1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2003 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.lustre.org.
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.
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.
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.
23 # define EXPORT_SYMTAB
25 #define DEBUG_SUBSYSTEM S_LDLM
28 # include <linux/module.h>
30 # include <liblustre.h>
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>
39 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
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;
50 struct obd_device *mgmt_obd;
51 mgmtcli_register_for_events_t register_f;
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;
69 CERROR("unknown client OBD type \"%s\", can't setup\n",
74 if (lcfg->lcfg_inllen1 < 1) {
75 CERROR("requires a TARGET UUID\n");
79 if (lcfg->lcfg_inllen1 > 37) {
80 CERROR("client UUID must be less than 38 characters\n");
84 if (lcfg->lcfg_inllen2 < 1) {
85 CERROR("setup requires a SERVER UUID\n");
89 if (lcfg->lcfg_inllen2 > 37) {
90 CERROR("target UUID must be less than 38 characters\n");
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)));
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;
117 CERROR("ldlm_get_ref failed: %d\n", rc);
121 conn = ptlrpc_uuid_to_connection(&server_uuid);
123 GOTO(err_ldlm, rc = -ENOENT);
125 ptlrpc_init_client(rq_portal, rp_portal, name,
126 &obddev->obd_ldlm_client);
128 imp = class_new_import();
130 ptlrpc_put_connection(conn);
131 GOTO(err_ldlm, rc = -ENOENT);
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,
142 class_import_put(imp);
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);
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;
156 if (lcfg->lcfg_inllen4 != 0)
157 mgmt_name = lcfg->lcfg_inlbuf4;
159 mgmt_name = lcfg->lcfg_inlbuf3;
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);
168 mgmt_obd = class_name2obd(mgmt_name);
170 CERROR("can't find mgmtcli %s to register\n",
172 GOTO(err_import, rc = -ENOSYS);
175 register_f = inter_module_get("mgmtcli_register_for_events");
177 CERROR("can't i_m_g mgmtcli_register_for_events\n");
178 GOTO(err_import, rc = -ENOSYS);
181 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
182 inter_module_put("mgmtcli_register_for_events");
185 cli->cl_mgmtcli_obd = mgmt_obd;
191 class_destroy_import(imp);
199 int client_obd_cleanup(struct obd_device *obddev, int flags)
201 struct client_obd *cli = &obddev->u.cli;
205 if (cli->cl_mgmtcli_obd) {
206 mgmtcli_deregister_for_events_t dereg_f;
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");
212 class_destroy_import(cli->cl_import);
213 cli->cl_import = NULL;
215 ldlm_put_ref(flags & OBD_OPT_FORCE);
220 int client_connect_import(struct lustre_handle *dlm_handle,
221 struct obd_device *obd,
222 struct obd_uuid *cluuid)
224 struct client_obd *cli = &obd->u.cli;
225 struct obd_import *imp = cli->cl_import;
226 struct obd_export *exp;
231 rc = class_connect(dlm_handle, obd, cluuid);
235 cli->cl_conn_count++;
236 if (cli->cl_conn_count > 1)
238 exp = class_conn2export(dlm_handle);
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);
247 imp->imp_dlm_handle = *dlm_handle;
248 rc = ptlrpc_init_import(imp);
252 exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
253 rc = ptlrpc_connect_import(imp, NULL);
255 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
259 ptlrpc_pinger_add_import(imp);
264 ldlm_namespace_free(obd->obd_namespace, 0);
265 obd->obd_namespace = NULL;
267 cli->cl_conn_count--;
268 class_disconnect(exp, 0);
270 class_export_put(exp);
277 int client_disconnect_export(struct obd_export *exp, int failover)
279 struct obd_device *obd = class_exp2obd(exp);
280 struct client_obd *cli = &obd->u.cli;
281 struct obd_import *imp = cli->cl_import;
286 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
287 exp, exp ? exp->exp_handle.h_cookie : -1);
292 if (!cli->cl_conn_count) {
293 CERROR("disconnecting disconnected device (%s)\n",
295 GOTO(out_sem, rc = -EINVAL);
298 cli->cl_conn_count--;
299 if (cli->cl_conn_count)
300 GOTO(out_no_disconnect, rc = 0);
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
305 (void)ptlrpc_pinger_del_import(imp);
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;
315 /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
316 if (obd->obd_no_recov)
317 ptlrpc_invalidate_import(imp, 0);
319 rc = ptlrpc_disconnect_import(imp);
323 err = class_disconnect(exp, 0);
331 /* --------------------------------------------------------------------------
332 * from old lib/target.c
333 * -------------------------------------------------------------------------- */
335 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
336 struct obd_uuid *cluuid)
338 if (exp->exp_connection) {
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;
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);
357 conn->cookie = exp->exp_handle.h_cookie;
358 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
360 CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
364 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
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;
375 int rc = 0, abort_recovery;
379 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
381 LASSERT_REQSWAB (req, 0);
382 str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
384 CERROR("bad target UUID for connect\n");
385 GOTO(out, rc = -EINVAL);
388 obd_str2uuid (&tgtuuid, str);
389 target = class_uuid2obd(&tgtuuid);
391 target = class_name2obd(str);
394 if (!target || target->obd_stopping || !target->obd_set_up) {
395 CERROR("UUID '%s' is not available for connect\n", str);
396 GOTO(out, rc = -ENODEV);
399 LASSERT_REQSWAB (req, 1);
400 str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
402 CERROR("bad client UUID for connect\n");
403 GOTO(out, rc = -EINVAL);
406 obd_str2uuid (&cluuid, str);
408 /* XXX extract a nettype and format accordingly */
409 snprintf(remote_uuid.uuid, sizeof remote_uuid,
410 "NET_"LPX64"_UUID", req->rq_peer.peer_nid);
412 spin_lock_bh(&target->obd_processing_task_lock);
413 abort_recovery = target->obd_abort_recovery;
414 spin_unlock_bh(&target->obd_processing_task_lock);
416 target_abort_recovery(target);
418 tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
420 GOTO(out, rc = -EPROTO);
422 memcpy(&conn, tmp, sizeof conn);
424 rc = lustre_pack_reply(req, 0, NULL, NULL);
428 /* lctl gets a backstage, all-access pass. */
429 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
430 goto dont_check_exports;
432 spin_lock(&target->obd_dev_lock);
433 list_for_each(p, &target->obd_exports) {
434 export = list_entry(p, struct obd_export, exp_obd_chain);
435 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
436 spin_unlock(&target->obd_dev_lock);
437 LASSERT(export->exp_obd == target);
439 rc = target_handle_reconnect(&conn, export, &cluuid);
444 /* If we found an export, we already unlocked. */
446 spin_unlock(&target->obd_dev_lock);
447 } else if (req->rq_reqmsg->conn_cnt == 1) {
448 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
450 GOTO(out, rc = -EALREADY);
453 /* Tell the client if we're in recovery. */
454 /* If this is the first client, start the recovery timer */
455 if (target->obd_recovering) {
456 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
457 target_start_recovery_timer(target, handler);
460 /* Tell the client if we support replayable requests */
461 if (target->obd_replayable)
462 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
464 if (export == NULL) {
465 if (target->obd_recovering) {
466 CERROR("denying connection for new client %s: "
467 "%d clients in recovery for %lds\n", cluuid.uuid,
468 target->obd_recoverable_clients,
469 (target->obd_recovery_timer.expires-jiffies)/HZ);
473 rc = obd_connect(&conn, target, &cluuid);
478 /* If all else goes well, this is our RPC return code. */
481 if (rc && rc != EALREADY)
484 /* XXX track this all the time? */
485 if (target->obd_recovering) {
486 target->obd_connected_clients++;
489 req->rq_repmsg->handle = conn;
491 /* If the client and the server are the same node, we will already
492 * have an export that really points to the client's DLM export,
493 * because we have a shared handles table.
495 * XXX this will go away when shaver stops sending the "connect" handle
496 * in the real "remote handle" field of the request --phik 24 Apr 2003
498 if (req->rq_export != NULL)
499 class_export_put(req->rq_export);
501 /* ownership of this export ref transfers to the request */
502 export = req->rq_export = class_conn2export(&conn);
503 LASSERT(export != NULL);
505 spin_lock_irqsave(&export->exp_lock, flags);
506 if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
507 CERROR("%s: already connected at a higher conn_cnt: %d > %d\n",
508 cluuid.uuid, export->exp_conn_cnt,
509 req->rq_reqmsg->conn_cnt);
510 spin_unlock_irqrestore(&export->exp_lock, flags);
511 GOTO(out, rc = -EALREADY);
513 export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
514 spin_unlock_irqrestore(&export->exp_lock, flags);
516 /* request from liblustre? */
517 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
518 export->exp_libclient = 1;
520 if (export->exp_connection != NULL)
521 ptlrpc_put_connection(export->exp_connection);
522 export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
525 if (rc == EALREADY) {
526 /* We indicate the reconnection in a flag, not an error code. */
527 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
531 memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
534 if (export->exp_imp_reverse != NULL)
535 class_destroy_import(export->exp_imp_reverse);
536 revimp = export->exp_imp_reverse = class_new_import();
537 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
538 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
539 revimp->imp_remote_handle = conn;
540 revimp->imp_obd = target;
541 revimp->imp_dlm_fake = 1;
542 revimp->imp_state = LUSTRE_IMP_FULL;
543 class_import_put(revimp);
550 int target_handle_disconnect(struct ptlrpc_request *req)
552 struct obd_export *exp;
556 rc = lustre_pack_reply(req, 0, NULL, NULL);
560 /* keep the rq_export around so we can send the reply */
561 exp = class_export_get(req->rq_export);
562 req->rq_status = obd_disconnect(exp, 0);
566 void target_destroy_export(struct obd_export *exp)
568 /* exports created from last_rcvd data, and "fake"
569 exports created by lctl don't have an import */
570 if (exp->exp_imp_reverse != NULL)
571 class_destroy_import(exp->exp_imp_reverse);
573 /* We cancel locks at disconnect time, but this will catch any locks
574 * granted in a race with recovery-induced disconnect. */
575 if (exp->exp_obd->obd_namespace != NULL)
576 ldlm_cancel_locks_for_export(exp);
583 static void abort_delayed_replies(struct obd_device *obd)
585 struct ptlrpc_request *req;
586 struct list_head *tmp, *n;
587 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
588 req = list_entry(tmp, struct ptlrpc_request, rq_list);
589 DEBUG_REQ(D_ERROR, req, "aborted:");
590 req->rq_status = -ENOTCONN;
591 req->rq_type = PTL_RPC_MSG_ERR;
593 class_export_put(req->rq_export);
594 list_del(&req->rq_list);
595 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
596 OBD_FREE(req, sizeof *req);
600 static void abort_recovery_queue(struct obd_device *obd)
602 struct ptlrpc_request *req;
603 struct list_head *tmp, *n;
606 list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
607 req = list_entry(tmp, struct ptlrpc_request, rq_list);
608 DEBUG_REQ(D_ERROR, req, "aborted:");
609 req->rq_status = -ENOTCONN;
610 req->rq_type = PTL_RPC_MSG_ERR;
611 rc = lustre_pack_reply(req, 0, NULL, NULL);
615 DEBUG_REQ(D_ERROR, req,
616 "packing failed for abort-reply; skipping");
618 list_del(&req->rq_list);
619 class_export_put(req->rq_export);
620 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
621 OBD_FREE(req, sizeof *req);
625 void target_abort_recovery(void *data)
627 struct obd_device *obd = data;
630 CERROR("disconnecting clients and aborting recovery\n");
631 spin_lock_bh(&obd->obd_processing_task_lock);
632 if (!obd->obd_recovering) {
633 spin_unlock_bh(&obd->obd_processing_task_lock);
638 obd->obd_recovering = obd->obd_abort_recovery = 0;
640 wake_up(&obd->obd_next_transno_waitq);
641 target_cancel_recovery_timer(obd);
642 spin_unlock_bh(&obd->obd_processing_task_lock);
644 class_disconnect_exports(obd, 0);
646 /* when recovery was aborted, cleanup orphans on mds and ost */
647 if (OBT(obd) && OBP(obd, postrecov)) {
648 rc = OBP(obd, postrecov)(obd);
650 CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
652 CERROR("postrecov failed %d\n", rc);
655 abort_delayed_replies(obd);
656 abort_recovery_queue(obd);
657 ptlrpc_run_recovery_over_upcall(obd);
660 static void target_recovery_expired(unsigned long castmeharder)
662 struct obd_device *obd = (struct obd_device *)castmeharder;
663 CERROR("recovery timed out, aborting\n");
664 spin_lock_bh(&obd->obd_processing_task_lock);
665 obd->obd_abort_recovery = 1;
666 wake_up(&obd->obd_next_transno_waitq);
667 spin_unlock_bh(&obd->obd_processing_task_lock);
671 /* obd_processing_task_lock should be held */
672 void target_cancel_recovery_timer(struct obd_device *obd)
674 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
675 del_timer(&obd->obd_recovery_timer);
678 static void reset_recovery_timer(struct obd_device *obd)
680 spin_lock_bh(&obd->obd_processing_task_lock);
681 if (!obd->obd_recovering) {
682 spin_unlock_bh(&obd->obd_processing_task_lock);
685 CDEBUG(D_HA, "timer will expire in %u seconds\n",
686 OBD_RECOVERY_TIMEOUT / HZ);
687 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
688 spin_unlock_bh(&obd->obd_processing_task_lock);
692 /* Only start it the first time called */
693 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
695 spin_lock_bh(&obd->obd_processing_task_lock);
696 if (obd->obd_recovery_handler) {
697 spin_unlock_bh(&obd->obd_processing_task_lock);
700 CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
701 OBD_RECOVERY_TIMEOUT / HZ);
702 obd->obd_recovery_handler = handler;
703 obd->obd_recovery_timer.function = target_recovery_expired;
704 obd->obd_recovery_timer.data = (unsigned long)obd;
705 init_timer(&obd->obd_recovery_timer);
706 spin_unlock_bh(&obd->obd_processing_task_lock);
708 reset_recovery_timer(obd);
711 static int check_for_next_transno(struct obd_device *obd)
713 struct ptlrpc_request *req;
714 int wake_up = 0, connected, completed, queue_len, max;
715 __u64 next_transno, req_transno;
717 spin_lock_bh(&obd->obd_processing_task_lock);
718 req = list_entry(obd->obd_recovery_queue.next,
719 struct ptlrpc_request, rq_list);
720 max = obd->obd_max_recoverable_clients;
721 req_transno = req->rq_reqmsg->transno;
722 connected = obd->obd_connected_clients;
723 completed = max - obd->obd_recoverable_clients;
724 queue_len = obd->obd_requests_queued_for_recovery;
725 next_transno = obd->obd_next_recovery_transno;
727 if (obd->obd_abort_recovery) {
728 CDEBUG(D_HA, "waking for aborted recovery\n");
730 } else if (!obd->obd_recovering) {
731 CDEBUG(D_HA, "waking for completed recovery (?)\n");
733 } else if (req_transno == next_transno) {
734 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
736 } else if (queue_len + completed == max) {
738 "waking for skipped transno (skip: "LPD64
739 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
740 next_transno, queue_len, completed, max, req_transno);
741 obd->obd_next_recovery_transno = req_transno;
744 spin_unlock_bh(&obd->obd_processing_task_lock);
745 LASSERT(req->rq_reqmsg->transno >= next_transno);
749 static void process_recovery_queue(struct obd_device *obd)
751 struct ptlrpc_request *req;
752 int abort_recovery = 0;
753 struct l_wait_info lwi = { 0 };
757 spin_lock_bh(&obd->obd_processing_task_lock);
758 LASSERT(obd->obd_processing_task == current->pid);
759 req = list_entry(obd->obd_recovery_queue.next,
760 struct ptlrpc_request, rq_list);
762 if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
763 spin_unlock_bh(&obd->obd_processing_task_lock);
764 CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
766 obd->obd_next_recovery_transno,
767 req->rq_reqmsg->transno);
768 l_wait_event(obd->obd_next_transno_waitq,
769 check_for_next_transno(obd), &lwi);
770 spin_lock_bh(&obd->obd_processing_task_lock);
771 abort_recovery = obd->obd_abort_recovery;
772 spin_unlock_bh(&obd->obd_processing_task_lock);
773 if (abort_recovery) {
774 target_abort_recovery(obd);
779 list_del_init(&req->rq_list);
780 obd->obd_requests_queued_for_recovery--;
781 spin_unlock_bh(&obd->obd_processing_task_lock);
783 DEBUG_REQ(D_HA, req, "processing: ");
784 (void)obd->obd_recovery_handler(req);
785 obd->obd_replayed_requests++;
786 reset_recovery_timer(obd);
787 /* bug 1580: decide how to properly sync() in recovery */
788 //mds_fsync_super(mds->mds_sb);
789 class_export_put(req->rq_export);
790 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
791 OBD_FREE(req, sizeof *req);
792 spin_lock_bh(&obd->obd_processing_task_lock);
793 obd->obd_next_recovery_transno++;
794 if (list_empty(&obd->obd_recovery_queue)) {
795 obd->obd_processing_task = 0;
796 spin_unlock_bh(&obd->obd_processing_task_lock);
799 spin_unlock_bh(&obd->obd_processing_task_lock);
804 int target_queue_recovery_request(struct ptlrpc_request *req,
805 struct obd_device *obd)
807 struct list_head *tmp;
809 __u64 transno = req->rq_reqmsg->transno;
810 struct ptlrpc_request *saved_req;
811 struct lustre_msg *reqmsg;
813 /* CAVEAT EMPTOR: The incoming request message has been swabbed
814 * (i.e. buflens etc are in my own byte order), but type-dependent
815 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
818 INIT_LIST_HEAD(&req->rq_list);
819 DEBUG_REQ(D_HA, req, "not queueing");
823 /* XXX If I were a real man, these LBUGs would be sane cleanups. */
824 /* XXX just like the request-dup code in queue_final_reply */
825 OBD_ALLOC(saved_req, sizeof *saved_req);
828 OBD_ALLOC(reqmsg, req->rq_reqlen);
832 spin_lock_bh(&obd->obd_processing_task_lock);
834 /* If we're processing the queue, we want don't want to queue this
837 * Also, if this request has a transno less than the one we're waiting
838 * for, we should process it now. It could (and currently always will)
839 * be an open request for a descriptor that was opened some time ago.
841 if (obd->obd_processing_task == current->pid ||
842 transno < obd->obd_next_recovery_transno) {
843 /* Processing the queue right now, don't re-add. */
844 LASSERT(list_empty(&req->rq_list));
845 spin_unlock_bh(&obd->obd_processing_task_lock);
846 OBD_FREE(reqmsg, req->rq_reqlen);
847 OBD_FREE(saved_req, sizeof *saved_req);
851 memcpy(saved_req, req, sizeof *req);
852 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
854 req->rq_reqmsg = reqmsg;
855 class_export_get(req->rq_export);
856 INIT_LIST_HEAD(&req->rq_list);
859 list_for_each(tmp, &obd->obd_recovery_queue) {
860 struct ptlrpc_request *reqiter =
861 list_entry(tmp, struct ptlrpc_request, rq_list);
863 if (reqiter->rq_reqmsg->transno > transno) {
864 list_add_tail(&req->rq_list, &reqiter->rq_list);
871 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
874 obd->obd_requests_queued_for_recovery++;
876 if (obd->obd_processing_task != 0) {
877 /* Someone else is processing this queue, we'll leave it to
880 wake_up(&obd->obd_next_transno_waitq);
881 spin_unlock_bh(&obd->obd_processing_task_lock);
885 /* Nobody is processing, and we know there's (at least) one to process
886 * now, so we'll do the honours.
888 obd->obd_processing_task = current->pid;
889 spin_unlock_bh(&obd->obd_processing_task_lock);
891 process_recovery_queue(obd);
895 struct obd_device * target_req2obd(struct ptlrpc_request *req)
897 return req->rq_export->exp_obd;
900 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
902 struct obd_device *obd = target_req2obd(req);
903 struct ptlrpc_request *saved_req;
904 struct lustre_msg *reqmsg;
905 int recovery_done = 0;
908 LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
911 /* Just like ptlrpc_error, but without the sending. */
912 rc = lustre_pack_reply(req, 0, NULL, NULL);
913 LASSERT(rc == 0); /* XXX handle this */
914 req->rq_type = PTL_RPC_MSG_ERR;
917 LASSERT (!req->rq_reply_state->rs_difficult);
918 LASSERT(list_empty(&req->rq_list));
919 /* XXX a bit like the request-dup code in queue_recovery_request */
920 OBD_ALLOC(saved_req, sizeof *saved_req);
923 OBD_ALLOC(reqmsg, req->rq_reqlen);
926 memcpy(saved_req, req, sizeof *saved_req);
927 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
928 /* the copied req takes over the reply state */
929 req->rq_reply_state = NULL;
931 req->rq_reqmsg = reqmsg;
932 class_export_get(req->rq_export);
933 list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
935 spin_lock_bh(&obd->obd_processing_task_lock);
936 --obd->obd_recoverable_clients;
937 recovery_done = (obd->obd_recoverable_clients == 0);
938 spin_unlock_bh(&obd->obd_processing_task_lock);
941 struct list_head *tmp, *n;
942 ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
943 CWARN("%s: all clients recovered, sending delayed replies\n",
945 spin_lock_bh(&obd->obd_processing_task_lock);
946 obd->obd_recovering = 0;
947 target_cancel_recovery_timer(obd);
948 spin_unlock_bh(&obd->obd_processing_task_lock);
950 /* when recovery finished, cleanup orphans on mds and ost */
951 if (OBT(obd) && OBP(obd, postrecov)) {
952 rc2 = OBP(obd, postrecov)(obd);
954 CWARN("%s: all clients recovered, %d MDS "
955 "orphans deleted\n", obd->obd_name, rc2);
957 CERROR("postrecov failed %d\n", rc2);
960 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
961 req = list_entry(tmp, struct ptlrpc_request, rq_list);
962 DEBUG_REQ(D_ERROR, req, "delayed:");
964 class_export_put(req->rq_export);
965 list_del(&req->rq_list);
966 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
967 OBD_FREE(req, sizeof *req);
969 ptlrpc_run_recovery_over_upcall(obd);
971 CWARN("%s: %d recoverable clients remain\n",
972 obd->obd_name, obd->obd_recoverable_clients);
973 wake_up(&obd->obd_next_transno_waitq);
980 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
982 if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
983 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
984 DEBUG_REQ(D_ERROR, req, "dropping reply");
985 /* NB this does _not_ send with ACK disabled, to simulate
986 * sending OK, but timing out for the ACK */
987 if (req->rq_reply_state != NULL) {
988 if (!req->rq_reply_state->rs_difficult) {
989 lustre_free_reply_state (req->rq_reply_state);
990 req->rq_reply_state = NULL;
992 struct ptlrpc_service *svc =
993 req->rq_rqbd->rqbd_srv_ni->sni_service;
994 atomic_inc(&svc->srv_outstanding_replies);
1001 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1002 if (req->rq_reply_state == NULL) {
1003 rc = lustre_pack_reply (req, 0, NULL, NULL);
1005 CERROR ("can't allocate reply\n");
1009 req->rq_type = PTL_RPC_MSG_ERR;
1011 DEBUG_REQ(D_NET, req, "sending reply");
1014 return (ptlrpc_send_reply(req, 1));
1018 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1021 unsigned long flags;
1022 struct ptlrpc_reply_state *rs;
1023 struct obd_device *obd;
1024 struct obd_export *exp;
1025 struct ptlrpc_srv_ni *sni;
1026 struct ptlrpc_service *svc;
1028 sni = req->rq_rqbd->rqbd_srv_ni;
1029 svc = sni->sni_service;
1031 rs = req->rq_reply_state;
1032 if (rs == NULL || !rs->rs_difficult) {
1033 /* The easy case; no notifiers and reply_out_callback()
1034 * cleans up (i.e. we can't look inside rs after a
1035 * successful send) */
1036 netrc = target_send_reply_msg (req, rc, fail_id);
1038 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1042 /* must be an export if locks saved */
1043 LASSERT (req->rq_export != NULL);
1044 /* req/reply consistent */
1045 LASSERT (rs->rs_srv_ni == sni);
1048 LASSERT (!rs->rs_scheduled);
1049 LASSERT (!rs->rs_scheduled_ever);
1050 LASSERT (!rs->rs_handled);
1051 LASSERT (!rs->rs_on_net);
1052 LASSERT (rs->rs_export == NULL);
1053 LASSERT (list_empty(&rs->rs_obd_list));
1054 LASSERT (list_empty(&rs->rs_exp_list));
1056 exp = class_export_get (req->rq_export);
1059 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1060 rs->rs_scheduled = 1;
1062 rs->rs_xid = req->rq_xid;
1063 rs->rs_transno = req->rq_transno;
1064 rs->rs_export = exp;
1066 spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1068 if (rs->rs_transno > obd->obd_last_committed) {
1069 /* not committed already */
1070 list_add_tail (&rs->rs_obd_list,
1071 &obd->obd_uncommitted_replies);
1074 spin_unlock (&obd->obd_uncommitted_replies_lock);
1075 spin_lock (&exp->exp_lock);
1077 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1079 spin_unlock_irqrestore (&exp->exp_lock, flags);
1081 netrc = target_send_reply_msg (req, rc, fail_id);
1083 spin_lock_irqsave (&svc->srv_lock, flags);
1085 svc->srv_n_difficult_replies++;
1087 if (netrc != 0) /* error sending: reply is off the net */
1090 if (!rs->rs_on_net || /* some notifier */
1091 list_empty(&rs->rs_exp_list) || /* completed already */
1092 list_empty(&rs->rs_obd_list)) {
1093 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1094 wake_up (&svc->srv_waitq);
1096 list_add (&rs->rs_list, &sni->sni_active_replies);
1097 rs->rs_scheduled = 0; /* allow notifier to schedule */
1100 spin_unlock_irqrestore (&svc->srv_lock, flags);
1103 int target_handle_ping(struct ptlrpc_request *req)
1105 return lustre_pack_reply(req, 0, NULL, NULL);