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>
38 #include <linux/lustre_sec.h>
40 /* @priority: if non-zero, move the selected to the list head
41 * @nocreate: if non-zero, only search in existed connections
43 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
44 int priority, int nocreate)
46 struct ptlrpc_connection *ptlrpc_conn;
47 struct obd_import_conn *imp_conn = NULL, *item;
51 LASSERT(!(nocreate && !priority));
53 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
55 CERROR("can't find connection %s\n", uuid->uuid);
60 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
62 CERROR("fail to alloc memory\n");
63 GOTO(out_put, rc = -ENOMEM);
67 spin_lock(&imp->imp_lock);
68 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
69 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
71 list_del(&item->oic_item);
72 list_add(&item->oic_item, &imp->imp_conn_list);
73 item->oic_last_attempt = 0;
75 CDEBUG(D_HA, "imp %p@%s: find existed conn %s%s\n",
76 imp, imp->imp_obd->obd_name, uuid->uuid,
77 (priority ? ", move to head." : ""));
78 spin_unlock(&imp->imp_lock);
79 GOTO(out_free, rc = 0);
84 imp_conn->oic_conn = ptlrpc_conn;
85 imp_conn->oic_uuid = *uuid;
86 imp_conn->oic_last_attempt = 0;
88 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
90 list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
91 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
92 imp, imp->imp_obd->obd_name, uuid->uuid,
93 (priority ? "head" : "tail"));
97 spin_unlock(&imp->imp_lock);
101 OBD_FREE(imp_conn, sizeof(*imp_conn));
103 ptlrpc_put_connection(ptlrpc_conn);
107 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
109 return import_set_conn(imp, uuid, 1, 1);
112 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
115 return import_set_conn(imp, uuid, priority, 0);
118 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
120 struct obd_import_conn *imp_conn;
121 struct obd_export *dlmexp;
125 spin_lock(&imp->imp_lock);
126 if (list_empty(&imp->imp_conn_list)) {
127 LASSERT(!imp->imp_conn_current);
128 LASSERT(!imp->imp_connection);
132 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
133 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
135 LASSERT(imp_conn->oic_conn);
137 /* is current conn? */
138 if (imp_conn == imp->imp_conn_current) {
139 LASSERT(imp_conn->oic_conn == imp->imp_connection);
141 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
142 imp->imp_state != LUSTRE_IMP_DISCON) {
143 CERROR("can't remove current connection\n");
144 GOTO(out, rc = -EBUSY);
147 ptlrpc_put_connection(imp->imp_connection);
148 imp->imp_connection = NULL;
150 dlmexp = class_conn2export(&imp->imp_dlm_handle);
151 if (dlmexp && dlmexp->exp_connection) {
152 LASSERT(dlmexp->exp_connection ==
154 ptlrpc_put_connection(dlmexp->exp_connection);
155 dlmexp->exp_connection = NULL;
159 list_del(&imp_conn->oic_item);
160 ptlrpc_put_connection(imp_conn->oic_conn);
161 OBD_FREE(imp_conn, sizeof(*imp_conn));
162 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
163 imp, imp->imp_obd->obd_name, uuid->uuid);
168 spin_unlock(&imp->imp_lock);
170 CERROR("connection %s not found\n", uuid->uuid);
174 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
176 struct lustre_cfg* lcfg = buf;
177 struct client_obd *cli = &obddev->u.cli;
178 struct obd_import *imp;
179 struct obd_uuid server_uuid;
180 int rq_portal, rp_portal, connect_op;
181 char *name = obddev->obd_type->typ_name;
182 char *mgmt_name = NULL;
186 /* In a more perfect world, we would hang a ptlrpc_client off of
187 * obd_type and just use the values from there. */
188 if (!strcmp(name, OBD_OSC_DEVICENAME)) {
189 rq_portal = OST_REQUEST_PORTAL;
190 rp_portal = OSC_REPLY_PORTAL;
191 connect_op = OST_CONNECT;
192 } else if (!strcmp(name, OBD_MDC_DEVICENAME)) {
193 rq_portal = MDS_REQUEST_PORTAL;
194 rp_portal = MDC_REPLY_PORTAL;
195 connect_op = MDS_CONNECT;
196 } else if (!strcmp(name, OBD_MGMTCLI_DEVICENAME)) {
197 rq_portal = MGMT_REQUEST_PORTAL;
198 rp_portal = MGMT_REPLY_PORTAL;
199 connect_op = MGMT_CONNECT;
201 CERROR("unknown client OBD type \"%s\", can't setup\n",
207 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
208 CERROR("requires a TARGET UUID\n");
212 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
213 CERROR("client UUID must be less than 38 characters\n");
217 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
218 CERROR("setup requires a SERVER UUID\n");
222 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
223 CERROR("target UUID must be less than 38 characters\n");
227 sema_init(&cli->cl_sem, 1);
228 cli->cl_conn_count = 0;
229 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
230 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
231 sizeof(server_uuid)));
234 cli->cl_avail_grant = 0;
235 /* FIXME: should limit this for the sum of all cl_dirty_max */
236 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
237 if (cli->cl_dirty_max >> PAGE_SHIFT > num_physpages / 8)
238 cli->cl_dirty_max = num_physpages << (PAGE_SHIFT - 3);
239 INIT_LIST_HEAD(&cli->cl_cache_waiters);
240 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
241 INIT_LIST_HEAD(&cli->cl_loi_write_list);
242 INIT_LIST_HEAD(&cli->cl_loi_read_list);
243 spin_lock_init(&cli->cl_loi_list_lock);
244 cli->cl_r_in_flight = 0;
245 cli->cl_w_in_flight = 0;
246 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
247 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
248 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
249 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
251 if (num_physpages >> (20 - PAGE_SHIFT) <= 128) { /* <= 128 MB */
252 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 4;
253 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 4;
255 } else if (num_physpages >> (20 - PAGE_SHIFT) <= 512) { /* <= 512 MB */
256 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 2;
257 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 2;
260 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
261 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
266 CERROR("ldlm_get_ref failed: %d\n", rc);
270 ptlrpc_init_client(rq_portal, rp_portal, name,
271 &obddev->obd_ldlm_client);
273 imp = class_new_import();
275 GOTO(err_ldlm, rc = -ENOENT);
276 imp->imp_client = &obddev->obd_ldlm_client;
277 imp->imp_obd = obddev;
278 imp->imp_connect_op = connect_op;
279 imp->imp_generation = 0;
280 imp->imp_initial_recov = 1;
281 INIT_LIST_HEAD(&imp->imp_pinger_chain);
282 memcpy(imp->imp_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
283 LUSTRE_CFG_BUFLEN(lcfg, 1));
284 class_import_put(imp);
286 rc = client_import_add_conn(imp, &server_uuid, 1);
288 CERROR("can't add initial connection\n");
289 GOTO(err_import, rc);
292 cli->cl_import = imp;
293 cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
294 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
295 cli->cl_sandev = to_kdev_t(0);
297 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
298 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
299 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
300 name, obddev->obd_name,
301 imp->imp_target_uuid.uuid);
302 imp->imp_invalid = 1;
304 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0)
305 mgmt_name = lustre_cfg_string(lcfg, 4);
307 mgmt_name = lustre_cfg_string(lcfg, 3);
311 if (mgmt_name != NULL) {
312 /* Register with management client if we need to. */
313 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
314 obddev->obd_name, mgmt_name, server_uuid.uuid);
316 mgmt_obd = class_name2obd(mgmt_name);
318 CERROR("can't find mgmtcli %s to register\n",
320 GOTO(err_import, rc = -ENOSYS);
323 register_f = (mgmtcli_register_for_events_t)symbol_get("mgmtcli_register_for_events");
325 CERROR("can't i_m_g mgmtcli_register_for_events\n");
326 GOTO(err_import, rc = -ENOSYS);
329 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
330 symbol_put("mgmtcli_register_for_events");
333 cli->cl_mgmtcli_obd = mgmt_obd;
339 class_destroy_import(imp);
347 int client_obd_cleanup(struct obd_device *obddev, int flags)
349 struct client_obd *cli = &obddev->u.cli;
354 if (cli->cl_mgmtcli_obd) {
355 mgmtcli_deregister_for_events_t dereg_f;
357 dereg_f = (mgmtcli_deregister_for_events_t)symbol_get("mgmtcli_deregister_for_events");
358 dereg_f(cli->cl_mgmtcli_obd, obddev);
359 symbol_put("mgmtcli_deregister_for_events");
362 /* Here we try to drop the security structure after destroy import,
363 * to avoid issue of "sleep in spinlock".
365 class_import_get(cli->cl_import);
366 class_destroy_import(cli->cl_import);
367 ptlrpcs_import_drop_sec(cli->cl_import);
368 class_import_put(cli->cl_import);
369 cli->cl_import = NULL;
371 ldlm_put_ref(flags & OBD_OPT_FORCE);
376 int client_connect_import(struct lustre_handle *dlm_handle,
377 struct obd_device *obd,
378 struct obd_uuid *cluuid,
379 struct obd_connect_data *conn_data,
380 unsigned long connect_flags)
382 struct client_obd *cli = &obd->u.cli;
383 struct obd_import *imp = cli->cl_import;
384 struct obd_export *exp;
389 rc = class_connect(dlm_handle, obd, cluuid);
393 cli->cl_conn_count++;
394 if (cli->cl_conn_count > 1)
396 exp = class_conn2export(dlm_handle);
398 if (obd->obd_namespace != NULL)
399 CERROR("already have namespace!\n");
400 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
401 LDLM_NAMESPACE_CLIENT);
402 if (obd->obd_namespace == NULL)
403 GOTO(out_disco, rc = -ENOMEM);
405 rc = ptlrpcs_import_get_sec(imp);
409 imp->imp_dlm_handle = *dlm_handle;
410 rc = ptlrpc_init_import(imp);
414 imp->imp_connect_flags = connect_flags;
416 memcpy(&imp->imp_connect_data, conn_data, sizeof(*conn_data));
418 rc = ptlrpc_connect_import(imp, NULL);
420 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
423 LASSERT(exp->exp_connection);
424 ptlrpc_pinger_add_import(imp);
429 ldlm_namespace_free(obd->obd_namespace, 0);
430 obd->obd_namespace = NULL;
432 cli->cl_conn_count--;
433 class_disconnect(exp, 0);
435 class_export_put(exp);
442 int client_disconnect_export(struct obd_export *exp, unsigned long flags)
444 struct obd_device *obd = class_exp2obd(exp);
445 struct client_obd *cli = &obd->u.cli;
446 struct obd_import *imp = cli->cl_import;
451 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
452 exp, exp ? exp->exp_handle.h_cookie : -1);
457 if (!cli->cl_conn_count) {
458 CERROR("disconnecting disconnected device (%s)\n",
460 GOTO(out_sem, rc = -EINVAL);
463 cli->cl_conn_count--;
464 if (cli->cl_conn_count)
465 GOTO(out_no_disconnect, rc = 0);
467 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
468 * delete it regardless. (It's safe to delete an import that was
470 (void)ptlrpc_pinger_del_import(imp);
472 if (obd->obd_namespace != NULL) {
473 /* obd_no_recov == local only */
474 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
475 obd->obd_no_recov, NULL);
476 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
477 obd->obd_namespace = NULL;
481 * Yeah, obd_no_recov also (mainly) means "forced shutdown".
483 if (obd->obd_no_recov)
484 ptlrpc_invalidate_import(imp, 0);
486 rc = ptlrpc_disconnect_import(imp);
490 err = class_disconnect(exp, 0);
498 /* --------------------------------------------------------------------------
499 * from old lib/target.c
500 * -------------------------------------------------------------------------- */
502 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
503 struct obd_uuid *cluuid, int initial_conn)
505 if (exp->exp_connection && !initial_conn) {
506 struct lustre_handle *hdl;
507 hdl = &exp->exp_imp_reverse->imp_remote_handle;
508 /* Might be a re-connect after a partition. */
509 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
510 CERROR("%s reconnecting\n", cluuid->uuid);
511 conn->cookie = exp->exp_handle.h_cookie;
514 CERROR("%s reconnecting from %s, "
515 "handle mismatch (ours "LPX64", theirs "
516 LPX64")\n", cluuid->uuid,
517 exp->exp_connection->c_remote_uuid.uuid,
518 hdl->cookie, conn->cookie);
519 memset(conn, 0, sizeof *conn);
524 conn->cookie = exp->exp_handle.h_cookie;
525 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
527 CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
531 static inline int ptlrpc_peer_is_local(struct ptlrpc_peer *peer)
533 ptl_process_id_t myid;
535 PtlGetId(peer->peer_ni->pni_ni_h, &myid);
536 return (memcmp(&peer->peer_id, &myid, sizeof(myid)) == 0);
539 /* To check whether the p_flavor is in deny list or not
542 * EPERM found, refuse
545 static int check_deny_list(struct list_head *head, __u32 flavor)
547 deny_sec_t *p_deny_sec = NULL;
548 deny_sec_t *n_deny_sec = NULL;
550 list_for_each_entry_safe(p_deny_sec, n_deny_sec, head, list) {
551 if (p_deny_sec->flavor == flavor)
557 int target_check_deny_sec(struct obd_device *target, struct ptlrpc_request *req)
562 flavor = req->rq_req_secflvr;
564 if (!strcmp(target->obd_type->typ_name, OBD_MDS_DEVICENAME)) {
565 spin_lock(&target->u.mds.mds_denylist_lock);
566 rc = check_deny_list(&target->u.mds.mds_denylist, flavor);
567 spin_unlock(&target->u.mds.mds_denylist_lock);
568 } else if (!strcmp(target->obd_type->typ_name, "obdfilter")) {
569 spin_lock(&target->u.filter.fo_denylist_lock);
570 rc = check_deny_list(&target->u.filter.fo_denylist, flavor);
571 spin_unlock(&target->u.filter.fo_denylist_lock);
577 int target_handle_connect(struct ptlrpc_request *req)
579 unsigned long connect_flags = 0, *cfp;
580 struct obd_device *target;
581 struct obd_export *export = NULL;
582 struct obd_import *revimp;
583 struct lustre_handle conn;
584 struct obd_uuid tgtuuid;
585 struct obd_uuid cluuid;
586 struct obd_uuid remote_uuid;
588 struct obd_connect_data *conn_data;
589 int conn_data_size = sizeof(*conn_data);
593 int initial_conn = 0;
594 char peer_str[PTL_NALFMT_SIZE];
595 const int offset = 1;
598 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
600 LASSERT_REQSWAB (req, offset + 0);
601 str = lustre_msg_string(req->rq_reqmsg, offset + 0,
602 sizeof(tgtuuid) - 1);
604 CERROR("bad target UUID for connect\n");
605 GOTO(out, rc = -EINVAL);
608 obd_str2uuid (&tgtuuid, str);
609 target = class_uuid2obd(&tgtuuid);
611 target = class_name2obd(str);
613 if (!target || target->obd_stopping || !target->obd_set_up) {
614 CERROR("UUID '%s' is not available for connect from %s\n",
615 str, req->rq_peerstr);
616 GOTO(out, rc = -ENODEV);
619 /* check the secure deny list of mds/ost */
620 rc = target_check_deny_sec(target, req);
624 LASSERT_REQSWAB (req, offset + 1);
625 str = lustre_msg_string(req->rq_reqmsg, offset + 1, sizeof(cluuid) - 1);
627 CERROR("bad client UUID for connect\n");
628 GOTO(out, rc = -EINVAL);
631 obd_str2uuid (&cluuid, str);
633 /* XXX extract a nettype and format accordingly */
634 switch (sizeof(ptl_nid_t)) {
635 /* NB the casts only avoid compiler warnings */
637 snprintf((char *)remote_uuid.uuid, sizeof(remote_uuid),
638 "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_id.nid);
641 snprintf((char *)remote_uuid.uuid, sizeof(remote_uuid),
642 "NET_%x_UUID", (__u32)req->rq_peer.peer_id.nid);
648 tmp = lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn));
650 GOTO(out, rc = -EPROTO);
652 memcpy(&conn, tmp, sizeof conn);
654 cfp = lustre_msg_buf(req->rq_reqmsg, offset + 3, sizeof(unsigned long));
655 LASSERT(cfp != NULL);
656 connect_flags = *cfp;
658 conn_data = lustre_swab_reqbuf(req, offset + 4, sizeof(*conn_data),
659 lustre_swab_connect);
661 GOTO(out, rc = -EPROTO);
663 rc = lustre_pack_reply(req, 1, &conn_data_size, NULL);
667 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
670 /* lctl gets a backstage, all-access pass. */
671 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
672 goto dont_check_exports;
674 spin_lock(&target->obd_dev_lock);
675 list_for_each(p, &target->obd_exports) {
676 export = list_entry(p, struct obd_export, exp_obd_chain);
677 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
678 spin_unlock(&target->obd_dev_lock);
679 LASSERT(export->exp_obd == target);
681 rc = target_handle_reconnect(&conn, export, &cluuid,
687 /* If we found an export, we already unlocked. */
689 spin_unlock(&target->obd_dev_lock);
690 } else if (req->rq_export == NULL &&
691 atomic_read(&export->exp_rpc_count) > 0) {
692 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
693 target->obd_name, cluuid.uuid,
694 ptlrpc_peernid2str(&req->rq_peer, peer_str),
695 export, atomic_read(&export->exp_refcount));
696 GOTO(out, rc = -EBUSY);
697 } else if (req->rq_export != NULL &&
698 atomic_read(&export->exp_rpc_count) > 1) {
699 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
700 target->obd_name, cluuid.uuid,
701 ptlrpc_peernid2str(&req->rq_peer, peer_str),
702 export, atomic_read(&export->exp_rpc_count));
703 GOTO(out, rc = -EBUSY);
704 } else if (req->rq_reqmsg->conn_cnt == 1 && !initial_conn) {
705 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
707 GOTO(out, rc = -EALREADY);
710 /* Tell the client if we're in recovery. */
711 /* If this is the first client, start the recovery timer */
712 CWARN("%s: connection from %s@%s/%lu %st"LPU64"\n", target->obd_name,
713 cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str), *cfp,
714 target->obd_recovering ? "recovering/" : "", conn_data->transno);
716 if (target->obd_recovering) {
717 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
718 target_start_recovery_timer(target);
722 /* Tell the client if we support replayable requests */
723 if (target->obd_replayable)
724 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
727 if (export == NULL) {
728 if (target->obd_recovering) {
729 CERROR("%s denying connection for new client %s@%s: "
730 "%d clients in recovery for %lds\n", target->obd_name,
732 ptlrpc_peernid2str(&req->rq_peer, peer_str),
733 target->obd_recoverable_clients,
734 (target->obd_recovery_timer.expires-jiffies)/HZ);
738 rc = obd_connect(&conn, target, &cluuid, conn_data,
743 /* Return only the parts of obd_connect_data that we understand, so the
744 * client knows that we don't understand the rest. */
745 conn_data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
746 memcpy(lustre_msg_buf(req->rq_repmsg, 0, sizeof(*conn_data)), conn_data,
749 /* Tell the client if we support replayable requests */
750 if (target->obd_replayable)
751 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
753 /* If all else goes well, this is our RPC return code. */
756 if (rc && rc != EALREADY)
759 req->rq_repmsg->handle = conn;
761 /* If the client and the server are the same node, we will already
762 * have an export that really points to the client's DLM export,
763 * because we have a shared handles table.
765 * XXX this will go away when shaver stops sending the "connect" handle
766 * in the real "remote handle" field of the request --phik 24 Apr 2003
768 if (req->rq_export != NULL)
769 class_export_put(req->rq_export);
771 /* ownership of this export ref transfers to the request */
772 export = req->rq_export = class_conn2export(&conn);
773 LASSERT(export != NULL);
775 spin_lock_irqsave(&export->exp_lock, flags);
777 req->rq_repmsg->conn_cnt = export->exp_conn_cnt + 1;
778 } else if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
779 CERROR("%s@%s: already connected at a higher conn_cnt: %d > %d\n",
780 cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str),
781 export->exp_conn_cnt,
782 req->rq_reqmsg->conn_cnt);
783 spin_unlock_irqrestore(&export->exp_lock, flags);
784 GOTO(out, rc = -EALREADY);
786 export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
787 spin_unlock_irqrestore(&export->exp_lock, flags);
789 /* request from liblustre? */
790 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
791 export->exp_libclient = 1;
793 if (!(lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_ASYNC) &&
794 ptlrpc_peer_is_local(&req->rq_peer)) {
795 CWARN("%s: exp %p set sync\n", target->obd_name, export);
796 export->exp_sync = 1;
798 CDEBUG(D_HA, "%s: exp %p set async\n",target->obd_name,export);
799 export->exp_sync = 0;
802 if (export->exp_connection != NULL)
803 ptlrpc_put_connection(export->exp_connection);
804 export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
807 if (rc == EALREADY) {
808 /* We indicate the reconnection in a flag, not an error code. */
809 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
813 spin_lock_bh(&target->obd_processing_task_lock);
814 if (target->obd_recovering && export->exp_connected == 0) {
815 __u64 t = conn_data->transno;
816 export->exp_connected = 1;
817 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
818 && t < target->obd_next_recovery_transno)
819 target->obd_next_recovery_transno = t;
820 target->obd_connected_clients++;
821 if (target->obd_connected_clients == target->obd_max_recoverable_clients)
822 wake_up(&target->obd_next_transno_waitq);
824 spin_unlock_bh(&target->obd_processing_task_lock);
826 memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn)),
829 if (export->exp_imp_reverse != NULL) {
830 /* same logic as client_obd_cleanup */
831 class_import_get(export->exp_imp_reverse);
832 class_destroy_import(export->exp_imp_reverse);
833 ptlrpcs_import_drop_sec(export->exp_imp_reverse);
834 class_import_put(export->exp_imp_reverse);
837 /* for the rest part, we return -ENOTCONN in case of errors
838 * in order to let client initialize connection again.
840 revimp = export->exp_imp_reverse = class_new_import();
842 CERROR("fail to alloc new reverse import.\n");
843 GOTO(out, rc = -ENOTCONN);
846 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
847 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
848 revimp->imp_remote_handle = conn;
849 revimp->imp_obd = target;
850 revimp->imp_dlm_fake = 1;
851 revimp->imp_state = LUSTRE_IMP_FULL;
853 rc = ptlrpcs_import_get_sec(revimp);
855 CERROR("reverse import can not get sec: %d\n", rc);
856 class_destroy_import(revimp);
857 export->exp_imp_reverse = NULL;
858 GOTO(out, rc = -ENOTCONN);
861 class_import_put(revimp);
863 rc = obd_connect_post(export, initial_conn, connect_flags);
870 int target_handle_disconnect(struct ptlrpc_request *req)
872 struct obd_export *exp;
876 rc = lustre_pack_reply(req, 0, NULL, NULL);
880 /* keep the rq_export around so we can send the reply */
881 exp = class_export_get(req->rq_export);
882 req->rq_status = obd_disconnect(exp, 0);
886 void target_destroy_export(struct obd_export *exp)
888 /* exports created from last_rcvd data, and "fake"
889 exports created by lctl don't have an import */
890 if (exp->exp_imp_reverse != NULL) {
891 ptlrpcs_import_drop_sec(exp->exp_imp_reverse);
892 class_destroy_import(exp->exp_imp_reverse);
895 /* We cancel locks at disconnect time, but this will catch any locks
896 * granted in a race with recovery-induced disconnect. */
897 if (exp->exp_obd->obd_namespace != NULL)
898 ldlm_cancel_locks_for_export(exp);
905 struct ptlrpc_request *
906 ptlrpc_clone_req( struct ptlrpc_request *orig_req)
908 struct ptlrpc_request *copy_req;
909 struct lustre_msg *copy_reqmsg;
911 OBD_ALLOC(copy_req, sizeof *copy_req);
914 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
916 OBD_FREE(copy_req, sizeof *copy_req);
920 memcpy(copy_req, orig_req, sizeof *copy_req);
921 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
922 /* the copied req takes over the reply state and security data */
923 orig_req->rq_reply_state = NULL;
924 orig_req->rq_svcsec_data = NULL;
926 copy_req->rq_reqmsg = copy_reqmsg;
927 class_export_get(copy_req->rq_export);
928 INIT_LIST_HEAD(©_req->rq_list);
933 void ptlrpc_free_clone( struct ptlrpc_request *req)
936 svcsec_cleanup_req(req);
938 class_export_put(req->rq_export);
939 list_del(&req->rq_list);
940 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
941 OBD_FREE(req, sizeof *req);
944 static void target_release_saved_req(struct ptlrpc_request *req)
947 svcsec_cleanup_req(req);
949 class_export_put(req->rq_export);
950 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
951 OBD_FREE(req, sizeof *req);
954 static void target_finish_recovery(struct obd_device *obd)
958 ldlm_reprocess_all_ns(obd->obd_namespace);
960 /* when recovery finished, cleanup orphans on mds and ost */
961 if (OBT(obd) && OBP(obd, postrecov)) {
962 rc = OBP(obd, postrecov)(obd);
964 CWARN("%s: all clients recovered, %d MDS "
965 "orphans deleted\n", obd->obd_name, rc);
967 CERROR("postrecov failed %d\n", rc);
970 obd->obd_recovery_end = LTIME_S(CURRENT_TIME);
974 static void abort_req_replay_queue(struct obd_device *obd)
976 struct ptlrpc_request *req;
977 struct list_head *tmp, *n;
980 list_for_each_safe(tmp, n, &obd->obd_req_replay_queue) {
981 req = list_entry(tmp, struct ptlrpc_request, rq_list);
982 list_del(&req->rq_list);
983 DEBUG_REQ(D_ERROR, req, "aborted:");
984 req->rq_status = -ENOTCONN;
985 req->rq_type = PTL_RPC_MSG_ERR;
986 rc = lustre_pack_reply(req, 0, NULL, NULL);
990 DEBUG_REQ(D_ERROR, req,
991 "packing failed for abort-reply; skipping");
993 target_release_saved_req(req);
997 static void abort_lock_replay_queue(struct obd_device *obd)
999 struct ptlrpc_request *req;
1000 struct list_head *tmp, *n;
1003 list_for_each_safe(tmp, n, &obd->obd_lock_replay_queue) {
1004 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1005 list_del(&req->rq_list);
1006 DEBUG_REQ(D_ERROR, req, "aborted:");
1007 req->rq_status = -ENOTCONN;
1008 req->rq_type = PTL_RPC_MSG_ERR;
1009 rc = lustre_pack_reply(req, 0, NULL, NULL);
1013 DEBUG_REQ(D_ERROR, req,
1014 "packing failed for abort-reply; skipping");
1016 target_release_saved_req(req);
1020 /* Called from a cleanup function if the device is being cleaned up
1021 forcefully. The exports should all have been disconnected already,
1022 the only thing left to do is
1023 - clear the recovery flags
1025 - free queued requests and replies, but don't send replies
1026 Because the obd_stopping flag is set, no new requests should be received.
1029 void target_cleanup_recovery(struct obd_device *obd)
1031 struct list_head *tmp, *n;
1032 struct ptlrpc_request *req;
1034 spin_lock_bh(&obd->obd_processing_task_lock);
1035 if (!obd->obd_recovering) {
1036 spin_unlock_bh(&obd->obd_processing_task_lock);
1040 obd->obd_recovering = obd->obd_abort_recovery = 0;
1041 target_cancel_recovery_timer(obd);
1042 spin_unlock_bh(&obd->obd_processing_task_lock);
1044 list_for_each_safe(tmp, n, &obd->obd_req_replay_queue) {
1045 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1046 list_del(&req->rq_list);
1047 LASSERT (req->rq_reply_state == 0);
1048 target_release_saved_req(req);
1050 list_for_each_safe(tmp, n, &obd->obd_lock_replay_queue) {
1051 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1052 list_del(&req->rq_list);
1053 LASSERT (req->rq_reply_state == 0);
1054 target_release_saved_req(req);
1056 list_for_each_safe(tmp, n, &obd->obd_final_req_queue) {
1057 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1058 list_del(&req->rq_list);
1059 LASSERT (req->rq_reply_state == 0);
1060 target_release_saved_req(req);
1065 static void target_abort_recovery(void *data)
1067 struct obd_device *obd = data;
1069 LASSERT(!obd->obd_recovering);
1071 class_disconnect_stale_exports(obd, 0);
1073 CERROR("%s: recovery period over; disconnecting unfinished clients.\n",
1076 abort_recovery_queue(obd);
1077 target_finish_recovery(obd);
1078 ptlrpc_run_recovery_over_upcall(obd);
1082 static void target_recovery_expired(unsigned long castmeharder)
1084 struct obd_device *obd = (struct obd_device *)castmeharder;
1085 spin_lock_bh(&obd->obd_processing_task_lock);
1086 if (obd->obd_recovering)
1087 obd->obd_abort_recovery = 1;
1089 wake_up(&obd->obd_next_transno_waitq);
1090 spin_unlock_bh(&obd->obd_processing_task_lock);
1094 /* obd_processing_task_lock should be held */
1095 void target_cancel_recovery_timer(struct obd_device *obd)
1097 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1098 del_timer(&obd->obd_recovery_timer);
1102 static void reset_recovery_timer(struct obd_device *obd)
1104 spin_lock_bh(&obd->obd_processing_task_lock);
1105 if (!obd->obd_recovering) {
1106 spin_unlock_bh(&obd->obd_processing_task_lock);
1109 CDEBUG(D_HA, "timer will expire in %u seconds\n",
1110 OBD_RECOVERY_TIMEOUT / HZ);
1111 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
1112 spin_unlock_bh(&obd->obd_processing_task_lock);
1116 /* Only start it the first time called */
1117 void target_start_recovery_timer(struct obd_device *obd)
1119 spin_lock_bh(&obd->obd_processing_task_lock);
1120 if (!obd->obd_recovering || timer_pending(&obd->obd_recovery_timer)) {
1121 spin_unlock_bh(&obd->obd_processing_task_lock);
1124 CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1125 OBD_RECOVERY_TIMEOUT / HZ);
1126 obd->obd_recovery_timer.function = target_recovery_expired;
1127 obd->obd_recovery_timer.data = (unsigned long)obd;
1128 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
1129 spin_unlock_bh(&obd->obd_processing_task_lock);
1133 static int check_for_next_transno(struct obd_device *obd)
1135 struct ptlrpc_request *req = NULL;
1136 int wake_up = 0, connected, completed, queue_len, max;
1137 __u64 next_transno, req_transno;
1139 spin_lock_bh(&obd->obd_processing_task_lock);
1140 if (!list_empty(&obd->obd_req_replay_queue)) {
1141 req = list_entry(obd->obd_req_replay_queue.next,
1142 struct ptlrpc_request, rq_list);
1143 req_transno = req->rq_reqmsg->transno;
1148 max = obd->obd_max_recoverable_clients;
1149 connected = obd->obd_connected_clients;
1150 completed = max - obd->obd_recoverable_clients;
1151 queue_len = obd->obd_requests_queued_for_recovery;
1152 next_transno = obd->obd_next_recovery_transno;
1154 CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
1155 "req_transno: "LPU64", next_transno: "LPU64"\n",
1156 max, connected, completed, queue_len, req_transno, next_transno);
1157 if (obd->obd_abort_recovery) {
1158 CDEBUG(D_HA, "waking for aborted recovery\n");
1160 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1161 CDEBUG(D_HA, "waking for completed recovery\n");
1163 } else if (req_transno == next_transno) {
1164 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1166 } else if (queue_len + completed == max) {
1167 LASSERT(req->rq_reqmsg->transno >= next_transno);
1168 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1169 "waking for skipped transno (skip: "LPD64
1170 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1171 next_transno, queue_len, completed, max, req_transno);
1172 obd->obd_next_recovery_transno = req_transno;
1174 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1175 /* some clients haven't connected in time, but we need
1176 * their requests to continue recovery. so, we abort ... */
1177 CDEBUG(D_ERROR, "abort due to missed clients: queue: %d max: %d\n",
1179 obd->obd_abort_recovery = 1;
1182 spin_unlock_bh(&obd->obd_processing_task_lock);
1187 static struct ptlrpc_request *
1188 target_next_replay_req(struct obd_device *obd)
1190 struct l_wait_info lwi = { 0 };
1191 struct ptlrpc_request *req;
1193 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1194 obd->obd_next_recovery_transno);
1195 l_wait_event(obd->obd_next_transno_waitq,
1196 check_for_next_transno(obd), &lwi);
1198 spin_lock_bh(&obd->obd_processing_task_lock);
1199 if (obd->obd_abort_recovery) {
1201 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1202 req = list_entry(obd->obd_req_replay_queue.next,
1203 struct ptlrpc_request, rq_list);
1204 list_del_init(&req->rq_list);
1205 obd->obd_requests_queued_for_recovery--;
1209 spin_unlock_bh(&obd->obd_processing_task_lock);
1213 static int check_for_next_lock(struct obd_device *obd)
1215 struct ptlrpc_request *req = NULL;
1218 spin_lock_bh(&obd->obd_processing_task_lock);
1219 if (!list_empty(&obd->obd_lock_replay_queue)) {
1220 req = list_entry(obd->obd_lock_replay_queue.next,
1221 struct ptlrpc_request, rq_list);
1222 CDEBUG(D_HA, "waking for next lock\n");
1224 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1225 CDEBUG(D_HA, "waking for completed lock replay\n");
1227 } else if (obd->obd_abort_recovery) {
1228 CDEBUG(D_HA, "waking for aborted recovery\n");
1231 spin_unlock_bh(&obd->obd_processing_task_lock);
1236 static struct ptlrpc_request *
1237 target_next_replay_lock(struct obd_device *obd)
1239 struct l_wait_info lwi = { 0 };
1240 struct ptlrpc_request *req;
1242 CDEBUG(D_HA, "Waiting for lock\n");
1243 l_wait_event(obd->obd_next_transno_waitq,
1244 check_for_next_lock(obd), &lwi);
1246 spin_lock_bh(&obd->obd_processing_task_lock);
1247 if (obd->obd_abort_recovery) {
1249 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1250 req = list_entry(obd->obd_lock_replay_queue.next,
1251 struct ptlrpc_request, rq_list);
1252 list_del_init(&req->rq_list);
1256 spin_unlock_bh(&obd->obd_processing_task_lock);
1260 static struct ptlrpc_request *
1261 target_next_final_ping(struct obd_device *obd)
1263 struct ptlrpc_request *req;
1265 spin_lock_bh(&obd->obd_processing_task_lock);
1266 if (!list_empty(&obd->obd_final_req_queue)) {
1267 req = list_entry(obd->obd_final_req_queue.next,
1268 struct ptlrpc_request, rq_list);
1269 list_del_init(&req->rq_list);
1273 spin_unlock_bh(&obd->obd_processing_task_lock);
1277 static int req_replay_done(struct obd_export *exp)
1279 if (exp->exp_req_replay_needed)
1284 static int lock_replay_done(struct obd_export *exp)
1286 if (exp->exp_lock_replay_needed)
1291 static int connect_done(struct obd_export *exp)
1293 if (exp->exp_connected)
1298 static int check_for_clients(struct obd_device *obd)
1300 if (obd->obd_abort_recovery)
1302 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1303 if (obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1308 static int target_recovery_thread(void *arg)
1310 struct obd_device *obd = arg;
1311 struct ptlrpc_request *req;
1312 struct target_recovery_data *trd = &obd->obd_recovery_data;
1313 char peer_str[PTL_NALFMT_SIZE];
1314 struct l_wait_info lwi = { 0 };
1315 unsigned long flags;
1318 kportal_daemonize("tgt-recov");
1320 SIGNAL_MASK_LOCK(current, flags);
1321 sigfillset(¤t->blocked);
1323 SIGNAL_MASK_UNLOCK(current, flags);
1325 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1327 trd->trd_processing_task = current->pid;
1329 obd->obd_recovering = 1;
1330 complete(&trd->trd_starting);
1332 /* first of all, we have to know the first transno to replay */
1333 obd->obd_abort_recovery = 0;
1334 l_wait_event(obd->obd_next_transno_waitq,
1335 check_for_clients(obd), &lwi);
1337 spin_lock_bh(&obd->obd_processing_task_lock);
1338 target_cancel_recovery_timer(obd);
1339 spin_unlock_bh(&obd->obd_processing_task_lock);
1341 /* If some clients haven't connected in time, evict them */
1342 if (obd->obd_abort_recovery) {
1344 CERROR("some clients haven't connect in time (%d/%d),"
1345 "evict them ...\n", obd->obd_connected_clients,
1346 obd->obd_max_recoverable_clients);
1347 obd->obd_abort_recovery = 0;
1348 stale = class_disconnect_stale_exports(obd, connect_done, 0);
1349 atomic_sub(stale, &obd->obd_req_replay_clients);
1350 atomic_sub(stale, &obd->obd_lock_replay_clients);
1353 /* next stage: replay requests */
1354 CWARN("1: request replay stage - %d clients from t"LPU64"\n",
1355 atomic_read(&obd->obd_req_replay_clients),
1356 obd->obd_next_recovery_transno);
1357 while ((req = target_next_replay_req(obd))) {
1358 LASSERT(trd->trd_processing_task == current->pid);
1359 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s: ",
1360 req->rq_reqmsg->transno,
1361 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1362 (void)trd->trd_recovery_handler(req);
1363 obd->obd_replayed_requests++;
1364 reset_recovery_timer(obd);
1365 /* bug 1580: decide how to properly sync() in recovery*/
1366 //mds_fsync_super(mds->mds_sb);
1367 ptlrpc_free_clone(req);
1368 spin_lock_bh(&obd->obd_processing_task_lock);
1369 obd->obd_next_recovery_transno++;
1370 spin_unlock_bh(&obd->obd_processing_task_lock);
1373 spin_lock_bh(&obd->obd_processing_task_lock);
1374 target_cancel_recovery_timer(obd);
1375 spin_unlock_bh(&obd->obd_processing_task_lock);
1377 /* If some clients haven't replayed requests in time, evict them */
1378 if (obd->obd_abort_recovery) {
1380 CERROR("req replay timed out, aborting ...\n");
1381 obd->obd_abort_recovery = 0;
1382 stale = class_disconnect_stale_exports(obd, req_replay_done, 0);
1383 atomic_sub(stale, &obd->obd_lock_replay_clients);
1384 abort_req_replay_queue(obd);
1387 /* The second stage: replay locks */
1388 CWARN("2: lock replay stage - %d clients\n",
1389 atomic_read(&obd->obd_lock_replay_clients));
1390 while ((req = target_next_replay_lock(obd))) {
1391 LASSERT(trd->trd_processing_task == current->pid);
1392 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1393 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1394 (void)trd->trd_recovery_handler(req);
1395 reset_recovery_timer(obd);
1396 ptlrpc_free_clone(req);
1397 obd->obd_replayed_locks++;
1400 spin_lock_bh(&obd->obd_processing_task_lock);
1401 target_cancel_recovery_timer(obd);
1402 spin_unlock_bh(&obd->obd_processing_task_lock);
1404 /* If some clients haven't replayed requests in time, evict them */
1405 if (obd->obd_abort_recovery) {
1407 CERROR("lock replay timed out, aborting ...\n");
1408 obd->obd_abort_recovery = 0;
1409 stale = class_disconnect_stale_exports(obd, lock_replay_done, 0);
1410 abort_lock_replay_queue(obd);
1413 /* We drop recoverying flag to forward all new requests
1414 * to regular mds_handle() since now */
1415 spin_lock_bh(&obd->obd_processing_task_lock);
1416 obd->obd_recovering = 0;
1417 spin_unlock_bh(&obd->obd_processing_task_lock);
1419 /* The third stage: reply on final pings */
1420 CWARN("3: final stage - process recovery completion pings\n");
1421 while ((req = target_next_final_ping(obd))) {
1422 LASSERT(trd->trd_processing_task == current->pid);
1423 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1424 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1425 (void)trd->trd_recovery_handler(req);
1426 ptlrpc_free_clone(req);
1429 CWARN("4: recovery completed - %d/%d reqs/locks replayed\n",
1430 obd->obd_replayed_requests, obd->obd_replayed_locks);
1431 target_finish_recovery(obd);
1433 trd->trd_processing_task = 0;
1434 complete(&trd->trd_finishing);
1438 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1441 struct target_recovery_data *trd = &obd->obd_recovery_data;
1443 memset(trd, 0, sizeof(*trd));
1444 init_completion(&trd->trd_starting);
1445 init_completion(&trd->trd_finishing);
1446 trd->trd_recovery_handler = handler;
1448 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1449 wait_for_completion(&trd->trd_starting);
1450 LASSERT(obd->obd_recovering != 0);
1457 void target_stop_recovery_thread(struct obd_device *obd)
1459 spin_lock_bh(&obd->obd_processing_task_lock);
1460 if (obd->obd_recovery_data.trd_processing_task > 0) {
1461 struct target_recovery_data *trd = &obd->obd_recovery_data;
1462 CERROR("%s: aborting recovery\n", obd->obd_name);
1463 obd->obd_abort_recovery = 1;
1464 wake_up(&obd->obd_next_transno_waitq);
1465 spin_unlock_bh(&obd->obd_processing_task_lock);
1466 wait_for_completion(&trd->trd_finishing);
1468 spin_unlock_bh(&obd->obd_processing_task_lock);
1473 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1475 struct obd_export *exp = req->rq_export;
1476 LASSERT(exp != NULL);
1477 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1478 /* client declares he's ready to replay locks */
1479 spin_lock_bh(&obd->obd_processing_task_lock);
1480 if (exp->exp_req_replay_needed) {
1481 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1482 exp->exp_req_replay_needed = 0;
1483 atomic_dec(&obd->obd_req_replay_clients);
1484 obd->obd_recoverable_clients--;
1485 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1486 CDEBUG(D_HA, "all clients have replayed reqs\n");
1487 wake_up(&obd->obd_next_transno_waitq);
1489 spin_unlock_bh(&obd->obd_processing_task_lock);
1491 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1492 /* client declares he's ready to complete recovery
1493 * so, we put the request on th final queue */
1494 spin_lock_bh(&obd->obd_processing_task_lock);
1495 if (exp->exp_lock_replay_needed) {
1496 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1497 exp->exp_lock_replay_needed = 0;
1498 atomic_dec(&obd->obd_lock_replay_clients);
1499 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1500 CDEBUG(D_HA, "all clients have replayed locks\n");
1501 wake_up(&obd->obd_next_transno_waitq);
1503 spin_unlock_bh(&obd->obd_processing_task_lock);
1509 int target_queue_recovery_request(struct ptlrpc_request *req,
1510 struct obd_device *obd)
1512 struct list_head *tmp;
1514 __u64 transno = req->rq_reqmsg->transno;
1516 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
1517 /* Processing the queue right now, don't re-add. */
1521 target_process_req_flags(obd, req);
1523 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1524 /* client declares he's ready to complete recovery
1525 * so, we put the request on th final queue */
1526 req = ptlrpc_clone_req(req);
1529 DEBUG_REQ(D_HA, req, "queue final req");
1530 spin_lock_bh(&obd->obd_processing_task_lock);
1531 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1532 spin_unlock_bh(&obd->obd_processing_task_lock);
1535 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1536 /* client declares he's ready to replay locks */
1537 req = ptlrpc_clone_req(req);
1540 DEBUG_REQ(D_HA, req, "queue lock replay req");
1541 spin_lock_bh(&obd->obd_processing_task_lock);
1542 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1543 spin_unlock_bh(&obd->obd_processing_task_lock);
1544 wake_up(&obd->obd_next_transno_waitq);
1549 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1550 * (i.e. buflens etc are in my own byte order), but type-dependent
1551 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1554 INIT_LIST_HEAD(&req->rq_list);
1555 DEBUG_REQ(D_HA, req, "not queueing");
1560 /* If we're processing the queue, we want don't want to queue this
1563 * Also, if this request has a transno less than the one we're waiting
1564 * for, we should process it now. It could (and currently always will)
1565 * be an open request for a descriptor that was opened some time ago.
1567 * Also, a resent, replayed request that has already been
1568 * handled will pass through here and be processed immediately.
1570 spin_lock_bh(&obd->obd_processing_task_lock);
1571 if (transno < obd->obd_next_recovery_transno && check_for_clients(obd)) {
1572 /* Processing the queue right now, don't re-add. */
1573 LASSERT(list_empty(&req->rq_list));
1574 spin_unlock_bh(&obd->obd_processing_task_lock);
1577 spin_unlock_bh(&obd->obd_processing_task_lock);
1579 /* A resent, replayed request that is still on the queue; just drop it.
1580 The queued request will handle this. */
1581 if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY))
1582 == (MSG_RESENT | MSG_REPLAY)) {
1583 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1587 req = ptlrpc_clone_req(req);
1591 spin_lock_bh(&obd->obd_processing_task_lock);
1594 list_for_each(tmp, &obd->obd_req_replay_queue) {
1595 struct ptlrpc_request *reqiter =
1596 list_entry(tmp, struct ptlrpc_request, rq_list);
1598 if (reqiter->rq_reqmsg->transno > transno) {
1599 list_add_tail(&req->rq_list, &reqiter->rq_list);
1606 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1608 obd->obd_requests_queued_for_recovery++;
1609 wake_up(&obd->obd_next_transno_waitq);
1610 spin_unlock_bh(&obd->obd_processing_task_lock);
1614 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1616 return req->rq_export->exp_obd;
1620 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1622 if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1623 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1624 DEBUG_REQ(D_ERROR, req, "dropping reply");
1625 /* NB this does _not_ send with ACK disabled, to simulate
1626 * sending OK, but timing out for the ACK */
1627 if (req->rq_reply_state != NULL) {
1628 if (!req->rq_reply_state->rs_difficult) {
1629 lustre_free_reply_state (req->rq_reply_state);
1630 req->rq_reply_state = NULL;
1632 struct ptlrpc_service *svc =
1633 req->rq_rqbd->rqbd_srv_ni->sni_service;
1634 atomic_inc(&svc->srv_outstanding_replies);
1641 req->rq_status = rc;
1642 return (ptlrpc_error(req));
1644 DEBUG_REQ(D_NET, req, "sending reply");
1647 return (ptlrpc_send_reply(req, 1));
1651 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1654 unsigned long flags;
1655 struct ptlrpc_reply_state *rs;
1656 struct obd_device *obd;
1657 struct obd_export *exp;
1658 struct ptlrpc_srv_ni *sni;
1659 struct ptlrpc_service *svc;
1661 sni = req->rq_rqbd->rqbd_srv_ni;
1662 svc = sni->sni_service;
1664 rs = req->rq_reply_state;
1665 if (rs == NULL || !rs->rs_difficult) {
1666 /* The easy case; no notifiers and reply_out_callback()
1667 * cleans up (i.e. we can't look inside rs after a
1668 * successful send) */
1669 netrc = target_send_reply_msg (req, rc, fail_id);
1671 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1675 /* must be an export if locks saved */
1676 LASSERT (req->rq_export != NULL);
1677 /* req/reply consistent */
1678 LASSERT (rs->rs_srv_ni == sni);
1681 LASSERT (!rs->rs_scheduled);
1682 LASSERT (!rs->rs_scheduled_ever);
1683 LASSERT (!rs->rs_handled);
1684 LASSERT (!rs->rs_on_net);
1685 LASSERT (rs->rs_export == NULL);
1686 LASSERT (list_empty(&rs->rs_obd_list));
1687 LASSERT (list_empty(&rs->rs_exp_list));
1689 exp = class_export_get (req->rq_export);
1692 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1693 rs->rs_scheduled = 1;
1695 rs->rs_xid = req->rq_xid;
1696 rs->rs_transno = req->rq_transno;
1697 rs->rs_export = exp;
1699 spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1701 if (rs->rs_transno > obd->obd_last_committed) {
1702 /* not committed already */
1703 list_add_tail (&rs->rs_obd_list,
1704 &obd->obd_uncommitted_replies);
1707 spin_unlock (&obd->obd_uncommitted_replies_lock);
1708 spin_lock (&exp->exp_lock);
1710 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1712 spin_unlock_irqrestore (&exp->exp_lock, flags);
1714 netrc = target_send_reply_msg (req, rc, fail_id);
1716 spin_lock_irqsave (&svc->srv_lock, flags);
1718 svc->srv_n_difficult_replies++;
1720 if (netrc != 0) /* error sending: reply is off the net */
1723 if (!rs->rs_on_net || /* some notifier */
1724 list_empty(&rs->rs_exp_list) || /* completed already */
1725 list_empty(&rs->rs_obd_list)) {
1726 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1727 wake_up (&svc->srv_waitq);
1729 list_add (&rs->rs_list, &sni->sni_active_replies);
1730 rs->rs_scheduled = 0; /* allow notifier to schedule */
1733 spin_unlock_irqrestore (&svc->srv_lock, flags);
1736 int target_handle_ping(struct ptlrpc_request *req)
1738 return lustre_pack_reply(req, 0, NULL, NULL);