Whamcloud - gitweb
59bc60873d7a54d38d97ac1d1786245abd1929a7
[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, 2016, 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 <linux/kthread.h>
43 #include <libcfs/libcfs.h>
44 #include <obd.h>
45 #include <obd_class.h>
46 #include <lustre_dlm.h>
47 #include <lustre_net.h>
48 #include <lustre_sec.h>
49 #include "ldlm_internal.h"
50
51 /* @priority: If non-zero, move the selected connection to the list head.
52  * @create: If zero, only search in existing connections.
53  */
54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
55                            int priority, int create)
56 {
57         struct ptlrpc_connection *ptlrpc_conn;
58         struct obd_import_conn *imp_conn = NULL, *item;
59         lnet_nid_t nid4refnet = LNET_NID_ANY;
60         int rc = 0;
61         ENTRY;
62
63         if (!create && !priority) {
64                 CDEBUG(D_HA, "Nothing to do\n");
65                 RETURN(-EINVAL);
66         }
67
68         if (imp->imp_connection &&
69             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
70                 /* nid4refnet is used to restrict network connections */
71                 nid4refnet = imp->imp_connection->c_self;
72         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
73         if (!ptlrpc_conn) {
74                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
75                 RETURN(-ENOENT);
76         }
77
78         if (create) {
79                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
80                 if (!imp_conn)
81                         GOTO(out_put, rc = -ENOMEM);
82         }
83
84         spin_lock(&imp->imp_lock);
85         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
86                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87                         if (priority) {
88                                 list_del(&item->oic_item);
89                                 list_add(&item->oic_item,
90                                          &imp->imp_conn_list);
91                                 item->oic_last_attempt = 0;
92                         }
93                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
94                                imp, imp->imp_obd->obd_name, uuid->uuid,
95                                (priority ? ", moved to head" : ""));
96                         spin_unlock(&imp->imp_lock);
97                         GOTO(out_free, rc = 0);
98                 }
99         }
100         /* No existing import connection found for \a uuid. */
101         if (create) {
102                 imp_conn->oic_conn = ptlrpc_conn;
103                 imp_conn->oic_uuid = *uuid;
104                 imp_conn->oic_last_attempt = 0;
105                 if (priority)
106                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107                 else
108                         list_add_tail(&imp_conn->oic_item,
109                                       &imp->imp_conn_list);
110                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
111                        imp, imp->imp_obd->obd_name, uuid->uuid,
112                        (priority ? "head" : "tail"));
113         } else {
114                 spin_unlock(&imp->imp_lock);
115                 GOTO(out_free, rc = -ENOENT);
116         }
117
118         spin_unlock(&imp->imp_lock);
119         RETURN(0);
120 out_free:
121         if (imp_conn)
122                 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 out_put:
124         ptlrpc_connection_put(ptlrpc_conn);
125         RETURN(rc);
126 }
127
128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 {
130         return import_set_conn(imp, uuid, 1, 0);
131 }
132
133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
134                            int priority)
135 {
136         return import_set_conn(imp, uuid, priority, 1);
137 }
138 EXPORT_SYMBOL(client_import_add_conn);
139
140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
141 {
142         struct obd_import_conn *imp_conn;
143         struct obd_export *dlmexp;
144         int rc = -ENOENT;
145         ENTRY;
146
147         spin_lock(&imp->imp_lock);
148         if (list_empty(&imp->imp_conn_list)) {
149                 LASSERT(!imp->imp_connection);
150                 GOTO(out, rc);
151         }
152
153         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
154                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
155                         continue;
156                 LASSERT(imp_conn->oic_conn);
157
158                 if (imp_conn == imp->imp_conn_current) {
159                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
160
161                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
162                             imp->imp_state != LUSTRE_IMP_DISCON) {
163                                 CERROR("can't remove current connection\n");
164                                 GOTO(out, rc = -EBUSY);
165                         }
166
167                         ptlrpc_connection_put(imp->imp_connection);
168                         imp->imp_connection = NULL;
169
170                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
171                         if (dlmexp && dlmexp->exp_connection) {
172                                 LASSERT(dlmexp->exp_connection ==
173                                         imp_conn->oic_conn);
174                                 ptlrpc_connection_put(dlmexp->exp_connection);
175                                 dlmexp->exp_connection = NULL;
176                         }
177
178                         if (dlmexp != NULL)
179                                 class_export_put(dlmexp);
180                 }
181
182                 list_del(&imp_conn->oic_item);
183                 ptlrpc_connection_put(imp_conn->oic_conn);
184                 OBD_FREE(imp_conn, sizeof(*imp_conn));
185                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
186                        imp, imp->imp_obd->obd_name, uuid->uuid);
187                 rc = 0;
188                 break;
189         }
190 out:
191         spin_unlock(&imp->imp_lock);
192         if (rc == -ENOENT)
193                 CERROR("connection %s not found\n", uuid->uuid);
194         RETURN(rc);
195 }
196 EXPORT_SYMBOL(client_import_del_conn);
197
198 /**
199  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
200  * to find a conn uuid of \a imp which can reach \a peer.
201  */
202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
203                             struct obd_uuid *uuid)
204 {
205         struct obd_import_conn *conn;
206         int rc = -ENOENT;
207         ENTRY;
208
209         spin_lock(&imp->imp_lock);
210         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
211                 /* Check if conn UUID does have this peer NID. */
212                 if (class_check_uuid(&conn->oic_uuid, peer)) {
213                         *uuid = conn->oic_uuid;
214                         rc = 0;
215                         break;
216                 }
217         }
218         spin_unlock(&imp->imp_lock);
219         RETURN(rc);
220 }
221 EXPORT_SYMBOL(client_import_find_conn);
222
223 void client_destroy_import(struct obd_import *imp)
224 {
225         /* Drop security policy instance after all RPCs have finished/aborted
226          * to let all busy contexts be released. */
227         class_import_get(imp);
228         class_destroy_import(imp);
229         sptlrpc_import_sec_put(imp);
230         class_import_put(imp);
231 }
232 EXPORT_SYMBOL(client_destroy_import);
233
234 /**
235  * Check whether or not the OSC is on MDT.
236  * In the config log,
237  * osc on MDT
238  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
239  * osc on client
240  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
241  *
242  **/
243 static int osc_on_mdt(char *obdname)
244 {
245         char *ptr;
246
247         ptr = strrchr(obdname, '-');
248         if (ptr == NULL)
249                 return 0;
250
251         if (strncmp(ptr + 1, "MDT", 3) == 0)
252                 return 1;
253
254         return 0;
255 }
256
257 /* Configure an RPC client OBD device.
258  *
259  * lcfg parameters:
260  * 1 - client UUID
261  * 2 - server UUID
262  * 3 - inactive-on-startup
263  * 4 - restrictive net
264  */
265 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
266 {
267         struct client_obd *cli = &obddev->u.cli;
268         struct obd_import *imp;
269         struct obd_uuid server_uuid;
270         int rq_portal, rp_portal, connect_op;
271         char *name = obddev->obd_type->typ_name;
272         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
273         char *cli_name = lustre_cfg_buf(lcfg, 0);
274         struct ptlrpc_connection fake_conn = { .c_self = 0,
275                                                .c_remote_uuid.uuid[0] = 0 };
276         int rc;
277         ENTRY;
278
279         /* In a more perfect world, we would hang a ptlrpc_client off of
280          * obd_type and just use the values from there. */
281         if (!strcmp(name, LUSTRE_OSC_NAME)) {
282                 rq_portal = OST_REQUEST_PORTAL;
283                 rp_portal = OSC_REPLY_PORTAL;
284                 connect_op = OST_CONNECT;
285                 cli->cl_sp_me = LUSTRE_SP_CLI;
286                 cli->cl_sp_to = LUSTRE_SP_OST;
287                 ns_type = LDLM_NS_TYPE_OSC;
288         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
289                    !strcmp(name, LUSTRE_LWP_NAME)) {
290                 rq_portal = MDS_REQUEST_PORTAL;
291                 rp_portal = MDC_REPLY_PORTAL;
292                 connect_op = MDS_CONNECT;
293                 if (is_lwp_on_ost(cli_name))
294                         cli->cl_sp_me = LUSTRE_SP_OST;
295                 else if (is_lwp_on_mdt(cli_name))
296                         cli->cl_sp_me = LUSTRE_SP_MDT;
297                 else
298                         cli->cl_sp_me = LUSTRE_SP_CLI;
299                 cli->cl_sp_to = LUSTRE_SP_MDT;
300                 ns_type = LDLM_NS_TYPE_MDC;
301         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
302                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
303                         /* OSP_on_MDT for other MDTs */
304                         connect_op = MDS_CONNECT;
305                         cli->cl_sp_to = LUSTRE_SP_MDT;
306                         ns_type = LDLM_NS_TYPE_MDC;
307                         rq_portal = OUT_PORTAL;
308                 } else {
309                         /* OSP on MDT for OST */
310                         connect_op = OST_CONNECT;
311                         cli->cl_sp_to = LUSTRE_SP_OST;
312                         ns_type = LDLM_NS_TYPE_OSC;
313                         rq_portal = OST_REQUEST_PORTAL;
314                 }
315                 rp_portal = OSC_REPLY_PORTAL;
316                 cli->cl_sp_me = LUSTRE_SP_MDT;
317         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
318                 rq_portal = MGS_REQUEST_PORTAL;
319                 rp_portal = MGC_REPLY_PORTAL;
320                 connect_op = MGS_CONNECT;
321                 cli->cl_sp_me = LUSTRE_SP_MGC;
322                 cli->cl_sp_to = LUSTRE_SP_MGS;
323                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
324                 ns_type = LDLM_NS_TYPE_MGC;
325         } else {
326                 CERROR("unknown client OBD type \"%s\", can't setup\n",
327                        name);
328                 RETURN(-EINVAL);
329         }
330
331         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
332                 CERROR("requires a TARGET UUID\n");
333                 RETURN(-EINVAL);
334         }
335
336         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
337                 CERROR("client UUID must be less than 38 characters\n");
338                 RETURN(-EINVAL);
339         }
340
341         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
342                 CERROR("setup requires a SERVER UUID\n");
343                 RETURN(-EINVAL);
344         }
345
346         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
347                 CERROR("target UUID must be less than 38 characters\n");
348                 RETURN(-EINVAL);
349         }
350
351         init_rwsem(&cli->cl_sem);
352         mutex_init(&cli->cl_mgc_mutex);
353         cli->cl_seq = NULL;
354         init_rwsem(&cli->cl_seq_rwsem);
355         cli->cl_conn_count = 0;
356         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
357                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
358                      sizeof(server_uuid)));
359
360         cli->cl_dirty_pages = 0;
361         cli->cl_avail_grant = 0;
362         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
363         /* cl_dirty_max_pages may be changed at connect time in
364          * ptlrpc_connect_interpret(). */
365         client_adjust_max_dirty(cli);
366         INIT_LIST_HEAD(&cli->cl_cache_waiters);
367         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
368         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
369         INIT_LIST_HEAD(&cli->cl_loi_write_list);
370         INIT_LIST_HEAD(&cli->cl_loi_read_list);
371         spin_lock_init(&cli->cl_loi_list_lock);
372         atomic_set(&cli->cl_pending_w_pages, 0);
373         atomic_set(&cli->cl_pending_r_pages, 0);
374         cli->cl_r_in_flight = 0;
375         cli->cl_w_in_flight = 0;
376
377         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
378         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
379         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
380         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
381         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
382         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
383
384         /* lru for osc. */
385         INIT_LIST_HEAD(&cli->cl_lru_osc);
386         atomic_set(&cli->cl_lru_shrinkers, 0);
387         atomic_long_set(&cli->cl_lru_busy, 0);
388         atomic_long_set(&cli->cl_lru_in_list, 0);
389         INIT_LIST_HEAD(&cli->cl_lru_list);
390         spin_lock_init(&cli->cl_lru_list_lock);
391         atomic_long_set(&cli->cl_unstable_count, 0);
392         INIT_LIST_HEAD(&cli->cl_shrink_list);
393
394         init_waitqueue_head(&cli->cl_destroy_waitq);
395         atomic_set(&cli->cl_destroy_in_flight, 0);
396 #ifdef ENABLE_CHECKSUM
397         /* Turn on checksumming by default. */
398         cli->cl_checksum = 1;
399         /*
400          * The supported checksum types will be worked out at connect time
401          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
402          * through procfs.
403          */
404         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
405 #endif
406         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
407
408         /* Set it to possible maximum size. It may be reduced by ocd_brw_size
409          * from OFD after connecting. */
410         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
411
412         /* set cl_chunkbits default value to PAGE_SHIFT,
413          * it will be updated at OSC connection time. */
414         cli->cl_chunkbits = PAGE_SHIFT;
415
416         if (!strcmp(name, LUSTRE_MDC_NAME)) {
417                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
418         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
419                 cli->cl_max_rpcs_in_flight = 2;
420         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
421                 cli->cl_max_rpcs_in_flight = 3;
422         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
423                 cli->cl_max_rpcs_in_flight = 4;
424         } else {
425                 if (osc_on_mdt(obddev->obd_name))
426                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
427                 else
428                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
429         }
430
431         spin_lock_init(&cli->cl_mod_rpcs_lock);
432         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
433         cli->cl_max_mod_rpcs_in_flight = 0;
434         cli->cl_mod_rpcs_in_flight = 0;
435         cli->cl_close_rpcs_in_flight = 0;
436         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
437         cli->cl_mod_tag_bitmap = NULL;
438
439         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
440
441         if (connect_op == MDS_CONNECT) {
442                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
443                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
444                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
445                 if (cli->cl_mod_tag_bitmap == NULL)
446                         GOTO(err, rc = -ENOMEM);
447         }
448
449         rc = ldlm_get_ref();
450         if (rc) {
451                 CERROR("ldlm_get_ref failed: %d\n", rc);
452                 GOTO(err, rc);
453         }
454
455         ptlrpc_init_client(rq_portal, rp_portal, name,
456                            &obddev->obd_ldlm_client);
457
458         imp = class_new_import(obddev);
459         if (imp == NULL)
460                 GOTO(err_ldlm, rc = -ENOENT);
461         imp->imp_client = &obddev->obd_ldlm_client;
462         imp->imp_connect_op = connect_op;
463         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
464                LUSTRE_CFG_BUFLEN(lcfg, 1));
465         class_import_put(imp);
466
467         if (lustre_cfg_buf(lcfg, 4)) {
468                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
469
470                 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
471                         rc = -EINVAL;
472                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
473                                obddev->obd_name, lustre_cfg_string(lcfg, 4),
474                                rc);
475                         GOTO(err_import, rc);
476                 }
477                 fake_conn.c_self = LNET_MKNID(refnet, 0);
478                 imp->imp_connection = &fake_conn;
479         }
480
481         rc = client_import_add_conn(imp, &server_uuid, 1);
482         if (rc) {
483                 CERROR("can't add initial connection\n");
484                 GOTO(err_import, rc);
485         }
486         imp->imp_connection = NULL;
487
488         cli->cl_import = imp;
489         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
490         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
491
492         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
493                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
494                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
495                                name, obddev->obd_name,
496                                cli->cl_target_uuid.uuid);
497                         spin_lock(&imp->imp_lock);
498                         imp->imp_deactive = 1;
499                         spin_unlock(&imp->imp_lock);
500                 }
501         }
502
503         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
504                                                    LDLM_NAMESPACE_CLIENT,
505                                                    LDLM_NAMESPACE_GREEDY,
506                                                    ns_type);
507         if (obddev->obd_namespace == NULL) {
508                 CERROR("Unable to create client namespace - %s\n",
509                        obddev->obd_name);
510                 GOTO(err_import, rc = -ENOMEM);
511         }
512
513         RETURN(rc);
514
515 err_import:
516         class_destroy_import(imp);
517 err_ldlm:
518         ldlm_put_ref();
519 err:
520         if (cli->cl_mod_tag_bitmap != NULL)
521                 OBD_FREE(cli->cl_mod_tag_bitmap,
522                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
523         cli->cl_mod_tag_bitmap = NULL;
524         RETURN(rc);
525
526 }
527 EXPORT_SYMBOL(client_obd_setup);
528
529 int client_obd_cleanup(struct obd_device *obddev)
530 {
531         struct client_obd *cli = &obddev->u.cli;
532         ENTRY;
533
534         ldlm_namespace_free_post(obddev->obd_namespace);
535         obddev->obd_namespace = NULL;
536
537         obd_cleanup_client_import(obddev);
538         LASSERT(obddev->u.cli.cl_import == NULL);
539
540         ldlm_put_ref();
541
542         if (cli->cl_mod_tag_bitmap != NULL)
543                 OBD_FREE(cli->cl_mod_tag_bitmap,
544                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
545         cli->cl_mod_tag_bitmap = NULL;
546
547         RETURN(0);
548 }
549 EXPORT_SYMBOL(client_obd_cleanup);
550
551 /* ->o_connect() method for client side (OSC and MDC and MGC) */
552 int client_connect_import(const struct lu_env *env,
553                           struct obd_export **exp,
554                           struct obd_device *obd, struct obd_uuid *cluuid,
555                           struct obd_connect_data *data, void *localdata)
556 {
557         struct client_obd       *cli    = &obd->u.cli;
558         struct obd_import       *imp    = cli->cl_import;
559         struct obd_connect_data *ocd;
560         struct lustre_handle    conn    = { 0 };
561         int                     rc;
562         ENTRY;
563
564         *exp = NULL;
565         down_write(&cli->cl_sem);
566         if (cli->cl_conn_count > 0)
567                 GOTO(out_sem, rc = -EALREADY);
568
569         rc = class_connect(&conn, obd, cluuid);
570         if (rc)
571                 GOTO(out_sem, rc);
572
573         cli->cl_conn_count++;
574         *exp = class_conn2export(&conn);
575
576         LASSERT(obd->obd_namespace);
577
578         imp->imp_dlm_handle = conn;
579         rc = ptlrpc_init_import(imp);
580         if (rc != 0)
581                 GOTO(out_ldlm, rc);
582
583         ocd = &imp->imp_connect_data;
584         if (data) {
585                 *ocd = *data;
586                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
587                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
588         }
589
590         rc = ptlrpc_connect_import(imp);
591         if (rc != 0) {
592                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
593                 GOTO(out_ldlm, rc);
594         }
595         LASSERT(*exp != NULL && (*exp)->exp_connection);
596
597         if (data) {
598                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
599                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
600                          data->ocd_connect_flags, ocd->ocd_connect_flags);
601                 data->ocd_connect_flags = ocd->ocd_connect_flags;
602                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
603         }
604
605         ptlrpc_pinger_add_import(imp);
606
607         EXIT;
608
609         if (rc) {
610 out_ldlm:
611                 cli->cl_conn_count--;
612                 class_disconnect(*exp);
613                 *exp = NULL;
614         }
615 out_sem:
616         up_write(&cli->cl_sem);
617
618         return rc;
619 }
620 EXPORT_SYMBOL(client_connect_import);
621
622 int client_disconnect_export(struct obd_export *exp)
623 {
624         struct obd_device *obd = class_exp2obd(exp);
625         struct client_obd *cli;
626         struct obd_import *imp;
627         int rc = 0, err;
628         ENTRY;
629
630         if (!obd) {
631                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
632                        exp, exp ? exp->exp_handle.h_cookie : -1);
633                 RETURN(-EINVAL);
634         }
635
636         cli = &obd->u.cli;
637         imp = cli->cl_import;
638
639         down_write(&cli->cl_sem);
640         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
641                 cli->cl_conn_count);
642
643         if (cli->cl_conn_count == 0) {
644                 CERROR("disconnecting disconnected device (%s)\n",
645                        obd->obd_name);
646                 GOTO(out_disconnect, rc = -EINVAL);
647         }
648
649         cli->cl_conn_count--;
650         if (cli->cl_conn_count != 0)
651                 GOTO(out_disconnect, rc = 0);
652
653         /* Mark import deactivated now, so we don't try to reconnect if any
654          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
655          * fully deactivate the import, or that would drop all requests. */
656         spin_lock(&imp->imp_lock);
657         imp->imp_deactive = 1;
658         spin_unlock(&imp->imp_lock);
659
660         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
661          * delete it regardless.  (It's safe to delete an import that was
662          * never added.) */
663         (void)ptlrpc_pinger_del_import(imp);
664
665         if (obd->obd_namespace != NULL) {
666                 /* obd_force == local only */
667                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
668                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
669                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
670         }
671
672         /* There's no need to hold sem while disconnecting an import,
673          * and it may actually cause deadlock in GSS. */
674         up_write(&cli->cl_sem);
675         rc = ptlrpc_disconnect_import(imp, 0);
676         down_write(&cli->cl_sem);
677
678         ptlrpc_invalidate_import(imp);
679
680         EXIT;
681
682 out_disconnect:
683         /* Use server style - class_disconnect should be always called for
684          * o_disconnect. */
685         err = class_disconnect(exp);
686         if (!rc && err)
687                 rc = err;
688
689         up_write(&cli->cl_sem);
690
691         RETURN(rc);
692 }
693 EXPORT_SYMBOL(client_disconnect_export);
694
695 #ifdef HAVE_SERVER_SUPPORT
696 int server_disconnect_export(struct obd_export *exp)
697 {
698         int rc;
699         ENTRY;
700
701         /* Disconnect early so that clients can't keep using export. */
702         rc = class_disconnect(exp);
703         /* Close import to avoid sending any requests. */
704         if (exp->exp_imp_reverse)
705                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
706
707         ldlm_bl_thread_wakeup();
708
709         /* complete all outstanding replies */
710         spin_lock(&exp->exp_lock);
711         while (!list_empty(&exp->exp_outstanding_replies)) {
712                 struct ptlrpc_reply_state *rs =
713                         list_entry(exp->exp_outstanding_replies.next,
714                                        struct ptlrpc_reply_state, rs_exp_list);
715                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
716
717                 spin_lock(&svcpt->scp_rep_lock);
718
719                 list_del_init(&rs->rs_exp_list);
720
721                 spin_lock(&rs->rs_lock);
722                 /* clear rs_convert_lock to make sure rs is handled and put */
723                 rs->rs_convert_lock = 0;
724                 ptlrpc_schedule_difficult_reply(rs);
725                 spin_unlock(&rs->rs_lock);
726
727                 spin_unlock(&svcpt->scp_rep_lock);
728         }
729         spin_unlock(&exp->exp_lock);
730
731         RETURN(rc);
732 }
733 EXPORT_SYMBOL(server_disconnect_export);
734
735 /* --------------------------------------------------------------------------
736  * from old lib/target.c
737  * -------------------------------------------------------------------------- */
738
739 static int target_handle_reconnect(struct lustre_handle *conn,
740                                    struct obd_export *exp,
741                                    struct obd_uuid *cluuid)
742 {
743         struct obd_device *target;
744         struct lustre_handle *hdl;
745         cfs_time_t now;
746         cfs_time_t deadline;
747         int timeout;
748         int rc = 0;
749         ENTRY;
750
751         hdl = &exp->exp_imp_reverse->imp_remote_handle;
752         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
753                 conn->cookie = exp->exp_handle.h_cookie;
754                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
755                        " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
756                 RETURN(0);
757         }
758
759         target = exp->exp_obd;
760
761         /* Might be a re-connect after a partition. */
762         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
763                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
764                               "with handle %#llx. Rejecting client "
765                               "with the same UUID trying to reconnect "
766                               "with handle %#llx\n", target->obd_name,
767                               obd_uuid2str(&exp->exp_client_uuid),
768                               obd_export_nid2str(exp),
769                               hdl->cookie, conn->cookie);
770                 memset(conn, 0, sizeof *conn);
771                 /* target_handle_connect() treats EALREADY and
772                  * -EALREADY differently.  -EALREADY is an error
773                  * (same UUID, different handle). */
774                 RETURN(-EALREADY);
775         }
776
777         if (!target->obd_recovering) {
778                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
779                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
780                         obd_export_nid2str(exp));
781                 GOTO(out_already, rc);
782         }
783
784         now = cfs_time_current();
785         deadline = target->obd_recovery_timer.expires;
786         if (cfs_time_before(now, deadline)) {
787                 struct target_distribute_txn_data *tdtd =
788                                         class_exp2tgt(exp)->lut_tdtd;
789                 int size = 0;
790                 int count = 0;
791                 char *buf = NULL;
792
793                 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
794                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
795                         buf = tdtd->tdtd_show_update_logs_retrievers(
796                                 tdtd->tdtd_show_retrievers_cbdata,
797                                 &size, &count);
798
799                 if (count > 0)
800                         LCONSOLE_WARN("%s: Recovery already passed deadline "
801                                       "%d:%.02d. It is due to DNE recovery "
802                                       "failed/stuck on the %d MDT(s):%s. "
803                                       "Please wait until all MDTs recovered "
804                                       "or abort the recovery by force.\n",
805                                       target->obd_name, timeout / 60,
806                                       timeout % 60, count,
807                                       buf ? buf : "unknown (not enough RAM)");
808                 else
809                         LCONSOLE_WARN("%s: Recovery already passed deadline "
810                                       "%d:%.02d. If you do not want to wait "
811                                       "more, please abort the recovery by "
812                                       "force.\n", target->obd_name,
813                                       timeout / 60, timeout % 60);
814
815                 if (buf != NULL)
816                         OBD_FREE(buf, size);
817         } else {
818                 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
819                 LCONSOLE_WARN("%s: Recovery already passed deadline"
820                         " %d:%.02d, It is most likely due to DNE"
821                         " recovery is failed or stuck, please wait a"
822                         " few more minutes or abort the recovery.\n",
823                         target->obd_name, timeout / 60, timeout % 60);
824         }
825
826 out_already:
827         conn->cookie = exp->exp_handle.h_cookie;
828         /* target_handle_connect() treats EALREADY and
829          * -EALREADY differently.  EALREADY means we are
830          * doing a valid reconnect from the same client. */
831         RETURN(EALREADY);
832 }
833
834 static void
835 check_and_start_recovery_timer(struct obd_device *obd,
836                                struct ptlrpc_request *req, int new_client);
837
838 /**
839  * update flags for import during reconnect process
840  */
841 static int rev_import_flags_update(struct obd_import *revimp,
842                                    struct ptlrpc_request *req)
843 {
844         int rc;
845         struct obd_connect_data *data;
846
847         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
848
849         if (data->ocd_connect_flags & OBD_CONNECT_AT)
850                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
851         else
852                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
853
854         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
855
856         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
857         if (rc) {
858                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
859                         revimp->imp_client->cli_name,
860                         libcfs_id2str(req->rq_peer), rc);
861                 return rc;
862         }
863
864         return 0;
865 }
866
867 /**
868  * Allocate a new reverse import for an export.
869  *
870  * \retval -errno in case error hit
871  * \retval 0 if reverse import correctly init
872  **/
873 int rev_import_init(struct obd_export *export)
874 {
875         struct obd_device *obd = export->exp_obd;
876         struct obd_import *revimp;
877
878         LASSERT(export->exp_imp_reverse == NULL);
879
880         revimp = class_new_import(obd);
881         if (revimp == NULL)
882                 return -ENOMEM;
883
884         revimp->imp_remote_handle.cookie = 0ULL;
885         revimp->imp_client = &obd->obd_ldlm_client;
886         revimp->imp_dlm_fake = 1;
887
888         /* it is safe to connect import in new state as no sends possible */
889         spin_lock(&export->exp_lock);
890         export->exp_imp_reverse = revimp;
891         spin_unlock(&export->exp_lock);
892         class_import_put(revimp);
893
894         return 0;
895 }
896 EXPORT_SYMBOL(rev_import_init);
897
898 /**
899  * Handle reconnect for an export.
900  *
901  * \param exp export to handle reconnect process
902  * \param req client reconnect request
903  *
904  * \retval -rc in case securitfy flavor can't be changed
905  * \retval 0 in case none problems
906  */
907 static int rev_import_reconnect(struct obd_export *exp,
908                                 struct ptlrpc_request *req)
909 {
910         struct obd_import *revimp = exp->exp_imp_reverse;
911         struct lustre_handle *lh;
912         int rc;
913
914         /* avoid sending a request until import flags are changed */
915         ptlrpc_import_enter_resend(revimp);
916
917         if (revimp->imp_connection != NULL)
918                 ptlrpc_connection_put(revimp->imp_connection);
919
920         /*
921          * client from recovery don't have a handle so we need to take from
922          * request. it may produce situation when wrong client connected
923          * to recovery as we trust a client uuid
924          */
925         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
926         revimp->imp_remote_handle = *lh;
927
928         /* unknown versions will be caught in
929          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
930         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
931
932         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
933
934         rc = rev_import_flags_update(revimp, req);
935         if (rc != 0) {
936                 /* it is safe to still be in RECOVERY phase as we are not able
937                  * to setup correct security flavor so requests are not able to
938                  * be delivered correctly */
939                 return rc;
940         }
941
942         /* resend all rpc's via new connection */
943         return ptlrpc_import_recovery_state_machine(revimp);
944 }
945
946 int target_handle_connect(struct ptlrpc_request *req)
947 {
948         struct obd_device *target = NULL;
949         struct obd_export *export = NULL;
950         /* connect handle - filled from target_handle_reconnect in
951          * reconnect case */
952         struct lustre_handle conn;
953         struct lustre_handle *tmp;
954         struct obd_uuid cluuid;
955         char *str;
956         int rc = 0;
957         char *target_start;
958         int target_len;
959         bool     mds_conn = false, lw_client = false, initial_conn = false;
960         bool     mds_mds_conn = false;
961         bool     new_mds_mds_conn = false;
962         struct obd_connect_data *data, *tmpdata;
963         int size, tmpsize;
964         lnet_nid_t *client_nid = NULL;
965         ENTRY;
966
967         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
968
969         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
970         if (str == NULL) {
971                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
972                 GOTO(out, rc = -EINVAL);
973         }
974
975         target = class_dev_by_str(str);
976         if (!target) {
977                 deuuidify(str, NULL, &target_start, &target_len);
978                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
979                                    "from %s (no target). If you are running "
980                                    "an HA pair check that the target is "
981                                    "mounted on the other server.\n", str,
982                                    libcfs_nid2str(req->rq_peer.nid));
983                 GOTO(out, rc = -ENODEV);
984         }
985
986         spin_lock(&target->obd_dev_lock);
987
988         target->obd_conn_inprogress++;
989
990         if (target->obd_stopping || !target->obd_set_up) {
991                 spin_unlock(&target->obd_dev_lock);
992
993                 deuuidify(str, NULL, &target_start, &target_len);
994                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
995                               target_len, target_start,
996                               libcfs_nid2str(req->rq_peer.nid),
997                               (target->obd_stopping ?
998                                "stopping" : "not set up"));
999                 GOTO(out, rc = -ENODEV);
1000         }
1001
1002         if (target->obd_no_conn) {
1003                 spin_unlock(&target->obd_dev_lock);
1004
1005                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1006                                "from %s\n", target->obd_name,
1007                                libcfs_nid2str(req->rq_peer.nid));
1008                 GOTO(out, rc = -EAGAIN);
1009         }
1010
1011         spin_unlock(&target->obd_dev_lock);
1012
1013         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1014         if (str == NULL) {
1015                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1016                 GOTO(out, rc = -EINVAL);
1017         }
1018
1019         obd_str2uuid(&cluuid, str);
1020
1021         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1022         if (tmp == NULL)
1023                 GOTO(out, rc = -EPROTO);
1024
1025         conn = *tmp;
1026
1027         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1028                                     RCL_CLIENT);
1029         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1030         if (!data)
1031                 GOTO(out, rc = -EPROTO);
1032
1033         rc = req_capsule_server_pack(&req->rq_pill);
1034         if (rc)
1035                 GOTO(out, rc);
1036
1037 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1038         /* Don't allow clients to connect that are using old 1.8 format
1039          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1040          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1041          * FULL20 flag should be set on all connections since 2.0, but no
1042          * longer affects behaviour.
1043          *
1044          * Later this check will be disabled and the flag can be retired
1045          * completely once interop with 3.0 is no longer needed.
1046          */
1047         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1048                 GOTO(out, rc = -EPROTO);
1049
1050         /* Don't allow liblustre clients to connect.
1051          * - testing was disabled in v2_2_50_0-61-g6a75d65
1052          * - building was disabled in v2_5_58_0-28-g7277179
1053          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1054          * - clients were refused connect for version difference > 0.0.1.32  */
1055         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1056                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1057                 GOTO(out, rc = -EPROTO);
1058         }
1059 #endif
1060
1061         /* Note: lw_client is needed in MDS-MDS failover during update log
1062          * processing, so we needs to allow lw_client to be connected at
1063          * anytime, instead of only the initial connection */
1064         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1065
1066         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1067                 initial_conn = true;
1068                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1069                 mds_mds_conn = (data->ocd_connect_flags &
1070                                 OBD_CONNECT_MDS_MDS) != 0;
1071
1072                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1073                  * for Imperative Recovery connection from MGC to MGS.
1074                  *
1075                  * Via check OBD_CONNECT_FID, we can distinguish whether
1076                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1077                  * MGC or MDT. */
1078                 if (!lw_client &&
1079                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1080                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1081                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1082                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1083                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1084                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1085
1086                         /* We do not support the MDT-MDT interoperations with
1087                          * different version MDT because of protocol changes. */
1088                         if (unlikely(major != LUSTRE_MAJOR ||
1089                                      minor != LUSTRE_MINOR ||
1090                                      abs(patch - LUSTRE_PATCH) > 3)) {
1091                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1092                                         "connection from different version MDT "
1093                                         "(%d.%d.%d.%d) %s %s\n",
1094                                         target->obd_name, LUSTRE_MAJOR,
1095                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1096                                         major, minor, patch,
1097                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1098                                         libcfs_nid2str(req->rq_peer.nid), str);
1099
1100                                 GOTO(out, rc = -EPROTO);
1101                         }
1102                 }
1103         }
1104
1105         /* lctl gets a backstage, all-access pass. */
1106         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1107                 goto dont_check_exports;
1108
1109         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1110         if (!export)
1111                 goto no_export;
1112
1113         /* We've found an export in the hash. */
1114
1115         spin_lock(&export->exp_lock);
1116
1117         if (export->exp_connecting) { /* bug 9635, et. al. */
1118                 spin_unlock(&export->exp_lock);
1119                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1120                               export->exp_obd->obd_name, export,
1121                               libcfs_nid2str(req->rq_peer.nid));
1122                 class_export_put(export);
1123                 export = NULL;
1124                 rc = -EALREADY;
1125         } else if ((mds_conn || (lw_client && initial_conn) ||
1126                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1127                    export->exp_connection != NULL) {
1128                 spin_unlock(&export->exp_lock);
1129                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1130                         /* MDS or LWP reconnected after failover. */
1131                         LCONSOLE_WARN("%s: Received %s connection from "
1132                             "%s, removing former export from %s\n",
1133                             target->obd_name, mds_conn ? "MDS" : "LWP",
1134                             libcfs_nid2str(req->rq_peer.nid),
1135                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1136                 } else {
1137                         /* New MDS connection from the same NID. */
1138                         LCONSOLE_WARN("%s: Received new %s connection from "
1139                                 "%s, removing former export from same NID\n",
1140                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1141                                 libcfs_nid2str(req->rq_peer.nid));
1142                 }
1143
1144                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1145                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1146                         /* Because exports between MDTs will always be
1147                          * kept, let's do not fail such export if they
1148                          * come from the same NID, otherwise it might
1149                          * cause eviction between MDTs, which might
1150                          * cause namespace inconsistency */
1151                         spin_lock(&export->exp_lock);
1152                         export->exp_connecting = 1;
1153                         spin_unlock(&export->exp_lock);
1154                         conn.cookie = export->exp_handle.h_cookie;
1155                         rc = EALREADY;
1156                 } else {
1157                         class_fail_export(export);
1158                         class_export_put(export);
1159                         export = NULL;
1160                         rc = 0;
1161                 }
1162         } else if (export->exp_connection != NULL && initial_conn &&
1163                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1164                 spin_unlock(&export->exp_lock);
1165                 /* In MDS failover we have static UUID but NID can change. */
1166                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1167                               "existing nid %s is already connected\n",
1168                               target->obd_name, cluuid.uuid,
1169                               libcfs_nid2str(req->rq_peer.nid),
1170                               libcfs_nid2str(
1171                                       export->exp_connection->c_peer.nid));
1172                 rc = -EALREADY;
1173                 class_export_put(export);
1174                 export = NULL;
1175         } else {
1176                 export->exp_connecting = 1;
1177                 spin_unlock(&export->exp_lock);
1178                 LASSERT(export->exp_obd == target);
1179
1180                 rc = target_handle_reconnect(&conn, export, &cluuid);
1181         }
1182
1183         /* If we found an export, we already unlocked. */
1184         if (!export) {
1185 no_export:
1186                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1187         } else if (req->rq_export == NULL &&
1188                    atomic_read(&export->exp_rpc_count) > 0) {
1189                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1190                               "still busy with %d references\n",
1191                               target->obd_name, cluuid.uuid,
1192                               libcfs_nid2str(req->rq_peer.nid),
1193                               atomic_read(&export->exp_refcount));
1194                 GOTO(out, rc = -EBUSY);
1195         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1196                 if (!strstr(cluuid.uuid, "mdt"))
1197                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1198                                       "known client %s (at %s) because it "
1199                                       "is indicating it is a new client",
1200                                       target->obd_name, cluuid.uuid,
1201                                       libcfs_nid2str(req->rq_peer.nid));
1202                 GOTO(out, rc = -EALREADY);
1203         } else {
1204                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1205         }
1206
1207         if (rc < 0) {
1208                 GOTO(out, rc);
1209         }
1210
1211         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1212                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1213                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1214                export, ktime_get_real_seconds(),
1215                export ? export->exp_last_request_time : 0);
1216
1217         /* If this is the first time a client connects, reset the recovery
1218          * timer. Discard lightweight connections which might be local. */
1219         if (!lw_client && rc == 0 && target->obd_recovering)
1220                 check_and_start_recovery_timer(target, req, export == NULL);
1221
1222         /* We want to handle EALREADY but *not* -EALREADY from
1223          * target_handle_reconnect(), return reconnection state in a flag. */
1224         if (rc == EALREADY) {
1225                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1226                 rc = 0;
1227         } else {
1228                 LASSERT(rc == 0);
1229         }
1230
1231         /* Tell the client if we support replayable requests. */
1232         if (target->obd_replayable)
1233                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1234         client_nid = &req->rq_peer.nid;
1235
1236         if (export == NULL) {
1237                 /* allow lightweight connections during recovery */
1238                 /* allow "new" MDT to be connected during recovery, since we
1239                  * need retrieve recovery update records from it */
1240                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1241                         cfs_time_t t;
1242                         int     c; /* connected */
1243                         int     i; /* in progress */
1244                         int     k; /* known */
1245                         int     s; /* stale/evicted */
1246
1247                         c = atomic_read(&target->obd_connected_clients);
1248                         i = atomic_read(&target->obd_lock_replay_clients);
1249                         k = target->obd_max_recoverable_clients;
1250                         s = target->obd_stale_clients;
1251                         t = target->obd_recovery_timer.expires;
1252                         t = cfs_time_sub(t, cfs_time_current());
1253                         t = cfs_duration_sec(t);
1254                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1255                                       "(at %s), waiting for %d known clients "
1256                                       "(%d recovered, %d in progress, and %d "
1257                                       "evicted) to recover in %d:%.02d\n",
1258                                       target->obd_name, cluuid.uuid,
1259                                       libcfs_nid2str(req->rq_peer.nid), k,
1260                                       c - i, i, s, (int)t / 60,
1261                                       (int)t % 60);
1262                         rc = -EBUSY;
1263                 } else {
1264 dont_check_exports:
1265                         rc = obd_connect(req->rq_svc_thread->t_env,
1266                                          &export, target, &cluuid, data,
1267                                          client_nid);
1268                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1269                                 lustre_msg_add_op_flags(req->rq_repmsg,
1270                                                         MSG_CONNECT_RECOVERING);
1271                         if (rc == 0) {
1272                                 conn.cookie = export->exp_handle.h_cookie;
1273                                 rc = rev_import_init(export);
1274                         }
1275
1276                         if (mds_mds_conn)
1277                                 new_mds_mds_conn = true;
1278                 }
1279         } else {
1280                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1281                                    export, target, &cluuid, data, client_nid);
1282         }
1283         if (rc)
1284                 GOTO(out, rc);
1285
1286         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1287         data->ocd_instance = target->u.obt.obt_instance;
1288
1289         /* Return only the parts of obd_connect_data that we understand, so the
1290          * client knows that we don't understand the rest. */
1291         if (data) {
1292                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1293                                                RCL_SERVER);
1294                 tmpdata = req_capsule_server_get(&req->rq_pill,
1295                                                  &RMF_CONNECT_DATA);
1296                 /* Don't use struct assignment here, because the client reply
1297                  * buffer may be smaller/larger than the local struct
1298                  * obd_connect_data. */
1299                 memcpy(tmpdata, data, min(tmpsize, size));
1300         }
1301
1302         /* If the client and the server are the same node, we will already
1303          * have an export that really points to the client's DLM export,
1304          * because we have a shared handles table.
1305          *
1306          * XXX this will go away when shaver stops sending the "connect" handle
1307          * in the real "remote handle" field of the request --phik 24 Apr 2003
1308          */
1309         ptlrpc_request_change_export(req, export);
1310
1311         spin_lock(&export->exp_lock);
1312         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1313                 spin_unlock(&export->exp_lock);
1314                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1315                        "or equal conn_cnt: %d >= %d\n",
1316                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1317                        export->exp_conn_cnt,
1318                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1319
1320                 GOTO(out, rc = -EALREADY);
1321         }
1322         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1323         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1324         spin_unlock(&export->exp_lock);
1325
1326         if (export->exp_connection != NULL) {
1327                 /* Check to see if connection came from another NID. */
1328                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1329                     !hlist_unhashed(&export->exp_nid_hash))
1330                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1331                                      &export->exp_connection->c_peer.nid,
1332                                      &export->exp_nid_hash);
1333
1334                 ptlrpc_connection_put(export->exp_connection);
1335         }
1336
1337         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1338                                                        req->rq_self,
1339                                                        &cluuid);
1340         if (hlist_unhashed(&export->exp_nid_hash))
1341                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1342                              &export->exp_connection->c_peer.nid,
1343                              &export->exp_nid_hash);
1344
1345         lustre_msg_set_handle(req->rq_repmsg, &conn);
1346
1347         rc = rev_import_reconnect(export, req);
1348         if (rc != 0)
1349                 GOTO(out, rc);
1350
1351         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1352                 int has_transno;
1353                 __u64 transno = data->ocd_transno;
1354
1355                 spin_lock(&export->exp_lock);
1356                 /* possible race with class_disconnect_stale_exports,
1357                  * export may be already in the eviction process */
1358                 if (export->exp_failed) {
1359                         spin_unlock(&export->exp_lock);
1360                         GOTO(out, rc = -ENODEV);
1361                 }
1362                 export->exp_in_recovery = 1;
1363                 export->exp_req_replay_needed = 1;
1364                 export->exp_lock_replay_needed = 1;
1365                 spin_unlock(&export->exp_lock);
1366
1367                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1368                                  MSG_CONNECT_TRANSNO);
1369                 if (has_transno && transno == 0)
1370                         CWARN("Connect with zero transno!\n");
1371
1372                 if (has_transno && transno > 0 &&
1373                     transno < target->obd_next_recovery_transno &&
1374                     transno > target->obd_last_committed) {
1375                         /* Another way is to use cmpxchg() to be lock-free. */
1376                         spin_lock(&target->obd_recovery_task_lock);
1377                         if (transno < target->obd_next_recovery_transno)
1378                                 target->obd_next_recovery_transno = transno;
1379                         spin_unlock(&target->obd_recovery_task_lock);
1380                 }
1381
1382                 atomic_inc(&target->obd_req_replay_clients);
1383                 atomic_inc(&target->obd_lock_replay_clients);
1384                 /* Note: MDS-MDS connection is allowed to be connected during
1385                  * recovery, no matter if the exports needs to be recoveried.
1386                  * Because we need retrieve updates logs from all other MDTs.
1387                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1388                  * also needs to be increased to match other recovery checking
1389                  * condition. */
1390                 if (new_mds_mds_conn)
1391                         target->obd_max_recoverable_clients++;
1392                 if (atomic_inc_return(&target->obd_connected_clients) ==
1393                     target->obd_max_recoverable_clients)
1394                         wake_up(&target->obd_next_transno_waitq);
1395         }
1396
1397         /* Tell the client we're in recovery, when client is involved in it. */
1398         if (target->obd_recovering && !lw_client)
1399                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1400
1401 out:
1402         if (export) {
1403                 spin_lock(&export->exp_lock);
1404                 export->exp_connecting = 0;
1405                 spin_unlock(&export->exp_lock);
1406
1407                 class_export_put(export);
1408         }
1409         if (target != NULL) {
1410                 spin_lock(&target->obd_dev_lock);
1411                 target->obd_conn_inprogress--;
1412                 spin_unlock(&target->obd_dev_lock);
1413                 class_decref(target, "find", current);
1414         }
1415         req->rq_status = rc;
1416         RETURN(rc);
1417 }
1418
1419 int target_handle_disconnect(struct ptlrpc_request *req)
1420 {
1421         int rc;
1422         ENTRY;
1423
1424         rc = req_capsule_server_pack(&req->rq_pill);
1425         if (rc)
1426                 RETURN(rc);
1427
1428         /* Keep the rq_export around so we can send the reply. */
1429         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1430
1431         RETURN(0);
1432 }
1433
1434 void target_destroy_export(struct obd_export *exp)
1435 {
1436         struct obd_import       *imp = NULL;
1437         /* exports created from last_rcvd data, and "fake"
1438            exports created by lctl don't have an import */
1439         spin_lock(&exp->exp_lock);
1440         if (exp->exp_imp_reverse != NULL) {
1441                 imp = exp->exp_imp_reverse;
1442                 exp->exp_imp_reverse = NULL;
1443         }
1444         spin_unlock(&exp->exp_lock);
1445         if (imp != NULL)
1446                 client_destroy_import(imp);
1447
1448         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1449         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1450         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1451         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1452 }
1453 EXPORT_SYMBOL(target_destroy_export);
1454
1455 /*
1456  * Recovery functions
1457  */
1458 static void target_request_copy_get(struct ptlrpc_request *req)
1459 {
1460         class_export_rpc_inc(req->rq_export);
1461         LASSERT(list_empty(&req->rq_list));
1462         INIT_LIST_HEAD(&req->rq_replay_list);
1463
1464         /* Increase refcount to keep request in queue. */
1465         atomic_inc(&req->rq_refcount);
1466         /* Let export know it has replays to be handled. */
1467         atomic_inc(&req->rq_export->exp_replay_count);
1468 }
1469
1470 static void target_request_copy_put(struct ptlrpc_request *req)
1471 {
1472         LASSERT(list_empty(&req->rq_replay_list));
1473         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1474
1475         atomic_dec(&req->rq_export->exp_replay_count);
1476         class_export_rpc_dec(req->rq_export);
1477         ptlrpc_server_drop_request(req);
1478 }
1479
1480 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1481 {
1482         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1483         struct obd_export     *exp = req->rq_export;
1484         struct ptlrpc_request *reqiter;
1485         struct ptlrpc_request *dup_req = NULL;
1486         int                    dup = 0;
1487
1488         LASSERT(exp);
1489
1490         spin_lock(&exp->exp_lock);
1491         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1492                                 rq_replay_list) {
1493                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1494                         dup_req = reqiter;
1495                         dup = 1;
1496                         break;
1497                 }
1498         }
1499
1500         if (dup) {
1501                 /* We expect it with RESENT and REPLAY flags. */
1502                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1503                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1504                         CERROR("invalid flags %x of resent replay\n",
1505                                lustre_msg_get_flags(req->rq_reqmsg));
1506
1507                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1508                         __u32 new_conn;
1509
1510                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1511                         if (new_conn >
1512                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1513                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1514                                                         new_conn);
1515                 }
1516         } else {
1517                 list_add_tail(&req->rq_replay_list,
1518                                   &exp->exp_req_replay_queue);
1519         }
1520
1521         spin_unlock(&exp->exp_lock);
1522         return dup;
1523 }
1524
1525 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1526 {
1527         LASSERT(!list_empty(&req->rq_replay_list));
1528         LASSERT(req->rq_export);
1529
1530         spin_lock(&req->rq_export->exp_lock);
1531         list_del_init(&req->rq_replay_list);
1532         spin_unlock(&req->rq_export->exp_lock);
1533 }
1534
1535 static void target_finish_recovery(struct lu_target *lut)
1536 {
1537         struct obd_device *obd = lut->lut_obd;
1538         ENTRY;
1539
1540         /* Only log a recovery message when recovery has occurred. */
1541         if (obd->obd_recovery_start) {
1542                 time64_t now = ktime_get_real_seconds();
1543                 time64_t elapsed_time;
1544
1545                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1546                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1547                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1548                         (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1549                         obd->obd_max_recoverable_clients,
1550                         atomic_read(&obd->obd_connected_clients),
1551                         obd->obd_stale_clients,
1552                         obd->obd_stale_clients == 1 ? "was" : "were");
1553         }
1554
1555         ldlm_reprocess_recovery_done(obd->obd_namespace);
1556         spin_lock(&obd->obd_recovery_task_lock);
1557         if (!list_empty(&obd->obd_req_replay_queue) ||
1558             !list_empty(&obd->obd_lock_replay_queue) ||
1559             !list_empty(&obd->obd_final_req_queue)) {
1560                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1561                        obd->obd_name,
1562                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1563                        list_empty(&obd->obd_lock_replay_queue) ? \
1564                                "" : "lock ",
1565                        list_empty(&obd->obd_final_req_queue) ? \
1566                                "" : "final ");
1567                 spin_unlock(&obd->obd_recovery_task_lock);
1568                 LBUG();
1569         }
1570         spin_unlock(&obd->obd_recovery_task_lock);
1571
1572         obd->obd_recovery_end = ktime_get_real_seconds();
1573
1574         /* When recovery finished, cleanup orphans on MDS and OST. */
1575         if (OBT(obd) && OBP(obd, postrecov)) {
1576                 int rc = OBP(obd, postrecov)(obd);
1577                 if (rc < 0)
1578                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1579                                       obd->obd_name, rc);
1580         }
1581         EXIT;
1582 }
1583
1584 static void abort_req_replay_queue(struct obd_device *obd)
1585 {
1586         struct ptlrpc_request *req, *n;
1587         struct list_head abort_list;
1588
1589         INIT_LIST_HEAD(&abort_list);
1590         spin_lock(&obd->obd_recovery_task_lock);
1591         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1592         spin_unlock(&obd->obd_recovery_task_lock);
1593         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1594                 DEBUG_REQ(D_WARNING, req, "aborted:");
1595                 req->rq_status = -ENOTCONN;
1596                 if (ptlrpc_error(req)) {
1597                         DEBUG_REQ(D_ERROR, req,
1598                                   "failed abort_req_reply; skipping");
1599                 }
1600                 target_exp_dequeue_req_replay(req);
1601                 target_request_copy_put(req);
1602         }
1603 }
1604
1605 static void abort_lock_replay_queue(struct obd_device *obd)
1606 {
1607         struct ptlrpc_request *req, *n;
1608         struct list_head abort_list;
1609
1610         INIT_LIST_HEAD(&abort_list);
1611         spin_lock(&obd->obd_recovery_task_lock);
1612         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1613         spin_unlock(&obd->obd_recovery_task_lock);
1614         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1615                 DEBUG_REQ(D_ERROR, req, "aborted:");
1616                 req->rq_status = -ENOTCONN;
1617                 if (ptlrpc_error(req)) {
1618                         DEBUG_REQ(D_ERROR, req,
1619                                   "failed abort_lock_reply; skipping");
1620                 }
1621                 target_request_copy_put(req);
1622         }
1623 }
1624
1625 /* Called from a cleanup function if the device is being cleaned up
1626    forcefully.  The exports should all have been disconnected already,
1627    the only thing left to do is
1628      - clear the recovery flags
1629      - cancel the timer
1630      - free queued requests and replies, but don't send replies
1631    Because the obd_stopping flag is set, no new requests should be received.
1632
1633 */
1634 void target_cleanup_recovery(struct obd_device *obd)
1635 {
1636         struct ptlrpc_request *req, *n;
1637         struct list_head clean_list;
1638
1639         INIT_LIST_HEAD(&clean_list);
1640         spin_lock(&obd->obd_dev_lock);
1641         if (!obd->obd_recovering) {
1642                 spin_unlock(&obd->obd_dev_lock);
1643                 EXIT;
1644                 return;
1645         }
1646         obd->obd_recovering = obd->obd_abort_recovery = 0;
1647         spin_unlock(&obd->obd_dev_lock);
1648
1649         spin_lock(&obd->obd_recovery_task_lock);
1650         target_cancel_recovery_timer(obd);
1651         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1652         spin_unlock(&obd->obd_recovery_task_lock);
1653
1654         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1655                 LASSERT(req->rq_reply_state == NULL);
1656                 target_exp_dequeue_req_replay(req);
1657                 target_request_copy_put(req);
1658         }
1659
1660         spin_lock(&obd->obd_recovery_task_lock);
1661         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1662         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1663         spin_unlock(&obd->obd_recovery_task_lock);
1664
1665         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1666                 LASSERT(req->rq_reply_state == NULL);
1667                 target_request_copy_put(req);
1668         }
1669
1670         EXIT;
1671 }
1672 EXPORT_SYMBOL(target_cleanup_recovery);
1673
1674 /* obd_recovery_task_lock should be held */
1675 void target_cancel_recovery_timer(struct obd_device *obd)
1676 {
1677         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1678         del_timer(&obd->obd_recovery_timer);
1679 }
1680
1681 static void target_start_recovery_timer(struct obd_device *obd)
1682 {
1683         if (obd->obd_recovery_start != 0)
1684                 return;
1685
1686         spin_lock(&obd->obd_dev_lock);
1687         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1688                 spin_unlock(&obd->obd_dev_lock);
1689                 return;
1690         }
1691
1692         LASSERT(obd->obd_recovery_timeout != 0);
1693
1694         if (obd->obd_recovery_start != 0) {
1695                 spin_unlock(&obd->obd_dev_lock);
1696                 return;
1697         }
1698
1699         mod_timer(&obd->obd_recovery_timer,
1700                   cfs_time_shift(obd->obd_recovery_timeout));
1701         obd->obd_recovery_start = ktime_get_real_seconds();
1702         spin_unlock(&obd->obd_dev_lock);
1703
1704         LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1705                       obd->obd_name,
1706                       obd->obd_recovery_timeout / 60,
1707                       obd->obd_recovery_timeout % 60,
1708                       obd->obd_max_recoverable_clients,
1709                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1710                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1711 }
1712
1713 /**
1714  * extend recovery window.
1715  *
1716  * if @extend is true, extend recovery window to have @drt remaining at least;
1717  * otherwise, make sure the recovery timeout value is not less than @drt.
1718  */
1719 static void extend_recovery_timer(struct obd_device *obd, int drt,
1720                                   bool extend)
1721 {
1722         time64_t now;
1723         time64_t end;
1724         time64_t left;
1725         time64_t to;
1726
1727         spin_lock(&obd->obd_dev_lock);
1728         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1729                 spin_unlock(&obd->obd_dev_lock);
1730                 return;
1731         }
1732         LASSERT(obd->obd_recovery_start != 0);
1733
1734         now = ktime_get_real_seconds();
1735         to = obd->obd_recovery_timeout;
1736         end = obd->obd_recovery_start + to;
1737         left = end - now;
1738
1739         if (extend && (drt > left)) {
1740                 to += drt - left;
1741         } else if (!extend && (drt > to)) {
1742                 to = drt;
1743         }
1744
1745         if (to > obd->obd_recovery_time_hard) {
1746                 to = obd->obd_recovery_time_hard;
1747                 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1748                       obd->obd_name, to, extend);
1749         }
1750
1751         if (obd->obd_recovery_timeout < to) {
1752                 obd->obd_recovery_timeout = to;
1753                 end = obd->obd_recovery_start + to;
1754                 mod_timer(&obd->obd_recovery_timer,
1755                           cfs_time_shift(end - now));
1756         }
1757         spin_unlock(&obd->obd_dev_lock);
1758
1759         CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1760                 obd->obd_name, (s64)(end - now));
1761 }
1762
1763 /* Reset the timer with each new client connection */
1764 /*
1765  * This timer is actually reconnect_timer, which is for making sure
1766  * the total recovery window is at least as big as my reconnect
1767  * attempt timing. So the initial recovery time_out will be set to
1768  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1769  * from client is bigger than this, then the recovery time_out will
1770  * be extended to make sure the client could be reconnected, in the
1771  * process, the timeout from the new client should be ignored.
1772  */
1773
1774 static void
1775 check_and_start_recovery_timer(struct obd_device *obd,
1776                                struct ptlrpc_request *req,
1777                                int new_client)
1778 {
1779         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1780         struct obd_device_target *obt = &obd->u.obt;
1781
1782         if (!new_client && service_time)
1783                 /* Teach server about old server's estimates, as first guess
1784                  * at how long new requests will take. */
1785                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1786                             service_time);
1787
1788         target_start_recovery_timer(obd);
1789
1790         /* Convert the service time to RPC timeout,
1791          * and reuse service_time to limit stack usage. */
1792         service_time = at_est2timeout(service_time);
1793
1794         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1795             service_time < at_extra)
1796                 service_time = at_extra;
1797
1798         /* We expect other clients to timeout within service_time, then try
1799          * to reconnect, then try the failover server.  The max delay between
1800          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1801         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1802
1803         LASSERT(obt->obt_magic == OBT_MAGIC);
1804         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1805         if (service_time > obd->obd_recovery_timeout && !new_client)
1806                 extend_recovery_timer(obd, service_time, false);
1807 }
1808
1809 /** Health checking routines */
1810 static inline int exp_connect_healthy(struct obd_export *exp)
1811 {
1812         return (exp->exp_in_recovery);
1813 }
1814
1815 /** if export done req_replay or has replay in queue */
1816 static inline int exp_req_replay_healthy(struct obd_export *exp)
1817 {
1818         return (!exp->exp_req_replay_needed ||
1819                 atomic_read(&exp->exp_replay_count) > 0);
1820 }
1821
1822
1823 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1824 {
1825         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1826                exp_req_replay_healthy(exp);
1827 }
1828
1829 /** if export done lock_replay or has replay in queue */
1830 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1831 {
1832         return (!exp->exp_lock_replay_needed ||
1833                 atomic_read(&exp->exp_replay_count) > 0);
1834 }
1835
1836 static inline int exp_vbr_healthy(struct obd_export *exp)
1837 {
1838         return (!exp->exp_vbr_failed);
1839 }
1840
1841 static inline int exp_finished(struct obd_export *exp)
1842 {
1843         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1844 }
1845
1846 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1847 {
1848         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1849                 exp_finished(exp);
1850 }
1851
1852 static int check_for_next_transno(struct lu_target *lut)
1853 {
1854         struct ptlrpc_request *req = NULL;
1855         struct obd_device *obd = lut->lut_obd;
1856         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1857         int wake_up = 0, connected, completed, queue_len;
1858         __u64 req_transno = 0;
1859         __u64 update_transno = 0;
1860         __u64 next_transno = 0;
1861         ENTRY;
1862
1863         spin_lock(&obd->obd_recovery_task_lock);
1864         if (!list_empty(&obd->obd_req_replay_queue)) {
1865                 req = list_entry(obd->obd_req_replay_queue.next,
1866                                      struct ptlrpc_request, rq_list);
1867                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1868         }
1869
1870         if (tdtd != NULL)
1871                 update_transno = distribute_txn_get_next_transno(tdtd);
1872
1873         connected = atomic_read(&obd->obd_connected_clients);
1874         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1875         queue_len = obd->obd_requests_queued_for_recovery;
1876         next_transno = obd->obd_next_recovery_transno;
1877
1878         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1879                "req_transno: %llu, next_transno: %llu\n",
1880                obd->obd_max_recoverable_clients, connected, completed,
1881                queue_len, req_transno, next_transno);
1882
1883         if (obd->obd_abort_recovery) {
1884                 CDEBUG(D_HA, "waking for aborted recovery\n");
1885                 wake_up = 1;
1886         } else if (obd->obd_recovery_expired) {
1887                 CDEBUG(D_HA, "waking for expired recovery\n");
1888                 wake_up = 1;
1889         } else if (tdtd != NULL && req != NULL &&
1890                    is_req_replayed_by_update(req)) {
1891                 LASSERTF(req_transno < next_transno, "req_transno %llu"
1892                          "next_transno%llu\n", req_transno, next_transno);
1893                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1894                        req_transno);
1895                 wake_up = 1;
1896         } else if (req_transno == next_transno ||
1897                    (update_transno != 0 && update_transno <= next_transno)) {
1898                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1899                 wake_up = 1;
1900         } else if (queue_len > 0 &&
1901                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1902                 /** handle gaps occured due to lost reply or VBR */
1903                 LASSERTF(req_transno >= next_transno,
1904                          "req_transno: %llu, next_transno: %llu\n",
1905                          req_transno, next_transno);
1906                 CDEBUG(D_HA,
1907                        "%s: waking for gap in transno, VBR is %s (skip: "
1908                        "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1909                        ", next_update %lld last_committed: %lld)\n",
1910                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1911                        next_transno, queue_len, completed, connected,
1912                        req_transno, update_transno, obd->obd_last_committed);
1913                 obd->obd_next_recovery_transno = req_transno;
1914                 wake_up = 1;
1915         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1916                 CDEBUG(D_HA, "waking for completed recovery\n");
1917                 wake_up = 1;
1918         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1919                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1920                        " by fail_lock, waking up (%lld)\n", next_transno);
1921                 obd->obd_next_recovery_transno = req_transno;
1922                 wake_up = 1;
1923         }
1924         spin_unlock(&obd->obd_recovery_task_lock);
1925         return wake_up;
1926 }
1927
1928 static int check_for_next_lock(struct lu_target *lut)
1929 {
1930         struct obd_device *obd = lut->lut_obd;
1931         int wake_up = 0;
1932
1933         spin_lock(&obd->obd_recovery_task_lock);
1934         if (!list_empty(&obd->obd_lock_replay_queue)) {
1935                 CDEBUG(D_HA, "waking for next lock\n");
1936                 wake_up = 1;
1937         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1938                 CDEBUG(D_HA, "waking for completed lock replay\n");
1939                 wake_up = 1;
1940         } else if (obd->obd_abort_recovery) {
1941                 CDEBUG(D_HA, "waking for aborted recovery\n");
1942                 wake_up = 1;
1943         } else if (obd->obd_recovery_expired) {
1944                 CDEBUG(D_HA, "waking for expired recovery\n");
1945                 wake_up = 1;
1946         }
1947         spin_unlock(&obd->obd_recovery_task_lock);
1948
1949         return wake_up;
1950 }
1951
1952 /**
1953  * wait for recovery events,
1954  * check its status with help of check_routine
1955  * evict dead clients via health_check
1956  */
1957 static int target_recovery_overseer(struct lu_target *lut,
1958                                     int (*check_routine)(struct lu_target *),
1959                                     int (*health_check)(struct obd_export *))
1960 {
1961         struct obd_device       *obd = lut->lut_obd;
1962         struct target_distribute_txn_data *tdtd;
1963 repeat:
1964         if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
1965               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
1966                 __u64 next_update_transno = 0;
1967
1968                 /* Only abort the recovery if there are no update recovery
1969                  * left in the queue */
1970                 spin_lock(&obd->obd_recovery_task_lock);
1971                 if (lut->lut_tdtd != NULL) {
1972                         next_update_transno =
1973                                 distribute_txn_get_next_transno(lut->lut_tdtd);
1974
1975                         tdtd = lut->lut_tdtd;
1976                         /* If next_update_transno == 0, it probably because
1977                          * updatelog retrieve threads did not get any records
1978                          * yet, let's wait those threads stopped */
1979                         if (next_update_transno == 0) {
1980                                 struct l_wait_info lwi = { 0 };
1981
1982                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1983                                        atomic_read(
1984                                        &tdtd->tdtd_recovery_threads_count) == 0,
1985                                        &lwi);
1986
1987                                 next_update_transno =
1988                                         distribute_txn_get_next_transno(
1989                                                                 lut->lut_tdtd);
1990                         }
1991                 }
1992
1993                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
1994                         obd->obd_next_recovery_transno = next_update_transno;
1995                         spin_unlock(&obd->obd_recovery_task_lock);
1996                         /* Disconnect unfinished exports from clients, and
1997                          * keep connection from MDT to make sure the update
1998                          * recovery will still keep trying until some one
1999                          * manually abort the recovery */
2000                         class_disconnect_stale_exports(obd,
2001                                                 exp_finished_or_from_mdt);
2002                         /* Abort all of replay and replay lock req from
2003                          * clients */
2004                         abort_req_replay_queue(obd);
2005                         abort_lock_replay_queue(obd);
2006                         CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2007                                ")in the queue.\n", obd->obd_name,
2008                                next_update_transno);
2009                 } else {
2010                         obd->obd_abort_recovery = 1;
2011                         spin_unlock(&obd->obd_recovery_task_lock);
2012                         CWARN("%s recovery is aborted by hard timeout\n",
2013                               obd->obd_name);
2014                 }
2015         }
2016
2017         while (wait_event_timeout(obd->obd_next_transno_waitq,
2018                                   check_routine(lut),
2019                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2020                 /* wait indefinitely for event, but don't trigger watchdog */;
2021
2022         if (obd->obd_abort_recovery) {
2023                 CWARN("recovery is aborted, evict exports in recovery\n");
2024                 if (lut->lut_tdtd != NULL) {
2025                         struct l_wait_info lwi = { 0 };
2026
2027                         tdtd = lut->lut_tdtd;
2028                         /* Let's wait all of the update log recovery thread
2029                          * finished */
2030                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2031                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2032                              &lwi);
2033                         /* Then abort the update recovery list */
2034                         dtrq_list_destroy(lut->lut_tdtd);
2035                 }
2036
2037                 /** evict exports which didn't finish recovery yet */
2038                 class_disconnect_stale_exports(obd, exp_finished);
2039                 return 1;
2040         } else if (obd->obd_recovery_expired) {
2041                 obd->obd_recovery_expired = 0;
2042                 /** If some clients died being recovered, evict them */
2043                 LCONSOLE_WARN("%s: recovery is timed out, "
2044                               "evict stale exports\n", obd->obd_name);
2045                 /** evict cexports with no replay in queue, they are stalled */
2046                 class_disconnect_stale_exports(obd, health_check);
2047
2048                 /** continue with VBR */
2049                 spin_lock(&obd->obd_dev_lock);
2050                 obd->obd_version_recov = 1;
2051                 spin_unlock(&obd->obd_dev_lock);
2052                 /**
2053                  * reset timer, recovery will proceed with versions now,
2054                  * timeout is set just to handle reconnection delays
2055                  */
2056                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2057                 /** Wait for recovery events again, after evicting bad clients */
2058                 goto repeat;
2059         }
2060         return 0;
2061 }
2062
2063 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2064 {
2065         struct obd_device       *obd = lut->lut_obd;
2066         struct ptlrpc_request *req = NULL;
2067
2068         CDEBUG(D_HA, "Waiting for lock\n");
2069         if (target_recovery_overseer(lut, check_for_next_lock,
2070                                      exp_lock_replay_healthy))
2071                 abort_lock_replay_queue(obd);
2072
2073         spin_lock(&obd->obd_recovery_task_lock);
2074         if (!list_empty(&obd->obd_lock_replay_queue)) {
2075                 req = list_entry(obd->obd_lock_replay_queue.next,
2076                                      struct ptlrpc_request, rq_list);
2077                 list_del_init(&req->rq_list);
2078                 spin_unlock(&obd->obd_recovery_task_lock);
2079         } else {
2080                 spin_unlock(&obd->obd_recovery_task_lock);
2081                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2082                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2083                 /** evict exports failed VBR */
2084                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2085         }
2086         return req;
2087 }
2088
2089 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2090 {
2091         struct ptlrpc_request *req = NULL;
2092
2093         spin_lock(&obd->obd_recovery_task_lock);
2094         if (!list_empty(&obd->obd_final_req_queue)) {
2095                 req = list_entry(obd->obd_final_req_queue.next,
2096                                      struct ptlrpc_request, rq_list);
2097                 list_del_init(&req->rq_list);
2098                 spin_unlock(&obd->obd_recovery_task_lock);
2099                 if (req->rq_export->exp_in_recovery) {
2100                         spin_lock(&req->rq_export->exp_lock);
2101                         req->rq_export->exp_in_recovery = 0;
2102                         spin_unlock(&req->rq_export->exp_lock);
2103                 }
2104         } else {
2105                 spin_unlock(&obd->obd_recovery_task_lock);
2106         }
2107         return req;
2108 }
2109
2110 static void handle_recovery_req(struct ptlrpc_thread *thread,
2111                                 struct ptlrpc_request *req,
2112                                 svc_handler_t handler)
2113 {
2114         ENTRY;
2115
2116         /**
2117          * export can be evicted during recovery, no need to handle replays for
2118          * it after that, discard such request silently
2119          */
2120         if (req->rq_export->exp_disconnected)
2121                 RETURN_EXIT;
2122
2123         req->rq_session.lc_thread = thread;
2124         req->rq_svc_thread = thread;
2125         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2126
2127         /* thread context */
2128         lu_context_enter(&thread->t_env->le_ctx);
2129         (void)handler(req);
2130         lu_context_exit(&thread->t_env->le_ctx);
2131
2132         /* don't reset timer for final stage */
2133         if (!exp_finished(req->rq_export)) {
2134                 int to = obd_timeout;
2135
2136                 /**
2137                  * Add request timeout to the recovery time so next request from
2138                  * this client may come in recovery time
2139                  */
2140                 if (!AT_OFF) {
2141                         struct ptlrpc_service_part *svcpt;
2142
2143                         svcpt = req->rq_rqbd->rqbd_svcpt;
2144                         /* If the server sent early reply for this request,
2145                          * the client will recalculate the timeout according to
2146                          * current server estimate service time, so we will
2147                          * use the maxium timeout here for waiting the client
2148                          * sending the next req */
2149                         to = max((int)at_est2timeout(
2150                                  at_get(&svcpt->scp_at_estimate)),
2151                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2152                         /* Add 2 net_latency, one for balance rq_deadline
2153                          * (see ptl_send_rpc), one for resend the req to server,
2154                          * Note: client will pack net_latency in replay req
2155                          * (see ptlrpc_replay_req) */
2156                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2157                 }
2158                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2159         }
2160         EXIT;
2161 }
2162
2163 /** Checking routines for recovery */
2164 static int check_for_recovery_ready(struct lu_target *lut)
2165 {
2166         struct obd_device *obd = lut->lut_obd;
2167         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2168
2169         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2170                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2171                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2172                obd->obd_recovery_expired);
2173
2174         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2175                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2176                 if (clnts + obd->obd_stale_clients <
2177                     obd->obd_max_recoverable_clients)
2178                         return 0;
2179         }
2180
2181         if (lut->lut_tdtd != NULL) {
2182                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2183                     !obd->obd_abort_recovery) {
2184                         /* Let's extend recovery timer, in case the recovery
2185                          * timer expired, and some clients got evicted */
2186                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2187                                               true);
2188                         CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2189                                obd->obd_name, obd->obd_recovery_timeout);
2190                         return 0;
2191                 }
2192         }
2193
2194         return 1;
2195 }
2196
2197 enum {
2198         REQUEST_RECOVERY = 1,
2199         UPDATE_RECOVERY = 2,
2200 };
2201
2202 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2203 {
2204         __u64 transno = 0;
2205
2206         if (!list_empty(&obd->obd_req_replay_queue)) {
2207                 struct ptlrpc_request *req;
2208
2209                 req = list_entry(obd->obd_req_replay_queue.next,
2210                                  struct ptlrpc_request, rq_list);
2211                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2212         }
2213
2214         return transno;
2215 }
2216
2217 static __u64 get_next_transno(struct lu_target *lut, int *type)
2218 {
2219         struct obd_device *obd = lut->lut_obd;
2220         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2221         __u64 transno = 0;
2222         __u64 update_transno;
2223         ENTRY;
2224
2225         transno = get_next_replay_req_transno(obd);
2226         if (type != NULL)
2227                 *type = REQUEST_RECOVERY;
2228
2229         if (tdtd == NULL)
2230                 RETURN(transno);
2231
2232         update_transno = distribute_txn_get_next_transno(tdtd);
2233         if (transno == 0 || (transno >= update_transno &&
2234                              update_transno != 0)) {
2235                 transno = update_transno;
2236                 if (type != NULL)
2237                         *type = UPDATE_RECOVERY;
2238         }
2239
2240         RETURN(transno);
2241 }
2242
2243 /**
2244  * drop duplicate replay request
2245  *
2246  * Because the operation has been replayed by update recovery, the request
2247  * with the same transno will be dropped and also notify the client to send
2248  * next replay request.
2249  *
2250  * \param[in] env       execution environment
2251  * \param[in] obd       failover obd device
2252  * \param[in] req       request to be dropped
2253  */
2254 static void drop_duplicate_replay_req(struct lu_env *env,
2255                                       struct obd_device *obd,
2256                                       struct ptlrpc_request *req)
2257 {
2258         DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2259                   " update records are found.\n",
2260                   lustre_msg_get_transno(req->rq_reqmsg),
2261                   libcfs_nid2str(req->rq_peer.nid));
2262
2263         /* Right now, only for MDS reint operation update replay and
2264          * normal request replay can have the same transno */
2265         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2266                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2267                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2268                 if (likely(req->rq_export))
2269                         target_committed_to_req(req);
2270                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2271                 target_send_reply(req, req->rq_status, 0);
2272         } else {
2273                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2274                 libcfs_nid2str(req->rq_peer.nid));
2275         }
2276         target_exp_dequeue_req_replay(req);
2277         target_request_copy_put(req);
2278         obd->obd_replayed_requests++;
2279 }
2280
2281 static void replay_request_or_update(struct lu_env *env,
2282                                      struct lu_target *lut,
2283                                      struct target_recovery_data *trd,
2284                                      struct ptlrpc_thread *thread)
2285 {
2286         struct obd_device *obd = lut->lut_obd;
2287         struct ptlrpc_request *req = NULL;
2288         int                     type;
2289         __u64                   transno;
2290         ENTRY;
2291
2292         CDEBUG(D_HA, "Waiting for transno %lld\n",
2293                obd->obd_next_recovery_transno);
2294
2295         /* Replay all of request and update by transno */
2296         do {
2297                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2298
2299                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2300
2301                 /** It is needed to extend recovery window above
2302                  *  recovery_time_soft. Extending is possible only in the
2303                  *  end of recovery window (see more details in
2304                  *  handle_recovery_req()).
2305                  */
2306                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2307
2308                 if (target_recovery_overseer(lut, check_for_next_transno,
2309                                         exp_req_replay_healthy_or_from_mdt)) {
2310                         abort_req_replay_queue(obd);
2311                         abort_lock_replay_queue(obd);
2312                         goto abort;
2313                 }
2314
2315                 spin_lock(&obd->obd_recovery_task_lock);
2316                 transno = get_next_transno(lut, &type);
2317                 if (type == REQUEST_RECOVERY && transno != 0) {
2318                         /* Drop replay request from client side, if the
2319                          * replay has been executed by update with the
2320                          * same transno */
2321                         req = list_entry(obd->obd_req_replay_queue.next,
2322                                         struct ptlrpc_request, rq_list);
2323
2324                         list_del_init(&req->rq_list);
2325                         obd->obd_requests_queued_for_recovery--;
2326                         spin_unlock(&obd->obd_recovery_task_lock);
2327
2328                         /* Let's check if the request has been redone by
2329                          * update replay */
2330                         if (is_req_replayed_by_update(req)) {
2331                                 struct distribute_txn_replay_req *dtrq;
2332
2333                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2334                                                                    req->rq_xid);
2335                                 LASSERT(dtrq != NULL);
2336                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2337                                 list_del_init(&dtrq->dtrq_list);
2338                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2339                                 dtrq_destroy(dtrq);
2340
2341                                 drop_duplicate_replay_req(env, obd, req);
2342
2343                                 continue;
2344                         }
2345
2346                         LASSERT(trd->trd_processing_task == current_pid());
2347                         DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2348                                   lustre_msg_get_transno(req->rq_reqmsg),
2349                                   libcfs_nid2str(req->rq_peer.nid));
2350
2351                         handle_recovery_req(thread, req,
2352                                             trd->trd_recovery_handler);
2353                         /**
2354                          * bz18031: increase next_recovery_transno before
2355                          * target_request_copy_put() will drop exp_rpc reference
2356                          */
2357                         spin_lock(&obd->obd_recovery_task_lock);
2358                         obd->obd_next_recovery_transno++;
2359                         spin_unlock(&obd->obd_recovery_task_lock);
2360                         target_exp_dequeue_req_replay(req);
2361                         target_request_copy_put(req);
2362                         obd->obd_replayed_requests++;
2363                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2364                         struct distribute_txn_replay_req *dtrq;
2365                         int rc;
2366
2367                         spin_unlock(&obd->obd_recovery_task_lock);
2368
2369                         LASSERT(tdtd != NULL);
2370                         dtrq = distribute_txn_get_next_req(tdtd);
2371                         lu_context_enter(&thread->t_env->le_ctx);
2372                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2373                         lu_context_exit(&thread->t_env->le_ctx);
2374                         extend_recovery_timer(obd, obd_timeout, true);
2375
2376                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2377                                 CDEBUG(D_HA, "Move x%llu t%llu"
2378                                        " to finish list\n", dtrq->dtrq_xid,
2379                                        dtrq->dtrq_master_transno);
2380
2381                                 /* Add it to the replay finish list */
2382                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2383                                 list_add(&dtrq->dtrq_list,
2384                                          &tdtd->tdtd_replay_finish_list);
2385                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2386
2387                                 spin_lock(&obd->obd_recovery_task_lock);
2388                                 if (transno == obd->obd_next_recovery_transno)
2389                                         obd->obd_next_recovery_transno++;
2390                                 else if (transno >
2391                                          obd->obd_next_recovery_transno)
2392                                         obd->obd_next_recovery_transno =
2393                                                                 transno + 1;
2394                                 spin_unlock(&obd->obd_recovery_task_lock);
2395                         } else {
2396                                 dtrq_destroy(dtrq);
2397                         }
2398                 } else {
2399                         spin_unlock(&obd->obd_recovery_task_lock);
2400 abort:
2401                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2402                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2403                         /** evict exports failed VBR */
2404                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2405                         break;
2406                 }
2407         } while (1);
2408 }
2409
2410 static int target_recovery_thread(void *arg)
2411 {
2412         struct lu_target *lut = arg;
2413         struct obd_device *obd = lut->lut_obd;
2414         struct ptlrpc_request *req;
2415         struct target_recovery_data *trd = &obd->obd_recovery_data;
2416         unsigned long delta;
2417         struct lu_env *env;
2418         struct ptlrpc_thread *thread = NULL;
2419         int rc = 0;
2420         ENTRY;
2421
2422         unshare_fs_struct();
2423         OBD_ALLOC_PTR(thread);
2424         if (thread == NULL)
2425                 RETURN(-ENOMEM);
2426
2427         OBD_ALLOC_PTR(env);
2428         if (env == NULL) {
2429                 OBD_FREE_PTR(thread);
2430                 RETURN(-ENOMEM);
2431         }
2432
2433         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2434         if (rc) {
2435                 OBD_FREE_PTR(thread);
2436                 OBD_FREE_PTR(env);
2437                 RETURN(rc);
2438         }
2439
2440         thread->t_env = env;
2441         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2442         env->le_ctx.lc_thread = thread;
2443         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2444         thread->t_watchdog = NULL;
2445
2446         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2447                current_pid());
2448         trd->trd_processing_task = current_pid();
2449
2450         spin_lock(&obd->obd_dev_lock);
2451         obd->obd_recovering = 1;
2452         spin_unlock(&obd->obd_dev_lock);
2453         complete(&trd->trd_starting);
2454
2455         /* first of all, we have to know the first transno to replay */
2456         if (target_recovery_overseer(lut, check_for_recovery_ready,
2457                                      exp_connect_healthy)) {
2458                 abort_req_replay_queue(obd);
2459                 abort_lock_replay_queue(obd);
2460                 if (lut->lut_tdtd != NULL)
2461                         dtrq_list_destroy(lut->lut_tdtd);
2462         }
2463
2464         /* next stage: replay requests or update */
2465         delta = jiffies;
2466         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2467                atomic_read(&obd->obd_req_replay_clients),
2468                obd->obd_next_recovery_transno);
2469         replay_request_or_update(env, lut, trd, thread);
2470
2471         /**
2472          * The second stage: replay locks
2473          */
2474         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2475                atomic_read(&obd->obd_lock_replay_clients));
2476         while ((req = target_next_replay_lock(lut))) {
2477                 LASSERT(trd->trd_processing_task == current_pid());
2478                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2479                           libcfs_nid2str(req->rq_peer.nid));
2480                 handle_recovery_req(thread, req,
2481                                     trd->trd_recovery_handler);
2482                 target_request_copy_put(req);
2483                 obd->obd_replayed_locks++;
2484         }
2485
2486         /**
2487          * The third stage: reply on final pings, at this moment all clients
2488          * must have request in final queue
2489          */
2490         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2491         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2492         /** Update server last boot epoch */
2493         tgt_boot_epoch_update(lut);
2494         /* We drop recoverying flag to forward all new requests
2495          * to regular mds_handle() since now */
2496         spin_lock(&obd->obd_dev_lock);
2497         obd->obd_recovering = obd->obd_abort_recovery = 0;
2498         spin_unlock(&obd->obd_dev_lock);
2499         spin_lock(&obd->obd_recovery_task_lock);
2500         target_cancel_recovery_timer(obd);
2501         spin_unlock(&obd->obd_recovery_task_lock);
2502         while ((req = target_next_final_ping(obd))) {
2503                 LASSERT(trd->trd_processing_task == current_pid());
2504                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2505                           libcfs_nid2str(req->rq_peer.nid));
2506                 handle_recovery_req(thread, req,
2507                                     trd->trd_recovery_handler);
2508                 /* Because the waiting client can not send ping to server,
2509                  * so we need refresh the last_request_time, to avoid the
2510                  * export is being evicted */
2511                 ptlrpc_update_export_timer(req->rq_export, 0);
2512                 target_request_copy_put(req);
2513         }
2514
2515         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2516         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2517               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2518         if (delta > OBD_RECOVERY_TIME_SOFT) {
2519                 CWARN("too long recovery - read logs\n");
2520                 libcfs_debug_dumplog();
2521         }
2522
2523         target_finish_recovery(lut);
2524
2525         lu_context_fini(&env->le_ctx);
2526         trd->trd_processing_task = 0;
2527         complete(&trd->trd_finishing);
2528
2529         tgt_io_thread_done(thread);
2530         OBD_FREE_PTR(thread);
2531         OBD_FREE_PTR(env);
2532         RETURN(rc);
2533 }
2534
2535 static int target_start_recovery_thread(struct lu_target *lut,
2536                                         svc_handler_t handler)
2537 {
2538         struct obd_device *obd = lut->lut_obd;
2539         int rc = 0;
2540         struct target_recovery_data *trd = &obd->obd_recovery_data;
2541         int index;
2542
2543         memset(trd, 0, sizeof(*trd));
2544         init_completion(&trd->trd_starting);
2545         init_completion(&trd->trd_finishing);
2546         trd->trd_recovery_handler = handler;
2547
2548         rc = server_name2index(obd->obd_name, &index, NULL);
2549         if (rc < 0)
2550                 return rc;
2551
2552         if (!IS_ERR(kthread_run(target_recovery_thread,
2553                                 lut, "tgt_recover_%d", index))) {
2554                 wait_for_completion(&trd->trd_starting);
2555                 LASSERT(obd->obd_recovering != 0);
2556         } else {
2557                 rc = -ECHILD;
2558         }
2559
2560         return rc;
2561 }
2562
2563 void target_stop_recovery_thread(struct obd_device *obd)
2564 {
2565         if (obd->obd_recovery_data.trd_processing_task > 0) {
2566                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2567                 /** recovery can be done but postrecovery is not yet */
2568                 spin_lock(&obd->obd_dev_lock);
2569                 if (obd->obd_recovering) {
2570                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2571                         obd->obd_abort_recovery = 1;
2572                         wake_up(&obd->obd_next_transno_waitq);
2573                 }
2574                 spin_unlock(&obd->obd_dev_lock);
2575                 wait_for_completion(&trd->trd_finishing);
2576         }
2577 }
2578 EXPORT_SYMBOL(target_stop_recovery_thread);
2579
2580 void target_recovery_fini(struct obd_device *obd)
2581 {
2582         class_disconnect_exports(obd);
2583         target_stop_recovery_thread(obd);
2584         target_cleanup_recovery(obd);
2585 }
2586 EXPORT_SYMBOL(target_recovery_fini);
2587
2588 static void target_recovery_expired(unsigned long castmeharder)
2589 {
2590         struct obd_device *obd = (struct obd_device *)castmeharder;
2591         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2592                " after %llus (%d clients connected)\n",
2593                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2594                (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2595                atomic_read(&obd->obd_connected_clients));
2596
2597         obd->obd_recovery_expired = 1;
2598         wake_up(&obd->obd_next_transno_waitq);
2599 }
2600
2601 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2602 {
2603         struct obd_device *obd = lut->lut_obd;
2604
2605         if (lut->lut_bottom->dd_rdonly)
2606                 return;
2607
2608         if (obd->obd_max_recoverable_clients == 0) {
2609                 /** Update server last boot epoch */
2610                 tgt_boot_epoch_update(lut);
2611                 return;
2612         }
2613
2614         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2615                "last_transno %llu\n", obd->obd_name,
2616                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2617         LASSERT(obd->obd_stopping == 0);
2618         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2619         obd->obd_recovery_start = 0;
2620         obd->obd_recovery_end = 0;
2621
2622         setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2623                     (unsigned long)obd);
2624         target_start_recovery_thread(lut, handler);
2625 }
2626 EXPORT_SYMBOL(target_recovery_init);
2627
2628 static int target_process_req_flags(struct obd_device *obd,
2629                                     struct ptlrpc_request *req)
2630 {
2631         struct obd_export *exp = req->rq_export;
2632         LASSERT(exp != NULL);
2633         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2634                 /* client declares he's ready to replay locks */
2635                 spin_lock(&exp->exp_lock);
2636                 if (exp->exp_req_replay_needed) {
2637                         exp->exp_req_replay_needed = 0;
2638                         spin_unlock(&exp->exp_lock);
2639
2640                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2641                         atomic_dec(&obd->obd_req_replay_clients);
2642                 } else {
2643                         spin_unlock(&exp->exp_lock);
2644                 }
2645         }
2646         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2647                 /* client declares he's ready to complete recovery
2648                  * so, we put the request on th final queue */
2649                 spin_lock(&exp->exp_lock);
2650                 if (exp->exp_lock_replay_needed) {
2651                         exp->exp_lock_replay_needed = 0;
2652                         spin_unlock(&exp->exp_lock);
2653
2654                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2655                         atomic_dec(&obd->obd_lock_replay_clients);
2656                 } else {
2657                         spin_unlock(&exp->exp_lock);
2658                 }
2659         }
2660         return 0;
2661 }
2662
2663 int target_queue_recovery_request(struct ptlrpc_request *req,
2664                                   struct obd_device *obd)
2665 {
2666         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2667         struct ptlrpc_request *reqiter;
2668         int inserted = 0;
2669         ENTRY;
2670
2671         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2672                 /* Processing the queue right now, don't re-add. */
2673                 RETURN(1);
2674         }
2675
2676         target_process_req_flags(obd, req);
2677
2678         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2679                 /* client declares he's ready to complete recovery
2680                  * so, we put the request on th final queue */
2681                 target_request_copy_get(req);
2682                 DEBUG_REQ(D_HA, req, "queue final req");
2683                 wake_up(&obd->obd_next_transno_waitq);
2684                 spin_lock(&obd->obd_recovery_task_lock);
2685                 if (obd->obd_recovering) {
2686                         struct ptlrpc_request *tmp;
2687                         struct ptlrpc_request *duplicate = NULL;
2688
2689                         if (likely(!req->rq_export->exp_replay_done)) {
2690                                 req->rq_export->exp_replay_done = 1;
2691                                 list_add_tail(&req->rq_list,
2692                                               &obd->obd_final_req_queue);
2693                                 spin_unlock(&obd->obd_recovery_task_lock);
2694                                 RETURN(0);
2695                         }
2696
2697                         /* XXX O(n), but only happens if final ping is
2698                          * timed out, probably reorganize the list as
2699                          * a hash list later */
2700                         list_for_each_entry_safe(reqiter, tmp,
2701                                                  &obd->obd_final_req_queue,
2702                                                  rq_list) {
2703                                 if (reqiter->rq_export == req->rq_export) {
2704                                         list_del_init(&reqiter->rq_list);
2705                                         duplicate = reqiter;
2706                                         break;
2707                                 }
2708                         }
2709
2710                         list_add_tail(&req->rq_list,
2711                                       &obd->obd_final_req_queue);
2712                         req->rq_export->exp_replay_done = 1;
2713                         spin_unlock(&obd->obd_recovery_task_lock);
2714
2715                         if (duplicate != NULL) {
2716                                 DEBUG_REQ(D_HA, duplicate,
2717                                           "put prev final req\n");
2718                                 target_request_copy_put(duplicate);
2719                         }
2720                         RETURN(0);
2721                 } else {
2722                         spin_unlock(&obd->obd_recovery_task_lock);
2723                         target_request_copy_put(req);
2724                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2725                 }
2726         }
2727         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2728                 /* client declares he's ready to replay locks */
2729                 target_request_copy_get(req);
2730                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2731                 wake_up(&obd->obd_next_transno_waitq);
2732                 spin_lock(&obd->obd_recovery_task_lock);
2733                 LASSERT(obd->obd_recovering);
2734                 /* usually due to recovery abort */
2735                 if (!req->rq_export->exp_in_recovery) {
2736                         spin_unlock(&obd->obd_recovery_task_lock);
2737                         target_request_copy_put(req);
2738                         RETURN(-ENOTCONN);
2739                 }
2740                 LASSERT(req->rq_export->exp_lock_replay_needed);
2741                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2742                 spin_unlock(&obd->obd_recovery_task_lock);
2743                 RETURN(0);
2744         }
2745
2746         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2747          * (i.e. buflens etc are in my own byte order), but type-dependent
2748          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2749
2750         if (!transno) {
2751                 INIT_LIST_HEAD(&req->rq_list);
2752                 DEBUG_REQ(D_HA, req, "not queueing");
2753                 RETURN(1);
2754         }
2755
2756         /* If we're processing the queue, we want don't want to queue this
2757          * message.
2758          *
2759          * Also, if this request has a transno less than the one we're waiting
2760          * for, we should process it now.  It could (and currently always will)
2761          * be an open request for a descriptor that was opened some time ago.
2762          *
2763          * Also, a resent, replayed request that has already been
2764          * handled will pass through here and be processed immediately.
2765          */
2766         CDEBUG(D_HA, "Next recovery transno: %llu"
2767                ", current: %llu, replaying\n",
2768                obd->obd_next_recovery_transno, transno);
2769
2770         /* If the request has been replayed by update replay, then sends this
2771          * request to the recovery thread (replay_request_or_update()), where
2772          * it will be handled */
2773         spin_lock(&obd->obd_recovery_task_lock);
2774         if (transno < obd->obd_next_recovery_transno &&
2775             !is_req_replayed_by_update(req)) {
2776                 /* Processing the queue right now, don't re-add. */
2777                 LASSERT(list_empty(&req->rq_list));
2778                 spin_unlock(&obd->obd_recovery_task_lock);
2779                 RETURN(1);
2780         }
2781         spin_unlock(&obd->obd_recovery_task_lock);
2782
2783         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2784                 RETURN(0);
2785
2786         target_request_copy_get(req);
2787         if (!req->rq_export->exp_in_recovery) {
2788                 target_request_copy_put(req);
2789                 RETURN(-ENOTCONN);
2790         }
2791         LASSERT(req->rq_export->exp_req_replay_needed);
2792
2793         if (target_exp_enqueue_req_replay(req)) {
2794                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2795                 target_request_copy_put(req);
2796                 RETURN(0);
2797         }
2798
2799         /* XXX O(n^2) */
2800         spin_lock(&obd->obd_recovery_task_lock);
2801         LASSERT(obd->obd_recovering);
2802         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2803                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2804                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2805                         inserted = 1;
2806                         goto added;
2807                 }
2808
2809                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2810                              transno)) {
2811                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2812                                   "has been claimed by another client");
2813                         spin_unlock(&obd->obd_recovery_task_lock);
2814                         target_exp_dequeue_req_replay(req);
2815                         target_request_copy_put(req);
2816                         RETURN(0);
2817                 }
2818         }
2819 added:
2820         if (!inserted)
2821                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2822
2823         obd->obd_requests_queued_for_recovery++;
2824         spin_unlock(&obd->obd_recovery_task_lock);
2825         wake_up(&obd->obd_next_transno_waitq);
2826         RETURN(0);
2827 }
2828
2829 int target_handle_ping(struct ptlrpc_request *req)
2830 {
2831         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2832         return req_capsule_server_pack(&req->rq_pill);
2833 }
2834
2835 void target_committed_to_req(struct ptlrpc_request *req)
2836 {
2837         struct obd_export *exp = req->rq_export;
2838
2839         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2840                 lustre_msg_set_last_committed(req->rq_repmsg,
2841                                               exp->exp_last_committed);
2842         else
2843                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2844                           "%d)", exp->exp_obd->obd_no_transno,
2845                           req->rq_repmsg == NULL);
2846
2847         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2848                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2849 }
2850
2851 #endif /* HAVE_SERVER_SUPPORT */
2852
2853 /**
2854  * Packs current SLV and Limit into \a req.
2855  */
2856 int target_pack_pool_reply(struct ptlrpc_request *req)
2857 {
2858         struct obd_device *obd;
2859         ENTRY;
2860
2861         /* Check that we still have all structures alive as this may
2862          * be some late RPC at shutdown time. */
2863         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2864                      !exp_connect_lru_resize(req->rq_export))) {
2865                 lustre_msg_set_slv(req->rq_repmsg, 0);
2866                 lustre_msg_set_limit(req->rq_repmsg, 0);
2867                 RETURN(0);
2868         }
2869
2870         /* OBD is alive here as export is alive, which we checked above. */
2871         obd = req->rq_export->exp_obd;
2872
2873         read_lock(&obd->obd_pool_lock);
2874         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2875         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2876         read_unlock(&obd->obd_pool_lock);
2877
2878         RETURN(0);
2879 }
2880
2881 static int target_send_reply_msg(struct ptlrpc_request *req,
2882                                  int rc, int fail_id)
2883 {
2884         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2885                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2886                 return -ECOMM;
2887         }
2888         /* We can have a null rq_reqmsg in the event of bad signature or
2889          * no context when unwrapping */
2890         if (req->rq_reqmsg &&
2891             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2892             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2893                 return -ECOMM;
2894
2895         if (unlikely(rc)) {
2896                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2897                 req->rq_status = rc;
2898                 return ptlrpc_send_error(req, 1);
2899         } else {
2900                 DEBUG_REQ(D_NET, req, "sending reply");
2901         }
2902
2903         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2904 }
2905
2906 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2907 {
2908         struct ptlrpc_service_part *svcpt;
2909         int                        netrc;
2910         struct ptlrpc_reply_state *rs;
2911         struct obd_export         *exp;
2912         ENTRY;
2913
2914         if (req->rq_no_reply) {
2915                 EXIT;
2916                 return;
2917         }
2918
2919         svcpt = req->rq_rqbd->rqbd_svcpt;
2920         rs = req->rq_reply_state;
2921         if (rs == NULL || !rs->rs_difficult) {
2922                 /* no notifiers */
2923                 target_send_reply_msg (req, rc, fail_id);
2924                 EXIT;
2925                 return;
2926         }
2927
2928         /* must be an export if locks saved */
2929         LASSERT(req->rq_export != NULL);
2930         /* req/reply consistent */
2931         LASSERT(rs->rs_svcpt == svcpt);
2932
2933         /* "fresh" reply */
2934         LASSERT(!rs->rs_scheduled);
2935         LASSERT(!rs->rs_scheduled_ever);
2936         LASSERT(!rs->rs_handled);
2937         LASSERT(!rs->rs_on_net);
2938         LASSERT(rs->rs_export == NULL);
2939         LASSERT(list_empty(&rs->rs_obd_list));
2940         LASSERT(list_empty(&rs->rs_exp_list));
2941
2942         exp = class_export_get(req->rq_export);
2943
2944         /* disable reply scheduling while I'm setting up */
2945         rs->rs_scheduled = 1;
2946         rs->rs_on_net    = 1;
2947         rs->rs_xid       = req->rq_xid;
2948         rs->rs_transno   = req->rq_transno;
2949         rs->rs_export    = exp;
2950         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2951
2952         spin_lock(&exp->exp_uncommitted_replies_lock);
2953         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2954                rs->rs_transno, exp->exp_last_committed);
2955         if (rs->rs_transno > exp->exp_last_committed) {
2956                 /* not committed already */
2957                 list_add_tail(&rs->rs_obd_list,
2958                                   &exp->exp_uncommitted_replies);
2959         }
2960         spin_unlock(&exp->exp_uncommitted_replies_lock);
2961
2962         spin_lock(&exp->exp_lock);
2963         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2964         spin_unlock(&exp->exp_lock);
2965
2966         netrc = target_send_reply_msg(req, rc, fail_id);
2967
2968         spin_lock(&svcpt->scp_rep_lock);
2969
2970         atomic_inc(&svcpt->scp_nreps_difficult);
2971
2972         if (netrc != 0) {
2973                 /* error sending: reply is off the net.  Also we need +1
2974                  * reply ref until ptlrpc_handle_rs() is done
2975                  * with the reply state (if the send was successful, there
2976                  * would have been +1 ref for the net, which
2977                  * reply_out_callback leaves alone) */
2978                 rs->rs_on_net = 0;
2979                 ptlrpc_rs_addref(rs);
2980         }
2981
2982         spin_lock(&rs->rs_lock);
2983         if (rs->rs_transno <= exp->exp_last_committed ||
2984             (!rs->rs_on_net && !rs->rs_no_ack) ||
2985             list_empty(&rs->rs_exp_list) ||     /* completed already */
2986             list_empty(&rs->rs_obd_list)) {
2987                 CDEBUG(D_HA, "Schedule reply immediately\n");
2988                 ptlrpc_dispatch_difficult_reply(rs);
2989         } else {
2990                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2991                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2992         }
2993         spin_unlock(&rs->rs_lock);
2994         spin_unlock(&svcpt->scp_rep_lock);
2995         EXIT;
2996 }
2997
2998 enum ldlm_mode lck_compat_array[] = {
2999         [LCK_EX]    = LCK_COMPAT_EX,
3000         [LCK_PW]    = LCK_COMPAT_PW,
3001         [LCK_PR]    = LCK_COMPAT_PR,
3002         [LCK_CW]    = LCK_COMPAT_CW,
3003         [LCK_CR]    = LCK_COMPAT_CR,
3004         [LCK_NL]    = LCK_COMPAT_NL,
3005         [LCK_GROUP] = LCK_COMPAT_GROUP,
3006         [LCK_COS]   = LCK_COMPAT_COS,
3007 };
3008
3009 /**
3010  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3011  * not escape to the user level.
3012  */
3013 int ldlm_error2errno(enum ldlm_error error)
3014 {
3015         int result;
3016
3017         switch (error) {
3018         case ELDLM_OK:
3019         case ELDLM_LOCK_MATCHED:
3020                 result = 0;
3021                 break;
3022         case ELDLM_LOCK_CHANGED:
3023                 result = -ESTALE;
3024                 break;
3025         case ELDLM_LOCK_ABORTED:
3026                 result = -ENAVAIL;
3027                 break;
3028         case ELDLM_LOCK_REPLACED:
3029                 result = -ESRCH;
3030                 break;
3031         case ELDLM_NO_LOCK_DATA:
3032                 result = -ENOENT;
3033                 break;
3034         case ELDLM_NAMESPACE_EXISTS:
3035                 result = -EEXIST;
3036                 break;
3037         case ELDLM_BAD_NAMESPACE:
3038                 result = -EBADF;
3039                 break;
3040         default:
3041                 if (((int)error) < 0) { /* cast to signed type */
3042                         result = error; /* as ldlm_error can be unsigned */
3043                 } else {
3044                         CERROR("Invalid DLM result code: %d\n", error);
3045                         result = -EPROTO;
3046                 }
3047         }
3048         return result;
3049 }
3050 EXPORT_SYMBOL(ldlm_error2errno);
3051
3052 /**
3053  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3054  */
3055 enum ldlm_error ldlm_errno2error(int err_no)
3056 {
3057         int error;
3058
3059         switch (err_no) {
3060         case 0:
3061                 error = ELDLM_OK;
3062                 break;
3063         case -ESTALE:
3064                 error = ELDLM_LOCK_CHANGED;
3065                 break;
3066         case -ENAVAIL:
3067                 error = ELDLM_LOCK_ABORTED;
3068                 break;
3069         case -ESRCH:
3070                 error = ELDLM_LOCK_REPLACED;
3071                 break;
3072         case -ENOENT:
3073                 error = ELDLM_NO_LOCK_DATA;
3074                 break;
3075         case -EEXIST:
3076                 error = ELDLM_NAMESPACE_EXISTS;
3077                 break;
3078         case -EBADF:
3079                 error = ELDLM_BAD_NAMESPACE;
3080                 break;
3081         default:
3082                 error = err_no;
3083         }
3084         return error;
3085 }
3086
3087 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3088 void ldlm_dump_export_locks(struct obd_export *exp)
3089 {
3090         spin_lock(&exp->exp_locks_list_guard);
3091         if (!list_empty(&exp->exp_locks_list)) {
3092                 struct ldlm_lock *lock;
3093
3094                 CERROR("dumping locks for export %p,"
3095                        "ignore if the unmount doesn't hang\n", exp);
3096                 list_for_each_entry(lock, &exp->exp_locks_list,
3097                                         l_exp_refs_link)
3098                         LDLM_ERROR(lock, "lock:");
3099         }
3100         spin_unlock(&exp->exp_locks_list_guard);
3101 }
3102 #endif
3103
3104 #ifdef HAVE_SERVER_SUPPORT
3105 static int target_bulk_timeout(void *data)
3106 {
3107         ENTRY;
3108         /* We don't fail the connection here, because having the export
3109          * killed makes the (vital) call to commitrw very sad.
3110          */
3111         RETURN(1);
3112 }
3113
3114 static inline const char *bulk2type(struct ptlrpc_request *req)
3115 {
3116         if (req->rq_bulk_read)
3117                 return "READ";
3118         if (req->rq_bulk_write)
3119                 return "WRITE";
3120         return "UNKNOWN";
3121 }
3122
3123 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3124                    struct l_wait_info *lwi)
3125 {
3126         struct ptlrpc_request   *req = desc->bd_req;
3127         time_t                   start = cfs_time_current_sec();
3128         time_t                   deadline;
3129         int                      rc = 0;
3130
3131         ENTRY;
3132
3133         /* If there is eviction in progress, wait for it to finish. */
3134         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3135                 *lwi = LWI_INTR(NULL, NULL);
3136                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3137                                   !atomic_read(&exp->exp_obd->
3138                                                    obd_evict_inprogress),
3139                                   lwi);
3140         }
3141
3142         /* Check if client was evicted or reconnected already. */
3143         if (exp->exp_failed ||
3144             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3145                 rc = -ENOTCONN;
3146         } else {
3147                 if (req->rq_bulk_read)
3148                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3149
3150                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3151                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3152                 else /* old version, bulk matchbits is rq_xid */
3153                         req->rq_mbits = req->rq_xid;
3154
3155                 if (rc == 0)
3156                         rc = ptlrpc_start_bulk_transfer(desc);
3157         }
3158
3159         if (rc < 0) {
3160                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3161                           bulk2type(req), rc);
3162                 RETURN(rc);
3163         }
3164
3165         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3166                 ptlrpc_abort_bulk(desc);
3167                 RETURN(0);
3168         }
3169
3170         /* limit actual bulk transfer to bulk_timeout seconds */
3171         deadline = start + bulk_timeout;
3172         if (deadline > req->rq_deadline)
3173                 deadline = req->rq_deadline;
3174
3175         do {
3176                 long timeoutl = deadline - cfs_time_current_sec();
3177                 cfs_duration_t timeout = timeoutl <= 0 ?
3178                                          CFS_TICK : cfs_time_seconds(timeoutl);
3179                 time_t  rq_deadline;
3180
3181                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3182                                             target_bulk_timeout, desc);
3183                 rc = l_wait_event(desc->bd_waitq,
3184                                   !ptlrpc_server_bulk_active(desc) ||
3185                                   exp->exp_failed ||
3186                                   exp->exp_conn_cnt >
3187                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3188                                   lwi);
3189                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3190                 /* Wait again if we changed rq_deadline. */
3191                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3192                 deadline = start + bulk_timeout;
3193                 if (deadline > rq_deadline)
3194                         deadline = rq_deadline;
3195         } while ((rc == -ETIMEDOUT) &&
3196                  (deadline > cfs_time_current_sec()));
3197
3198         if (rc == -ETIMEDOUT) {
3199                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3200                           bulk2type(req), deadline - start,
3201                           cfs_time_current_sec() - deadline);
3202                 ptlrpc_abort_bulk(desc);
3203         } else if (exp->exp_failed) {
3204                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3205                           bulk2type(req));
3206                 rc = -ENOTCONN;
3207                 ptlrpc_abort_bulk(desc);
3208         } else if (exp->exp_conn_cnt >
3209                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3210                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3211                           bulk2type(req));
3212                 /* We don't reply anyway. */
3213                 rc = -ETIMEDOUT;
3214                 ptlrpc_abort_bulk(desc);
3215         } else if (desc->bd_failure) {
3216                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3217                           bulk2type(req));
3218                 /* XXX should this be a different errno? */
3219                 rc = -ETIMEDOUT;
3220         } else {
3221                 if (req->rq_bulk_write)
3222                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3223                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3224                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3225                                   bulk2type(req), desc->bd_nob_transferred,
3226                                   desc->bd_nob);
3227                         /* XXX should this be a different errno? */
3228                         rc = -ETIMEDOUT;
3229                 }
3230         }
3231
3232         RETURN(rc);
3233 }
3234 EXPORT_SYMBOL(target_bulk_io);
3235
3236 #endif /* HAVE_SERVER_SUPPORT */