imp->imp_last_success_conn)) {
531                         imp_conn = conn;
532                         tried_all = 0;
533                         break;
534                 }
535
536                 /* If all of the connections have already been tried
537                    since the last successful connection; just choose the
538                    least recently used */
539                 if (!imp_conn)
540                         imp_conn = conn;
541                 else if (cfs_time_before_64(conn->oic_last_attempt,
542                                             imp_conn->oic_last_attempt))
543                         imp_conn = conn;
544         }
545
546         /* if not found, simply choose the current one */
547         if (!imp_conn || imp->imp_force_reconnect) {
548                 LASSERT(imp->imp_conn_current);
549                 imp_conn = imp->imp_conn_current;
550                 tried_all = 0;
551         }
552         LASSERT(imp_conn->oic_conn);
553
554         /* If we've tried everything, and we're back to the beginning of the
555            list, increase our timeout and try again. It will be reset when
556            we do finally connect. (FIXME: really we should wait for all network
557            state associated with the last connection attempt to drain before
558            trying to reconnect on it.) */
559         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
560                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
561                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
562                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
563                         if (at_get(at) > CONNECTION_SWITCH_MAX)
564                                 at_reset(at, CONNECTION_SWITCH_MAX);
565                 }
566                 LASSERT(imp_conn->oic_last_attempt);
567                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
568                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
569         }
570
571         imp_conn->oic_last_attempt = cfs_time_current_64();
572
573         /* switch connection, don't mind if it's same as the current one */
574         if (imp->imp_connection)
575                 ptlrpc_connection_put(imp->imp_connection);
576         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
577
578         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
579         LASSERT(dlmexp != NULL);
580         if (dlmexp->exp_connection)
581                 ptlrpc_connection_put(dlmexp->exp_connection);
582         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
583         class_export_put(dlmexp);
584
585         if (imp->imp_conn_current != imp_conn) {
586                 if (imp->imp_conn_current) {
587                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
588                                   &target_start, &target_len);
589
590                         CDEBUG(D_HA, "%s: Connection changing to"
591                                " %.*s (at %s)\n",
592                                imp->imp_obd->obd_name,
593                                target_len, target_start,
594                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
595                 }
596
597                 imp->imp_conn_current = imp_conn;
598         }
599
600         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
601                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
602                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
603
604         spin_unlock(&imp->imp_lock);
605
606         RETURN(0);
607 }
608
609 /*
610  * must be called under imp_lock
611  */
612 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
613 {
614         struct ptlrpc_request   *req;
615         struct list_head        *tmp;
616
617         /* The requests in committed_list always have smaller transnos than
618          * the requests in replay_list */
619         if (!list_empty(&imp->imp_committed_list)) {
620                 tmp = imp->imp_committed_list.next;
621                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
622                 *transno = req->rq_transno;
623                 if (req->rq_transno == 0) {
624                         DEBUG_REQ(D_ERROR, req, "zero transno in committed_list");
625                         LBUG();
626                 }
627                 return 1;
628         }
629         if (!list_empty(&imp->imp_replay_list)) {
630                 tmp = imp->imp_replay_list.next;
631                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
632                 *transno = req->rq_transno;
633                 if (req->rq_transno == 0) {
634                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
635                         LBUG();
636                 }
637                 return 1;
638         }
639         return 0;
640 }
641
642 /**
643  * Attempt to (re)connect import \a imp. This includes all preparations,
644  * initializing CONNECT RPC request and passing it to ptlrpcd for
645  * actual sending.
646  * Returns 0 on success or error code.
647  */
648 int ptlrpc_connect_import(struct obd_import *imp)
649 {
650         struct obd_device *obd = imp->imp_obd;
651         int initial_connect = 0;
652         int set_transno = 0;
653         __u64 committed_before_reconnect = 0;
654         struct ptlrpc_request *request;
655         char *bufs[] = { NULL,
656                          obd2cli_tgt(imp->imp_obd),
657                          obd->obd_uuid.uuid,
658                          (char *)&imp->imp_dlm_handle,
659                          (char *)&imp->imp_connect_data };
660         struct ptlrpc_connect_async_args *aa;
661         int rc;
662         ENTRY;
663
664         spin_lock(&imp->imp_lock);
665         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
666                 spin_unlock(&imp->imp_lock);
667                 CERROR("can't connect to a closed import\n");
668                 RETURN(-EINVAL);
669         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
670                 spin_unlock(&imp->imp_lock);
671                 CERROR("already connected\n");
672                 RETURN(0);
673         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
674                 spin_unlock(&imp->imp_lock);
675                 CERROR("already connecting\n");
676                 RETURN(-EALREADY);
677         }
678
679         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
680
681         imp->imp_conn_cnt++;
682         imp->imp_resend_replay = 0;
683
684         if (!lustre_handle_is_used(&imp->imp_remote_handle))
685                 initial_connect = 1;
686         else
687                 committed_before_reconnect = imp->imp_peer_committed_transno;
688
689         set_transno = ptlrpc_first_transno(imp,
690                                            &imp->imp_connect_data.ocd_transno);
691         spin_unlock(&imp->imp_lock);
692
693         rc = import_select_connection(imp);
694         if (rc)
695                 GOTO(out, rc);
696
697         rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
698         if (rc)
699                 GOTO(out, rc);
700
701         /* Reset connect flags to the originally requested flags, in case
702          * the server is updated on-the-fly we will get the new features. */
703         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
704         /* Reset ocd_version each time so the server knows the exact versions */
705         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
706         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
707         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
708
709         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
710                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
711         if (rc)
712                 GOTO(out, rc);
713
714         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
715         if (request == NULL)
716                 GOTO(out, rc = -ENOMEM);
717
718         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
719                                       imp->imp_connect_op, bufs, NULL);
720         if (rc) {
721                 ptlrpc_request_free(request);
722                 GOTO(out, rc);
723         }
724
725         /* Report the rpc service time to the server so that it knows how long
726          * to wait for clients to join recovery */
727         lustre_msg_set_service_time(request->rq_reqmsg,
728                                     at_timeout2est(request->rq_timeout));
729
730         /* The amount of time we give the server to process the connect req.
731          * import_select_connection will increase the net latency on
732          * repeated reconnect attempts to cover slow networks.
733          * We override/ignore the server rpc completion estimate here,
734          * which may be large if this is a reconnect attempt */
735         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
736         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
737
738         request->rq_no_resend = request->rq_no_delay = 1;
739         request->rq_send_state = LUSTRE_IMP_CONNECTING;
740         /* Allow a slightly larger reply for future growth compatibility */
741         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
742                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
743         ptlrpc_request_set_replen(request);
744         request->rq_interpret_reply = ptlrpc_connect_interpret;
745
746         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
747         aa = ptlrpc_req_async_args(request);
748         memset(aa, 0, sizeof *aa);
749
750         aa->pcaa_peer_committed = committed_before_reconnect;
751         aa->pcaa_initial_connect = initial_connect;
752
753         if (aa->pcaa_initial_connect) {
754                 spin_lock(&imp->imp_lock);
755                 imp->imp_replayable = 1;
756                 spin_unlock(&imp->imp_lock);
757                 lustre_msg_add_op_flags(request->rq_reqmsg,
758                                         MSG_CONNECT_INITIAL);
759         }
760
761         if (set_transno)
762                 lustre_msg_add_op_flags(request->rq_reqmsg,
763                                         MSG_CONNECT_TRANSNO);
764
765         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
766                   request->rq_timeout);
767         ptlrpcd_add_req(request);
768         rc = 0;
769 out:
770         if (rc != 0)
771                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
772
773         RETURN(rc);
774 }
775 EXPORT_SYMBOL(ptlrpc_connect_import);
776
777 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
778 {
779         int force_verify;
780
781         spin_lock(&imp->imp_lock);
782         force_verify = imp->imp_force_verify != 0;
783         spin_unlock(&imp->imp_lock);
784
785         if (force_verify)
786                 ptlrpc_pinger_wake_up();
787 }
788
789 static int ptlrpc_busy_reconnect(int rc)
790 {
791         return (rc == -EBUSY) || (rc == -EAGAIN);
792 }
793
794 static int ptlrpc_connect_set_flags(struct obd_import *imp,
795                                      struct obd_connect_data *ocd,
796                                      __u64 old_connect_flags,
797                                      struct obd_export *exp, int init_connect)
798 {
799         static bool warned;
800         struct client_obd *cli = &imp->imp_obd->u.cli;
801
802         if ((imp->imp_connect_flags_orig & OBD_CONNECT_IBITS) &&
803             !(ocd->ocd_connect_flags & OBD_CONNECT_IBITS)) {
804                 LCONSOLE_WARN("%s: MDS %s does not support ibits "
805                               "lock, either very old or invalid: "
806                               "requested "LPX64", replied "LPX64"\n",
807                               imp->imp_obd->obd_name,
808                               imp->imp_connection->c_remote_uuid.uuid,
809                               imp->imp_connect_flags_orig,
810                               ocd->ocd_connect_flags);
811                 return -EPROTO;
812         }
813
814         spin_lock(&imp->imp_lock);
815         list_del(&imp->imp_conn_current->oic_item);
816         list_add(&imp->imp_conn_current->oic_item,
817                  &imp->imp_conn_list);
818         imp->imp_last_success_conn =
819                 imp->imp_conn_current->oic_last_attempt;
820
821         spin_unlock(&imp->imp_lock);
822
823
824         if (!warned && (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
825             (ocd->ocd_version > LUSTRE_VERSION_CODE +
826                                 LUSTRE_VERSION_OFFSET_WARN ||
827              ocd->ocd_version < LUSTRE_VERSION_CODE -
828                                 LUSTRE_VERSION_OFFSET_WARN)) {
829                 /* Sigh, some compilers do not like #ifdef in the middle
830                    of macro arguments */
831                 const char *older = "older than client. "
832                                     "Consider upgrading server";
833                 const char *newer = "newer than client. "
834                                     "Consider recompiling application";
835
836                 LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
837                               "is much %s (%s)\n",
838                               obd2cli_tgt(imp->imp_obd),
839                               OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
840                               OBD_OCD_VERSION_MINOR(ocd->ocd_version),
841                               OBD_OCD_VERSION_PATCH(ocd->ocd_version),
842                               OBD_OCD_VERSION_FIX(ocd->ocd_version),
843                               ocd->ocd_version > LUSTRE_VERSION_CODE ?
844                               newer : older, LUSTRE_VERSION_STRING);
845                 warned = true;
846         }
847
848 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
849         /* Check if server has LU-1252 fix applied to not always swab
850          * the IR MNE entries. Do this only once per connection.  This
851          * fixup is version-limited, because we don't want to carry the
852          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
853          * need interop with unpatched 2.2 servers.  For newer servers,
854          * the client will do MNE swabbing only as needed.  LU-1644 */
855         if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
856                      !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
857                      OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
858                      OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
859                      OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
860                      strcmp(imp->imp_obd->obd_type->typ_name,
861                             LUSTRE_MGC_NAME) == 0))
862                 imp->imp_need_mne_swab = 1;
863         else /* clear if server was upgraded since last connect */
864                 imp->imp_need_mne_swab = 0;
865 #endif
866
867         if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
868                 /* We sent to the server ocd_cksum_types with bits set
869                  * for algorithms we understand. The server masked off
870                  * the checksum types it doesn't support */
871                 if ((ocd->ocd_cksum_types &
872                      cksum_types_supported_client()) == 0) {
873                         LCONSOLE_WARN("The negotiation of the checksum "
874                                       "alogrithm to use with server %s "
875                                       "failed (%x/%x), disabling "
876                                       "checksums\n",
877                                       obd2cli_tgt(imp->imp_obd),
878                                       ocd->ocd_cksum_types,
879                                       cksum_types_supported_client());
880                         cli->cl_checksum = 0;
881                         cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
882                 } else {
883                         cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
884                 }
885         } else {
886                 /* The server does not support OBD_CONNECT_CKSUM.
887                  * Enforce ADLER for backward compatibility*/
888                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
889         }
890         cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
891
892         if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
893                 cli->cl_max_pages_per_rpc =
894                         min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
895                             cli->cl_max_pages_per_rpc);
896         else if (imp->imp_connect_op == MDS_CONNECT ||
897                  imp->imp_connect_op == MGS_CONNECT)
898                 cli->cl_max_pages_per_rpc = 1;
899
900         LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
901                 (cli->cl_max_pages_per_rpc > 0));
902
903         client_adjust_max_dirty(cli);
904
905         /* Update client max modify RPCs in flight with value returned
906          * by the server */
907         if (ocd->ocd_connect_flags & OBD_CONNECT_MULTIMODRPCS)
908                 cli->cl_max_mod_rpcs_in_flight = min(
909                                         cli->cl_max_mod_rpcs_in_flight,
910                                         ocd->ocd_maxmodrpcs);
911         else
912                 cli->cl_max_mod_rpcs_in_flight = 1;
913
914         /* Reset ns_connect_flags only for initial connect. It might be
915          * changed in while using FS and if we reset it in reconnect
916          * this leads to losing user settings done before such as
917          * disable lru_resize, etc. */
918         if (old_connect_flags != exp_connect_flags(exp) || init_connect) {
919                 CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
920                              "flags: "LPX64"\n", imp->imp_obd->obd_name,
921                              ocd->ocd_connect_flags);
922                 imp->imp_obd->obd_namespace->ns_connect_flags =
923                         ocd->ocd_connect_flags;
924                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
925                         ocd->ocd_connect_flags;
926         }
927
928         if (ocd->ocd_connect_flags & OBD_CONNECT_AT)
929                 /* We need a per-message support flag, because
930                  * a. we don't know if the incoming connect reply
931                  *    supports AT or not (in reply_in_callback)
932                  *    until we unpack it.
933                  * b. failovered server means export and flags are gone
934                  *    (in ptlrpc_send_reply).
935                  *    Can only be set when we know AT is supported at
936                  *    both ends */
937                 imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
938         else
939                 imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
940
941         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
942
943         return 0;
944 }
945
946 /**
947  * interpret_reply callback for connect RPCs.
948  * Looks into returned status of connect operation and decides
949  * what to do with the import - i.e enter recovery, promote it to
950  * full state for normal operations of disconnect it due to an error.
951  */
952 static int ptlrpc_connect_interpret(const struct lu_env *env,
953                                     struct ptlrpc_request *request,
954                                     void *data, int rc)
955 {
956         struct ptlrpc_connect_async_args *aa = data;
957         struct obd_import *imp = request->rq_import;
958         struct lustre_handle old_hdl;
959         __u64 old_connect_flags;
960         int msg_flags;
961         struct obd_connect_data *ocd;
962         struct obd_export *exp;
963         int ret;
964         ENTRY;
965
966         spin_lock(&imp->imp_lock);
967         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
968                 imp->imp_connect_tried = 1;
969                 spin_unlock(&imp->imp_lock);
970                 RETURN(0);
971         }
972
973         if (rc) {
974                 /* if this reconnect to busy export - not need select new target
975                  * for connecting*/
976                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
977                 spin_unlock(&imp->imp_lock);
978                 ptlrpc_maybe_ping_import_soon(imp);
979                 GOTO(out, rc);
980         }
981         spin_unlock(&imp->imp_lock);
982
983         LASSERT(imp->imp_conn_current);
984
985         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
986
987         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
988                                    RCL_SERVER);
989         /* server replied obd_connect_data is always bigger */
990         ocd = req_capsule_server_sized_get(&request->rq_pill,
991                                            &RMF_CONNECT_DATA, ret);
992
993         if (ocd == NULL) {
994                 CERROR("%s: no connect data from server\n",
995                        imp->imp_obd->obd_name);
996                 rc = -EPROTO;
997                 GOTO(out, rc);
998         }
999
1000         spin_lock(&imp->imp_lock);
1001
1002         /* All imports are pingable */
1003         imp->imp_pingable = 1;
1004         imp->imp_force_reconnect = 0;
1005         imp->imp_force_verify = 0;
1006
1007         imp->imp_connect_data = *ocd;
1008
1009         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
1010                imp->imp_obd->obd_name, ocd->ocd_instance);
1011         exp = class_conn2export(&imp->imp_dlm_handle);
1012
1013         spin_unlock(&imp->imp_lock);
1014
1015         /* check that server granted subset of flags we asked for. */
1016         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
1017             ocd->ocd_connect_flags) {
1018                 CERROR("%s: Server didn't granted asked subset of flags: "
1019                        "asked="LPX64" grranted="LPX64"\n",
1020                        imp->imp_obd->obd_name,imp->imp_connect_flags_orig,
1021                        ocd->ocd_connect_flags);
1022                 GOTO(out, rc = -EPROTO);
1023         }
1024
1025         if (!(imp->imp_connect_flags_orig & OBD_CONNECT_LIGHTWEIGHT) &&
1026             (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) &&
1027             (imp->imp_connect_flags_orig & OBD_CONNECT_FID) &&
1028             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1029                 __u32 major = OBD_OCD_VERSION_MAJOR(ocd->ocd_version);
1030                 __u32 minor = OBD_OCD_VERSION_MINOR(ocd->ocd_version);
1031                 __u32 patch = OBD_OCD_VERSION_PATCH(ocd->ocd_version);
1032
1033                 /* We do not support the MDT-MDT interoperations with
1034                  * different version MDT because of protocol changes. */
1035                 if (unlikely(major != LUSTRE_MAJOR ||
1036                              minor != LUSTRE_MINOR ||
1037                              abs(patch - LUSTRE_PATCH) > 3)) {
1038                         LCONSOLE_WARN("%s: import %p (%u.%u.%u.%u) tried the "
1039                                       "connection to different version MDT "
1040                                       "(%d.%d.%d.%d) %s\n",
1041                                       imp->imp_obd->obd_name, imp, LUSTRE_MAJOR,
1042                                       LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1043                                       major, minor, patch,
1044                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1045                                       imp->imp_connection->c_remote_uuid.uuid);
1046
1047                         GOTO(out, rc = -EPROTO);
1048                 }
1049         }
1050
1051         if (!exp) {
1052                 /* This could happen if export is cleaned during the
1053                    connect attempt */
1054                 CERROR("%s: missing export after connect\n",
1055                        imp->imp_obd->obd_name);
1056                 GOTO(out, rc = -ENODEV);
1057         }
1058         old_connect_flags = exp_connect_flags(exp);
1059         exp->exp_connect_data = *ocd;
1060         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
1061
1062         /* The net statistics after (re-)connect is not valid anymore,
1063          * because may reflect other routing, etc. */
1064         at_reinit(&imp->imp_at.iat_net_latency, 0, 0);
1065         ptlrpc_at_adj_net_latency(request,
1066                         lustre_msg_get_service_time(request->rq_repmsg));
1067
1068         /* Import flags should be updated before waking import at FULL state */
1069         rc = ptlrpc_connect_set_flags(imp, ocd, old_connect_flags, exp,
1070                                  aa->pcaa_initial_connect);
1071         class_export_put(exp);
1072         if (rc != 0)
1073                 GOTO(out, rc);
1074
1075         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1076
1077         if (aa->pcaa_initial_connect) {
1078                 spin_lock(&imp->imp_lock);
1079                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
1080                         imp->imp_replayable = 1;
1081                         spin_unlock(&imp->imp_lock);
1082                         CDEBUG(D_HA, "connected to replayable target: %s\n",
1083                                obd2cli_tgt(imp->imp_obd));
1084                 } else {
1085                         imp->imp_replayable = 0;
1086                         spin_unlock(&imp->imp_lock);
1087                 }
1088
1089                 /* if applies, adjust the imp->imp_msg_magic here
1090                  * according to reply flags */
1091
1092                 imp->imp_remote_handle =
1093                                 *lustre_msg_get_handle(request->rq_repmsg);
1094
1095                 /* Initial connects are allowed for clients with non-random
1096                  * uuids when servers are in recovery.  Simply signal the
1097                  * servers replay is complete and wait in REPLAY_WAIT. */
1098                 if (msg_flags & MSG_CONNECT_RECOVERING) {
1099                         CDEBUG(D_HA, "connect to %s during recovery\n",
1100                                obd2cli_tgt(imp->imp_obd));
1101                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1102                 } else {
1103                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1104                         ptlrpc_activate_import(imp);
1105                 }
1106
1107                 GOTO(finish, rc = 0);
1108         }
1109
1110         /* Determine what recovery state to move the import to. */
1111         if (MSG_CONNECT_RECONNECT & msg_flags) {
1112                 memset(&old_hdl, 0, sizeof(old_hdl));
1113                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
1114                             sizeof (old_hdl))) {
1115                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
1116                                       "bad handle "LPX64"\n",
1117                                       obd2cli_tgt(imp->imp_obd),
1118                                       imp->imp_connection->c_remote_uuid.uuid,
1119                                       imp->imp_dlm_handle.cookie);
1120                         GOTO(out, rc = -ENOTCONN);
1121                 }
1122
1123                 if (memcmp(&imp->imp_remote_handle,
1124                            lustre_msg_get_handle(request->rq_repmsg),
1125                            sizeof(imp->imp_remote_handle))) {
1126                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
1127                                 D_HA : D_WARNING;
1128
1129                         /* Bug 16611/14775: if server handle have changed,
1130                          * that means some sort of disconnection happened.
1131                          * If the server is not in recovery, that also means it
1132                          * already erased all of our state because of previous
1133                          * eviction. If it is in recovery - we are safe to
1134                          * participate since we can reestablish all of our state
1135                          * with server again */
1136                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
1137                                 CDEBUG(level,"%s@%s changed server handle from "
1138                                        LPX64" to "LPX64
1139                                        " but is still in recovery\n",
1140                                        obd2cli_tgt(imp->imp_obd),
1141                                        imp->imp_connection->c_remote_uuid.uuid,
1142                                        imp->imp_remote_handle.cookie,
1143                                        lustre_msg_get_handle(
1144                                        request->rq_repmsg)->cookie);
1145                         } else {
1146                                 LCONSOLE_WARN("Evicted from %s (at %s) "
1147                                               "after server handle changed from "
1148                                               LPX64" to "LPX64"\n",
1149                                               obd2cli_tgt(imp->imp_obd),
1150                                               imp->imp_connection-> \
1151                                               c_remote_uuid.uuid,
1152                                               imp->imp_remote_handle.cookie,
1153                                               lustre_msg_get_handle(
1154                                               request->rq_repmsg)->cookie);
1155                         }
1156
1157
1158                         imp->imp_remote_handle =
1159                                      *lustre_msg_get_handle(request->rq_repmsg);
1160
1161                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
1162                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1163                                 GOTO(finish, rc = 0);
1164                         }
1165
1166                 } else {
1167                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
1168                                obd2cli_tgt(imp->imp_obd),
1169                                imp->imp_connection->c_remote_uuid.uuid);
1170                 }
1171
1172                 if (imp->imp_invalid) {
1173                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
1174                                "marking evicted\n", imp->imp_obd->obd_name);
1175                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1176                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
1177                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
1178                                imp->imp_obd->obd_name,
1179                                obd2cli_tgt(imp->imp_obd));
1180
1181                         spin_lock(&imp->imp_lock);
1182                         imp->imp_resend_replay = 1;
1183                         spin_unlock(&imp->imp_lock);
1184
1185                         IMPORT_SET_STATE(imp, imp->imp_replay_state);
1186                 } else {
1187                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1188                 }
1189         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
1190                 LASSERT(imp->imp_replayable);
1191                 imp->imp_remote_handle =
1192                                 *lustre_msg_get_handle(request->rq_repmsg);
1193                 imp->imp_last_replay_transno = 0;
1194                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
1195         } else {
1196                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
1197                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
1198                 imp->imp_remote_handle =
1199                                 *lustre_msg_get_handle(request->rq_repmsg);
1200                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
1201         }
1202
1203         /* Sanity checks for a reconnected import. */
1204         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
1205                 CERROR("imp_replayable flag does not match server "
1206                        "after reconnect. We should LBUG right here.\n");
1207         }
1208
1209         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
1210             lustre_msg_get_last_committed(request->rq_repmsg) <
1211             aa->pcaa_peer_committed) {
1212                 CERROR("%s went back in time (transno "LPD64
1213                        " was previously committed, server now claims "LPD64
1214                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
1215                        "id=9646\n",
1216                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1217                        lustre_msg_get_last_committed(request->rq_repmsg));
1218         }
1219
1220 finish:
1221         rc = ptlrpc_import_recovery_state_machine(imp);
1222         if (rc == -ENOTCONN) {
1223                 CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
1224                        "invalidating and reconnecting\n",
1225                        obd2cli_tgt(imp->imp_obd),
1226                        imp->imp_connection->c_remote_uuid.uuid);
1227                 ptlrpc_connect_import(imp);
1228                 imp->imp_connect_tried = 1;
1229                 RETURN(0);
1230         }
1231
1232 out:
1233         imp->imp_connect_tried = 1;
1234
1235         if (rc != 0) {
1236                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1237                 if (rc == -EACCES) {
1238                         /*
1239                          * Give up trying to reconnect
1240                          * EACCES means client has no permission for connection
1241                          */
1242                         imp->imp_obd->obd_no_recov = 1;
1243                         ptlrpc_deactivate_import(imp);
1244                 }
1245
1246                 if (rc == -EPROTO) {
1247                         struct obd_connect_data *ocd;
1248
1249                         /* reply message might not be ready */
1250                         if (request->rq_repmsg == NULL)
1251                                 RETURN(-EPROTO);
1252
1253                         ocd = req_capsule_server_get(&request->rq_pill,
1254                                                      &RMF_CONNECT_DATA);
1255                         if (ocd &&
1256                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1257                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1258                            /* Actually servers are only supposed to refuse
1259                               connection from liblustre clients, so we should
1260                               never see this from VFS context */
1261                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1262                                         "(%d.%d.%d.%d)"
1263                                         " refused connection from this client "
1264                                         "with an incompatible version (%s).  "
1265                                         "Client must be recompiled\n",
1266                                         obd2cli_tgt(imp->imp_obd),
1267                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1268                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1269                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1270                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1271                                         LUSTRE_VERSION_STRING);
1272                                 ptlrpc_deactivate_import(imp);
1273                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1274                         }
1275                         RETURN(-EPROTO);
1276                 }
1277
1278                 ptlrpc_maybe_ping_import_soon(imp);
1279
1280                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1281                        obd2cli_tgt(imp->imp_obd),
1282                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1283         }
1284
1285         wake_up_all(&imp->imp_recovery_waitq);
1286         RETURN(rc);
1287 }
1288
1289 /**
1290  * interpret callback for "completed replay" RPCs.
1291  * \see signal_completed_replay
1292  */
1293 static int completed_replay_interpret(const struct lu_env *env,
1294                                       struct ptlrpc_request *req,
1295                                       void * data, int rc)
1296 {
1297         ENTRY;
1298         atomic_dec(&req->rq_import->imp_replay_inflight);
1299         if (req->rq_status == 0 &&
1300             !req->rq_import->imp_vbr_failed) {
1301                 ptlrpc_import_recovery_state_machine(req->rq_import);
1302         } else {
1303                 if (req->rq_import->imp_vbr_failed) {
1304                         CDEBUG(D_WARNING,
1305                                "%s: version recovery fails, reconnecting\n",
1306                                req->rq_import->imp_obd->obd_name);
1307                 } else {
1308                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1309                                      "reconnecting\n",
1310                                req->rq_import->imp_obd->obd_name,
1311                                req->rq_status);
1312                 }
1313                 ptlrpc_connect_import(req->rq_import);
1314         }
1315
1316         RETURN(0);
1317 }
1318
1319 /**
1320  * Let server know that we have no requests to replay anymore.
1321  * Achieved by just sending a PING request
1322  */
1323 static int signal_completed_replay(struct obd_import *imp)
1324 {
1325         struct ptlrpc_request *req;
1326         ENTRY;
1327
1328         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1329                 RETURN(0);
1330
1331         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1332         atomic_inc(&imp->imp_replay_inflight);
1333
1334         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1335                                         OBD_PING);
1336         if (req == NULL) {
1337                 atomic_dec(&imp->imp_replay_inflight);
1338                 RETURN(-ENOMEM);
1339         }
1340
1341         ptlrpc_request_set_replen(req);
1342         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1343         lustre_msg_add_flags(req->rq_reqmsg,
1344                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1345         if (AT_OFF)
1346                 req->rq_timeout *= 3;
1347         req->rq_interpret_reply = completed_replay_interpret;
1348
1349         ptlrpcd_add_req(req);
1350         RETURN(0);
1351 }
1352
1353 /**
1354  * In kernel code all import invalidation happens in its own
1355  * separate thread, so that whatever application happened to encounter
1356  * a problem could still be killed or otherwise continue
1357  */
1358 static int ptlrpc_invalidate_import_thread(void *data)
1359 {
1360         struct obd_import *imp = data;
1361
1362         ENTRY;
1363
1364         unshare_fs_struct();
1365
1366         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1367                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1368                imp->imp_connection->c_remote_uuid.uuid);
1369
1370         ptlrpc_invalidate_import(imp);
1371
1372         if (obd_dump_on_eviction) {
1373                 CERROR("dump the log upon eviction\n");
1374                 libcfs_debug_dumplog();
1375         }
1376
1377         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1378         ptlrpc_import_recovery_state_machine(imp);
1379
1380         class_import_put(imp);
1381         RETURN(0);
1382 }
1383
1384 /**
1385  * This is the state machine for client-side recovery on import.
1386  *
1387  * Typicaly we have two possibly paths. If we came to server and it is not
1388  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1389  * state and reconnect from scratch.
1390  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1391  * We go through our list of requests to replay and send them to server one by
1392  * one.
1393  * After sending all request from the list we change import state to
1394  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1395  * and also all the locks we don't yet have and wait for server to grant us.
1396  * After that we send a special "replay completed" request and change import
1397  * state to IMP_REPLAY_WAIT.
1398  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1399  * state and resend all requests from sending list.
1400  * After that we promote import to FULL state and send all delayed requests
1401  * and import is fully operational after that.
1402  *
1403  */
1404 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1405 {
1406         int rc = 0;
1407         int inflight;
1408         char *target_start;
1409         int target_len;
1410
1411         ENTRY;
1412         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1413                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1414                           &target_start, &target_len);
1415                 /* Don't care about MGC eviction */
1416                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1417                            LUSTRE_MGC_NAME) != 0) {
1418                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1419                                            "by %.*s; in progress operations "
1420                                            "using this service will fail.\n",
1421                                            imp->imp_obd->obd_name, target_len,
1422                                            target_start);
1423                 }
1424                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1425                        obd2cli_tgt(imp->imp_obd),
1426                        imp->imp_connection->c_remote_uuid.uuid);
1427                 /* reset vbr_failed flag upon eviction */
1428                 spin_lock(&imp->imp_lock);
1429                 imp->imp_vbr_failed = 0;
1430                 spin_unlock(&imp->imp_lock);
1431
1432                 {
1433                 struct task_struct *task;
1434                 /* bug 17802:  XXX client_disconnect_export vs connect request
1435                  * race. if client is evicted at this time then we start
1436                  * invalidate thread without reference to import and import can
1437                  * be freed at same time. */
1438                 class_import_get(imp);
1439                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1440                                      "ll_imp_inval");
1441                 if (IS_ERR(task)) {
1442                         class_import_put(imp);
1443                         CERROR("error starting invalidate thread: %d\n", rc);
1444                         rc = PTR_ERR(task);
1445                 } else {
1446                         rc = 0;
1447                 }
1448                 RETURN(rc);
1449                 }
1450         }
1451
1452         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1453                 CDEBUG(D_HA, "replay requested by %s\n",
1454                        obd2cli_tgt(imp->imp_obd));
1455                 rc = ptlrpc_replay_next(imp, &inflight);
1456                 if (inflight == 0 &&
1457                     atomic_read(&imp->imp_replay_inflight) == 0) {
1458                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1459                         rc = ldlm_replay_locks(imp);
1460                         if (rc)
1461                                 GOTO(out, rc);
1462                 }
1463                 rc = 0;
1464         }
1465
1466         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1467                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1468                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1469                         rc = signal_completed_replay(imp);
1470                         if (rc)
1471                                 GOTO(out, rc);
1472                 }
1473         }
1474
1475         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1476                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1477                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1478                 }
1479         }
1480
1481         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1482                 struct ptlrpc_connection *conn = imp->imp_connection;
1483
1484                 rc = ptlrpc_resend(imp);
1485                 if (rc)
1486                         GOTO(out, rc);
1487                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1488                 ptlrpc_activate_import(imp);
1489
1490                 LCONSOLE_INFO("%s: Connection restored to %s (at %s)\n",
1491                               imp->imp_obd->obd_name,
1492                               obd_uuid2str(&conn->c_remote_uuid),
1493                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1494         }
1495
1496         if (imp->imp_state == LUSTRE_IMP_FULL) {
1497                 wake_up_all(&imp->imp_recovery_waitq);
1498                 ptlrpc_wake_delayed(imp);
1499         }
1500
1501 out:
1502         RETURN(rc);
1503 }
1504
1505 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1506 {
1507         struct ptlrpc_request *req;
1508         int rq_opc, rc = 0;
1509         ENTRY;
1510
1511         if (imp->imp_obd->obd_force)
1512                 GOTO(set_state, rc);
1513
1514         switch (imp->imp_connect_op) {
1515         case OST_CONNECT:
1516                 rq_opc = OST_DISCONNECT;
1517                 break;
1518         case MDS_CONNECT:
1519                 rq_opc = MDS_DISCONNECT;
1520                 break;
1521         case MGS_CONNECT:
1522                 rq_opc = MGS_DISCONNECT;
1523                 break;
1524         default:
1525                 rc = -EINVAL;
1526                 CERROR("%s: don't know how to disconnect from %s "
1527                        "(connect_op %d): rc = %d\n",
1528                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1529                        imp->imp_connect_op, rc);
1530                 RETURN(rc);
1531         }
1532
1533         if (ptlrpc_import_in_recovery(imp)) {
1534                 struct l_wait_info lwi;
1535                 cfs_duration_t timeout;
1536
1537                 if (AT_OFF) {
1538                         if (imp->imp_server_timeout)
1539                                 timeout = cfs_time_seconds(obd_timeout / 2);
1540                         else
1541                                 timeout = cfs_time_seconds(obd_timeout);
1542                 } else {
1543                         int idx = import_at_get_index(imp,
1544                                 imp->imp_client->cli_request_portal);
1545                         timeout = cfs_time_seconds(
1546                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1547                 }
1548
1549                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1550                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1551                 rc = l_wait_event(imp->imp_recovery_waitq,
1552                                   !ptlrpc_import_in_recovery(imp), &lwi);
1553
1554         }
1555
1556         spin_lock(&imp->imp_lock);
1557         if (imp->imp_state != LUSTRE_IMP_FULL)
1558                 GOTO(out, rc);
1559         spin_unlock(&imp->imp_lock);
1560
1561         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1562                                         LUSTRE_OBD_VERSION, rq_opc);
1563         if (req) {
1564                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1565                  * it fails.  We can get through the above with a down server
1566                  * if the client doesn't know the server is gone yet. */
1567                 req->rq_no_resend = 1;
1568
1569                 /* We want client umounts to happen quickly, no matter the
1570                    server state... */
1571                 req->rq_timeout = min_t(int, req->rq_timeout,
1572                                         INITIAL_CONNECT_TIMEOUT);
1573
1574                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1575                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1576                 ptlrpc_request_set_replen(req);
1577                 rc = ptlrpc_queue_wait(req);
1578                 ptlrpc_req_finished(req);
1579         }
1580
1581 set_state:
1582         spin_lock(&imp->imp_lock);
1583 out:
1584         if (noclose)
1585                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1586         else
1587                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1588         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1589         spin_unlock(&imp->imp_lock);
1590
1591         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1592                 rc = 0;
1593         RETURN(rc);
1594 }
1595 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1596
1597 void ptlrpc_cleanup_imp(struct obd_import *imp)
1598 {
1599         ENTRY;
1600
1601         spin_lock(&imp->imp_lock);
1602         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1603         imp->imp_generation++;
1604         spin_unlock(&imp->imp_lock);
1605         ptlrpc_abort_inflight(imp);
1606
1607         EXIT;
1608 }
1609
1610 /* Adaptive Timeout utils */
1611 extern unsigned int at_min, at_max, at_history;
1612
1613 /* Update at_current with the specified value (bounded by at_min and at_max),
1614  * as well as the AT history "bins".
1615  *  - Bin into timeslices using AT_BINS bins.
1616  *  - This gives us a max of the last at_history seconds without the storage,
1617  *    but still smoothing out a return to normalcy from a slow response.
1618  *  - (E.g. remember the maximum latency in each minute of the last 4 minutes.)
1619  */
1620 int at_measured(struct adaptive_timeout *at, unsigned int val)
1621 {
1622         unsigned int old = at->at_current;
1623         time_t now = cfs_time_current_sec();
1624         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1625
1626         LASSERT(at);
1627         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1628                val, at, now - at->at_binstart, at->at_current,
1629                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1630
1631         if (val == 0)
1632                 /* 0's don't count, because we never want our timeout to
1633                    drop to 0, and because 0 could mean an error */
1634                 return 0;
1635
1636         spin_lock(&at->at_lock);
1637
1638         if (unlikely(at->at_binstart == 0)) {
1639                 /* Special case to remove default from history */
1640                 at->at_current = val;
1641                 at->at_worst_ever = val;
1642                 at->at_worst_time = now;
1643                 at->at_hist[0] = val;
1644                 at->at_binstart = now;
1645         } else if (now - at->at_binstart < binlimit ) {
1646                 /* in bin 0 */
1647                 at->at_hist[0] = max(val, at->at_hist[0]);
1648                 at->at_current = max(val, at->at_current);
1649         } else {
1650                 int i, shift;
1651                 unsigned int maxv = val;
1652                 /* move bins over */
1653                 shift = (now - at->at_binstart) / binlimit;
1654                 LASSERT(shift > 0);
1655                 for(i = AT_BINS - 1; i >= 0; i--) {
1656                         if (i >= shift) {
1657                                 at->at_hist[i] = at->at_hist[i - shift];
1658                                 maxv = max(maxv, at->at_hist[i]);
1659                         } else {
1660                                 at->at_hist[i] = 0;
1661                         }
1662                 }
1663                 at->at_hist[0] = val;
1664                 at->at_current = maxv;
1665                 at->at_binstart += shift * binlimit;
1666         }
1667
1668         if (at->at_current > at->at_worst_ever) {
1669                 at->at_worst_ever = at->at_current;
1670                 at->at_worst_time = now;
1671         }
1672
1673         if (at->at_flags & AT_FLG_NOHIST)
1674                 /* Only keep last reported val; keeping the rest of the history
1675                    for proc only */
1676                 at->at_current = val;
1677
1678         if (at_max > 0)
1679                 at->at_current =  min(at->at_current, at_max);
1680         at->at_current =  max(at->at_current, at_min);
1681
1682         if (at->at_current != old)
1683                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1684                        "(val=%u) hist %u %u %u %u\n", at,
1685                        old, at->at_current, at->at_current - old, val,
1686                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1687                        at->at_hist[3]);
1688
1689         /* if we changed, report the old value */
1690         old = (at->at_current != old) ? old : 0;
1691
1692         spin_unlock(&at->at_lock);
1693         return old;
1694 }
1695
1696 /* Find the imp_at index for a given portal; assign if space available */
1697 int import_at_get_index(struct obd_import *imp, int portal)
1698 {
1699         struct imp_at *at = &imp->imp_at;
1700         int i;
1701
1702         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1703                 if (at->iat_portal[i] == portal)
1704                         return i;
1705                 if (at->iat_portal[i] == 0)
1706                         /* unused */
1707                         break;
1708         }
1709
1710         /* Not found in list, add it under a lock */
1711         spin_lock(&imp->imp_lock);
1712
1713         /* Check unused under lock */
1714         for (; i < IMP_AT_MAX_PORTALS; i++) {
1715                 if (at->iat_portal[i] == portal)
1716                         goto out;
1717                 if (at->iat_portal[i] == 0)
1718                         /* unused */
1719                         break;
1720         }
1721
1722         /* Not enough portals? */
1723         LASSERT(i < IMP_AT_MAX_PORTALS);
1724
1725         at->iat_portal[i] = portal;
1726 out:
1727         spin_unlock(&imp->imp_lock);
1728         return i;
1729 }