Whamcloud - gitweb
350f3a954a71ebef92ad11e2efb4a66d34109872
[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  2008 Sun Microsystems, Inc. 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         int rc;
218         ENTRY;
219
220         /* In a more perfect world, we would hang a ptlrpc_client off of
221          * obd_type and just use the values from there. */
222         if (!strcmp(name, LUSTRE_OSC_NAME)) {
223                 rq_portal = OST_REQUEST_PORTAL;
224                 rp_portal = OSC_REPLY_PORTAL;
225                 connect_op = OST_CONNECT;
226                 cli->cl_sp_me = LUSTRE_SP_CLI;
227                 cli->cl_sp_to = LUSTRE_SP_OST;
228         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
229                 rq_portal = MDS_REQUEST_PORTAL;
230                 rp_portal = MDC_REPLY_PORTAL;
231                 connect_op = MDS_CONNECT;
232                 cli->cl_sp_me = LUSTRE_SP_CLI;
233                 cli->cl_sp_to = LUSTRE_SP_MDT;
234         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
235                 rq_portal = MGS_REQUEST_PORTAL;
236                 rp_portal = MGC_REPLY_PORTAL;
237                 connect_op = MGS_CONNECT;
238                 cli->cl_sp_me = LUSTRE_SP_MGC;
239                 cli->cl_sp_to = LUSTRE_SP_MGS;
240                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
241         } else {
242                 CERROR("unknown client OBD type \"%s\", can't setup\n",
243                        name);
244                 RETURN(-EINVAL);
245         }
246
247         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
248                 CERROR("requires a TARGET UUID\n");
249                 RETURN(-EINVAL);
250         }
251
252         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
253                 CERROR("client UUID must be less than 38 characters\n");
254                 RETURN(-EINVAL);
255         }
256
257         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
258                 CERROR("setup requires a SERVER UUID\n");
259                 RETURN(-EINVAL);
260         }
261
262         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
263                 CERROR("target UUID must be less than 38 characters\n");
264                 RETURN(-EINVAL);
265         }
266
267         cfs_init_rwsem(&cli->cl_sem);
268         cfs_sema_init(&cli->cl_mgc_sem, 1);
269         cli->cl_conn_count = 0;
270         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
271                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
272                      sizeof(server_uuid)));
273
274         cli->cl_dirty = 0;
275         cli->cl_avail_grant = 0;
276         /* FIXME: should limit this for the sum of all cl_dirty_max */
277         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
278         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
279                 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
280         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
281         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
282         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
283         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
284         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
285         client_obd_list_lock_init(&cli->cl_loi_list_lock);
286         cli->cl_r_in_flight = 0;
287         cli->cl_w_in_flight = 0;
288
289         cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
290         cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
291         cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
292         cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
293         cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
294         cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
295         cfs_waitq_init(&cli->cl_destroy_waitq);
296         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
297 #ifdef ENABLE_CHECKSUM
298         /* Turn on checksumming by default. */
299         cli->cl_checksum = 1;
300         /*
301          * The supported checksum types will be worked out at connect time
302          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
303          * through procfs.
304          */
305         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
306 #endif
307         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
308         cfs_atomic_set(&cli->cl_quota_resends, CLIENT_QUOTA_DEFAULT_RESENDS);
309
310         /* This value may be changed at connect time in
311            ptlrpc_connect_interpret. */
312         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
313                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
314
315         if (!strcmp(name, LUSTRE_MDC_NAME)) {
316                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
317         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
318                 cli->cl_max_rpcs_in_flight = 2;
319         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
320                 cli->cl_max_rpcs_in_flight = 3;
321         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
322                 cli->cl_max_rpcs_in_flight = 4;
323         } else {
324                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
325         }
326
327         rc = ldlm_get_ref();
328         if (rc) {
329                 CERROR("ldlm_get_ref failed: %d\n", rc);
330                 GOTO(err, rc);
331         }
332
333         ptlrpc_init_client(rq_portal, rp_portal, name,
334                            &obddev->obd_ldlm_client);
335
336         imp = class_new_import(obddev);
337         if (imp == NULL)
338                 GOTO(err_ldlm, rc = -ENOENT);
339         imp->imp_client = &obddev->obd_ldlm_client;
340         imp->imp_connect_op = connect_op;
341         imp->imp_initial_recov = 1;
342         imp->imp_initial_recov_bk = 0;
343         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
344         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
345                LUSTRE_CFG_BUFLEN(lcfg, 1));
346         class_import_put(imp);
347
348         rc = client_import_add_conn(imp, &server_uuid, 1);
349         if (rc) {
350                 CERROR("can't add initial connection\n");
351                 GOTO(err_import, rc);
352         }
353
354         cli->cl_import = imp;
355         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
356         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
357         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
358
359         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
360                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
361                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
362                                name, obddev->obd_name,
363                                cli->cl_target_uuid.uuid);
364                         cfs_spin_lock(&imp->imp_lock);
365                         imp->imp_deactive = 1;
366                         cfs_spin_unlock(&imp->imp_lock);
367                 }
368         }
369
370         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
371                                                    LDLM_NAMESPACE_CLIENT,
372                                                    LDLM_NAMESPACE_GREEDY);
373         if (obddev->obd_namespace == NULL) {
374                 CERROR("Unable to create client namespace - %s\n",
375                        obddev->obd_name);
376                 GOTO(err_import, rc = -ENOMEM);
377         }
378
379         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
380
381         RETURN(rc);
382
383 err_import:
384         class_destroy_import(imp);
385 err_ldlm:
386         ldlm_put_ref();
387 err:
388         RETURN(rc);
389
390 }
391
392 int client_obd_cleanup(struct obd_device *obddev)
393 {
394         ENTRY;
395
396         ldlm_namespace_free_post(obddev->obd_namespace);
397         obddev->obd_namespace = NULL;
398
399         ldlm_put_ref();
400         RETURN(0);
401 }
402
403 /* ->o_connect() method for client side (OSC and MDC and MGC) */
404 int client_connect_import(const struct lu_env *env,
405                           struct obd_export **exp,
406                           struct obd_device *obd, struct obd_uuid *cluuid,
407                           struct obd_connect_data *data, void *localdata)
408 {
409         struct client_obd *cli = &obd->u.cli;
410         struct obd_import *imp = cli->cl_import;
411         struct obd_connect_data *ocd;
412         struct lustre_handle conn = { 0 };
413         int rc;
414         ENTRY;
415
416         *exp = NULL;
417         cfs_down_write(&cli->cl_sem);
418         if (cli->cl_conn_count > 0 )
419                 GOTO(out_sem, rc = -EALREADY);
420
421         rc = class_connect(&conn, obd, cluuid);
422         if (rc)
423                 GOTO(out_sem, rc);
424
425         cli->cl_conn_count++;
426         *exp = class_conn2export(&conn);
427
428         LASSERT(obd->obd_namespace);
429
430         imp->imp_dlm_handle = conn;
431         rc = ptlrpc_init_import(imp);
432         if (rc != 0)
433                 GOTO(out_ldlm, rc);
434
435         ocd = &imp->imp_connect_data;
436         if (data) {
437                 *ocd = *data;
438                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
439         }
440
441         rc = ptlrpc_connect_import(imp, NULL);
442         if (rc != 0) {
443                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
444                 GOTO(out_ldlm, rc);
445         }
446         LASSERT((*exp)->exp_connection);
447
448         if (data) {
449                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
450                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
451                          data->ocd_connect_flags, ocd->ocd_connect_flags);
452                 data->ocd_connect_flags = ocd->ocd_connect_flags;
453         }
454
455         ptlrpc_pinger_add_import(imp);
456
457         EXIT;
458
459         if (rc) {
460 out_ldlm:
461                 cli->cl_conn_count--;
462                 class_disconnect(*exp);
463                 *exp = NULL;
464         }
465 out_sem:
466         cfs_up_write(&cli->cl_sem);
467
468         return rc;
469 }
470
471 int client_disconnect_export(struct obd_export *exp)
472 {
473         struct obd_device *obd = class_exp2obd(exp);
474         struct client_obd *cli;
475         struct obd_import *imp;
476         int rc = 0, err;
477         ENTRY;
478
479         if (!obd) {
480                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
481                        exp, exp ? exp->exp_handle.h_cookie : -1);
482                 RETURN(-EINVAL);
483         }
484
485         cli = &obd->u.cli;
486         imp = cli->cl_import;
487
488         cfs_down_write(&cli->cl_sem);
489         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
490                cli->cl_conn_count);
491
492         if (!cli->cl_conn_count) {
493                 CERROR("disconnecting disconnected device (%s)\n",
494                        obd->obd_name);
495                 GOTO(out_disconnect, rc = -EINVAL);
496         }
497
498         cli->cl_conn_count--;
499         if (cli->cl_conn_count)
500                 GOTO(out_disconnect, rc = 0);
501
502         /* Mark import deactivated now, so we don't try to reconnect if any
503          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
504          * fully deactivate the import, or that would drop all requests. */
505         cfs_spin_lock(&imp->imp_lock);
506         imp->imp_deactive = 1;
507         cfs_spin_unlock(&imp->imp_lock);
508
509         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
510          * delete it regardless.  (It's safe to delete an import that was
511          * never added.) */
512         (void)ptlrpc_pinger_del_import(imp);
513
514         if (obd->obd_namespace != NULL) {
515                 /* obd_force == local only */
516                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
517                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
518                                        NULL);
519                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
520         }
521
522         /*
523          * there's no need to hold sem during disconnecting an import,
524          * and actually it may cause deadlock in gss.
525          */
526         cfs_up_write(&cli->cl_sem);
527         rc = ptlrpc_disconnect_import(imp, 0);
528         cfs_down_write(&cli->cl_sem);
529
530         ptlrpc_invalidate_import(imp);
531
532         if (imp->imp_rq_pool) {
533                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
534                 imp->imp_rq_pool = NULL;
535         }
536         client_destroy_import(imp);
537         cli->cl_import = NULL;
538
539         EXIT;
540
541  out_disconnect:
542         /* use server style - class_disconnect should be always called for
543          * o_disconnect */
544         err = class_disconnect(exp);
545         if (!rc && err)
546                 rc = err;
547
548         cfs_up_write(&cli->cl_sem);
549
550         RETURN(rc);
551 }
552
553 int server_disconnect_export(struct obd_export *exp)
554 {
555         int rc;
556         ENTRY;
557
558         /* Disconnect early so that clients can't keep using export */
559         rc = class_disconnect(exp);
560         /* close import for avoid sending any requests */
561         if (exp->exp_imp_reverse)
562                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
563
564         if (exp->exp_obd->obd_namespace != NULL)
565                 ldlm_cancel_locks_for_export(exp);
566
567         /* complete all outstanding replies */
568         cfs_spin_lock(&exp->exp_lock);
569         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
570                 struct ptlrpc_reply_state *rs =
571                         cfs_list_entry(exp->exp_outstanding_replies.next,
572                                        struct ptlrpc_reply_state, rs_exp_list);
573                 struct ptlrpc_service *svc = rs->rs_service;
574
575                 cfs_spin_lock(&svc->srv_lock);
576                 cfs_list_del_init(&rs->rs_exp_list);
577                 cfs_spin_lock(&rs->rs_lock);
578                 ptlrpc_schedule_difficult_reply(rs);
579                 cfs_spin_unlock(&rs->rs_lock);
580                 cfs_spin_unlock(&svc->srv_lock);
581         }
582         cfs_spin_unlock(&exp->exp_lock);
583
584         /* release nid stat refererence */
585         lprocfs_exp_cleanup(exp);
586
587         RETURN(rc);
588 }
589
590 /* --------------------------------------------------------------------------
591  * from old lib/target.c
592  * -------------------------------------------------------------------------- */
593
594 static int target_handle_reconnect(struct lustre_handle *conn,
595                                    struct obd_export *exp,
596                                    struct obd_uuid *cluuid)
597 {
598         ENTRY;
599         if (exp->exp_connection && exp->exp_imp_reverse) {
600                 struct lustre_handle *hdl;
601                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
602                 /* Might be a re-connect after a partition. */
603                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
604                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
605                               cluuid->uuid);
606                         conn->cookie = exp->exp_handle.h_cookie;
607                         /* target_handle_connect() treats EALREADY and
608                          * -EALREADY differently.  EALREADY means we are
609                          * doing a valid reconnect from the same client. */
610                         RETURN(EALREADY);
611                 } else {
612                         CERROR("%s reconnecting from %s, "
613                                "handle mismatch (ours "LPX64", theirs "
614                                LPX64")\n", cluuid->uuid,
615                                exp->exp_connection->c_remote_uuid.uuid,
616                                hdl->cookie, conn->cookie);
617                         memset(conn, 0, sizeof *conn);
618                         /* target_handle_connect() treats EALREADY and
619                          * -EALREADY differently.  -EALREADY is an error
620                          * (same UUID, different handle). */
621                         RETURN(-EALREADY);
622                 }
623         }
624
625         conn->cookie = exp->exp_handle.h_cookie;
626         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
627                cluuid->uuid, exp, conn->cookie);
628         RETURN(0);
629 }
630
631 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
632                           int error)
633 {
634         struct obd_export *exp = cb_data;
635
636         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
637                obd->obd_name, exp->exp_client_uuid.uuid);
638
639         cfs_spin_lock(&exp->exp_lock);
640         exp->exp_need_sync = 0;
641         cfs_spin_unlock(&exp->exp_lock);
642         class_export_cb_put(exp);
643 }
644 EXPORT_SYMBOL(target_client_add_cb);
645
646 static void
647 target_start_and_reset_recovery_timer(struct obd_device *obd,
648                                       struct ptlrpc_request *req,
649                                       int new_client);
650
651 int target_handle_connect(struct ptlrpc_request *req)
652 {
653         struct obd_device *target, *targref = NULL;
654         struct obd_export *export = NULL;
655         struct obd_import *revimp;
656         struct lustre_handle conn;
657         struct lustre_handle *tmp;
658         struct obd_uuid tgtuuid;
659         struct obd_uuid cluuid;
660         struct obd_uuid remote_uuid;
661         char *str;
662         int rc = 0;
663         int mds_conn = 0;
664         struct obd_connect_data *data, *tmpdata;
665         lnet_nid_t *client_nid = NULL;
666         ENTRY;
667
668         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
669
670         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
671         if (str == NULL) {
672                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
673                 GOTO(out, rc = -EINVAL);
674         }
675
676         obd_str2uuid(&tgtuuid, str);
677         target = class_uuid2obd(&tgtuuid);
678         if (!target)
679                 target = class_name2obd(str);
680
681         if (!target || target->obd_stopping || !target->obd_set_up) {
682                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
683                                    " for connect (%s)\n", str,
684                                    !target ? "no target" :
685                                    (target->obd_stopping ? "stopping" :
686                                    "not set up"));
687                 GOTO(out, rc = -ENODEV);
688         }
689
690         if (target->obd_no_conn) {
691                 LCONSOLE_WARN("%s: temporarily refusing client connection "
692                               "from %s\n", target->obd_name,
693                               libcfs_nid2str(req->rq_peer.nid));
694                 GOTO(out, rc = -EAGAIN);
695         }
696
697         /* Make sure the target isn't cleaned up while we're here. Yes,
698            there's still a race between the above check and our incref here.
699            Really, class_uuid2obd should take the ref. */
700         targref = class_incref(target, __FUNCTION__, cfs_current());
701
702
703         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
704         if (str == NULL) {
705                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
706                 GOTO(out, rc = -EINVAL);
707         }
708
709         obd_str2uuid(&cluuid, str);
710
711         /* XXX extract a nettype and format accordingly */
712         switch (sizeof(lnet_nid_t)) {
713                 /* NB the casts only avoid compiler warnings */
714         case 8:
715                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
716                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
717                 break;
718         case 4:
719                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
720                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
721                 break;
722         default:
723                 LBUG();
724         }
725
726         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
727         if (tmp == NULL)
728                 GOTO(out, rc = -EPROTO);
729
730         conn = *tmp;
731
732         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
733         if (!data)
734                 GOTO(out, rc = -EPROTO);
735
736         rc = req_capsule_server_pack(&req->rq_pill);
737         if (rc)
738                 GOTO(out, rc);
739
740         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
741                 if (!data) {
742                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
743                                   "libclient connection attempt");
744                         GOTO(out, rc = -EPROTO);
745                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
746                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
747                            data->ocd_version > LUSTRE_VERSION_CODE +
748                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
749                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
750                                   "libclient connection attempt",
751                                   data->ocd_version < LUSTRE_VERSION_CODE ?
752                                   "old" : "new",
753                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
754                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
755                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
756                                   OBD_OCD_VERSION_FIX(data->ocd_version));
757                         data = req_capsule_server_sized_get(&req->rq_pill,
758                                                         &RMF_CONNECT_DATA,
759                                     offsetof(typeof(*data), ocd_version) +
760                                              sizeof(data->ocd_version));
761                         if (data) {
762                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
763                                 data->ocd_version = LUSTRE_VERSION_CODE;
764                         }
765                         GOTO(out, rc = -EPROTO);
766                 }
767         }
768
769         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
770             (data->ocd_connect_flags & OBD_CONNECT_MDS))
771                 mds_conn = 1;
772
773         /* lctl gets a backstage, all-access pass. */
774         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
775                 goto dont_check_exports;
776
777         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
778         if (!export)
779                 goto no_export;
780
781         /* we've found an export in the hash */
782         if (export->exp_connecting) { /* bug 9635, et. al. */
783                 CWARN("%s: exp %p already connecting\n",
784                       export->exp_obd->obd_name, export);
785                 class_export_put(export);
786                 export = NULL;
787                 rc = -EALREADY;
788         } else if (mds_conn && export->exp_connection) {
789                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
790                         /* mds reconnected after failover */
791                         CWARN("%s: received MDS connection from NID %s,"
792                               " removing former export from NID %s\n",
793                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
794                             libcfs_nid2str(export->exp_connection->c_peer.nid));
795                 else
796                         /* new mds connection from the same nid */
797                         CWARN("%s: received new MDS connection from NID %s,"
798                               " removing former export from same NID\n",
799                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
800                 class_fail_export(export);
801                 class_export_put(export);
802                 export = NULL;
803                 rc = 0;
804         } else if (export->exp_connection != NULL &&
805                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
806                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
807                     MSG_CONNECT_INITIAL)) {
808                 /* in mds failover we have static uuid but nid can be
809                  * changed*/
810                 CWARN("%s: cookie %s seen on new NID %s when "
811                       "existing NID %s is already connected\n",
812                       target->obd_name, cluuid.uuid,
813                       libcfs_nid2str(req->rq_peer.nid),
814                       libcfs_nid2str(export->exp_connection->c_peer.nid));
815                 rc = -EALREADY;
816                 class_export_put(export);
817                 export = NULL;
818         } else {
819                 cfs_spin_lock(&export->exp_lock);
820                 export->exp_connecting = 1;
821                 cfs_spin_unlock(&export->exp_lock);
822                 class_export_put(export);
823                 LASSERT(export->exp_obd == target);
824
825                 rc = target_handle_reconnect(&conn, export, &cluuid);
826         }
827
828         /* If we found an export, we already unlocked. */
829         if (!export) {
830 no_export:
831                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
832         } else if (req->rq_export == NULL &&
833                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
834                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
835                       target->obd_name, cluuid.uuid,
836                       libcfs_nid2str(req->rq_peer.nid),
837                       export, cfs_atomic_read(&export->exp_refcount));
838                 GOTO(out, rc = -EBUSY);
839         } else if (req->rq_export != NULL &&
840                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
841                 /* the current connect rpc has increased exp_rpc_count */
842                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
843                       target->obd_name, cluuid.uuid,
844                       libcfs_nid2str(req->rq_peer.nid),
845                       export, cfs_atomic_read(&export->exp_rpc_count) - 1);
846                 cfs_spin_lock(&export->exp_lock);
847                 if (req->rq_export->exp_conn_cnt <
848                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
849                         /* try to abort active requests */
850                         req->rq_export->exp_abort_active_req = 1;
851                 cfs_spin_unlock(&export->exp_lock);
852                 GOTO(out, rc = -EBUSY);
853         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
854                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
855                        "cookies not random?\n", target->obd_name,
856                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
857                 GOTO(out, rc = -EALREADY);
858         } else {
859                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
860         }
861
862         if (rc < 0) {
863                 GOTO(out, rc);
864         }
865
866         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
867                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
868               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
869               export, (long)cfs_time_current_sec(),
870               export ? (long)export->exp_last_request_time : 0);
871
872         /* Tell the client if we're in recovery. */
873         if (target->obd_recovering) {
874                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
875                 /* If this is the first time a client connects,
876                    reset the recovery timer */
877                 if (rc == 0)
878                         target_start_and_reset_recovery_timer(target, req,
879                                                               !export);
880         }
881
882         /* We want to handle EALREADY but *not* -EALREADY from
883          * target_handle_reconnect(), return reconnection state in a flag */
884         if (rc == EALREADY) {
885                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
886                 rc = 0;
887         } else {
888                 LASSERT(rc == 0);
889         }
890
891         /* Tell the client if we support replayable requests */
892         if (target->obd_replayable)
893                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
894         client_nid = &req->rq_peer.nid;
895
896         if (export == NULL) {
897                 if (target->obd_recovering) {
898                         cfs_time_t t;
899
900                         t = cfs_timer_deadline(&target->obd_recovery_timer);
901                         t = cfs_time_sub(t, cfs_time_current());
902                         CERROR("%s: denying connection for new client %s (%s): "
903                                "%d clients in recovery for "CFS_TIME_T"s\n",
904                                target->obd_name,
905                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
906                                cfs_atomic_read(&target-> \
907                                                obd_lock_replay_clients),
908                                cfs_duration_sec(t));
909                         rc = -EBUSY;
910                 } else {
911 dont_check_exports:
912                         rc = obd_connect(req->rq_svc_thread->t_env,
913                                          &export, target, &cluuid, data,
914                                          client_nid);
915                         if (rc == 0)
916                                 conn.cookie = export->exp_handle.h_cookie;
917                 }
918         } else {
919                 rc = obd_reconnect(req->rq_svc_thread->t_env,
920                                    export, target, &cluuid, data, client_nid);
921                 if (rc == 0)
922                         /* prevous done via class_conn2export */
923                         class_export_get(export);
924         }
925         if (rc)
926                 GOTO(out, rc);
927         /* Return only the parts of obd_connect_data that we understand, so the
928          * client knows that we don't understand the rest. */
929         if (data) {
930                  tmpdata = req_capsule_server_get(&req->rq_pill,
931                                                   &RMF_CONNECT_DATA);
932                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
933                  *tmpdata = *data;
934         }
935
936         /* If all else goes well, this is our RPC return code. */
937         req->rq_status = 0;
938
939         lustre_msg_set_handle(req->rq_repmsg, &conn);
940
941         /* If the client and the server are the same node, we will already
942          * have an export that really points to the client's DLM export,
943          * because we have a shared handles table.
944          *
945          * XXX this will go away when shaver stops sending the "connect" handle
946          * in the real "remote handle" field of the request --phik 24 Apr 2003
947          */
948         if (req->rq_export != NULL)
949                 class_export_put(req->rq_export);
950
951         req->rq_export = export;
952
953         cfs_spin_lock(&export->exp_lock);
954         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
955                 cfs_spin_unlock(&export->exp_lock);
956                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
957                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
958                        export->exp_conn_cnt,
959                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
960
961                 GOTO(out, rc = -EALREADY);
962         }
963         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
964         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
965         export->exp_abort_active_req = 0;
966
967         /* request from liblustre?  Don't evict it for not pinging. */
968         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
969                 export->exp_libclient = 1;
970                 cfs_spin_unlock(&export->exp_lock);
971
972                 cfs_spin_lock(&target->obd_dev_lock);
973                 cfs_list_del_init(&export->exp_obd_chain_timed);
974                 cfs_spin_unlock(&target->obd_dev_lock);
975         } else {
976                 cfs_spin_unlock(&export->exp_lock);
977         }
978
979         if (export->exp_connection != NULL) {
980                 /* Check to see if connection came from another NID */
981                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
982                     !cfs_hlist_unhashed(&export->exp_nid_hash))
983                         cfs_hash_del(export->exp_obd->obd_nid_hash,
984                                      &export->exp_connection->c_peer.nid,
985                                      &export->exp_nid_hash);
986
987                 ptlrpc_connection_put(export->exp_connection);
988         }
989
990         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
991                                                        req->rq_self,
992                                                        &remote_uuid);
993         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
994                 cfs_hash_add_unique(export->exp_obd->obd_nid_hash,
995                                     &export->exp_connection->c_peer.nid,
996                                     &export->exp_nid_hash);
997         }
998
999         cfs_spin_lock_bh(&target->obd_processing_task_lock);
1000         if (target->obd_recovering && !export->exp_in_recovery) {
1001                 cfs_spin_lock(&export->exp_lock);
1002                 export->exp_in_recovery = 1;
1003                 export->exp_req_replay_needed = 1;
1004                 export->exp_lock_replay_needed = 1;
1005                 cfs_spin_unlock(&export->exp_lock);
1006                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1007                      && (data->ocd_transno == 0))
1008                         CWARN("Connect with zero transno!\n");
1009
1010                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1011                      && data->ocd_transno < target->obd_next_recovery_transno)
1012                         target->obd_next_recovery_transno = data->ocd_transno;
1013                 target->obd_connected_clients++;
1014                 cfs_atomic_inc(&target->obd_req_replay_clients);
1015                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1016                 if (target->obd_connected_clients ==
1017                     target->obd_max_recoverable_clients)
1018                         cfs_waitq_signal(&target->obd_next_transno_waitq);
1019         }
1020         cfs_spin_unlock_bh(&target->obd_processing_task_lock);
1021         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1022         conn = *tmp;
1023
1024         if (export->exp_imp_reverse != NULL) {
1025                 /* destroyed import can be still referenced in ctxt */
1026                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1027                                    KEY_REVIMP_UPD, 0, NULL, NULL);
1028
1029                 client_destroy_import(export->exp_imp_reverse);
1030         }
1031
1032         /* for the rest part, we return -ENOTCONN in case of errors
1033          * in order to let client initialize connection again.
1034          */
1035         revimp = export->exp_imp_reverse = class_new_import(target);
1036         if (!revimp) {
1037                 CERROR("fail to alloc new reverse import.\n");
1038                 GOTO(out, rc = -ENOTCONN);
1039         }
1040
1041         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1042         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1043         revimp->imp_remote_handle = conn;
1044         revimp->imp_dlm_fake = 1;
1045         revimp->imp_state = LUSTRE_IMP_FULL;
1046
1047         /* unknown versions will be caught in
1048          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1049         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1050
1051         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1052             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1053                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1054         else
1055                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1056
1057         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1058         if (rc) {
1059                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1060                 export->exp_imp_reverse = NULL;
1061                 class_destroy_import(revimp);
1062         }
1063
1064         class_import_put(revimp);
1065 out:
1066         if (export) {
1067                 cfs_spin_lock(&export->exp_lock);
1068                 export->exp_connecting = 0;
1069                 cfs_spin_unlock(&export->exp_lock);
1070         }
1071         if (targref)
1072                 class_decref(targref, __FUNCTION__, cfs_current());
1073         if (rc)
1074                 req->rq_status = rc;
1075         RETURN(rc);
1076 }
1077
1078 int target_handle_disconnect(struct ptlrpc_request *req)
1079 {
1080         int rc;
1081         ENTRY;
1082
1083         rc = req_capsule_server_pack(&req->rq_pill);
1084         if (rc)
1085                 RETURN(rc);
1086
1087         /* keep the rq_export around so we can send the reply */
1088         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1089
1090         RETURN(0);
1091 }
1092
1093 void target_destroy_export(struct obd_export *exp)
1094 {
1095         /* exports created from last_rcvd data, and "fake"
1096            exports created by lctl don't have an import */
1097         if (exp->exp_imp_reverse != NULL)
1098                 client_destroy_import(exp->exp_imp_reverse);
1099
1100         LASSERT(cfs_atomic_read(&exp->exp_locks_count) == 0);
1101         LASSERT(cfs_atomic_read(&exp->exp_rpc_count) == 0);
1102         LASSERT(cfs_atomic_read(&exp->exp_cb_count) == 0);
1103         LASSERT(cfs_atomic_read(&exp->exp_replay_count) == 0);
1104 }
1105
1106 /*
1107  * Recovery functions
1108  */
1109 static void target_request_copy_get(struct ptlrpc_request *req)
1110 {
1111         class_export_rpc_get(req->rq_export);
1112         LASSERT(cfs_list_empty(&req->rq_list));
1113         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1114         /* increase refcount to keep request in queue */
1115         LASSERT(cfs_atomic_read(&req->rq_refcount));
1116         cfs_atomic_inc(&req->rq_refcount);
1117         /** let export know it has replays to be handled */
1118         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1119 }
1120
1121 static void target_request_copy_put(struct ptlrpc_request *req)
1122 {
1123         LASSERT(cfs_list_empty(&req->rq_replay_list));
1124         LASSERT(cfs_atomic_read(&req->rq_export->exp_replay_count) > 0);
1125         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1126         class_export_rpc_put(req->rq_export);
1127         ptlrpc_server_drop_request(req);
1128 }
1129
1130 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1131 {
1132         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1133         struct obd_export     *exp = req->rq_export;
1134         struct ptlrpc_request *reqiter;
1135         int                    dup = 0;
1136
1137         LASSERT(exp);
1138
1139         cfs_spin_lock(&exp->exp_lock);
1140         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1141                                 rq_replay_list) {
1142                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1143                         dup = 1;
1144                         break;
1145                 }
1146         }
1147
1148         if (dup) {
1149                 /* we expect it with RESENT and REPLAY flags */
1150                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1151                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1152                         CERROR("invalid flags %x of resent replay\n",
1153                                lustre_msg_get_flags(req->rq_reqmsg));
1154         } else {
1155                 cfs_list_add_tail(&req->rq_replay_list,
1156                                   &exp->exp_req_replay_queue);
1157         }
1158
1159         cfs_spin_unlock(&exp->exp_lock);
1160         return dup;
1161 }
1162
1163 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1164 {
1165         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1166         LASSERT(req->rq_export);
1167
1168         cfs_spin_lock(&req->rq_export->exp_lock);
1169         cfs_list_del_init(&req->rq_replay_list);
1170         cfs_spin_unlock(&req->rq_export->exp_lock);
1171 }
1172
1173 #ifdef __KERNEL__
1174 static void target_finish_recovery(struct obd_device *obd)
1175 {
1176         ENTRY;
1177         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1178                       obd->obd_name);
1179
1180         ldlm_reprocess_all_ns(obd->obd_namespace);
1181         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1182         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1183             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1184             !cfs_list_empty(&obd->obd_final_req_queue)) {
1185                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1186                        obd->obd_name,
1187                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1188                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1189                                "" : "lock ",
1190                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1191                                "" : "final ");
1192                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1193                 LBUG();
1194         }
1195         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1196
1197         obd->obd_recovery_end = cfs_time_current_sec();
1198
1199         /* when recovery finished, cleanup orphans on mds and ost */
1200         if (OBT(obd) && OBP(obd, postrecov)) {
1201                 int rc = OBP(obd, postrecov)(obd);
1202                 if (rc < 0)
1203                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1204                                       obd->obd_name, rc);
1205         }
1206         EXIT;
1207 }
1208
1209 static void abort_req_replay_queue(struct obd_device *obd)
1210 {
1211         struct ptlrpc_request *req, *n;
1212         cfs_list_t abort_list;
1213
1214         CFS_INIT_LIST_HEAD(&abort_list);
1215         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1216         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1217         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1218         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1219                 DEBUG_REQ(D_WARNING, req, "aborted:");
1220                 req->rq_status = -ENOTCONN;
1221                 if (ptlrpc_error(req)) {
1222                         DEBUG_REQ(D_ERROR, req,
1223                                   "failed abort_req_reply; skipping");
1224                 }
1225                 target_exp_dequeue_req_replay(req);
1226                 target_request_copy_put(req);
1227         }
1228 }
1229
1230 static void abort_lock_replay_queue(struct obd_device *obd)
1231 {
1232         struct ptlrpc_request *req, *n;
1233         cfs_list_t abort_list;
1234
1235         CFS_INIT_LIST_HEAD(&abort_list);
1236         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1237         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1238         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1239         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1240                 DEBUG_REQ(D_ERROR, req, "aborted:");
1241                 req->rq_status = -ENOTCONN;
1242                 if (ptlrpc_error(req)) {
1243                         DEBUG_REQ(D_ERROR, req,
1244                                   "failed abort_lock_reply; skipping");
1245                 }
1246                 target_request_copy_put(req);
1247         }
1248 }
1249 #endif
1250
1251 /* Called from a cleanup function if the device is being cleaned up
1252    forcefully.  The exports should all have been disconnected already,
1253    the only thing left to do is
1254      - clear the recovery flags
1255      - cancel the timer
1256      - free queued requests and replies, but don't send replies
1257    Because the obd_stopping flag is set, no new requests should be received.
1258
1259 */
1260 void target_cleanup_recovery(struct obd_device *obd)
1261 {
1262         struct ptlrpc_request *req, *n;
1263         cfs_list_t clean_list;
1264         ENTRY;
1265
1266         CFS_INIT_LIST_HEAD(&clean_list);
1267         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1268         if (!obd->obd_recovering) {
1269                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1270                 EXIT;
1271                 return;
1272         }
1273         obd->obd_recovering = obd->obd_abort_recovery = 0;
1274         target_cancel_recovery_timer(obd);
1275
1276         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1277         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1278
1279         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1280                 LASSERT(req->rq_reply_state == 0);
1281                 target_exp_dequeue_req_replay(req);
1282                 target_request_copy_put(req);
1283         }
1284
1285         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1286         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1287         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1288         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1289
1290         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1291                 LASSERT(req->rq_reply_state == 0);
1292                 target_request_copy_put(req);
1293         }
1294
1295         EXIT;
1296 }
1297
1298 /* obd_processing_task_lock should be held */
1299 void target_cancel_recovery_timer(struct obd_device *obd)
1300 {
1301         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1302         cfs_timer_disarm(&obd->obd_recovery_timer);
1303 }
1304
1305 /* extend = 1 means require at least "duration" seconds left in the timer,
1306    extend = 0 means set the total duration (start_recovery_timer) */
1307 static void reset_recovery_timer(struct obd_device *obd, int duration,
1308                                  int extend)
1309 {
1310         cfs_time_t now = cfs_time_current_sec();
1311         cfs_duration_t left;
1312
1313         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1314         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1315                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1316                 return;
1317         }
1318
1319         left = cfs_time_sub(obd->obd_recovery_end, now);
1320
1321         if (extend && (duration > left))
1322                 obd->obd_recovery_timeout += duration - left;
1323         else if (!extend && (duration > obd->obd_recovery_timeout))
1324                 /* Track the client's largest expected replay time */
1325                 obd->obd_recovery_timeout = duration;
1326 #ifdef CRAY_XT3
1327         /*
1328          * If total recovery time already exceed the
1329          * obd_recovery_max_time, then CRAY XT3 will
1330          * abort the recovery
1331          */
1332         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1333                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1334 #endif
1335         obd->obd_recovery_end = obd->obd_recovery_start +
1336                                 obd->obd_recovery_timeout;
1337         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1338             cfs_time_before(now, obd->obd_recovery_end)) {
1339                 left = cfs_time_sub(obd->obd_recovery_end, now);
1340                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1341         }
1342         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1343         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1344                obd->obd_name, (unsigned)left);
1345 }
1346
1347 static void check_and_start_recovery_timer(struct obd_device *obd)
1348 {
1349         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1350         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1351                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1352                 return;
1353         }
1354         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1355         obd->obd_recovery_start = cfs_time_current_sec();
1356         /* minimum */
1357         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1358         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1359
1360         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1361 }
1362
1363 /* Reset the timer with each new client connection */
1364 /*
1365  * This timer is actually reconnect_timer, which is for making sure
1366  * the total recovery window is at least as big as my reconnect
1367  * attempt timing. So the initial recovery time_out will be set to
1368  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1369  * from client is bigger than this, then the recovery time_out will
1370  * be extended to make sure the client could be reconnected, in the
1371  * process, the timeout from the new client should be ignored.
1372  */
1373
1374 static void
1375 target_start_and_reset_recovery_timer(struct obd_device *obd,
1376                                       struct ptlrpc_request *req,
1377                                       int new_client)
1378 {
1379         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1380
1381         if (!new_client && service_time)
1382                 /* Teach server about old server's estimates, as first guess
1383                  * at how long new requests will take. */
1384                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1385                        service_time);
1386
1387         check_and_start_recovery_timer(obd);
1388
1389         /* convert the service time to rpc timeout,
1390          * reuse service_time to limit stack usage */
1391         service_time = at_est2timeout(service_time);
1392
1393         /* We expect other clients to timeout within service_time, then try
1394          * to reconnect, then try the failover server.  The max delay between
1395          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1396         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1397                              INITIAL_CONNECT_TIMEOUT);
1398         if (service_time > obd->obd_recovery_timeout && !new_client)
1399                 reset_recovery_timer(obd, service_time, 0);
1400 }
1401
1402 #ifdef __KERNEL__
1403
1404 /** Health checking routines */
1405 static inline int exp_connect_healthy(struct obd_export *exp)
1406 {
1407         return (exp->exp_in_recovery);
1408 }
1409
1410 /** if export done req_replay or has replay in queue */
1411 static inline int exp_req_replay_healthy(struct obd_export *exp)
1412 {
1413         return (!exp->exp_req_replay_needed ||
1414                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1415 }
1416 /** if export done lock_replay or has replay in queue */
1417 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1418 {
1419         return (!exp->exp_lock_replay_needed ||
1420                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1421 }
1422
1423 static inline int exp_vbr_healthy(struct obd_export *exp)
1424 {
1425         return (!exp->exp_vbr_failed);
1426 }
1427
1428 static inline int exp_finished(struct obd_export *exp)
1429 {
1430         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1431 }
1432
1433 /** Checking routines for recovery */
1434 static int check_for_clients(struct obd_device *obd)
1435 {
1436         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1437                 return 1;
1438         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1439         if (obd->obd_no_conn == 0 &&
1440             obd->obd_connected_clients + obd->obd_stale_clients ==
1441             obd->obd_max_recoverable_clients)
1442                 return 1;
1443         return 0;
1444 }
1445
1446 static int check_for_next_transno(struct obd_device *obd)
1447 {
1448         struct ptlrpc_request *req = NULL;
1449         int wake_up = 0, connected, completed, queue_len;
1450         __u64 next_transno, req_transno;
1451         ENTRY;
1452         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1453
1454         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1455                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1456                                      struct ptlrpc_request, rq_list);
1457                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1458         } else {
1459                 req_transno = 0;
1460         }
1461
1462         connected = obd->obd_connected_clients;
1463         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1464         queue_len = obd->obd_requests_queued_for_recovery;
1465         next_transno = obd->obd_next_recovery_transno;
1466
1467         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1468                "req_transno: "LPU64", next_transno: "LPU64"\n",
1469                obd->obd_max_recoverable_clients, connected, completed,
1470                queue_len, req_transno, next_transno);
1471
1472         if (obd->obd_abort_recovery) {
1473                 CDEBUG(D_HA, "waking for aborted recovery\n");
1474                 wake_up = 1;
1475         } else if (obd->obd_recovery_expired) {
1476                 CDEBUG(D_HA, "waking for expired recovery\n");
1477                 wake_up = 1;
1478         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1479                 CDEBUG(D_HA, "waking for completed recovery\n");
1480                 wake_up = 1;
1481         } else if (req_transno == next_transno) {
1482                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1483                 wake_up = 1;
1484         } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1485                 int d_lvl = D_HA;
1486                 /** handle gaps occured due to lost reply or VBR */
1487                 LASSERTF(req_transno >= next_transno,
1488                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1489                          req_transno, next_transno);
1490                 if (req_transno > obd->obd_last_committed &&
1491                     !obd->obd_version_recov)
1492                         d_lvl = D_ERROR;
1493                 CDEBUG(d_lvl,
1494                        "%s: waking for gap in transno, VBR is %s (skip: "
1495                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1496                        ", last_committed: "LPD64")\n",
1497                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1498                        next_transno, queue_len, completed, connected,
1499                        req_transno, obd->obd_last_committed);
1500                 obd->obd_next_recovery_transno = req_transno;
1501                 wake_up = 1;
1502         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1503                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1504                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1505                 obd->obd_next_recovery_transno = req_transno;
1506                 wake_up = 1;
1507         }
1508         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1509         return wake_up;
1510 }
1511
1512 static int check_for_next_lock(struct obd_device *obd)
1513 {
1514         int wake_up = 0;
1515
1516         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1517         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1518                 CDEBUG(D_HA, "waking for next lock\n");
1519                 wake_up = 1;
1520         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1521                 CDEBUG(D_HA, "waking for completed lock replay\n");
1522                 wake_up = 1;
1523         } else if (obd->obd_abort_recovery) {
1524                 CDEBUG(D_HA, "waking for aborted recovery\n");
1525                 wake_up = 1;
1526         } else if (obd->obd_recovery_expired) {
1527                 CDEBUG(D_HA, "waking for expired recovery\n");
1528                 wake_up = 1;
1529         }
1530         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1531
1532         return wake_up;
1533 }
1534
1535 /**
1536  * wait for recovery events,
1537  * check its status with help of check_routine
1538  * evict dead clients via health_check
1539  */
1540 static int target_recovery_overseer(struct obd_device *obd,
1541                                     int (*check_routine)(struct obd_device *),
1542                                     int (*health_check)(struct obd_export *))
1543 {
1544         int abort = 0, expired = 1;
1545
1546         do {
1547                 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1548                 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1549                 abort = obd->obd_abort_recovery;
1550                 expired = obd->obd_recovery_expired;
1551                 obd->obd_recovery_expired = 0;
1552                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1553                 if (abort) {
1554                         CWARN("recovery is aborted, evict exports in recovery\n");
1555                         /** evict exports which didn't finish recovery yet */
1556                         class_disconnect_stale_exports(obd, exp_finished);
1557                 } else if (expired) {
1558                         /** If some clients died being recovered, evict them */
1559                         CDEBUG(D_WARNING, "recovery is timed out, evict stale exports\n");
1560                         /** evict cexports with no replay in queue, they are stalled */
1561                         class_disconnect_stale_exports(obd, health_check);
1562                         /** continue with VBR */
1563                         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1564                         obd->obd_version_recov = 1;
1565                         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1566                         /**
1567                          * reset timer, recovery will proceed with versions now,
1568                          * timeout is set just to handle reconnection delays
1569                          */
1570                         reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1571                         /** Wait for recovery events again, after evicting bad clients */
1572                 }
1573         } while (!abort && expired);
1574
1575         return abort;
1576 }
1577
1578 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1579 {
1580         struct ptlrpc_request *req = NULL;
1581         ENTRY;
1582
1583         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1584                obd->obd_next_recovery_transno);
1585
1586         if (target_recovery_overseer(obd, check_for_next_transno,
1587                                      exp_req_replay_healthy)) {
1588                 abort_req_replay_queue(obd);
1589                 abort_lock_replay_queue(obd);
1590         }
1591
1592         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1593         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1594                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1595                                      struct ptlrpc_request, rq_list);
1596                 cfs_list_del_init(&req->rq_list);
1597                 obd->obd_requests_queued_for_recovery--;
1598                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1599         } else {
1600                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1601                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1602                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1603                 /** evict exports failed VBR */
1604                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1605         }
1606         RETURN(req);
1607 }
1608
1609 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1610 {
1611         struct ptlrpc_request *req = NULL;
1612
1613         CDEBUG(D_HA, "Waiting for lock\n");
1614         if (target_recovery_overseer(obd, check_for_next_lock,
1615                                      exp_lock_replay_healthy))
1616                 abort_lock_replay_queue(obd);
1617
1618         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1619         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1620                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1621                                      struct ptlrpc_request, rq_list);
1622                 cfs_list_del_init(&req->rq_list);
1623                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1624         } else {
1625                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1626                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1627                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1628                 /** evict exports failed VBR */
1629                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1630         }
1631         return req;
1632 }
1633
1634 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1635 {
1636         struct ptlrpc_request *req = NULL;
1637
1638         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1639         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1640                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1641                                      struct ptlrpc_request, rq_list);
1642                 cfs_list_del_init(&req->rq_list);
1643                 if (req->rq_export->exp_in_recovery) {
1644                         cfs_spin_lock(&req->rq_export->exp_lock);
1645                         req->rq_export->exp_in_recovery = 0;
1646                         cfs_spin_unlock(&req->rq_export->exp_lock);
1647                 }
1648         }
1649         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1650         return req;
1651 }
1652
1653 static int handle_recovery_req(struct ptlrpc_thread *thread,
1654                                struct ptlrpc_request *req,
1655                                svc_handler_t handler)
1656 {
1657         int rc;
1658         ENTRY;
1659
1660         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1661         if (rc) {
1662                 CERROR("Failure to initialize session: %d\n", rc);
1663                 GOTO(reqcopy_put, rc);
1664         }
1665         /**
1666          * export can be evicted during recovery, no need to handle replays for
1667          * it after that, discard such request silently
1668          */
1669         if (req->rq_export->exp_disconnected)
1670                 GOTO(reqcopy_put, rc);
1671
1672         req->rq_recov_session.lc_thread = thread;
1673         lu_context_enter(&req->rq_recov_session);
1674         req->rq_svc_thread = thread;
1675         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1676
1677         /* thread context */
1678         lu_context_enter(&thread->t_env->le_ctx);
1679         (void)handler(req);
1680         lu_context_exit(&thread->t_env->le_ctx);
1681
1682         lu_context_exit(&req->rq_recov_session);
1683         lu_context_fini(&req->rq_recov_session);
1684         /* don't reset timer for final stage */
1685         if (!exp_finished(req->rq_export)) {
1686                 /**
1687                  * Add request timeout to the recovery time so next request from
1688                  * this client may come in recovery time
1689                  */
1690                  reset_recovery_timer(class_exp2obd(req->rq_export),
1691                                       AT_OFF ? obd_timeout :
1692                                       lustre_msg_get_timeout(req->rq_reqmsg), 1);
1693         }
1694 reqcopy_put:
1695         RETURN(rc);
1696 }
1697
1698 static int target_recovery_thread(void *arg)
1699 {
1700         struct lu_target *lut = arg;
1701         struct obd_device *obd = lut->lut_obd;
1702         struct ptlrpc_request *req;
1703         struct target_recovery_data *trd = &obd->obd_recovery_data;
1704         unsigned long delta;
1705         unsigned long flags;
1706         struct lu_env env;
1707         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1708         int rc = 0;
1709         ENTRY;
1710
1711         cfs_daemonize("tgt_recov");
1712
1713         SIGNAL_MASK_LOCK(current, flags);
1714         sigfillset(&current->blocked);
1715         RECALC_SIGPENDING;
1716         SIGNAL_MASK_UNLOCK(current, flags);
1717
1718         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1719         if (rc)
1720                 RETURN(rc);
1721
1722         thread->t_env = &env;
1723         env.le_ctx.lc_thread = thread;
1724
1725         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1726                cfs_curproc_pid());
1727         trd->trd_processing_task = cfs_curproc_pid();
1728
1729         obd->obd_recovering = 1;
1730         cfs_complete(&trd->trd_starting);
1731
1732         /* first of all, we have to know the first transno to replay */
1733         if (target_recovery_overseer(obd, check_for_clients,
1734                                      exp_connect_healthy)) {
1735                 abort_req_replay_queue(obd);
1736                 abort_lock_replay_queue(obd);
1737         }
1738
1739         /* next stage: replay requests */
1740         delta = jiffies;
1741         obd->obd_req_replaying = 1;
1742         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1743                cfs_atomic_read(&obd->obd_req_replay_clients),
1744                obd->obd_next_recovery_transno);
1745         while ((req = target_next_replay_req(obd))) {
1746                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1747                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1748                           lustre_msg_get_transno(req->rq_reqmsg),
1749                           libcfs_nid2str(req->rq_peer.nid));
1750                 handle_recovery_req(thread, req,
1751                                     trd->trd_recovery_handler);
1752                 /**
1753                  * bz18031: increase next_recovery_transno before
1754                  * target_request_copy_put() will drop exp_rpc reference
1755                  */
1756                 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1757                 obd->obd_next_recovery_transno++;
1758                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1759                 target_exp_dequeue_req_replay(req);
1760                 target_request_copy_put(req);
1761                 obd->obd_replayed_requests++;
1762         }
1763
1764         /**
1765          * The second stage: replay locks
1766          */
1767         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1768                cfs_atomic_read(&obd->obd_lock_replay_clients));
1769         while ((req = target_next_replay_lock(obd))) {
1770                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1771                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1772                           libcfs_nid2str(req->rq_peer.nid));
1773                 handle_recovery_req(thread, req,
1774                                     trd->trd_recovery_handler);
1775                 target_request_copy_put(req);
1776                 obd->obd_replayed_locks++;
1777         }
1778
1779         /**
1780          * The third stage: reply on final pings, at this moment all clients
1781          * must have request in final queue
1782          */
1783         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1784         /** Update server last boot epoch */
1785         lut_boot_epoch_update(lut);
1786         /* We drop recoverying flag to forward all new requests
1787          * to regular mds_handle() since now */
1788         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1789         obd->obd_recovering = obd->obd_abort_recovery = 0;
1790         target_cancel_recovery_timer(obd);
1791         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1792         while ((req = target_next_final_ping(obd))) {
1793                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1794                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1795                           libcfs_nid2str(req->rq_peer.nid));
1796                 handle_recovery_req(thread, req,
1797                                     trd->trd_recovery_handler);
1798                 target_request_copy_put(req);
1799         }
1800
1801         delta = (jiffies - delta) / CFS_HZ;
1802         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1803               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1804         if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1805                 CWARN("too long recovery - read logs\n");
1806                 libcfs_debug_dumplog();
1807         }
1808
1809         target_finish_recovery(obd);
1810
1811         lu_context_fini(&env.le_ctx);
1812         trd->trd_processing_task = 0;
1813         cfs_complete(&trd->trd_finishing);
1814         RETURN(rc);
1815 }
1816
1817 static int target_start_recovery_thread(struct lu_target *lut,
1818                                         svc_handler_t handler)
1819 {
1820         struct obd_device *obd = lut->lut_obd;
1821         int rc = 0;
1822         struct target_recovery_data *trd = &obd->obd_recovery_data;
1823
1824         memset(trd, 0, sizeof(*trd));
1825         cfs_init_completion(&trd->trd_starting);
1826         cfs_init_completion(&trd->trd_finishing);
1827         trd->trd_recovery_handler = handler;
1828
1829         if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
1830                 cfs_wait_for_completion(&trd->trd_starting);
1831                 LASSERT(obd->obd_recovering != 0);
1832         } else
1833                 rc = -ECHILD;
1834
1835         return rc;
1836 }
1837
1838 void target_stop_recovery_thread(struct obd_device *obd)
1839 {
1840         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1841         if (obd->obd_recovery_data.trd_processing_task > 0) {
1842                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1843                 /** recovery can be done but postrecovery is not yet */
1844                 if (obd->obd_recovering) {
1845                         CERROR("%s: Aborting recovery\n", obd->obd_name);
1846                         obd->obd_abort_recovery = 1;
1847                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1848                 }
1849                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1850                 cfs_wait_for_completion(&trd->trd_finishing);
1851         } else {
1852                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1853         }
1854 }
1855
1856 void target_recovery_fini(struct obd_device *obd)
1857 {
1858         class_disconnect_exports(obd);
1859         target_stop_recovery_thread(obd);
1860         target_cleanup_recovery(obd);
1861 }
1862 EXPORT_SYMBOL(target_recovery_fini);
1863
1864 static void target_recovery_expired(unsigned long castmeharder)
1865 {
1866         struct obd_device *obd = (struct obd_device *)castmeharder;
1867         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1868                " after %lds (%d clients connected)\n",
1869                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1870                cfs_time_current_sec()- obd->obd_recovery_start,
1871                obd->obd_connected_clients);
1872
1873         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1874         obd->obd_recovery_expired = 1;
1875         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1876         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1877 }
1878
1879 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1880 {
1881         struct obd_device *obd = lut->lut_obd;
1882         if (obd->obd_max_recoverable_clients == 0) {
1883                 /** Update server last boot epoch */
1884                 lut_boot_epoch_update(lut);
1885                 return;
1886         }
1887
1888         CWARN("RECOVERY: service %s, %d recoverable clients, "
1889               "last_transno "LPU64"\n", obd->obd_name,
1890               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1891         LASSERT(obd->obd_stopping == 0);
1892         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1893         obd->obd_recovery_start = 0;
1894         obd->obd_recovery_end = 0;
1895         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1896         /* bz13079: this should be set to desired value for ost but not for mds */
1897         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1898         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1899         target_start_recovery_thread(lut, handler);
1900 }
1901 EXPORT_SYMBOL(target_recovery_init);
1902
1903 #endif
1904
1905 static int target_process_req_flags(struct obd_device *obd,
1906                                     struct ptlrpc_request *req)
1907 {
1908         struct obd_export *exp = req->rq_export;
1909         LASSERT(exp != NULL);
1910         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1911                 /* client declares he's ready to replay locks */
1912                 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1913                 if (exp->exp_req_replay_needed) {
1914                         LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) >
1915                                 0);
1916                         cfs_spin_lock(&exp->exp_lock);
1917                         exp->exp_req_replay_needed = 0;
1918                         cfs_spin_unlock(&exp->exp_lock);
1919                         cfs_atomic_dec(&obd->obd_req_replay_clients);
1920                 }
1921                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1922         }
1923         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1924                 /* client declares he's ready to complete recovery
1925                  * so, we put the request on th final queue */
1926                 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1927                 if (exp->exp_lock_replay_needed) {
1928                         LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) >
1929                                 0);
1930                         cfs_spin_lock(&exp->exp_lock);
1931                         exp->exp_lock_replay_needed = 0;
1932                         cfs_spin_unlock(&exp->exp_lock);
1933                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
1934                 }
1935                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1936         }
1937
1938         return 0;
1939 }
1940
1941 int target_queue_recovery_request(struct ptlrpc_request *req,
1942                                   struct obd_device *obd)
1943 {
1944         cfs_list_t *tmp;
1945         int inserted = 0;
1946         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1947         ENTRY;
1948
1949         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1950                 /* Processing the queue right now, don't re-add. */
1951                 RETURN(1);
1952         }
1953
1954         target_process_req_flags(obd, req);
1955
1956         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1957                 /* client declares he's ready to complete recovery
1958                  * so, we put the request on th final queue */
1959                 target_request_copy_get(req);
1960                 DEBUG_REQ(D_HA, req, "queue final req");
1961                 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1962                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1963                 if (obd->obd_recovering) {
1964                         cfs_list_add_tail(&req->rq_list,
1965                                           &obd->obd_final_req_queue);
1966                 } else {
1967                         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1968                         target_request_copy_put(req);
1969                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1970                 }
1971                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1972                 RETURN(0);
1973         }
1974         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1975                 /* client declares he's ready to replay locks */
1976                 target_request_copy_get(req);
1977                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1978                 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1979                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1980                 LASSERT(obd->obd_recovering);
1981                 /* usually due to recovery abort */
1982                 if (!req->rq_export->exp_in_recovery) {
1983                         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1984                         target_request_copy_put(req);
1985                         RETURN(-ENOTCONN);
1986                 }
1987                 LASSERT(req->rq_export->exp_lock_replay_needed);
1988                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1989                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1990                 RETURN(0);
1991         }
1992
1993         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1994          * (i.e. buflens etc are in my own byte order), but type-dependent
1995          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1996
1997         if (!transno) {
1998                 CFS_INIT_LIST_HEAD(&req->rq_list);
1999                 DEBUG_REQ(D_HA, req, "not queueing");
2000                 RETURN(1);
2001         }
2002
2003         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2004
2005         /* If we're processing the queue, we want don't want to queue this
2006          * message.
2007          *
2008          * Also, if this request has a transno less than the one we're waiting
2009          * for, we should process it now.  It could (and currently always will)
2010          * be an open request for a descriptor that was opened some time ago.
2011          *
2012          * Also, a resent, replayed request that has already been
2013          * handled will pass through here and be processed immediately.
2014          */
2015         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2016               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2017         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2018                 /* Processing the queue right now, don't re-add. */
2019                 LASSERT(cfs_list_empty(&req->rq_list));
2020                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2021                 RETURN(1);
2022         }
2023         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2024
2025         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2026                 RETURN(0);
2027
2028         target_request_copy_get(req);
2029         cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2030         LASSERT(obd->obd_recovering);
2031         if (!req->rq_export->exp_in_recovery) {
2032                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2033                 target_request_copy_put(req);
2034                 RETURN(-ENOTCONN);
2035         }
2036         LASSERT(req->rq_export->exp_req_replay_needed);
2037
2038         if (target_exp_enqueue_req_replay(req)) {
2039                 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2040                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2041                 target_request_copy_put(req);
2042                 RETURN(0);
2043         }
2044
2045         /* XXX O(n^2) */
2046         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2047                 struct ptlrpc_request *reqiter =
2048                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2049
2050                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2051                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2052                         inserted = 1;
2053                         break;
2054                 }
2055
2056                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2057                              transno)) {
2058                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2059                                   "has been claimed by another client");
2060                         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2061                         target_exp_dequeue_req_replay(req);
2062                         target_request_copy_put(req);
2063                         RETURN(0);
2064                 }
2065         }
2066
2067         if (!inserted)
2068                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2069
2070         obd->obd_requests_queued_for_recovery++;
2071         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2072         cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2073         RETURN(0);
2074 }
2075
2076 /**
2077  * Packs current SLV and Limit into \a req.
2078  */
2079 int target_pack_pool_reply(struct ptlrpc_request *req)
2080 {
2081         struct obd_device *obd;
2082         ENTRY;
2083
2084         /*
2085          * Check that we still have all structures alive as this may
2086          * be some late rpc in shutdown time.
2087          */
2088         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2089                      !exp_connect_lru_resize(req->rq_export))) {
2090                 lustre_msg_set_slv(req->rq_repmsg, 0);
2091                 lustre_msg_set_limit(req->rq_repmsg, 0);
2092                 RETURN(0);
2093         }
2094
2095         /*
2096          * OBD is alive here as export is alive, which we checked above.
2097          */
2098         obd = req->rq_export->exp_obd;
2099
2100         cfs_read_lock(&obd->obd_pool_lock);
2101         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2102         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2103         cfs_read_unlock(&obd->obd_pool_lock);
2104
2105         RETURN(0);
2106 }
2107
2108 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2109 {
2110         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2111                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2112                 return (-ECOMM);
2113         }
2114
2115         if (unlikely(rc)) {
2116                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2117                 req->rq_status = rc;
2118                 return (ptlrpc_send_error(req, 1));
2119         } else {
2120                 DEBUG_REQ(D_NET, req, "sending reply");
2121         }
2122
2123         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2124 }
2125
2126 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2127 {
2128         int                        netrc;
2129         struct ptlrpc_reply_state *rs;
2130         struct obd_device         *obd;
2131         struct obd_export         *exp;
2132         struct ptlrpc_service     *svc;
2133         ENTRY;
2134
2135         if (req->rq_no_reply) {
2136                 EXIT;
2137                 return;
2138         }
2139
2140         svc = req->rq_rqbd->rqbd_service;
2141         rs = req->rq_reply_state;
2142         if (rs == NULL || !rs->rs_difficult) {
2143                 /* no notifiers */
2144                 target_send_reply_msg (req, rc, fail_id);
2145                 EXIT;
2146                 return;
2147         }
2148
2149         /* must be an export if locks saved */
2150         LASSERT (req->rq_export != NULL);
2151         /* req/reply consistent */
2152         LASSERT (rs->rs_service == svc);
2153
2154         /* "fresh" reply */
2155         LASSERT (!rs->rs_scheduled);
2156         LASSERT (!rs->rs_scheduled_ever);
2157         LASSERT (!rs->rs_handled);
2158         LASSERT (!rs->rs_on_net);
2159         LASSERT (rs->rs_export == NULL);
2160         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2161         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2162
2163         exp = class_export_get (req->rq_export);
2164         obd = exp->exp_obd;
2165
2166         /* disable reply scheduling while I'm setting up */
2167         rs->rs_scheduled = 1;
2168         rs->rs_on_net    = 1;
2169         rs->rs_xid       = req->rq_xid;
2170         rs->rs_transno   = req->rq_transno;
2171         rs->rs_export    = exp;
2172         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2173
2174         cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2175         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2176                rs->rs_transno, exp->exp_last_committed);
2177         if (rs->rs_transno > exp->exp_last_committed) {
2178                 /* not committed already */
2179                 cfs_list_add_tail(&rs->rs_obd_list,
2180                                   &exp->exp_uncommitted_replies);
2181         }
2182         cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2183
2184         cfs_spin_lock(&exp->exp_lock);
2185         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2186         cfs_spin_unlock(&exp->exp_lock);
2187
2188         netrc = target_send_reply_msg (req, rc, fail_id);
2189
2190         cfs_spin_lock(&svc->srv_lock);
2191
2192         cfs_atomic_inc(&svc->srv_n_difficult_replies);
2193
2194         if (netrc != 0) {
2195                 /* error sending: reply is off the net.  Also we need +1
2196                  * reply ref until ptlrpc_handle_rs() is done
2197                  * with the reply state (if the send was successful, there
2198                  * would have been +1 ref for the net, which
2199                  * reply_out_callback leaves alone) */
2200                 rs->rs_on_net = 0;
2201                 ptlrpc_rs_addref(rs);
2202                 cfs_atomic_inc (&svc->srv_outstanding_replies);
2203         }
2204
2205         cfs_spin_lock(&rs->rs_lock);
2206         if (rs->rs_transno <= exp->exp_last_committed ||
2207             (!rs->rs_on_net && !rs->rs_no_ack) ||
2208              cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2209              cfs_list_empty(&rs->rs_obd_list)) {
2210                 CDEBUG(D_HA, "Schedule reply immediately\n");
2211                 ptlrpc_dispatch_difficult_reply(rs);
2212         } else {
2213                 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2214                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2215         }
2216         cfs_spin_unlock(&rs->rs_lock);
2217         cfs_spin_unlock(&svc->srv_lock);
2218         EXIT;
2219 }
2220
2221 int target_handle_ping(struct ptlrpc_request *req)
2222 {
2223         obd_ping(req->rq_export);
2224         return req_capsule_server_pack(&req->rq_pill);
2225 }
2226
2227 void target_committed_to_req(struct ptlrpc_request *req)
2228 {
2229         struct obd_export *exp = req->rq_export;
2230
2231         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2232                 lustre_msg_set_last_committed(req->rq_repmsg,
2233                                               exp->exp_last_committed);
2234         else
2235                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2236                           "%d)", exp->exp_obd->obd_no_transno,
2237                           req->rq_repmsg == NULL);
2238
2239         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2240                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2241 }
2242 EXPORT_SYMBOL(target_committed_to_req);
2243
2244 int target_handle_qc_callback(struct ptlrpc_request *req)
2245 {
2246         struct obd_quotactl *oqctl;
2247         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2248
2249         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2250         if (oqctl == NULL) {
2251                 CERROR("Can't unpack obd_quotactl\n");
2252                 RETURN(-EPROTO);
2253         }
2254
2255         cli->cl_qchk_stat = oqctl->qc_stat;
2256
2257         return 0;
2258 }
2259
2260 #ifdef HAVE_QUOTA_SUPPORT
2261 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2262 {
2263 #ifdef __KERNEL__
2264         struct obd_device *obd = req->rq_export->exp_obd;
2265         struct obd_device *master_obd = NULL, *lov_obd = NULL;
2266         struct obd_device_target *obt;
2267         struct lustre_quota_ctxt *qctxt;
2268         struct qunit_data *qdata = NULL;
2269         int rc = 0;
2270         ENTRY;
2271
2272         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2273                 RETURN(rc);
2274
2275         rc = req_capsule_server_pack(&req->rq_pill);
2276         if (rc) {
2277                 CERROR("packing reply failed!: rc = %d\n", rc);
2278                 RETURN(rc);
2279         }
2280
2281         LASSERT(req->rq_export);
2282
2283         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2284         if (IS_ERR(qdata)) {
2285                 rc = PTR_ERR(qdata);
2286                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2287                 req->rq_status = rc;
2288                 GOTO(out, rc);
2289         }
2290
2291         /* we use the observer */
2292         if (obd_pin_observer(obd, &lov_obd) ||
2293             obd_pin_observer(lov_obd, &master_obd)) {
2294                 CERROR("Can't find the observer, it is recovering\n");
2295                 req->rq_status = -EAGAIN;
2296                 GOTO(out, rc);
2297         }
2298
2299         obt = &master_obd->u.obt;
2300         qctxt = &obt->obt_qctxt;
2301
2302         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2303                 /* quota_type has not been processed yet, return EAGAIN
2304                  * until we know whether or not quotas are supposed to
2305                  * be enabled */
2306                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2307                        "-EAGAIN\n");
2308                 req->rq_status = -EAGAIN;
2309                 GOTO(out, rc);
2310         }
2311
2312         cfs_down_read(&obt->obt_rwsem);
2313         if (qctxt->lqc_lqs_hash == NULL) {
2314                 cfs_up_read(&obt->obt_rwsem);
2315                 /* quota_type has not been processed yet, return EAGAIN
2316                  * until we know whether or not quotas are supposed to
2317                  * be enabled */
2318                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2319                        "-EAGAIN\n");
2320                 req->rq_status = -EAGAIN;
2321                 GOTO(out, rc);
2322         }
2323
2324         LASSERT(qctxt->lqc_handler);
2325         rc = qctxt->lqc_handler(master_obd, qdata,
2326                                 lustre_msg_get_opc(req->rq_reqmsg));
2327         cfs_up_read(&obt->obt_rwsem);
2328         if (rc && rc != -EDQUOT)
2329                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2330                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2331         req->rq_status = rc;
2332
2333         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2334         if (rc < 0) {
2335                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2336                 GOTO(out, rc);
2337         }
2338
2339         /* Block the quota req. b=14840 */
2340         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2341         EXIT;
2342
2343 out:
2344         if (master_obd)
2345                 obd_unpin_observer(lov_obd);
2346         if (lov_obd)
2347                 obd_unpin_observer(obd);
2348
2349         rc = ptlrpc_reply(req);
2350         return rc;
2351 #else
2352         return 0;
2353 #endif /* !__KERNEL__ */
2354 }
2355 #endif /* HAVE_QUOTA_SUPPORT */
2356
2357 /* Send a remote set_info_async.
2358  * This may go from client to server or server to client
2359  */
2360 int target_set_info_rpc(struct obd_import *imp, int opcode,
2361                         obd_count keylen, void *key,
2362                         obd_count vallen, void *val,
2363                         struct ptlrpc_request_set *set)
2364 {
2365         struct ptlrpc_request *req;
2366         char                  *tmp;
2367         int                    rc;
2368         ENTRY;
2369
2370         req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
2371         if (req == NULL)
2372                 RETURN(-ENOMEM);
2373
2374         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2375                              RCL_CLIENT, keylen);
2376         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2377                              RCL_CLIENT, vallen);
2378         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, opcode);
2379         if (rc) {
2380                 ptlrpc_request_free(req);
2381                 RETURN(rc);
2382         }
2383
2384         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2385         memcpy(tmp, key, keylen);
2386         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2387         memcpy(tmp, val, vallen);
2388
2389         ptlrpc_request_set_replen(req);
2390
2391         if (set) {
2392                 ptlrpc_set_add_req(set, req);
2393                 ptlrpc_check_set(NULL, set);
2394         } else {
2395                 rc = ptlrpc_queue_wait(req);
2396                 ptlrpc_req_finished(req);
2397         }
2398
2399         RETURN(rc);
2400 }
2401 EXPORT_SYMBOL(target_set_info_rpc);
2402
2403
2404 ldlm_mode_t lck_compat_array[] = {
2405         [LCK_EX] LCK_COMPAT_EX,
2406         [LCK_PW] LCK_COMPAT_PW,
2407         [LCK_PR] LCK_COMPAT_PR,
2408         [LCK_CW] LCK_COMPAT_CW,
2409         [LCK_CR] LCK_COMPAT_CR,
2410         [LCK_NL] LCK_COMPAT_NL,
2411         [LCK_GROUP] LCK_COMPAT_GROUP,
2412         [LCK_COS] LCK_COMPAT_COS,
2413 };
2414
2415 /**
2416  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2417  * not escape to the user level.
2418  */
2419 int ldlm_error2errno(ldlm_error_t error)
2420 {
2421         int result;
2422
2423         switch (error) {
2424         case ELDLM_OK:
2425                 result = 0;
2426                 break;
2427         case ELDLM_LOCK_CHANGED:
2428                 result = -ESTALE;
2429                 break;
2430         case ELDLM_LOCK_ABORTED:
2431                 result = -ENAVAIL;
2432                 break;
2433         case ELDLM_LOCK_REPLACED:
2434                 result = -ESRCH;
2435                 break;
2436         case ELDLM_NO_LOCK_DATA:
2437                 result = -ENOENT;
2438                 break;
2439         case ELDLM_NAMESPACE_EXISTS:
2440                 result = -EEXIST;
2441                 break;
2442         case ELDLM_BAD_NAMESPACE:
2443                 result = -EBADF;
2444                 break;
2445         default:
2446                 if (((int)error) < 0)  /* cast to signed type */
2447                         result = error; /* as ldlm_error_t can be unsigned */
2448                 else {
2449                         CERROR("Invalid DLM result code: %i\n", error);
2450                         result = -EPROTO;
2451                 }
2452         }
2453         return result;
2454 }
2455 EXPORT_SYMBOL(ldlm_error2errno);
2456
2457 /**
2458  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2459  */
2460 ldlm_error_t ldlm_errno2error(int err_no)
2461 {
2462         int error;
2463
2464         switch (err_no) {
2465         case 0:
2466                 error = ELDLM_OK;
2467                 break;
2468         case -ESTALE:
2469                 error = ELDLM_LOCK_CHANGED;
2470                 break;
2471         case -ENAVAIL:
2472                 error = ELDLM_LOCK_ABORTED;
2473                 break;
2474         case -ESRCH:
2475                 error = ELDLM_LOCK_REPLACED;
2476                 break;
2477         case -ENOENT:
2478                 error = ELDLM_NO_LOCK_DATA;
2479                 break;
2480         case -EEXIST:
2481                 error = ELDLM_NAMESPACE_EXISTS;
2482                 break;
2483         case -EBADF:
2484                 error = ELDLM_BAD_NAMESPACE;
2485                 break;
2486         default:
2487                 error = err_no;
2488         }
2489         return error;
2490 }
2491 EXPORT_SYMBOL(ldlm_errno2error);
2492
2493 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2494 void ldlm_dump_export_locks(struct obd_export *exp)
2495 {
2496         cfs_spin_lock(&exp->exp_locks_list_guard);
2497         if (!cfs_list_empty(&exp->exp_locks_list)) {
2498             struct ldlm_lock *lock;
2499
2500             CERROR("dumping locks for export %p,"
2501                    "ignore if the unmount doesn't hang\n", exp);
2502             cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2503                 ldlm_lock_dump(D_ERROR, lock, 0);
2504         }
2505         cfs_spin_unlock(&exp->exp_locks_list_guard);
2506 }
2507 #endif