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,
223 unsigned long connect_flags)
225 struct client_obd *cli = &obd->u.cli;
226 struct obd_import *imp = cli->cl_import;
227 struct obd_export *exp;
232 rc = class_connect(dlm_handle, obd, cluuid);
236 cli->cl_conn_count++;
237 if (cli->cl_conn_count > 1)
239 exp = class_conn2export(dlm_handle);
241 if (obd->obd_namespace != NULL)
242 CERROR("already have namespace!\n");
243 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
244 LDLM_NAMESPACE_CLIENT);
245 if (obd->obd_namespace == NULL)
246 GOTO(out_disco, rc = -ENOMEM);
248 imp->imp_dlm_handle = *dlm_handle;
249 rc = ptlrpc_init_import(imp);
253 exp->exp_connection = ptlrpc_connection_addref(imp->imp_connection);
254 imp->imp_connect_flags = connect_flags;
255 rc = ptlrpc_connect_import(imp, NULL);
257 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
261 ptlrpc_pinger_add_import(imp);
266 ldlm_namespace_free(obd->obd_namespace, 0);
267 obd->obd_namespace = NULL;
269 cli->cl_conn_count--;
270 class_disconnect(exp, 0);
272 class_export_put(exp);
279 int client_disconnect_export(struct obd_export *exp, int failover)
281 struct obd_device *obd = class_exp2obd(exp);
282 struct client_obd *cli = &obd->u.cli;
283 struct obd_import *imp = cli->cl_import;
288 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
289 exp, exp ? exp->exp_handle.h_cookie : -1);
294 if (!cli->cl_conn_count) {
295 CERROR("disconnecting disconnected device (%s)\n",
297 GOTO(out_sem, rc = -EINVAL);
300 cli->cl_conn_count--;
301 if (cli->cl_conn_count)
302 GOTO(out_no_disconnect, rc = 0);
304 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
305 * delete it regardless. (It's safe to delete an import that was
307 (void)ptlrpc_pinger_del_import(imp);
309 if (obd->obd_namespace != NULL) {
310 /* obd_no_recov == local only */
311 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
312 obd->obd_no_recov, NULL);
313 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
314 obd->obd_namespace = NULL;
317 /* Yeah, obd_no_recov also (mainly) means "forced shutdown". */
318 if (obd->obd_no_recov)
319 ptlrpc_invalidate_import(imp, 0);
321 rc = ptlrpc_disconnect_import(imp);
325 err = class_disconnect(exp, 0);
333 /* --------------------------------------------------------------------------
334 * from old lib/target.c
335 * -------------------------------------------------------------------------- */
337 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
338 struct obd_uuid *cluuid, int initial_conn)
340 if (exp->exp_connection && !initial_conn) {
341 struct lustre_handle *hdl;
342 hdl = &exp->exp_imp_reverse->imp_remote_handle;
343 /* Might be a re-connect after a partition. */
344 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
345 CERROR("%s reconnecting\n", cluuid->uuid);
346 conn->cookie = exp->exp_handle.h_cookie;
349 CERROR("%s reconnecting from %s, "
350 "handle mismatch (ours "LPX64", theirs "
351 LPX64")\n", cluuid->uuid,
352 exp->exp_connection->c_remote_uuid.uuid,
353 hdl->cookie, conn->cookie);
354 memset(conn, 0, sizeof *conn);
359 conn->cookie = exp->exp_handle.h_cookie;
360 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
362 CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
366 int target_handle_connect(struct ptlrpc_request *req)
368 unsigned long connect_flags = 0, *cfp;
369 struct obd_device *target;
370 struct obd_export *export = NULL;
371 struct obd_import *revimp;
372 struct lustre_handle conn;
373 struct obd_uuid tgtuuid;
374 struct obd_uuid cluuid;
375 struct obd_uuid remote_uuid;
380 int initial_conn = 0;
381 char peer_str[PTL_NALFMT_SIZE];
384 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
386 LASSERT_REQSWAB (req, 0);
387 str = lustre_msg_string(req->rq_reqmsg, 0, sizeof(tgtuuid) - 1);
389 CERROR("bad target UUID for connect\n");
390 GOTO(out, rc = -EINVAL);
393 obd_str2uuid (&tgtuuid, str);
394 target = class_uuid2obd(&tgtuuid);
396 target = class_name2obd(str);
399 if (!target || target->obd_stopping || !target->obd_set_up) {
400 CERROR("UUID '%s' is not available for connect\n", str);
402 GOTO(out, rc = -ENODEV);
405 LASSERT_REQSWAB (req, 1);
406 str = lustre_msg_string(req->rq_reqmsg, 1, sizeof(cluuid) - 1);
408 CERROR("bad client UUID for connect\n");
409 GOTO(out, rc = -EINVAL);
412 obd_str2uuid (&cluuid, str);
414 /* XXX extract a nettype and format accordingly */
415 switch (sizeof(ptl_nid_t)) {
416 /* NB the casts only avoid compiler warnings */
418 snprintf(remote_uuid.uuid, sizeof remote_uuid,
419 "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_nid);
422 snprintf(remote_uuid.uuid, sizeof remote_uuid,
423 "NET_%x_UUID", (__u32)req->rq_peer.peer_nid);
429 tmp = lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn);
431 GOTO(out, rc = -EPROTO);
433 memcpy(&conn, tmp, sizeof conn);
435 cfp = lustre_msg_buf(req->rq_reqmsg, 3, sizeof(unsigned long));
436 LASSERT(cfp != NULL);
437 connect_flags = *cfp;
439 rc = lustre_pack_reply(req, 0, NULL, NULL);
443 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
446 /* lctl gets a backstage, all-access pass. */
447 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
448 goto dont_check_exports;
450 spin_lock(&target->obd_dev_lock);
451 list_for_each(p, &target->obd_exports) {
452 export = list_entry(p, struct obd_export, exp_obd_chain);
453 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
454 spin_unlock(&target->obd_dev_lock);
455 LASSERT(export->exp_obd == target);
457 rc = target_handle_reconnect(&conn, export, &cluuid,
463 /* If we found an export, we already unlocked. */
465 spin_unlock(&target->obd_dev_lock);
466 } else if (req->rq_export == NULL &&
467 atomic_read(&export->exp_rpc_count) > 0) {
468 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
469 target->obd_name, cluuid.uuid,
470 ptlrpc_peernid2str(&req->rq_peer, peer_str),
471 export, atomic_read(&export->exp_refcount));
472 GOTO(out, rc = -EBUSY);
473 } else if (req->rq_export != NULL &&
474 atomic_read(&export->exp_rpc_count) > 1) {
475 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
476 target->obd_name, cluuid.uuid,
477 ptlrpc_peernid2str(&req->rq_peer, peer_str),
478 export, atomic_read(&export->exp_rpc_count));
479 GOTO(out, rc = -EBUSY);
481 else if (req->rq_reqmsg->conn_cnt == 1 && !initial_conn) {
482 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
484 GOTO(out, rc = -EALREADY);
487 /* Tell the client if we're in recovery. */
488 /* If this is the first client, start the recovery timer */
489 CWARN("%s: connection from %s@%s/%lu %s\n", target->obd_name, cluuid.uuid,
490 ptlrpc_peernid2str(&req->rq_peer, peer_str), *cfp,
491 target->obd_recovering ? "(recovering)" : "");
492 if (target->obd_recovering) {
493 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
494 target_start_recovery_timer(target);
497 /* Tell the client if we support replayable requests */
498 if (target->obd_replayable)
499 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
501 if (export == NULL) {
502 if (target->obd_recovering) {
503 CERROR("denying connection for new client %s@%s: "
504 "%d clients in recovery for %lds\n", cluuid.uuid,
505 ptlrpc_peernid2str(&req->rq_peer, peer_str),
506 target->obd_recoverable_clients,
507 (target->obd_recovery_timer.expires-jiffies)/HZ);
511 rc = obd_connect(&conn, target, &cluuid, connect_flags);
514 /* Tell the client if we support replayable requests */
515 if (target->obd_replayable)
516 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
518 /* If all else goes well, this is our RPC return code. */
521 if (rc && rc != EALREADY)
524 req->rq_repmsg->handle = conn;
526 /* If the client and the server are the same node, we will already
527 * have an export that really points to the client's DLM export,
528 * because we have a shared handles table.
530 * XXX this will go away when shaver stops sending the "connect" handle
531 * in the real "remote handle" field of the request --phik 24 Apr 2003
533 if (req->rq_export != NULL)
534 class_export_put(req->rq_export);
536 /* ownership of this export ref transfers to the request */
537 export = req->rq_export = class_conn2export(&conn);
538 LASSERT(export != NULL);
540 spin_lock_irqsave(&export->exp_lock, flags);
542 req->rq_repmsg->conn_cnt = export->exp_conn_cnt + 1;
543 } else if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
544 CERROR("%s@%s: already connected at a higher conn_cnt: %d > %d\n",
545 cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str),
546 export->exp_conn_cnt,
547 req->rq_reqmsg->conn_cnt);
548 spin_unlock_irqrestore(&export->exp_lock, flags);
549 GOTO(out, rc = -EALREADY);
551 export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
552 spin_unlock_irqrestore(&export->exp_lock, flags);
554 /* request from liblustre? */
555 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
556 export->exp_libclient = 1;
558 if (export->exp_connection != NULL)
559 ptlrpc_put_connection(export->exp_connection);
560 export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
563 if (rc == EALREADY) {
564 /* We indicate the reconnection in a flag, not an error code. */
565 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
569 if (target->obd_recovering) {
570 target->obd_connected_clients++;
573 memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, 2, sizeof conn),
576 if (export->exp_imp_reverse != NULL)
577 class_destroy_import(export->exp_imp_reverse);
578 revimp = export->exp_imp_reverse = class_new_import();
579 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
580 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
581 revimp->imp_remote_handle = conn;
582 revimp->imp_obd = target;
583 revimp->imp_dlm_fake = 1;
584 revimp->imp_state = LUSTRE_IMP_FULL;
585 class_import_put(revimp);
587 rc = obd_connect_post(export);
594 int target_handle_disconnect(struct ptlrpc_request *req)
596 struct obd_export *exp;
600 rc = lustre_pack_reply(req, 0, NULL, NULL);
604 /* keep the rq_export around so we can send the reply */
605 exp = class_export_get(req->rq_export);
606 req->rq_status = obd_disconnect(exp, 0);
610 void target_destroy_export(struct obd_export *exp)
612 /* exports created from last_rcvd data, and "fake"
613 exports created by lctl don't have an import */
614 if (exp->exp_imp_reverse != NULL)
615 class_destroy_import(exp->exp_imp_reverse);
617 /* We cancel locks at disconnect time, but this will catch any locks
618 * granted in a race with recovery-induced disconnect. */
619 if (exp->exp_obd->obd_namespace != NULL)
620 ldlm_cancel_locks_for_export(exp);
627 struct ptlrpc_request *
628 ptlrpc_clone_req( struct ptlrpc_request *orig_req)
630 struct ptlrpc_request *copy_req;
631 struct lustre_msg *copy_reqmsg;
633 OBD_ALLOC(copy_req, sizeof *copy_req);
636 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
638 OBD_FREE(copy_req, sizeof *copy_req);
642 memcpy(copy_req, orig_req, sizeof *copy_req);
643 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
644 /* the copied req takes over the reply state */
645 orig_req->rq_reply_state = NULL;
647 copy_req->rq_reqmsg = copy_reqmsg;
648 class_export_get(copy_req->rq_export);
649 INIT_LIST_HEAD(©_req->rq_list);
653 void ptlrpc_free_clone( struct ptlrpc_request *req)
655 class_export_put(req->rq_export);
656 list_del(&req->rq_list);
657 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
658 OBD_FREE(req, sizeof *req);
661 static void abort_delayed_replies(struct obd_device *obd)
663 struct ptlrpc_request *req;
664 struct list_head *tmp, *n;
665 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
666 req = list_entry(tmp, struct ptlrpc_request, rq_list);
667 DEBUG_REQ(D_ERROR, req, "aborted:");
668 req->rq_status = -ENOTCONN;
669 req->rq_type = PTL_RPC_MSG_ERR;
671 class_export_put(req->rq_export);
672 list_del(&req->rq_list);
673 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
674 OBD_FREE(req, sizeof *req);
678 static void abort_recovery_queue(struct obd_device *obd)
680 struct ptlrpc_request *req;
681 struct list_head *tmp, *n;
684 list_for_each_safe(tmp, n, &obd->obd_recovery_queue) {
685 req = list_entry(tmp, struct ptlrpc_request, rq_list);
686 DEBUG_REQ(D_ERROR, req, "aborted:");
687 req->rq_status = -ENOTCONN;
688 req->rq_type = PTL_RPC_MSG_ERR;
689 rc = lustre_pack_reply(req, 0, NULL, NULL);
693 DEBUG_REQ(D_ERROR, req,
694 "packing failed for abort-reply; skipping");
696 list_del(&req->rq_list);
697 class_export_put(req->rq_export);
698 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
699 OBD_FREE(req, sizeof *req);
703 static void target_abort_recovery(void *data)
705 struct obd_device *obd = data;
708 CERROR("disconnecting clients and aborting recovery\n");
709 LASSERT(!obd->obd_recovering);
711 class_disconnect_exports(obd, 0);
713 /* when recovery was aborted, cleanup orphans on mds and ost */
714 if (OBT(obd) && OBP(obd, postrecov)) {
715 rc = OBP(obd, postrecov)(obd);
717 CWARN("Cleanup %d orphans after recovery was aborted\n",
720 CERROR("postrecov failed %d\n", rc);
723 abort_delayed_replies(obd);
724 abort_recovery_queue(obd);
725 ptlrpc_run_recovery_over_upcall(obd);
728 static void target_recovery_expired(unsigned long castmeharder)
730 struct obd_device *obd = (struct obd_device *)castmeharder;
731 CERROR("recovery timed out, aborting\n");
732 spin_lock_bh(&obd->obd_processing_task_lock);
733 obd->obd_abort_recovery = 1;
734 wake_up(&obd->obd_next_transno_waitq);
735 spin_unlock_bh(&obd->obd_processing_task_lock);
739 /* obd_processing_task_lock should be held */
740 void target_cancel_recovery_timer(struct obd_device *obd)
742 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
743 del_timer(&obd->obd_recovery_timer);
746 static void reset_recovery_timer(struct obd_device *obd)
748 spin_lock_bh(&obd->obd_processing_task_lock);
749 if (!obd->obd_recovering) {
750 spin_unlock_bh(&obd->obd_processing_task_lock);
753 CDEBUG(D_HA, "timer will expire in %u seconds\n",
754 OBD_RECOVERY_TIMEOUT / HZ);
755 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
756 spin_unlock_bh(&obd->obd_processing_task_lock);
760 /* Only start it the first time called */
761 void target_start_recovery_timer(struct obd_device *obd)
763 spin_lock_bh(&obd->obd_processing_task_lock);
764 if (!obd->obd_recovering || timer_pending(&obd->obd_recovery_timer)) {
765 spin_unlock_bh(&obd->obd_processing_task_lock);
768 CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
769 OBD_RECOVERY_TIMEOUT / HZ);
770 obd->obd_recovery_timer.function = target_recovery_expired;
771 obd->obd_recovery_timer.data = (unsigned long)obd;
772 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
773 spin_unlock_bh(&obd->obd_processing_task_lock);
776 static void target_finish_recovery(struct obd_device *obd)
778 struct list_head *tmp, *n;
781 ldlm_reprocess_all_ns(obd->obd_namespace);
783 CWARN("%s: all clients recovered, calling postrecov\n",
785 /* when recovery finished, cleanup orphans on mds and ost */
786 if (OBT(obd) && OBP(obd, postrecov)) {
787 rc2 = OBP(obd, postrecov)(obd);
789 CWARN("%s: all clients recovered, %d MDS "
790 "orphans deleted\n", obd->obd_name, rc2);
792 CERROR("postrecov failed %d\n", rc2);
795 CWARN("%s: recovery over, sending delayed replies\n",
797 list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
798 struct ptlrpc_request *req;
799 req = list_entry(tmp, struct ptlrpc_request, rq_list);
800 DEBUG_REQ(D_ERROR, req, "delayed:");
802 ptlrpc_free_clone(req);
804 ptlrpc_run_recovery_over_upcall(obd);
807 static int check_for_next_transno(struct obd_device *obd)
809 struct ptlrpc_request *req;
810 int wake_up = 0, connected, completed, queue_len, max;
811 __u64 next_transno, req_transno;
813 spin_lock_bh(&obd->obd_processing_task_lock);
814 if (!list_empty(&obd->obd_recovery_queue)) {
815 req = list_entry(obd->obd_recovery_queue.next,
816 struct ptlrpc_request, rq_list);
817 req_transno = req->rq_reqmsg->transno;
822 max = obd->obd_max_recoverable_clients;
823 connected = obd->obd_connected_clients;
824 completed = max - obd->obd_recoverable_clients;
825 queue_len = obd->obd_requests_queued_for_recovery;
826 next_transno = obd->obd_next_recovery_transno;
828 CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
829 "req_transno: "LPU64", next_transno: "LPU64"\n",
830 max, connected, completed, queue_len, req_transno, next_transno);
831 if (obd->obd_abort_recovery) {
832 CDEBUG(D_HA, "waking for aborted recovery\n");
834 } else if (max == completed) {
835 CDEBUG(D_HA, "waking for completed recovery\n");
837 } else if (req_transno == next_transno) {
838 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
840 } else if (queue_len + completed == max) {
841 LASSERT(req->rq_reqmsg->transno >= next_transno);
843 "waking for skipped transno (skip: "LPD64
844 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
845 next_transno, queue_len, completed, max, req_transno);
846 obd->obd_next_recovery_transno = req_transno;
849 spin_unlock_bh(&obd->obd_processing_task_lock);
854 static struct ptlrpc_request *
855 target_next_replay_req(struct obd_device *obd)
857 struct l_wait_info lwi = { 0 };
858 struct ptlrpc_request *req;
860 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
861 obd->obd_next_recovery_transno);
862 l_wait_event(obd->obd_next_transno_waitq,
863 check_for_next_transno(obd), &lwi);
865 spin_lock_bh(&obd->obd_processing_task_lock);
866 if (obd->obd_abort_recovery) {
868 } else if (!list_empty(&obd->obd_recovery_queue)) {
869 req = list_entry(obd->obd_recovery_queue.next,
870 struct ptlrpc_request, rq_list);
871 list_del_init(&req->rq_list);
872 obd->obd_requests_queued_for_recovery--;
876 spin_unlock_bh(&obd->obd_processing_task_lock);
881 static int target_recovery_thread(void *arg)
883 struct obd_device *obd = arg;
884 struct ptlrpc_request *req;
885 struct target_recovery_data *trd = &obd->obd_recovery_data;
889 kportal_daemonize("tgt-recov");
891 SIGNAL_MASK_LOCK(current, flags);
892 sigfillset(¤t->blocked);
894 SIGNAL_MASK_UNLOCK(current, flags);
896 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
898 trd->trd_processing_task = current->pid;
900 obd->obd_recovering = 1;
901 complete(&trd->trd_starting);
903 while (obd->obd_recovering) {
904 LASSERT(trd->trd_processing_task == current->pid);
905 req = target_next_replay_req(obd);
907 char peer_str[PTL_NALFMT_SIZE];
908 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s: ",
909 req->rq_reqmsg->transno,
910 ptlrpc_peernid2str(&req->rq_peer, peer_str));
911 (void)trd->trd_recovery_handler(req);
912 obd->obd_replayed_requests++;
913 reset_recovery_timer(obd);
914 /* bug 1580: decide how to properly sync() in recovery*/
915 //mds_fsync_super(mds->mds_sb);
916 ptlrpc_free_clone(req);
917 spin_lock_bh(&obd->obd_processing_task_lock);
918 obd->obd_next_recovery_transno++;
919 spin_unlock_bh(&obd->obd_processing_task_lock);
921 /* recovery is over */
922 spin_lock_bh(&obd->obd_processing_task_lock);
923 obd->obd_recovering = 0;
924 target_cancel_recovery_timer(obd);
925 if (obd->obd_abort_recovery) {
926 obd->obd_abort_recovery = 0;
927 spin_unlock_bh(&obd->obd_processing_task_lock);
928 target_abort_recovery(obd);
930 LASSERT(obd->obd_recoverable_clients == 0);
931 spin_unlock_bh(&obd->obd_processing_task_lock);
932 target_finish_recovery(obd);
937 trd->trd_processing_task = 0;
938 complete(&trd->trd_finishing);
942 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
945 struct target_recovery_data *trd = &obd->obd_recovery_data;
947 memset(trd, 0, sizeof(*trd));
948 init_completion(&trd->trd_starting);
949 init_completion(&trd->trd_finishing);
950 trd->trd_recovery_handler = handler;
952 if (kernel_thread(target_recovery_thread, obd, 0) == 0)
953 wait_for_completion(&trd->trd_starting);
960 void target_stop_recovery_thread(struct obd_device *obd)
962 spin_lock_bh(&obd->obd_processing_task_lock);
963 if (obd->obd_recovery_data.trd_processing_task > 0) {
964 struct target_recovery_data *trd = &obd->obd_recovery_data;
965 CERROR("%s: aborting recovery\n", obd->obd_name);
966 obd->obd_abort_recovery = 1;
967 wake_up(&obd->obd_next_transno_waitq);
968 spin_unlock_bh(&obd->obd_processing_task_lock);
969 wait_for_completion(&trd->trd_finishing);
971 spin_unlock_bh(&obd->obd_processing_task_lock);
975 int target_queue_recovery_request(struct ptlrpc_request *req,
976 struct obd_device *obd)
978 struct list_head *tmp;
980 __u64 transno = req->rq_reqmsg->transno;
982 /* CAVEAT EMPTOR: The incoming request message has been swabbed
983 * (i.e. buflens etc are in my own byte order), but type-dependent
984 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
987 INIT_LIST_HEAD(&req->rq_list);
988 DEBUG_REQ(D_HA, req, "not queueing");
993 /* If we're processing the queue, we want don't want to queue this
996 * Also, if this request has a transno less than the one we're waiting
997 * for, we should process it now. It could (and currently always will)
998 * be an open request for a descriptor that was opened some time ago.
1000 * Also, a resent, replayed request that has already been
1001 * handled will pass through here and be processed immediately.
1003 spin_lock_bh(&obd->obd_processing_task_lock);
1004 if (obd->obd_recovery_data.trd_processing_task == current->pid ||
1005 transno < obd->obd_next_recovery_transno) {
1006 /* Processing the queue right now, don't re-add. */
1007 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1008 LASSERT(list_empty(&req->rq_list));
1009 spin_unlock_bh(&obd->obd_processing_task_lock);
1012 spin_unlock_bh(&obd->obd_processing_task_lock);
1014 /* A resent, replayed request that is still on the queue; just drop it.
1015 The queued request will handle this. */
1016 if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY))
1017 == (MSG_RESENT | MSG_REPLAY)) {
1018 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1022 req = ptlrpc_clone_req(req);
1026 spin_lock_bh(&obd->obd_processing_task_lock);
1029 list_for_each(tmp, &obd->obd_recovery_queue) {
1030 struct ptlrpc_request *reqiter =
1031 list_entry(tmp, struct ptlrpc_request, rq_list);
1033 if (reqiter->rq_reqmsg->transno > transno) {
1034 list_add_tail(&req->rq_list, &reqiter->rq_list);
1041 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
1044 obd->obd_requests_queued_for_recovery++;
1045 wake_up(&obd->obd_next_transno_waitq);
1046 spin_unlock_bh(&obd->obd_processing_task_lock);
1051 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1053 return req->rq_export->exp_obd;
1056 int target_queue_final_reply(struct ptlrpc_request *req, int rc)
1058 struct obd_device *obd = target_req2obd(req);
1060 LASSERT ((rc == 0) == (req->rq_reply_state != NULL));
1063 /* Just like ptlrpc_error, but without the sending. */
1064 rc = lustre_pack_reply(req, 0, NULL, NULL);
1065 LASSERT(rc == 0); /* XXX handle this */
1066 req->rq_type = PTL_RPC_MSG_ERR;
1069 LASSERT (!req->rq_reply_state->rs_difficult);
1070 LASSERT(list_empty(&req->rq_list));
1072 req = ptlrpc_clone_req(req);
1074 spin_lock_bh(&obd->obd_processing_task_lock);
1076 list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
1078 /* only count the first "replay over" request from each
1080 if (req->rq_export->exp_replay_needed) {
1081 --obd->obd_recoverable_clients;
1082 req->rq_export->exp_replay_needed = 0;
1083 CWARN("%s: %d recoverable clients remain\n",
1084 obd->obd_name, obd->obd_recoverable_clients);
1086 wake_up(&obd->obd_next_transno_waitq);
1087 spin_unlock_bh(&obd->obd_processing_task_lock);
1092 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1094 if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1095 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1096 DEBUG_REQ(D_ERROR, req, "dropping reply");
1097 /* NB this does _not_ send with ACK disabled, to simulate
1098 * sending OK, but timing out for the ACK */
1099 if (req->rq_reply_state != NULL) {
1100 if (!req->rq_reply_state->rs_difficult) {
1101 lustre_free_reply_state (req->rq_reply_state);
1102 req->rq_reply_state = NULL;
1104 struct ptlrpc_service *svc =
1105 req->rq_rqbd->rqbd_srv_ni->sni_service;
1106 atomic_inc(&svc->srv_outstanding_replies);
1113 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1114 if (req->rq_reply_state == NULL) {
1115 rc = lustre_pack_reply (req, 0, NULL, NULL);
1117 CERROR ("can't allocate reply\n");
1121 req->rq_type = PTL_RPC_MSG_ERR;
1123 DEBUG_REQ(D_NET, req, "sending reply");
1126 return (ptlrpc_send_reply(req, 1));
1130 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1133 unsigned long flags;
1134 struct ptlrpc_reply_state *rs;
1135 struct obd_device *obd;
1136 struct obd_export *exp;
1137 struct ptlrpc_srv_ni *sni;
1138 struct ptlrpc_service *svc;
1140 sni = req->rq_rqbd->rqbd_srv_ni;
1141 svc = sni->sni_service;
1143 rs = req->rq_reply_state;
1144 if (rs == NULL || !rs->rs_difficult) {
1145 /* The easy case; no notifiers and reply_out_callback()
1146 * cleans up (i.e. we can't look inside rs after a
1147 * successful send) */
1148 netrc = target_send_reply_msg (req, rc, fail_id);
1150 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1154 /* must be an export if locks saved */
1155 LASSERT (req->rq_export != NULL);
1156 /* req/reply consistent */
1157 LASSERT (rs->rs_srv_ni == sni);
1160 LASSERT (!rs->rs_scheduled);
1161 LASSERT (!rs->rs_scheduled_ever);
1162 LASSERT (!rs->rs_handled);
1163 LASSERT (!rs->rs_on_net);
1164 LASSERT (rs->rs_export == NULL);
1165 LASSERT (list_empty(&rs->rs_obd_list));
1166 LASSERT (list_empty(&rs->rs_exp_list));
1168 exp = class_export_get (req->rq_export);
1171 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1172 rs->rs_scheduled = 1;
1174 rs->rs_xid = req->rq_xid;
1175 rs->rs_transno = req->rq_transno;
1176 rs->rs_export = exp;
1178 spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1180 if (rs->rs_transno > obd->obd_last_committed) {
1181 /* not committed already */
1182 list_add_tail (&rs->rs_obd_list,
1183 &obd->obd_uncommitted_replies);
1186 spin_unlock (&obd->obd_uncommitted_replies_lock);
1187 spin_lock (&exp->exp_lock);
1189 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1191 spin_unlock_irqrestore (&exp->exp_lock, flags);
1193 netrc = target_send_reply_msg (req, rc, fail_id);
1195 spin_lock_irqsave (&svc->srv_lock, flags);
1197 svc->srv_n_difficult_replies++;
1199 if (netrc != 0) /* error sending: reply is off the net */
1202 if (!rs->rs_on_net || /* some notifier */
1203 list_empty(&rs->rs_exp_list) || /* completed already */
1204 list_empty(&rs->rs_obd_list)) {
1205 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1206 wake_up (&svc->srv_waitq);
1208 list_add (&rs->rs_list, &sni->sni_active_replies);
1209 rs->rs_scheduled = 0; /* allow notifier to schedule */
1212 spin_unlock_irqrestore (&svc->srv_lock, flags);
1215 int target_handle_ping(struct ptlrpc_request *req)
1217 return lustre_pack_reply(req, 0, NULL, NULL);