Whamcloud - gitweb
c8c06970c0b64815407a7d7ecf4586bbbb06988b
[fs/lustre-release.git] / lustre / ldlm / ldlm_lib.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
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 #define DEBUG_SUBSYSTEM S_LDLM
38
39 #ifdef __KERNEL__
40 # include <libcfs/libcfs.h>
41 #else
42 # include <liblustre.h>
43 #endif
44 #include <obd.h>
45 #include <lustre_mds.h>
46 #include <lustre_dlm.h>
47 #include <lustre_net.h>
48 #include <lustre_sec.h>
49 #include "ldlm_internal.h"
50
51 /* @priority: if non-zero, move the selected to the list head
52  * @create: if zero, only search in existed connections
53  */
54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
55                            int priority, int create)
56 {
57         struct ptlrpc_connection *ptlrpc_conn;
58         struct obd_import_conn *imp_conn = NULL, *item;
59         int rc = 0;
60         ENTRY;
61
62         if (!create && !priority) {
63                 CDEBUG(D_HA, "Nothing to do\n");
64                 RETURN(-EINVAL);
65         }
66
67         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
68         if (!ptlrpc_conn) {
69                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
70                 RETURN (-ENOENT);
71         }
72
73         if (create) {
74                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
75                 if (!imp_conn) {
76                         GOTO(out_put, rc = -ENOMEM);
77                 }
78         }
79
80         cfs_spin_lock(&imp->imp_lock);
81         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
82                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
83                         if (priority) {
84                                 cfs_list_del(&item->oic_item);
85                                 cfs_list_add(&item->oic_item,
86                                              &imp->imp_conn_list);
87                                 item->oic_last_attempt = 0;
88                         }
89                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
90                                imp, imp->imp_obd->obd_name, uuid->uuid,
91                                (priority ? ", moved to head" : ""));
92                         cfs_spin_unlock(&imp->imp_lock);
93                         GOTO(out_free, rc = 0);
94                 }
95         }
96         /* not found */
97         if (create) {
98                 imp_conn->oic_conn = ptlrpc_conn;
99                 imp_conn->oic_uuid = *uuid;
100                 imp_conn->oic_last_attempt = 0;
101                 if (priority)
102                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
103                 else
104                         cfs_list_add_tail(&imp_conn->oic_item,
105                                           &imp->imp_conn_list);
106                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
107                        imp, imp->imp_obd->obd_name, uuid->uuid,
108                        (priority ? "head" : "tail"));
109         } else {
110                 cfs_spin_unlock(&imp->imp_lock);
111                 GOTO(out_free, rc = -ENOENT);
112         }
113
114         cfs_spin_unlock(&imp->imp_lock);
115         RETURN(0);
116 out_free:
117         if (imp_conn)
118                 OBD_FREE(imp_conn, sizeof(*imp_conn));
119 out_put:
120         ptlrpc_connection_put(ptlrpc_conn);
121         RETURN(rc);
122 }
123
124 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
125 {
126         return import_set_conn(imp, uuid, 1, 0);
127 }
128
129 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
130                            int priority)
131 {
132         return import_set_conn(imp, uuid, priority, 1);
133 }
134
135 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
136 {
137         struct obd_import_conn *imp_conn;
138         struct obd_export *dlmexp;
139         int rc = -ENOENT;
140         ENTRY;
141
142         cfs_spin_lock(&imp->imp_lock);
143         if (cfs_list_empty(&imp->imp_conn_list)) {
144                 LASSERT(!imp->imp_connection);
145                 GOTO(out, rc);
146         }
147
148         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
149                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
150                         continue;
151                 LASSERT(imp_conn->oic_conn);
152
153                 /* is current conn? */
154                 if (imp_conn == imp->imp_conn_current) {
155                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
156
157                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
158                             imp->imp_state != LUSTRE_IMP_DISCON) {
159                                 CERROR("can't remove current connection\n");
160                                 GOTO(out, rc = -EBUSY);
161                         }
162
163                         ptlrpc_connection_put(imp->imp_connection);
164                         imp->imp_connection = NULL;
165
166                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
167                         if (dlmexp && dlmexp->exp_connection) {
168                                 LASSERT(dlmexp->exp_connection ==
169                                         imp_conn->oic_conn);
170                                 ptlrpc_connection_put(dlmexp->exp_connection);
171                                 dlmexp->exp_connection = NULL;
172                         }
173                 }
174
175                 cfs_list_del(&imp_conn->oic_item);
176                 ptlrpc_connection_put(imp_conn->oic_conn);
177                 OBD_FREE(imp_conn, sizeof(*imp_conn));
178                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
179                        imp, imp->imp_obd->obd_name, uuid->uuid);
180                 rc = 0;
181                 break;
182         }
183 out:
184         cfs_spin_unlock(&imp->imp_lock);
185         if (rc == -ENOENT)
186                 CERROR("connection %s not found\n", uuid->uuid);
187         RETURN(rc);
188 }
189
190 /**
191  * Find conn uuid by peer nid. @peer is a server nid. This function is used
192  * to find a conn uuid of @imp which can reach @peer.
193  */
194 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
195                             struct obd_uuid *uuid)
196 {
197         struct obd_import_conn *conn;
198         int rc = -ENOENT;
199         ENTRY;
200
201         cfs_spin_lock(&imp->imp_lock);
202         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
203                 /* check if conn uuid does have this peer nid */
204                 if (class_check_uuid(&conn->oic_uuid, peer)) {
205                         *uuid = conn->oic_uuid;
206                         rc = 0;
207                         break;
208                 }
209         }
210         cfs_spin_unlock(&imp->imp_lock);
211         RETURN(rc);
212 }
213 EXPORT_SYMBOL(client_import_find_conn);
214
215 void client_destroy_import(struct obd_import *imp)
216 {
217         /* drop security policy instance after all rpc finished/aborted
218          * to let all busy contexts be released. */
219         class_import_get(imp);
220         class_destroy_import(imp);
221         sptlrpc_import_sec_put(imp);
222         class_import_put(imp);
223 }
224 EXPORT_SYMBOL(client_destroy_import);
225
226 /**
227  * check whether the osc is on MDT or not
228  * In the config log,
229  * osc on MDT
230  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
231  * osc on client
232  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
233  *
234  **/
235 static int osc_on_mdt(char *obdname)
236 {
237         char *ptr;
238
239         ptr = strrchr(obdname, '-');
240         if (ptr == NULL)
241                 return 0;
242
243         if (strncmp(ptr + 1, "MDT", 3) == 0)
244                 return 1;
245
246         return 0;
247 }
248
249 /* configure an RPC client OBD device
250  *
251  * lcfg parameters:
252  * 1 - client UUID
253  * 2 - server UUID
254  * 3 - inactive-on-startup
255  */
256 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
257 {
258         struct client_obd *cli = &obddev->u.cli;
259         struct obd_import *imp;
260         struct obd_uuid server_uuid;
261         int rq_portal, rp_portal, connect_op;
262         char *name = obddev->obd_type->typ_name;
263         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
264         int rc;
265         ENTRY;
266
267         /* In a more perfect world, we would hang a ptlrpc_client off of
268          * obd_type and just use the values from there. */
269         if (!strcmp(name, LUSTRE_OSC_NAME)) {
270                 rq_portal = OST_REQUEST_PORTAL;
271                 rp_portal = OSC_REPLY_PORTAL;
272                 connect_op = OST_CONNECT;
273                 cli->cl_sp_me = LUSTRE_SP_CLI;
274                 cli->cl_sp_to = LUSTRE_SP_OST;
275                 ns_type = LDLM_NS_TYPE_OSC;
276
277         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
278                 rq_portal = MDS_REQUEST_PORTAL;
279                 rp_portal = MDC_REPLY_PORTAL;
280                 connect_op = MDS_CONNECT;
281                 cli->cl_sp_me = LUSTRE_SP_CLI;
282                 cli->cl_sp_to = LUSTRE_SP_MDT;
283                 ns_type = LDLM_NS_TYPE_MDC;
284
285         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
286                 rq_portal = MGS_REQUEST_PORTAL;
287                 rp_portal = MGC_REPLY_PORTAL;
288                 connect_op = MGS_CONNECT;
289                 cli->cl_sp_me = LUSTRE_SP_MGC;
290                 cli->cl_sp_to = LUSTRE_SP_MGS;
291                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
292                 ns_type = LDLM_NS_TYPE_MGC;
293
294         } else {
295                 CERROR("unknown client OBD type \"%s\", can't setup\n",
296                        name);
297                 RETURN(-EINVAL);
298         }
299
300         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
301                 CERROR("requires a TARGET UUID\n");
302                 RETURN(-EINVAL);
303         }
304
305         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
306                 CERROR("client UUID must be less than 38 characters\n");
307                 RETURN(-EINVAL);
308         }
309
310         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
311                 CERROR("setup requires a SERVER UUID\n");
312                 RETURN(-EINVAL);
313         }
314
315         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
316                 CERROR("target UUID must be less than 38 characters\n");
317                 RETURN(-EINVAL);
318         }
319
320         cfs_init_rwsem(&cli->cl_sem);
321         cfs_sema_init(&cli->cl_mgc_sem, 1);
322         cli->cl_conn_count = 0;
323         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
324                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
325                      sizeof(server_uuid)));
326
327         cli->cl_dirty = 0;
328         cli->cl_avail_grant = 0;
329         /* FIXME: should limit this for the sum of all cl_dirty_max */
330         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
331         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
332                 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
333         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
334         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
335         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
336         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
337         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
338         client_obd_list_lock_init(&cli->cl_loi_list_lock);
339         cfs_atomic_set(&cli->cl_pending_w_pages, 0);
340         cfs_atomic_set(&cli->cl_pending_r_pages, 0);
341         cli->cl_r_in_flight = 0;
342         cli->cl_w_in_flight = 0;
343
344         cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
345         cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
346         cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
347         cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
348         cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
349         cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
350         cfs_waitq_init(&cli->cl_destroy_waitq);
351         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
352 #ifdef ENABLE_CHECKSUM
353         /* Turn on checksumming by default. */
354         cli->cl_checksum = 1;
355         /*
356          * The supported checksum types will be worked out at connect time
357          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
358          * through procfs.
359          */
360         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
361 #endif
362         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
363
364         /* This value may be changed at connect time in
365            ptlrpc_connect_interpret. */
366         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
367                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
368
369         if (!strcmp(name, LUSTRE_MDC_NAME)) {
370                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
371         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
372                 cli->cl_max_rpcs_in_flight = 2;
373         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
374                 cli->cl_max_rpcs_in_flight = 3;
375         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
376                 cli->cl_max_rpcs_in_flight = 4;
377         } else {
378                 if (osc_on_mdt(obddev->obd_name))
379                         cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
380                 else
381                         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
382         }
383         rc = ldlm_get_ref();
384         if (rc) {
385                 CERROR("ldlm_get_ref failed: %d\n", rc);
386                 GOTO(err, rc);
387         }
388
389         ptlrpc_init_client(rq_portal, rp_portal, name,
390                            &obddev->obd_ldlm_client);
391
392         imp = class_new_import(obddev);
393         if (imp == NULL)
394                 GOTO(err_ldlm, rc = -ENOENT);
395         imp->imp_client = &obddev->obd_ldlm_client;
396         imp->imp_connect_op = connect_op;
397         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
398         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
399                LUSTRE_CFG_BUFLEN(lcfg, 1));
400         class_import_put(imp);
401
402         rc = client_import_add_conn(imp, &server_uuid, 1);
403         if (rc) {
404                 CERROR("can't add initial connection\n");
405                 GOTO(err_import, rc);
406         }
407
408         cli->cl_import = imp;
409         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
410         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
411         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
412
413         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
414                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
415                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
416                                name, obddev->obd_name,
417                                cli->cl_target_uuid.uuid);
418                         cfs_spin_lock(&imp->imp_lock);
419                         imp->imp_deactive = 1;
420                         cfs_spin_unlock(&imp->imp_lock);
421                 }
422         }
423
424         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
425                                                    LDLM_NAMESPACE_CLIENT,
426                                                    LDLM_NAMESPACE_GREEDY,
427                                                    ns_type);
428         if (obddev->obd_namespace == NULL) {
429                 CERROR("Unable to create client namespace - %s\n",
430                        obddev->obd_name);
431                 GOTO(err_import, rc = -ENOMEM);
432         }
433
434         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
435
436         RETURN(rc);
437
438 err_import:
439         class_destroy_import(imp);
440 err_ldlm:
441         ldlm_put_ref();
442 err:
443         RETURN(rc);
444
445 }
446
447 int client_obd_cleanup(struct obd_device *obddev)
448 {
449         ENTRY;
450
451         ldlm_namespace_free_post(obddev->obd_namespace);
452         obddev->obd_namespace = NULL;
453
454         LASSERT(obddev->u.cli.cl_import == NULL);
455
456         ldlm_put_ref();
457         RETURN(0);
458 }
459
460 /* ->o_connect() method for client side (OSC and MDC and MGC) */
461 int client_connect_import(const struct lu_env *env,
462                           struct obd_export **exp,
463                           struct obd_device *obd, struct obd_uuid *cluuid,
464                           struct obd_connect_data *data, void *localdata)
465 {
466         struct client_obd *cli = &obd->u.cli;
467         struct obd_import *imp = cli->cl_import;
468         struct obd_connect_data *ocd;
469         struct lustre_handle conn = { 0 };
470         int rc;
471         ENTRY;
472
473         *exp = NULL;
474         cfs_down_write(&cli->cl_sem);
475         if (cli->cl_conn_count > 0 )
476                 GOTO(out_sem, rc = -EALREADY);
477
478         rc = class_connect(&conn, obd, cluuid);
479         if (rc)
480                 GOTO(out_sem, rc);
481
482         cli->cl_conn_count++;
483         *exp = class_conn2export(&conn);
484
485         LASSERT(obd->obd_namespace);
486
487         imp->imp_dlm_handle = conn;
488         rc = ptlrpc_init_import(imp);
489         if (rc != 0)
490                 GOTO(out_ldlm, rc);
491
492         ocd = &imp->imp_connect_data;
493         if (data) {
494                 *ocd = *data;
495                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
496         }
497
498         rc = ptlrpc_connect_import(imp);
499         if (rc != 0) {
500                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
501                 GOTO(out_ldlm, rc);
502         }
503         LASSERT((*exp)->exp_connection);
504
505         if (data) {
506                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
507                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
508                          data->ocd_connect_flags, ocd->ocd_connect_flags);
509                 data->ocd_connect_flags = ocd->ocd_connect_flags;
510         }
511
512         ptlrpc_pinger_add_import(imp);
513
514         EXIT;
515
516         if (rc) {
517 out_ldlm:
518                 cli->cl_conn_count--;
519                 class_disconnect(*exp);
520                 *exp = NULL;
521         }
522 out_sem:
523         cfs_up_write(&cli->cl_sem);
524
525         return rc;
526 }
527
528 int client_disconnect_export(struct obd_export *exp)
529 {
530         struct obd_device *obd = class_exp2obd(exp);
531         struct client_obd *cli;
532         struct obd_import *imp;
533         int rc = 0, err;
534         ENTRY;
535
536         if (!obd) {
537                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
538                        exp, exp ? exp->exp_handle.h_cookie : -1);
539                 RETURN(-EINVAL);
540         }
541
542         cli = &obd->u.cli;
543         imp = cli->cl_import;
544
545         cfs_down_write(&cli->cl_sem);
546         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
547                cli->cl_conn_count);
548
549         if (!cli->cl_conn_count) {
550                 CERROR("disconnecting disconnected device (%s)\n",
551                        obd->obd_name);
552                 GOTO(out_disconnect, rc = -EINVAL);
553         }
554
555         cli->cl_conn_count--;
556         if (cli->cl_conn_count)
557                 GOTO(out_disconnect, rc = 0);
558
559         /* Mark import deactivated now, so we don't try to reconnect if any
560          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
561          * fully deactivate the import, or that would drop all requests. */
562         cfs_spin_lock(&imp->imp_lock);
563         imp->imp_deactive = 1;
564         cfs_spin_unlock(&imp->imp_lock);
565
566         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
567          * delete it regardless.  (It's safe to delete an import that was
568          * never added.) */
569         (void)ptlrpc_pinger_del_import(imp);
570
571         if (obd->obd_namespace != NULL) {
572                 /* obd_force == local only */
573                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
574                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
575                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
576         }
577
578         /*
579          * there's no need to hold sem during disconnecting an import,
580          * and actually it may cause deadlock in gss.
581          */
582         cfs_up_write(&cli->cl_sem);
583         rc = ptlrpc_disconnect_import(imp, 0);
584         cfs_down_write(&cli->cl_sem);
585
586         ptlrpc_invalidate_import(imp);
587
588         EXIT;
589
590  out_disconnect:
591         /* use server style - class_disconnect should be always called for
592          * o_disconnect */
593         err = class_disconnect(exp);
594         if (!rc && err)
595                 rc = err;
596
597         cfs_up_write(&cli->cl_sem);
598
599         RETURN(rc);
600 }
601
602 #ifdef HAVE_SERVER_SUPPORT
603 int server_disconnect_export(struct obd_export *exp)
604 {
605         int rc;
606         ENTRY;
607
608         /* Disconnect early so that clients can't keep using export */
609         rc = class_disconnect(exp);
610         /* close import for avoid sending any requests */
611         if (exp->exp_imp_reverse)
612                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
613
614         if (exp->exp_obd->obd_namespace != NULL)
615                 ldlm_cancel_locks_for_export(exp);
616
617         /* complete all outstanding replies */
618         cfs_spin_lock(&exp->exp_lock);
619         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
620                 struct ptlrpc_reply_state *rs =
621                         cfs_list_entry(exp->exp_outstanding_replies.next,
622                                        struct ptlrpc_reply_state, rs_exp_list);
623                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
624
625                 cfs_spin_lock(&svcpt->scp_rep_lock);
626
627                 cfs_list_del_init(&rs->rs_exp_list);
628                 cfs_spin_lock(&rs->rs_lock);
629                 ptlrpc_schedule_difficult_reply(rs);
630                 cfs_spin_unlock(&rs->rs_lock);
631
632                 cfs_spin_unlock(&svcpt->scp_rep_lock);
633         }
634         cfs_spin_unlock(&exp->exp_lock);
635
636         RETURN(rc);
637 }
638
639 /* --------------------------------------------------------------------------
640  * from old lib/target.c
641  * -------------------------------------------------------------------------- */
642
643 static int target_handle_reconnect(struct lustre_handle *conn,
644                                    struct obd_export *exp,
645                                    struct obd_uuid *cluuid)
646 {
647         ENTRY;
648
649         if (exp->exp_connection && exp->exp_imp_reverse) {
650                 struct lustre_handle *hdl;
651                 struct obd_device *target;
652
653                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
654                 target = exp->exp_obd;
655
656                 /* Might be a re-connect after a partition. */
657                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
658                         if (target->obd_recovering) {
659                                 int timeout = cfs_duration_sec(cfs_time_sub(
660                                         cfs_timer_deadline(
661                                         &target->obd_recovery_timer),
662                                         cfs_time_current()));
663
664                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
665                                         "ing, waiting for %d clients in recov"
666                                         "ery for %d:%.02d\n", target->obd_name,
667                                         obd_uuid2str(&exp->exp_client_uuid),
668                                         obd_export_nid2str(exp),
669                                         target->obd_max_recoverable_clients,
670                                         timeout / 60, timeout % 60);
671                         } else {
672                                 LCONSOLE_WARN("%s: Client %s (at %s) "
673                                         "reconnecting\n", target->obd_name,
674                                         obd_uuid2str(&exp->exp_client_uuid),
675                                         obd_export_nid2str(exp));
676                         }
677
678                         conn->cookie = exp->exp_handle.h_cookie;
679                         /* target_handle_connect() treats EALREADY and
680                          * -EALREADY differently.  EALREADY means we are
681                          * doing a valid reconnect from the same client. */
682                         RETURN(EALREADY);
683                 } else {
684                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
685                                       "with handle "LPX64". Rejecting client "
686                                       "with the same UUID trying to reconnect "
687                                       "with handle "LPX64"\n", target->obd_name,
688                                       obd_uuid2str(&exp->exp_client_uuid),
689                                       obd_export_nid2str(exp),
690                                       hdl->cookie, conn->cookie);
691                         memset(conn, 0, sizeof *conn);
692                         /* target_handle_connect() treats EALREADY and
693                          * -EALREADY differently.  -EALREADY is an error
694                          * (same UUID, different handle). */
695                         RETURN(-EALREADY);
696                 }
697         }
698
699         conn->cookie = exp->exp_handle.h_cookie;
700         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
701                cluuid->uuid, exp, conn->cookie);
702         RETURN(0);
703 }
704
705 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
706                           int error)
707 {
708         struct obd_export *exp = cb_data;
709
710         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
711                obd->obd_name, exp->exp_client_uuid.uuid);
712
713         cfs_spin_lock(&exp->exp_lock);
714         exp->exp_need_sync = 0;
715         cfs_spin_unlock(&exp->exp_lock);
716         class_export_cb_put(exp);
717 }
718 EXPORT_SYMBOL(target_client_add_cb);
719
720 #ifdef __KERNEL__
721 static void
722 check_and_start_recovery_timer(struct obd_device *obd,
723                                struct ptlrpc_request *req, int new_client);
724 #else
725 static inline void
726 check_and_start_recovery_timer(struct obd_device *obd,
727                                struct ptlrpc_request *req, int new_client)
728 {
729 }
730 #endif
731
732 int target_handle_connect(struct ptlrpc_request *req)
733 {
734         struct obd_device *target = NULL, *targref = NULL;
735         struct obd_export *export = NULL;
736         struct obd_import *revimp;
737         struct lustre_handle conn;
738         struct lustre_handle *tmp;
739         struct obd_uuid tgtuuid;
740         struct obd_uuid cluuid;
741         struct obd_uuid remote_uuid;
742         char *str;
743         int rc = 0;
744         char *target_start;
745         int target_len;
746         int mds_conn = 0;
747         struct obd_connect_data *data, *tmpdata;
748         int size, tmpsize;
749         lnet_nid_t *client_nid = NULL;
750         bool mne_swab_client_ver;
751         ENTRY;
752
753         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
754
755         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
756         if (str == NULL) {
757                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
758                 GOTO(out, rc = -EINVAL);
759         }
760
761         obd_str2uuid(&tgtuuid, str);
762         target = class_uuid2obd(&tgtuuid);
763         if (!target)
764                 target = class_name2obd(str);
765
766         if (!target) {
767                 deuuidify(str, NULL, &target_start, &target_len);
768                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available for "
769                                    "connect (no target)\n", str);
770                 GOTO(out, rc = -ENODEV);
771         }
772
773         cfs_spin_lock(&target->obd_dev_lock);
774         if (target->obd_stopping || !target->obd_set_up) {
775                 cfs_spin_unlock(&target->obd_dev_lock);
776
777                 deuuidify(str, NULL, &target_start, &target_len);
778                 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
779                                    "from %s (%s)\n", target_len, target_start,
780                                    libcfs_nid2str(req->rq_peer.nid), 
781                                    (target->obd_stopping ?
782                                    "stopping" : "not set up"));
783                 GOTO(out, rc = -ENODEV);
784         }
785
786         if (target->obd_no_conn) {
787                 cfs_spin_unlock(&target->obd_dev_lock);
788
789                 LCONSOLE_WARN("%s: Temporarily refusing client connection "
790                               "from %s\n", target->obd_name,
791                               libcfs_nid2str(req->rq_peer.nid));
792                 GOTO(out, rc = -EAGAIN);
793         }
794
795         /* Make sure the target isn't cleaned up while we're here. Yes,
796            there's still a race between the above check and our incref here.
797            Really, class_uuid2obd should take the ref. */
798         targref = class_incref(target, __FUNCTION__, cfs_current());
799
800         target->obd_conn_inprogress++;
801         cfs_spin_unlock(&target->obd_dev_lock);
802
803         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
804         if (str == NULL) {
805                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
806                 GOTO(out, rc = -EINVAL);
807         }
808
809         obd_str2uuid(&cluuid, str);
810
811         /* XXX extract a nettype and format accordingly */
812         switch (sizeof(lnet_nid_t)) {
813                 /* NB the casts only avoid compiler warnings */
814         case 8:
815                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
816                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
817                 break;
818         case 4:
819                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
820                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
821                 break;
822         default:
823                 LBUG();
824         }
825
826         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
827         if (tmp == NULL)
828                 GOTO(out, rc = -EPROTO);
829
830         conn = *tmp;
831
832         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
833                                     RCL_CLIENT);
834         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
835         if (!data)
836                 GOTO(out, rc = -EPROTO);
837
838         rc = req_capsule_server_pack(&req->rq_pill);
839         if (rc)
840                 GOTO(out, rc);
841
842 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
843         /* Check if the client might be missing the LU-1252 fix to swab
844          * the IR mne_length entries. Do this as early as possible in case
845          * the version code is modified. See LU-1644 for details. */
846         mne_swab_client_ver = OBD_OCD_VERSION_MAJOR(data->ocd_version) == 2 &&
847                               OBD_OCD_VERSION_MINOR(data->ocd_version) == 2 &&
848                               OBD_OCD_VERSION_PATCH(data->ocd_version) < 55;
849 #endif
850
851         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
852                 if (!data) {
853                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
854                                   "libclient connection attempt");
855                         GOTO(out, rc = -EPROTO);
856                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
857                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
858                            data->ocd_version > LUSTRE_VERSION_CODE +
859                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
860                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
861                                   "libclient connection attempt",
862                                   data->ocd_version < LUSTRE_VERSION_CODE ?
863                                   "old" : "new",
864                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
865                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
866                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
867                                   OBD_OCD_VERSION_FIX(data->ocd_version));
868                         data = req_capsule_server_sized_get(&req->rq_pill,
869                                                         &RMF_CONNECT_DATA,
870                                     offsetof(typeof(*data), ocd_version) +
871                                              sizeof(data->ocd_version));
872                         if (data) {
873                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
874                                 data->ocd_version = LUSTRE_VERSION_CODE;
875                         }
876                         GOTO(out, rc = -EPROTO);
877                 }
878         }
879
880         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
881             (data->ocd_connect_flags & OBD_CONNECT_MDS))
882                 mds_conn = 1;
883
884         /* lctl gets a backstage, all-access pass. */
885         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
886                 goto dont_check_exports;
887
888         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
889         if (!export)
890                 goto no_export;
891
892         /* we've found an export in the hash */
893         if (export->exp_connecting) { /* bug 9635, et. al. */
894                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
895                               export->exp_obd->obd_name, export,
896                               libcfs_nid2str(req->rq_peer.nid));
897                 class_export_put(export);
898                 export = NULL;
899                 rc = -EALREADY;
900         } else if (mds_conn && export->exp_connection) {
901                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
902                         /* mds reconnected after failover */
903                         LCONSOLE_WARN("%s: Received MDS connection from "
904                             "%s, removing former export from %s\n",
905                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
906                             libcfs_nid2str(export->exp_connection->c_peer.nid));
907                 else
908                         /* new mds connection from the same nid */
909                         LCONSOLE_WARN("%s: Received new MDS connection from "
910                             "%s, removing former export from same NID\n",
911                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
912                 class_fail_export(export);
913                 class_export_put(export);
914                 export = NULL;
915                 rc = 0;
916         } else if (export->exp_connection != NULL &&
917                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
918                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
919                     MSG_CONNECT_INITIAL)) {
920                 /* in mds failover we have static uuid but nid can be
921                  * changed*/
922                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
923                               "existing nid %s is already connected\n",
924                               target->obd_name, cluuid.uuid,
925                               libcfs_nid2str(req->rq_peer.nid),
926                               libcfs_nid2str(
927                                       export->exp_connection->c_peer.nid));
928                 rc = -EALREADY;
929                 class_export_put(export);
930                 export = NULL;
931         } else {
932                 cfs_spin_lock(&export->exp_lock);
933                 export->exp_connecting = 1;
934                 cfs_spin_unlock(&export->exp_lock);
935                 LASSERT(export->exp_obd == target);
936
937                 rc = target_handle_reconnect(&conn, export, &cluuid);
938         }
939
940         /* If we found an export, we already unlocked. */
941         if (!export) {
942 no_export:
943                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
944         } else if (req->rq_export == NULL &&
945                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
946                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
947                               "still busy with %d references\n",
948                               target->obd_name, cluuid.uuid,
949                               libcfs_nid2str(req->rq_peer.nid),
950                               cfs_atomic_read(&export->exp_refcount));
951                 GOTO(out, rc = -EBUSY);
952         } else if (req->rq_export != NULL &&
953                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
954                 /* the current connect rpc has increased exp_rpc_count */
955                 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
956                               "still busy with %d active RPCs\n",
957                               target->obd_name, cluuid.uuid,
958                               libcfs_nid2str(req->rq_peer.nid),
959                               cfs_atomic_read(&export->exp_rpc_count) - 1);
960                 cfs_spin_lock(&export->exp_lock);
961                 if (req->rq_export->exp_conn_cnt <
962                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
963                         /* try to abort active requests */
964                         req->rq_export->exp_abort_active_req = 1;
965                 cfs_spin_unlock(&export->exp_lock);
966                 GOTO(out, rc = -EBUSY);
967         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
968                 if (!strstr(cluuid.uuid, "mdt"))
969                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
970                                       "known client %s (at %s) because it "
971                                       "is indicating it is a new client",
972                                       target->obd_name, cluuid.uuid,
973                                       libcfs_nid2str(req->rq_peer.nid));
974                 GOTO(out, rc = -EALREADY);
975         } else {
976                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
977         }
978
979         if (rc < 0) {
980                 GOTO(out, rc);
981         }
982
983         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
984                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
985               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
986               export, (long)cfs_time_current_sec(),
987               export ? (long)export->exp_last_request_time : 0);
988
989         /* If this is the first time a client connects, reset the recovery
990          * timer */
991         if (rc == 0 && target->obd_recovering)
992                 check_and_start_recovery_timer(target, req, export == NULL);
993
994         /* We want to handle EALREADY but *not* -EALREADY from
995          * target_handle_reconnect(), return reconnection state in a flag */
996         if (rc == EALREADY) {
997                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
998                 rc = 0;
999         } else {
1000                 LASSERT(rc == 0);
1001         }
1002
1003         /* Tell the client if we support replayable requests */
1004         if (target->obd_replayable)
1005                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1006         client_nid = &req->rq_peer.nid;
1007
1008         if (export == NULL) {
1009                 if (target->obd_recovering) {
1010                         cfs_time_t t;
1011                         int        c; /* connected */
1012                         int        i; /* in progress */
1013                         int        k; /* known */
1014
1015                         c = cfs_atomic_read(&target->obd_connected_clients);
1016                         i = cfs_atomic_read(&target->obd_lock_replay_clients);
1017                         k = target->obd_max_recoverable_clients;
1018                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1019                         t = cfs_time_sub(t, cfs_time_current());
1020                         t = cfs_duration_sec(t);
1021                         LCONSOLE_WARN("%s: Denying connection for new client "
1022                                       "%s (at %s), waiting for all %d known "
1023                                       "clients (%d recovered, %d in progress, "
1024                                       "and %d unseen) to recover in %d:%.02d\n",
1025                                       target->obd_name, cluuid.uuid,
1026                                       libcfs_nid2str(req->rq_peer.nid), k,
1027                                       c - i, i, k - c, (int)t / 60,
1028                                       (int)t % 60);
1029                         rc = -EBUSY;
1030                 } else {
1031 dont_check_exports:
1032                         rc = obd_connect(req->rq_svc_thread->t_env,
1033                                          &export, target, &cluuid, data,
1034                                          client_nid);
1035                         if (rc == 0)
1036                                 conn.cookie = export->exp_handle.h_cookie;
1037                 }
1038         } else {
1039                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1040                                    export, target, &cluuid, data, client_nid);
1041         }
1042         if (rc)
1043                 GOTO(out, rc);
1044
1045 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
1046         /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
1047          * will do the swabbing for if the client is using the same endianness.
1048          *
1049          * This fixup is version-limited, because we don't want to carry the
1050          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
1051          * interop with unpatched 2.2 clients.  For newer clients, servers
1052          * will never do MNE swabbing, let the client handle that.  LU-1644 */
1053         export->exp_need_mne_swab =
1054                 !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
1055                 mne_swab_client_ver && !ptlrpc_req_need_swab(req);
1056 #else
1057 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
1058 #endif
1059
1060         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1061         data->ocd_instance = target->u.obt.obt_instance;
1062
1063         /* Return only the parts of obd_connect_data that we understand, so the
1064          * client knows that we don't understand the rest. */
1065         if (data) {
1066                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1067                                                RCL_SERVER);
1068                 tmpdata = req_capsule_server_get(&req->rq_pill,
1069                                                  &RMF_CONNECT_DATA);
1070                 /* Don't use struct assignment here, because the client reply
1071                  * buffer may be smaller/larger than the local struct
1072                  * obd_connect_data. */
1073                 memcpy(tmpdata, data, min(tmpsize, size));
1074         }
1075
1076         /* If all else goes well, this is our RPC return code. */
1077         req->rq_status = 0;
1078
1079         lustre_msg_set_handle(req->rq_repmsg, &conn);
1080
1081         /* If the client and the server are the same node, we will already
1082          * have an export that really points to the client's DLM export,
1083          * because we have a shared handles table.
1084          *
1085          * XXX this will go away when shaver stops sending the "connect" handle
1086          * in the real "remote handle" field of the request --phik 24 Apr 2003
1087          */
1088         if (req->rq_export != NULL)
1089                 class_export_put(req->rq_export);
1090
1091         /* request takes one export refcount */
1092         req->rq_export = class_export_get(export);
1093
1094         cfs_spin_lock(&export->exp_lock);
1095         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1096                 cfs_spin_unlock(&export->exp_lock);
1097                 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1098                        "conn_cnt: %d > %d\n",
1099                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1100                        export->exp_conn_cnt,
1101                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1102
1103                 GOTO(out, rc = -EALREADY);
1104         }
1105         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1106         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1107         export->exp_abort_active_req = 0;
1108
1109         /* request from liblustre?  Don't evict it for not pinging. */
1110         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1111                 export->exp_libclient = 1;
1112                 cfs_spin_unlock(&export->exp_lock);
1113
1114                 cfs_spin_lock(&target->obd_dev_lock);
1115                 cfs_list_del_init(&export->exp_obd_chain_timed);
1116                 cfs_spin_unlock(&target->obd_dev_lock);
1117         } else {
1118                 cfs_spin_unlock(&export->exp_lock);
1119         }
1120
1121         if (export->exp_connection != NULL) {
1122                 /* Check to see if connection came from another NID */
1123                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1124                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1125                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1126                                      &export->exp_connection->c_peer.nid,
1127                                      &export->exp_nid_hash);
1128
1129                 ptlrpc_connection_put(export->exp_connection);
1130         }
1131
1132         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1133                                                        req->rq_self,
1134                                                        &remote_uuid);
1135         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1136                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1137                              &export->exp_connection->c_peer.nid,
1138                              &export->exp_nid_hash);
1139         }
1140
1141         if (target->obd_recovering && !export->exp_in_recovery) {
1142                 int has_transno;
1143                 __u64 transno = data->ocd_transno;
1144
1145                 cfs_spin_lock(&export->exp_lock);
1146                 /* possible race with class_disconnect_stale_exports,
1147                  * export may be already in the eviction process */
1148                 if (export->exp_failed) {
1149                         cfs_spin_unlock(&export->exp_lock);
1150                         GOTO(out, rc = -ENODEV);
1151                 }
1152                 export->exp_in_recovery = 1;
1153                 export->exp_req_replay_needed = 1;
1154                 export->exp_lock_replay_needed = 1;
1155                 cfs_spin_unlock(&export->exp_lock);
1156
1157                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1158                                  MSG_CONNECT_TRANSNO);
1159                 if (has_transno && transno == 0)
1160                         CWARN("Connect with zero transno!\n");
1161
1162                 if (has_transno && transno > 0 &&
1163                     transno < target->obd_next_recovery_transno &&
1164                     transno > target->obd_last_committed) {
1165                         /* another way is to use cmpxchg() so it will be
1166                          * lock free */
1167                         cfs_spin_lock(&target->obd_recovery_task_lock);
1168                         if (transno < target->obd_next_recovery_transno)
1169                                 target->obd_next_recovery_transno = transno;
1170                         cfs_spin_unlock(&target->obd_recovery_task_lock);
1171                 }
1172
1173                 cfs_atomic_inc(&target->obd_req_replay_clients);
1174                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1175                 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1176                     target->obd_max_recoverable_clients)
1177                         cfs_waitq_signal(&target->obd_next_transno_waitq);
1178         }
1179
1180         /* Tell the client we're in recovery, when client is involved in it. */
1181         if (target->obd_recovering)
1182                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1183
1184         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1185         conn = *tmp;
1186
1187         if (export->exp_imp_reverse != NULL) {
1188                 /* destroyed import can be still referenced in ctxt */
1189                 obd_set_info_async(req->rq_svc_thread->t_env, export,
1190                                    sizeof(KEY_REVIMP_UPD), KEY_REVIMP_UPD,
1191                                    0, NULL, NULL);
1192
1193                 client_destroy_import(export->exp_imp_reverse);
1194         }
1195
1196         /* for the rest part, we return -ENOTCONN in case of errors
1197          * in order to let client initialize connection again.
1198          */
1199         revimp = export->exp_imp_reverse = class_new_import(target);
1200         if (!revimp) {
1201                 CERROR("fail to alloc new reverse import.\n");
1202                 GOTO(out, rc = -ENOTCONN);
1203         }
1204
1205         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1206         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1207         revimp->imp_remote_handle = conn;
1208         revimp->imp_dlm_fake = 1;
1209         revimp->imp_state = LUSTRE_IMP_FULL;
1210
1211         /* unknown versions will be caught in
1212          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1213         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1214
1215         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1216             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1217                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1218         else
1219                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1220
1221         if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1222             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1223                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1224         else
1225                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1226
1227         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1228         if (rc) {
1229                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1230                 export->exp_imp_reverse = NULL;
1231                 class_destroy_import(revimp);
1232         }
1233
1234         class_import_put(revimp);
1235 out:
1236         if (export) {
1237                 cfs_spin_lock(&export->exp_lock);
1238                 export->exp_connecting = 0;
1239                 cfs_spin_unlock(&export->exp_lock);
1240
1241                 class_export_put(export);
1242         }
1243         if (targref) {
1244                 cfs_spin_lock(&target->obd_dev_lock);
1245                 target->obd_conn_inprogress--;
1246                 cfs_spin_unlock(&target->obd_dev_lock);
1247
1248                 class_decref(targref, __FUNCTION__, cfs_current());
1249         }
1250         if (rc)
1251                 req->rq_status = rc;
1252         RETURN(rc);
1253 }
1254
1255 int target_handle_disconnect(struct ptlrpc_request *req)
1256 {
1257         int rc;
1258         ENTRY;
1259
1260         rc = req_capsule_server_pack(&req->rq_pill);
1261         if (rc)
1262                 RETURN(rc);
1263
1264         /* keep the rq_export around so we can send the reply */
1265         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1266
1267         RETURN(0);
1268 }
1269
1270 void target_destroy_export(struct obd_export *exp)
1271 {
1272         /* exports created from last_rcvd data, and "fake"
1273            exports created by lctl don't have an import */
1274         if (exp->exp_imp_reverse != NULL)
1275                 client_destroy_import(exp->exp_imp_reverse);
1276
1277         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1278         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1279         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1280         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1281 }
1282
1283 /*
1284  * Recovery functions
1285  */
1286 static void target_request_copy_get(struct ptlrpc_request *req)
1287 {
1288         class_export_rpc_get(req->rq_export);
1289         LASSERT(cfs_list_empty(&req->rq_list));
1290         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1291
1292         /* increase refcount to keep request in queue */
1293         cfs_atomic_inc(&req->rq_refcount);
1294         /** let export know it has replays to be handled */
1295         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1296 }
1297
1298 static void target_request_copy_put(struct ptlrpc_request *req)
1299 {
1300         LASSERT(cfs_list_empty(&req->rq_replay_list));
1301         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1302
1303         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1304         class_export_rpc_put(req->rq_export);
1305         ptlrpc_server_drop_request(req);
1306 }
1307
1308 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1309 {
1310         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1311         struct obd_export     *exp = req->rq_export;
1312         struct ptlrpc_request *reqiter;
1313         int                    dup = 0;
1314
1315         LASSERT(exp);
1316
1317         cfs_spin_lock(&exp->exp_lock);
1318         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1319                                 rq_replay_list) {
1320                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1321                         dup = 1;
1322                         break;
1323                 }
1324         }
1325
1326         if (dup) {
1327                 /* we expect it with RESENT and REPLAY flags */
1328                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1329                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1330                         CERROR("invalid flags %x of resent replay\n",
1331                                lustre_msg_get_flags(req->rq_reqmsg));
1332         } else {
1333                 cfs_list_add_tail(&req->rq_replay_list,
1334                                   &exp->exp_req_replay_queue);
1335         }
1336
1337         cfs_spin_unlock(&exp->exp_lock);
1338         return dup;
1339 }
1340
1341 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1342 {
1343         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1344         LASSERT(req->rq_export);
1345
1346         cfs_spin_lock(&req->rq_export->exp_lock);
1347         cfs_list_del_init(&req->rq_replay_list);
1348         cfs_spin_unlock(&req->rq_export->exp_lock);
1349 }
1350
1351 #ifdef __KERNEL__
1352 static void target_finish_recovery(struct obd_device *obd)
1353 {
1354         time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1355                                     obd->obd_recovery_start);
1356         ENTRY;
1357
1358         LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1359                       "%d recovered and %d %s evicted.\n", obd->obd_name,
1360                       (int)elapsed_time / 60, (int)elapsed_time % 60,
1361                       obd->obd_max_recoverable_clients,
1362                       cfs_atomic_read(&obd->obd_connected_clients),
1363                       obd->obd_stale_clients,
1364                       obd->obd_stale_clients == 1 ? "was" : "were");
1365
1366         ldlm_reprocess_all_ns(obd->obd_namespace);
1367         cfs_spin_lock(&obd->obd_recovery_task_lock);
1368         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1369             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1370             !cfs_list_empty(&obd->obd_final_req_queue)) {
1371                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1372                        obd->obd_name,
1373                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1374                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1375                                "" : "lock ",
1376                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1377                                "" : "final ");
1378                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1379                 LBUG();
1380         }
1381         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1382
1383         obd->obd_recovery_end = cfs_time_current_sec();
1384
1385         /* when recovery finished, cleanup orphans on mds and ost */
1386         if (OBT(obd) && OBP(obd, postrecov)) {
1387                 int rc = OBP(obd, postrecov)(obd);
1388                 if (rc < 0)
1389                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1390                                       obd->obd_name, rc);
1391         }
1392         EXIT;
1393 }
1394
1395 static void abort_req_replay_queue(struct obd_device *obd)
1396 {
1397         struct ptlrpc_request *req, *n;
1398         cfs_list_t abort_list;
1399
1400         CFS_INIT_LIST_HEAD(&abort_list);
1401         cfs_spin_lock(&obd->obd_recovery_task_lock);
1402         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1403         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1404         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1405                 DEBUG_REQ(D_WARNING, req, "aborted:");
1406                 req->rq_status = -ENOTCONN;
1407                 if (ptlrpc_error(req)) {
1408                         DEBUG_REQ(D_ERROR, req,
1409                                   "failed abort_req_reply; skipping");
1410                 }
1411                 target_exp_dequeue_req_replay(req);
1412                 target_request_copy_put(req);
1413         }
1414 }
1415
1416 static void abort_lock_replay_queue(struct obd_device *obd)
1417 {
1418         struct ptlrpc_request *req, *n;
1419         cfs_list_t abort_list;
1420
1421         CFS_INIT_LIST_HEAD(&abort_list);
1422         cfs_spin_lock(&obd->obd_recovery_task_lock);
1423         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1424         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1425         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1426                 DEBUG_REQ(D_ERROR, req, "aborted:");
1427                 req->rq_status = -ENOTCONN;
1428                 if (ptlrpc_error(req)) {
1429                         DEBUG_REQ(D_ERROR, req,
1430                                   "failed abort_lock_reply; skipping");
1431                 }
1432                 target_request_copy_put(req);
1433         }
1434 }
1435
1436 /* Called from a cleanup function if the device is being cleaned up
1437    forcefully.  The exports should all have been disconnected already,
1438    the only thing left to do is
1439      - clear the recovery flags
1440      - cancel the timer
1441      - free queued requests and replies, but don't send replies
1442    Because the obd_stopping flag is set, no new requests should be received.
1443
1444 */
1445 void target_cleanup_recovery(struct obd_device *obd)
1446 {
1447         struct ptlrpc_request *req, *n;
1448         cfs_list_t clean_list;
1449         ENTRY;
1450
1451         CFS_INIT_LIST_HEAD(&clean_list);
1452         cfs_spin_lock(&obd->obd_dev_lock);
1453         if (!obd->obd_recovering) {
1454                 cfs_spin_unlock(&obd->obd_dev_lock);
1455                 EXIT;
1456                 return;
1457         }
1458         obd->obd_recovering = obd->obd_abort_recovery = 0;
1459         cfs_spin_unlock(&obd->obd_dev_lock);
1460
1461         cfs_spin_lock(&obd->obd_recovery_task_lock);
1462         target_cancel_recovery_timer(obd);
1463         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1464         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1465
1466         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1467                 LASSERT(req->rq_reply_state == 0);
1468                 target_exp_dequeue_req_replay(req);
1469                 target_request_copy_put(req);
1470         }
1471
1472         cfs_spin_lock(&obd->obd_recovery_task_lock);
1473         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1474         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1475         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1476
1477         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1478                 LASSERT(req->rq_reply_state == 0);
1479                 target_request_copy_put(req);
1480         }
1481
1482         EXIT;
1483 }
1484
1485 /* obd_recovery_task_lock should be held */
1486 void target_cancel_recovery_timer(struct obd_device *obd)
1487 {
1488         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1489         cfs_timer_disarm(&obd->obd_recovery_timer);
1490 }
1491
1492 static void target_start_recovery_timer(struct obd_device *obd)
1493 {
1494         if (obd->obd_recovery_start != 0)
1495                 return;
1496
1497         cfs_spin_lock(&obd->obd_dev_lock);
1498         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1499                 cfs_spin_unlock(&obd->obd_dev_lock);
1500                 return;
1501         }
1502
1503         LASSERT(obd->obd_recovery_timeout != 0);
1504
1505         if (obd->obd_recovery_start != 0) {
1506                 cfs_spin_unlock(&obd->obd_dev_lock);
1507                 return;
1508         }
1509
1510         cfs_timer_arm(&obd->obd_recovery_timer,
1511                       cfs_time_shift(obd->obd_recovery_timeout));
1512         obd->obd_recovery_start = cfs_time_current_sec();
1513         cfs_spin_unlock(&obd->obd_dev_lock);
1514
1515         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1516                       "or until %d client%s reconnect%s\n",
1517                       obd->obd_name,
1518                       obd->obd_recovery_timeout / 60,
1519                       obd->obd_recovery_timeout % 60,
1520                       obd->obd_max_recoverable_clients,
1521                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1522                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1523 }
1524
1525 /**
1526  * extend recovery window.
1527  *
1528  * if @extend is true, extend recovery window to have @drt remaining at least;
1529  * otherwise, make sure the recovery timeout value is not less than @drt.
1530  */
1531 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1532 {
1533         cfs_time_t now;
1534         cfs_time_t end;
1535         cfs_duration_t left;
1536         int to;
1537
1538         cfs_spin_lock(&obd->obd_dev_lock);
1539         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1540                 cfs_spin_unlock(&obd->obd_dev_lock);
1541                 return;
1542         }
1543         LASSERT(obd->obd_recovery_start != 0);
1544
1545         now  = cfs_time_current_sec();
1546         to   = obd->obd_recovery_timeout;
1547         end  = obd->obd_recovery_start + to;
1548         left = cfs_time_sub(end, now);
1549
1550         if (extend && (drt > left)) {
1551                 to += drt - left;
1552         } else if (!extend && (drt > to)) {
1553                 to = drt;
1554                 /* reduce drt by already passed time */
1555                 drt -= obd->obd_recovery_timeout - left;
1556         }
1557
1558         if (to > obd->obd_recovery_time_hard)
1559                 to = obd->obd_recovery_time_hard;
1560         if (obd->obd_recovery_timeout < to) {
1561                 obd->obd_recovery_timeout = to;
1562                 cfs_timer_arm(&obd->obd_recovery_timer,
1563                               cfs_time_shift(drt));
1564         }
1565         cfs_spin_unlock(&obd->obd_dev_lock);
1566
1567         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1568                obd->obd_name, (unsigned)drt);
1569 }
1570
1571 /* Reset the timer with each new client connection */
1572 /*
1573  * This timer is actually reconnect_timer, which is for making sure
1574  * the total recovery window is at least as big as my reconnect
1575  * attempt timing. So the initial recovery time_out will be set to
1576  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1577  * from client is bigger than this, then the recovery time_out will
1578  * be extended to make sure the client could be reconnected, in the
1579  * process, the timeout from the new client should be ignored.
1580  */
1581
1582 static void
1583 check_and_start_recovery_timer(struct obd_device *obd,
1584                                struct ptlrpc_request *req,
1585                                int new_client)
1586 {
1587         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1588         struct obd_device_target *obt = &obd->u.obt;
1589         struct lustre_sb_info *lsi;
1590
1591         if (!new_client && service_time)
1592                 /* Teach server about old server's estimates, as first guess
1593                  * at how long new requests will take. */
1594                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1595                             service_time);
1596
1597         target_start_recovery_timer(obd);
1598
1599         /* convert the service time to rpc timeout,
1600          * reuse service_time to limit stack usage */
1601         service_time = at_est2timeout(service_time);
1602
1603         /* We expect other clients to timeout within service_time, then try
1604          * to reconnect, then try the failover server.  The max delay between
1605          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1606         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1607
1608         LASSERT(obt->obt_magic == OBT_MAGIC);
1609         lsi = s2lsi(obt->obt_sb);
1610         if (!(lsi->lsi_flags | LSI_IR_CAPABLE))
1611                 service_time += 2 * (CONNECTION_SWITCH_MAX +
1612                                      CONNECTION_SWITCH_INC);
1613         if (service_time > obd->obd_recovery_timeout && !new_client)
1614                 extend_recovery_timer(obd, service_time, false);
1615 }
1616
1617 /** Health checking routines */
1618 static inline int exp_connect_healthy(struct obd_export *exp)
1619 {
1620         return (exp->exp_in_recovery);
1621 }
1622
1623 /** if export done req_replay or has replay in queue */
1624 static inline int exp_req_replay_healthy(struct obd_export *exp)
1625 {
1626         return (!exp->exp_req_replay_needed ||
1627                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1628 }
1629 /** if export done lock_replay or has replay in queue */
1630 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1631 {
1632         return (!exp->exp_lock_replay_needed ||
1633                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1634 }
1635
1636 static inline int exp_vbr_healthy(struct obd_export *exp)
1637 {
1638         return (!exp->exp_vbr_failed);
1639 }
1640
1641 static inline int exp_finished(struct obd_export *exp)
1642 {
1643         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1644 }
1645
1646 /** Checking routines for recovery */
1647 static int check_for_clients(struct obd_device *obd)
1648 {
1649         unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1650
1651         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1652                 return 1;
1653         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1654         if (obd->obd_no_conn == 0 &&
1655             clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1656                 return 1;
1657         return 0;
1658 }
1659
1660 static int check_for_next_transno(struct obd_device *obd)
1661 {
1662         struct ptlrpc_request *req = NULL;
1663         int wake_up = 0, connected, completed, queue_len;
1664         __u64 next_transno, req_transno;
1665         ENTRY;
1666
1667         cfs_spin_lock(&obd->obd_recovery_task_lock);
1668         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1669                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1670                                      struct ptlrpc_request, rq_list);
1671                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1672         } else {
1673                 req_transno = 0;
1674         }
1675
1676         connected = cfs_atomic_read(&obd->obd_connected_clients);
1677         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1678         queue_len = obd->obd_requests_queued_for_recovery;
1679         next_transno = obd->obd_next_recovery_transno;
1680
1681         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1682                "req_transno: "LPU64", next_transno: "LPU64"\n",
1683                obd->obd_max_recoverable_clients, connected, completed,
1684                queue_len, req_transno, next_transno);
1685
1686         if (obd->obd_abort_recovery) {
1687                 CDEBUG(D_HA, "waking for aborted recovery\n");
1688                 wake_up = 1;
1689         } else if (obd->obd_recovery_expired) {
1690                 CDEBUG(D_HA, "waking for expired recovery\n");
1691                 wake_up = 1;
1692         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1693                 CDEBUG(D_HA, "waking for completed recovery\n");
1694                 wake_up = 1;
1695         } else if (req_transno == next_transno) {
1696                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1697                 wake_up = 1;
1698         } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1699                 int d_lvl = D_HA;
1700                 /** handle gaps occured due to lost reply or VBR */
1701                 LASSERTF(req_transno >= next_transno,
1702                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1703                          req_transno, next_transno);
1704                 if (req_transno > obd->obd_last_committed &&
1705                     !obd->obd_version_recov)
1706                         d_lvl = D_ERROR;
1707                 CDEBUG(d_lvl,
1708                        "%s: waking for gap in transno, VBR is %s (skip: "
1709                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1710                        ", last_committed: "LPD64")\n",
1711                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1712                        next_transno, queue_len, completed, connected,
1713                        req_transno, obd->obd_last_committed);
1714                 obd->obd_next_recovery_transno = req_transno;
1715                 wake_up = 1;
1716         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1717                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1718                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1719                 obd->obd_next_recovery_transno = req_transno;
1720                 wake_up = 1;
1721         }
1722         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1723         return wake_up;
1724 }
1725
1726 static int check_for_next_lock(struct obd_device *obd)
1727 {
1728         int wake_up = 0;
1729
1730         cfs_spin_lock(&obd->obd_recovery_task_lock);
1731         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1732                 CDEBUG(D_HA, "waking for next lock\n");
1733                 wake_up = 1;
1734         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1735                 CDEBUG(D_HA, "waking for completed lock replay\n");
1736                 wake_up = 1;
1737         } else if (obd->obd_abort_recovery) {
1738                 CDEBUG(D_HA, "waking for aborted recovery\n");
1739                 wake_up = 1;
1740         } else if (obd->obd_recovery_expired) {
1741                 CDEBUG(D_HA, "waking for expired recovery\n");
1742                 wake_up = 1;
1743         }
1744         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1745
1746         return wake_up;
1747 }
1748
1749 /**
1750  * wait for recovery events,
1751  * check its status with help of check_routine
1752  * evict dead clients via health_check
1753  */
1754 static int target_recovery_overseer(struct obd_device *obd,
1755                                     int (*check_routine)(struct obd_device *),
1756                                     int (*health_check)(struct obd_export *))
1757 {
1758 repeat:
1759         cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1760         if (obd->obd_abort_recovery) {
1761                 CWARN("recovery is aborted, evict exports in recovery\n");
1762                 /** evict exports which didn't finish recovery yet */
1763                 class_disconnect_stale_exports(obd, exp_finished);
1764                 return 1;
1765         } else if (obd->obd_recovery_expired) {
1766                 obd->obd_recovery_expired = 0;
1767                 /** If some clients died being recovered, evict them */
1768                 LCONSOLE_WARN("%s: recovery is timed out, "
1769                               "evict stale exports\n", obd->obd_name);
1770                 /** evict cexports with no replay in queue, they are stalled */
1771                 class_disconnect_stale_exports(obd, health_check);
1772                 /** continue with VBR */
1773                 cfs_spin_lock(&obd->obd_dev_lock);
1774                 obd->obd_version_recov = 1;
1775                 cfs_spin_unlock(&obd->obd_dev_lock);
1776                 /**
1777                  * reset timer, recovery will proceed with versions now,
1778                  * timeout is set just to handle reconnection delays
1779                  */
1780                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1781                 /** Wait for recovery events again, after evicting bad clients */
1782                 goto repeat;
1783         }
1784         return 0;
1785 }
1786
1787 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1788 {
1789         struct ptlrpc_request *req = NULL;
1790         ENTRY;
1791
1792         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1793                obd->obd_next_recovery_transno);
1794
1795         if (target_recovery_overseer(obd, check_for_next_transno,
1796                                      exp_req_replay_healthy)) {
1797                 abort_req_replay_queue(obd);
1798                 abort_lock_replay_queue(obd);
1799         }
1800
1801         cfs_spin_lock(&obd->obd_recovery_task_lock);
1802         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1803                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1804                                      struct ptlrpc_request, rq_list);
1805                 cfs_list_del_init(&req->rq_list);
1806                 obd->obd_requests_queued_for_recovery--;
1807                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1808         } else {
1809                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1810                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1811                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1812                 /** evict exports failed VBR */
1813                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1814         }
1815         RETURN(req);
1816 }
1817
1818 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1819 {
1820         struct ptlrpc_request *req = NULL;
1821
1822         CDEBUG(D_HA, "Waiting for lock\n");
1823         if (target_recovery_overseer(obd, check_for_next_lock,
1824                                      exp_lock_replay_healthy))
1825                 abort_lock_replay_queue(obd);
1826
1827         cfs_spin_lock(&obd->obd_recovery_task_lock);
1828         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1829                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1830                                      struct ptlrpc_request, rq_list);
1831                 cfs_list_del_init(&req->rq_list);
1832                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1833         } else {
1834                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1835                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1836                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1837                 /** evict exports failed VBR */
1838                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1839         }
1840         return req;
1841 }
1842
1843 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1844 {
1845         struct ptlrpc_request *req = NULL;
1846
1847         cfs_spin_lock(&obd->obd_recovery_task_lock);
1848         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1849                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1850                                      struct ptlrpc_request, rq_list);
1851                 cfs_list_del_init(&req->rq_list);
1852                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1853                 if (req->rq_export->exp_in_recovery) {
1854                         cfs_spin_lock(&req->rq_export->exp_lock);
1855                         req->rq_export->exp_in_recovery = 0;
1856                         cfs_spin_unlock(&req->rq_export->exp_lock);
1857                 }
1858         } else {
1859                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1860         }
1861         return req;
1862 }
1863
1864 static int handle_recovery_req(struct ptlrpc_thread *thread,
1865                                struct ptlrpc_request *req,
1866                                svc_handler_t handler)
1867 {
1868         int rc;
1869         ENTRY;
1870
1871         /**
1872          * export can be evicted during recovery, no need to handle replays for
1873          * it after that, discard such request silently
1874          */
1875         if (req->rq_export->exp_disconnected)
1876                 GOTO(reqcopy_put, rc = 0);
1877
1878         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1879         if (rc) {
1880                 CERROR("Failure to initialize session: %d\n", rc);
1881                 GOTO(reqcopy_put, rc);
1882         }
1883
1884         req->rq_recov_session.lc_thread = thread;
1885         lu_context_enter(&req->rq_recov_session);
1886         req->rq_svc_thread = thread;
1887         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1888
1889         /* thread context */
1890         lu_context_enter(&thread->t_env->le_ctx);
1891         (void)handler(req);
1892         lu_context_exit(&thread->t_env->le_ctx);
1893
1894         lu_context_exit(&req->rq_recov_session);
1895         lu_context_fini(&req->rq_recov_session);
1896         /* don't reset timer for final stage */
1897         if (!exp_finished(req->rq_export)) {
1898                 int to = obd_timeout;
1899
1900                 /**
1901                  * Add request timeout to the recovery time so next request from
1902                  * this client may come in recovery time
1903                  */
1904                 if (!AT_OFF) {
1905                         struct ptlrpc_service_part *svcpt;
1906
1907                         svcpt = req->rq_rqbd->rqbd_svcpt;
1908                         /* If the server sent early reply for this request,
1909                          * the client will recalculate the timeout according to
1910                          * current server estimate service time, so we will
1911                          * use the maxium timeout here for waiting the client
1912                          * sending the next req */
1913                         to = max((int)at_est2timeout(
1914                                  at_get(&svcpt->scp_at_estimate)),
1915                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1916                         /* Add net_latency (see ptlrpc_replay_req) */
1917                         to += lustre_msg_get_service_time(req->rq_reqmsg);
1918                 }
1919                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1920         }
1921 reqcopy_put:
1922         RETURN(rc);
1923 }
1924
1925 static int target_recovery_thread(void *arg)
1926 {
1927         struct lu_target *lut = arg;
1928         struct obd_device *obd = lut->lut_obd;
1929         struct ptlrpc_request *req;
1930         struct target_recovery_data *trd = &obd->obd_recovery_data;
1931         unsigned long delta;
1932         unsigned long flags;
1933         struct lu_env *env;
1934         struct ptlrpc_thread *thread = NULL;
1935         int rc = 0;
1936         ENTRY;
1937
1938         cfs_daemonize_ctxt("tgt_recov");
1939
1940         SIGNAL_MASK_LOCK(current, flags);
1941         sigfillset(&current->blocked);
1942         RECALC_SIGPENDING;
1943         SIGNAL_MASK_UNLOCK(current, flags);
1944
1945         OBD_ALLOC_PTR(thread);
1946         if (thread == NULL)
1947                 RETURN(-ENOMEM);
1948
1949         OBD_ALLOC_PTR(env);
1950         if (env == NULL) {
1951                 OBD_FREE_PTR(thread);
1952                 RETURN(-ENOMEM);
1953         }
1954
1955         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1956         if (rc) {
1957                 OBD_FREE_PTR(thread);
1958                 OBD_FREE_PTR(env);
1959                 RETURN(rc);
1960         }
1961
1962         thread->t_env = env;
1963         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1964         env->le_ctx.lc_thread = thread;
1965         thread->t_data = NULL;
1966         thread->t_watchdog = NULL;
1967
1968         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1969                cfs_curproc_pid());
1970         trd->trd_processing_task = cfs_curproc_pid();
1971
1972         cfs_spin_lock(&obd->obd_dev_lock);
1973         obd->obd_recovering = 1;
1974         cfs_spin_unlock(&obd->obd_dev_lock);
1975         cfs_complete(&trd->trd_starting);
1976
1977         /* first of all, we have to know the first transno to replay */
1978         if (target_recovery_overseer(obd, check_for_clients,
1979                                      exp_connect_healthy)) {
1980                 abort_req_replay_queue(obd);
1981                 abort_lock_replay_queue(obd);
1982         }
1983
1984         /* next stage: replay requests */
1985         delta = jiffies;
1986         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1987                cfs_atomic_read(&obd->obd_req_replay_clients),
1988                obd->obd_next_recovery_transno);
1989         while ((req = target_next_replay_req(obd))) {
1990                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1991                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1992                           lustre_msg_get_transno(req->rq_reqmsg),
1993                           libcfs_nid2str(req->rq_peer.nid));
1994                 handle_recovery_req(thread, req,
1995                                     trd->trd_recovery_handler);
1996                 /**
1997                  * bz18031: increase next_recovery_transno before
1998                  * target_request_copy_put() will drop exp_rpc reference
1999                  */
2000                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2001                 obd->obd_next_recovery_transno++;
2002                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2003                 target_exp_dequeue_req_replay(req);
2004                 target_request_copy_put(req);
2005                 obd->obd_replayed_requests++;
2006         }
2007
2008         /**
2009          * The second stage: replay locks
2010          */
2011         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2012                cfs_atomic_read(&obd->obd_lock_replay_clients));
2013         while ((req = target_next_replay_lock(obd))) {
2014                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2015                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2016                           libcfs_nid2str(req->rq_peer.nid));
2017                 handle_recovery_req(thread, req,
2018                                     trd->trd_recovery_handler);
2019                 target_request_copy_put(req);
2020                 obd->obd_replayed_locks++;
2021         }
2022
2023         /**
2024          * The third stage: reply on final pings, at this moment all clients
2025          * must have request in final queue
2026          */
2027         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2028         /** Update server last boot epoch */
2029         lut_boot_epoch_update(lut);
2030         /* We drop recoverying flag to forward all new requests
2031          * to regular mds_handle() since now */
2032         cfs_spin_lock(&obd->obd_dev_lock);
2033         obd->obd_recovering = obd->obd_abort_recovery = 0;
2034         cfs_spin_unlock(&obd->obd_dev_lock);
2035         cfs_spin_lock(&obd->obd_recovery_task_lock);
2036         target_cancel_recovery_timer(obd);
2037         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2038         while ((req = target_next_final_ping(obd))) {
2039                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2040                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2041                           libcfs_nid2str(req->rq_peer.nid));
2042                 handle_recovery_req(thread, req,
2043                                     trd->trd_recovery_handler);
2044                 target_request_copy_put(req);
2045         }
2046
2047         delta = (jiffies - delta) / CFS_HZ;
2048         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2049               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2050         if (delta > OBD_RECOVERY_TIME_SOFT) {
2051                 CWARN("too long recovery - read logs\n");
2052                 libcfs_debug_dumplog();
2053         }
2054
2055         target_finish_recovery(obd);
2056
2057         lu_context_fini(&env->le_ctx);
2058         trd->trd_processing_task = 0;
2059         cfs_complete(&trd->trd_finishing);
2060
2061         OBD_FREE_PTR(thread);
2062         OBD_FREE_PTR(env);
2063         RETURN(rc);
2064 }
2065
2066 static int target_start_recovery_thread(struct lu_target *lut,
2067                                         svc_handler_t handler)
2068 {
2069         struct obd_device *obd = lut->lut_obd;
2070         int rc = 0;
2071         struct target_recovery_data *trd = &obd->obd_recovery_data;
2072
2073         memset(trd, 0, sizeof(*trd));
2074         cfs_init_completion(&trd->trd_starting);
2075         cfs_init_completion(&trd->trd_finishing);
2076         trd->trd_recovery_handler = handler;
2077
2078         if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
2079                 cfs_wait_for_completion(&trd->trd_starting);
2080                 LASSERT(obd->obd_recovering != 0);
2081         } else
2082                 rc = -ECHILD;
2083
2084         return rc;
2085 }
2086
2087 void target_stop_recovery_thread(struct obd_device *obd)
2088 {
2089         if (obd->obd_recovery_data.trd_processing_task > 0) {
2090                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2091                 /** recovery can be done but postrecovery is not yet */
2092                 cfs_spin_lock(&obd->obd_dev_lock);
2093                 if (obd->obd_recovering) {
2094                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2095                         obd->obd_abort_recovery = 1;
2096                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2097                 }
2098                 cfs_spin_unlock(&obd->obd_dev_lock);
2099                 cfs_wait_for_completion(&trd->trd_finishing);
2100         }
2101 }
2102
2103 void target_recovery_fini(struct obd_device *obd)
2104 {
2105         class_disconnect_exports(obd);
2106         target_stop_recovery_thread(obd);
2107         target_cleanup_recovery(obd);
2108 }
2109 EXPORT_SYMBOL(target_recovery_fini);
2110
2111 static void target_recovery_expired(unsigned long castmeharder)
2112 {
2113         struct obd_device *obd = (struct obd_device *)castmeharder;
2114         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2115                " after %lds (%d clients connected)\n",
2116                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2117                cfs_time_current_sec()- obd->obd_recovery_start,
2118                cfs_atomic_read(&obd->obd_connected_clients));
2119
2120         obd->obd_recovery_expired = 1;
2121         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2122 }
2123
2124 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2125 {
2126         struct obd_device *obd = lut->lut_obd;
2127         if (obd->obd_max_recoverable_clients == 0) {
2128                 /** Update server last boot epoch */
2129                 lut_boot_epoch_update(lut);
2130                 return;
2131         }
2132
2133         CWARN("RECOVERY: service %s, %d recoverable clients, "
2134               "last_transno "LPU64"\n", obd->obd_name,
2135               obd->obd_max_recoverable_clients, obd->obd_last_committed);
2136         LASSERT(obd->obd_stopping == 0);
2137         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2138         obd->obd_recovery_start = 0;
2139         obd->obd_recovery_end = 0;
2140
2141         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2142         target_start_recovery_thread(lut, handler);
2143 }
2144 EXPORT_SYMBOL(target_recovery_init);
2145
2146 #endif /* __KERNEL__ */
2147
2148 static int target_process_req_flags(struct obd_device *obd,
2149                                     struct ptlrpc_request *req)
2150 {
2151         struct obd_export *exp = req->rq_export;
2152         LASSERT(exp != NULL);
2153         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2154                 /* client declares he's ready to replay locks */
2155                 cfs_spin_lock(&exp->exp_lock);
2156                 if (exp->exp_req_replay_needed) {
2157                         exp->exp_req_replay_needed = 0;
2158                         cfs_spin_unlock(&exp->exp_lock);
2159
2160                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2161                         cfs_atomic_dec(&obd->obd_req_replay_clients);
2162                 } else {
2163                         cfs_spin_unlock(&exp->exp_lock);
2164                 }
2165         }
2166         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2167                 /* client declares he's ready to complete recovery
2168                  * so, we put the request on th final queue */
2169                 cfs_spin_lock(&exp->exp_lock);
2170                 if (exp->exp_lock_replay_needed) {
2171                         exp->exp_lock_replay_needed = 0;
2172                         cfs_spin_unlock(&exp->exp_lock);
2173
2174                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2175                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
2176                 } else {
2177                         cfs_spin_unlock(&exp->exp_lock);
2178                 }
2179         }
2180         return 0;
2181 }
2182
2183 int target_queue_recovery_request(struct ptlrpc_request *req,
2184                                   struct obd_device *obd)
2185 {
2186         cfs_list_t *tmp;
2187         int inserted = 0;
2188         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2189         ENTRY;
2190
2191         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2192                 /* Processing the queue right now, don't re-add. */
2193                 RETURN(1);
2194         }
2195
2196         target_process_req_flags(obd, req);
2197
2198         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2199                 /* client declares he's ready to complete recovery
2200                  * so, we put the request on th final queue */
2201                 target_request_copy_get(req);
2202                 DEBUG_REQ(D_HA, req, "queue final req");
2203                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2204                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2205                 if (obd->obd_recovering) {
2206                         cfs_list_add_tail(&req->rq_list,
2207                                           &obd->obd_final_req_queue);
2208                 } else {
2209                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2210                         target_request_copy_put(req);
2211                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2212                 }
2213                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2214                 RETURN(0);
2215         }
2216         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2217                 /* client declares he's ready to replay locks */
2218                 target_request_copy_get(req);
2219                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2220                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2221                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2222                 LASSERT(obd->obd_recovering);
2223                 /* usually due to recovery abort */
2224                 if (!req->rq_export->exp_in_recovery) {
2225                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2226                         target_request_copy_put(req);
2227                         RETURN(-ENOTCONN);
2228                 }
2229                 LASSERT(req->rq_export->exp_lock_replay_needed);
2230                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2231                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2232                 RETURN(0);
2233         }
2234
2235         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2236          * (i.e. buflens etc are in my own byte order), but type-dependent
2237          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2238
2239         if (!transno) {
2240                 CFS_INIT_LIST_HEAD(&req->rq_list);
2241                 DEBUG_REQ(D_HA, req, "not queueing");
2242                 RETURN(1);
2243         }
2244
2245         /* If we're processing the queue, we want don't want to queue this
2246          * message.
2247          *
2248          * Also, if this request has a transno less than the one we're waiting
2249          * for, we should process it now.  It could (and currently always will)
2250          * be an open request for a descriptor that was opened some time ago.
2251          *
2252          * Also, a resent, replayed request that has already been
2253          * handled will pass through here and be processed immediately.
2254          */
2255         CDEBUG(D_HA, "Next recovery transno: "LPU64
2256                ", current: "LPU64", replaying\n",
2257                obd->obd_next_recovery_transno, transno);
2258         cfs_spin_lock(&obd->obd_recovery_task_lock);
2259         if (transno < obd->obd_next_recovery_transno) {
2260                 /* Processing the queue right now, don't re-add. */
2261                 LASSERT(cfs_list_empty(&req->rq_list));
2262                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2263                 RETURN(1);
2264         }
2265         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2266
2267         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2268                 RETURN(0);
2269
2270         target_request_copy_get(req);
2271         if (!req->rq_export->exp_in_recovery) {
2272                 target_request_copy_put(req);
2273                 RETURN(-ENOTCONN);
2274         }
2275         LASSERT(req->rq_export->exp_req_replay_needed);
2276
2277         if (target_exp_enqueue_req_replay(req)) {
2278                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2279                 target_request_copy_put(req);
2280                 RETURN(0);
2281         }
2282
2283         /* XXX O(n^2) */
2284         cfs_spin_lock(&obd->obd_recovery_task_lock);
2285         LASSERT(obd->obd_recovering);
2286         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2287                 struct ptlrpc_request *reqiter =
2288                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2289
2290                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2291                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2292                         inserted = 1;
2293                         break;
2294                 }
2295
2296                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2297                              transno)) {
2298                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2299                                   "has been claimed by another client");
2300                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2301                         target_exp_dequeue_req_replay(req);
2302                         target_request_copy_put(req);
2303                         RETURN(0);
2304                 }
2305         }
2306
2307         if (!inserted)
2308                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2309
2310         obd->obd_requests_queued_for_recovery++;
2311         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2312         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2313         RETURN(0);
2314 }
2315
2316 int target_handle_ping(struct ptlrpc_request *req)
2317 {
2318         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2319         return req_capsule_server_pack(&req->rq_pill);
2320 }
2321
2322 void target_committed_to_req(struct ptlrpc_request *req)
2323 {
2324         struct obd_export *exp = req->rq_export;
2325
2326         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2327                 lustre_msg_set_last_committed(req->rq_repmsg,
2328                                               exp->exp_last_committed);
2329         else
2330                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2331                           "%d)", exp->exp_obd->obd_no_transno,
2332                           req->rq_repmsg == NULL);
2333
2334         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2335                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2336 }
2337 EXPORT_SYMBOL(target_committed_to_req);
2338
2339 #endif /* HAVE_SERVER_SUPPORT */
2340
2341 /**
2342  * Packs current SLV and Limit into \a req.
2343  */
2344 int target_pack_pool_reply(struct ptlrpc_request *req)
2345 {
2346         struct obd_device *obd;
2347         ENTRY;
2348
2349         /*
2350          * Check that we still have all structures alive as this may
2351          * be some late rpc in shutdown time.
2352          */
2353         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2354                      !exp_connect_lru_resize(req->rq_export))) {
2355                 lustre_msg_set_slv(req->rq_repmsg, 0);
2356                 lustre_msg_set_limit(req->rq_repmsg, 0);
2357                 RETURN(0);
2358         }
2359
2360         /*
2361          * OBD is alive here as export is alive, which we checked above.
2362          */
2363         obd = req->rq_export->exp_obd;
2364
2365         cfs_read_lock(&obd->obd_pool_lock);
2366         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2367         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2368         cfs_read_unlock(&obd->obd_pool_lock);
2369
2370         RETURN(0);
2371 }
2372
2373 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2374 {
2375         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2376                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2377                 return (-ECOMM);
2378         }
2379
2380         if (unlikely(rc)) {
2381                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2382                 req->rq_status = rc;
2383                 return (ptlrpc_send_error(req, 1));
2384         } else {
2385                 DEBUG_REQ(D_NET, req, "sending reply");
2386         }
2387
2388         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2389 }
2390
2391 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2392 {
2393         struct ptlrpc_service_part *svcpt;
2394         int                        netrc;
2395         struct ptlrpc_reply_state *rs;
2396         struct obd_export         *exp;
2397         ENTRY;
2398
2399         if (req->rq_no_reply) {
2400                 EXIT;
2401                 return;
2402         }
2403
2404         svcpt = req->rq_rqbd->rqbd_svcpt;
2405         rs = req->rq_reply_state;
2406         if (rs == NULL || !rs->rs_difficult) {
2407                 /* no notifiers */
2408                 target_send_reply_msg (req, rc, fail_id);
2409                 EXIT;
2410                 return;
2411         }
2412
2413         /* must be an export if locks saved */
2414         LASSERT (req->rq_export != NULL);
2415         /* req/reply consistent */
2416         LASSERT(rs->rs_svcpt == svcpt);
2417
2418         /* "fresh" reply */
2419         LASSERT (!rs->rs_scheduled);
2420         LASSERT (!rs->rs_scheduled_ever);
2421         LASSERT (!rs->rs_handled);
2422         LASSERT (!rs->rs_on_net);
2423         LASSERT (rs->rs_export == NULL);
2424         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2425         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2426
2427         exp = class_export_get (req->rq_export);
2428
2429         /* disable reply scheduling while I'm setting up */
2430         rs->rs_scheduled = 1;
2431         rs->rs_on_net    = 1;
2432         rs->rs_xid       = req->rq_xid;
2433         rs->rs_transno   = req->rq_transno;
2434         rs->rs_export    = exp;
2435         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2436
2437         cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2438         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2439                rs->rs_transno, exp->exp_last_committed);
2440         if (rs->rs_transno > exp->exp_last_committed) {
2441                 /* not committed already */
2442                 cfs_list_add_tail(&rs->rs_obd_list,
2443                                   &exp->exp_uncommitted_replies);
2444         }
2445         cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2446
2447         cfs_spin_lock(&exp->exp_lock);
2448         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2449         cfs_spin_unlock(&exp->exp_lock);
2450
2451         netrc = target_send_reply_msg (req, rc, fail_id);
2452
2453         cfs_spin_lock(&svcpt->scp_rep_lock);
2454
2455         cfs_atomic_inc(&svcpt->scp_nreps_difficult);
2456
2457         if (netrc != 0) {
2458                 /* error sending: reply is off the net.  Also we need +1
2459                  * reply ref until ptlrpc_handle_rs() is done
2460                  * with the reply state (if the send was successful, there
2461                  * would have been +1 ref for the net, which
2462                  * reply_out_callback leaves alone) */
2463                 rs->rs_on_net = 0;
2464                 ptlrpc_rs_addref(rs);
2465         }
2466
2467         cfs_spin_lock(&rs->rs_lock);
2468         if (rs->rs_transno <= exp->exp_last_committed ||
2469             (!rs->rs_on_net && !rs->rs_no_ack) ||
2470              cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2471              cfs_list_empty(&rs->rs_obd_list)) {
2472                 CDEBUG(D_HA, "Schedule reply immediately\n");
2473                 ptlrpc_dispatch_difficult_reply(rs);
2474         } else {
2475                 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
2476                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2477         }
2478         cfs_spin_unlock(&rs->rs_lock);
2479         cfs_spin_unlock(&svcpt->scp_rep_lock);
2480         EXIT;
2481 }
2482
2483 int target_handle_qc_callback(struct ptlrpc_request *req)
2484 {
2485         struct obd_quotactl *oqctl;
2486         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2487
2488         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2489         if (oqctl == NULL) {
2490                 CERROR("Can't unpack obd_quotactl\n");
2491                 RETURN(-EPROTO);
2492         }
2493
2494         cli->cl_qchk_stat = oqctl->qc_stat;
2495
2496         return 0;
2497 }
2498
2499 #ifdef HAVE_QUOTA_SUPPORT
2500 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2501 {
2502 #ifdef __KERNEL__
2503         struct obd_device *obd = req->rq_export->exp_obd;
2504         struct obd_device *master_obd = NULL, *lov_obd = NULL;
2505         struct obd_device_target *obt;
2506         struct lustre_quota_ctxt *qctxt;
2507         struct qunit_data *qdata = NULL;
2508         int rc = 0;
2509         ENTRY;
2510
2511         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2512                 RETURN(rc);
2513
2514         rc = req_capsule_server_pack(&req->rq_pill);
2515         if (rc) {
2516                 CERROR("packing reply failed!: rc = %d\n", rc);
2517                 RETURN(rc);
2518         }
2519
2520         LASSERT(req->rq_export);
2521
2522         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2523         if (IS_ERR(qdata)) {
2524                 rc = PTR_ERR(qdata);
2525                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2526                 req->rq_status = rc;
2527                 GOTO(out, rc);
2528         }
2529
2530         /* we use the observer */
2531         if (obd_pin_observer(obd, &lov_obd) ||
2532             obd_pin_observer(lov_obd, &master_obd)) {
2533                 CERROR("Can't find the observer, it is recovering\n");
2534                 req->rq_status = -EAGAIN;
2535                 GOTO(out, rc);
2536         }
2537
2538         obt = &master_obd->u.obt;
2539         qctxt = &obt->obt_qctxt;
2540
2541         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2542                 /* quota_type has not been processed yet, return EAGAIN
2543                  * until we know whether or not quotas are supposed to
2544                  * be enabled */
2545                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2546                        "-EAGAIN\n");
2547                 req->rq_status = -EAGAIN;
2548                 GOTO(out, rc);
2549         }
2550
2551         cfs_down_read(&obt->obt_rwsem);
2552         if (qctxt->lqc_lqs_hash == NULL) {
2553                 cfs_up_read(&obt->obt_rwsem);
2554                 /* quota_type has not been processed yet, return EAGAIN
2555                  * until we know whether or not quotas are supposed to
2556                  * be enabled */
2557                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2558                        "-EAGAIN\n");
2559                 req->rq_status = -EAGAIN;
2560                 GOTO(out, rc);
2561         }
2562
2563         LASSERT(qctxt->lqc_handler);
2564         rc = qctxt->lqc_handler(master_obd, qdata,
2565                                 lustre_msg_get_opc(req->rq_reqmsg));
2566         cfs_up_read(&obt->obt_rwsem);
2567         if (rc && rc != -EDQUOT)
2568                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2569                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2570         req->rq_status = rc;
2571
2572         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2573         if (rc < 0) {
2574                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2575                 GOTO(out, rc);
2576         }
2577
2578         /* Block the quota req. b=14840 */
2579         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2580         EXIT;
2581
2582 out:
2583         if (master_obd)
2584                 obd_unpin_observer(lov_obd);
2585         if (lov_obd)
2586                 obd_unpin_observer(obd);
2587
2588         rc = ptlrpc_reply(req);
2589         return rc;
2590 #else
2591         return 0;
2592 #endif /* !__KERNEL__ */
2593 }
2594 #endif /* HAVE_QUOTA_SUPPORT */
2595
2596 ldlm_mode_t lck_compat_array[] = {
2597         [LCK_EX] LCK_COMPAT_EX,
2598         [LCK_PW] LCK_COMPAT_PW,
2599         [LCK_PR] LCK_COMPAT_PR,
2600         [LCK_CW] LCK_COMPAT_CW,
2601         [LCK_CR] LCK_COMPAT_CR,
2602         [LCK_NL] LCK_COMPAT_NL,
2603         [LCK_GROUP] LCK_COMPAT_GROUP,
2604         [LCK_COS] LCK_COMPAT_COS,
2605 };
2606
2607 /**
2608  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2609  * not escape to the user level.
2610  */
2611 int ldlm_error2errno(ldlm_error_t error)
2612 {
2613         int result;
2614
2615         switch (error) {
2616         case ELDLM_OK:
2617                 result = 0;
2618                 break;
2619         case ELDLM_LOCK_CHANGED:
2620                 result = -ESTALE;
2621                 break;
2622         case ELDLM_LOCK_ABORTED:
2623                 result = -ENAVAIL;
2624                 break;
2625         case ELDLM_LOCK_REPLACED:
2626                 result = -ESRCH;
2627                 break;
2628         case ELDLM_NO_LOCK_DATA:
2629                 result = -ENOENT;
2630                 break;
2631         case ELDLM_NAMESPACE_EXISTS:
2632                 result = -EEXIST;
2633                 break;
2634         case ELDLM_BAD_NAMESPACE:
2635                 result = -EBADF;
2636                 break;
2637         default:
2638                 if (((int)error) < 0)  /* cast to signed type */
2639                         result = error; /* as ldlm_error_t can be unsigned */
2640                 else {
2641                         CERROR("Invalid DLM result code: %d\n", error);
2642                         result = -EPROTO;
2643                 }
2644         }
2645         return result;
2646 }
2647 EXPORT_SYMBOL(ldlm_error2errno);
2648
2649 /**
2650  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2651  */
2652 ldlm_error_t ldlm_errno2error(int err_no)
2653 {
2654         int error;
2655
2656         switch (err_no) {
2657         case 0:
2658                 error = ELDLM_OK;
2659                 break;
2660         case -ESTALE:
2661                 error = ELDLM_LOCK_CHANGED;
2662                 break;
2663         case -ENAVAIL:
2664                 error = ELDLM_LOCK_ABORTED;
2665                 break;
2666         case -ESRCH:
2667                 error = ELDLM_LOCK_REPLACED;
2668                 break;
2669         case -ENOENT:
2670                 error = ELDLM_NO_LOCK_DATA;
2671                 break;
2672         case -EEXIST:
2673                 error = ELDLM_NAMESPACE_EXISTS;
2674                 break;
2675         case -EBADF:
2676                 error = ELDLM_BAD_NAMESPACE;
2677                 break;
2678         default:
2679                 error = err_no;
2680         }
2681         return error;
2682 }
2683 EXPORT_SYMBOL(ldlm_errno2error);
2684
2685 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2686 void ldlm_dump_export_locks(struct obd_export *exp)
2687 {
2688         cfs_spin_lock(&exp->exp_locks_list_guard);
2689         if (!cfs_list_empty(&exp->exp_locks_list)) {
2690             struct ldlm_lock *lock;
2691
2692             CERROR("dumping locks for export %p,"
2693                    "ignore if the unmount doesn't hang\n", exp);
2694             cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2695                 LDLM_ERROR(lock, "lock:");
2696         }
2697         cfs_spin_unlock(&exp->exp_locks_list_guard);
2698 }
2699 #endif
2700
2701 #ifdef HAVE_SERVER_SUPPORT
2702 static int target_bulk_timeout(void *data)
2703 {
2704         ENTRY;
2705         /* We don't fail the connection here, because having the export
2706          * killed makes the (vital) call to commitrw very sad.
2707          */
2708         RETURN(1);
2709 }
2710
2711 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2712 {
2713         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2714 }
2715
2716 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2717                    struct l_wait_info *lwi)
2718 {
2719         struct ptlrpc_request *req = desc->bd_req;
2720         int rc = 0;
2721         ENTRY;
2722
2723         /* Check if there is eviction in progress, and if so, wait for
2724          * it to finish */
2725         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2726                 *lwi = LWI_INTR(NULL, NULL);
2727                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2728                                   !cfs_atomic_read(&exp->exp_obd->
2729                                                    obd_evict_inprogress),
2730                                   lwi);
2731         }
2732
2733         /* Check if client was evicted or tried to reconnect already */
2734         if (exp->exp_failed || exp->exp_abort_active_req) {
2735                 rc = -ENOTCONN;
2736         } else {
2737                 if (desc->bd_type == BULK_PUT_SINK)
2738                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2739                 if (rc == 0)
2740                         rc = ptlrpc_start_bulk_transfer(desc);
2741         }
2742
2743         if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2744                 ptlrpc_abort_bulk(desc);
2745         } else if (rc == 0) {
2746                 time_t start = cfs_time_current_sec();
2747                 do {
2748                         long timeoutl = req->rq_deadline - cfs_time_current_sec();
2749                         cfs_duration_t timeout = timeoutl <= 0 ?
2750                                 CFS_TICK : cfs_time_seconds(timeoutl);
2751                         *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2752                                                     cfs_time_seconds(1),
2753                                                    target_bulk_timeout,
2754                                                    desc);
2755                         rc = l_wait_event(desc->bd_waitq,
2756                                           !ptlrpc_server_bulk_active(desc) ||
2757                                           exp->exp_failed ||
2758                                           exp->exp_abort_active_req,
2759                                           lwi);
2760                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
2761                         /* Wait again if we changed deadline */
2762                 } while ((rc == -ETIMEDOUT) &&
2763                          (req->rq_deadline > cfs_time_current_sec()));
2764
2765                 if (rc == -ETIMEDOUT) {
2766                         DEBUG_REQ(D_ERROR, req,
2767                                   "timeout on bulk %s after %ld%+lds",
2768                                   bulk2type(desc),
2769                                   req->rq_deadline - start,
2770                                   cfs_time_current_sec() -
2771                                   req->rq_deadline);
2772                         ptlrpc_abort_bulk(desc);
2773                 } else if (exp->exp_failed) {
2774                         DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2775                                   bulk2type(desc));
2776                         rc = -ENOTCONN;
2777                         ptlrpc_abort_bulk(desc);
2778                 } else if (exp->exp_abort_active_req) {
2779                         DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2780                                   bulk2type(desc));
2781                         /* we don't reply anyway */
2782                         rc = -ETIMEDOUT;
2783                         ptlrpc_abort_bulk(desc);
2784                 } else if (!desc->bd_success ||
2785                            desc->bd_nob_transferred != desc->bd_nob) {
2786                         DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2787                                   desc->bd_success ?
2788                                   "truncated" : "network error on",
2789                                   bulk2type(desc),
2790                                   desc->bd_nob_transferred,
2791                                   desc->bd_nob);
2792                         /* XXX should this be a different errno? */
2793                         rc = -ETIMEDOUT;
2794                 } else if (desc->bd_type == BULK_GET_SINK) {
2795                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
2796                 }
2797         } else {
2798                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2799                           bulk2type(desc), rc);
2800         }
2801
2802         RETURN(rc);
2803 }
2804 EXPORT_SYMBOL(target_bulk_io);
2805
2806 #endif /* HAVE_SERVER_SUPPORT */