Whamcloud - gitweb
LU-7990 clio: revise readahead to support 16MB IO
[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
179                 list_del(&imp_conn->oic_item);
180                 ptlrpc_connection_put(imp_conn->oic_conn);
181                 OBD_FREE(imp_conn, sizeof(*imp_conn));
182                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
183                        imp, imp->imp_obd->obd_name, uuid->uuid);
184                 rc = 0;
185                 break;
186         }
187 out:
188         spin_unlock(&imp->imp_lock);
189         if (rc == -ENOENT)
190                 CERROR("connection %s not found\n", uuid->uuid);
191         RETURN(rc);
192 }
193 EXPORT_SYMBOL(client_import_del_conn);
194
195 /**
196  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
197  * to find a conn uuid of \a imp which can reach \a peer.
198  */
199 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
200                             struct obd_uuid *uuid)
201 {
202         struct obd_import_conn *conn;
203         int rc = -ENOENT;
204         ENTRY;
205
206         spin_lock(&imp->imp_lock);
207         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
208                 /* Check if conn UUID does have this peer NID. */
209                 if (class_check_uuid(&conn->oic_uuid, peer)) {
210                         *uuid = conn->oic_uuid;
211                         rc = 0;
212                         break;
213                 }
214         }
215         spin_unlock(&imp->imp_lock);
216         RETURN(rc);
217 }
218 EXPORT_SYMBOL(client_import_find_conn);
219
220 void client_destroy_import(struct obd_import *imp)
221 {
222         /* Drop security policy instance after all RPCs have finished/aborted
223          * to let all busy contexts be released. */
224         class_import_get(imp);
225         class_destroy_import(imp);
226         sptlrpc_import_sec_put(imp);
227         class_import_put(imp);
228 }
229 EXPORT_SYMBOL(client_destroy_import);
230
231 /**
232  * Check whether or not the OSC is on MDT.
233  * In the config log,
234  * osc on MDT
235  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
236  * osc on client
237  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
238  *
239  **/
240 static int osc_on_mdt(char *obdname)
241 {
242         char *ptr;
243
244         ptr = strrchr(obdname, '-');
245         if (ptr == NULL)
246                 return 0;
247
248         if (strncmp(ptr + 1, "MDT", 3) == 0)
249                 return 1;
250
251         return 0;
252 }
253
254 /* Configure an RPC client OBD device.
255  *
256  * lcfg parameters:
257  * 1 - client UUID
258  * 2 - server UUID
259  * 3 - inactive-on-startup
260  */
261 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
262 {
263         struct client_obd *cli = &obddev->u.cli;
264         struct obd_import *imp;
265         struct obd_uuid server_uuid;
266         int rq_portal, rp_portal, connect_op;
267         char *name = obddev->obd_type->typ_name;
268         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
269         char *cli_name = lustre_cfg_buf(lcfg, 0);
270         int rc;
271         ENTRY;
272
273         /* In a more perfect world, we would hang a ptlrpc_client off of
274          * obd_type and just use the values from there. */
275         if (!strcmp(name, LUSTRE_OSC_NAME)) {
276                 rq_portal = OST_REQUEST_PORTAL;
277                 rp_portal = OSC_REPLY_PORTAL;
278                 connect_op = OST_CONNECT;
279                 cli->cl_sp_me = LUSTRE_SP_CLI;
280                 cli->cl_sp_to = LUSTRE_SP_OST;
281                 ns_type = LDLM_NS_TYPE_OSC;
282         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
283                    !strcmp(name, LUSTRE_LWP_NAME)) {
284                 rq_portal = MDS_REQUEST_PORTAL;
285                 rp_portal = MDC_REPLY_PORTAL;
286                 connect_op = MDS_CONNECT;
287                 if (is_lwp_on_ost(cli_name))
288                         cli->cl_sp_me = LUSTRE_SP_OST;
289                 else if (is_lwp_on_mdt(cli_name))
290                         cli->cl_sp_me = LUSTRE_SP_MDT;
291                 else
292                         cli->cl_sp_me = LUSTRE_SP_CLI;
293                 cli->cl_sp_to = LUSTRE_SP_MDT;
294                 ns_type = LDLM_NS_TYPE_MDC;
295         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
296                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
297                         /* OSP_on_MDT for other MDTs */
298                         connect_op = MDS_CONNECT;
299                         cli->cl_sp_to = LUSTRE_SP_MDT;
300                         ns_type = LDLM_NS_TYPE_MDC;
301                         rq_portal = OUT_PORTAL;
302                 } else {
303                         /* OSP on MDT for OST */
304                         connect_op = OST_CONNECT;
305                         cli->cl_sp_to = LUSTRE_SP_OST;
306                         ns_type = LDLM_NS_TYPE_OSC;
307                         rq_portal = OST_REQUEST_PORTAL;
308                 }
309                 rp_portal = OSC_REPLY_PORTAL;
310                 cli->cl_sp_me = LUSTRE_SP_MDT;
311         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
312                 rq_portal = MGS_REQUEST_PORTAL;
313                 rp_portal = MGC_REPLY_PORTAL;
314                 connect_op = MGS_CONNECT;
315                 cli->cl_sp_me = LUSTRE_SP_MGC;
316                 cli->cl_sp_to = LUSTRE_SP_MGS;
317                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
318                 ns_type = LDLM_NS_TYPE_MGC;
319         } else {
320                 CERROR("unknown client OBD type \"%s\", can't setup\n",
321                        name);
322                 RETURN(-EINVAL);
323         }
324
325         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
326                 CERROR("requires a TARGET UUID\n");
327                 RETURN(-EINVAL);
328         }
329
330         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
331                 CERROR("client UUID must be less than 38 characters\n");
332                 RETURN(-EINVAL);
333         }
334
335         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
336                 CERROR("setup requires a SERVER UUID\n");
337                 RETURN(-EINVAL);
338         }
339
340         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
341                 CERROR("target UUID must be less than 38 characters\n");
342                 RETURN(-EINVAL);
343         }
344
345         init_rwsem(&cli->cl_sem);
346         mutex_init(&cli->cl_mgc_mutex);
347         cli->cl_conn_count = 0;
348         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
349                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
350                      sizeof(server_uuid)));
351
352         cli->cl_dirty_pages = 0;
353         cli->cl_avail_grant = 0;
354         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
355         /* cl_dirty_max_pages may be changed at connect time in
356          * ptlrpc_connect_interpret(). */
357         client_adjust_max_dirty(cli);
358         INIT_LIST_HEAD(&cli->cl_cache_waiters);
359         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
360         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
361         INIT_LIST_HEAD(&cli->cl_loi_write_list);
362         INIT_LIST_HEAD(&cli->cl_loi_read_list);
363         spin_lock_init(&cli->cl_loi_list_lock);
364         atomic_set(&cli->cl_pending_w_pages, 0);
365         atomic_set(&cli->cl_pending_r_pages, 0);
366         cli->cl_r_in_flight = 0;
367         cli->cl_w_in_flight = 0;
368
369         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
370         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
371         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
372         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
373         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
374         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
375
376         /* lru for osc. */
377         INIT_LIST_HEAD(&cli->cl_lru_osc);
378         atomic_set(&cli->cl_lru_shrinkers, 0);
379         atomic_long_set(&cli->cl_lru_busy, 0);
380         atomic_long_set(&cli->cl_lru_in_list, 0);
381         INIT_LIST_HEAD(&cli->cl_lru_list);
382         spin_lock_init(&cli->cl_lru_list_lock);
383         atomic_long_set(&cli->cl_unstable_count, 0);
384         INIT_LIST_HEAD(&cli->cl_shrink_list);
385
386         init_waitqueue_head(&cli->cl_destroy_waitq);
387         atomic_set(&cli->cl_destroy_in_flight, 0);
388 #ifdef ENABLE_CHECKSUM
389         /* Turn on checksumming by default. */
390         cli->cl_checksum = 1;
391         /*
392          * The supported checksum types will be worked out at connect time
393          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
394          * through procfs.
395          */
396         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
397 #endif
398         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
399
400         /* Set it to possible maximum size. It may be reduced by ocd_brw_size
401          * from OFD after connecting. */
402         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
403
404         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
405          * it will be updated at OSC connection time. */
406         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
407
408         if (!strcmp(name, LUSTRE_MDC_NAME)) {
409                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
410         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
411                 cli->cl_max_rpcs_in_flight = 2;
412         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
413                 cli->cl_max_rpcs_in_flight = 3;
414         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
415                 cli->cl_max_rpcs_in_flight = 4;
416         } else {
417                 if (osc_on_mdt(obddev->obd_name))
418                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
419                 else
420                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
421         }
422
423         spin_lock_init(&cli->cl_mod_rpcs_lock);
424         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
425         cli->cl_max_mod_rpcs_in_flight = 0;
426         cli->cl_mod_rpcs_in_flight = 0;
427         cli->cl_close_rpcs_in_flight = 0;
428         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
429         cli->cl_mod_tag_bitmap = NULL;
430
431         if (connect_op == MDS_CONNECT) {
432                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
433                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
434                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
435                 if (cli->cl_mod_tag_bitmap == NULL)
436                         GOTO(err, rc = -ENOMEM);
437         }
438
439         rc = ldlm_get_ref();
440         if (rc) {
441                 CERROR("ldlm_get_ref failed: %d\n", rc);
442                 GOTO(err, rc);
443         }
444
445         ptlrpc_init_client(rq_portal, rp_portal, name,
446                            &obddev->obd_ldlm_client);
447
448         imp = class_new_import(obddev);
449         if (imp == NULL)
450                 GOTO(err_ldlm, rc = -ENOENT);
451         imp->imp_client = &obddev->obd_ldlm_client;
452         imp->imp_connect_op = connect_op;
453         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
454                LUSTRE_CFG_BUFLEN(lcfg, 1));
455         class_import_put(imp);
456
457         rc = client_import_add_conn(imp, &server_uuid, 1);
458         if (rc) {
459                 CERROR("can't add initial connection\n");
460                 GOTO(err_import, rc);
461         }
462
463         cli->cl_import = imp;
464         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
465         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
466
467         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
468                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
469                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
470                                name, obddev->obd_name,
471                                cli->cl_target_uuid.uuid);
472                         spin_lock(&imp->imp_lock);
473                         imp->imp_deactive = 1;
474                         spin_unlock(&imp->imp_lock);
475                 }
476         }
477
478         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
479                                                    LDLM_NAMESPACE_CLIENT,
480                                                    LDLM_NAMESPACE_GREEDY,
481                                                    ns_type);
482         if (obddev->obd_namespace == NULL) {
483                 CERROR("Unable to create client namespace - %s\n",
484                        obddev->obd_name);
485                 GOTO(err_import, rc = -ENOMEM);
486         }
487
488         RETURN(rc);
489
490 err_import:
491         class_destroy_import(imp);
492 err_ldlm:
493         ldlm_put_ref();
494 err:
495         if (cli->cl_mod_tag_bitmap != NULL)
496                 OBD_FREE(cli->cl_mod_tag_bitmap,
497                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
498         cli->cl_mod_tag_bitmap = NULL;
499         RETURN(rc);
500
501 }
502 EXPORT_SYMBOL(client_obd_setup);
503
504 int client_obd_cleanup(struct obd_device *obddev)
505 {
506         struct client_obd *cli = &obddev->u.cli;
507         ENTRY;
508
509         ldlm_namespace_free_post(obddev->obd_namespace);
510         obddev->obd_namespace = NULL;
511
512         obd_cleanup_client_import(obddev);
513         LASSERT(obddev->u.cli.cl_import == NULL);
514
515         ldlm_put_ref();
516
517         if (cli->cl_mod_tag_bitmap != NULL)
518                 OBD_FREE(cli->cl_mod_tag_bitmap,
519                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
520         cli->cl_mod_tag_bitmap = NULL;
521
522         RETURN(0);
523 }
524 EXPORT_SYMBOL(client_obd_cleanup);
525
526 /* ->o_connect() method for client side (OSC and MDC and MGC) */
527 int client_connect_import(const struct lu_env *env,
528                           struct obd_export **exp,
529                           struct obd_device *obd, struct obd_uuid *cluuid,
530                           struct obd_connect_data *data, void *localdata)
531 {
532         struct client_obd       *cli    = &obd->u.cli;
533         struct obd_import       *imp    = cli->cl_import;
534         struct obd_connect_data *ocd;
535         struct lustre_handle    conn    = { 0 };
536         int                     rc;
537         bool                    is_mdc = false;
538         ENTRY;
539
540         *exp = NULL;
541         down_write(&cli->cl_sem);
542         if (cli->cl_conn_count > 0)
543                 GOTO(out_sem, rc = -EALREADY);
544
545         rc = class_connect(&conn, obd, cluuid);
546         if (rc)
547                 GOTO(out_sem, rc);
548
549         cli->cl_conn_count++;
550         *exp = class_conn2export(&conn);
551
552         LASSERT(obd->obd_namespace);
553
554         imp->imp_dlm_handle = conn;
555         rc = ptlrpc_init_import(imp);
556         if (rc != 0)
557                 GOTO(out_ldlm, rc);
558
559         ocd = &imp->imp_connect_data;
560         if (data) {
561                 *ocd = *data;
562                 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
563                                  LUSTRE_MDC_NAME, 3) == 0;
564                 if (is_mdc)
565                         data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
566                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
567         }
568
569         rc = ptlrpc_connect_import(imp);
570         if (rc != 0) {
571                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
572                 GOTO(out_ldlm, rc);
573         }
574         LASSERT(*exp != NULL && (*exp)->exp_connection);
575
576         if (data) {
577                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
578                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
579                          data->ocd_connect_flags, ocd->ocd_connect_flags);
580                 data->ocd_connect_flags = ocd->ocd_connect_flags;
581                 /* clear the flag as it was not set and is not known
582                  * by upper layers */
583                 if (is_mdc)
584                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
585         }
586
587         ptlrpc_pinger_add_import(imp);
588
589         EXIT;
590
591         if (rc) {
592 out_ldlm:
593                 cli->cl_conn_count--;
594                 class_disconnect(*exp);
595                 *exp = NULL;
596         }
597 out_sem:
598         up_write(&cli->cl_sem);
599
600         return rc;
601 }
602 EXPORT_SYMBOL(client_connect_import);
603
604 int client_disconnect_export(struct obd_export *exp)
605 {
606         struct obd_device *obd = class_exp2obd(exp);
607         struct client_obd *cli;
608         struct obd_import *imp;
609         int rc = 0, err;
610         ENTRY;
611
612         if (!obd) {
613                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
614                        exp, exp ? exp->exp_handle.h_cookie : -1);
615                 RETURN(-EINVAL);
616         }
617
618         cli = &obd->u.cli;
619         imp = cli->cl_import;
620
621         down_write(&cli->cl_sem);
622         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
623                 cli->cl_conn_count);
624
625         if (cli->cl_conn_count == 0) {
626                 CERROR("disconnecting disconnected device (%s)\n",
627                        obd->obd_name);
628                 GOTO(out_disconnect, rc = -EINVAL);
629         }
630
631         cli->cl_conn_count--;
632         if (cli->cl_conn_count != 0)
633                 GOTO(out_disconnect, rc = 0);
634
635         /* Mark import deactivated now, so we don't try to reconnect if any
636          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
637          * fully deactivate the import, or that would drop all requests. */
638         spin_lock(&imp->imp_lock);
639         imp->imp_deactive = 1;
640         spin_unlock(&imp->imp_lock);
641
642         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
643          * delete it regardless.  (It's safe to delete an import that was
644          * never added.) */
645         (void)ptlrpc_pinger_del_import(imp);
646
647         if (obd->obd_namespace != NULL) {
648                 /* obd_force == local only */
649                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
650                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
651                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
652         }
653
654         /* There's no need to hold sem while disconnecting an import,
655          * and it may actually cause deadlock in GSS. */
656         up_write(&cli->cl_sem);
657         rc = ptlrpc_disconnect_import(imp, 0);
658         down_write(&cli->cl_sem);
659
660         ptlrpc_invalidate_import(imp);
661
662         EXIT;
663
664 out_disconnect:
665         /* Use server style - class_disconnect should be always called for
666          * o_disconnect. */
667         err = class_disconnect(exp);
668         if (!rc && err)
669                 rc = err;
670
671         up_write(&cli->cl_sem);
672
673         RETURN(rc);
674 }
675 EXPORT_SYMBOL(client_disconnect_export);
676
677 #ifdef HAVE_SERVER_SUPPORT
678 int server_disconnect_export(struct obd_export *exp)
679 {
680         int rc;
681         ENTRY;
682
683         /* Disconnect early so that clients can't keep using export. */
684         rc = class_disconnect(exp);
685         /* Close import to avoid sending any requests. */
686         if (exp->exp_imp_reverse)
687                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
688
689         ldlm_bl_thread_wakeup();
690
691         /* complete all outstanding replies */
692         spin_lock(&exp->exp_lock);
693         while (!list_empty(&exp->exp_outstanding_replies)) {
694                 struct ptlrpc_reply_state *rs =
695                         list_entry(exp->exp_outstanding_replies.next,
696                                        struct ptlrpc_reply_state, rs_exp_list);
697                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
698
699                 spin_lock(&svcpt->scp_rep_lock);
700
701                 list_del_init(&rs->rs_exp_list);
702                 spin_lock(&rs->rs_lock);
703                 ptlrpc_schedule_difficult_reply(rs);
704                 spin_unlock(&rs->rs_lock);
705
706                 spin_unlock(&svcpt->scp_rep_lock);
707         }
708         spin_unlock(&exp->exp_lock);
709
710         RETURN(rc);
711 }
712 EXPORT_SYMBOL(server_disconnect_export);
713
714 /* --------------------------------------------------------------------------
715  * from old lib/target.c
716  * -------------------------------------------------------------------------- */
717
718 static int target_handle_reconnect(struct lustre_handle *conn,
719                                    struct obd_export *exp,
720                                    struct obd_uuid *cluuid)
721 {
722         struct obd_device *target;
723         struct lustre_handle *hdl;
724         cfs_time_t now;
725         cfs_time_t deadline;
726         int timeout;
727         int rc = 0;
728         ENTRY;
729
730         hdl = &exp->exp_imp_reverse->imp_remote_handle;
731         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
732                 conn->cookie = exp->exp_handle.h_cookie;
733                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
734                        " cookie "LPX64"\n", cluuid->uuid, exp, conn->cookie);
735                 RETURN(0);
736         }
737
738         target = exp->exp_obd;
739
740         /* Might be a re-connect after a partition. */
741         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
742                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
743                               "with handle "LPX64". Rejecting client "
744                               "with the same UUID trying to reconnect "
745                               "with handle "LPX64"\n", target->obd_name,
746                               obd_uuid2str(&exp->exp_client_uuid),
747                               obd_export_nid2str(exp),
748                               hdl->cookie, conn->cookie);
749                 memset(conn, 0, sizeof *conn);
750                 /* target_handle_connect() treats EALREADY and
751                  * -EALREADY differently.  -EALREADY is an error
752                  * (same UUID, different handle). */
753                 RETURN(-EALREADY);
754         }
755
756         if (!target->obd_recovering) {
757                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
758                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
759                         obd_export_nid2str(exp));
760                 GOTO(out_already, rc);
761         }
762
763         now = cfs_time_current();
764         deadline = cfs_timer_deadline(&target->obd_recovery_timer);
765         if (cfs_time_before(now, deadline)) {
766                 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
767                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting,"
768                         " waiting for %d clients in recovery for"
769                         " %d:%.02d\n", target->obd_name,
770                         obd_uuid2str(&exp->exp_client_uuid),
771                         obd_export_nid2str(exp),
772                         target->obd_max_recoverable_clients,
773                         timeout / 60, timeout % 60);
774         } else {
775                 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
776                 LCONSOLE_WARN("%s: Recovery already passed deadline"
777                         " %d:%.02d, It is most likely due to DNE"
778                         " recovery is failed or stuck, please wait a"
779                         " few more minutes or abort the recovery.\n",
780                         target->obd_name, timeout / 60, timeout % 60);
781         }
782
783 out_already:
784         conn->cookie = exp->exp_handle.h_cookie;
785         /* target_handle_connect() treats EALREADY and
786          * -EALREADY differently.  EALREADY means we are
787          * doing a valid reconnect from the same client. */
788         RETURN(EALREADY);
789 }
790
791 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
792                           int error)
793 {
794         struct obd_export *exp = cb_data;
795
796         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
797                obd->obd_name, exp->exp_client_uuid.uuid);
798
799         spin_lock(&exp->exp_lock);
800         exp->exp_need_sync = 0;
801         spin_unlock(&exp->exp_lock);
802         class_export_cb_put(exp);
803 }
804
805 static void
806 check_and_start_recovery_timer(struct obd_device *obd,
807                                struct ptlrpc_request *req, int new_client);
808
809 /**
810  * update flags for import during reconnect process
811  */
812 static int rev_import_flags_update(struct obd_import *revimp,
813                                    struct ptlrpc_request *req)
814 {
815         int rc;
816         struct obd_connect_data *data;
817
818         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
819
820         if (data->ocd_connect_flags & OBD_CONNECT_AT)
821                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
822         else
823                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
824
825         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
826
827         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
828         if (rc) {
829                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
830                         revimp->imp_client->cli_name,
831                         libcfs_id2str(req->rq_peer), rc);
832                 return rc;
833         }
834
835         return 0;
836 }
837
838 /**
839  * Allocate a new reverse import for an export.
840  *
841  * \retval -errno in case error hit
842  * \retval 0 if reverse import correctly init
843  **/
844 int rev_import_init(struct obd_export *export)
845 {
846         struct obd_device *obd = export->exp_obd;
847         struct obd_import *revimp;
848
849         LASSERT(export->exp_imp_reverse == NULL);
850
851         revimp = class_new_import(obd);
852         if (revimp == NULL)
853                 return -ENOMEM;
854
855         revimp->imp_remote_handle.cookie = 0ULL;
856         revimp->imp_client = &obd->obd_ldlm_client;
857         revimp->imp_dlm_fake = 1;
858
859         /* it is safe to connect import in new state as no sends possible */
860         spin_lock(&export->exp_lock);
861         export->exp_imp_reverse = revimp;
862         spin_unlock(&export->exp_lock);
863         class_import_put(revimp);
864
865         return 0;
866 }
867 EXPORT_SYMBOL(rev_import_init);
868
869 /**
870  * Handle reconnect for an export.
871  *
872  * \param exp export to handle reconnect process
873  * \param req client reconnect request
874  *
875  * \retval -rc in case securitfy flavor can't be changed
876  * \retval 0 in case none problems
877  */
878 static int rev_import_reconnect(struct obd_export *exp,
879                                 struct ptlrpc_request *req)
880 {
881         struct obd_import *revimp = exp->exp_imp_reverse;
882         struct lustre_handle *lh;
883         int rc;
884
885         /* avoid sending a request until import flags are changed */
886         ptlrpc_import_enter_resend(revimp);
887
888         if (revimp->imp_connection != NULL)
889                 ptlrpc_connection_put(revimp->imp_connection);
890
891         /*
892          * client from recovery don't have a handle so we need to take from
893          * request. it may produce situation when wrong client connected
894          * to recovery as we trust a client uuid
895          */
896         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
897         revimp->imp_remote_handle = *lh;
898
899         /* unknown versions will be caught in
900          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
901         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
902
903         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
904
905         rc = rev_import_flags_update(revimp, req);
906         if (rc != 0) {
907                 /* it is safe to still be in RECOVERY phase as we are not able
908                  * to setup correct security flavor so requests are not able to
909                  * be delivered correctly */
910                 return rc;
911         }
912
913         /* resend all rpc's via new connection */
914         return ptlrpc_import_recovery_state_machine(revimp);
915 }
916
917 int target_handle_connect(struct ptlrpc_request *req)
918 {
919         struct obd_device *target = NULL;
920         struct obd_export *export = NULL;
921         /* connect handle - filled from target_handle_reconnect in
922          * reconnect case */
923         struct lustre_handle conn;
924         struct lustre_handle *tmp;
925         struct obd_uuid tgtuuid;
926         struct obd_uuid cluuid;
927         char *str;
928         int rc = 0;
929         char *target_start;
930         int target_len;
931         bool     mds_conn = false, lw_client = false;
932         bool     mds_mds_conn = false;
933         bool     new_mds_mds_conn = false;
934         bool     target_referenced = false;
935         struct obd_connect_data *data, *tmpdata;
936         int size, tmpsize;
937         lnet_nid_t *client_nid = NULL;
938         ENTRY;
939
940         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
941
942         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
943         if (str == NULL) {
944                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
945                 GOTO(out, rc = -EINVAL);
946         }
947
948         obd_str2uuid(&tgtuuid, str);
949         target = class_uuid2obd(&tgtuuid);
950         if (!target)
951                 target = class_name2obd(str);
952
953         if (!target) {
954                 deuuidify(str, NULL, &target_start, &target_len);
955                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
956                                    "from %s (no target). If you are running "
957                                    "an HA pair check that the target is "
958                                    "mounted on the other server.\n", str,
959                                    libcfs_nid2str(req->rq_peer.nid));
960                 GOTO(out, rc = -ENODEV);
961         }
962
963         spin_lock(&target->obd_dev_lock);
964         if (target->obd_stopping || !target->obd_set_up) {
965                 spin_unlock(&target->obd_dev_lock);
966
967                 deuuidify(str, NULL, &target_start, &target_len);
968                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
969                               target_len, target_start,
970                               libcfs_nid2str(req->rq_peer.nid),
971                               (target->obd_stopping ?
972                                "stopping" : "not set up"));
973                 GOTO(out, rc = -ENODEV);
974         }
975
976         if (target->obd_no_conn) {
977                 spin_unlock(&target->obd_dev_lock);
978
979                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
980                                "from %s\n", target->obd_name,
981                                libcfs_nid2str(req->rq_peer.nid));
982                 GOTO(out, rc = -EAGAIN);
983         }
984
985         /* Make sure the target isn't cleaned up while we're here. Yes,
986          * there's still a race between the above check and our incref here.
987          * Really, class_uuid2obd should take the ref. */
988         class_incref(target, __func__, current);
989         target_referenced = true;
990
991         target->obd_conn_inprogress++;
992         spin_unlock(&target->obd_dev_lock);
993
994         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
995         if (str == NULL) {
996                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
997                 GOTO(out, rc = -EINVAL);
998         }
999
1000         obd_str2uuid(&cluuid, str);
1001
1002         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1003         if (tmp == NULL)
1004                 GOTO(out, rc = -EPROTO);
1005
1006         conn = *tmp;
1007
1008         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1009                                     RCL_CLIENT);
1010         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1011         if (!data)
1012                 GOTO(out, rc = -EPROTO);
1013
1014         rc = req_capsule_server_pack(&req->rq_pill);
1015         if (rc)
1016                 GOTO(out, rc);
1017
1018 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1019         /* Don't allow clients to connect that are using old 1.8 format
1020          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1021          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1022          * FULL20 flag should be set on all connections since 2.0, but no
1023          * longer affects behaviour.
1024          *
1025          * Later this check will be disabled and the flag can be retired
1026          * completely once interop with 3.0 is no longer needed.
1027          */
1028         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1029                 GOTO(out, rc = -EPROTO);
1030 #endif
1031
1032         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1033                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1034                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1035                     data->ocd_version > LUSTRE_VERSION_CODE +
1036                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1037                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1038                                   "libclient connection attempt",
1039                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1040                                   "old" : "new",
1041                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1042                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1043                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1044                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1045                         data = req_capsule_server_sized_get(&req->rq_pill,
1046                                                             &RMF_CONNECT_DATA,
1047                                     offsetof(typeof(*data), ocd_version) +
1048                                              sizeof(data->ocd_version));
1049                         if (data) {
1050                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1051                                 data->ocd_version = LUSTRE_VERSION_CODE;
1052                         }
1053                         GOTO(out, rc = -EPROTO);
1054                 }
1055         }
1056
1057         /* Note: lw_client is needed in MDS-MDS failover during update log
1058          * processing, so we needs to allow lw_client to be connected at
1059          * anytime, instead of only the initial connection */
1060         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1061
1062         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1063                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1064                 mds_mds_conn = (data->ocd_connect_flags &
1065                                 OBD_CONNECT_MDS_MDS) != 0;
1066
1067                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1068                  * for Imperative Recovery connection from MGC to MGS.
1069                  *
1070                  * Via check OBD_CONNECT_FID, we can distinguish whether
1071                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1072                  * MGC or MDT. */
1073                 if (!lw_client &&
1074                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1075                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1076                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1077                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1078                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1079                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1080
1081                         /* We do not support the MDT-MDT interoperations with
1082                          * different version MDT because of protocol changes. */
1083                         if (unlikely(major != LUSTRE_MAJOR ||
1084                                      minor != LUSTRE_MINOR ||
1085                                      abs(patch - LUSTRE_PATCH) > 3)) {
1086                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1087                                         "connection from different version MDT "
1088                                         "(%d.%d.%d.%d) %s %s\n",
1089                                         target->obd_name, LUSTRE_MAJOR,
1090                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1091                                         major, minor, patch,
1092                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1093                                         libcfs_nid2str(req->rq_peer.nid), str);
1094
1095                                 GOTO(out, rc = -EPROTO);
1096                         }
1097                 }
1098         }
1099
1100         /* lctl gets a backstage, all-access pass. */
1101         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1102                 goto dont_check_exports;
1103
1104         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1105         if (!export)
1106                 goto no_export;
1107
1108         /* We've found an export in the hash. */
1109
1110         spin_lock(&export->exp_lock);
1111
1112         if (export->exp_connecting) { /* bug 9635, et. al. */
1113                 spin_unlock(&export->exp_lock);
1114                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1115                               export->exp_obd->obd_name, export,
1116                               libcfs_nid2str(req->rq_peer.nid));
1117                 class_export_put(export);
1118                 export = NULL;
1119                 rc = -EALREADY;
1120         } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
1121                 spin_unlock(&export->exp_lock);
1122                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
1123                         /* MDS or LWP reconnected after failover. */
1124                         LCONSOLE_WARN("%s: Received %s connection from "
1125                             "%s, removing former export from %s\n",
1126                             target->obd_name, mds_conn ? "MDS" : "LWP",
1127                             libcfs_nid2str(req->rq_peer.nid),
1128                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1129                 else
1130                         /* New MDS connection from the same NID. */
1131                         LCONSOLE_WARN("%s: Received new %s connection from "
1132                                 "%s, removing former export from same NID\n",
1133                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1134                                 libcfs_nid2str(req->rq_peer.nid));
1135                 class_fail_export(export);
1136                 class_export_put(export);
1137                 export = NULL;
1138                 rc = 0;
1139         } else if (export->exp_connection != NULL &&
1140                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1141                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
1142                     MSG_CONNECT_INITIAL)) {
1143                 spin_unlock(&export->exp_lock);
1144                 /* In MDS failover we have static UUID but NID can change. */
1145                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1146                               "existing nid %s is already connected\n",
1147                               target->obd_name, cluuid.uuid,
1148                               libcfs_nid2str(req->rq_peer.nid),
1149                               libcfs_nid2str(
1150                                       export->exp_connection->c_peer.nid));
1151                 rc = -EALREADY;
1152                 class_export_put(export);
1153                 export = NULL;
1154         } else {
1155                 export->exp_connecting = 1;
1156                 spin_unlock(&export->exp_lock);
1157                 LASSERT(export->exp_obd == target);
1158
1159                 rc = target_handle_reconnect(&conn, export, &cluuid);
1160         }
1161
1162         /* If we found an export, we already unlocked. */
1163         if (!export) {
1164 no_export:
1165                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1166         } else if (req->rq_export == NULL &&
1167                    atomic_read(&export->exp_rpc_count) > 0) {
1168                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1169                               "still busy with %d references\n",
1170                               target->obd_name, cluuid.uuid,
1171                               libcfs_nid2str(req->rq_peer.nid),
1172                               atomic_read(&export->exp_refcount));
1173                 GOTO(out, rc = -EBUSY);
1174         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1175                 if (!strstr(cluuid.uuid, "mdt"))
1176                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1177                                       "known client %s (at %s) because it "
1178                                       "is indicating it is a new client",
1179                                       target->obd_name, cluuid.uuid,
1180                                       libcfs_nid2str(req->rq_peer.nid));
1181                 GOTO(out, rc = -EALREADY);
1182         } else {
1183                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1184         }
1185
1186         if (rc < 0) {
1187                 GOTO(out, rc);
1188         }
1189
1190         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1191                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1192               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1193               export, (long)cfs_time_current_sec(),
1194               export ? (long)export->exp_last_request_time : 0);
1195
1196         /* If this is the first time a client connects, reset the recovery
1197          * timer. Discard lightweight connections which might be local. */
1198         if (!lw_client && rc == 0 && target->obd_recovering)
1199                 check_and_start_recovery_timer(target, req, export == NULL);
1200
1201         /* We want to handle EALREADY but *not* -EALREADY from
1202          * target_handle_reconnect(), return reconnection state in a flag. */
1203         if (rc == EALREADY) {
1204                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1205                 rc = 0;
1206         } else {
1207                 LASSERT(rc == 0);
1208         }
1209
1210         /* Tell the client if we support replayable requests. */
1211         if (target->obd_replayable)
1212                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1213         client_nid = &req->rq_peer.nid;
1214
1215         if (export == NULL) {
1216                 /* allow lightweight connections during recovery */
1217                 /* allow "new" MDT to be connected during recovery, since we
1218                  * need retrieve recovery update records from it */
1219                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1220                         cfs_time_t t;
1221                         int     c; /* connected */
1222                         int     i; /* in progress */
1223                         int     k; /* known */
1224                         int     s; /* stale/evicted */
1225
1226                         c = atomic_read(&target->obd_connected_clients);
1227                         i = atomic_read(&target->obd_lock_replay_clients);
1228                         k = target->obd_max_recoverable_clients;
1229                         s = target->obd_stale_clients;
1230                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1231                         t = cfs_time_sub(t, cfs_time_current());
1232                         t = cfs_duration_sec(t);
1233                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1234                                       "(at %s), waiting for %d known clients "
1235                                       "(%d recovered, %d in progress, and %d "
1236                                       "evicted) to recover in %d:%.02d\n",
1237                                       target->obd_name, cluuid.uuid,
1238                                       libcfs_nid2str(req->rq_peer.nid), k,
1239                                       c - i, i, s, (int)t / 60,
1240                                       (int)t % 60);
1241                         rc = -EBUSY;
1242                 } else {
1243 dont_check_exports:
1244                         rc = obd_connect(req->rq_svc_thread->t_env,
1245                                          &export, target, &cluuid, data,
1246                                          client_nid);
1247                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1248                                 lustre_msg_add_op_flags(req->rq_repmsg,
1249                                                         MSG_CONNECT_RECOVERING);
1250                         if (rc == 0) {
1251                                 conn.cookie = export->exp_handle.h_cookie;
1252                                 rc = rev_import_init(export);
1253                         }
1254
1255                         if (mds_mds_conn)
1256                                 new_mds_mds_conn = true;
1257                 }
1258         } else {
1259                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1260                                    export, target, &cluuid, data, client_nid);
1261         }
1262         if (rc)
1263                 GOTO(out, rc);
1264
1265         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1266         data->ocd_instance = target->u.obt.obt_instance;
1267
1268         /* Return only the parts of obd_connect_data that we understand, so the
1269          * client knows that we don't understand the rest. */
1270         if (data) {
1271                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1272                                                RCL_SERVER);
1273                 tmpdata = req_capsule_server_get(&req->rq_pill,
1274                                                  &RMF_CONNECT_DATA);
1275                 /* Don't use struct assignment here, because the client reply
1276                  * buffer may be smaller/larger than the local struct
1277                  * obd_connect_data. */
1278                 memcpy(tmpdata, data, min(tmpsize, size));
1279         }
1280
1281         /* If the client and the server are the same node, we will already
1282          * have an export that really points to the client's DLM export,
1283          * because we have a shared handles table.
1284          *
1285          * XXX this will go away when shaver stops sending the "connect" handle
1286          * in the real "remote handle" field of the request --phik 24 Apr 2003
1287          */
1288         ptlrpc_request_change_export(req, export);
1289
1290         spin_lock(&export->exp_lock);
1291         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1292                 spin_unlock(&export->exp_lock);
1293                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1294                        "or equal conn_cnt: %d >= %d\n",
1295                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1296                        export->exp_conn_cnt,
1297                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1298
1299                 GOTO(out, rc = -EALREADY);
1300         }
1301         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1302         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1303
1304         /* Don't evict liblustre clients for not pinging. */
1305         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1306                 export->exp_libclient = 1;
1307                 spin_unlock(&export->exp_lock);
1308
1309                 spin_lock(&target->obd_dev_lock);
1310                 list_del_init(&export->exp_obd_chain_timed);
1311                 spin_unlock(&target->obd_dev_lock);
1312         } else {
1313                 spin_unlock(&export->exp_lock);
1314         }
1315
1316         if (export->exp_connection != NULL) {
1317                 /* Check to see if connection came from another NID. */
1318                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1319                     !hlist_unhashed(&export->exp_nid_hash))
1320                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1321                                      &export->exp_connection->c_peer.nid,
1322                                      &export->exp_nid_hash);
1323
1324                 ptlrpc_connection_put(export->exp_connection);
1325         }
1326
1327         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1328                                                        req->rq_self,
1329                                                        &cluuid);
1330         if (hlist_unhashed(&export->exp_nid_hash))
1331                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1332                              &export->exp_connection->c_peer.nid,
1333                              &export->exp_nid_hash);
1334
1335         lustre_msg_set_handle(req->rq_repmsg, &conn);
1336
1337         rc = rev_import_reconnect(export, req);
1338         if (rc != 0)
1339                 GOTO(out, rc);
1340
1341         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1342                 int has_transno;
1343                 __u64 transno = data->ocd_transno;
1344
1345                 spin_lock(&export->exp_lock);
1346                 /* possible race with class_disconnect_stale_exports,
1347                  * export may be already in the eviction process */
1348                 if (export->exp_failed) {
1349                         spin_unlock(&export->exp_lock);
1350                         GOTO(out, rc = -ENODEV);
1351                 }
1352                 export->exp_in_recovery = 1;
1353                 export->exp_req_replay_needed = 1;
1354                 export->exp_lock_replay_needed = 1;
1355                 spin_unlock(&export->exp_lock);
1356
1357                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1358                                  MSG_CONNECT_TRANSNO);
1359                 if (has_transno && transno == 0)
1360                         CWARN("Connect with zero transno!\n");
1361
1362                 if (has_transno && transno > 0 &&
1363                     transno < target->obd_next_recovery_transno &&
1364                     transno > target->obd_last_committed) {
1365                         /* Another way is to use cmpxchg() to be lock-free. */
1366                         spin_lock(&target->obd_recovery_task_lock);
1367                         if (transno < target->obd_next_recovery_transno)
1368                                 target->obd_next_recovery_transno = transno;
1369                         spin_unlock(&target->obd_recovery_task_lock);
1370                 }
1371
1372                 atomic_inc(&target->obd_req_replay_clients);
1373                 atomic_inc(&target->obd_lock_replay_clients);
1374                 /* Note: MDS-MDS connection is allowed to be connected during
1375                  * recovery, no matter if the exports needs to be recoveried.
1376                  * Because we need retrieve updates logs from all other MDTs.
1377                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1378                  * also needs to be increased to match other recovery checking
1379                  * condition. */
1380                 if (new_mds_mds_conn)
1381                         target->obd_max_recoverable_clients++;
1382                 if (atomic_inc_return(&target->obd_connected_clients) ==
1383                     target->obd_max_recoverable_clients)
1384                         wake_up(&target->obd_next_transno_waitq);
1385         }
1386
1387         /* Tell the client we're in recovery, when client is involved in it. */
1388         if (target->obd_recovering && !lw_client)
1389                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1390
1391 out:
1392         if (export) {
1393                 spin_lock(&export->exp_lock);
1394                 export->exp_connecting = 0;
1395                 spin_unlock(&export->exp_lock);
1396
1397                 class_export_put(export);
1398         }
1399         if (target_referenced == true && target != NULL) {
1400                 spin_lock(&target->obd_dev_lock);
1401                 target->obd_conn_inprogress--;
1402                 spin_unlock(&target->obd_dev_lock);
1403
1404                 class_decref(target, __func__, current);
1405         }
1406         req->rq_status = rc;
1407         RETURN(rc);
1408 }
1409
1410 int target_handle_disconnect(struct ptlrpc_request *req)
1411 {
1412         int rc;
1413         ENTRY;
1414
1415         rc = req_capsule_server_pack(&req->rq_pill);
1416         if (rc)
1417                 RETURN(rc);
1418
1419         /* Keep the rq_export around so we can send the reply. */
1420         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1421
1422         RETURN(0);
1423 }
1424
1425 void target_destroy_export(struct obd_export *exp)
1426 {
1427         struct obd_import       *imp = NULL;
1428         /* exports created from last_rcvd data, and "fake"
1429            exports created by lctl don't have an import */
1430         spin_lock(&exp->exp_lock);
1431         if (exp->exp_imp_reverse != NULL) {
1432                 imp = exp->exp_imp_reverse;
1433                 exp->exp_imp_reverse = NULL;
1434         }
1435         spin_unlock(&exp->exp_lock);
1436         if (imp != NULL)
1437                 client_destroy_import(imp);
1438
1439         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1440         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1441         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1442         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1443 }
1444 EXPORT_SYMBOL(target_destroy_export);
1445
1446 /*
1447  * Recovery functions
1448  */
1449 static void target_request_copy_get(struct ptlrpc_request *req)
1450 {
1451         class_export_rpc_inc(req->rq_export);
1452         LASSERT(list_empty(&req->rq_list));
1453         INIT_LIST_HEAD(&req->rq_replay_list);
1454
1455         /* Increase refcount to keep request in queue. */
1456         atomic_inc(&req->rq_refcount);
1457         /* Let export know it has replays to be handled. */
1458         atomic_inc(&req->rq_export->exp_replay_count);
1459 }
1460
1461 static void target_request_copy_put(struct ptlrpc_request *req)
1462 {
1463         LASSERT(list_empty(&req->rq_replay_list));
1464         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1465
1466         atomic_dec(&req->rq_export->exp_replay_count);
1467         class_export_rpc_dec(req->rq_export);
1468         ptlrpc_server_drop_request(req);
1469 }
1470
1471 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1472 {
1473         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1474         struct obd_export     *exp = req->rq_export;
1475         struct ptlrpc_request *reqiter;
1476         struct ptlrpc_request *dup_req = NULL;
1477         int                    dup = 0;
1478
1479         LASSERT(exp);
1480
1481         spin_lock(&exp->exp_lock);
1482         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1483                                 rq_replay_list) {
1484                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1485                         dup_req = reqiter;
1486                         dup = 1;
1487                         break;
1488                 }
1489         }
1490
1491         if (dup) {
1492                 /* We expect it with RESENT and REPLAY flags. */
1493                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1494                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1495                         CERROR("invalid flags %x of resent replay\n",
1496                                lustre_msg_get_flags(req->rq_reqmsg));
1497
1498                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1499                         __u32 new_conn;
1500
1501                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1502                         if (new_conn >
1503                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1504                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1505                                                         new_conn);
1506                 }
1507         } else {
1508                 list_add_tail(&req->rq_replay_list,
1509                                   &exp->exp_req_replay_queue);
1510         }
1511
1512         spin_unlock(&exp->exp_lock);
1513         return dup;
1514 }
1515
1516 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1517 {
1518         LASSERT(!list_empty(&req->rq_replay_list));
1519         LASSERT(req->rq_export);
1520
1521         spin_lock(&req->rq_export->exp_lock);
1522         list_del_init(&req->rq_replay_list);
1523         spin_unlock(&req->rq_export->exp_lock);
1524 }
1525
1526 static void target_finish_recovery(struct lu_target *lut)
1527 {
1528         struct obd_device *obd = lut->lut_obd;
1529         ENTRY;
1530
1531         /* Only log a recovery message when recovery has occurred. */
1532         if (obd->obd_recovery_start) {
1533                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1534                                         obd->obd_recovery_start);
1535                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1536                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1537                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1538                         obd->obd_max_recoverable_clients,
1539                         atomic_read(&obd->obd_connected_clients),
1540                         obd->obd_stale_clients,
1541                         obd->obd_stale_clients == 1 ? "was" : "were");
1542         }
1543
1544         ldlm_reprocess_all_ns(obd->obd_namespace);
1545         spin_lock(&obd->obd_recovery_task_lock);
1546         if (!list_empty(&obd->obd_req_replay_queue) ||
1547             !list_empty(&obd->obd_lock_replay_queue) ||
1548             !list_empty(&obd->obd_final_req_queue)) {
1549                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1550                        obd->obd_name,
1551                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1552                        list_empty(&obd->obd_lock_replay_queue) ? \
1553                                "" : "lock ",
1554                        list_empty(&obd->obd_final_req_queue) ? \
1555                                "" : "final ");
1556                 spin_unlock(&obd->obd_recovery_task_lock);
1557                 LBUG();
1558         }
1559         spin_unlock(&obd->obd_recovery_task_lock);
1560
1561         obd->obd_recovery_end = cfs_time_current_sec();
1562
1563         /* When recovery finished, cleanup orphans on MDS and OST. */
1564         if (OBT(obd) && OBP(obd, postrecov)) {
1565                 int rc = OBP(obd, postrecov)(obd);
1566                 if (rc < 0)
1567                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1568                                       obd->obd_name, rc);
1569         }
1570         EXIT;
1571 }
1572
1573 static void abort_req_replay_queue(struct obd_device *obd)
1574 {
1575         struct ptlrpc_request *req, *n;
1576         struct list_head abort_list;
1577
1578         INIT_LIST_HEAD(&abort_list);
1579         spin_lock(&obd->obd_recovery_task_lock);
1580         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1581         spin_unlock(&obd->obd_recovery_task_lock);
1582         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1583                 DEBUG_REQ(D_WARNING, req, "aborted:");
1584                 req->rq_status = -ENOTCONN;
1585                 if (ptlrpc_error(req)) {
1586                         DEBUG_REQ(D_ERROR, req,
1587                                   "failed abort_req_reply; skipping");
1588                 }
1589                 target_exp_dequeue_req_replay(req);
1590                 target_request_copy_put(req);
1591         }
1592 }
1593
1594 static void abort_lock_replay_queue(struct obd_device *obd)
1595 {
1596         struct ptlrpc_request *req, *n;
1597         struct list_head abort_list;
1598
1599         INIT_LIST_HEAD(&abort_list);
1600         spin_lock(&obd->obd_recovery_task_lock);
1601         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1602         spin_unlock(&obd->obd_recovery_task_lock);
1603         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1604                 DEBUG_REQ(D_ERROR, req, "aborted:");
1605                 req->rq_status = -ENOTCONN;
1606                 if (ptlrpc_error(req)) {
1607                         DEBUG_REQ(D_ERROR, req,
1608                                   "failed abort_lock_reply; skipping");
1609                 }
1610                 target_request_copy_put(req);
1611         }
1612 }
1613
1614 /* Called from a cleanup function if the device is being cleaned up
1615    forcefully.  The exports should all have been disconnected already,
1616    the only thing left to do is
1617      - clear the recovery flags
1618      - cancel the timer
1619      - free queued requests and replies, but don't send replies
1620    Because the obd_stopping flag is set, no new requests should be received.
1621
1622 */
1623 void target_cleanup_recovery(struct obd_device *obd)
1624 {
1625         struct ptlrpc_request *req, *n;
1626         struct list_head clean_list;
1627
1628         INIT_LIST_HEAD(&clean_list);
1629         spin_lock(&obd->obd_dev_lock);
1630         if (!obd->obd_recovering) {
1631                 spin_unlock(&obd->obd_dev_lock);
1632                 EXIT;
1633                 return;
1634         }
1635         obd->obd_recovering = obd->obd_abort_recovery = 0;
1636         spin_unlock(&obd->obd_dev_lock);
1637
1638         spin_lock(&obd->obd_recovery_task_lock);
1639         target_cancel_recovery_timer(obd);
1640         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1641         spin_unlock(&obd->obd_recovery_task_lock);
1642
1643         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1644                 LASSERT(req->rq_reply_state == NULL);
1645                 target_exp_dequeue_req_replay(req);
1646                 target_request_copy_put(req);
1647         }
1648
1649         spin_lock(&obd->obd_recovery_task_lock);
1650         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1651         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1652         spin_unlock(&obd->obd_recovery_task_lock);
1653
1654         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1655                 LASSERT(req->rq_reply_state == NULL);
1656                 target_request_copy_put(req);
1657         }
1658
1659         EXIT;
1660 }
1661 EXPORT_SYMBOL(target_cleanup_recovery);
1662
1663 /* obd_recovery_task_lock should be held */
1664 void target_cancel_recovery_timer(struct obd_device *obd)
1665 {
1666         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1667         cfs_timer_disarm(&obd->obd_recovery_timer);
1668 }
1669
1670 static void target_start_recovery_timer(struct obd_device *obd)
1671 {
1672         if (obd->obd_recovery_start != 0)
1673                 return;
1674
1675         spin_lock(&obd->obd_dev_lock);
1676         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1677                 spin_unlock(&obd->obd_dev_lock);
1678                 return;
1679         }
1680
1681         LASSERT(obd->obd_recovery_timeout != 0);
1682
1683         if (obd->obd_recovery_start != 0) {
1684                 spin_unlock(&obd->obd_dev_lock);
1685                 return;
1686         }
1687
1688         cfs_timer_arm(&obd->obd_recovery_timer,
1689                       cfs_time_shift(obd->obd_recovery_timeout));
1690         obd->obd_recovery_start = cfs_time_current_sec();
1691         spin_unlock(&obd->obd_dev_lock);
1692
1693         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1694                       "or until %d client%s reconnect%s\n",
1695                       obd->obd_name,
1696                       obd->obd_recovery_timeout / 60,
1697                       obd->obd_recovery_timeout % 60,
1698                       obd->obd_max_recoverable_clients,
1699                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1700                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1701 }
1702
1703 /**
1704  * extend recovery window.
1705  *
1706  * if @extend is true, extend recovery window to have @drt remaining at least;
1707  * otherwise, make sure the recovery timeout value is not less than @drt.
1708  */
1709 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1710 {
1711         cfs_time_t now;
1712         cfs_time_t end;
1713         cfs_duration_t left;
1714         int to;
1715
1716         spin_lock(&obd->obd_dev_lock);
1717         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1718                 spin_unlock(&obd->obd_dev_lock);
1719                 return;
1720         }
1721         LASSERT(obd->obd_recovery_start != 0);
1722
1723         now  = cfs_time_current_sec();
1724         to   = obd->obd_recovery_timeout;
1725         end  = obd->obd_recovery_start + to;
1726         left = cfs_time_sub(end, now);
1727
1728         if (extend && (drt > left)) {
1729                 to += drt - left;
1730         } else if (!extend && (drt > to)) {
1731                 to = drt;
1732         }
1733
1734         if (to > obd->obd_recovery_time_hard)
1735                 to = obd->obd_recovery_time_hard;
1736         if (obd->obd_recovery_timeout < to) {
1737                 obd->obd_recovery_timeout = to;
1738                 end = obd->obd_recovery_start + to;
1739                 cfs_timer_arm(&obd->obd_recovery_timer,
1740                                 cfs_time_shift(end - now));
1741         }
1742         spin_unlock(&obd->obd_dev_lock);
1743
1744         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1745                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1746 }
1747
1748 /* Reset the timer with each new client connection */
1749 /*
1750  * This timer is actually reconnect_timer, which is for making sure
1751  * the total recovery window is at least as big as my reconnect
1752  * attempt timing. So the initial recovery time_out will be set to
1753  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1754  * from client is bigger than this, then the recovery time_out will
1755  * be extended to make sure the client could be reconnected, in the
1756  * process, the timeout from the new client should be ignored.
1757  */
1758
1759 static void
1760 check_and_start_recovery_timer(struct obd_device *obd,
1761                                struct ptlrpc_request *req,
1762                                int new_client)
1763 {
1764         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1765         struct obd_device_target *obt = &obd->u.obt;
1766
1767         if (!new_client && service_time)
1768                 /* Teach server about old server's estimates, as first guess
1769                  * at how long new requests will take. */
1770                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1771                             service_time);
1772
1773         target_start_recovery_timer(obd);
1774
1775         /* Convert the service time to RPC timeout,
1776          * and reuse service_time to limit stack usage. */
1777         service_time = at_est2timeout(service_time);
1778
1779         /* We expect other clients to timeout within service_time, then try
1780          * to reconnect, then try the failover server.  The max delay between
1781          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1782         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1783
1784         LASSERT(obt->obt_magic == OBT_MAGIC);
1785         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1786         if (service_time > obd->obd_recovery_timeout && !new_client)
1787                 extend_recovery_timer(obd, service_time, false);
1788 }
1789
1790 /** Health checking routines */
1791 static inline int exp_connect_healthy(struct obd_export *exp)
1792 {
1793         return (exp->exp_in_recovery);
1794 }
1795
1796 /** if export done req_replay or has replay in queue */
1797 static inline int exp_req_replay_healthy(struct obd_export *exp)
1798 {
1799         return (!exp->exp_req_replay_needed ||
1800                 atomic_read(&exp->exp_replay_count) > 0);
1801 }
1802
1803
1804 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1805 {
1806         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1807                exp_req_replay_healthy(exp);
1808 }
1809
1810 /** if export done lock_replay or has replay in queue */
1811 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1812 {
1813         return (!exp->exp_lock_replay_needed ||
1814                 atomic_read(&exp->exp_replay_count) > 0);
1815 }
1816
1817 static inline int exp_vbr_healthy(struct obd_export *exp)
1818 {
1819         return (!exp->exp_vbr_failed);
1820 }
1821
1822 static inline int exp_finished(struct obd_export *exp)
1823 {
1824         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1825 }
1826
1827 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1828 {
1829         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1830                 exp_finished(exp);
1831 }
1832
1833 static int check_for_next_transno(struct lu_target *lut)
1834 {
1835         struct ptlrpc_request *req = NULL;
1836         struct obd_device *obd = lut->lut_obd;
1837         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1838         int wake_up = 0, connected, completed, queue_len;
1839         __u64 req_transno = 0;
1840         __u64 update_transno = 0;
1841         __u64 next_transno = 0;
1842         ENTRY;
1843
1844         spin_lock(&obd->obd_recovery_task_lock);
1845         if (!list_empty(&obd->obd_req_replay_queue)) {
1846                 req = list_entry(obd->obd_req_replay_queue.next,
1847                                      struct ptlrpc_request, rq_list);
1848                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1849         }
1850
1851         if (tdtd != NULL)
1852                 update_transno = distribute_txn_get_next_transno(tdtd);
1853
1854         connected = atomic_read(&obd->obd_connected_clients);
1855         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1856         queue_len = obd->obd_requests_queued_for_recovery;
1857         next_transno = obd->obd_next_recovery_transno;
1858
1859         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1860                "req_transno: "LPU64", next_transno: "LPU64"\n",
1861                obd->obd_max_recoverable_clients, connected, completed,
1862                queue_len, req_transno, next_transno);
1863
1864         if (obd->obd_abort_recovery) {
1865                 CDEBUG(D_HA, "waking for aborted recovery\n");
1866                 wake_up = 1;
1867         } else if (obd->obd_recovery_expired) {
1868                 CDEBUG(D_HA, "waking for expired recovery\n");
1869                 wake_up = 1;
1870         } else if (tdtd != NULL && req != NULL &&
1871                    is_req_replayed_by_update(req)) {
1872                 LASSERTF(req_transno < next_transno, "req_transno "LPU64
1873                          "next_transno"LPU64"\n", req_transno, next_transno);
1874                 CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n",
1875                        req_transno);
1876                 wake_up = 1;
1877         } else if (req_transno == next_transno ||
1878                    (update_transno != 0 && update_transno <= next_transno)) {
1879                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1880                 wake_up = 1;
1881         } else if (queue_len > 0 &&
1882                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1883                 int d_lvl = D_HA;
1884                 /** handle gaps occured due to lost reply or VBR */
1885                 LASSERTF(req_transno >= next_transno,
1886                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1887                          req_transno, next_transno);
1888                 if (req_transno > obd->obd_last_committed &&
1889                     !obd->obd_version_recov)
1890                         d_lvl = D_ERROR;
1891                 CDEBUG(d_lvl,
1892                        "%s: waking for gap in transno, VBR is %s (skip: "
1893                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1894                        ", next_update "LPD64" last_committed: "LPD64")\n",
1895                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1896                        next_transno, queue_len, completed, connected,
1897                        req_transno, update_transno, obd->obd_last_committed);
1898                 obd->obd_next_recovery_transno = req_transno;
1899                 wake_up = 1;
1900         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1901                 CDEBUG(D_HA, "waking for completed recovery\n");
1902                 wake_up = 1;
1903         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1904                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1905                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1906                 obd->obd_next_recovery_transno = req_transno;
1907                 wake_up = 1;
1908         }
1909         spin_unlock(&obd->obd_recovery_task_lock);
1910         return wake_up;
1911 }
1912
1913 static int check_for_next_lock(struct lu_target *lut)
1914 {
1915         struct obd_device *obd = lut->lut_obd;
1916         int wake_up = 0;
1917
1918         spin_lock(&obd->obd_recovery_task_lock);
1919         if (!list_empty(&obd->obd_lock_replay_queue)) {
1920                 CDEBUG(D_HA, "waking for next lock\n");
1921                 wake_up = 1;
1922         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1923                 CDEBUG(D_HA, "waking for completed lock replay\n");
1924                 wake_up = 1;
1925         } else if (obd->obd_abort_recovery) {
1926                 CDEBUG(D_HA, "waking for aborted recovery\n");
1927                 wake_up = 1;
1928         } else if (obd->obd_recovery_expired) {
1929                 CDEBUG(D_HA, "waking for expired recovery\n");
1930                 wake_up = 1;
1931         }
1932         spin_unlock(&obd->obd_recovery_task_lock);
1933
1934         return wake_up;
1935 }
1936
1937 /**
1938  * wait for recovery events,
1939  * check its status with help of check_routine
1940  * evict dead clients via health_check
1941  */
1942 static int target_recovery_overseer(struct lu_target *lut,
1943                                     int (*check_routine)(struct lu_target *),
1944                                     int (*health_check)(struct obd_export *))
1945 {
1946         struct obd_device       *obd = lut->lut_obd;
1947         struct target_distribute_txn_data *tdtd;
1948 repeat:
1949         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1950               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1951                 __u64 next_update_transno = 0;
1952
1953                 /* Only abort the recovery if there are no update recovery
1954                  * left in the queue */
1955                 spin_lock(&obd->obd_recovery_task_lock);
1956                 if (lut->lut_tdtd != NULL) {
1957                         next_update_transno =
1958                                 distribute_txn_get_next_transno(lut->lut_tdtd);
1959
1960                         tdtd = lut->lut_tdtd;
1961                         /* If next_update_transno == 0, it probably because
1962                          * updatelog retrieve threads did not get any records
1963                          * yet, let's wait those threads stopped */
1964                         if (next_update_transno == 0) {
1965                                 struct l_wait_info lwi = { 0 };
1966
1967                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1968                                        atomic_read(
1969                                        &tdtd->tdtd_recovery_threads_count) == 0,
1970                                        &lwi);
1971
1972                                 next_update_transno =
1973                                         distribute_txn_get_next_transno(
1974                                                                 lut->lut_tdtd);
1975                         }
1976                 }
1977
1978                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
1979                         obd->obd_next_recovery_transno = next_update_transno;
1980                         spin_unlock(&obd->obd_recovery_task_lock);
1981                         /* Disconnect unfinished exports from clients, and
1982                          * keep connection from MDT to make sure the update
1983                          * recovery will still keep trying until some one
1984                          * manually abort the recovery */
1985                         class_disconnect_stale_exports(obd,
1986                                                 exp_finished_or_from_mdt);
1987                         /* Abort all of replay and replay lock req from
1988                          * clients */
1989                         abort_req_replay_queue(obd);
1990                         abort_lock_replay_queue(obd);
1991                         CDEBUG(D_HA, "%s: there are still update replay ("LPX64
1992                                ")in the queue.\n", obd->obd_name,
1993                                next_update_transno);
1994                 } else {
1995                         obd->obd_abort_recovery = 1;
1996                         spin_unlock(&obd->obd_recovery_task_lock);
1997                         CWARN("%s recovery is aborted by hard timeout\n",
1998                               obd->obd_name);
1999                 }
2000         }
2001
2002         while (wait_event_timeout(obd->obd_next_transno_waitq,
2003                                   check_routine(lut),
2004                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2005                 /* wait indefinitely for event, but don't trigger watchdog */;
2006
2007         if (obd->obd_abort_recovery) {
2008                 CWARN("recovery is aborted, evict exports in recovery\n");
2009                 if (lut->lut_tdtd != NULL) {
2010                         struct l_wait_info lwi = { 0 };
2011
2012                         tdtd = lut->lut_tdtd;
2013                         /* Let's wait all of the update log recovery thread
2014                          * finished */
2015                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2016                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2017                              &lwi);
2018                         /* Then abort the update recovery list */
2019                         dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2020                         dtrq_list_destroy(lut->lut_tdtd);
2021                 }
2022
2023                 /** evict exports which didn't finish recovery yet */
2024                 class_disconnect_stale_exports(obd, exp_finished);
2025                 return 1;
2026         } else if (obd->obd_recovery_expired) {
2027                 obd->obd_recovery_expired = 0;
2028                 /** If some clients died being recovered, evict them */
2029                 LCONSOLE_WARN("%s: recovery is timed out, "
2030                               "evict stale exports\n", obd->obd_name);
2031                 /** evict cexports with no replay in queue, they are stalled */
2032                 class_disconnect_stale_exports(obd, health_check);
2033
2034                 /** continue with VBR */
2035                 spin_lock(&obd->obd_dev_lock);
2036                 obd->obd_version_recov = 1;
2037                 spin_unlock(&obd->obd_dev_lock);
2038                 /**
2039                  * reset timer, recovery will proceed with versions now,
2040                  * timeout is set just to handle reconnection delays
2041                  */
2042                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2043                 /** Wait for recovery events again, after evicting bad clients */
2044                 goto repeat;
2045         }
2046         return 0;
2047 }
2048
2049 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2050 {
2051         struct obd_device       *obd = lut->lut_obd;
2052         struct ptlrpc_request *req = NULL;
2053
2054         CDEBUG(D_HA, "Waiting for lock\n");
2055         if (target_recovery_overseer(lut, check_for_next_lock,
2056                                      exp_lock_replay_healthy))
2057                 abort_lock_replay_queue(obd);
2058
2059         spin_lock(&obd->obd_recovery_task_lock);
2060         if (!list_empty(&obd->obd_lock_replay_queue)) {
2061                 req = list_entry(obd->obd_lock_replay_queue.next,
2062                                      struct ptlrpc_request, rq_list);
2063                 list_del_init(&req->rq_list);
2064                 spin_unlock(&obd->obd_recovery_task_lock);
2065         } else {
2066                 spin_unlock(&obd->obd_recovery_task_lock);
2067                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2068                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2069                 /** evict exports failed VBR */
2070                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2071         }
2072         return req;
2073 }
2074
2075 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2076 {
2077         struct ptlrpc_request *req = NULL;
2078
2079         spin_lock(&obd->obd_recovery_task_lock);
2080         if (!list_empty(&obd->obd_final_req_queue)) {
2081                 req = list_entry(obd->obd_final_req_queue.next,
2082                                      struct ptlrpc_request, rq_list);
2083                 list_del_init(&req->rq_list);
2084                 spin_unlock(&obd->obd_recovery_task_lock);
2085                 if (req->rq_export->exp_in_recovery) {
2086                         spin_lock(&req->rq_export->exp_lock);
2087                         req->rq_export->exp_in_recovery = 0;
2088                         spin_unlock(&req->rq_export->exp_lock);
2089                 }
2090         } else {
2091                 spin_unlock(&obd->obd_recovery_task_lock);
2092         }
2093         return req;
2094 }
2095
2096 static void handle_recovery_req(struct ptlrpc_thread *thread,
2097                                 struct ptlrpc_request *req,
2098                                 svc_handler_t handler)
2099 {
2100         ENTRY;
2101
2102         /**
2103          * export can be evicted during recovery, no need to handle replays for
2104          * it after that, discard such request silently
2105          */
2106         if (req->rq_export->exp_disconnected)
2107                 RETURN_EXIT;
2108
2109         req->rq_session.lc_thread = thread;
2110         req->rq_svc_thread = thread;
2111         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2112
2113         /* thread context */
2114         lu_context_enter(&thread->t_env->le_ctx);
2115         (void)handler(req);
2116         lu_context_exit(&thread->t_env->le_ctx);
2117
2118         /* don't reset timer for final stage */
2119         if (!exp_finished(req->rq_export)) {
2120                 int to = obd_timeout;
2121
2122                 /**
2123                  * Add request timeout to the recovery time so next request from
2124                  * this client may come in recovery time
2125                  */
2126                 if (!AT_OFF) {
2127                         struct ptlrpc_service_part *svcpt;
2128
2129                         svcpt = req->rq_rqbd->rqbd_svcpt;
2130                         /* If the server sent early reply for this request,
2131                          * the client will recalculate the timeout according to
2132                          * current server estimate service time, so we will
2133                          * use the maxium timeout here for waiting the client
2134                          * sending the next req */
2135                         to = max((int)at_est2timeout(
2136                                  at_get(&svcpt->scp_at_estimate)),
2137                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2138                         /* Add 2 net_latency, one for balance rq_deadline
2139                          * (see ptl_send_rpc), one for resend the req to server,
2140                          * Note: client will pack net_latency in replay req
2141                          * (see ptlrpc_replay_req) */
2142                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2143                 }
2144                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2145         }
2146         EXIT;
2147 }
2148
2149 /** Checking routines for recovery */
2150 static int check_for_recovery_ready(struct lu_target *lut)
2151 {
2152         struct obd_device *obd = lut->lut_obd;
2153         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2154
2155         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2156                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2157                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2158                obd->obd_recovery_expired);
2159
2160         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2161                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2162                 if (clnts + obd->obd_stale_clients <
2163                     obd->obd_max_recoverable_clients)
2164                         return 0;
2165         }
2166
2167         if (lut->lut_tdtd != NULL) {
2168                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2169                     !obd->obd_abort_recovery) {
2170                         /* Let's extend recovery timer, in case the recovery
2171                          * timer expired, and some clients got evicted */
2172                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2173                                               true);
2174                         CDEBUG(D_HA, "%s update recovery is not ready,"
2175                                " extend recovery %d\n", obd->obd_name,
2176                                obd->obd_recovery_timeout);
2177                         return 0;
2178                 } else {
2179                         dtrq_list_dump(lut->lut_tdtd, D_HA);
2180                 }
2181         }
2182
2183         return 1;
2184 }
2185
2186 enum {
2187         REQUEST_RECOVERY = 1,
2188         UPDATE_RECOVERY = 2,
2189 };
2190
2191 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2192 {
2193         __u64 transno = 0;
2194
2195         if (!list_empty(&obd->obd_req_replay_queue)) {
2196                 struct ptlrpc_request *req;
2197
2198                 req = list_entry(obd->obd_req_replay_queue.next,
2199                                  struct ptlrpc_request, rq_list);
2200                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2201         }
2202
2203         return transno;
2204 }
2205
2206 static __u64 get_next_transno(struct lu_target *lut, int *type)
2207 {
2208         struct obd_device *obd = lut->lut_obd;
2209         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2210         __u64 transno = 0;
2211         __u64 update_transno;
2212         ENTRY;
2213
2214         transno = get_next_replay_req_transno(obd);
2215         if (type != NULL)
2216                 *type = REQUEST_RECOVERY;
2217
2218         if (tdtd == NULL)
2219                 RETURN(transno);
2220
2221         update_transno = distribute_txn_get_next_transno(tdtd);
2222         if (transno == 0 || (transno >= update_transno &&
2223                              update_transno != 0)) {
2224                 transno = update_transno;
2225                 if (type != NULL)
2226                         *type = UPDATE_RECOVERY;
2227         }
2228
2229         RETURN(transno);
2230 }
2231
2232 /**
2233  * drop duplicate replay request
2234  *
2235  * Because the operation has been replayed by update recovery, the request
2236  * with the same transno will be dropped and also notify the client to send
2237  * next replay request.
2238  *
2239  * \param[in] env       execution environment
2240  * \param[in] obd       failover obd device
2241  * \param[in] req       request to be dropped
2242  */
2243 static void drop_duplicate_replay_req(struct lu_env *env,
2244                                       struct obd_device *obd,
2245                                       struct ptlrpc_request *req)
2246 {
2247         DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2248                   " update records are found.\n",
2249                   lustre_msg_get_transno(req->rq_reqmsg),
2250                   libcfs_nid2str(req->rq_peer.nid));
2251
2252         /* Right now, only for MDS reint operation update replay and
2253          * normal request replay can have the same transno */
2254         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2255                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2256                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2257                 if (likely(req->rq_export))
2258                         target_committed_to_req(req);
2259                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2260                 target_send_reply(req, req->rq_status, 0);
2261         } else {
2262                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2263                 libcfs_nid2str(req->rq_peer.nid));
2264         }
2265         target_exp_dequeue_req_replay(req);
2266         target_request_copy_put(req);
2267         obd->obd_replayed_requests++;
2268 }
2269
2270 static void replay_request_or_update(struct lu_env *env,
2271                                      struct lu_target *lut,
2272                                      struct target_recovery_data *trd,
2273                                      struct ptlrpc_thread *thread)
2274 {
2275         struct obd_device *obd = lut->lut_obd;
2276         struct ptlrpc_request *req = NULL;
2277         int                     type;
2278         __u64                   transno;
2279         ENTRY;
2280
2281         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2282                obd->obd_next_recovery_transno);
2283
2284         /* Replay all of request and update by transno */
2285         do {
2286                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2287
2288                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2289
2290                 /** It is needed to extend recovery window above
2291                  *  recovery_time_soft. Extending is possible only in the
2292                  *  end of recovery window (see more details in
2293                  *  handle_recovery_req()).
2294                  */
2295                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2296
2297                 if (target_recovery_overseer(lut, check_for_next_transno,
2298                                         exp_req_replay_healthy_or_from_mdt)) {
2299                         abort_req_replay_queue(obd);
2300                         abort_lock_replay_queue(obd);
2301                         goto abort;
2302                 }
2303
2304                 spin_lock(&obd->obd_recovery_task_lock);
2305                 transno = get_next_transno(lut, &type);
2306                 if (type == REQUEST_RECOVERY && transno != 0) {
2307                         /* Drop replay request from client side, if the
2308                          * replay has been executed by update with the
2309                          * same transno */
2310                         req = list_entry(obd->obd_req_replay_queue.next,
2311                                         struct ptlrpc_request, rq_list);
2312
2313                         list_del_init(&req->rq_list);
2314                         obd->obd_requests_queued_for_recovery--;
2315                         spin_unlock(&obd->obd_recovery_task_lock);
2316
2317                         /* Let's check if the request has been redone by
2318                          * update replay */
2319                         if (is_req_replayed_by_update(req)) {
2320                                 struct distribute_txn_replay_req *dtrq;
2321
2322                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2323                                                                    req->rq_xid);
2324                                 LASSERT(dtrq != NULL);
2325                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2326                                 list_del_init(&dtrq->dtrq_list);
2327                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2328                                 dtrq_destroy(dtrq);
2329
2330                                 drop_duplicate_replay_req(env, obd, req);
2331
2332                                 continue;
2333                         }
2334
2335                         LASSERT(trd->trd_processing_task == current_pid());
2336                         DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2337                                   lustre_msg_get_transno(req->rq_reqmsg),
2338                                   libcfs_nid2str(req->rq_peer.nid));
2339
2340                         handle_recovery_req(thread, req,
2341                                             trd->trd_recovery_handler);
2342                         /**
2343                          * bz18031: increase next_recovery_transno before
2344                          * target_request_copy_put() will drop exp_rpc reference
2345                          */
2346                         spin_lock(&obd->obd_recovery_task_lock);
2347                         obd->obd_next_recovery_transno++;
2348                         spin_unlock(&obd->obd_recovery_task_lock);
2349                         target_exp_dequeue_req_replay(req);
2350                         target_request_copy_put(req);
2351                         obd->obd_replayed_requests++;
2352                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2353                         struct distribute_txn_replay_req *dtrq;
2354                         int rc;
2355
2356                         spin_unlock(&obd->obd_recovery_task_lock);
2357
2358                         LASSERT(tdtd != NULL);
2359                         dtrq = distribute_txn_get_next_req(tdtd);
2360                         lu_context_enter(&thread->t_env->le_ctx);
2361                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2362                         lu_context_exit(&thread->t_env->le_ctx);
2363                         extend_recovery_timer(obd, obd_timeout, true);
2364
2365                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2366                                 CDEBUG(D_HA, "Move x"LPU64" t"LPU64
2367                                        " to finish list\n", dtrq->dtrq_xid,
2368                                        dtrq->dtrq_master_transno);
2369
2370                                 /* Add it to the replay finish list */
2371                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2372                                 list_add(&dtrq->dtrq_list,
2373                                          &tdtd->tdtd_replay_finish_list);
2374                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2375
2376                                 spin_lock(&obd->obd_recovery_task_lock);
2377                                 if (transno == obd->obd_next_recovery_transno)
2378                                         obd->obd_next_recovery_transno++;
2379                                 else if (transno >
2380                                          obd->obd_next_recovery_transno)
2381                                         obd->obd_next_recovery_transno =
2382                                                                 transno + 1;
2383                                 spin_unlock(&obd->obd_recovery_task_lock);
2384                         } else {
2385                                 dtrq_destroy(dtrq);
2386                         }
2387                 } else {
2388                         spin_unlock(&obd->obd_recovery_task_lock);
2389 abort:
2390                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2391                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2392                         /** evict exports failed VBR */
2393                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2394                         break;
2395                 }
2396         } while (1);
2397 }
2398
2399 static int target_recovery_thread(void *arg)
2400 {
2401         struct lu_target *lut = arg;
2402         struct obd_device *obd = lut->lut_obd;
2403         struct ptlrpc_request *req;
2404         struct target_recovery_data *trd = &obd->obd_recovery_data;
2405         unsigned long delta;
2406         struct lu_env *env;
2407         struct ptlrpc_thread *thread = NULL;
2408         int rc = 0;
2409         ENTRY;
2410
2411         unshare_fs_struct();
2412         OBD_ALLOC_PTR(thread);
2413         if (thread == NULL)
2414                 RETURN(-ENOMEM);
2415
2416         OBD_ALLOC_PTR(env);
2417         if (env == NULL) {
2418                 OBD_FREE_PTR(thread);
2419                 RETURN(-ENOMEM);
2420         }
2421
2422         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2423         if (rc) {
2424                 OBD_FREE_PTR(thread);
2425                 OBD_FREE_PTR(env);
2426                 RETURN(rc);
2427         }
2428
2429         thread->t_env = env;
2430         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2431         env->le_ctx.lc_thread = thread;
2432         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2433         thread->t_watchdog = NULL;
2434
2435         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2436                current_pid());
2437         trd->trd_processing_task = current_pid();
2438
2439         spin_lock(&obd->obd_dev_lock);
2440         obd->obd_recovering = 1;
2441         spin_unlock(&obd->obd_dev_lock);
2442         complete(&trd->trd_starting);
2443
2444         /* first of all, we have to know the first transno to replay */
2445         if (target_recovery_overseer(lut, check_for_recovery_ready,
2446                                      exp_connect_healthy)) {
2447                 abort_req_replay_queue(obd);
2448                 abort_lock_replay_queue(obd);
2449                 if (lut->lut_tdtd != NULL)
2450                         dtrq_list_destroy(lut->lut_tdtd);
2451         }
2452
2453         /* next stage: replay requests or update */
2454         delta = jiffies;
2455         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2456                atomic_read(&obd->obd_req_replay_clients),
2457                obd->obd_next_recovery_transno);
2458         replay_request_or_update(env, lut, trd, thread);
2459
2460         /**
2461          * The second stage: replay locks
2462          */
2463         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2464                atomic_read(&obd->obd_lock_replay_clients));
2465         while ((req = target_next_replay_lock(lut))) {
2466                 LASSERT(trd->trd_processing_task == current_pid());
2467                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2468                           libcfs_nid2str(req->rq_peer.nid));
2469                 handle_recovery_req(thread, req,
2470                                     trd->trd_recovery_handler);
2471                 target_request_copy_put(req);
2472                 obd->obd_replayed_locks++;
2473         }
2474
2475         /**
2476          * The third stage: reply on final pings, at this moment all clients
2477          * must have request in final queue
2478          */
2479         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2480         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2481         /** Update server last boot epoch */
2482         tgt_boot_epoch_update(lut);
2483         /* We drop recoverying flag to forward all new requests
2484          * to regular mds_handle() since now */
2485         spin_lock(&obd->obd_dev_lock);
2486         obd->obd_recovering = obd->obd_abort_recovery = 0;
2487         spin_unlock(&obd->obd_dev_lock);
2488         spin_lock(&obd->obd_recovery_task_lock);
2489         target_cancel_recovery_timer(obd);
2490         spin_unlock(&obd->obd_recovery_task_lock);
2491         while ((req = target_next_final_ping(obd))) {
2492                 LASSERT(trd->trd_processing_task == current_pid());
2493                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2494                           libcfs_nid2str(req->rq_peer.nid));
2495                 handle_recovery_req(thread, req,
2496                                     trd->trd_recovery_handler);
2497                 /* Because the waiting client can not send ping to server,
2498                  * so we need refresh the last_request_time, to avoid the
2499                  * export is being evicted */
2500                 ptlrpc_update_export_timer(req->rq_export, 0);
2501                 target_request_copy_put(req);
2502         }
2503
2504         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2505         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2506               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2507         if (delta > OBD_RECOVERY_TIME_SOFT) {
2508                 CWARN("too long recovery - read logs\n");
2509                 libcfs_debug_dumplog();
2510         }
2511
2512         target_finish_recovery(lut);
2513
2514         lu_context_fini(&env->le_ctx);
2515         trd->trd_processing_task = 0;
2516         complete(&trd->trd_finishing);
2517
2518         tgt_io_thread_done(thread);
2519         OBD_FREE_PTR(thread);
2520         OBD_FREE_PTR(env);
2521         RETURN(rc);
2522 }
2523
2524 static int target_start_recovery_thread(struct lu_target *lut,
2525                                         svc_handler_t handler)
2526 {
2527         struct obd_device *obd = lut->lut_obd;
2528         int rc = 0;
2529         struct target_recovery_data *trd = &obd->obd_recovery_data;
2530         int index;
2531
2532         memset(trd, 0, sizeof(*trd));
2533         init_completion(&trd->trd_starting);
2534         init_completion(&trd->trd_finishing);
2535         trd->trd_recovery_handler = handler;
2536
2537         rc = server_name2index(obd->obd_name, &index, NULL);
2538         if (rc < 0)
2539                 return rc;
2540
2541         if (!IS_ERR(kthread_run(target_recovery_thread,
2542                                 lut, "tgt_recover_%d", index))) {
2543                 wait_for_completion(&trd->trd_starting);
2544                 LASSERT(obd->obd_recovering != 0);
2545         } else {
2546                 rc = -ECHILD;
2547         }
2548
2549         return rc;
2550 }
2551
2552 void target_stop_recovery_thread(struct obd_device *obd)
2553 {
2554         if (obd->obd_recovery_data.trd_processing_task > 0) {
2555                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2556                 /** recovery can be done but postrecovery is not yet */
2557                 spin_lock(&obd->obd_dev_lock);
2558                 if (obd->obd_recovering) {
2559                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2560                         obd->obd_abort_recovery = 1;
2561                         wake_up(&obd->obd_next_transno_waitq);
2562                 }
2563                 spin_unlock(&obd->obd_dev_lock);
2564                 wait_for_completion(&trd->trd_finishing);
2565         }
2566 }
2567 EXPORT_SYMBOL(target_stop_recovery_thread);
2568
2569 void target_recovery_fini(struct obd_device *obd)
2570 {
2571         class_disconnect_exports(obd);
2572         target_stop_recovery_thread(obd);
2573         target_cleanup_recovery(obd);
2574 }
2575 EXPORT_SYMBOL(target_recovery_fini);
2576
2577 static void target_recovery_expired(unsigned long castmeharder)
2578 {
2579         struct obd_device *obd = (struct obd_device *)castmeharder;
2580         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2581                " after %lds (%d clients connected)\n",
2582                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2583                cfs_time_current_sec()- obd->obd_recovery_start,
2584                atomic_read(&obd->obd_connected_clients));
2585
2586         obd->obd_recovery_expired = 1;
2587         wake_up(&obd->obd_next_transno_waitq);
2588 }
2589
2590 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2591 {
2592         struct obd_device *obd = lut->lut_obd;
2593
2594         if (obd->obd_max_recoverable_clients == 0) {
2595                 /** Update server last boot epoch */
2596                 tgt_boot_epoch_update(lut);
2597                 return;
2598         }
2599
2600         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2601                "last_transno "LPU64"\n", obd->obd_name,
2602                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2603         LASSERT(obd->obd_stopping == 0);
2604         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2605         obd->obd_recovery_start = 0;
2606         obd->obd_recovery_end = 0;
2607
2608         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2609         target_start_recovery_thread(lut, handler);
2610 }
2611 EXPORT_SYMBOL(target_recovery_init);
2612
2613 static int target_process_req_flags(struct obd_device *obd,
2614                                     struct ptlrpc_request *req)
2615 {
2616         struct obd_export *exp = req->rq_export;
2617         LASSERT(exp != NULL);
2618         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2619                 /* client declares he's ready to replay locks */
2620                 spin_lock(&exp->exp_lock);
2621                 if (exp->exp_req_replay_needed) {
2622                         exp->exp_req_replay_needed = 0;
2623                         spin_unlock(&exp->exp_lock);
2624
2625                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2626                         atomic_dec(&obd->obd_req_replay_clients);
2627                 } else {
2628                         spin_unlock(&exp->exp_lock);
2629                 }
2630         }
2631         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2632                 /* client declares he's ready to complete recovery
2633                  * so, we put the request on th final queue */
2634                 spin_lock(&exp->exp_lock);
2635                 if (exp->exp_lock_replay_needed) {
2636                         exp->exp_lock_replay_needed = 0;
2637                         spin_unlock(&exp->exp_lock);
2638
2639                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2640                         atomic_dec(&obd->obd_lock_replay_clients);
2641                 } else {
2642                         spin_unlock(&exp->exp_lock);
2643                 }
2644         }
2645         return 0;
2646 }
2647
2648 int target_queue_recovery_request(struct ptlrpc_request *req,
2649                                   struct obd_device *obd)
2650 {
2651         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2652         struct ptlrpc_request *reqiter;
2653         int inserted = 0;
2654         ENTRY;
2655
2656         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2657                 /* Processing the queue right now, don't re-add. */
2658                 RETURN(1);
2659         }
2660
2661         target_process_req_flags(obd, req);
2662
2663         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2664                 /* client declares he's ready to complete recovery
2665                  * so, we put the request on th final queue */
2666                 target_request_copy_get(req);
2667                 DEBUG_REQ(D_HA, req, "queue final req");
2668                 wake_up(&obd->obd_next_transno_waitq);
2669                 spin_lock(&obd->obd_recovery_task_lock);
2670                 if (obd->obd_recovering) {
2671                         list_add_tail(&req->rq_list,
2672                                           &obd->obd_final_req_queue);
2673                 } else {
2674                         spin_unlock(&obd->obd_recovery_task_lock);
2675                         target_request_copy_put(req);
2676                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2677                 }
2678                 spin_unlock(&obd->obd_recovery_task_lock);
2679                 RETURN(0);
2680         }
2681         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2682                 /* client declares he's ready to replay locks */
2683                 target_request_copy_get(req);
2684                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2685                 wake_up(&obd->obd_next_transno_waitq);
2686                 spin_lock(&obd->obd_recovery_task_lock);
2687                 LASSERT(obd->obd_recovering);
2688                 /* usually due to recovery abort */
2689                 if (!req->rq_export->exp_in_recovery) {
2690                         spin_unlock(&obd->obd_recovery_task_lock);
2691                         target_request_copy_put(req);
2692                         RETURN(-ENOTCONN);
2693                 }
2694                 LASSERT(req->rq_export->exp_lock_replay_needed);
2695                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2696                 spin_unlock(&obd->obd_recovery_task_lock);
2697                 RETURN(0);
2698         }
2699
2700         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2701          * (i.e. buflens etc are in my own byte order), but type-dependent
2702          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2703
2704         if (!transno) {
2705                 INIT_LIST_HEAD(&req->rq_list);
2706                 DEBUG_REQ(D_HA, req, "not queueing");
2707                 RETURN(1);
2708         }
2709
2710         /* If we're processing the queue, we want don't want to queue this
2711          * message.
2712          *
2713          * Also, if this request has a transno less than the one we're waiting
2714          * for, we should process it now.  It could (and currently always will)
2715          * be an open request for a descriptor that was opened some time ago.
2716          *
2717          * Also, a resent, replayed request that has already been
2718          * handled will pass through here and be processed immediately.
2719          */
2720         CDEBUG(D_HA, "Next recovery transno: "LPU64
2721                ", current: "LPU64", replaying\n",
2722                obd->obd_next_recovery_transno, transno);
2723
2724         /* If the request has been replayed by update replay, then sends this
2725          * request to the recovery thread (replay_request_or_update()), where
2726          * it will be handled */
2727         spin_lock(&obd->obd_recovery_task_lock);
2728         if (transno < obd->obd_next_recovery_transno &&
2729             !is_req_replayed_by_update(req)) {
2730                 /* Processing the queue right now, don't re-add. */
2731                 LASSERT(list_empty(&req->rq_list));
2732                 spin_unlock(&obd->obd_recovery_task_lock);
2733                 RETURN(1);
2734         }
2735         spin_unlock(&obd->obd_recovery_task_lock);
2736
2737         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2738                 RETURN(0);
2739
2740         target_request_copy_get(req);
2741         if (!req->rq_export->exp_in_recovery) {
2742                 target_request_copy_put(req);
2743                 RETURN(-ENOTCONN);
2744         }
2745         LASSERT(req->rq_export->exp_req_replay_needed);
2746
2747         if (target_exp_enqueue_req_replay(req)) {
2748                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2749                 target_request_copy_put(req);
2750                 RETURN(0);
2751         }
2752
2753         /* XXX O(n^2) */
2754         spin_lock(&obd->obd_recovery_task_lock);
2755         LASSERT(obd->obd_recovering);
2756         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2757                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2758                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2759                         inserted = 1;
2760                         goto added;
2761                 }
2762
2763                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2764                              transno)) {
2765                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2766                                   "has been claimed by another client");
2767                         spin_unlock(&obd->obd_recovery_task_lock);
2768                         target_exp_dequeue_req_replay(req);
2769                         target_request_copy_put(req);
2770                         RETURN(0);
2771                 }
2772         }
2773 added:
2774         if (!inserted)
2775                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2776
2777         obd->obd_requests_queued_for_recovery++;
2778         spin_unlock(&obd->obd_recovery_task_lock);
2779         wake_up(&obd->obd_next_transno_waitq);
2780         RETURN(0);
2781 }
2782
2783 int target_handle_ping(struct ptlrpc_request *req)
2784 {
2785         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2786         return req_capsule_server_pack(&req->rq_pill);
2787 }
2788
2789 void target_committed_to_req(struct ptlrpc_request *req)
2790 {
2791         struct obd_export *exp = req->rq_export;
2792
2793         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2794                 lustre_msg_set_last_committed(req->rq_repmsg,
2795                                               exp->exp_last_committed);
2796         else
2797                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2798                           "%d)", exp->exp_obd->obd_no_transno,
2799                           req->rq_repmsg == NULL);
2800
2801         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2802                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2803 }
2804
2805 #endif /* HAVE_SERVER_SUPPORT */
2806
2807 /**
2808  * Packs current SLV and Limit into \a req.
2809  */
2810 int target_pack_pool_reply(struct ptlrpc_request *req)
2811 {
2812         struct obd_device *obd;
2813         ENTRY;
2814
2815         /* Check that we still have all structures alive as this may
2816          * be some late RPC at shutdown time. */
2817         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2818                      !exp_connect_lru_resize(req->rq_export))) {
2819                 lustre_msg_set_slv(req->rq_repmsg, 0);
2820                 lustre_msg_set_limit(req->rq_repmsg, 0);
2821                 RETURN(0);
2822         }
2823
2824         /* OBD is alive here as export is alive, which we checked above. */
2825         obd = req->rq_export->exp_obd;
2826
2827         read_lock(&obd->obd_pool_lock);
2828         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2829         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2830         read_unlock(&obd->obd_pool_lock);
2831
2832         RETURN(0);
2833 }
2834
2835 static int target_send_reply_msg(struct ptlrpc_request *req,
2836                                  int rc, int fail_id)
2837 {
2838         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2839                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2840                 return -ECOMM;
2841         }
2842         /* We can have a null rq_reqmsg in the event of bad signature or
2843          * no context when unwrapping */
2844         if (req->rq_reqmsg &&
2845             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2846             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2847                 return -ECOMM;
2848
2849         if (unlikely(rc)) {
2850                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2851                 req->rq_status = rc;
2852                 return ptlrpc_send_error(req, 1);
2853         } else {
2854                 DEBUG_REQ(D_NET, req, "sending reply");
2855         }
2856
2857         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2858 }
2859
2860 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2861 {
2862         struct ptlrpc_service_part *svcpt;
2863         int                        netrc;
2864         struct ptlrpc_reply_state *rs;
2865         struct obd_export         *exp;
2866         ENTRY;
2867
2868         if (req->rq_no_reply) {
2869                 EXIT;
2870                 return;
2871         }
2872
2873         svcpt = req->rq_rqbd->rqbd_svcpt;
2874         rs = req->rq_reply_state;
2875         if (rs == NULL || !rs->rs_difficult) {
2876                 /* no notifiers */
2877                 target_send_reply_msg (req, rc, fail_id);
2878                 EXIT;
2879                 return;
2880         }
2881
2882         /* must be an export if locks saved */
2883         LASSERT(req->rq_export != NULL);
2884         /* req/reply consistent */
2885         LASSERT(rs->rs_svcpt == svcpt);
2886
2887         /* "fresh" reply */
2888         LASSERT(!rs->rs_scheduled);
2889         LASSERT(!rs->rs_scheduled_ever);
2890         LASSERT(!rs->rs_handled);
2891         LASSERT(!rs->rs_on_net);
2892         LASSERT(rs->rs_export == NULL);
2893         LASSERT(list_empty(&rs->rs_obd_list));
2894         LASSERT(list_empty(&rs->rs_exp_list));
2895
2896         exp = class_export_get(req->rq_export);
2897
2898         /* disable reply scheduling while I'm setting up */
2899         rs->rs_scheduled = 1;
2900         rs->rs_on_net    = 1;
2901         rs->rs_xid       = req->rq_xid;
2902         rs->rs_transno   = req->rq_transno;
2903         rs->rs_export    = exp;
2904         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2905
2906         spin_lock(&exp->exp_uncommitted_replies_lock);
2907         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2908                rs->rs_transno, exp->exp_last_committed);
2909         if (rs->rs_transno > exp->exp_last_committed) {
2910                 /* not committed already */
2911                 list_add_tail(&rs->rs_obd_list,
2912                                   &exp->exp_uncommitted_replies);
2913         }
2914         spin_unlock(&exp->exp_uncommitted_replies_lock);
2915
2916         spin_lock(&exp->exp_lock);
2917         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2918         spin_unlock(&exp->exp_lock);
2919
2920         netrc = target_send_reply_msg(req, rc, fail_id);
2921
2922         spin_lock(&svcpt->scp_rep_lock);
2923
2924         atomic_inc(&svcpt->scp_nreps_difficult);
2925
2926         if (netrc != 0) {
2927                 /* error sending: reply is off the net.  Also we need +1
2928                  * reply ref until ptlrpc_handle_rs() is done
2929                  * with the reply state (if the send was successful, there
2930                  * would have been +1 ref for the net, which
2931                  * reply_out_callback leaves alone) */
2932                 rs->rs_on_net = 0;
2933                 ptlrpc_rs_addref(rs);
2934         }
2935
2936         spin_lock(&rs->rs_lock);
2937         if (rs->rs_transno <= exp->exp_last_committed ||
2938             (!rs->rs_on_net && !rs->rs_no_ack) ||
2939             list_empty(&rs->rs_exp_list) ||     /* completed already */
2940             list_empty(&rs->rs_obd_list)) {
2941                 CDEBUG(D_HA, "Schedule reply immediately\n");
2942                 ptlrpc_dispatch_difficult_reply(rs);
2943         } else {
2944                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2945                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2946         }
2947         spin_unlock(&rs->rs_lock);
2948         spin_unlock(&svcpt->scp_rep_lock);
2949         EXIT;
2950 }
2951
2952 enum ldlm_mode lck_compat_array[] = {
2953         [LCK_EX]    = LCK_COMPAT_EX,
2954         [LCK_PW]    = LCK_COMPAT_PW,
2955         [LCK_PR]    = LCK_COMPAT_PR,
2956         [LCK_CW]    = LCK_COMPAT_CW,
2957         [LCK_CR]    = LCK_COMPAT_CR,
2958         [LCK_NL]    = LCK_COMPAT_NL,
2959         [LCK_GROUP] = LCK_COMPAT_GROUP,
2960         [LCK_COS]   = LCK_COMPAT_COS,
2961 };
2962
2963 /**
2964  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2965  * not escape to the user level.
2966  */
2967 int ldlm_error2errno(enum ldlm_error error)
2968 {
2969         int result;
2970
2971         switch (error) {
2972         case ELDLM_OK:
2973         case ELDLM_LOCK_MATCHED:
2974                 result = 0;
2975                 break;
2976         case ELDLM_LOCK_CHANGED:
2977                 result = -ESTALE;
2978                 break;
2979         case ELDLM_LOCK_ABORTED:
2980                 result = -ENAVAIL;
2981                 break;
2982         case ELDLM_LOCK_REPLACED:
2983                 result = -ESRCH;
2984                 break;
2985         case ELDLM_NO_LOCK_DATA:
2986                 result = -ENOENT;
2987                 break;
2988         case ELDLM_NAMESPACE_EXISTS:
2989                 result = -EEXIST;
2990                 break;
2991         case ELDLM_BAD_NAMESPACE:
2992                 result = -EBADF;
2993                 break;
2994         default:
2995                 if (((int)error) < 0) { /* cast to signed type */
2996                         result = error; /* as ldlm_error can be unsigned */
2997                 } else {
2998                         CERROR("Invalid DLM result code: %d\n", error);
2999                         result = -EPROTO;
3000                 }
3001         }
3002         return result;
3003 }
3004 EXPORT_SYMBOL(ldlm_error2errno);
3005
3006 /**
3007  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3008  */
3009 enum ldlm_error ldlm_errno2error(int err_no)
3010 {
3011         int error;
3012
3013         switch (err_no) {
3014         case 0:
3015                 error = ELDLM_OK;
3016                 break;
3017         case -ESTALE:
3018                 error = ELDLM_LOCK_CHANGED;
3019                 break;
3020         case -ENAVAIL:
3021                 error = ELDLM_LOCK_ABORTED;
3022                 break;
3023         case -ESRCH:
3024                 error = ELDLM_LOCK_REPLACED;
3025                 break;
3026         case -ENOENT:
3027                 error = ELDLM_NO_LOCK_DATA;
3028                 break;
3029         case -EEXIST:
3030                 error = ELDLM_NAMESPACE_EXISTS;
3031                 break;
3032         case -EBADF:
3033                 error = ELDLM_BAD_NAMESPACE;
3034                 break;
3035         default:
3036                 error = err_no;
3037         }
3038         return error;
3039 }
3040
3041 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3042 void ldlm_dump_export_locks(struct obd_export *exp)
3043 {
3044         spin_lock(&exp->exp_locks_list_guard);
3045         if (!list_empty(&exp->exp_locks_list)) {
3046                 struct ldlm_lock *lock;
3047
3048                 CERROR("dumping locks for export %p,"
3049                        "ignore if the unmount doesn't hang\n", exp);
3050                 list_for_each_entry(lock, &exp->exp_locks_list,
3051                                         l_exp_refs_link)
3052                         LDLM_ERROR(lock, "lock:");
3053         }
3054         spin_unlock(&exp->exp_locks_list_guard);
3055 }
3056 #endif
3057
3058 #ifdef HAVE_SERVER_SUPPORT
3059 static int target_bulk_timeout(void *data)
3060 {
3061         ENTRY;
3062         /* We don't fail the connection here, because having the export
3063          * killed makes the (vital) call to commitrw very sad.
3064          */
3065         RETURN(1);
3066 }
3067
3068 static inline const char *bulk2type(struct ptlrpc_request *req)
3069 {
3070         if (req->rq_bulk_read)
3071                 return "READ";
3072         if (req->rq_bulk_write)
3073                 return "WRITE";
3074         return "UNKNOWN";
3075 }
3076
3077 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3078                    struct l_wait_info *lwi)
3079 {
3080         struct ptlrpc_request   *req = desc->bd_req;
3081         time_t                   start = cfs_time_current_sec();
3082         time_t                   deadline;
3083         int                      rc = 0;
3084
3085         ENTRY;
3086
3087         /* If there is eviction in progress, wait for it to finish. */
3088         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3089                 *lwi = LWI_INTR(NULL, NULL);
3090                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3091                                   !atomic_read(&exp->exp_obd->
3092                                                    obd_evict_inprogress),
3093                                   lwi);
3094         }
3095
3096         /* Check if client was evicted or reconnected already. */
3097         if (exp->exp_failed ||
3098             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3099                 rc = -ENOTCONN;
3100         } else {
3101                 if (req->rq_bulk_read)
3102                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3103
3104                 if ((exp->exp_connect_data.ocd_connect_flags &
3105                      OBD_CONNECT_BULK_MBITS) != 0)
3106                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3107                 else /* old version, bulk matchbits is rq_xid */
3108                         req->rq_mbits = req->rq_xid;
3109
3110                 if (rc == 0)
3111                         rc = ptlrpc_start_bulk_transfer(desc);
3112         }
3113
3114         if (rc < 0) {
3115                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3116                           bulk2type(req), rc);
3117                 RETURN(rc);
3118         }
3119
3120         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3121                 ptlrpc_abort_bulk(desc);
3122                 RETURN(0);
3123         }
3124
3125         /* limit actual bulk transfer to bulk_timeout seconds */
3126         deadline = start + bulk_timeout;
3127         if (deadline > req->rq_deadline)
3128                 deadline = req->rq_deadline;
3129
3130         do {
3131                 long timeoutl = deadline - cfs_time_current_sec();
3132                 cfs_duration_t timeout = timeoutl <= 0 ?
3133                                          CFS_TICK : cfs_time_seconds(timeoutl);
3134                 time_t  rq_deadline;
3135
3136                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3137                                             target_bulk_timeout, desc);
3138                 rc = l_wait_event(desc->bd_waitq,
3139                                   !ptlrpc_server_bulk_active(desc) ||
3140                                   exp->exp_failed ||
3141                                   exp->exp_conn_cnt >
3142                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3143                                   lwi);
3144                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3145                 /* Wait again if we changed rq_deadline. */
3146                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3147                 deadline = start + bulk_timeout;
3148                 if (deadline > rq_deadline)
3149                         deadline = rq_deadline;
3150         } while ((rc == -ETIMEDOUT) &&
3151                  (deadline > cfs_time_current_sec()));
3152
3153         if (rc == -ETIMEDOUT) {
3154                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3155                           bulk2type(req), deadline - start,
3156                           cfs_time_current_sec() - deadline);
3157                 ptlrpc_abort_bulk(desc);
3158         } else if (exp->exp_failed) {
3159                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3160                           bulk2type(req));
3161                 rc = -ENOTCONN;
3162                 ptlrpc_abort_bulk(desc);
3163         } else if (exp->exp_conn_cnt >
3164                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3165                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3166                           bulk2type(req));
3167                 /* We don't reply anyway. */
3168                 rc = -ETIMEDOUT;
3169                 ptlrpc_abort_bulk(desc);
3170         } else if (desc->bd_failure) {
3171                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3172                           bulk2type(req));
3173                 /* XXX should this be a different errno? */
3174                 rc = -ETIMEDOUT;
3175         } else {