Whamcloud - gitweb
facc9758789c845e1238fae0da77e73f323076c6
[fs/lustre-release.git] / lustre / ldlm / ldlm_lib.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2013, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  */
36
37 /**
38  * This file deals with various client/target related logic including recovery.
39  *
40  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
41  * should be moved.
42  */
43
44 #define DEBUG_SUBSYSTEM S_LDLM
45
46 #ifdef __KERNEL__
47 # include <libcfs/libcfs.h>
48 #else
49 # include <liblustre.h>
50 #endif
51 #include <obd.h>
52 #include <obd_class.h>
53 #include <lustre_dlm.h>
54 #include <lustre_net.h>
55 #include <lustre_sec.h>
56 #include "ldlm_internal.h"
57
58 /* @priority: If non-zero, move the selected connection to the list head.
59  * @create: If zero, only search in existing connections.
60  */
61 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
62                            int priority, int create)
63 {
64         struct ptlrpc_connection *ptlrpc_conn;
65         struct obd_import_conn *imp_conn = NULL, *item;
66         int rc = 0;
67         ENTRY;
68
69         if (!create && !priority) {
70                 CDEBUG(D_HA, "Nothing to do\n");
71                 RETURN(-EINVAL);
72         }
73
74         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
75         if (!ptlrpc_conn) {
76                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77                 RETURN (-ENOENT);
78         }
79
80         if (create) {
81                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
82                 if (!imp_conn) {
83                         GOTO(out_put, rc = -ENOMEM);
84                 }
85         }
86
87         spin_lock(&imp->imp_lock);
88         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
89                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
90                         if (priority) {
91                                 cfs_list_del(&item->oic_item);
92                                 cfs_list_add(&item->oic_item,
93                                              &imp->imp_conn_list);
94                                 item->oic_last_attempt = 0;
95                         }
96                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
97                                imp, imp->imp_obd->obd_name, uuid->uuid,
98                                (priority ? ", moved to head" : ""));
99                         spin_unlock(&imp->imp_lock);
100                         GOTO(out_free, rc = 0);
101                 }
102         }
103         /* No existing import connection found for \a uuid. */
104         if (create) {
105                 imp_conn->oic_conn = ptlrpc_conn;
106                 imp_conn->oic_uuid = *uuid;
107                 imp_conn->oic_last_attempt = 0;
108                 if (priority)
109                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
110                 else
111                         cfs_list_add_tail(&imp_conn->oic_item,
112                                           &imp->imp_conn_list);
113                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
114                        imp, imp->imp_obd->obd_name, uuid->uuid,
115                        (priority ? "head" : "tail"));
116         } else {
117                 spin_unlock(&imp->imp_lock);
118                 GOTO(out_free, rc = -ENOENT);
119         }
120
121         spin_unlock(&imp->imp_lock);
122         RETURN(0);
123 out_free:
124         if (imp_conn)
125                 OBD_FREE(imp_conn, sizeof(*imp_conn));
126 out_put:
127         ptlrpc_connection_put(ptlrpc_conn);
128         RETURN(rc);
129 }
130
131 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
132 {
133         return import_set_conn(imp, uuid, 1, 0);
134 }
135
136 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
137                            int priority)
138 {
139         return import_set_conn(imp, uuid, priority, 1);
140 }
141 EXPORT_SYMBOL(client_import_add_conn);
142
143 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
144 {
145         struct obd_import_conn *imp_conn;
146         struct obd_export *dlmexp;
147         int rc = -ENOENT;
148         ENTRY;
149
150         spin_lock(&imp->imp_lock);
151         if (cfs_list_empty(&imp->imp_conn_list)) {
152                 LASSERT(!imp->imp_connection);
153                 GOTO(out, rc);
154         }
155
156         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
157                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
158                         continue;
159                 LASSERT(imp_conn->oic_conn);
160
161                 if (imp_conn == imp->imp_conn_current) {
162                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
163
164                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
165                             imp->imp_state != LUSTRE_IMP_DISCON) {
166                                 CERROR("can't remove current connection\n");
167                                 GOTO(out, rc = -EBUSY);
168                         }
169
170                         ptlrpc_connection_put(imp->imp_connection);
171                         imp->imp_connection = NULL;
172
173                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
174                         if (dlmexp && dlmexp->exp_connection) {
175                                 LASSERT(dlmexp->exp_connection ==
176                                         imp_conn->oic_conn);
177                                 ptlrpc_connection_put(dlmexp->exp_connection);
178                                 dlmexp->exp_connection = NULL;
179                         }
180                 }
181
182                 cfs_list_del(&imp_conn->oic_item);
183                 ptlrpc_connection_put(imp_conn->oic_conn);
184                 OBD_FREE(imp_conn, sizeof(*imp_conn));
185                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
186                        imp, imp->imp_obd->obd_name, uuid->uuid);
187                 rc = 0;
188                 break;
189         }
190 out:
191         spin_unlock(&imp->imp_lock);
192         if (rc == -ENOENT)
193                 CERROR("connection %s not found\n", uuid->uuid);
194         RETURN(rc);
195 }
196 EXPORT_SYMBOL(client_import_del_conn);
197
198 /**
199  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
200  * to find a conn uuid of \a imp which can reach \a peer.
201  */
202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
203                             struct obd_uuid *uuid)
204 {
205         struct obd_import_conn *conn;
206         int rc = -ENOENT;
207         ENTRY;
208
209         spin_lock(&imp->imp_lock);
210         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
211                 /* Check if conn UUID does have this peer NID. */
212                 if (class_check_uuid(&conn->oic_uuid, peer)) {
213                         *uuid = conn->oic_uuid;
214                         rc = 0;
215                         break;
216                 }
217         }
218         spin_unlock(&imp->imp_lock);
219         RETURN(rc);
220 }
221 EXPORT_SYMBOL(client_import_find_conn);
222
223 void client_destroy_import(struct obd_import *imp)
224 {
225         /* Drop security policy instance after all RPCs have finished/aborted
226          * to let all busy contexts be released. */
227         class_import_get(imp);
228         class_destroy_import(imp);
229         sptlrpc_import_sec_put(imp);
230         class_import_put(imp);
231 }
232 EXPORT_SYMBOL(client_destroy_import);
233
234 /**
235  * Check whether or not the OSC is on MDT.
236  * In the config log,
237  * osc on MDT
238  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
239  * osc on client
240  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
241  *
242  **/
243 static int osc_on_mdt(char *obdname)
244 {
245         char *ptr;
246
247         ptr = strrchr(obdname, '-');
248         if (ptr == NULL)
249                 return 0;
250
251         if (strncmp(ptr + 1, "MDT", 3) == 0)
252                 return 1;
253
254         return 0;
255 }
256
257 /* Configure an RPC client OBD device.
258  *
259  * lcfg parameters:
260  * 1 - client UUID
261  * 2 - server UUID
262  * 3 - inactive-on-startup
263  */
264 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
265 {
266         struct client_obd *cli = &obddev->u.cli;
267         struct obd_import *imp;
268         struct obd_uuid server_uuid;
269         int rq_portal, rp_portal, connect_op;
270         char *name = obddev->obd_type->typ_name;
271         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
272         int rc;
273         char    *cli_name = lustre_cfg_buf(lcfg, 0);
274         ENTRY;
275
276         /* In a more perfect world, we would hang a ptlrpc_client off of
277          * obd_type and just use the values from there. */
278         if (!strcmp(name, LUSTRE_OSC_NAME) ||
279             (!(strcmp(name, LUSTRE_OSP_NAME)) &&
280              (is_osp_on_mdt(cli_name) &&
281                strstr(lustre_cfg_buf(lcfg, 1), "OST") != NULL))) {
282                 /* OSC or OSP_on_MDT for OSTs */
283                 rq_portal = OST_REQUEST_PORTAL;
284                 rp_portal = OSC_REPLY_PORTAL;
285                 connect_op = OST_CONNECT;
286                 cli->cl_sp_me = LUSTRE_SP_CLI;
287                 cli->cl_sp_to = LUSTRE_SP_OST;
288                 ns_type = LDLM_NS_TYPE_OSC;
289         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
290                    !strcmp(name, LUSTRE_LWP_NAME) ||
291                    (!strcmp(name, LUSTRE_OSP_NAME) &&
292                     (is_osp_on_mdt(cli_name) &&
293                      strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL))) {
294                 /* MDC or OSP_on_MDT for other MDTs */
295                 rq_portal = MDS_REQUEST_PORTAL;
296                 rp_portal = MDC_REPLY_PORTAL;
297                 connect_op = MDS_CONNECT;
298                 cli->cl_sp_me = LUSTRE_SP_CLI;
299                 cli->cl_sp_to = LUSTRE_SP_MDT;
300                 ns_type = LDLM_NS_TYPE_MDC;
301         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
302                 rq_portal = MGS_REQUEST_PORTAL;
303                 rp_portal = MGC_REPLY_PORTAL;
304                 connect_op = MGS_CONNECT;
305                 cli->cl_sp_me = LUSTRE_SP_MGC;
306                 cli->cl_sp_to = LUSTRE_SP_MGS;
307                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
308                 ns_type = LDLM_NS_TYPE_MGC;
309         } else {
310                 CERROR("unknown client OBD type \"%s\", can't setup\n",
311                        name);
312                 RETURN(-EINVAL);
313         }
314
315         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
316                 CERROR("requires a TARGET UUID\n");
317                 RETURN(-EINVAL);
318         }
319
320         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
321                 CERROR("client UUID must be less than 38 characters\n");
322                 RETURN(-EINVAL);
323         }
324
325         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
326                 CERROR("setup requires a SERVER UUID\n");
327                 RETURN(-EINVAL);
328         }
329
330         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
331                 CERROR("target UUID must be less than 38 characters\n");
332                 RETURN(-EINVAL);
333         }
334
335         init_rwsem(&cli->cl_sem);
336         sema_init(&cli->cl_mgc_sem, 1);
337         cli->cl_conn_count = 0;
338         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
339                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
340                      sizeof(server_uuid)));
341
342         cli->cl_dirty = 0;
343         cli->cl_avail_grant = 0;
344         /* FIXME: Should limit this for the sum of all cl_dirty_max. */
345         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
346         if (cli->cl_dirty_max >> PAGE_CACHE_SHIFT > totalram_pages / 8)
347                 cli->cl_dirty_max = totalram_pages << (PAGE_CACHE_SHIFT - 3);
348         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
349         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
350         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
351         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
352         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
353         client_obd_list_lock_init(&cli->cl_loi_list_lock);
354         cfs_atomic_set(&cli->cl_pending_w_pages, 0);
355         cfs_atomic_set(&cli->cl_pending_r_pages, 0);
356         cli->cl_r_in_flight = 0;
357         cli->cl_w_in_flight = 0;
358
359         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
360         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
361         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
362         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
363         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
364         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
365
366         /* lru for osc. */
367         CFS_INIT_LIST_HEAD(&cli->cl_lru_osc);
368         cfs_atomic_set(&cli->cl_lru_shrinkers, 0);
369         cfs_atomic_set(&cli->cl_lru_busy, 0);
370         cfs_atomic_set(&cli->cl_lru_in_list, 0);
371         CFS_INIT_LIST_HEAD(&cli->cl_lru_list);
372         client_obd_list_lock_init(&cli->cl_lru_list_lock);
373         cfs_atomic_set(&cli->cl_unstable_count, 0);
374
375         init_waitqueue_head(&cli->cl_destroy_waitq);
376         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
377 #ifdef ENABLE_CHECKSUM
378         /* Turn on checksumming by default. */
379         cli->cl_checksum = 1;
380         /*
381          * The supported checksum types will be worked out at connect time
382          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
383          * through procfs.
384          */
385         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
386 #endif
387         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
388
389         /* This value may be reduced at connect time in
390          * ptlrpc_connect_interpret() . We initialize it to only
391          * 1MB until we know what the performance looks like.
392          * In the future this should likely be increased. LU-1431 */
393         cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
394                                           LNET_MTU >> PAGE_CACHE_SHIFT);
395
396         if (!strcmp(name, LUSTRE_MDC_NAME)) {
397                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
398         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
399                 cli->cl_max_rpcs_in_flight = 2;
400         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
401                 cli->cl_max_rpcs_in_flight = 3;
402         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
403                 cli->cl_max_rpcs_in_flight = 4;
404         } else {
405                 if (osc_on_mdt(obddev->obd_name))
406                         cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
407                 else
408                         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
409         }
410         rc = ldlm_get_ref();
411         if (rc) {
412                 CERROR("ldlm_get_ref failed: %d\n", rc);
413                 GOTO(err, rc);
414         }
415
416         ptlrpc_init_client(rq_portal, rp_portal, name,
417                            &obddev->obd_ldlm_client);
418
419         imp = class_new_import(obddev);
420         if (imp == NULL)
421                 GOTO(err_ldlm, rc = -ENOENT);
422         imp->imp_client = &obddev->obd_ldlm_client;
423         imp->imp_connect_op = connect_op;
424         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
425                LUSTRE_CFG_BUFLEN(lcfg, 1));
426         class_import_put(imp);
427
428         rc = client_import_add_conn(imp, &server_uuid, 1);
429         if (rc) {
430                 CERROR("can't add initial connection\n");
431                 GOTO(err_import, rc);
432         }
433
434         cli->cl_import = imp;
435         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
436         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
437         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
438
439         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
440                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
441                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
442                                name, obddev->obd_name,
443                                cli->cl_target_uuid.uuid);
444                         spin_lock(&imp->imp_lock);
445                         imp->imp_deactive = 1;
446                         spin_unlock(&imp->imp_lock);
447                 }
448         }
449
450         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
451                                                    LDLM_NAMESPACE_CLIENT,
452                                                    LDLM_NAMESPACE_GREEDY,
453                                                    ns_type);
454         if (obddev->obd_namespace == NULL) {
455                 CERROR("Unable to create client namespace - %s\n",
456                        obddev->obd_name);
457                 GOTO(err_import, rc = -ENOMEM);
458         }
459
460         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
461
462         RETURN(rc);
463
464 err_import:
465         class_destroy_import(imp);
466 err_ldlm:
467         ldlm_put_ref();
468 err:
469         RETURN(rc);
470
471 }
472 EXPORT_SYMBOL(client_obd_setup);
473
474 int client_obd_cleanup(struct obd_device *obddev)
475 {
476         ENTRY;
477
478         ldlm_namespace_free_post(obddev->obd_namespace);
479         obddev->obd_namespace = NULL;
480
481         LASSERT(obddev->u.cli.cl_import == NULL);
482
483         ldlm_put_ref();
484         RETURN(0);
485 }
486 EXPORT_SYMBOL(client_obd_cleanup);
487
488 /* ->o_connect() method for client side (OSC and MDC and MGC) */
489 int client_connect_import(const struct lu_env *env,
490                           struct obd_export **exp,
491                           struct obd_device *obd, struct obd_uuid *cluuid,
492                           struct obd_connect_data *data, void *localdata)
493 {
494         struct client_obd       *cli    = &obd->u.cli;
495         struct obd_import       *imp    = cli->cl_import;
496         struct obd_connect_data *ocd;
497         struct lustre_handle    conn    = { 0 };
498         int                     rc;
499         ENTRY;
500
501         *exp = NULL;
502         down_write(&cli->cl_sem);
503         if (cli->cl_conn_count > 0 )
504                 GOTO(out_sem, rc = -EALREADY);
505
506         rc = class_connect(&conn, obd, cluuid);
507         if (rc)
508                 GOTO(out_sem, rc);
509
510         cli->cl_conn_count++;
511         *exp = class_conn2export(&conn);
512
513         LASSERT(obd->obd_namespace);
514
515         imp->imp_dlm_handle = conn;
516         rc = ptlrpc_init_import(imp);
517         if (rc != 0)
518                 GOTO(out_ldlm, rc);
519
520         ocd = &imp->imp_connect_data;
521         if (data) {
522                 *ocd = *data;
523                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
524         }
525
526         rc = ptlrpc_connect_import(imp);
527         if (rc != 0) {
528                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
529                 GOTO(out_ldlm, rc);
530         }
531         LASSERT((*exp)->exp_connection);
532
533         if (data) {
534                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
535                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
536                          data->ocd_connect_flags, ocd->ocd_connect_flags);
537                 data->ocd_connect_flags = ocd->ocd_connect_flags;
538         }
539
540         ptlrpc_pinger_add_import(imp);
541
542         EXIT;
543
544         if (rc) {
545 out_ldlm:
546                 cli->cl_conn_count--;
547                 class_disconnect(*exp);
548                 *exp = NULL;
549         }
550 out_sem:
551         up_write(&cli->cl_sem);
552
553         return rc;
554 }
555 EXPORT_SYMBOL(client_connect_import);
556
557 int client_disconnect_export(struct obd_export *exp)
558 {
559         struct obd_device *obd = class_exp2obd(exp);
560         struct client_obd *cli;
561         struct obd_import *imp;
562         int rc = 0, err;
563         ENTRY;
564
565         if (!obd) {
566                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
567                        exp, exp ? exp->exp_handle.h_cookie : -1);
568                 RETURN(-EINVAL);
569         }
570
571         cli = &obd->u.cli;
572         imp = cli->cl_import;
573
574         down_write(&cli->cl_sem);
575         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
576                cli->cl_conn_count);
577
578         if (!cli->cl_conn_count) {
579                 CERROR("disconnecting disconnected device (%s)\n",
580                        obd->obd_name);
581                 GOTO(out_disconnect, rc = -EINVAL);
582         }
583
584         cli->cl_conn_count--;
585         if (cli->cl_conn_count)
586                 GOTO(out_disconnect, rc = 0);
587
588         /* Mark import deactivated now, so we don't try to reconnect if any
589          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
590          * fully deactivate the import, or that would drop all requests. */
591         spin_lock(&imp->imp_lock);
592         imp->imp_deactive = 1;
593         spin_unlock(&imp->imp_lock);
594
595         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
596          * delete it regardless.  (It's safe to delete an import that was
597          * never added.) */
598         (void)ptlrpc_pinger_del_import(imp);
599
600         if (obd->obd_namespace != NULL) {
601                 /* obd_force == local only */
602                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
603                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
604                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
605         }
606
607         /* There's no need to hold sem while disconnecting an import,
608          * and it may actually cause deadlock in GSS. */
609         up_write(&cli->cl_sem);
610         rc = ptlrpc_disconnect_import(imp, 0);
611         down_write(&cli->cl_sem);
612
613         ptlrpc_invalidate_import(imp);
614
615         EXIT;
616
617 out_disconnect:
618         /* Use server style - class_disconnect should be always called for
619          * o_disconnect. */
620         err = class_disconnect(exp);
621         if (!rc && err)
622                 rc = err;
623
624         up_write(&cli->cl_sem);
625
626         RETURN(rc);
627 }
628 EXPORT_SYMBOL(client_disconnect_export);
629
630 #ifdef HAVE_SERVER_SUPPORT
631 int server_disconnect_export(struct obd_export *exp)
632 {
633         int rc;
634         ENTRY;
635
636         /* Disconnect early so that clients can't keep using export. */
637         rc = class_disconnect(exp);
638         /* Close import to avoid sending any requests. */
639         if (exp->exp_imp_reverse)
640                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
641
642         if (exp->exp_obd->obd_namespace != NULL)
643                 ldlm_cancel_locks_for_export(exp);
644
645         /* complete all outstanding replies */
646         spin_lock(&exp->exp_lock);
647         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
648                 struct ptlrpc_reply_state *rs =
649                         cfs_list_entry(exp->exp_outstanding_replies.next,
650                                        struct ptlrpc_reply_state, rs_exp_list);
651                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
652
653                 spin_lock(&svcpt->scp_rep_lock);
654
655                 cfs_list_del_init(&rs->rs_exp_list);
656                 spin_lock(&rs->rs_lock);
657                 ptlrpc_schedule_difficult_reply(rs);
658                 spin_unlock(&rs->rs_lock);
659
660                 spin_unlock(&svcpt->scp_rep_lock);
661         }
662         spin_unlock(&exp->exp_lock);
663
664         RETURN(rc);
665 }
666 EXPORT_SYMBOL(server_disconnect_export);
667
668 /* --------------------------------------------------------------------------
669  * from old lib/target.c
670  * -------------------------------------------------------------------------- */
671
672 static int target_handle_reconnect(struct lustre_handle *conn,
673                                    struct obd_export *exp,
674                                    struct obd_uuid *cluuid)
675 {
676         ENTRY;
677
678         if (exp->exp_connection && exp->exp_imp_reverse) {
679                 struct lustre_handle *hdl;
680                 struct obd_device *target;
681
682                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
683                 target = exp->exp_obd;
684
685                 /* Might be a re-connect after a partition. */
686                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
687                         if (target->obd_recovering) {
688                                 int timeout = cfs_duration_sec(cfs_time_sub(
689                                         cfs_timer_deadline(
690                                         &target->obd_recovery_timer),
691                                         cfs_time_current()));
692
693                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
694                                         "ing, waiting for %d clients in recov"
695                                         "ery for %d:%.02d\n", target->obd_name,
696                                         obd_uuid2str(&exp->exp_client_uuid),
697                                         obd_export_nid2str(exp),
698                                         target->obd_max_recoverable_clients,
699                                         timeout / 60, timeout % 60);
700                         } else {
701                                 LCONSOLE_WARN("%s: Client %s (at %s) "
702                                         "reconnecting\n", target->obd_name,
703                                         obd_uuid2str(&exp->exp_client_uuid),
704                                         obd_export_nid2str(exp));
705                         }
706
707                         conn->cookie = exp->exp_handle.h_cookie;
708                         /* target_handle_connect() treats EALREADY and
709                          * -EALREADY differently.  EALREADY means we are
710                          * doing a valid reconnect from the same client. */
711                         RETURN(EALREADY);
712                 } else {
713                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
714                                       "with handle "LPX64". Rejecting client "
715                                       "with the same UUID trying to reconnect "
716                                       "with handle "LPX64"\n", target->obd_name,
717                                       obd_uuid2str(&exp->exp_client_uuid),
718                                       obd_export_nid2str(exp),
719                                       hdl->cookie, conn->cookie);
720                         memset(conn, 0, sizeof *conn);
721                         /* target_handle_connect() treats EALREADY and
722                          * -EALREADY differently.  -EALREADY is an error
723                          * (same UUID, different handle). */
724                         RETURN(-EALREADY);
725                 }
726         }
727
728         conn->cookie = exp->exp_handle.h_cookie;
729         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
730                cluuid->uuid, exp, conn->cookie);
731         RETURN(0);
732 }
733
734 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
735                           int error)
736 {
737         struct obd_export *exp = cb_data;
738
739         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
740                obd->obd_name, exp->exp_client_uuid.uuid);
741
742         spin_lock(&exp->exp_lock);
743         exp->exp_need_sync = 0;
744         spin_unlock(&exp->exp_lock);
745         class_export_cb_put(exp);
746 }
747 EXPORT_SYMBOL(target_client_add_cb);
748
749 #ifdef __KERNEL__
750 static void
751 check_and_start_recovery_timer(struct obd_device *obd,
752                                struct ptlrpc_request *req, int new_client);
753 #else
754 static inline void
755 check_and_start_recovery_timer(struct obd_device *obd,
756                                struct ptlrpc_request *req, int new_client)
757 {
758 }
759 #endif
760
761 int target_handle_connect(struct ptlrpc_request *req)
762 {
763         struct obd_device *target = NULL, *targref = NULL;
764         struct obd_export *export = NULL;
765         struct obd_import *revimp;
766         struct obd_import *tmp_imp = NULL;
767         struct lustre_handle conn;
768         struct lustre_handle *tmp;
769         struct obd_uuid tgtuuid;
770         struct obd_uuid cluuid;
771         struct obd_uuid remote_uuid;
772         char *str;
773         int rc = 0;
774         char *target_start;
775         int target_len;
776         bool     mds_conn = false, lw_client = false;
777         struct obd_connect_data *data, *tmpdata;
778         int size, tmpsize;
779         lnet_nid_t *client_nid = NULL;
780         ENTRY;
781
782         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
783
784         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
785         if (str == NULL) {
786                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
787                 GOTO(out, rc = -EINVAL);
788         }
789
790         obd_str2uuid(&tgtuuid, str);
791         target = class_uuid2obd(&tgtuuid);
792         if (!target)
793                 target = class_name2obd(str);
794
795         if (!target) {
796                 deuuidify(str, NULL, &target_start, &target_len);
797                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
798                                    "from %s (no target). If you are running "
799                                    "an HA pair check that the target is "
800                                    "mounted on the other server.\n", str,
801                                    libcfs_nid2str(req->rq_peer.nid));
802                 GOTO(out, rc = -ENODEV);
803         }
804
805         spin_lock(&target->obd_dev_lock);
806         if (target->obd_stopping || !target->obd_set_up) {
807                 spin_unlock(&target->obd_dev_lock);
808
809                 deuuidify(str, NULL, &target_start, &target_len);
810                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
811                               target_len, target_start,
812                               libcfs_nid2str(req->rq_peer.nid),
813                               (target->obd_stopping ?
814                                "stopping" : "not set up"));
815                 GOTO(out, rc = -ENODEV);
816         }
817
818         if (target->obd_no_conn) {
819                 spin_unlock(&target->obd_dev_lock);
820
821                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
822                                "from %s\n", target->obd_name,
823                                libcfs_nid2str(req->rq_peer.nid));
824                 GOTO(out, rc = -EAGAIN);
825         }
826
827         /* Make sure the target isn't cleaned up while we're here. Yes,
828          * there's still a race between the above check and our incref here.
829          * Really, class_uuid2obd should take the ref. */
830         targref = class_incref(target, __FUNCTION__, current);
831
832         target->obd_conn_inprogress++;
833         spin_unlock(&target->obd_dev_lock);
834
835         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
836         if (str == NULL) {
837                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
838                 GOTO(out, rc = -EINVAL);
839         }
840
841         obd_str2uuid(&cluuid, str);
842
843         /* XXX Extract a nettype and format accordingly. */
844         switch (sizeof(lnet_nid_t)) {
845         /* NB the casts only avoid compiler warnings. */
846         case 8:
847                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
848                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
849                 break;
850         case 4:
851                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
852                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
853                 break;
854         default:
855                 LBUG();
856         }
857
858         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
859         if (tmp == NULL)
860                 GOTO(out, rc = -EPROTO);
861
862         conn = *tmp;
863
864         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
865                                     RCL_CLIENT);
866         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
867         if (!data)
868                 GOTO(out, rc = -EPROTO);
869
870         rc = req_capsule_server_pack(&req->rq_pill);
871         if (rc)
872                 GOTO(out, rc);
873
874         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
875                 if (data->ocd_version < LUSTRE_VERSION_CODE -
876                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
877                     data->ocd_version > LUSTRE_VERSION_CODE +
878                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
879                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
880                                   "libclient connection attempt",
881                                   data->ocd_version < LUSTRE_VERSION_CODE ?
882                                   "old" : "new",
883                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
884                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
885                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
886                                   OBD_OCD_VERSION_FIX(data->ocd_version));
887                         data = req_capsule_server_sized_get(&req->rq_pill,
888                                                             &RMF_CONNECT_DATA,
889                                     offsetof(typeof(*data), ocd_version) +
890                                              sizeof(data->ocd_version));
891                         if (data) {
892                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
893                                 data->ocd_version = LUSTRE_VERSION_CODE;
894                         }
895                         GOTO(out, rc = -EPROTO);
896                 }
897         }
898
899         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
900             (data->ocd_connect_flags & OBD_CONNECT_MDS))
901                 mds_conn = true;
902
903         if ((data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0)
904                 lw_client = true;
905
906         /* lctl gets a backstage, all-access pass. */
907         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
908                 goto dont_check_exports;
909
910         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
911         if (!export)
912                 goto no_export;
913
914         /* We've found an export in the hash. */
915
916         spin_lock(&export->exp_lock);
917
918         if (export->exp_connecting) { /* bug 9635, et. al. */
919                 spin_unlock(&export->exp_lock);
920                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
921                               export->exp_obd->obd_name, export,
922                               libcfs_nid2str(req->rq_peer.nid));
923                 class_export_put(export);
924                 export = NULL;
925                 rc = -EALREADY;
926         } else if (mds_conn && export->exp_connection) {
927                 spin_unlock(&export->exp_lock);
928                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
929                         /* MDS reconnected after failover. */
930                         LCONSOLE_WARN("%s: Received MDS connection from "
931                             "%s, removing former export from %s\n",
932                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
933                             libcfs_nid2str(export->exp_connection->c_peer.nid));
934                 else
935                         /* New MDS connection from the same NID. */
936                         LCONSOLE_WARN("%s: Received new MDS connection from "
937                             "%s, removing former export from same NID\n",
938                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
939                 class_fail_export(export);
940                 class_export_put(export);
941                 export = NULL;
942                 rc = 0;
943         } else if (export->exp_connection != NULL &&
944                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
945                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
946                     MSG_CONNECT_INITIAL)) {
947                 spin_unlock(&export->exp_lock);
948                 /* In MDS failover we have static UUID but NID can change. */
949                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
950                               "existing nid %s is already connected\n",
951                               target->obd_name, cluuid.uuid,
952                               libcfs_nid2str(req->rq_peer.nid),
953                               libcfs_nid2str(
954                                       export->exp_connection->c_peer.nid));
955                 rc = -EALREADY;
956                 class_export_put(export);
957                 export = NULL;
958         } else {
959                 export->exp_connecting = 1;
960                 spin_unlock(&export->exp_lock);
961                 LASSERT(export->exp_obd == target);
962
963                 rc = target_handle_reconnect(&conn, export, &cluuid);
964         }
965
966         /* If we found an export, we already unlocked. */
967         if (!export) {
968 no_export:
969                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
970         } else if (req->rq_export == NULL &&
971                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
972                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
973                               "still busy with %d references\n",
974                               target->obd_name, cluuid.uuid,
975                               libcfs_nid2str(req->rq_peer.nid),
976                               cfs_atomic_read(&export->exp_refcount));
977                 GOTO(out, rc = -EBUSY);
978         } else if (req->rq_export != NULL &&
979                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
980                 /* The current connect RPC has increased exp_rpc_count. */
981                 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
982                               "still busy with %d active RPCs\n",
983                               target->obd_name, cluuid.uuid,
984                               libcfs_nid2str(req->rq_peer.nid),
985                               cfs_atomic_read(&export->exp_rpc_count) - 1);
986                 spin_lock(&export->exp_lock);
987                 if (req->rq_export->exp_conn_cnt <
988                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
989                         /* try to abort active requests */
990                         req->rq_export->exp_abort_active_req = 1;
991                 spin_unlock(&export->exp_lock);
992                 GOTO(out, rc = -EBUSY);
993         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
994                 if (!strstr(cluuid.uuid, "mdt"))
995                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
996                                       "known client %s (at %s) because it "
997                                       "is indicating it is a new client",
998                                       target->obd_name, cluuid.uuid,
999                                       libcfs_nid2str(req->rq_peer.nid));
1000                 GOTO(out, rc = -EALREADY);
1001         } else {
1002                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1003         }
1004
1005         if (rc < 0) {
1006                 GOTO(out, rc);
1007         }
1008
1009         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1010                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1011               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1012               export, (long)cfs_time_current_sec(),
1013               export ? (long)export->exp_last_request_time : 0);
1014
1015         /* If this is the first time a client connects, reset the recovery
1016          * timer. Discard lightweight connections which might be local. */
1017         if (!lw_client && rc == 0 && target->obd_recovering)
1018                 check_and_start_recovery_timer(target, req, export == NULL);
1019
1020         /* We want to handle EALREADY but *not* -EALREADY from
1021          * target_handle_reconnect(), return reconnection state in a flag. */
1022         if (rc == EALREADY) {
1023                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1024                 rc = 0;
1025         } else {
1026                 LASSERT(rc == 0);
1027         }
1028
1029         /* Tell the client if we support replayable requests. */
1030         if (target->obd_replayable)
1031                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1032         client_nid = &req->rq_peer.nid;
1033
1034         if (export == NULL) {
1035                 /* allow lightweight connections during recovery */
1036                 if (target->obd_recovering && !lw_client) {
1037                         cfs_time_t t;
1038                         int     c; /* connected */
1039                         int     i; /* in progress */
1040                         int     k; /* known */
1041                         int     s; /* stale/evicted */
1042
1043                         c = cfs_atomic_read(&target->obd_connected_clients);
1044                         i = cfs_atomic_read(&target->obd_lock_replay_clients);
1045                         k = target->obd_max_recoverable_clients;
1046                         s = target->obd_stale_clients;
1047                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1048                         t = cfs_time_sub(t, cfs_time_current());
1049                         t = cfs_duration_sec(t);
1050                         LCONSOLE_WARN("%s: Denying connection for new client "
1051                                       "%s (at %s), waiting for all %d known "
1052                                       "clients (%d recovered, %d in progress, "
1053                                       "and %d evicted) to recover in %d:%.02d\n",
1054                                       target->obd_name, cluuid.uuid,
1055                                       libcfs_nid2str(req->rq_peer.nid), k,
1056                                       c - i, i, s, (int)t / 60,
1057                                       (int)t % 60);
1058                         rc = -EBUSY;
1059                 } else {
1060 dont_check_exports:
1061                         rc = obd_connect(req->rq_svc_thread->t_env,
1062                                          &export, target, &cluuid, data,
1063                                          client_nid);
1064                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1065                                 lustre_msg_add_op_flags(req->rq_repmsg,
1066                                                 MSG_CONNECT_RECOVERING);
1067                         if (rc == 0)
1068                                 conn.cookie = export->exp_handle.h_cookie;
1069                 }
1070         } else {
1071                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1072                                    export, target, &cluuid, data, client_nid);
1073         }
1074         if (rc)
1075                 GOTO(out, rc);
1076
1077 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
1078         /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
1079          * will do the swabbing for if the client is using the same endianness.
1080          *
1081          * This fixup is version-limited, because we don't want to carry the
1082          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
1083          * interop with unpatched 2.2 clients.  For newer clients, servers
1084          * will never do MNE swabbing, let the client handle that.  LU-1644 */
1085         export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) &&
1086                         !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB);
1087 #else
1088 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
1089 #endif
1090
1091         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1092         data->ocd_instance = target->u.obt.obt_instance;
1093
1094         /* Return only the parts of obd_connect_data that we understand, so the
1095          * client knows that we don't understand the rest. */
1096         if (data) {
1097                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1098                                                RCL_SERVER);
1099                 tmpdata = req_capsule_server_get(&req->rq_pill,
1100                                                  &RMF_CONNECT_DATA);
1101                 /* Don't use struct assignment here, because the client reply
1102                  * buffer may be smaller/larger than the local struct
1103                  * obd_connect_data. */
1104                 memcpy(tmpdata, data, min(tmpsize, size));
1105         }
1106
1107         /* If all else goes well, this is our RPC return code. */
1108         req->rq_status = 0;
1109
1110         lustre_msg_set_handle(req->rq_repmsg, &conn);
1111
1112         /* If the client and the server are the same node, we will already
1113          * have an export that really points to the client's DLM export,
1114          * because we have a shared handles table.
1115          *
1116          * XXX this will go away when shaver stops sending the "connect" handle
1117          * in the real "remote handle" field of the request --phik 24 Apr 2003
1118          */
1119         ptlrpc_request_change_export(req, export);
1120
1121         spin_lock(&export->exp_lock);
1122         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1123                 spin_unlock(&export->exp_lock);
1124                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1125                        "or equal conn_cnt: %d >= %d\n",
1126                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1127                        export->exp_conn_cnt,
1128                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1129
1130                 GOTO(out, rc = -EALREADY);
1131         }
1132         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1133         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1134         export->exp_abort_active_req = 0;
1135
1136         /* Don't evict liblustre clients for not pinging. */
1137         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1138                 export->exp_libclient = 1;
1139                 spin_unlock(&export->exp_lock);
1140
1141                 spin_lock(&target->obd_dev_lock);
1142                 cfs_list_del_init(&export->exp_obd_chain_timed);
1143                 spin_unlock(&target->obd_dev_lock);
1144         } else {
1145                 spin_unlock(&export->exp_lock);
1146         }
1147
1148         if (export->exp_connection != NULL) {
1149                 /* Check to see if connection came from another NID. */
1150                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1151                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1152                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1153                                      &export->exp_connection->c_peer.nid,
1154                                      &export->exp_nid_hash);
1155
1156                 ptlrpc_connection_put(export->exp_connection);
1157         }
1158
1159         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1160                                                        req->rq_self,
1161                                                        &remote_uuid);
1162         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1163                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1164                              &export->exp_connection->c_peer.nid,
1165                              &export->exp_nid_hash);
1166         }
1167
1168         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1169                 int has_transno;
1170                 __u64 transno = data->ocd_transno;
1171
1172                 spin_lock(&export->exp_lock);
1173                 /* possible race with class_disconnect_stale_exports,
1174                  * export may be already in the eviction process */
1175                 if (export->exp_failed) {
1176                         spin_unlock(&export->exp_lock);
1177                         GOTO(out, rc = -ENODEV);
1178                 }
1179                 export->exp_in_recovery = 1;
1180                 export->exp_req_replay_needed = 1;
1181                 export->exp_lock_replay_needed = 1;
1182                 spin_unlock(&export->exp_lock);
1183
1184                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1185                                  MSG_CONNECT_TRANSNO);
1186                 if (has_transno && transno == 0)
1187                         CWARN("Connect with zero transno!\n");
1188
1189                 if (has_transno && transno > 0 &&
1190                     transno < target->obd_next_recovery_transno &&
1191                     transno > target->obd_last_committed) {
1192                         /* Another way is to use cmpxchg() to be lock-free. */
1193                         spin_lock(&target->obd_recovery_task_lock);
1194                         if (transno < target->obd_next_recovery_transno)
1195                                 target->obd_next_recovery_transno = transno;
1196                         spin_unlock(&target->obd_recovery_task_lock);
1197                 }
1198
1199                 cfs_atomic_inc(&target->obd_req_replay_clients);
1200                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1201                 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1202                     target->obd_max_recoverable_clients)
1203                         wake_up(&target->obd_next_transno_waitq);
1204         }
1205
1206         /* Tell the client we're in recovery, when client is involved in it. */
1207         if (target->obd_recovering && !lw_client)
1208                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1209
1210         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1211         conn = *tmp;
1212
1213         /* Return -ENOTCONN in case of errors to let client reconnect. */
1214         revimp = class_new_import(target);
1215         if (revimp == NULL) {
1216                 CERROR("fail to alloc new reverse import.\n");
1217                 GOTO(out, rc = -ENOTCONN);
1218         }
1219
1220         spin_lock(&export->exp_lock);
1221         if (export->exp_imp_reverse != NULL)
1222                 /* destroyed import can be still referenced in ctxt */
1223                 tmp_imp = export->exp_imp_reverse;
1224         export->exp_imp_reverse = revimp;
1225         spin_unlock(&export->exp_lock);
1226
1227         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1228         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1229         revimp->imp_remote_handle = conn;
1230         revimp->imp_dlm_fake = 1;
1231         revimp->imp_state = LUSTRE_IMP_FULL;
1232
1233         /* Unknown versions will be caught in
1234          * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1235         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1236
1237         if ((data->ocd_connect_flags & OBD_CONNECT_AT) &&
1238             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1239                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1240         else
1241                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1242
1243         if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1244             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1245                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1246         else
1247                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1248
1249         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1250         if (rc) {
1251                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1252                 spin_lock(&export->exp_lock);
1253                 export->exp_imp_reverse = NULL;
1254                 spin_unlock(&export->exp_lock);
1255                 class_destroy_import(revimp);
1256         }
1257
1258         class_import_put(revimp);
1259
1260 out:
1261         if (tmp_imp != NULL)
1262                 client_destroy_import(tmp_imp);
1263         if (export) {
1264                 spin_lock(&export->exp_lock);
1265                 export->exp_connecting = 0;
1266                 spin_unlock(&export->exp_lock);
1267
1268                 class_export_put(export);
1269         }
1270         if (targref) {
1271                 spin_lock(&target->obd_dev_lock);
1272                 target->obd_conn_inprogress--;
1273                 spin_unlock(&target->obd_dev_lock);
1274
1275                 class_decref(targref, __func__, current);
1276         }
1277         if (rc)
1278                 req->rq_status = rc;
1279         RETURN(rc);
1280 }
1281 EXPORT_SYMBOL(target_handle_connect);
1282
1283 int target_handle_disconnect(struct ptlrpc_request *req)
1284 {
1285         int rc;
1286         ENTRY;
1287
1288         rc = req_capsule_server_pack(&req->rq_pill);
1289         if (rc)
1290                 RETURN(rc);
1291
1292         /* Keep the rq_export around so we can send the reply. */
1293         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1294
1295         RETURN(0);
1296 }
1297 EXPORT_SYMBOL(target_handle_disconnect);
1298
1299 void target_destroy_export(struct obd_export *exp)
1300 {
1301         struct obd_import       *imp = NULL;
1302         /* exports created from last_rcvd data, and "fake"
1303            exports created by lctl don't have an import */
1304         spin_lock(&exp->exp_lock);
1305         if (exp->exp_imp_reverse != NULL) {
1306                 imp = exp->exp_imp_reverse;
1307                 exp->exp_imp_reverse = NULL;
1308         }
1309         spin_unlock(&exp->exp_lock);
1310         if (imp != NULL)
1311                 client_destroy_import(imp);
1312
1313         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1314         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1315         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1316         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1317 }
1318 EXPORT_SYMBOL(target_destroy_export);
1319
1320 /*
1321  * Recovery functions
1322  */
1323 static void target_request_copy_get(struct ptlrpc_request *req)
1324 {
1325         class_export_rpc_inc(req->rq_export);
1326         LASSERT(cfs_list_empty(&req->rq_list));
1327         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1328
1329         /* Increase refcount to keep request in queue. */
1330         cfs_atomic_inc(&req->rq_refcount);
1331         /* Let export know it has replays to be handled. */
1332         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1333 }
1334
1335 static void target_request_copy_put(struct ptlrpc_request *req)
1336 {
1337         LASSERT(cfs_list_empty(&req->rq_replay_list));
1338         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1339
1340         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1341         class_export_rpc_dec(req->rq_export);
1342         ptlrpc_server_drop_request(req);
1343 }
1344
1345 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1346 {
1347         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1348         struct obd_export     *exp = req->rq_export;
1349         struct ptlrpc_request *reqiter;
1350         int                    dup = 0;
1351
1352         LASSERT(exp);
1353
1354         spin_lock(&exp->exp_lock);
1355         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1356                                 rq_replay_list) {
1357                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1358                         dup = 1;
1359                         break;
1360                 }
1361         }
1362
1363         if (dup) {
1364                 /* We expect it with RESENT and REPLAY flags. */
1365                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1366                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1367                         CERROR("invalid flags %x of resent replay\n",
1368                                lustre_msg_get_flags(req->rq_reqmsg));
1369         } else {
1370                 cfs_list_add_tail(&req->rq_replay_list,
1371                                   &exp->exp_req_replay_queue);
1372         }
1373
1374         spin_unlock(&exp->exp_lock);
1375         return dup;
1376 }
1377
1378 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1379 {
1380         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1381         LASSERT(req->rq_export);
1382
1383         spin_lock(&req->rq_export->exp_lock);
1384         cfs_list_del_init(&req->rq_replay_list);
1385         spin_unlock(&req->rq_export->exp_lock);
1386 }
1387
1388 #ifdef __KERNEL__
1389 static void target_finish_recovery(struct obd_device *obd)
1390 {
1391         ENTRY;
1392
1393         /* Only log a recovery message when recovery has occurred. */
1394         if (obd->obd_recovery_start) {
1395                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1396                                         obd->obd_recovery_start);
1397                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1398                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1399                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1400                         obd->obd_max_recoverable_clients,
1401                         cfs_atomic_read(&obd->obd_connected_clients),
1402                         obd->obd_stale_clients,
1403                         obd->obd_stale_clients == 1 ? "was" : "were");
1404         }
1405
1406         ldlm_reprocess_all_ns(obd->obd_namespace);
1407         spin_lock(&obd->obd_recovery_task_lock);
1408         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1409             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1410             !cfs_list_empty(&obd->obd_final_req_queue)) {
1411                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1412                        obd->obd_name,
1413                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1414                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1415                                "" : "lock ",
1416                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1417                                "" : "final ");
1418                 spin_unlock(&obd->obd_recovery_task_lock);
1419                 LBUG();
1420         }
1421         spin_unlock(&obd->obd_recovery_task_lock);
1422
1423         obd->obd_recovery_end = cfs_time_current_sec();
1424
1425         /* When recovery finished, cleanup orphans on MDS and OST. */
1426         if (OBT(obd) && OBP(obd, postrecov)) {
1427                 int rc = OBP(obd, postrecov)(obd);
1428                 if (rc < 0)
1429                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1430                                       obd->obd_name, rc);
1431         }
1432         EXIT;
1433 }
1434
1435 static void abort_req_replay_queue(struct obd_device *obd)
1436 {
1437         struct ptlrpc_request *req, *n;
1438         cfs_list_t abort_list;
1439
1440         CFS_INIT_LIST_HEAD(&abort_list);
1441         spin_lock(&obd->obd_recovery_task_lock);
1442         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1443         spin_unlock(&obd->obd_recovery_task_lock);
1444         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1445                 DEBUG_REQ(D_WARNING, req, "aborted:");
1446                 req->rq_status = -ENOTCONN;
1447                 if (ptlrpc_error(req)) {
1448                         DEBUG_REQ(D_ERROR, req,
1449                                   "failed abort_req_reply; skipping");
1450                 }
1451                 target_exp_dequeue_req_replay(req);
1452                 target_request_copy_put(req);
1453         }
1454 }
1455
1456 static void abort_lock_replay_queue(struct obd_device *obd)
1457 {
1458         struct ptlrpc_request *req, *n;
1459         cfs_list_t abort_list;
1460
1461         CFS_INIT_LIST_HEAD(&abort_list);
1462         spin_lock(&obd->obd_recovery_task_lock);
1463         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1464         spin_unlock(&obd->obd_recovery_task_lock);
1465         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1466                 DEBUG_REQ(D_ERROR, req, "aborted:");
1467                 req->rq_status = -ENOTCONN;
1468                 if (ptlrpc_error(req)) {
1469                         DEBUG_REQ(D_ERROR, req,
1470                                   "failed abort_lock_reply; skipping");
1471                 }
1472                 target_request_copy_put(req);
1473         }
1474 }
1475
1476 /* Called from a cleanup function if the device is being cleaned up
1477    forcefully.  The exports should all have been disconnected already,
1478    the only thing left to do is
1479      - clear the recovery flags
1480      - cancel the timer
1481      - free queued requests and replies, but don't send replies
1482    Because the obd_stopping flag is set, no new requests should be received.
1483
1484 */
1485 void target_cleanup_recovery(struct obd_device *obd)
1486 {
1487         struct ptlrpc_request *req, *n;
1488         cfs_list_t clean_list;
1489         ENTRY;
1490
1491         CFS_INIT_LIST_HEAD(&clean_list);
1492         spin_lock(&obd->obd_dev_lock);
1493         if (!obd->obd_recovering) {
1494                 spin_unlock(&obd->obd_dev_lock);
1495                 EXIT;
1496                 return;
1497         }
1498         obd->obd_recovering = obd->obd_abort_recovery = 0;
1499         spin_unlock(&obd->obd_dev_lock);
1500
1501         spin_lock(&obd->obd_recovery_task_lock);
1502         target_cancel_recovery_timer(obd);
1503         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1504         spin_unlock(&obd->obd_recovery_task_lock);
1505
1506         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1507                 LASSERT(req->rq_reply_state == 0);
1508                 target_exp_dequeue_req_replay(req);
1509                 target_request_copy_put(req);
1510         }
1511
1512         spin_lock(&obd->obd_recovery_task_lock);
1513         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1514         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1515         spin_unlock(&obd->obd_recovery_task_lock);
1516
1517         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1518                 LASSERT(req->rq_reply_state == 0);
1519                 target_request_copy_put(req);
1520         }
1521
1522         EXIT;
1523 }
1524 EXPORT_SYMBOL(target_cleanup_recovery);
1525
1526 /* obd_recovery_task_lock should be held */
1527 void target_cancel_recovery_timer(struct obd_device *obd)
1528 {
1529         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1530         cfs_timer_disarm(&obd->obd_recovery_timer);
1531 }
1532 EXPORT_SYMBOL(target_cancel_recovery_timer);
1533
1534 static void target_start_recovery_timer(struct obd_device *obd)
1535 {
1536         if (obd->obd_recovery_start != 0)
1537                 return;
1538
1539         spin_lock(&obd->obd_dev_lock);
1540         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1541                 spin_unlock(&obd->obd_dev_lock);
1542                 return;
1543         }
1544
1545         LASSERT(obd->obd_recovery_timeout != 0);
1546
1547         if (obd->obd_recovery_start != 0) {
1548                 spin_unlock(&obd->obd_dev_lock);
1549                 return;
1550         }
1551
1552         cfs_timer_arm(&obd->obd_recovery_timer,
1553                       cfs_time_shift(obd->obd_recovery_timeout));
1554         obd->obd_recovery_start = cfs_time_current_sec();
1555         spin_unlock(&obd->obd_dev_lock);
1556
1557         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1558                       "or until %d client%s reconnect%s\n",
1559                       obd->obd_name,
1560                       obd->obd_recovery_timeout / 60,
1561                       obd->obd_recovery_timeout % 60,
1562                       obd->obd_max_recoverable_clients,
1563                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1564                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1565 }
1566
1567 /**
1568  * extend recovery window.
1569  *
1570  * if @extend is true, extend recovery window to have @drt remaining at least;
1571  * otherwise, make sure the recovery timeout value is not less than @drt.
1572  */
1573 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1574 {
1575         cfs_time_t now;
1576         cfs_time_t end;
1577         cfs_duration_t left;
1578         int to;
1579
1580         spin_lock(&obd->obd_dev_lock);
1581         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1582                 spin_unlock(&obd->obd_dev_lock);
1583                 return;
1584         }
1585         LASSERT(obd->obd_recovery_start != 0);
1586
1587         now  = cfs_time_current_sec();
1588         to   = obd->obd_recovery_timeout;
1589         end  = obd->obd_recovery_start + to;
1590         left = cfs_time_sub(end, now);
1591
1592         if (extend && (drt > left)) {
1593                 to += drt - left;
1594         } else if (!extend && (drt > to)) {
1595                 to = drt;
1596                 /* reduce drt by already passed time */
1597                 drt -= obd->obd_recovery_timeout - left;
1598         }
1599
1600         if (to > obd->obd_recovery_time_hard)
1601                 to = obd->obd_recovery_time_hard;
1602         if (obd->obd_recovery_timeout < to ||
1603             obd->obd_recovery_timeout == obd->obd_recovery_time_hard) {
1604                 obd->obd_recovery_timeout = to;
1605                 cfs_timer_arm(&obd->obd_recovery_timer,
1606                               cfs_time_shift(drt));
1607         }
1608         spin_unlock(&obd->obd_dev_lock);
1609
1610         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1611                obd->obd_name, (unsigned)drt);
1612 }
1613
1614 /* Reset the timer with each new client connection */
1615 /*
1616  * This timer is actually reconnect_timer, which is for making sure
1617  * the total recovery window is at least as big as my reconnect
1618  * attempt timing. So the initial recovery time_out will be set to
1619  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1620  * from client is bigger than this, then the recovery time_out will
1621  * be extended to make sure the client could be reconnected, in the
1622  * process, the timeout from the new client should be ignored.
1623  */
1624
1625 static void
1626 check_and_start_recovery_timer(struct obd_device *obd,
1627                                struct ptlrpc_request *req,
1628                                int new_client)
1629 {
1630         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1631         struct obd_device_target *obt = &obd->u.obt;
1632
1633         if (!new_client && service_time)
1634                 /* Teach server about old server's estimates, as first guess
1635                  * at how long new requests will take. */
1636                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1637                             service_time);
1638
1639         target_start_recovery_timer(obd);
1640
1641         /* Convert the service time to RPC timeout,
1642          * and reuse service_time to limit stack usage. */
1643         service_time = at_est2timeout(service_time);
1644
1645         /* We expect other clients to timeout within service_time, then try
1646          * to reconnect, then try the failover server.  The max delay between
1647          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1648         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1649
1650         LASSERT(obt->obt_magic == OBT_MAGIC);
1651         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1652         if (service_time > obd->obd_recovery_timeout && !new_client)
1653                 extend_recovery_timer(obd, service_time, false);
1654 }
1655
1656 /** Health checking routines */
1657 static inline int exp_connect_healthy(struct obd_export *exp)
1658 {
1659         return (exp->exp_in_recovery);
1660 }
1661
1662 /** if export done req_replay or has replay in queue */
1663 static inline int exp_req_replay_healthy(struct obd_export *exp)
1664 {
1665         return (!exp->exp_req_replay_needed ||
1666                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1667 }
1668 /** if export done lock_replay or has replay in queue */
1669 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1670 {
1671         return (!exp->exp_lock_replay_needed ||
1672                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1673 }
1674
1675 static inline int exp_vbr_healthy(struct obd_export *exp)
1676 {
1677         return (!exp->exp_vbr_failed);
1678 }
1679
1680 static inline int exp_finished(struct obd_export *exp)
1681 {
1682         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1683 }
1684
1685 /** Checking routines for recovery */
1686 static int check_for_clients(struct obd_device *obd)
1687 {
1688         unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1689
1690         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1691                 return 1;
1692         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1693         return (clnts + obd->obd_stale_clients ==
1694                 obd->obd_max_recoverable_clients);
1695 }
1696
1697 static int check_for_next_transno(struct obd_device *obd)
1698 {
1699         struct ptlrpc_request *req = NULL;
1700         int wake_up = 0, connected, completed, queue_len;
1701         __u64 next_transno, req_transno;
1702         ENTRY;
1703
1704         spin_lock(&obd->obd_recovery_task_lock);
1705         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1706                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1707                                      struct ptlrpc_request, rq_list);
1708                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1709         } else {
1710                 req_transno = 0;
1711         }
1712
1713         connected = cfs_atomic_read(&obd->obd_connected_clients);
1714         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1715         queue_len = obd->obd_requests_queued_for_recovery;
1716         next_transno = obd->obd_next_recovery_transno;
1717
1718         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1719                "req_transno: "LPU64", next_transno: "LPU64"\n",
1720                obd->obd_max_recoverable_clients, connected, completed,
1721                queue_len, req_transno, next_transno);
1722
1723         if (obd->obd_abort_recovery) {
1724                 CDEBUG(D_HA, "waking for aborted recovery\n");
1725                 wake_up = 1;
1726         } else if (obd->obd_recovery_expired) {
1727                 CDEBUG(D_HA, "waking for expired recovery\n");
1728                 wake_up = 1;
1729         } else if (req_transno == next_transno) {
1730                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1731                 wake_up = 1;
1732         } else if (queue_len > 0 &&
1733                    queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1734                 int d_lvl = D_HA;
1735                 /** handle gaps occured due to lost reply or VBR */
1736                 LASSERTF(req_transno >= next_transno,
1737                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1738                          req_transno, next_transno);
1739                 if (req_transno > obd->obd_last_committed &&
1740                     !obd->obd_version_recov)
1741                         d_lvl = D_ERROR;
1742                 CDEBUG(d_lvl,
1743                        "%s: waking for gap in transno, VBR is %s (skip: "
1744                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1745                        ", last_committed: "LPD64")\n",
1746                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1747                        next_transno, queue_len, completed, connected,
1748                        req_transno, obd->obd_last_committed);
1749                 obd->obd_next_recovery_transno = req_transno;
1750                 wake_up = 1;
1751         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1752                 CDEBUG(D_HA, "waking for completed recovery\n");
1753                 wake_up = 1;
1754         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1755                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1756                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1757                 obd->obd_next_recovery_transno = req_transno;
1758                 wake_up = 1;
1759         }
1760         spin_unlock(&obd->obd_recovery_task_lock);
1761         return wake_up;
1762 }
1763
1764 static int check_for_next_lock(struct obd_device *obd)
1765 {
1766         int wake_up = 0;
1767
1768         spin_lock(&obd->obd_recovery_task_lock);
1769         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1770                 CDEBUG(D_HA, "waking for next lock\n");
1771                 wake_up = 1;
1772         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1773                 CDEBUG(D_HA, "waking for completed lock replay\n");
1774                 wake_up = 1;
1775         } else if (obd->obd_abort_recovery) {
1776                 CDEBUG(D_HA, "waking for aborted recovery\n");
1777                 wake_up = 1;
1778         } else if (obd->obd_recovery_expired) {
1779                 CDEBUG(D_HA, "waking for expired recovery\n");
1780                 wake_up = 1;
1781         }
1782         spin_unlock(&obd->obd_recovery_task_lock);
1783
1784         return wake_up;
1785 }
1786
1787 /**
1788  * wait for recovery events,
1789  * check its status with help of check_routine
1790  * evict dead clients via health_check
1791  */
1792 static int target_recovery_overseer(struct obd_device *obd,
1793                                     int (*check_routine)(struct obd_device *),
1794                                     int (*health_check)(struct obd_export *))
1795 {
1796 repeat:
1797         wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1798         if (obd->obd_abort_recovery) {
1799                 CWARN("recovery is aborted, evict exports in recovery\n");
1800                 /** evict exports which didn't finish recovery yet */
1801                 class_disconnect_stale_exports(obd, exp_finished);
1802                 return 1;
1803         } else if (obd->obd_recovery_expired) {
1804                 obd->obd_recovery_expired = 0;
1805                 /** If some clients died being recovered, evict them */
1806                 LCONSOLE_WARN("%s: recovery is timed out, "
1807                               "evict stale exports\n", obd->obd_name);
1808                 /** evict cexports with no replay in queue, they are stalled */
1809                 class_disconnect_stale_exports(obd, health_check);
1810                 /** continue with VBR */
1811                 spin_lock(&obd->obd_dev_lock);
1812                 obd->obd_version_recov = 1;
1813                 spin_unlock(&obd->obd_dev_lock);
1814                 /**
1815                  * reset timer, recovery will proceed with versions now,
1816                  * timeout is set just to handle reconnection delays
1817                  */
1818                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1819                 /** Wait for recovery events again, after evicting bad clients */
1820                 goto repeat;
1821         }
1822         return 0;
1823 }
1824
1825 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1826 {
1827         struct ptlrpc_request *req = NULL;
1828         ENTRY;
1829
1830         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1831                obd->obd_next_recovery_transno);
1832
1833         if (target_recovery_overseer(obd, check_for_next_transno,
1834                                      exp_req_replay_healthy)) {
1835                 abort_req_replay_queue(obd);
1836                 abort_lock_replay_queue(obd);
1837         }
1838
1839         spin_lock(&obd->obd_recovery_task_lock);
1840         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1841                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1842                                      struct ptlrpc_request, rq_list);
1843                 cfs_list_del_init(&req->rq_list);
1844                 obd->obd_requests_queued_for_recovery--;
1845                 spin_unlock(&obd->obd_recovery_task_lock);
1846         } else {
1847                 spin_unlock(&obd->obd_recovery_task_lock);
1848                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1849                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1850                 /** evict exports failed VBR */
1851                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1852         }
1853         RETURN(req);
1854 }
1855
1856 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1857 {
1858         struct ptlrpc_request *req = NULL;
1859
1860         CDEBUG(D_HA, "Waiting for lock\n");
1861         if (target_recovery_overseer(obd, check_for_next_lock,
1862                                      exp_lock_replay_healthy))
1863                 abort_lock_replay_queue(obd);
1864
1865         spin_lock(&obd->obd_recovery_task_lock);
1866         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1867                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1868                                      struct ptlrpc_request, rq_list);
1869                 cfs_list_del_init(&req->rq_list);
1870                 spin_unlock(&obd->obd_recovery_task_lock);
1871         } else {
1872                 spin_unlock(&obd->obd_recovery_task_lock);
1873                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1874                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1875                 /** evict exports failed VBR */
1876                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1877         }
1878         return req;
1879 }
1880
1881 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1882 {
1883         struct ptlrpc_request *req = NULL;
1884
1885         spin_lock(&obd->obd_recovery_task_lock);
1886         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1887                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1888                                      struct ptlrpc_request, rq_list);
1889                 cfs_list_del_init(&req->rq_list);
1890                 spin_unlock(&obd->obd_recovery_task_lock);
1891                 if (req->rq_export->exp_in_recovery) {
1892                         spin_lock(&req->rq_export->exp_lock);
1893                         req->rq_export->exp_in_recovery = 0;
1894                         spin_unlock(&req->rq_export->exp_lock);
1895                 }
1896         } else {
1897                 spin_unlock(&obd->obd_recovery_task_lock);
1898         }
1899         return req;
1900 }
1901
1902 static int handle_recovery_req(struct ptlrpc_thread *thread,
1903                                struct ptlrpc_request *req,
1904                                svc_handler_t handler)
1905 {
1906         int rc;
1907         ENTRY;
1908
1909         /**
1910          * export can be evicted during recovery, no need to handle replays for
1911          * it after that, discard such request silently
1912          */
1913         if (req->rq_export->exp_disconnected)
1914                 GOTO(reqcopy_put, rc = 0);
1915
1916         rc = lu_context_init(&req->rq_recov_session, LCT_SERVER_SESSION);
1917         if (rc) {
1918                 CERROR("Failure to initialize session: %d\n", rc);
1919                 GOTO(reqcopy_put, rc);
1920         }
1921
1922         req->rq_recov_session.lc_thread = thread;
1923         lu_context_enter(&req->rq_recov_session);
1924         req->rq_svc_thread = thread;
1925         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1926
1927         /* thread context */
1928         lu_context_enter(&thread->t_env->le_ctx);
1929         (void)handler(req);
1930         lu_context_exit(&thread->t_env->le_ctx);
1931
1932         lu_context_exit(&req->rq_recov_session);
1933         lu_context_fini(&req->rq_recov_session);
1934         /* don't reset timer for final stage */
1935         if (!exp_finished(req->rq_export)) {
1936                 int to = obd_timeout;
1937
1938                 /**
1939                  * Add request timeout to the recovery time so next request from
1940                  * this client may come in recovery time
1941                  */
1942                 if (!AT_OFF) {
1943                         struct ptlrpc_service_part *svcpt;
1944
1945                         svcpt = req->rq_rqbd->rqbd_svcpt;
1946                         /* If the server sent early reply for this request,
1947                          * the client will recalculate the timeout according to
1948                          * current server estimate service time, so we will
1949                          * use the maxium timeout here for waiting the client
1950                          * sending the next req */
1951                         to = max((int)at_est2timeout(
1952                                  at_get(&svcpt->scp_at_estimate)),
1953                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1954                         /* Add 2 net_latency, one for balance rq_deadline
1955                          * (see ptl_send_rpc), one for resend the req to server,
1956                          * Note: client will pack net_latency in replay req
1957                          * (see ptlrpc_replay_req) */
1958                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
1959                 }
1960                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1961         }
1962 reqcopy_put:
1963         RETURN(rc);
1964 }
1965
1966 static int target_recovery_thread(void *arg)
1967 {
1968         struct lu_target *lut = arg;
1969         struct obd_device *obd = lut->lut_obd;
1970         struct ptlrpc_request *req;
1971         struct target_recovery_data *trd = &obd->obd_recovery_data;
1972         unsigned long delta;
1973         struct lu_env *env;
1974         struct ptlrpc_thread *thread = NULL;
1975         int rc = 0;
1976         ENTRY;
1977
1978         unshare_fs_struct();
1979         OBD_ALLOC_PTR(thread);
1980         if (thread == NULL)
1981                 RETURN(-ENOMEM);
1982
1983         OBD_ALLOC_PTR(env);
1984         if (env == NULL) {
1985                 OBD_FREE_PTR(thread);
1986                 RETURN(-ENOMEM);
1987         }
1988
1989         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1990         if (rc) {
1991                 OBD_FREE_PTR(thread);
1992                 OBD_FREE_PTR(env);
1993                 RETURN(rc);
1994         }
1995
1996         thread->t_env = env;
1997         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1998         env->le_ctx.lc_thread = thread;
1999         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2000         thread->t_watchdog = NULL;
2001
2002         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2003                current_pid());
2004         trd->trd_processing_task = current_pid();
2005
2006         spin_lock(&obd->obd_dev_lock);
2007         obd->obd_recovering = 1;
2008         spin_unlock(&obd->obd_dev_lock);
2009         complete(&trd->trd_starting);
2010
2011         /* first of all, we have to know the first transno to replay */
2012         if (target_recovery_overseer(obd, check_for_clients,
2013                                      exp_connect_healthy)) {
2014                 abort_req_replay_queue(obd);
2015                 abort_lock_replay_queue(obd);
2016         }
2017
2018         /* next stage: replay requests */
2019         delta = jiffies;
2020         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2021                cfs_atomic_read(&obd->obd_req_replay_clients),
2022                obd->obd_next_recovery_transno);
2023         while ((req = target_next_replay_req(obd))) {
2024                 LASSERT(trd->trd_processing_task == current_pid());
2025                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2026                           lustre_msg_get_transno(req->rq_reqmsg),
2027                           libcfs_nid2str(req->rq_peer.nid));
2028                 handle_recovery_req(thread, req,
2029                                     trd->trd_recovery_handler);
2030                 /**
2031                  * bz18031: increase next_recovery_transno before
2032                  * target_request_copy_put() will drop exp_rpc reference
2033                  */
2034                 spin_lock(&obd->obd_recovery_task_lock);
2035                 obd->obd_next_recovery_transno++;
2036                 spin_unlock(&obd->obd_recovery_task_lock);
2037                 target_exp_dequeue_req_replay(req);
2038                 target_request_copy_put(req);
2039                 obd->obd_replayed_requests++;
2040         }
2041
2042         /**
2043          * The second stage: replay locks
2044          */
2045         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2046                cfs_atomic_read(&obd->obd_lock_replay_clients));
2047         while ((req = target_next_replay_lock(obd))) {
2048                 LASSERT(trd->trd_processing_task == current_pid());
2049                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2050                           libcfs_nid2str(req->rq_peer.nid));
2051                 handle_recovery_req(thread, req,
2052                                     trd->trd_recovery_handler);
2053                 target_request_copy_put(req);
2054                 obd->obd_replayed_locks++;
2055         }
2056
2057         /**
2058          * The third stage: reply on final pings, at this moment all clients
2059          * must have request in final queue
2060          */
2061         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2062         /** Update server last boot epoch */
2063         tgt_boot_epoch_update(lut);
2064         /* We drop recoverying flag to forward all new requests
2065          * to regular mds_handle() since now */
2066         spin_lock(&obd->obd_dev_lock);
2067         obd->obd_recovering = obd->obd_abort_recovery = 0;
2068         spin_unlock(&obd->obd_dev_lock);
2069         spin_lock(&obd->obd_recovery_task_lock);
2070         target_cancel_recovery_timer(obd);
2071         spin_unlock(&obd->obd_recovery_task_lock);
2072         while ((req = target_next_final_ping(obd))) {
2073                 LASSERT(trd->trd_processing_task == current_pid());
2074                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2075                           libcfs_nid2str(req->rq_peer.nid));
2076                 handle_recovery_req(thread, req,
2077                                     trd->trd_recovery_handler);
2078                 target_request_copy_put(req);
2079         }
2080
2081         delta = (jiffies - delta) / HZ;
2082         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2083               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2084         if (delta > OBD_RECOVERY_TIME_SOFT) {
2085                 CWARN("too long recovery - read logs\n");
2086                 libcfs_debug_dumplog();
2087         }
2088
2089         target_finish_recovery(obd);
2090
2091         lu_context_fini(&env->le_ctx);
2092         trd->trd_processing_task = 0;
2093         complete(&trd->trd_finishing);
2094
2095         tgt_io_thread_done(thread);
2096         OBD_FREE_PTR(thread);
2097         OBD_FREE_PTR(env);
2098         RETURN(rc);
2099 }
2100
2101 static int target_start_recovery_thread(struct lu_target *lut,
2102                                         svc_handler_t handler)
2103 {
2104         struct obd_device *obd = lut->lut_obd;
2105         int rc = 0;
2106         struct target_recovery_data *trd = &obd->obd_recovery_data;
2107
2108         memset(trd, 0, sizeof(*trd));
2109         init_completion(&trd->trd_starting);
2110         init_completion(&trd->trd_finishing);
2111         trd->trd_recovery_handler = handler;
2112
2113         if (!IS_ERR(kthread_run(target_recovery_thread,
2114                                 lut, "tgt_recov"))) {
2115                 wait_for_completion(&trd->trd_starting);
2116                 LASSERT(obd->obd_recovering != 0);
2117         } else {
2118                 rc = -ECHILD;
2119         }
2120
2121         return rc;
2122 }
2123
2124 void target_stop_recovery_thread(struct obd_device *obd)
2125 {
2126         if (obd->obd_recovery_data.trd_processing_task > 0) {
2127                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2128                 /** recovery can be done but postrecovery is not yet */
2129                 spin_lock(&obd->obd_dev_lock);
2130                 if (obd->obd_recovering) {
2131                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2132                         obd->obd_abort_recovery = 1;
2133                         wake_up(&obd->obd_next_transno_waitq);
2134                 }
2135                 spin_unlock(&obd->obd_dev_lock);
2136                 wait_for_completion(&trd->trd_finishing);
2137         }
2138 }
2139 EXPORT_SYMBOL(target_stop_recovery_thread);
2140
2141 void target_recovery_fini(struct obd_device *obd)
2142 {
2143         class_disconnect_exports(obd);
2144         target_stop_recovery_thread(obd);
2145         target_cleanup_recovery(obd);
2146 }
2147 EXPORT_SYMBOL(target_recovery_fini);
2148
2149 static void target_recovery_expired(unsigned long castmeharder)
2150 {
2151         struct obd_device *obd = (struct obd_device *)castmeharder;
2152         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2153                " after %lds (%d clients connected)\n",
2154                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2155                cfs_time_current_sec()- obd->obd_recovery_start,
2156                cfs_atomic_read(&obd->obd_connected_clients));
2157
2158         obd->obd_recovery_expired = 1;
2159         wake_up(&obd->obd_next_transno_waitq);
2160 }
2161
2162 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2163 {
2164         struct obd_device *obd = lut->lut_obd;
2165         if (obd->obd_max_recoverable_clients == 0) {
2166                 /** Update server last boot epoch */
2167                 tgt_boot_epoch_update(lut);
2168                 return;
2169         }
2170
2171         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2172                "last_transno "LPU64"\n", obd->obd_name,
2173                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2174         LASSERT(obd->obd_stopping == 0);
2175         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2176         obd->obd_recovery_start = 0;
2177         obd->obd_recovery_end = 0;
2178
2179         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2180         target_start_recovery_thread(lut, handler);
2181 }
2182 EXPORT_SYMBOL(target_recovery_init);
2183
2184 #endif /* __KERNEL__ */
2185
2186 static int target_process_req_flags(struct obd_device *obd,
2187                                     struct ptlrpc_request *req)
2188 {
2189         struct obd_export *exp = req->rq_export;
2190         LASSERT(exp != NULL);
2191         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2192                 /* client declares he's ready to replay locks */
2193                 spin_lock(&exp->exp_lock);
2194                 if (exp->exp_req_replay_needed) {
2195                         exp->exp_req_replay_needed = 0;
2196                         spin_unlock(&exp->exp_lock);
2197
2198                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2199                         cfs_atomic_dec(&obd->obd_req_replay_clients);
2200                 } else {
2201                         spin_unlock(&exp->exp_lock);
2202                 }
2203         }
2204         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2205                 /* client declares he's ready to complete recovery
2206                  * so, we put the request on th final queue */
2207                 spin_lock(&exp->exp_lock);
2208                 if (exp->exp_lock_replay_needed) {
2209                         exp->exp_lock_replay_needed = 0;
2210                         spin_unlock(&exp->exp_lock);
2211
2212                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2213                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
2214                 } else {
2215                         spin_unlock(&exp->exp_lock);
2216                 }
2217         }
2218         return 0;
2219 }
2220
2221 int target_queue_recovery_request(struct ptlrpc_request *req,
2222                                   struct obd_device *obd)
2223 {
2224         cfs_list_t *tmp;
2225         int inserted = 0;
2226         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2227         ENTRY;
2228
2229         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2230                 /* Processing the queue right now, don't re-add. */
2231                 RETURN(1);
2232         }
2233
2234         target_process_req_flags(obd, req);
2235
2236         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2237                 /* client declares he's ready to complete recovery
2238                  * so, we put the request on th final queue */
2239                 target_request_copy_get(req);
2240                 DEBUG_REQ(D_HA, req, "queue final req");
2241                 wake_up(&obd->obd_next_transno_waitq);
2242                 spin_lock(&obd->obd_recovery_task_lock);
2243                 if (obd->obd_recovering) {
2244                         cfs_list_add_tail(&req->rq_list,
2245                                           &obd->obd_final_req_queue);
2246                 } else {
2247                         spin_unlock(&obd->obd_recovery_task_lock);
2248                         target_request_copy_put(req);
2249                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2250                 }
2251                 spin_unlock(&obd->obd_recovery_task_lock);
2252                 RETURN(0);
2253         }
2254         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2255                 /* client declares he's ready to replay locks */
2256                 target_request_copy_get(req);
2257                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2258                 wake_up(&obd->obd_next_transno_waitq);
2259                 spin_lock(&obd->obd_recovery_task_lock);
2260                 LASSERT(obd->obd_recovering);
2261                 /* usually due to recovery abort */
2262                 if (!req->rq_export->exp_in_recovery) {
2263                         spin_unlock(&obd->obd_recovery_task_lock);
2264                         target_request_copy_put(req);
2265                         RETURN(-ENOTCONN);
2266                 }
2267                 LASSERT(req->rq_export->exp_lock_replay_needed);
2268                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2269                 spin_unlock(&obd->obd_recovery_task_lock);
2270                 RETURN(0);
2271         }
2272
2273         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2274          * (i.e. buflens etc are in my own byte order), but type-dependent
2275          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2276
2277         if (!transno) {
2278                 CFS_INIT_LIST_HEAD(&req->rq_list);
2279                 DEBUG_REQ(D_HA, req, "not queueing");
2280                 RETURN(1);
2281         }
2282
2283         /* If we're processing the queue, we want don't want to queue this
2284          * message.
2285          *
2286          * Also, if this request has a transno less than the one we're waiting
2287          * for, we should process it now.  It could (and currently always will)
2288          * be an open request for a descriptor that was opened some time ago.
2289          *
2290          * Also, a resent, replayed request that has already been
2291          * handled will pass through here and be processed immediately.
2292          */
2293         CDEBUG(D_HA, "Next recovery transno: "LPU64
2294                ", current: "LPU64", replaying\n",
2295                obd->obd_next_recovery_transno, transno);
2296         spin_lock(&obd->obd_recovery_task_lock);
2297         if (transno < obd->obd_next_recovery_transno) {
2298                 /* Processing the queue right now, don't re-add. */
2299                 LASSERT(cfs_list_empty(&req->rq_list));
2300                 spin_unlock(&obd->obd_recovery_task_lock);
2301                 RETURN(1);
2302         }
2303         spin_unlock(&obd->obd_recovery_task_lock);
2304
2305         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2306                 RETURN(0);
2307
2308         target_request_copy_get(req);
2309         if (!req->rq_export->exp_in_recovery) {
2310                 target_request_copy_put(req);
2311                 RETURN(-ENOTCONN);
2312         }
2313         LASSERT(req->rq_export->exp_req_replay_needed);
2314
2315         if (target_exp_enqueue_req_replay(req)) {
2316                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2317                 target_request_copy_put(req);
2318                 RETURN(0);
2319         }
2320
2321         /* XXX O(n^2) */
2322         spin_lock(&obd->obd_recovery_task_lock);
2323         LASSERT(obd->obd_recovering);
2324         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2325                 struct ptlrpc_request *reqiter =
2326                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2327
2328                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2329                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2330                         inserted = 1;
2331                         break;
2332                 }
2333
2334                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2335                              transno)) {
2336                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2337                                   "has been claimed by another client");
2338                         spin_unlock(&obd->obd_recovery_task_lock);
2339                         target_exp_dequeue_req_replay(req);
2340                         target_request_copy_put(req);
2341                         RETURN(0);
2342                 }
2343         }
2344
2345         if (!inserted)
2346                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2347
2348         obd->obd_requests_queued_for_recovery++;
2349         spin_unlock(&obd->obd_recovery_task_lock);
2350         wake_up(&obd->obd_next_transno_waitq);
2351         RETURN(0);
2352 }
2353 EXPORT_SYMBOL(target_queue_recovery_request);
2354
2355 int target_handle_ping(struct ptlrpc_request *req)
2356 {
2357         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2358         return req_capsule_server_pack(&req->rq_pill);
2359 }
2360 EXPORT_SYMBOL(target_handle_ping);
2361
2362 void target_committed_to_req(struct ptlrpc_request *req)
2363 {
2364         struct obd_export *exp = req->rq_export;
2365
2366         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2367                 lustre_msg_set_last_committed(req->rq_repmsg,
2368                                               exp->exp_last_committed);
2369         else
2370                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2371                           "%d)", exp->exp_obd->obd_no_transno,
2372                           req->rq_repmsg == NULL);
2373
2374         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2375                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2376 }
2377 EXPORT_SYMBOL(target_committed_to_req);
2378
2379 #endif /* HAVE_SERVER_SUPPORT */
2380
2381 /**
2382  * Packs current SLV and Limit into \a req.
2383  */
2384 int target_pack_pool_reply(struct ptlrpc_request *req)
2385 {
2386         struct obd_device *obd;
2387         ENTRY;
2388
2389         /* Check that we still have all structures alive as this may
2390          * be some late RPC at shutdown time. */
2391         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2392                      !exp_connect_lru_resize(req->rq_export))) {
2393                 lustre_msg_set_slv(req->rq_repmsg, 0);
2394                 lustre_msg_set_limit(req->rq_repmsg, 0);
2395                 RETURN(0);
2396         }
2397
2398         /* OBD is alive here as export is alive, which we checked above. */
2399         obd = req->rq_export->exp_obd;
2400
2401         read_lock(&obd->obd_pool_lock);
2402         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2403         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2404         read_unlock(&obd->obd_pool_lock);
2405
2406         RETURN(0);
2407 }
2408 EXPORT_SYMBOL(target_pack_pool_reply);
2409
2410 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2411 {
2412         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2413                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2414                 return (-ECOMM);
2415         }
2416
2417         if (unlikely(rc)) {
2418                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2419                 req->rq_status = rc;
2420                 return (ptlrpc_send_error(req, 1));
2421         } else {
2422                 DEBUG_REQ(D_NET, req, "sending reply");
2423         }
2424
2425         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2426 }
2427
2428 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2429 {
2430         struct ptlrpc_service_part *svcpt;
2431         int                        netrc;
2432         struct ptlrpc_reply_state *rs;
2433         struct obd_export         *exp;
2434         ENTRY;
2435
2436         if (req->rq_no_reply) {
2437                 EXIT;
2438                 return;
2439         }
2440
2441         svcpt = req->rq_rqbd->rqbd_svcpt;
2442         rs = req->rq_reply_state;
2443         if (rs == NULL || !rs->rs_difficult) {
2444                 /* no notifiers */
2445                 target_send_reply_msg (req, rc, fail_id);
2446                 EXIT;
2447                 return;
2448         }
2449
2450         /* must be an export if locks saved */
2451         LASSERT (req->rq_export != NULL);
2452         /* req/reply consistent */
2453         LASSERT(rs->rs_svcpt == svcpt);
2454
2455         /* "fresh" reply */
2456         LASSERT (!rs->rs_scheduled);
2457         LASSERT (!rs->rs_scheduled_ever);
2458         LASSERT (!rs->rs_handled);
2459         LASSERT (!rs->rs_on_net);
2460         LASSERT (rs->rs_export == NULL);
2461         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2462         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2463
2464         exp = class_export_get (req->rq_export);
2465
2466         /* disable reply scheduling while I'm setting up */
2467         rs->rs_scheduled = 1;
2468         rs->rs_on_net    = 1;
2469         rs->rs_xid       = req->rq_xid;
2470         rs->rs_transno   = req->rq_transno;
2471         rs->rs_export    = exp;
2472         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2473
2474         spin_lock(&exp->exp_uncommitted_replies_lock);
2475         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2476                rs->rs_transno, exp->exp_last_committed);
2477         if (rs->rs_transno > exp->exp_last_committed) {
2478                 /* not committed already */
2479                 cfs_list_add_tail(&rs->rs_obd_list,
2480                                   &exp->exp_uncommitted_replies);
2481         }
2482         spin_unlock(&exp->exp_uncommitted_replies_lock);
2483
2484         spin_lock(&exp->exp_lock);
2485         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2486         spin_unlock(&exp->exp_lock);
2487
2488         netrc = target_send_reply_msg(req, rc, fail_id);
2489
2490         spin_lock(&svcpt->scp_rep_lock);
2491
2492         cfs_atomic_inc(&svcpt->scp_nreps_difficult);
2493
2494         if (netrc != 0) {
2495                 /* error sending: reply is off the net.  Also we need +1
2496                  * reply ref until ptlrpc_handle_rs() is done
2497                  * with the reply state (if the send was successful, there
2498                  * would have been +1 ref for the net, which
2499                  * reply_out_callback leaves alone) */
2500                 rs->rs_on_net = 0;
2501                 ptlrpc_rs_addref(rs);
2502         }
2503
2504         spin_lock(&rs->rs_lock);
2505         if (rs->rs_transno <= exp->exp_last_committed ||
2506             (!rs->rs_on_net && !rs->rs_no_ack) ||
2507             cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2508             cfs_list_empty(&rs->rs_obd_list)) {
2509                 CDEBUG(D_HA, "Schedule reply immediately\n");
2510                 ptlrpc_dispatch_difficult_reply(rs);
2511         } else {
2512                 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
2513                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2514         }
2515         spin_unlock(&rs->rs_lock);
2516         spin_unlock(&svcpt->scp_rep_lock);
2517         EXIT;
2518 }
2519 EXPORT_SYMBOL(target_send_reply);
2520
2521 ldlm_mode_t lck_compat_array[] = {
2522         [LCK_EX] LCK_COMPAT_EX,
2523         [LCK_PW] LCK_COMPAT_PW,
2524         [LCK_PR] LCK_COMPAT_PR,
2525         [LCK_CW] LCK_COMPAT_CW,
2526         [LCK_CR] LCK_COMPAT_CR,
2527         [LCK_NL] LCK_COMPAT_NL,
2528         [LCK_GROUP] LCK_COMPAT_GROUP,
2529         [LCK_COS] LCK_COMPAT_COS,
2530 };
2531
2532 /**
2533  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2534  * not escape to the user level.
2535  */
2536 int ldlm_error2errno(ldlm_error_t error)
2537 {
2538         int result;
2539
2540         switch (error) {
2541         case ELDLM_OK:
2542                 result = 0;
2543                 break;
2544         case ELDLM_LOCK_CHANGED:
2545                 result = -ESTALE;
2546                 break;
2547         case ELDLM_LOCK_ABORTED:
2548                 result = -ENAVAIL;
2549                 break;
2550         case ELDLM_LOCK_REPLACED:
2551                 result = -ESRCH;
2552                 break;
2553         case ELDLM_NO_LOCK_DATA:
2554                 result = -ENOENT;
2555                 break;
2556         case ELDLM_NAMESPACE_EXISTS:
2557                 result = -EEXIST;
2558                 break;
2559         case ELDLM_BAD_NAMESPACE:
2560                 result = -EBADF;
2561                 break;
2562         default:
2563                 if (((int)error) < 0)  /* cast to signed type */
2564                         result = error; /* as ldlm_error_t can be unsigned */
2565                 else {
2566                         CERROR("Invalid DLM result code: %d\n", error);
2567                         result = -EPROTO;
2568                 }
2569         }
2570         return result;
2571 }
2572 EXPORT_SYMBOL(ldlm_error2errno);
2573
2574 /**
2575  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2576  */
2577 ldlm_error_t ldlm_errno2error(int err_no)
2578 {
2579         int error;
2580
2581         switch (err_no) {
2582         case 0:
2583                 error = ELDLM_OK;
2584                 break;
2585         case -ESTALE:
2586                 error = ELDLM_LOCK_CHANGED;
2587                 break;
2588         case -ENAVAIL:
2589                 error = ELDLM_LOCK_ABORTED;
2590                 break;
2591         case -ESRCH:
2592                 error = ELDLM_LOCK_REPLACED;
2593                 break;
2594         case -ENOENT:
2595                 error = ELDLM_NO_LOCK_DATA;
2596                 break;
2597         case -EEXIST:
2598                 error = ELDLM_NAMESPACE_EXISTS;
2599                 break;
2600         case -EBADF:
2601                 error = ELDLM_BAD_NAMESPACE;
2602                 break;
2603         default:
2604                 error = err_no;
2605         }
2606         return error;
2607 }
2608 EXPORT_SYMBOL(ldlm_errno2error);
2609
2610 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2611 void ldlm_dump_export_locks(struct obd_export *exp)
2612 {
2613         spin_lock(&exp->exp_locks_list_guard);
2614         if (!cfs_list_empty(&exp->exp_locks_list)) {
2615                 struct ldlm_lock *lock;
2616
2617                 CERROR("dumping locks for export %p,"
2618                        "ignore if the unmount doesn't hang\n", exp);
2619                 cfs_list_for_each_entry(lock, &exp->exp_locks_list,
2620                                         l_exp_refs_link)
2621                         LDLM_ERROR(lock, "lock:");
2622         }
2623         spin_unlock(&exp->exp_locks_list_guard);
2624 }
2625 #endif
2626
2627 #ifdef HAVE_SERVER_SUPPORT
2628 static int target_bulk_timeout(void *data)
2629 {
2630         ENTRY;
2631         /* We don't fail the connection here, because having the export
2632          * killed makes the (vital) call to commitrw very sad.
2633          */
2634         RETURN(1);
2635 }
2636
2637 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2638 {
2639         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2640 }
2641
2642 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2643                    struct l_wait_info *lwi)
2644 {
2645         struct ptlrpc_request *req = desc->bd_req;
2646         int rc = 0;
2647         ENTRY;
2648
2649         /* If there is eviction in progress, wait for it to finish. */
2650         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2651                 *lwi = LWI_INTR(NULL, NULL);
2652                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2653                                   !cfs_atomic_read(&exp->exp_obd->
2654                                                    obd_evict_inprogress),
2655                                   lwi);
2656         }
2657
2658         /* Check if client was evicted or tried to reconnect already. */
2659         if (exp->exp_failed || exp->exp_abort_active_req) {
2660                 rc = -ENOTCONN;
2661         } else {
2662                 if (desc->bd_type == BULK_PUT_SINK)
2663                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2664                 if (rc == 0)
2665                         rc = ptlrpc_start_bulk_transfer(desc);
2666         }
2667
2668         if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2669                 ptlrpc_abort_bulk(desc);
2670         } else if (rc == 0) {
2671                 time_t start = cfs_time_current_sec();
2672                 do {
2673                         long timeoutl = req->rq_deadline - cfs_time_current_sec();
2674                         cfs_duration_t timeout = timeoutl <= 0 ?
2675                                 CFS_TICK : cfs_time_seconds(timeoutl);
2676                         *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2677                                                     cfs_time_seconds(1),
2678                                                    target_bulk_timeout,
2679                                                    desc);
2680                         rc = l_wait_event(desc->bd_waitq,
2681                                           !ptlrpc_server_bulk_active(desc) ||
2682                                           exp->exp_failed ||
2683                                           exp->exp_abort_active_req,
2684                                           lwi);
2685                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
2686                         /* Wait again if we changed deadline. */
2687                 } while ((rc == -ETIMEDOUT) &&
2688                          (req->rq_deadline > cfs_time_current_sec()));
2689
2690                 if (rc == -ETIMEDOUT) {
2691                         DEBUG_REQ(D_ERROR, req,
2692                                   "timeout on bulk %s after %ld%+lds",
2693                                   bulk2type(desc),
2694                                   req->rq_deadline - start,
2695                                   cfs_time_current_sec() -
2696                                   req->rq_deadline);
2697                         ptlrpc_abort_bulk(desc);
2698                 } else if (exp->exp_failed) {
2699                         DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2700                                   bulk2type(desc));
2701                         rc = -ENOTCONN;
2702                         ptlrpc_abort_bulk(desc);
2703                 } else if (exp->exp_abort_active_req) {
2704                         DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2705                                   bulk2type(desc));
2706                         /* We don't reply anyway. */
2707                         rc = -ETIMEDOUT;
2708                         ptlrpc_abort_bulk(desc);
2709                 } else if (desc->bd_failure ||
2710                            desc->bd_nob_transferred != desc->bd_nob) {
2711                         DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2712                                   desc->bd_failure ?
2713                                   "network error on" : "truncated",
2714                                   bulk2type(desc),
2715                                   desc->bd_nob_transferred,
2716                                   desc->bd_nob);
2717                         /* XXX Should this be a different errno? */
2718                         rc = -ETIMEDOUT;
2719                 } else if (desc->bd_type == BULK_GET_SINK) {
2720                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
2721                 }
2722         } else {
2723                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2724                           bulk2type(desc), rc);
2725         }
2726
2727         RETURN(rc);
2728 }
2729 EXPORT_SYMBOL(target_bulk_io);
2730
2731 #endif /* HAVE_SERVER_SUPPORT */