Whamcloud - gitweb
b=1451
[fs/lustre-release.git] / lustre / ldlm / ldlm_lib.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  *  Copyright (c) 2003 Cluster File Systems, Inc.
5  *
6  *   This file is part of Lustre, http://www.lustre.org.
7  *
8  *   Lustre is free software; you can redistribute it and/or
9  *   modify it under the terms of version 2 of the GNU General Public
10  *   License as published by the Free Software Foundation.
11  *
12  *   Lustre is distributed in the hope that it will be useful,
13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  *   GNU General Public License for more details.
16  *
17  *   You should have received a copy of the GNU General Public License
18  *   along with Lustre; if not, write to the Free Software
19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20  */
21
22 #ifndef EXPORT_SYMTAB
23 # define EXPORT_SYMTAB
24 #endif
25 #define DEBUG_SUBSYSTEM S_LDLM
26
27 #ifdef __KERNEL__
28 # include <linux/module.h>
29 #else
30 # include <liblustre.h>
31 #endif
32 #include <linux/obd.h>
33 #include <linux/obd_ost.h> /* for LUSTRE_OSC_NAME */
34 #include <linux/lustre_mds.h> /* for LUSTRE_MDC_NAME */
35 #include <linux/lustre_mgmt.h>
36 #include <linux/lustre_dlm.h>
37 #include <linux/lustre_net.h>
38
39 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
40 {
41         struct ptlrpc_connection *conn;
42         struct lustre_cfg* lcfg = buf;
43         struct client_obd *cli = &obddev->u.cli;
44         struct obd_import *imp;
45         struct obd_uuid server_uuid;
46         int rq_portal, rp_portal, connect_op;
47         char *name = obddev->obd_type->typ_name;
48         char *mgmt_name = NULL;
49         int rc;
50         struct obd_device *mgmt_obd;
51         mgmtcli_register_for_events_t register_f;
52         ENTRY;
53
54         /* In a more perfect world, we would hang a ptlrpc_client off of
55          * obd_type and just use the values from there. */
56         if (!strcmp(name, LUSTRE_OSC_NAME)) {
57                 rq_portal = OST_REQUEST_PORTAL;
58                 rp_portal = OSC_REPLY_PORTAL;
59                 connect_op = OST_CONNECT;
60         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
61                 rq_portal = MDS_REQUEST_PORTAL;
62                 rp_portal = MDC_REPLY_PORTAL;
63                 connect_op = MDS_CONNECT;
64         } else if (!strcmp(name, LUSTRE_MGMTCLI_NAME)) {
65                 rq_portal = MGMT_REQUEST_PORTAL;
66                 rp_portal = MGMT_REPLY_PORTAL;
67                 connect_op = MGMT_CONNECT;
68         } else {
69                 CERROR("unknown client OBD type \"%s\", can't setup\n",
70                        name);
71                 RETURN(-EINVAL);
72         }
73
74         if (lcfg->lcfg_inllen1 < 1) {
75                 CERROR("requires a TARGET UUID\n");
76                 RETURN(-EINVAL);
77         }
78
79         if (lcfg->lcfg_inllen1 > 37) {
80                 CERROR("client UUID must be less than 38 characters\n");
81                 RETURN(-EINVAL);
82         }
83
84         if (lcfg->lcfg_inllen2 < 1) {
85                 CERROR("setup requires a SERVER UUID\n");
86                 RETURN(-EINVAL);
87         }
88
89         if (lcfg->lcfg_inllen2 > 37) {
90                 CERROR("target UUID must be less than 38 characters\n");
91                 RETURN(-EINVAL);
92         }
93
94         sema_init(&cli->cl_sem, 1);
95         cli->cl_conn_count = 0;
96         memcpy(server_uuid.uuid, lcfg->lcfg_inlbuf2,
97                min_t(unsigned int, lcfg->lcfg_inllen2, sizeof(server_uuid)));
98
99         cli->cl_dirty = 0;
100         cli->cl_avail_grant = 0;
101         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
102         INIT_LIST_HEAD(&cli->cl_cache_waiters);
103         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
104         INIT_LIST_HEAD(&cli->cl_loi_write_list);
105         INIT_LIST_HEAD(&cli->cl_loi_read_list);
106         spin_lock_init(&cli->cl_loi_list_lock);
107         cli->cl_brw_in_flight = 0;
108         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
109         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
110         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
111         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
112         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
113         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
114
115         rc = ldlm_get_ref();
116         if (rc) {
117                 CERROR("ldlm_get_ref failed: %d\n", rc);
118                 GOTO(err, rc);
119         }
120
121         conn = ptlrpc_uuid_to_connection(&server_uuid);
122         if (conn == NULL)
123                 GOTO(err_ldlm, rc = -ENOENT);
124
125         ptlrpc_init_client(rq_portal, rp_portal, name,
126                            &obddev->obd_ldlm_client);
127
128         imp = class_new_import();
129         if (imp == NULL) {
130                 ptlrpc_put_connection(conn);
131                 GOTO(err_ldlm, rc = -ENOENT);
132         }
133         imp->imp_connection = conn;
134         imp->imp_client = &obddev->obd_ldlm_client;
135         imp->imp_obd = obddev;
136         imp->imp_connect_op = connect_op;
137         imp->imp_generation = 0;
138         imp->imp_initial_recov = 1;
139         INIT_LIST_HEAD(&imp->imp_pinger_chain);
140         memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
141               lcfg->lcfg_inllen1);
142         class_import_put(imp);
143
144         cli->cl_import = imp;
145         cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
146         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
147         cli->cl_sandev = to_kdev_t(0);
148
149         if (lcfg->lcfg_inllen3 != 0) {
150                 if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
151                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
152                                name, obddev->obd_name,
153                                imp->imp_target_uuid.uuid);
154                         imp->imp_invalid = 1;
155
156                         if (lcfg->lcfg_inllen4 != 0)
157                                 mgmt_name = lcfg->lcfg_inlbuf4;
158                 } else {
159                         mgmt_name = lcfg->lcfg_inlbuf3;
160                 }
161         }
162
163         if (mgmt_name != NULL) {
164                 /* Register with management client if we need to. */
165                 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
166                        obddev->obd_name, mgmt_name, server_uuid.uuid);
167
168                 mgmt_obd = class_name2obd(mgmt_name);
169                 if (!mgmt_obd) {
170                         CERROR("can't find mgmtcli %s to register\n",
171                                mgmt_name);
172                         GOTO(err_import, rc = -ENOSYS);
173                 }
174
175                 register_f = inter_module_get("mgmtcli_register_for_events");
176                 if (!register_f) {
177                         CERROR("can't i_m_g mgmtcli_register_for_events\n");
178                         GOTO(err_import, rc = -ENOSYS);
179                 }
180
181                 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
182                 inter_module_put("mgmtcli_register_for_events");
183
184                 if (!rc)
185                         cli->cl_mgmtcli_obd = mgmt_obd;
186         }
187
188         RETURN(rc);
189
190 err_import:
191         class_destroy_import(imp);
192 err_ldlm:
193         ldlm_put_ref(0);
194 err:
195         RETURN(rc);
196
197 }
198
199 int client_obd_cleanup(struct obd_device *obddev, int flags)
200 {
201         struct client_obd *cli = &obddev->u.cli;
202
203         if (!cli->cl_import)
204                 RETURN(-EINVAL);
205         if (cli->cl_mgmtcli_obd) {
206                 mgmtcli_deregister_for_events_t dereg_f;
207
208                 dereg_f = inter_module_get("mgmtcli_deregister_for_events");
209                 dereg_f(cli->cl_mgmtcli_obd, obddev);
210                 inter_module_put("mgmtcli_deregister_for_events");
211         }
212         class_destroy_import(cli->cl_import);
213         cli->cl_import = NULL;
214
215         ldlm_put_ref(flags & OBD_OPT_FORCE);
216
217         RETURN(0);
218 }
219
220 int client_connect_import(struct lustre_handle *dlm_handle,
221                           struct obd_device *obd,
222                           struct obd_uuid *cluuid)
223 {
224         struct client_obd *cli = &obd->u.cli;
225         struct obd_import *imp = cli->cl_import;
226         struct obd_export *exp;
227         int rc;
228         ENTRY;
229
230         down(&cli->cl_sem);
231         rc = class_connect(dlm_handle, obd, cluuid);
232         if (rc)
233                 GOTO(out_sem, rc);
234
235         cli->cl_conn_count++;
236         if (cli->cl_conn_count > 1)
237                 GOTO(out_sem, rc);
238         exp = class_conn2export(dlm_handle);
239
240         if (obd->obd_namespace != NULL)
241                 CERROR("already have namespace!\n");
242         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
243                                                 LDLM_NAMESPACE_CLIENT);
244         if (obd->obd_namespace == NULL)
245                 GOTO(out_disco, rc = -ENOMEM);
246
247         imp->imp_dlm_handle = *dlm_handle;
248         rc = ptlrpc_init_import(imp);
249         if (rc != 0) 
250                 GOTO(out_ldlm, rc);
251
252         exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
253         rc = ptlrpc_connect_import(imp, NULL);
254         if (rc != 0) {
255                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
256                 GOTO(out_ldlm, rc);
257         }
258
259         ptlrpc_pinger_add_import(imp);
260         EXIT;
261
262         if (rc) {
263 out_ldlm:
264                 ldlm_namespace_free(obd->obd_namespace, 0);
265                 obd->obd_namespace = NULL;
266 out_disco:
267                 cli->cl_conn_count--;
268                 class_disconnect(exp, 0);
269         } else {
270                 class_export_put(exp);
271         }
272 out_sem:
273         up(&cli->cl_sem);
274         return rc;
275 }
276
277 int client_disconnect_export(struct obd_export *exp, int failover)
278 {
279         struct obd_device *obd = class_exp2obd(exp);
280         struct client_obd *cli = &obd->u.cli;
281         struct obd_import *imp = cli->cl_import;
282         int rc = 0, err;
283         ENTRY;
284
285         if (!obd) {
286                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
287                        exp, exp ? exp->exp_handle.h_cookie : -1);
288                 RETURN(-EINVAL);
289         }
290
291         down(&cli->cl_sem);
292         if (!cli->cl_conn_count) {
293                 CERROR("disconnecting disconnected device (%s)\n",
294                        obd->obd_name);
295                 GOTO(out_sem, rc = -EINVAL);
296         }
297
298         cli->cl_conn_count--;
299         if (cli->cl_conn_count)
300                 GOTO(out_no_disconnect, rc = 0);
301
302         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
303          * delete it regardless.  (It's safe to delete an import that was
304          * never added.) */
305         (void)ptlrpc_pinger_del_import(imp);
306
307         if (obd->obd_namespace != NULL) {
308                 /* obd_no_recov == local only */
309                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
310                                        obd->obd_no_recov, NULL);
311                 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
312                 obd->obd_namespace = NULL;
313         }
314
315         /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
316         if (obd->obd_no_recov)
317                 ptlrpc_invalidate_import(imp, 0);
318         else
319                 rc = ptlrpc_disconnect_import(imp);
320
321         EXIT;
322  out_no_disconnect:
323         err = class_disconnect(exp, 0);
324         if (!rc && err)
325                 rc = err;
326  out_sem:
327         up(&cli->cl_sem);
328         RETURN(rc);
329 }
330
331 /* --------------------------------------------------------------------------
332  * from old lib/target.c
333  * -------------------------------------------------------------------------- */
334
335 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
336                             struct obd_uuid *cluuid, int initial_conn)
337 {
338         if (exp->exp_connection && !initial_conn) {
339                 struct lustre_handle *hdl;
340                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
341                 /* Might be a re-connect after a partition. */
342                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
343                         CERROR("%s reconnecting\n", cluuid->uuid);
344                         conn->cookie = exp->exp_handle.h_cookie;
345                         RETURN(EALREADY);
346                 } else {
347                         CERROR("%s reconnecting from %s, "
348                                "handle mismatch (ours "LPX64", theirs "
349                                LPX64")\n", cluuid->uuid,
350                                exp->exp_connection->c_remote_uuid.uuid,
351                                hdl->cookie, conn->cookie);
352                         memset(conn, 0, sizeof *conn);
353                         RETURN(-EALREADY);
354                 }
355         }
356
357         conn->cookie = exp->exp_handle.h_cookie;
358         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
359                cluuid->uuid, exp);
360         CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
361         RETURN(0);
362 }
363
364 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
365 {
366         struct obd_device *target;
367         struct obd_export *export = NULL;
368         struct obd_import *revimp;
369         struct lustre_handle conn;
370         struct obd_uuid tgtuuid;
371         struct obd_uuid cluuid;
372         struct obd_uuid remote_uuid;
373         struct list_head *p;
374         char *str, *tmp;
375         int rc = 0, abort_recovery;
376         unsigned long flags;
377         int initial_conn = 0;
378         char peer_str[PTL_NALFMT_SIZE];
379         ENTRY;
380
381         OBD_RACE(OBD_FAIL_TGT_CONN_RACE); 
382
383         LASSERT_REQSWAB (req, 0);
384         str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
385         if (str == NULL) {
386                 CERROR("bad target UUID for connect\n");
387                 GOTO(out, rc = -EINVAL);
388         }
389
390         obd_str2uuid (&tgtuuid, str);
391         target = class_uuid2obd(&tgtuuid);
392         if (!target) {
393                 target = class_name2obd(str);
394         }
395         
396         if (!target || target->obd_stopping || !target->obd_set_up) {
397                 CERROR("UUID '%s' is not available for connect\n", str);
398
399                 GOTO(out, rc = -ENODEV);
400         }
401
402         LASSERT_REQSWAB (req, 1);
403         str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
404         if (str == NULL) {
405                 CERROR("bad client UUID for connect\n");
406                 GOTO(out, rc = -EINVAL);
407         }
408
409         obd_str2uuid (&cluuid, str);
410
411         /* XXX extract a nettype and format accordingly */
412         switch (sizeof(ptl_nid_t)) {
413                 /* NB the casts only avoid compiler warnings */
414         case 8:
415                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
416                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_nid);
417                 break;
418         case 4:
419                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
420                          "NET_%x_UUID", (__u32)req->rq_peer.peer_nid);
421                 break;
422         default:
423                 LBUG();
424         }
425
426         spin_lock_bh(&target->obd_processing_task_lock);
427         abort_recovery = target->obd_abort_recovery;
428         spin_unlock_bh(&target->obd_processing_task_lock);
429         if (abort_recovery)
430                 target_abort_recovery(target);
431
432         tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
433         if (tmp == NULL)
434                 GOTO(out, rc = -EPROTO);
435
436         memcpy(&conn, tmp, sizeof conn);
437
438         rc = lustre_pack_reply(req, 0, NULL, NULL);
439         if (rc)
440                 GOTO(out, rc);
441         
442         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
443                 initial_conn = 1;
444         
445         /* lctl gets a backstage, all-access pass. */
446         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
447                 goto dont_check_exports;
448
449         spin_lock(&target->obd_dev_lock);
450         list_for_each(p, &target->obd_exports) {
451                 export = list_entry(p, struct obd_export, exp_obd_chain);
452                 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
453                         spin_unlock(&target->obd_dev_lock);
454                         LASSERT(export->exp_obd == target);
455
456                         rc = target_handle_reconnect(&conn, export, &cluuid,
457                                                      initial_conn);
458                         break;
459                 }
460                 export = NULL;
461         }
462         /* If we found an export, we already unlocked. */
463         if (!export) {
464                 spin_unlock(&target->obd_dev_lock);
465         } else if (req->rq_export == NULL && 
466                    atomic_read(&export->exp_rpc_count) > 0) {
467                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
468                       target->obd_name, cluuid.uuid,
469                       ptlrpc_peernid2str(&req->rq_peer, peer_str),
470                       export, atomic_read(&export->exp_refcount));
471                 GOTO(out, rc = -EBUSY);
472         } else if (req->rq_export != NULL &&
473                    atomic_read(&export->exp_rpc_count) > 1) {
474                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
475                       target->obd_name, cluuid.uuid,
476                       ptlrpc_peernid2str(&req->rq_peer, peer_str),
477                       export, atomic_read(&export->exp_rpc_count));
478                 GOTO(out, rc = -EBUSY);
479         }
480         else if (req->rq_reqmsg->conn_cnt == 1 && !initial_conn) {
481                 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
482                        cluuid.uuid);
483                 GOTO(out, rc = -EALREADY);
484         }
485
486         /* Tell the client if we're in recovery. */
487         /* If this is the first client, start the recovery timer */
488         CWARN("%s: connection from %s@%s %s\n", target->obd_name, cluuid.uuid,
489               ptlrpc_peernid2str(&req->rq_peer, peer_str),
490               target->obd_recovering ? "(recovering)" : "");
491         if (target->obd_recovering) {
492                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
493                 target_start_recovery_timer(target, handler);
494         }
495 #if 0
496         /* Tell the client if we support replayable requests */
497         if (target->obd_replayable)
498                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
499 #endif
500         if (export == NULL) {
501                 if (target->obd_recovering) {
502                         CERROR("denying connection for new client %s@%s: "
503                                "%d clients in recovery for %lds\n", cluuid.uuid,
504                                ptlrpc_peernid2str(&req->rq_peer, peer_str),
505                                target->obd_recoverable_clients,
506                                (target->obd_recovery_timer.expires-jiffies)/HZ);
507                         rc = -EBUSY;
508                 } else {
509  dont_check_exports:
510                         rc = obd_connect(&conn, target, &cluuid);
511                 }
512         }
513         /* Tell the client if we support replayable requests */
514         if (target->obd_replayable)
515                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
516
517         /* If all else goes well, this is our RPC return code. */
518         req->rq_status = 0;
519
520         if (rc && rc != EALREADY)
521                 GOTO(out, rc);
522
523         req->rq_repmsg->handle = conn;
524
525         /* If the client and the server are the same node, we will already
526          * have an export that really points to the client's DLM export,
527          * because we have a shared handles table.
528          *
529          * XXX this will go away when shaver stops sending the "connect" handle
530          * in the real "remote handle" field of the request --phik 24 Apr 2003
531          */
532         if (req->rq_export != NULL)
533                 class_export_put(req->rq_export);
534
535         /* ownership of this export ref transfers to the request */
536         export = req->rq_export = class_conn2export(&conn);
537         LASSERT(export != NULL);
538
539         spin_lock_irqsave(&export->exp_lock, flags);
540         if (initial_conn) {
541                 req->rq_repmsg->conn_cnt = export->exp_conn_cnt + 1;
542         } else if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
543                 CERROR("%s@%s: already connected at a higher conn_cnt: %d > %d\n",
544                        cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str),
545                        export->exp_conn_cnt, 
546                        req->rq_reqmsg->conn_cnt);
547                 spin_unlock_irqrestore(&export->exp_lock, flags);
548                 GOTO(out, rc = -EALREADY);
549         } 
550         export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
551         spin_unlock_irqrestore(&export->exp_lock, flags);
552
553         /* request from liblustre? */
554         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
555                 export->exp_libclient = 1;
556
557         if (export->exp_connection != NULL)
558                 ptlrpc_put_connection(export->exp_connection);
559         export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
560                                                        &remote_uuid);
561
562         if (rc == EALREADY) {
563                 /* We indicate the reconnection in a flag, not an error code. */
564                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
565                 GOTO(out, rc = 0);
566         }
567
568         if (target->obd_recovering) {
569                 target->obd_connected_clients++;
570         }
571
572         memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
573                sizeof conn);
574
575         if (export->exp_imp_reverse != NULL)
576                 class_destroy_import(export->exp_imp_reverse);
577         revimp = export->exp_imp_reverse = class_new_import();
578         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
579         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
580         revimp->imp_remote_handle = conn;
581         revimp->imp_obd = target;
582         revimp->imp_dlm_fake = 1;
583         revimp->imp_state = LUSTRE_IMP_FULL;
584         class_import_put(revimp);
585 out:
586         if (rc)
587                 req->rq_status = rc;
588         RETURN(rc);
589 }
590
591 int target_handle_disconnect(struct ptlrpc_request *req)
592 {
593         struct obd_export *exp;
594         int rc;
595         ENTRY;
596
597         rc = lustre_pack_reply(req, 0, NULL, NULL);
598         if (rc)
599                 RETURN(rc);
600
601         /* keep the rq_export around so we can send the reply */
602         exp = class_export_get(req->rq_export);
603         req->rq_status = obd_disconnect(exp, 0);
604         RETURN(0);
605 }
606
607 void target_destroy_export(struct obd_export *exp)
608 {
609         /* exports created from last_rcvd data, and "fake"
610            exports created by lctl don't have an import */
611         if (exp->exp_imp_reverse != NULL)
612                 class_destroy_import(exp->exp_imp_reverse);
613
614         /* We cancel locks at disconnect time, but this will catch any locks
615          * granted in a race with recovery-induced disconnect. */
616         if (exp->exp_obd->obd_namespace != NULL)
617                 ldlm_cancel_locks_for_export(exp);
618 }
619
620 /*
621  * Recovery functions
622  */
623
624 static void abort_delayed_replies(struct obd_device *obd)
625 {
626         struct ptlrpc_request *req;
627         struct list_head *tmp, *n;
628         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
629                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
630                 DEBUG_REQ(D_ERROR, req, "aborted:");
631                 req->rq_status = -ENOTCONN;
632                 req->rq_type = PTL_RPC_MSG_ERR;
633                 ptlrpc_reply(req);
634                 class_export_put(req->rq_export);
635                 list_del(&req->rq_list);
636                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
637                 OBD_FREE(req, sizeof *req);
638         }
639 }
640
641 static void abort_recovery_queue(struct obd_device *obd)
642 {
643         struct ptlrpc_request *req;
644         struct list_head *tmp, *n;
645         int rc;
646
647         list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
648                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
649                 DEBUG_REQ(D_ERROR, req, "aborted:");
650                 req->rq_status = -ENOTCONN;
651                 req->rq_type = PTL_RPC_MSG_ERR;
652                 rc = lustre_pack_reply(req, 0, NULL, NULL);
653                 if (rc == 0) {
654                         ptlrpc_reply(req);
655                 } else {
656                         DEBUG_REQ(D_ERROR, req,
657                                   "packing failed for abort-reply; skipping");
658                 }
659                 list_del(&req->rq_list);
660                 class_export_put(req->rq_export);
661                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
662                 OBD_FREE(req, sizeof *req);
663         }
664 }
665
666 void target_abort_recovery(void *data)
667 {
668         struct obd_device *obd = data;
669         int rc;
670
671         CERROR("disconnecting clients and aborting recovery\n");
672         spin_lock_bh(&obd->obd_processing_task_lock);
673         if (!obd->obd_recovering) {
674                 spin_unlock_bh(&obd->obd_processing_task_lock);
675                 EXIT;
676                 return;
677         }
678
679         obd->obd_recovering = obd->obd_abort_recovery = 0;
680
681         wake_up(&obd->obd_next_transno_waitq);
682         target_cancel_recovery_timer(obd);
683         spin_unlock_bh(&obd->obd_processing_task_lock);
684
685         class_disconnect_exports(obd, 0);
686
687         /* when recovery was aborted, cleanup orphans on mds and ost */
688         if (OBT(obd) && OBP(obd, postrecov)) {
689                 rc = OBP(obd, postrecov)(obd);
690                 if (rc >= 0)
691                         CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
692                 else
693                         CERROR("postrecov failed %d\n", rc);
694         }
695
696         abort_delayed_replies(obd);
697         abort_recovery_queue(obd);
698         ptlrpc_run_recovery_over_upcall(obd);
699 }
700
701 static void target_recovery_expired(unsigned long castmeharder)
702 {
703         struct obd_device *obd = (struct obd_device *)castmeharder;
704         CERROR("recovery timed out, aborting\n");
705         spin_lock_bh(&obd->obd_processing_task_lock);
706         obd->obd_abort_recovery = 1;
707         wake_up(&obd->obd_next_transno_waitq);
708         spin_unlock_bh(&obd->obd_processing_task_lock);
709 }
710
711
712 /* obd_processing_task_lock should be held */
713 void target_cancel_recovery_timer(struct obd_device *obd)
714 {
715         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
716         del_timer(&obd->obd_recovery_timer);
717 }
718
719 static void reset_recovery_timer(struct obd_device *obd)
720 {
721         spin_lock_bh(&obd->obd_processing_task_lock);
722         if (!obd->obd_recovering) {
723                 spin_unlock_bh(&obd->obd_processing_task_lock);
724                 return;
725         }                
726         CDEBUG(D_HA, "timer will expire in %u seconds\n",
727                OBD_RECOVERY_TIMEOUT / HZ);
728         mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
729         spin_unlock_bh(&obd->obd_processing_task_lock);
730 }
731
732
733 /* Only start it the first time called */
734 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
735 {
736         spin_lock_bh(&obd->obd_processing_task_lock);
737         if (obd->obd_recovery_handler) {
738                 spin_unlock_bh(&obd->obd_processing_task_lock);
739                 return;
740         }
741         CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
742                OBD_RECOVERY_TIMEOUT / HZ);
743         obd->obd_recovery_handler = handler;
744         obd->obd_recovery_timer.function = target_recovery_expired;
745         obd->obd_recovery_timer.data = (unsigned long)obd;
746         spin_unlock_bh(&obd->obd_processing_task_lock);
747
748         reset_recovery_timer(obd);
749 }
750
751 static int check_for_next_transno(struct obd_device *obd)
752 {
753         struct ptlrpc_request *req;
754         int wake_up = 0, connected, completed, queue_len, max;
755         __u64 next_transno, req_transno;
756
757         spin_lock_bh(&obd->obd_processing_task_lock);
758         req = list_entry(obd->obd_recovery_queue.next,
759                          struct ptlrpc_request, rq_list);
760         max = obd->obd_max_recoverable_clients;
761         req_transno = req->rq_reqmsg->transno;
762         connected = obd->obd_connected_clients;
763         completed = max - obd->obd_recoverable_clients;
764         queue_len = obd->obd_requests_queued_for_recovery;
765         next_transno = obd->obd_next_recovery_transno;
766
767         CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
768                "req_transno: "LPU64", next_transno: "LPU64"\n",
769                max, connected, completed, queue_len, req_transno, next_transno);
770         if (obd->obd_abort_recovery) {
771                 CDEBUG(D_HA, "waking for aborted recovery\n");
772                 wake_up = 1;
773         } else if (!obd->obd_recovering) {
774                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
775                 wake_up = 1;
776         } else if (req_transno == next_transno) {
777                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
778                 wake_up = 1;
779         } else if (queue_len + completed == max) {
780                 CDEBUG(D_ERROR,
781                        "waking for skipped transno (skip: "LPD64
782                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
783                        next_transno, queue_len, completed, max, req_transno);
784                 obd->obd_next_recovery_transno = req_transno;
785                 wake_up = 1;
786         }
787         spin_unlock_bh(&obd->obd_processing_task_lock);
788         LASSERT(req->rq_reqmsg->transno >= next_transno);
789         return wake_up;
790 }
791
792 static void process_recovery_queue(struct obd_device *obd)
793 {
794         struct ptlrpc_request *req;
795         int abort_recovery = 0;
796         struct l_wait_info lwi = { 0 };
797         ENTRY;
798
799         for (;;) {
800                 spin_lock_bh(&obd->obd_processing_task_lock);
801                 LASSERT(obd->obd_processing_task == current->pid);
802                 req = list_entry(obd->obd_recovery_queue.next,
803                                  struct ptlrpc_request, rq_list);
804
805                 if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
806                         spin_unlock_bh(&obd->obd_processing_task_lock);
807                         CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
808                                LPD64")\n",
809                                obd->obd_next_recovery_transno,
810                                req->rq_reqmsg->transno);
811                         l_wait_event(obd->obd_next_transno_waitq,
812                                      check_for_next_transno(obd), &lwi);
813                         spin_lock_bh(&obd->obd_processing_task_lock);
814                         abort_recovery = obd->obd_abort_recovery;
815                         spin_unlock_bh(&obd->obd_processing_task_lock);
816                         if (abort_recovery) {
817                                 target_abort_recovery(obd);
818                                 return;
819                         }
820                         continue;
821                 }
822                 list_del_init(&req->rq_list);
823                 obd->obd_requests_queued_for_recovery--;
824                 spin_unlock_bh(&obd->obd_processing_task_lock);
825
826                 DEBUG_REQ(D_HA, req, "processing: ");
827                 (void)obd->obd_recovery_handler(req);
828                 obd->obd_replayed_requests++;
829                 reset_recovery_timer(obd);
830                 /* bug 1580: decide how to properly sync() in recovery */
831                 //mds_fsync_super(mds->mds_sb);
832                 class_export_put(req->rq_export);
833                 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
834                 OBD_FREE(req, sizeof *req);
835                 spin_lock_bh(&obd->obd_processing_task_lock);
836                 obd->obd_next_recovery_transno++;
837                 if (list_empty(&obd->obd_recovery_queue)) {
838                         obd->obd_processing_task = 0;
839                         spin_unlock_bh(&obd->obd_processing_task_lock);
840                         break;
841                 }
842                 spin_unlock_bh(&obd->obd_processing_task_lock);
843         }
844         EXIT;
845 }
846
847 int target_queue_recovery_request(struct ptlrpc_request *req,
848                                   struct obd_device *obd)
849 {
850         struct list_head *tmp;
851         int inserted = 0;
852         __u64 transno = req->rq_reqmsg->transno;
853         struct ptlrpc_request *saved_req;
854         struct lustre_msg *reqmsg;
855
856         /* CAVEAT EMPTOR: The incoming request message has been swabbed
857          * (i.e. buflens etc are in my own byte order), but type-dependent
858          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
859
860         if (!transno) {
861                 INIT_LIST_HEAD(&req->rq_list);
862                 DEBUG_REQ(D_HA, req, "not queueing");
863                 return 1;
864         }
865
866         /* XXX If I were a real man, these LBUGs would be sane cleanups. */
867         /* XXX just like the request-dup code in queue_final_reply */
868         OBD_ALLOC(saved_req, sizeof *saved_req);
869         if (!saved_req)
870                 LBUG();
871         OBD_ALLOC(reqmsg, req->rq_reqlen);
872         if (!reqmsg)
873                 LBUG();
874
875         spin_lock_bh(&obd->obd_processing_task_lock);
876
877         /* If we're processing the queue, we want don't want to queue this
878          * message.
879          *
880          * Also, if this request has a transno less than the one we're waiting
881          * for, we should process it now.  It could (and currently always will)
882          * be an open request for a descriptor that was opened some time ago.
883          *
884          * Also, a resent, replayed request that has already been
885          * handled will pass through here and be processed immediately.
886          */
887         if (obd->obd_processing_task == current->pid ||
888             transno < obd->obd_next_recovery_transno) {
889                 /* Processing the queue right now, don't re-add. */
890                 LASSERT(list_empty(&req->rq_list));
891                 spin_unlock_bh(&obd->obd_processing_task_lock);
892                 OBD_FREE(reqmsg, req->rq_reqlen);
893                 OBD_FREE(saved_req, sizeof *saved_req);
894                 return 1;
895         }
896
897         /* A resent, replayed request that is still on the queue; just drop it.
898            The queued request will handle this. */
899         if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) ==
900             (MSG_RESENT | MSG_REPLAY)) {
901                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
902                 spin_unlock_bh(&obd->obd_processing_task_lock);
903                 OBD_FREE(reqmsg, req->rq_reqlen);
904                 OBD_FREE(saved_req, sizeof *saved_req);
905                 return 0;
906         }
907
908         memcpy(saved_req, req, sizeof *req);
909         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
910         req = saved_req;
911         req->rq_reqmsg = reqmsg;
912         class_export_get(req->rq_export);
913         INIT_LIST_HEAD(&req->rq_list);
914
915         /* XXX O(n^2) */
916         list_for_each(tmp, &obd->obd_recovery_queue) {
917                 struct ptlrpc_request *reqiter =
918                         list_entry(tmp, struct ptlrpc_request, rq_list);
919
920                 if (reqiter->rq_reqmsg->transno > transno) {
921                         list_add_tail(&req->rq_list, &reqiter->rq_list);
922                         inserted = 1;
923                         break;
924                 }
925         }
926
927         if (!inserted) {
928                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
929         }
930
931         obd->obd_requests_queued_for_recovery++;
932
933         if (obd->obd_processing_task != 0) {
934                 /* Someone else is processing this queue, we'll leave it to
935                  * them.
936                  */
937                 wake_up(&obd->obd_next_transno_waitq);
938                 spin_unlock_bh(&obd->obd_processing_task_lock);
939                 return 0;
940         }
941
942         /* Nobody is processing, and we know there's (at least) one to process
943          * now, so we'll do the honours.
944          */
945         obd->obd_processing_task = current->pid;
946         spin_unlock_bh(&obd->obd_processing_task_lock);
947
948         process_recovery_queue(obd);
949         return 0;
950 }
951
952 struct obd_device * target_req2obd(struct ptlrpc_request *req)
953 {
954         return req->rq_export->exp_obd;
955 }
956
957 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
958 {
959         struct obd_device *obd = target_req2obd(req);
960         struct ptlrpc_request *saved_req;
961         struct lustre_msg *reqmsg;
962         int recovery_done = 0;
963         int rc2;
964
965         LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
966
967         if (rc) {
968                 /* Just like ptlrpc_error, but without the sending. */
969                 rc = lustre_pack_reply(req, 0, NULL, NULL);
970                 LASSERT(rc == 0); /* XXX handle this */
971                 req->rq_type = PTL_RPC_MSG_ERR;
972         }
973
974         LASSERT (!req->rq_reply_state->rs_difficult);
975         LASSERT(list_empty(&req->rq_list));
976         /* XXX a bit like the request-dup code in queue_recovery_request */
977         OBD_ALLOC(saved_req, sizeof *saved_req);
978         if (!saved_req)
979                 LBUG();
980         OBD_ALLOC(reqmsg, req->rq_reqlen);
981         if (!reqmsg)
982                 LBUG();
983         memcpy(saved_req, req, sizeof *saved_req);
984         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
985         /* the copied req takes over the reply state */
986         req->rq_reply_state = NULL;
987         req = saved_req;
988         req->rq_reqmsg = reqmsg;
989         class_export_get(req->rq_export);
990         list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
991
992         spin_lock_bh(&obd->obd_processing_task_lock);
993         --obd->obd_recoverable_clients;
994         recovery_done = (obd->obd_recoverable_clients == 0);
995         spin_unlock_bh(&obd->obd_processing_task_lock);
996
997         if (recovery_done) {
998                 struct list_head *tmp, *n;
999                 ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
1000                 CWARN("%s: all clients recovered, sending delayed replies\n",
1001                        obd->obd_name);
1002                 spin_lock_bh(&obd->obd_processing_task_lock);
1003                 obd->obd_recovering = 0;
1004                 target_cancel_recovery_timer(obd);
1005                 spin_unlock_bh(&obd->obd_processing_task_lock);
1006
1007                 /* when recovery finished, cleanup orphans on mds and ost */
1008                 if (OBT(obd) && OBP(obd, postrecov)) {
1009                         rc2 = OBP(obd, postrecov)(obd);
1010                         if (rc2 >= 0)
1011                                 CWARN("%s: all clients recovered, %d MDS "
1012                                       "orphans deleted\n", obd->obd_name, rc2);
1013                         else
1014                                 CERROR("postrecov failed %d\n", rc2);
1015                 }
1016
1017                 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
1018                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
1019                         DEBUG_REQ(D_ERROR, req, "delayed:");
1020                         ptlrpc_reply(req);
1021                         class_export_put(req->rq_export);
1022                         list_del(&req->rq_list);
1023                         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1024                         OBD_FREE(req, sizeof *req);
1025                 }
1026                 ptlrpc_run_recovery_over_upcall(obd);
1027         } else {
1028                 CWARN("%s: %d recoverable clients remain\n",
1029                        obd->obd_name, obd->obd_recoverable_clients);
1030                 wake_up(&obd->obd_next_transno_waitq);
1031         }
1032
1033         return 1;
1034 }
1035
1036 int
1037 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1038 {
1039         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1040                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1041                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1042                 /* NB this does _not_ send with ACK disabled, to simulate
1043                  * sending OK, but timing out for the ACK */
1044                 if (req->rq_reply_state != NULL) {
1045                         if (!req->rq_reply_state->rs_difficult) {
1046                                 lustre_free_reply_state (req->rq_reply_state);
1047                                 req->rq_reply_state = NULL;
1048                         } else {
1049                                 struct ptlrpc_service *svc =
1050                                         req->rq_rqbd->rqbd_srv_ni->sni_service;
1051                                 atomic_inc(&svc->srv_outstanding_replies);
1052                         }
1053                 }
1054                 return (-ECOMM);
1055         }
1056
1057         if (rc) {
1058                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1059                 if (req->rq_reply_state == NULL) {
1060                         rc = lustre_pack_reply (req, 0, NULL, NULL);
1061                         if (rc != 0) {
1062                                 CERROR ("can't allocate reply\n");
1063                                 return (rc);
1064                         }
1065                 }
1066                 req->rq_type = PTL_RPC_MSG_ERR;
1067         } else {
1068                 DEBUG_REQ(D_NET, req, "sending reply");
1069         }
1070         
1071         return (ptlrpc_send_reply(req, 1));
1072 }
1073
1074 void 
1075 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1076 {
1077         int                        netrc;
1078         unsigned long              flags;
1079         struct ptlrpc_reply_state *rs;
1080         struct obd_device         *obd;
1081         struct obd_export         *exp;
1082         struct ptlrpc_srv_ni      *sni;
1083         struct ptlrpc_service     *svc;
1084
1085         sni = req->rq_rqbd->rqbd_srv_ni;
1086         svc = sni->sni_service;
1087         
1088         rs = req->rq_reply_state;
1089         if (rs == NULL || !rs->rs_difficult) {
1090                 /* The easy case; no notifiers and reply_out_callback()
1091                  * cleans up (i.e. we can't look inside rs after a
1092                  * successful send) */
1093                 netrc = target_send_reply_msg (req, rc, fail_id);
1094
1095                 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1096                 return;
1097         }
1098
1099         /* must be an export if locks saved */
1100         LASSERT (req->rq_export != NULL);
1101         /* req/reply consistent */
1102         LASSERT (rs->rs_srv_ni == sni);
1103
1104         /* "fresh" reply */
1105         LASSERT (!rs->rs_scheduled);
1106         LASSERT (!rs->rs_scheduled_ever);
1107         LASSERT (!rs->rs_handled);
1108         LASSERT (!rs->rs_on_net);
1109         LASSERT (rs->rs_export == NULL);
1110         LASSERT (list_empty(&rs->rs_obd_list));
1111         LASSERT (list_empty(&rs->rs_exp_list));
1112
1113         exp = class_export_get (req->rq_export);
1114         obd = exp->exp_obd;
1115
1116         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1117         rs->rs_scheduled = 1;
1118         rs->rs_on_net    = 1;
1119         rs->rs_xid       = req->rq_xid;
1120         rs->rs_transno   = req->rq_transno;
1121         rs->rs_export    = exp;
1122         
1123         spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1124
1125         if (rs->rs_transno > obd->obd_last_committed) {
1126                 /* not committed already */ 
1127                 list_add_tail (&rs->rs_obd_list, 
1128                                &obd->obd_uncommitted_replies);
1129         }
1130
1131         spin_unlock (&obd->obd_uncommitted_replies_lock);
1132         spin_lock (&exp->exp_lock);
1133
1134         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1135
1136         spin_unlock_irqrestore (&exp->exp_lock, flags);
1137
1138         netrc = target_send_reply_msg (req, rc, fail_id);
1139
1140         spin_lock_irqsave (&svc->srv_lock, flags);
1141
1142         svc->srv_n_difficult_replies++;
1143
1144         if (netrc != 0) /* error sending: reply is off the net */
1145                 rs->rs_on_net = 0;
1146
1147         if (!rs->rs_on_net ||                   /* some notifier */
1148             list_empty(&rs->rs_exp_list) ||     /* completed already */
1149             list_empty(&rs->rs_obd_list)) {
1150                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1151                 wake_up (&svc->srv_waitq);
1152         } else {
1153                 list_add (&rs->rs_list, &sni->sni_active_replies);
1154                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
1155         }
1156
1157         spin_unlock_irqrestore (&svc->srv_lock, flags);
1158 }
1159
1160 int target_handle_ping(struct ptlrpc_request *req)
1161 {
1162         return lustre_pack_reply(req, 0, NULL, NULL);
1163 }