Whamcloud - gitweb
LU-14552 ptlrpc: NULL pointer dereference in ptlrpc_watchdog_fire
[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  * Lustre is a trademark of Sun Microsystems, Inc.
31  */
32
33 /**
34  * This file deals with various client/target related logic including recovery.
35  *
36  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
37  * should be moved.
38  */
39
40 #define DEBUG_SUBSYSTEM S_LDLM
41
42 #include <cl_object.h>
43 #include <linux/jiffies.h>
44 #include <linux/kernel.h>
45 #include <linux/kthread.h>
46 #include <libcfs/libcfs.h>
47 #include <obd.h>
48 #include <obd_class.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
53
54 /*
55  * @priority: If non-zero, move the selected connection to the list head.
56  * @create: If zero, only search in existing connections.
57  */
58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
59                            int priority, int create)
60 {
61         struct ptlrpc_connection *ptlrpc_conn;
62         struct obd_import_conn *imp_conn = NULL, *item;
63         lnet_nid_t nid4refnet = LNET_NID_ANY;
64         int rc = 0;
65
66         ENTRY;
67
68         if (!create && !priority) {
69                 CDEBUG(D_HA, "Nothing to do\n");
70                 RETURN(-EINVAL);
71         }
72
73         if (imp->imp_connection &&
74             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
75                 /* nid4refnet is used to restrict network connections */
76                 nid4refnet = imp->imp_connection->c_self;
77         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
78         if (!ptlrpc_conn) {
79                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
80                 RETURN(-ENOENT);
81         }
82
83         if (create) {
84                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
85                 if (!imp_conn)
86                         GOTO(out_put, rc = -ENOMEM);
87         }
88
89         spin_lock(&imp->imp_lock);
90         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
91                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
92                         if (priority) {
93                                 list_move(&item->oic_item,
94                                           &imp->imp_conn_list);
95                                 item->oic_last_attempt = 0;
96                         }
97                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
98                                imp, imp->imp_obd->obd_name, uuid->uuid,
99                                (priority ? ", moved to head" : ""));
100                         spin_unlock(&imp->imp_lock);
101                         GOTO(out_free, rc = 0);
102                 }
103         }
104         /* No existing import connection found for \a uuid. */
105         if (create) {
106                 imp_conn->oic_conn = ptlrpc_conn;
107                 imp_conn->oic_uuid = *uuid;
108                 imp_conn->oic_last_attempt = 0;
109                 if (priority)
110                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
111                 else
112                         list_add_tail(&imp_conn->oic_item,
113                                       &imp->imp_conn_list);
114                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
115                        imp, imp->imp_obd->obd_name, uuid->uuid,
116                        (priority ? "head" : "tail"));
117         } else {
118                 spin_unlock(&imp->imp_lock);
119                 GOTO(out_free, rc = -ENOENT);
120         }
121
122         spin_unlock(&imp->imp_lock);
123         RETURN(0);
124 out_free:
125         if (imp_conn)
126                 OBD_FREE(imp_conn, sizeof(*imp_conn));
127 out_put:
128         ptlrpc_connection_put(ptlrpc_conn);
129         RETURN(rc);
130 }
131
132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
133 {
134         return import_set_conn(imp, uuid, 1, 0);
135 }
136
137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
138                            int priority)
139 {
140         return import_set_conn(imp, uuid, priority, 1);
141 }
142 EXPORT_SYMBOL(client_import_add_conn);
143
144 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
145                                lnet_nid_t prim_nid, int priority)
146 {
147         struct ptlrpc_connection *ptlrpc_conn;
148         int rc;
149
150         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, prim_nid);
151         if (!ptlrpc_conn) {
152                 const char *str_uuid = obd_uuid2str(uuid);
153
154                 rc = class_add_uuid(str_uuid, prim_nid);
155                 if (rc) {
156                         CERROR("%s: failed to add UUID '%s': rc = %d\n",
157                                imp->imp_obd->obd_name, str_uuid, rc);
158                         return rc;
159                 }
160         }
161         return import_set_conn(imp, uuid, priority, 1);
162 }
163 EXPORT_SYMBOL(client_import_dyn_add_conn);
164
165 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
166                                    int nid_count, struct obd_uuid *uuid)
167 {
168         struct obd_import_conn *conn;
169         int rc = -ENOENT;
170
171         ENTRY;
172         if (nid_count <= 0 || !nids)
173                 return rc;
174
175         spin_lock(&imp->imp_lock);
176         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
177                 if (class_check_uuid(&conn->oic_uuid, nids[0])) {
178                         *uuid = conn->oic_uuid;
179                         rc = class_add_nids_to_uuid(&conn->oic_uuid, nids,
180                                                     nid_count);
181                         break;
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         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
987         if (rc) {
988                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
989                         revimp->imp_client->cli_name,
990                         libcfs_id2str(req->rq_peer), rc);
991                 return rc;
992         }
993
994         return 0;
995 }
996
997 /**
998  * Allocate a new reverse import for an export.
999  *
1000  * \retval -errno in case error hit
1001  * \retval 0 if reverse import correctly init
1002  **/
1003 int rev_import_init(struct obd_export *export)
1004 {
1005         struct obd_device *obd = export->exp_obd;
1006         struct obd_import *revimp;
1007
1008         LASSERT(export->exp_imp_reverse == NULL);
1009
1010         revimp = class_new_import(obd);
1011         if (revimp == NULL)
1012                 return -ENOMEM;
1013
1014         revimp->imp_remote_handle.cookie = 0ULL;
1015         revimp->imp_client = &obd->obd_ldlm_client;
1016         revimp->imp_dlm_fake = 1;
1017
1018         /* it is safe to connect import in new state as no sends possible */
1019         spin_lock(&export->exp_lock);
1020         export->exp_imp_reverse = revimp;
1021         spin_unlock(&export->exp_lock);
1022         class_import_put(revimp);
1023
1024         return 0;
1025 }
1026 EXPORT_SYMBOL(rev_import_init);
1027
1028 /**
1029  * Handle reconnect for an export.
1030  *
1031  * \param exp export to handle reconnect process
1032  * \param req client reconnect request
1033  *
1034  * \retval -rc in case securitfy flavor can't be changed
1035  * \retval 0 in case none problems
1036  */
1037 static int rev_import_reconnect(struct obd_export *exp,
1038                                 struct ptlrpc_request *req)
1039 {
1040         struct obd_import *revimp = exp->exp_imp_reverse;
1041         struct lustre_handle *lh;
1042         int rc;
1043
1044         /* avoid sending a request until import flags are changed */
1045         ptlrpc_import_enter_resend(revimp);
1046
1047         ptlrpc_connection_put(revimp->imp_connection);
1048
1049         /*
1050          * client from recovery don't have a handle so we need to take from
1051          * request. it may produce situation when wrong client connected
1052          * to recovery as we trust a client uuid
1053          */
1054         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1055         revimp->imp_remote_handle = *lh;
1056
1057         /*
1058          * unknown versions will be caught in
1059          * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1060          */
1061         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1062
1063         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1064
1065         rc = rev_import_flags_update(revimp, req);
1066         if (rc != 0) {
1067                 /*
1068                  * it is safe to still be in RECOVERY phase as we are not able
1069                  * to setup correct security flavor so requests are not able to
1070                  * be delivered correctly
1071                  */
1072                 return rc;
1073         }
1074
1075         /* resend all rpc's via new connection */
1076         return ptlrpc_import_recovery_state_machine(revimp);
1077 }
1078
1079 int target_handle_connect(struct ptlrpc_request *req)
1080 {
1081         struct obd_device *target = NULL;
1082         struct obd_export *export = NULL;
1083         /*
1084          * connect handle - filled from target_handle_reconnect in
1085          * reconnect case
1086          */
1087         struct lustre_handle conn;
1088         struct lustre_handle *tmp;
1089         struct obd_uuid cluuid;
1090         char *str;
1091         int rc = 0;
1092         char *target_start;
1093         int target_len;
1094         bool mds_conn = false, lw_client = false, initial_conn = false;
1095         bool mds_mds_conn = false;
1096         bool new_mds_mds_conn = false;
1097         struct obd_connect_data *data, *tmpdata;
1098         int size, tmpsize;
1099         lnet_nid_t *client_nid = NULL;
1100
1101         ENTRY;
1102
1103         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1104
1105         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1106         if (str == NULL) {
1107                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1108                 GOTO(out, rc = -EINVAL);
1109         }
1110
1111         target = class_dev_by_str(str);
1112         if (!target) {
1113                 deuuidify(str, NULL, &target_start, &target_len);
1114                 LCONSOLE_ERROR_MSG(0x137,
1115                                    "%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",
1116                                    str, libcfs_nid2str(req->rq_peer.nid));
1117                 GOTO(out, rc = -ENODEV);
1118         }
1119
1120         spin_lock(&target->obd_dev_lock);
1121
1122         target->obd_conn_inprogress++;
1123
1124         if (target->obd_stopping || !target->obd_set_up) {
1125                 spin_unlock(&target->obd_dev_lock);
1126
1127                 deuuidify(str, NULL, &target_start, &target_len);
1128                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1129                               target_len, target_start,
1130                               libcfs_nid2str(req->rq_peer.nid),
1131                               (target->obd_stopping ?
1132                                "stopping" : "not set up"));
1133                 GOTO(out, rc = -ENODEV);
1134         }
1135
1136         if (target->obd_no_conn) {
1137                 spin_unlock(&target->obd_dev_lock);
1138
1139                 CDEBUG(D_INFO,
1140                        "%s: Temporarily refusing client connection from %s\n",
1141                        target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1142                 GOTO(out, rc = -EAGAIN);
1143         }
1144
1145         spin_unlock(&target->obd_dev_lock);
1146
1147         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1148         if (str == NULL) {
1149                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1150                 GOTO(out, rc = -EINVAL);
1151         }
1152
1153         obd_str2uuid(&cluuid, str);
1154
1155         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1156         if (tmp == NULL)
1157                 GOTO(out, rc = -EPROTO);
1158
1159         conn = *tmp;
1160
1161         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1162                                     RCL_CLIENT);
1163         if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1164                 GOTO(out, rc = -EPROTO);
1165         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1166         if (!data)
1167                 GOTO(out, rc = -EPROTO);
1168
1169         rc = req_capsule_server_pack(&req->rq_pill);
1170         if (rc)
1171                 GOTO(out, rc);
1172
1173 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1174         /*
1175          * Don't allow clients to connect that are using old 1.8 format
1176          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1177          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1178          * FULL20 flag should be set on all connections since 2.0, but no
1179          * longer affects behaviour.
1180          *
1181          * Later this check will be disabled and the flag can be retired
1182          * completely once interop with 3.0 is no longer needed.
1183          */
1184         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1185                 GOTO(out, rc = -EPROTO);
1186
1187         /*
1188          * Don't allow liblustre clients to connect.
1189          * - testing was disabled in v2_2_50_0-61-g6a75d65
1190          * - building was disabled in v2_5_58_0-28-g7277179
1191          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1192          * - clients were refused connect for version difference > 0.0.1.32
1193          */
1194         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1195                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1196                 GOTO(out, rc = -EPROTO);
1197         }
1198 #endif
1199
1200         /*
1201          * Note: lw_client is needed in MDS-MDS failover during update log
1202          * processing, so we needs to allow lw_client to be connected at
1203          * anytime, instead of only the initial connection
1204          */
1205         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1206
1207         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1208                 initial_conn = true;
1209                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1210                 mds_mds_conn = (data->ocd_connect_flags &
1211                                 OBD_CONNECT_MDS_MDS) != 0;
1212
1213                 /*
1214                  * OBD_CONNECT_MNE_SWAB is removed at 2.14
1215                  * Checking OBD_CONNECT_FID can be removed in the future.
1216                  *
1217                  * Via check OBD_CONNECT_FID, we can distinguish whether
1218                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1219                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1220                  */
1221                 if (!lw_client &&
1222                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1223                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1224                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1225                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1226                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1227                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1228
1229                         /*
1230                          * We do not support the MDT-MDT interoperations with
1231                          * different version MDT because of protocol changes.
1232                          */
1233                         if (unlikely(major != LUSTRE_MAJOR ||
1234                                      minor != LUSTRE_MINOR ||
1235                                      abs(patch - LUSTRE_PATCH) > 3)) {
1236                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1237                                               target->obd_name, LUSTRE_MAJOR,
1238                                               LUSTRE_MINOR, LUSTRE_PATCH,
1239                                               LUSTRE_FIX, major, minor, patch,
1240                                               OBD_OCD_VERSION_FIX(data->ocd_version),
1241                                               libcfs_nid2str(req->rq_peer.nid),
1242                                               str);
1243                                 GOTO(out, rc = -EPROTO);
1244                         }
1245                 }
1246         }
1247
1248         /* lctl gets a backstage, all-access pass. */
1249         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1250                 goto dont_check_exports;
1251
1252         export = obd_uuid_lookup(target, &cluuid);
1253         if (!export)
1254                 goto no_export;
1255
1256         /* We've found an export in the hash. */
1257
1258         spin_lock(&export->exp_lock);
1259
1260         if (export->exp_connecting) { /* b=9635, et. al. */
1261                 spin_unlock(&export->exp_lock);
1262                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1263                               export->exp_obd->obd_name, export,
1264                               libcfs_nid2str(req->rq_peer.nid));
1265                 class_export_put(export);
1266                 export = NULL;
1267                 rc = -EALREADY;
1268         } else if ((mds_conn || (lw_client && initial_conn) ||
1269                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1270                    export->exp_connection != NULL) {
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                                       mds_conn ? "MDS" : "LWP",
1277                                       libcfs_nid2str(req->rq_peer.nid),
1278                                       libcfs_nid2str(export->exp_connection->c_peer.nid));
1279                 } else {
1280                         /* New MDS connection from the same NID. */
1281                         LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1282                                       target->obd_name,
1283                                       mds_conn ? "MDS" : "LWP",
1284                                       libcfs_nid2str(req->rq_peer.nid));
1285                 }
1286
1287                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1288                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1289                         /*
1290                          * Because exports between MDTs will always be
1291                          * kept, let's do not fail such export if they
1292                          * come from the same NID, otherwise it might
1293                          * cause eviction between MDTs, which might
1294                          * cause namespace inconsistency
1295                          */
1296                         spin_lock(&export->exp_lock);
1297                         export->exp_connecting = 1;
1298                         export->exp_conn_cnt = 0;
1299                         spin_unlock(&export->exp_lock);
1300                         conn.cookie = export->exp_handle.h_cookie;
1301                         rc = EALREADY;
1302                 } else {
1303                         class_fail_export(export);
1304                         class_export_put(export);
1305                         export = NULL;
1306                         rc = 0;
1307                 }
1308         } else if (export->exp_connection != NULL && initial_conn &&
1309                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1310                 spin_unlock(&export->exp_lock);
1311                 /* In MDS failover we have static UUID but NID can change. */
1312                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1313                               target->obd_name, cluuid.uuid,
1314                               libcfs_nid2str(req->rq_peer.nid),
1315                               libcfs_nid2str(
1316                                         export->exp_connection->c_peer.nid));
1317                 rc = -EALREADY;
1318                 class_export_put(export);
1319                 export = NULL;
1320         } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1321                    !lw_client) {
1322                 spin_unlock(&export->exp_lock);
1323                 rc = -EAGAIN;
1324         } else {
1325                 export->exp_connecting = 1;
1326                 spin_unlock(&export->exp_lock);
1327                 LASSERT(export->exp_obd == target);
1328
1329                 rc = target_handle_reconnect(&conn, export, &cluuid);
1330         }
1331
1332         /* If we found an export, we already unlocked. */
1333         if (!export) {
1334 no_export:
1335                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1336         } else if (req->rq_export == NULL &&
1337                    atomic_read(&export->exp_rpc_count) > 0) {
1338                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1339                               target->obd_name, cluuid.uuid,
1340                               libcfs_nid2str(req->rq_peer.nid),
1341                               refcount_read(&export->exp_handle.h_ref));
1342                         GOTO(out, rc = -EBUSY);
1343         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1344                    rc != EALREADY) {
1345                 if (!strstr(cluuid.uuid, "mdt"))
1346                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1347                                       target->obd_name, cluuid.uuid,
1348                                       libcfs_nid2str(req->rq_peer.nid));
1349                 GOTO(out, rc = -EALREADY);
1350         } else {
1351                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1352         }
1353
1354         if (rc < 0)
1355                 GOTO(out, rc);
1356
1357         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1358                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1359                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1360                export, ktime_get_seconds(),
1361                export ? export->exp_last_request_time : 0);
1362
1363         /*
1364          * If this is the first time a client connects, reset the recovery
1365          * timer. Discard lightweight connections which might be local.
1366          */
1367         if (!lw_client && rc == 0 && target->obd_recovering)
1368                 check_and_start_recovery_timer(target, req, export == NULL);
1369
1370         /*
1371          * We want to handle EALREADY but *not* -EALREADY from
1372          * target_handle_reconnect(), return reconnection state in a flag.
1373          */
1374         if (rc == EALREADY) {
1375                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1376                 rc = 0;
1377         } else {
1378                 LASSERT(rc == 0);
1379         }
1380
1381         /* Tell the client if we support replayable requests. */
1382         if (target->obd_replayable)
1383                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1384         client_nid = &req->rq_peer.nid;
1385
1386         if (export == NULL) {
1387                 /* allow lightweight connections during recovery */
1388                 /*
1389                  * allow "new" MDT to be connected during recovery, since we
1390                  * need retrieve recovery update records from it
1391                  */
1392                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1393                         struct hrtimer *timer = &target->obd_recovery_timer;
1394                         ktime_t remaining;
1395                         s64 timeout, left;
1396                         int in_progress;
1397                         int connected;
1398                         int known;
1399                         int stale;
1400                         char *msg;
1401
1402                         connected = atomic_read(&target->obd_connected_clients);
1403                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1404                         known =
1405                            atomic_read(&target->obd_max_recoverable_clients);
1406                         stale = target->obd_stale_clients;
1407                         remaining = hrtimer_get_remaining(timer);
1408                         left = ktime_divns(remaining, NSEC_PER_SEC);
1409
1410                         if (ktime_to_ns(remaining) > 0) {
1411                                 msg = "to recover in";
1412                                 timeout = left;
1413                         } else {
1414                                 msg = "already passed deadline";
1415                                 timeout = -left;
1416
1417                                 target_check_recovery_timer(target);
1418                         }
1419
1420                         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",
1421                                       target->obd_name, cluuid.uuid,
1422                                       libcfs_nid2str(req->rq_peer.nid), known,
1423                                       connected - in_progress, in_progress,
1424                                       stale, msg, timeout / 60, timeout % 60);
1425                         rc = -EBUSY;
1426                 } else {
1427 dont_check_exports:
1428                         rc = obd_connect(req->rq_svc_thread->t_env,
1429                                          &export, target, &cluuid, data,
1430                                          client_nid);
1431                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1432                                 lustre_msg_add_op_flags(req->rq_repmsg,
1433                                                         MSG_CONNECT_RECOVERING);
1434                         if (rc == 0) {
1435                                 conn.cookie = export->exp_handle.h_cookie;
1436                                 rc = rev_import_init(export);
1437                         }
1438
1439                         if (mds_mds_conn)
1440                                 new_mds_mds_conn = true;
1441                 }
1442         } else {
1443                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1444                                    export, target, &cluuid, data, client_nid);
1445         }
1446         if (rc)
1447                 GOTO(out, rc);
1448
1449         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1450         data->ocd_instance = target->u.obt.obt_instance;
1451
1452         /*
1453          * Return only the parts of obd_connect_data that we understand, so the
1454          * client knows that we don't understand the rest.
1455          */
1456         if (data) {
1457                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1458                                                RCL_SERVER);
1459                 tmpdata = req_capsule_server_get(&req->rq_pill,
1460                                                  &RMF_CONNECT_DATA);
1461                 /*
1462                  * Don't use struct assignment here, because the client reply
1463                  * buffer may be smaller/larger than the local struct
1464                  * obd_connect_data.
1465                  */
1466                 memcpy(tmpdata, data, min(tmpsize, size));
1467         }
1468
1469         /*
1470          * If the client and the server are the same node, we will already
1471          * have an export that really points to the client's DLM export,
1472          * because we have a shared handles table.
1473          *
1474          * XXX this will go away when shaver stops sending the "connect" handle
1475          * in the real "remote handle" field of the request --phik 24 Apr 2003
1476          */
1477         ptlrpc_request_change_export(req, export);
1478
1479         spin_lock(&export->exp_lock);
1480         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1481                 spin_unlock(&export->exp_lock);
1482                 CDEBUG(D_RPCTRACE,
1483                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1484                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1485                        export->exp_conn_cnt,
1486                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1487
1488                 GOTO(out, rc = -EALREADY);
1489         }
1490         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1491         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1492         spin_unlock(&export->exp_lock);
1493
1494         if (export->exp_connection != NULL) {
1495                 /* Check to see if connection came from another NID. */
1496                 if (export->exp_connection->c_peer.nid != req->rq_peer.nid)
1497                         obd_nid_del(export->exp_obd, export);
1498
1499                 ptlrpc_connection_put(export->exp_connection);
1500         }
1501
1502         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1503                                                        req->rq_self,
1504                                                        &cluuid);
1505         if (!export->exp_connection)
1506                 GOTO(out, rc = -ENOTCONN);
1507
1508         obd_nid_add(export->exp_obd, export);
1509
1510         lustre_msg_set_handle(req->rq_repmsg, &conn);
1511
1512         rc = rev_import_reconnect(export, req);
1513         if (rc != 0)
1514                 GOTO(out, rc);
1515
1516         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1517                 int has_transno;
1518                 __u64 transno = data->ocd_transno;
1519
1520                 spin_lock(&export->exp_lock);
1521                 /*
1522                  * possible race with class_disconnect_stale_exports,
1523                  * export may be already in the eviction process
1524                  */
1525                 if (export->exp_failed) {
1526                         spin_unlock(&export->exp_lock);
1527                         GOTO(out, rc = -ENODEV);
1528                 }
1529                 export->exp_in_recovery = 1;
1530                 export->exp_req_replay_needed = 1;
1531                 export->exp_lock_replay_needed = 1;
1532                 spin_unlock(&export->exp_lock);
1533
1534                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1535                                  MSG_CONNECT_TRANSNO);
1536                 if (has_transno && transno == 0)
1537                         CWARN("Connect with zero transno!\n");
1538
1539                 if (has_transno && transno > 0 &&
1540                     transno < target->obd_next_recovery_transno &&
1541                     transno > target->obd_last_committed) {
1542                         /* Another way is to use cmpxchg() to be lock-free. */
1543                         spin_lock(&target->obd_recovery_task_lock);
1544                         if (transno < target->obd_next_recovery_transno)
1545                                 target->obd_next_recovery_transno = transno;
1546                         spin_unlock(&target->obd_recovery_task_lock);
1547                 }
1548
1549                 atomic_inc(&target->obd_req_replay_clients);
1550                 atomic_inc(&target->obd_lock_replay_clients);
1551                 /*
1552                  * Note: MDS-MDS connection is allowed to be connected during
1553                  * recovery, no matter if the exports needs to be recoveried.
1554                  * Because we need retrieve updates logs from all other MDTs.
1555                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1556                  * also needs to be increased to match other recovery checking
1557                  * condition.
1558                  */
1559                 if (new_mds_mds_conn)
1560                         atomic_inc(&target->obd_max_recoverable_clients);
1561
1562                 if (atomic_inc_return(&target->obd_connected_clients) ==
1563                     atomic_read(&target->obd_max_recoverable_clients))
1564                         wake_up(&target->obd_next_transno_waitq);
1565         }
1566
1567         /* Tell the client we're in recovery, when client is involved in it. */
1568         if (target->obd_recovering && !lw_client)
1569                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1570
1571 out:
1572         if (export) {
1573                 spin_lock(&export->exp_lock);
1574                 export->exp_connecting = 0;
1575                 spin_unlock(&export->exp_lock);
1576
1577                 class_export_put(export);
1578         }
1579         if (target != NULL) {
1580                 spin_lock(&target->obd_dev_lock);
1581                 target->obd_conn_inprogress--;
1582                 spin_unlock(&target->obd_dev_lock);
1583                 class_decref(target, "find", current);
1584         }
1585         req->rq_status = rc;
1586         RETURN(rc);
1587 }
1588
1589 int target_handle_disconnect(struct ptlrpc_request *req)
1590 {
1591         int rc;
1592
1593         ENTRY;
1594
1595         rc = req_capsule_server_pack(&req->rq_pill);
1596         if (rc)
1597                 RETURN(rc);
1598
1599         /* In case of target disconnect, updating sec ctx immediately is
1600          * required in order to record latest sequence number used.
1601          * Sequence is normally updated on export destroy, but this event
1602          * can occur too late, ie after a new target connect request has
1603          * been processed.
1604          * Maintaining correct sequence when client connection becomes idle
1605          * ensures that GSS does not erroneously consider requests as replays.
1606          */
1607         rc = sptlrpc_export_update_ctx(req->rq_export);
1608         if (rc)
1609                 RETURN(rc);
1610
1611         /* Keep the rq_export around so we can send the reply. */
1612         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1613
1614         RETURN(0);
1615 }
1616
1617 void target_destroy_export(struct obd_export *exp)
1618 {
1619         struct obd_import *imp = NULL;
1620         /*
1621          * exports created from last_rcvd data, and "fake"
1622          * exports created by lctl don't have an import
1623          */
1624         spin_lock(&exp->exp_lock);
1625         if (exp->exp_imp_reverse != NULL) {
1626                 imp = exp->exp_imp_reverse;
1627                 exp->exp_imp_reverse = NULL;
1628         }
1629         spin_unlock(&exp->exp_lock);
1630         if (imp != NULL)
1631                 client_destroy_import(imp);
1632
1633         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1634         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1635         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1636         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1637 }
1638 EXPORT_SYMBOL(target_destroy_export);
1639
1640 /*
1641  * Recovery functions
1642  */
1643 static void target_request_copy_get(struct ptlrpc_request *req)
1644 {
1645         class_export_rpc_inc(req->rq_export);
1646         LASSERT(list_empty(&req->rq_list));
1647         INIT_LIST_HEAD(&req->rq_replay_list);
1648
1649         /* Increase refcount to keep request in queue. */
1650         atomic_inc(&req->rq_refcount);
1651         /* Let export know it has replays to be handled. */
1652         atomic_inc(&req->rq_export->exp_replay_count);
1653 }
1654
1655 static void target_request_copy_put(struct ptlrpc_request *req)
1656 {
1657         LASSERT(list_empty(&req->rq_replay_list));
1658         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1659
1660         atomic_dec(&req->rq_export->exp_replay_count);
1661         class_export_rpc_dec(req->rq_export);
1662         ptlrpc_server_drop_request(req);
1663 }
1664
1665 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1666 {
1667         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1668         struct obd_export *exp = req->rq_export;
1669         struct ptlrpc_request *reqiter;
1670         struct ptlrpc_request *dup_req = NULL;
1671         int dup = 0;
1672
1673         LASSERT(exp);
1674
1675         spin_lock(&exp->exp_lock);
1676         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1677                             rq_replay_list) {
1678                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1679                         dup_req = reqiter;
1680                         dup = 1;
1681                         break;
1682                 }
1683         }
1684
1685         if (dup) {
1686                 /* We expect it with RESENT and REPLAY flags. */
1687                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1688                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1689                         CERROR("invalid flags %x of resent replay\n",
1690                                lustre_msg_get_flags(req->rq_reqmsg));
1691
1692                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1693                         __u32 new_conn;
1694
1695                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1696                         if (new_conn >
1697                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1698                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1699                                                         new_conn);
1700                 }
1701         } else {
1702                 list_add_tail(&req->rq_replay_list,
1703                               &exp->exp_req_replay_queue);
1704         }
1705
1706         spin_unlock(&exp->exp_lock);
1707         return dup;
1708 }
1709
1710 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1711 {
1712         LASSERT(!list_empty(&req->rq_replay_list));
1713         LASSERT(req->rq_export);
1714
1715         spin_lock(&req->rq_export->exp_lock);
1716         list_del_init(&req->rq_replay_list);
1717         spin_unlock(&req->rq_export->exp_lock);
1718 }
1719
1720 static void target_finish_recovery(struct lu_target *lut)
1721 {
1722         struct obd_device *obd = lut->lut_obd;
1723
1724         ENTRY;
1725
1726         /* Only log a recovery message when recovery has occurred. */
1727         if (obd->obd_recovery_start) {
1728                 time64_t now = ktime_get_seconds();
1729                 time64_t elapsed_time;
1730
1731                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1732                                      1);
1733                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1734                               obd->obd_name, elapsed_time / 60,
1735                               elapsed_time % 60,
1736                               atomic_read(&obd->obd_max_recoverable_clients),
1737                               atomic_read(&obd->obd_connected_clients),
1738                               obd->obd_stale_clients,
1739                               obd->obd_stale_clients == 1 ? "was" : "were");
1740         }
1741
1742         ldlm_reprocess_recovery_done(obd->obd_namespace);
1743         spin_lock(&obd->obd_recovery_task_lock);
1744         if (!list_empty(&obd->obd_req_replay_queue) ||
1745             !list_empty(&obd->obd_lock_replay_queue) ||
1746             !list_empty(&obd->obd_final_req_queue)) {
1747                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1748                        obd->obd_name,
1749                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1750                        list_empty(&obd->obd_lock_replay_queue) ? \
1751                                   "" : "lock ",
1752                        list_empty(&obd->obd_final_req_queue) ? \
1753                                   "" : "final ");
1754                 spin_unlock(&obd->obd_recovery_task_lock);
1755                 LBUG();
1756         }
1757         spin_unlock(&obd->obd_recovery_task_lock);
1758
1759         obd->obd_recovery_end = ktime_get_seconds();
1760
1761         /* When recovery finished, cleanup orphans on MDS and OST. */
1762         if (obd->obd_type && OBP(obd, postrecov)) {
1763                 int rc = OBP(obd, postrecov)(obd);
1764
1765                 if (rc < 0)
1766                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1767                                       obd->obd_name, rc);
1768         }
1769         EXIT;
1770 }
1771
1772 static void abort_req_replay_queue(struct obd_device *obd)
1773 {
1774         struct ptlrpc_request *req, *n;
1775         LIST_HEAD(abort_list);
1776
1777         spin_lock(&obd->obd_recovery_task_lock);
1778         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1779         spin_unlock(&obd->obd_recovery_task_lock);
1780         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1781                 DEBUG_REQ(D_WARNING, req, "aborted:");
1782                 req->rq_status = -ENOTCONN;
1783                 if (ptlrpc_error(req)) {
1784                         DEBUG_REQ(D_ERROR, req,
1785                                   "failed abort_req_reply; skipping");
1786                 }
1787                 target_exp_dequeue_req_replay(req);
1788                 target_request_copy_put(req);
1789         }
1790 }
1791
1792 static void abort_lock_replay_queue(struct obd_device *obd)
1793 {
1794         struct ptlrpc_request *req, *n;
1795         LIST_HEAD(abort_list);
1796
1797         spin_lock(&obd->obd_recovery_task_lock);
1798         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1799         spin_unlock(&obd->obd_recovery_task_lock);
1800         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1801                 DEBUG_REQ(D_ERROR, req, "aborted:");
1802                 req->rq_status = -ENOTCONN;
1803                 if (ptlrpc_error(req)) {
1804                         DEBUG_REQ(D_ERROR, req,
1805                                   "failed abort_lock_reply; skipping");
1806                 }
1807                 target_request_copy_put(req);
1808         }
1809 }
1810
1811 /*
1812  * Called from a cleanup function if the device is being cleaned up
1813  * forcefully.  The exports should all have been disconnected already,
1814  * the only thing left to do is
1815  * - clear the recovery flags
1816  * - cancel the timer
1817  * - free queued requests and replies, but don't send replies
1818  * Because the obd_stopping flag is set, no new requests should be received.
1819  */
1820 void target_cleanup_recovery(struct obd_device *obd)
1821 {
1822         struct ptlrpc_request *req, *n;
1823         LIST_HEAD(clean_list);
1824
1825         spin_lock(&obd->obd_dev_lock);
1826         if (!obd->obd_recovering) {
1827                 spin_unlock(&obd->obd_dev_lock);
1828                 EXIT;
1829                 return;
1830         }
1831         obd->obd_recovering = obd->obd_abort_recovery = 0;
1832         spin_unlock(&obd->obd_dev_lock);
1833
1834         spin_lock(&obd->obd_recovery_task_lock);
1835         target_cancel_recovery_timer(obd);
1836         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1837         spin_unlock(&obd->obd_recovery_task_lock);
1838
1839         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1840                 LASSERT(req->rq_reply_state == NULL);
1841                 target_exp_dequeue_req_replay(req);
1842                 target_request_copy_put(req);
1843         }
1844
1845         spin_lock(&obd->obd_recovery_task_lock);
1846         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1847         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1848         spin_unlock(&obd->obd_recovery_task_lock);
1849
1850         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1851                 LASSERT(req->rq_reply_state == NULL);
1852                 target_request_copy_put(req);
1853         }
1854
1855         EXIT;
1856 }
1857 EXPORT_SYMBOL(target_cleanup_recovery);
1858
1859 /* obd_recovery_task_lock should be held */
1860 void target_cancel_recovery_timer(struct obd_device *obd)
1861 {
1862         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1863         hrtimer_cancel(&obd->obd_recovery_timer);
1864 }
1865
1866 static void target_start_recovery_timer(struct obd_device *obd)
1867 {
1868         ktime_t delay;
1869
1870         if (obd->obd_recovery_start != 0)
1871                 return;
1872
1873         spin_lock(&obd->obd_dev_lock);
1874         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1875                 spin_unlock(&obd->obd_dev_lock);
1876                 return;
1877         }
1878
1879         LASSERT(obd->obd_recovery_timeout != 0);
1880
1881         if (obd->obd_recovery_start != 0) {
1882                 spin_unlock(&obd->obd_dev_lock);
1883                 return;
1884         }
1885
1886         obd->obd_recovery_start = ktime_get_seconds();
1887         delay = ktime_set(obd->obd_recovery_start +
1888                           obd->obd_recovery_timeout, 0);
1889         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1890         spin_unlock(&obd->obd_dev_lock);
1891
1892         LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1893                       obd->obd_name,
1894                       obd->obd_recovery_timeout / 60,
1895                       obd->obd_recovery_timeout % 60,
1896                       atomic_read(&obd->obd_max_recoverable_clients),
1897                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1898                       "" : "s",
1899                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1900                       "s" : "");
1901 }
1902
1903 /**
1904  * extend recovery window.
1905  *
1906  * if @extend is true, extend recovery window to have @dr_timeout remaining
1907  * at least; otherwise, make sure the recovery timeout value is not less
1908  * than @dr_timeout.
1909  */
1910 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1911                                   bool extend)
1912 {
1913         ktime_t left_ns;
1914         timeout_t timeout;
1915         timeout_t left;
1916
1917         spin_lock(&obd->obd_dev_lock);
1918         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1919             obd->obd_stopping) {
1920                 spin_unlock(&obd->obd_dev_lock);
1921                 return;
1922         }
1923         LASSERT(obd->obd_recovery_start != 0);
1924
1925         left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1926         left = ktime_divns(left_ns, NSEC_PER_SEC);
1927
1928         if (extend) {
1929                 timeout = obd->obd_recovery_timeout;
1930                 /* dr_timeout will happen after the hrtimer has expired.
1931                  * Add the excess time to the soft recovery timeout without
1932                  * exceeding the hard recovery timeout.
1933                  */
1934                 if (dr_timeout > left) {
1935                         timeout += dr_timeout - left;
1936                         timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1937                                         timeout);
1938                 }
1939         } else {
1940                 timeout = clamp_t(timeout_t, dr_timeout,
1941                                   obd->obd_recovery_timeout,
1942                                   obd->obd_recovery_time_hard);
1943         }
1944
1945         if (timeout == obd->obd_recovery_time_hard)
1946                 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1947                       obd->obd_name, timeout, extend);
1948
1949         if (obd->obd_recovery_timeout < timeout) {
1950                 ktime_t end, now;
1951
1952                 obd->obd_recovery_timeout = timeout;
1953                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1954                 now = ktime_set(ktime_get_seconds(), 0);
1955                 left_ns = ktime_sub(end, now);
1956                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1957                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1958         }
1959         spin_unlock(&obd->obd_dev_lock);
1960
1961         CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1962                 obd->obd_name, left);
1963 }
1964
1965 /* Reset the timer with each new client connection */
1966 /*
1967  * This timer is actually reconnect_timer, which is for making sure
1968  * the total recovery window is at least as big as my reconnect
1969  * attempt timing. So the initial recovery time_out will be set to
1970  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1971  * from client is bigger than this, then the recovery time_out will
1972  * be extended to make sure the client could be reconnected, in the
1973  * process, the timeout from the new client should be ignored.
1974  */
1975 static void
1976 check_and_start_recovery_timer(struct obd_device *obd,
1977                                struct ptlrpc_request *req,
1978                                int new_client)
1979 {
1980         timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
1981         struct obd_device_target *obt = &obd->u.obt;
1982
1983         if (!new_client && service_timeout)
1984                 /*
1985                  * Teach server about old server's estimates, as first guess
1986                  * at how long new requests will take.
1987                  */
1988                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1989                             service_timeout);
1990
1991         target_start_recovery_timer(obd);
1992
1993         /*
1994          * Convert the service time to RPC timeout,
1995          * and reuse service_timeout to limit stack usage.
1996          */
1997         service_timeout = at_est2timeout(service_timeout);
1998
1999         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2000             service_timeout < at_extra)
2001                 service_timeout = at_extra;
2002
2003         /*
2004          * We expect other clients to timeout within service_timeout, then try
2005          * to reconnect, then try the failover server.  The max delay between
2006          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2007          */
2008         service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2009
2010         LASSERT(obt->obt_magic == OBT_MAGIC);
2011         service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2012         if (service_timeout > obd->obd_recovery_timeout && !new_client)
2013                 extend_recovery_timer(obd, service_timeout, false);
2014 }
2015
2016 /** Health checking routines */
2017 static inline int exp_connect_healthy(struct obd_export *exp)
2018 {
2019         return exp->exp_in_recovery;
2020 }
2021
2022 /** if export done req_replay or has replay in queue */
2023 static inline int exp_req_replay_healthy(struct obd_export *exp)
2024 {
2025         return (!exp->exp_req_replay_needed ||
2026                 atomic_read(&exp->exp_replay_count) > 0);
2027 }
2028
2029
2030 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2031 {
2032         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2033                exp_req_replay_healthy(exp);
2034 }
2035
2036 /** if export done lock_replay or has replay in queue */
2037 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2038 {
2039         return (!exp->exp_lock_replay_needed ||
2040                 atomic_read(&exp->exp_replay_count) > 0);
2041 }
2042
2043 static inline int exp_vbr_healthy(struct obd_export *exp)
2044 {
2045         return !exp->exp_vbr_failed;
2046 }
2047
2048 static inline int exp_finished(struct obd_export *exp)
2049 {
2050         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2051 }
2052
2053 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2054 {
2055         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2056                 exp_finished(exp);
2057 }
2058
2059 static int check_for_next_transno(struct lu_target *lut)
2060 {
2061         struct ptlrpc_request *req = NULL;
2062         struct obd_device *obd = lut->lut_obd;
2063         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2064         int wake_up = 0, connected, completed, queue_len;
2065         __u64 req_transno = 0;
2066         __u64 update_transno = 0;
2067         __u64 next_transno = 0;
2068
2069         ENTRY;
2070
2071         spin_lock(&obd->obd_recovery_task_lock);
2072         if (!list_empty(&obd->obd_req_replay_queue)) {
2073                 req = list_entry(obd->obd_req_replay_queue.next,
2074                                      struct ptlrpc_request, rq_list);
2075                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2076         }
2077
2078         if (!obd->obd_abort_recov_mdt && tdtd)
2079                 update_transno = distribute_txn_get_next_transno(tdtd);
2080
2081         connected = atomic_read(&obd->obd_connected_clients);
2082         completed = connected - atomic_read(&obd->obd_req_replay_clients);
2083         queue_len = obd->obd_requests_queued_for_recovery;
2084         next_transno = obd->obd_next_recovery_transno;
2085
2086         CDEBUG(D_HA,
2087                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2088                atomic_read(&obd->obd_max_recoverable_clients),
2089                connected, completed,
2090                queue_len, req_transno, next_transno);
2091
2092         if (obd->obd_abort_recovery) {
2093                 CDEBUG(D_HA, "waking for aborted recovery\n");
2094                 wake_up = 1;
2095         } else if (obd->obd_recovery_expired) {
2096                 CDEBUG(D_HA, "waking for expired recovery\n");
2097                 wake_up = 1;
2098         } else if (!obd->obd_abort_recov_mdt && tdtd && req &&
2099                    is_req_replayed_by_update(req)) {
2100                 LASSERTF(req_transno < next_transno,
2101                          "req_transno %llu next_transno%llu\n", req_transno,
2102                          next_transno);
2103                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2104                        req_transno);
2105                 wake_up = 1;
2106         } else if (req_transno == next_transno ||
2107                    (update_transno != 0 && update_transno <= next_transno)) {
2108                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2109                 wake_up = 1;
2110         } else if (queue_len > 0 &&
2111                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2112                 /** handle gaps occured due to lost reply or VBR */
2113                 LASSERTF(req_transno >= next_transno,
2114                          "req_transno: %llu, next_transno: %llu\n",
2115                          req_transno, next_transno);
2116                 CDEBUG(D_HA,
2117                        "%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",
2118                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2119                        next_transno, queue_len, completed, connected,
2120                        req_transno, update_transno, obd->obd_last_committed);
2121                 obd->obd_next_recovery_transno = req_transno;
2122                 wake_up = 1;
2123         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2124                 CDEBUG(D_HA, "waking for completed recovery\n");
2125                 wake_up = 1;
2126         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2127                 CDEBUG(D_HA,
2128                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2129                        next_transno);
2130                 obd->obd_next_recovery_transno = req_transno;
2131                 wake_up = 1;
2132         }
2133         spin_unlock(&obd->obd_recovery_task_lock);
2134         return wake_up;
2135 }
2136
2137 static int check_for_next_lock(struct lu_target *lut)
2138 {
2139         struct obd_device *obd = lut->lut_obd;
2140         int wake_up = 0;
2141
2142         spin_lock(&obd->obd_recovery_task_lock);
2143         if (!list_empty(&obd->obd_lock_replay_queue)) {
2144                 CDEBUG(D_HA, "waking for next lock\n");
2145                 wake_up = 1;
2146         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2147                 CDEBUG(D_HA, "waking for completed lock replay\n");
2148                 wake_up = 1;
2149         } else if (obd->obd_abort_recovery) {
2150                 CDEBUG(D_HA, "waking for aborted recovery\n");
2151                 wake_up = 1;
2152         } else if (obd->obd_recovery_expired) {
2153                 CDEBUG(D_HA, "waking for expired recovery\n");
2154                 wake_up = 1;
2155         }
2156         spin_unlock(&obd->obd_recovery_task_lock);
2157
2158         return wake_up;
2159 }
2160
2161 static int check_update_llog(struct lu_target *lut)
2162 {
2163         struct obd_device *obd = lut->lut_obd;
2164         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2165
2166         if (obd->obd_abort_recovery) {
2167                 CDEBUG(D_HA, "waking for aborted recovery\n");
2168                 return 1;
2169         }
2170
2171         if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2172                 CDEBUG(D_HA, "waking for completion of reading update log\n");
2173                 return 1;
2174         }
2175
2176         return 0;
2177 }
2178
2179 /**
2180  * wait for recovery events,
2181  * check its status with help of check_routine
2182  * evict dead clients via health_check
2183  */
2184 static int target_recovery_overseer(struct lu_target *lut,
2185                                     int (*check_routine)(struct lu_target *),
2186                                     int (*health_check)(struct obd_export *))
2187 {
2188         struct obd_device *obd = lut->lut_obd;
2189         struct target_distribute_txn_data *tdtd;
2190         time64_t last = 0;
2191         time64_t now;
2192 repeat:
2193         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2194                 now = ktime_get_seconds();
2195                 if (now - last > 600) {
2196                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2197                                       obd->obd_name);
2198                         last = now;
2199                 }
2200         }
2201         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2202               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2203                 __u64 next_update_transno = 0;
2204
2205                 /*
2206                  * Only abort the recovery if there are no update recovery
2207                  * left in the queue
2208                  */
2209                 spin_lock(&obd->obd_recovery_task_lock);
2210                 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd) {
2211                         next_update_transno =
2212                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2213
2214                         tdtd = lut->lut_tdtd;
2215                         /*
2216                          * If next_update_transno == 0, it probably because
2217                          * updatelog retrieve threads did not get any records
2218                          * yet, let's wait those threads stopped
2219                          */
2220                         if (next_update_transno == 0) {
2221                                 spin_unlock(&obd->obd_recovery_task_lock);
2222
2223                                 while (wait_event_timeout(
2224                                         tdtd->tdtd_recovery_threads_waitq,
2225                                         check_update_llog(lut),
2226                                         cfs_time_seconds(60)) == 0);
2227
2228                                 spin_lock(&obd->obd_recovery_task_lock);
2229                                 next_update_transno =
2230                                         distribute_txn_get_next_transno(tdtd);
2231                         }
2232                 }
2233
2234                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2235                         obd->obd_next_recovery_transno = next_update_transno;
2236                         spin_unlock(&obd->obd_recovery_task_lock);
2237                         /*
2238                          * Disconnect unfinished exports from clients, and
2239                          * keep connection from MDT to make sure the update
2240                          * recovery will still keep trying until some one
2241                          * manually abort the recovery
2242                          */
2243                         class_disconnect_stale_exports(obd,
2244                                                 exp_finished_or_from_mdt);
2245                         /* Abort all of replay & replay lock req from clients */
2246                         abort_req_replay_queue(obd);
2247                         abort_lock_replay_queue(obd);
2248                         CDEBUG(D_HA,
2249                                "%s: there are still update replay (%#llx)in the queue.\n",
2250                                obd->obd_name, next_update_transno);
2251                 } else {
2252                         obd->obd_abort_recovery = 1;
2253                         spin_unlock(&obd->obd_recovery_task_lock);
2254                         CWARN("%s recovery is aborted by hard timeout\n",
2255                               obd->obd_name);
2256                 }
2257         }
2258
2259         while (wait_event_timeout(obd->obd_next_transno_waitq,
2260                                   check_routine(lut),
2261                                   cfs_time_seconds(60)) == 0)
2262                 ; /* wait indefinitely for event, but don't trigger watchdog */
2263
2264         if (obd->obd_abort_recovery) {
2265                 CWARN("recovery is aborted, evict exports in recovery\n");
2266                 if (lut->lut_tdtd != NULL) {
2267                         tdtd = lut->lut_tdtd;
2268                         /*
2269                          * Let's wait all of the update log recovery thread
2270                          * finished
2271                          */
2272                         wait_event_idle(
2273                                 tdtd->tdtd_recovery_threads_waitq,
2274                                 atomic_read(&tdtd->tdtd_recovery_threads_count)
2275                                 == 0);
2276                         /* Then abort the update recovery list */
2277                         dtrq_list_destroy(lut->lut_tdtd);
2278                 }
2279
2280                 /** evict exports which didn't finish recovery yet */
2281                 class_disconnect_stale_exports(obd, exp_finished);
2282                 return 1;
2283         } else if (obd->obd_recovery_expired) {
2284                 obd->obd_recovery_expired = 0;
2285
2286                 /** If some clients died being recovered, evict them */
2287                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2288                               obd->obd_name);
2289                 /** evict cexports with no replay in queue, they are stalled */
2290                 class_disconnect_stale_exports(obd, health_check);
2291
2292                 /** continue with VBR */
2293                 spin_lock(&obd->obd_dev_lock);
2294                 obd->obd_version_recov = 1;
2295                 spin_unlock(&obd->obd_dev_lock);
2296                 /**
2297                  * reset timer, recovery will proceed with versions now,
2298                  * timeout is set just to handle reconnection delays
2299                  */
2300                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2301                 /**
2302                  * Wait for recovery events again, after evicting bad clients
2303                  */
2304                 goto repeat;
2305         }
2306         return 0;
2307 }
2308
2309 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2310 {
2311         struct obd_device *obd = lut->lut_obd;
2312         struct ptlrpc_request *req = NULL;
2313
2314         CDEBUG(D_HA, "Waiting for lock\n");
2315         if (target_recovery_overseer(lut, check_for_next_lock,
2316                                      exp_lock_replay_healthy))
2317                 abort_lock_replay_queue(obd);
2318
2319         spin_lock(&obd->obd_recovery_task_lock);
2320         if (!list_empty(&obd->obd_lock_replay_queue)) {
2321                 req = list_entry(obd->obd_lock_replay_queue.next,
2322                                      struct ptlrpc_request, rq_list);
2323                 list_del_init(&req->rq_list);
2324                 spin_unlock(&obd->obd_recovery_task_lock);
2325         } else {
2326                 spin_unlock(&obd->obd_recovery_task_lock);
2327                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2328                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2329                 /** evict exports failed VBR */
2330                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2331         }
2332         return req;
2333 }
2334
2335 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2336 {
2337         struct ptlrpc_request *req = NULL;
2338
2339         spin_lock(&obd->obd_recovery_task_lock);
2340         if (!list_empty(&obd->obd_final_req_queue)) {
2341                 req = list_entry(obd->obd_final_req_queue.next,
2342                                      struct ptlrpc_request, rq_list);
2343                 list_del_init(&req->rq_list);
2344                 spin_unlock(&obd->obd_recovery_task_lock);
2345                 if (req->rq_export->exp_in_recovery) {
2346                         spin_lock(&req->rq_export->exp_lock);
2347                         req->rq_export->exp_in_recovery = 0;
2348                         spin_unlock(&req->rq_export->exp_lock);
2349                 }
2350         } else {
2351                 spin_unlock(&obd->obd_recovery_task_lock);
2352         }
2353         return req;
2354 }
2355
2356 static void handle_recovery_req(struct ptlrpc_thread *thread,
2357                                 struct ptlrpc_request *req,
2358                                 svc_handler_t handler)
2359 {
2360         ENTRY;
2361
2362         /**
2363          * export can be evicted during recovery, no need to handle replays for
2364          * it after that, discard such request silently
2365          */
2366         if (req->rq_export->exp_disconnected)
2367                 RETURN_EXIT;
2368
2369         req->rq_session.lc_thread = thread;
2370         req->rq_svc_thread = thread;
2371         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2372
2373         /* thread context */
2374         lu_context_enter(&thread->t_env->le_ctx);
2375         (void)handler(req);
2376         lu_context_exit(&thread->t_env->le_ctx);
2377
2378         req->rq_svc_thread->t_env->le_ses = NULL;
2379
2380         /* don't reset timer for final stage */
2381         if (!exp_finished(req->rq_export)) {
2382                 timeout_t timeout = obd_timeout;
2383
2384                 /**
2385                  * Add request @timeout to the recovery time so next request from
2386                  * this client may come in recovery time
2387                  */
2388                 if (!AT_OFF) {
2389                         struct ptlrpc_service_part *svcpt;
2390                         timeout_t est_timeout;
2391
2392                         svcpt = req->rq_rqbd->rqbd_svcpt;
2393                         /*
2394                          * If the server sent early reply for this request,
2395                          * the client will recalculate the timeout according to
2396                          * current server estimate service time, so we will
2397                          * use the maxium timeout here for waiting the client
2398                          * sending the next req
2399                          */
2400                         est_timeout = at_get(&svcpt->scp_at_estimate);
2401                         timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2402                                         lustre_msg_get_timeout(req->rq_reqmsg));
2403                         /*
2404                          * Add 2 net_latency, one for balance rq_deadline
2405                          * (see ptl_send_rpc), one for resend the req to server,
2406                          * Note: client will pack net_latency in replay req
2407                          * (see ptlrpc_replay_req)
2408                          */
2409                         timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2410                 }
2411                 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2412                                       true);
2413         }
2414         EXIT;
2415 }
2416
2417 /** Checking routines for recovery */
2418 static int check_for_recovery_ready(struct lu_target *lut)
2419 {
2420         struct obd_device *obd = lut->lut_obd;
2421         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2422
2423         CDEBUG(D_HA,
2424                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2425                clnts, obd->obd_stale_clients,
2426                atomic_read(&obd->obd_max_recoverable_clients),
2427                obd->obd_abort_recovery, obd->obd_recovery_expired);
2428
2429         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2430                 LASSERT(clnts <=
2431                         atomic_read(&obd->obd_max_recoverable_clients));
2432                 if (clnts + obd->obd_stale_clients <
2433                     atomic_read(&obd->obd_max_recoverable_clients))
2434                         return 0;
2435         }
2436
2437         if (!obd->obd_abort_recov_mdt && lut->lut_tdtd != NULL) {
2438                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2439                     !obd->obd_abort_recovery && !obd->obd_stopping) {
2440                         /*
2441                          * Let's extend recovery timer, in case the recovery
2442                          * timer expired, and some clients got evicted
2443                          */
2444                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2445                                               true);
2446                         CDEBUG(D_HA,
2447                                "%s update recovery is not ready, extend recovery %d\n",
2448                                obd->obd_name, obd->obd_recovery_timeout);
2449                         return 0;
2450                 }
2451         }
2452
2453         return 1;
2454 }
2455
2456 enum {
2457         REQUEST_RECOVERY = 1,
2458         UPDATE_RECOVERY = 2,
2459 };
2460
2461 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2462 {
2463         __u64 transno = 0;
2464
2465         if (!list_empty(&obd->obd_req_replay_queue)) {
2466                 struct ptlrpc_request *req;
2467
2468                 req = list_entry(obd->obd_req_replay_queue.next,
2469                                  struct ptlrpc_request, rq_list);
2470                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2471         }
2472
2473         return transno;
2474 }
2475
2476 static __u64 get_next_transno(struct lu_target *lut, int *type)
2477 {
2478         struct obd_device *obd = lut->lut_obd;
2479         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2480         __u64 transno = 0;
2481         __u64 update_transno;
2482
2483         ENTRY;
2484
2485         transno = get_next_replay_req_transno(obd);
2486         if (type != NULL)
2487                 *type = REQUEST_RECOVERY;
2488
2489         if (!tdtd || obd->obd_abort_recov_mdt)
2490                 RETURN(transno);
2491
2492         update_transno = distribute_txn_get_next_transno(tdtd);
2493         if (transno == 0 || (transno >= update_transno &&
2494                              update_transno != 0)) {
2495                 transno = update_transno;
2496                 if (type != NULL)
2497                         *type = UPDATE_RECOVERY;
2498         }
2499
2500         RETURN(transno);
2501 }
2502
2503 /**
2504  * drop duplicate replay request
2505  *
2506  * Because the operation has been replayed by update recovery, the request
2507  * with the same transno will be dropped and also notify the client to send
2508  * next replay request.
2509  *
2510  * \param[in] env       execution environment
2511  * \param[in] obd       failover obd device
2512  * \param[in] req       request to be dropped
2513  */
2514 static void drop_duplicate_replay_req(struct lu_env *env,
2515                                       struct obd_device *obd,
2516                                       struct ptlrpc_request *req)
2517 {
2518         DEBUG_REQ(D_HA, req,
2519                   "remove t%lld from %s because duplicate update records found",
2520                   lustre_msg_get_transno(req->rq_reqmsg),
2521                   libcfs_nid2str(req->rq_peer.nid));
2522
2523         /*
2524          * Right now, only for MDS reint operation update replay and
2525          * normal request replay can have the same transno
2526          */
2527         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2528                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2529                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2530                 if (likely(req->rq_export))
2531                         target_committed_to_req(req);
2532                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2533                 target_send_reply(req, req->rq_status, 0);
2534         } else {
2535                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2536                 libcfs_nid2str(req->rq_peer.nid));
2537         }
2538         target_exp_dequeue_req_replay(req);
2539         target_request_copy_put(req);
2540         obd->obd_replayed_requests++;
2541 }
2542
2543 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2544
2545 static void replay_request_or_update(struct lu_env *env,
2546                                      struct lu_target *lut,
2547                                      struct target_recovery_data *trd,
2548                                      struct ptlrpc_thread *thread)
2549 {
2550         struct obd_device *obd = lut->lut_obd;
2551         struct ptlrpc_request *req = NULL;
2552         int type;
2553         __u64 transno;
2554
2555         ENTRY;
2556
2557         CDEBUG(D_HA, "Waiting for transno %lld\n",
2558                obd->obd_next_recovery_transno);
2559
2560         /* Replay all of request and update by transno */
2561         do {
2562                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2563
2564                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2565
2566                 /**
2567                  * It is needed to extend recovery window above
2568                  *  recovery_time_soft. Extending is possible only in the
2569                  *  end of recovery window (see more details in
2570                  *  handle_recovery_req()).
2571                  */
2572                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2573
2574                 if (target_recovery_overseer(lut, check_for_next_transno,
2575                                         exp_req_replay_healthy_or_from_mdt)) {
2576                         abort_req_replay_queue(obd);
2577                         abort_lock_replay_queue(obd);
2578                         goto abort;
2579                 }
2580
2581                 spin_lock(&obd->obd_recovery_task_lock);
2582                 transno = get_next_transno(lut, &type);
2583                 if (type == REQUEST_RECOVERY && transno != 0) {
2584                         /*
2585                          * Drop replay request from client side, if the
2586                          * replay has been executed by update with the
2587                          * same transno
2588                          */
2589                         req = list_entry(obd->obd_req_replay_queue.next,
2590                                         struct ptlrpc_request, rq_list);
2591
2592                         list_del_init(&req->rq_list);
2593                         obd->obd_requests_queued_for_recovery--;
2594                         spin_unlock(&obd->obd_recovery_task_lock);
2595
2596                         /*
2597                          * Let's check if the request has been redone by
2598                          * update replay
2599                          */
2600                         if (is_req_replayed_by_update(req)) {
2601                                 struct distribute_txn_replay_req *dtrq;
2602
2603                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2604                                                                       transno);
2605                                 LASSERT(dtrq != NULL);
2606                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2607                                 list_del_init(&dtrq->dtrq_list);
2608                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2609                                 dtrq_destroy(dtrq);
2610
2611                                 drop_duplicate_replay_req(env, obd, req);
2612
2613                                 continue;
2614                         }
2615
2616                         LASSERT(trd->trd_processing_task == current->pid);
2617                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2618                                   req->rq_xid,
2619                                   lustre_msg_get_transno(req->rq_reqmsg),
2620                                   libcfs_nid2str(req->rq_peer.nid));
2621
2622                         ptlrpc_watchdog_init(&thread->t_watchdog,
2623                                              WATCHDOG_TIMEOUT);
2624                         handle_recovery_req(thread, req,
2625                                             trd->trd_recovery_handler);
2626                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2627
2628                         /**
2629                          * bz18031: increase next_recovery_transno before
2630                          * target_request_copy_put() will drop exp_rpc reference
2631                          */
2632                         spin_lock(&obd->obd_recovery_task_lock);
2633                         obd->obd_next_recovery_transno++;
2634                         spin_unlock(&obd->obd_recovery_task_lock);
2635                         target_exp_dequeue_req_replay(req);
2636                         target_request_copy_put(req);
2637                         obd->obd_replayed_requests++;
2638                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2639                         struct distribute_txn_replay_req *dtrq;
2640                         int rc;
2641
2642                         spin_unlock(&obd->obd_recovery_task_lock);
2643
2644                         LASSERT(tdtd != NULL);
2645                         dtrq = distribute_txn_get_next_req(tdtd);
2646                         lu_context_enter(&thread->t_env->le_ctx);
2647                         ptlrpc_watchdog_init(&thread->t_watchdog,
2648                                              WATCHDOG_TIMEOUT);
2649                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2650                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2651                         lu_context_exit(&thread->t_env->le_ctx);
2652                         extend_recovery_timer(obd, obd_timeout, true);
2653
2654                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2655                                 CDEBUG(D_HA,
2656                                        "Move x%llu t%llu to finish list\n",
2657                                        dtrq->dtrq_xid,
2658                                        dtrq->dtrq_master_transno);
2659
2660                                 /* Add it to the replay finish list */
2661                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2662                                 list_add(&dtrq->dtrq_list,
2663                                          &tdtd->tdtd_replay_finish_list);
2664                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2665
2666                                 spin_lock(&obd->obd_recovery_task_lock);
2667                                 if (transno == obd->obd_next_recovery_transno)
2668                                         obd->obd_next_recovery_transno++;
2669                                 else if (transno >
2670                                          obd->obd_next_recovery_transno)
2671                                         obd->obd_next_recovery_transno =
2672                                                                 transno + 1;
2673                                 spin_unlock(&obd->obd_recovery_task_lock);
2674                         } else {
2675                                 dtrq_destroy(dtrq);
2676                         }
2677                 } else {
2678                         spin_unlock(&obd->obd_recovery_task_lock);
2679 abort:
2680                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2681                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2682                         /** evict exports failed VBR */
2683                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2684                         break;
2685                 }
2686         } while (1);
2687 }
2688
2689 static int target_recovery_thread(void *arg)
2690 {
2691         struct lu_target *lut = arg;
2692         struct obd_device *obd = lut->lut_obd;
2693         struct ptlrpc_request *req;
2694         struct target_recovery_data *trd = &obd->obd_recovery_data;
2695         unsigned long delta;
2696         struct lu_env *env;
2697         struct ptlrpc_thread *thread = NULL;
2698         int rc = 0;
2699
2700         ENTRY;
2701         OBD_ALLOC_PTR(thread);
2702         if (thread == NULL)
2703                 RETURN(-ENOMEM);
2704
2705         OBD_ALLOC_PTR(env);
2706         if (env == NULL)
2707                 GOTO(out_thread, rc = -ENOMEM);
2708         rc = lu_env_add(env);
2709         if (rc)
2710                 GOTO(out_env, rc);
2711
2712         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2713         if (rc)
2714                 GOTO(out_env_remove, rc);
2715
2716         thread->t_env = env;
2717         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2718         thread->t_task = current;
2719         env->le_ctx.lc_thread = thread;
2720         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2721
2722         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2723                current->pid);
2724         trd->trd_processing_task = current->pid;
2725
2726         spin_lock(&obd->obd_dev_lock);
2727         obd->obd_recovering = 1;
2728         spin_unlock(&obd->obd_dev_lock);
2729         complete(&trd->trd_starting);
2730
2731         /* first of all, we have to know the first transno to replay */
2732         if (target_recovery_overseer(lut, check_for_recovery_ready,
2733                                      exp_connect_healthy)) {
2734                 abort_req_replay_queue(obd);
2735                 abort_lock_replay_queue(obd);
2736                 if (lut->lut_tdtd != NULL)
2737                         dtrq_list_destroy(lut->lut_tdtd);
2738         }
2739
2740         /* next stage: replay requests or update */
2741         delta = jiffies;
2742         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2743                atomic_read(&obd->obd_req_replay_clients),
2744                obd->obd_next_recovery_transno);
2745         replay_request_or_update(env, lut, trd, thread);
2746
2747         /**
2748          * The second stage: replay locks
2749          */
2750         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2751                atomic_read(&obd->obd_lock_replay_clients));
2752         while ((req = target_next_replay_lock(lut))) {
2753                 LASSERT(trd->trd_processing_task == current->pid);
2754                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2755                           libcfs_nid2str(req->rq_peer.nid));
2756                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2757                         req->rq_status = -ENODEV;
2758                         target_request_copy_put(req);
2759                         continue;
2760                 }
2761                 handle_recovery_req(thread, req,
2762                                     trd->trd_recovery_handler);
2763                 target_request_copy_put(req);
2764                 obd->obd_replayed_locks++;
2765         }
2766
2767         /**
2768          * The third stage: reply on final pings, at this moment all clients
2769          * must have request in final queue
2770          */
2771         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2772         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2773         /** Update server last boot epoch */
2774         tgt_boot_epoch_update(lut);
2775         /*
2776          * We drop recoverying flag to forward all new requests
2777          * to regular mds_handle() since now
2778          */
2779         spin_lock(&obd->obd_dev_lock);
2780         obd->obd_recovering = obd->obd_abort_recovery = 0;
2781         spin_unlock(&obd->obd_dev_lock);
2782         spin_lock(&obd->obd_recovery_task_lock);
2783         target_cancel_recovery_timer(obd);
2784         spin_unlock(&obd->obd_recovery_task_lock);
2785         while ((req = target_next_final_ping(obd))) {
2786                 LASSERT(trd->trd_processing_task == current->pid);
2787                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2788                           libcfs_nid2str(req->rq_peer.nid));
2789                 handle_recovery_req(thread, req,
2790                                     trd->trd_recovery_handler);
2791                 /*
2792                  * Because the waiting client can not send ping to server,
2793                  * so we need refresh the last_request_time, to avoid the
2794                  * export is being evicted
2795                  */
2796                 ptlrpc_update_export_timer(req->rq_export, 0);
2797                 target_request_copy_put(req);
2798         }
2799
2800         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2801         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2802                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2803         if (delta > OBD_RECOVERY_TIME_SOFT) {
2804                 CWARN("too long recovery - read logs\n");
2805                 libcfs_debug_dumplog();
2806         }
2807
2808         target_finish_recovery(lut);
2809         lu_context_fini(&env->le_ctx);
2810         trd->trd_processing_task = 0;
2811         complete_all(&trd->trd_finishing);
2812         tgt_io_thread_done(thread);
2813 out_env_remove:
2814         lu_env_remove(env);
2815 out_env:
2816         OBD_FREE_PTR(env);
2817 out_thread:
2818         OBD_FREE_PTR(thread);
2819         RETURN(rc);
2820 }
2821
2822 static int target_start_recovery_thread(struct lu_target *lut,
2823                                         svc_handler_t handler)
2824 {
2825         struct obd_device *obd = lut->lut_obd;
2826         int rc = 0;
2827         struct target_recovery_data *trd = &obd->obd_recovery_data;
2828         int index;
2829
2830         memset(trd, 0, sizeof(*trd));
2831         init_completion(&trd->trd_starting);
2832         init_completion(&trd->trd_finishing);
2833         trd->trd_recovery_handler = handler;
2834
2835         rc = server_name2index(obd->obd_name, &index, NULL);
2836         if (rc < 0)
2837                 return rc;
2838
2839         if (!IS_ERR(kthread_run(target_recovery_thread,
2840                                 lut, "tgt_recover_%d", index))) {
2841                 wait_for_completion(&trd->trd_starting);
2842                 LASSERT(obd->obd_recovering != 0);
2843         } else {
2844                 rc = -ECHILD;
2845         }
2846
2847         return rc;
2848 }
2849
2850 void target_stop_recovery_thread(struct obd_device *obd)
2851 {
2852         if (obd->obd_recovery_data.trd_processing_task > 0) {
2853                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2854                 /** recovery can be done but postrecovery is not yet */
2855                 spin_lock(&obd->obd_dev_lock);
2856                 if (obd->obd_recovering) {
2857                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2858                         obd->obd_abort_recovery = 1;
2859                         wake_up(&obd->obd_next_transno_waitq);
2860                 }
2861                 spin_unlock(&obd->obd_dev_lock);
2862                 wait_for_completion(&trd->trd_finishing);
2863         }
2864 }
2865 EXPORT_SYMBOL(target_stop_recovery_thread);
2866
2867 void target_recovery_fini(struct obd_device *obd)
2868 {
2869         class_disconnect_exports(obd);
2870         target_stop_recovery_thread(obd);
2871         target_cleanup_recovery(obd);
2872 }
2873 EXPORT_SYMBOL(target_recovery_fini);
2874
2875 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2876 {
2877         struct obd_device *obd = container_of(timer, struct obd_device,
2878                                               obd_recovery_timer);
2879
2880         CDEBUG(D_HA,
2881                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2882                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2883                ktime_get_seconds() - obd->obd_recovery_start,
2884                atomic_read(&obd->obd_connected_clients));
2885
2886         obd->obd_recovery_expired = 1;
2887         wake_up(&obd->obd_next_transno_waitq);
2888         return HRTIMER_NORESTART;
2889 }
2890
2891 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2892 {
2893         struct obd_device *obd = lut->lut_obd;
2894
2895         if (lut->lut_bottom->dd_rdonly)
2896                 return;
2897
2898         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2899                 /** Update server last boot epoch */
2900                 tgt_boot_epoch_update(lut);
2901                 return;
2902         }
2903
2904         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2905                "last_transno %llu\n", obd->obd_name,
2906                atomic_read(&obd->obd_max_recoverable_clients),
2907                obd->obd_last_committed);
2908         LASSERT(obd->obd_stopping == 0);
2909         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2910         obd->obd_recovery_start = 0;
2911         obd->obd_recovery_end = 0;
2912
2913         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2914                      HRTIMER_MODE_ABS);
2915         obd->obd_recovery_timer.function = &target_recovery_expired;
2916         target_start_recovery_thread(lut, handler);
2917 }
2918 EXPORT_SYMBOL(target_recovery_init);
2919
2920 static int target_process_req_flags(struct obd_device *obd,
2921                                     struct ptlrpc_request *req)
2922 {
2923         struct obd_export *exp = req->rq_export;
2924
2925         LASSERT(exp != NULL);
2926         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2927                 /* client declares he's ready to replay locks */
2928                 spin_lock(&exp->exp_lock);
2929                 if (exp->exp_req_replay_needed) {
2930                         exp->exp_req_replay_needed = 0;
2931                         spin_unlock(&exp->exp_lock);
2932
2933                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2934                         atomic_dec(&obd->obd_req_replay_clients);
2935                 } else {
2936                         spin_unlock(&exp->exp_lock);
2937                 }
2938         }
2939         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2940                 /*
2941                  * client declares he's ready to complete recovery
2942                  * so, we put the request on th final queue
2943                  */
2944                 spin_lock(&exp->exp_lock);
2945                 if (exp->exp_lock_replay_needed) {
2946                         exp->exp_lock_replay_needed = 0;
2947                         spin_unlock(&exp->exp_lock);
2948
2949                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2950                         atomic_dec(&obd->obd_lock_replay_clients);
2951                 } else {
2952                         spin_unlock(&exp->exp_lock);
2953                 }
2954         }
2955         return 0;
2956 }
2957
2958 int target_queue_recovery_request(struct ptlrpc_request *req,
2959                                   struct obd_device *obd)
2960 {
2961         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2962         struct ptlrpc_request *reqiter;
2963         int inserted = 0;
2964
2965         ENTRY;
2966
2967         if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2968                 /* Processing the queue right now, don't re-add. */
2969                 RETURN(1);
2970         }
2971
2972         target_process_req_flags(obd, req);
2973
2974         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2975                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2976                         if (cfs_fail_val == 1) {
2977                                 cfs_race_state = 1;
2978                                 cfs_fail_val = 0;
2979                                 wake_up(&cfs_race_waitq);
2980
2981                                 schedule_timeout_interruptible(
2982                                         cfs_time_seconds(1));
2983                         }
2984                 }
2985
2986                 /*
2987                  * client declares he's ready to complete recovery
2988                  * so, we put the request on th final queue
2989                  */
2990                 target_request_copy_get(req);
2991                 DEBUG_REQ(D_HA, req, "queue final req");
2992                 wake_up(&obd->obd_next_transno_waitq);
2993                 spin_lock(&obd->obd_recovery_task_lock);
2994                 if (obd->obd_recovering) {
2995                         struct ptlrpc_request *tmp;
2996                         struct ptlrpc_request *duplicate = NULL;
2997
2998                         if (likely(!req->rq_export->exp_replay_done)) {
2999                                 req->rq_export->exp_replay_done = 1;
3000                                 list_add_tail(&req->rq_list,
3001                                               &obd->obd_final_req_queue);
3002                                 spin_unlock(&obd->obd_recovery_task_lock);
3003                                 RETURN(0);
3004                         }
3005
3006                         /*
3007                          * XXX O(n), but only happens if final ping is
3008                          * timed out, probably reorganize the list as
3009                          * a hash list later
3010                          */
3011                         list_for_each_entry_safe(reqiter, tmp,
3012                                                  &obd->obd_final_req_queue,
3013                                                  rq_list) {
3014                                 if (reqiter->rq_export == req->rq_export) {
3015                                         list_del_init(&reqiter->rq_list);
3016                                         duplicate = reqiter;
3017                                         break;
3018                                 }
3019                         }
3020
3021                         list_add_tail(&req->rq_list,
3022                                       &obd->obd_final_req_queue);
3023                         req->rq_export->exp_replay_done = 1;
3024                         spin_unlock(&obd->obd_recovery_task_lock);
3025
3026                         if (duplicate != NULL) {
3027                                 DEBUG_REQ(D_HA, duplicate,
3028                                           "put prev final req");
3029                                 target_request_copy_put(duplicate);
3030                         }
3031                         RETURN(0);
3032                 } else {
3033                         spin_unlock(&obd->obd_recovery_task_lock);
3034                         target_request_copy_put(req);
3035                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3036                 }
3037         }
3038         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3039                 /* client declares he's ready to replay locks */
3040                 target_request_copy_get(req);
3041                 DEBUG_REQ(D_HA, req, "queue lock replay req");
3042                 wake_up(&obd->obd_next_transno_waitq);
3043                 spin_lock(&obd->obd_recovery_task_lock);
3044                 LASSERT(obd->obd_recovering);
3045                 /* usually due to recovery abort */
3046                 if (!req->rq_export->exp_in_recovery) {
3047                         spin_unlock(&obd->obd_recovery_task_lock);
3048                         target_request_copy_put(req);
3049                         RETURN(-ENOTCONN);
3050                 }
3051                 LASSERT(req->rq_export->exp_lock_replay_needed);
3052                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3053                 spin_unlock(&obd->obd_recovery_task_lock);
3054                 RETURN(0);
3055         }
3056
3057         /*
3058          * CAVEAT EMPTOR: The incoming request message has been swabbed
3059          * (i.e. buflens etc are in my own byte order), but type-dependent
3060          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3061          */
3062
3063         if (!transno) {
3064                 INIT_LIST_HEAD(&req->rq_list);
3065                 DEBUG_REQ(D_HA, req, "not queueing");
3066                 RETURN(1);
3067         }
3068
3069         /*
3070          * If we're processing the queue, we want don't want to queue this
3071          * message.
3072          *
3073          * Also, if this request has a transno less than the one we're waiting
3074          * for, we should process it now.  It could (and currently always will)
3075          * be an open request for a descriptor that was opened some time ago.
3076          *
3077          * Also, a resent, replayed request that has already been
3078          * handled will pass through here and be processed immediately.
3079          */
3080         CDEBUG(D_HA,
3081                "Next recovery transno: %llu, current: %llu, replaying\n",
3082                obd->obd_next_recovery_transno, transno);
3083
3084         /*
3085          * If the request has been replayed by update replay, then sends this
3086          * request to the recovery thread (replay_request_or_update()), where
3087          * it will be handled
3088          */
3089         spin_lock(&obd->obd_recovery_task_lock);
3090         if (transno < obd->obd_next_recovery_transno &&
3091             !is_req_replayed_by_update(req)) {
3092                 /* Processing the queue right now, don't re-add. */
3093                 LASSERT(list_empty(&req->rq_list));
3094                 spin_unlock(&obd->obd_recovery_task_lock);
3095                 RETURN(1);
3096         }
3097         spin_unlock(&obd->obd_recovery_task_lock);
3098
3099         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3100                 RETURN(0);
3101
3102         target_request_copy_get(req);
3103         if (!req->rq_export->exp_in_recovery) {
3104                 target_request_copy_put(req);
3105                 RETURN(-ENOTCONN);
3106         }
3107         LASSERT(req->rq_export->exp_req_replay_needed);
3108
3109         if (target_exp_enqueue_req_replay(req)) {
3110                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3111                 target_request_copy_put(req);
3112                 RETURN(0);
3113         }
3114
3115         /* XXX O(n^2) */
3116         spin_lock(&obd->obd_recovery_task_lock);
3117         LASSERT(obd->obd_recovering);
3118         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3119                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3120                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3121                         inserted = 1;
3122                         goto added;
3123                 }
3124
3125                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3126                              transno)) {
3127                         DEBUG_REQ(D_ERROR, req,
3128                                   "dropping replay: transno has been claimed by another client");
3129                         spin_unlock(&obd->obd_recovery_task_lock);
3130                         target_exp_dequeue_req_replay(req);
3131                         target_request_copy_put(req);
3132                         RETURN(0);
3133                 }
3134         }
3135 added:
3136         if (!inserted)
3137                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3138
3139         obd->obd_requests_queued_for_recovery++;
3140         spin_unlock(&obd->obd_recovery_task_lock);
3141         wake_up(&obd->obd_next_transno_waitq);
3142         RETURN(0);
3143 }
3144
3145 void target_committed_to_req(struct ptlrpc_request *req)
3146 {
3147         struct obd_export *exp = req->rq_export;
3148
3149         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3150                 lustre_msg_set_last_committed(req->rq_repmsg,
3151                                               exp->exp_last_committed);
3152         else
3153                 DEBUG_REQ(D_IOCTL, req,
3154                           "not sending last_committed update (%d/%d)",
3155                           exp->exp_obd->obd_no_transno,
3156                           req->rq_repmsg == NULL);
3157
3158         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3159                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3160 }
3161
3162 #endif /* HAVE_SERVER_SUPPORT */
3163
3164 /**
3165  * Packs current SLV and Limit into \a req.
3166  */
3167 int target_pack_pool_reply(struct ptlrpc_request *req)
3168 {
3169         struct obd_device *obd;
3170
3171         ENTRY;
3172