Whamcloud - gitweb
b49c28a3f14b008144028ad64d2f958c85b3fbb1
[fs/lustre-release.git] / lustre / ldlm / ldlm_lib.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
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 #ifndef EXPORT_SYMTAB
38 # define EXPORT_SYMTAB
39 #endif
40 #define DEBUG_SUBSYSTEM S_LDLM
41
42 #ifdef __KERNEL__
43 # include <libcfs/libcfs.h>
44 #else
45 # include <liblustre.h>
46 #endif
47 #include <obd.h>
48 #include <lustre_mds.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 to the list head
55  * @create: if zero, only search in existed 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         cfs_spin_lock(&imp->imp_lock);
84         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
86                         if (priority) {
87                                 cfs_list_del(&item->oic_item);
88                                 cfs_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                         cfs_spin_unlock(&imp->imp_lock);
96                         GOTO(out_free, rc = 0);
97                 }
98         }
99         /* not found */
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                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
106                 else
107                         cfs_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                 cfs_spin_unlock(&imp->imp_lock);
114                 GOTO(out_free, rc = -ENOENT);
115         }
116
117         cfs_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
138 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
139 {
140         struct obd_import_conn *imp_conn;
141         struct obd_export *dlmexp;
142         int rc = -ENOENT;
143         ENTRY;
144
145         cfs_spin_lock(&imp->imp_lock);
146         if (cfs_list_empty(&imp->imp_conn_list)) {
147                 LASSERT(!imp->imp_connection);
148                 GOTO(out, rc);
149         }
150
151         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
152                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
153                         continue;
154                 LASSERT(imp_conn->oic_conn);
155
156                 /* is current conn? */
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                 cfs_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         cfs_spin_unlock(&imp->imp_lock);
188         if (rc == -ENOENT)
189                 CERROR("connection %s not found\n", uuid->uuid);
190         RETURN(rc);
191 }
192
193 void client_destroy_import(struct obd_import *imp)
194 {
195         /* drop security policy instance after all rpc finished/aborted
196          * to let all busy contexts be released. */
197         class_import_get(imp);
198         class_destroy_import(imp);
199         sptlrpc_import_sec_put(imp);
200         class_import_put(imp);
201 }
202
203 /* configure an RPC client OBD device
204  *
205  * lcfg parameters:
206  * 1 - client UUID
207  * 2 - server UUID
208  * 3 - inactive-on-startup
209  */
210 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
211 {
212         struct client_obd *cli = &obddev->u.cli;
213         struct obd_import *imp;
214         struct obd_uuid server_uuid;
215         int rq_portal, rp_portal, connect_op;
216         char *name = obddev->obd_type->typ_name;
217         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
218         int rc;
219         ENTRY;
220
221         /* In a more perfect world, we would hang a ptlrpc_client off of
222          * obd_type and just use the values from there. */
223         if (!strcmp(name, LUSTRE_OSC_NAME)) {
224                 rq_portal = OST_REQUEST_PORTAL;
225                 rp_portal = OSC_REPLY_PORTAL;
226                 connect_op = OST_CONNECT;
227                 cli->cl_sp_me = LUSTRE_SP_CLI;
228                 cli->cl_sp_to = LUSTRE_SP_OST;
229                 ns_type = LDLM_NS_TYPE_OSC;
230
231         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
232                 rq_portal = MDS_REQUEST_PORTAL;
233                 rp_portal = MDC_REPLY_PORTAL;
234                 connect_op = MDS_CONNECT;
235                 cli->cl_sp_me = LUSTRE_SP_CLI;
236                 cli->cl_sp_to = LUSTRE_SP_MDT;
237                 ns_type = LDLM_NS_TYPE_MDC;
238
239         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
240                 rq_portal = MGS_REQUEST_PORTAL;
241                 rp_portal = MGC_REPLY_PORTAL;
242                 connect_op = MGS_CONNECT;
243                 cli->cl_sp_me = LUSTRE_SP_MGC;
244                 cli->cl_sp_to = LUSTRE_SP_MGS;
245                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
246                 ns_type = LDLM_NS_TYPE_MGC;
247
248         } else {
249                 CERROR("unknown client OBD type \"%s\", can't setup\n",
250                        name);
251                 RETURN(-EINVAL);
252         }
253
254         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
255                 CERROR("requires a TARGET UUID\n");
256                 RETURN(-EINVAL);
257         }
258
259         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
260                 CERROR("client UUID must be less than 38 characters\n");
261                 RETURN(-EINVAL);
262         }
263
264         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
265                 CERROR("setup requires a SERVER UUID\n");
266                 RETURN(-EINVAL);
267         }
268
269         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
270                 CERROR("target UUID must be less than 38 characters\n");
271                 RETURN(-EINVAL);
272         }
273
274         cfs_init_rwsem(&cli->cl_sem);
275         cfs_sema_init(&cli->cl_mgc_sem, 1);
276         cli->cl_conn_count = 0;
277         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
278                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
279                      sizeof(server_uuid)));
280
281         cli->cl_dirty = 0;
282         cli->cl_avail_grant = 0;
283         /* FIXME: should limit this for the sum of all cl_dirty_max */
284         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
285         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
286                 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
287         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
288         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
289         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
290         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
291         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
292         client_obd_list_lock_init(&cli->cl_loi_list_lock);
293         cli->cl_r_in_flight = 0;
294         cli->cl_w_in_flight = 0;
295
296         cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
297         cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
298         cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
299         cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
300         cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
301         cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
302         cfs_waitq_init(&cli->cl_destroy_waitq);
303         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
304 #ifdef ENABLE_CHECKSUM
305         /* Turn on checksumming by default. */
306         cli->cl_checksum = 1;
307         /*
308          * The supported checksum types will be worked out at connect time
309          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
310          * through procfs.
311          */
312         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
313 #endif
314         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
315
316         /* This value may be changed at connect time in
317            ptlrpc_connect_interpret. */
318         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
319                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
320
321         if (!strcmp(name, LUSTRE_MDC_NAME)) {
322                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
323         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
324                 cli->cl_max_rpcs_in_flight = 2;
325         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
326                 cli->cl_max_rpcs_in_flight = 3;
327         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
328                 cli->cl_max_rpcs_in_flight = 4;
329         } else {
330                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
331         }
332
333         rc = ldlm_get_ref();
334         if (rc) {
335                 CERROR("ldlm_get_ref failed: %d\n", rc);
336                 GOTO(err, rc);
337         }
338
339         ptlrpc_init_client(rq_portal, rp_portal, name,
340                            &obddev->obd_ldlm_client);
341
342         imp = class_new_import(obddev);
343         if (imp == NULL)
344                 GOTO(err_ldlm, rc = -ENOENT);
345         imp->imp_client = &obddev->obd_ldlm_client;
346         imp->imp_connect_op = connect_op;
347         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
348         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
349                LUSTRE_CFG_BUFLEN(lcfg, 1));
350         class_import_put(imp);
351
352         rc = client_import_add_conn(imp, &server_uuid, 1);
353         if (rc) {
354                 CERROR("can't add initial connection\n");
355                 GOTO(err_import, rc);
356         }
357
358         cli->cl_import = imp;
359         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
360         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
361         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
362
363         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
364                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
365                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
366                                name, obddev->obd_name,
367                                cli->cl_target_uuid.uuid);
368                         cfs_spin_lock(&imp->imp_lock);
369                         imp->imp_deactive = 1;
370                         cfs_spin_unlock(&imp->imp_lock);
371                 }
372         }
373
374         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
375                                                    LDLM_NAMESPACE_CLIENT,
376                                                    LDLM_NAMESPACE_GREEDY,
377                                                    ns_type);
378         if (obddev->obd_namespace == NULL) {
379                 CERROR("Unable to create client namespace - %s\n",
380                        obddev->obd_name);
381                 GOTO(err_import, rc = -ENOMEM);
382         }
383
384         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
385
386         RETURN(rc);
387
388 err_import:
389         class_destroy_import(imp);
390 err_ldlm:
391         ldlm_put_ref();
392 err:
393         RETURN(rc);
394
395 }
396
397 int client_obd_cleanup(struct obd_device *obddev)
398 {
399         ENTRY;
400
401         ldlm_namespace_free_post(obddev->obd_namespace);
402         obddev->obd_namespace = NULL;
403
404         ldlm_put_ref();
405         RETURN(0);
406 }
407
408 /* ->o_connect() method for client side (OSC and MDC and MGC) */
409 int client_connect_import(const struct lu_env *env,
410                           struct obd_export **exp,
411                           struct obd_device *obd, struct obd_uuid *cluuid,
412                           struct obd_connect_data *data, void *localdata)
413 {
414         struct client_obd *cli = &obd->u.cli;
415         struct obd_import *imp = cli->cl_import;
416         struct obd_connect_data *ocd;
417         struct lustre_handle conn = { 0 };
418         int rc;
419         ENTRY;
420
421         *exp = NULL;
422         cfs_down_write(&cli->cl_sem);
423         if (cli->cl_conn_count > 0 )
424                 GOTO(out_sem, rc = -EALREADY);
425
426         rc = class_connect(&conn, obd, cluuid);
427         if (rc)
428                 GOTO(out_sem, rc);
429
430         cli->cl_conn_count++;
431         *exp = class_conn2export(&conn);
432
433         LASSERT(obd->obd_namespace);
434
435         imp->imp_dlm_handle = conn;
436         rc = ptlrpc_init_import(imp);
437         if (rc != 0)
438                 GOTO(out_ldlm, rc);
439
440         ocd = &imp->imp_connect_data;
441         if (data) {
442                 *ocd = *data;
443                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
444         }
445
446         rc = ptlrpc_connect_import(imp, NULL);
447         if (rc != 0) {
448                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
449                 GOTO(out_ldlm, rc);
450         }
451         LASSERT((*exp)->exp_connection);
452
453         if (data) {
454                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
455                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
456                          data->ocd_connect_flags, ocd->ocd_connect_flags);
457                 data->ocd_connect_flags = ocd->ocd_connect_flags;
458         }
459
460         ptlrpc_pinger_add_import(imp);
461
462         EXIT;
463
464         if (rc) {
465 out_ldlm:
466                 cli->cl_conn_count--;
467                 class_disconnect(*exp);
468                 *exp = NULL;
469         }
470 out_sem:
471         cfs_up_write(&cli->cl_sem);
472
473         return rc;
474 }
475
476 int client_disconnect_export(struct obd_export *exp)
477 {
478         struct obd_device *obd = class_exp2obd(exp);
479         struct client_obd *cli;
480         struct obd_import *imp;
481         int rc = 0, err;
482         ENTRY;
483
484         if (!obd) {
485                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
486                        exp, exp ? exp->exp_handle.h_cookie : -1);
487                 RETURN(-EINVAL);
488         }
489
490         cli = &obd->u.cli;
491         imp = cli->cl_import;
492
493         cfs_down_write(&cli->cl_sem);
494         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
495                cli->cl_conn_count);
496
497         if (!cli->cl_conn_count) {
498                 CERROR("disconnecting disconnected device (%s)\n",
499                        obd->obd_name);
500                 GOTO(out_disconnect, rc = -EINVAL);
501         }
502
503         cli->cl_conn_count--;
504         if (cli->cl_conn_count)
505                 GOTO(out_disconnect, rc = 0);
506
507         /* Mark import deactivated now, so we don't try to reconnect if any
508          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
509          * fully deactivate the import, or that would drop all requests. */
510         cfs_spin_lock(&imp->imp_lock);
511         imp->imp_deactive = 1;
512         cfs_spin_unlock(&imp->imp_lock);
513
514         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
515          * delete it regardless.  (It's safe to delete an import that was
516          * never added.) */
517         (void)ptlrpc_pinger_del_import(imp);
518
519         if (obd->obd_namespace != NULL) {
520                 /* obd_force == local only */
521                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
522                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
523                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
524         }
525
526         /*
527          * there's no need to hold sem during disconnecting an import,
528          * and actually it may cause deadlock in gss.
529          */
530         cfs_up_write(&cli->cl_sem);
531         rc = ptlrpc_disconnect_import(imp, 0);
532         cfs_down_write(&cli->cl_sem);
533
534         ptlrpc_invalidate_import(imp);
535
536         if (imp->imp_rq_pool) {
537                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
538                 imp->imp_rq_pool = NULL;
539         }
540         client_destroy_import(imp);
541         cli->cl_import = NULL;
542
543         EXIT;
544
545  out_disconnect:
546         /* use server style - class_disconnect should be always called for
547          * o_disconnect */
548         err = class_disconnect(exp);
549         if (!rc && err)
550                 rc = err;
551
552         cfs_up_write(&cli->cl_sem);
553
554         RETURN(rc);
555 }
556
557 int server_disconnect_export(struct obd_export *exp)
558 {
559         int rc;
560         ENTRY;
561
562         /* Disconnect early so that clients can't keep using export */
563         rc = class_disconnect(exp);
564         /* close import for avoid sending any requests */
565         if (exp->exp_imp_reverse)
566                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
567
568         if (exp->exp_obd->obd_namespace != NULL)
569                 ldlm_cancel_locks_for_export(exp);
570
571         /* complete all outstanding replies */
572         cfs_spin_lock(&exp->exp_lock);
573         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
574                 struct ptlrpc_reply_state *rs =
575                         cfs_list_entry(exp->exp_outstanding_replies.next,
576                                        struct ptlrpc_reply_state, rs_exp_list);
577                 struct ptlrpc_service *svc = rs->rs_service;
578
579                 cfs_spin_lock(&svc->srv_rs_lock);
580                 cfs_list_del_init(&rs->rs_exp_list);
581                 cfs_spin_lock(&rs->rs_lock);
582                 ptlrpc_schedule_difficult_reply(rs);
583                 cfs_spin_unlock(&rs->rs_lock);
584                 cfs_spin_unlock(&svc->srv_rs_lock);
585         }
586         cfs_spin_unlock(&exp->exp_lock);
587
588         RETURN(rc);
589 }
590
591 /* --------------------------------------------------------------------------
592  * from old lib/target.c
593  * -------------------------------------------------------------------------- */
594
595 static int target_handle_reconnect(struct lustre_handle *conn,
596                                    struct obd_export *exp,
597                                    struct obd_uuid *cluuid)
598 {
599         ENTRY;
600         if (exp->exp_connection && exp->exp_imp_reverse) {
601                 struct lustre_handle *hdl;
602                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
603                 /* Might be a re-connect after a partition. */
604                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
605                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
606                               cluuid->uuid);
607                         conn->cookie = exp->exp_handle.h_cookie;
608                         /* target_handle_connect() treats EALREADY and
609                          * -EALREADY differently.  EALREADY means we are
610                          * doing a valid reconnect from the same client. */
611                         RETURN(EALREADY);
612                 } else {
613                         CERROR("%s reconnecting from %s, "
614                                "handle mismatch (ours "LPX64", theirs "
615                                LPX64")\n", cluuid->uuid,
616                                exp->exp_connection->c_remote_uuid.uuid,
617                                hdl->cookie, conn->cookie);
618                         memset(conn, 0, sizeof *conn);
619                         /* target_handle_connect() treats EALREADY and
620                          * -EALREADY differently.  -EALREADY is an error
621                          * (same UUID, different handle). */
622                         RETURN(-EALREADY);
623                 }
624         }
625
626         conn->cookie = exp->exp_handle.h_cookie;
627         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
628                cluuid->uuid, exp, conn->cookie);
629         RETURN(0);
630 }
631
632 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
633                           int error)
634 {
635         struct obd_export *exp = cb_data;
636
637         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
638                obd->obd_name, exp->exp_client_uuid.uuid);
639
640         cfs_spin_lock(&exp->exp_lock);
641         exp->exp_need_sync = 0;
642         cfs_spin_unlock(&exp->exp_lock);
643         class_export_cb_put(exp);
644 }
645 EXPORT_SYMBOL(target_client_add_cb);
646
647 static void
648 target_start_and_reset_recovery_timer(struct obd_device *obd,
649                                       struct ptlrpc_request *req,
650                                       int new_client);
651
652 int target_handle_connect(struct ptlrpc_request *req)
653 {
654         struct obd_device *target, *targref = NULL;
655         struct obd_export *export = NULL;
656         struct obd_import *revimp;
657         struct lustre_handle conn;
658         struct lustre_handle *tmp;
659         struct obd_uuid tgtuuid;
660         struct obd_uuid cluuid;
661         struct obd_uuid remote_uuid;
662         char *str;
663         int rc = 0;
664         int mds_conn = 0;
665         struct obd_connect_data *data, *tmpdata;
666         lnet_nid_t *client_nid = NULL;
667         ENTRY;
668
669         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
670
671         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
672         if (str == NULL) {
673                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
674                 GOTO(out, rc = -EINVAL);
675         }
676
677         obd_str2uuid(&tgtuuid, str);
678         target = class_uuid2obd(&tgtuuid);
679         if (!target)
680                 target = class_name2obd(str);
681
682         if (!target || target->obd_stopping || !target->obd_set_up) {
683                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
684                                    " for connect (%s)\n", str,
685                                    !target ? "no target" :
686                                    (target->obd_stopping ? "stopping" :
687                                    "not set up"));
688                 GOTO(out, rc = -ENODEV);
689         }
690
691         if (target->obd_no_conn) {
692                 LCONSOLE_WARN("%s: temporarily refusing client connection "
693                               "from %s\n", target->obd_name,
694                               libcfs_nid2str(req->rq_peer.nid));
695                 GOTO(out, rc = -EAGAIN);
696         }
697
698         /* Make sure the target isn't cleaned up while we're here. Yes,
699            there's still a race between the above check and our incref here.
700            Really, class_uuid2obd should take the ref. */
701         targref = class_incref(target, __FUNCTION__, cfs_current());
702
703
704         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
705         if (str == NULL) {
706                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
707                 GOTO(out, rc = -EINVAL);
708         }
709
710         obd_str2uuid(&cluuid, str);
711
712         /* XXX extract a nettype and format accordingly */
713         switch (sizeof(lnet_nid_t)) {
714                 /* NB the casts only avoid compiler warnings */
715         case 8:
716                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
717                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
718                 break;
719         case 4:
720                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
721                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
722                 break;
723         default:
724                 LBUG();
725         }
726
727         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
728         if (tmp == NULL)
729                 GOTO(out, rc = -EPROTO);
730
731         conn = *tmp;
732
733         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
734         if (!data)
735                 GOTO(out, rc = -EPROTO);
736
737         rc = req_capsule_server_pack(&req->rq_pill);
738         if (rc)
739                 GOTO(out, rc);
740
741         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
742                 if (!data) {
743                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
744                                   "libclient connection attempt");
745                         GOTO(out, rc = -EPROTO);
746                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
747                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
748                            data->ocd_version > LUSTRE_VERSION_CODE +
749                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
750                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
751                                   "libclient connection attempt",
752                                   data->ocd_version < LUSTRE_VERSION_CODE ?
753                                   "old" : "new",
754                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
755                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
756                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
757                                   OBD_OCD_VERSION_FIX(data->ocd_version));
758                         data = req_capsule_server_sized_get(&req->rq_pill,
759                                                         &RMF_CONNECT_DATA,
760                                     offsetof(typeof(*data), ocd_version) +
761                                              sizeof(data->ocd_version));
762                         if (data) {
763                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
764                                 data->ocd_version = LUSTRE_VERSION_CODE;
765                         }
766                         GOTO(out, rc = -EPROTO);
767                 }
768         }
769
770         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
771             (data->ocd_connect_flags & OBD_CONNECT_MDS))
772                 mds_conn = 1;
773
774         /* lctl gets a backstage, all-access pass. */
775         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
776                 goto dont_check_exports;
777
778         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
779         if (!export)
780                 goto no_export;
781
782         /* we've found an export in the hash */
783         if (export->exp_connecting) { /* bug 9635, et. al. */
784                 CWARN("%s: exp %p already connecting\n",
785                       export->exp_obd->obd_name, export);
786                 class_export_put(export);
787                 export = NULL;
788                 rc = -EALREADY;
789         } else if (mds_conn && export->exp_connection) {
790                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
791                         /* mds reconnected after failover */
792                         CWARN("%s: received MDS connection from NID %s,"
793                               " removing former export from NID %s\n",
794                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
795                             libcfs_nid2str(export->exp_connection->c_peer.nid));
796                 else
797                         /* new mds connection from the same nid */
798                         CWARN("%s: received new MDS connection from NID %s,"
799                               " removing former export from same NID\n",
800                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
801                 class_fail_export(export);
802                 class_export_put(export);
803                 export = NULL;
804                 rc = 0;
805         } else if (export->exp_connection != NULL &&
806                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
807                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
808                     MSG_CONNECT_INITIAL)) {
809                 /* in mds failover we have static uuid but nid can be
810                  * changed*/
811                 CWARN("%s: cookie %s seen on new NID %s when "
812                       "existing NID %s is already connected\n",
813                       target->obd_name, cluuid.uuid,
814                       libcfs_nid2str(req->rq_peer.nid),
815                       libcfs_nid2str(export->exp_connection->c_peer.nid));
816                 rc = -EALREADY;
817                 class_export_put(export);
818                 export = NULL;
819         } else {
820                 cfs_spin_lock(&export->exp_lock);
821                 export->exp_connecting = 1;
822                 cfs_spin_unlock(&export->exp_lock);
823                 class_export_put(export);
824                 LASSERT(export->exp_obd == target);
825
826                 rc = target_handle_reconnect(&conn, export, &cluuid);
827         }
828
829         /* If we found an export, we already unlocked. */
830         if (!export) {
831 no_export:
832                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
833         } else if (req->rq_export == NULL &&
834                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
835                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
836                       target->obd_name, cluuid.uuid,
837                       libcfs_nid2str(req->rq_peer.nid),
838                       export, cfs_atomic_read(&export->exp_refcount));
839                 GOTO(out, rc = -EBUSY);
840         } else if (req->rq_export != NULL &&
841                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
842                 /* the current connect rpc has increased exp_rpc_count */
843                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
844                       target->obd_name, cluuid.uuid,
845                       libcfs_nid2str(req->rq_peer.nid),
846                       export, cfs_atomic_read(&export->exp_rpc_count) - 1);
847                 cfs_spin_lock(&export->exp_lock);
848                 if (req->rq_export->exp_conn_cnt <
849                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
850                         /* try to abort active requests */
851                         req->rq_export->exp_abort_active_req = 1;
852                 cfs_spin_unlock(&export->exp_lock);
853                 GOTO(out, rc = -EBUSY);
854         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
855                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
856                        "cookies not random?\n", target->obd_name,
857                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
858                 GOTO(out, rc = -EALREADY);
859         } else {
860                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
861         }
862
863         if (rc < 0) {
864                 GOTO(out, rc);
865         }
866
867         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
868                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
869               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
870               export, (long)cfs_time_current_sec(),
871               export ? (long)export->exp_last_request_time : 0);
872
873         /* Tell the client if we're in recovery. */
874         if (target->obd_recovering) {
875                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
876                 /* If this is the first time a client connects,
877                    reset the recovery timer */
878                 if (rc == 0)
879                         target_start_and_reset_recovery_timer(target, req,
880                                                               !export);
881         }
882
883         /* We want to handle EALREADY but *not* -EALREADY from
884          * target_handle_reconnect(), return reconnection state in a flag */
885         if (rc == EALREADY) {
886                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
887                 rc = 0;
888         } else {
889                 LASSERT(rc == 0);
890         }
891
892         /* Tell the client if we support replayable requests */
893         if (target->obd_replayable)
894                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
895         client_nid = &req->rq_peer.nid;
896
897         if (export == NULL) {
898                 if (target->obd_recovering) {
899                         cfs_time_t t;
900
901                         t = cfs_timer_deadline(&target->obd_recovery_timer);
902                         t = cfs_time_sub(t, cfs_time_current());
903                         CERROR("%s: denying connection for new client %s (%s): "
904                                "%d clients in recovery for "CFS_TIME_T"s\n",
905                                target->obd_name,
906                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
907                                cfs_atomic_read(&target-> \
908                                                obd_lock_replay_clients),
909                                cfs_duration_sec(t));
910                         rc = -EBUSY;
911                 } else {
912 dont_check_exports:
913                         rc = obd_connect(req->rq_svc_thread->t_env,
914                                          &export, target, &cluuid, data,
915                                          client_nid);
916                         if (rc == 0)
917                                 conn.cookie = export->exp_handle.h_cookie;
918                 }
919         } else {
920                 rc = obd_reconnect(req->rq_svc_thread->t_env,
921                                    export, target, &cluuid, data, client_nid);
922                 if (rc == 0)
923                         /* prevous done via class_conn2export */
924                         class_export_get(export);
925         }
926         if (rc)
927                 GOTO(out, rc);
928         /* Return only the parts of obd_connect_data that we understand, so the
929          * client knows that we don't understand the rest. */
930         if (data) {
931                  tmpdata = req_capsule_server_get(&req->rq_pill,
932                                                   &RMF_CONNECT_DATA);
933                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
934                  *tmpdata = *data;
935         }
936
937         /* If all else goes well, this is our RPC return code. */
938         req->rq_status = 0;
939
940         lustre_msg_set_handle(req->rq_repmsg, &conn);
941
942         /* If the client and the server are the same node, we will already
943          * have an export that really points to the client's DLM export,
944          * because we have a shared handles table.
945          *
946          * XXX this will go away when shaver stops sending the "connect" handle
947          * in the real "remote handle" field of the request --phik 24 Apr 2003
948          */
949         if (req->rq_export != NULL)
950                 class_export_put(req->rq_export);
951
952         req->rq_export = export;
953
954         cfs_spin_lock(&export->exp_lock);
955         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
956                 cfs_spin_unlock(&export->exp_lock);
957                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
958                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
959                        export->exp_conn_cnt,
960                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
961
962                 GOTO(out, rc = -EALREADY);
963         }
964         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
965         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
966         export->exp_abort_active_req = 0;
967
968         /* request from liblustre?  Don't evict it for not pinging. */
969         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
970                 export->exp_libclient = 1;
971                 cfs_spin_unlock(&export->exp_lock);
972
973                 cfs_spin_lock(&target->obd_dev_lock);
974                 cfs_list_del_init(&export->exp_obd_chain_timed);
975                 cfs_spin_unlock(&target->obd_dev_lock);
976         } else {
977                 cfs_spin_unlock(&export->exp_lock);
978         }
979
980         if (export->exp_connection != NULL) {
981                 /* Check to see if connection came from another NID */
982                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
983                     !cfs_hlist_unhashed(&export->exp_nid_hash))
984                         cfs_hash_del(export->exp_obd->obd_nid_hash,
985                                      &export->exp_connection->c_peer.nid,
986                                      &export->exp_nid_hash);
987
988                 ptlrpc_connection_put(export->exp_connection);
989         }
990
991         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
992                                                        req->rq_self,
993                                                        &remote_uuid);
994         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
995                 cfs_hash_add(export->exp_obd->obd_nid_hash,
996                              &export->exp_connection->c_peer.nid,
997                              &export->exp_nid_hash);
998         }
999
1000         cfs_spin_lock(&target->obd_recovery_task_lock);
1001         if (target->obd_recovering && !export->exp_in_recovery) {
1002                 cfs_spin_lock(&export->exp_lock);
1003                 export->exp_in_recovery = 1;
1004                 export->exp_req_replay_needed = 1;
1005                 export->exp_lock_replay_needed = 1;
1006                 cfs_spin_unlock(&export->exp_lock);
1007                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1008                      && (data->ocd_transno == 0))
1009                         CWARN("Connect with zero transno!\n");
1010
1011                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1012                      && data->ocd_transno < target->obd_next_recovery_transno)
1013                         target->obd_next_recovery_transno = data->ocd_transno;
1014                 target->obd_connected_clients++;
1015                 cfs_atomic_inc(&target->obd_req_replay_clients);
1016                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1017                 if (target->obd_connected_clients ==
1018                     target->obd_max_recoverable_clients)
1019                         cfs_waitq_signal(&target->obd_next_transno_waitq);
1020         }
1021         cfs_spin_unlock(&target->obd_recovery_task_lock);
1022         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1023         conn = *tmp;
1024
1025         if (export->exp_imp_reverse != NULL) {
1026                 /* destroyed import can be still referenced in ctxt */
1027                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1028                                    KEY_REVIMP_UPD, 0, NULL, NULL);
1029
1030                 client_destroy_import(export->exp_imp_reverse);
1031         }
1032
1033         /* for the rest part, we return -ENOTCONN in case of errors
1034          * in order to let client initialize connection again.
1035          */
1036         revimp = export->exp_imp_reverse = class_new_import(target);
1037         if (!revimp) {
1038                 CERROR("fail to alloc new reverse import.\n");
1039                 GOTO(out, rc = -ENOTCONN);
1040         }
1041
1042         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1043         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1044         revimp->imp_remote_handle = conn;
1045         revimp->imp_dlm_fake = 1;
1046         revimp->imp_state = LUSTRE_IMP_FULL;
1047
1048         /* unknown versions will be caught in
1049          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1050         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1051
1052         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1053             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1054                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1055         else
1056                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1057
1058         if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1059             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1060                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1061         else
1062                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1063
1064         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1065         if (rc) {
1066                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1067                 export->exp_imp_reverse = NULL;
1068                 class_destroy_import(revimp);
1069         }
1070
1071         class_import_put(revimp);
1072 out:
1073         if (export) {
1074                 cfs_spin_lock(&export->exp_lock);
1075                 export->exp_connecting = 0;
1076                 cfs_spin_unlock(&export->exp_lock);
1077         }
1078         if (targref)
1079                 class_decref(targref, __FUNCTION__, cfs_current());
1080         if (rc)
1081                 req->rq_status = rc;
1082         RETURN(rc);
1083 }
1084
1085 int target_handle_disconnect(struct ptlrpc_request *req)
1086 {
1087         int rc;
1088         ENTRY;
1089
1090         rc = req_capsule_server_pack(&req->rq_pill);
1091         if (rc)
1092                 RETURN(rc);
1093
1094         /* keep the rq_export around so we can send the reply */
1095         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1096
1097         RETURN(0);
1098 }
1099
1100 void target_destroy_export(struct obd_export *exp)
1101 {
1102         /* exports created from last_rcvd data, and "fake"
1103            exports created by lctl don't have an import */
1104         if (exp->exp_imp_reverse != NULL)
1105                 client_destroy_import(exp->exp_imp_reverse);
1106
1107         LASSERT(cfs_atomic_read(&exp->exp_locks_count) == 0);
1108         LASSERT(cfs_atomic_read(&exp->exp_rpc_count) == 0);
1109         LASSERT(cfs_atomic_read(&exp->exp_cb_count) == 0);
1110         LASSERT(cfs_atomic_read(&exp->exp_replay_count) == 0);
1111 }
1112
1113 /*
1114  * Recovery functions
1115  */
1116 static void target_request_copy_get(struct ptlrpc_request *req)
1117 {
1118         class_export_rpc_get(req->rq_export);
1119         LASSERT(cfs_list_empty(&req->rq_list));
1120         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1121         /* increase refcount to keep request in queue */
1122         LASSERT(cfs_atomic_read(&req->rq_refcount));
1123         cfs_atomic_inc(&req->rq_refcount);
1124         /** let export know it has replays to be handled */
1125         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1126         /* release service thread while request is queued 
1127          * we are moving the request from active processing
1128          * to waiting on the replay queue */
1129         ptlrpc_server_active_request_dec(req);
1130 }
1131
1132 static void target_request_copy_put(struct ptlrpc_request *req)
1133 {
1134         LASSERT(cfs_list_empty(&req->rq_replay_list));
1135         LASSERT(cfs_atomic_read(&req->rq_export->exp_replay_count) > 0);
1136         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1137         class_export_rpc_put(req->rq_export);
1138         /* ptlrpc_server_drop_request() assumes the request is active */
1139         ptlrpc_server_active_request_inc(req);
1140         ptlrpc_server_drop_request(req);
1141 }
1142
1143 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1144 {
1145         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1146         struct obd_export     *exp = req->rq_export;
1147         struct ptlrpc_request *reqiter;
1148         int                    dup = 0;
1149
1150         LASSERT(exp);
1151
1152         cfs_spin_lock(&exp->exp_lock);
1153         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1154                                 rq_replay_list) {
1155                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1156                         dup = 1;
1157                         break;
1158                 }
1159         }
1160
1161         if (dup) {
1162                 /* we expect it with RESENT and REPLAY flags */
1163                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1164                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1165                         CERROR("invalid flags %x of resent replay\n",
1166                                lustre_msg_get_flags(req->rq_reqmsg));
1167         } else {
1168                 cfs_list_add_tail(&req->rq_replay_list,
1169                                   &exp->exp_req_replay_queue);
1170         }
1171
1172         cfs_spin_unlock(&exp->exp_lock);
1173         return dup;
1174 }
1175
1176 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1177 {
1178         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1179         LASSERT(req->rq_export);
1180
1181         cfs_spin_lock(&req->rq_export->exp_lock);
1182         cfs_list_del_init(&req->rq_replay_list);
1183         cfs_spin_unlock(&req->rq_export->exp_lock);
1184 }
1185
1186 #ifdef __KERNEL__
1187 static void target_finish_recovery(struct obd_device *obd)
1188 {
1189         ENTRY;
1190         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1191                       obd->obd_name);
1192
1193         ldlm_reprocess_all_ns(obd->obd_namespace);
1194         cfs_spin_lock(&obd->obd_recovery_task_lock);
1195         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1196             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1197             !cfs_list_empty(&obd->obd_final_req_queue)) {
1198                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1199                        obd->obd_name,
1200                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1201                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1202                                "" : "lock ",
1203                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1204                                "" : "final ");
1205                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1206                 LBUG();
1207         }
1208         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1209
1210         obd->obd_recovery_end = cfs_time_current_sec();
1211
1212         /* when recovery finished, cleanup orphans on mds and ost */
1213         if (OBT(obd) && OBP(obd, postrecov)) {
1214                 int rc = OBP(obd, postrecov)(obd);
1215                 if (rc < 0)
1216                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1217                                       obd->obd_name, rc);
1218         }
1219         EXIT;
1220 }
1221
1222 static void abort_req_replay_queue(struct obd_device *obd)
1223 {
1224         struct ptlrpc_request *req, *n;
1225         cfs_list_t abort_list;
1226
1227         CFS_INIT_LIST_HEAD(&abort_list);
1228         cfs_spin_lock(&obd->obd_recovery_task_lock);
1229         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1230         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1231         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1232                 DEBUG_REQ(D_WARNING, req, "aborted:");
1233                 req->rq_status = -ENOTCONN;
1234                 if (ptlrpc_error(req)) {
1235                         DEBUG_REQ(D_ERROR, req,
1236                                   "failed abort_req_reply; skipping");
1237                 }
1238                 target_exp_dequeue_req_replay(req);
1239                 target_request_copy_put(req);
1240         }
1241 }
1242
1243 static void abort_lock_replay_queue(struct obd_device *obd)
1244 {
1245         struct ptlrpc_request *req, *n;
1246         cfs_list_t abort_list;
1247
1248         CFS_INIT_LIST_HEAD(&abort_list);
1249         cfs_spin_lock(&obd->obd_recovery_task_lock);
1250         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1251         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1252         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1253                 DEBUG_REQ(D_ERROR, req, "aborted:");
1254                 req->rq_status = -ENOTCONN;
1255                 if (ptlrpc_error(req)) {
1256                         DEBUG_REQ(D_ERROR, req,
1257                                   "failed abort_lock_reply; skipping");
1258                 }
1259                 target_request_copy_put(req);
1260         }
1261 }
1262 #endif
1263
1264 /* Called from a cleanup function if the device is being cleaned up
1265    forcefully.  The exports should all have been disconnected already,
1266    the only thing left to do is
1267      - clear the recovery flags
1268      - cancel the timer
1269      - free queued requests and replies, but don't send replies
1270    Because the obd_stopping flag is set, no new requests should be received.
1271
1272 */
1273 void target_cleanup_recovery(struct obd_device *obd)
1274 {
1275         struct ptlrpc_request *req, *n;
1276         cfs_list_t clean_list;
1277         ENTRY;
1278
1279         CFS_INIT_LIST_HEAD(&clean_list);
1280         cfs_spin_lock(&obd->obd_dev_lock);
1281         if (!obd->obd_recovering) {
1282                 cfs_spin_unlock(&obd->obd_dev_lock);
1283                 EXIT;
1284                 return;
1285         }
1286         obd->obd_recovering = obd->obd_abort_recovery = 0;
1287         cfs_spin_unlock(&obd->obd_dev_lock);
1288
1289         cfs_spin_lock(&obd->obd_recovery_task_lock);
1290         target_cancel_recovery_timer(obd);
1291         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1292         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1293
1294         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1295                 LASSERT(req->rq_reply_state == 0);
1296                 target_exp_dequeue_req_replay(req);
1297                 target_request_copy_put(req);
1298         }
1299
1300         cfs_spin_lock(&obd->obd_recovery_task_lock);
1301         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1302         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1303         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1304
1305         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1306                 LASSERT(req->rq_reply_state == 0);
1307                 target_request_copy_put(req);
1308         }
1309
1310         EXIT;
1311 }
1312
1313 /* obd_recovery_task_lock should be held */
1314 void target_cancel_recovery_timer(struct obd_device *obd)
1315 {
1316         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1317         cfs_timer_disarm(&obd->obd_recovery_timer);
1318 }
1319
1320 /* extend = 1 means require at least "duration" seconds left in the timer,
1321    extend = 0 means set the total duration (start_recovery_timer) */
1322 static void reset_recovery_timer(struct obd_device *obd, int duration,
1323                                  int extend)
1324 {
1325         cfs_time_t now = cfs_time_current_sec();
1326         cfs_duration_t left;
1327
1328         cfs_spin_lock(&obd->obd_recovery_task_lock);
1329         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1330                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1331                 return;
1332         }
1333
1334         left = cfs_time_sub(obd->obd_recovery_end, now);
1335
1336         if (extend && (duration > left))
1337                 obd->obd_recovery_timeout += duration - left;
1338         else if (!extend && (duration > obd->obd_recovery_timeout))
1339                 /* Track the client's largest expected replay time */
1340                 obd->obd_recovery_timeout = duration;
1341
1342         /* Hard limit of obd_recovery_time_hard which should not happen */
1343         if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1344                 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1345
1346         obd->obd_recovery_end = obd->obd_recovery_start +
1347                                 obd->obd_recovery_timeout;
1348         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1349             cfs_time_before(now, obd->obd_recovery_end)) {
1350                 left = cfs_time_sub(obd->obd_recovery_end, now);
1351                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1352         }
1353         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1354         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1355                obd->obd_name, (unsigned)left);
1356 }
1357
1358 static void check_and_start_recovery_timer(struct obd_device *obd)
1359 {
1360         cfs_spin_lock(&obd->obd_recovery_task_lock);
1361         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1362                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1363                 return;
1364         }
1365         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1366         obd->obd_recovery_start = cfs_time_current_sec();
1367         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1368
1369         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1370 }
1371
1372 /* Reset the timer with each new client connection */
1373 /*
1374  * This timer is actually reconnect_timer, which is for making sure
1375  * the total recovery window is at least as big as my reconnect
1376  * attempt timing. So the initial recovery time_out will be set to
1377  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1378  * from client is bigger than this, then the recovery time_out will
1379  * be extended to make sure the client could be reconnected, in the
1380  * process, the timeout from the new client should be ignored.
1381  */
1382
1383 static void
1384 target_start_and_reset_recovery_timer(struct obd_device *obd,
1385                                       struct ptlrpc_request *req,
1386                                       int new_client)
1387 {
1388         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1389
1390         if (!new_client && service_time)
1391                 /* Teach server about old server's estimates, as first guess
1392                  * at how long new requests will take. */
1393                 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1394                             service_time);
1395
1396         check_and_start_recovery_timer(obd);
1397
1398         /* convert the service time to rpc timeout,
1399          * reuse service_time to limit stack usage */
1400         service_time = at_est2timeout(service_time);
1401
1402         /* We expect other clients to timeout within service_time, then try
1403          * to reconnect, then try the failover server.  The max delay between
1404          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1405         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1406                              INITIAL_CONNECT_TIMEOUT);
1407         if (service_time > obd->obd_recovery_timeout && !new_client)
1408                 reset_recovery_timer(obd, service_time, 0);
1409 }
1410
1411 #ifdef __KERNEL__
1412
1413 /** Health checking routines */
1414 static inline int exp_connect_healthy(struct obd_export *exp)
1415 {
1416         return (exp->exp_in_recovery);
1417 }
1418
1419 /** if export done req_replay or has replay in queue */
1420 static inline int exp_req_replay_healthy(struct obd_export *exp)
1421 {
1422         return (!exp->exp_req_replay_needed ||
1423                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1424 }
1425 /** if export done lock_replay or has replay in queue */
1426 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1427 {
1428         return (!exp->exp_lock_replay_needed ||
1429                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1430 }
1431
1432 static inline int exp_vbr_healthy(struct obd_export *exp)
1433 {
1434         return (!exp->exp_vbr_failed);
1435 }
1436
1437 static inline int exp_finished(struct obd_export *exp)
1438 {
1439         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1440 }
1441
1442 /** Checking routines for recovery */
1443 static int check_for_clients(struct obd_device *obd)
1444 {
1445         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1446                 return 1;
1447         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1448         if (obd->obd_no_conn == 0 &&
1449             obd->obd_connected_clients + obd->obd_stale_clients ==
1450             obd->obd_max_recoverable_clients)
1451                 return 1;
1452         return 0;
1453 }
1454
1455 static int check_for_next_transno(struct obd_device *obd)
1456 {
1457         struct ptlrpc_request *req = NULL;
1458         int wake_up = 0, connected, completed, queue_len;
1459         __u64 next_transno, req_transno;
1460         ENTRY;
1461
1462         cfs_spin_lock(&obd->obd_recovery_task_lock);
1463         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1464                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1465                                      struct ptlrpc_request, rq_list);
1466                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1467         } else {
1468                 req_transno = 0;
1469         }
1470
1471         connected = obd->obd_connected_clients;
1472         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1473         queue_len = obd->obd_requests_queued_for_recovery;
1474         next_transno = obd->obd_next_recovery_transno;
1475
1476         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1477                "req_transno: "LPU64", next_transno: "LPU64"\n",
1478                obd->obd_max_recoverable_clients, connected, completed,
1479                queue_len, req_transno, next_transno);
1480
1481         if (obd->obd_abort_recovery) {
1482                 CDEBUG(D_HA, "waking for aborted recovery\n");
1483                 wake_up = 1;
1484         } else if (obd->obd_recovery_expired) {
1485                 CDEBUG(D_HA, "waking for expired recovery\n");
1486                 wake_up = 1;
1487         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1488                 CDEBUG(D_HA, "waking for completed recovery\n");
1489                 wake_up = 1;
1490         } else if (req_transno == next_transno) {
1491                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1492                 wake_up = 1;
1493         } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1494                 int d_lvl = D_HA;
1495                 /** handle gaps occured due to lost reply or VBR */
1496                 LASSERTF(req_transno >= next_transno,
1497                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1498                          req_transno, next_transno);
1499                 if (req_transno > obd->obd_last_committed &&
1500                     !obd->obd_version_recov)
1501                         d_lvl = D_ERROR;
1502                 CDEBUG(d_lvl,
1503                        "%s: waking for gap in transno, VBR is %s (skip: "
1504                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1505                        ", last_committed: "LPD64")\n",
1506                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1507                        next_transno, queue_len, completed, connected,
1508                        req_transno, obd->obd_last_committed);
1509                 obd->obd_next_recovery_transno = req_transno;
1510                 wake_up = 1;
1511         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1512                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1513                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1514                 obd->obd_next_recovery_transno = req_transno;
1515                 wake_up = 1;
1516         }
1517         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1518         return wake_up;
1519 }
1520
1521 static int check_for_next_lock(struct obd_device *obd)
1522 {
1523         int wake_up = 0;
1524
1525         cfs_spin_lock(&obd->obd_recovery_task_lock);
1526         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1527                 CDEBUG(D_HA, "waking for next lock\n");
1528                 wake_up = 1;
1529         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1530                 CDEBUG(D_HA, "waking for completed lock replay\n");
1531                 wake_up = 1;
1532         } else if (obd->obd_abort_recovery) {
1533                 CDEBUG(D_HA, "waking for aborted recovery\n");
1534                 wake_up = 1;
1535         } else if (obd->obd_recovery_expired) {
1536                 CDEBUG(D_HA, "waking for expired recovery\n");
1537                 wake_up = 1;
1538         }
1539         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1540
1541         return wake_up;
1542 }
1543
1544 /**
1545  * wait for recovery events,
1546  * check its status with help of check_routine
1547  * evict dead clients via health_check
1548  */
1549 static int target_recovery_overseer(struct obd_device *obd,
1550                                     int (*check_routine)(struct obd_device *),
1551                                     int (*health_check)(struct obd_export *))
1552 {
1553 repeat:
1554         cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1555         if (obd->obd_abort_recovery) {
1556                 CWARN("recovery is aborted, evict exports in recovery\n");
1557                 /** evict exports which didn't finish recovery yet */
1558                 class_disconnect_stale_exports(obd, exp_finished);
1559                 return 1;
1560         } else if (obd->obd_recovery_expired) {
1561                 obd->obd_recovery_expired = 0;
1562                 /** If some clients died being recovered, evict them */
1563                 CDEBUG(D_WARNING,
1564                        "recovery is timed out, evict stale exports\n");
1565                 /** evict cexports with no replay in queue, they are stalled */
1566                 class_disconnect_stale_exports(obd, health_check);
1567                 /** continue with VBR */
1568                 cfs_spin_lock(&obd->obd_dev_lock);
1569                 obd->obd_version_recov = 1;
1570                 cfs_spin_unlock(&obd->obd_dev_lock);
1571                 /**
1572                  * reset timer, recovery will proceed with versions now,
1573                  * timeout is set just to handle reconnection delays
1574                  */
1575                 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1576                 /** Wait for recovery events again, after evicting bad clients */
1577                 goto repeat;
1578         }
1579         return 0;
1580 }
1581
1582 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1583 {
1584         struct ptlrpc_request *req = NULL;
1585         ENTRY;
1586
1587         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1588                obd->obd_next_recovery_transno);
1589
1590         if (target_recovery_overseer(obd, check_for_next_transno,
1591                                      exp_req_replay_healthy)) {
1592                 abort_req_replay_queue(obd);
1593                 abort_lock_replay_queue(obd);
1594         }
1595
1596         cfs_spin_lock(&obd->obd_recovery_task_lock);
1597         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1598                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1599                                      struct ptlrpc_request, rq_list);
1600                 cfs_list_del_init(&req->rq_list);
1601                 obd->obd_requests_queued_for_recovery--;
1602                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1603         } else {
1604                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1605                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1606                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1607                 /** evict exports failed VBR */
1608                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1609         }
1610         RETURN(req);
1611 }
1612
1613 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1614 {
1615         struct ptlrpc_request *req = NULL;
1616
1617         CDEBUG(D_HA, "Waiting for lock\n");
1618         if (target_recovery_overseer(obd, check_for_next_lock,
1619                                      exp_lock_replay_healthy))
1620                 abort_lock_replay_queue(obd);
1621
1622         cfs_spin_lock(&obd->obd_recovery_task_lock);
1623         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1624                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1625                                      struct ptlrpc_request, rq_list);
1626                 cfs_list_del_init(&req->rq_list);
1627                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1628         } else {
1629                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1630                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1631                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1632                 /** evict exports failed VBR */
1633                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1634         }
1635         return req;
1636 }
1637
1638 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1639 {
1640         struct ptlrpc_request *req = NULL;
1641
1642         cfs_spin_lock(&obd->obd_recovery_task_lock);
1643         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1644                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1645                                      struct ptlrpc_request, rq_list);
1646                 cfs_list_del_init(&req->rq_list);
1647                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1648                 if (req->rq_export->exp_in_recovery) {
1649                         cfs_spin_lock(&req->rq_export->exp_lock);
1650                         req->rq_export->exp_in_recovery = 0;
1651                         cfs_spin_unlock(&req->rq_export->exp_lock);
1652                 }
1653         } else {
1654                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1655         }
1656         return req;
1657 }
1658
1659 static int handle_recovery_req(struct ptlrpc_thread *thread,
1660                                struct ptlrpc_request *req,
1661                                svc_handler_t handler)
1662 {
1663         int rc;
1664         ENTRY;
1665
1666         /**
1667          * export can be evicted during recovery, no need to handle replays for
1668          * it after that, discard such request silently
1669          */
1670         if (req->rq_export->exp_disconnected)
1671                 GOTO(reqcopy_put, rc = 0);
1672
1673         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1674         if (rc) {
1675                 CERROR("Failure to initialize session: %d\n", rc);
1676                 GOTO(reqcopy_put, rc);
1677         }
1678
1679         req->rq_recov_session.lc_thread = thread;
1680         lu_context_enter(&req->rq_recov_session);
1681         req->rq_svc_thread = thread;
1682         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1683
1684         /* thread context */
1685         lu_context_enter(&thread->t_env->le_ctx);
1686         (void)handler(req);
1687         lu_context_exit(&thread->t_env->le_ctx);
1688
1689         lu_context_exit(&req->rq_recov_session);
1690         lu_context_fini(&req->rq_recov_session);
1691         /* don't reset timer for final stage */
1692         if (!exp_finished(req->rq_export)) {
1693                 /**
1694                  * Add request timeout to the recovery time so next request from
1695                  * this client may come in recovery time
1696                  */
1697                  reset_recovery_timer(class_exp2obd(req->rq_export),
1698                                       AT_OFF ? obd_timeout :
1699                                       lustre_msg_get_timeout(req->rq_reqmsg), 1);
1700         }
1701 reqcopy_put:
1702         RETURN(rc);
1703 }
1704
1705 static int target_recovery_thread(void *arg)
1706 {
1707         struct lu_target *lut = arg;
1708         struct obd_device *obd = lut->lut_obd;
1709         struct ptlrpc_request *req;
1710         struct target_recovery_data *trd = &obd->obd_recovery_data;
1711         unsigned long delta;
1712         unsigned long flags;
1713         struct lu_env env;
1714         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1715         int rc = 0;
1716         ENTRY;
1717
1718         cfs_daemonize_ctxt("tgt_recov");
1719
1720         SIGNAL_MASK_LOCK(current, flags);
1721         sigfillset(&current->blocked);
1722         RECALC_SIGPENDING;
1723         SIGNAL_MASK_UNLOCK(current, flags);
1724
1725         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1726         if (rc)
1727                 RETURN(rc);
1728
1729         thread->t_env = &env;
1730         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1731         env.le_ctx.lc_thread = thread;
1732         thread->t_data = NULL;
1733
1734         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1735                cfs_curproc_pid());
1736         trd->trd_processing_task = cfs_curproc_pid();
1737
1738         cfs_spin_lock(&obd->obd_dev_lock);
1739         obd->obd_recovering = 1;
1740         cfs_spin_unlock(&obd->obd_dev_lock);
1741         cfs_complete(&trd->trd_starting);
1742
1743         /* first of all, we have to know the first transno to replay */
1744         if (target_recovery_overseer(obd, check_for_clients,
1745                                      exp_connect_healthy)) {
1746                 abort_req_replay_queue(obd);
1747                 abort_lock_replay_queue(obd);
1748         }
1749
1750         /* next stage: replay requests */
1751         delta = jiffies;
1752         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1753                cfs_atomic_read(&obd->obd_req_replay_clients),
1754                obd->obd_next_recovery_transno);
1755         while ((req = target_next_replay_req(obd))) {
1756                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1757                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1758                           lustre_msg_get_transno(req->rq_reqmsg),
1759                           libcfs_nid2str(req->rq_peer.nid));
1760                 handle_recovery_req(thread, req,
1761                                     trd->trd_recovery_handler);
1762                 /**
1763                  * bz18031: increase next_recovery_transno before
1764                  * target_request_copy_put() will drop exp_rpc reference
1765                  */
1766                 cfs_spin_lock(&obd->obd_recovery_task_lock);
1767                 obd->obd_next_recovery_transno++;
1768                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1769                 target_exp_dequeue_req_replay(req);
1770                 target_request_copy_put(req);
1771                 obd->obd_replayed_requests++;
1772         }
1773
1774         /**
1775          * The second stage: replay locks
1776          */
1777         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1778                cfs_atomic_read(&obd->obd_lock_replay_clients));
1779         while ((req = target_next_replay_lock(obd))) {
1780                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1781                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1782                           libcfs_nid2str(req->rq_peer.nid));
1783                 handle_recovery_req(thread, req,
1784                                     trd->trd_recovery_handler);
1785                 target_request_copy_put(req);
1786                 obd->obd_replayed_locks++;
1787         }
1788
1789         /**
1790          * The third stage: reply on final pings, at this moment all clients
1791          * must have request in final queue
1792          */
1793         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1794         /** Update server last boot epoch */
1795         lut_boot_epoch_update(lut);
1796         /* We drop recoverying flag to forward all new requests
1797          * to regular mds_handle() since now */
1798         cfs_spin_lock(&obd->obd_dev_lock);
1799         obd->obd_recovering = obd->obd_abort_recovery = 0;
1800         cfs_spin_unlock(&obd->obd_dev_lock);
1801         cfs_spin_lock(&obd->obd_recovery_task_lock);
1802         target_cancel_recovery_timer(obd);
1803         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1804         while ((req = target_next_final_ping(obd))) {
1805                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1806                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1807                           libcfs_nid2str(req->rq_peer.nid));
1808                 handle_recovery_req(thread, req,
1809                                     trd->trd_recovery_handler);
1810                 target_request_copy_put(req);
1811         }
1812
1813         delta = (jiffies - delta) / CFS_HZ;
1814         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1815               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1816         if (delta > OBD_RECOVERY_TIME_SOFT) {
1817                 CWARN("too long recovery - read logs\n");
1818                 libcfs_debug_dumplog();
1819         }
1820
1821         target_finish_recovery(obd);
1822
1823         lu_context_fini(&env.le_ctx);
1824         trd->trd_processing_task = 0;
1825         cfs_complete(&trd->trd_finishing);
1826         RETURN(rc);
1827 }
1828
1829 static int target_start_recovery_thread(struct lu_target *lut,
1830                                         svc_handler_t handler)
1831 {
1832         struct obd_device *obd = lut->lut_obd;
1833         int rc = 0;
1834         struct target_recovery_data *trd = &obd->obd_recovery_data;
1835
1836         memset(trd, 0, sizeof(*trd));
1837         cfs_init_completion(&trd->trd_starting);
1838         cfs_init_completion(&trd->trd_finishing);
1839         trd->trd_recovery_handler = handler;
1840
1841         if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
1842                 cfs_wait_for_completion(&trd->trd_starting);
1843                 LASSERT(obd->obd_recovering != 0);
1844         } else
1845                 rc = -ECHILD;
1846
1847         return rc;
1848 }
1849
1850 void target_stop_recovery_thread(struct obd_device *obd)
1851 {
1852         if (obd->obd_recovery_data.trd_processing_task > 0) {
1853                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1854                 /** recovery can be done but postrecovery is not yet */
1855                 cfs_spin_lock(&obd->obd_dev_lock);
1856                 if (obd->obd_recovering) {
1857                         CERROR("%s: Aborting recovery\n", obd->obd_name);
1858                         obd->obd_abort_recovery = 1;
1859                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1860                 }
1861                 cfs_spin_unlock(&obd->obd_dev_lock);
1862                 cfs_wait_for_completion(&trd->trd_finishing);
1863         }
1864 }
1865
1866 void target_recovery_fini(struct obd_device *obd)
1867 {
1868         class_disconnect_exports(obd);
1869         target_stop_recovery_thread(obd);
1870         target_cleanup_recovery(obd);
1871 }
1872 EXPORT_SYMBOL(target_recovery_fini);
1873
1874 static void target_recovery_expired(unsigned long castmeharder)
1875 {
1876         struct obd_device *obd = (struct obd_device *)castmeharder;
1877         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1878                " after %lds (%d clients connected)\n",
1879                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1880                cfs_time_current_sec()- obd->obd_recovery_start,
1881                obd->obd_connected_clients);
1882
1883         obd->obd_recovery_expired = 1;
1884         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1885 }
1886
1887 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1888 {
1889         struct obd_device *obd = lut->lut_obd;
1890         if (obd->obd_max_recoverable_clients == 0) {
1891                 /** Update server last boot epoch */
1892                 lut_boot_epoch_update(lut);
1893                 return;
1894         }
1895
1896         CWARN("RECOVERY: service %s, %d recoverable clients, "
1897               "last_transno "LPU64"\n", obd->obd_name,
1898               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1899         LASSERT(obd->obd_stopping == 0);
1900         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1901         obd->obd_recovery_start = 0;
1902         obd->obd_recovery_end = 0;
1903
1904         /* both values can be get from mount data already */
1905         if (obd->obd_recovery_timeout == 0)
1906                 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
1907         if (obd->obd_recovery_time_hard == 0)
1908                 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
1909         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1910         target_start_recovery_thread(lut, handler);
1911 }
1912 EXPORT_SYMBOL(target_recovery_init);
1913
1914 #endif
1915
1916 static int target_process_req_flags(struct obd_device *obd,
1917                                     struct ptlrpc_request *req)
1918 {
1919         struct obd_export *exp = req->rq_export;
1920         LASSERT(exp != NULL);
1921         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1922                 /* client declares he's ready to replay locks */
1923                 cfs_spin_lock(&exp->exp_lock);
1924                 if (exp->exp_req_replay_needed) {
1925                         exp->exp_req_replay_needed = 0;
1926                         cfs_spin_unlock(&exp->exp_lock);
1927                         LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients));
1928                         cfs_atomic_dec(&obd->obd_req_replay_clients);
1929                 } else {
1930                         cfs_spin_unlock(&exp->exp_lock);
1931                 }
1932         }
1933         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1934                 /* client declares he's ready to complete recovery
1935                  * so, we put the request on th final queue */
1936                 cfs_spin_lock(&exp->exp_lock);
1937                 if (exp->exp_lock_replay_needed) {
1938                         exp->exp_lock_replay_needed = 0;
1939                         cfs_spin_unlock(&exp->exp_lock);
1940                         LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients));
1941                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
1942                 } else {
1943                         cfs_spin_unlock(&exp->exp_lock);
1944                 }
1945         }
1946         return 0;
1947 }
1948
1949 int target_queue_recovery_request(struct ptlrpc_request *req,
1950                                   struct obd_device *obd)
1951 {
1952         cfs_list_t *tmp;
1953         int inserted = 0;
1954         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1955         ENTRY;
1956
1957         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1958                 /* Processing the queue right now, don't re-add. */
1959                 RETURN(1);
1960         }
1961
1962         target_process_req_flags(obd, req);
1963
1964         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1965                 /* client declares he's ready to complete recovery
1966                  * so, we put the request on th final queue */
1967                 target_request_copy_get(req);
1968                 DEBUG_REQ(D_HA, req, "queue final req");
1969                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1970                 cfs_spin_lock(&obd->obd_recovery_task_lock);
1971                 if (obd->obd_recovering) {
1972                         cfs_list_add_tail(&req->rq_list,
1973                                           &obd->obd_final_req_queue);
1974                 } else {
1975                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1976                         target_request_copy_put(req);
1977                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1978                 }
1979                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1980                 RETURN(0);
1981         }
1982         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1983                 /* client declares he's ready to replay locks */
1984                 target_request_copy_get(req);
1985                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1986                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1987                 cfs_spin_lock(&obd->obd_recovery_task_lock);
1988                 LASSERT(obd->obd_recovering);
1989                 /* usually due to recovery abort */
1990                 if (!req->rq_export->exp_in_recovery) {
1991                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1992                         target_request_copy_put(req);
1993                         RETURN(-ENOTCONN);
1994                 }
1995                 LASSERT(req->rq_export->exp_lock_replay_needed);
1996                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1997                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1998                 RETURN(0);
1999         }
2000
2001         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2002          * (i.e. buflens etc are in my own byte order), but type-dependent
2003          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2004
2005         if (!transno) {
2006                 CFS_INIT_LIST_HEAD(&req->rq_list);
2007                 DEBUG_REQ(D_HA, req, "not queueing");
2008                 RETURN(1);
2009         }
2010
2011         /* If we're processing the queue, we want don't want to queue this
2012          * message.
2013          *
2014          * Also, if this request has a transno less than the one we're waiting
2015          * for, we should process it now.  It could (and currently always will)
2016          * be an open request for a descriptor that was opened some time ago.
2017          *
2018          * Also, a resent, replayed request that has already been
2019          * handled will pass through here and be processed immediately.
2020          */
2021         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2022               obd->obd_next_recovery_transno, transno);
2023         cfs_spin_lock(&obd->obd_recovery_task_lock);
2024         if (transno < obd->obd_next_recovery_transno) {
2025                 /* Processing the queue right now, don't re-add. */
2026                 LASSERT(cfs_list_empty(&req->rq_list));
2027                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2028                 RETURN(1);
2029         }
2030         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2031
2032         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2033                 RETURN(0);
2034
2035         target_request_copy_get(req);
2036         if (!req->rq_export->exp_in_recovery) {
2037                 target_request_copy_put(req);
2038                 RETURN(-ENOTCONN);
2039         }
2040         LASSERT(req->rq_export->exp_req_replay_needed);
2041
2042         if (target_exp_enqueue_req_replay(req)) {
2043                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2044                 target_request_copy_put(req);
2045                 RETURN(0);
2046         }
2047
2048         /* XXX O(n^2) */
2049         cfs_spin_lock(&obd->obd_recovery_task_lock);
2050         LASSERT(obd->obd_recovering);
2051         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2052                 struct ptlrpc_request *reqiter =
2053                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2054
2055                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2056                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2057                         inserted = 1;
2058                         break;
2059                 }
2060
2061                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2062                              transno)) {
2063                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2064                                   "has been claimed by another client");
2065                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2066                         target_exp_dequeue_req_replay(req);
2067                         target_request_copy_put(req);
2068                         RETURN(0);
2069                 }
2070         }
2071
2072         if (!inserted)
2073                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2074
2075         obd->obd_requests_queued_for_recovery++;
2076         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2077         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2078         RETURN(0);
2079 }
2080
2081 /**
2082  * Packs current SLV and Limit into \a req.
2083  */
2084 int target_pack_pool_reply(struct ptlrpc_request *req)
2085 {
2086         struct obd_device *obd;
2087         ENTRY;
2088
2089         /*
2090          * Check that we still have all structures alive as this may
2091          * be some late rpc in shutdown time.
2092          */
2093         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2094                      !exp_connect_lru_resize(req->rq_export))) {
2095                 lustre_msg_set_slv(req->rq_repmsg, 0);
2096                 lustre_msg_set_limit(req->rq_repmsg, 0);
2097                 RETURN(0);
2098         }
2099
2100         /*
2101          * OBD is alive here as export is alive, which we checked above.
2102          */
2103         obd = req->rq_export->exp_obd;
2104
2105         cfs_read_lock(&obd->obd_pool_lock);
2106         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2107         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2108         cfs_read_unlock(&obd->obd_pool_lock);
2109
2110         RETURN(0);
2111 }
2112
2113 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2114 {
2115         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2116                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2117                 return (-ECOMM);
2118         }
2119
2120         if (unlikely(rc)) {
2121                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2122                 req->rq_status = rc;
2123                 return (ptlrpc_send_error(req, 1));
2124         } else {
2125                 DEBUG_REQ(D_NET, req, "sending reply");
2126         }
2127
2128         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2129 }
2130
2131 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2132 {
2133         int                        netrc;
2134         struct ptlrpc_reply_state *rs;
2135         struct obd_device         *obd;
2136         struct obd_export         *exp;
2137         struct ptlrpc_service     *svc;
2138         ENTRY;
2139
2140         if (req->rq_no_reply) {
2141                 EXIT;
2142                 return;
2143         }
2144
2145         svc = req->rq_rqbd->rqbd_service;
2146         rs = req->rq_reply_state;
2147         if (rs == NULL || !rs->rs_difficult) {
2148                 /* no notifiers */
2149                 target_send_reply_msg (req, rc, fail_id);
2150                 EXIT;
2151                 return;
2152         }
2153
2154         /* must be an export if locks saved */
2155         LASSERT (req->rq_export != NULL);
2156         /* req/reply consistent */
2157         LASSERT (rs->rs_service == svc);
2158
2159         /* "fresh" reply */
2160         LASSERT (!rs->rs_scheduled);
2161         LASSERT (!rs->rs_scheduled_ever);
2162         LASSERT (!rs->rs_handled);
2163         LASSERT (!rs->rs_on_net);
2164         LASSERT (rs->rs_export == NULL);
2165         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2166         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2167
2168         exp = class_export_get (req->rq_export);
2169         obd = exp->exp_obd;
2170
2171         /* disable reply scheduling while I'm setting up */
2172         rs->rs_scheduled = 1;
2173         rs->rs_on_net    = 1;
2174         rs->rs_xid       = req->rq_xid;
2175         rs->rs_transno   = req->rq_transno;
2176         rs->rs_export    = exp;
2177         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2178
2179         cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2180         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2181                rs->rs_transno, exp->exp_last_committed);
2182         if (rs->rs_transno > exp->exp_last_committed) {
2183                 /* not committed already */
2184                 cfs_list_add_tail(&rs->rs_obd_list,
2185                                   &exp->exp_uncommitted_replies);
2186         }
2187         cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2188
2189         cfs_spin_lock(&exp->exp_lock);
2190         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2191         cfs_spin_unlock(&exp->exp_lock);
2192
2193         netrc = target_send_reply_msg (req, rc, fail_id);
2194
2195         cfs_spin_lock(&svc->srv_rs_lock);
2196
2197         cfs_atomic_inc(&svc->srv_n_difficult_replies);
2198
2199         if (netrc != 0) {
2200                 /* error sending: reply is off the net.  Also we need +1
2201                  * reply ref until ptlrpc_handle_rs() is done
2202                  * with the reply state (if the send was successful, there
2203                  * would have been +1 ref for the net, which
2204                  * reply_out_callback leaves alone) */
2205                 rs->rs_on_net = 0;
2206                 ptlrpc_rs_addref(rs);
2207         }
2208
2209         cfs_spin_lock(&rs->rs_lock);
2210         if (rs->rs_transno <= exp->exp_last_committed ||
2211             (!rs->rs_on_net && !rs->rs_no_ack) ||
2212              cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2213              cfs_list_empty(&rs->rs_obd_list)) {
2214                 CDEBUG(D_HA, "Schedule reply immediately\n");
2215                 ptlrpc_dispatch_difficult_reply(rs);
2216         } else {
2217                 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2218                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2219         }
2220         cfs_spin_unlock(&rs->rs_lock);
2221         cfs_spin_unlock(&svc->srv_rs_lock);
2222         EXIT;
2223 }
2224
2225 int target_handle_ping(struct ptlrpc_request *req)
2226 {
2227         obd_ping(req->rq_export);
2228         return req_capsule_server_pack(&req->rq_pill);
2229 }
2230
2231 void target_committed_to_req(struct ptlrpc_request *req)
2232 {
2233         struct obd_export *exp = req->rq_export;
2234
2235         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2236                 lustre_msg_set_last_committed(req->rq_repmsg,
2237                                               exp->exp_last_committed);
2238         else
2239                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2240                           "%d)", exp->exp_obd->obd_no_transno,
2241                           req->rq_repmsg == NULL);
2242
2243         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2244                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2245 }
2246 EXPORT_SYMBOL(target_committed_to_req);
2247
2248 int target_handle_qc_callback(struct ptlrpc_request *req)
2249 {
2250         struct obd_quotactl *oqctl;
2251         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2252
2253         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2254         if (oqctl == NULL) {
2255                 CERROR("Can't unpack obd_quotactl\n");
2256                 RETURN(-EPROTO);
2257         }
2258
2259         cli->cl_qchk_stat = oqctl->qc_stat;
2260
2261         return 0;
2262 }
2263
2264 #ifdef HAVE_QUOTA_SUPPORT
2265 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2266 {
2267 #ifdef __KERNEL__
2268         struct obd_device *obd = req->rq_export->exp_obd;
2269         struct obd_device *master_obd = NULL, *lov_obd = NULL;
2270         struct obd_device_target *obt;
2271         struct lustre_quota_ctxt *qctxt;
2272         struct qunit_data *qdata = NULL;
2273         int rc = 0;
2274         ENTRY;
2275
2276         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2277                 RETURN(rc);
2278
2279         rc = req_capsule_server_pack(&req->rq_pill);
2280         if (rc) {
2281                 CERROR("packing reply failed!: rc = %d\n", rc);
2282                 RETURN(rc);
2283         }
2284
2285         LASSERT(req->rq_export);
2286
2287         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2288         if (IS_ERR(qdata)) {
2289                 rc = PTR_ERR(qdata);
2290                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2291                 req->rq_status = rc;
2292                 GOTO(out, rc);
2293         }
2294
2295         /* we use the observer */
2296         if (obd_pin_observer(obd, &lov_obd) ||
2297             obd_pin_observer(lov_obd, &master_obd)) {
2298                 CERROR("Can't find the observer, it is recovering\n");
2299                 req->rq_status = -EAGAIN;
2300                 GOTO(out, rc);
2301         }
2302
2303         obt = &master_obd->u.obt;
2304         qctxt = &obt->obt_qctxt;
2305
2306         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2307                 /* quota_type has not been processed yet, return EAGAIN
2308                  * until we know whether or not quotas are supposed to
2309                  * be enabled */
2310                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2311                        "-EAGAIN\n");
2312                 req->rq_status = -EAGAIN;
2313                 GOTO(out, rc);
2314         }
2315
2316         cfs_down_read(&obt->obt_rwsem);
2317         if (qctxt->lqc_lqs_hash == NULL) {
2318                 cfs_up_read(&obt->obt_rwsem);
2319                 /* quota_type has not been processed yet, return EAGAIN
2320                  * until we know whether or not quotas are supposed to
2321                  * be enabled */
2322                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2323                        "-EAGAIN\n");
2324                 req->rq_status = -EAGAIN;
2325                 GOTO(out, rc);
2326         }
2327
2328         LASSERT(qctxt->lqc_handler);
2329         rc = qctxt->lqc_handler(master_obd, qdata,
2330                                 lustre_msg_get_opc(req->rq_reqmsg));
2331         cfs_up_read(&obt->obt_rwsem);
2332         if (rc && rc != -EDQUOT)
2333                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2334                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2335         req->rq_status = rc;
2336
2337         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2338         if (rc < 0) {
2339                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2340                 GOTO(out, rc);
2341         }
2342
2343         /* Block the quota req. b=14840 */
2344         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2345         EXIT;
2346
2347 out:
2348         if (master_obd)
2349                 obd_unpin_observer(lov_obd);
2350         if (lov_obd)
2351                 obd_unpin_observer(obd);
2352
2353         rc = ptlrpc_reply(req);
2354         return rc;
2355 #else
2356         return 0;
2357 #endif /* !__KERNEL__ */
2358 }
2359 #endif /* HAVE_QUOTA_SUPPORT */
2360
2361 ldlm_mode_t lck_compat_array[] = {
2362         [LCK_EX] LCK_COMPAT_EX,
2363         [LCK_PW] LCK_COMPAT_PW,
2364         [LCK_PR] LCK_COMPAT_PR,
2365         [LCK_CW] LCK_COMPAT_CW,
2366         [LCK_CR] LCK_COMPAT_CR,
2367         [LCK_NL] LCK_COMPAT_NL,
2368         [LCK_GROUP] LCK_COMPAT_GROUP,
2369         [LCK_COS] LCK_COMPAT_COS,
2370 };
2371
2372 /**
2373  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2374  * not escape to the user level.
2375  */
2376 int ldlm_error2errno(ldlm_error_t error)
2377 {
2378         int result;
2379
2380         switch (error) {
2381         case ELDLM_OK:
2382                 result = 0;
2383                 break;
2384         case ELDLM_LOCK_CHANGED:
2385                 result = -ESTALE;
2386                 break;
2387         case ELDLM_LOCK_ABORTED:
2388                 result = -ENAVAIL;
2389                 break;
2390         case ELDLM_LOCK_REPLACED:
2391                 result = -ESRCH;
2392                 break;
2393         case ELDLM_NO_LOCK_DATA:
2394                 result = -ENOENT;
2395                 break;
2396         case ELDLM_NAMESPACE_EXISTS:
2397                 result = -EEXIST;
2398                 break;
2399         case ELDLM_BAD_NAMESPACE:
2400                 result = -EBADF;
2401                 break;
2402         default:
2403                 if (((int)error) < 0)  /* cast to signed type */
2404                         result = error; /* as ldlm_error_t can be unsigned */
2405                 else {
2406                         CERROR("Invalid DLM result code: %d\n", error);
2407                         result = -EPROTO;
2408                 }
2409         }
2410         return result;
2411 }
2412 EXPORT_SYMBOL(ldlm_error2errno);
2413
2414 /**
2415  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2416  */
2417 ldlm_error_t ldlm_errno2error(int err_no)
2418 {
2419         int error;
2420
2421         switch (err_no) {
2422         case 0:
2423                 error = ELDLM_OK;
2424                 break;
2425         case -ESTALE:
2426                 error = ELDLM_LOCK_CHANGED;
2427                 break;
2428         case -ENAVAIL:
2429                 error = ELDLM_LOCK_ABORTED;
2430                 break;
2431         case -ESRCH:
2432                 error = ELDLM_LOCK_REPLACED;
2433                 break;
2434         case -ENOENT:
2435                 error = ELDLM_NO_LOCK_DATA;
2436                 break;
2437         case -EEXIST:
2438                 error = ELDLM_NAMESPACE_EXISTS;
2439                 break;
2440         case -EBADF:
2441                 error = ELDLM_BAD_NAMESPACE;
2442                 break;
2443         default:
2444                 error = err_no;
2445         }
2446         return error;
2447 }
2448 EXPORT_SYMBOL(ldlm_errno2error);
2449
2450 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2451 void ldlm_dump_export_locks(struct obd_export *exp)
2452 {
2453         cfs_spin_lock(&exp->exp_locks_list_guard);
2454         if (!cfs_list_empty(&exp->exp_locks_list)) {
2455             struct ldlm_lock *lock;
2456
2457             CERROR("dumping locks for export %p,"
2458                    "ignore if the unmount doesn't hang\n", exp);
2459             cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2460                 ldlm_lock_dump(D_ERROR, lock, 0);
2461         }
2462         cfs_spin_unlock(&exp->exp_locks_list_guard);
2463 }
2464 #endif