Whamcloud - gitweb
LU-14661 obdclass: Add peer/peer NI when processing llog
[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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2010, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  */
31
32 /**
33  * This file deals with various client/target related logic including recovery.
34  *
35  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
36  * should be moved.
37  */
38
39 #define DEBUG_SUBSYSTEM S_LDLM
40
41 #include <cl_object.h>
42 #include <linux/jiffies.h>
43 #include <linux/kernel.h>
44 #include <linux/kthread.h>
45 #include <libcfs/libcfs.h>
46 #include <obd.h>
47 #include <obd_class.h>
48 #include <lustre_dlm.h>
49 #include <lustre_net.h>
50 #include <lustre_sec.h>
51 #include "ldlm_internal.h"
52
53 /*
54  * @priority: If non-zero, move the selected connection to the list head.
55  * @create: If zero, only search in existing connections.
56  */
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58                            int priority, int create)
59 {
60         struct ptlrpc_connection *ptlrpc_conn;
61         struct obd_import_conn *imp_conn = NULL, *item;
62         lnet_nid_t nid4refnet = LNET_NID_ANY;
63         int rc = 0;
64
65         ENTRY;
66
67         if (!create && !priority) {
68                 CDEBUG(D_HA, "Nothing to do\n");
69                 RETURN(-EINVAL);
70         }
71
72         if (imp->imp_connection &&
73             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
74                 /* nid4refnet is used to restrict network connections */
75                 nid4refnet = imp->imp_connection->c_self;
76         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
77         if (!ptlrpc_conn) {
78                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
79                 RETURN(-ENOENT);
80         }
81
82         if (create) {
83                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
84                 if (!imp_conn)
85                         GOTO(out_put, rc = -ENOMEM);
86         }
87
88         spin_lock(&imp->imp_lock);
89         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
90                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
91                         if (priority) {
92                                 list_move(&item->oic_item,
93                                           &imp->imp_conn_list);
94                                 item->oic_last_attempt = 0;
95                         }
96                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
97                                imp, imp->imp_obd->obd_name, uuid->uuid,
98                                (priority ? ", moved to head" : ""));
99                         spin_unlock(&imp->imp_lock);
100                         GOTO(out_free, rc = 0);
101                 }
102         }
103         /* No existing import connection found for \a uuid. */
104         if (create) {
105                 imp_conn->oic_conn = ptlrpc_conn;
106                 imp_conn->oic_uuid = *uuid;
107                 imp_conn->oic_last_attempt = 0;
108                 if (priority)
109                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
110                 else
111                         list_add_tail(&imp_conn->oic_item,
112                                       &imp->imp_conn_list);
113                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
114                        imp, imp->imp_obd->obd_name, uuid->uuid,
115                        (priority ? "head" : "tail"));
116         } else {
117                 spin_unlock(&imp->imp_lock);
118                 GOTO(out_free, rc = -ENOENT);
119         }
120
121         spin_unlock(&imp->imp_lock);
122         RETURN(0);
123 out_free:
124         if (imp_conn)
125                 OBD_FREE(imp_conn, sizeof(*imp_conn));
126 out_put:
127         ptlrpc_connection_put(ptlrpc_conn);
128         RETURN(rc);
129 }
130
131 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
132 {
133         return import_set_conn(imp, uuid, 1, 0);
134 }
135
136 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
137                            int priority)
138 {
139         return import_set_conn(imp, uuid, priority, 1);
140 }
141 EXPORT_SYMBOL(client_import_add_conn);
142
143 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
144                                lnet_nid_t prim_nid, int priority)
145 {
146         struct ptlrpc_connection *ptlrpc_conn;
147         int rc;
148
149         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, prim_nid);
150         if (!ptlrpc_conn) {
151                 const char *str_uuid = obd_uuid2str(uuid);
152
153                 rc = class_add_uuid(str_uuid, prim_nid);
154                 if (rc) {
155                         CERROR("%s: failed to add UUID '%s': rc = %d\n",
156                                imp->imp_obd->obd_name, str_uuid, rc);
157                         return rc;
158                 }
159         }
160         return import_set_conn(imp, uuid, priority, 1);
161 }
162 EXPORT_SYMBOL(client_import_dyn_add_conn);
163
164 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
165                                    int nid_count, struct obd_uuid *uuid)
166 {
167         struct obd_import_conn *conn;
168         int rc = -ENOENT;
169
170         ENTRY;
171         if (nid_count <= 0 || !nids)
172                 return rc;
173
174         spin_lock(&imp->imp_lock);
175         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
176                 if (class_check_uuid(&conn->oic_uuid, nids[0])) {
177                         *uuid = conn->oic_uuid;
178                         spin_unlock(&imp->imp_lock);
179                         rc = class_add_nids_to_uuid(&conn->oic_uuid, nids,
180                                                     nid_count);
181                         RETURN(rc);
182                 }
183         }
184         spin_unlock(&imp->imp_lock);
185         RETURN(rc);
186 }
187 EXPORT_SYMBOL(client_import_add_nids_to_conn);
188
189 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
190 {
191         struct obd_import_conn *imp_conn;
192         struct obd_export *dlmexp;
193         int rc = -ENOENT;
194
195         ENTRY;
196
197         spin_lock(&imp->imp_lock);
198         if (list_empty(&imp->imp_conn_list)) {
199                 LASSERT(!imp->imp_connection);
200                 GOTO(out, rc);
201         }
202
203         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
204                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
205                         continue;
206                 LASSERT(imp_conn->oic_conn);
207
208                 if (imp_conn == imp->imp_conn_current) {
209                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
210
211                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
212                             imp->imp_state != LUSTRE_IMP_DISCON) {
213                                 CERROR("can't remove current connection\n");
214                                 GOTO(out, rc = -EBUSY);
215                         }
216
217                         ptlrpc_connection_put(imp->imp_connection);
218                         imp->imp_connection = NULL;
219
220                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
221                         if (dlmexp && dlmexp->exp_connection) {
222                                 LASSERT(dlmexp->exp_connection ==
223                                         imp_conn->oic_conn);
224                                 ptlrpc_connection_put(dlmexp->exp_connection);
225                                 dlmexp->exp_connection = NULL;
226                         }
227
228                         if (dlmexp != NULL)
229                                 class_export_put(dlmexp);
230                 }
231
232                 list_del(&imp_conn->oic_item);
233                 ptlrpc_connection_put(imp_conn->oic_conn);
234                 OBD_FREE(imp_conn, sizeof(*imp_conn));
235                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
236                        imp, imp->imp_obd->obd_name, uuid->uuid);
237                 rc = 0;
238                 break;
239         }
240 out:
241         spin_unlock(&imp->imp_lock);
242         if (rc == -ENOENT)
243                 CERROR("connection %s not found\n", uuid->uuid);
244         RETURN(rc);
245 }
246 EXPORT_SYMBOL(client_import_del_conn);
247
248 /**
249  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
250  * to find a conn uuid of \a imp which can reach \a peer.
251  */
252 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
253                             struct obd_uuid *uuid)
254 {
255         struct obd_import_conn *conn;
256         int rc = -ENOENT;
257
258         ENTRY;
259
260         spin_lock(&imp->imp_lock);
261         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
262                 /* Check if conn UUID does have this peer NID. */
263                 if (class_check_uuid(&conn->oic_uuid, peer)) {
264                         *uuid = conn->oic_uuid;
265                         rc = 0;
266                         break;
267                 }
268         }
269         spin_unlock(&imp->imp_lock);
270         RETURN(rc);
271 }
272 EXPORT_SYMBOL(client_import_find_conn);
273
274 void client_destroy_import(struct obd_import *imp)
275 {
276         /*
277          * Drop security policy instance after all RPCs have finished/aborted
278          * to let all busy contexts be released.
279          */
280         class_import_get(imp);
281         class_destroy_import(imp);
282         sptlrpc_import_sec_put(imp);
283         class_import_put(imp);
284 }
285 EXPORT_SYMBOL(client_destroy_import);
286
287 /**
288  * Check whether or not the OSC is on MDT.
289  * In the config log,
290  * osc on MDT
291  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
292  * osc on client
293  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
294  *
295  **/
296 static int osc_on_mdt(char *obdname)
297 {
298         char *ptr;
299
300         ptr = strrchr(obdname, '-');
301         if (ptr == NULL)
302                 return 0;
303
304         if (strncmp(ptr + 1, "MDT", 3) == 0)
305                 return 1;
306
307         return 0;
308 }
309
310 /*
311  * Configure an RPC client OBD device.
312  *
313  * lcfg parameters:
314  * 1 - client UUID
315  * 2 - server UUID
316  * 3 - inactive-on-startup
317  * 4 - restrictive net
318  */
319 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
320 {
321         struct client_obd *cli = &obd->u.cli;
322         struct obd_import *imp;
323         struct obd_uuid server_uuid;
324         int rq_portal, rp_portal, connect_op;
325         const char *name = obd->obd_type->typ_name;
326         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
327         char *cli_name = lustre_cfg_buf(lcfg, 0);
328         struct ptlrpc_connection fake_conn = { .c_self = 0,
329                                                .c_remote_uuid.uuid[0] = 0 };
330         int rc;
331
332         ENTRY;
333
334         /*
335          * In a more perfect world, we would hang a ptlrpc_client off of
336          * obd_type and just use the values from there.
337          */
338         if (!strcmp(name, LUSTRE_OSC_NAME)) {
339                 rq_portal = OST_REQUEST_PORTAL;
340                 rp_portal = OSC_REPLY_PORTAL;
341                 connect_op = OST_CONNECT;
342                 cli->cl_sp_me = LUSTRE_SP_CLI;
343                 cli->cl_sp_to = LUSTRE_SP_OST;
344                 ns_type = LDLM_NS_TYPE_OSC;
345         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
346                    !strcmp(name, LUSTRE_LWP_NAME)) {
347                 rq_portal = MDS_REQUEST_PORTAL;
348                 rp_portal = MDC_REPLY_PORTAL;
349                 connect_op = MDS_CONNECT;
350                 if (is_lwp_on_ost(cli_name))
351                         cli->cl_sp_me = LUSTRE_SP_OST;
352                 else if (is_lwp_on_mdt(cli_name))
353                         cli->cl_sp_me = LUSTRE_SP_MDT;
354                 else
355                         cli->cl_sp_me = LUSTRE_SP_CLI;
356                 cli->cl_sp_to = LUSTRE_SP_MDT;
357                 ns_type = LDLM_NS_TYPE_MDC;
358         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
359                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
360                         /* OSP_on_MDT for other MDTs */
361                         connect_op = MDS_CONNECT;
362                         cli->cl_sp_to = LUSTRE_SP_MDT;
363                         ns_type = LDLM_NS_TYPE_MDC;
364                         rq_portal = OUT_PORTAL;
365                 } else {
366                         /* OSP on MDT for OST */
367                         connect_op = OST_CONNECT;
368                         cli->cl_sp_to = LUSTRE_SP_OST;
369                         ns_type = LDLM_NS_TYPE_OSC;
370                         rq_portal = OST_REQUEST_PORTAL;
371                 }
372                 rp_portal = OSC_REPLY_PORTAL;
373                 cli->cl_sp_me = LUSTRE_SP_MDT;
374         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
375                 rq_portal = MGS_REQUEST_PORTAL;
376                 rp_portal = MGC_REPLY_PORTAL;
377                 connect_op = MGS_CONNECT;
378                 cli->cl_sp_me = LUSTRE_SP_MGC;
379                 cli->cl_sp_to = LUSTRE_SP_MGS;
380                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
381                 ns_type = LDLM_NS_TYPE_MGC;
382         } else {
383                 CERROR("unknown client OBD type \"%s\", can't setup\n",
384                        name);
385                 RETURN(-EINVAL);
386         }
387
388         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
389                 CERROR("requires a TARGET UUID\n");
390                 RETURN(-EINVAL);
391         }
392
393         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
394                 CERROR("client UUID must be less than 38 characters\n");
395                 RETURN(-EINVAL);
396         }
397
398         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
399                 CERROR("setup requires a SERVER UUID\n");
400                 RETURN(-EINVAL);
401         }
402
403         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
404                 CERROR("target UUID must be less than 38 characters\n");
405                 RETURN(-EINVAL);
406         }
407
408         init_rwsem(&cli->cl_sem);
409         mutex_init(&cli->cl_mgc_mutex);
410         cli->cl_seq = NULL;
411         init_rwsem(&cli->cl_seq_rwsem);
412         cli->cl_conn_count = 0;
413         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
414                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
415                      sizeof(server_uuid)));
416
417         cli->cl_dirty_pages = 0;
418         cli->cl_dirty_max_pages = 0;
419         cli->cl_avail_grant = 0;
420         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
421         /*
422          * cl_dirty_max_pages may be changed at connect time in
423          * ptlrpc_connect_interpret().
424          */
425         client_adjust_max_dirty(cli);
426         init_waitqueue_head(&cli->cl_cache_waiters);
427         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
428         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
429         INIT_LIST_HEAD(&cli->cl_loi_write_list);
430         INIT_LIST_HEAD(&cli->cl_loi_read_list);
431         spin_lock_init(&cli->cl_loi_list_lock);
432         atomic_set(&cli->cl_pending_w_pages, 0);
433         atomic_set(&cli->cl_pending_r_pages, 0);
434         cli->cl_r_in_flight = 0;
435         cli->cl_w_in_flight = 0;
436
437         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
438         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
439         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
440         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
441         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
442         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
443
444         /* lru for osc. */
445         INIT_LIST_HEAD(&cli->cl_lru_osc);
446         atomic_set(&cli->cl_lru_shrinkers, 0);
447         atomic_long_set(&cli->cl_lru_busy, 0);
448         atomic_long_set(&cli->cl_lru_in_list, 0);
449         INIT_LIST_HEAD(&cli->cl_lru_list);
450         spin_lock_init(&cli->cl_lru_list_lock);
451         atomic_long_set(&cli->cl_unstable_count, 0);
452         INIT_LIST_HEAD(&cli->cl_shrink_list);
453         INIT_LIST_HEAD(&cli->cl_grant_chain);
454
455         INIT_LIST_HEAD(&cli->cl_flight_waiters);
456         cli->cl_rpcs_in_flight = 0;
457
458         init_waitqueue_head(&cli->cl_destroy_waitq);
459         atomic_set(&cli->cl_destroy_in_flight, 0);
460
461
462         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
463         cli->cl_preferred_cksum_type = 0;
464 #ifdef ENABLE_CHECKSUM
465         /* Turn on checksumming by default. */
466         cli->cl_checksum = 1;
467         /*
468          * The supported checksum types will be worked out at connect time
469          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
470          * through procfs.
471          */
472         cli->cl_cksum_type = cli->cl_supp_cksum_types;
473 #endif
474         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
475
476         /*
477          * Set it to possible maximum size. It may be reduced by ocd_brw_size
478          * from OFD after connecting.
479          */
480         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
481
482         cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
483
484         /*
485          * set cl_chunkbits default value to PAGE_SHIFT,
486          * it will be updated at OSC connection time.
487          */
488         cli->cl_chunkbits = PAGE_SHIFT;
489
490         if (!strcmp(name, LUSTRE_MDC_NAME)) {
491                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
492         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
493                 cli->cl_max_rpcs_in_flight = 2;
494         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
495                 cli->cl_max_rpcs_in_flight = 3;
496         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
497                 cli->cl_max_rpcs_in_flight = 4;
498         } else {
499                 if (osc_on_mdt(obd->obd_name))
500                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
501                 else
502                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
503         }
504
505         spin_lock_init(&cli->cl_mod_rpcs_lock);
506         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
507         cli->cl_max_mod_rpcs_in_flight = 0;
508         cli->cl_mod_rpcs_in_flight = 0;
509         cli->cl_close_rpcs_in_flight = 0;
510         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
511         cli->cl_mod_tag_bitmap = NULL;
512
513         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
514
515         if (connect_op == MDS_CONNECT) {
516                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
517                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
518                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
519                 if (cli->cl_mod_tag_bitmap == NULL)
520                         GOTO(err, rc = -ENOMEM);
521         }
522
523         rc = ldlm_get_ref();
524         if (rc) {
525                 CERROR("ldlm_get_ref failed: %d\n", rc);
526                 GOTO(err, rc);
527         }
528
529         ptlrpc_init_client(rq_portal, rp_portal, name,
530                            &obd->obd_ldlm_client);
531
532         imp = class_new_import(obd);
533         if (imp == NULL)
534                 GOTO(err_ldlm, rc = -ENOENT);
535         imp->imp_client = &obd->obd_ldlm_client;
536         imp->imp_connect_op = connect_op;
537         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
538                LUSTRE_CFG_BUFLEN(lcfg, 1));
539         class_import_put(imp);
540
541         if (lustre_cfg_buf(lcfg, 4)) {
542                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
543
544                 if (refnet == LNET_NET_ANY) {
545                         rc = -EINVAL;
546                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
547                                obd->obd_name, lustre_cfg_string(lcfg, 4),
548                                rc);
549                         GOTO(err_import, rc);
550                 }
551                 fake_conn.c_self = LNET_MKNID(refnet, 0);
552                 imp->imp_connection = &fake_conn;
553         }
554
555         rc = client_import_add_conn(imp, &server_uuid, 1);
556         if (rc) {
557                 CERROR("can't add initial connection\n");
558                 GOTO(err_import, rc);
559         }
560         imp->imp_connection = NULL;
561
562         cli->cl_import = imp;
563         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
564         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
565
566         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
567                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
568                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
569                                name, obd->obd_name,
570                                cli->cl_target_uuid.uuid);
571                         spin_lock(&imp->imp_lock);
572                         imp->imp_deactive = 1;
573                         spin_unlock(&imp->imp_lock);
574                 }
575         }
576
577         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
578                                                 LDLM_NAMESPACE_CLIENT,
579                                                 LDLM_NAMESPACE_GREEDY,
580                                                 ns_type);
581         if (IS_ERR(obd->obd_namespace)) {
582                 rc = PTR_ERR(obd->obd_namespace);
583                 CERROR("%s: unable to create client namespace: rc = %d\n",
584                        obd->obd_name, rc);
585                 obd->obd_namespace = NULL;
586                 GOTO(err_import, rc);
587         }
588
589         RETURN(rc);
590
591 err_import:
592         class_destroy_import(imp);
593 err_ldlm:
594         ldlm_put_ref();
595 err:
596         if (cli->cl_mod_tag_bitmap != NULL)
597                 OBD_FREE(cli->cl_mod_tag_bitmap,
598                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
599         cli->cl_mod_tag_bitmap = NULL;
600
601         RETURN(rc);
602 }
603 EXPORT_SYMBOL(client_obd_setup);
604
605 int client_obd_cleanup(struct obd_device *obd)
606 {
607         struct client_obd *cli = &obd->u.cli;
608
609         ENTRY;
610
611         ldlm_namespace_free_post(obd->obd_namespace);
612         obd->obd_namespace = NULL;
613
614         obd_cleanup_client_import(obd);
615         LASSERT(obd->u.cli.cl_import == NULL);
616
617         ldlm_put_ref();
618
619         if (cli->cl_mod_tag_bitmap != NULL)
620                 OBD_FREE(cli->cl_mod_tag_bitmap,
621                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
622         cli->cl_mod_tag_bitmap = NULL;
623
624         RETURN(0);
625 }
626 EXPORT_SYMBOL(client_obd_cleanup);
627
628 /* ->o_connect() method for client side (OSC and MDC and MGC) */
629 int client_connect_import(const struct lu_env *env,
630                           struct obd_export **exp,
631                           struct obd_device *obd, struct obd_uuid *cluuid,
632                           struct obd_connect_data *data, void *localdata)
633 {
634         struct client_obd *cli = &obd->u.cli;
635         struct obd_import *imp = cli->cl_import;
636         struct obd_connect_data *ocd;
637         struct lustre_handle conn = { 0 };
638         int rc;
639
640         ENTRY;
641
642         *exp = NULL;
643         down_write(&cli->cl_sem);
644         if (cli->cl_conn_count > 0)
645                 GOTO(out_sem, rc = -EALREADY);
646
647         rc = class_connect(&conn, obd, cluuid);
648         if (rc)
649                 GOTO(out_sem, rc);
650
651         cli->cl_conn_count++;
652         *exp = class_conn2export(&conn);
653
654         LASSERT(obd->obd_namespace);
655
656         imp->imp_dlm_handle = conn;
657         rc = ptlrpc_init_import(imp);
658         if (rc != 0)
659                 GOTO(out_ldlm, rc);
660
661         ocd = &imp->imp_connect_data;
662         if (data) {
663                 *ocd = *data;
664                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
665                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
666         }
667
668         rc = ptlrpc_connect_import(imp);
669         if (rc != 0) {
670                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
671                 GOTO(out_ldlm, rc);
672         }
673         LASSERT(*exp != NULL && (*exp)->exp_connection);
674
675         if (data) {
676                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
677                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
678                          data->ocd_connect_flags, ocd->ocd_connect_flags);
679                 data->ocd_connect_flags = ocd->ocd_connect_flags;
680                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
681         }
682
683         ptlrpc_pinger_add_import(imp);
684
685         EXIT;
686
687         if (rc) {
688 out_ldlm:
689                 cli->cl_conn_count--;
690                 class_disconnect(*exp);
691                 *exp = NULL;
692         }
693 out_sem:
694         up_write(&cli->cl_sem);
695
696         if (!rc && localdata) {
697                 LASSERT(cli->cl_cache == NULL); /* only once */
698                 cli->cl_cache = (struct cl_client_cache *)localdata;
699                 cl_cache_incref(cli->cl_cache);
700                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
701
702                 /* add this osc into entity list */
703                 LASSERT(list_empty(&cli->cl_lru_osc));
704                 spin_lock(&cli->cl_cache->ccc_lru_lock);
705                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
706                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
707         }
708
709         return rc;
710 }
711 EXPORT_SYMBOL(client_connect_import);
712
713 int client_disconnect_export(struct obd_export *exp)
714 {
715         struct obd_device *obd = class_exp2obd(exp);
716         struct client_obd *cli;
717         struct obd_import *imp;
718         int rc = 0, err;
719
720         ENTRY;
721
722         if (!obd) {
723                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
724                        exp, exp ? exp->exp_handle.h_cookie : -1);
725                 RETURN(-EINVAL);
726         }
727
728         cli = &obd->u.cli;
729         imp = cli->cl_import;
730
731         down_write(&cli->cl_sem);
732         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
733                 cli->cl_conn_count);
734
735         if (cli->cl_conn_count == 0) {
736                 CERROR("disconnecting disconnected device (%s)\n",
737                        obd->obd_name);
738                 GOTO(out_disconnect, rc = -EINVAL);
739         }
740
741         cli->cl_conn_count--;
742         if (cli->cl_conn_count != 0)
743                 GOTO(out_disconnect, rc = 0);
744
745         /*
746          * Mark import deactivated now, so we don't try to reconnect if any
747          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
748          * fully deactivate the import, or that would drop all requests.
749          */
750         spin_lock(&imp->imp_lock);
751         imp->imp_deactive = 1;
752         spin_unlock(&imp->imp_lock);
753
754         /*
755          * Some non-replayable imports (MDS's OSCs) are pinged, so just
756          * delete it regardless.  (It's safe to delete an import that was
757          * never added.)
758          */
759         (void)ptlrpc_pinger_del_import(imp);
760
761         if (obd->obd_namespace != NULL) {
762                 /* obd_force == local only */
763                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
764                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
765                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
766                                           obd->obd_force);
767         }
768
769         /*
770          * There's no need to hold sem while disconnecting an import,
771          * and it may actually cause deadlock in GSS.
772          */
773         up_write(&cli->cl_sem);
774         rc = ptlrpc_disconnect_import(imp, 0);
775         down_write(&cli->cl_sem);
776
777         ptlrpc_invalidate_import(imp);
778
779         EXIT;
780
781 out_disconnect:
782         /*
783          * Use server style - class_disconnect should be always called for
784          * o_disconnect.
785          */
786         err = class_disconnect(exp);
787         if (!rc && err)
788                 rc = err;
789
790         up_write(&cli->cl_sem);
791
792         RETURN(rc);
793 }
794 EXPORT_SYMBOL(client_disconnect_export);
795
796 #ifdef HAVE_SERVER_SUPPORT
797 int server_disconnect_export(struct obd_export *exp)
798 {
799         int rc;
800
801         ENTRY;
802
803         /* Disconnect early so that clients can't keep using export. */
804         rc = class_disconnect(exp);
805         /* Close import to avoid sending any requests. */
806         if (exp->exp_imp_reverse)
807                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
808
809         ldlm_bl_thread_wakeup();
810
811         /* complete all outstanding replies */
812         spin_lock(&exp->exp_lock);
813         while (!list_empty(&exp->exp_outstanding_replies)) {
814                 struct ptlrpc_reply_state *rs =
815                         list_entry(exp->exp_outstanding_replies.next,
816                                        struct ptlrpc_reply_state, rs_exp_list);
817                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
818
819                 spin_lock(&svcpt->scp_rep_lock);
820
821                 list_del_init(&rs->rs_exp_list);
822
823                 spin_lock(&rs->rs_lock);
824                 /* clear rs_convert_lock to make sure rs is handled and put */
825                 rs->rs_convert_lock = 0;
826                 ptlrpc_schedule_difficult_reply(rs);
827                 spin_unlock(&rs->rs_lock);
828
829                 spin_unlock(&svcpt->scp_rep_lock);
830         }
831         spin_unlock(&exp->exp_lock);
832
833         RETURN(rc);
834 }
835 EXPORT_SYMBOL(server_disconnect_export);
836
837 static inline int target_check_recovery_timer(struct obd_device *target)
838 {
839         ktime_t remaining;
840         s64 timeout;
841
842         if (!target->obd_recovering || target->obd_recovery_start == 0)
843                 return 0;
844
845         remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
846         timeout = ktime_divns(remaining, NSEC_PER_SEC);
847         if (timeout > -30)
848                 return 0;
849
850         /* the recovery timer should expire, but it isn't triggered,
851          * it's better to abort the recovery of this target to speed up
852          * the recovery of the whole cluster. */
853         spin_lock(&target->obd_dev_lock);
854         if (target->obd_recovering) {
855                 CERROR("%s: Aborting recovery\n", target->obd_name);
856                 target->obd_abort_recovery = 1;
857                 wake_up(&target->obd_next_transno_waitq);
858         }
859         spin_unlock(&target->obd_dev_lock);
860         return 0;
861 }
862
863 /*
864  * --------------------------------------------------------------------------
865  * from old lib/target.c
866  * --------------------------------------------------------------------------
867  */
868 static int target_handle_reconnect(struct lustre_handle *conn,
869                                    struct obd_export *exp,
870                                    struct obd_uuid *cluuid)
871 {
872         struct obd_device *target;
873         struct lustre_handle *hdl;
874         ktime_t remaining;
875         s64 timeout;
876         int rc = 0;
877
878         ENTRY;
879         hdl = &exp->exp_imp_reverse->imp_remote_handle;
880         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
881                 conn->cookie = exp->exp_handle.h_cookie;
882                 CDEBUG(D_HA,
883                        "connect export for UUID '%s' at %p, cookie %#llx\n",
884                        cluuid->uuid, exp, conn->cookie);
885                 RETURN(0);
886         }
887
888         target = exp->exp_obd;
889
890         /* Might be a re-connect after a partition. */
891         if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
892                 LCONSOLE_WARN("%s: already connected client %s (at %s) with handle %#llx. Rejecting client with the same UUID trying to reconnect with handle %#llx\n",
893                               target->obd_name,
894                               obd_uuid2str(&exp->exp_client_uuid),
895                               obd_export_nid2str(exp),
896                               hdl->cookie, conn->cookie);
897                 memset(conn, 0, sizeof(*conn));
898                 /*
899                  * target_handle_connect() treats EALREADY and
900                  * -EALREADY differently.  -EALREADY is an error
901                  * (same UUID, different handle).
902                  */
903                 RETURN(-EALREADY);
904         }
905
906         if (!target->obd_recovering) {
907                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
908                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
909                         obd_export_nid2str(exp));
910                 GOTO(out_already, rc);
911         }
912
913         remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
914         timeout = ktime_divns(remaining, NSEC_PER_SEC);
915         if (timeout > 0) {
916                 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
917                               target->obd_name,
918                               obd_uuid2str(&exp->exp_client_uuid),
919                               obd_export_nid2str(exp),
920                               atomic_read(&target->obd_max_recoverable_clients),
921                               timeout / 60, timeout % 60);
922         } else {
923                 struct target_distribute_txn_data *tdtd;
924                 int size = 0;
925                 int count = 0;
926                 char *buf = NULL;
927
928                 target_check_recovery_timer(target);
929
930                 tdtd = class_exp2tgt(exp)->lut_tdtd;
931                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
932                         buf = tdtd->tdtd_show_update_logs_retrievers(
933                                 tdtd->tdtd_show_retrievers_cbdata,
934                                 &size, &count);
935
936                 if (count > 0)
937                         LCONSOLE_WARN("%s: Client %s (at %s) reconnecting, waiting for %d MDTs (%s) in recovery for %lld:%.02lld. Please wait until all MDTs recovered or you may force MDT evicition via 'lctl --device %s abort_recovery.\n",
938                                       target->obd_name,
939                                       obd_uuid2str(&exp->exp_client_uuid),
940                                       obd_export_nid2str(exp), count,
941                                       buf ? buf : "unknown (not enough RAM)",
942                                       (abs(timeout) + target->obd_recovery_timeout) / 60,
943                                       (abs(timeout) + target->obd_recovery_timeout) % 60,
944                                       target->obd_name);
945                 else
946                         LCONSOLE_WARN("%s: Recovery already passed deadline %lld:%.02lld. If you do not want to wait more, you may force taget eviction via 'lctl --device %s abort_recovery.\n",
947                                       target->obd_name, abs(timeout) / 60,
948                                       abs(timeout) % 60, target->obd_name);
949
950                 if (buf != NULL)
951                         OBD_FREE(buf, size);
952         }
953
954 out_already:
955         conn->cookie = exp->exp_handle.h_cookie;
956         /*
957          * target_handle_connect() treats EALREADY and
958          * -EALREADY differently.  EALREADY means we are
959          * doing a valid reconnect from the same client.
960          */
961         RETURN(EALREADY);
962 }
963
964 static void
965 check_and_start_recovery_timer(struct obd_device *obd,
966                                struct ptlrpc_request *req, int new_client);
967
968 /**
969  * update flags for import during reconnect process
970  */
971 static int rev_import_flags_update(struct obd_import *revimp,
972                                    struct ptlrpc_request *req)
973 {
974         int rc;
975         struct obd_connect_data *data;
976
977         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
978
979         if (data->ocd_connect_flags & OBD_CONNECT_AT)
980                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
981         else
982                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
983
984         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
985
986         revimp->imp_connect_data = *data;
987         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
988         if (rc) {
989                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
990                         revimp->imp_client->cli_name,
991                         libcfs_id2str(req->rq_peer), rc);
992                 return rc;
993         }
994
995         return 0;
996 }
997
998 /**
999  * Allocate a new reverse import for an export.
1000  *
1001  * \retval -errno in case error hit
1002  * \retval 0 if reverse import correctly init
1003  **/
1004 int rev_import_init(struct obd_export *export)
1005 {
1006         struct obd_device *obd = export->exp_obd;
1007         struct obd_import *revimp;
1008
1009         LASSERT(export->exp_imp_reverse == NULL);
1010
1011         revimp = class_new_import(obd);
1012         if (revimp == NULL)
1013                 return -ENOMEM;
1014
1015         revimp->imp_remote_handle.cookie = 0ULL;
1016         revimp->imp_client = &obd->obd_ldlm_client;
1017         revimp->imp_dlm_fake = 1;
1018
1019         /* it is safe to connect import in new state as no sends possible */
1020         spin_lock(&export->exp_lock);
1021         export->exp_imp_reverse = revimp;
1022         spin_unlock(&export->exp_lock);
1023         class_import_put(revimp);
1024
1025         return 0;
1026 }
1027 EXPORT_SYMBOL(rev_import_init);
1028
1029 /**
1030  * Handle reconnect for an export.
1031  *
1032  * \param exp export to handle reconnect process
1033  * \param req client reconnect request
1034  *
1035  * \retval -rc in case securitfy flavor can't be changed
1036  * \retval 0 in case none problems
1037  */
1038 static int rev_import_reconnect(struct obd_export *exp,
1039                                 struct ptlrpc_request *req)
1040 {
1041         struct obd_import *revimp = exp->exp_imp_reverse;
1042         struct lustre_handle *lh;
1043         int rc;
1044
1045         /* avoid sending a request until import flags are changed */
1046         ptlrpc_import_enter_resend(revimp);
1047
1048         ptlrpc_connection_put(revimp->imp_connection);
1049
1050         /*
1051          * client from recovery don't have a handle so we need to take from
1052          * request. it may produce situation when wrong client connected
1053          * to recovery as we trust a client uuid
1054          */
1055         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1056         revimp->imp_remote_handle = *lh;
1057
1058         /*
1059          * unknown versions will be caught in
1060          * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1061          */
1062         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1063
1064         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1065
1066         rc = rev_import_flags_update(revimp, req);
1067         if (rc != 0) {
1068                 /*
1069                  * it is safe to still be in RECOVERY phase as we are not able
1070                  * to setup correct security flavor so requests are not able to
1071                  * be delivered correctly
1072                  */
1073                 return rc;
1074         }
1075
1076         /* resend all rpc's via new connection */
1077         return ptlrpc_import_recovery_state_machine(revimp);
1078 }
1079
1080 int target_handle_connect(struct ptlrpc_request *req)
1081 {
1082         struct obd_device *target = NULL;
1083         struct obd_export *export = NULL;
1084         /*
1085          * connect handle - filled from target_handle_reconnect in
1086          * reconnect case
1087          */
1088         struct lustre_handle conn;
1089         struct lustre_handle *tmp;
1090         struct obd_uuid cluuid;
1091         char *str;
1092         int rc = 0;
1093         char *target_start;
1094         int target_len;
1095         bool mds_conn = false, lw_client = false, initial_conn = false;
1096         bool mds_mds_conn = false;
1097         bool new_mds_mds_conn = false;
1098         struct obd_connect_data *data, *tmpdata;
1099         int size, tmpsize;
1100         lnet_nid_t *client_nid = NULL;
1101         struct ptlrpc_connection *pcon = NULL;
1102
1103         ENTRY;
1104
1105         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1106
1107         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1108         if (str == NULL) {
1109                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1110                 GOTO(out, rc = -EINVAL);
1111         }
1112
1113         target = class_dev_by_str(str);
1114         if (!target) {
1115                 deuuidify(str, NULL, &target_start, &target_len);
1116                 LCONSOLE_ERROR_MSG(0x137,
1117                                    "%s: not available for connect from %s (no target). If you are running an HA pair check that the target is mounted on the other server.\n",
1118                                    str, libcfs_nid2str(req->rq_peer.nid));
1119                 GOTO(out, rc = -ENODEV);
1120         }
1121
1122         spin_lock(&target->obd_dev_lock);
1123
1124         target->obd_conn_inprogress++;
1125
1126         if (target->obd_stopping || !target->obd_set_up) {
1127                 spin_unlock(&target->obd_dev_lock);
1128
1129                 deuuidify(str, NULL, &target_start, &target_len);
1130                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1131                               target_len, target_start,
1132                               libcfs_nid2str(req->rq_peer.nid),
1133                               (target->obd_stopping ?
1134                                "stopping" : "not set up"));
1135                 GOTO(out, rc = -ENODEV);
1136         }
1137
1138         if (target->obd_no_conn) {
1139                 spin_unlock(&target->obd_dev_lock);
1140
1141                 CDEBUG(D_INFO,
1142                        "%s: Temporarily refusing client connection from %s\n",
1143                        target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1144                 GOTO(out, rc = -EAGAIN);
1145         }
1146
1147         spin_unlock(&target->obd_dev_lock);
1148
1149         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1150         if (str == NULL) {
1151                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1152                 GOTO(out, rc = -EINVAL);
1153         }
1154
1155         obd_str2uuid(&cluuid, str);
1156
1157         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1158         if (tmp == NULL)
1159                 GOTO(out, rc = -EPROTO);
1160
1161         conn = *tmp;
1162
1163         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1164                                     RCL_CLIENT);
1165         if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1166                 GOTO(out, rc = -EPROTO);
1167         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1168         if (!data)
1169                 GOTO(out, rc = -EPROTO);
1170
1171         rc = req_capsule_server_pack(&req->rq_pill);
1172         if (rc)
1173                 GOTO(out, rc);
1174
1175 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1176         /*
1177          * Don't allow clients to connect that are using old 1.8 format
1178          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1179          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1180          * FULL20 flag should be set on all connections since 2.0, but no
1181          * longer affects behaviour.
1182          *
1183          * Later this check will be disabled and the flag can be retired
1184          * completely once interop with 3.0 is no longer needed.
1185          */
1186         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1187                 GOTO(out, rc = -EPROTO);
1188
1189         /*
1190          * Don't allow liblustre clients to connect.
1191          * - testing was disabled in v2_2_50_0-61-g6a75d65
1192          * - building was disabled in v2_5_58_0-28-g7277179
1193          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1194          * - clients were refused connect for version difference > 0.0.1.32
1195          */
1196         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1197                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1198                 GOTO(out, rc = -EPROTO);
1199         }
1200 #endif
1201
1202         /*
1203          * Note: lw_client is needed in MDS-MDS failover during update log
1204          * processing, so we needs to allow lw_client to be connected at
1205          * anytime, instead of only the initial connection
1206          */
1207         lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1208
1209         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1210                 initial_conn = true;
1211                 mds_conn = OCD_HAS_FLAG(data, MDS);
1212                 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1213
1214                 /*
1215                  * OBD_CONNECT_MNE_SWAB is removed at 2.14
1216                  * Checking OBD_CONNECT_FID can be removed in the future.
1217                  *
1218                  * Via check OBD_CONNECT_FID, we can distinguish whether
1219                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1220                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1221                  */
1222                 if (!lw_client &&
1223                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1224                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1225                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1226                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1227                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1228                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1229
1230                         /*
1231                          * We do not support the MDT-MDT interoperations with
1232                          * different version MDT because of protocol changes.
1233                          */
1234                         if (unlikely(major != LUSTRE_MAJOR ||
1235                                      minor != LUSTRE_MINOR ||
1236                                      abs(patch - LUSTRE_PATCH) > 3)) {
1237                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1238                                               target->obd_name, LUSTRE_MAJOR,
1239                                               LUSTRE_MINOR, LUSTRE_PATCH,
1240                                               LUSTRE_FIX, major, minor, patch,
1241                                               OBD_OCD_VERSION_FIX(data->ocd_version),
1242                                               libcfs_nid2str(req->rq_peer.nid),
1243                                               str);
1244                                 GOTO(out, rc = -EPROTO);
1245                         }
1246                 }
1247         }
1248
1249         /* lctl gets a backstage, all-access pass. */
1250         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1251                 goto dont_check_exports;
1252
1253         export = obd_uuid_lookup(target, &cluuid);
1254         if (!export)
1255                 goto no_export;
1256
1257         /* We've found an export in the hash. */
1258
1259         spin_lock(&export->exp_lock);
1260
1261         if (export->exp_connecting) { /* b=9635, et. al. */
1262                 spin_unlock(&export->exp_lock);
1263                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1264                               export->exp_obd->obd_name, export,
1265                               libcfs_nid2str(req->rq_peer.nid));
1266                 class_export_put(export);
1267                 export = NULL;
1268                 rc = -EALREADY;
1269         } else if ((mds_conn || (lw_client && initial_conn) ||
1270                    OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
1271                 spin_unlock(&export->exp_lock);
1272                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1273                         /* MDS or LWP reconnected after failover. */
1274                         LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1275                                       target->obd_name,
1276                                       lw_client ? "LWP" : "MDS",
1277                                       libcfs_nid2str(req->rq_peer.nid),
1278                                       libcfs_nid2str(export->exp_connection->c_peer.nid));
1279                 } else {
1280                         /* New connection from the same NID. */
1281                         LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1282                                       target->obd_name,
1283                                       lw_client ? "LWP" : "MDS",
1284                                       libcfs_nid2str(req->rq_peer.nid),
1285                                       OCD_HAS_FLAG(data, MDS_MDS) ?
1286                                       "keep" : "remove");
1287                 }
1288
1289                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1290                     OCD_HAS_FLAG(data, MDS_MDS)) {
1291                         /*
1292                          * Because exports between MDTs will always be
1293                          * kept, let's do not fail such export if they
1294                          * come from the same NID, otherwise it might
1295                          * cause eviction between MDTs, which might
1296                          * cause namespace inconsistency
1297                          */
1298                         spin_lock(&export->exp_lock);
1299                         export->exp_connecting = 1;
1300                         export->exp_conn_cnt = 0;
1301                         spin_unlock(&export->exp_lock);
1302                         conn.cookie = export->exp_handle.h_cookie;
1303                         rc = EALREADY;
1304                 } else {
1305                         class_fail_export(export);
1306                         class_export_put(export);
1307                         export = NULL;
1308                         rc = 0;
1309                 }
1310         } else if (export->exp_connection != NULL && initial_conn &&
1311                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1312                 spin_unlock(&export->exp_lock);
1313                 /* In MDS failover we have static UUID but NID can change. */
1314                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1315                               target->obd_name, cluuid.uuid,
1316                               libcfs_nid2str(req->rq_peer.nid),
1317                               libcfs_nid2str(
1318                                         export->exp_connection->c_peer.nid));
1319                 rc = -EALREADY;
1320                 class_export_put(export);
1321                 export = NULL;
1322         } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1323                    !lw_client) {
1324                 spin_unlock(&export->exp_lock);
1325                 rc = -EAGAIN;
1326         } else {
1327                 export->exp_connecting = 1;
1328                 spin_unlock(&export->exp_lock);
1329                 LASSERT(export->exp_obd == target);
1330
1331                 rc = target_handle_reconnect(&conn, export, &cluuid);
1332         }
1333
1334         /* If we found an export, we already unlocked. */
1335         if (!export) {
1336 no_export:
1337                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1338         } else if (req->rq_export == NULL &&
1339                    atomic_read(&export->exp_rpc_count) > 0) {
1340                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1341                               target->obd_name, cluuid.uuid,
1342                               libcfs_nid2str(req->rq_peer.nid),
1343                               refcount_read(&export->exp_handle.h_ref));
1344                         GOTO(out, rc = -EBUSY);
1345         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1346                    rc != EALREADY) {
1347                 if (!strstr(cluuid.uuid, "mdt"))
1348                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1349                                       target->obd_name, cluuid.uuid,
1350                                       libcfs_nid2str(req->rq_peer.nid));
1351                 GOTO(out, rc = -EALREADY);
1352         } else {
1353                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1354         }
1355
1356         if (rc < 0)
1357                 GOTO(out, rc);
1358
1359         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1360                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1361                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1362                export, ktime_get_seconds(),
1363                export ? export->exp_last_request_time : 0);
1364
1365         /*
1366          * If this is the first time a client connects, reset the recovery
1367          * timer. Discard lightweight connections which might be local.
1368          */
1369         if (!lw_client && rc == 0 && target->obd_recovering)
1370                 check_and_start_recovery_timer(target, req, export == NULL);
1371
1372         /*
1373          * We want to handle EALREADY but *not* -EALREADY from
1374          * target_handle_reconnect(), return reconnection state in a flag.
1375          */
1376         if (rc == EALREADY) {
1377                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1378                 rc = 0;
1379         } else {
1380                 LASSERT(rc == 0);
1381         }
1382
1383         /* Tell the client if we support replayable requests. */
1384         if (target->obd_replayable)
1385                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1386         client_nid = &req->rq_peer.nid;
1387
1388         if (export == NULL) {
1389                 /* allow lightweight connections during recovery */
1390                 /*
1391                  * allow "new" MDT to be connected during recovery, since we
1392                  * need retrieve recovery update records from it
1393                  */
1394                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1395                         struct hrtimer *timer = &target->obd_recovery_timer;
1396                         ktime_t remaining;
1397                         s64 timeout, left;
1398                         int in_progress;
1399                         int connected;
1400                         int known;
1401                         int stale;
1402                         char *msg;
1403
1404                         connected = atomic_read(&target->obd_connected_clients);
1405                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1406                         known =
1407                            atomic_read(&target->obd_max_recoverable_clients);
1408                         stale = target->obd_stale_clients;
1409                         remaining = hrtimer_get_remaining(timer);
1410                         left = ktime_divns(remaining, NSEC_PER_SEC);
1411
1412                         if (ktime_to_ns(remaining) > 0) {
1413                                 msg = "to recover in";
1414                                 timeout = left;
1415                         } else {
1416                                 msg = "already passed deadline";
1417                                 timeout = -left;
1418
1419                                 target_check_recovery_timer(target);
1420                         }
1421
1422                         LCONSOLE_WARN("%s: Denying connection for new client %s (at %s), waiting for %d known clients (%d recovered, %d in progress, and %d evicted) %s %lld:%.02lld\n",
1423                                       target->obd_name, cluuid.uuid,
1424                                       libcfs_nid2str(req->rq_peer.nid), known,
1425                                       connected - in_progress, in_progress,
1426                                       stale, msg, timeout / 60, timeout % 60);
1427                         rc = -EBUSY;
1428                 } else {
1429 dont_check_exports:
1430                         rc = obd_connect(req->rq_svc_thread->t_env,
1431                                          &export, target, &cluuid, data,
1432                                          client_nid);
1433                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1434                                 lustre_msg_add_op_flags(req->rq_repmsg,
1435                                                         MSG_CONNECT_RECOVERING);
1436                         if (rc == 0) {
1437                                 conn.cookie = export->exp_handle.h_cookie;
1438                                 rc = rev_import_init(export);
1439                         }
1440
1441                         if (mds_mds_conn)
1442                                 new_mds_mds_conn = true;
1443                 }
1444         } else {
1445                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1446                                    export, target, &cluuid, data, client_nid);
1447         }
1448         if (rc)
1449                 GOTO(out, rc);
1450
1451         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1452         data->ocd_instance = target->u.obt.obt_instance;
1453
1454         /*
1455          * Return only the parts of obd_connect_data that we understand, so the
1456          * client knows that we don't understand the rest.
1457          */
1458         if (data) {
1459                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1460                                                RCL_SERVER);
1461                 tmpdata = req_capsule_server_get(&req->rq_pill,
1462                                                  &RMF_CONNECT_DATA);
1463                 /*
1464                  * Don't use struct assignment here, because the client reply
1465                  * buffer may be smaller/larger than the local struct
1466                  * obd_connect_data.
1467                  */
1468                 memcpy(tmpdata, data, min(tmpsize, size));
1469         }
1470
1471         /*
1472          * If the client and the server are the same node, we will already
1473          * have an export that really points to the client's DLM export,
1474          * because we have a shared handles table.
1475          *
1476          * XXX this will go away when shaver stops sending the "connect" handle
1477          * in the real "remote handle" field of the request --phik 24 Apr 2003
1478          */
1479         ptlrpc_request_change_export(req, export);
1480
1481         pcon = ptlrpc_connection_get(req->rq_peer, req->rq_self, &cluuid);
1482         if (pcon == NULL)
1483                 GOTO(out, rc = -ENOTCONN);
1484
1485         spin_lock(&export->exp_lock);
1486
1487         if (export->exp_disconnected) {
1488                 spin_unlock(&export->exp_lock);
1489                 GOTO(out, rc = -ENODEV);
1490         }
1491         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1492                 spin_unlock(&export->exp_lock);
1493                 CDEBUG(D_RPCTRACE,
1494                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1495                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1496                        export->exp_conn_cnt,
1497                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1498
1499                 GOTO(out, rc = -EALREADY);
1500         }
1501         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1502         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1503
1504         /* Check to see if connection came from another NID. */
1505         if (export->exp_connection != NULL &&
1506             export->exp_connection->c_peer.nid != req->rq_peer.nid) {
1507                 obd_nid_del(export->exp_obd, export);
1508                 ptlrpc_connection_put(export->exp_connection);
1509                 export->exp_connection = NULL;
1510         }
1511
1512         if (export->exp_connection == NULL) {
1513                 export->exp_connection = pcon;
1514                 pcon = NULL;
1515         }
1516         obd_nid_add(export->exp_obd, export);
1517
1518         spin_unlock(&export->exp_lock);
1519
1520         lustre_msg_set_handle(req->rq_repmsg, &conn);
1521
1522         rc = rev_import_reconnect(export, req);
1523         if (rc != 0)
1524                 GOTO(out, rc);
1525
1526         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1527                 int has_transno;
1528                 __u64 transno = data->ocd_transno;
1529
1530                 spin_lock(&export->exp_lock);
1531                 /*
1532                  * possible race with class_disconnect_stale_exports,
1533                  * export may be already in the eviction process
1534                  */
1535                 if (export->exp_failed) {
1536                         spin_unlock(&export->exp_lock);
1537                         GOTO(out, rc = -ENODEV);
1538                 }
1539                 export->exp_in_recovery = 1;
1540                 export->exp_req_replay_needed = 1;
1541                 export->exp_lock_replay_needed = 1;
1542                 spin_unlock(&export->exp_lock);
1543
1544                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1545                                  MSG_CONNECT_TRANSNO);
1546                 if (has_transno && transno == 0)
1547                         CWARN("Connect with zero transno!\n");
1548
1549                 if (has_transno && transno > 0 &&
1550                     transno < target->obd_next_recovery_transno &&
1551                     transno > target->obd_last_committed) {
1552                         /* Another way is to use cmpxchg() to be lock-free. */
1553                         spin_lock(&target->obd_recovery_task_lock);
1554                         if (transno < target->obd_next_recovery_transno)
1555                                 target->obd_next_recovery_transno = transno;
1556                         spin_unlock(&target->obd_recovery_task_lock);
1557                 }
1558
1559                 atomic_inc(&target->obd_req_replay_clients);
1560                 atomic_inc(&target->obd_lock_replay_clients);
1561                 /*
1562                  * Note: MDS-MDS connection is allowed to be connected during
1563                  * recovery, no matter if the exports needs to be recoveried.
1564                  * Because we need retrieve updates logs from all other MDTs.
1565                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1566                  * also needs to be increased to match other recovery checking
1567                  * condition.
1568                  */
1569                 if (new_mds_mds_conn)
1570                         atomic_inc(&target->obd_max_recoverable_clients);
1571
1572                 if (atomic_inc_return(&target->obd_connected_clients) ==
1573                     atomic_read(&target->obd_max_recoverable_clients))
1574                         wake_up(&target->obd_next_transno_waitq);
1575         }
1576
1577         /* Tell the client we're in recovery, when client is involved in it. */
1578         if (target->obd_recovering && !lw_client)
1579                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1580
1581 out:
1582         if (export) {
1583                 spin_lock(&export->exp_lock);
1584                 export->exp_connecting = 0;
1585                 spin_unlock(&export->exp_lock);
1586
1587                 class_export_put(export);
1588         }
1589         if (target != NULL) {
1590                 spin_lock(&target->obd_dev_lock);
1591                 target->obd_conn_inprogress--;
1592                 spin_unlock(&target->obd_dev_lock);
1593                 class_decref(target, "find", current);
1594         }
1595         if (pcon)
1596                 ptlrpc_connection_put(pcon);
1597         req->rq_status = rc;
1598         RETURN(rc);
1599 }
1600
1601 int target_handle_disconnect(struct ptlrpc_request *req)
1602 {
1603         int rc;
1604
1605         ENTRY;
1606
1607         rc = req_capsule_server_pack(&req->rq_pill);
1608         if (rc)
1609                 RETURN(rc);
1610
1611         /* In case of target disconnect, updating sec ctx immediately is
1612          * required in order to record latest sequence number used.
1613          * Sequence is normally updated on export destroy, but this event
1614          * can occur too late, ie after a new target connect request has
1615          * been processed.
1616          * Maintaining correct sequence when client connection becomes idle
1617          * ensures that GSS does not erroneously consider requests as replays.
1618          */
1619         rc = sptlrpc_export_update_ctx(req->rq_export);
1620         if (rc)
1621                 RETURN(rc);
1622
1623         /* Keep the rq_export around so we can send the reply. */
1624         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1625
1626         RETURN(0);
1627 }
1628
1629 void target_destroy_export(struct obd_export *exp)
1630 {
1631         struct obd_import *imp = NULL;
1632         /*
1633          * exports created from last_rcvd data, and "fake"
1634          * exports created by lctl don't have an import
1635          */
1636         spin_lock(&exp->exp_lock);
1637         if (exp->exp_imp_reverse != NULL) {
1638                 imp = exp->exp_imp_reverse;
1639                 exp->exp_imp_reverse = NULL;
1640         }
1641         spin_unlock(&exp->exp_lock);
1642         if (imp != NULL)
1643                 client_destroy_import(imp);
1644
1645         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1646         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1647         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1648         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1649 }
1650 EXPORT_SYMBOL(target_destroy_export);
1651
1652 /*
1653  * Recovery functions
1654  */
1655 static void target_request_copy_get(struct ptlrpc_request *req)
1656 {
1657         class_export_rpc_inc(req->rq_export);
1658         LASSERT(list_empty(&req->rq_list));
1659         INIT_LIST_HEAD(&req->rq_replay_list);
1660
1661         /* Increase refcount to keep request in queue. */
1662         atomic_inc(&req->rq_refcount);
1663         /* Let export know it has replays to be handled. */
1664         atomic_inc(&req->rq_export->exp_replay_count);
1665 }
1666
1667 static void target_request_copy_put(struct ptlrpc_request *req)
1668 {
1669         LASSERT(list_empty(&req->rq_replay_list));
1670         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1671
1672         atomic_dec(&req->rq_export->exp_replay_count);
1673         class_export_rpc_dec(req->rq_export);
1674         ptlrpc_server_drop_request(req);
1675 }
1676
1677 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1678 {
1679         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1680         struct obd_export *exp = req->rq_export;
1681         struct ptlrpc_request *reqiter;
1682         struct ptlrpc_request *dup_req = NULL;
1683         int dup = 0;
1684
1685         LASSERT(exp);
1686
1687         spin_lock(&exp->exp_lock);
1688         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1689                             rq_replay_list) {
1690                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1691                         dup_req = reqiter;
1692                         dup = 1;
1693                         break;
1694                 }
1695         }
1696
1697         if (dup) {
1698                 /* We expect it with RESENT and REPLAY flags. */
1699                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1700                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1701                         CERROR("invalid flags %x of resent replay\n",
1702                                lustre_msg_get_flags(req->rq_reqmsg));
1703
1704                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1705                         __u32 new_conn;
1706
1707                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1708                         if (new_conn >
1709                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1710                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1711                                                         new_conn);
1712                 }
1713         } else {
1714                 list_add_tail(&req->rq_replay_list,
1715                               &exp->exp_req_replay_queue);
1716         }
1717
1718         spin_unlock(&exp->exp_lock);
1719         return dup;
1720 }
1721
1722 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1723 {
1724         LASSERT(!list_empty(&req->rq_replay_list));
1725         LASSERT(req->rq_export);
1726
1727         spin_lock(&req->rq_export->exp_lock);
1728         list_del_init(&req->rq_replay_list);
1729         spin_unlock(&req->rq_export->exp_lock);
1730 }
1731
1732 static void target_finish_recovery(struct lu_target *lut)
1733 {
1734         struct obd_device *obd = lut->lut_obd;
1735
1736         ENTRY;
1737
1738         /* Only log a recovery message when recovery has occurred. */
1739         if (obd->obd_recovery_start) {
1740                 time64_t now = ktime_get_seconds();
1741                 time64_t elapsed_time;
1742
1743                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1744                                      1);
1745                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1746                               obd->obd_name, elapsed_time / 60,
1747                               elapsed_time % 60,
1748                               atomic_read(&obd->obd_max_recoverable_clients),
1749                               atomic_read(&obd->obd_connected_clients),
1750                               obd->obd_stale_clients,
1751                               obd->obd_stale_clients == 1 ? "was" : "were");
1752                 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1753                         libcfs_debug_dumplog();
1754         }
1755
1756         ldlm_reprocess_recovery_done(obd->obd_namespace);
1757         spin_lock(&obd->obd_recovery_task_lock);
1758         if (!list_empty(&obd->obd_req_replay_queue) ||
1759             !list_empty(&obd->obd_lock_replay_queue) ||
1760             !list_empty(&obd->obd_final_req_queue)) {
1761                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1762                        obd->obd_name,
1763                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1764                        list_empty(&obd->obd_lock_replay_queue) ? \
1765                                   "" : "lock ",
1766                        list_empty(&obd->obd_final_req_queue) ? \
1767                                   "" : "final ");
1768                 spin_unlock(&obd->obd_recovery_task_lock);
1769                 LBUG();
1770         }
1771         spin_unlock(&obd->obd_recovery_task_lock);
1772
1773         obd->obd_recovery_end = ktime_get_seconds();
1774
1775         /* When recovery finished, cleanup orphans on MDS and OST. */
1776         if (obd->obd_type && OBP(obd, postrecov)) {
1777                 int rc = OBP(obd, postrecov)(obd);
1778
1779                 if (rc < 0)
1780                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1781                                       obd->obd_name, rc);
1782         }
1783         EXIT;
1784 }
1785
1786 static void abort_req_replay_queue(struct obd_device *obd)
1787 {
1788         struct ptlrpc_request *req, *n;
1789         LIST_HEAD(abort_list);
1790
1791         spin_lock(&obd->obd_recovery_task_lock);
1792         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1793         spin_unlock(&obd->obd_recovery_task_lock);
1794         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1795                 DEBUG_REQ(D_WARNING, req, "aborted:");
1796                 req->rq_status = -ENOTCONN;
1797                 if (ptlrpc_error(req)) {
1798                         DEBUG_REQ(D_ERROR, req,
1799                                   "failed abort_req_reply; skipping");
1800                 }
1801                 target_exp_dequeue_req_replay(req);
1802                 target_request_copy_put(req);
1803         }
1804 }
1805
1806 static void abort_lock_replay_queue(struct obd_device *obd)
1807 {
1808         struct ptlrpc_request *req, *n;
1809         LIST_HEAD(abort_list);
1810
1811         spin_lock(&obd->obd_recovery_task_lock);
1812         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1813         spin_unlock(&obd->obd_recovery_task_lock);
1814         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1815                 DEBUG_REQ(D_ERROR, req, "aborted:");
1816                 req->rq_status = -ENOTCONN;
1817                 if (ptlrpc_error(req)) {
1818                         DEBUG_REQ(D_ERROR, req,
1819                                   "failed abort_lock_reply; skipping");
1820                 }
1821                 target_request_copy_put(req);
1822         }
1823 }
1824
1825 /*
1826  * Called from a cleanup function if the device is being cleaned up
1827  * forcefully.  The exports should all have been disconnected already,
1828  * the only thing left to do is
1829  * - clear the recovery flags
1830  * - cancel the timer
1831  * - free queued requests and replies, but don't send replies
1832  * Because the obd_stopping flag is set, no new requests should be received.
1833  */
1834 void target_cleanup_recovery(struct obd_device *obd)
1835 {
1836         struct ptlrpc_request *req, *n;
1837         LIST_HEAD(clean_list);
1838
1839         spin_lock(&obd->obd_dev_lock);
1840         if (!obd->obd_recovering) {
1841                 spin_unlock(&obd->obd_dev_lock);
1842                 EXIT;
1843                 return;
1844         }
1845         obd->obd_recovering = obd->obd_abort_recovery = 0;
1846         spin_unlock(&obd->obd_dev_lock);
1847
1848         spin_lock(&obd->obd_recovery_task_lock);
1849         target_cancel_recovery_timer(obd);
1850         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1851         spin_unlock(&obd->obd_recovery_task_lock);
1852
1853         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1854                 LASSERT(req->rq_reply_state == NULL);
1855                 target_exp_dequeue_req_replay(req);
1856                 target_request_copy_put(req);
1857         }
1858
1859         spin_lock(&obd->obd_recovery_task_lock);
1860         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1861         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1862         spin_unlock(&obd->obd_recovery_task_lock);
1863
1864         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1865                 LASSERT(req->rq_reply_state == NULL);
1866                 target_request_copy_put(req);
1867         }
1868
1869         EXIT;
1870 }
1871 EXPORT_SYMBOL(target_cleanup_recovery);
1872
1873 /* obd_recovery_task_lock should be held */
1874 void target_cancel_recovery_timer(struct obd_device *obd)
1875 {
1876         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1877         hrtimer_cancel(&obd->obd_recovery_timer);
1878 }
1879
1880 static void target_start_recovery_timer(struct obd_device *obd)
1881 {
1882         ktime_t delay;
1883
1884         if (obd->obd_recovery_start != 0)
1885                 return;
1886
1887         spin_lock(&obd->obd_dev_lock);
1888         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1889                 spin_unlock(&obd->obd_dev_lock);
1890                 return;
1891         }
1892
1893         LASSERT(obd->obd_recovery_timeout != 0);
1894
1895         if (obd->obd_recovery_start != 0) {
1896                 spin_unlock(&obd->obd_dev_lock);
1897                 return;
1898         }
1899
1900         obd->obd_recovery_start = ktime_get_seconds();
1901         delay = ktime_set(obd->obd_recovery_start +
1902                           obd->obd_recovery_timeout, 0);
1903         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1904         spin_unlock(&obd->obd_dev_lock);
1905
1906         LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1907                       obd->obd_name,
1908                       obd->obd_recovery_timeout / 60,
1909                       obd->obd_recovery_timeout % 60,
1910                       atomic_read(&obd->obd_max_recoverable_clients),
1911                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1912                       "" : "s",
1913                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1914                       "s" : "");
1915 }
1916
1917 /**
1918  * extend recovery window.
1919  *
1920  * if @extend is true, extend recovery window to have @dr_timeout remaining
1921  * at least; otherwise, make sure the recovery timeout value is not less
1922  * than @dr_timeout.
1923  */
1924 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1925                                   bool extend)
1926 {
1927         ktime_t left_ns;
1928         timeout_t timeout;
1929         timeout_t left;
1930
1931         spin_lock(&obd->obd_dev_lock);
1932         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1933             obd->obd_stopping) {
1934                 spin_unlock(&obd->obd_dev_lock);
1935                 return;
1936         }
1937         LASSERT(obd->obd_recovery_start != 0);
1938
1939         left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1940         left = ktime_divns(left_ns, NSEC_PER_SEC);
1941
1942         if (extend) {
1943                 timeout = obd->obd_recovery_timeout;
1944                 /* dr_timeout will happen after the hrtimer has expired.
1945                  * Add the excess time to the soft recovery timeout without
1946                  * exceeding the hard recovery timeout.
1947                  */
1948                 if (dr_timeout > left) {
1949                         timeout += dr_timeout - left;
1950                         timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1951                                         timeout);
1952                 }
1953         } else {
1954                 timeout = clamp_t(timeout_t, dr_timeout,
1955                                   obd->obd_recovery_timeout,
1956                                   obd->obd_recovery_time_hard);
1957         }
1958
1959         if (timeout == obd->obd_recovery_time_hard)
1960                 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1961                       obd->obd_name, timeout, extend);
1962
1963         if (obd->obd_recovery_timeout < timeout) {
1964                 ktime_t end, now;
1965
1966                 obd->obd_recovery_timeout = timeout;
1967                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1968                 now = ktime_set(ktime_get_seconds(), 0);
1969                 left_ns = ktime_sub(end, now);
1970                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1971                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1972         }
1973         spin_unlock(&obd->obd_dev_lock);
1974
1975         CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1976                 obd->obd_name, left);
1977 }
1978
1979 /* Reset the timer with each new client connection */
1980 /*
1981  * This timer is actually reconnect_timer, which is for making sure
1982  * the total recovery window is at least as big as my reconnect
1983  * attempt timing. So the initial recovery time_out will be set to
1984  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1985  * from client is bigger than this, then the recovery time_out will
1986  * be extended to make sure the client could be reconnected, in the
1987  * process, the timeout from the new client should be ignored.
1988  */
1989 static void
1990 check_and_start_recovery_timer(struct obd_device *obd,
1991                                struct ptlrpc_request *req,
1992                                int new_client)
1993 {
1994         timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
1995         struct obd_device_target *obt = &obd->u.obt;
1996
1997         if (!new_client && service_timeout)
1998                 /*
1999                  * Teach server about old server's estimates, as first guess
2000                  * at how long new requests will take.
2001                  */
2002                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2003                             service_timeout);
2004
2005         target_start_recovery_timer(obd);
2006
2007         /*
2008          * Convert the service time to RPC timeout,
2009          * and reuse service_timeout to limit stack usage.
2010          */
2011         service_timeout = at_est2timeout(service_timeout);
2012
2013         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2014             service_timeout < at_extra)
2015                 service_timeout = at_extra;
2016
2017         /*
2018          * We expect other clients to timeout within service_timeout, then try
2019          * to reconnect, then try the failover server.  The max delay between
2020          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2021          */
2022         service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2023
2024         LASSERT(obt->obt_magic == OBT_MAGIC);
2025         service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2026         if (service_timeout > obd->obd_recovery_timeout && !new_client)
2027                 extend_recovery_timer(obd, service_timeout, false);
2028 }
2029
2030 /** Health checking routines */
2031 static inline int exp_connect_healthy(struct obd_export *exp)
2032 {
2033         return exp->exp_in_recovery;
2034 }
2035
2036 /** if export done req_replay or has replay in queue */
2037 static inline int exp_req_replay_healthy(struct obd_export *exp)
2038 {
2039         return (!exp->exp_req_replay_needed ||
2040                 atomic_read(&exp->exp_replay_count) > 0);
2041 }
2042
2043
2044 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2045 {
2046         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2047                exp_req_replay_healthy(exp);
2048 }
2049
2050 /** if export done lock_replay or has replay in queue */
2051 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2052 {
2053         return (!exp->exp_lock_replay_needed ||
2054                 atomic_read(&exp->exp_replay_count) > 0);
2055 }
2056
2057 static inline int exp_vbr_healthy(struct obd_export *exp)
2058 {
2059         return !exp->exp_vbr_failed;
2060 }
2061
2062 static inline int exp_finished(struct obd_export *exp)
2063 {
2064         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2065 }
2066
2067 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2068 {
2069         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2070                 exp_finished(exp);
2071 }
2072
2073 static int check_for_next_transno(struct lu_target *lut)
2074 {
2075         struct ptlrpc_request *req = NULL;
2076         struct obd_device *obd = lut->lut_obd;
2077         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2078         int wake_up = 0, connected, completed, queue_len;
2079         __u64 req_transno = 0;
2080         __u64 update_transno = 0;
2081         __u64 next_transno = 0;
2082
2083         ENTRY;
2084
2085         spin_lock(&obd->obd_recovery_task_lock);
2086         if (!list_empty(&obd->obd_req_replay_queue)) {
2087                 req = list_entry(obd->obd_req_replay_queue.next,
2088                                      struct ptlrpc_request, rq_list);
2089                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2090         }
2091
2092         if (!obd->obd_abort_recov_mdt && tdtd)
2093                 update_transno = distribute_txn_get_next_transno(tdtd);
2094
2095         connected = atomic_read(&obd->obd_connected_clients);
2096         completed = connected - atomic_read(&obd->obd_req_replay_clients);
2097         queue_len = obd->obd_requests_queued_for_recovery;
2098         next_transno = obd->obd_next_recovery_transno;
2099
2100         CDEBUG(D_HA,
2101                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2102                atomic_read(&obd->obd_max_recoverable_clients),
2103                connected, completed,
2104                queue_len, req_transno, next_transno);
2105
2106         if (obd->obd_abort_recovery) {
2107                 CDEBUG(D_HA, "waking for aborted recovery\n");
2108                 wake_up = 1;
2109         } else if (obd->obd_recovery_expired) {
2110                 CDEBUG(D_HA, "waking for expired recovery\n");
2111                 wake_up = 1;
2112         } else if (!obd->obd_abort_recov_mdt && tdtd && req &&
2113                    is_req_replayed_by_update(req)) {
2114                 LASSERTF(req_transno < next_transno,
2115                          "req_transno %llu next_transno%llu\n", req_transno,
2116                          next_transno);
2117                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2118                        req_transno);
2119                 wake_up = 1;
2120         } else if (req_transno == next_transno ||
2121                    (update_transno != 0 && update_transno <= next_transno)) {
2122                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2123                 wake_up = 1;
2124         } else if (queue_len > 0 &&
2125                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2126                 /** handle gaps occured due to lost reply or VBR */
2127                 LASSERTF(req_transno >= next_transno,
2128                          "req_transno: %llu, next_transno: %llu\n",
2129                          req_transno, next_transno);
2130                 CDEBUG(D_HA,
2131                        "%s: waking for gap in transno, VBR is %s (skip: %lld, ql: %d, comp: %d, conn: %d, next: %lld, next_update %lld last_committed: %lld)\n",
2132                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2133                        next_transno, queue_len, completed, connected,
2134                        req_transno, update_transno, obd->obd_last_committed);
2135                 obd->obd_next_recovery_transno = req_transno;
2136                 wake_up = 1;
2137         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2138                 CDEBUG(D_HA, "waking for completed recovery\n");
2139                 wake_up = 1;
2140         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2141                 CDEBUG(D_HA,
2142                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2143                        next_transno);
2144                 obd->obd_next_recovery_transno = req_transno;
2145                 wake_up = 1;
2146         }
2147         spin_unlock(&obd->obd_recovery_task_lock);
2148         return wake_up;
2149 }
2150
2151 static int check_for_next_lock(struct lu_target *lut)
2152 {
2153         struct obd_device *obd = lut->lut_obd;
2154         int wake_up = 0;
2155
2156         spin_lock(&obd->obd_recovery_task_lock);
2157         if (!list_empty(&obd->obd_lock_replay_queue)) {
2158                 CDEBUG(D_HA, "waking for next lock\n");
2159                 wake_up = 1;
2160         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2161                 CDEBUG(D_HA, "waking for completed lock replay\n");
2162                 wake_up = 1;
2163         } else if (obd->obd_abort_recovery) {
2164                 CDEBUG(D_HA, "waking for aborted recovery\n");
2165                 wake_up = 1;
2166         } else if (obd->obd_recovery_expired) {
2167                 CDEBUG(D_HA, "waking for expired recovery\n");
2168                 wake_up = 1;
2169         }
2170         spin_unlock(&obd->obd_recovery_task_lock);
2171
2172         return wake_up;
2173 }
2174
2175 static int check_update_llog(struct lu_target *lut)
2176 {
2177         struct obd_device *obd = lut->lut_obd;
2178         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2179
2180         if (obd->obd_abort_recovery) {
2181                 CDEBUG(D_HA, "waking for aborted recovery\n");
2182                 return 1;
2183         }
2184
2185         if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2186                 CDEBUG(D_HA, "waking for completion of reading update log\n");
2187                 return 1;
2188         }
2189
2190         return 0;
2191 }
2192
2193 /**
2194  * wait for recovery events,
2195  * check its status with help of check_routine
2196  * evict dead clients via health_check
2197  */
2198 static int target_recovery_overseer(struct lu_target *lut,
2199                                     int (*check_routine)(struct lu_target *),
2200                                     int (*health_check)(struct obd_export *))
2201 {
2202         struct obd_device *obd = lut->lut_obd;
2203         struct target_distribute_txn_data *tdtd;
2204         time64_t last = 0;
2205         time64_t now;
2206 repeat:
2207         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2208                 now = ktime_get_seconds();
2209                 if (now - last > 600) {
2210                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2211                                       obd->obd_name);
2212                         last = now;
2213                 }
2214         }
2215         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2216               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2217                 __u64 next_update_transno = 0;
2218
2219                 /*
2220                  * Only abort the recovery if there are no update recovery
2221                  * left in the queue
2222                  */
2223                 spin_lock(&obd->obd_recovery_task_lock);
2224                 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd) {
2225                         next_update_transno =
2226                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2227
2228                         tdtd = lut->lut_tdtd;
2229                         /*
2230                          * If next_update_transno == 0, it probably because
2231                          * updatelog retrieve threads did not get any records
2232                          * yet, let's wait those threads stopped
2233                          */
2234                         if (next_update_transno == 0) {
2235                                 spin_unlock(&obd->obd_recovery_task_lock);
2236
2237                                 while (wait_event_timeout(
2238                                         tdtd->tdtd_recovery_threads_waitq,
2239                                         check_update_llog(lut),
2240                                         cfs_time_seconds(60)) == 0);
2241
2242                                 spin_lock(&obd->obd_recovery_task_lock);
2243                                 next_update_transno =
2244                                         distribute_txn_get_next_transno(tdtd);
2245                         }
2246                 }
2247
2248                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2249                         obd->obd_next_recovery_transno = next_update_transno;
2250                         spin_unlock(&obd->obd_recovery_task_lock);
2251                         /*
2252                          * Disconnect unfinished exports from clients, and
2253                          * keep connection from MDT to make sure the update
2254                          * recovery will still keep trying until some one
2255                          * manually abort the recovery
2256                          */
2257                         class_disconnect_stale_exports(obd,
2258                                                 exp_finished_or_from_mdt);
2259                         /* Abort all of replay & replay lock req from clients */
2260                         abort_req_replay_queue(obd);
2261                         abort_lock_replay_queue(obd);
2262                         CDEBUG(D_HA,
2263                                "%s: there are still update replay (%#llx)in the queue.\n",
2264                                obd->obd_name, next_update_transno);
2265                 } else {
2266                         obd->obd_abort_recovery = 1;
2267                         spin_unlock(&obd->obd_recovery_task_lock);
2268                         CWARN("%s recovery is aborted by hard timeout\n",
2269                               obd->obd_name);
2270                 }
2271         }
2272
2273         while (wait_event_timeout(obd->obd_next_transno_waitq,
2274                                   check_routine(lut),
2275                                   cfs_time_seconds(60)) == 0)
2276                 ; /* wait indefinitely for event, but don't trigger watchdog */
2277
2278         if (obd->obd_abort_recovery) {
2279                 CWARN("recovery is aborted, evict exports in recovery\n");
2280                 if (lut->lut_tdtd != NULL) {
2281                         tdtd = lut->lut_tdtd;
2282                         /*
2283                          * Let's wait all of the update log recovery thread
2284                          * finished
2285                          */
2286                         wait_event_idle(
2287                                 tdtd->tdtd_recovery_threads_waitq,
2288                                 atomic_read(&tdtd->tdtd_recovery_threads_count)
2289                                 == 0);
2290                         /* Then abort the update recovery list */
2291                         dtrq_list_destroy(lut->lut_tdtd);
2292                 }
2293
2294                 /** evict exports which didn't finish recovery yet */
2295                 class_disconnect_stale_exports(obd, exp_finished);
2296                 return 1;
2297         } else if (obd->obd_recovery_expired) {
2298                 obd->obd_recovery_expired = 0;
2299
2300                 /** If some clients died being recovered, evict them */
2301                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2302                               obd->obd_name);
2303                 /** evict cexports with no replay in queue, they are stalled */
2304                 class_disconnect_stale_exports(obd, health_check);
2305
2306                 /** continue with VBR */
2307                 spin_lock(&obd->obd_dev_lock);
2308                 obd->obd_version_recov = 1;
2309                 spin_unlock(&obd->obd_dev_lock);
2310                 /**
2311                  * reset timer, recovery will proceed with versions now,
2312                  * timeout is set just to handle reconnection delays
2313                  */
2314                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2315                 /**
2316                  * Wait for recovery events again, after evicting bad clients
2317                  */
2318                 goto repeat;
2319         }
2320         return 0;
2321 }
2322
2323 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2324 {
2325         struct obd_device *obd = lut->lut_obd;
2326         struct ptlrpc_request *req = NULL;
2327
2328         CDEBUG(D_HA, "Waiting for lock\n");
2329         if (target_recovery_overseer(lut, check_for_next_lock,
2330                                      exp_lock_replay_healthy))
2331                 abort_lock_replay_queue(obd);
2332
2333         spin_lock(&obd->obd_recovery_task_lock);
2334         if (!list_empty(&obd->obd_lock_replay_queue)) {
2335                 req = list_entry(obd->obd_lock_replay_queue.next,
2336                                      struct ptlrpc_request, rq_list);
2337                 list_del_init(&req->rq_list);
2338                 spin_unlock(&obd->obd_recovery_task_lock);
2339         } else {
2340                 spin_unlock(&obd->obd_recovery_task_lock);
2341                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2342                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2343                 /** evict exports failed VBR */
2344                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2345         }
2346         return req;
2347 }
2348
2349 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2350 {
2351         struct ptlrpc_request *req = NULL;
2352
2353         spin_lock(&obd->obd_recovery_task_lock);
2354         if (!list_empty(&obd->obd_final_req_queue)) {
2355                 req = list_entry(obd->obd_final_req_queue.next,
2356                                      struct ptlrpc_request, rq_list);
2357                 list_del_init(&req->rq_list);
2358                 spin_unlock(&obd->obd_recovery_task_lock);
2359                 if (req->rq_export->exp_in_recovery) {
2360                         spin_lock(&req->rq_export->exp_lock);
2361                         req->rq_export->exp_in_recovery = 0;
2362                         spin_unlock(&req->rq_export->exp_lock);
2363                 }
2364         } else {
2365                 spin_unlock(&obd->obd_recovery_task_lock);
2366         }
2367         return req;
2368 }
2369
2370 static void handle_recovery_req(struct ptlrpc_thread *thread,
2371                                 struct ptlrpc_request *req,
2372                                 svc_handler_t handler)
2373 {
2374         ENTRY;
2375
2376         /**
2377          * export can be evicted during recovery, no need to handle replays for
2378          * it after that, discard such request silently
2379          */
2380         if (req->rq_export->exp_disconnected)
2381                 RETURN_EXIT;
2382
2383         req->rq_session.lc_thread = thread;
2384         req->rq_svc_thread = thread;
2385         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2386
2387         /* thread context */
2388         lu_context_enter(&thread->t_env->le_ctx);
2389         (void)handler(req);
2390         lu_context_exit(&thread->t_env->le_ctx);
2391
2392         req->rq_svc_thread->t_env->le_ses = NULL;
2393
2394         /* don't reset timer for final stage */
2395         if (!exp_finished(req->rq_export)) {
2396                 timeout_t timeout = obd_timeout;
2397
2398                 /**
2399                  * Add request @timeout to the recovery time so next request from
2400                  * this client may come in recovery time
2401                  */
2402                 if (!AT_OFF) {
2403                         struct ptlrpc_service_part *svcpt;
2404                         timeout_t est_timeout;
2405
2406                         svcpt = req->rq_rqbd->rqbd_svcpt;
2407                         /*
2408                          * If the server sent early reply for this request,
2409                          * the client will recalculate the timeout according to
2410                          * current server estimate service time, so we will
2411                          * use the maxium timeout here for waiting the client
2412                          * sending the next req
2413                          */
2414                         est_timeout = at_get(&svcpt->scp_at_estimate);
2415                         timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2416                                         lustre_msg_get_timeout(req->rq_reqmsg));
2417                         /*
2418                          * Add 2 net_latency, one for balance rq_deadline
2419                          * (see ptl_send_rpc), one for resend the req to server,
2420                          * Note: client will pack net_latency in replay req
2421                          * (see ptlrpc_replay_req)
2422                          */
2423                         timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2424                 }
2425                 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2426                                       true);
2427         }
2428         EXIT;
2429 }
2430
2431 /** Checking routines for recovery */
2432 static int check_for_recovery_ready(struct lu_target *lut)
2433 {
2434         struct obd_device *obd = lut->lut_obd;
2435         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2436
2437         CDEBUG(D_HA,
2438                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2439                clnts, obd->obd_stale_clients,
2440                atomic_read(&obd->obd_max_recoverable_clients),
2441                obd->obd_abort_recovery, obd->obd_recovery_expired);
2442
2443         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2444                 LASSERT(clnts <=
2445                         atomic_read(&obd->obd_max_recoverable_clients));
2446                 if (clnts + obd->obd_stale_clients <
2447                     atomic_read(&obd->obd_max_recoverable_clients))
2448                         return 0;
2449         }
2450
2451         if (!obd->obd_abort_recov_mdt && lut->lut_tdtd != NULL) {
2452                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2453                     !obd->obd_abort_recovery && !obd->obd_stopping) {
2454                         /*
2455                          * Let's extend recovery timer, in case the recovery
2456                          * timer expired, and some clients got evicted
2457                          */
2458                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2459                                               true);
2460                         CDEBUG(D_HA,
2461                                "%s update recovery is not ready, extend recovery %d\n",
2462                                obd->obd_name, obd->obd_recovery_timeout);
2463                         return 0;
2464                 }
2465         }
2466
2467         return 1;
2468 }
2469
2470 enum {
2471         REQUEST_RECOVERY = 1,
2472         UPDATE_RECOVERY = 2,
2473 };
2474
2475 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2476 {
2477         __u64 transno = 0;
2478
2479         if (!list_empty(&obd->obd_req_replay_queue)) {
2480                 struct ptlrpc_request *req;
2481
2482                 req = list_entry(obd->obd_req_replay_queue.next,
2483                                  struct ptlrpc_request, rq_list);
2484                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2485         }
2486
2487         return transno;
2488 }
2489
2490 static __u64 get_next_transno(struct lu_target *lut, int *type)
2491 {
2492         struct obd_device *obd = lut->lut_obd;
2493         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2494         __u64 transno = 0;
2495         __u64 update_transno;
2496
2497         ENTRY;
2498
2499         transno = get_next_replay_req_transno(obd);
2500         if (type != NULL)
2501                 *type = REQUEST_RECOVERY;
2502
2503         if (!tdtd || obd->obd_abort_recov_mdt)
2504                 RETURN(transno);
2505
2506         update_transno = distribute_txn_get_next_transno(tdtd);
2507         if (transno == 0 || (transno >= update_transno &&
2508                              update_transno != 0)) {
2509                 transno = update_transno;
2510                 if (type != NULL)
2511                         *type = UPDATE_RECOVERY;
2512         }
2513
2514         RETURN(transno);
2515 }
2516
2517 /**
2518  * drop duplicate replay request
2519  *
2520  * Because the operation has been replayed by update recovery, the request
2521  * with the same transno will be dropped and also notify the client to send
2522  * next replay request.
2523  *
2524  * \param[in] env       execution environment
2525  * \param[in] obd       failover obd device
2526  * \param[in] req       request to be dropped
2527  */
2528 static void drop_duplicate_replay_req(struct lu_env *env,
2529                                       struct obd_device *obd,
2530                                       struct ptlrpc_request *req)
2531 {
2532         DEBUG_REQ(D_HA, req,
2533                   "remove t%lld from %s because duplicate update records found",
2534                   lustre_msg_get_transno(req->rq_reqmsg),
2535                   libcfs_nid2str(req->rq_peer.nid));
2536
2537         /*
2538          * Right now, only for MDS reint operation update replay and
2539          * normal request replay can have the same transno
2540          */
2541         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2542                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2543                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2544                 if (likely(req->rq_export))
2545                         target_committed_to_req(req);
2546                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2547                 target_send_reply(req, req->rq_status, 0);
2548         } else {
2549                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2550                 libcfs_nid2str(req->rq_peer.nid));
2551         }
2552         target_exp_dequeue_req_replay(req);
2553         target_request_copy_put(req);
2554         obd->obd_replayed_requests++;
2555 }
2556
2557 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2558
2559 static void replay_request_or_update(struct lu_env *env,
2560                                      struct lu_target *lut,
2561                                      struct target_recovery_data *trd,
2562                                      struct ptlrpc_thread *thread)
2563 {
2564         struct obd_device *obd = lut->lut_obd;
2565         struct ptlrpc_request *req = NULL;
2566         int type;
2567         __u64 transno;
2568
2569         ENTRY;
2570
2571         CDEBUG(D_HA, "Waiting for transno %lld\n",
2572                obd->obd_next_recovery_transno);
2573
2574         /* Replay all of request and update by transno */
2575         do {
2576                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2577
2578                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2579
2580                 /**
2581                  * It is needed to extend recovery window above
2582                  *  recovery_time_soft. Extending is possible only in the
2583                  *  end of recovery window (see more details in
2584                  *  handle_recovery_req()).
2585                  */
2586                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2587
2588                 if (target_recovery_overseer(lut, check_for_next_transno,
2589                                         exp_req_replay_healthy_or_from_mdt)) {
2590                         abort_req_replay_queue(obd);
2591                         abort_lock_replay_queue(obd);
2592                         goto abort;
2593                 }
2594
2595                 spin_lock(&obd->obd_recovery_task_lock);
2596                 transno = get_next_transno(lut, &type);
2597                 if (type == REQUEST_RECOVERY && transno != 0) {
2598                         /*
2599                          * Drop replay request from client side, if the
2600                          * replay has been executed by update with the
2601                          * same transno
2602                          */
2603                         req = list_entry(obd->obd_req_replay_queue.next,
2604                                         struct ptlrpc_request, rq_list);
2605
2606                         list_del_init(&req->rq_list);
2607                         obd->obd_requests_queued_for_recovery--;
2608                         spin_unlock(&obd->obd_recovery_task_lock);
2609
2610                         /*
2611                          * Let's check if the request has been redone by
2612                          * update replay
2613                          */
2614                         if (is_req_replayed_by_update(req)) {
2615                                 struct distribute_txn_replay_req *dtrq;
2616
2617                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2618                                                                       transno);
2619                                 LASSERT(dtrq != NULL);
2620                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2621                                 list_del_init(&dtrq->dtrq_list);
2622                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2623                                 dtrq_destroy(dtrq);
2624
2625                                 drop_duplicate_replay_req(env, obd, req);
2626
2627                                 continue;
2628                         }
2629
2630                         LASSERT(trd->trd_processing_task == current->pid);
2631                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2632                                   req->rq_xid,
2633                                   lustre_msg_get_transno(req->rq_reqmsg),
2634                                   libcfs_nid2str(req->rq_peer.nid));
2635
2636                         ptlrpc_watchdog_init(&thread->t_watchdog,
2637                                              WATCHDOG_TIMEOUT);
2638                         handle_recovery_req(thread, req,
2639                                             trd->trd_recovery_handler);
2640                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2641
2642                         /**
2643                          * bz18031: increase next_recovery_transno before
2644                          * target_request_copy_put() will drop exp_rpc reference
2645                          */
2646                         spin_lock(&obd->obd_recovery_task_lock);
2647                         obd->obd_next_recovery_transno++;
2648                         spin_unlock(&obd->obd_recovery_task_lock);
2649                         target_exp_dequeue_req_replay(req);
2650                         target_request_copy_put(req);
2651                         obd->obd_replayed_requests++;
2652                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2653                         struct distribute_txn_replay_req *dtrq;
2654                         int rc;
2655
2656                         spin_unlock(&obd->obd_recovery_task_lock);
2657
2658                         LASSERT(tdtd != NULL);
2659                         dtrq = distribute_txn_get_next_req(tdtd);
2660                         lu_context_enter(&thread->t_env->le_ctx);
2661                         ptlrpc_watchdog_init(&thread->t_watchdog,
2662                                              WATCHDOG_TIMEOUT);
2663                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2664                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2665                         lu_context_exit(&thread->t_env->le_ctx);
2666                         extend_recovery_timer(obd, obd_timeout, true);
2667
2668                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2669                                 CDEBUG(D_HA,
2670                                        "Move x%llu t%llu to finish list\n",
2671                                        dtrq->dtrq_xid,
2672                                        dtrq->dtrq_master_transno);
2673
2674                                 /* Add it to the replay finish list */
2675                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2676                                 list_add(&dtrq->dtrq_list,
2677                                          &tdtd->tdtd_replay_finish_list);
2678                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2679
2680                                 spin_lock(&obd->obd_recovery_task_lock);
2681                                 if (transno == obd->obd_next_recovery_transno)
2682                                         obd->obd_next_recovery_transno++;
2683                                 else if (transno >
2684                                          obd->obd_next_recovery_transno)
2685                                         obd->obd_next_recovery_transno =
2686                                                                 transno + 1;
2687                                 spin_unlock(&obd->obd_recovery_task_lock);
2688                         } else {
2689                                 dtrq_destroy(dtrq);
2690                         }
2691                 } else {
2692                         spin_unlock(&obd->obd_recovery_task_lock);
2693 abort:
2694                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2695                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2696                         /** evict exports failed VBR */
2697                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2698                         break;
2699                 }
2700         } while (1);
2701 }
2702
2703 static int target_recovery_thread(void *arg)
2704 {
2705         struct lu_target *lut = arg;
2706         struct obd_device *obd = lut->lut_obd;
2707         struct ptlrpc_request *req;
2708         struct target_recovery_data *trd = &obd->obd_recovery_data;
2709         unsigned long delta;
2710         struct lu_env *env;
2711         struct ptlrpc_thread *thread = NULL;
2712         int rc = 0;
2713
2714         ENTRY;
2715         OBD_ALLOC_PTR(thread);
2716         if (thread == NULL)
2717                 RETURN(-ENOMEM);
2718
2719         OBD_ALLOC_PTR(env);
2720         if (env == NULL)
2721                 GOTO(out_thread, rc = -ENOMEM);
2722         rc = lu_env_add(env);
2723         if (rc)
2724                 GOTO(out_env, rc);
2725
2726         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2727         if (rc)
2728                 GOTO(out_env_remove, rc);
2729
2730         thread->t_env = env;
2731         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2732         thread->t_task = current;
2733         env->le_ctx.lc_thread = thread;
2734         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2735
2736         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2737                current->pid);
2738         trd->trd_processing_task = current->pid;
2739
2740         spin_lock(&obd->obd_dev_lock);
2741         obd->obd_recovering = 1;
2742         spin_unlock(&obd->obd_dev_lock);
2743         complete(&trd->trd_starting);
2744
2745         /* first of all, we have to know the first transno to replay */
2746         if (target_recovery_overseer(lut, check_for_recovery_ready,
2747                                      exp_connect_healthy)) {
2748                 abort_req_replay_queue(obd);
2749                 abort_lock_replay_queue(obd);
2750                 if (lut->lut_tdtd != NULL)
2751                         dtrq_list_destroy(lut->lut_tdtd);
2752         }
2753
2754         /* next stage: replay requests or update */
2755         delta = jiffies;
2756         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2757                atomic_read(&obd->obd_req_replay_clients),
2758                obd->obd_next_recovery_transno);
2759         replay_request_or_update(env, lut, trd, thread);
2760
2761         /**
2762          * The second stage: replay locks
2763          */
2764         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2765                atomic_read(&obd->obd_lock_replay_clients));
2766         while ((req = target_next_replay_lock(lut))) {
2767                 LASSERT(trd->trd_processing_task == current->pid);
2768                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2769                           libcfs_nid2str(req->rq_peer.nid));
2770                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2771                         req->rq_status = -ENODEV;
2772                         target_request_copy_put(req);
2773                         continue;
2774                 }
2775                 handle_recovery_req(thread, req,
2776                                     trd->trd_recovery_handler);
2777                 target_request_copy_put(req);
2778                 obd->obd_replayed_locks++;
2779         }
2780
2781         /**
2782          * The third stage: reply on final pings, at this moment all clients
2783          * must have request in final queue
2784          */
2785         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2786         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2787         /** Update server last boot epoch */
2788         tgt_boot_epoch_update(lut);
2789         /*
2790          * We drop recoverying flag to forward all new requests
2791          * to regular mds_handle() since now
2792          */
2793         spin_lock(&obd->obd_dev_lock);
2794         obd->obd_recovering = obd->obd_abort_recovery = 0;
2795         spin_unlock(&obd->obd_dev_lock);
2796         spin_lock(&obd->obd_recovery_task_lock);
2797         target_cancel_recovery_timer(obd);
2798         spin_unlock(&obd->obd_recovery_task_lock);
2799         while ((req = target_next_final_ping(obd))) {
2800                 LASSERT(trd->trd_processing_task == current->pid);
2801                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2802                           libcfs_nid2str(req->rq_peer.nid));
2803                 handle_recovery_req(thread, req,
2804                                     trd->trd_recovery_handler);
2805                 /*
2806                  * Because the waiting client can not send ping to server,
2807                  * so we need refresh the last_request_time, to avoid the
2808                  * export is being evicted
2809                  */
2810                 ptlrpc_update_export_timer(req->rq_export, 0);
2811                 target_request_copy_put(req);
2812         }
2813
2814         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2815         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2816                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2817         if (delta > OBD_RECOVERY_TIME_SOFT) {
2818                 CWARN("too long recovery - read logs\n");
2819                 libcfs_debug_dumplog();
2820         }
2821
2822         target_finish_recovery(lut);
2823         lu_context_fini(&env->le_ctx);
2824         trd->trd_processing_task = 0;
2825         complete_all(&trd->trd_finishing);
2826         tgt_io_thread_done(thread);
2827 out_env_remove:
2828         lu_env_remove(env);
2829 out_env:
2830         OBD_FREE_PTR(env);
2831 out_thread:
2832         OBD_FREE_PTR(thread);
2833         RETURN(rc);
2834 }
2835
2836 static int target_start_recovery_thread(struct lu_target *lut,
2837                                         svc_handler_t handler)
2838 {
2839         struct obd_device *obd = lut->lut_obd;
2840         int rc = 0;
2841         struct target_recovery_data *trd = &obd->obd_recovery_data;
2842         int index;
2843
2844         memset(trd, 0, sizeof(*trd));
2845         init_completion(&trd->trd_starting);
2846         init_completion(&trd->trd_finishing);
2847         trd->trd_recovery_handler = handler;
2848
2849         rc = server_name2index(obd->obd_name, &index, NULL);
2850         if (rc < 0)
2851                 return rc;
2852
2853         if (!IS_ERR(kthread_run(target_recovery_thread,
2854                                 lut, "tgt_recover_%d", index))) {
2855                 wait_for_completion(&trd->trd_starting);
2856                 LASSERT(obd->obd_recovering != 0);
2857         } else {
2858                 rc = -ECHILD;
2859         }
2860
2861         return rc;
2862 }
2863
2864 void target_stop_recovery_thread(struct obd_device *obd)
2865 {
2866         if (obd->obd_recovery_data.trd_processing_task > 0) {
2867                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2868                 /** recovery can be done but postrecovery is not yet */
2869                 spin_lock(&obd->obd_dev_lock);
2870                 if (obd->obd_recovering) {
2871                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2872                         obd->obd_abort_recovery = 1;
2873                         wake_up(&obd->obd_next_transno_waitq);
2874                 }
2875                 spin_unlock(&obd->obd_dev_lock);
2876                 wait_for_completion(&trd->trd_finishing);
2877         }
2878 }
2879 EXPORT_SYMBOL(target_stop_recovery_thread);
2880
2881 void target_recovery_fini(struct obd_device *obd)
2882 {
2883         class_disconnect_exports(obd);
2884         target_stop_recovery_thread(obd);
2885         target_cleanup_recovery(obd);
2886 }
2887 EXPORT_SYMBOL(target_recovery_fini);
2888
2889 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2890 {
2891         struct obd_device *obd = container_of(timer, struct obd_device,
2892                                               obd_recovery_timer);
2893
2894         CDEBUG(D_HA,
2895                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2896                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2897                ktime_get_seconds() - obd->obd_recovery_start,
2898                atomic_read(&obd->obd_connected_clients));
2899
2900         obd->obd_recovery_expired = 1;
2901         wake_up(&obd->obd_next_transno_waitq);
2902         return HRTIMER_NORESTART;
2903 }
2904
2905 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2906 {
2907         struct obd_device *obd = lut->lut_obd;
2908
2909         if (lut->lut_bottom->dd_rdonly)
2910                 return;
2911
2912         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2913                 /** Update server last boot epoch */
2914                 tgt_boot_epoch_update(lut);
2915                 return;
2916         }
2917
2918         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2919                "last_transno %llu\n", obd->obd_name,
2920                atomic_read(&obd->obd_max_recoverable_clients),
2921                obd->obd_last_committed);
2922         LASSERT(obd->obd_stopping == 0);
2923         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2924         obd->obd_recovery_start = 0;
2925         obd->obd_recovery_end = 0;
2926
2927         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2928                      HRTIMER_MODE_ABS);
2929         obd->obd_recovery_timer.function = &target_recovery_expired;
2930         target_start_recovery_thread(lut, handler);
2931 }
2932 EXPORT_SYMBOL(target_recovery_init);
2933
2934 static int target_process_req_flags(struct obd_device *obd,
2935                                     struct ptlrpc_request *req)
2936 {
2937         struct obd_export *exp = req->rq_export;
2938
2939         LASSERT(exp != NULL);
2940         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2941                 /* client declares he's ready to replay locks */
2942                 spin_lock(&exp->exp_lock);
2943                 if (exp->exp_req_replay_needed) {
2944                         exp->exp_req_replay_needed = 0;
2945                         spin_unlock(&exp->exp_lock);
2946
2947                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2948                         atomic_dec(&obd->obd_req_replay_clients);
2949                 } else {
2950                         spin_unlock(&exp->exp_lock);
2951                 }
2952         }
2953         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2954                 /*
2955                  * client declares he's ready to complete recovery
2956                  * so, we put the request on th final queue
2957                  */
2958                 spin_lock(&exp->exp_lock);
2959                 if (exp->exp_lock_replay_needed) {
2960                         exp->exp_lock_replay_needed = 0;
2961                         spin_unlock(&exp->exp_lock);
2962
2963                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2964                         atomic_dec(&obd->obd_lock_replay_clients);
2965                 } else {
2966                         spin_unlock(&exp->exp_lock);
2967                 }
2968         }
2969         return 0;
2970 }
2971
2972 int target_queue_recovery_request(struct ptlrpc_request *req,
2973                                   struct obd_device *obd)
2974 {
2975         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2976         struct ptlrpc_request *reqiter;
2977         int inserted = 0;
2978
2979         ENTRY;
2980
2981         if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2982                 /* Processing the queue right now, don't re-add. */
2983                 RETURN(1);
2984         }
2985
2986         target_process_req_flags(obd, req);
2987
2988         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2989                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2990                         if (cfs_fail_val == 1) {
2991                                 cfs_race_state = 1;
2992                                 cfs_fail_val = 0;
2993                                 wake_up(&cfs_race_waitq);
2994
2995                                 schedule_timeout_interruptible(
2996                                         cfs_time_seconds(1));
2997                         }
2998                 }
2999
3000                 /*
3001                  * client declares he's ready to complete recovery
3002                  * so, we put the request on th final queue
3003                  */
3004                 target_request_copy_get(req);
3005                 DEBUG_REQ(D_HA, req, "queue final req");
3006                 wake_up(&obd->obd_next_transno_waitq);
3007                 spin_lock(&obd->obd_recovery_task_lock);
3008                 if (obd->obd_recovering) {
3009                         struct ptlrpc_request *tmp;
3010                         struct ptlrpc_request *duplicate = NULL;
3011
3012                         if (likely(!req->rq_export->exp_replay_done)) {
3013                                 req->rq_export->exp_replay_done = 1;
3014                                 list_add_tail(&req->rq_list,
3015                                               &obd->obd_final_req_queue);
3016                                 spin_unlock(&obd->obd_recovery_task_lock);
3017                                 RETURN(0);
3018                         }
3019
3020                         /*
3021                          * XXX O(n), but only happens if final ping is
3022                          * timed out, probably reorganize the list as
3023                          * a hash list later
3024                          */
3025                         list_for_each_entry_safe(reqiter, tmp,
3026                                                  &obd->obd_final_req_queue,
3027                                                  rq_list) {
3028                                 if (reqiter->rq_export == req->rq_export) {
3029                                         list_del_init(&reqiter->rq_list);
3030                                         duplicate = reqiter;
3031                                         break;
3032                                 }
3033                         }
3034
3035                         list_add_tail(&req->rq_list,
3036                                       &obd->obd_final_req_queue);
3037                         req->rq_export->exp_replay_done = 1;
3038                         spin_unlock(&obd->obd_recovery_task_lock);
3039
3040                         if (duplicate != NULL) {
3041                                 DEBUG_REQ(D_HA, duplicate,
3042                                           "put prev final req");
3043                                 target_request_copy_put(duplicate);
3044                         }
3045                         RETURN(0);
3046                 } else {
3047                         spin_unlock(&obd->obd_recovery_task_lock);
3048                         target_request_copy_put(req);
3049                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3050                 }
3051         }
3052         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3053                 /* client declares he's ready to replay locks */
3054                 target_request_copy_get(req);
3055                 DEBUG_REQ(D_HA, req, "queue lock replay req");
3056                 wake_up(&obd->obd_next_transno_waitq);
3057                 spin_lock(&obd->obd_recovery_task_lock);
3058                 LASSERT(obd->obd_recovering);
3059                 /* usually due to recovery abort */
3060                 if (!req->rq_export->exp_in_recovery) {
3061                         spin_unlock(&obd->obd_recovery_task_lock);
3062                         target_request_copy_put(req);
3063                         RETURN(-ENOTCONN);
3064                 }
3065                 LASSERT(req->rq_export->exp_lock_replay_needed);
3066                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3067                 spin_unlock(&obd->obd_recovery_task_lock);
3068                 RETURN(0);
3069         }
3070
3071         /*
3072          * CAVEAT EMPTOR: The incoming request message has been swabbed
3073          * (i.e. buflens etc are in my own byte order), but type-dependent
3074          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3075          */
3076
3077         if (!transno) {
3078                 INIT_LIST_HEAD(&req->rq_list);
3079                 DEBUG_REQ(D_HA, req, "not queueing");
3080                 RETURN(1);
3081         }
3082
3083         /*
3084          * If we're processing the queue, we want don't want to queue this
3085          * message.
3086          *
3087          * Also, if this request has a transno less than the one we're waiting
3088          * for, we should process it now.  It could (and currently always will)
3089          * be an open request for a descriptor that was opened some time ago.
3090          *
3091          * Also, a resent, replayed request that has already been
3092          * handled will pass through here and be processed immediately.
3093          */
3094         CDEBUG(D_HA,
3095                "Next recovery transno: %llu, current: %llu, replaying\n",
3096                obd->obd_next_recovery_transno, transno);
3097
3098         /*
3099          * If the request has been replayed by update replay, then sends this
3100          * request to the recovery thread (replay_request_or_update()), where
3101          * it will be handled
3102          */
3103         spin_lock(&obd->obd_recovery_task_lock);
3104         if (transno < obd->obd_next_recovery_transno &&
3105             !is_req_replayed_by_update(req)) {
3106                 /* Processing the queue right now, don't re-add. */
3107                 LASSERT(list_empty(&req->rq_list));
3108                 spin_unlock(&obd->obd_recovery_task_lock);
3109                 RETURN(1);
3110         }
3111         spin_unlock(&obd->obd_recovery_task_lock);
3112
3113         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3114                 RETURN(0);
3115
3116         target_request_copy_get(req);
3117         if (!req->rq_export->exp_in_recovery) {
3118                 target_request_copy_put(req);
3119                 RETURN(-ENOTCONN);
3120         }
3121         LASSERT(req->rq_export->exp_req_replay_needed);
3122
3123         if (target_exp_enqueue_req_replay(req)) {
3124                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3125                 target_request_copy_put(req);
3126                 RETURN(0);
3127         }
3128
3129         /* XXX O(n^2) */
3130         spin_lock(&obd->obd_recovery_task_lock);
3131         LASSERT(obd->obd_recovering);
3132         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3133                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3134                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3135                         inserted = 1;
3136                         goto added;
3137                 }
3138
3139                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3140                              transno)) {
3141                         DEBUG_REQ(D_ERROR, req,
3142                                   "dropping replay: transno has been claimed by another client");
3143                         spin_unlock(&obd->obd_recovery_task_lock);
3144                         target_exp_dequeue_req_replay(req);
3145                         target_request_copy_put(req);
3146                         RETURN(0);
3147                 }
3148         }
3149 added:
3150         if (!inserted)
3151                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3152
3153         obd->obd_requests_queued_for_recovery++;
3154         spin_unlock(&obd->obd_recovery_task_lock);
3155         wake_up(&obd->obd_next_transno_waitq);
3156         RETURN(0);
3157 }
3158
3159 void target_committed_to_req(struct ptlrpc_request *req)
3160 {
3161         struct obd_export *exp = req->rq_export;
3162
3163         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3164                 lustre_msg_set_last_committed(req->rq_repmsg,
3165                                               exp->exp_last_committed);
3166         else
3167                 DEBUG_REQ(D_IOCTL, req,
3168                           "not sending last_committed update (%d/%d)",
3169                           exp->exp_obd->obd_no_transno,
3170                           req->rq_repmsg == NULL);