Whamcloud - gitweb
LU-8407 recovery: more clear message about recovery failure
[fs/lustre-release.git] / lustre / ldlm / ldlm_lib.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) 2010, 2015, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
37 /**
38  * This file deals with various client/target related logic including recovery.
39  *
40  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
41  * should be moved.
42  */
43
44 #define DEBUG_SUBSYSTEM S_LDLM
45
46 #include <linux/kthread.h>
47 #include <libcfs/libcfs.h>
48 #include <obd.h>
49 #include <obd_class.h>
50 #include <lustre_dlm.h>
51 #include <lustre_net.h>
52 #include <lustre_sec.h>
53 #include "ldlm_internal.h"
54
55 /* @priority: If non-zero, move the selected connection to the list head.
56  * @create: If zero, only search in existing connections.
57  */
58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
59                            int priority, int create)
60 {
61         struct ptlrpc_connection *ptlrpc_conn;
62         struct obd_import_conn *imp_conn = NULL, *item;
63         int rc = 0;
64         ENTRY;
65
66         if (!create && !priority) {
67                 CDEBUG(D_HA, "Nothing to do\n");
68                 RETURN(-EINVAL);
69         }
70
71         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72         if (!ptlrpc_conn) {
73                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
74                 RETURN (-ENOENT);
75         }
76
77         if (create) {
78                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79                 if (!imp_conn) {
80                         GOTO(out_put, rc = -ENOMEM);
81                 }
82         }
83
84         spin_lock(&imp->imp_lock);
85         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
86                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87                         if (priority) {
88                                 list_del(&item->oic_item);
89                                 list_add(&item->oic_item,
90                                              &imp->imp_conn_list);
91                                 item->oic_last_attempt = 0;
92                         }
93                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
94                                imp, imp->imp_obd->obd_name, uuid->uuid,
95                                (priority ? ", moved to head" : ""));
96                         spin_unlock(&imp->imp_lock);
97                         GOTO(out_free, rc = 0);
98                 }
99         }
100         /* No existing import connection found for \a uuid. */
101         if (create) {
102                 imp_conn->oic_conn = ptlrpc_conn;
103                 imp_conn->oic_uuid = *uuid;
104                 imp_conn->oic_last_attempt = 0;
105                 if (priority)
106                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107                 else
108                         list_add_tail(&imp_conn->oic_item,
109                                           &imp->imp_conn_list);
110                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
111                        imp, imp->imp_obd->obd_name, uuid->uuid,
112                        (priority ? "head" : "tail"));
113         } else {
114                 spin_unlock(&imp->imp_lock);
115                 GOTO(out_free, rc = -ENOENT);
116         }
117
118         spin_unlock(&imp->imp_lock);
119         RETURN(0);
120 out_free:
121         if (imp_conn)
122                 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 out_put:
124         ptlrpc_connection_put(ptlrpc_conn);
125         RETURN(rc);
126 }
127
128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 {
130         return import_set_conn(imp, uuid, 1, 0);
131 }
132
133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
134                            int priority)
135 {
136         return import_set_conn(imp, uuid, priority, 1);
137 }
138 EXPORT_SYMBOL(client_import_add_conn);
139
140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
141 {
142         struct obd_import_conn *imp_conn;
143         struct obd_export *dlmexp;
144         int rc = -ENOENT;
145         ENTRY;
146
147         spin_lock(&imp->imp_lock);
148         if (list_empty(&imp->imp_conn_list)) {
149                 LASSERT(!imp->imp_connection);
150                 GOTO(out, rc);
151         }
152
153         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
154                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
155                         continue;
156                 LASSERT(imp_conn->oic_conn);
157
158                 if (imp_conn == imp->imp_conn_current) {
159                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
160
161                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
162                             imp->imp_state != LUSTRE_IMP_DISCON) {
163                                 CERROR("can't remove current connection\n");
164                                 GOTO(out, rc = -EBUSY);
165                         }
166
167                         ptlrpc_connection_put(imp->imp_connection);
168                         imp->imp_connection = NULL;
169
170                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
171                         if (dlmexp && dlmexp->exp_connection) {
172                                 LASSERT(dlmexp->exp_connection ==
173                                         imp_conn->oic_conn);
174                                 ptlrpc_connection_put(dlmexp->exp_connection);
175                                 dlmexp->exp_connection = NULL;
176                         }
177
178                         if (dlmexp != NULL)
179                                 class_export_put(dlmexp);
180                 }
181
182                 list_del(&imp_conn->oic_item);
183                 ptlrpc_connection_put(imp_conn->oic_conn);
184                 OBD_FREE(imp_conn, sizeof(*imp_conn));
185                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
186                        imp, imp->imp_obd->obd_name, uuid->uuid);
187                 rc = 0;
188                 break;
189         }
190 out:
191         spin_unlock(&imp->imp_lock);
192         if (rc == -ENOENT)
193                 CERROR("connection %s not found\n", uuid->uuid);
194         RETURN(rc);
195 }
196 EXPORT_SYMBOL(client_import_del_conn);
197
198 /**
199  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
200  * to find a conn uuid of \a imp which can reach \a peer.
201  */
202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
203                             struct obd_uuid *uuid)
204 {
205         struct obd_import_conn *conn;
206         int rc = -ENOENT;
207         ENTRY;
208
209         spin_lock(&imp->imp_lock);
210         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
211                 /* Check if conn UUID does have this peer NID. */
212                 if (class_check_uuid(&conn->oic_uuid, peer)) {
213                         *uuid = conn->oic_uuid;
214                         rc = 0;
215                         break;
216                 }
217         }
218         spin_unlock(&imp->imp_lock);
219         RETURN(rc);
220 }
221 EXPORT_SYMBOL(client_import_find_conn);
222
223 void client_destroy_import(struct obd_import *imp)
224 {
225         /* Drop security policy instance after all RPCs have finished/aborted
226          * to let all busy contexts be released. */
227         class_import_get(imp);
228         class_destroy_import(imp);
229         sptlrpc_import_sec_put(imp);
230         class_import_put(imp);
231 }
232 EXPORT_SYMBOL(client_destroy_import);
233
234 /**
235  * Check whether or not the OSC is on MDT.
236  * In the config log,
237  * osc on MDT
238  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
239  * osc on client
240  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
241  *
242  **/
243 static int osc_on_mdt(char *obdname)
244 {
245         char *ptr;
246
247         ptr = strrchr(obdname, '-');
248         if (ptr == NULL)
249                 return 0;
250
251         if (strncmp(ptr + 1, "MDT", 3) == 0)
252                 return 1;
253
254         return 0;
255 }
256
257 /* Configure an RPC client OBD device.
258  *
259  * lcfg parameters:
260  * 1 - client UUID
261  * 2 - server UUID
262  * 3 - inactive-on-startup
263  */
264 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
265 {
266         struct client_obd *cli = &obddev->u.cli;
267         struct obd_import *imp;
268         struct obd_uuid server_uuid;
269         int rq_portal, rp_portal, connect_op;
270         char *name = obddev->obd_type->typ_name;
271         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
272         char *cli_name = lustre_cfg_buf(lcfg, 0);
273         int rc;
274         ENTRY;
275
276         /* In a more perfect world, we would hang a ptlrpc_client off of
277          * obd_type and just use the values from there. */
278         if (!strcmp(name, LUSTRE_OSC_NAME)) {
279                 rq_portal = OST_REQUEST_PORTAL;
280                 rp_portal = OSC_REPLY_PORTAL;
281                 connect_op = OST_CONNECT;
282                 cli->cl_sp_me = LUSTRE_SP_CLI;
283                 cli->cl_sp_to = LUSTRE_SP_OST;
284                 ns_type = LDLM_NS_TYPE_OSC;
285         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
286                    !strcmp(name, LUSTRE_LWP_NAME)) {
287                 rq_portal = MDS_REQUEST_PORTAL;
288                 rp_portal = MDC_REPLY_PORTAL;
289                 connect_op = MDS_CONNECT;
290                 if (is_lwp_on_ost(cli_name))
291                         cli->cl_sp_me = LUSTRE_SP_OST;
292                 else if (is_lwp_on_mdt(cli_name))
293                         cli->cl_sp_me = LUSTRE_SP_MDT;
294                 else
295                         cli->cl_sp_me = LUSTRE_SP_CLI;
296                 cli->cl_sp_to = LUSTRE_SP_MDT;
297                 ns_type = LDLM_NS_TYPE_MDC;
298         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
299                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
300                         /* OSP_on_MDT for other MDTs */
301                         connect_op = MDS_CONNECT;
302                         cli->cl_sp_to = LUSTRE_SP_MDT;
303                         ns_type = LDLM_NS_TYPE_MDC;
304                         rq_portal = OUT_PORTAL;
305                 } else {
306                         /* OSP on MDT for OST */
307                         connect_op = OST_CONNECT;
308                         cli->cl_sp_to = LUSTRE_SP_OST;
309                         ns_type = LDLM_NS_TYPE_OSC;
310                         rq_portal = OST_REQUEST_PORTAL;
311                 }
312                 rp_portal = OSC_REPLY_PORTAL;
313                 cli->cl_sp_me = LUSTRE_SP_MDT;
314         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
315                 rq_portal = MGS_REQUEST_PORTAL;
316                 rp_portal = MGC_REPLY_PORTAL;
317                 connect_op = MGS_CONNECT;
318                 cli->cl_sp_me = LUSTRE_SP_MGC;
319                 cli->cl_sp_to = LUSTRE_SP_MGS;
320                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
321                 ns_type = LDLM_NS_TYPE_MGC;
322         } else {
323                 CERROR("unknown client OBD type \"%s\", can't setup\n",
324                        name);
325                 RETURN(-EINVAL);
326         }
327
328         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
329                 CERROR("requires a TARGET UUID\n");
330                 RETURN(-EINVAL);
331         }
332
333         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
334                 CERROR("client UUID must be less than 38 characters\n");
335                 RETURN(-EINVAL);
336         }
337
338         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
339                 CERROR("setup requires a SERVER UUID\n");
340                 RETURN(-EINVAL);
341         }
342
343         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
344                 CERROR("target UUID must be less than 38 characters\n");
345                 RETURN(-EINVAL);
346         }
347
348         init_rwsem(&cli->cl_sem);
349         mutex_init(&cli->cl_mgc_mutex);
350         cli->cl_conn_count = 0;
351         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
352                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
353                      sizeof(server_uuid)));
354
355         cli->cl_dirty_pages = 0;
356         cli->cl_avail_grant = 0;
357         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
358         /* cl_dirty_max_pages may be changed at connect time in
359          * ptlrpc_connect_interpret(). */
360         client_adjust_max_dirty(cli);
361         INIT_LIST_HEAD(&cli->cl_cache_waiters);
362         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
363         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
364         INIT_LIST_HEAD(&cli->cl_loi_write_list);
365         INIT_LIST_HEAD(&cli->cl_loi_read_list);
366         spin_lock_init(&cli->cl_loi_list_lock);
367         atomic_set(&cli->cl_pending_w_pages, 0);
368         atomic_set(&cli->cl_pending_r_pages, 0);
369         cli->cl_r_in_flight = 0;
370         cli->cl_w_in_flight = 0;
371
372         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
373         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
374         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
375         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
376         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
377         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
378
379         /* lru for osc. */
380         INIT_LIST_HEAD(&cli->cl_lru_osc);
381         atomic_set(&cli->cl_lru_shrinkers, 0);
382         atomic_long_set(&cli->cl_lru_busy, 0);
383         atomic_long_set(&cli->cl_lru_in_list, 0);
384         INIT_LIST_HEAD(&cli->cl_lru_list);
385         spin_lock_init(&cli->cl_lru_list_lock);
386         atomic_long_set(&cli->cl_unstable_count, 0);
387         INIT_LIST_HEAD(&cli->cl_shrink_list);
388
389         init_waitqueue_head(&cli->cl_destroy_waitq);
390         atomic_set(&cli->cl_destroy_in_flight, 0);
391 #ifdef ENABLE_CHECKSUM
392         /* Turn on checksumming by default. */
393         cli->cl_checksum = 1;
394         /*
395          * The supported checksum types will be worked out at connect time
396          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
397          * through procfs.
398          */
399         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
400 #endif
401         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
402
403         /* Set it to possible maximum size. It may be reduced by ocd_brw_size
404          * from OFD after connecting. */
405         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
406
407         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
408          * it will be updated at OSC connection time. */
409         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
410
411         if (!strcmp(name, LUSTRE_MDC_NAME)) {
412                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
413         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
414                 cli->cl_max_rpcs_in_flight = 2;
415         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
416                 cli->cl_max_rpcs_in_flight = 3;
417         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
418                 cli->cl_max_rpcs_in_flight = 4;
419         } else {
420                 if (osc_on_mdt(obddev->obd_name))
421                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
422                 else
423                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
424         }
425
426         spin_lock_init(&cli->cl_mod_rpcs_lock);
427         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
428         cli->cl_max_mod_rpcs_in_flight = 0;
429         cli->cl_mod_rpcs_in_flight = 0;
430         cli->cl_close_rpcs_in_flight = 0;
431         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
432         cli->cl_mod_tag_bitmap = NULL;
433
434         if (connect_op == MDS_CONNECT) {
435                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
436                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
437                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
438                 if (cli->cl_mod_tag_bitmap == NULL)
439                         GOTO(err, rc = -ENOMEM);
440         }
441
442         rc = ldlm_get_ref();
443         if (rc) {
444                 CERROR("ldlm_get_ref failed: %d\n", rc);
445                 GOTO(err, rc);
446         }
447
448         ptlrpc_init_client(rq_portal, rp_portal, name,
449                            &obddev->obd_ldlm_client);
450
451         imp = class_new_import(obddev);
452         if (imp == NULL)
453                 GOTO(err_ldlm, rc = -ENOENT);
454         imp->imp_client = &obddev->obd_ldlm_client;
455         imp->imp_connect_op = connect_op;
456         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
457                LUSTRE_CFG_BUFLEN(lcfg, 1));
458         class_import_put(imp);
459
460         rc = client_import_add_conn(imp, &server_uuid, 1);
461         if (rc) {
462                 CERROR("can't add initial connection\n");
463                 GOTO(err_import, rc);
464         }
465
466         cli->cl_import = imp;
467         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
468         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
469
470         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
471                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
472                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
473                                name, obddev->obd_name,
474                                cli->cl_target_uuid.uuid);
475                         spin_lock(&imp->imp_lock);
476                         imp->imp_deactive = 1;
477                         spin_unlock(&imp->imp_lock);
478                 }
479         }
480
481         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
482                                                    LDLM_NAMESPACE_CLIENT,
483                                                    LDLM_NAMESPACE_GREEDY,
484                                                    ns_type);
485         if (obddev->obd_namespace == NULL) {
486                 CERROR("Unable to create client namespace - %s\n",
487                        obddev->obd_name);
488                 GOTO(err_import, rc = -ENOMEM);
489         }
490
491         RETURN(rc);
492
493 err_import:
494         class_destroy_import(imp);
495 err_ldlm:
496         ldlm_put_ref();
497 err:
498         if (cli->cl_mod_tag_bitmap != NULL)
499                 OBD_FREE(cli->cl_mod_tag_bitmap,
500                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
501         cli->cl_mod_tag_bitmap = NULL;
502         RETURN(rc);
503
504 }
505 EXPORT_SYMBOL(client_obd_setup);
506
507 int client_obd_cleanup(struct obd_device *obddev)
508 {
509         struct client_obd *cli = &obddev->u.cli;
510         ENTRY;
511
512         ldlm_namespace_free_post(obddev->obd_namespace);
513         obddev->obd_namespace = NULL;
514
515         obd_cleanup_client_import(obddev);
516         LASSERT(obddev->u.cli.cl_import == NULL);
517
518         ldlm_put_ref();
519
520         if (cli->cl_mod_tag_bitmap != NULL)
521                 OBD_FREE(cli->cl_mod_tag_bitmap,
522                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
523         cli->cl_mod_tag_bitmap = NULL;
524
525         RETURN(0);
526 }
527 EXPORT_SYMBOL(client_obd_cleanup);
528
529 /* ->o_connect() method for client side (OSC and MDC and MGC) */
530 int client_connect_import(const struct lu_env *env,
531                           struct obd_export **exp,
532                           struct obd_device *obd, struct obd_uuid *cluuid,
533                           struct obd_connect_data *data, void *localdata)
534 {
535         struct client_obd       *cli    = &obd->u.cli;
536         struct obd_import       *imp    = cli->cl_import;
537         struct obd_connect_data *ocd;
538         struct lustre_handle    conn    = { 0 };
539         int                     rc;
540         bool                    is_mdc = false;
541         ENTRY;
542
543         *exp = NULL;
544         down_write(&cli->cl_sem);
545         if (cli->cl_conn_count > 0)
546                 GOTO(out_sem, rc = -EALREADY);
547
548         rc = class_connect(&conn, obd, cluuid);
549         if (rc)
550                 GOTO(out_sem, rc);
551
552         cli->cl_conn_count++;
553         *exp = class_conn2export(&conn);
554
555         LASSERT(obd->obd_namespace);
556
557         imp->imp_dlm_handle = conn;
558         rc = ptlrpc_init_import(imp);
559         if (rc != 0)
560                 GOTO(out_ldlm, rc);
561
562         ocd = &imp->imp_connect_data;
563         if (data) {
564                 *ocd = *data;
565                 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
566                                  LUSTRE_MDC_NAME, 3) == 0;
567                 if (is_mdc)
568                         data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
569                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
570                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
571         }
572
573         rc = ptlrpc_connect_import(imp);
574         if (rc != 0) {
575                 if (data && is_mdc)
576                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
577                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
578                 GOTO(out_ldlm, rc);
579         }
580         LASSERT(*exp != NULL && (*exp)->exp_connection);
581
582         if (data) {
583                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
584                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
585                          data->ocd_connect_flags, ocd->ocd_connect_flags);
586                 data->ocd_connect_flags = ocd->ocd_connect_flags;
587                 /* clear the flag as it was not set and is not known
588                  * by upper layers */
589                 if (is_mdc)
590                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
591         }
592
593         ptlrpc_pinger_add_import(imp);
594
595         EXIT;
596
597         if (rc) {
598 out_ldlm:
599                 cli->cl_conn_count--;
600                 class_disconnect(*exp);
601                 *exp = NULL;
602         }
603 out_sem:
604         up_write(&cli->cl_sem);
605
606         return rc;
607 }
608 EXPORT_SYMBOL(client_connect_import);
609
610 int client_disconnect_export(struct obd_export *exp)
611 {
612         struct obd_device *obd = class_exp2obd(exp);
613         struct client_obd *cli;
614         struct obd_import *imp;
615         int rc = 0, err;
616         ENTRY;
617
618         if (!obd) {
619                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
620                        exp, exp ? exp->exp_handle.h_cookie : -1);
621                 RETURN(-EINVAL);
622         }
623
624         cli = &obd->u.cli;
625         imp = cli->cl_import;
626
627         down_write(&cli->cl_sem);
628         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
629                 cli->cl_conn_count);
630
631         if (cli->cl_conn_count == 0) {
632                 CERROR("disconnecting disconnected device (%s)\n",
633                        obd->obd_name);
634                 GOTO(out_disconnect, rc = -EINVAL);
635         }
636
637         cli->cl_conn_count--;
638         if (cli->cl_conn_count != 0)
639                 GOTO(out_disconnect, rc = 0);
640
641         /* Mark import deactivated now, so we don't try to reconnect if any
642          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
643          * fully deactivate the import, or that would drop all requests. */
644         spin_lock(&imp->imp_lock);
645         imp->imp_deactive = 1;
646         spin_unlock(&imp->imp_lock);
647
648         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
649          * delete it regardless.  (It's safe to delete an import that was
650          * never added.) */
651         (void)ptlrpc_pinger_del_import(imp);
652
653         if (obd->obd_namespace != NULL) {
654                 /* obd_force == local only */
655                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
656                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
657                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
658         }
659
660         /* There's no need to hold sem while disconnecting an import,
661          * and it may actually cause deadlock in GSS. */
662         up_write(&cli->cl_sem);
663         rc = ptlrpc_disconnect_import(imp, 0);
664         down_write(&cli->cl_sem);
665
666         ptlrpc_invalidate_import(imp);
667
668         EXIT;
669
670 out_disconnect:
671         /* Use server style - class_disconnect should be always called for
672          * o_disconnect. */
673         err = class_disconnect(exp);
674         if (!rc && err)
675                 rc = err;
676
677         up_write(&cli->cl_sem);
678
679         RETURN(rc);
680 }
681 EXPORT_SYMBOL(client_disconnect_export);
682
683 #ifdef HAVE_SERVER_SUPPORT
684 int server_disconnect_export(struct obd_export *exp)
685 {
686         int rc;
687         ENTRY;
688
689         /* Disconnect early so that clients can't keep using export. */
690         rc = class_disconnect(exp);
691         /* Close import to avoid sending any requests. */
692         if (exp->exp_imp_reverse)
693                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
694
695         ldlm_bl_thread_wakeup();
696
697         /* complete all outstanding replies */
698         spin_lock(&exp->exp_lock);
699         while (!list_empty(&exp->exp_outstanding_replies)) {
700                 struct ptlrpc_reply_state *rs =
701                         list_entry(exp->exp_outstanding_replies.next,
702                                        struct ptlrpc_reply_state, rs_exp_list);
703                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
704
705                 spin_lock(&svcpt->scp_rep_lock);
706
707                 list_del_init(&rs->rs_exp_list);
708                 spin_lock(&rs->rs_lock);
709                 ptlrpc_schedule_difficult_reply(rs);
710                 spin_unlock(&rs->rs_lock);
711
712                 spin_unlock(&svcpt->scp_rep_lock);
713         }
714         spin_unlock(&exp->exp_lock);
715
716         RETURN(rc);
717 }
718 EXPORT_SYMBOL(server_disconnect_export);
719
720 /* --------------------------------------------------------------------------
721  * from old lib/target.c
722  * -------------------------------------------------------------------------- */
723
724 static int target_handle_reconnect(struct lustre_handle *conn,
725                                    struct obd_export *exp,
726                                    struct obd_uuid *cluuid)
727 {
728         struct obd_device *target;
729         struct lustre_handle *hdl;
730         cfs_time_t now;
731         cfs_time_t deadline;
732         int timeout;
733         int rc = 0;
734         ENTRY;
735
736         hdl = &exp->exp_imp_reverse->imp_remote_handle;
737         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
738                 conn->cookie = exp->exp_handle.h_cookie;
739                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
740                        " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
741                 RETURN(0);
742         }
743
744         target = exp->exp_obd;
745
746         /* Might be a re-connect after a partition. */
747         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
748                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
749                               "with handle %#llx. Rejecting client "
750                               "with the same UUID trying to reconnect "
751                               "with handle %#llx\n", target->obd_name,
752                               obd_uuid2str(&exp->exp_client_uuid),
753                               obd_export_nid2str(exp),
754                               hdl->cookie, conn->cookie);
755                 memset(conn, 0, sizeof *conn);
756                 /* target_handle_connect() treats EALREADY and
757                  * -EALREADY differently.  -EALREADY is an error
758                  * (same UUID, different handle). */
759                 RETURN(-EALREADY);
760         }
761
762         if (!target->obd_recovering) {
763                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
764                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
765                         obd_export_nid2str(exp));
766                 GOTO(out_already, rc);
767         }
768
769         now = cfs_time_current();
770         deadline = cfs_timer_deadline(&target->obd_recovery_timer);
771         if (cfs_time_before(now, deadline)) {
772                 struct target_distribute_txn_data *tdtd =
773                                         class_exp2tgt(exp)->lut_tdtd;
774                 int size = 0;
775                 int count = 0;
776                 char *buf = NULL;
777
778                 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
779                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
780                         buf = tdtd->tdtd_show_update_logs_retrievers(
781                                 tdtd->tdtd_show_retrievers_cbdata,
782                                 &size, &count);
783
784                 if (count > 0)
785                         LCONSOLE_WARN("%s: Recovery already passed deadline "
786                                       "%d:%.02d. It is due to DNE recovery "
787                                       "failed/stuck on the %d MDT(s):%s. "
788                                       "Please wait until all MDTs recovered "
789                                       "or abort the recovery by force.\n",
790                                       target->obd_name, timeout / 60,
791                                       timeout % 60, count,
792                                       buf ? buf : "unknown (not enough RAM)");
793                 else
794                         LCONSOLE_WARN("%s: Recovery already passed deadline "
795                                       "%d:%.02d. If you do not want to wait "
796                                       "more, please abort the recovery by "
797                                       "force.\n", target->obd_name,
798                                       timeout / 60, timeout % 60);
799
800                 if (buf != NULL)
801                         OBD_FREE(buf, size);
802         } else {
803                 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
804                 LCONSOLE_WARN("%s: Recovery already passed deadline"
805                         " %d:%.02d, It is most likely due to DNE"
806                         " recovery is failed or stuck, please wait a"
807                         " few more minutes or abort the recovery.\n",
808                         target->obd_name, timeout / 60, timeout % 60);
809         }
810
811 out_already:
812         conn->cookie = exp->exp_handle.h_cookie;
813         /* target_handle_connect() treats EALREADY and
814          * -EALREADY differently.  EALREADY means we are
815          * doing a valid reconnect from the same client. */
816         RETURN(EALREADY);
817 }
818
819 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
820                           int error)
821 {
822         struct obd_export *exp = cb_data;
823
824         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
825                obd->obd_name, exp->exp_client_uuid.uuid);
826
827         spin_lock(&exp->exp_lock);
828         exp->exp_need_sync = 0;
829         spin_unlock(&exp->exp_lock);
830         class_export_cb_put(exp);
831 }
832
833 static void
834 check_and_start_recovery_timer(struct obd_device *obd,
835                                struct ptlrpc_request *req, int new_client);
836
837 /**
838  * update flags for import during reconnect process
839  */
840 static int rev_import_flags_update(struct obd_import *revimp,
841                                    struct ptlrpc_request *req)
842 {
843         int rc;
844         struct obd_connect_data *data;
845
846         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
847
848         if (data->ocd_connect_flags & OBD_CONNECT_AT)
849                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
850         else
851                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
852
853         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
854
855         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
856         if (rc) {
857                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
858                         revimp->imp_client->cli_name,
859                         libcfs_id2str(req->rq_peer), rc);
860                 return rc;
861         }
862
863         return 0;
864 }
865
866 /**
867  * Allocate a new reverse import for an export.
868  *
869  * \retval -errno in case error hit
870  * \retval 0 if reverse import correctly init
871  **/
872 int rev_import_init(struct obd_export *export)
873 {
874         struct obd_device *obd = export->exp_obd;
875         struct obd_import *revimp;
876
877         LASSERT(export->exp_imp_reverse == NULL);
878
879         revimp = class_new_import(obd);
880         if (revimp == NULL)
881                 return -ENOMEM;
882
883         revimp->imp_remote_handle.cookie = 0ULL;
884         revimp->imp_client = &obd->obd_ldlm_client;
885         revimp->imp_dlm_fake = 1;
886
887         /* it is safe to connect import in new state as no sends possible */
888         spin_lock(&export->exp_lock);
889         export->exp_imp_reverse = revimp;
890         spin_unlock(&export->exp_lock);
891         class_import_put(revimp);
892
893         return 0;
894 }
895 EXPORT_SYMBOL(rev_import_init);
896
897 /**
898  * Handle reconnect for an export.
899  *
900  * \param exp export to handle reconnect process
901  * \param req client reconnect request
902  *
903  * \retval -rc in case securitfy flavor can't be changed
904  * \retval 0 in case none problems
905  */
906 static int rev_import_reconnect(struct obd_export *exp,
907                                 struct ptlrpc_request *req)
908 {
909         struct obd_import *revimp = exp->exp_imp_reverse;
910         struct lustre_handle *lh;
911         int rc;
912
913         /* avoid sending a request until import flags are changed */
914         ptlrpc_import_enter_resend(revimp);
915
916         if (revimp->imp_connection != NULL)
917                 ptlrpc_connection_put(revimp->imp_connection);
918
919         /*
920          * client from recovery don't have a handle so we need to take from
921          * request. it may produce situation when wrong client connected
922          * to recovery as we trust a client uuid
923          */
924         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
925         revimp->imp_remote_handle = *lh;
926
927         /* unknown versions will be caught in
928          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
929         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
930
931         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
932
933         rc = rev_import_flags_update(revimp, req);
934         if (rc != 0) {
935                 /* it is safe to still be in RECOVERY phase as we are not able
936                  * to setup correct security flavor so requests are not able to
937                  * be delivered correctly */
938                 return rc;
939         }
940
941         /* resend all rpc's via new connection */
942         return ptlrpc_import_recovery_state_machine(revimp);
943 }
944
945 int target_handle_connect(struct ptlrpc_request *req)
946 {
947         struct obd_device *target = NULL;
948         struct obd_export *export = NULL;
949         /* connect handle - filled from target_handle_reconnect in
950          * reconnect case */
951         struct lustre_handle conn;
952         struct lustre_handle *tmp;
953         struct obd_uuid tgtuuid;
954         struct obd_uuid cluuid;
955         char *str;
956         int rc = 0;
957         char *target_start;
958         int target_len;
959         bool     mds_conn = false, lw_client = false;
960         bool     mds_mds_conn = false;
961         bool     new_mds_mds_conn = false;
962         bool     target_referenced = false;
963         struct obd_connect_data *data, *tmpdata;
964         int size, tmpsize;
965         lnet_nid_t *client_nid = NULL;
966         ENTRY;
967
968         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
969
970         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
971         if (str == NULL) {
972                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
973                 GOTO(out, rc = -EINVAL);
974         }
975
976         obd_str2uuid(&tgtuuid, str);
977         target = class_uuid2obd(&tgtuuid);
978         if (!target)
979                 target = class_name2obd(str);
980
981         if (!target) {
982                 deuuidify(str, NULL, &target_start, &target_len);
983                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
984                                    "from %s (no target). If you are running "
985                                    "an HA pair check that the target is "
986                                    "mounted on the other server.\n", str,
987                                    libcfs_nid2str(req->rq_peer.nid));
988                 GOTO(out, rc = -ENODEV);
989         }
990
991         spin_lock(&target->obd_dev_lock);
992         if (target->obd_stopping || !target->obd_set_up) {
993                 spin_unlock(&target->obd_dev_lock);
994
995                 deuuidify(str, NULL, &target_start, &target_len);
996                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
997                               target_len, target_start,
998                               libcfs_nid2str(req->rq_peer.nid),
999                               (target->obd_stopping ?
1000                                "stopping" : "not set up"));
1001                 GOTO(out, rc = -ENODEV);
1002         }
1003
1004         if (target->obd_no_conn) {
1005                 spin_unlock(&target->obd_dev_lock);
1006
1007                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1008                                "from %s\n", target->obd_name,
1009                                libcfs_nid2str(req->rq_peer.nid));
1010                 GOTO(out, rc = -EAGAIN);
1011         }
1012
1013         /* Make sure the target isn't cleaned up while we're here. Yes,
1014          * there's still a race between the above check and our incref here.
1015          * Really, class_uuid2obd should take the ref. */
1016         class_incref(target, __func__, current);
1017         target_referenced = true;
1018
1019         target->obd_conn_inprogress++;
1020         spin_unlock(&target->obd_dev_lock);
1021
1022         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1023         if (str == NULL) {
1024                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1025                 GOTO(out, rc = -EINVAL);
1026         }
1027
1028         obd_str2uuid(&cluuid, str);
1029
1030         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1031         if (tmp == NULL)
1032                 GOTO(out, rc = -EPROTO);
1033
1034         conn = *tmp;
1035
1036         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1037                                     RCL_CLIENT);
1038         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1039         if (!data)
1040                 GOTO(out, rc = -EPROTO);
1041
1042         rc = req_capsule_server_pack(&req->rq_pill);
1043         if (rc)
1044                 GOTO(out, rc);
1045
1046 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1047         /* Don't allow clients to connect that are using old 1.8 format
1048          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1049          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1050          * FULL20 flag should be set on all connections since 2.0, but no
1051          * longer affects behaviour.
1052          *
1053          * Later this check will be disabled and the flag can be retired
1054          * completely once interop with 3.0 is no longer needed.
1055          */
1056         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1057                 GOTO(out, rc = -EPROTO);
1058 #endif
1059
1060         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1061                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1062                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1063                     data->ocd_version > LUSTRE_VERSION_CODE +
1064                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1065                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1066                                   "libclient connection attempt",
1067                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1068                                   "old" : "new",
1069                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1070                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1071                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1072                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1073                         data = req_capsule_server_sized_get(&req->rq_pill,
1074                                                             &RMF_CONNECT_DATA,
1075                                     offsetof(typeof(*data), ocd_version) +
1076                                              sizeof(data->ocd_version));
1077                         if (data) {
1078                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1079                                 data->ocd_version = LUSTRE_VERSION_CODE;
1080                         }
1081                         GOTO(out, rc = -EPROTO);
1082                 }
1083         }
1084
1085         /* Note: lw_client is needed in MDS-MDS failover during update log
1086          * processing, so we needs to allow lw_client to be connected at
1087          * anytime, instead of only the initial connection */
1088         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1089
1090         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1091                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1092                 mds_mds_conn = (data->ocd_connect_flags &
1093                                 OBD_CONNECT_MDS_MDS) != 0;
1094
1095                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1096                  * for Imperative Recovery connection from MGC to MGS.
1097                  *
1098                  * Via check OBD_CONNECT_FID, we can distinguish whether
1099                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1100                  * MGC or MDT. */
1101                 if (!lw_client &&
1102                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1103                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1104                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1105                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1106                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1107                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1108
1109                         /* We do not support the MDT-MDT interoperations with
1110                          * different version MDT because of protocol changes. */
1111                         if (unlikely(major != LUSTRE_MAJOR ||
1112                                      minor != LUSTRE_MINOR ||
1113                                      abs(patch - LUSTRE_PATCH) > 3)) {
1114                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1115                                         "connection from different version MDT "
1116                                         "(%d.%d.%d.%d) %s %s\n",
1117                                         target->obd_name, LUSTRE_MAJOR,
1118                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1119                                         major, minor, patch,
1120                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1121                                         libcfs_nid2str(req->rq_peer.nid), str);
1122
1123                                 GOTO(out, rc = -EPROTO);
1124                         }
1125                 }
1126         }
1127
1128         /* lctl gets a backstage, all-access pass. */
1129         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1130                 goto dont_check_exports;
1131
1132         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1133         if (!export)
1134                 goto no_export;
1135
1136         /* We've found an export in the hash. */
1137
1138         spin_lock(&export->exp_lock);
1139
1140         if (export->exp_connecting) { /* bug 9635, et. al. */
1141                 spin_unlock(&export->exp_lock);
1142                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1143                               export->exp_obd->obd_name, export,
1144                               libcfs_nid2str(req->rq_peer.nid));
1145                 class_export_put(export);
1146                 export = NULL;
1147                 rc = -EALREADY;
1148         } else if ((mds_conn || lw_client ||
1149                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1150                    export->exp_connection != NULL) {
1151                 spin_unlock(&export->exp_lock);
1152                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1153                         /* MDS or LWP reconnected after failover. */
1154                         LCONSOLE_WARN("%s: Received %s connection from "
1155                             "%s, removing former export from %s\n",
1156                             target->obd_name, mds_conn ? "MDS" : "LWP",
1157                             libcfs_nid2str(req->rq_peer.nid),
1158                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1159                 } else {
1160                         /* New MDS connection from the same NID. */
1161                         LCONSOLE_WARN("%s: Received new %s connection from "
1162                                 "%s, removing former export from same NID\n",
1163                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1164                                 libcfs_nid2str(req->rq_peer.nid));
1165                 }
1166
1167                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1168                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1169                         /* Because exports between MDTs will always be
1170                          * kept, let's do not fail such export if they
1171                          * come from the same NID, otherwise it might
1172                          * cause eviction between MDTs, which might
1173                          * cause namespace inconsistency */
1174                         spin_lock(&export->exp_lock);
1175                         export->exp_connecting = 1;
1176                         spin_unlock(&export->exp_lock);
1177                         conn.cookie = export->exp_handle.h_cookie;
1178                         rc = EALREADY;
1179                 } else {
1180                         class_fail_export(export);
1181                         class_export_put(export);
1182                         export = NULL;
1183                         rc = 0;
1184                 }
1185         } else if (export->exp_connection != NULL &&
1186                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1187                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
1188                     MSG_CONNECT_INITIAL)) {
1189                 spin_unlock(&export->exp_lock);
1190                 /* In MDS failover we have static UUID but NID can change. */
1191                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1192                               "existing nid %s is already connected\n",
1193                               target->obd_name, cluuid.uuid,
1194                               libcfs_nid2str(req->rq_peer.nid),
1195                               libcfs_nid2str(
1196                                       export->exp_connection->c_peer.nid));
1197                 rc = -EALREADY;
1198                 class_export_put(export);
1199                 export = NULL;
1200         } else {
1201                 export->exp_connecting = 1;
1202                 spin_unlock(&export->exp_lock);
1203                 LASSERT(export->exp_obd == target);
1204
1205                 rc = target_handle_reconnect(&conn, export, &cluuid);
1206         }
1207
1208         /* If we found an export, we already unlocked. */
1209         if (!export) {
1210 no_export:
1211                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1212         } else if (req->rq_export == NULL &&
1213                    atomic_read(&export->exp_rpc_count) > 0) {
1214                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1215                               "still busy with %d references\n",
1216                               target->obd_name, cluuid.uuid,
1217                               libcfs_nid2str(req->rq_peer.nid),
1218                               atomic_read(&export->exp_refcount));
1219                 GOTO(out, rc = -EBUSY);
1220         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1221                 if (!strstr(cluuid.uuid, "mdt"))
1222                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1223                                       "known client %s (at %s) because it "
1224                                       "is indicating it is a new client",
1225                                       target->obd_name, cluuid.uuid,
1226                                       libcfs_nid2str(req->rq_peer.nid));
1227                 GOTO(out, rc = -EALREADY);
1228         } else {
1229                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1230         }
1231
1232         if (rc < 0) {
1233                 GOTO(out, rc);
1234         }
1235
1236         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1237                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1238               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1239               export, (long)cfs_time_current_sec(),
1240               export ? (long)export->exp_last_request_time : 0);
1241
1242         /* If this is the first time a client connects, reset the recovery
1243          * timer. Discard lightweight connections which might be local. */
1244         if (!lw_client && rc == 0 && target->obd_recovering)
1245                 check_and_start_recovery_timer(target, req, export == NULL);
1246
1247         /* We want to handle EALREADY but *not* -EALREADY from
1248          * target_handle_reconnect(), return reconnection state in a flag. */
1249         if (rc == EALREADY) {
1250                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1251                 rc = 0;
1252         } else {
1253                 LASSERT(rc == 0);
1254         }
1255
1256         /* Tell the client if we support replayable requests. */
1257         if (target->obd_replayable)
1258                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1259         client_nid = &req->rq_peer.nid;
1260
1261         if (export == NULL) {
1262                 /* allow lightweight connections during recovery */
1263                 /* allow "new" MDT to be connected during recovery, since we
1264                  * need retrieve recovery update records from it */
1265                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1266                         cfs_time_t t;
1267                         int     c; /* connected */
1268                         int     i; /* in progress */
1269                         int     k; /* known */
1270                         int     s; /* stale/evicted */
1271
1272                         c = atomic_read(&target->obd_connected_clients);
1273                         i = atomic_read(&target->obd_lock_replay_clients);
1274                         k = target->obd_max_recoverable_clients;
1275                         s = target->obd_stale_clients;
1276                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1277                         t = cfs_time_sub(t, cfs_time_current());
1278                         t = cfs_duration_sec(t);
1279                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1280                                       "(at %s), waiting for %d known clients "
1281                                       "(%d recovered, %d in progress, and %d "
1282                                       "evicted) to recover in %d:%.02d\n",
1283                                       target->obd_name, cluuid.uuid,
1284                                       libcfs_nid2str(req->rq_peer.nid), k,
1285                                       c - i, i, s, (int)t / 60,
1286                                       (int)t % 60);
1287                         rc = -EBUSY;
1288                 } else {
1289 dont_check_exports:
1290                         rc = obd_connect(req->rq_svc_thread->t_env,
1291                                          &export, target, &cluuid, data,
1292                                          client_nid);
1293                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1294                                 lustre_msg_add_op_flags(req->rq_repmsg,
1295                                                         MSG_CONNECT_RECOVERING);
1296                         if (rc == 0) {
1297                                 conn.cookie = export->exp_handle.h_cookie;
1298                                 rc = rev_import_init(export);
1299                         }
1300
1301                         if (mds_mds_conn)
1302                                 new_mds_mds_conn = true;
1303                 }
1304         } else {
1305                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1306                                    export, target, &cluuid, data, client_nid);
1307         }
1308         if (rc)
1309                 GOTO(out, rc);
1310
1311         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1312         data->ocd_instance = target->u.obt.obt_instance;
1313
1314         /* Return only the parts of obd_connect_data that we understand, so the
1315          * client knows that we don't understand the rest. */
1316         if (data) {
1317                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1318                                                RCL_SERVER);
1319                 tmpdata = req_capsule_server_get(&req->rq_pill,
1320                                                  &RMF_CONNECT_DATA);
1321                 /* Don't use struct assignment here, because the client reply
1322                  * buffer may be smaller/larger than the local struct
1323                  * obd_connect_data. */
1324                 memcpy(tmpdata, data, min(tmpsize, size));
1325         }
1326
1327         /* If the client and the server are the same node, we will already
1328          * have an export that really points to the client's DLM export,
1329          * because we have a shared handles table.
1330          *
1331          * XXX this will go away when shaver stops sending the "connect" handle
1332          * in the real "remote handle" field of the request --phik 24 Apr 2003
1333          */
1334         ptlrpc_request_change_export(req, export);
1335
1336         spin_lock(&export->exp_lock);
1337         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1338                 spin_unlock(&export->exp_lock);
1339                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1340                        "or equal conn_cnt: %d >= %d\n",
1341                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1342                        export->exp_conn_cnt,
1343                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1344
1345                 GOTO(out, rc = -EALREADY);
1346         }
1347         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1348         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1349
1350         /* Don't evict liblustre clients for not pinging. */
1351         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1352                 export->exp_libclient = 1;
1353                 spin_unlock(&export->exp_lock);
1354
1355                 spin_lock(&target->obd_dev_lock);
1356                 list_del_init(&export->exp_obd_chain_timed);
1357                 spin_unlock(&target->obd_dev_lock);
1358         } else {
1359                 spin_unlock(&export->exp_lock);
1360         }
1361
1362         if (export->exp_connection != NULL) {
1363                 /* Check to see if connection came from another NID. */
1364                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1365                     !hlist_unhashed(&export->exp_nid_hash))
1366                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1367                                      &export->exp_connection->c_peer.nid,
1368                                      &export->exp_nid_hash);
1369
1370                 ptlrpc_connection_put(export->exp_connection);
1371         }
1372
1373         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1374                                                        req->rq_self,
1375                                                        &cluuid);
1376         if (hlist_unhashed(&export->exp_nid_hash))
1377                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1378                              &export->exp_connection->c_peer.nid,
1379                              &export->exp_nid_hash);
1380
1381         lustre_msg_set_handle(req->rq_repmsg, &conn);
1382
1383         rc = rev_import_reconnect(export, req);
1384         if (rc != 0)
1385                 GOTO(out, rc);
1386
1387         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1388                 int has_transno;
1389                 __u64 transno = data->ocd_transno;
1390
1391                 spin_lock(&export->exp_lock);
1392                 /* possible race with class_disconnect_stale_exports,
1393                  * export may be already in the eviction process */
1394                 if (export->exp_failed) {
1395                         spin_unlock(&export->exp_lock);
1396                         GOTO(out, rc = -ENODEV);
1397                 }
1398                 export->exp_in_recovery = 1;
1399                 export->exp_req_replay_needed = 1;
1400                 export->exp_lock_replay_needed = 1;
1401                 spin_unlock(&export->exp_lock);
1402
1403                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1404                                  MSG_CONNECT_TRANSNO);
1405                 if (has_transno && transno == 0)
1406                         CWARN("Connect with zero transno!\n");
1407
1408                 if (has_transno && transno > 0 &&
1409                     transno < target->obd_next_recovery_transno &&
1410                     transno > target->obd_last_committed) {
1411                         /* Another way is to use cmpxchg() to be lock-free. */
1412                         spin_lock(&target->obd_recovery_task_lock);
1413                         if (transno < target->obd_next_recovery_transno)
1414                                 target->obd_next_recovery_transno = transno;
1415                         spin_unlock(&target->obd_recovery_task_lock);
1416                 }
1417
1418                 atomic_inc(&target->obd_req_replay_clients);
1419                 atomic_inc(&target->obd_lock_replay_clients);
1420                 /* Note: MDS-MDS connection is allowed to be connected during
1421                  * recovery, no matter if the exports needs to be recoveried.
1422                  * Because we need retrieve updates logs from all other MDTs.
1423                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1424                  * also needs to be increased to match other recovery checking
1425                  * condition. */
1426                 if (new_mds_mds_conn)
1427                         target->obd_max_recoverable_clients++;
1428                 if (atomic_inc_return(&target->obd_connected_clients) ==
1429                     target->obd_max_recoverable_clients)
1430                         wake_up(&target->obd_next_transno_waitq);
1431         }
1432
1433         /* Tell the client we're in recovery, when client is involved in it. */
1434         if (target->obd_recovering && !lw_client)
1435                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1436
1437 out:
1438         if (export) {
1439                 spin_lock(&export->exp_lock);
1440                 export->exp_connecting = 0;
1441                 spin_unlock(&export->exp_lock);
1442
1443                 class_export_put(export);
1444         }
1445         if (target_referenced == true && target != NULL) {
1446                 spin_lock(&target->obd_dev_lock);
1447                 target->obd_conn_inprogress--;
1448                 spin_unlock(&target->obd_dev_lock);
1449
1450                 class_decref(target, __func__, current);
1451         }
1452         req->rq_status = rc;
1453         RETURN(rc);
1454 }
1455
1456 int target_handle_disconnect(struct ptlrpc_request *req)
1457 {
1458         int rc;
1459         ENTRY;
1460
1461         rc = req_capsule_server_pack(&req->rq_pill);
1462         if (rc)
1463                 RETURN(rc);
1464
1465         /* Keep the rq_export around so we can send the reply. */
1466         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1467
1468         RETURN(0);
1469 }
1470
1471 void target_destroy_export(struct obd_export *exp)
1472 {
1473         struct obd_import       *imp = NULL;
1474         /* exports created from last_rcvd data, and "fake"
1475            exports created by lctl don't have an import */
1476         spin_lock(&exp->exp_lock);
1477         if (exp->exp_imp_reverse != NULL) {
1478                 imp = exp->exp_imp_reverse;
1479                 exp->exp_imp_reverse = NULL;
1480         }
1481         spin_unlock(&exp->exp_lock);
1482         if (imp != NULL)
1483                 client_destroy_import(imp);
1484
1485         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1486         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1487         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1488         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1489 }
1490 EXPORT_SYMBOL(target_destroy_export);
1491
1492 /*
1493  * Recovery functions
1494  */
1495 static void target_request_copy_get(struct ptlrpc_request *req)
1496 {
1497         class_export_rpc_inc(req->rq_export);
1498         LASSERT(list_empty(&req->rq_list));
1499         INIT_LIST_HEAD(&req->rq_replay_list);
1500
1501         /* Increase refcount to keep request in queue. */
1502         atomic_inc(&req->rq_refcount);
1503         /* Let export know it has replays to be handled. */
1504         atomic_inc(&req->rq_export->exp_replay_count);
1505 }
1506
1507 static void target_request_copy_put(struct ptlrpc_request *req)
1508 {
1509         LASSERT(list_empty(&req->rq_replay_list));
1510         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1511
1512         atomic_dec(&req->rq_export->exp_replay_count);
1513         class_export_rpc_dec(req->rq_export);
1514         ptlrpc_server_drop_request(req);
1515 }
1516
1517 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1518 {
1519         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1520         struct obd_export     *exp = req->rq_export;
1521         struct ptlrpc_request *reqiter;
1522         struct ptlrpc_request *dup_req = NULL;
1523         int                    dup = 0;
1524
1525         LASSERT(exp);
1526
1527         spin_lock(&exp->exp_lock);
1528         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1529                                 rq_replay_list) {
1530                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1531                         dup_req = reqiter;
1532                         dup = 1;
1533                         break;
1534                 }
1535         }
1536
1537         if (dup) {
1538                 /* We expect it with RESENT and REPLAY flags. */
1539                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1540                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1541                         CERROR("invalid flags %x of resent replay\n",
1542                                lustre_msg_get_flags(req->rq_reqmsg));
1543
1544                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1545                         __u32 new_conn;
1546
1547                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1548                         if (new_conn >
1549                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1550                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1551                                                         new_conn);
1552                 }
1553         } else {
1554                 list_add_tail(&req->rq_replay_list,
1555                                   &exp->exp_req_replay_queue);
1556         }
1557
1558         spin_unlock(&exp->exp_lock);
1559         return dup;
1560 }
1561
1562 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1563 {
1564         LASSERT(!list_empty(&req->rq_replay_list));
1565         LASSERT(req->rq_export);
1566
1567         spin_lock(&req->rq_export->exp_lock);
1568         list_del_init(&req->rq_replay_list);
1569         spin_unlock(&req->rq_export->exp_lock);
1570 }
1571
1572 static void target_finish_recovery(struct lu_target *lut)
1573 {
1574         struct obd_device *obd = lut->lut_obd;
1575         ENTRY;
1576
1577         /* Only log a recovery message when recovery has occurred. */
1578         if (obd->obd_recovery_start) {
1579                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1580                                         obd->obd_recovery_start);
1581                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1582                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1583                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1584                         obd->obd_max_recoverable_clients,
1585                         atomic_read(&obd->obd_connected_clients),
1586                         obd->obd_stale_clients,
1587                         obd->obd_stale_clients == 1 ? "was" : "were");
1588         }
1589
1590         ldlm_reprocess_all_ns(obd->obd_namespace);
1591         spin_lock(&obd->obd_recovery_task_lock);
1592         if (!list_empty(&obd->obd_req_replay_queue) ||
1593             !list_empty(&obd->obd_lock_replay_queue) ||
1594             !list_empty(&obd->obd_final_req_queue)) {
1595                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1596                        obd->obd_name,
1597                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1598                        list_empty(&obd->obd_lock_replay_queue) ? \
1599                                "" : "lock ",
1600                        list_empty(&obd->obd_final_req_queue) ? \
1601                                "" : "final ");
1602                 spin_unlock(&obd->obd_recovery_task_lock);
1603                 LBUG();
1604         }
1605         spin_unlock(&obd->obd_recovery_task_lock);
1606
1607         obd->obd_recovery_end = cfs_time_current_sec();
1608
1609         /* When recovery finished, cleanup orphans on MDS and OST. */
1610         if (OBT(obd) && OBP(obd, postrecov)) {
1611                 int rc = OBP(obd, postrecov)(obd);
1612                 if (rc < 0)
1613                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1614                                       obd->obd_name, rc);
1615         }
1616         EXIT;
1617 }
1618
1619 static void abort_req_replay_queue(struct obd_device *obd)
1620 {
1621         struct ptlrpc_request *req, *n;
1622         struct list_head abort_list;
1623
1624         INIT_LIST_HEAD(&abort_list);
1625         spin_lock(&obd->obd_recovery_task_lock);
1626         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1627         spin_unlock(&obd->obd_recovery_task_lock);
1628         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1629                 DEBUG_REQ(D_WARNING, req, "aborted:");
1630                 req->rq_status = -ENOTCONN;
1631                 if (ptlrpc_error(req)) {
1632                         DEBUG_REQ(D_ERROR, req,
1633                                   "failed abort_req_reply; skipping");
1634                 }
1635                 target_exp_dequeue_req_replay(req);
1636                 target_request_copy_put(req);
1637         }
1638 }
1639
1640 static void abort_lock_replay_queue(struct obd_device *obd)
1641 {
1642         struct ptlrpc_request *req, *n;
1643         struct list_head abort_list;
1644
1645         INIT_LIST_HEAD(&abort_list);
1646         spin_lock(&obd->obd_recovery_task_lock);
1647         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1648         spin_unlock(&obd->obd_recovery_task_lock);
1649         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1650                 DEBUG_REQ(D_ERROR, req, "aborted:");
1651                 req->rq_status = -ENOTCONN;
1652                 if (ptlrpc_error(req)) {
1653                         DEBUG_REQ(D_ERROR, req,
1654                                   "failed abort_lock_reply; skipping");
1655                 }
1656                 target_request_copy_put(req);
1657         }
1658 }
1659
1660 /* Called from a cleanup function if the device is being cleaned up
1661    forcefully.  The exports should all have been disconnected already,
1662    the only thing left to do is
1663      - clear the recovery flags
1664      - cancel the timer
1665      - free queued requests and replies, but don't send replies
1666    Because the obd_stopping flag is set, no new requests should be received.
1667
1668 */
1669 void target_cleanup_recovery(struct obd_device *obd)
1670 {
1671         struct ptlrpc_request *req, *n;
1672         struct list_head clean_list;
1673
1674         INIT_LIST_HEAD(&clean_list);
1675         spin_lock(&obd->obd_dev_lock);
1676         if (!obd->obd_recovering) {
1677                 spin_unlock(&obd->obd_dev_lock);
1678                 EXIT;
1679                 return;
1680         }
1681         obd->obd_recovering = obd->obd_abort_recovery = 0;
1682         spin_unlock(&obd->obd_dev_lock);
1683
1684         spin_lock(&obd->obd_recovery_task_lock);
1685         target_cancel_recovery_timer(obd);
1686         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1687         spin_unlock(&obd->obd_recovery_task_lock);
1688
1689         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1690                 LASSERT(req->rq_reply_state == NULL);
1691                 target_exp_dequeue_req_replay(req);
1692                 target_request_copy_put(req);
1693         }
1694
1695         spin_lock(&obd->obd_recovery_task_lock);
1696         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1697         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1698         spin_unlock(&obd->obd_recovery_task_lock);
1699
1700         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1701                 LASSERT(req->rq_reply_state == NULL);
1702                 target_request_copy_put(req);
1703         }
1704
1705         EXIT;
1706 }
1707 EXPORT_SYMBOL(target_cleanup_recovery);
1708
1709 /* obd_recovery_task_lock should be held */
1710 void target_cancel_recovery_timer(struct obd_device *obd)
1711 {
1712         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1713         cfs_timer_disarm(&obd->obd_recovery_timer);
1714 }
1715
1716 static void target_start_recovery_timer(struct obd_device *obd)
1717 {
1718         if (obd->obd_recovery_start != 0)
1719                 return;
1720
1721         spin_lock(&obd->obd_dev_lock);
1722         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1723                 spin_unlock(&obd->obd_dev_lock);
1724                 return;
1725         }
1726
1727         LASSERT(obd->obd_recovery_timeout != 0);
1728
1729         if (obd->obd_recovery_start != 0) {
1730                 spin_unlock(&obd->obd_dev_lock);
1731                 return;
1732         }
1733
1734         cfs_timer_arm(&obd->obd_recovery_timer,
1735                       cfs_time_shift(obd->obd_recovery_timeout));
1736         obd->obd_recovery_start = cfs_time_current_sec();
1737         spin_unlock(&obd->obd_dev_lock);
1738
1739         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1740                       "or until %d client%s reconnect%s\n",
1741                       obd->obd_name,
1742                       obd->obd_recovery_timeout / 60,
1743                       obd->obd_recovery_timeout % 60,
1744                       obd->obd_max_recoverable_clients,
1745                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1746                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1747 }
1748
1749 /**
1750  * extend recovery window.
1751  *
1752  * if @extend is true, extend recovery window to have @drt remaining at least;
1753  * otherwise, make sure the recovery timeout value is not less than @drt.
1754  */
1755 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1756 {
1757         cfs_time_t now;
1758         cfs_time_t end;
1759         cfs_duration_t left;
1760         int to;
1761
1762         spin_lock(&obd->obd_dev_lock);
1763         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1764                 spin_unlock(&obd->obd_dev_lock);
1765                 return;
1766         }
1767         LASSERT(obd->obd_recovery_start != 0);
1768
1769         now  = cfs_time_current_sec();
1770         to   = obd->obd_recovery_timeout;
1771         end  = obd->obd_recovery_start + to;
1772         left = cfs_time_sub(end, now);
1773
1774         if (extend && (drt > left)) {
1775                 to += drt - left;
1776         } else if (!extend && (drt > to)) {
1777                 to = drt;
1778         }
1779
1780         if (to > obd->obd_recovery_time_hard)
1781                 to = obd->obd_recovery_time_hard;
1782         if (obd->obd_recovery_timeout < to) {
1783                 obd->obd_recovery_timeout = to;
1784                 end = obd->obd_recovery_start + to;
1785                 cfs_timer_arm(&obd->obd_recovery_timer,
1786                                 cfs_time_shift(end - now));
1787         }
1788         spin_unlock(&obd->obd_dev_lock);
1789
1790         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1791                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1792 }
1793
1794 /* Reset the timer with each new client connection */
1795 /*
1796  * This timer is actually reconnect_timer, which is for making sure
1797  * the total recovery window is at least as big as my reconnect
1798  * attempt timing. So the initial recovery time_out will be set to
1799  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1800  * from client is bigger than this, then the recovery time_out will
1801  * be extended to make sure the client could be reconnected, in the
1802  * process, the timeout from the new client should be ignored.
1803  */
1804
1805 static void
1806 check_and_start_recovery_timer(struct obd_device *obd,
1807                                struct ptlrpc_request *req,
1808                                int new_client)
1809 {
1810         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1811         struct obd_device_target *obt = &obd->u.obt;
1812
1813         if (!new_client && service_time)
1814                 /* Teach server about old server's estimates, as first guess
1815                  * at how long new requests will take. */
1816                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1817                             service_time);
1818
1819         target_start_recovery_timer(obd);
1820
1821         /* Convert the service time to RPC timeout,
1822          * and reuse service_time to limit stack usage. */
1823         service_time = at_est2timeout(service_time);
1824
1825         /* We expect other clients to timeout within service_time, then try
1826          * to reconnect, then try the failover server.  The max delay between
1827          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1828         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1829
1830         LASSERT(obt->obt_magic == OBT_MAGIC);
1831         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1832         if (service_time > obd->obd_recovery_timeout && !new_client)
1833                 extend_recovery_timer(obd, service_time, false);
1834 }
1835
1836 /** Health checking routines */
1837 static inline int exp_connect_healthy(struct obd_export *exp)
1838 {
1839         return (exp->exp_in_recovery);
1840 }
1841
1842 /** if export done req_replay or has replay in queue */
1843 static inline int exp_req_replay_healthy(struct obd_export *exp)
1844 {
1845         return (!exp->exp_req_replay_needed ||
1846                 atomic_read(&exp->exp_replay_count) > 0);
1847 }
1848
1849
1850 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1851 {
1852         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1853                exp_req_replay_healthy(exp);
1854 }
1855
1856 /** if export done lock_replay or has replay in queue */
1857 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1858 {
1859         return (!exp->exp_lock_replay_needed ||
1860                 atomic_read(&exp->exp_replay_count) > 0);
1861 }
1862
1863 static inline int exp_vbr_healthy(struct obd_export *exp)
1864 {
1865         return (!exp->exp_vbr_failed);
1866 }
1867
1868 static inline int exp_finished(struct obd_export *exp)
1869 {
1870         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1871 }
1872
1873 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1874 {
1875         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1876                 exp_finished(exp);
1877 }
1878
1879 static int check_for_next_transno(struct lu_target *lut)
1880 {
1881         struct ptlrpc_request *req = NULL;
1882         struct obd_device *obd = lut->lut_obd;
1883         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1884         int wake_up = 0, connected, completed, queue_len;
1885         __u64 req_transno = 0;
1886         __u64 update_transno = 0;
1887         __u64 next_transno = 0;
1888         ENTRY;
1889
1890         spin_lock(&obd->obd_recovery_task_lock);
1891         if (!list_empty(&obd->obd_req_replay_queue)) {
1892                 req = list_entry(obd->obd_req_replay_queue.next,
1893                                      struct ptlrpc_request, rq_list);
1894                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1895         }
1896
1897         if (tdtd != NULL)
1898                 update_transno = distribute_txn_get_next_transno(tdtd);
1899
1900         connected = atomic_read(&obd->obd_connected_clients);
1901         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1902         queue_len = obd->obd_requests_queued_for_recovery;
1903         next_transno = obd->obd_next_recovery_transno;
1904
1905         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1906                "req_transno: %llu, next_transno: %llu\n",
1907                obd->obd_max_recoverable_clients, connected, completed,
1908                queue_len, req_transno, next_transno);
1909
1910         if (obd->obd_abort_recovery) {
1911                 CDEBUG(D_HA, "waking for aborted recovery\n");
1912                 wake_up = 1;
1913         } else if (obd->obd_recovery_expired) {
1914                 CDEBUG(D_HA, "waking for expired recovery\n");
1915                 wake_up = 1;
1916         } else if (tdtd != NULL && req != NULL &&
1917                    is_req_replayed_by_update(req)) {
1918                 LASSERTF(req_transno < next_transno, "req_transno %llu"
1919                          "next_transno%llu\n", req_transno, next_transno);
1920                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1921                        req_transno);
1922                 wake_up = 1;
1923         } else if (req_transno == next_transno ||
1924                    (update_transno != 0 && update_transno <= next_transno)) {
1925                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1926                 wake_up = 1;
1927         } else if (queue_len > 0 &&
1928                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1929                 /** handle gaps occured due to lost reply or VBR */
1930                 LASSERTF(req_transno >= next_transno,
1931                          "req_transno: %llu, next_transno: %llu\n",
1932                          req_transno, next_transno);
1933                 CDEBUG(D_HA,
1934                        "%s: waking for gap in transno, VBR is %s (skip: "
1935                        "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1936                        ", next_update %lld last_committed: %lld)\n",
1937                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1938                        next_transno, queue_len, completed, connected,
1939                        req_transno, update_transno, obd->obd_last_committed);
1940                 obd->obd_next_recovery_transno = req_transno;
1941                 wake_up = 1;
1942         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1943                 CDEBUG(D_HA, "waking for completed recovery\n");
1944                 wake_up = 1;
1945         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1946                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1947                        " by fail_lock, waking up (%lld)\n", next_transno);
1948                 obd->obd_next_recovery_transno = req_transno;
1949                 wake_up = 1;
1950         }
1951         spin_unlock(&obd->obd_recovery_task_lock);
1952         return wake_up;
1953 }
1954
1955 static int check_for_next_lock(struct lu_target *lut)
1956 {
1957         struct obd_device *obd = lut->lut_obd;
1958         int wake_up = 0;
1959
1960         spin_lock(&obd->obd_recovery_task_lock);
1961         if (!list_empty(&obd->obd_lock_replay_queue)) {
1962                 CDEBUG(D_HA, "waking for next lock\n");
1963                 wake_up = 1;
1964         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1965                 CDEBUG(D_HA, "waking for completed lock replay\n");
1966                 wake_up = 1;
1967         } else if (obd->obd_abort_recovery) {
1968                 CDEBUG(D_HA, "waking for aborted recovery\n");
1969                 wake_up = 1;
1970         } else if (obd->obd_recovery_expired) {
1971                 CDEBUG(D_HA, "waking for expired recovery\n");
1972                 wake_up = 1;
1973         }
1974         spin_unlock(&obd->obd_recovery_task_lock);
1975
1976         return wake_up;
1977 }
1978
1979 /**
1980  * wait for recovery events,
1981  * check its status with help of check_routine
1982  * evict dead clients via health_check
1983  */
1984 static int target_recovery_overseer(struct lu_target *lut,
1985                                     int (*check_routine)(struct lu_target *),
1986                                     int (*health_check)(struct obd_export *))
1987 {
1988         struct obd_device       *obd = lut->lut_obd;
1989         struct target_distribute_txn_data *tdtd;
1990 repeat:
1991         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1992               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1993                 __u64 next_update_transno = 0;
1994
1995                 /* Only abort the recovery if there are no update recovery
1996                  * left in the queue */
1997                 spin_lock(&obd->obd_recovery_task_lock);
1998                 if (lut->lut_tdtd != NULL) {
1999                         next_update_transno =
2000                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2001
2002                         tdtd = lut->lut_tdtd;
2003                         /* If next_update_transno == 0, it probably because
2004                          * updatelog retrieve threads did not get any records
2005                          * yet, let's wait those threads stopped */
2006                         if (next_update_transno == 0) {
2007                                 struct l_wait_info lwi = { 0 };
2008
2009                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2010                                        atomic_read(
2011                                        &tdtd->tdtd_recovery_threads_count) == 0,
2012                                        &lwi);
2013
2014                                 next_update_transno =
2015                                         distribute_txn_get_next_transno(
2016                                                                 lut->lut_tdtd);
2017                         }
2018                 }
2019
2020                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2021                         obd->obd_next_recovery_transno = next_update_transno;
2022                         spin_unlock(&obd->obd_recovery_task_lock);
2023                         /* Disconnect unfinished exports from clients, and
2024                          * keep connection from MDT to make sure the update
2025                          * recovery will still keep trying until some one
2026                          * manually abort the recovery */
2027                         class_disconnect_stale_exports(obd,
2028                                                 exp_finished_or_from_mdt);
2029                         /* Abort all of replay and replay lock req from
2030                          * clients */
2031                         abort_req_replay_queue(obd);
2032                         abort_lock_replay_queue(obd);
2033                         CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2034                                ")in the queue.\n", obd->obd_name,
2035                                next_update_transno);
2036                 } else {
2037                         obd->obd_abort_recovery = 1;
2038                         spin_unlock(&obd->obd_recovery_task_lock);
2039                         CWARN("%s recovery is aborted by hard timeout\n",
2040                               obd->obd_name);
2041                 }
2042         }
2043
2044         while (wait_event_timeout(obd->obd_next_transno_waitq,
2045                                   check_routine(lut),
2046                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2047                 /* wait indefinitely for event, but don't trigger watchdog */;
2048
2049         if (obd->obd_abort_recovery) {
2050                 CWARN("recovery is aborted, evict exports in recovery\n");
2051                 if (lut->lut_tdtd != NULL) {
2052                         struct l_wait_info lwi = { 0 };
2053
2054                         tdtd = lut->lut_tdtd;
2055                         /* Let's wait all of the update log recovery thread
2056                          * finished */
2057                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2058                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2059                              &lwi);
2060                         /* Then abort the update recovery list */
2061                         dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2062                         dtrq_list_destroy(lut->lut_tdtd);
2063                 }
2064
2065                 /** evict exports which didn't finish recovery yet */
2066                 class_disconnect_stale_exports(obd, exp_finished);
2067                 return 1;
2068         } else if (obd->obd_recovery_expired) {
2069                 obd->obd_recovery_expired = 0;
2070                 /** If some clients died being recovered, evict them */
2071                 LCONSOLE_WARN("%s: recovery is timed out, "
2072                               "evict stale exports\n", obd->obd_name);
2073                 /** evict cexports with no replay in queue, they are stalled */
2074                 class_disconnect_stale_exports(obd, health_check);
2075
2076                 /** continue with VBR */
2077                 spin_lock(&obd->obd_dev_lock);
2078                 obd->obd_version_recov = 1;
2079                 spin_unlock(&obd->obd_dev_lock);
2080                 /**
2081                  * reset timer, recovery will proceed with versions now,
2082                  * timeout is set just to handle reconnection delays
2083                  */
2084                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2085                 /** Wait for recovery events again, after evicting bad clients */
2086                 goto repeat;
2087         }
2088         return 0;
2089 }
2090
2091 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2092 {
2093         struct obd_device       *obd = lut->lut_obd;
2094         struct ptlrpc_request *req = NULL;
2095
2096         CDEBUG(D_HA, "Waiting for lock\n");
2097         if (target_recovery_overseer(lut, check_for_next_lock,
2098                                      exp_lock_replay_healthy))
2099                 abort_lock_replay_queue(obd);
2100
2101         spin_lock(&obd->obd_recovery_task_lock);
2102         if (!list_empty(&obd->obd_lock_replay_queue)) {
2103                 req = list_entry(obd->obd_lock_replay_queue.next,
2104                                      struct ptlrpc_request, rq_list);
2105                 list_del_init(&req->rq_list);
2106                 spin_unlock(&obd->obd_recovery_task_lock);
2107         } else {
2108                 spin_unlock(&obd->obd_recovery_task_lock);
2109                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2110                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2111                 /** evict exports failed VBR */
2112                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2113         }
2114         return req;
2115 }
2116
2117 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2118 {
2119         struct ptlrpc_request *req = NULL;
2120
2121         spin_lock(&obd->obd_recovery_task_lock);
2122         if (!list_empty(&obd->obd_final_req_queue)) {
2123                 req = list_entry(obd->obd_final_req_queue.next,
2124                                      struct ptlrpc_request, rq_list);
2125                 list_del_init(&req->rq_list);
2126                 spin_unlock(&obd->obd_recovery_task_lock);
2127                 if (req->rq_export->exp_in_recovery) {
2128                         spin_lock(&req->rq_export->exp_lock);
2129                         req->rq_export->exp_in_recovery = 0;
2130                         spin_unlock(&req->rq_export->exp_lock);
2131                 }
2132         } else {
2133                 spin_unlock(&obd->obd_recovery_task_lock);
2134         }
2135         return req;
2136 }
2137
2138 static void handle_recovery_req(struct ptlrpc_thread *thread,
2139                                 struct ptlrpc_request *req,
2140                                 svc_handler_t handler)
2141 {
2142         ENTRY;
2143
2144         /**
2145          * export can be evicted during recovery, no need to handle replays for
2146          * it after that, discard such request silently
2147          */
2148         if (req->rq_export->exp_disconnected)
2149                 RETURN_EXIT;
2150
2151         req->rq_session.lc_thread = thread;
2152         req->rq_svc_thread = thread;
2153         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2154
2155         /* thread context */
2156         lu_context_enter(&thread->t_env->le_ctx);
2157         (void)handler(req);
2158         lu_context_exit(&thread->t_env->le_ctx);
2159
2160         /* don't reset timer for final stage */
2161         if (!exp_finished(req->rq_export)) {
2162                 int to = obd_timeout;
2163
2164                 /**
2165                  * Add request timeout to the recovery time so next request from
2166                  * this client may come in recovery time
2167                  */
2168                 if (!AT_OFF) {
2169                         struct ptlrpc_service_part *svcpt;
2170
2171                         svcpt = req->rq_rqbd->rqbd_svcpt;
2172                         /* If the server sent early reply for this request,
2173                          * the client will recalculate the timeout according to
2174                          * current server estimate service time, so we will
2175                          * use the maxium timeout here for waiting the client
2176                          * sending the next req */
2177                         to = max((int)at_est2timeout(
2178                                  at_get(&svcpt->scp_at_estimate)),
2179                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2180                         /* Add 2 net_latency, one for balance rq_deadline
2181                          * (see ptl_send_rpc), one for resend the req to server,
2182                          * Note: client will pack net_latency in replay req
2183                          * (see ptlrpc_replay_req) */
2184                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2185                 }
2186                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2187         }
2188         EXIT;
2189 }
2190
2191 /** Checking routines for recovery */
2192 static int check_for_recovery_ready(struct lu_target *lut)
2193 {
2194         struct obd_device *obd = lut->lut_obd;
2195         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2196
2197         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2198                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2199                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2200                obd->obd_recovery_expired);
2201
2202         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2203                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2204                 if (clnts + obd->obd_stale_clients <
2205                     obd->obd_max_recoverable_clients)
2206                         return 0;
2207         }
2208
2209         if (lut->lut_tdtd != NULL) {
2210                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2211                     !obd->obd_abort_recovery) {
2212                         /* Let's extend recovery timer, in case the recovery
2213                          * timer expired, and some clients got evicted */
2214                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2215                                               true);
2216                         CDEBUG(D_HA, "%s update recovery is not ready,"
2217                                " extend recovery %d\n", obd->obd_name,
2218                                obd->obd_recovery_timeout);
2219                         return 0;
2220                 } else {
2221                         dtrq_list_dump(lut->lut_tdtd, D_HA);
2222                 }
2223         }
2224
2225         return 1;
2226 }
2227
2228 enum {
2229         REQUEST_RECOVERY = 1,
2230         UPDATE_RECOVERY = 2,
2231 };
2232
2233 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2234 {
2235         __u64 transno = 0;
2236
2237         if (!list_empty(&obd->obd_req_replay_queue)) {
2238                 struct ptlrpc_request *req;
2239
2240                 req = list_entry(obd->obd_req_replay_queue.next,
2241                                  struct ptlrpc_request, rq_list);
2242                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2243         }
2244
2245         return transno;
2246 }
2247
2248 static __u64 get_next_transno(struct lu_target *lut, int *type)
2249 {
2250         struct obd_device *obd = lut->lut_obd;
2251         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2252         __u64 transno = 0;
2253         __u64 update_transno;
2254         ENTRY;
2255
2256         transno = get_next_replay_req_transno(obd);
2257         if (type != NULL)
2258                 *type = REQUEST_RECOVERY;
2259
2260         if (tdtd == NULL)
2261                 RETURN(transno);
2262
2263         update_transno = distribute_txn_get_next_transno(tdtd);
2264         if (transno == 0 || (transno >= update_transno &&
2265                              update_transno != 0)) {
2266                 transno = update_transno;
2267                 if (type != NULL)
2268                         *type = UPDATE_RECOVERY;
2269         }
2270
2271         RETURN(transno);
2272 }
2273
2274 /**
2275  * drop duplicate replay request
2276  *
2277  * Because the operation has been replayed by update recovery, the request
2278  * with the same transno will be dropped and also notify the client to send
2279  * next replay request.
2280  *
2281  * \param[in] env       execution environment
2282  * \param[in] obd       failover obd device
2283  * \param[in] req       request to be dropped
2284  */
2285 static void drop_duplicate_replay_req(struct lu_env *env,
2286                                       struct obd_device *obd,
2287                                       struct ptlrpc_request *req)
2288 {
2289         DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2290                   " update records are found.\n",
2291                   lustre_msg_get_transno(req->rq_reqmsg),
2292                   libcfs_nid2str(req->rq_peer.nid));
2293
2294         /* Right now, only for MDS reint operation update replay and
2295          * normal request replay can have the same transno */
2296         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2297                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2298                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2299                 if (likely(req->rq_export))
2300                         target_committed_to_req(req);
2301                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2302                 target_send_reply(req, req->rq_status, 0);
2303         } else {
2304                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2305                 libcfs_nid2str(req->rq_peer.nid));
2306         }
2307         target_exp_dequeue_req_replay(req);
2308         target_request_copy_put(req);
2309         obd->obd_replayed_requests++;
2310 }
2311
2312 static void replay_request_or_update(struct lu_env *env,
2313                                      struct lu_target *lut,
2314                                      struct target_recovery_data *trd,
2315                                      struct ptlrpc_thread *thread)
2316 {
2317         struct obd_device *obd = lut->lut_obd;
2318         struct ptlrpc_request *req = NULL;
2319         int                     type;
2320         __u64                   transno;
2321         ENTRY;
2322
2323         CDEBUG(D_HA, "Waiting for transno %lld\n",
2324                obd->obd_next_recovery_transno);
2325
2326         /* Replay all of request and update by transno */
2327         do {
2328                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2329
2330                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2331
2332                 /** It is needed to extend recovery window above
2333                  *  recovery_time_soft. Extending is possible only in the
2334                  *  end of recovery window (see more details in
2335                  *  handle_recovery_req()).
2336                  */
2337                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2338
2339                 if (target_recovery_overseer(lut, check_for_next_transno,
2340                                         exp_req_replay_healthy_or_from_mdt)) {
2341                         abort_req_replay_queue(obd);
2342                         abort_lock_replay_queue(obd);
2343                         goto abort;
2344                 }
2345
2346                 spin_lock(&obd->obd_recovery_task_lock);
2347                 transno = get_next_transno(lut, &type);
2348                 if (type == REQUEST_RECOVERY && transno != 0) {
2349                         /* Drop replay request from client side, if the
2350                          * replay has been executed by update with the
2351                          * same transno */
2352                         req = list_entry(obd->obd_req_replay_queue.next,
2353                                         struct ptlrpc_request, rq_list);
2354
2355                         list_del_init(&req->rq_list);
2356                         obd->obd_requests_queued_for_recovery--;
2357                         spin_unlock(&obd->obd_recovery_task_lock);
2358
2359                         /* Let's check if the request has been redone by
2360                          * update replay */
2361                         if (is_req_replayed_by_update(req)) {
2362                                 struct distribute_txn_replay_req *dtrq;
2363
2364                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2365                                                                    req->rq_xid);
2366                                 LASSERT(dtrq != NULL);
2367                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2368                                 list_del_init(&dtrq->dtrq_list);
2369                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2370                                 dtrq_destroy(dtrq);
2371
2372                                 drop_duplicate_replay_req(env, obd, req);
2373
2374                                 continue;
2375                         }
2376
2377                         LASSERT(trd->trd_processing_task == current_pid());
2378                         DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2379                                   lustre_msg_get_transno(req->rq_reqmsg),
2380                                   libcfs_nid2str(req->rq_peer.nid));
2381
2382                         handle_recovery_req(thread, req,
2383                                             trd->trd_recovery_handler);
2384                         /**
2385                          * bz18031: increase next_recovery_transno before
2386                          * target_request_copy_put() will drop exp_rpc reference
2387                          */
2388                         spin_lock(&obd->obd_recovery_task_lock);
2389                         obd->obd_next_recovery_transno++;
2390                         spin_unlock(&obd->obd_recovery_task_lock);
2391                         target_exp_dequeue_req_replay(req);
2392                         target_request_copy_put(req);
2393                         obd->obd_replayed_requests++;
2394                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2395                         struct distribute_txn_replay_req *dtrq;
2396                         int rc;
2397
2398                         spin_unlock(&obd->obd_recovery_task_lock);
2399
2400                         LASSERT(tdtd != NULL);
2401                         dtrq = distribute_txn_get_next_req(tdtd);
2402                         lu_context_enter(&thread->t_env->le_ctx);
2403                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2404                         lu_context_exit(&thread->t_env->le_ctx);
2405                         extend_recovery_timer(obd, obd_timeout, true);
2406
2407                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2408                                 CDEBUG(D_HA, "Move x%llu t%llu"
2409                                        " to finish list\n", dtrq->dtrq_xid,
2410                                        dtrq->dtrq_master_transno);
2411
2412                                 /* Add it to the replay finish list */
2413                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2414                                 list_add(&dtrq->dtrq_list,
2415                                          &tdtd->tdtd_replay_finish_list);
2416                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2417
2418                                 spin_lock(&obd->obd_recovery_task_lock);
2419                                 if (transno == obd->obd_next_recovery_transno)
2420                                         obd->obd_next_recovery_transno++;
2421                                 else if (transno >
2422                                          obd->obd_next_recovery_transno)
2423                                         obd->obd_next_recovery_transno =
2424                                                                 transno + 1;
2425                                 spin_unlock(&obd->obd_recovery_task_lock);
2426                         } else {
2427                                 dtrq_destroy(dtrq);
2428                         }
2429                 } else {
2430                         spin_unlock(&obd->obd_recovery_task_lock);
2431 abort:
2432                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2433                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2434                         /** evict exports failed VBR */
2435                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2436                         break;
2437                 }
2438         } while (1);
2439 }
2440
2441 static int target_recovery_thread(void *arg)
2442 {
2443         struct lu_target *lut = arg;
2444         struct obd_device *obd = lut->lut_obd;
2445         struct ptlrpc_request *req;
2446         struct target_recovery_data *trd = &obd->obd_recovery_data;
2447         unsigned long delta;
2448         struct lu_env *env;
2449         struct ptlrpc_thread *thread = NULL;
2450         int rc = 0;
2451         ENTRY;
2452
2453         unshare_fs_struct();
2454         OBD_ALLOC_PTR(thread);
2455         if (thread == NULL)
2456                 RETURN(-ENOMEM);
2457
2458         OBD_ALLOC_PTR(env);
2459         if (env == NULL) {
2460                 OBD_FREE_PTR(thread);
2461                 RETURN(-ENOMEM);
2462         }
2463
2464         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2465         if (rc) {
2466                 OBD_FREE_PTR(thread);
2467                 OBD_FREE_PTR(env);
2468                 RETURN(rc);
2469         }
2470
2471         thread->t_env = env;
2472         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2473         env->le_ctx.lc_thread = thread;
2474         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2475         thread->t_watchdog = NULL;
2476
2477         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2478                current_pid());
2479         trd->trd_processing_task = current_pid();
2480
2481         spin_lock(&obd->obd_dev_lock);
2482         obd->obd_recovering = 1;
2483         spin_unlock(&obd->obd_dev_lock);
2484         complete(&trd->trd_starting);
2485
2486         /* first of all, we have to know the first transno to replay */
2487         if (target_recovery_overseer(lut, check_for_recovery_ready,
2488                                      exp_connect_healthy)) {
2489                 abort_req_replay_queue(obd);
2490                 abort_lock_replay_queue(obd);
2491                 if (lut->lut_tdtd != NULL)
2492                         dtrq_list_destroy(lut->lut_tdtd);
2493         }
2494
2495         /* next stage: replay requests or update */
2496         delta = jiffies;
2497         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2498                atomic_read(&obd->obd_req_replay_clients),
2499                obd->obd_next_recovery_transno);
2500         replay_request_or_update(env, lut, trd, thread);
2501
2502         /**
2503          * The second stage: replay locks
2504          */
2505         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2506                atomic_read(&obd->obd_lock_replay_clients));
2507         while ((req = target_next_replay_lock(lut))) {
2508                 LASSERT(trd->trd_processing_task == current_pid());
2509                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2510                           libcfs_nid2str(req->rq_peer.nid));
2511                 handle_recovery_req(thread, req,
2512                                     trd->trd_recovery_handler);
2513                 target_request_copy_put(req);
2514                 obd->obd_replayed_locks++;
2515         }
2516
2517         /**
2518          * The third stage: reply on final pings, at this moment all clients
2519          * must have request in final queue
2520          */
2521         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2522         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2523         /** Update server last boot epoch */
2524         tgt_boot_epoch_update(lut);
2525         /* We drop recoverying flag to forward all new requests
2526          * to regular mds_handle() since now */
2527         spin_lock(&obd->obd_dev_lock);
2528         obd->obd_recovering = obd->obd_abort_recovery = 0;
2529         spin_unlock(&obd->obd_dev_lock);
2530         spin_lock(&obd->obd_recovery_task_lock);
2531         target_cancel_recovery_timer(obd);
2532         spin_unlock(&obd->obd_recovery_task_lock);
2533         while ((req = target_next_final_ping(obd))) {
2534                 LASSERT(trd->trd_processing_task == current_pid());
2535                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2536                           libcfs_nid2str(req->rq_peer.nid));
2537                 handle_recovery_req(thread, req,
2538                                     trd->trd_recovery_handler);
2539                 /* Because the waiting client can not send ping to server,
2540                  * so we need refresh the last_request_time, to avoid the
2541                  * export is being evicted */
2542                 ptlrpc_update_export_timer(req->rq_export, 0);
2543                 target_request_copy_put(req);
2544         }
2545
2546         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2547         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2548               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2549         if (delta > OBD_RECOVERY_TIME_SOFT) {
2550                 CWARN("too long recovery - read logs\n");
2551                 libcfs_debug_dumplog();
2552         }
2553
2554         target_finish_recovery(lut);
2555
2556         lu_context_fini(&env->le_ctx);
2557         trd->trd_processing_task = 0;
2558         complete(&trd->trd_finishing);
2559
2560         tgt_io_thread_done(thread);
2561         OBD_FREE_PTR(thread);
2562         OBD_FREE_PTR(env);
2563         RETURN(rc);
2564 }
2565
2566 static int target_start_recovery_thread(struct lu_target *lut,
2567                                         svc_handler_t handler)
2568 {
2569         struct obd_device *obd = lut->lut_obd;
2570         int rc = 0;
2571         struct target_recovery_data *trd = &obd->obd_recovery_data;
2572         int index;
2573
2574         memset(trd, 0, sizeof(*trd));
2575         init_completion(&trd->trd_starting);
2576         init_completion(&trd->trd_finishing);
2577         trd->trd_recovery_handler = handler;
2578
2579         rc = server_name2index(obd->obd_name, &index, NULL);
2580         if (rc < 0)
2581                 return rc;
2582
2583         if (!IS_ERR(kthread_run(target_recovery_thread,
2584                                 lut, "tgt_recover_%d", index))) {
2585                 wait_for_completion(&trd->trd_starting);
2586                 LASSERT(obd->obd_recovering != 0);
2587         } else {
2588                 rc = -ECHILD;
2589         }
2590
2591         return rc;
2592 }
2593
2594 void target_stop_recovery_thread(struct obd_device *obd)
2595 {
2596         if (obd->obd_recovery_data.trd_processing_task > 0) {
2597                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2598                 /** recovery can be done but postrecovery is not yet */
2599                 spin_lock(&obd->obd_dev_lock);
2600                 if (obd->obd_recovering) {
2601                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2602                         obd->obd_abort_recovery = 1;
2603                         wake_up(&obd->obd_next_transno_waitq);
2604                 }
2605                 spin_unlock(&obd->obd_dev_lock);
2606                 wait_for_completion(&trd->trd_finishing);
2607         }
2608 }
2609 EXPORT_SYMBOL(target_stop_recovery_thread);
2610
2611 void target_recovery_fini(struct obd_device *obd)
2612 {
2613         class_disconnect_exports(obd);
2614         target_stop_recovery_thread(obd);
2615         target_cleanup_recovery(obd);
2616 }
2617 EXPORT_SYMBOL(target_recovery_fini);
2618
2619 static void target_recovery_expired(unsigned long castmeharder)
2620 {
2621         struct obd_device *obd = (struct obd_device *)castmeharder;
2622         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2623                " after %lds (%d clients connected)\n",
2624                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2625                cfs_time_current_sec()- obd->obd_recovery_start,
2626                atomic_read(&obd->obd_connected_clients));
2627
2628         obd->obd_recovery_expired = 1;
2629         wake_up(&obd->obd_next_transno_waitq);
2630 }
2631
2632 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2633 {
2634         struct obd_device *obd = lut->lut_obd;
2635
2636         if (obd->obd_max_recoverable_clients == 0) {
2637                 /** Update server last boot epoch */
2638                 tgt_boot_epoch_update(lut);
2639                 return;
2640         }
2641
2642         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2643                "last_transno %llu\n", obd->obd_name,
2644                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2645         LASSERT(obd->obd_stopping == 0);
2646         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2647         obd->obd_recovery_start = 0;
2648         obd->obd_recovery_end = 0;
2649
2650         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2651         target_start_recovery_thread(lut, handler);
2652 }
2653 EXPORT_SYMBOL(target_recovery_init);
2654
2655 static int target_process_req_flags(struct obd_device *obd,
2656                                     struct ptlrpc_request *req)
2657 {
2658         struct obd_export *exp = req->rq_export;
2659         LASSERT(exp != NULL);
2660         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2661                 /* client declares he's ready to replay locks */
2662                 spin_lock(&exp->exp_lock);
2663                 if (exp->exp_req_replay_needed) {
2664                         exp->exp_req_replay_needed = 0;
2665                         spin_unlock(&exp->exp_lock);
2666
2667                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2668                         atomic_dec(&obd->obd_req_replay_clients);
2669                 } else {
2670                         spin_unlock(&exp->exp_lock);
2671                 }
2672         }
2673         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2674                 /* client declares he's ready to complete recovery
2675                  * so, we put the request on th final queue */
2676                 spin_lock(&exp->exp_lock);
2677                 if (exp->exp_lock_replay_needed) {
2678                         exp->exp_lock_replay_needed = 0;
2679                         spin_unlock(&exp->exp_lock);
2680
2681                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2682                         atomic_dec(&obd->obd_lock_replay_clients);
2683                 } else {
2684                         spin_unlock(&exp->exp_lock);
2685                 }
2686         }
2687         return 0;
2688 }
2689
2690 int target_queue_recovery_request(struct ptlrpc_request *req,
2691                                   struct obd_device *obd)
2692 {
2693         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2694         struct ptlrpc_request *reqiter;
2695         int inserted = 0;
2696         ENTRY;
2697
2698         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2699                 /* Processing the queue right now, don't re-add. */
2700                 RETURN(1);
2701         }
2702
2703         target_process_req_flags(obd, req);
2704
2705         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2706                 /* client declares he's ready to complete recovery
2707                  * so, we put the request on th final queue */
2708                 target_request_copy_get(req);
2709                 DEBUG_REQ(D_HA, req, "queue final req");
2710                 wake_up(&obd->obd_next_transno_waitq);
2711                 spin_lock(&obd->obd_recovery_task_lock);
2712                 if (obd->obd_recovering) {
2713                         struct ptlrpc_request *tmp;
2714                         struct ptlrpc_request *duplicate = NULL;
2715
2716                         if (likely(!req->rq_export->exp_replay_done)) {
2717                                 req->rq_export->exp_replay_done = 1;
2718                                 list_add_tail(&req->rq_list,
2719                                               &obd->obd_final_req_queue);
2720                                 spin_unlock(&obd->obd_recovery_task_lock);
2721                                 RETURN(0);
2722                         }
2723
2724                         /* XXX O(n), but only happens if final ping is
2725                          * timed out, probably reorganize the list as
2726                          * a hash list later */
2727                         list_for_each_entry_safe(reqiter, tmp,
2728                                                  &obd->obd_final_req_queue,
2729                                                  rq_list) {
2730                                 if (reqiter->rq_export == req->rq_export) {
2731                                         list_del_init(&reqiter->rq_list);
2732                                         duplicate = reqiter;
2733                                         break;
2734                                 }
2735                         }
2736
2737                         list_add_tail(&req->rq_list,
2738                                       &obd->obd_final_req_queue);
2739                         req->rq_export->exp_replay_done = 1;
2740                         spin_unlock(&obd->obd_recovery_task_lock);
2741
2742                         if (duplicate != NULL) {
2743                                 DEBUG_REQ(D_HA, duplicate,
2744                                           "put prev final req\n");
2745                                 target_request_copy_put(duplicate);
2746                         }
2747                         RETURN(0);
2748                 } else {
2749                         spin_unlock(&obd->obd_recovery_task_lock);
2750                         target_request_copy_put(req);
2751                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2752                 }
2753         }
2754         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2755                 /* client declares he's ready to replay locks */
2756                 target_request_copy_get(req);
2757                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2758                 wake_up(&obd->obd_next_transno_waitq);
2759                 spin_lock(&obd->obd_recovery_task_lock);
2760                 LASSERT(obd->obd_recovering);
2761                 /* usually due to recovery abort */
2762                 if (!req->rq_export->exp_in_recovery) {
2763                         spin_unlock(&obd->obd_recovery_task_lock);
2764                         target_request_copy_put(req);
2765                         RETURN(-ENOTCONN);
2766                 }
2767                 LASSERT(req->rq_export->exp_lock_replay_needed);
2768                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2769                 spin_unlock(&obd->obd_recovery_task_lock);
2770                 RETURN(0);
2771         }
2772
2773         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2774          * (i.e. buflens etc are in my own byte order), but type-dependent
2775          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2776
2777         if (!transno) {
2778                 INIT_LIST_HEAD(&req->rq_list);
2779                 DEBUG_REQ(D_HA, req, "not queueing");
2780                 RETURN(1);
2781         }
2782
2783         /* If we're processing the queue, we want don't want to queue this
2784          * message.
2785          *
2786          * Also, if this request has a transno less than the one we're waiting
2787          * for, we should process it now.  It could (and currently always will)
2788          * be an open request for a descriptor that was opened some time ago.
2789          *
2790          * Also, a resent, replayed request that has already been
2791          * handled will pass through here and be processed immediately.
2792          */
2793         CDEBUG(D_HA, "Next recovery transno: %llu"
2794                ", current: %llu, replaying\n",
2795                obd->obd_next_recovery_transno, transno);
2796
2797         /* If the request has been replayed by update replay, then sends this
2798          * request to the recovery thread (replay_request_or_update()), where
2799          * it will be handled */
2800         spin_lock(&obd->obd_recovery_task_lock);
2801         if (transno < obd->obd_next_recovery_transno &&
2802             !is_req_replayed_by_update(req)) {
2803                 /* Processing the queue right now, don't re-add. */
2804                 LASSERT(list_empty(&req->rq_list));
2805                 spin_unlock(&obd->obd_recovery_task_lock);
2806                 RETURN(1);
2807         }
2808         spin_unlock(&obd->obd_recovery_task_lock);
2809
2810         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2811                 RETURN(0);
2812
2813         target_request_copy_get(req);
2814         if (!req->rq_export->exp_in_recovery) {
2815                 target_request_copy_put(req);
2816                 RETURN(-ENOTCONN);
2817         }
2818         LASSERT(req->rq_export->exp_req_replay_needed);
2819
2820         if (target_exp_enqueue_req_replay(req)) {
2821                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2822                 target_request_copy_put(req);
2823                 RETURN(0);
2824         }
2825
2826         /* XXX O(n^2) */
2827         spin_lock(&obd->obd_recovery_task_lock);
2828         LASSERT(obd->obd_recovering);
2829         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2830                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2831                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2832                         inserted = 1;
2833                         goto added;
2834                 }
2835
2836                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2837                              transno)) {
2838                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2839                                   "has been claimed by another client");
2840                         spin_unlock(&obd->obd_recovery_task_lock);
2841                         target_exp_dequeue_req_replay(req);
2842                         target_request_copy_put(req);
2843                         RETURN(0);
2844                 }
2845         }
2846 added:
2847         if (!inserted)
2848                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2849
2850         obd->obd_requests_queued_for_recovery++;
2851         spin_unlock(&obd->obd_recovery_task_lock);
2852         wake_up(&obd->obd_next_transno_waitq);
2853         RETURN(0);
2854 }
2855
2856 int target_handle_ping(struct ptlrpc_request *req)
2857 {
2858         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2859         return req_capsule_server_pack(&req->rq_pill);
2860 }
2861
2862 void target_committed_to_req(struct ptlrpc_request *req)
2863 {
2864         struct obd_export *exp = req->rq_export;
2865
2866         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2867                 lustre_msg_set_last_committed(req->rq_repmsg,
2868                                               exp->exp_last_committed);
2869         else
2870                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2871                           "%d)", exp->exp_obd->obd_no_transno,
2872                           req->rq_repmsg == NULL);
2873
2874         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2875                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2876 }
2877
2878 #endif /* HAVE_SERVER_SUPPORT */
2879
2880 /**
2881  * Packs current SLV and Limit into \a req.
2882  */
2883 int target_pack_pool_reply(struct ptlrpc_request *req)
2884 {
2885         struct obd_device *obd;
2886         ENTRY;
2887
2888         /* Check that we still have all structures alive as this may
2889          * be some late RPC at shutdown time. */
2890         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2891                      !exp_connect_lru_resize(req->rq_export))) {
2892                 lustre_msg_set_slv(req->rq_repmsg, 0);
2893                 lustre_msg_set_limit(req->rq_repmsg, 0);
2894                 RETURN(0);
2895         }
2896
2897         /* OBD is alive here as export is alive, which we checked above. */
2898         obd = req->rq_export->exp_obd;
2899
2900         read_lock(&obd->obd_pool_lock);
2901         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2902         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2903         read_unlock(&obd->obd_pool_lock);
2904
2905         RETURN(0);
2906 }
2907
2908 static int target_send_reply_msg(struct ptlrpc_request *req,
2909                                  int rc, int fail_id)
2910 {
2911         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2912                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2913                 return -ECOMM;
2914         }
2915         /* We can have a null rq_reqmsg in the event of bad signature or
2916          * no context when unwrapping */
2917         if (req->rq_reqmsg &&
2918             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2919             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2920                 return -ECOMM;
2921
2922         if (unlikely(rc)) {
2923                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2924                 req->rq_status = rc;
2925                 return ptlrpc_send_error(req, 1);
2926         } else {
2927                 DEBUG_REQ(D_NET, req, "sending reply");
2928         }
2929
2930         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2931 }
2932
2933 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2934 {
2935         struct ptlrpc_service_part *svcpt;
2936         int                        netrc;
2937         struct ptlrpc_reply_state *rs;
2938         struct obd_export         *exp;
2939         ENTRY;
2940
2941         if (req->rq_no_reply) {
2942                 EXIT;
2943                 return;
2944         }
2945
2946         svcpt = req->rq_rqbd->rqbd_svcpt;
2947         rs = req->rq_reply_state;
2948         if (rs == NULL || !rs->rs_difficult) {
2949                 /* no notifiers */
2950                 target_send_reply_msg (req, rc, fail_id);
2951                 EXIT;
2952                 return;
2953         }
2954
2955         /* must be an export if locks saved */
2956         LASSERT(req->rq_export != NULL);
2957         /* req/reply consistent */
2958         LASSERT(rs->rs_svcpt == svcpt);
2959
2960         /* "fresh" reply */
2961         LASSERT(!rs->rs_scheduled);
2962         LASSERT(!rs->rs_scheduled_ever);
2963         LASSERT(!rs->rs_handled);
2964         LASSERT(!rs->rs_on_net);
2965         LASSERT(rs->rs_export == NULL);
2966         LASSERT(list_empty(&rs->rs_obd_list));
2967         LASSERT(list_empty(&rs->rs_exp_list));
2968
2969         exp = class_export_get(req->rq_export);
2970
2971         /* disable reply scheduling while I'm setting up */
2972         rs->rs_scheduled = 1;
2973         rs->rs_on_net    = 1;
2974         rs->rs_xid       = req->rq_xid;
2975         rs->rs_transno   = req->rq_transno;
2976         rs->rs_export    = exp;
2977         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2978
2979         spin_lock(&exp->exp_uncommitted_replies_lock);
2980         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2981                rs->rs_transno, exp->exp_last_committed);
2982         if (rs->rs_transno > exp->exp_last_committed) {
2983                 /* not committed already */
2984                 list_add_tail(&rs->rs_obd_list,
2985                                   &exp->exp_uncommitted_replies);
2986         }
2987         spin_unlock(&exp->exp_uncommitted_replies_lock);
2988
2989         spin_lock(&exp->exp_lock);
2990         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2991         spin_unlock(&exp->exp_lock);
2992
2993         netrc = target_send_reply_msg(req, rc, fail_id);
2994
2995         spin_lock(&svcpt->scp_rep_lock);
2996
2997         atomic_inc(&svcpt->scp_nreps_difficult);
2998
2999         if (netrc != 0) {
3000                 /* error sending: reply is off the net.  Also we need +1
3001                  * reply ref until ptlrpc_handle_rs() is done
3002                  * with the reply state (if the send was successful, there
3003                  * would have been +1 ref for the net, which
3004                  * reply_out_callback leaves alone) */
3005                 rs->rs_on_net = 0;
3006                 ptlrpc_rs_addref(rs);
3007         }
3008
3009         spin_lock(&rs->rs_lock);
3010         if (rs->rs_transno <= exp->exp_last_committed ||
3011             (!rs->rs_on_net && !rs->rs_no_ack) ||
3012             list_empty(&rs->rs_exp_list) ||     /* completed already */
3013             list_empty(&rs->rs_obd_list)) {
3014                 CDEBUG(D_HA, "Schedule reply immediately\n");
3015                 ptlrpc_dispatch_difficult_reply(rs);
3016         } else {
3017                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3018                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3019         }
3020         spin_unlock(&rs->rs_lock);
3021         spin_unlock(&svcpt->scp_rep_lock);
3022         EXIT;
3023 }
3024
3025 enum ldlm_mode lck_compat_array[] = {
3026         [LCK_EX]    = LCK_COMPAT_EX,
3027         [LCK_PW]    = LCK_COMPAT_PW,
3028         [LCK_PR]    = LCK_COMPAT_PR,
3029         [LCK_CW]    = LCK_COMPAT_CW,
3030         [LCK_CR]    = LCK_COMPAT_CR,
3031         [LCK_NL]    = LCK_COMPAT_NL,
3032         [LCK_GROUP] = LCK_COMPAT_GROUP,
3033         [LCK_COS]   = LCK_COMPAT_COS,
3034 };
3035
3036 /**
3037  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3038  * not escape to the user level.
3039  */
3040 int ldlm_error2errno(enum ldlm_error error)
3041 {
3042         int result;
3043
3044         switch (error) {
3045         case ELDLM_OK:
3046         case ELDLM_LOCK_MATCHED:
3047                 result = 0;
3048                 break;
3049         case ELDLM_LOCK_CHANGED:
3050                 result = -ESTALE;
3051                 break;
3052         case ELDLM_LOCK_ABORTED:
3053                 result = -ENAVAIL;
3054                 break;
3055         case ELDLM_LOCK_REPLACED:
3056                 result = -ESRCH;
3057                 break;
3058         case ELDLM_NO_LOCK_DATA:
3059                 result = -ENOENT;
3060                 break;
3061         case ELDLM_NAMESPACE_EXISTS:
3062                 result = -EEXIST;
3063                 break;
3064         case ELDLM_BAD_NAMESPACE:
3065                 result = -EBADF;
3066                 break;
3067         default:
3068                 if (((int)error) < 0) { /* cast to signed type */
3069                         result = error; /* as ldlm_error can be unsigned */
3070                 } else {
3071                         CERROR("Invalid DLM result code: %d\n", error);
3072                         result = -EPROTO;
3073                 }
3074         }
3075         return result;
3076 }
3077 EXPORT_SYMBOL(ldlm_error2errno);
3078
3079 /**
3080  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3081  */
3082 enum ldlm_error ldlm_errno2error(int err_no)
3083 {
3084         int error;
3085
3086         switch (err_no) {
3087         case 0:
3088                 error = ELDLM_OK;
3089                 break;
3090         case -ESTALE:
3091                 error = ELDLM_LOCK_CHANGED;
3092                 break;
3093         case -ENAVAIL:
3094                 error = ELDLM_LOCK_ABORTED;
3095                 break;
3096         case -ESRCH:
3097                 error = ELDLM_LOCK_REPLACED;
3098                 break;
3099         case -ENOENT:
3100                 error = ELDLM_NO_LOCK_DATA;
3101                 break;
3102         case -EEXIST:
3103                 error = ELDLM_NAMESPACE_EXISTS;
3104                 break;
3105         case -EBADF:
3106                 error = ELDLM_BAD_NAMESPACE;
3107                 break;
3108         default:
3109                 error = err_no;
3110         }
3111         return error;
3112 }
3113
3114 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3115 void ldlm_dump_export_locks(struct obd_export *exp)
3116 {
3117         spin_lock(&exp->exp_locks_list_guard);
3118         if (!list_empty(&exp->exp_locks_list)) {
3119                 struct ldlm_lock *lock;
3120
3121                 CERROR("dumping locks for export %p,"
3122                        "ignore if the unmount doesn't hang\n", exp);
3123                 list_for_each_entry(lock, &exp->exp_locks_list,
3124                                         l_exp_refs_link)
3125                         LDLM_ERROR(lock, "lock:");
3126         }
3127         spin_unlock(&exp->exp_locks_list_guard);
3128 }
3129 #endif
3130
3131 #ifdef HAVE_SERVER_SUPPORT
3132 static int target_bulk_timeout(void *data)
3133 {
3134         ENTRY;
3135         /* We don't fail the connection here, because having the export
3136          * killed makes the (vital) call to commitrw very sad.
3137          */
3138         RETURN(1);
3139 }
3140
3141 static inline const char *bulk2type(struct ptlrpc_request *req)
3142 {
3143         if (req->rq_bulk_read)
3144                 return "READ";
3145         if (req->rq_bulk_write)
3146                 return "WRITE";
3147         return "UNKNOWN";
3148 }
3149
3150 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3151                    struct l_wait_info *lwi)
3152 {
3153         struct ptlrpc_request   *req = desc->bd_req;
3154         time_t                   start = cfs_time_current_sec();
3155         time_t                   deadline;
3156         int                      rc = 0;
3157
3158         ENTRY;
3159
3160         /* If there is eviction in progress, wait for it to finish. */
3161         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3162                 *lwi = LWI_INTR(NULL, NULL);
3163                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3164                                   !atomic_read(&exp->exp_obd->
<