Whamcloud - gitweb
move ldlm namespace creation in setup phase to avoid grab
[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         spin_lock(&imp->imp_lock);
84         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
86                         if (priority) {
87                                 list_del(&item->oic_item);
88                                 list_add(&item->oic_item, &imp->imp_conn_list);
89                                 item->oic_last_attempt = 0;
90                         }
91                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
92                                imp, imp->imp_obd->obd_name, uuid->uuid,
93                                (priority ? ", moved to head" : ""));
94                         spin_unlock(&imp->imp_lock);
95                         GOTO(out_free, rc = 0);
96                 }
97         }
98         /* not found */
99         if (create) {
100                 imp_conn->oic_conn = ptlrpc_conn;
101                 imp_conn->oic_uuid = *uuid;
102                 imp_conn->oic_last_attempt = 0;
103                 if (priority)
104                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
105                 else
106                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
107                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
108                        imp, imp->imp_obd->obd_name, uuid->uuid,
109                        (priority ? "head" : "tail"));
110         } else {
111                 spin_unlock(&imp->imp_lock);
112                 GOTO(out_free, rc = -ENOENT);
113         }
114
115         spin_unlock(&imp->imp_lock);
116         RETURN(0);
117 out_free:
118         if (imp_conn)
119                 OBD_FREE(imp_conn, sizeof(*imp_conn));
120 out_put:
121         ptlrpc_connection_put(ptlrpc_conn);
122         RETURN(rc);
123 }
124
125 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
126 {
127         return import_set_conn(imp, uuid, 1, 0);
128 }
129
130 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
131                            int priority)
132 {
133         return import_set_conn(imp, uuid, priority, 1);
134 }
135
136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
137 {
138         struct obd_import_conn *imp_conn;
139         struct obd_export *dlmexp;
140         int rc = -ENOENT;
141         ENTRY;
142
143         spin_lock(&imp->imp_lock);
144         if (list_empty(&imp->imp_conn_list)) {
145                 LASSERT(!imp->imp_connection);
146                 GOTO(out, rc);
147         }
148
149         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
150                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
151                         continue;
152                 LASSERT(imp_conn->oic_conn);
153
154                 /* is current conn? */
155                 if (imp_conn == imp->imp_conn_current) {
156                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
157
158                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
159                             imp->imp_state != LUSTRE_IMP_DISCON) {
160                                 CERROR("can't remove current connection\n");
161                                 GOTO(out, rc = -EBUSY);
162                         }
163
164                         ptlrpc_connection_put(imp->imp_connection);
165                         imp->imp_connection = NULL;
166
167                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
168                         if (dlmexp && dlmexp->exp_connection) {
169                                 LASSERT(dlmexp->exp_connection ==
170                                         imp_conn->oic_conn);
171                                 ptlrpc_connection_put(dlmexp->exp_connection);
172                                 dlmexp->exp_connection = NULL;
173                         }
174                 }
175
176                 list_del(&imp_conn->oic_item);
177                 ptlrpc_connection_put(imp_conn->oic_conn);
178                 OBD_FREE(imp_conn, sizeof(*imp_conn));
179                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
180                        imp, imp->imp_obd->obd_name, uuid->uuid);
181                 rc = 0;
182                 break;
183         }
184 out:
185         spin_unlock(&imp->imp_lock);
186         if (rc == -ENOENT)
187                 CERROR("connection %s not found\n", uuid->uuid);
188         RETURN(rc);
189 }
190
191 void client_destroy_import(struct obd_import *imp)
192 {
193         /* drop security policy instance after all rpc finished/aborted
194          * to let all busy contexts be released. */
195         class_import_get(imp);
196         class_destroy_import(imp);
197         sptlrpc_import_sec_put(imp);
198         class_import_put(imp);
199 }
200
201 /* configure an RPC client OBD device
202  *
203  * lcfg parameters:
204  * 1 - client UUID
205  * 2 - server UUID
206  * 3 - inactive-on-startup
207  */
208 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
209 {
210         struct client_obd *cli = &obddev->u.cli;
211         struct obd_import *imp;
212         struct obd_uuid server_uuid;
213         int rq_portal, rp_portal, connect_op;
214         char *name = obddev->obd_type->typ_name;
215         int rc;
216         ENTRY;
217
218         /* In a more perfect world, we would hang a ptlrpc_client off of
219          * obd_type and just use the values from there. */
220         if (!strcmp(name, LUSTRE_OSC_NAME)) {
221                 rq_portal = OST_REQUEST_PORTAL;
222                 rp_portal = OSC_REPLY_PORTAL;
223                 connect_op = OST_CONNECT;
224                 cli->cl_sp_me = LUSTRE_SP_CLI;
225                 cli->cl_sp_to = LUSTRE_SP_OST;
226         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
227                 rq_portal = MDS_REQUEST_PORTAL;
228                 rp_portal = MDC_REPLY_PORTAL;
229                 connect_op = MDS_CONNECT;
230                 cli->cl_sp_me = LUSTRE_SP_CLI;
231                 cli->cl_sp_to = LUSTRE_SP_MDT;
232         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
233                 rq_portal = MGS_REQUEST_PORTAL;
234                 rp_portal = MGC_REPLY_PORTAL;
235                 connect_op = MGS_CONNECT;
236                 cli->cl_sp_me = LUSTRE_SP_MGC;
237                 cli->cl_sp_to = LUSTRE_SP_MGS;
238                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
239         } else {
240                 CERROR("unknown client OBD type \"%s\", can't setup\n",
241                        name);
242                 RETURN(-EINVAL);
243         }
244
245         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
246                 CERROR("requires a TARGET UUID\n");
247                 RETURN(-EINVAL);
248         }
249
250         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
251                 CERROR("client UUID must be less than 38 characters\n");
252                 RETURN(-EINVAL);
253         }
254
255         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
256                 CERROR("setup requires a SERVER UUID\n");
257                 RETURN(-EINVAL);
258         }
259
260         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
261                 CERROR("target UUID must be less than 38 characters\n");
262                 RETURN(-EINVAL);
263         }
264
265         init_rwsem(&cli->cl_sem);
266         sema_init(&cli->cl_mgc_sem, 1);
267         cli->cl_conn_count = 0;
268         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
269                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
270                      sizeof(server_uuid)));
271
272         cli->cl_dirty = 0;
273         cli->cl_avail_grant = 0;
274         /* FIXME: should limit this for the sum of all cl_dirty_max */
275         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
276         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
277                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
278         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
279         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
280         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
281         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
282         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
283         client_obd_list_lock_init(&cli->cl_loi_list_lock);
284         cli->cl_r_in_flight = 0;
285         cli->cl_w_in_flight = 0;
286
287         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
288         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
289         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
290         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
291         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
292         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
293         cfs_waitq_init(&cli->cl_destroy_waitq);
294         atomic_set(&cli->cl_destroy_in_flight, 0);
295 #ifdef ENABLE_CHECKSUM
296         /* Turn on checksumming by default. */
297         cli->cl_checksum = 1;
298         /*
299          * The supported checksum types will be worked out at connect time
300          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
301          * through procfs.
302          */
303         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
304 #endif
305         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
306
307         /* This value may be changed at connect time in
308            ptlrpc_connect_interpret. */
309         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
310                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
311
312         if (!strcmp(name, LUSTRE_MDC_NAME)) {
313                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
314         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
315                 cli->cl_max_rpcs_in_flight = 2;
316         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
317                 cli->cl_max_rpcs_in_flight = 3;
318         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
319                 cli->cl_max_rpcs_in_flight = 4;
320         } else {
321                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
322         }
323
324         rc = ldlm_get_ref();
325         if (rc) {
326                 CERROR("ldlm_get_ref failed: %d\n", rc);
327                 GOTO(err, rc);
328         }
329
330         ptlrpc_init_client(rq_portal, rp_portal, name,
331                            &obddev->obd_ldlm_client);
332
333         imp = class_new_import(obddev);
334         if (imp == NULL)
335                 GOTO(err_ldlm, rc = -ENOENT);
336         imp->imp_client = &obddev->obd_ldlm_client;
337         imp->imp_connect_op = connect_op;
338         imp->imp_initial_recov = 1;
339         imp->imp_initial_recov_bk = 0;
340         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
341         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
342                LUSTRE_CFG_BUFLEN(lcfg, 1));
343         class_import_put(imp);
344
345         rc = client_import_add_conn(imp, &server_uuid, 1);
346         if (rc) {
347                 CERROR("can't add initial connection\n");
348                 GOTO(err_import, rc);
349         }
350
351         cli->cl_import = imp;
352         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
353         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
354         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
355
356         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
357                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
358                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
359                                name, obddev->obd_name,
360                                cli->cl_target_uuid.uuid);
361                         spin_lock(&imp->imp_lock);
362                         imp->imp_deactive = 1;
363                         spin_unlock(&imp->imp_lock);
364                 }
365         }
366
367         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
368                                                    LDLM_NAMESPACE_CLIENT,
369                                                    LDLM_NAMESPACE_GREEDY);
370         if (obddev->obd_namespace == NULL) {
371                 CERROR("Unable to create client namespace - %s\n",
372                        obddev->obd_name);
373                 GOTO(err_import, rc = -ENOMEM);
374         }
375
376         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
377
378         RETURN(rc);
379
380 err_import:
381         class_destroy_import(imp);
382 err_ldlm:
383         ldlm_put_ref();
384 err:
385         RETURN(rc);
386
387 }
388
389 int client_obd_cleanup(struct obd_device *obddev)
390 {
391         ENTRY;
392
393         ldlm_namespace_free_post(obddev->obd_namespace);
394         obddev->obd_namespace = NULL;
395
396         ldlm_put_ref();
397         RETURN(0);
398 }
399
400 /* ->o_connect() method for client side (OSC and MDC and MGC) */
401 int client_connect_import(const struct lu_env *env,
402                           struct obd_export **exp,
403                           struct obd_device *obd, struct obd_uuid *cluuid,
404                           struct obd_connect_data *data, void *localdata)
405 {
406         struct client_obd *cli = &obd->u.cli;
407         struct obd_import *imp = cli->cl_import;
408         struct obd_connect_data *ocd;
409         struct lustre_handle conn = { 0 };
410         int rc;
411         ENTRY;
412
413         *exp = NULL;
414         down_write(&cli->cl_sem);
415         if (cli->cl_conn_count > 0 )
416                 GOTO(out_sem, rc = -EALREADY);
417
418         rc = class_connect(&conn, obd, cluuid);
419         if (rc)
420                 GOTO(out_sem, rc);
421
422         cli->cl_conn_count++;
423         *exp = class_conn2export(&conn);
424
425         LASSERT(obd->obd_namespace);
426
427         imp->imp_dlm_handle = conn;
428         rc = ptlrpc_init_import(imp);
429         if (rc != 0)
430                 GOTO(out_ldlm, rc);
431
432         ocd = &imp->imp_connect_data;
433         if (data) {
434                 *ocd = *data;
435                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
436         }
437
438         rc = ptlrpc_connect_import(imp, NULL);
439         if (rc != 0) {
440                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
441                 GOTO(out_ldlm, rc);
442         }
443         LASSERT((*exp)->exp_connection);
444
445         if (data) {
446                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
447                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
448                          data->ocd_connect_flags, ocd->ocd_connect_flags);
449                 data->ocd_connect_flags = ocd->ocd_connect_flags;
450         }
451
452         ptlrpc_pinger_add_import(imp);
453
454         EXIT;
455
456         if (rc) {
457 out_ldlm:
458                 cli->cl_conn_count--;
459                 class_disconnect(*exp);
460                 *exp = NULL;
461         }
462 out_sem:
463         up_write(&cli->cl_sem);
464
465         return rc;
466 }
467
468 int client_disconnect_export(struct obd_export *exp)
469 {
470         struct obd_device *obd = class_exp2obd(exp);
471         struct client_obd *cli;
472         struct obd_import *imp;
473         int rc = 0, err;
474         ENTRY;
475
476         if (!obd) {
477                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
478                        exp, exp ? exp->exp_handle.h_cookie : -1);
479                 RETURN(-EINVAL);
480         }
481
482         cli = &obd->u.cli;
483         imp = cli->cl_import;
484
485         down_write(&cli->cl_sem);
486         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
487                cli->cl_conn_count);
488
489         if (!cli->cl_conn_count) {
490                 CERROR("disconnecting disconnected device (%s)\n",
491                        obd->obd_name);
492                 GOTO(out_disconnect, rc = -EINVAL);
493         }
494
495         cli->cl_conn_count--;
496         if (cli->cl_conn_count)
497                 GOTO(out_disconnect, rc = 0);
498
499         /* Mark import deactivated now, so we don't try to reconnect if any
500          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
501          * fully deactivate the import, or that would drop all requests. */
502         spin_lock(&imp->imp_lock);
503         imp->imp_deactive = 1;
504         spin_unlock(&imp->imp_lock);
505
506         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
507          * delete it regardless.  (It's safe to delete an import that was
508          * never added.) */
509         (void)ptlrpc_pinger_del_import(imp);
510
511         if (obd->obd_namespace != NULL) {
512                 /* obd_force == local only */
513                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
514                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
515                                        NULL);
516                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
517         }
518
519         /*
520          * there's no necessary to hold sem during diconnecting an import,
521          * and actually it may cause deadlock in gss.
522          */
523         up_write(&cli->cl_sem);
524         rc = ptlrpc_disconnect_import(imp, 0);
525         down_write(&cli->cl_sem);
526
527         ptlrpc_invalidate_import(imp);
528
529         if (imp->imp_rq_pool) {
530                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
531                 imp->imp_rq_pool = NULL;
532         }
533         client_destroy_import(imp);
534         cli->cl_import = NULL;
535
536         EXIT;
537
538  out_disconnect:
539         /* use server style - class_disconnect should be always called for
540          * o_disconnect */
541         err = class_disconnect(exp);
542         if (!rc && err)
543                 rc = err;
544
545         up_write(&cli->cl_sem);
546
547         RETURN(rc);
548 }
549
550 /* --------------------------------------------------------------------------
551  * from old lib/target.c
552  * -------------------------------------------------------------------------- */
553
554 static int target_handle_reconnect(struct lustre_handle *conn,
555                                    struct obd_export *exp,
556                                    struct obd_uuid *cluuid)
557 {
558         ENTRY;
559         if (exp->exp_connection && exp->exp_imp_reverse) {
560                 struct lustre_handle *hdl;
561                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
562                 /* Might be a re-connect after a partition. */
563                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
564                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
565                               cluuid->uuid);
566                         conn->cookie = exp->exp_handle.h_cookie;
567                         /* target_handle_connect() treats EALREADY and
568                          * -EALREADY differently.  EALREADY means we are
569                          * doing a valid reconnect from the same client. */
570                         RETURN(EALREADY);
571                 } else {
572                         CERROR("%s reconnecting from %s, "
573                                "handle mismatch (ours "LPX64", theirs "
574                                LPX64")\n", cluuid->uuid,
575                                exp->exp_connection->c_remote_uuid.uuid,
576                                hdl->cookie, conn->cookie);
577                         memset(conn, 0, sizeof *conn);
578                         /* target_handle_connect() treats EALREADY and
579                          * -EALREADY differently.  -EALREADY is an error
580                          * (same UUID, different handle). */
581                         RETURN(-EALREADY);
582                 }
583         }
584
585         conn->cookie = exp->exp_handle.h_cookie;
586         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
587                cluuid->uuid, exp, conn->cookie);
588         RETURN(0);
589 }
590
591 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
592                           int error)
593 {
594         struct obd_export *exp = cb_data;
595
596         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
597                obd->obd_name, exp->exp_client_uuid.uuid);
598
599         spin_lock(&exp->exp_lock);
600         exp->exp_need_sync = 0;
601         spin_unlock(&exp->exp_lock);
602 }
603 EXPORT_SYMBOL(target_client_add_cb);
604 static void
605 target_start_and_reset_recovery_timer(struct obd_device *obd,
606                                       struct ptlrpc_request *req,
607                                       int new_client);
608
609 int target_handle_connect(struct ptlrpc_request *req)
610 {
611         struct obd_device *target, *targref = NULL;
612         struct obd_export *export = NULL;
613         struct obd_import *revimp;
614         struct lustre_handle conn;
615         struct lustre_handle *tmp;
616         struct obd_uuid tgtuuid;
617         struct obd_uuid cluuid;
618         struct obd_uuid remote_uuid;
619         char *str;
620         int rc = 0;
621         int mds_conn = 0;
622         struct obd_connect_data *data, *tmpdata;
623         lnet_nid_t *client_nid = NULL;
624         ENTRY;
625
626         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
627
628         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
629         if (str == NULL) {
630                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
631                 GOTO(out, rc = -EINVAL);
632         }
633
634         obd_str2uuid(&tgtuuid, str);
635         target = class_uuid2obd(&tgtuuid);
636         if (!target)
637                 target = class_name2obd(str);
638
639         if (!target || target->obd_stopping || !target->obd_set_up) {
640                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
641                                    " for connect (%s)\n", str,
642                                    !target ? "no target" :
643                                    (target->obd_stopping ? "stopping" :
644                                    "not set up"));
645                 GOTO(out, rc = -ENODEV);
646         }
647
648         if (target->obd_no_conn) {
649                 LCONSOLE_WARN("%s: temporarily refusing client connection "
650                               "from %s\n", target->obd_name,
651                               libcfs_nid2str(req->rq_peer.nid));
652                 GOTO(out, rc = -EAGAIN);
653         }
654
655         /* Make sure the target isn't cleaned up while we're here. Yes,
656            there's still a race between the above check and our incref here.
657            Really, class_uuid2obd should take the ref. */
658         targref = class_incref(target, __FUNCTION__, cfs_current());
659
660
661         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
662         if (str == NULL) {
663                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
664                 GOTO(out, rc = -EINVAL);
665         }
666
667         obd_str2uuid(&cluuid, str);
668
669         /* XXX extract a nettype and format accordingly */
670         switch (sizeof(lnet_nid_t)) {
671                 /* NB the casts only avoid compiler warnings */
672         case 8:
673                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
674                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
675                 break;
676         case 4:
677                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
678                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
679                 break;
680         default:
681                 LBUG();
682         }
683
684         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
685         if (tmp == NULL)
686                 GOTO(out, rc = -EPROTO);
687
688         conn = *tmp;
689
690         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
691         if (!data)
692                 GOTO(out, rc = -EPROTO);
693
694         rc = req_capsule_server_pack(&req->rq_pill);
695         if (rc)
696                 GOTO(out, rc);
697
698         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
699                 if (!data) {
700                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
701                                   "libclient connection attempt");
702                         GOTO(out, rc = -EPROTO);
703                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
704                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
705                            data->ocd_version > LUSTRE_VERSION_CODE +
706                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
707                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
708                                   "libclient connection attempt",
709                                   data->ocd_version < LUSTRE_VERSION_CODE ?
710                                   "old" : "new",
711                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
712                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
713                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
714                                   OBD_OCD_VERSION_FIX(data->ocd_version));
715                         data = req_capsule_server_sized_get(&req->rq_pill,
716                                                         &RMF_CONNECT_DATA,
717                                     offsetof(typeof(*data), ocd_version) +
718                                              sizeof(data->ocd_version));
719                         if (data) {
720                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
721                                 data->ocd_version = LUSTRE_VERSION_CODE;
722                         }
723                         GOTO(out, rc = -EPROTO);
724                 }
725         }
726
727         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
728             (data->ocd_connect_flags & OBD_CONNECT_MDS))
729                 mds_conn = 1;
730
731         /* lctl gets a backstage, all-access pass. */
732         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
733                 goto dont_check_exports;
734
735         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
736         if (!export)
737                 goto no_export;
738
739         /* we've found an export in the hash */
740         if (export->exp_connecting) { /* bug 9635, et. al. */
741                 CWARN("%s: exp %p already connecting\n",
742                       export->exp_obd->obd_name, export);
743                 class_export_put(export);
744                 export = NULL;
745                 rc = -EALREADY;
746         } else if (mds_conn && export->exp_connection) {
747                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
748                         /* mds reconnected after failover */
749                         CWARN("%s: received MDS connection from NID %s,"
750                               " removing former export from NID %s\n",
751                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
752                             libcfs_nid2str(export->exp_connection->c_peer.nid));
753                 else
754                         /* new mds connection from the same nid */
755                         CWARN("%s: received new MDS connection from NID %s,"
756                               " removing former export from same NID\n",
757                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
758                 class_fail_export(export);
759                 class_export_put(export);
760                 export = NULL;
761                 rc = 0;
762         } else if (export->exp_connection != NULL &&
763                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
764                 /* in mds failover we have static uuid but nid can be
765                  * changed*/
766                 CWARN("%s: cookie %s seen on new NID %s when "
767                       "existing NID %s is already connected\n",
768                       target->obd_name, cluuid.uuid,
769                       libcfs_nid2str(req->rq_peer.nid),
770                       libcfs_nid2str(export->exp_connection->c_peer.nid));
771                 rc = -EALREADY;
772                 class_export_put(export);
773                 export = NULL;
774         } else {
775                 spin_lock(&export->exp_lock);
776                 export->exp_connecting = 1;
777                 spin_unlock(&export->exp_lock);
778                 class_export_put(export);
779                 LASSERT(export->exp_obd == target);
780
781                 rc = target_handle_reconnect(&conn, export, &cluuid);
782         }
783
784         /* If we found an export, we already unlocked. */
785         if (!export) {
786 no_export:
787                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
788         } else if (req->rq_export == NULL &&
789                    atomic_read(&export->exp_rpc_count) > 0) {
790                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
791                       target->obd_name, cluuid.uuid,
792                       libcfs_nid2str(req->rq_peer.nid),
793                       export, atomic_read(&export->exp_refcount));
794                 GOTO(out, rc = -EBUSY);
795         } else if (req->rq_export != NULL &&
796                    (atomic_read(&export->exp_rpc_count) > 1)) {
797                 /* the current connect rpc has increased exp_rpc_count */
798                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
799                       target->obd_name, cluuid.uuid,
800                       libcfs_nid2str(req->rq_peer.nid),
801                       export, atomic_read(&export->exp_rpc_count) - 1);
802                 spin_lock(&export->exp_lock);
803                 if (req->rq_export->exp_conn_cnt <
804                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
805                         /* try to abort active requests */
806                         req->rq_export->exp_abort_active_req = 1;
807                 spin_unlock(&export->exp_lock);
808                 GOTO(out, rc = -EBUSY);
809         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
810                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
811                        "cookies not random?\n", target->obd_name,
812                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
813                 GOTO(out, rc = -EALREADY);
814         } else {
815                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
816         }
817
818         if (rc < 0) {
819                 GOTO(out, rc);
820         }
821
822         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
823                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
824               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
825               export, (long)cfs_time_current_sec(),
826               export ? (long)export->exp_last_request_time : 0);
827
828         /* Tell the client if we're in recovery. */
829         if (target->obd_recovering) {
830                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
831                 /* If this is the first time a client connects,
832                    reset the recovery timer */
833                 if (rc == 0)
834                         target_start_and_reset_recovery_timer(target, req,
835                                                               !export);
836         }
837
838         /* We want to handle EALREADY but *not* -EALREADY from
839          * target_handle_reconnect(), return reconnection state in a flag */
840         if (rc == EALREADY) {
841                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
842                 rc = 0;
843         } else {
844                 LASSERT(rc == 0);
845         }
846
847         /* Tell the client if we support replayable requests */
848         if (target->obd_replayable)
849                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
850         client_nid = &req->rq_peer.nid;
851
852         if (export == NULL) {
853                 if (target->obd_recovering) {
854                         cfs_time_t t;
855
856                         t = cfs_timer_deadline(&target->obd_recovery_timer);
857                         t = cfs_time_sub(t, cfs_time_current());
858                         CERROR("%s: denying connection for new client %s (%s): "
859                                "%d clients in recovery for "CFS_TIME_T"s\n",
860                                target->obd_name,
861                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
862                                target->obd_recoverable_clients,
863                                cfs_duration_sec(t));
864                         rc = -EBUSY;
865                 } else {
866 dont_check_exports:
867                         rc = obd_connect(req->rq_svc_thread->t_env,
868                                          &export, target, &cluuid, data,
869                                          client_nid);
870                         if (rc == 0)
871                                 conn.cookie = export->exp_handle.h_cookie;
872                 }
873         } else {
874                 rc = obd_reconnect(req->rq_svc_thread->t_env,
875                                    export, target, &cluuid, data, client_nid);
876                 if (rc == 0)
877                         /* prevous done via class_conn2export */
878                         class_export_get(export);
879         }
880         if (rc)
881                 GOTO(out, rc);
882         /* Return only the parts of obd_connect_data that we understand, so the
883          * client knows that we don't understand the rest. */
884         if (data) {
885                  tmpdata = req_capsule_server_get(&req->rq_pill,
886                                                   &RMF_CONNECT_DATA);
887                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
888                  *tmpdata = *data;
889         }
890
891         /* If all else goes well, this is our RPC return code. */
892         req->rq_status = 0;
893
894         lustre_msg_set_handle(req->rq_repmsg, &conn);
895
896         /* If the client and the server are the same node, we will already
897          * have an export that really points to the client's DLM export,
898          * because we have a shared handles table.
899          *
900          * XXX this will go away when shaver stops sending the "connect" handle
901          * in the real "remote handle" field of the request --phik 24 Apr 2003
902          */
903         if (req->rq_export != NULL)
904                 class_export_put(req->rq_export);
905
906         req->rq_export = export;
907
908         spin_lock(&export->exp_lock);
909         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
910                 spin_unlock(&export->exp_lock);
911                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
912                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
913                        export->exp_conn_cnt,
914                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
915
916                 GOTO(out, rc = -EALREADY);
917         }
918         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
919         export->exp_abort_active_req = 0;
920
921         /* request from liblustre?  Don't evict it for not pinging. */
922         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
923                 export->exp_libclient = 1;
924                 spin_unlock(&export->exp_lock);
925
926                 spin_lock(&target->obd_dev_lock);
927                 list_del_init(&export->exp_obd_chain_timed);
928                 spin_unlock(&target->obd_dev_lock);
929         } else {
930                 spin_unlock(&export->exp_lock);
931         }
932
933         if (export->exp_connection != NULL)
934                 ptlrpc_connection_put(export->exp_connection);
935         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
936                                                        req->rq_self,
937                                                        &remote_uuid);
938         if (hlist_unhashed(&export->exp_nid_hash)) {
939                 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
940                                        &export->exp_connection->c_peer.nid,
941                                        &export->exp_nid_hash);
942         }
943
944         spin_lock_bh(&target->obd_processing_task_lock);
945         if (target->obd_recovering && !export->exp_in_recovery) {
946                 spin_lock(&export->exp_lock);
947                 export->exp_in_recovery = 1;
948                 export->exp_req_replay_needed = 1;
949                 export->exp_lock_replay_needed = 1;
950                 spin_unlock(&export->exp_lock);
951                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
952                      && (data->ocd_transno == 0))
953                         CWARN("Connect with zero transno!\n");
954
955                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
956                      && data->ocd_transno < target->obd_next_recovery_transno)
957                         target->obd_next_recovery_transno = data->ocd_transno;
958                 target->obd_connected_clients++;
959                 /* each connected client is counted as recoverable */
960                 target->obd_recoverable_clients++;
961                 atomic_inc(&target->obd_req_replay_clients);
962                 atomic_inc(&target->obd_lock_replay_clients);
963                 if (target->obd_connected_clients ==
964                     target->obd_max_recoverable_clients)
965                         cfs_waitq_signal(&target->obd_next_transno_waitq);
966         }
967         spin_unlock_bh(&target->obd_processing_task_lock);
968         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
969         conn = *tmp;
970
971         if (export->exp_imp_reverse != NULL) {
972                 /* destroyed import can be still referenced in ctxt */
973                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
974                                    KEY_REVIMP_UPD, 0, NULL, NULL);
975
976                 /* in some recovery senarios, previous ctx init rpc handled
977                  * in sptlrpc_target_export_check() might be used to install
978                  * a reverse ctx in this reverse import, and later OBD_CONNECT
979                  * using the same gss ctx could reach here and following new
980                  * reverse import. note all reverse ctx in new/old import are
981                  * actually based on the same gss ctx. so we invalidate ctx
982                  * here before destroy import, otherwise flush old import will
983                  * lead to remote reverse ctx be destroied, thus the reverse
984                  * ctx of new import will lost its peer.
985                  * there might be a better way to deal with this???
986                  */
987                 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
988
989                 client_destroy_import(export->exp_imp_reverse);
990         }
991
992         /* for the rest part, we return -ENOTCONN in case of errors
993          * in order to let client initialize connection again.
994          */
995         revimp = export->exp_imp_reverse = class_new_import(target);
996         if (!revimp) {
997                 CERROR("fail to alloc new reverse import.\n");
998                 GOTO(out, rc = -ENOTCONN);
999         }
1000
1001         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1002         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1003         revimp->imp_remote_handle = conn;
1004         revimp->imp_dlm_fake = 1;
1005         revimp->imp_state = LUSTRE_IMP_FULL;
1006
1007         /* unknown versions will be caught in
1008          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1009         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1010
1011         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1012             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1013                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1014         else
1015                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1016
1017         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1018         if (rc) {
1019                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1020                 export->exp_imp_reverse = NULL;
1021                 class_destroy_import(revimp);
1022         }
1023
1024         class_import_put(revimp);
1025 out:
1026         if (export) {
1027                 spin_lock(&export->exp_lock);
1028                 export->exp_connecting = 0;
1029                 spin_unlock(&export->exp_lock);
1030         }
1031         if (targref)
1032                 class_decref(targref, __FUNCTION__, cfs_current());
1033         if (rc)
1034                 req->rq_status = rc;
1035         RETURN(rc);
1036 }
1037
1038 int target_handle_disconnect(struct ptlrpc_request *req)
1039 {
1040         int rc;
1041         ENTRY;
1042
1043         rc = req_capsule_server_pack(&req->rq_pill);
1044         if (rc)
1045                 RETURN(rc);
1046
1047         /* keep the rq_export around so we can send the reply */
1048         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1049
1050         RETURN(0);
1051 }
1052
1053 void target_destroy_export(struct obd_export *exp)
1054 {
1055         /* exports created from last_rcvd data, and "fake"
1056            exports created by lctl don't have an import */
1057         if (exp->exp_imp_reverse != NULL)
1058                 client_destroy_import(exp->exp_imp_reverse);
1059
1060         /* We cancel locks at disconnect time, but this will catch any locks
1061          * granted in a race with recovery-induced disconnect. */
1062         if (exp->exp_obd->obd_namespace != NULL)
1063                 ldlm_cancel_locks_for_export(exp);
1064 }
1065
1066 /*
1067  * Recovery functions
1068  */
1069
1070 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1071 {
1072         struct ptlrpc_request *copy_req;
1073         struct lustre_msg *copy_reqmsg;
1074         struct ptlrpc_user_desc *udesc = NULL;
1075
1076         OBD_ALLOC_PTR(copy_req);
1077         if (!copy_req)
1078                 return NULL;
1079         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1080         if (!copy_reqmsg){
1081                 OBD_FREE_PTR(copy_req);
1082                 return NULL;
1083         }
1084
1085         if (orig_req->rq_user_desc) {
1086                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1087
1088                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1089                 if (!udesc) {
1090                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1091                         OBD_FREE_PTR(copy_req);
1092                         return NULL;
1093                 }
1094                 memcpy(udesc, orig_req->rq_user_desc,
1095                        sptlrpc_user_desc_size(ngroups));
1096         }
1097
1098         *copy_req = *orig_req;
1099         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1100         copy_req->rq_reqmsg = copy_reqmsg;
1101         copy_req->rq_user_desc = udesc;
1102
1103         class_export_rpc_get(copy_req->rq_export);
1104         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1105         CFS_INIT_LIST_HEAD(&copy_req->rq_replay_list);
1106         sptlrpc_svc_ctx_addref(copy_req);
1107
1108         if (copy_req->rq_reply_state) {
1109                 /* the copied req takes over the reply state */
1110                 orig_req->rq_reply_state = NULL;
1111                 /* to catch further access */
1112                 orig_req->rq_repmsg = NULL;
1113                 orig_req->rq_replen = 0;
1114         }
1115
1116         return copy_req;
1117 }
1118
1119 void ptlrpc_free_clone(struct ptlrpc_request *req)
1120 {
1121         LASSERT(list_empty(&req->rq_replay_list));
1122
1123         ptlrpc_req_drop_rs(req);
1124         sptlrpc_svc_ctx_decref(req);
1125         class_export_rpc_put(req->rq_export);
1126         list_del_init(&req->rq_list);
1127
1128         if (req->rq_user_desc) {
1129                 int ngroups = req->rq_user_desc->pud_ngroups;
1130                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1131         }
1132         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1133         OBD_FREE_PTR(req);
1134 }
1135
1136 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1137 {
1138         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1139         struct obd_export     *exp = req->rq_export;
1140         struct ptlrpc_request *reqiter;
1141         int                    dup = 0;
1142
1143         LASSERT(exp);
1144
1145         spin_lock(&exp->exp_lock);
1146         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1147                             rq_replay_list) {
1148                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1149                         dup = 1;
1150                         break;
1151                 }
1152         }
1153
1154         if (dup) {
1155                 /* we expect it with RESENT and REPLAY flags */
1156                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1157                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1158                         CERROR("invalid flags %x of resent replay\n",
1159                                lustre_msg_get_flags(req->rq_reqmsg));
1160         } else {
1161                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1162         }
1163
1164         spin_unlock(&exp->exp_lock);
1165         return dup;
1166 }
1167
1168 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1169 {
1170         LASSERT(!list_empty(&req->rq_replay_list));
1171         LASSERT(req->rq_export);
1172
1173         spin_lock(&req->rq_export->exp_lock);
1174         list_del_init(&req->rq_replay_list);
1175         spin_unlock(&req->rq_export->exp_lock);
1176 }
1177
1178 #ifdef __KERNEL__
1179 static void target_finish_recovery(struct obd_device *obd)
1180 {
1181         ENTRY;
1182         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1183                       obd->obd_name);
1184
1185         ldlm_reprocess_all_ns(obd->obd_namespace);
1186         spin_lock_bh(&obd->obd_processing_task_lock);
1187         if (list_empty(&obd->obd_req_replay_queue) &&
1188             list_empty(&obd->obd_lock_replay_queue) &&
1189             list_empty(&obd->obd_final_req_queue)) {
1190                 obd->obd_processing_task = 0;
1191         } else {
1192                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1193                        obd->obd_name,
1194                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1195                        list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1196                        list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1197                 spin_unlock_bh(&obd->obd_processing_task_lock);
1198                 LBUG();
1199         }
1200         spin_unlock_bh(&obd->obd_processing_task_lock);
1201
1202         /* when recovery finished, cleanup orphans on mds and ost */
1203         if (OBT(obd) && OBP(obd, postrecov)) {
1204                 int rc = OBP(obd, postrecov)(obd);
1205                 if (rc < 0)
1206                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1207                                       obd->obd_name, rc);
1208         }
1209
1210         obd->obd_recovery_end = cfs_time_current_sec();
1211         EXIT;
1212 }
1213
1214 static void abort_req_replay_queue(struct obd_device *obd)
1215 {
1216         struct ptlrpc_request *req, *n;
1217         struct list_head abort_list;
1218
1219         CFS_INIT_LIST_HEAD(&abort_list);
1220         spin_lock_bh(&obd->obd_processing_task_lock);
1221         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1222         spin_unlock_bh(&obd->obd_processing_task_lock);
1223         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1224                 DEBUG_REQ(D_WARNING, req, "aborted:");
1225                 req->rq_status = -ENOTCONN;
1226                 if (ptlrpc_error(req)) {
1227                         DEBUG_REQ(D_ERROR, req,
1228                                   "failed abort_req_reply; skipping");
1229                 }
1230                 target_exp_dequeue_req_replay(req);
1231                 ptlrpc_free_clone(req);
1232         }
1233 }
1234
1235 static void abort_lock_replay_queue(struct obd_device *obd)
1236 {
1237         struct ptlrpc_request *req, *n;
1238         struct list_head abort_list;
1239
1240         CFS_INIT_LIST_HEAD(&abort_list);
1241         spin_lock_bh(&obd->obd_processing_task_lock);
1242         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1243         spin_unlock_bh(&obd->obd_processing_task_lock);
1244         list_for_each_entry_safe(req, n, &abort_list, rq_list){
1245                 DEBUG_REQ(D_ERROR, req, "aborted:");
1246                 req->rq_status = -ENOTCONN;
1247                 if (ptlrpc_error(req)) {
1248                         DEBUG_REQ(D_ERROR, req,
1249                                   "failed abort_lock_reply; skipping");
1250                 }
1251                 ptlrpc_free_clone(req);
1252         }
1253 }
1254 #endif
1255
1256 /* Called from a cleanup function if the device is being cleaned up
1257    forcefully.  The exports should all have been disconnected already,
1258    the only thing left to do is
1259      - clear the recovery flags
1260      - cancel the timer
1261      - free queued requests and replies, but don't send replies
1262    Because the obd_stopping flag is set, no new requests should be received.
1263
1264 */
1265 void target_cleanup_recovery(struct obd_device *obd)
1266 {
1267         struct ptlrpc_request *req, *n;
1268         struct list_head clean_list;
1269         ENTRY;
1270
1271         LASSERT(obd->obd_stopping);
1272
1273         CFS_INIT_LIST_HEAD(&clean_list);
1274         spin_lock_bh(&obd->obd_processing_task_lock);
1275         if (!obd->obd_recovering) {
1276                 spin_unlock_bh(&obd->obd_processing_task_lock);
1277                 EXIT;
1278                 return;
1279         }
1280         obd->obd_recovering = obd->obd_abort_recovery = 0;
1281         target_cancel_recovery_timer(obd);
1282
1283         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1284         spin_unlock_bh(&obd->obd_processing_task_lock);
1285
1286         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1287                 LASSERT(req->rq_reply_state == 0);
1288                 target_exp_dequeue_req_replay(req);
1289                 ptlrpc_free_clone(req);
1290         }
1291
1292         spin_lock_bh(&obd->obd_processing_task_lock);
1293         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1294         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1295         spin_unlock_bh(&obd->obd_processing_task_lock);
1296
1297         list_for_each_entry_safe(req, n, &clean_list, rq_list){
1298                 LASSERT(req->rq_reply_state == 0);
1299                 ptlrpc_free_clone(req);
1300         }
1301
1302         EXIT;
1303 }
1304
1305 /* obd_processing_task_lock should be held */
1306 void target_cancel_recovery_timer(struct obd_device *obd)
1307 {
1308         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1309         cfs_timer_disarm(&obd->obd_recovery_timer);
1310 }
1311
1312 /* extend = 1 means require at least "duration" seconds left in the timer,
1313    extend = 0 means set the total duration (start_recovery_timer) */
1314 static void reset_recovery_timer(struct obd_device *obd, int duration,
1315                                  int extend)
1316 {
1317         cfs_time_t now = cfs_time_current_sec();
1318         cfs_duration_t left;
1319
1320         spin_lock_bh(&obd->obd_processing_task_lock);
1321         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1322                 spin_unlock_bh(&obd->obd_processing_task_lock);
1323                 return;
1324         }
1325
1326         left = cfs_time_sub(obd->obd_recovery_end, now);
1327
1328         if (extend && (duration > left))
1329                 obd->obd_recovery_timeout += duration - left;
1330         else if (!extend && (duration > obd->obd_recovery_timeout))
1331                 /* Track the client's largest expected replay time */
1332                 obd->obd_recovery_timeout = duration;
1333 #ifdef CRAY_XT3
1334         /*
1335          * If total recovery time already exceed the
1336          * obd_recovery_max_time, then CRAY XT3 will
1337          * abort the recovery
1338          */
1339         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1340                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1341 #endif
1342         obd->obd_recovery_end = obd->obd_recovery_start +
1343                                 obd->obd_recovery_timeout;
1344         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1345             cfs_time_before(now, obd->obd_recovery_end)) {
1346                 left = cfs_time_sub(obd->obd_recovery_end, now);
1347                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1348         }
1349         spin_unlock_bh(&obd->obd_processing_task_lock);
1350         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1351                obd->obd_name, (unsigned)left);
1352 }
1353
1354 static void check_and_start_recovery_timer(struct obd_device *obd)
1355 {
1356         spin_lock_bh(&obd->obd_processing_task_lock);
1357         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1358                 spin_unlock_bh(&obd->obd_processing_task_lock);
1359                 return;
1360         }
1361         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1362         obd->obd_recovery_start = cfs_time_current_sec();
1363         /* minimum */
1364         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1365         spin_unlock_bh(&obd->obd_processing_task_lock);
1366
1367         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1368 }
1369
1370 /* Reset the timer with each new client connection */
1371 /*
1372  * This timer is actually reconnect_timer, which is for making sure
1373  * the total recovery window is at least as big as my reconnect
1374  * attempt timing. So the initial recovery time_out will be set to
1375  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1376  * from client is bigger than this, then the recovery time_out will
1377  * be extend to make sure the client could be reconnected, in the
1378  * process, the timeout from the new client should be ignored.
1379  */
1380
1381 static void
1382 target_start_and_reset_recovery_timer(struct obd_device *obd,
1383                                       struct ptlrpc_request *req,
1384                                       int new_client)
1385 {
1386         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1387
1388         if (!new_client && service_time)
1389                 /* Teach server about old server's estimates, as first guess
1390                  * at how long new requests will take. */
1391                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1392                        service_time);
1393
1394         check_and_start_recovery_timer(obd);
1395
1396         /* convert the service time to rpc timeout,
1397          * reuse service_time to limit stack usage */
1398         service_time = at_est2timeout(service_time);
1399
1400         /* We expect other clients to timeout within service_time, then try
1401          * to reconnect, then try the failover server.  The max delay between
1402          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1403         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1404                              INITIAL_CONNECT_TIMEOUT);
1405         if (service_time > obd->obd_recovery_timeout && !new_client)
1406                 reset_recovery_timer(obd, service_time, 0);
1407 }
1408
1409 #ifdef __KERNEL__
1410 static int check_for_next_transno(struct obd_device *obd)
1411 {
1412         struct ptlrpc_request *req = NULL;
1413         int wake_up = 0, connected, completed, queue_len, max;
1414         __u64 next_transno, req_transno;
1415         ENTRY;
1416         spin_lock_bh(&obd->obd_processing_task_lock);
1417
1418         if (!list_empty(&obd->obd_req_replay_queue)) {
1419                 req = list_entry(obd->obd_req_replay_queue.next,
1420                                  struct ptlrpc_request, rq_list);
1421                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1422         } else {
1423                 req_transno = 0;
1424         }
1425
1426         max = obd->obd_max_recoverable_clients;
1427         connected = obd->obd_connected_clients;
1428         completed = connected - obd->obd_recoverable_clients;
1429         queue_len = obd->obd_requests_queued_for_recovery;
1430         next_transno = obd->obd_next_recovery_transno;
1431
1432         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1433                "req_transno: "LPU64", next_transno: "LPU64"\n",
1434                max, connected, completed, queue_len, req_transno, next_transno);
1435
1436         if (obd->obd_abort_recovery) {
1437                 CDEBUG(D_HA, "waking for aborted recovery\n");
1438                 wake_up = 1;
1439         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1440                 CDEBUG(D_HA, "waking for completed recovery\n");
1441                 wake_up = 1;
1442         } else if (req_transno == next_transno) {
1443                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1444                 wake_up = 1;
1445         } else if (queue_len + completed == max) {
1446                 /* handle gaps occured due to lost reply. It is allowed gaps
1447                  * because all clients are connected and there will be resend
1448                  * for missed transaction */
1449                 LASSERTF(req_transno >= next_transno,
1450                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1451                          req_transno, next_transno);
1452
1453                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1454                        "waking for skipped transno (skip: "LPD64
1455                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1456                        next_transno, queue_len, completed, connected, req_transno);
1457                 obd->obd_next_recovery_transno = req_transno;
1458                 wake_up = 1;
1459         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1460                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1461                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1462                 obd->obd_next_recovery_transno = req_transno;
1463                 wake_up = 1;
1464         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1465                 /* some clients haven't connected in time, but we can try
1466                  * to replay requests that demand on already committed ones
1467                  * also, we can replay first non-committed transation */
1468                 LASSERT(req_transno != 0);
1469                 if (obd->obd_version_recov ||
1470                     req_transno == obd->obd_last_committed + 1) {
1471                         obd->obd_next_recovery_transno = req_transno;
1472                 } else if (req_transno > obd->obd_last_committed) {
1473                         /* can't continue recovery: have no needed transno */
1474                         obd->obd_abort_recovery = 1;
1475                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1476                                "connected: %d, completed: %d, queue_len: %d, "
1477                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1478                                max, connected, completed, queue_len,
1479                                req_transno, next_transno);
1480                 }
1481                 wake_up = 1;
1482         }
1483
1484         spin_unlock_bh(&obd->obd_processing_task_lock);
1485         return wake_up;
1486 }
1487
1488 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1489 {
1490         struct l_wait_info lwi = { 0 };
1491         struct ptlrpc_request *req;
1492
1493         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1494                obd->obd_next_recovery_transno);
1495         l_wait_event(obd->obd_next_transno_waitq,
1496                      check_for_next_transno(obd), &lwi);
1497
1498         spin_lock_bh(&obd->obd_processing_task_lock);
1499         if (obd->obd_abort_recovery) {
1500                 req = NULL;
1501         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1502                 req = list_entry(obd->obd_req_replay_queue.next,
1503                                  struct ptlrpc_request, rq_list);
1504                 list_del_init(&req->rq_list);
1505                 obd->obd_requests_queued_for_recovery--;
1506         } else {
1507                 req = NULL;
1508         }
1509         spin_unlock_bh(&obd->obd_processing_task_lock);
1510         RETURN(req);
1511 }
1512
1513 static int check_for_next_lock(struct obd_device *obd)
1514 {
1515         struct ptlrpc_request *req = NULL;
1516         int wake_up = 0;
1517
1518         spin_lock_bh(&obd->obd_processing_task_lock);
1519         if (!list_empty(&obd->obd_lock_replay_queue)) {
1520                 req = list_entry(obd->obd_lock_replay_queue.next,
1521                                  struct ptlrpc_request, rq_list);
1522                 CDEBUG(D_HA, "waking for next lock\n");
1523                 wake_up = 1;
1524         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1525                 CDEBUG(D_HA, "waking for completed lock replay\n");
1526                 wake_up = 1;
1527         } else if (obd->obd_abort_recovery) {
1528                 CDEBUG(D_HA, "waking for aborted recovery\n");
1529                 wake_up = 1;
1530         }
1531         spin_unlock_bh(&obd->obd_processing_task_lock);
1532
1533         return wake_up;
1534 }
1535
1536 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1537 {
1538         struct l_wait_info lwi = { 0 };
1539         struct ptlrpc_request *req;
1540
1541         CDEBUG(D_HA, "Waiting for lock\n");
1542         l_wait_event(obd->obd_next_transno_waitq,
1543                      check_for_next_lock(obd), &lwi);
1544
1545         spin_lock_bh(&obd->obd_processing_task_lock);
1546         if (obd->obd_abort_recovery) {
1547                 req = NULL;
1548         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1549                 req = list_entry(obd->obd_lock_replay_queue.next,
1550                                  struct ptlrpc_request, rq_list);
1551                 list_del_init(&req->rq_list);
1552         } else {
1553                 req = NULL;
1554         }
1555         spin_unlock_bh(&obd->obd_processing_task_lock);
1556         return req;
1557 }
1558
1559 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1560 {
1561         struct ptlrpc_request *req;
1562
1563         spin_lock_bh(&obd->obd_processing_task_lock);
1564         if (!list_empty(&obd->obd_final_req_queue)) {
1565                 req = list_entry(obd->obd_final_req_queue.next,
1566                                  struct ptlrpc_request, rq_list);
1567                 list_del_init(&req->rq_list);
1568                 if (req->rq_export->exp_in_recovery) {
1569                         spin_lock(&req->rq_export->exp_lock);
1570                         req->rq_export->exp_in_recovery = 0;
1571                         spin_unlock(&req->rq_export->exp_lock);
1572                 }
1573         } else {
1574                 req = NULL;
1575         }
1576         spin_unlock_bh(&obd->obd_processing_task_lock);
1577         return req;
1578 }
1579
1580 static inline int req_vbr_done(struct obd_export *exp)
1581 {
1582         return (exp->exp_vbr_failed == 0);
1583 }
1584
1585 static inline int req_replay_done(struct obd_export *exp)
1586 {
1587         return (exp->exp_req_replay_needed == 0);
1588 }
1589
1590 static inline int lock_replay_done(struct obd_export *exp)
1591 {
1592         return (exp->exp_lock_replay_needed == 0);
1593 }
1594
1595 static inline int connect_done(struct obd_export *exp)
1596 {
1597         return (exp->exp_in_recovery != 0);
1598 }
1599
1600 static int check_for_clients(struct obd_device *obd)
1601 {
1602         if (obd->obd_abort_recovery || obd->obd_version_recov)
1603                 return 1;
1604         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1605         if (obd->obd_no_conn == 0 &&
1606             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1607                 return 1;
1608         return 0;
1609 }
1610
1611 static int handle_recovery_req(struct ptlrpc_thread *thread,
1612                                struct ptlrpc_request *req,
1613                                svc_handler_t handler)
1614 {
1615         int rc;
1616         ENTRY;
1617
1618         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1619         if (rc) {
1620                 CERROR("Failure to initialize session: %d\n", rc);
1621                 return rc;
1622         }
1623         req->rq_session.lc_thread = thread;
1624         lu_context_enter(&req->rq_session);
1625         req->rq_svc_thread = thread;
1626         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1627
1628         /* thread context */
1629         lu_context_enter(&thread->t_env->le_ctx);
1630         (void)handler(req);
1631         lu_context_exit(&thread->t_env->le_ctx);
1632
1633         lu_context_exit(&req->rq_session);
1634         lu_context_fini(&req->rq_session);
1635         /* don't reset timer for final stage */
1636         if (!req_replay_done(req->rq_export) ||
1637             !lock_replay_done(req->rq_export))
1638                 reset_recovery_timer(class_exp2obd(req->rq_export),
1639                                      AT_OFF ? obd_timeout :
1640                        at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1641
1642         /**
1643          * bz18031: increase next_recovery_transno before ptlrpc_free_clone()
1644          * will drop exp_rpc reference
1645          */
1646         if (!req_replay_done(req->rq_export)) {
1647                 spin_lock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1648                 req->rq_export->exp_obd->obd_next_recovery_transno++;
1649                 spin_unlock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1650                 target_exp_dequeue_req_replay(req);
1651         }
1652         ptlrpc_free_clone(req);
1653         RETURN(0);
1654 }
1655
1656 static int target_recovery_thread(void *arg)
1657 {
1658         struct lu_target *lut = arg;
1659         struct obd_device *obd = lut->lut_obd;
1660         struct ptlrpc_request *req;
1661         struct target_recovery_data *trd = &obd->obd_recovery_data;
1662         struct l_wait_info lwi = { 0 };
1663         unsigned long delta;
1664         unsigned long flags;
1665         struct lu_env env;
1666         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1667         int rc = 0;
1668         ENTRY;
1669
1670         cfs_daemonize("tgt_recov");
1671
1672         SIGNAL_MASK_LOCK(current, flags);
1673         sigfillset(&current->blocked);
1674         RECALC_SIGPENDING;
1675         SIGNAL_MASK_UNLOCK(current, flags);
1676
1677         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1678         if (rc)
1679                 RETURN(rc);
1680
1681         thread->t_env = &env;
1682         env.le_ctx.lc_thread = thread;
1683
1684         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1685                cfs_curproc_pid());
1686         trd->trd_processing_task = cfs_curproc_pid();
1687
1688         obd->obd_recovering = 1;
1689         complete(&trd->trd_starting);
1690
1691         /* first of all, we have to know the first transno to replay */
1692         obd->obd_abort_recovery = 0;
1693         l_wait_event(obd->obd_next_transno_waitq,
1694                      check_for_clients(obd), &lwi);
1695
1696         /* If some clients haven't connected in time, evict them */
1697         if (obd->obd_connected_clients < obd->obd_max_recoverable_clients) {
1698                 CWARN("Some clients haven't connect in time (%d/%d),"
1699                        "evict them\n", obd->obd_connected_clients,
1700                        obd->obd_max_recoverable_clients);
1701                 class_disconnect_stale_exports(obd, connect_done,
1702                                                exp_flags_from_obd(obd) |
1703                                                OBD_OPT_ABORT_RECOV);
1704                 /**
1705                  * if recovery proceeds with versions then some clients may be
1706                  * timed out waiting for others and trying to reconnect.
1707                  * Extend timer for such reconnect cases.
1708                  */
1709                 if (obd->obd_version_recov)
1710                         reset_recovery_timer(obd, RECONNECT_DELAY_MAX * 2, 1);
1711         }
1712
1713         /* next stage: replay requests */
1714         delta = jiffies;
1715         obd->obd_req_replaying = 1;
1716         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1717                atomic_read(&obd->obd_req_replay_clients),
1718                obd->obd_next_recovery_transno);
1719         while ((req = target_next_replay_req(obd))) {
1720                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1721                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1722                           lustre_msg_get_transno(req->rq_reqmsg),
1723                           libcfs_nid2str(req->rq_peer.nid));
1724                 handle_recovery_req(thread, req,
1725                                     trd->trd_recovery_handler);
1726                 obd->obd_replayed_requests++;
1727         }
1728
1729         /* If some clients haven't replayed requests in time, evict them */
1730         if (obd->obd_abort_recovery) {
1731                 CDEBUG(D_WARNING, "req replay is aborted\n");
1732                 class_disconnect_stale_exports(obd, req_replay_done,
1733                                                exp_flags_from_obd(obd) |
1734                                                OBD_OPT_ABORT_RECOV);
1735                 abort_req_replay_queue(obd);
1736         }
1737         LASSERT(list_empty(&obd->obd_req_replay_queue));
1738
1739         /* The second stage: replay locks */
1740         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1741                atomic_read(&obd->obd_lock_replay_clients));
1742         while ((req = target_next_replay_lock(obd))) {
1743                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1744                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1745                           libcfs_nid2str(req->rq_peer.nid));
1746                 handle_recovery_req(thread, req,
1747                                     trd->trd_recovery_handler);
1748                 obd->obd_replayed_locks++;
1749         }
1750
1751         /* If some clients haven't replayed requests in time, evict them */
1752         if (obd->obd_abort_recovery) {
1753                 CERROR("lock replay is aborted\n");
1754                 class_disconnect_stale_exports(obd, lock_replay_done,
1755                                                exp_flags_from_obd(obd) |
1756                                                OBD_OPT_ABORT_RECOV);
1757                 abort_lock_replay_queue(obd);
1758         }
1759         LASSERT(list_empty(&obd->obd_lock_replay_queue));
1760
1761         /* The third stage: reply on final pings */
1762         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1763         /** evict exports failed VBR */
1764         class_disconnect_stale_exports(obd, req_vbr_done,
1765                                        exp_flags_from_obd(obd) |
1766                                        OBD_OPT_ABORT_RECOV);
1767         /** Update server last boot epoch */
1768         lut_boot_epoch_update(lut);
1769         /* We drop recoverying flag to forward all new requests
1770          * to regular mds_handle() since now */
1771         spin_lock_bh(&obd->obd_processing_task_lock);
1772         obd->obd_recovering = obd->obd_abort_recovery = 0;
1773         target_cancel_recovery_timer(obd);
1774         spin_unlock_bh(&obd->obd_processing_task_lock);
1775         while ((req = target_next_final_ping(obd))) {
1776                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1777                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1778                           libcfs_nid2str(req->rq_peer.nid));
1779                 handle_recovery_req(thread, req,
1780                                     trd->trd_recovery_handler);
1781         }
1782
1783         delta = (jiffies - delta) / HZ;
1784         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1785               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1786         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1787         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1788         if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1789                 CWARN("too long recovery - read logs\n");
1790                 libcfs_debug_dumplog();
1791         }
1792
1793         target_finish_recovery(obd);
1794
1795         lu_context_fini(&env.le_ctx);
1796         trd->trd_processing_task = 0;
1797         complete(&trd->trd_finishing);
1798         RETURN(rc);
1799 }
1800
1801 static int target_start_recovery_thread(struct lu_target *lut,
1802                                         svc_handler_t handler)
1803 {
1804         struct obd_device *obd = lut->lut_obd;
1805         int rc = 0;
1806         struct target_recovery_data *trd = &obd->obd_recovery_data;
1807
1808         memset(trd, 0, sizeof(*trd));
1809         init_completion(&trd->trd_starting);
1810         init_completion(&trd->trd_finishing);
1811         trd->trd_recovery_handler = handler;
1812
1813         if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1814                 wait_for_completion(&trd->trd_starting);
1815                 LASSERT(obd->obd_recovering != 0);
1816         } else
1817                 rc = -ECHILD;
1818
1819         return rc;
1820 }
1821
1822 void target_stop_recovery_thread(struct obd_device *obd)
1823 {
1824         spin_lock_bh(&obd->obd_processing_task_lock);
1825         if (obd->obd_recovery_data.trd_processing_task > 0) {
1826                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1827                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1828                 obd->obd_abort_recovery = 1;
1829                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1830                 spin_unlock_bh(&obd->obd_processing_task_lock);
1831                 wait_for_completion(&trd->trd_finishing);
1832         } else {
1833                 spin_unlock_bh(&obd->obd_processing_task_lock);
1834         }
1835 }
1836
1837 void target_recovery_fini(struct obd_device *obd)
1838 {
1839         class_disconnect_exports(obd);
1840         target_stop_recovery_thread(obd);
1841         target_cleanup_recovery(obd);
1842 }
1843 EXPORT_SYMBOL(target_recovery_fini);
1844
1845 static void target_recovery_expired(unsigned long castmeharder)
1846 {
1847         struct obd_device *obd = (struct obd_device *)castmeharder;
1848         CDEBUG(D_HA, "%s: recovery timed out; %d clients never reconnected "
1849                "after %lds (%d clients did)\n",
1850                obd->obd_name, obd->obd_recoverable_clients,
1851                cfs_time_current_sec()- obd->obd_recovery_start,
1852                obd->obd_connected_clients);
1853
1854         spin_lock_bh(&obd->obd_processing_task_lock);
1855         obd->obd_version_recov = 1;
1856         CDEBUG(D_INFO, "VBR is used for %d clients from t"LPU64"\n",
1857                atomic_read(&obd->obd_req_replay_clients),
1858                obd->obd_next_recovery_transno);
1859         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1860         spin_unlock_bh(&obd->obd_processing_task_lock);
1861 }
1862
1863 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1864 {
1865         struct obd_device *obd = lut->lut_obd;
1866         if (obd->obd_max_recoverable_clients == 0) {
1867                 /** Update server last boot epoch */
1868                 lut_boot_epoch_update(lut);
1869                 return;
1870         }
1871
1872         CWARN("RECOVERY: service %s, %d recoverable clients, "
1873               "last_transno "LPU64"\n", obd->obd_name,
1874               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1875         LASSERT(obd->obd_stopping == 0);
1876         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1877         obd->obd_recovery_start = 0;
1878         obd->obd_recovery_end = 0;
1879         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1880         /* bz13079: this should be set to desired value for ost but not for mds */
1881         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1882         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1883         target_start_recovery_thread(lut, handler);
1884 }
1885 EXPORT_SYMBOL(target_recovery_init);
1886
1887 #endif
1888
1889 static int target_process_req_flags(struct obd_device *obd,
1890                                     struct ptlrpc_request *req)
1891 {
1892         struct obd_export *exp = req->rq_export;
1893         LASSERT(exp != NULL);
1894         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1895                 /* client declares he's ready to replay locks */
1896                 spin_lock_bh(&obd->obd_processing_task_lock);
1897                 if (exp->exp_req_replay_needed) {
1898                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1899                         spin_lock(&exp->exp_lock);
1900                         exp->exp_req_replay_needed = 0;
1901                         spin_unlock(&exp->exp_lock);
1902                         atomic_dec(&obd->obd_req_replay_clients);
1903                         LASSERT(obd->obd_recoverable_clients > 0);
1904                         obd->obd_recoverable_clients--;
1905                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1906                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1907                 }
1908                 spin_unlock_bh(&obd->obd_processing_task_lock);
1909         }
1910         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1911                 /* client declares he's ready to complete recovery
1912                  * so, we put the request on th final queue */
1913                 spin_lock_bh(&obd->obd_processing_task_lock);
1914                 if (exp->exp_lock_replay_needed) {
1915                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1916                         spin_lock(&exp->exp_lock);
1917                         exp->exp_lock_replay_needed = 0;
1918                         spin_unlock(&exp->exp_lock);
1919                         atomic_dec(&obd->obd_lock_replay_clients);
1920                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1921                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1922                 }
1923                 spin_unlock_bh(&obd->obd_processing_task_lock);
1924         }
1925
1926         return 0;
1927 }
1928
1929 int target_queue_recovery_request(struct ptlrpc_request *req,
1930                                   struct obd_device *obd)
1931 {
1932         struct list_head *tmp;
1933         int inserted = 0;
1934         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1935         ENTRY;
1936
1937         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1938                 /* Processing the queue right now, don't re-add. */
1939                 RETURN(1);
1940         }
1941
1942         target_process_req_flags(obd, req);
1943
1944         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1945                 /* client declares he's ready to complete recovery
1946                  * so, we put the request on th final queue */
1947                 req = ptlrpc_clone_req(req);
1948                 if (req == NULL)
1949                         RETURN(-ENOMEM);
1950                 DEBUG_REQ(D_HA, req, "queue final req");
1951                 spin_lock_bh(&obd->obd_processing_task_lock);
1952                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1953                 if (obd->obd_recovering)
1954                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1955                 else {
1956                         spin_unlock_bh(&obd->obd_processing_task_lock);
1957                         ptlrpc_free_clone(req);
1958                         if (obd->obd_stopping) {
1959                                 RETURN(-ENOTCONN);
1960                         } else {
1961                                 RETURN(1);
1962                         }
1963                 }
1964                 spin_unlock_bh(&obd->obd_processing_task_lock);
1965                 RETURN(0);
1966         }
1967         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1968                 /* client declares he's ready to replay locks */
1969                 req = ptlrpc_clone_req(req);
1970                 if (req == NULL)
1971                         RETURN(-ENOMEM);
1972                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1973                 spin_lock_bh(&obd->obd_processing_task_lock);
1974                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1975                 LASSERT(obd->obd_recovering);
1976                 /* usually due to recovery abort */
1977                 if (!req->rq_export->exp_in_recovery) {
1978                         spin_unlock_bh(&obd->obd_processing_task_lock);
1979                         ptlrpc_free_clone(req);
1980                         RETURN(-ENOTCONN);
1981                 }
1982                 LASSERT(req->rq_export->exp_lock_replay_needed);
1983                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1984                 spin_unlock_bh(&obd->obd_processing_task_lock);
1985                 RETURN(0);
1986         }
1987
1988         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1989          * (i.e. buflens etc are in my own byte order), but type-dependent
1990          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1991
1992         if (!transno) {
1993                 CFS_INIT_LIST_HEAD(&req->rq_list);
1994                 DEBUG_REQ(D_HA, req, "not queueing");
1995                 RETURN(1);
1996         }
1997
1998         spin_lock_bh(&obd->obd_processing_task_lock);
1999
2000         /* If we're processing the queue, we want don't want to queue this
2001          * message.
2002          *
2003          * Also, if this request has a transno less than the one we're waiting
2004          * for, we should process it now.  It could (and currently always will)
2005          * be an open request for a descriptor that was opened some time ago.
2006          *
2007          * Also, a resent, replayed request that has already been
2008          * handled will pass through here and be processed immediately.
2009          */
2010         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2011               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2012         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2013                 /* Processing the queue right now, don't re-add. */
2014                 LASSERT(list_empty(&req->rq_list));
2015                 spin_unlock_bh(&obd->obd_processing_task_lock);
2016                 RETURN(1);
2017         }
2018         spin_unlock_bh(&obd->obd_processing_task_lock);
2019
2020         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2021                 RETURN(0);
2022
2023         req = ptlrpc_clone_req(req);
2024         if (req == NULL)
2025                 RETURN(-ENOMEM);
2026
2027         spin_lock_bh(&obd->obd_processing_task_lock);
2028         LASSERT(obd->obd_recovering);
2029         if (!req->rq_export->exp_in_recovery) {
2030                 spin_unlock_bh(&obd->obd_processing_task_lock);
2031                 ptlrpc_free_clone(req);
2032                 RETURN(-ENOTCONN);
2033         }
2034         LASSERT(req->rq_export->exp_req_replay_needed);
2035
2036         if (target_exp_enqueue_req_replay(req)) {
2037                 spin_unlock_bh(&obd->obd_processing_task_lock);
2038                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2039                 ptlrpc_free_clone(req);
2040                 RETURN(0);
2041         }
2042
2043         /* XXX O(n^2) */
2044         list_for_each(tmp, &obd->obd_req_replay_queue) {
2045                 struct ptlrpc_request *reqiter =
2046                         list_entry(tmp, struct ptlrpc_request, rq_list);
2047
2048                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2049                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2050                         inserted = 1;
2051                         break;
2052                 }
2053
2054                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2055                              transno)) {
2056                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2057                                   "has been claimed by another client");
2058                         spin_unlock_bh(&obd->obd_processing_task_lock);
2059                         target_exp_dequeue_req_replay(req);
2060                         ptlrpc_free_clone(req);
2061                         RETURN(0);
2062                 }
2063         }
2064
2065         if (!inserted)
2066                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2067
2068         obd->obd_requests_queued_for_recovery++;
2069         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2070         spin_unlock_bh(&obd->obd_processing_task_lock);
2071         RETURN(0);
2072 }
2073
2074 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2075 {
2076         return req->rq_export->exp_obd;
2077 }
2078
2079 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2080 {
2081         LASSERT(exp != NULL);
2082         return &exp->exp_obd->obd_namespace->ns_pool;
2083 }
2084
2085 /**
2086  * Packs current SLV and Limit into \a req.
2087  */
2088 int target_pack_pool_reply(struct ptlrpc_request *req)
2089 {
2090         struct obd_device *obd;
2091         ENTRY;
2092
2093         /*
2094          * Check that we still have all structures alive as this may
2095          * be some late rpc in shutdown time.
2096          */
2097         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2098                      !exp_connect_lru_resize(req->rq_export))) {
2099                 lustre_msg_set_slv(req->rq_repmsg, 0);
2100                 lustre_msg_set_limit(req->rq_repmsg, 0);
2101                 RETURN(0);
2102         }
2103
2104         /*
2105          * OBD is alive here as export is alive, which we checked above.
2106          */
2107         obd = req->rq_export->exp_obd;
2108
2109         read_lock(&obd->obd_pool_lock);
2110         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2111         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2112         read_unlock(&obd->obd_pool_lock);
2113
2114         RETURN(0);
2115 }
2116
2117 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2118 {
2119         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2120                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2121                 return (-ECOMM);
2122         }
2123
2124         if (unlikely(rc)) {
2125                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2126                 req->rq_status = rc;
2127                 return (ptlrpc_send_error(req, 1));
2128         } else {
2129                 DEBUG_REQ(D_NET, req, "sending reply");
2130         }
2131
2132         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2133 }
2134
2135 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2136 {
2137         int                        netrc;
2138         struct ptlrpc_reply_state *rs;
2139         struct obd_device         *obd;
2140         struct obd_export         *exp;
2141         struct ptlrpc_service     *svc;
2142         ENTRY;
2143
2144         if (req->rq_no_reply) {
2145                 EXIT;
2146                 return;
2147         }
2148
2149         svc = req->rq_rqbd->rqbd_service;
2150         rs = req->rq_reply_state;
2151         if (rs == NULL || !rs->rs_difficult) {
2152                 /* no notifiers */
2153                 target_send_reply_msg (req, rc, fail_id);
2154                 EXIT;
2155                 return;
2156         }
2157
2158         /* must be an export if locks saved */
2159         LASSERT (req->rq_export != NULL);
2160         /* req/reply consistent */
2161         LASSERT (rs->rs_service == svc);
2162
2163         /* "fresh" reply */
2164         LASSERT (!rs->rs_scheduled);
2165         LASSERT (!rs->rs_scheduled_ever);
2166         LASSERT (!rs->rs_handled);
2167         LASSERT (!rs->rs_on_net);
2168         LASSERT (rs->rs_export == NULL);
2169         LASSERT (list_empty(&rs->rs_obd_list));
2170         LASSERT (list_empty(&rs->rs_exp_list));
2171
2172         exp = class_export_get (req->rq_export);
2173         obd = exp->exp_obd;
2174
2175         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2176         rs->rs_scheduled = 1;
2177         rs->rs_on_net    = 1;
2178         rs->rs_xid       = req->rq_xid;
2179         rs->rs_transno   = req->rq_transno;
2180         rs->rs_export    = exp;
2181         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2182
2183         spin_lock(&exp->exp_uncommitted_replies_lock);
2184         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2185                rs->rs_transno, exp->exp_last_committed);
2186         if (rs->rs_transno > exp->exp_last_committed) {
2187                 /* not committed already */
2188                 list_add_tail(&rs->rs_obd_list,
2189                               &exp->exp_uncommitted_replies);
2190         }
2191         spin_unlock (&exp->exp_uncommitted_replies_lock);
2192
2193         spin_lock(&exp->exp_lock);
2194         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2195         spin_unlock(&exp->exp_lock);
2196
2197         netrc = target_send_reply_msg (req, rc, fail_id);
2198
2199         spin_lock(&svc->srv_lock);
2200
2201         atomic_inc(&svc->srv_n_difficult_replies);
2202
2203         if (netrc != 0) {
2204                 /* error sending: reply is off the net.  Also we need +1
2205                  * reply ref until ptlrpc_server_handle_reply() is done
2206                  * with the reply state (if the send was successful, there
2207                  * would have been +1 ref for the net, which
2208                  * reply_out_callback leaves alone) */
2209                 rs->rs_on_net = 0;
2210                 ptlrpc_rs_addref(rs);
2211                 atomic_inc (&svc->srv_outstanding_replies);
2212         }
2213
2214         spin_lock(&rs->rs_lock);
2215         if (rs->rs_transno <= exp->exp_last_committed ||
2216             (!rs->rs_on_net && !rs->rs_no_ack) ||
2217              list_empty(&rs->rs_exp_list) ||     /* completed already */
2218              list_empty(&rs->rs_obd_list)) {
2219                 CDEBUG(D_HA, "Schedule reply immediately\n");
2220                 ptlrpc_dispatch_difficult_reply(rs);
2221         } else {
2222                 list_add (&rs->rs_list, &svc->srv_active_replies);
2223                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2224         }
2225         spin_unlock(&rs->rs_lock);
2226         spin_unlock(&svc->srv_lock);
2227         EXIT;
2228 }
2229
2230 int target_handle_ping(struct ptlrpc_request *req)
2231 {
2232         obd_ping(req->rq_export);
2233         return req_capsule_server_pack(&req->rq_pill);
2234 }
2235
2236 void target_committed_to_req(struct ptlrpc_request *req)
2237 {
2238         struct obd_export *exp = req->rq_export;
2239
2240         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2241                 lustre_msg_set_last_committed(req->rq_repmsg,
2242                                               exp->exp_last_committed);
2243         else
2244                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2245                           "%d)", exp->exp_obd->obd_no_transno,
2246                           req->rq_repmsg == NULL);
2247
2248         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2249                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2250 }
2251 EXPORT_SYMBOL(target_committed_to_req);
2252
2253 int target_handle_qc_callback(struct ptlrpc_request *req)
2254 {
2255         struct obd_quotactl *oqctl;
2256         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2257
2258         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2259         if (oqctl == NULL) {
2260                 CERROR("Can't unpack obd_quotactl\n");
2261                 RETURN(-EPROTO);
2262         }
2263
2264         cli->cl_qchk_stat = oqctl->qc_stat;
2265
2266         return 0;
2267 }
2268
2269 #ifdef HAVE_QUOTA_SUPPORT
2270 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2271 {
2272 #ifdef __KERNEL__
2273         struct obd_device *obd = req->rq_export->exp_obd;
2274         struct obd_device *master_obd;
2275         struct obd_device_target *obt;
2276         struct lustre_quota_ctxt *qctxt;
2277         struct qunit_data *qdata = NULL;
2278         int rc = 0;
2279         ENTRY;
2280
2281         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2282                 RETURN(rc);
2283
2284         rc = req_capsule_server_pack(&req->rq_pill);
2285         if (rc) {
2286                 CERROR("packing reply failed!: rc = %d\n", rc);
2287                 RETURN(rc);
2288         }
2289
2290         LASSERT(req->rq_export);
2291
2292         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2293         if (IS_ERR(qdata)) {
2294                 rc = PTR_ERR(qdata);
2295                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2296                 req->rq_status = rc;
2297                 GOTO(out, rc);
2298         }
2299
2300         /* we use the observer */
2301         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2302                 CERROR("Can't find the observer, it is recovering\n");
2303                 req->rq_status = -EAGAIN;
2304                 GOTO(out, rc);
2305         }
2306
2307         master_obd = obd->obd_observer->obd_observer;
2308         obt = &master_obd->u.obt;
2309         qctxt = &obt->obt_qctxt;
2310
2311         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2312                 /* quota_type has not been processed yet, return EAGAIN
2313                  * until we know whether or not quotas are supposed to
2314                  * be enabled */
2315                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2316                        "-EAGAIN\n");
2317                 req->rq_status = -EAGAIN;
2318                 GOTO(out, rc);
2319         }
2320
2321         down_read(&obt->obt_rwsem);
2322         if (qctxt->lqc_lqs_hash == NULL) {
2323                 up_read(&obt->obt_rwsem);
2324                 /* quota_type has not been processed yet, return EAGAIN
2325                  * until we know whether or not quotas are supposed to
2326                  * be enabled */
2327                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2328                        "-EAGAIN\n");
2329                 req->rq_status = -EAGAIN;
2330                 GOTO(out, rc);
2331         }
2332
2333         LASSERT(qctxt->lqc_handler);
2334         rc = qctxt->lqc_handler(master_obd, qdata,
2335                                 lustre_msg_get_opc(req->rq_reqmsg));
2336         up_read(&obt->obt_rwsem);
2337         if (rc && rc != -EDQUOT)
2338                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2339                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2340         req->rq_status = rc;
2341
2342         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2343         if (rc < 0) {
2344                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2345                 GOTO(out, rc);
2346         }
2347
2348         /* Block the quota req. b=14840 */
2349         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2350         EXIT;
2351
2352 out:
2353         rc = ptlrpc_reply(req);
2354         return rc;
2355 #else
2356         return 0;
2357 #endif /* !__KERNEL__ */
2358 }
2359 #endif /* HAVE_QUOTA_SUPPORT */
2360
2361 ldlm_mode_t lck_compat_array[] = {
2362         [LCK_EX] LCK_COMPAT_EX,
2363         [LCK_PW] LCK_COMPAT_PW,
2364         [LCK_PR] LCK_COMPAT_PR,
2365         [LCK_CW] LCK_COMPAT_CW,
2366         [LCK_CR] LCK_COMPAT_CR,
2367         [LCK_NL] LCK_COMPAT_NL,
2368         [LCK_GROUP] LCK_COMPAT_GROUP,
2369         [LCK_COS] LCK_COMPAT_COS,
2370 };
2371
2372 /**
2373  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2374  * not escape to the user level.
2375  */
2376 int ldlm_error2errno(ldlm_error_t error)
2377 {
2378         int result;
2379
2380         switch (error) {
2381         case ELDLM_OK:
2382                 result = 0;
2383                 break;
2384         case ELDLM_LOCK_CHANGED:
2385                 result = -ESTALE;
2386                 break;
2387         case ELDLM_LOCK_ABORTED:
2388                 result = -ENAVAIL;
2389                 break;
2390         case ELDLM_LOCK_REPLACED:
2391                 result = -ESRCH;
2392                 break;
2393         case ELDLM_NO_LOCK_DATA:
2394                 result = -ENOENT;
2395                 break;
2396         case ELDLM_NAMESPACE_EXISTS:
2397                 result = -EEXIST;
2398                 break;
2399         case ELDLM_BAD_NAMESPACE:
2400                 result = -EBADF;
2401                 break;
2402         default:
2403                 if (((int)error) < 0)  /* cast to signed type */
2404                         result = error; /* as ldlm_error_t can be unsigned */
2405                 else {
2406                         CERROR("Invalid DLM result code: %i\n", error);
2407                         result = -EPROTO;
2408                 }
2409         }
2410         return result;
2411 }
2412 EXPORT_SYMBOL(ldlm_error2errno);
2413
2414 /**
2415  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2416  */
2417 ldlm_error_t ldlm_errno2error(int err_no)
2418 {
2419         int error;
2420
2421         switch (err_no) {
2422         case 0:
2423                 error = ELDLM_OK;
2424                 break;
2425         case -ESTALE:
2426                 error = ELDLM_LOCK_CHANGED;
2427                 break;
2428         case -ENAVAIL:
2429                 error = ELDLM_LOCK_ABORTED;
2430                 break;
2431         case -ESRCH:
2432                 error = ELDLM_LOCK_REPLACED;
2433                 break;
2434         case -ENOENT:
2435                 error = ELDLM_NO_LOCK_DATA;
2436                 break;
2437         case -EEXIST:
2438                 error = ELDLM_NAMESPACE_EXISTS;
2439                 break;
2440         case -EBADF:
2441                 error = ELDLM_BAD_NAMESPACE;
2442                 break;
2443         default:
2444                 error = err_no;
2445         }
2446         return error;
2447 }
2448 EXPORT_SYMBOL(ldlm_errno2error);
2449