Whamcloud - gitweb
LU-6086 obdclass: check peer's version for MDT-MDT connection
[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, 2014, 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 <libcfs/libcfs.h>
47 #include <obd.h>
48 #include <obd_class.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
53
54 /* @priority: If non-zero, move the selected connection to the list head.
55  * @create: If zero, only search in existing connections.
56  */
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58                            int priority, int create)
59 {
60         struct ptlrpc_connection *ptlrpc_conn;
61         struct obd_import_conn *imp_conn = NULL, *item;
62         int rc = 0;
63         ENTRY;
64
65         if (!create && !priority) {
66                 CDEBUG(D_HA, "Nothing to do\n");
67                 RETURN(-EINVAL);
68         }
69
70         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
71         if (!ptlrpc_conn) {
72                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
73                 RETURN (-ENOENT);
74         }
75
76         if (create) {
77                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
78                 if (!imp_conn) {
79                         GOTO(out_put, rc = -ENOMEM);
80                 }
81         }
82
83         spin_lock(&imp->imp_lock);
84         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
86                         if (priority) {
87                                 list_del(&item->oic_item);
88                                 list_add(&item->oic_item,
89                                              &imp->imp_conn_list);
90                                 item->oic_last_attempt = 0;
91                         }
92                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93                                imp, imp->imp_obd->obd_name, uuid->uuid,
94                                (priority ? ", moved to head" : ""));
95                         spin_unlock(&imp->imp_lock);
96                         GOTO(out_free, rc = 0);
97                 }
98         }
99         /* No existing import connection found for \a uuid. */
100         if (create) {
101                 imp_conn->oic_conn = ptlrpc_conn;
102                 imp_conn->oic_uuid = *uuid;
103                 imp_conn->oic_last_attempt = 0;
104                 if (priority)
105                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
106                 else
107                         list_add_tail(&imp_conn->oic_item,
108                                           &imp->imp_conn_list);
109                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
110                        imp, imp->imp_obd->obd_name, uuid->uuid,
111                        (priority ? "head" : "tail"));
112         } else {
113                 spin_unlock(&imp->imp_lock);
114                 GOTO(out_free, rc = -ENOENT);
115         }
116
117         spin_unlock(&imp->imp_lock);
118         RETURN(0);
119 out_free:
120         if (imp_conn)
121                 OBD_FREE(imp_conn, sizeof(*imp_conn));
122 out_put:
123         ptlrpc_connection_put(ptlrpc_conn);
124         RETURN(rc);
125 }
126
127 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
128 {
129         return import_set_conn(imp, uuid, 1, 0);
130 }
131
132 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
133                            int priority)
134 {
135         return import_set_conn(imp, uuid, priority, 1);
136 }
137 EXPORT_SYMBOL(client_import_add_conn);
138
139 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
140 {
141         struct obd_import_conn *imp_conn;
142         struct obd_export *dlmexp;
143         int rc = -ENOENT;
144         ENTRY;
145
146         spin_lock(&imp->imp_lock);
147         if (list_empty(&imp->imp_conn_list)) {
148                 LASSERT(!imp->imp_connection);
149                 GOTO(out, rc);
150         }
151
152         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
153                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
154                         continue;
155                 LASSERT(imp_conn->oic_conn);
156
157                 if (imp_conn == imp->imp_conn_current) {
158                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
159
160                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
161                             imp->imp_state != LUSTRE_IMP_DISCON) {
162                                 CERROR("can't remove current connection\n");
163                                 GOTO(out, rc = -EBUSY);
164                         }
165
166                         ptlrpc_connection_put(imp->imp_connection);
167                         imp->imp_connection = NULL;
168
169                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
170                         if (dlmexp && dlmexp->exp_connection) {
171                                 LASSERT(dlmexp->exp_connection ==
172                                         imp_conn->oic_conn);
173                                 ptlrpc_connection_put(dlmexp->exp_connection);
174                                 dlmexp->exp_connection = NULL;
175                         }
176                 }
177
178                 list_del(&imp_conn->oic_item);
179                 ptlrpc_connection_put(imp_conn->oic_conn);
180                 OBD_FREE(imp_conn, sizeof(*imp_conn));
181                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182                        imp, imp->imp_obd->obd_name, uuid->uuid);
183                 rc = 0;
184                 break;
185         }
186 out:
187         spin_unlock(&imp->imp_lock);
188         if (rc == -ENOENT)
189                 CERROR("connection %s not found\n", uuid->uuid);
190         RETURN(rc);
191 }
192 EXPORT_SYMBOL(client_import_del_conn);
193
194 /**
195  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
196  * to find a conn uuid of \a imp which can reach \a peer.
197  */
198 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
199                             struct obd_uuid *uuid)
200 {
201         struct obd_import_conn *conn;
202         int rc = -ENOENT;
203         ENTRY;
204
205         spin_lock(&imp->imp_lock);
206         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
207                 /* Check if conn UUID does have this peer NID. */
208                 if (class_check_uuid(&conn->oic_uuid, peer)) {
209                         *uuid = conn->oic_uuid;
210                         rc = 0;
211                         break;
212                 }
213         }
214         spin_unlock(&imp->imp_lock);
215         RETURN(rc);
216 }
217 EXPORT_SYMBOL(client_import_find_conn);
218
219 void client_destroy_import(struct obd_import *imp)
220 {
221         /* Drop security policy instance after all RPCs have finished/aborted
222          * to let all busy contexts be released. */
223         class_import_get(imp);
224         class_destroy_import(imp);
225         sptlrpc_import_sec_put(imp);
226         class_import_put(imp);
227 }
228 EXPORT_SYMBOL(client_destroy_import);
229
230 /**
231  * Check whether or not the OSC is on MDT.
232  * In the config log,
233  * osc on MDT
234  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
235  * osc on client
236  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
237  *
238  **/
239 static int osc_on_mdt(char *obdname)
240 {
241         char *ptr;
242
243         ptr = strrchr(obdname, '-');
244         if (ptr == NULL)
245                 return 0;
246
247         if (strncmp(ptr + 1, "MDT", 3) == 0)
248                 return 1;
249
250         return 0;
251 }
252
253 /* Configure an RPC client OBD device.
254  *
255  * lcfg parameters:
256  * 1 - client UUID
257  * 2 - server UUID
258  * 3 - inactive-on-startup
259  */
260 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
261 {
262         struct client_obd *cli = &obddev->u.cli;
263         struct obd_import *imp;
264         struct obd_uuid server_uuid;
265         int rq_portal, rp_portal, connect_op;
266         char *name = obddev->obd_type->typ_name;
267         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
268         int rc;
269         ENTRY;
270
271         /* In a more perfect world, we would hang a ptlrpc_client off of
272          * obd_type and just use the values from there. */
273         if (!strcmp(name, LUSTRE_OSC_NAME)) {
274                 rq_portal = OST_REQUEST_PORTAL;
275                 rp_portal = OSC_REPLY_PORTAL;
276                 connect_op = OST_CONNECT;
277                 cli->cl_sp_me = LUSTRE_SP_CLI;
278                 cli->cl_sp_to = LUSTRE_SP_OST;
279                 ns_type = LDLM_NS_TYPE_OSC;
280         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
281                    !strcmp(name, LUSTRE_LWP_NAME)) {
282                 rq_portal = MDS_REQUEST_PORTAL;
283                 rp_portal = MDC_REPLY_PORTAL;
284                 connect_op = MDS_CONNECT;
285                 cli->cl_sp_me = LUSTRE_SP_CLI;
286                 cli->cl_sp_to = LUSTRE_SP_MDT;
287                 ns_type = LDLM_NS_TYPE_MDC;
288         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
289                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
290                         /* OSP_on_MDT for other MDTs */
291                         connect_op = MDS_CONNECT;
292                         cli->cl_sp_to = LUSTRE_SP_MDT;
293                         ns_type = LDLM_NS_TYPE_MDC;
294                         rq_portal = OUT_PORTAL;
295                 } else {
296                         /* OSP on MDT for OST */
297                         connect_op = OST_CONNECT;
298                         cli->cl_sp_to = LUSTRE_SP_OST;
299                         ns_type = LDLM_NS_TYPE_OSC;
300                         rq_portal = OST_REQUEST_PORTAL;
301                 }
302                 rp_portal = OSC_REPLY_PORTAL;
303                 cli->cl_sp_me = LUSTRE_SP_CLI;
304         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
305                 rq_portal = MGS_REQUEST_PORTAL;
306                 rp_portal = MGC_REPLY_PORTAL;
307                 connect_op = MGS_CONNECT;
308                 cli->cl_sp_me = LUSTRE_SP_MGC;
309                 cli->cl_sp_to = LUSTRE_SP_MGS;
310                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
311                 ns_type = LDLM_NS_TYPE_MGC;
312         } else {
313                 CERROR("unknown client OBD type \"%s\", can't setup\n",
314                        name);
315                 RETURN(-EINVAL);
316         }
317
318         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
319                 CERROR("requires a TARGET UUID\n");
320                 RETURN(-EINVAL);
321         }
322
323         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
324                 CERROR("client UUID must be less than 38 characters\n");
325                 RETURN(-EINVAL);
326         }
327
328         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
329                 CERROR("setup requires a SERVER UUID\n");
330                 RETURN(-EINVAL);
331         }
332
333         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
334                 CERROR("target UUID must be less than 38 characters\n");
335                 RETURN(-EINVAL);
336         }
337
338         init_rwsem(&cli->cl_sem);
339         mutex_init(&cli->cl_mgc_mutex);
340         cli->cl_conn_count = 0;
341         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
342                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
343                      sizeof(server_uuid)));
344
345         cli->cl_dirty_pages = 0;
346         cli->cl_avail_grant = 0;
347         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
348         /* cl_dirty_max_pages may be changed at connect time in
349          * ptlrpc_connect_interpret(). */
350         client_adjust_max_dirty(cli);
351         INIT_LIST_HEAD(&cli->cl_cache_waiters);
352         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
353         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
354         INIT_LIST_HEAD(&cli->cl_loi_write_list);
355         INIT_LIST_HEAD(&cli->cl_loi_read_list);
356         spin_lock_init(&cli->cl_loi_list_lock);
357         atomic_set(&cli->cl_pending_w_pages, 0);
358         atomic_set(&cli->cl_pending_r_pages, 0);
359         cli->cl_r_in_flight = 0;
360         cli->cl_w_in_flight = 0;
361
362         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
363         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
364         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
365         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
366         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
367         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
368
369         /* lru for osc. */
370         INIT_LIST_HEAD(&cli->cl_lru_osc);
371         atomic_set(&cli->cl_lru_shrinkers, 0);
372         atomic_long_set(&cli->cl_lru_busy, 0);
373         atomic_long_set(&cli->cl_lru_in_list, 0);
374         INIT_LIST_HEAD(&cli->cl_lru_list);
375         spin_lock_init(&cli->cl_lru_list_lock);
376         atomic_long_set(&cli->cl_unstable_count, 0);
377
378         init_waitqueue_head(&cli->cl_destroy_waitq);
379         atomic_set(&cli->cl_destroy_in_flight, 0);
380 #ifdef ENABLE_CHECKSUM
381         /* Turn on checksumming by default. */
382         cli->cl_checksum = 1;
383         /*
384          * The supported checksum types will be worked out at connect time
385          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
386          * through procfs.
387          */
388         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
389 #endif
390         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
391
392         /* This value may be reduced at connect time in
393          * ptlrpc_connect_interpret() . We initialize it to only
394          * 1MB until we know what the performance looks like.
395          * In the future this should likely be increased. LU-1431 */
396         cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
397                                           LNET_MTU >> PAGE_CACHE_SHIFT);
398
399         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
400          * it will be updated at OSC connection time. */
401         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
402
403         if (!strcmp(name, LUSTRE_MDC_NAME)) {
404                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
405         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
406                 cli->cl_max_rpcs_in_flight = 2;
407         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
408                 cli->cl_max_rpcs_in_flight = 3;
409         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
410                 cli->cl_max_rpcs_in_flight = 4;
411         } else {
412                 if (osc_on_mdt(obddev->obd_name))
413                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
414                 else
415                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
416         }
417         rc = ldlm_get_ref();
418         if (rc) {
419                 CERROR("ldlm_get_ref failed: %d\n", rc);
420                 GOTO(err, rc);
421         }
422
423         ptlrpc_init_client(rq_portal, rp_portal, name,
424                            &obddev->obd_ldlm_client);
425
426         imp = class_new_import(obddev);
427         if (imp == NULL)
428                 GOTO(err_ldlm, rc = -ENOENT);
429         imp->imp_client = &obddev->obd_ldlm_client;
430         imp->imp_connect_op = connect_op;
431         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
432                LUSTRE_CFG_BUFLEN(lcfg, 1));
433         class_import_put(imp);
434
435         rc = client_import_add_conn(imp, &server_uuid, 1);
436         if (rc) {
437                 CERROR("can't add initial connection\n");
438                 GOTO(err_import, rc);
439         }
440
441         cli->cl_import = imp;
442         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
443         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
444         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
445
446         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
447                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
448                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
449                                name, obddev->obd_name,
450                                cli->cl_target_uuid.uuid);
451                         spin_lock(&imp->imp_lock);
452                         imp->imp_deactive = 1;
453                         spin_unlock(&imp->imp_lock);
454                 }
455         }
456
457         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
458                                                    LDLM_NAMESPACE_CLIENT,
459                                                    LDLM_NAMESPACE_GREEDY,
460                                                    ns_type);
461         if (obddev->obd_namespace == NULL) {
462                 CERROR("Unable to create client namespace - %s\n",
463                        obddev->obd_name);
464                 GOTO(err_import, rc = -ENOMEM);
465         }
466
467         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
468
469         RETURN(rc);
470
471 err_import:
472         class_destroy_import(imp);
473 err_ldlm:
474         ldlm_put_ref();
475 err:
476         RETURN(rc);
477
478 }
479 EXPORT_SYMBOL(client_obd_setup);
480
481 int client_obd_cleanup(struct obd_device *obddev)
482 {
483         ENTRY;
484
485         ldlm_namespace_free_post(obddev->obd_namespace);
486         obddev->obd_namespace = NULL;
487
488         obd_cleanup_client_import(obddev);
489         LASSERT(obddev->u.cli.cl_import == NULL);
490
491         ldlm_put_ref();
492         RETURN(0);
493 }
494 EXPORT_SYMBOL(client_obd_cleanup);
495
496 /* ->o_connect() method for client side (OSC and MDC and MGC) */
497 int client_connect_import(const struct lu_env *env,
498                           struct obd_export **exp,
499                           struct obd_device *obd, struct obd_uuid *cluuid,
500                           struct obd_connect_data *data, void *localdata)
501 {
502         struct client_obd       *cli    = &obd->u.cli;
503         struct obd_import       *imp    = cli->cl_import;
504         struct obd_connect_data *ocd;
505         struct lustre_handle    conn    = { 0 };
506         int                     rc;
507         ENTRY;
508
509         *exp = NULL;
510         down_write(&cli->cl_sem);
511         if (cli->cl_conn_count > 0)
512                 GOTO(out_sem, rc = -EALREADY);
513
514         rc = class_connect(&conn, obd, cluuid);
515         if (rc)
516                 GOTO(out_sem, rc);
517
518         cli->cl_conn_count++;
519         *exp = class_conn2export(&conn);
520
521         LASSERT(obd->obd_namespace);
522
523         imp->imp_dlm_handle = conn;
524         rc = ptlrpc_init_import(imp);
525         if (rc != 0)
526                 GOTO(out_ldlm, rc);
527
528         ocd = &imp->imp_connect_data;
529         if (data) {
530                 *ocd = *data;
531                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
532         }
533
534         rc = ptlrpc_connect_import(imp);
535         if (rc != 0) {
536                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
537                 GOTO(out_ldlm, rc);
538         }
539         LASSERT(*exp != NULL && (*exp)->exp_connection);
540
541         if (data) {
542                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
543                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
544                          data->ocd_connect_flags, ocd->ocd_connect_flags);
545                 data->ocd_connect_flags = ocd->ocd_connect_flags;
546         }
547
548         ptlrpc_pinger_add_import(imp);
549
550         EXIT;
551
552         if (rc) {
553 out_ldlm:
554                 cli->cl_conn_count--;
555                 class_disconnect(*exp);
556                 *exp = NULL;
557         }
558 out_sem:
559         up_write(&cli->cl_sem);
560
561         return rc;
562 }
563 EXPORT_SYMBOL(client_connect_import);
564
565 int client_disconnect_export(struct obd_export *exp)
566 {
567         struct obd_device *obd = class_exp2obd(exp);
568         struct client_obd *cli;
569         struct obd_import *imp;
570         int rc = 0, err;
571         ENTRY;
572
573         if (!obd) {
574                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
575                        exp, exp ? exp->exp_handle.h_cookie : -1);
576                 RETURN(-EINVAL);
577         }
578
579         cli = &obd->u.cli;
580         imp = cli->cl_import;
581
582         down_write(&cli->cl_sem);
583         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
584                 cli->cl_conn_count);
585
586         if (cli->cl_conn_count == 0) {
587                 CERROR("disconnecting disconnected device (%s)\n",
588                        obd->obd_name);
589                 GOTO(out_disconnect, rc = -EINVAL);
590         }
591
592         cli->cl_conn_count--;
593         if (cli->cl_conn_count != 0)
594                 GOTO(out_disconnect, rc = 0);
595
596         /* Mark import deactivated now, so we don't try to reconnect if any
597          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
598          * fully deactivate the import, or that would drop all requests. */
599         spin_lock(&imp->imp_lock);
600         imp->imp_deactive = 1;
601         spin_unlock(&imp->imp_lock);
602
603         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
604          * delete it regardless.  (It's safe to delete an import that was
605          * never added.) */
606         (void)ptlrpc_pinger_del_import(imp);
607
608         if (obd->obd_namespace != NULL) {
609                 /* obd_force == local only */
610                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
611                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
612                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
613         }
614
615         /* There's no need to hold sem while disconnecting an import,
616          * and it may actually cause deadlock in GSS. */
617         up_write(&cli->cl_sem);
618         rc = ptlrpc_disconnect_import(imp, 0);
619         down_write(&cli->cl_sem);
620
621         ptlrpc_invalidate_import(imp);
622
623         EXIT;
624
625 out_disconnect:
626         /* Use server style - class_disconnect should be always called for
627          * o_disconnect. */
628         err = class_disconnect(exp);
629         if (!rc && err)
630                 rc = err;
631
632         up_write(&cli->cl_sem);
633
634         RETURN(rc);
635 }
636 EXPORT_SYMBOL(client_disconnect_export);
637
638 #ifdef HAVE_SERVER_SUPPORT
639 int server_disconnect_export(struct obd_export *exp)
640 {
641         int rc;
642         ENTRY;
643
644         /* Disconnect early so that clients can't keep using export. */
645         rc = class_disconnect(exp);
646         /* Close import to avoid sending any requests. */
647         if (exp->exp_imp_reverse)
648                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
649
650         if (exp->exp_obd->obd_namespace != NULL)
651                 ldlm_cancel_locks_for_export(exp);
652
653         /* complete all outstanding replies */
654         spin_lock(&exp->exp_lock);
655         while (!list_empty(&exp->exp_outstanding_replies)) {
656                 struct ptlrpc_reply_state *rs =
657                         list_entry(exp->exp_outstanding_replies.next,
658                                        struct ptlrpc_reply_state, rs_exp_list);
659                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
660
661                 spin_lock(&svcpt->scp_rep_lock);
662
663                 list_del_init(&rs->rs_exp_list);
664                 spin_lock(&rs->rs_lock);
665                 ptlrpc_schedule_difficult_reply(rs);
666                 spin_unlock(&rs->rs_lock);
667
668                 spin_unlock(&svcpt->scp_rep_lock);
669         }
670         spin_unlock(&exp->exp_lock);
671
672         RETURN(rc);
673 }
674 EXPORT_SYMBOL(server_disconnect_export);
675
676 /* --------------------------------------------------------------------------
677  * from old lib/target.c
678  * -------------------------------------------------------------------------- */
679
680 static int target_handle_reconnect(struct lustre_handle *conn,
681                                    struct obd_export *exp,
682                                    struct obd_uuid *cluuid)
683 {
684         ENTRY;
685
686         if (exp->exp_connection && exp->exp_imp_reverse) {
687                 struct lustre_handle *hdl;
688                 struct obd_device *target;
689
690                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
691                 target = exp->exp_obd;
692
693                 /* Might be a re-connect after a partition. */
694                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
695                         if (target->obd_recovering) {
696                                 int timeout = cfs_duration_sec(cfs_time_sub(
697                                         cfs_timer_deadline(
698                                         &target->obd_recovery_timer),
699                                         cfs_time_current()));
700
701                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
702                                         "ing, waiting for %d clients in recov"
703                                         "ery for %d:%.02d\n", target->obd_name,
704                                         obd_uuid2str(&exp->exp_client_uuid),
705                                         obd_export_nid2str(exp),
706                                         target->obd_max_recoverable_clients,
707                                         timeout / 60, timeout % 60);
708                         } else {
709                                 LCONSOLE_WARN("%s: Client %s (at %s) "
710                                         "reconnecting\n", target->obd_name,
711                                         obd_uuid2str(&exp->exp_client_uuid),
712                                         obd_export_nid2str(exp));
713                         }
714
715                         conn->cookie = exp->exp_handle.h_cookie;
716                         /* target_handle_connect() treats EALREADY and
717                          * -EALREADY differently.  EALREADY means we are
718                          * doing a valid reconnect from the same client. */
719                         RETURN(EALREADY);
720                 } else {
721                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
722                                       "with handle "LPX64". Rejecting client "
723                                       "with the same UUID trying to reconnect "
724                                       "with handle "LPX64"\n", target->obd_name,
725                                       obd_uuid2str(&exp->exp_client_uuid),
726                                       obd_export_nid2str(exp),
727                                       hdl->cookie, conn->cookie);
728                         memset(conn, 0, sizeof *conn);
729                         /* target_handle_connect() treats EALREADY and
730                          * -EALREADY differently.  -EALREADY is an error
731                          * (same UUID, different handle). */
732                         RETURN(-EALREADY);
733                 }
734         }
735
736         conn->cookie = exp->exp_handle.h_cookie;
737         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
738                cluuid->uuid, exp, conn->cookie);
739         RETURN(0);
740 }
741
742 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
743                           int error)
744 {
745         struct obd_export *exp = cb_data;
746
747         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
748                obd->obd_name, exp->exp_client_uuid.uuid);
749
750         spin_lock(&exp->exp_lock);
751         exp->exp_need_sync = 0;
752         spin_unlock(&exp->exp_lock);
753         class_export_cb_put(exp);
754 }
755 EXPORT_SYMBOL(target_client_add_cb);
756
757 static void
758 check_and_start_recovery_timer(struct obd_device *obd,
759                                struct ptlrpc_request *req, int new_client);
760
761 int target_handle_connect(struct ptlrpc_request *req)
762 {
763         struct obd_device *target = NULL, *targref = NULL;
764         struct obd_export *export = NULL;
765         struct obd_import *revimp;
766         struct obd_import *tmp_imp = NULL;
767         struct lustre_handle conn;
768         struct lustre_handle *tmp;
769         struct obd_uuid tgtuuid;
770         struct obd_uuid cluuid;
771         struct obd_uuid remote_uuid;
772         char *str;
773         int rc = 0;
774         char *target_start;
775         int target_len;
776         bool     mds_conn = false, lw_client = false;
777         struct obd_connect_data *data, *tmpdata;
778         int size, tmpsize;
779         lnet_nid_t *client_nid = NULL;
780         ENTRY;
781
782         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
783
784         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
785         if (str == NULL) {
786                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
787                 GOTO(out, rc = -EINVAL);
788         }
789
790         obd_str2uuid(&tgtuuid, str);
791         target = class_uuid2obd(&tgtuuid);
792         if (!target)
793                 target = class_name2obd(str);
794
795         if (!target) {
796                 deuuidify(str, NULL, &target_start, &target_len);
797                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
798                                    "from %s (no target). If you are running "
799                                    "an HA pair check that the target is "
800                                    "mounted on the other server.\n", str,
801                                    libcfs_nid2str(req->rq_peer.nid));
802                 GOTO(out, rc = -ENODEV);
803         }
804
805         spin_lock(&target->obd_dev_lock);
806         if (target->obd_stopping || !target->obd_set_up) {
807                 spin_unlock(&target->obd_dev_lock);
808
809                 deuuidify(str, NULL, &target_start, &target_len);
810                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
811                               target_len, target_start,
812                               libcfs_nid2str(req->rq_peer.nid),
813                               (target->obd_stopping ?
814                                "stopping" : "not set up"));
815                 GOTO(out, rc = -ENODEV);
816         }
817
818         if (target->obd_no_conn) {
819                 spin_unlock(&target->obd_dev_lock);
820
821                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
822                                "from %s\n", target->obd_name,
823                                libcfs_nid2str(req->rq_peer.nid));
824                 GOTO(out, rc = -EAGAIN);
825         }
826
827         /* Make sure the target isn't cleaned up while we're here. Yes,
828          * there's still a race between the above check and our incref here.
829          * Really, class_uuid2obd should take the ref. */
830         targref = class_incref(target, __FUNCTION__, current);
831
832         target->obd_conn_inprogress++;
833         spin_unlock(&target->obd_dev_lock);
834
835         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
836         if (str == NULL) {
837                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
838                 GOTO(out, rc = -EINVAL);
839         }
840
841         obd_str2uuid(&cluuid, str);
842
843         /* XXX Extract a nettype and format accordingly. */
844         switch (sizeof(lnet_nid_t)) {
845         /* NB the casts only avoid compiler warnings. */
846         case 8:
847                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
848                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
849                 break;
850         case 4:
851                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
852                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
853                 break;
854         default:
855                 LBUG();
856         }
857
858         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
859         if (tmp == NULL)
860                 GOTO(out, rc = -EPROTO);
861
862         conn = *tmp;
863
864         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
865                                     RCL_CLIENT);
866         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
867         if (!data)
868                 GOTO(out, rc = -EPROTO);
869
870         rc = req_capsule_server_pack(&req->rq_pill);
871         if (rc)
872                 GOTO(out, rc);
873
874         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
875                 if (data->ocd_version < LUSTRE_VERSION_CODE -
876                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
877                     data->ocd_version > LUSTRE_VERSION_CODE +
878                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
879                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
880                                   "libclient connection attempt",
881                                   data->ocd_version < LUSTRE_VERSION_CODE ?
882                                   "old" : "new",
883                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
884                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
885                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
886                                   OBD_OCD_VERSION_FIX(data->ocd_version));
887                         data = req_capsule_server_sized_get(&req->rq_pill,
888                                                             &RMF_CONNECT_DATA,
889                                     offsetof(typeof(*data), ocd_version) +
890                                              sizeof(data->ocd_version));
891                         if (data) {
892                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
893                                 data->ocd_version = LUSTRE_VERSION_CODE;
894                         }
895                         GOTO(out, rc = -EPROTO);
896                 }
897         }
898
899         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
900                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
901                 lw_client = (data->ocd_connect_flags &
902                              OBD_CONNECT_LIGHTWEIGHT) != 0;
903
904                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
905                  * for Imperative Recovery connection from MGC to MGS. */
906                 if (!lw_client &&
907                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
908                     !(data->ocd_connect_flags & OBD_CONNECT_IMP_RECOV) &&
909                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
910                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
911                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
912                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
913
914                         /* We do not support the MDT-MDT interoperations with
915                          * different version MDT because of protocol changes. */
916                         if (unlikely(major != LUSTRE_MAJOR ||
917                                      minor != LUSTRE_MINOR ||
918                                      abs(patch - LUSTRE_PATCH) > 3)) {
919                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
920                                         "connection from different version MDT "
921                                         "(%d.%d.%d.%d) %s %s\n",
922                                         target->obd_name, LUSTRE_MAJOR,
923                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
924                                         major, minor, patch,
925                                         OBD_OCD_VERSION_FIX(data->ocd_version),
926                                         libcfs_nid2str(req->rq_peer.nid), str);
927
928                                 GOTO(out, rc = -EPROTO);
929                         }
930                 }
931         }
932
933         /* lctl gets a backstage, all-access pass. */
934         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
935                 goto dont_check_exports;
936
937         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
938         if (!export)
939                 goto no_export;
940
941         /* We've found an export in the hash. */
942
943         spin_lock(&export->exp_lock);
944
945         if (export->exp_connecting) { /* bug 9635, et. al. */
946                 spin_unlock(&export->exp_lock);
947                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
948                               export->exp_obd->obd_name, export,
949                               libcfs_nid2str(req->rq_peer.nid));
950                 class_export_put(export);
951                 export = NULL;
952                 rc = -EALREADY;
953         } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
954                 spin_unlock(&export->exp_lock);
955                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
956                         /* MDS or LWP reconnected after failover. */
957                         LCONSOLE_WARN("%s: Received %s connection from "
958                             "%s, removing former export from %s\n",
959                             target->obd_name, mds_conn ? "MDS" : "LWP",
960                             libcfs_nid2str(req->rq_peer.nid),
961                             libcfs_nid2str(export->exp_connection->c_peer.nid));
962                 else
963                         /* New MDS connection from the same NID. */
964                         LCONSOLE_WARN("%s: Received new %s connection from "
965                                 "%s, removing former export from same NID\n",
966                                 target->obd_name, mds_conn ? "MDS" : "LWP",
967                                 libcfs_nid2str(req->rq_peer.nid));
968                 class_fail_export(export);
969                 class_export_put(export);
970                 export = NULL;
971                 rc = 0;
972         } else if (export->exp_connection != NULL &&
973                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
974                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
975                     MSG_CONNECT_INITIAL)) {
976                 spin_unlock(&export->exp_lock);
977                 /* In MDS failover we have static UUID but NID can change. */
978                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
979                               "existing nid %s is already connected\n",
980                               target->obd_name, cluuid.uuid,
981                               libcfs_nid2str(req->rq_peer.nid),
982                               libcfs_nid2str(
983                                       export->exp_connection->c_peer.nid));
984                 rc = -EALREADY;
985                 class_export_put(export);
986                 export = NULL;
987         } else {
988                 export->exp_connecting = 1;
989                 spin_unlock(&export->exp_lock);
990                 LASSERT(export->exp_obd == target);
991
992                 rc = target_handle_reconnect(&conn, export, &cluuid);
993         }
994
995         /* If we found an export, we already unlocked. */
996         if (!export) {
997 no_export:
998                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
999         } else if (req->rq_export == NULL &&
1000                    atomic_read(&export->exp_rpc_count) > 0) {
1001                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1002                               "still busy with %d references\n",
1003                               target->obd_name, cluuid.uuid,
1004                               libcfs_nid2str(req->rq_peer.nid),
1005                               atomic_read(&export->exp_refcount));
1006                 GOTO(out, rc = -EBUSY);
1007         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1008                 if (!strstr(cluuid.uuid, "mdt"))
1009                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1010                                       "known client %s (at %s) because it "
1011                                       "is indicating it is a new client",
1012                                       target->obd_name, cluuid.uuid,
1013                                       libcfs_nid2str(req->rq_peer.nid));
1014                 GOTO(out, rc = -EALREADY);
1015         } else {
1016                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1017         }
1018
1019         if (rc < 0) {
1020                 GOTO(out, rc);
1021         }
1022
1023         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1024                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1025               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1026               export, (long)cfs_time_current_sec(),
1027               export ? (long)export->exp_last_request_time : 0);
1028
1029         /* If this is the first time a client connects, reset the recovery
1030          * timer. Discard lightweight connections which might be local. */
1031         if (!lw_client && rc == 0 && target->obd_recovering)
1032                 check_and_start_recovery_timer(target, req, export == NULL);
1033
1034         /* We want to handle EALREADY but *not* -EALREADY from
1035          * target_handle_reconnect(), return reconnection state in a flag. */
1036         if (rc == EALREADY) {
1037                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1038                 rc = 0;
1039         } else {
1040                 LASSERT(rc == 0);
1041         }
1042
1043         /* Tell the client if we support replayable requests. */
1044         if (target->obd_replayable)
1045                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1046         client_nid = &req->rq_peer.nid;
1047
1048         if (export == NULL) {
1049                 /* allow lightweight connections during recovery */
1050                 if (target->obd_recovering && !lw_client) {
1051                         cfs_time_t t;
1052                         int     c; /* connected */
1053                         int     i; /* in progress */
1054                         int     k; /* known */
1055                         int     s; /* stale/evicted */
1056
1057                         c = atomic_read(&target->obd_connected_clients);
1058                         i = atomic_read(&target->obd_lock_replay_clients);
1059                         k = target->obd_max_recoverable_clients;
1060                         s = target->obd_stale_clients;
1061                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1062                         t = cfs_time_sub(t, cfs_time_current());
1063                         t = cfs_duration_sec(t);
1064                         LCONSOLE_WARN("%s: Denying connection for new client "
1065                                       "%s (at %s), waiting for all %d known "
1066                                       "clients (%d recovered, %d in progress, "
1067                                       "and %d evicted) to recover in %d:%.02d\n",
1068                                       target->obd_name, cluuid.uuid,
1069                                       libcfs_nid2str(req->rq_peer.nid), k,
1070                                       c - i, i, s, (int)t / 60,
1071                                       (int)t % 60);
1072                         rc = -EBUSY;
1073                 } else {
1074 dont_check_exports:
1075                         rc = obd_connect(req->rq_svc_thread->t_env,
1076                                          &export, target, &cluuid, data,
1077                                          client_nid);
1078                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1079                                 lustre_msg_add_op_flags(req->rq_repmsg,
1080                                                 MSG_CONNECT_RECOVERING);
1081                         if (rc == 0)
1082                                 conn.cookie = export->exp_handle.h_cookie;
1083                 }
1084         } else {
1085                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1086                                    export, target, &cluuid, data, client_nid);
1087         }
1088         if (rc)
1089                 GOTO(out, rc);
1090
1091         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1092         data->ocd_instance = target->u.obt.obt_instance;
1093
1094         /* Return only the parts of obd_connect_data that we understand, so the
1095          * client knows that we don't understand the rest. */
1096         if (data) {
1097                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1098                                                RCL_SERVER);
1099                 tmpdata = req_capsule_server_get(&req->rq_pill,
1100                                                  &RMF_CONNECT_DATA);
1101                 /* Don't use struct assignment here, because the client reply
1102                  * buffer may be smaller/larger than the local struct
1103                  * obd_connect_data. */
1104                 memcpy(tmpdata, data, min(tmpsize, size));
1105         }
1106
1107         /* If all else goes well, this is our RPC return code. */
1108         req->rq_status = 0;
1109
1110         lustre_msg_set_handle(req->rq_repmsg, &conn);
1111
1112         /* If the client and the server are the same node, we will already
1113          * have an export that really points to the client's DLM export,
1114          * because we have a shared handles table.
1115          *
1116          * XXX this will go away when shaver stops sending the "connect" handle
1117          * in the real "remote handle" field of the request --phik 24 Apr 2003
1118          */
1119         ptlrpc_request_change_export(req, export);
1120
1121         spin_lock(&export->exp_lock);
1122         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1123                 spin_unlock(&export->exp_lock);
1124                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1125                        "or equal conn_cnt: %d >= %d\n",
1126                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1127                        export->exp_conn_cnt,
1128                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1129
1130                 GOTO(out, rc = -EALREADY);
1131         }
1132         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1133         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1134
1135         /* Don't evict liblustre clients for not pinging. */
1136         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1137                 export->exp_libclient = 1;
1138                 spin_unlock(&export->exp_lock);
1139
1140                 spin_lock(&target->obd_dev_lock);
1141                 list_del_init(&export->exp_obd_chain_timed);
1142                 spin_unlock(&target->obd_dev_lock);
1143         } else {
1144                 spin_unlock(&export->exp_lock);
1145         }
1146
1147         if (export->exp_connection != NULL) {
1148                 /* Check to see if connection came from another NID. */
1149                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1150                     !hlist_unhashed(&export->exp_nid_hash))
1151                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1152                                      &export->exp_connection->c_peer.nid,
1153                                      &export->exp_nid_hash);
1154
1155                 ptlrpc_connection_put(export->exp_connection);
1156         }
1157
1158         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1159                                                        req->rq_self,
1160                                                        &remote_uuid);
1161         if (hlist_unhashed(&export->exp_nid_hash)) {
1162                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1163                              &export->exp_connection->c_peer.nid,
1164                              &export->exp_nid_hash);
1165         }
1166
1167         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1168                 int has_transno;
1169                 __u64 transno = data->ocd_transno;
1170
1171                 spin_lock(&export->exp_lock);
1172                 /* possible race with class_disconnect_stale_exports,
1173                  * export may be already in the eviction process */
1174                 if (export->exp_failed) {
1175                         spin_unlock(&export->exp_lock);
1176                         GOTO(out, rc = -ENODEV);
1177                 }
1178                 export->exp_in_recovery = 1;
1179                 export->exp_req_replay_needed = 1;
1180                 export->exp_lock_replay_needed = 1;
1181                 spin_unlock(&export->exp_lock);
1182
1183                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1184                                  MSG_CONNECT_TRANSNO);
1185                 if (has_transno && transno == 0)
1186                         CWARN("Connect with zero transno!\n");
1187
1188                 if (has_transno && transno > 0 &&
1189                     transno < target->obd_next_recovery_transno &&
1190                     transno > target->obd_last_committed) {
1191                         /* Another way is to use cmpxchg() to be lock-free. */
1192                         spin_lock(&target->obd_recovery_task_lock);
1193                         if (transno < target->obd_next_recovery_transno)
1194                                 target->obd_next_recovery_transno = transno;
1195                         spin_unlock(&target->obd_recovery_task_lock);
1196                 }
1197
1198                 atomic_inc(&target->obd_req_replay_clients);
1199                 atomic_inc(&target->obd_lock_replay_clients);
1200                 if (atomic_inc_return(&target->obd_connected_clients) ==
1201                     target->obd_max_recoverable_clients)
1202                         wake_up(&target->obd_next_transno_waitq);
1203         }
1204
1205         /* Tell the client we're in recovery, when client is involved in it. */
1206         if (target->obd_recovering && !lw_client)
1207                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1208
1209         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1210         conn = *tmp;
1211
1212         /* Return -ENOTCONN in case of errors to let client reconnect. */
1213         revimp = class_new_import(target);
1214         if (revimp == NULL) {
1215                 CERROR("fail to alloc new reverse import.\n");
1216                 GOTO(out, rc = -ENOTCONN);
1217         }
1218
1219         spin_lock(&export->exp_lock);
1220         if (export->exp_imp_reverse != NULL)
1221                 /* destroyed import can be still referenced in ctxt */
1222                 tmp_imp = export->exp_imp_reverse;
1223         export->exp_imp_reverse = revimp;
1224         spin_unlock(&export->exp_lock);
1225
1226         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1227         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1228         revimp->imp_remote_handle = conn;
1229         revimp->imp_dlm_fake = 1;
1230         revimp->imp_state = LUSTRE_IMP_FULL;
1231
1232         /* Unknown versions will be caught in
1233          * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1234         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1235
1236         if ((data->ocd_connect_flags & OBD_CONNECT_AT) &&
1237             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1238                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1239         else
1240                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1241
1242         if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1243             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1244                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1245         else
1246                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1247
1248         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1249         if (rc) {
1250                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1251                 spin_lock(&export->exp_lock);
1252                 export->exp_imp_reverse = NULL;
1253                 spin_unlock(&export->exp_lock);
1254                 class_destroy_import(revimp);
1255         }
1256
1257         class_import_put(revimp);
1258
1259 out:
1260         if (tmp_imp != NULL)
1261                 client_destroy_import(tmp_imp);
1262         if (export) {
1263                 spin_lock(&export->exp_lock);
1264                 export->exp_connecting = 0;
1265                 spin_unlock(&export->exp_lock);
1266
1267                 class_export_put(export);
1268         }
1269         if (targref) {
1270                 spin_lock(&target->obd_dev_lock);
1271                 target->obd_conn_inprogress--;
1272                 spin_unlock(&target->obd_dev_lock);
1273
1274                 class_decref(targref, __func__, current);
1275         }
1276         if (rc)
1277                 req->rq_status = rc;
1278         RETURN(rc);
1279 }
1280 EXPORT_SYMBOL(target_handle_connect);
1281
1282 int target_handle_disconnect(struct ptlrpc_request *req)
1283 {
1284         int rc;
1285         ENTRY;
1286
1287         rc = req_capsule_server_pack(&req->rq_pill);
1288         if (rc)
1289                 RETURN(rc);
1290
1291         /* Keep the rq_export around so we can send the reply. */
1292         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1293
1294         RETURN(0);
1295 }
1296 EXPORT_SYMBOL(target_handle_disconnect);
1297
1298 void target_destroy_export(struct obd_export *exp)
1299 {
1300         struct obd_import       *imp = NULL;
1301         /* exports created from last_rcvd data, and "fake"
1302            exports created by lctl don't have an import */
1303         spin_lock(&exp->exp_lock);
1304         if (exp->exp_imp_reverse != NULL) {
1305                 imp = exp->exp_imp_reverse;
1306                 exp->exp_imp_reverse = NULL;
1307         }
1308         spin_unlock(&exp->exp_lock);
1309         if (imp != NULL)
1310                 client_destroy_import(imp);
1311
1312         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1313         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1314         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1315         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1316 }
1317 EXPORT_SYMBOL(target_destroy_export);
1318
1319 /*
1320  * Recovery functions
1321  */
1322 static void target_request_copy_get(struct ptlrpc_request *req)
1323 {
1324         class_export_rpc_inc(req->rq_export);
1325         LASSERT(list_empty(&req->rq_list));
1326         INIT_LIST_HEAD(&req->rq_replay_list);
1327
1328         /* Increase refcount to keep request in queue. */
1329         atomic_inc(&req->rq_refcount);
1330         /* Let export know it has replays to be handled. */
1331         atomic_inc(&req->rq_export->exp_replay_count);
1332 }
1333
1334 static void target_request_copy_put(struct ptlrpc_request *req)
1335 {
1336         LASSERT(list_empty(&req->rq_replay_list));
1337         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1338
1339         atomic_dec(&req->rq_export->exp_replay_count);
1340         class_export_rpc_dec(req->rq_export);
1341         ptlrpc_server_drop_request(req);
1342 }
1343
1344 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1345 {
1346         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1347         struct obd_export     *exp = req->rq_export;
1348         struct ptlrpc_request *reqiter;
1349         int                    dup = 0;
1350
1351         LASSERT(exp);
1352
1353         spin_lock(&exp->exp_lock);
1354         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1355                                 rq_replay_list) {
1356                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1357                         dup = 1;
1358                         break;
1359                 }
1360         }
1361
1362         if (dup) {
1363                 /* We expect it with RESENT and REPLAY flags. */
1364                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1365                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1366                         CERROR("invalid flags %x of resent replay\n",
1367                                lustre_msg_get_flags(req->rq_reqmsg));
1368         } else {
1369                 list_add_tail(&req->rq_replay_list,
1370                                   &exp->exp_req_replay_queue);
1371         }
1372
1373         spin_unlock(&exp->exp_lock);
1374         return dup;
1375 }
1376
1377 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1378 {
1379         LASSERT(!list_empty(&req->rq_replay_list));
1380         LASSERT(req->rq_export);
1381
1382         spin_lock(&req->rq_export->exp_lock);
1383         list_del_init(&req->rq_replay_list);
1384         spin_unlock(&req->rq_export->exp_lock);
1385 }
1386
1387 static void target_finish_recovery(struct obd_device *obd)
1388 {
1389         ENTRY;
1390
1391         /* Only log a recovery message when recovery has occurred. */
1392         if (obd->obd_recovery_start) {
1393                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1394                                         obd->obd_recovery_start);
1395                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1396                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1397                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1398                         obd->obd_max_recoverable_clients,
1399                         atomic_read(&obd->obd_connected_clients),
1400                         obd->obd_stale_clients,
1401                         obd->obd_stale_clients == 1 ? "was" : "were");
1402         }
1403
1404         ldlm_reprocess_all_ns(obd->obd_namespace);
1405         spin_lock(&obd->obd_recovery_task_lock);
1406         if (!list_empty(&obd->obd_req_replay_queue) ||
1407             !list_empty(&obd->obd_lock_replay_queue) ||
1408             !list_empty(&obd->obd_final_req_queue)) {
1409                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1410                        obd->obd_name,
1411                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1412                        list_empty(&obd->obd_lock_replay_queue) ? \
1413                                "" : "lock ",
1414                        list_empty(&obd->obd_final_req_queue) ? \
1415                                "" : "final ");
1416                 spin_unlock(&obd->obd_recovery_task_lock);
1417                 LBUG();
1418         }
1419         spin_unlock(&obd->obd_recovery_task_lock);
1420
1421         obd->obd_recovery_end = cfs_time_current_sec();
1422
1423         /* When recovery finished, cleanup orphans on MDS and OST. */
1424         if (OBT(obd) && OBP(obd, postrecov)) {
1425                 int rc = OBP(obd, postrecov)(obd);
1426                 if (rc < 0)
1427                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1428                                       obd->obd_name, rc);
1429         }
1430         EXIT;
1431 }
1432
1433 static void abort_req_replay_queue(struct obd_device *obd)
1434 {
1435         struct ptlrpc_request *req, *n;
1436         struct list_head abort_list;
1437
1438         INIT_LIST_HEAD(&abort_list);
1439         spin_lock(&obd->obd_recovery_task_lock);
1440         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1441         spin_unlock(&obd->obd_recovery_task_lock);
1442         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1443                 DEBUG_REQ(D_WARNING, req, "aborted:");
1444                 req->rq_status = -ENOTCONN;
1445                 if (ptlrpc_error(req)) {
1446                         DEBUG_REQ(D_ERROR, req,
1447                                   "failed abort_req_reply; skipping");
1448                 }
1449                 target_exp_dequeue_req_replay(req);
1450                 target_request_copy_put(req);
1451         }
1452 }
1453
1454 static void abort_lock_replay_queue(struct obd_device *obd)
1455 {
1456         struct ptlrpc_request *req, *n;
1457         struct list_head abort_list;
1458
1459         INIT_LIST_HEAD(&abort_list);
1460         spin_lock(&obd->obd_recovery_task_lock);
1461         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1462         spin_unlock(&obd->obd_recovery_task_lock);
1463         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1464                 DEBUG_REQ(D_ERROR, req, "aborted:");
1465                 req->rq_status = -ENOTCONN;
1466                 if (ptlrpc_error(req)) {
1467                         DEBUG_REQ(D_ERROR, req,
1468                                   "failed abort_lock_reply; skipping");
1469                 }
1470                 target_request_copy_put(req);
1471         }
1472 }
1473
1474 /* Called from a cleanup function if the device is being cleaned up
1475    forcefully.  The exports should all have been disconnected already,
1476    the only thing left to do is
1477      - clear the recovery flags
1478      - cancel the timer
1479      - free queued requests and replies, but don't send replies
1480    Because the obd_stopping flag is set, no new requests should be received.
1481
1482 */
1483 void target_cleanup_recovery(struct obd_device *obd)
1484 {
1485         struct ptlrpc_request *req, *n;
1486         struct list_head clean_list;
1487         ENTRY;
1488
1489         INIT_LIST_HEAD(&clean_list);
1490         spin_lock(&obd->obd_dev_lock);
1491         if (!obd->obd_recovering) {
1492                 spin_unlock(&obd->obd_dev_lock);
1493                 EXIT;
1494                 return;
1495         }
1496         obd->obd_recovering = obd->obd_abort_recovery = 0;
1497         spin_unlock(&obd->obd_dev_lock);
1498
1499         spin_lock(&obd->obd_recovery_task_lock);
1500         target_cancel_recovery_timer(obd);
1501         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1502         spin_unlock(&obd->obd_recovery_task_lock);
1503
1504         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1505                 LASSERT(req->rq_reply_state == NULL);
1506                 target_exp_dequeue_req_replay(req);
1507                 target_request_copy_put(req);
1508         }
1509
1510         spin_lock(&obd->obd_recovery_task_lock);
1511         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1512         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1513         spin_unlock(&obd->obd_recovery_task_lock);
1514
1515         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1516                 LASSERT(req->rq_reply_state == NULL);
1517                 target_request_copy_put(req);
1518         }
1519
1520         EXIT;
1521 }
1522 EXPORT_SYMBOL(target_cleanup_recovery);
1523
1524 /* obd_recovery_task_lock should be held */
1525 void target_cancel_recovery_timer(struct obd_device *obd)
1526 {
1527         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1528         cfs_timer_disarm(&obd->obd_recovery_timer);
1529 }
1530 EXPORT_SYMBOL(target_cancel_recovery_timer);
1531
1532 static void target_start_recovery_timer(struct obd_device *obd)
1533 {
1534         if (obd->obd_recovery_start != 0)
1535                 return;
1536
1537         spin_lock(&obd->obd_dev_lock);
1538         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1539                 spin_unlock(&obd->obd_dev_lock);
1540                 return;
1541         }
1542
1543         LASSERT(obd->obd_recovery_timeout != 0);
1544
1545         if (obd->obd_recovery_start != 0) {
1546                 spin_unlock(&obd->obd_dev_lock);
1547                 return;
1548         }
1549
1550         cfs_timer_arm(&obd->obd_recovery_timer,
1551                       cfs_time_shift(obd->obd_recovery_timeout));
1552         obd->obd_recovery_start = cfs_time_current_sec();
1553         spin_unlock(&obd->obd_dev_lock);
1554
1555         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1556                       "or until %d client%s reconnect%s\n",
1557                       obd->obd_name,
1558                       obd->obd_recovery_timeout / 60,
1559                       obd->obd_recovery_timeout % 60,
1560                       obd->obd_max_recoverable_clients,
1561                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1562                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1563 }
1564
1565 /**
1566  * extend recovery window.
1567  *
1568  * if @extend is true, extend recovery window to have @drt remaining at least;
1569  * otherwise, make sure the recovery timeout value is not less than @drt.
1570  */
1571 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1572 {
1573         cfs_time_t now;
1574         cfs_time_t end;
1575         cfs_duration_t left;
1576         int to;
1577
1578         spin_lock(&obd->obd_dev_lock);
1579         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1580                 spin_unlock(&obd->obd_dev_lock);
1581                 return;
1582         }
1583         LASSERT(obd->obd_recovery_start != 0);
1584
1585         now  = cfs_time_current_sec();
1586         to   = obd->obd_recovery_timeout;
1587         end  = obd->obd_recovery_start + to;
1588         left = cfs_time_sub(end, now);
1589
1590         if (extend && (drt > left)) {
1591                 to += drt - left;
1592         } else if (!extend && (drt > to)) {
1593                 to = drt;
1594         }
1595
1596         if (to > obd->obd_recovery_time_hard)
1597                 to = obd->obd_recovery_time_hard;
1598         if (obd->obd_recovery_timeout < to) {
1599                 obd->obd_recovery_timeout = to;
1600                 end = obd->obd_recovery_start + to;
1601                 cfs_timer_arm(&obd->obd_recovery_timer,
1602                                 cfs_time_shift(end - now));
1603         }
1604         spin_unlock(&obd->obd_dev_lock);
1605
1606         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1607                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1608 }
1609
1610 /* Reset the timer with each new client connection */
1611 /*
1612  * This timer is actually reconnect_timer, which is for making sure
1613  * the total recovery window is at least as big as my reconnect
1614  * attempt timing. So the initial recovery time_out will be set to
1615  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1616  * from client is bigger than this, then the recovery time_out will
1617  * be extended to make sure the client could be reconnected, in the
1618  * process, the timeout from the new client should be ignored.
1619  */
1620
1621 static void
1622 check_and_start_recovery_timer(struct obd_device *obd,
1623                                struct ptlrpc_request *req,
1624                                int new_client)
1625 {
1626         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1627         struct obd_device_target *obt = &obd->u.obt;
1628
1629         if (!new_client && service_time)
1630                 /* Teach server about old server's estimates, as first guess
1631                  * at how long new requests will take. */
1632                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1633                             service_time);
1634
1635         target_start_recovery_timer(obd);
1636
1637         /* Convert the service time to RPC timeout,
1638          * and reuse service_time to limit stack usage. */
1639         service_time = at_est2timeout(service_time);
1640
1641         /* We expect other clients to timeout within service_time, then try
1642          * to reconnect, then try the failover server.  The max delay between
1643          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1644         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1645
1646         LASSERT(obt->obt_magic == OBT_MAGIC);
1647         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1648         if (service_time > obd->obd_recovery_timeout && !new_client)
1649                 extend_recovery_timer(obd, service_time, false);
1650 }
1651
1652 /** Health checking routines */
1653 static inline int exp_connect_healthy(struct obd_export *exp)
1654 {
1655         return (exp->exp_in_recovery);
1656 }
1657
1658 /** if export done req_replay or has replay in queue */
1659 static inline int exp_req_replay_healthy(struct obd_export *exp)
1660 {
1661         return (!exp->exp_req_replay_needed ||
1662                 atomic_read(&exp->exp_replay_count) > 0);
1663 }
1664 /** if export done lock_replay or has replay in queue */
1665 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1666 {
1667         return (!exp->exp_lock_replay_needed ||
1668                 atomic_read(&exp->exp_replay_count) > 0);
1669 }
1670
1671 static inline int exp_vbr_healthy(struct obd_export *exp)
1672 {
1673         return (!exp->exp_vbr_failed);
1674 }
1675
1676 static inline int exp_finished(struct obd_export *exp)
1677 {
1678         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1679 }
1680
1681 /** Checking routines for recovery */
1682 static int check_for_clients(struct obd_device *obd)
1683 {
1684         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
1685
1686         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1687                 return 1;
1688         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1689         return (clnts + obd->obd_stale_clients ==
1690                 obd->obd_max_recoverable_clients);
1691 }
1692
1693 static int check_for_next_transno(struct obd_device *obd)
1694 {
1695         struct ptlrpc_request *req = NULL;
1696         int wake_up = 0, connected, completed, queue_len;
1697         __u64 next_transno, req_transno;
1698         ENTRY;
1699
1700         spin_lock(&obd->obd_recovery_task_lock);
1701         if (!list_empty(&obd->obd_req_replay_queue)) {
1702                 req = list_entry(obd->obd_req_replay_queue.next,
1703                                      struct ptlrpc_request, rq_list);
1704                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1705         } else {
1706                 req_transno = 0;
1707         }
1708
1709         connected = atomic_read(&obd->obd_connected_clients);
1710         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1711         queue_len = obd->obd_requests_queued_for_recovery;
1712         next_transno = obd->obd_next_recovery_transno;
1713
1714         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1715                "req_transno: "LPU64", next_transno: "LPU64"\n",
1716                obd->obd_max_recoverable_clients, connected, completed,
1717                queue_len, req_transno, next_transno);
1718
1719         if (obd->obd_abort_recovery) {
1720                 CDEBUG(D_HA, "waking for aborted recovery\n");
1721                 wake_up = 1;
1722         } else if (obd->obd_recovery_expired) {
1723                 CDEBUG(D_HA, "waking for expired recovery\n");
1724                 wake_up = 1;
1725         } else if (req_transno == next_transno) {
1726                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1727                 wake_up = 1;
1728         } else if (queue_len > 0 &&
1729                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1730                 int d_lvl = D_HA;
1731                 /** handle gaps occured due to lost reply or VBR */
1732                 LASSERTF(req_transno >= next_transno,
1733                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1734                          req_transno, next_transno);
1735                 if (req_transno > obd->obd_last_committed &&
1736                     !obd->obd_version_recov)
1737                         d_lvl = D_ERROR;
1738                 CDEBUG(d_lvl,
1739                        "%s: waking for gap in transno, VBR is %s (skip: "
1740                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1741                        ", last_committed: "LPD64")\n",
1742                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1743                        next_transno, queue_len, completed, connected,
1744                        req_transno, obd->obd_last_committed);
1745                 obd->obd_next_recovery_transno = req_transno;
1746                 wake_up = 1;
1747         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1748                 CDEBUG(D_HA, "waking for completed recovery\n");
1749                 wake_up = 1;
1750         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1751                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1752                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1753                 obd->obd_next_recovery_transno = req_transno;
1754                 wake_up = 1;
1755         }
1756         spin_unlock(&obd->obd_recovery_task_lock);
1757         return wake_up;
1758 }
1759
1760 static int check_for_next_lock(struct obd_device *obd)
1761 {
1762         int wake_up = 0;
1763
1764         spin_lock(&obd->obd_recovery_task_lock);
1765         if (!list_empty(&obd->obd_lock_replay_queue)) {
1766                 CDEBUG(D_HA, "waking for next lock\n");
1767                 wake_up = 1;
1768         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1769                 CDEBUG(D_HA, "waking for completed lock replay\n");
1770                 wake_up = 1;
1771         } else if (obd->obd_abort_recovery) {
1772                 CDEBUG(D_HA, "waking for aborted recovery\n");
1773                 wake_up = 1;
1774         } else if (obd->obd_recovery_expired) {
1775                 CDEBUG(D_HA, "waking for expired recovery\n");
1776                 wake_up = 1;
1777         }
1778         spin_unlock(&obd->obd_recovery_task_lock);
1779
1780         return wake_up;
1781 }
1782
1783 /**
1784  * wait for recovery events,
1785  * check its status with help of check_routine
1786  * evict dead clients via health_check
1787  */
1788 static int target_recovery_overseer(struct obd_device *obd,
1789                                     int (*check_routine)(struct obd_device *),
1790                                     int (*health_check)(struct obd_export *))
1791 {
1792 repeat:
1793         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1794               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1795                 CWARN("recovery is aborted by hard timeout\n");
1796                 obd->obd_abort_recovery = 1;
1797         }
1798
1799         while (wait_event_timeout(obd->obd_next_transno_waitq,
1800                                   check_routine(obd),
1801                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
1802                 /* wait indefinitely for event, but don't trigger watchdog */;
1803
1804         if (obd->obd_abort_recovery) {
1805                 CWARN("recovery is aborted, evict exports in recovery\n");
1806                 /** evict exports which didn't finish recovery yet */
1807                 class_disconnect_stale_exports(obd, exp_finished);
1808                 return 1;
1809         } else if (obd->obd_recovery_expired) {
1810                 obd->obd_recovery_expired = 0;
1811                 /** If some clients died being recovered, evict them */
1812                 LCONSOLE_WARN("%s: recovery is timed out, "
1813                               "evict stale exports\n", obd->obd_name);
1814                 /** evict cexports with no replay in queue, they are stalled */
1815                 class_disconnect_stale_exports(obd, health_check);
1816                 /** continue with VBR */
1817                 spin_lock(&obd->obd_dev_lock);
1818                 obd->obd_version_recov = 1;
1819                 spin_unlock(&obd->obd_dev_lock);
1820                 /**
1821                  * reset timer, recovery will proceed with versions now,
1822                  * timeout is set just to handle reconnection delays
1823                  */
1824                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1825                 /** Wait for recovery events again, after evicting bad clients */
1826                 goto repeat;
1827         }
1828         return 0;
1829 }
1830
1831 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1832 {
1833         struct ptlrpc_request *req = NULL;
1834         ENTRY;
1835
1836         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1837                 obd->obd_next_recovery_transno);
1838
1839         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
1840         /** It is needed to extend recovery window above recovery_time_soft.
1841          *  Extending is possible only in the end of recovery window
1842          *  (see more details in handle_recovery_req).
1843          */
1844         CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
1845
1846         if (target_recovery_overseer(obd, check_for_next_transno,
1847                                      exp_req_replay_healthy)) {
1848                 abort_req_replay_queue(obd);
1849                 abort_lock_replay_queue(obd);
1850         }
1851
1852         spin_lock(&obd->obd_recovery_task_lock);
1853         if (!list_empty(&obd->obd_req_replay_queue)) {
1854                 req = list_entry(obd->obd_req_replay_queue.next,
1855                                      struct ptlrpc_request, rq_list);
1856                 list_del_init(&req->rq_list);
1857                 obd->obd_requests_queued_for_recovery--;
1858                 spin_unlock(&obd->obd_recovery_task_lock);
1859         } else {
1860                 spin_unlock(&obd->obd_recovery_task_lock);
1861                 LASSERT(list_empty(&obd->obd_req_replay_queue));
1862                 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1863                 /** evict exports failed VBR */
1864                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1865         }
1866         RETURN(req);
1867 }
1868
1869 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1870 {
1871         struct ptlrpc_request *req = NULL;
1872
1873         CDEBUG(D_HA, "Waiting for lock\n");
1874         if (target_recovery_overseer(obd, check_for_next_lock,
1875                                      exp_lock_replay_healthy))
1876                 abort_lock_replay_queue(obd);
1877
1878         spin_lock(&obd->obd_recovery_task_lock);
1879         if (!list_empty(&obd->obd_lock_replay_queue)) {
1880                 req = list_entry(obd->obd_lock_replay_queue.next,
1881                                      struct ptlrpc_request, rq_list);
1882                 list_del_init(&req->rq_list);
1883                 spin_unlock(&obd->obd_recovery_task_lock);
1884         } else {
1885                 spin_unlock(&obd->obd_recovery_task_lock);
1886                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1887                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1888                 /** evict exports failed VBR */
1889                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1890         }
1891         return req;
1892 }
1893
1894 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1895 {
1896         struct ptlrpc_request *req = NULL;
1897
1898         spin_lock(&obd->obd_recovery_task_lock);
1899         if (!list_empty(&obd->obd_final_req_queue)) {
1900                 req = list_entry(obd->obd_final_req_queue.next,
1901                                      struct ptlrpc_request, rq_list);
1902                 list_del_init(&req->rq_list);
1903                 spin_unlock(&obd->obd_recovery_task_lock);
1904                 if (req->rq_export->exp_in_recovery) {
1905                         spin_lock(&req->rq_export->exp_lock);
1906                         req->rq_export->exp_in_recovery = 0;
1907                         spin_unlock(&req->rq_export->exp_lock);
1908                 }
1909         } else {
1910                 spin_unlock(&obd->obd_recovery_task_lock);
1911         }
1912         return req;
1913 }
1914
1915 static void handle_recovery_req(struct ptlrpc_thread *thread,
1916                                 struct ptlrpc_request *req,
1917                                 svc_handler_t handler)
1918 {
1919         ENTRY;
1920
1921         /**
1922          * export can be evicted during recovery, no need to handle replays for
1923          * it after that, discard such request silently
1924          */
1925         if (req->rq_export->exp_disconnected)
1926                 RETURN_EXIT;
1927
1928         req->rq_session.lc_thread = thread;
1929         req->rq_svc_thread = thread;
1930         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1931
1932         /* thread context */
1933         lu_context_enter(&thread->t_env->le_ctx);
1934         (void)handler(req);
1935         lu_context_exit(&thread->t_env->le_ctx);
1936
1937         /* don't reset timer for final stage */
1938         if (!exp_finished(req->rq_export)) {
1939                 int to = obd_timeout;
1940
1941                 /**
1942                  * Add request timeout to the recovery time so next request from
1943                  * this client may come in recovery time
1944                  */
1945                 if (!AT_OFF) {
1946                         struct ptlrpc_service_part *svcpt;
1947
1948                         svcpt = req->rq_rqbd->rqbd_svcpt;
1949                         /* If the server sent early reply for this request,
1950                          * the client will recalculate the timeout according to
1951                          * current server estimate service time, so we will
1952                          * use the maxium timeout here for waiting the client
1953                          * sending the next req */
1954                         to = max((int)at_est2timeout(
1955                                  at_get(&svcpt->scp_at_estimate)),
1956                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1957                         /* Add 2 net_latency, one for balance rq_deadline
1958                          * (see ptl_send_rpc), one for resend the req to server,
1959                          * Note: client will pack net_latency in replay req
1960                          * (see ptlrpc_replay_req) */
1961                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
1962                 }
1963                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1964         }
1965         EXIT;
1966 }
1967
1968 static int target_recovery_thread(void *arg)
1969 {
1970         struct lu_target *lut = arg;
1971         struct obd_device *obd = lut->lut_obd;
1972         struct ptlrpc_request *req;
1973         struct target_recovery_data *trd = &obd->obd_recovery_data;
1974         unsigned long delta;
1975         struct lu_env *env;
1976         struct ptlrpc_thread *thread = NULL;
1977         int rc = 0;
1978         ENTRY;
1979
1980         unshare_fs_struct();
1981         OBD_ALLOC_PTR(thread);
1982         if (thread == NULL)
1983                 RETURN(-ENOMEM);
1984
1985         OBD_ALLOC_PTR(env);
1986         if (env == NULL) {
1987                 OBD_FREE_PTR(thread);
1988                 RETURN(-ENOMEM);
1989         }
1990
1991         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1992         if (rc) {
1993                 OBD_FREE_PTR(thread);
1994                 OBD_FREE_PTR(env);
1995                 RETURN(rc);
1996         }
1997
1998         thread->t_env = env;
1999         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2000         env->le_ctx.lc_thread = thread;
2001         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2002         thread->t_watchdog = NULL;
2003
2004         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2005                current_pid());
2006         trd->trd_processing_task = current_pid();
2007
2008         spin_lock(&obd->obd_dev_lock);
2009         obd->obd_recovering = 1;
2010         spin_unlock(&obd->obd_dev_lock);
2011         complete(&trd->trd_starting);
2012
2013         /* first of all, we have to know the first transno to replay */
2014         if (target_recovery_overseer(obd, check_for_clients,
2015                                      exp_connect_healthy)) {
2016                 abort_req_replay_queue(obd);
2017                 abort_lock_replay_queue(obd);
2018         }
2019
2020         /* next stage: replay requests */
2021         delta = jiffies;
2022         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2023                atomic_read(&obd->obd_req_replay_clients),
2024                obd->obd_next_recovery_transno);
2025         while ((req = target_next_replay_req(obd))) {
2026                 LASSERT(trd->trd_processing_task == current_pid());
2027                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2028                           lustre_msg_get_transno(req->rq_reqmsg),
2029                           libcfs_nid2str(req->rq_peer.nid));
2030                 handle_recovery_req(thread, req,
2031                                     trd->trd_recovery_handler);
2032                 /**
2033                  * bz18031: increase next_recovery_transno before
2034                  * target_request_copy_put() will drop exp_rpc reference
2035                  */
2036                 spin_lock(&obd->obd_recovery_task_lock);
2037                 obd->obd_next_recovery_transno++;
2038                 spin_unlock(&obd->obd_recovery_task_lock);
2039                 target_exp_dequeue_req_replay(req);
2040                 target_request_copy_put(req);
2041                 obd->obd_replayed_requests++;
2042         }
2043
2044         /**
2045          * The second stage: replay locks
2046          */
2047         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2048                atomic_read(&obd->obd_lock_replay_clients));
2049         while ((req = target_next_replay_lock(obd))) {
2050                 LASSERT(trd->trd_processing_task == current_pid());
2051                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2052                           libcfs_nid2str(req->rq_peer.nid));
2053                 handle_recovery_req(thread, req,
2054                                     trd->trd_recovery_handler);
2055                 target_request_copy_put(req);
2056                 obd->obd_replayed_locks++;
2057         }
2058
2059         /**
2060          * The third stage: reply on final pings, at this moment all clients
2061          * must have request in final queue
2062          */
2063         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2064         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2065         /** Update server last boot epoch */
2066         tgt_boot_epoch_update(lut);
2067         /* We drop recoverying flag to forward all new requests
2068          * to regular mds_handle() since now */
2069         spin_lock(&obd->obd_dev_lock);
2070         obd->obd_recovering = obd->obd_abort_recovery = 0;
2071         spin_unlock(&obd->obd_dev_lock);
2072         spin_lock(&obd->obd_recovery_task_lock);
2073         target_cancel_recovery_timer(obd);
2074         spin_unlock(&obd->obd_recovery_task_lock);
2075         while ((req = target_next_final_ping(obd))) {
2076                 LASSERT(trd->trd_processing_task == current_pid());
2077                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2078                           libcfs_nid2str(req->rq_peer.nid));
2079                 handle_recovery_req(thread, req,
2080                                     trd->trd_recovery_handler);
2081                 /* Because the waiting client can not send ping to server,
2082                  * so we need refresh the last_request_time, to avoid the
2083                  * export is being evicted */
2084                 ptlrpc_update_export_timer(req->rq_export, 0);
2085                 target_request_copy_put(req);
2086         }
2087
2088         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2089         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2090               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2091         if (delta > OBD_RECOVERY_TIME_SOFT) {
2092                 CWARN("too long recovery - read logs\n");
2093                 libcfs_debug_dumplog();
2094         }
2095
2096         target_finish_recovery(obd);
2097
2098         lu_context_fini(&env->le_ctx);
2099         trd->trd_processing_task = 0;
2100         complete(&trd->trd_finishing);
2101
2102         tgt_io_thread_done(thread);
2103         OBD_FREE_PTR(thread);
2104         OBD_FREE_PTR(env);
2105         RETURN(rc);
2106 }
2107
2108 static int target_start_recovery_thread(struct lu_target *lut,
2109                                         svc_handler_t handler)
2110 {
2111         struct obd_device *obd = lut->lut_obd;
2112         int rc = 0;
2113         struct target_recovery_data *trd = &obd->obd_recovery_data;
2114
2115         memset(trd, 0, sizeof(*trd));
2116         init_completion(&trd->trd_starting);
2117         init_completion(&trd->trd_finishing);
2118         trd->trd_recovery_handler = handler;
2119
2120         if (!IS_ERR(kthread_run(target_recovery_thread,
2121                                 lut, "tgt_recov"))) {
2122                 wait_for_completion(&trd->trd_starting);
2123                 LASSERT(obd->obd_recovering != 0);
2124         } else {
2125                 rc = -ECHILD;
2126         }
2127
2128         return rc;
2129 }
2130
2131 void target_stop_recovery_thread(struct obd_device *obd)
2132 {
2133         if (obd->obd_recovery_data.trd_processing_task > 0) {
2134                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2135                 /** recovery can be done but postrecovery is not yet */
2136                 spin_lock(&obd->obd_dev_lock);
2137                 if (obd->obd_recovering) {
2138                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2139                         obd->obd_abort_recovery = 1;
2140                         wake_up(&obd->obd_next_transno_waitq);
2141                 }
2142                 spin_unlock(&obd->obd_dev_lock);
2143                 wait_for_completion(&trd->trd_finishing);
2144         }
2145 }
2146 EXPORT_SYMBOL(target_stop_recovery_thread);
2147
2148 void target_recovery_fini(struct obd_device *obd)
2149 {
2150         class_disconnect_exports(obd);
2151         target_stop_recovery_thread(obd);
2152         target_cleanup_recovery(obd);
2153 }
2154 EXPORT_SYMBOL(target_recovery_fini);
2155
2156 static void target_recovery_expired(unsigned long castmeharder)
2157 {
2158         struct obd_device *obd = (struct obd_device *)castmeharder;
2159         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2160                " after %lds (%d clients connected)\n",
2161                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2162                cfs_time_current_sec()- obd->obd_recovery_start,
2163                atomic_read(&obd->obd_connected_clients));
2164
2165         obd->obd_recovery_expired = 1;
2166         wake_up(&obd->obd_next_transno_waitq);
2167 }
2168
2169 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2170 {
2171         struct obd_device *obd = lut->lut_obd;
2172         if (obd->obd_max_recoverable_clients == 0) {
2173                 /** Update server last boot epoch */
2174                 tgt_boot_epoch_update(lut);
2175                 return;
2176         }
2177
2178         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2179                "last_transno "LPU64"\n", obd->obd_name,
2180                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2181         LASSERT(obd->obd_stopping == 0);
2182         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2183         obd->obd_recovery_start = 0;
2184         obd->obd_recovery_end = 0;
2185
2186         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2187         target_start_recovery_thread(lut, handler);
2188 }
2189 EXPORT_SYMBOL(target_recovery_init);
2190
2191
2192 static int target_process_req_flags(struct obd_device *obd,
2193                                     struct ptlrpc_request *req)
2194 {
2195         struct obd_export *exp = req->rq_export;
2196         LASSERT(exp != NULL);
2197         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2198                 /* client declares he's ready to replay locks */
2199                 spin_lock(&exp->exp_lock);
2200                 if (exp->exp_req_replay_needed) {
2201                         exp->exp_req_replay_needed = 0;
2202                         spin_unlock(&exp->exp_lock);
2203
2204                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2205                         atomic_dec(&obd->obd_req_replay_clients);
2206                 } else {
2207                         spin_unlock(&exp->exp_lock);
2208                 }
2209         }
2210         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2211                 /* client declares he's ready to complete recovery
2212                  * so, we put the request on th final queue */
2213                 spin_lock(&exp->exp_lock);
2214                 if (exp->exp_lock_replay_needed) {
2215                         exp->exp_lock_replay_needed = 0;
2216                         spin_unlock(&exp->exp_lock);
2217
2218                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2219                         atomic_dec(&obd->obd_lock_replay_clients);
2220                 } else {
2221                         spin_unlock(&exp->exp_lock);
2222                 }
2223         }
2224         return 0;
2225 }
2226
2227 int target_queue_recovery_request(struct ptlrpc_request *req,
2228                                   struct obd_device *obd)
2229 {
2230         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2231         struct ptlrpc_request *reqiter;
2232         int inserted = 0;
2233         ENTRY;
2234
2235         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2236                 /* Processing the queue right now, don't re-add. */
2237                 RETURN(1);
2238         }
2239
2240         target_process_req_flags(obd, req);
2241
2242         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2243                 /* client declares he's ready to complete recovery
2244                  * so, we put the request on th final queue */
2245                 target_request_copy_get(req);
2246                 DEBUG_REQ(D_HA, req, "queue final req");
2247                 wake_up(&obd->obd_next_transno_waitq);
2248                 spin_lock(&obd->obd_recovery_task_lock);
2249                 if (obd->obd_recovering) {
2250                         list_add_tail(&req->rq_list,
2251                                           &obd->obd_final_req_queue);
2252                 } else {
2253                         spin_unlock(&obd->obd_recovery_task_lock);
2254                         target_request_copy_put(req);
2255                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2256                 }
2257                 spin_unlock(&obd->obd_recovery_task_lock);
2258                 RETURN(0);
2259         }
2260         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2261                 /* client declares he's ready to replay locks */
2262                 target_request_copy_get(req);
2263                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2264                 wake_up(&obd->obd_next_transno_waitq);
2265                 spin_lock(&obd->obd_recovery_task_lock);
2266                 LASSERT(obd->obd_recovering);
2267                 /* usually due to recovery abort */
2268                 if (!req->rq_export->exp_in_recovery) {
2269                         spin_unlock(&obd->obd_recovery_task_lock);
2270                         target_request_copy_put(req);
2271                         RETURN(-ENOTCONN);
2272                 }
2273                 LASSERT(req->rq_export->exp_lock_replay_needed);
2274                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2275                 spin_unlock(&obd->obd_recovery_task_lock);
2276                 RETURN(0);
2277         }
2278
2279         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2280          * (i.e. buflens etc are in my own byte order), but type-dependent
2281          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2282
2283         if (!transno) {
2284                 INIT_LIST_HEAD(&req->rq_list);
2285                 DEBUG_REQ(D_HA, req, "not queueing");
2286                 RETURN(1);
2287         }
2288
2289         /* If we're processing the queue, we want don't want to queue this
2290          * message.
2291          *
2292          * Also, if this request has a transno less than the one we're waiting
2293          * for, we should process it now.  It could (and currently always will)
2294          * be an open request for a descriptor that was opened some time ago.
2295          *
2296          * Also, a resent, replayed request that has already been
2297          * handled will pass through here and be processed immediately.
2298          */
2299         CDEBUG(D_HA, "Next recovery transno: "LPU64
2300                ", current: "LPU64", replaying\n",
2301                obd->obd_next_recovery_transno, transno);
2302         spin_lock(&obd->obd_recovery_task_lock);
2303         if (transno < obd->obd_next_recovery_transno) {
2304                 /* Processing the queue right now, don't re-add. */
2305                 LASSERT(list_empty(&req->rq_list));
2306                 spin_unlock(&obd->obd_recovery_task_lock);
2307                 RETURN(1);
2308         }
2309         spin_unlock(&obd->obd_recovery_task_lock);
2310
2311         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2312                 RETURN(0);
2313
2314         target_request_copy_get(req);
2315         if (!req->rq_export->exp_in_recovery) {
2316                 target_request_copy_put(req);
2317                 RETURN(-ENOTCONN);
2318         }
2319         LASSERT(req->rq_export->exp_req_replay_needed);
2320
2321         if (target_exp_enqueue_req_replay(req)) {
2322                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2323                 target_request_copy_put(req);
2324                 RETURN(0);
2325         }
2326
2327         /* XXX O(n^2) */
2328         spin_lock(&obd->obd_recovery_task_lock);
2329         LASSERT(obd->obd_recovering);
2330         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2331                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2332                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2333                         inserted = 1;
2334                         goto added;
2335                 }
2336
2337                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2338                              transno)) {
2339                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2340                                   "has been claimed by another client");
2341                         spin_unlock(&obd->obd_recovery_task_lock);
2342                         target_exp_dequeue_req_replay(req);
2343                         target_request_copy_put(req);
2344                         RETURN(0);
2345                 }
2346         }
2347 added:
2348         if (!inserted)
2349                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2350
2351         obd->obd_requests_queued_for_recovery++;
2352         spin_unlock(&obd->obd_recovery_task_lock);
2353         wake_up(&obd->obd_next_transno_waitq);
2354         RETURN(0);
2355 }
2356 EXPORT_SYMBOL(target_queue_recovery_request);
2357
2358 int target_handle_ping(struct ptlrpc_request *req)
2359 {
2360         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2361         return req_capsule_server_pack(&req->rq_pill);
2362 }
2363 EXPORT_SYMBOL(target_handle_ping);
2364
2365 void target_committed_to_req(struct ptlrpc_request *req)
2366 {
2367         struct obd_export *exp = req->rq_export;
2368
2369         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2370                 lustre_msg_set_last_committed(req->rq_repmsg,
2371                                               exp->exp_last_committed);
2372         else
2373                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2374                           "%d)", exp->exp_obd->obd_no_transno,
2375                           req->rq_repmsg == NULL);
2376
2377         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2378                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2379 }
2380 EXPORT_SYMBOL(target_committed_to_req);
2381
2382 #endif /* HAVE_SERVER_SUPPORT */
2383
2384 /**
2385  * Packs current SLV and Limit into \a req.
2386  */
2387 int target_pack_pool_reply(struct ptlrpc_request *req)
2388 {
2389         struct obd_device *obd;
2390         ENTRY;
2391
2392         /* Check that we still have all structures alive as this may
2393          * be some late RPC at shutdown time. */
2394         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2395                      !exp_connect_lru_resize(req->rq_export))) {
2396                 lustre_msg_set_slv(req->rq_repmsg, 0);
2397                 lustre_msg_set_limit(req->rq_repmsg, 0);
2398                 RETURN(0);
2399         }
2400
2401         /* OBD is alive here as export is alive, which we checked above. */
2402         obd = req->rq_export->exp_obd;
2403
2404         read_lock(&obd->obd_pool_lock);
2405         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2406         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2407         read_unlock(&obd->obd_pool_lock);
2408
2409         RETURN(0);
2410 }
2411 EXPORT_SYMBOL(target_pack_pool_reply);
2412
2413 static int target_send_reply_msg(struct ptlrpc_request *req,
2414                                  int rc, int fail_id)
2415 {
2416         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2417                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2418                 return (-ECOMM);
2419         }
2420
2421         if (unlikely(rc)) {
2422                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2423                 req->rq_status = rc;
2424                 return (ptlrpc_send_error(req, 1));
2425         } else {
2426                 DEBUG_REQ(D_NET, req, "sending reply");
2427         }
2428
2429         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2430 }
2431
2432 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2433 {
2434         struct ptlrpc_service_part *svcpt;
2435         int                        netrc;
2436         struct ptlrpc_reply_state *rs;
2437         struct obd_export         *exp;
2438         ENTRY;
2439
2440         if (req->rq_no_reply) {
2441                 EXIT;
2442                 return;
2443         }
2444
2445         svcpt = req->rq_rqbd->rqbd_svcpt;
2446         rs = req->rq_reply_state;
2447         if (rs == NULL || !rs->rs_difficult) {
2448                 /* no notifiers */
2449                 target_send_reply_msg (req, rc, fail_id);
2450                 EXIT;
2451                 return;
2452         }
2453
2454         /* must be an export if locks saved */
2455         LASSERT(req->rq_export != NULL);
2456         /* req/reply consistent */
2457         LASSERT(rs->rs_svcpt == svcpt);
2458
2459         /* "fresh" reply */
2460         LASSERT(!rs->rs_scheduled);
2461         LASSERT(!rs->rs_scheduled_ever);
2462         LASSERT(!rs->rs_handled);
2463         LASSERT(!rs->rs_on_net);
2464         LASSERT(rs->rs_export == NULL);
2465         LASSERT(list_empty(&rs->rs_obd_list));
2466         LASSERT(list_empty(&rs->rs_exp_list));
2467
2468         exp = class_export_get(req->rq_export);
2469
2470         /* disable reply scheduling while I'm setting up */
2471         rs->rs_scheduled = 1;
2472         rs->rs_on_net    = 1;
2473         rs->rs_xid       = req->rq_xid;
2474         rs->rs_transno   = req->rq_transno;
2475         rs->rs_export    = exp;
2476         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2477
2478         spin_lock(&exp->exp_uncommitted_replies_lock);
2479         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2480                rs->rs_transno, exp->exp_last_committed);
2481         if (rs->rs_transno > exp->exp_last_committed) {
2482                 /* not committed already */
2483                 list_add_tail(&rs->rs_obd_list,
2484                                   &exp->exp_uncommitted_replies);
2485         }
2486         spin_unlock(&exp->exp_uncommitted_replies_lock);
2487
2488         spin_lock(&exp->exp_lock);
2489         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2490         spin_unlock(&exp->exp_lock);
2491
2492         netrc = target_send_reply_msg(req, rc, fail_id);
2493
2494         spin_lock(&svcpt->scp_rep_lock);
2495
2496         atomic_inc(&svcpt->scp_nreps_difficult);
2497
2498         if (netrc != 0) {
2499                 /* error sending: reply is off the net.  Also we need +1
2500                  * reply ref until ptlrpc_handle_rs() is done
2501                  * with the reply state (if the send was successful, there
2502                  * would have been +1 ref for the net, which
2503                  * reply_out_callback leaves alone) */
2504                 rs->rs_on_net = 0;
2505                 ptlrpc_rs_addref(rs);
2506         }
2507
2508         spin_lock(&rs->rs_lock);
2509         if (rs->rs_transno <= exp->exp_last_committed ||
2510             (!rs->rs_on_net && !rs->rs_no_ack) ||
2511             list_empty(&rs->rs_exp_list) ||     /* completed already */
2512             list_empty(&rs->rs_obd_list)) {
2513                 CDEBUG(D_HA, "Schedule reply immediately\n");
2514                 ptlrpc_dispatch_difficult_reply(rs);
2515         } else {
2516                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2517                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2518         }
2519         spin_unlock(&rs->rs_lock);
2520         spin_unlock(&svcpt->scp_rep_lock);
2521         EXIT;
2522 }
2523 EXPORT_SYMBOL(target_send_reply);
2524
2525 ldlm_mode_t lck_compat_array[] = {
2526         [LCK_EX]    = LCK_COMPAT_EX,
2527         [LCK_PW]    = LCK_COMPAT_PW,
2528         [LCK_PR]    = LCK_COMPAT_PR,
2529         [LCK_CW]    = LCK_COMPAT_CW,
2530         [LCK_CR]    = LCK_COMPAT_CR,
2531         [LCK_NL]    = LCK_COMPAT_NL,
2532         [LCK_GROUP] = LCK_COMPAT_GROUP,
2533         [LCK_COS]   = LCK_COMPAT_COS,
2534 };
2535
2536 /**
2537  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2538  * not escape to the user level.
2539  */
2540 int ldlm_error2errno(ldlm_error_t error)
2541 {
2542         int result;
2543
2544         switch (error) {
2545         case ELDLM_OK:
2546         case ELDLM_LOCK_MATCHED:
2547                 result = 0;
2548                 break;
2549         case ELDLM_LOCK_CHANGED:
2550                 result = -ESTALE;
2551                 break;
2552         case ELDLM_LOCK_ABORTED:
2553                 result = -ENAVAIL;
2554                 break;
2555         case ELDLM_LOCK_REPLACED:
2556                 result = -ESRCH;
2557                 break;
2558         case ELDLM_NO_LOCK_DATA:
2559                 result = -ENOENT;
2560                 break;
2561         case ELDLM_NAMESPACE_EXISTS:
2562                 result = -EEXIST;
2563                 break;
2564         case ELDLM_BAD_NAMESPACE:
2565                 result = -EBADF;
2566                 break;
2567         default:
2568                 if (((int)error) < 0)  /* cast to signed type */
2569                         result = error; /* as ldlm_error_t can be unsigned */
2570                 else {
2571                         CERROR("Invalid DLM result code: %d\n", error);
2572                         result = -EPROTO;
2573                 }
2574         }
2575         return result;
2576 }
2577 EXPORT_SYMBOL(ldlm_error2errno);
2578
2579 /**
2580  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2581  */
2582 ldlm_error_t ldlm_errno2error(int err_no)
2583 {
2584         int error;
2585
2586         switch (err_no) {
2587         case 0:
2588                 error = ELDLM_OK;
2589                 break;
2590         case -ESTALE:
2591                 error = ELDLM_LOCK_CHANGED;
2592                 break;
2593         case -ENAVAIL:
2594                 error = ELDLM_LOCK_ABORTED;
2595                 break;
2596         case -ESRCH:
2597                 error = ELDLM_LOCK_REPLACED;
2598                 break;
2599         case -ENOENT:
2600                 error = ELDLM_NO_LOCK_DATA;
2601                 break;
2602         case -EEXIST:
2603                 error = ELDLM_NAMESPACE_EXISTS;
2604                 break;
2605         case -EBADF:
2606                 error = ELDLM_BAD_NAMESPACE;
2607                 break;
2608         default:
2609                 error = err_no;
2610         }
2611         return error;
2612 }
2613 EXPORT_SYMBOL(ldlm_errno2error);
2614
2615 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2616 void ldlm_dump_export_locks(struct obd_export *exp)
2617 {
2618         spin_lock(&exp->exp_locks_list_guard);
2619         if (!list_empty(&exp->exp_locks_list)) {
2620                 struct ldlm_lock *lock;
2621
2622                 CERROR("dumping locks for export %p,"
2623                        "ignore if the unmount doesn't hang\n", exp);
2624                 list_for_each_entry(lock, &exp->exp_locks_list,
2625                                         l_exp_refs_link)
2626                         LDLM_ERROR(lock, "lock:");
2627         }
2628         spin_unlock(&exp->exp_locks_list_guard);
2629 }
2630 #endif
2631
2632 #ifdef HAVE_SERVER_SUPPORT
2633 static int target_bulk_timeout(void *data)
2634 {
2635         ENTRY;
2636         /* We don't fail the connection here, because having the export
2637          * killed makes the (vital) call to commitrw very sad.
2638          */
2639         RETURN(1);
2640 }
2641
2642 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2643 {
2644         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2645 }
2646
2647 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2648                    struct l_wait_info *lwi)
2649 {
2650         struct ptlrpc_request   *req = desc->bd_req;
2651         time_t                   start = cfs_time_current_sec();
2652         time_t                   deadline;
2653         int                      rc = 0;
2654
2655         ENTRY;
2656
2657         /* If there is eviction in progress, wait for it to finish. */
2658         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2659                 *lwi = LWI_INTR(NULL, NULL);
2660                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2661                                   !atomic_read(&exp->exp_obd->
2662                                                    obd_evict_inprogress),
2663                                   lwi);
2664         }
2665
2666         /* Check if client was evicted or reconnected already. */
2667         if (exp->exp_failed ||
2668             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2669                 rc = -ENOTCONN;
2670         } else {
2671                 if (desc->bd_type == BULK_PUT_SINK)
2672                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2673                 if (rc == 0)
2674                         rc = ptlrpc_start_bulk_transfer(desc);
2675         }
2676
2677         if (rc < 0) {
2678                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2679                           bulk2type(desc), rc);
2680                 RETURN(rc);
2681         }
2682
2683         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2684                 ptlrpc_abort_bulk(desc);
2685                 RETURN(0);
2686         }
2687
2688         /* limit actual bulk transfer to bulk_timeout seconds */
2689         deadline = start + bulk_timeout;
2690         if (deadline > req->rq_deadline)
2691                 deadline = req->rq_deadline;
2692
2693         do {
2694                 long timeoutl = deadline - cfs_time_current_sec();
2695                 cfs_duration_t timeout = timeoutl <= 0 ?
2696                                          CFS_TICK : cfs_time_seconds(timeoutl);
2697                 time_t  rq_deadline;
2698
2699                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
2700                                             target_bulk_timeout, desc);
2701                 rc = l_wait_event(desc->bd_waitq,
2702                                   !ptlrpc_server_bulk_active(desc) ||
2703                                   exp->exp_failed ||
2704                                   exp->exp_conn_cnt >
2705                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
2706                                   lwi);
2707                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2708                 /* Wait again if we changed rq_deadline. */
2709                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
2710                 deadline = start + bulk_timeout;
2711                 if (deadline > rq_deadline)
2712                         deadline = rq_deadline;
2713         } while ((rc == -ETIMEDOUT) &&
2714                  (deadline > cfs_time_current_sec()));
2715
2716         if (rc == -ETIMEDOUT) {
2717                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
2718                           bulk2type(desc), deadline - start,
2719                           cfs_time_current_sec() - deadline);
2720                 ptlrpc_abort_bulk(desc);
2721         } else if (exp->exp_failed) {
2722                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2723                           bulk2type(desc));
2724                 rc = -ENOTCONN;
2725                 ptlrpc_abort_bulk(desc);
2726         } else if (exp->exp_conn_cnt >
2727                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2728                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2729                           bulk2type(desc));
2730                 /* We don't reply anyway. */
2731                 rc = -ETIMEDOUT;
2732                 ptlrpc_abort_bulk(desc);
2733         } else if (desc->bd_failure ||
2734                    desc->bd_nob_transferred != desc->bd_nob) {
2735                 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2736                           desc->bd_failure ? "network error on" : "truncated",
2737                           bulk2type(desc), desc->bd_nob_transferred,
2738                           desc->bd_nob);
2739                 /* XXX Should this be a different errno? */
2740                 rc = -ETIMEDOUT;
2741         } else if (desc->bd_type == BULK_GET_SINK) {
2742                 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2743         }
2744
2745         RETURN(rc);
2746 }
2747 EXPORT_SYMBOL(target_bulk_io);
2748
2749 #endif /* HAVE_SERVER_SUPPORT */