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>
39 #include <linux/lustre_gs.h>
41 /* @priority: if non-zero, move the selected to the list head
42 * @nocreate: if non-zero, only search in existed connections
44 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
45 int priority, int nocreate)
47 struct ptlrpc_connection *ptlrpc_conn;
48 struct obd_import_conn *imp_conn = NULL, *item;
52 LASSERT(!(nocreate && !priority));
54 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
56 CERROR("can't find connection %s\n", uuid->uuid);
61 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
63 CERROR("fail to alloc memory\n");
64 GOTO(out_put, rc = -ENOMEM);
68 spin_lock(&imp->imp_lock);
69 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
70 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
72 list_del(&item->oic_item);
73 list_add(&item->oic_item, &imp->imp_conn_list);
74 item->oic_last_attempt = 0;
76 CDEBUG(D_HA, "imp %p@%s: find existed conn %s%s\n",
77 imp, imp->imp_obd->obd_name, uuid->uuid,
78 (priority ? ", move to head." : ""));
79 spin_unlock(&imp->imp_lock);
80 GOTO(out_free, rc = 0);
85 imp_conn->oic_conn = ptlrpc_conn;
86 imp_conn->oic_uuid = *uuid;
87 imp_conn->oic_last_attempt = 0;
89 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
91 list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
92 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? "head" : "tail"));
98 spin_unlock(&imp->imp_lock);
102 OBD_FREE(imp_conn, sizeof(*imp_conn));
104 ptlrpc_put_connection(ptlrpc_conn);
108 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
110 return import_set_conn(imp, uuid, 1, 1);
113 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
116 return import_set_conn(imp, uuid, priority, 0);
119 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
121 struct obd_import_conn *imp_conn;
122 struct obd_export *dlmexp;
126 spin_lock(&imp->imp_lock);
127 if (list_empty(&imp->imp_conn_list)) {
128 LASSERT(!imp->imp_conn_current);
129 LASSERT(!imp->imp_connection);
133 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
134 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
136 LASSERT(imp_conn->oic_conn);
138 /* is current conn? */
139 if (imp_conn == imp->imp_conn_current) {
140 LASSERT(imp_conn->oic_conn == imp->imp_connection);
142 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
143 imp->imp_state != LUSTRE_IMP_DISCON) {
144 CERROR("can't remove current connection\n");
145 GOTO(out, rc = -EBUSY);
148 ptlrpc_put_connection(imp->imp_connection);
149 imp->imp_connection = NULL;
151 dlmexp = class_conn2export(&imp->imp_dlm_handle);
152 if (dlmexp && dlmexp->exp_connection) {
153 LASSERT(dlmexp->exp_connection ==
155 ptlrpc_put_connection(dlmexp->exp_connection);
156 dlmexp->exp_connection = NULL;
160 list_del(&imp_conn->oic_item);
161 ptlrpc_put_connection(imp_conn->oic_conn);
162 OBD_FREE(imp_conn, sizeof(*imp_conn));
163 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
164 imp, imp->imp_obd->obd_name, uuid->uuid);
169 spin_unlock(&imp->imp_lock);
171 CERROR("connection %s not found\n", uuid->uuid);
175 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
177 struct lustre_cfg* lcfg = buf;
178 struct client_obd *cli = &obddev->u.cli;
179 struct obd_import *imp;
180 struct obd_uuid server_uuid;
181 int rq_portal, rp_portal, connect_op;
182 char *name = obddev->obd_type->typ_name;
183 char *mgmt_name = NULL;
187 /* In a more perfect world, we would hang a ptlrpc_client off of
188 * obd_type and just use the values from there. */
189 if (!strcmp(name, OBD_OSC_DEVICENAME)) {
190 rq_portal = OST_REQUEST_PORTAL;
191 rp_portal = OSC_REPLY_PORTAL;
192 connect_op = OST_CONNECT;
193 } else if (!strcmp(name, OBD_MDC_DEVICENAME)) {
194 rq_portal = MDS_REQUEST_PORTAL;
195 rp_portal = MDC_REPLY_PORTAL;
196 connect_op = MDS_CONNECT;
197 } else if (!strcmp(name, OBD_MGMTCLI_DEVICENAME)) {
198 rq_portal = MGMT_REQUEST_PORTAL;
199 rp_portal = MGMT_REPLY_PORTAL;
200 connect_op = MGMT_CONNECT;
201 } else if (!strcmp(name, LUSTRE_GKC_NAME)) {
202 rq_portal = GKS_REQUEST_PORTAL;
203 rp_portal = GKC_REPLY_PORTAL;
204 connect_op = GKS_CONNECT;
207 CERROR("unknown client OBD type \"%s\", can't setup\n",
213 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
214 CERROR("requires a TARGET UUID\n");
218 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
219 CERROR("client UUID must be less than 38 characters\n");
223 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
224 CERROR("setup requires a SERVER UUID\n");
228 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
229 CERROR("target UUID must be less than 38 characters\n");
233 sema_init(&cli->cl_sem, 1);
234 cli->cl_conn_count = 0;
235 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
236 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
237 sizeof(server_uuid)));
240 cli->cl_avail_grant = 0;
241 /* FIXME: should limit this for the sum of all cl_dirty_max */
242 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
243 if (cli->cl_dirty_max >> PAGE_SHIFT > num_physpages / 8)
244 cli->cl_dirty_max = num_physpages << (PAGE_SHIFT - 3);
245 INIT_LIST_HEAD(&cli->cl_cache_waiters);
246 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
247 INIT_LIST_HEAD(&cli->cl_loi_write_list);
248 INIT_LIST_HEAD(&cli->cl_loi_read_list);
249 spin_lock_init(&cli->cl_loi_list_lock);
250 cli->cl_r_in_flight = 0;
251 cli->cl_w_in_flight = 0;
252 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
253 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
254 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
255 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
257 if (num_physpages >> (20 - PAGE_SHIFT) <= 128) { /* <= 128 MB */
258 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 4;
259 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 4;
261 } else if (num_physpages >> (20 - PAGE_SHIFT) <= 512) { /* <= 512 MB */
262 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 2;
263 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 2;
266 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
267 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
272 CERROR("ldlm_get_ref failed: %d\n", rc);
276 ptlrpc_init_client(rq_portal, rp_portal, name,
277 &obddev->obd_ldlm_client);
279 imp = class_new_import();
281 GOTO(err_ldlm, rc = -ENOENT);
282 imp->imp_client = &obddev->obd_ldlm_client;
283 imp->imp_obd = obddev;
284 imp->imp_connect_op = connect_op;
285 imp->imp_generation = 0;
286 imp->imp_initial_recov = 1;
287 INIT_LIST_HEAD(&imp->imp_pinger_chain);
288 memcpy(imp->imp_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
289 LUSTRE_CFG_BUFLEN(lcfg, 1));
290 class_import_put(imp);
292 rc = client_import_add_conn(imp, &server_uuid, 1);
294 CERROR("can't add initial connection\n");
295 GOTO(err_import, rc);
298 cli->cl_import = imp;
299 cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
300 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
301 cli->cl_sandev = to_kdev_t(0);
303 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
304 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
305 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
306 name, obddev->obd_name,
307 imp->imp_target_uuid.uuid);
308 imp->imp_invalid = 1;
310 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0)
311 mgmt_name = lustre_cfg_string(lcfg, 4);
313 mgmt_name = lustre_cfg_string(lcfg, 3);
317 if (mgmt_name != NULL) {
318 /* Register with management client if we need to. */
319 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
320 obddev->obd_name, mgmt_name, server_uuid.uuid);
322 mgmt_obd = class_name2obd(mgmt_name);
324 CERROR("can't find mgmtcli %s to register\n",
326 GOTO(err_import, rc = -ENOSYS);
329 register_f = (mgmtcli_register_for_events_t)symbol_get("mgmtcli_register_for_events");
331 CERROR("can't i_m_g mgmtcli_register_for_events\n");
332 GOTO(err_import, rc = -ENOSYS);
335 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
336 symbol_put("mgmtcli_register_for_events");
339 cli->cl_mgmtcli_obd = mgmt_obd;
345 class_destroy_import(imp);
353 int client_obd_cleanup(struct obd_device *obddev, int flags)
355 struct client_obd *cli = &obddev->u.cli;
360 if (cli->cl_mgmtcli_obd) {
361 mgmtcli_deregister_for_events_t dereg_f;
363 dereg_f = (mgmtcli_deregister_for_events_t)symbol_get("mgmtcli_deregister_for_events");
364 dereg_f(cli->cl_mgmtcli_obd, obddev);
365 symbol_put("mgmtcli_deregister_for_events");
368 /* Here we try to drop the security structure after destroy import,
369 * to avoid issue of "sleep in spinlock".
371 class_import_get(cli->cl_import);
372 class_destroy_import(cli->cl_import);
373 ptlrpcs_import_drop_sec(cli->cl_import);
374 class_import_put(cli->cl_import);
375 cli->cl_import = NULL;
377 ldlm_put_ref(flags & OBD_OPT_FORCE);
382 int client_connect_import(struct lustre_handle *dlm_handle,
383 struct obd_device *obd,
384 struct obd_uuid *cluuid,
385 struct obd_connect_data *conn_data,
386 unsigned long connect_flags)
388 struct client_obd *cli = &obd->u.cli;
389 struct obd_import *imp = cli->cl_import;
390 struct obd_export *exp;
395 rc = class_connect(dlm_handle, obd, cluuid);
399 cli->cl_conn_count++;
400 if (cli->cl_conn_count > 1)
402 exp = class_conn2export(dlm_handle);
404 if (obd->obd_namespace != NULL)
405 CERROR("already have namespace!\n");
406 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
407 LDLM_NAMESPACE_CLIENT);
408 if (obd->obd_namespace == NULL)
409 GOTO(out_disco, rc = -ENOMEM);
411 rc = ptlrpcs_import_get_sec(imp);
415 imp->imp_dlm_handle = *dlm_handle;
416 rc = ptlrpc_init_import(imp);
420 imp->imp_connect_flags = connect_flags;
422 memcpy(&imp->imp_connect_data, conn_data, sizeof(*conn_data));
424 rc = ptlrpc_connect_import(imp, NULL);
426 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
429 LASSERT(exp->exp_connection);
430 ptlrpc_pinger_add_import(imp);
435 ldlm_namespace_free(obd->obd_namespace, 0);
436 obd->obd_namespace = NULL;
438 cli->cl_conn_count--;
439 class_disconnect(exp, 0);
441 class_export_put(exp);
448 int client_disconnect_export(struct obd_export *exp, unsigned long flags)
450 struct obd_device *obd = class_exp2obd(exp);
451 struct client_obd *cli = &obd->u.cli;
452 struct obd_import *imp = cli->cl_import;
457 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
458 exp, exp ? exp->exp_handle.h_cookie : -1);
463 if (!cli->cl_conn_count) {
464 CERROR("disconnecting disconnected device (%s)\n",
466 GOTO(out_sem, rc = -EINVAL);
469 cli->cl_conn_count--;
470 if (cli->cl_conn_count)
471 GOTO(out_no_disconnect, rc = 0);
473 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
474 * delete it regardless. (It's safe to delete an import that was
476 (void)ptlrpc_pinger_del_import(imp);
478 if (obd->obd_namespace != NULL) {
479 /* obd_no_recov == local only */
480 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
481 obd->obd_no_recov, NULL);
482 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
483 obd->obd_namespace = NULL;
487 * Yeah, obd_no_recov also (mainly) means "forced shutdown".
489 if (obd->obd_no_recov)
490 ptlrpc_invalidate_import(imp, 0);
492 rc = ptlrpc_disconnect_import(imp);
496 err = class_disconnect(exp, 0);
504 /* --------------------------------------------------------------------------
505 * from old lib/target.c
506 * -------------------------------------------------------------------------- */
508 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
509 struct obd_uuid *cluuid, int initial_conn)
511 if (exp->exp_connection && !initial_conn) {
512 struct lustre_handle *hdl;
513 hdl = &exp->exp_imp_reverse->imp_remote_handle;
514 /* Might be a re-connect after a partition. */
515 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
516 CERROR("%s reconnecting\n", cluuid->uuid);
517 conn->cookie = exp->exp_handle.h_cookie;
520 CERROR("%s reconnecting from %s, "
521 "handle mismatch (ours "LPX64", theirs "
522 LPX64")\n", cluuid->uuid,
523 exp->exp_connection->c_remote_uuid.uuid,
524 hdl->cookie, conn->cookie);
525 memset(conn, 0, sizeof *conn);
530 conn->cookie = exp->exp_handle.h_cookie;
531 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
533 CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
537 static inline int ptlrpc_peer_is_local(struct ptlrpc_peer *peer)
539 ptl_process_id_t myid;
541 PtlGetId(peer->peer_ni->pni_ni_h, &myid);
542 return (memcmp(&peer->peer_id, &myid, sizeof(myid)) == 0);
545 /* To check whether the p_flavor is in deny list or not
548 * EPERM found, refuse
551 static int check_deny_list(struct list_head *head, __u32 flavor)
553 deny_sec_t *p_deny_sec = NULL;
554 deny_sec_t *n_deny_sec = NULL;
556 list_for_each_entry_safe(p_deny_sec, n_deny_sec, head, list) {
557 if (p_deny_sec->flavor == flavor)
563 int target_check_deny_sec(struct obd_device *target, struct ptlrpc_request *req)
568 flavor = req->rq_req_secflvr;
570 if (!strcmp(target->obd_type->typ_name, OBD_MDS_DEVICENAME)) {
571 spin_lock(&target->u.mds.mds_denylist_lock);
572 rc = check_deny_list(&target->u.mds.mds_denylist, flavor);
573 spin_unlock(&target->u.mds.mds_denylist_lock);
574 } else if (!strcmp(target->obd_type->typ_name, "obdfilter")) {
575 spin_lock(&target->u.filter.fo_denylist_lock);
576 rc = check_deny_list(&target->u.filter.fo_denylist, flavor);
577 spin_unlock(&target->u.filter.fo_denylist_lock);
583 int target_handle_connect(struct ptlrpc_request *req)
585 unsigned long connect_flags = 0, *cfp;
586 struct obd_device *target;
587 struct obd_export *export = NULL;
588 struct obd_import *revimp;
589 struct lustre_handle conn;
590 struct obd_uuid tgtuuid;
591 struct obd_uuid cluuid;
592 struct obd_uuid remote_uuid;
594 struct obd_connect_data *conn_data;
595 int conn_data_size = sizeof(*conn_data);
599 int initial_conn = 0;
600 char peer_str[PTL_NALFMT_SIZE];
601 const int offset = 1;
604 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
606 LASSERT_REQSWAB (req, offset + 0);
607 str = lustre_msg_string(req->rq_reqmsg, offset + 0,
608 sizeof(tgtuuid) - 1);
610 CERROR("bad target UUID for connect\n");
611 GOTO(out, rc = -EINVAL);
614 obd_str2uuid (&tgtuuid, str);
615 target = class_uuid2obd(&tgtuuid);
617 target = class_name2obd(str);
619 if (!target || target->obd_stopping || !target->obd_set_up) {
620 CERROR("UUID '%s' is not available for connect from %s\n",
621 str, req->rq_peerstr);
622 GOTO(out, rc = -ENODEV);
625 /* check the secure deny list of mds/ost */
626 rc = target_check_deny_sec(target, req);
630 LASSERT_REQSWAB (req, offset + 1);
631 str = lustre_msg_string(req->rq_reqmsg, offset + 1, sizeof(cluuid) - 1);
633 CERROR("bad client UUID for connect\n");
634 GOTO(out, rc = -EINVAL);
637 obd_str2uuid (&cluuid, str);
639 /* XXX extract a nettype and format accordingly */
640 switch (sizeof(ptl_nid_t)) {
641 /* NB the casts only avoid compiler warnings */
643 snprintf((char *)remote_uuid.uuid, sizeof(remote_uuid),
644 "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_id.nid);
647 snprintf((char *)remote_uuid.uuid, sizeof(remote_uuid),
648 "NET_%x_UUID", (__u32)req->rq_peer.peer_id.nid);
654 tmp = lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn));
656 GOTO(out, rc = -EPROTO);
658 memcpy(&conn, tmp, sizeof conn);
660 cfp = lustre_msg_buf(req->rq_reqmsg, offset + 3, sizeof(unsigned long));
661 LASSERT(cfp != NULL);
662 connect_flags = *cfp;
664 conn_data = lustre_swab_reqbuf(req, offset + 4, sizeof(*conn_data),
665 lustre_swab_connect);
667 GOTO(out, rc = -EPROTO);
669 rc = lustre_pack_reply(req, 1, &conn_data_size, NULL);
673 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
676 /* lctl gets a backstage, all-access pass. */
677 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
678 goto dont_check_exports;
680 spin_lock(&target->obd_dev_lock);
681 list_for_each(p, &target->obd_exports) {
682 export = list_entry(p, struct obd_export, exp_obd_chain);
683 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
684 spin_unlock(&target->obd_dev_lock);
685 LASSERT(export->exp_obd == target);
687 rc = target_handle_reconnect(&conn, export, &cluuid,
693 /* If we found an export, we already unlocked. */
695 spin_unlock(&target->obd_dev_lock);
696 } else if (req->rq_export == NULL &&
697 atomic_read(&export->exp_rpc_count) > 0) {
698 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
699 target->obd_name, cluuid.uuid,
700 ptlrpc_peernid2str(&req->rq_peer, peer_str),
701 export, atomic_read(&export->exp_refcount));
702 GOTO(out, rc = -EBUSY);
703 } else if (req->rq_export != NULL &&
704 atomic_read(&export->exp_rpc_count) > 1) {
705 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
706 target->obd_name, cluuid.uuid,
707 ptlrpc_peernid2str(&req->rq_peer, peer_str),
708 export, atomic_read(&export->exp_rpc_count));
709 GOTO(out, rc = -EBUSY);
710 } else if (req->rq_reqmsg->conn_cnt == 1 && !initial_conn) {
711 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
713 GOTO(out, rc = -EALREADY);
716 /* Tell the client if we're in recovery. */
717 /* If this is the first client, start the recovery timer */
718 CWARN("%s: connection from %s@%s/%lu %st"LPU64"\n", target->obd_name,
719 cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str), *cfp,
720 target->obd_recovering ? "recovering/" : "", conn_data->transno);
722 if (target->obd_recovering) {
723 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
724 target_start_recovery_timer(target);
728 /* Tell the client if we support replayable requests */
729 if (target->obd_replayable)
730 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
733 if (export == NULL) {
734 if (target->obd_recovering) {
735 CERROR("%s denying connection for new client %s@%s: "
736 "%d clients in recovery for %lds\n", target->obd_name,
738 ptlrpc_peernid2str(&req->rq_peer, peer_str),
739 target->obd_recoverable_clients,
740 (target->obd_recovery_timer.expires-jiffies)/HZ);
744 rc = obd_connect(&conn, target, &cluuid, conn_data,
749 /* Return only the parts of obd_connect_data that we understand, so the
750 * client knows that we don't understand the rest. */
751 conn_data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
752 memcpy(lustre_msg_buf(req->rq_repmsg, 0, sizeof(*conn_data)), conn_data,
755 /* Tell the client if we support replayable requests */
756 if (target->obd_replayable)
757 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
759 /* If all else goes well, this is our RPC return code. */
762 if (rc && rc != EALREADY)
765 req->rq_repmsg->handle = conn;
767 /* If the client and the server are the same node, we will already
768 * have an export that really points to the client's DLM export,
769 * because we have a shared handles table.
771 * XXX this will go away when shaver stops sending the "connect" handle
772 * in the real "remote handle" field of the request --phik 24 Apr 2003
774 if (req->rq_export != NULL)
775 class_export_put(req->rq_export);
777 /* ownership of this export ref transfers to the request */
778 export = req->rq_export = class_conn2export(&conn);
779 LASSERT(export != NULL);
781 spin_lock_irqsave(&export->exp_lock, flags);
783 req->rq_repmsg->conn_cnt = export->exp_conn_cnt + 1;
784 } else if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
785 CERROR("%s@%s: already connected at a higher conn_cnt: %d > %d\n",
786 cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str),
787 export->exp_conn_cnt,
788 req->rq_reqmsg->conn_cnt);
789 spin_unlock_irqrestore(&export->exp_lock, flags);
790 GOTO(out, rc = -EALREADY);
792 export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
793 spin_unlock_irqrestore(&export->exp_lock, flags);
795 /* request from liblustre? */
796 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
797 export->exp_libclient = 1;
799 if (!(lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_ASYNC) &&
800 ptlrpc_peer_is_local(&req->rq_peer)) {
801 CWARN("%s: exp %p set sync\n", target->obd_name, export);
802 export->exp_sync = 1;
804 CDEBUG(D_HA, "%s: exp %p set async\n",target->obd_name,export);
805 export->exp_sync = 0;
808 if (export->exp_connection != NULL)
809 ptlrpc_put_connection(export->exp_connection);
810 export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
813 if (rc == EALREADY) {
814 /* We indicate the reconnection in a flag, not an error code. */
815 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
819 spin_lock_bh(&target->obd_processing_task_lock);
820 if (target->obd_recovering && export->exp_connected == 0) {
821 __u64 t = conn_data->transno;
822 export->exp_connected = 1;
823 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
824 && t < target->obd_next_recovery_transno)
825 target->obd_next_recovery_transno = t;
826 target->obd_connected_clients++;
827 if (target->obd_connected_clients == target->obd_max_recoverable_clients)
828 wake_up(&target->obd_next_transno_waitq);
830 spin_unlock_bh(&target->obd_processing_task_lock);
832 memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn)),
835 if (export->exp_imp_reverse != NULL) {
836 /* same logic as client_obd_cleanup */
837 class_import_get(export->exp_imp_reverse);
838 class_destroy_import(export->exp_imp_reverse);
839 ptlrpcs_import_drop_sec(export->exp_imp_reverse);
840 class_import_put(export->exp_imp_reverse);
843 /* for the rest part, we return -ENOTCONN in case of errors
844 * in order to let client initialize connection again.
846 revimp = export->exp_imp_reverse = class_new_import();
848 CERROR("fail to alloc new reverse import.\n");
849 GOTO(out, rc = -ENOTCONN);
852 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
853 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
854 revimp->imp_remote_handle = conn;
855 revimp->imp_obd = target;
856 revimp->imp_dlm_fake = 1;
857 revimp->imp_state = LUSTRE_IMP_FULL;
859 rc = ptlrpcs_import_get_sec(revimp);
861 CERROR("reverse import can not get sec: %d\n", rc);
862 class_destroy_import(revimp);
863 export->exp_imp_reverse = NULL;
864 GOTO(out, rc = -ENOTCONN);
867 class_import_put(revimp);
869 rc = obd_connect_post(export, initial_conn, connect_flags);
876 int target_handle_disconnect(struct ptlrpc_request *req)
878 struct obd_export *exp;
882 rc = lustre_pack_reply(req, 0, NULL, NULL);
886 /* keep the rq_export around so we can send the reply */
887 exp = class_export_get(req->rq_export);
888 req->rq_status = obd_disconnect(exp, 0);
892 void target_destroy_export(struct obd_export *exp)
894 /* exports created from last_rcvd data, and "fake"
895 exports created by lctl don't have an import */
896 if (exp->exp_imp_reverse != NULL) {
897 ptlrpcs_import_drop_sec(exp->exp_imp_reverse);
898 class_destroy_import(exp->exp_imp_reverse);
901 /* We cancel locks at disconnect time, but this will catch any locks
902 * granted in a race with recovery-induced disconnect. */
903 if (exp->exp_obd->obd_namespace != NULL)
904 ldlm_cancel_locks_for_export(exp);
911 struct ptlrpc_request *
912 ptlrpc_clone_req( struct ptlrpc_request *orig_req)
914 struct ptlrpc_request *copy_req;
915 struct lustre_msg *copy_reqmsg;
917 OBD_ALLOC(copy_req, sizeof *copy_req);
920 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
922 OBD_FREE(copy_req, sizeof *copy_req);
926 memcpy(copy_req, orig_req, sizeof *copy_req);
927 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
928 /* the copied req takes over the reply state and security data */
929 orig_req->rq_reply_state = NULL;
930 orig_req->rq_svcsec_data = NULL;
932 copy_req->rq_reqmsg = copy_reqmsg;
933 class_export_get(copy_req->rq_export);
934 INIT_LIST_HEAD(©_req->rq_list);
939 void ptlrpc_free_clone( struct ptlrpc_request *req)
942 svcsec_cleanup_req(req);
944 class_export_put(req->rq_export);
945 list_del(&req->rq_list);
946 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
947 OBD_FREE(req, sizeof *req);
950 static void target_release_saved_req(struct ptlrpc_request *req)
953 svcsec_cleanup_req(req);
955 class_export_put(req->rq_export);
956 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
957 OBD_FREE(req, sizeof *req);
960 static void target_finish_recovery(struct obd_device *obd)
964 ldlm_reprocess_all_ns(obd->obd_namespace);
966 /* when recovery finished, cleanup orphans on mds and ost */
967 if (OBT(obd) && OBP(obd, postrecov)) {
968 rc = OBP(obd, postrecov)(obd);
970 CWARN("%s: all clients recovered, %d MDS "
971 "orphans deleted\n", obd->obd_name, rc);
973 CERROR("postrecov failed %d\n", rc);
976 obd->obd_recovery_end = LTIME_S(CURRENT_TIME);
980 static void abort_req_replay_queue(struct obd_device *obd)
982 struct ptlrpc_request *req;
983 struct list_head *tmp, *n;
986 list_for_each_safe(tmp, n, &obd->obd_req_replay_queue) {
987 req = list_entry(tmp, struct ptlrpc_request, rq_list);
988 list_del(&req->rq_list);
989 DEBUG_REQ(D_ERROR, req, "aborted:");
990 req->rq_status = -ENOTCONN;
991 req->rq_type = PTL_RPC_MSG_ERR;
992 rc = lustre_pack_reply(req, 0, NULL, NULL);
996 DEBUG_REQ(D_ERROR, req,
997 "packing failed for abort-reply; skipping");
999 target_release_saved_req(req);
1003 static void abort_lock_replay_queue(struct obd_device *obd)
1005 struct ptlrpc_request *req;
1006 struct list_head *tmp, *n;
1009 list_for_each_safe(tmp, n, &obd->obd_lock_replay_queue) {
1010 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1011 list_del(&req->rq_list);
1012 DEBUG_REQ(D_ERROR, req, "aborted:");
1013 req->rq_status = -ENOTCONN;
1014 req->rq_type = PTL_RPC_MSG_ERR;
1015 rc = lustre_pack_reply(req, 0, NULL, NULL);
1019 DEBUG_REQ(D_ERROR, req,
1020 "packing failed for abort-reply; skipping");
1022 target_release_saved_req(req);
1026 /* Called from a cleanup function if the device is being cleaned up
1027 forcefully. The exports should all have been disconnected already,
1028 the only thing left to do is
1029 - clear the recovery flags
1031 - free queued requests and replies, but don't send replies
1032 Because the obd_stopping flag is set, no new requests should be received.
1035 void target_cleanup_recovery(struct obd_device *obd)
1037 struct list_head *tmp, *n;
1038 struct ptlrpc_request *req;
1040 spin_lock_bh(&obd->obd_processing_task_lock);
1041 if (!obd->obd_recovering) {
1042 spin_unlock_bh(&obd->obd_processing_task_lock);
1046 obd->obd_recovering = obd->obd_abort_recovery = 0;
1047 target_cancel_recovery_timer(obd);
1048 spin_unlock_bh(&obd->obd_processing_task_lock);
1050 list_for_each_safe(tmp, n, &obd->obd_req_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_lock_replay_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);
1062 list_for_each_safe(tmp, n, &obd->obd_final_req_queue) {
1063 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1064 list_del(&req->rq_list);
1065 LASSERT (req->rq_reply_state == 0);
1066 target_release_saved_req(req);
1071 static void target_abort_recovery(void *data)
1073 struct obd_device *obd = data;
1075 LASSERT(!obd->obd_recovering);
1077 class_disconnect_stale_exports(obd, 0);
1079 CERROR("%s: recovery period over; disconnecting unfinished clients.\n",
1082 abort_recovery_queue(obd);
1083 target_finish_recovery(obd);
1084 ptlrpc_run_recovery_over_upcall(obd);
1088 static void target_recovery_expired(unsigned long castmeharder)
1090 struct obd_device *obd = (struct obd_device *)castmeharder;
1091 spin_lock_bh(&obd->obd_processing_task_lock);
1092 if (obd->obd_recovering)
1093 obd->obd_abort_recovery = 1;
1095 wake_up(&obd->obd_next_transno_waitq);
1096 spin_unlock_bh(&obd->obd_processing_task_lock);
1100 /* obd_processing_task_lock should be held */
1101 void target_cancel_recovery_timer(struct obd_device *obd)
1103 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1104 del_timer(&obd->obd_recovery_timer);
1108 static void reset_recovery_timer(struct obd_device *obd)
1110 spin_lock_bh(&obd->obd_processing_task_lock);
1111 if (!obd->obd_recovering) {
1112 spin_unlock_bh(&obd->obd_processing_task_lock);
1115 CDEBUG(D_HA, "timer will expire in %u seconds\n",
1116 OBD_RECOVERY_TIMEOUT / HZ);
1117 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
1118 spin_unlock_bh(&obd->obd_processing_task_lock);
1122 /* Only start it the first time called */
1123 void target_start_recovery_timer(struct obd_device *obd)
1125 spin_lock_bh(&obd->obd_processing_task_lock);
1126 if (!obd->obd_recovering || timer_pending(&obd->obd_recovery_timer)) {
1127 spin_unlock_bh(&obd->obd_processing_task_lock);
1130 CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1131 OBD_RECOVERY_TIMEOUT / HZ);
1132 obd->obd_recovery_timer.function = target_recovery_expired;
1133 obd->obd_recovery_timer.data = (unsigned long)obd;
1134 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
1135 spin_unlock_bh(&obd->obd_processing_task_lock);
1139 static int check_for_next_transno(struct obd_device *obd)
1141 struct ptlrpc_request *req = NULL;
1142 int wake_up = 0, connected, completed, queue_len, max;
1143 __u64 next_transno, req_transno;
1145 spin_lock_bh(&obd->obd_processing_task_lock);
1146 if (!list_empty(&obd->obd_req_replay_queue)) {
1147 req = list_entry(obd->obd_req_replay_queue.next,
1148 struct ptlrpc_request, rq_list);
1149 req_transno = req->rq_reqmsg->transno;
1154 max = obd->obd_max_recoverable_clients;
1155 connected = obd->obd_connected_clients;
1156 completed = max - obd->obd_recoverable_clients;
1157 queue_len = obd->obd_requests_queued_for_recovery;
1158 next_transno = obd->obd_next_recovery_transno;
1160 CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
1161 "req_transno: "LPU64", next_transno: "LPU64"\n",
1162 max, connected, completed, queue_len, req_transno, next_transno);
1163 if (obd->obd_abort_recovery) {
1164 CDEBUG(D_HA, "waking for aborted recovery\n");
1166 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1167 CDEBUG(D_HA, "waking for completed recovery\n");
1169 } else if (req_transno == next_transno) {
1170 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1172 } else if (queue_len + completed == max) {
1173 LASSERT(req->rq_reqmsg->transno >= next_transno);
1174 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1175 "waking for skipped transno (skip: "LPD64
1176 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1177 next_transno, queue_len, completed, max, req_transno);
1178 obd->obd_next_recovery_transno = req_transno;
1180 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1181 /* some clients haven't connected in time, but we can try
1182 * to replay requests that demand on already committed ones
1183 * also, we can replay first non-committed transation */
1184 LASSERT(req_transno != 0);
1185 if (req_transno == obd->obd_last_committed + 1) {
1186 obd->obd_next_recovery_transno = req_transno;
1187 } else if (req_transno > obd->obd_last_committed) {
1188 /* can't continue recovery: have no needed transno */
1189 obd->obd_abort_recovery = 1;
1190 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1191 "connected: %d, completed: %d, queue_len: %d, "
1192 "req_transno: "LPU64", next_transno: "LPU64"\n",
1193 max, connected, completed, queue_len,
1194 req_transno, next_transno);
1198 spin_unlock_bh(&obd->obd_processing_task_lock);
1203 static struct ptlrpc_request *
1204 target_next_replay_req(struct obd_device *obd)
1206 struct l_wait_info lwi = { 0 };
1207 struct ptlrpc_request *req;
1209 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1210 obd->obd_next_recovery_transno);
1211 l_wait_event(obd->obd_next_transno_waitq,
1212 check_for_next_transno(obd), &lwi);
1214 spin_lock_bh(&obd->obd_processing_task_lock);
1215 if (obd->obd_abort_recovery) {
1217 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1218 req = list_entry(obd->obd_req_replay_queue.next,
1219 struct ptlrpc_request, rq_list);
1220 list_del_init(&req->rq_list);
1221 obd->obd_requests_queued_for_recovery--;
1225 spin_unlock_bh(&obd->obd_processing_task_lock);
1229 static int check_for_next_lock(struct obd_device *obd)
1231 struct ptlrpc_request *req = NULL;
1234 spin_lock_bh(&obd->obd_processing_task_lock);
1235 if (!list_empty(&obd->obd_lock_replay_queue)) {
1236 req = list_entry(obd->obd_lock_replay_queue.next,
1237 struct ptlrpc_request, rq_list);
1238 CDEBUG(D_HA, "waking for next lock\n");
1240 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1241 CDEBUG(D_HA, "waking for completed lock replay\n");
1243 } else if (obd->obd_abort_recovery) {
1244 CDEBUG(D_HA, "waking for aborted recovery\n");
1247 spin_unlock_bh(&obd->obd_processing_task_lock);
1252 static struct ptlrpc_request *
1253 target_next_replay_lock(struct obd_device *obd)
1255 struct l_wait_info lwi = { 0 };
1256 struct ptlrpc_request *req;
1258 CDEBUG(D_HA, "Waiting for lock\n");
1259 l_wait_event(obd->obd_next_transno_waitq,
1260 check_for_next_lock(obd), &lwi);
1262 spin_lock_bh(&obd->obd_processing_task_lock);
1263 if (obd->obd_abort_recovery) {
1265 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1266 req = list_entry(obd->obd_lock_replay_queue.next,
1267 struct ptlrpc_request, rq_list);
1268 list_del_init(&req->rq_list);
1272 spin_unlock_bh(&obd->obd_processing_task_lock);
1276 static struct ptlrpc_request *
1277 target_next_final_ping(struct obd_device *obd)
1279 struct ptlrpc_request *req;
1281 spin_lock_bh(&obd->obd_processing_task_lock);
1282 if (!list_empty(&obd->obd_final_req_queue)) {
1283 req = list_entry(obd->obd_final_req_queue.next,
1284 struct ptlrpc_request, rq_list);
1285 list_del_init(&req->rq_list);
1289 spin_unlock_bh(&obd->obd_processing_task_lock);
1293 static int req_replay_done(struct obd_export *exp)
1295 if (exp->exp_req_replay_needed)
1300 static int lock_replay_done(struct obd_export *exp)
1302 if (exp->exp_lock_replay_needed)
1307 static int connect_done(struct obd_export *exp)
1309 if (exp->exp_connected)
1314 static int check_for_clients(struct obd_device *obd)
1316 if (obd->obd_abort_recovery)
1318 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1319 if (obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1324 static int target_recovery_thread(void *arg)
1326 struct obd_device *obd = arg;
1327 struct ptlrpc_request *req;
1328 struct target_recovery_data *trd = &obd->obd_recovery_data;
1329 char peer_str[PTL_NALFMT_SIZE];
1330 struct l_wait_info lwi = { 0 };
1331 unsigned long delta;
1332 unsigned long flags;
1335 kportal_daemonize("tgt-recov");
1337 SIGNAL_MASK_LOCK(current, flags);
1338 sigfillset(¤t->blocked);
1340 SIGNAL_MASK_UNLOCK(current, flags);
1342 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1344 trd->trd_processing_task = current->pid;
1346 obd->obd_recovering = 1;
1347 complete(&trd->trd_starting);
1349 /* first of all, we have to know the first transno to replay */
1350 obd->obd_abort_recovery = 0;
1351 l_wait_event(obd->obd_next_transno_waitq,
1352 check_for_clients(obd), &lwi);
1354 spin_lock_bh(&obd->obd_processing_task_lock);
1355 target_cancel_recovery_timer(obd);
1356 spin_unlock_bh(&obd->obd_processing_task_lock);
1358 /* If some clients haven't connected in time, evict them */
1359 if (obd->obd_abort_recovery) {
1361 CDEBUG(D_ERROR, "few clients haven't connect in time (%d/%d),"
1362 "evict them ...\n", obd->obd_connected_clients,
1363 obd->obd_max_recoverable_clients);
1364 obd->obd_abort_recovery = 0;
1365 stale = class_disconnect_stale_exports(obd, connect_done, 0);
1366 atomic_sub(stale, &obd->obd_req_replay_clients);
1367 atomic_sub(stale, &obd->obd_lock_replay_clients);
1370 /* next stage: replay requests */
1372 CDEBUG(D_ERROR, "1: request replay stage - %d clients from t"LPU64"\n",
1373 atomic_read(&obd->obd_req_replay_clients),
1374 obd->obd_next_recovery_transno);
1375 while ((req = target_next_replay_req(obd))) {
1376 LASSERT(trd->trd_processing_task == current->pid);
1377 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s: ",
1378 req->rq_reqmsg->transno,
1379 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1380 (void)trd->trd_recovery_handler(req);
1381 obd->obd_replayed_requests++;
1382 reset_recovery_timer(obd);
1383 /* bug 1580: decide how to properly sync() in recovery*/
1384 //mds_fsync_super(mds->mds_sb);
1385 ptlrpc_free_clone(req);
1386 spin_lock_bh(&obd->obd_processing_task_lock);
1387 obd->obd_next_recovery_transno++;
1388 spin_unlock_bh(&obd->obd_processing_task_lock);
1391 spin_lock_bh(&obd->obd_processing_task_lock);
1392 target_cancel_recovery_timer(obd);
1393 spin_unlock_bh(&obd->obd_processing_task_lock);
1395 /* If some clients haven't replayed requests in time, evict them */
1396 if (obd->obd_abort_recovery) {
1398 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1399 obd->obd_abort_recovery = 0;
1400 stale = class_disconnect_stale_exports(obd, req_replay_done, 0);
1401 atomic_sub(stale, &obd->obd_lock_replay_clients);
1402 abort_req_replay_queue(obd);
1403 /* XXX for debuggin tests 11 and 17 */
1407 /* The second stage: replay locks */
1408 CDEBUG(D_ERROR, "2: lock replay stage - %d clients\n",
1409 atomic_read(&obd->obd_lock_replay_clients));
1410 while ((req = target_next_replay_lock(obd))) {
1411 LASSERT(trd->trd_processing_task == current->pid);
1412 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1413 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1414 (void)trd->trd_recovery_handler(req);
1415 reset_recovery_timer(obd);
1416 ptlrpc_free_clone(req);
1417 obd->obd_replayed_locks++;
1420 spin_lock_bh(&obd->obd_processing_task_lock);
1421 target_cancel_recovery_timer(obd);
1422 spin_unlock_bh(&obd->obd_processing_task_lock);
1424 /* If some clients haven't replayed requests in time, evict them */
1425 if (obd->obd_abort_recovery) {
1427 CERROR("lock replay timed out, aborting ...\n");
1428 obd->obd_abort_recovery = 0;
1429 stale = class_disconnect_stale_exports(obd, lock_replay_done, 0);
1430 abort_lock_replay_queue(obd);
1433 /* We drop recoverying flag to forward all new requests
1434 * to regular mds_handle() since now */
1435 spin_lock_bh(&obd->obd_processing_task_lock);
1436 obd->obd_recovering = 0;
1437 spin_unlock_bh(&obd->obd_processing_task_lock);
1439 /* The third stage: reply on final pings */
1440 CDEBUG(D_ERROR, "3: final stage - process recovery completion pings\n");
1441 while ((req = target_next_final_ping(obd))) {
1442 LASSERT(trd->trd_processing_task == current->pid);
1443 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1444 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1445 (void)trd->trd_recovery_handler(req);
1446 ptlrpc_free_clone(req);
1449 delta = (jiffies - delta) / HZ;
1450 CDEBUG(D_ERROR,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1451 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1452 if (delta > obd_timeout * 2) {
1453 CWARN("too long recovery - read logs\n");
1454 portals_debug_dumplog();
1456 target_finish_recovery(obd);
1458 trd->trd_processing_task = 0;
1459 complete(&trd->trd_finishing);
1463 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1466 struct target_recovery_data *trd = &obd->obd_recovery_data;
1468 memset(trd, 0, sizeof(*trd));
1469 init_completion(&trd->trd_starting);
1470 init_completion(&trd->trd_finishing);
1471 trd->trd_recovery_handler = handler;
1473 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1474 wait_for_completion(&trd->trd_starting);
1475 LASSERT(obd->obd_recovering != 0);
1482 void target_stop_recovery_thread(struct obd_device *obd)
1484 spin_lock_bh(&obd->obd_processing_task_lock);
1485 if (obd->obd_recovery_data.trd_processing_task > 0) {
1486 struct target_recovery_data *trd = &obd->obd_recovery_data;
1487 CERROR("%s: aborting recovery\n", obd->obd_name);
1488 obd->obd_abort_recovery = 1;
1489 wake_up(&obd->obd_next_transno_waitq);
1490 spin_unlock_bh(&obd->obd_processing_task_lock);
1491 wait_for_completion(&trd->trd_finishing);
1493 spin_unlock_bh(&obd->obd_processing_task_lock);
1498 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1500 struct obd_export *exp = req->rq_export;
1501 LASSERT(exp != NULL);
1502 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1503 /* client declares he's ready to replay locks */
1504 spin_lock_bh(&obd->obd_processing_task_lock);
1505 if (exp->exp_req_replay_needed) {
1506 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1507 exp->exp_req_replay_needed = 0;
1508 atomic_dec(&obd->obd_req_replay_clients);
1509 obd->obd_recoverable_clients--;
1510 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1511 CDEBUG(D_HA, "all clients have replayed reqs\n");
1512 wake_up(&obd->obd_next_transno_waitq);
1514 spin_unlock_bh(&obd->obd_processing_task_lock);
1516 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1517 /* client declares he's ready to complete recovery
1518 * so, we put the request on th final queue */
1519 spin_lock_bh(&obd->obd_processing_task_lock);
1520 if (exp->exp_lock_replay_needed) {
1521 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1522 exp->exp_lock_replay_needed = 0;
1523 atomic_dec(&obd->obd_lock_replay_clients);
1524 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1525 CDEBUG(D_HA, "all clients have replayed locks\n");
1526 wake_up(&obd->obd_next_transno_waitq);
1528 spin_unlock_bh(&obd->obd_processing_task_lock);
1534 int target_queue_recovery_request(struct ptlrpc_request *req,
1535 struct obd_device *obd)
1537 struct list_head *tmp;
1539 __u64 transno = req->rq_reqmsg->transno;
1541 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
1542 /* Processing the queue right now, don't re-add. */
1546 target_process_req_flags(obd, req);
1548 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1549 /* client declares he's ready to complete recovery
1550 * so, we put the request on th final queue */
1551 req = ptlrpc_clone_req(req);
1554 DEBUG_REQ(D_HA, req, "queue final req");
1555 spin_lock_bh(&obd->obd_processing_task_lock);
1556 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1557 spin_unlock_bh(&obd->obd_processing_task_lock);
1560 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1561 /* client declares he's ready to replay locks */
1562 req = ptlrpc_clone_req(req);
1565 DEBUG_REQ(D_HA, req, "queue lock replay req");
1566 spin_lock_bh(&obd->obd_processing_task_lock);
1567 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1568 spin_unlock_bh(&obd->obd_processing_task_lock);
1569 wake_up(&obd->obd_next_transno_waitq);
1574 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1575 * (i.e. buflens etc are in my own byte order), but type-dependent
1576 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1579 INIT_LIST_HEAD(&req->rq_list);
1580 DEBUG_REQ(D_HA, req, "not queueing");
1585 /* If we're processing the queue, we want don't want to queue this
1588 * Also, if this request has a transno less than the one we're waiting
1589 * for, we should process it now. It could (and currently always will)
1590 * be an open request for a descriptor that was opened some time ago.
1592 * Also, a resent, replayed request that has already been
1593 * handled will pass through here and be processed immediately.
1595 spin_lock_bh(&obd->obd_processing_task_lock);
1596 if (transno < obd->obd_next_recovery_transno && check_for_clients(obd)) {
1597 /* Processing the queue right now, don't re-add. */
1598 LASSERT(list_empty(&req->rq_list));
1599 spin_unlock_bh(&obd->obd_processing_task_lock);
1602 spin_unlock_bh(&obd->obd_processing_task_lock);
1604 /* A resent, replayed request that is still on the queue; just drop it.
1605 The queued request will handle this. */
1606 if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY))
1607 == (MSG_RESENT | MSG_REPLAY)) {
1608 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1612 req = ptlrpc_clone_req(req);
1616 spin_lock_bh(&obd->obd_processing_task_lock);
1619 list_for_each(tmp, &obd->obd_req_replay_queue) {
1620 struct ptlrpc_request *reqiter =
1621 list_entry(tmp, struct ptlrpc_request, rq_list);
1623 if (reqiter->rq_reqmsg->transno > transno) {
1624 list_add_tail(&req->rq_list, &reqiter->rq_list);
1631 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1633 obd->obd_requests_queued_for_recovery++;
1634 wake_up(&obd->obd_next_transno_waitq);
1635 spin_unlock_bh(&obd->obd_processing_task_lock);
1639 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1641 return req->rq_export->exp_obd;
1645 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1647 if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1648 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1649 DEBUG_REQ(D_ERROR, req, "dropping reply");
1650 /* NB this does _not_ send with ACK disabled, to simulate
1651 * sending OK, but timing out for the ACK */
1652 if (req->rq_reply_state != NULL) {
1653 if (!req->rq_reply_state->rs_difficult) {
1654 lustre_free_reply_state (req->rq_reply_state);
1655 req->rq_reply_state = NULL;
1657 struct ptlrpc_service *svc =
1658 req->rq_rqbd->rqbd_srv_ni->sni_service;
1659 atomic_inc(&svc->srv_outstanding_replies);
1666 req->rq_status = rc;
1667 return (ptlrpc_error(req));
1669 DEBUG_REQ(D_NET, req, "sending reply");
1672 return (ptlrpc_send_reply(req, 1));
1676 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1679 unsigned long flags;
1680 struct ptlrpc_reply_state *rs;
1681 struct obd_device *obd;
1682 struct obd_export *exp;
1683 struct ptlrpc_srv_ni *sni;
1684 struct ptlrpc_service *svc;
1686 sni = req->rq_rqbd->rqbd_srv_ni;
1687 svc = sni->sni_service;
1689 rs = req->rq_reply_state;
1690 if (rs == NULL || !rs->rs_difficult) {
1691 /* The easy case; no notifiers and reply_out_callback()
1692 * cleans up (i.e. we can't look inside rs after a
1693 * successful send) */
1694 netrc = target_send_reply_msg (req, rc, fail_id);
1696 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1700 /* must be an export if locks saved */
1701 LASSERT (req->rq_export != NULL);
1702 /* req/reply consistent */
1703 LASSERT (rs->rs_srv_ni == sni);
1706 LASSERT (!rs->rs_scheduled);
1707 LASSERT (!rs->rs_scheduled_ever);
1708 LASSERT (!rs->rs_handled);
1709 LASSERT (!rs->rs_on_net);
1710 LASSERT (rs->rs_export == NULL);
1711 LASSERT (list_empty(&rs->rs_obd_list));
1712 LASSERT (list_empty(&rs->rs_exp_list));
1714 exp = class_export_get (req->rq_export);
1717 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1718 rs->rs_scheduled = 1;
1720 rs->rs_xid = req->rq_xid;
1721 rs->rs_transno = req->rq_transno;
1722 rs->rs_export = exp;
1724 spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1726 if (rs->rs_transno > obd->obd_last_committed) {
1727 /* not committed already */
1728 list_add_tail (&rs->rs_obd_list,
1729 &obd->obd_uncommitted_replies);
1732 spin_unlock (&obd->obd_uncommitted_replies_lock);
1733 spin_lock (&exp->exp_lock);
1735 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1737 spin_unlock_irqrestore (&exp->exp_lock, flags);
1739 netrc = target_send_reply_msg (req, rc, fail_id);
1741 spin_lock_irqsave (&svc->srv_lock, flags);
1743 svc->srv_n_difficult_replies++;
1745 if (netrc != 0) /* error sending: reply is off the net */
1748 if (!rs->rs_on_net || /* some notifier */
1749 list_empty(&rs->rs_exp_list) || /* completed already */
1750 list_empty(&rs->rs_obd_list)) {
1751 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1752 wake_up (&svc->srv_waitq);
1754 list_add (&rs->rs_list, &sni->sni_active_replies);
1755 rs->rs_scheduled = 0; /* allow notifier to schedule */
1758 spin_unlock_irqrestore (&svc->srv_lock, flags);
1761 int target_handle_ping(struct ptlrpc_request *req)
1763 return lustre_pack_reply(req, 0, NULL, NULL);