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 spin_lock_init(&cli->cl_loi_list_lock);
106 cli->cl_brw_in_flight = 0;
107 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
108 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
109 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
110 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
111 cli->cl_max_pages_per_rpc = PTL_MD_MAX_PAGES;
112 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
116 CERROR("ldlm_get_ref failed: %d\n", rc);
120 conn = ptlrpc_uuid_to_connection(&server_uuid);
122 GOTO(err_ldlm, rc = -ENOENT);
124 ptlrpc_init_client(rq_portal, rp_portal, name,
125 &obddev->obd_ldlm_client);
127 imp = class_new_import();
129 ptlrpc_put_connection(conn);
130 GOTO(err_ldlm, rc = -ENOENT);
132 imp->imp_connection = conn;
133 imp->imp_client = &obddev->obd_ldlm_client;
134 imp->imp_obd = obddev;
135 imp->imp_connect_op = connect_op;
136 imp->imp_generation = 0;
137 imp->imp_initial_recov = 1;
138 INIT_LIST_HEAD(&imp->imp_pinger_chain);
139 memcpy(imp->imp_target_uuid.uuid, lcfg->lcfg_inlbuf1,
141 class_import_put(imp);
143 cli->cl_import = imp;
144 cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
145 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
146 cli->cl_sandev = to_kdev_t(0);
148 if (lcfg->lcfg_inllen3 != 0) {
149 if (!strcmp(lcfg->lcfg_inlbuf3, "inactive")) {
150 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
151 name, obddev->obd_name,
152 imp->imp_target_uuid.uuid);
153 imp->imp_invalid = 1;
155 if (lcfg->lcfg_inllen4 != 0)
156 mgmt_name = lcfg->lcfg_inlbuf4;
158 mgmt_name = lcfg->lcfg_inlbuf3;
162 if (mgmt_name != NULL) {
163 /* Register with management client if we need to. */
164 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
165 obddev->obd_name, mgmt_name, server_uuid.uuid);
167 mgmt_obd = class_name2obd(mgmt_name);
169 CERROR("can't find mgmtcli %s to register\n",
171 GOTO(err_import, rc = -ENOSYS);
174 register_f = inter_module_get("mgmtcli_register_for_events");
176 CERROR("can't i_m_g mgmtcli_register_for_events\n");
177 GOTO(err_import, rc = -ENOSYS);
180 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
181 inter_module_put("mgmtcli_register_for_events");
184 cli->cl_mgmtcli_obd = mgmt_obd;
190 class_destroy_import(imp);
198 int client_obd_cleanup(struct obd_device *obddev, int flags)
200 struct client_obd *cli = &obddev->u.cli;
204 if (cli->cl_mgmtcli_obd) {
205 mgmtcli_deregister_for_events_t dereg_f;
207 dereg_f = inter_module_get("mgmtcli_deregister_for_events");
208 dereg_f(cli->cl_mgmtcli_obd, obddev);
209 inter_module_put("mgmtcli_deregister_for_events");
211 class_destroy_import(cli->cl_import);
212 cli->cl_import = NULL;
214 ldlm_put_ref(flags & OBD_OPT_FORCE);
219 int client_connect_import(struct lustre_handle *dlm_handle,
220 struct obd_device *obd,
221 struct obd_uuid *cluuid)
223 struct client_obd *cli = &obd->u.cli;
224 struct obd_import *imp = cli->cl_import;
225 struct obd_export *exp;
230 rc = class_connect(dlm_handle, obd, cluuid);
234 cli->cl_conn_count++;
235 if (cli->cl_conn_count > 1)
237 exp = class_conn2export(dlm_handle);
239 if (obd->obd_namespace != NULL)
240 CERROR("already have namespace!\n");
241 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
242 LDLM_NAMESPACE_CLIENT);
243 if (obd->obd_namespace == NULL)
244 GOTO(out_disco, rc = -ENOMEM);
246 imp->imp_dlm_handle = *dlm_handle;
247 rc = ptlrpc_init_import(imp);
251 exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
252 rc = ptlrpc_connect_import(imp, NULL);
254 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
258 ptlrpc_pinger_add_import(imp);
263 ldlm_namespace_free(obd->obd_namespace, 0);
264 obd->obd_namespace = NULL;
266 cli->cl_conn_count--;
267 class_disconnect(exp, 0);
269 class_export_put(exp);
276 int client_disconnect_export(struct obd_export *exp, int failover)
278 struct obd_device *obd = class_exp2obd(exp);
279 struct client_obd *cli = &obd->u.cli;
280 struct obd_import *imp = cli->cl_import;
285 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
286 exp, exp ? exp->exp_handle.h_cookie : -1);
291 if (!cli->cl_conn_count) {
292 CERROR("disconnecting disconnected device (%s)\n",
294 GOTO(out_sem, rc = -EINVAL);
297 cli->cl_conn_count--;
298 if (cli->cl_conn_count)
299 GOTO(out_no_disconnect, rc = 0);
301 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
302 * delete it regardless. (It's safe to delete an import that was
304 (void)ptlrpc_pinger_del_import(imp);
306 if (obd->obd_namespace != NULL) {
307 /* obd_no_recov == local only */
308 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
309 obd->obd_no_recov, NULL);
310 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
311 obd->obd_namespace = NULL;
314 /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
315 if (obd->obd_no_recov)
316 ptlrpc_invalidate_import(imp);
318 rc = ptlrpc_disconnect_import(imp);
322 err = class_disconnect(exp, 0);
330 /* --------------------------------------------------------------------------
331 * from old lib/target.c
332 * -------------------------------------------------------------------------- */
334 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
335 struct obd_uuid *cluuid)
337 if (exp->exp_connection) {
338 struct lustre_handle *hdl;
339 hdl = &exp->exp_imp_reverse->imp_remote_handle;
340 /* Might be a re-connect after a partition. */
341 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
342 CERROR("%s reconnecting\n", cluuid->uuid);
343 conn->cookie = exp->exp_handle.h_cookie;
346 CERROR("%s reconnecting from %s, "
347 "handle mismatch (ours "LPX64", theirs "
348 LPX64")\n", cluuid->uuid,
349 exp->exp_connection->c_remote_uuid.uuid,
350 hdl->cookie, conn->cookie);
351 memset(conn, 0, sizeof *conn);
356 conn->cookie = exp->exp_handle.h_cookie;
357 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
359 CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
363 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
365 struct obd_device *target;
366 struct obd_export *export = NULL;
367 struct obd_import *revimp;
368 struct lustre_handle conn;
369 struct obd_uuid tgtuuid;
370 struct obd_uuid cluuid;
371 struct obd_uuid remote_uuid;
374 int rc = 0, abort_recovery;
377 LASSERT_REQSWAB (req, 0);
378 str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
380 CERROR("bad target UUID for connect\n");
381 GOTO(out, rc = -EINVAL);
384 obd_str2uuid (&tgtuuid, str);
385 target = class_uuid2obd(&tgtuuid);
387 target = class_name2obd(str);
390 if (!target || target->obd_stopping || !target->obd_set_up) {
391 CERROR("UUID '%s' is not available for connect\n", str);
392 GOTO(out, rc = -ENODEV);
395 LASSERT_REQSWAB (req, 1);
396 str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
398 CERROR("bad client UUID for connect\n");
399 GOTO(out, rc = -EINVAL);
402 obd_str2uuid (&cluuid, str);
404 /* XXX extract a nettype and format accordingly */
405 snprintf(remote_uuid.uuid, sizeof remote_uuid,
406 "NET_"LPX64"_UUID", req->rq_peer.peer_nid);
408 spin_lock_bh(&target->obd_processing_task_lock);
409 abort_recovery = target->obd_abort_recovery;
410 spin_unlock_bh(&target->obd_processing_task_lock);
412 target_abort_recovery(target);
414 tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
416 GOTO(out, rc = -EPROTO);
418 memcpy(&conn, tmp, sizeof conn);
420 rc = lustre_pack_reply(req, 0, NULL, NULL);
424 /* lctl gets a backstage, all-access pass. */
425 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
426 goto dont_check_exports;
428 spin_lock(&target->obd_dev_lock);
429 list_for_each(p, &target->obd_exports) {
430 export = list_entry(p, struct obd_export, exp_obd_chain);
431 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
432 spin_unlock(&target->obd_dev_lock);
433 LASSERT(export->exp_obd == target);
435 rc = target_handle_reconnect(&conn, export, &cluuid);
440 /* If we found an export, we already unlocked. */
442 spin_unlock(&target->obd_dev_lock);
443 } else if (req->rq_reqmsg->conn_cnt == 1) {
444 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
446 GOTO(out, rc = -EALREADY);
449 /* Tell the client if we're in recovery. */
450 /* If this is the first client, start the recovery timer */
451 if (target->obd_recovering) {
452 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
453 target_start_recovery_timer(target, handler);
456 /* Tell the client if we support replayable requests */
457 if (target->obd_replayable)
458 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
460 if (export == NULL) {
461 if (target->obd_recovering) {
462 CERROR("denying connection for new client %s: "
463 "%d clients in recovery for %lds\n", cluuid.uuid,
464 target->obd_recoverable_clients,
465 (target->obd_recovery_timer.expires-jiffies)/HZ);
469 rc = obd_connect(&conn, target, &cluuid);
474 /* If all else goes well, this is our RPC return code. */
477 if (rc && rc != EALREADY)
480 /* XXX track this all the time? */
481 if (target->obd_recovering) {
482 target->obd_connected_clients++;
485 req->rq_repmsg->handle = conn;
487 /* If the client and the server are the same node, we will already
488 * have an export that really points to the client's DLM export,
489 * because we have a shared handles table.
491 * XXX this will go away when shaver stops sending the "connect" handle
492 * in the real "remote handle" field of the request --phik 24 Apr 2003
494 if (req->rq_export != NULL)
495 class_export_put(req->rq_export);
497 /* ownership of this export ref transfers to the request */
498 export = req->rq_export = class_conn2export(&conn);
499 LASSERT(export != NULL);
501 /* request from liblustre? */
502 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
503 export->exp_libclient = 1;
505 if (export->exp_connection != NULL)
506 ptlrpc_put_connection(export->exp_connection);
507 export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
510 LASSERT(export->exp_conn_cnt < req->rq_reqmsg->conn_cnt);
511 export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
513 if (rc == EALREADY) {
514 /* We indicate the reconnection in a flag, not an error code. */
515 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
519 memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
522 if (export->exp_imp_reverse != NULL)
523 class_destroy_import(export->exp_imp_reverse);
524 revimp = export->exp_imp_reverse = class_new_import();
525 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
526 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
527 revimp->imp_remote_handle = conn;
528 revimp->imp_obd = target;
529 revimp->imp_dlm_fake = 1;
530 revimp->imp_state = LUSTRE_IMP_FULL;
531 class_import_put(revimp);
538 int target_handle_disconnect(struct ptlrpc_request *req)
540 struct obd_export *exp;
544 rc = lustre_pack_reply(req, 0, NULL, NULL);
548 /* keep the rq_export around so we can send the reply */
549 exp = class_export_get(req->rq_export);
550 req->rq_status = obd_disconnect(exp, 0);
554 void target_destroy_export(struct obd_export *exp)
556 /* exports created from last_rcvd data, and "fake"
557 exports created by lctl don't have an import */
558 if (exp->exp_imp_reverse != NULL)
559 class_destroy_import(exp->exp_imp_reverse);
561 /* We cancel locks at disconnect time, but this will catch any locks
562 * granted in a race with recovery-induced disconnect. */
563 if (exp->exp_obd->obd_namespace != NULL)
564 ldlm_cancel_locks_for_export(exp);
571 static void abort_delayed_replies(struct obd_device *obd)
573 struct ptlrpc_request *req;
574 struct list_head *tmp, *n;
575 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
576 req = list_entry(tmp, struct ptlrpc_request, rq_list);
577 DEBUG_REQ(D_ERROR, req, "aborted:");
578 req->rq_status = -ENOTCONN;
579 req->rq_type = PTL_RPC_MSG_ERR;
581 class_export_put(req->rq_export);
582 list_del(&req->rq_list);
583 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
584 OBD_FREE(req, sizeof *req);
588 static void abort_recovery_queue(struct obd_device *obd)
590 struct ptlrpc_request *req;
591 struct list_head *tmp, *n;
594 list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
595 req = list_entry(tmp, struct ptlrpc_request, rq_list);
596 DEBUG_REQ(D_ERROR, req, "aborted:");
597 req->rq_status = -ENOTCONN;
598 req->rq_type = PTL_RPC_MSG_ERR;
599 rc = lustre_pack_reply(req, 0, NULL, NULL);
603 DEBUG_REQ(D_ERROR, req,
604 "packing failed for abort-reply; skipping");
606 list_del(&req->rq_list);
607 class_export_put(req->rq_export);
608 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
609 OBD_FREE(req, sizeof *req);
613 void target_abort_recovery(void *data)
615 struct obd_device *obd = data;
618 CERROR("disconnecting clients and aborting recovery\n");
619 spin_lock_bh(&obd->obd_processing_task_lock);
620 if (!obd->obd_recovering) {
621 spin_unlock_bh(&obd->obd_processing_task_lock);
626 obd->obd_recovering = obd->obd_abort_recovery = 0;
628 wake_up(&obd->obd_next_transno_waitq);
629 target_cancel_recovery_timer(obd);
630 spin_unlock_bh(&obd->obd_processing_task_lock);
632 class_disconnect_exports(obd, 0);
634 /* when recovery was aborted, cleanup orphans on mds and ost */
635 if (OBT(obd) && OBP(obd, postrecov)) {
636 rc = OBP(obd, postrecov)(obd);
638 CWARN("Cleanup %d orphans after recovery was aborted\n", rc);
640 CERROR("postrecov failed %d\n", rc);
643 abort_delayed_replies(obd);
644 abort_recovery_queue(obd);
645 ptlrpc_run_recovery_over_upcall(obd);
648 static void target_recovery_expired(unsigned long castmeharder)
650 struct obd_device *obd = (struct obd_device *)castmeharder;
651 CERROR("recovery timed out, aborting\n");
652 spin_lock_bh(&obd->obd_processing_task_lock);
653 obd->obd_abort_recovery = 1;
654 wake_up(&obd->obd_next_transno_waitq);
655 spin_unlock_bh(&obd->obd_processing_task_lock);
659 /* obd_processing_task_lock should be held */
660 void target_cancel_recovery_timer(struct obd_device *obd)
662 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
663 del_timer(&obd->obd_recovery_timer);
666 static void reset_recovery_timer(struct obd_device *obd)
668 spin_lock_bh(&obd->obd_processing_task_lock);
669 if (!obd->obd_recovering) {
670 spin_unlock_bh(&obd->obd_processing_task_lock);
673 CDEBUG(D_HA, "timer will expire in %u seconds\n",
674 OBD_RECOVERY_TIMEOUT / HZ);
675 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
676 spin_unlock_bh(&obd->obd_processing_task_lock);
680 /* Only start it the first time called */
681 void target_start_recovery_timer(struct obd_device *obd, svc_handler_t handler)
683 spin_lock_bh(&obd->obd_processing_task_lock);
684 if (obd->obd_recovery_handler) {
685 spin_unlock_bh(&obd->obd_processing_task_lock);
688 CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
689 OBD_RECOVERY_TIMEOUT / HZ);
690 obd->obd_recovery_handler = handler;
691 obd->obd_recovery_timer.function = target_recovery_expired;
692 obd->obd_recovery_timer.data = (unsigned long)obd;
693 init_timer(&obd->obd_recovery_timer);
694 spin_unlock_bh(&obd->obd_processing_task_lock);
696 reset_recovery_timer(obd);
699 static int check_for_next_transno(struct obd_device *obd)
701 struct ptlrpc_request *req;
702 int wake_up = 0, connected, completed, queue_len, max;
703 __u64 next_transno, req_transno;
705 spin_lock_bh(&obd->obd_processing_task_lock);
706 req = list_entry(obd->obd_recovery_queue.next,
707 struct ptlrpc_request, rq_list);
708 max = obd->obd_max_recoverable_clients;
709 req_transno = req->rq_reqmsg->transno;
710 connected = obd->obd_connected_clients;
711 completed = max - obd->obd_recoverable_clients;
712 queue_len = obd->obd_requests_queued_for_recovery;
713 next_transno = obd->obd_next_recovery_transno;
715 if (obd->obd_abort_recovery) {
716 CDEBUG(D_HA, "waking for aborted recovery\n");
718 } else if (!obd->obd_recovering) {
719 CDEBUG(D_HA, "waking for completed recovery (?)\n");
721 } else if (req_transno == next_transno) {
722 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
724 } else if (queue_len + completed == max) {
726 "waking for skipped transno (skip: "LPD64
727 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
728 next_transno, queue_len, completed, max, req_transno);
729 obd->obd_next_recovery_transno = req_transno;
732 spin_unlock_bh(&obd->obd_processing_task_lock);
733 LASSERT(req->rq_reqmsg->transno >= next_transno);
737 static void process_recovery_queue(struct obd_device *obd)
739 struct ptlrpc_request *req;
740 int abort_recovery = 0;
741 struct l_wait_info lwi = { 0 };
745 spin_lock_bh(&obd->obd_processing_task_lock);
746 LASSERT(obd->obd_processing_task == current->pid);
747 req = list_entry(obd->obd_recovery_queue.next,
748 struct ptlrpc_request, rq_list);
750 if (req->rq_reqmsg->transno != obd->obd_next_recovery_transno) {
751 spin_unlock_bh(&obd->obd_processing_task_lock);
752 CDEBUG(D_HA, "Waiting for transno "LPD64" (1st is "
754 obd->obd_next_recovery_transno,
755 req->rq_reqmsg->transno);
756 l_wait_event(obd->obd_next_transno_waitq,
757 check_for_next_transno(obd), &lwi);
758 spin_lock_bh(&obd->obd_processing_task_lock);
759 abort_recovery = obd->obd_abort_recovery;
760 spin_unlock_bh(&obd->obd_processing_task_lock);
761 if (abort_recovery) {
762 target_abort_recovery(obd);
767 list_del_init(&req->rq_list);
768 obd->obd_requests_queued_for_recovery--;
769 spin_unlock_bh(&obd->obd_processing_task_lock);
771 DEBUG_REQ(D_HA, req, "processing: ");
772 (void)obd->obd_recovery_handler(req);
773 obd->obd_replayed_requests++;
774 reset_recovery_timer(obd);
775 /* bug 1580: decide how to properly sync() in recovery */
776 //mds_fsync_super(mds->mds_sb);
777 class_export_put(req->rq_export);
778 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
779 OBD_FREE(req, sizeof *req);
780 spin_lock_bh(&obd->obd_processing_task_lock);
781 obd->obd_next_recovery_transno++;
782 if (list_empty(&obd->obd_recovery_queue)) {
783 obd->obd_processing_task = 0;
784 spin_unlock_bh(&obd->obd_processing_task_lock);
787 spin_unlock_bh(&obd->obd_processing_task_lock);
792 int target_queue_recovery_request(struct ptlrpc_request *req,
793 struct obd_device *obd)
795 struct list_head *tmp;
797 __u64 transno = req->rq_reqmsg->transno;
798 struct ptlrpc_request *saved_req;
799 struct lustre_msg *reqmsg;
801 /* CAVEAT EMPTOR: The incoming request message has been swabbed
802 * (i.e. buflens etc are in my own byte order), but type-dependent
803 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
806 INIT_LIST_HEAD(&req->rq_list);
807 DEBUG_REQ(D_HA, req, "not queueing");
811 /* XXX If I were a real man, these LBUGs would be sane cleanups. */
812 /* XXX just like the request-dup code in queue_final_reply */
813 OBD_ALLOC(saved_req, sizeof *saved_req);
816 OBD_ALLOC(reqmsg, req->rq_reqlen);
820 spin_lock_bh(&obd->obd_processing_task_lock);
822 /* If we're processing the queue, we want don't want to queue this
825 * Also, if this request has a transno less than the one we're waiting
826 * for, we should process it now. It could (and currently always will)
827 * be an open request for a descriptor that was opened some time ago.
829 if (obd->obd_processing_task == current->pid ||
830 transno < obd->obd_next_recovery_transno) {
831 /* Processing the queue right now, don't re-add. */
832 LASSERT(list_empty(&req->rq_list));
833 spin_unlock_bh(&obd->obd_processing_task_lock);
834 OBD_FREE(reqmsg, req->rq_reqlen);
835 OBD_FREE(saved_req, sizeof *saved_req);
839 memcpy(saved_req, req, sizeof *req);
840 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
842 req->rq_reqmsg = reqmsg;
843 class_export_get(req->rq_export);
844 INIT_LIST_HEAD(&req->rq_list);
847 list_for_each(tmp, &obd->obd_recovery_queue) {
848 struct ptlrpc_request *reqiter =
849 list_entry(tmp, struct ptlrpc_request, rq_list);
851 if (reqiter->rq_reqmsg->transno > transno) {
852 list_add_tail(&req->rq_list, &reqiter->rq_list);
859 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
862 obd->obd_requests_queued_for_recovery++;
864 if (obd->obd_processing_task != 0) {
865 /* Someone else is processing this queue, we'll leave it to
868 wake_up(&obd->obd_next_transno_waitq);
869 spin_unlock_bh(&obd->obd_processing_task_lock);
873 /* Nobody is processing, and we know there's (at least) one to process
874 * now, so we'll do the honours.
876 obd->obd_processing_task = current->pid;
877 spin_unlock_bh(&obd->obd_processing_task_lock);
879 process_recovery_queue(obd);
883 struct obd_device * target_req2obd(struct ptlrpc_request *req)
885 return req->rq_export->exp_obd;
888 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
890 struct obd_device *obd = target_req2obd(req);
891 struct ptlrpc_request *saved_req;
892 struct lustre_msg *reqmsg;
893 int recovery_done = 0;
896 LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
899 /* Just like ptlrpc_error, but without the sending. */
900 rc = lustre_pack_reply(req, 0, NULL, NULL);
901 LASSERT(rc == 0); /* XXX handle this */
902 req->rq_type = PTL_RPC_MSG_ERR;
905 LASSERT (!req->rq_reply_state->rs_difficult);
906 LASSERT(list_empty(&req->rq_list));
907 /* XXX a bit like the request-dup code in queue_recovery_request */
908 OBD_ALLOC(saved_req, sizeof *saved_req);
911 OBD_ALLOC(reqmsg, req->rq_reqlen);
914 memcpy(saved_req, req, sizeof *saved_req);
915 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
916 /* the copied req takes over the reply state */
917 req->rq_reply_state = NULL;
919 req->rq_reqmsg = reqmsg;
920 class_export_get(req->rq_export);
921 list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
923 spin_lock_bh(&obd->obd_processing_task_lock);
924 --obd->obd_recoverable_clients;
925 recovery_done = (obd->obd_recoverable_clients == 0);
926 spin_unlock_bh(&obd->obd_processing_task_lock);
929 struct list_head *tmp, *n;
930 ldlm_reprocess_all_ns(req->rq_export->exp_obd->obd_namespace);
931 CWARN("%s: all clients recovered, sending delayed replies\n",
933 spin_lock_bh(&obd->obd_processing_task_lock);
934 obd->obd_recovering = 0;
935 target_cancel_recovery_timer(obd);
936 spin_unlock_bh(&obd->obd_processing_task_lock);
938 /* when recovery finished, cleanup orphans on mds and ost */
939 if (OBT(obd) && OBP(obd, postrecov)) {
940 rc2 = OBP(obd, postrecov)(obd);
942 CWARN("%s: all clients recovered, %d MDS "
943 "orphans deleted\n", obd->obd_name, rc2);
945 CERROR("postrecov failed %d\n", rc2);
948 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
949 req = list_entry(tmp, struct ptlrpc_request, rq_list);
950 DEBUG_REQ(D_ERROR, req, "delayed:");
952 class_export_put(req->rq_export);
953 list_del(&req->rq_list);
954 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
955 OBD_FREE(req, sizeof *req);
957 ptlrpc_run_recovery_over_upcall(obd);
959 CWARN("%s: %d recoverable clients remain\n",
960 obd->obd_name, obd->obd_recoverable_clients);
961 wake_up(&obd->obd_next_transno_waitq);
968 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
970 if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
971 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
972 DEBUG_REQ(D_ERROR, req, "dropping reply");
973 /* NB this does _not_ send with ACK disabled, to simulate
974 * sending OK, but timing out for the ACK */
975 if (req->rq_reply_state != NULL) {
976 if (!req->rq_reply_state->rs_difficult) {
977 lustre_free_reply_state (req->rq_reply_state);
978 req->rq_reply_state = NULL;
980 struct ptlrpc_service *svc =
981 req->rq_rqbd->rqbd_srv_ni->sni_service;
982 atomic_inc(&svc->srv_outstanding_replies);
989 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
990 if (req->rq_reply_state == NULL) {
991 rc = lustre_pack_reply (req, 0, NULL, NULL);
993 CERROR ("can't allocate reply\n");
997 req->rq_type = PTL_RPC_MSG_ERR;
999 DEBUG_REQ(D_NET, req, "sending reply");
1002 return (ptlrpc_send_reply(req, 1));
1006 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1009 unsigned long flags;
1010 struct ptlrpc_reply_state *rs;
1011 struct obd_device *obd;
1012 struct obd_export *exp;
1013 struct ptlrpc_srv_ni *sni;
1014 struct ptlrpc_service *svc;
1016 sni = req->rq_rqbd->rqbd_srv_ni;
1017 svc = sni->sni_service;
1019 rs = req->rq_reply_state;
1020 if (rs == NULL || !rs->rs_difficult) {
1021 /* The easy case; no notifiers and reply_out_callback()
1022 * cleans up (i.e. we can't look inside rs after a
1023 * successful send) */
1024 netrc = target_send_reply_msg (req, rc, fail_id);
1026 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1030 /* must be an export if locks saved */
1031 LASSERT (req->rq_export != NULL);
1032 /* req/reply consistent */
1033 LASSERT (rs->rs_srv_ni == sni);
1036 LASSERT (!rs->rs_scheduled);
1037 LASSERT (!rs->rs_scheduled_ever);
1038 LASSERT (!rs->rs_handled);
1039 LASSERT (!rs->rs_on_net);
1040 LASSERT (rs->rs_export == NULL);
1041 LASSERT (list_empty(&rs->rs_obd_list));
1042 LASSERT (list_empty(&rs->rs_exp_list));
1044 exp = class_export_get (req->rq_export);
1047 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1048 rs->rs_scheduled = 1;
1050 rs->rs_xid = req->rq_xid;
1051 rs->rs_transno = req->rq_transno;
1052 rs->rs_export = exp;
1054 spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1056 if (rs->rs_transno > obd->obd_last_committed) {
1057 /* not committed already */
1058 list_add_tail (&rs->rs_obd_list,
1059 &obd->obd_uncommitted_replies);
1062 spin_unlock (&obd->obd_uncommitted_replies_lock);
1063 spin_lock (&exp->exp_lock);
1065 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1067 spin_unlock_irqrestore (&exp->exp_lock, flags);
1069 netrc = target_send_reply_msg (req, rc, fail_id);
1071 spin_lock_irqsave (&svc->srv_lock, flags);
1073 svc->srv_n_difficult_replies++;
1075 if (netrc != 0) /* error sending: reply is off the net */
1078 if (!rs->rs_on_net || /* some notifier */
1079 list_empty(&rs->rs_exp_list) || /* completed already */
1080 list_empty(&rs->rs_obd_list)) {
1081 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1082 wake_up (&svc->srv_waitq);
1084 list_add (&rs->rs_list, &sni->sni_active_replies);
1085 rs->rs_scheduled = 0; /* allow notifier to schedule */
1088 spin_unlock_irqrestore (&svc->srv_lock, flags);
1091 int target_handle_ping(struct ptlrpc_request *req)
1093 return lustre_pack_reply(req, 0, NULL, NULL);