Whamcloud - gitweb
LU-8500 ldlm: fix export reference problem
[fs/lustre-release.git] / lustre / ldlm / ldlm_lib.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2015, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
37 /**
38  * This file deals with various client/target related logic including recovery.
39  *
40  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
41  * should be moved.
42  */
43
44 #define DEBUG_SUBSYSTEM S_LDLM
45
46 #include <linux/kthread.h>
47 #include <libcfs/libcfs.h>
48 #include <obd.h>
49 #include <obd_class.h>
50 #include <lustre_dlm.h>
51 #include <lustre_net.h>
52 #include <lustre_sec.h>
53 #include "ldlm_internal.h"
54
55 /* @priority: If non-zero, move the selected connection to the list head.
56  * @create: If zero, only search in existing connections.
57  */
58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
59                            int priority, int create)
60 {
61         struct ptlrpc_connection *ptlrpc_conn;
62         struct obd_import_conn *imp_conn = NULL, *item;
63         int rc = 0;
64         ENTRY;
65
66         if (!create && !priority) {
67                 CDEBUG(D_HA, "Nothing to do\n");
68                 RETURN(-EINVAL);
69         }
70
71         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72         if (!ptlrpc_conn) {
73                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
74                 RETURN (-ENOENT);
75         }
76
77         if (create) {
78                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79                 if (!imp_conn) {
80                         GOTO(out_put, rc = -ENOMEM);
81                 }
82         }
83
84         spin_lock(&imp->imp_lock);
85         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
86                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87                         if (priority) {
88                                 list_del(&item->oic_item);
89                                 list_add(&item->oic_item,
90                                              &imp->imp_conn_list);
91                                 item->oic_last_attempt = 0;
92                         }
93                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
94                                imp, imp->imp_obd->obd_name, uuid->uuid,
95                                (priority ? ", moved to head" : ""));
96                         spin_unlock(&imp->imp_lock);
97                         GOTO(out_free, rc = 0);
98                 }
99         }
100         /* No existing import connection found for \a uuid. */
101         if (create) {
102                 imp_conn->oic_conn = ptlrpc_conn;
103                 imp_conn->oic_uuid = *uuid;
104                 imp_conn->oic_last_attempt = 0;
105                 if (priority)
106                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107                 else
108                         list_add_tail(&imp_conn->oic_item,
109                                           &imp->imp_conn_list);
110                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
111                        imp, imp->imp_obd->obd_name, uuid->uuid,
112                        (priority ? "head" : "tail"));
113         } else {
114                 spin_unlock(&imp->imp_lock);
115                 GOTO(out_free, rc = -ENOENT);
116         }
117
118         spin_unlock(&imp->imp_lock);
119         RETURN(0);
120 out_free:
121         if (imp_conn)
122                 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 out_put:
124         ptlrpc_connection_put(ptlrpc_conn);
125         RETURN(rc);
126 }
127
128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 {
130         return import_set_conn(imp, uuid, 1, 0);
131 }
132
133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
134                            int priority)
135 {
136         return import_set_conn(imp, uuid, priority, 1);
137 }
138 EXPORT_SYMBOL(client_import_add_conn);
139
140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
141 {
142         struct obd_import_conn *imp_conn;
143         struct obd_export *dlmexp;
144         int rc = -ENOENT;
145         ENTRY;
146
147         spin_lock(&imp->imp_lock);
148         if (list_empty(&imp->imp_conn_list)) {
149                 LASSERT(!imp->imp_connection);
150                 GOTO(out, rc);
151         }
152
153         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
154                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
155                         continue;
156                 LASSERT(imp_conn->oic_conn);
157
158                 if (imp_conn == imp->imp_conn_current) {
159                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
160
161                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
162                             imp->imp_state != LUSTRE_IMP_DISCON) {
163                                 CERROR("can't remove current connection\n");
164                                 GOTO(out, rc = -EBUSY);
165                         }
166
167                         ptlrpc_connection_put(imp->imp_connection);
168                         imp->imp_connection = NULL;
169
170                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
171                         if (dlmexp && dlmexp->exp_connection) {
172                                 LASSERT(dlmexp->exp_connection ==
173                                         imp_conn->oic_conn);
174                                 ptlrpc_connection_put(dlmexp->exp_connection);
175                                 dlmexp->exp_connection = NULL;
176                         }
177
178                         if (dlmexp != NULL)
179                                 class_export_put(dlmexp);
180                 }
181
182                 list_del(&imp_conn->oic_item);
183                 ptlrpc_connection_put(imp_conn->oic_conn);
184                 OBD_FREE(imp_conn, sizeof(*imp_conn));
185                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
186                        imp, imp->imp_obd->obd_name, uuid->uuid);
187                 rc = 0;
188                 break;
189         }
190 out:
191         spin_unlock(&imp->imp_lock);
192         if (rc == -ENOENT)
193                 CERROR("connection %s not found\n", uuid->uuid);
194         RETURN(rc);
195 }
196 EXPORT_SYMBOL(client_import_del_conn);
197
198 /**
199  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
200  * to find a conn uuid of \a imp which can reach \a peer.
201  */
202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
203                             struct obd_uuid *uuid)
204 {
205         struct obd_import_conn *conn;
206         int rc = -ENOENT;
207         ENTRY;
208
209         spin_lock(&imp->imp_lock);
210         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
211                 /* Check if conn UUID does have this peer NID. */
212                 if (class_check_uuid(&conn->oic_uuid, peer)) {
213                         *uuid = conn->oic_uuid;
214                         rc = 0;
215                         break;
216                 }
217         }
218         spin_unlock(&imp->imp_lock);
219         RETURN(rc);
220 }
221 EXPORT_SYMBOL(client_import_find_conn);
222
223 void client_destroy_import(struct obd_import *imp)
224 {
225         /* Drop security policy instance after all RPCs have finished/aborted
226          * to let all busy contexts be released. */
227         class_import_get(imp);
228         class_destroy_import(imp);
229         sptlrpc_import_sec_put(imp);
230         class_import_put(imp);
231 }
232 EXPORT_SYMBOL(client_destroy_import);
233
234 /**
235  * Check whether or not the OSC is on MDT.
236  * In the config log,
237  * osc on MDT
238  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
239  * osc on client
240  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
241  *
242  **/
243 static int osc_on_mdt(char *obdname)
244 {
245         char *ptr;
246
247         ptr = strrchr(obdname, '-');
248         if (ptr == NULL)
249                 return 0;
250
251         if (strncmp(ptr + 1, "MDT", 3) == 0)
252                 return 1;
253
254         return 0;
255 }
256
257 /* Configure an RPC client OBD device.
258  *
259  * lcfg parameters:
260  * 1 - client UUID
261  * 2 - server UUID
262  * 3 - inactive-on-startup
263  */
264 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
265 {
266         struct client_obd *cli = &obddev->u.cli;
267         struct obd_import *imp;
268         struct obd_uuid server_uuid;
269         int rq_portal, rp_portal, connect_op;
270         char *name = obddev->obd_type->typ_name;
271         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
272         char *cli_name = lustre_cfg_buf(lcfg, 0);
273         int rc;
274         ENTRY;
275
276         /* In a more perfect world, we would hang a ptlrpc_client off of
277          * obd_type and just use the values from there. */
278         if (!strcmp(name, LUSTRE_OSC_NAME)) {
279                 rq_portal = OST_REQUEST_PORTAL;
280                 rp_portal = OSC_REPLY_PORTAL;
281                 connect_op = OST_CONNECT;
282                 cli->cl_sp_me = LUSTRE_SP_CLI;
283                 cli->cl_sp_to = LUSTRE_SP_OST;
284                 ns_type = LDLM_NS_TYPE_OSC;
285         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
286                    !strcmp(name, LUSTRE_LWP_NAME)) {
287                 rq_portal = MDS_REQUEST_PORTAL;
288                 rp_portal = MDC_REPLY_PORTAL;
289                 connect_op = MDS_CONNECT;
290                 if (is_lwp_on_ost(cli_name))
291                         cli->cl_sp_me = LUSTRE_SP_OST;
292                 else if (is_lwp_on_mdt(cli_name))
293                         cli->cl_sp_me = LUSTRE_SP_MDT;
294                 else
295                         cli->cl_sp_me = LUSTRE_SP_CLI;
296                 cli->cl_sp_to = LUSTRE_SP_MDT;
297                 ns_type = LDLM_NS_TYPE_MDC;
298         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
299                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
300                         /* OSP_on_MDT for other MDTs */
301                         connect_op = MDS_CONNECT;
302                         cli->cl_sp_to = LUSTRE_SP_MDT;
303                         ns_type = LDLM_NS_TYPE_MDC;
304                         rq_portal = OUT_PORTAL;
305                 } else {
306                         /* OSP on MDT for OST */
307                         connect_op = OST_CONNECT;
308                         cli->cl_sp_to = LUSTRE_SP_OST;
309                         ns_type = LDLM_NS_TYPE_OSC;
310                         rq_portal = OST_REQUEST_PORTAL;
311                 }
312                 rp_portal = OSC_REPLY_PORTAL;
313                 cli->cl_sp_me = LUSTRE_SP_MDT;
314         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
315                 rq_portal = MGS_REQUEST_PORTAL;
316                 rp_portal = MGC_REPLY_PORTAL;
317                 connect_op = MGS_CONNECT;
318                 cli->cl_sp_me = LUSTRE_SP_MGC;
319                 cli->cl_sp_to = LUSTRE_SP_MGS;
320                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
321                 ns_type = LDLM_NS_TYPE_MGC;
322         } else {
323                 CERROR("unknown client OBD type \"%s\", can't setup\n",
324                        name);
325                 RETURN(-EINVAL);
326         }
327
328         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
329                 CERROR("requires a TARGET UUID\n");
330                 RETURN(-EINVAL);
331         }
332
333         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
334                 CERROR("client UUID must be less than 38 characters\n");
335                 RETURN(-EINVAL);
336         }
337
338         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
339                 CERROR("setup requires a SERVER UUID\n");
340                 RETURN(-EINVAL);
341         }
342
343         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
344                 CERROR("target UUID must be less than 38 characters\n");
345                 RETURN(-EINVAL);
346         }
347
348         init_rwsem(&cli->cl_sem);
349         mutex_init(&cli->cl_mgc_mutex);
350         cli->cl_conn_count = 0;
351         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
352                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
353                      sizeof(server_uuid)));
354
355         cli->cl_dirty_pages = 0;
356         cli->cl_avail_grant = 0;
357         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
358         /* cl_dirty_max_pages may be changed at connect time in
359          * ptlrpc_connect_interpret(). */
360         client_adjust_max_dirty(cli);
361         INIT_LIST_HEAD(&cli->cl_cache_waiters);
362         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
363         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
364         INIT_LIST_HEAD(&cli->cl_loi_write_list);
365         INIT_LIST_HEAD(&cli->cl_loi_read_list);
366         spin_lock_init(&cli->cl_loi_list_lock);
367         atomic_set(&cli->cl_pending_w_pages, 0);
368         atomic_set(&cli->cl_pending_r_pages, 0);
369         cli->cl_r_in_flight = 0;
370         cli->cl_w_in_flight = 0;
371
372         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
373         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
374         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
375         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
376         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
377         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
378
379         /* lru for osc. */
380         INIT_LIST_HEAD(&cli->cl_lru_osc);
381         atomic_set(&cli->cl_lru_shrinkers, 0);
382         atomic_long_set(&cli->cl_lru_busy, 0);
383         atomic_long_set(&cli->cl_lru_in_list, 0);
384         INIT_LIST_HEAD(&cli->cl_lru_list);
385         spin_lock_init(&cli->cl_lru_list_lock);
386         atomic_long_set(&cli->cl_unstable_count, 0);
387         INIT_LIST_HEAD(&cli->cl_shrink_list);
388
389         init_waitqueue_head(&cli->cl_destroy_waitq);
390         atomic_set(&cli->cl_destroy_in_flight, 0);
391 #ifdef ENABLE_CHECKSUM
392         /* Turn on checksumming by default. */
393         cli->cl_checksum = 1;
394         /*
395          * The supported checksum types will be worked out at connect time
396          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
397          * through procfs.
398          */
399         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
400 #endif
401         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
402
403         /* Set it to possible maximum size. It may be reduced by ocd_brw_size
404          * from OFD after connecting. */
405         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
406
407         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
408          * it will be updated at OSC connection time. */
409         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
410
411         if (!strcmp(name, LUSTRE_MDC_NAME)) {
412                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
413         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
414                 cli->cl_max_rpcs_in_flight = 2;
415         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
416                 cli->cl_max_rpcs_in_flight = 3;
417         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
418                 cli->cl_max_rpcs_in_flight = 4;
419         } else {
420                 if (osc_on_mdt(obddev->obd_name))
421                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
422                 else
423                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
424         }
425
426         spin_lock_init(&cli->cl_mod_rpcs_lock);
427         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
428         cli->cl_max_mod_rpcs_in_flight = 0;
429         cli->cl_mod_rpcs_in_flight = 0;
430         cli->cl_close_rpcs_in_flight = 0;
431         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
432         cli->cl_mod_tag_bitmap = NULL;
433
434         if (connect_op == MDS_CONNECT) {
435                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
436                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
437                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
438                 if (cli->cl_mod_tag_bitmap == NULL)
439                         GOTO(err, rc = -ENOMEM);
440         }
441
442         rc = ldlm_get_ref();
443         if (rc) {
444                 CERROR("ldlm_get_ref failed: %d\n", rc);
445                 GOTO(err, rc);
446         }
447
448         ptlrpc_init_client(rq_portal, rp_portal, name,
449                            &obddev->obd_ldlm_client);
450
451         imp = class_new_import(obddev);
452         if (imp == NULL)
453                 GOTO(err_ldlm, rc = -ENOENT);
454         imp->imp_client = &obddev->obd_ldlm_client;
455         imp->imp_connect_op = connect_op;
456         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
457                LUSTRE_CFG_BUFLEN(lcfg, 1));
458         class_import_put(imp);
459
460         rc = client_import_add_conn(imp, &server_uuid, 1);
461         if (rc) {
462                 CERROR("can't add initial connection\n");
463                 GOTO(err_import, rc);
464         }
465
466         cli->cl_import = imp;
467         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
468         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
469
470         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
471                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
472                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
473                                name, obddev->obd_name,
474                                cli->cl_target_uuid.uuid);
475                         spin_lock(&imp->imp_lock);
476                         imp->imp_deactive = 1;
477                         spin_unlock(&imp->imp_lock);
478                 }
479         }
480
481         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
482                                                    LDLM_NAMESPACE_CLIENT,
483                                                    LDLM_NAMESPACE_GREEDY,
484                                                    ns_type);
485         if (obddev->obd_namespace == NULL) {
486                 CERROR("Unable to create client namespace - %s\n",
487                        obddev->obd_name);
488                 GOTO(err_import, rc = -ENOMEM);
489         }
490
491         RETURN(rc);
492
493 err_import:
494         class_destroy_import(imp);
495 err_ldlm:
496         ldlm_put_ref();
497 err:
498         if (cli->cl_mod_tag_bitmap != NULL)
499                 OBD_FREE(cli->cl_mod_tag_bitmap,
500                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
501         cli->cl_mod_tag_bitmap = NULL;
502         RETURN(rc);
503
504 }
505 EXPORT_SYMBOL(client_obd_setup);
506
507 int client_obd_cleanup(struct obd_device *obddev)
508 {
509         struct client_obd *cli = &obddev->u.cli;
510         ENTRY;
511
512         ldlm_namespace_free_post(obddev->obd_namespace);
513         obddev->obd_namespace = NULL;
514
515         obd_cleanup_client_import(obddev);
516         LASSERT(obddev->u.cli.cl_import == NULL);
517
518         ldlm_put_ref();
519
520         if (cli->cl_mod_tag_bitmap != NULL)
521                 OBD_FREE(cli->cl_mod_tag_bitmap,
522                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
523         cli->cl_mod_tag_bitmap = NULL;
524
525         RETURN(0);
526 }
527 EXPORT_SYMBOL(client_obd_cleanup);
528
529 /* ->o_connect() method for client side (OSC and MDC and MGC) */
530 int client_connect_import(const struct lu_env *env,
531                           struct obd_export **exp,
532                           struct obd_device *obd, struct obd_uuid *cluuid,
533                           struct obd_connect_data *data, void *localdata)
534 {
535         struct client_obd       *cli    = &obd->u.cli;
536         struct obd_import       *imp    = cli->cl_import;
537         struct obd_connect_data *ocd;
538         struct lustre_handle    conn    = { 0 };
539         int                     rc;
540         bool                    is_mdc = false;
541         ENTRY;
542
543         *exp = NULL;
544         down_write(&cli->cl_sem);
545         if (cli->cl_conn_count > 0)
546                 GOTO(out_sem, rc = -EALREADY);
547
548         rc = class_connect(&conn, obd, cluuid);
549         if (rc)
550                 GOTO(out_sem, rc);
551
552         cli->cl_conn_count++;
553         *exp = class_conn2export(&conn);
554
555         LASSERT(obd->obd_namespace);
556
557         imp->imp_dlm_handle = conn;
558         rc = ptlrpc_init_import(imp);
559         if (rc != 0)
560                 GOTO(out_ldlm, rc);
561
562         ocd = &imp->imp_connect_data;
563         if (data) {
564                 *ocd = *data;
565                 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
566                                  LUSTRE_MDC_NAME, 3) == 0;
567                 if (is_mdc)
568                         data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
569                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
570                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
571         }
572
573         rc = ptlrpc_connect_import(imp);
574         if (rc != 0) {
575                 if (data && is_mdc)
576                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
577                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
578                 GOTO(out_ldlm, rc);
579         }
580         LASSERT(*exp != NULL && (*exp)->exp_connection);
581
582         if (data) {
583                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
584                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
585                          data->ocd_connect_flags, ocd->ocd_connect_flags);
586                 data->ocd_connect_flags = ocd->ocd_connect_flags;
587                 /* clear the flag as it was not set and is not known
588                  * by upper layers */
589                 if (is_mdc)
590                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
591         }
592
593         ptlrpc_pinger_add_import(imp);
594
595         EXIT;
596
597         if (rc) {
598 out_ldlm:
599                 cli->cl_conn_count--;
600                 class_disconnect(*exp);
601                 *exp = NULL;
602         }
603 out_sem:
604         up_write(&cli->cl_sem);
605
606         return rc;
607 }
608 EXPORT_SYMBOL(client_connect_import);
609
610 int client_disconnect_export(struct obd_export *exp)
611 {
612         struct obd_device *obd = class_exp2obd(exp);
613         struct client_obd *cli;
614         struct obd_import *imp;
615         int rc = 0, err;
616         ENTRY;
617
618         if (!obd) {
619                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
620                        exp, exp ? exp->exp_handle.h_cookie : -1);
621                 RETURN(-EINVAL);
622         }
623
624         cli = &obd->u.cli;
625         imp = cli->cl_import;
626
627         down_write(&cli->cl_sem);
628         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
629                 cli->cl_conn_count);
630
631         if (cli->cl_conn_count == 0) {
632                 CERROR("disconnecting disconnected device (%s)\n",
633                        obd->obd_name);
634                 GOTO(out_disconnect, rc = -EINVAL);
635         }
636
637         cli->cl_conn_count--;
638         if (cli->cl_conn_count != 0)
639                 GOTO(out_disconnect, rc = 0);
640
641         /* Mark import deactivated now, so we don't try to reconnect if any
642          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
643          * fully deactivate the import, or that would drop all requests. */
644         spin_lock(&imp->imp_lock);
645         imp->imp_deactive = 1;
646         spin_unlock(&imp->imp_lock);
647
648         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
649          * delete it regardless.  (It's safe to delete an import that was
650          * never added.) */
651         (void)ptlrpc_pinger_del_import(imp);
652
653         if (obd->obd_namespace != NULL) {
654                 /* obd_force == local only */
655                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
656                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
657                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
658         }
659
660         /* There's no need to hold sem while disconnecting an import,
661          * and it may actually cause deadlock in GSS. */
662         up_write(&cli->cl_sem);
663         rc = ptlrpc_disconnect_import(imp, 0);
664         down_write(&cli->cl_sem);
665
666         ptlrpc_invalidate_import(imp);
667
668         EXIT;
669
670 out_disconnect:
671         /* Use server style - class_disconnect should be always called for
672          * o_disconnect. */
673         err = class_disconnect(exp);
674         if (!rc && err)
675                 rc = err;
676
677         up_write(&cli->cl_sem);
678
679         RETURN(rc);
680 }
681 EXPORT_SYMBOL(client_disconnect_export);
682
683 #ifdef HAVE_SERVER_SUPPORT
684 int server_disconnect_export(struct obd_export *exp)
685 {
686         int rc;
687         ENTRY;
688
689         /* Disconnect early so that clients can't keep using export. */
690         rc = class_disconnect(exp);
691         /* Close import to avoid sending any requests. */
692         if (exp->exp_imp_reverse)
693                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
694
695         ldlm_bl_thread_wakeup();
696
697         /* complete all outstanding replies */
698         spin_lock(&exp->exp_lock);
699         while (!list_empty(&exp->exp_outstanding_replies)) {
700                 struct ptlrpc_reply_state *rs =
701                         list_entry(exp->exp_outstanding_replies.next,
702                                        struct ptlrpc_reply_state, rs_exp_list);
703                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
704
705                 spin_lock(&svcpt->scp_rep_lock);
706
707                 list_del_init(&rs->rs_exp_list);
708                 spin_lock(&rs->rs_lock);
709                 ptlrpc_schedule_difficult_reply(rs);
710                 spin_unlock(&rs->rs_lock);
711
712                 spin_unlock(&svcpt->scp_rep_lock);
713         }
714         spin_unlock(&exp->exp_lock);
715
716         RETURN(rc);
717 }
718 EXPORT_SYMBOL(server_disconnect_export);
719
720 /* --------------------------------------------------------------------------
721  * from old lib/target.c
722  * -------------------------------------------------------------------------- */
723
724 static int target_handle_reconnect(struct lustre_handle *conn,
725                                    struct obd_export *exp,
726                                    struct obd_uuid *cluuid)
727 {
728         struct obd_device *target;
729         struct lustre_handle *hdl;
730         cfs_time_t now;
731         cfs_time_t deadline;
732         int timeout;
733         int rc = 0;
734         ENTRY;
735
736         hdl = &exp->exp_imp_reverse->imp_remote_handle;
737         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
738                 conn->cookie = exp->exp_handle.h_cookie;
739                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
740                        " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
741                 RETURN(0);
742         }
743
744         target = exp->exp_obd;
745
746         /* Might be a re-connect after a partition. */
747         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
748                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
749                               "with handle %#llx. Rejecting client "
750                               "with the same UUID trying to reconnect "
751                               "with handle %#llx\n", target->obd_name,
752                               obd_uuid2str(&exp->exp_client_uuid),
753                               obd_export_nid2str(exp),
754                               hdl->cookie, conn->cookie);
755                 memset(conn, 0, sizeof *conn);
756                 /* target_handle_connect() treats EALREADY and
757                  * -EALREADY differently.  -EALREADY is an error
758                  * (same UUID, different handle). */
759                 RETURN(-EALREADY);
760         }
761
762         if (!target->obd_recovering) {
763                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
764                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
765                         obd_export_nid2str(exp));
766                 GOTO(out_already, rc);
767         }
768
769         now = cfs_time_current();
770         deadline = cfs_timer_deadline(&target->obd_recovery_timer);
771         if (cfs_time_before(now, deadline)) {
772                 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
773                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting,"
774                         " waiting for %d clients in recovery for"
775                         " %d:%.02d\n", target->obd_name,
776                         obd_uuid2str(&exp->exp_client_uuid),
777                         obd_export_nid2str(exp),
778                         target->obd_max_recoverable_clients,
779                         timeout / 60, timeout % 60);
780         } else {
781                 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
782                 LCONSOLE_WARN("%s: Recovery already passed deadline"
783                         " %d:%.02d, It is most likely due to DNE"
784                         " recovery is failed or stuck, please wait a"
785                         " few more minutes or abort the recovery.\n",
786                         target->obd_name, timeout / 60, timeout % 60);
787         }
788
789 out_already:
790         conn->cookie = exp->exp_handle.h_cookie;
791         /* target_handle_connect() treats EALREADY and
792          * -EALREADY differently.  EALREADY means we are
793          * doing a valid reconnect from the same client. */
794         RETURN(EALREADY);
795 }
796
797 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
798                           int error)
799 {
800         struct obd_export *exp = cb_data;
801
802         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
803                obd->obd_name, exp->exp_client_uuid.uuid);
804
805         spin_lock(&exp->exp_lock);
806         exp->exp_need_sync = 0;
807         spin_unlock(&exp->exp_lock);
808         class_export_cb_put(exp);
809 }
810
811 static void
812 check_and_start_recovery_timer(struct obd_device *obd,
813                                struct ptlrpc_request *req, int new_client);
814
815 /**
816  * update flags for import during reconnect process
817  */
818 static int rev_import_flags_update(struct obd_import *revimp,
819                                    struct ptlrpc_request *req)
820 {
821         int rc;
822         struct obd_connect_data *data;
823
824         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
825
826         if (data->ocd_connect_flags & OBD_CONNECT_AT)
827                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
828         else
829                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
830
831         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
832
833         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
834         if (rc) {
835                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
836                         revimp->imp_client->cli_name,
837                         libcfs_id2str(req->rq_peer), rc);
838                 return rc;
839         }
840
841         return 0;
842 }
843
844 /**
845  * Allocate a new reverse import for an export.
846  *
847  * \retval -errno in case error hit
848  * \retval 0 if reverse import correctly init
849  **/
850 int rev_import_init(struct obd_export *export)
851 {
852         struct obd_device *obd = export->exp_obd;
853         struct obd_import *revimp;
854
855         LASSERT(export->exp_imp_reverse == NULL);
856
857         revimp = class_new_import(obd);
858         if (revimp == NULL)
859                 return -ENOMEM;
860
861         revimp->imp_remote_handle.cookie = 0ULL;
862         revimp->imp_client = &obd->obd_ldlm_client;
863         revimp->imp_dlm_fake = 1;
864
865         /* it is safe to connect import in new state as no sends possible */
866         spin_lock(&export->exp_lock);
867         export->exp_imp_reverse = revimp;
868         spin_unlock(&export->exp_lock);
869         class_import_put(revimp);
870
871         return 0;
872 }
873 EXPORT_SYMBOL(rev_import_init);
874
875 /**
876  * Handle reconnect for an export.
877  *
878  * \param exp export to handle reconnect process
879  * \param req client reconnect request
880  *
881  * \retval -rc in case securitfy flavor can't be changed
882  * \retval 0 in case none problems
883  */
884 static int rev_import_reconnect(struct obd_export *exp,
885                                 struct ptlrpc_request *req)
886 {
887         struct obd_import *revimp = exp->exp_imp_reverse;
888         struct lustre_handle *lh;
889         int rc;
890
891         /* avoid sending a request until import flags are changed */
892         ptlrpc_import_enter_resend(revimp);
893
894         if (revimp->imp_connection != NULL)
895                 ptlrpc_connection_put(revimp->imp_connection);
896
897         /*
898          * client from recovery don't have a handle so we need to take from
899          * request. it may produce situation when wrong client connected
900          * to recovery as we trust a client uuid
901          */
902         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
903         revimp->imp_remote_handle = *lh;
904
905         /* unknown versions will be caught in
906          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
907         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
908
909         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
910
911         rc = rev_import_flags_update(revimp, req);
912         if (rc != 0) {
913                 /* it is safe to still be in RECOVERY phase as we are not able
914                  * to setup correct security flavor so requests are not able to
915                  * be delivered correctly */
916                 return rc;
917         }
918
919         /* resend all rpc's via new connection */
920         return ptlrpc_import_recovery_state_machine(revimp);
921 }
922
923 int target_handle_connect(struct ptlrpc_request *req)
924 {
925         struct obd_device *target = NULL;
926         struct obd_export *export = NULL;
927         /* connect handle - filled from target_handle_reconnect in
928          * reconnect case */
929         struct lustre_handle conn;
930         struct lustre_handle *tmp;
931         struct obd_uuid tgtuuid;
932         struct obd_uuid cluuid;
933         char *str;
934         int rc = 0;
935         char *target_start;
936         int target_len;
937         bool     mds_conn = false, lw_client = false;
938         bool     mds_mds_conn = false;
939         bool     new_mds_mds_conn = false;
940         bool     target_referenced = false;
941         struct obd_connect_data *data, *tmpdata;
942         int size, tmpsize;
943         lnet_nid_t *client_nid = NULL;
944         ENTRY;
945
946         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
947
948         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
949         if (str == NULL) {
950                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
951                 GOTO(out, rc = -EINVAL);
952         }
953
954         obd_str2uuid(&tgtuuid, str);
955         target = class_uuid2obd(&tgtuuid);
956         if (!target)
957                 target = class_name2obd(str);
958
959         if (!target) {
960                 deuuidify(str, NULL, &target_start, &target_len);
961                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
962                                    "from %s (no target). If you are running "
963                                    "an HA pair check that the target is "
964                                    "mounted on the other server.\n", str,
965                                    libcfs_nid2str(req->rq_peer.nid));
966                 GOTO(out, rc = -ENODEV);
967         }
968
969         spin_lock(&target->obd_dev_lock);
970         if (target->obd_stopping || !target->obd_set_up) {
971                 spin_unlock(&target->obd_dev_lock);
972
973                 deuuidify(str, NULL, &target_start, &target_len);
974                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
975                               target_len, target_start,
976                               libcfs_nid2str(req->rq_peer.nid),
977                               (target->obd_stopping ?
978                                "stopping" : "not set up"));
979                 GOTO(out, rc = -ENODEV);
980         }
981
982         if (target->obd_no_conn) {
983                 spin_unlock(&target->obd_dev_lock);
984
985                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
986                                "from %s\n", target->obd_name,
987                                libcfs_nid2str(req->rq_peer.nid));
988                 GOTO(out, rc = -EAGAIN);
989         }
990
991         /* Make sure the target isn't cleaned up while we're here. Yes,
992          * there's still a race between the above check and our incref here.
993          * Really, class_uuid2obd should take the ref. */
994         class_incref(target, __func__, current);
995         target_referenced = true;
996
997         target->obd_conn_inprogress++;
998         spin_unlock(&target->obd_dev_lock);
999
1000         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1001         if (str == NULL) {
1002                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1003                 GOTO(out, rc = -EINVAL);
1004         }
1005
1006         obd_str2uuid(&cluuid, str);
1007
1008         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1009         if (tmp == NULL)
1010                 GOTO(out, rc = -EPROTO);
1011
1012         conn = *tmp;
1013
1014         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1015                                     RCL_CLIENT);
1016         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1017         if (!data)
1018                 GOTO(out, rc = -EPROTO);
1019
1020         rc = req_capsule_server_pack(&req->rq_pill);
1021         if (rc)
1022                 GOTO(out, rc);
1023
1024 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1025         /* Don't allow clients to connect that are using old 1.8 format
1026          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1027          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1028          * FULL20 flag should be set on all connections since 2.0, but no
1029          * longer affects behaviour.
1030          *
1031          * Later this check will be disabled and the flag can be retired
1032          * completely once interop with 3.0 is no longer needed.
1033          */
1034         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1035                 GOTO(out, rc = -EPROTO);
1036 #endif
1037
1038         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1039                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1040                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1041                     data->ocd_version > LUSTRE_VERSION_CODE +
1042                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1043                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1044                                   "libclient connection attempt",
1045                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1046                                   "old" : "new",
1047                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1048                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1049                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1050                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1051                         data = req_capsule_server_sized_get(&req->rq_pill,
1052                                                             &RMF_CONNECT_DATA,
1053                                     offsetof(typeof(*data), ocd_version) +
1054                                              sizeof(data->ocd_version));
1055                         if (data) {
1056                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1057                                 data->ocd_version = LUSTRE_VERSION_CODE;
1058                         }
1059                         GOTO(out, rc = -EPROTO);
1060                 }
1061         }
1062
1063         /* Note: lw_client is needed in MDS-MDS failover during update log
1064          * processing, so we needs to allow lw_client to be connected at
1065          * anytime, instead of only the initial connection */
1066         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1067
1068         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1069                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1070                 mds_mds_conn = (data->ocd_connect_flags &
1071                                 OBD_CONNECT_MDS_MDS) != 0;
1072
1073                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1074                  * for Imperative Recovery connection from MGC to MGS.
1075                  *
1076                  * Via check OBD_CONNECT_FID, we can distinguish whether
1077                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1078                  * MGC or MDT. */
1079                 if (!lw_client &&
1080                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1081                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1082                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1083                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1084                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1085                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1086
1087                         /* We do not support the MDT-MDT interoperations with
1088                          * different version MDT because of protocol changes. */
1089                         if (unlikely(major != LUSTRE_MAJOR ||
1090                                      minor != LUSTRE_MINOR ||
1091                                      abs(patch - LUSTRE_PATCH) > 3)) {
1092                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1093                                         "connection from different version MDT "
1094                                         "(%d.%d.%d.%d) %s %s\n",
1095                                         target->obd_name, LUSTRE_MAJOR,
1096                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1097                                         major, minor, patch,
1098                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1099                                         libcfs_nid2str(req->rq_peer.nid), str);
1100
1101                                 GOTO(out, rc = -EPROTO);
1102                         }
1103                 }
1104         }
1105
1106         /* lctl gets a backstage, all-access pass. */
1107         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1108                 goto dont_check_exports;
1109
1110         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1111         if (!export)
1112                 goto no_export;
1113
1114         /* We've found an export in the hash. */
1115
1116         spin_lock(&export->exp_lock);
1117
1118         if (export->exp_connecting) { /* bug 9635, et. al. */
1119                 spin_unlock(&export->exp_lock);
1120                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1121                               export->exp_obd->obd_name, export,
1122                               libcfs_nid2str(req->rq_peer.nid));
1123                 class_export_put(export);
1124                 export = NULL;
1125                 rc = -EALREADY;
1126         } else if ((mds_conn || lw_client ||
1127                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1128                    export->exp_connection != NULL) {
1129                 spin_unlock(&export->exp_lock);
1130                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1131                         /* MDS or LWP reconnected after failover. */
1132                         LCONSOLE_WARN("%s: Received %s connection from "
1133                             "%s, removing former export from %s\n",
1134                             target->obd_name, mds_conn ? "MDS" : "LWP",
1135                             libcfs_nid2str(req->rq_peer.nid),
1136                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1137                 } else {
1138                         /* New MDS connection from the same NID. */
1139                         LCONSOLE_WARN("%s: Received new %s connection from "
1140                                 "%s, removing former export from same NID\n",
1141                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1142                                 libcfs_nid2str(req->rq_peer.nid));
1143                 }
1144
1145                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1146                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1147                         /* Because exports between MDTs will always be
1148                          * kept, let's do not fail such export if they
1149                          * come from the same NID, otherwise it might
1150                          * cause eviction between MDTs, which might
1151                          * cause namespace inconsistency */
1152                         spin_lock(&export->exp_lock);
1153                         export->exp_connecting = 1;
1154                         spin_unlock(&export->exp_lock);
1155                         conn.cookie = export->exp_handle.h_cookie;
1156                         rc = EALREADY;
1157                 } else {
1158                         class_fail_export(export);
1159                         class_export_put(export);
1160                         export = NULL;
1161                         rc = 0;
1162                 }
1163         } else if (export->exp_connection != NULL &&
1164                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1165                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
1166                     MSG_CONNECT_INITIAL)) {
1167                 spin_unlock(&export->exp_lock);
1168                 /* In MDS failover we have static UUID but NID can change. */
1169                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1170                               "existing nid %s is already connected\n",
1171                               target->obd_name, cluuid.uuid,
1172                               libcfs_nid2str(req->rq_peer.nid),
1173                               libcfs_nid2str(
1174                                       export->exp_connection->c_peer.nid));
1175                 rc = -EALREADY;
1176                 class_export_put(export);
1177                 export = NULL;
1178         } else {
1179                 export->exp_connecting = 1;
1180                 spin_unlock(&export->exp_lock);
1181                 LASSERT(export->exp_obd == target);
1182
1183                 rc = target_handle_reconnect(&conn, export, &cluuid);
1184         }
1185
1186         /* If we found an export, we already unlocked. */
1187         if (!export) {
1188 no_export:
1189                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1190         } else if (req->rq_export == NULL &&
1191                    atomic_read(&export->exp_rpc_count) > 0) {
1192                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1193                               "still busy with %d references\n",
1194                               target->obd_name, cluuid.uuid,
1195                               libcfs_nid2str(req->rq_peer.nid),
1196                               atomic_read(&export->exp_refcount));
1197                 GOTO(out, rc = -EBUSY);
1198         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1199                 if (!strstr(cluuid.uuid, "mdt"))
1200                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1201                                       "known client %s (at %s) because it "
1202                                       "is indicating it is a new client",
1203                                       target->obd_name, cluuid.uuid,
1204                                       libcfs_nid2str(req->rq_peer.nid));
1205                 GOTO(out, rc = -EALREADY);
1206         } else {
1207                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1208         }
1209
1210         if (rc < 0) {
1211                 GOTO(out, rc);
1212         }
1213
1214         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1215                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1216               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1217               export, (long)cfs_time_current_sec(),
1218               export ? (long)export->exp_last_request_time : 0);
1219
1220         /* If this is the first time a client connects, reset the recovery
1221          * timer. Discard lightweight connections which might be local. */
1222         if (!lw_client && rc == 0 && target->obd_recovering)
1223                 check_and_start_recovery_timer(target, req, export == NULL);
1224
1225         /* We want to handle EALREADY but *not* -EALREADY from
1226          * target_handle_reconnect(), return reconnection state in a flag. */
1227         if (rc == EALREADY) {
1228                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1229                 rc = 0;
1230         } else {
1231                 LASSERT(rc == 0);
1232         }
1233
1234         /* Tell the client if we support replayable requests. */
1235         if (target->obd_replayable)
1236                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1237         client_nid = &req->rq_peer.nid;
1238
1239         if (export == NULL) {
1240                 /* allow lightweight connections during recovery */
1241                 /* allow "new" MDT to be connected during recovery, since we
1242                  * need retrieve recovery update records from it */
1243                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1244                         cfs_time_t t;
1245                         int     c; /* connected */
1246                         int     i; /* in progress */
1247                         int     k; /* known */
1248                         int     s; /* stale/evicted */
1249
1250                         c = atomic_read(&target->obd_connected_clients);
1251                         i = atomic_read(&target->obd_lock_replay_clients);
1252                         k = target->obd_max_recoverable_clients;
1253                         s = target->obd_stale_clients;
1254                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1255                         t = cfs_time_sub(t, cfs_time_current());
1256                         t = cfs_duration_sec(t);
1257                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1258                                       "(at %s), waiting for %d known clients "
1259                                       "(%d recovered, %d in progress, and %d "
1260                                       "evicted) to recover in %d:%.02d\n",
1261                                       target->obd_name, cluuid.uuid,
1262                                       libcfs_nid2str(req->rq_peer.nid), k,
1263                                       c - i, i, s, (int)t / 60,
1264                                       (int)t % 60);
1265                         rc = -EBUSY;
1266                 } else {
1267 dont_check_exports:
1268                         rc = obd_connect(req->rq_svc_thread->t_env,
1269                                          &export, target, &cluuid, data,
1270                                          client_nid);
1271                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1272                                 lustre_msg_add_op_flags(req->rq_repmsg,
1273                                                         MSG_CONNECT_RECOVERING);
1274                         if (rc == 0) {
1275                                 conn.cookie = export->exp_handle.h_cookie;
1276                                 rc = rev_import_init(export);
1277                         }
1278
1279                         if (mds_mds_conn)
1280                                 new_mds_mds_conn = true;
1281                 }
1282         } else {
1283                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1284                                    export, target, &cluuid, data, client_nid);
1285         }
1286         if (rc)
1287                 GOTO(out, rc);
1288
1289         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1290         data->ocd_instance = target->u.obt.obt_instance;
1291
1292         /* Return only the parts of obd_connect_data that we understand, so the
1293          * client knows that we don't understand the rest. */
1294         if (data) {
1295                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1296                                                RCL_SERVER);
1297                 tmpdata = req_capsule_server_get(&req->rq_pill,
1298                                                  &RMF_CONNECT_DATA);
1299                 /* Don't use struct assignment here, because the client reply
1300                  * buffer may be smaller/larger than the local struct
1301                  * obd_connect_data. */
1302                 memcpy(tmpdata, data, min(tmpsize, size));
1303         }
1304
1305         /* If the client and the server are the same node, we will already
1306          * have an export that really points to the client's DLM export,
1307          * because we have a shared handles table.
1308          *
1309          * XXX this will go away when shaver stops sending the "connect" handle
1310          * in the real "remote handle" field of the request --phik 24 Apr 2003
1311          */
1312         ptlrpc_request_change_export(req, export);
1313
1314         spin_lock(&export->exp_lock);
1315         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1316                 spin_unlock(&export->exp_lock);
1317                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1318                        "or equal conn_cnt: %d >= %d\n",
1319                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1320                        export->exp_conn_cnt,
1321                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1322
1323                 GOTO(out, rc = -EALREADY);
1324         }
1325         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1326         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1327
1328         /* Don't evict liblustre clients for not pinging. */
1329         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1330                 export->exp_libclient = 1;
1331                 spin_unlock(&export->exp_lock);
1332
1333                 spin_lock(&target->obd_dev_lock);
1334                 list_del_init(&export->exp_obd_chain_timed);
1335                 spin_unlock(&target->obd_dev_lock);
1336         } else {
1337                 spin_unlock(&export->exp_lock);
1338         }
1339
1340         if (export->exp_connection != NULL) {
1341                 /* Check to see if connection came from another NID. */
1342                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1343                     !hlist_unhashed(&export->exp_nid_hash))
1344                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1345                                      &export->exp_connection->c_peer.nid,
1346                                      &export->exp_nid_hash);
1347
1348                 ptlrpc_connection_put(export->exp_connection);
1349         }
1350
1351         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1352                                                        req->rq_self,
1353                                                        &cluuid);
1354         if (hlist_unhashed(&export->exp_nid_hash))
1355                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1356                              &export->exp_connection->c_peer.nid,
1357                              &export->exp_nid_hash);
1358
1359         lustre_msg_set_handle(req->rq_repmsg, &conn);
1360
1361         rc = rev_import_reconnect(export, req);
1362         if (rc != 0)
1363                 GOTO(out, rc);
1364
1365         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1366                 int has_transno;
1367                 __u64 transno = data->ocd_transno;
1368
1369                 spin_lock(&export->exp_lock);
1370                 /* possible race with class_disconnect_stale_exports,
1371                  * export may be already in the eviction process */
1372                 if (export->exp_failed) {
1373                         spin_unlock(&export->exp_lock);
1374                         GOTO(out, rc = -ENODEV);
1375                 }
1376                 export->exp_in_recovery = 1;
1377                 export->exp_req_replay_needed = 1;
1378                 export->exp_lock_replay_needed = 1;
1379                 spin_unlock(&export->exp_lock);
1380
1381                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1382                                  MSG_CONNECT_TRANSNO);
1383                 if (has_transno && transno == 0)
1384                         CWARN("Connect with zero transno!\n");
1385
1386                 if (has_transno && transno > 0 &&
1387                     transno < target->obd_next_recovery_transno &&
1388                     transno > target->obd_last_committed) {
1389                         /* Another way is to use cmpxchg() to be lock-free. */
1390                         spin_lock(&target->obd_recovery_task_lock);
1391                         if (transno < target->obd_next_recovery_transno)
1392                                 target->obd_next_recovery_transno = transno;
1393                         spin_unlock(&target->obd_recovery_task_lock);
1394                 }
1395
1396                 atomic_inc(&target->obd_req_replay_clients);
1397                 atomic_inc(&target->obd_lock_replay_clients);
1398                 /* Note: MDS-MDS connection is allowed to be connected during
1399                  * recovery, no matter if the exports needs to be recoveried.
1400                  * Because we need retrieve updates logs from all other MDTs.
1401                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1402                  * also needs to be increased to match other recovery checking
1403                  * condition. */
1404                 if (new_mds_mds_conn)
1405                         target->obd_max_recoverable_clients++;
1406                 if (atomic_inc_return(&target->obd_connected_clients) ==
1407                     target->obd_max_recoverable_clients)
1408                         wake_up(&target->obd_next_transno_waitq);
1409         }
1410
1411         /* Tell the client we're in recovery, when client is involved in it. */
1412         if (target->obd_recovering && !lw_client)
1413                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1414
1415 out:
1416         if (export) {
1417                 spin_lock(&export->exp_lock);
1418                 export->exp_connecting = 0;
1419                 spin_unlock(&export->exp_lock);
1420
1421                 class_export_put(export);
1422         }
1423         if (target_referenced == true && target != NULL) {
1424                 spin_lock(&target->obd_dev_lock);
1425                 target->obd_conn_inprogress--;
1426                 spin_unlock(&target->obd_dev_lock);
1427
1428                 class_decref(target, __func__, current);
1429         }
1430         req->rq_status = rc;
1431         RETURN(rc);
1432 }
1433
1434 int target_handle_disconnect(struct ptlrpc_request *req)
1435 {
1436         int rc;
1437         ENTRY;
1438
1439         rc = req_capsule_server_pack(&req->rq_pill);
1440         if (rc)
1441                 RETURN(rc);
1442
1443         /* Keep the rq_export around so we can send the reply. */
1444         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1445
1446         RETURN(0);
1447 }
1448
1449 void target_destroy_export(struct obd_export *exp)
1450 {
1451         struct obd_import       *imp = NULL;
1452         /* exports created from last_rcvd data, and "fake"
1453            exports created by lctl don't have an import */
1454         spin_lock(&exp->exp_lock);
1455         if (exp->exp_imp_reverse != NULL) {
1456                 imp = exp->exp_imp_reverse;
1457                 exp->exp_imp_reverse = NULL;
1458         }
1459         spin_unlock(&exp->exp_lock);
1460         if (imp != NULL)
1461                 client_destroy_import(imp);
1462
1463         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1464         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1465         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1466         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1467 }
1468 EXPORT_SYMBOL(target_destroy_export);
1469
1470 /*
1471  * Recovery functions
1472  */
1473 static void target_request_copy_get(struct ptlrpc_request *req)
1474 {
1475         class_export_rpc_inc(req->rq_export);
1476         LASSERT(list_empty(&req->rq_list));
1477         INIT_LIST_HEAD(&req->rq_replay_list);
1478
1479         /* Increase refcount to keep request in queue. */
1480         atomic_inc(&req->rq_refcount);
1481         /* Let export know it has replays to be handled. */
1482         atomic_inc(&req->rq_export->exp_replay_count);
1483 }
1484
1485 static void target_request_copy_put(struct ptlrpc_request *req)
1486 {
1487         LASSERT(list_empty(&req->rq_replay_list));
1488         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1489
1490         atomic_dec(&req->rq_export->exp_replay_count);
1491         class_export_rpc_dec(req->rq_export);
1492         ptlrpc_server_drop_request(req);
1493 }
1494
1495 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1496 {
1497         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1498         struct obd_export     *exp = req->rq_export;
1499         struct ptlrpc_request *reqiter;
1500         struct ptlrpc_request *dup_req = NULL;
1501         int                    dup = 0;
1502
1503         LASSERT(exp);
1504
1505         spin_lock(&exp->exp_lock);
1506         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1507                                 rq_replay_list) {
1508                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1509                         dup_req = reqiter;
1510                         dup = 1;
1511                         break;
1512                 }
1513         }
1514
1515         if (dup) {
1516                 /* We expect it with RESENT and REPLAY flags. */
1517                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1518                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1519                         CERROR("invalid flags %x of resent replay\n",
1520                                lustre_msg_get_flags(req->rq_reqmsg));
1521
1522                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1523                         __u32 new_conn;
1524
1525                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1526                         if (new_conn >
1527                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1528                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1529                                                         new_conn);
1530                 }
1531         } else {
1532                 list_add_tail(&req->rq_replay_list,
1533                                   &exp->exp_req_replay_queue);
1534         }
1535
1536         spin_unlock(&exp->exp_lock);
1537         return dup;
1538 }
1539
1540 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1541 {
1542         LASSERT(!list_empty(&req->rq_replay_list));
1543         LASSERT(req->rq_export);
1544
1545         spin_lock(&req->rq_export->exp_lock);
1546         list_del_init(&req->rq_replay_list);
1547         spin_unlock(&req->rq_export->exp_lock);
1548 }
1549
1550 static void target_finish_recovery(struct lu_target *lut)
1551 {
1552         struct obd_device *obd = lut->lut_obd;
1553         ENTRY;
1554
1555         /* Only log a recovery message when recovery has occurred. */
1556         if (obd->obd_recovery_start) {
1557                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1558                                         obd->obd_recovery_start);
1559                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1560                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1561                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1562                         obd->obd_max_recoverable_clients,
1563                         atomic_read(&obd->obd_connected_clients),
1564                         obd->obd_stale_clients,
1565                         obd->obd_stale_clients == 1 ? "was" : "were");
1566         }
1567
1568         ldlm_reprocess_all_ns(obd->obd_namespace);
1569         spin_lock(&obd->obd_recovery_task_lock);
1570         if (!list_empty(&obd->obd_req_replay_queue) ||
1571             !list_empty(&obd->obd_lock_replay_queue) ||
1572             !list_empty(&obd->obd_final_req_queue)) {
1573                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1574                        obd->obd_name,
1575                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1576                        list_empty(&obd->obd_lock_replay_queue) ? \
1577                                "" : "lock ",
1578                        list_empty(&obd->obd_final_req_queue) ? \
1579                                "" : "final ");
1580                 spin_unlock(&obd->obd_recovery_task_lock);
1581                 LBUG();
1582         }
1583         spin_unlock(&obd->obd_recovery_task_lock);
1584
1585         obd->obd_recovery_end = cfs_time_current_sec();
1586
1587         /* When recovery finished, cleanup orphans on MDS and OST. */
1588         if (OBT(obd) && OBP(obd, postrecov)) {
1589                 int rc = OBP(obd, postrecov)(obd);
1590                 if (rc < 0)
1591                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1592                                       obd->obd_name, rc);
1593         }
1594         EXIT;
1595 }
1596
1597 static void abort_req_replay_queue(struct obd_device *obd)
1598 {
1599         struct ptlrpc_request *req, *n;
1600         struct list_head abort_list;
1601
1602         INIT_LIST_HEAD(&abort_list);
1603         spin_lock(&obd->obd_recovery_task_lock);
1604         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1605         spin_unlock(&obd->obd_recovery_task_lock);
1606         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1607                 DEBUG_REQ(D_WARNING, req, "aborted:");
1608                 req->rq_status = -ENOTCONN;
1609                 if (ptlrpc_error(req)) {
1610                         DEBUG_REQ(D_ERROR, req,
1611                                   "failed abort_req_reply; skipping");
1612                 }
1613                 target_exp_dequeue_req_replay(req);
1614                 target_request_copy_put(req);
1615         }
1616 }
1617
1618 static void abort_lock_replay_queue(struct obd_device *obd)
1619 {
1620         struct ptlrpc_request *req, *n;
1621         struct list_head abort_list;
1622
1623         INIT_LIST_HEAD(&abort_list);
1624         spin_lock(&obd->obd_recovery_task_lock);
1625         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1626         spin_unlock(&obd->obd_recovery_task_lock);
1627         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1628                 DEBUG_REQ(D_ERROR, req, "aborted:");
1629                 req->rq_status = -ENOTCONN;
1630                 if (ptlrpc_error(req)) {
1631                         DEBUG_REQ(D_ERROR, req,
1632                                   "failed abort_lock_reply; skipping");
1633                 }
1634                 target_request_copy_put(req);
1635         }
1636 }
1637
1638 /* Called from a cleanup function if the device is being cleaned up
1639    forcefully.  The exports should all have been disconnected already,
1640    the only thing left to do is
1641      - clear the recovery flags
1642      - cancel the timer
1643      - free queued requests and replies, but don't send replies
1644    Because the obd_stopping flag is set, no new requests should be received.
1645
1646 */
1647 void target_cleanup_recovery(struct obd_device *obd)
1648 {
1649         struct ptlrpc_request *req, *n;
1650         struct list_head clean_list;
1651
1652         INIT_LIST_HEAD(&clean_list);
1653         spin_lock(&obd->obd_dev_lock);
1654         if (!obd->obd_recovering) {
1655                 spin_unlock(&obd->obd_dev_lock);
1656                 EXIT;
1657                 return;
1658         }
1659         obd->obd_recovering = obd->obd_abort_recovery = 0;
1660         spin_unlock(&obd->obd_dev_lock);
1661
1662         spin_lock(&obd->obd_recovery_task_lock);
1663         target_cancel_recovery_timer(obd);
1664         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1665         spin_unlock(&obd->obd_recovery_task_lock);
1666
1667         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1668                 LASSERT(req->rq_reply_state == NULL);
1669                 target_exp_dequeue_req_replay(req);
1670                 target_request_copy_put(req);
1671         }
1672
1673         spin_lock(&obd->obd_recovery_task_lock);
1674         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1675         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1676         spin_unlock(&obd->obd_recovery_task_lock);
1677
1678         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1679                 LASSERT(req->rq_reply_state == NULL);
1680                 target_request_copy_put(req);
1681         }
1682
1683         EXIT;
1684 }
1685 EXPORT_SYMBOL(target_cleanup_recovery);
1686
1687 /* obd_recovery_task_lock should be held */
1688 void target_cancel_recovery_timer(struct obd_device *obd)
1689 {
1690         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1691         cfs_timer_disarm(&obd->obd_recovery_timer);
1692 }
1693
1694 static void target_start_recovery_timer(struct obd_device *obd)
1695 {
1696         if (obd->obd_recovery_start != 0)
1697                 return;
1698
1699         spin_lock(&obd->obd_dev_lock);
1700         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1701                 spin_unlock(&obd->obd_dev_lock);
1702                 return;
1703         }
1704
1705         LASSERT(obd->obd_recovery_timeout != 0);
1706
1707         if (obd->obd_recovery_start != 0) {
1708                 spin_unlock(&obd->obd_dev_lock);
1709                 return;
1710         }
1711
1712         cfs_timer_arm(&obd->obd_recovery_timer,
1713                       cfs_time_shift(obd->obd_recovery_timeout));
1714         obd->obd_recovery_start = cfs_time_current_sec();
1715         spin_unlock(&obd->obd_dev_lock);
1716
1717         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1718                       "or until %d client%s reconnect%s\n",
1719                       obd->obd_name,
1720                       obd->obd_recovery_timeout / 60,
1721                       obd->obd_recovery_timeout % 60,
1722                       obd->obd_max_recoverable_clients,
1723                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1724                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1725 }
1726
1727 /**
1728  * extend recovery window.
1729  *
1730  * if @extend is true, extend recovery window to have @drt remaining at least;
1731  * otherwise, make sure the recovery timeout value is not less than @drt.
1732  */
1733 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1734 {
1735         cfs_time_t now;
1736         cfs_time_t end;
1737         cfs_duration_t left;
1738         int to;
1739
1740         spin_lock(&obd->obd_dev_lock);
1741         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1742                 spin_unlock(&obd->obd_dev_lock);
1743                 return;
1744         }
1745         LASSERT(obd->obd_recovery_start != 0);
1746
1747         now  = cfs_time_current_sec();
1748         to   = obd->obd_recovery_timeout;
1749         end  = obd->obd_recovery_start + to;
1750         left = cfs_time_sub(end, now);
1751
1752         if (extend && (drt > left)) {
1753                 to += drt - left;
1754         } else if (!extend && (drt > to)) {
1755                 to = drt;
1756         }
1757
1758         if (to > obd->obd_recovery_time_hard)
1759                 to = obd->obd_recovery_time_hard;
1760         if (obd->obd_recovery_timeout < to) {
1761                 obd->obd_recovery_timeout = to;
1762                 end = obd->obd_recovery_start + to;
1763                 cfs_timer_arm(&obd->obd_recovery_timer,
1764                                 cfs_time_shift(end - now));
1765         }
1766         spin_unlock(&obd->obd_dev_lock);
1767
1768         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1769                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1770 }
1771
1772 /* Reset the timer with each new client connection */
1773 /*
1774  * This timer is actually reconnect_timer, which is for making sure
1775  * the total recovery window is at least as big as my reconnect
1776  * attempt timing. So the initial recovery time_out will be set to
1777  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1778  * from client is bigger than this, then the recovery time_out will
1779  * be extended to make sure the client could be reconnected, in the
1780  * process, the timeout from the new client should be ignored.
1781  */
1782
1783 static void
1784 check_and_start_recovery_timer(struct obd_device *obd,
1785                                struct ptlrpc_request *req,
1786                                int new_client)
1787 {
1788         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1789         struct obd_device_target *obt = &obd->u.obt;
1790
1791         if (!new_client && service_time)
1792                 /* Teach server about old server's estimates, as first guess
1793                  * at how long new requests will take. */
1794                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1795                             service_time);
1796
1797         target_start_recovery_timer(obd);
1798
1799         /* Convert the service time to RPC timeout,
1800          * and reuse service_time to limit stack usage. */
1801         service_time = at_est2timeout(service_time);
1802
1803         /* We expect other clients to timeout within service_time, then try
1804          * to reconnect, then try the failover server.  The max delay between
1805          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1806         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1807
1808         LASSERT(obt->obt_magic == OBT_MAGIC);
1809         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1810         if (service_time > obd->obd_recovery_timeout && !new_client)
1811                 extend_recovery_timer(obd, service_time, false);
1812 }
1813
1814 /** Health checking routines */
1815 static inline int exp_connect_healthy(struct obd_export *exp)
1816 {
1817         return (exp->exp_in_recovery);
1818 }
1819
1820 /** if export done req_replay or has replay in queue */
1821 static inline int exp_req_replay_healthy(struct obd_export *exp)
1822 {
1823         return (!exp->exp_req_replay_needed ||
1824                 atomic_read(&exp->exp_replay_count) > 0);
1825 }
1826
1827
1828 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1829 {
1830         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1831                exp_req_replay_healthy(exp);
1832 }
1833
1834 /** if export done lock_replay or has replay in queue */
1835 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1836 {
1837         return (!exp->exp_lock_replay_needed ||
1838                 atomic_read(&exp->exp_replay_count) > 0);
1839 }
1840
1841 static inline int exp_vbr_healthy(struct obd_export *exp)
1842 {
1843         return (!exp->exp_vbr_failed);
1844 }
1845
1846 static inline int exp_finished(struct obd_export *exp)
1847 {
1848         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1849 }
1850
1851 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1852 {
1853         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1854                 exp_finished(exp);
1855 }
1856
1857 static int check_for_next_transno(struct lu_target *lut)
1858 {
1859         struct ptlrpc_request *req = NULL;
1860         struct obd_device *obd = lut->lut_obd;
1861         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1862         int wake_up = 0, connected, completed, queue_len;
1863         __u64 req_transno = 0;
1864         __u64 update_transno = 0;
1865         __u64 next_transno = 0;
1866         ENTRY;
1867
1868         spin_lock(&obd->obd_recovery_task_lock);
1869         if (!list_empty(&obd->obd_req_replay_queue)) {
1870                 req = list_entry(obd->obd_req_replay_queue.next,
1871                                      struct ptlrpc_request, rq_list);
1872                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1873         }
1874
1875         if (tdtd != NULL)
1876                 update_transno = distribute_txn_get_next_transno(tdtd);
1877
1878         connected = atomic_read(&obd->obd_connected_clients);
1879         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1880         queue_len = obd->obd_requests_queued_for_recovery;
1881         next_transno = obd->obd_next_recovery_transno;
1882
1883         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1884                "req_transno: %llu, next_transno: %llu\n",
1885                obd->obd_max_recoverable_clients, connected, completed,
1886                queue_len, req_transno, next_transno);
1887
1888         if (obd->obd_abort_recovery) {
1889                 CDEBUG(D_HA, "waking for aborted recovery\n");
1890                 wake_up = 1;
1891         } else if (obd->obd_recovery_expired) {
1892                 CDEBUG(D_HA, "waking for expired recovery\n");
1893                 wake_up = 1;
1894         } else if (tdtd != NULL && req != NULL &&
1895                    is_req_replayed_by_update(req)) {
1896                 LASSERTF(req_transno < next_transno, "req_transno %llu"
1897                          "next_transno%llu\n", req_transno, next_transno);
1898                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1899                        req_transno);
1900                 wake_up = 1;
1901         } else if (req_transno == next_transno ||
1902                    (update_transno != 0 && update_transno <= next_transno)) {
1903                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1904                 wake_up = 1;
1905         } else if (queue_len > 0 &&
1906                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1907                 /** handle gaps occured due to lost reply or VBR */
1908                 LASSERTF(req_transno >= next_transno,
1909                          "req_transno: %llu, next_transno: %llu\n",
1910                          req_transno, next_transno);
1911                 CDEBUG(D_HA,
1912                        "%s: waking for gap in transno, VBR is %s (skip: "
1913                        "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1914                        ", next_update %lld last_committed: %lld)\n",
1915                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1916                        next_transno, queue_len, completed, connected,
1917                        req_transno, update_transno, obd->obd_last_committed);
1918                 obd->obd_next_recovery_transno = req_transno;
1919                 wake_up = 1;
1920         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1921                 CDEBUG(D_HA, "waking for completed recovery\n");
1922                 wake_up = 1;
1923         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1924                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1925                        " by fail_lock, waking up (%lld)\n", next_transno);
1926                 obd->obd_next_recovery_transno = req_transno;
1927                 wake_up = 1;
1928         }
1929         spin_unlock(&obd->obd_recovery_task_lock);
1930         return wake_up;
1931 }
1932
1933 static int check_for_next_lock(struct lu_target *lut)
1934 {
1935         struct obd_device *obd = lut->lut_obd;
1936         int wake_up = 0;
1937
1938         spin_lock(&obd->obd_recovery_task_lock);
1939         if (!list_empty(&obd->obd_lock_replay_queue)) {
1940                 CDEBUG(D_HA, "waking for next lock\n");
1941                 wake_up = 1;
1942         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1943                 CDEBUG(D_HA, "waking for completed lock replay\n");
1944                 wake_up = 1;
1945         } else if (obd->obd_abort_recovery) {
1946                 CDEBUG(D_HA, "waking for aborted recovery\n");
1947                 wake_up = 1;
1948         } else if (obd->obd_recovery_expired) {
1949                 CDEBUG(D_HA, "waking for expired recovery\n");
1950                 wake_up = 1;
1951         }
1952         spin_unlock(&obd->obd_recovery_task_lock);
1953
1954         return wake_up;
1955 }
1956
1957 /**
1958  * wait for recovery events,
1959  * check its status with help of check_routine
1960  * evict dead clients via health_check
1961  */
1962 static int target_recovery_overseer(struct lu_target *lut,
1963                                     int (*check_routine)(struct lu_target *),
1964                                     int (*health_check)(struct obd_export *))
1965 {
1966         struct obd_device       *obd = lut->lut_obd;
1967         struct target_distribute_txn_data *tdtd;
1968 repeat:
1969         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1970               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1971                 __u64 next_update_transno = 0;
1972
1973                 /* Only abort the recovery if there are no update recovery
1974                  * left in the queue */
1975                 spin_lock(&obd->obd_recovery_task_lock);
1976                 if (lut->lut_tdtd != NULL) {
1977                         next_update_transno =
1978                                 distribute_txn_get_next_transno(lut->lut_tdtd);
1979
1980                         tdtd = lut->lut_tdtd;
1981                         /* If next_update_transno == 0, it probably because
1982                          * updatelog retrieve threads did not get any records
1983                          * yet, let's wait those threads stopped */
1984                         if (next_update_transno == 0) {
1985                                 struct l_wait_info lwi = { 0 };
1986
1987                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1988                                        atomic_read(
1989                                        &tdtd->tdtd_recovery_threads_count) == 0,
1990                                        &lwi);
1991
1992                                 next_update_transno =
1993                                         distribute_txn_get_next_transno(
1994                                                                 lut->lut_tdtd);
1995                         }
1996                 }
1997
1998                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
1999                         obd->obd_next_recovery_transno = next_update_transno;
2000                         spin_unlock(&obd->obd_recovery_task_lock);
2001                         /* Disconnect unfinished exports from clients, and
2002                          * keep connection from MDT to make sure the update
2003                          * recovery will still keep trying until some one
2004                          * manually abort the recovery */
2005                         class_disconnect_stale_exports(obd,
2006                                                 exp_finished_or_from_mdt);
2007                         /* Abort all of replay and replay lock req from
2008                          * clients */
2009                         abort_req_replay_queue(obd);
2010                         abort_lock_replay_queue(obd);
2011                         CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2012                                ")in the queue.\n", obd->obd_name,
2013                                next_update_transno);
2014                 } else {
2015                         obd->obd_abort_recovery = 1;
2016                         spin_unlock(&obd->obd_recovery_task_lock);
2017                         CWARN("%s recovery is aborted by hard timeout\n",
2018                               obd->obd_name);
2019                 }
2020         }
2021
2022         while (wait_event_timeout(obd->obd_next_transno_waitq,
2023                                   check_routine(lut),
2024                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2025                 /* wait indefinitely for event, but don't trigger watchdog */;
2026
2027         if (obd->obd_abort_recovery) {
2028                 CWARN("recovery is aborted, evict exports in recovery\n");
2029                 if (lut->lut_tdtd != NULL) {
2030                         struct l_wait_info lwi = { 0 };
2031
2032                         tdtd = lut->lut_tdtd;
2033                         /* Let's wait all of the update log recovery thread
2034                          * finished */
2035                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2036                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2037                              &lwi);
2038                         /* Then abort the update recovery list */
2039                         dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2040                         dtrq_list_destroy(lut->lut_tdtd);
2041                 }
2042
2043                 /** evict exports which didn't finish recovery yet */
2044                 class_disconnect_stale_exports(obd, exp_finished);
2045                 return 1;
2046         } else if (obd->obd_recovery_expired) {
2047                 obd->obd_recovery_expired = 0;
2048                 /** If some clients died being recovered, evict them */
2049                 LCONSOLE_WARN("%s: recovery is timed out, "
2050                               "evict stale exports\n", obd->obd_name);
2051                 /** evict cexports with no replay in queue, they are stalled */
2052                 class_disconnect_stale_exports(obd, health_check);
2053
2054                 /** continue with VBR */
2055                 spin_lock(&obd->obd_dev_lock);
2056                 obd->obd_version_recov = 1;
2057                 spin_unlock(&obd->obd_dev_lock);
2058                 /**
2059                  * reset timer, recovery will proceed with versions now,
2060                  * timeout is set just to handle reconnection delays
2061                  */
2062                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2063                 /** Wait for recovery events again, after evicting bad clients */
2064                 goto repeat;
2065         }
2066         return 0;
2067 }
2068
2069 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2070 {
2071         struct obd_device       *obd = lut->lut_obd;
2072         struct ptlrpc_request *req = NULL;
2073
2074         CDEBUG(D_HA, "Waiting for lock\n");
2075         if (target_recovery_overseer(lut, check_for_next_lock,
2076                                      exp_lock_replay_healthy))
2077                 abort_lock_replay_queue(obd);
2078
2079         spin_lock(&obd->obd_recovery_task_lock);
2080         if (!list_empty(&obd->obd_lock_replay_queue)) {
2081                 req = list_entry(obd->obd_lock_replay_queue.next,
2082                                      struct ptlrpc_request, rq_list);
2083                 list_del_init(&req->rq_list);
2084                 spin_unlock(&obd->obd_recovery_task_lock);
2085         } else {
2086                 spin_unlock(&obd->obd_recovery_task_lock);
2087                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2088                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2089                 /** evict exports failed VBR */
2090                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2091         }
2092         return req;
2093 }
2094
2095 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2096 {
2097         struct ptlrpc_request *req = NULL;
2098
2099         spin_lock(&obd->obd_recovery_task_lock);
2100         if (!list_empty(&obd->obd_final_req_queue)) {
2101                 req = list_entry(obd->obd_final_req_queue.next,
2102                                      struct ptlrpc_request, rq_list);
2103                 list_del_init(&req->rq_list);
2104                 spin_unlock(&obd->obd_recovery_task_lock);
2105                 if (req->rq_export->exp_in_recovery) {
2106                         spin_lock(&req->rq_export->exp_lock);
2107                         req->rq_export->exp_in_recovery = 0;
2108                         spin_unlock(&req->rq_export->exp_lock);
2109                 }
2110         } else {
2111                 spin_unlock(&obd->obd_recovery_task_lock);
2112         }
2113         return req;
2114 }
2115
2116 static void handle_recovery_req(struct ptlrpc_thread *thread,
2117                                 struct ptlrpc_request *req,
2118                                 svc_handler_t handler)
2119 {
2120         ENTRY;
2121
2122         /**
2123          * export can be evicted during recovery, no need to handle replays for
2124          * it after that, discard such request silently
2125          */
2126         if (req->rq_export->exp_disconnected)
2127                 RETURN_EXIT;
2128
2129         req->rq_session.lc_thread = thread;
2130         req->rq_svc_thread = thread;
2131         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2132
2133         /* thread context */
2134         lu_context_enter(&thread->t_env->le_ctx);
2135         (void)handler(req);
2136         lu_context_exit(&thread->t_env->le_ctx);
2137
2138         /* don't reset timer for final stage */
2139         if (!exp_finished(req->rq_export)) {
2140                 int to = obd_timeout;
2141
2142                 /**
2143                  * Add request timeout to the recovery time so next request from
2144                  * this client may come in recovery time
2145                  */
2146                 if (!AT_OFF) {
2147                         struct ptlrpc_service_part *svcpt;
2148
2149                         svcpt = req->rq_rqbd->rqbd_svcpt;
2150                         /* If the server sent early reply for this request,
2151                          * the client will recalculate the timeout according to
2152                          * current server estimate service time, so we will
2153                          * use the maxium timeout here for waiting the client
2154                          * sending the next req */
2155                         to = max((int)at_est2timeout(
2156                                  at_get(&svcpt->scp_at_estimate)),
2157                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2158                         /* Add 2 net_latency, one for balance rq_deadline
2159                          * (see ptl_send_rpc), one for resend the req to server,
2160                          * Note: client will pack net_latency in replay req
2161                          * (see ptlrpc_replay_req) */
2162                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2163                 }
2164                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2165         }
2166         EXIT;
2167 }
2168
2169 /** Checking routines for recovery */
2170 static int check_for_recovery_ready(struct lu_target *lut)
2171 {
2172         struct obd_device *obd = lut->lut_obd;
2173         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2174
2175         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2176                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2177                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2178                obd->obd_recovery_expired);
2179
2180         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2181                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2182                 if (clnts + obd->obd_stale_clients <
2183                     obd->obd_max_recoverable_clients)
2184                         return 0;
2185         }
2186
2187         if (lut->lut_tdtd != NULL) {
2188                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2189                     !obd->obd_abort_recovery) {
2190                         /* Let's extend recovery timer, in case the recovery
2191                          * timer expired, and some clients got evicted */
2192                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2193                                               true);
2194                         CDEBUG(D_HA, "%s update recovery is not ready,"
2195                                " extend recovery %d\n", obd->obd_name,
2196                                obd->obd_recovery_timeout);
2197                         return 0;
2198                 } else {
2199                         dtrq_list_dump(lut->lut_tdtd, D_HA);
2200                 }
2201         }
2202
2203         return 1;
2204 }
2205
2206 enum {
2207         REQUEST_RECOVERY = 1,
2208         UPDATE_RECOVERY = 2,
2209 };
2210
2211 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2212 {
2213         __u64 transno = 0;
2214
2215         if (!list_empty(&obd->obd_req_replay_queue)) {
2216                 struct ptlrpc_request *req;
2217
2218                 req = list_entry(obd->obd_req_replay_queue.next,
2219                                  struct ptlrpc_request, rq_list);
2220                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2221         }
2222
2223         return transno;
2224 }
2225
2226 static __u64 get_next_transno(struct lu_target *lut, int *type)
2227 {
2228         struct obd_device *obd = lut->lut_obd;
2229         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2230         __u64 transno = 0;
2231         __u64 update_transno;
2232         ENTRY;
2233
2234         transno = get_next_replay_req_transno(obd);
2235         if (type != NULL)
2236                 *type = REQUEST_RECOVERY;
2237
2238         if (tdtd == NULL)
2239                 RETURN(transno);
2240
2241         update_transno = distribute_txn_get_next_transno(tdtd);
2242         if (transno == 0 || (transno >= update_transno &&
2243                              update_transno != 0)) {
2244                 transno = update_transno;
2245                 if (type != NULL)
2246                         *type = UPDATE_RECOVERY;
2247         }
2248
2249         RETURN(transno);
2250 }
2251
2252 /**
2253  * drop duplicate replay request
2254  *
2255  * Because the operation has been replayed by update recovery, the request
2256  * with the same transno will be dropped and also notify the client to send
2257  * next replay request.
2258  *
2259  * \param[in] env       execution environment
2260  * \param[in] obd       failover obd device
2261  * \param[in] req       request to be dropped
2262  */
2263 static void drop_duplicate_replay_req(struct lu_env *env,
2264                                       struct obd_device *obd,
2265                                       struct ptlrpc_request *req)
2266 {
2267         DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2268                   " update records are found.\n",
2269                   lustre_msg_get_transno(req->rq_reqmsg),
2270                   libcfs_nid2str(req->rq_peer.nid));
2271
2272         /* Right now, only for MDS reint operation update replay and
2273          * normal request replay can have the same transno */
2274         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2275                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2276                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2277                 if (likely(req->rq_export))
2278                         target_committed_to_req(req);
2279                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2280                 target_send_reply(req, req->rq_status, 0);
2281         } else {
2282                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2283                 libcfs_nid2str(req->rq_peer.nid));
2284         }
2285         target_exp_dequeue_req_replay(req);
2286         target_request_copy_put(req);
2287         obd->obd_replayed_requests++;
2288 }
2289
2290 static void replay_request_or_update(struct lu_env *env,
2291                                      struct lu_target *lut,
2292                                      struct target_recovery_data *trd,
2293                                      struct ptlrpc_thread *thread)
2294 {
2295         struct obd_device *obd = lut->lut_obd;
2296         struct ptlrpc_request *req = NULL;
2297         int                     type;
2298         __u64                   transno;
2299         ENTRY;
2300
2301         CDEBUG(D_HA, "Waiting for transno %lld\n",
2302                obd->obd_next_recovery_transno);
2303
2304         /* Replay all of request and update by transno */
2305         do {
2306                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2307
2308                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2309
2310                 /** It is needed to extend recovery window above
2311                  *  recovery_time_soft. Extending is possible only in the
2312                  *  end of recovery window (see more details in
2313                  *  handle_recovery_req()).
2314                  */
2315                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2316
2317                 if (target_recovery_overseer(lut, check_for_next_transno,
2318                                         exp_req_replay_healthy_or_from_mdt)) {
2319                         abort_req_replay_queue(obd);
2320                         abort_lock_replay_queue(obd);
2321                         goto abort;
2322                 }
2323
2324                 spin_lock(&obd->obd_recovery_task_lock);
2325                 transno = get_next_transno(lut, &type);
2326                 if (type == REQUEST_RECOVERY && transno != 0) {
2327                         /* Drop replay request from client side, if the
2328                          * replay has been executed by update with the
2329                          * same transno */
2330                         req = list_entry(obd->obd_req_replay_queue.next,
2331                                         struct ptlrpc_request, rq_list);
2332
2333                         list_del_init(&req->rq_list);
2334                         obd->obd_requests_queued_for_recovery--;
2335                         spin_unlock(&obd->obd_recovery_task_lock);
2336
2337                         /* Let's check if the request has been redone by
2338                          * update replay */
2339                         if (is_req_replayed_by_update(req)) {
2340                                 struct distribute_txn_replay_req *dtrq;
2341
2342                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2343                                                                    req->rq_xid);
2344                                 LASSERT(dtrq != NULL);
2345                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2346                                 list_del_init(&dtrq->dtrq_list);
2347                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2348                                 dtrq_destroy(dtrq);
2349
2350                                 drop_duplicate_replay_req(env, obd, req);
2351
2352                                 continue;
2353                         }
2354
2355                         LASSERT(trd->trd_processing_task == current_pid());
2356                         DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2357                                   lustre_msg_get_transno(req->rq_reqmsg),
2358                                   libcfs_nid2str(req->rq_peer.nid));
2359
2360                         handle_recovery_req(thread, req,
2361                                             trd->trd_recovery_handler);
2362                         /**
2363                          * bz18031: increase next_recovery_transno before
2364                          * target_request_copy_put() will drop exp_rpc reference
2365                          */
2366                         spin_lock(&obd->obd_recovery_task_lock);
2367                         obd->obd_next_recovery_transno++;
2368                         spin_unlock(&obd->obd_recovery_task_lock);
2369                         target_exp_dequeue_req_replay(req);
2370                         target_request_copy_put(req);
2371                         obd->obd_replayed_requests++;
2372                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2373                         struct distribute_txn_replay_req *dtrq;
2374                         int rc;
2375
2376                         spin_unlock(&obd->obd_recovery_task_lock);
2377
2378                         LASSERT(tdtd != NULL);
2379                         dtrq = distribute_txn_get_next_req(tdtd);
2380                         lu_context_enter(&thread->t_env->le_ctx);
2381                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2382                         lu_context_exit(&thread->t_env->le_ctx);
2383                         extend_recovery_timer(obd, obd_timeout, true);
2384
2385                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2386                                 CDEBUG(D_HA, "Move x%llu t%llu"
2387                                        " to finish list\n", dtrq->dtrq_xid,
2388                                        dtrq->dtrq_master_transno);
2389
2390                                 /* Add it to the replay finish list */
2391                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2392                                 list_add(&dtrq->dtrq_list,
2393                                          &tdtd->tdtd_replay_finish_list);
2394                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2395
2396                                 spin_lock(&obd->obd_recovery_task_lock);
2397                                 if (transno == obd->obd_next_recovery_transno)
2398                                         obd->obd_next_recovery_transno++;
2399                                 else if (transno >
2400                                          obd->obd_next_recovery_transno)
2401                                         obd->obd_next_recovery_transno =
2402                                                                 transno + 1;
2403                                 spin_unlock(&obd->obd_recovery_task_lock);
2404                         } else {
2405                                 dtrq_destroy(dtrq);
2406                         }
2407                 } else {
2408                         spin_unlock(&obd->obd_recovery_task_lock);
2409 abort:
2410                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2411                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2412                         /** evict exports failed VBR */
2413                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2414                         break;
2415                 }
2416         } while (1);
2417 }
2418
2419 static int target_recovery_thread(void *arg)
2420 {
2421         struct lu_target *lut = arg;
2422         struct obd_device *obd = lut->lut_obd;
2423         struct ptlrpc_request *req;
2424         struct target_recovery_data *trd = &obd->obd_recovery_data;
2425         unsigned long delta;
2426         struct lu_env *env;
2427         struct ptlrpc_thread *thread = NULL;
2428         int rc = 0;
2429         ENTRY;
2430
2431         unshare_fs_struct();
2432         OBD_ALLOC_PTR(thread);
2433         if (thread == NULL)
2434                 RETURN(-ENOMEM);
2435
2436         OBD_ALLOC_PTR(env);
2437         if (env == NULL) {
2438                 OBD_FREE_PTR(thread);
2439                 RETURN(-ENOMEM);
2440         }
2441
2442         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2443         if (rc) {
2444                 OBD_FREE_PTR(thread);
2445                 OBD_FREE_PTR(env);
2446                 RETURN(rc);
2447         }
2448
2449         thread->t_env = env;
2450         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2451         env->le_ctx.lc_thread = thread;
2452         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2453         thread->t_watchdog = NULL;
2454
2455         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2456                current_pid());
2457         trd->trd_processing_task = current_pid();
2458
2459         spin_lock(&obd->obd_dev_lock);
2460         obd->obd_recovering = 1;
2461         spin_unlock(&obd->obd_dev_lock);
2462         complete(&trd->trd_starting);
2463
2464         /* first of all, we have to know the first transno to replay */
2465         if (target_recovery_overseer(lut, check_for_recovery_ready,
2466                                      exp_connect_healthy)) {
2467                 abort_req_replay_queue(obd);
2468                 abort_lock_replay_queue(obd);
2469                 if (lut->lut_tdtd != NULL)
2470                         dtrq_list_destroy(lut->lut_tdtd);
2471         }
2472
2473         /* next stage: replay requests or update */
2474         delta = jiffies;
2475         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2476                atomic_read(&obd->obd_req_replay_clients),
2477                obd->obd_next_recovery_transno);
2478         replay_request_or_update(env, lut, trd, thread);
2479
2480         /**
2481          * The second stage: replay locks
2482          */
2483         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2484                atomic_read(&obd->obd_lock_replay_clients));
2485         while ((req = target_next_replay_lock(lut))) {
2486                 LASSERT(trd->trd_processing_task == current_pid());
2487                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2488                           libcfs_nid2str(req->rq_peer.nid));
2489                 handle_recovery_req(thread, req,
2490                                     trd->trd_recovery_handler);
2491                 target_request_copy_put(req);
2492                 obd->obd_replayed_locks++;
2493         }
2494
2495         /**
2496          * The third stage: reply on final pings, at this moment all clients
2497          * must have request in final queue
2498          */
2499         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2500         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2501         /** Update server last boot epoch */
2502         tgt_boot_epoch_update(lut);
2503         /* We drop recoverying flag to forward all new requests
2504          * to regular mds_handle() since now */
2505         spin_lock(&obd->obd_dev_lock);
2506         obd->obd_recovering = obd->obd_abort_recovery = 0;
2507         spin_unlock(&obd->obd_dev_lock);
2508         spin_lock(&obd->obd_recovery_task_lock);
2509         target_cancel_recovery_timer(obd);
2510         spin_unlock(&obd->obd_recovery_task_lock);
2511         while ((req = target_next_final_ping(obd))) {
2512                 LASSERT(trd->trd_processing_task == current_pid());
2513                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2514                           libcfs_nid2str(req->rq_peer.nid));
2515                 handle_recovery_req(thread, req,
2516                                     trd->trd_recovery_handler);
2517                 /* Because the waiting client can not send ping to server,
2518                  * so we need refresh the last_request_time, to avoid the
2519                  * export is being evicted */
2520                 ptlrpc_update_export_timer(req->rq_export, 0);
2521                 target_request_copy_put(req);
2522         }
2523
2524         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2525         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2526               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2527         if (delta > OBD_RECOVERY_TIME_SOFT) {
2528                 CWARN("too long recovery - read logs\n");
2529                 libcfs_debug_dumplog();
2530         }
2531
2532         target_finish_recovery(lut);
2533
2534         lu_context_fini(&env->le_ctx);
2535         trd->trd_processing_task = 0;
2536         complete(&trd->trd_finishing);
2537
2538         tgt_io_thread_done(thread);
2539         OBD_FREE_PTR(thread);
2540         OBD_FREE_PTR(env);
2541         RETURN(rc);
2542 }
2543
2544 static int target_start_recovery_thread(struct lu_target *lut,
2545                                         svc_handler_t handler)
2546 {
2547         struct obd_device *obd = lut->lut_obd;
2548         int rc = 0;
2549         struct target_recovery_data *trd = &obd->obd_recovery_data;
2550         int index;
2551
2552         memset(trd, 0, sizeof(*trd));
2553         init_completion(&trd->trd_starting);
2554         init_completion(&trd->trd_finishing);
2555         trd->trd_recovery_handler = handler;
2556
2557         rc = server_name2index(obd->obd_name, &index, NULL);
2558         if (rc < 0)
2559                 return rc;
2560
2561         if (!IS_ERR(kthread_run(target_recovery_thread,
2562                                 lut, "tgt_recover_%d", index))) {
2563                 wait_for_completion(&trd->trd_starting);
2564                 LASSERT(obd->obd_recovering != 0);
2565         } else {
2566                 rc = -ECHILD;
2567         }
2568
2569         return rc;
2570 }
2571
2572 void target_stop_recovery_thread(struct obd_device *obd)
2573 {
2574         if (obd->obd_recovery_data.trd_processing_task > 0) {
2575                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2576                 /** recovery can be done but postrecovery is not yet */
2577                 spin_lock(&obd->obd_dev_lock);
2578                 if (obd->obd_recovering) {
2579                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2580                         obd->obd_abort_recovery = 1;
2581                         wake_up(&obd->obd_next_transno_waitq);
2582                 }
2583                 spin_unlock(&obd->obd_dev_lock);
2584                 wait_for_completion(&trd->trd_finishing);
2585         }
2586 }
2587 EXPORT_SYMBOL(target_stop_recovery_thread);
2588
2589 void target_recovery_fini(struct obd_device *obd)
2590 {
2591         class_disconnect_exports(obd);
2592         target_stop_recovery_thread(obd);
2593         target_cleanup_recovery(obd);
2594 }
2595 EXPORT_SYMBOL(target_recovery_fini);
2596
2597 static void target_recovery_expired(unsigned long castmeharder)
2598 {
2599         struct obd_device *obd = (struct obd_device *)castmeharder;
2600         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2601                " after %lds (%d clients connected)\n",
2602                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2603                cfs_time_current_sec()- obd->obd_recovery_start,
2604                atomic_read(&obd->obd_connected_clients));
2605
2606         obd->obd_recovery_expired = 1;
2607         wake_up(&obd->obd_next_transno_waitq);
2608 }
2609
2610 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2611 {
2612         struct obd_device *obd = lut->lut_obd;
2613
2614         if (obd->obd_max_recoverable_clients == 0) {
2615                 /** Update server last boot epoch */
2616                 tgt_boot_epoch_update(lut);
2617                 return;
2618         }
2619
2620         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2621                "last_transno %llu\n", obd->obd_name,
2622                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2623         LASSERT(obd->obd_stopping == 0);
2624         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2625         obd->obd_recovery_start = 0;
2626         obd->obd_recovery_end = 0;
2627
2628         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2629         target_start_recovery_thread(lut, handler);
2630 }
2631 EXPORT_SYMBOL(target_recovery_init);
2632
2633 static int target_process_req_flags(struct obd_device *obd,
2634                                     struct ptlrpc_request *req)
2635 {
2636         struct obd_export *exp = req->rq_export;
2637         LASSERT(exp != NULL);
2638         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2639                 /* client declares he's ready to replay locks */
2640                 spin_lock(&exp->exp_lock);
2641                 if (exp->exp_req_replay_needed) {
2642                         exp->exp_req_replay_needed = 0;
2643                         spin_unlock(&exp->exp_lock);
2644
2645                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2646                         atomic_dec(&obd->obd_req_replay_clients);
2647                 } else {
2648                         spin_unlock(&exp->exp_lock);
2649                 }
2650         }
2651         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2652                 /* client declares he's ready to complete recovery
2653                  * so, we put the request on th final queue */
2654                 spin_lock(&exp->exp_lock);
2655                 if (exp->exp_lock_replay_needed) {
2656                         exp->exp_lock_replay_needed = 0;
2657                         spin_unlock(&exp->exp_lock);
2658
2659                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2660                         atomic_dec(&obd->obd_lock_replay_clients);
2661                 } else {
2662                         spin_unlock(&exp->exp_lock);
2663                 }
2664         }
2665         return 0;
2666 }
2667
2668 int target_queue_recovery_request(struct ptlrpc_request *req,
2669                                   struct obd_device *obd)
2670 {
2671         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2672         struct ptlrpc_request *reqiter;
2673         int inserted = 0;
2674         ENTRY;
2675
2676         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2677                 /* Processing the queue right now, don't re-add. */
2678                 RETURN(1);
2679         }
2680
2681         target_process_req_flags(obd, req);
2682
2683         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2684                 /* client declares he's ready to complete recovery
2685                  * so, we put the request on th final queue */
2686                 target_request_copy_get(req);
2687                 DEBUG_REQ(D_HA, req, "queue final req");
2688                 wake_up(&obd->obd_next_transno_waitq);
2689                 spin_lock(&obd->obd_recovery_task_lock);
2690                 if (obd->obd_recovering) {
2691                         struct ptlrpc_request *tmp;
2692                         struct ptlrpc_request *duplicate = NULL;
2693
2694                         if (likely(!req->rq_export->exp_replay_done)) {
2695                                 req->rq_export->exp_replay_done = 1;
2696                                 list_add_tail(&req->rq_list,
2697                                               &obd->obd_final_req_queue);
2698                                 spin_unlock(&obd->obd_recovery_task_lock);
2699                                 RETURN(0);
2700                         }
2701
2702                         /* XXX O(n), but only happens if final ping is
2703                          * timed out, probably reorganize the list as
2704                          * a hash list later */
2705                         list_for_each_entry_safe(reqiter, tmp,
2706                                                  &obd->obd_final_req_queue,
2707                                                  rq_list) {
2708                                 if (reqiter->rq_export == req->rq_export) {
2709                                         list_del_init(&reqiter->rq_list);
2710                                         duplicate = reqiter;
2711                                         break;
2712                                 }
2713                         }
2714
2715                         list_add_tail(&req->rq_list,
2716                                       &obd->obd_final_req_queue);
2717                         req->rq_export->exp_replay_done = 1;
2718                         spin_unlock(&obd->obd_recovery_task_lock);
2719
2720                         if (duplicate != NULL) {
2721                                 DEBUG_REQ(D_HA, duplicate,
2722                                           "put prev final req\n");
2723                                 target_request_copy_put(duplicate);
2724                         }
2725                         RETURN(0);
2726                 } else {
2727                         spin_unlock(&obd->obd_recovery_task_lock);
2728                         target_request_copy_put(req);
2729                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2730                 }
2731         }
2732         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2733                 /* client declares he's ready to replay locks */
2734                 target_request_copy_get(req);
2735                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2736                 wake_up(&obd->obd_next_transno_waitq);
2737                 spin_lock(&obd->obd_recovery_task_lock);
2738                 LASSERT(obd->obd_recovering);
2739                 /* usually due to recovery abort */
2740                 if (!req->rq_export->exp_in_recovery) {
2741                         spin_unlock(&obd->obd_recovery_task_lock);
2742                         target_request_copy_put(req);
2743                         RETURN(-ENOTCONN);
2744                 }
2745                 LASSERT(req->rq_export->exp_lock_replay_needed);
2746                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2747                 spin_unlock(&obd->obd_recovery_task_lock);
2748                 RETURN(0);
2749         }
2750
2751         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2752          * (i.e. buflens etc are in my own byte order), but type-dependent
2753          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2754
2755         if (!transno) {
2756                 INIT_LIST_HEAD(&req->rq_list);
2757                 DEBUG_REQ(D_HA, req, "not queueing");
2758                 RETURN(1);
2759         }
2760
2761         /* If we're processing the queue, we want don't want to queue this
2762          * message.
2763          *
2764          * Also, if this request has a transno less than the one we're waiting
2765          * for, we should process it now.  It could (and currently always will)
2766          * be an open request for a descriptor that was opened some time ago.
2767          *
2768          * Also, a resent, replayed request that has already been
2769          * handled will pass through here and be processed immediately.
2770          */
2771         CDEBUG(D_HA, "Next recovery transno: %llu"
2772                ", current: %llu, replaying\n",
2773                obd->obd_next_recovery_transno, transno);
2774
2775         /* If the request has been replayed by update replay, then sends this
2776          * request to the recovery thread (replay_request_or_update()), where
2777          * it will be handled */
2778         spin_lock(&obd->obd_recovery_task_lock);
2779         if (transno < obd->obd_next_recovery_transno &&
2780             !is_req_replayed_by_update(req)) {
2781                 /* Processing the queue right now, don't re-add. */
2782                 LASSERT(list_empty(&req->rq_list));
2783                 spin_unlock(&obd->obd_recovery_task_lock);
2784                 RETURN(1);
2785         }
2786         spin_unlock(&obd->obd_recovery_task_lock);
2787
2788         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2789                 RETURN(0);
2790
2791         target_request_copy_get(req);
2792         if (!req->rq_export->exp_in_recovery) {
2793                 target_request_copy_put(req);
2794                 RETURN(-ENOTCONN);
2795         }
2796         LASSERT(req->rq_export->exp_req_replay_needed);
2797
2798         if (target_exp_enqueue_req_replay(req)) {
2799                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2800                 target_request_copy_put(req);
2801                 RETURN(0);
2802         }
2803
2804         /* XXX O(n^2) */
2805         spin_lock(&obd->obd_recovery_task_lock);
2806         LASSERT(obd->obd_recovering);
2807         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2808                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2809                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2810                         inserted = 1;
2811                         goto added;
2812                 }
2813
2814                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2815                              transno)) {
2816                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2817                                   "has been claimed by another client");
2818                         spin_unlock(&obd->obd_recovery_task_lock);
2819                         target_exp_dequeue_req_replay(req);
2820                         target_request_copy_put(req);
2821                         RETURN(0);
2822                 }
2823         }
2824 added:
2825         if (!inserted)
2826                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2827
2828         obd->obd_requests_queued_for_recovery++;
2829         spin_unlock(&obd->obd_recovery_task_lock);
2830         wake_up(&obd->obd_next_transno_waitq);
2831         RETURN(0);
2832 }
2833
2834 int target_handle_ping(struct ptlrpc_request *req)
2835 {
2836         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2837         return req_capsule_server_pack(&req->rq_pill);
2838 }
2839
2840 void target_committed_to_req(struct ptlrpc_request *req)
2841 {
2842         struct obd_export *exp = req->rq_export;
2843
2844         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2845                 lustre_msg_set_last_committed(req->rq_repmsg,
2846                                               exp->exp_last_committed);
2847         else
2848                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2849                           "%d)", exp->exp_obd->obd_no_transno,
2850                           req->rq_repmsg == NULL);
2851
2852         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2853                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2854 }
2855
2856 #endif /* HAVE_SERVER_SUPPORT */
2857
2858 /**
2859  * Packs current SLV and Limit into \a req.
2860  */
2861 int target_pack_pool_reply(struct ptlrpc_request *req)
2862 {
2863         struct obd_device *obd;
2864         ENTRY;
2865
2866         /* Check that we still have all structures alive as this may
2867          * be some late RPC at shutdown time. */
2868         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2869                      !exp_connect_lru_resize(req->rq_export))) {
2870                 lustre_msg_set_slv(req->rq_repmsg, 0);
2871                 lustre_msg_set_limit(req->rq_repmsg, 0);
2872                 RETURN(0);
2873         }
2874
2875         /* OBD is alive here as export is alive, which we checked above. */
2876         obd = req->rq_export->exp_obd;
2877
2878         read_lock(&obd->obd_pool_lock);
2879         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2880         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2881         read_unlock(&obd->obd_pool_lock);
2882
2883         RETURN(0);
2884 }
2885
2886 static int target_send_reply_msg(struct ptlrpc_request *req,
2887                                  int rc, int fail_id)
2888 {
2889         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2890                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2891                 return -ECOMM;
2892         }
2893         /* We can have a null rq_reqmsg in the event of bad signature or
2894          * no context when unwrapping */
2895         if (req->rq_reqmsg &&
2896             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2897             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2898                 return -ECOMM;
2899
2900         if (unlikely(rc)) {
2901                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2902                 req->rq_status = rc;
2903                 return ptlrpc_send_error(req, 1);
2904         } else {
2905                 DEBUG_REQ(D_NET, req, "sending reply");
2906         }
2907
2908         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2909 }
2910
2911 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2912 {
2913         struct ptlrpc_service_part *svcpt;
2914         int                        netrc;
2915         struct ptlrpc_reply_state *rs;
2916         struct obd_export         *exp;
2917         ENTRY;
2918
2919         if (req->rq_no_reply) {
2920                 EXIT;
2921                 return;
2922         }
2923
2924         svcpt = req->rq_rqbd->rqbd_svcpt;
2925         rs = req->rq_reply_state;
2926         if (rs == NULL || !rs->rs_difficult) {
2927                 /* no notifiers */
2928                 target_send_reply_msg (req, rc, fail_id);
2929                 EXIT;
2930                 return;
2931         }
2932
2933         /* must be an export if locks saved */
2934         LASSERT(req->rq_export != NULL);
2935         /* req/reply consistent */
2936         LASSERT(rs->rs_svcpt == svcpt);
2937
2938         /* "fresh" reply */
2939         LASSERT(!rs->rs_scheduled);
2940         LASSERT(!rs->rs_scheduled_ever);
2941         LASSERT(!rs->rs_handled);
2942         LASSERT(!rs->rs_on_net);
2943         LASSERT(rs->rs_export == NULL);
2944         LASSERT(list_empty(&rs->rs_obd_list));
2945         LASSERT(list_empty(&rs->rs_exp_list));
2946
2947         exp = class_export_get(req->rq_export);
2948
2949         /* disable reply scheduling while I'm setting up */
2950         rs->rs_scheduled = 1;
2951         rs->rs_on_net    = 1;
2952         rs->rs_xid       = req->rq_xid;
2953         rs->rs_transno   = req->rq_transno;
2954         rs->rs_export    = exp;
2955         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2956
2957         spin_lock(&exp->exp_uncommitted_replies_lock);
2958         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2959                rs->rs_transno, exp->exp_last_committed);
2960         if (rs->rs_transno > exp->exp_last_committed) {
2961                 /* not committed already */
2962                 list_add_tail(&rs->rs_obd_list,
2963                                   &exp->exp_uncommitted_replies);
2964         }
2965         spin_unlock(&exp->exp_uncommitted_replies_lock);
2966
2967         spin_lock(&exp->exp_lock);
2968         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2969         spin_unlock(&exp->exp_lock);
2970
2971         netrc = target_send_reply_msg(req, rc, fail_id);
2972
2973         spin_lock(&svcpt->scp_rep_lock);
2974
2975         atomic_inc(&svcpt->scp_nreps_difficult);
2976
2977         if (netrc != 0) {
2978                 /* error sending: reply is off the net.  Also we need +1
2979                  * reply ref until ptlrpc_handle_rs() is done
2980                  * with the reply state (if the send was successful, there
2981                  * would have been +1 ref for the net, which
2982                  * reply_out_callback leaves alone) */
2983                 rs->rs_on_net = 0;
2984                 ptlrpc_rs_addref(rs);
2985         }
2986
2987         spin_lock(&rs->rs_lock);
2988         if (rs->rs_transno <= exp->exp_last_committed ||
2989             (!rs->rs_on_net && !rs->rs_no_ack) ||
2990             list_empty(&rs->rs_exp_list) ||     /* completed already */
2991             list_empty(&rs->rs_obd_list)) {
2992                 CDEBUG(D_HA, "Schedule reply immediately\n");
2993                 ptlrpc_dispatch_difficult_reply(rs);
2994         } else {
2995                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2996                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2997         }
2998         spin_unlock(&rs->rs_lock);
2999         spin_unlock(&svcpt->scp_rep_lock);
3000         EXIT;
3001 }
3002
3003 enum ldlm_mode lck_compat_array[] = {
3004         [LCK_EX]    = LCK_COMPAT_EX,
3005         [LCK_PW]    = LCK_COMPAT_PW,
3006         [LCK_PR]    = LCK_COMPAT_PR,
3007         [LCK_CW]    = LCK_COMPAT_CW,
3008         [LCK_CR]    = LCK_COMPAT_CR,
3009         [LCK_NL]    = LCK_COMPAT_NL,
3010         [LCK_GROUP] = LCK_COMPAT_GROUP,
3011         [LCK_COS]   = LCK_COMPAT_COS,
3012 };
3013
3014 /**
3015  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3016  * not escape to the user level.
3017  */
3018 int ldlm_error2errno(enum ldlm_error error)
3019 {
3020         int result;
3021
3022         switch (error) {
3023         case ELDLM_OK:
3024         case ELDLM_LOCK_MATCHED:
3025                 result = 0;
3026                 break;
3027         case ELDLM_LOCK_CHANGED:
3028                 result = -ESTALE;
3029                 break;
3030         case ELDLM_LOCK_ABORTED:
3031                 result = -ENAVAIL;
3032                 break;
3033         case ELDLM_LOCK_REPLACED:
3034                 result = -ESRCH;
3035                 break;
3036         case ELDLM_NO_LOCK_DATA:
3037                 result = -ENOENT;
3038                 break;
3039         case ELDLM_NAMESPACE_EXISTS:
3040                 result = -EEXIST;
3041                 break;
3042         case ELDLM_BAD_NAMESPACE:
3043                 result = -EBADF;
3044                 break;
3045         default:
3046                 if (((int)error) < 0) { /* cast to signed type */
3047                         result = error; /* as ldlm_error can be unsigned */
3048                 } else {
3049                         CERROR("Invalid DLM result code: %d\n", error);
3050                         result = -EPROTO;
3051                 }
3052         }
3053         return result;
3054 }
3055 EXPORT_SYMBOL(ldlm_error2errno);
3056
3057 /**
3058  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3059  */
3060 enum ldlm_error ldlm_errno2error(int err_no)
3061 {
3062         int error;
3063
3064         switch (err_no) {
3065         case 0:
3066                 error = ELDLM_OK;
3067                 break;
3068         case -ESTALE:
3069                 error = ELDLM_LOCK_CHANGED;
3070                 break;
3071         case -ENAVAIL:
3072                 error = ELDLM_LOCK_ABORTED;
3073                 break;
3074         case -ESRCH:
3075                 error = ELDLM_LOCK_REPLACED;
3076                 break;
3077         case -ENOENT:
3078                 error = ELDLM_NO_LOCK_DATA;
3079                 break;
3080         case -EEXIST:
3081                 error = ELDLM_NAMESPACE_EXISTS;
3082                 break;
3083         case -EBADF:
3084                 error = ELDLM_BAD_NAMESPACE;
3085                 break;
3086         default:
3087                 error = err_no;
3088         }
3089         return error;
3090 }
3091
3092 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3093 void ldlm_dump_export_locks(struct obd_export *exp)
3094 {
3095         spin_lock(&exp->exp_locks_list_guard);
3096         if (!list_empty(&exp->exp_locks_list)) {
3097                 struct ldlm_lock *lock;
3098
3099                 CERROR("dumping locks for export %p,"
3100                        "ignore if the unmount doesn't hang\n", exp);
3101                 list_for_each_entry(lock, &exp->exp_locks_list,
3102                                         l_exp_refs_link)
3103                         LDLM_ERROR(lock, "lock:");
3104         }
3105         spin_unlock(&exp->exp_locks_list_guard);
3106 }
3107 #endif
3108
3109 #ifdef HAVE_SERVER_SUPPORT
3110 static int target_bulk_timeout(void *data)
3111 {
3112         ENTRY;
3113         /* We don't fail the connection here, because having the export
3114          * killed makes the (vital) call to commitrw very sad.
3115          */
3116         RETURN(1);
3117 }
3118
3119 static inline const char *bulk2type(struct ptlrpc_request *req)
3120 {
3121         if (req->rq_bulk_read)
3122                 return "READ";
3123         if (req->rq_bulk_write)
3124                 return "WRITE";
3125         return "UNKNOWN";
3126 }
3127
3128 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3129                    struct l_wait_info *lwi)
3130 {
3131         struct ptlrpc_request   *req = desc->bd_req;
3132         time_t                   start = cfs_time_current_sec();
3133         time_t                   deadline;
3134         int                      rc = 0;
3135
3136         ENTRY;
3137
3138         /* If there is eviction in progress, wait for it to finish. */
3139         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3140                 *lwi = LWI_INTR(NULL, NULL);
3141                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3142                                   !atomic_read(&exp->exp_obd->
3143                                                    obd_evict_inprogress),
3144                                   lwi);
3145         }
3146
3147         /* Check if client was evicted or reconnected already. */
3148         if (exp->exp_failed ||
3149             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3150                 rc = -ENOTCONN;
3151         } else {
3152                 if (req->rq_bulk_read)
3153                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3154
3155                 if ((exp->exp_connect_data.ocd_connect_flags &
3156                      OBD_CONNECT_BULK_MBITS) != 0)
3157                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3158                 else /* old version, bulk matchbits is rq_xid */
3159                         req->rq_mbits = req->rq_xid;
3160
3161                 if (rc == 0)
3162                         rc = ptlrpc_start_bulk_transfer(desc);
3163         }
3164
3165         if (rc < 0) {
3166                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3167                           bulk2type(req), rc);
3168                 RETURN(rc);
3169         }