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 memset(&cli->cl_last_write_time, 0,
258 sizeof(cli->cl_last_write_time));
259 cli->cl_write_gap_sum = 0;
260 cli->cl_write_gaps = 0;
262 if (num_physpages >> (20 - PAGE_SHIFT) <= 128) { /* <= 128 MB */
263 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 4;
264 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 4;
266 } else if (num_physpages >> (20 - PAGE_SHIFT) <= 512) { /* <= 512 MB */
267 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 2;
268 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 2;
271 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
272 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
277 CERROR("ldlm_get_ref failed: %d\n", rc);
281 ptlrpc_init_client(rq_portal, rp_portal, name,
282 &obddev->obd_ldlm_client);
284 imp = class_new_import();
286 GOTO(err_ldlm, rc = -ENOENT);
287 imp->imp_client = &obddev->obd_ldlm_client;
288 imp->imp_obd = obddev;
289 imp->imp_connect_op = connect_op;
290 imp->imp_generation = 0;
291 imp->imp_initial_recov = 1;
292 INIT_LIST_HEAD(&imp->imp_pinger_chain);
293 memcpy(imp->imp_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
294 LUSTRE_CFG_BUFLEN(lcfg, 1));
295 class_import_put(imp);
297 rc = client_import_add_conn(imp, &server_uuid, 1);
299 CERROR("can't add initial connection\n");
300 GOTO(err_import, rc);
303 cli->cl_import = imp;
304 cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
305 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
306 cli->cl_sandev = to_kdev_t(0);
308 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
309 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
310 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
311 name, obddev->obd_name,
312 imp->imp_target_uuid.uuid);
313 imp->imp_invalid = 1;
315 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0)
316 mgmt_name = lustre_cfg_string(lcfg, 4);
318 mgmt_name = lustre_cfg_string(lcfg, 3);
322 if (mgmt_name != NULL) {
323 /* Register with management client if we need to. */
324 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
325 obddev->obd_name, mgmt_name, server_uuid.uuid);
327 mgmt_obd = class_name2obd(mgmt_name);
329 CERROR("can't find mgmtcli %s to register\n",
331 GOTO(err_import, rc = -ENOSYS);
334 register_f = (mgmtcli_register_for_events_t)symbol_get("mgmtcli_register_for_events");
336 CERROR("can't i_m_g mgmtcli_register_for_events\n");
337 GOTO(err_import, rc = -ENOSYS);
340 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
341 symbol_put("mgmtcli_register_for_events");
344 cli->cl_mgmtcli_obd = mgmt_obd;
350 class_destroy_import(imp);
358 int client_obd_cleanup(struct obd_device *obddev, int flags)
360 struct client_obd *cli = &obddev->u.cli;
365 if (cli->cl_mgmtcli_obd) {
366 mgmtcli_deregister_for_events_t dereg_f;
368 dereg_f = (mgmtcli_deregister_for_events_t)symbol_get("mgmtcli_deregister_for_events");
369 dereg_f(cli->cl_mgmtcli_obd, obddev);
370 symbol_put("mgmtcli_deregister_for_events");
373 /* Here we try to drop the security structure after destroy import,
374 * to avoid issue of "sleep in spinlock".
376 class_import_get(cli->cl_import);
377 class_destroy_import(cli->cl_import);
378 ptlrpcs_import_drop_sec(cli->cl_import);
379 class_import_put(cli->cl_import);
380 cli->cl_import = NULL;
382 if (cli->cl_write_gap_sum) {
383 CWARN("%s: %lu write gaps: %lu av. (usec), %lu total "
384 "(usec)\n", obddev->obd_name, cli->cl_write_gaps,
385 cli->cl_write_gap_sum / cli->cl_write_gaps,
386 cli->cl_write_gap_sum);
389 ldlm_put_ref(flags & OBD_OPT_FORCE);
393 int client_connect_import(struct lustre_handle *dlm_handle,
394 struct obd_device *obd,
395 struct obd_uuid *cluuid,
396 struct obd_connect_data *conn_data,
397 unsigned long connect_flags)
399 struct client_obd *cli = &obd->u.cli;
400 struct obd_import *imp = cli->cl_import;
401 struct obd_export *exp;
406 rc = class_connect(dlm_handle, obd, cluuid);
410 cli->cl_conn_count++;
411 if (cli->cl_conn_count > 1)
413 exp = class_conn2export(dlm_handle);
415 if (obd->obd_namespace != NULL)
416 CERROR("already have namespace!\n");
417 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
418 LDLM_NAMESPACE_CLIENT);
419 if (obd->obd_namespace == NULL)
420 GOTO(out_disco, rc = -ENOMEM);
422 rc = ptlrpcs_import_get_sec(imp);
426 imp->imp_dlm_handle = *dlm_handle;
427 rc = ptlrpc_init_import(imp);
431 imp->imp_connect_flags = connect_flags;
433 memcpy(&imp->imp_connect_data, conn_data, sizeof(*conn_data));
435 rc = ptlrpc_connect_import(imp, NULL);
437 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
440 LASSERT(exp->exp_connection);
441 ptlrpc_pinger_add_import(imp);
446 ldlm_namespace_free(obd->obd_namespace, 0);
447 obd->obd_namespace = NULL;
449 cli->cl_conn_count--;
450 class_disconnect(exp, 0);
452 class_export_put(exp);
459 int client_disconnect_export(struct obd_export *exp, unsigned long flags)
461 struct obd_device *obd = class_exp2obd(exp);
462 struct client_obd *cli = &obd->u.cli;
463 struct obd_import *imp = cli->cl_import;
468 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
469 exp, exp ? exp->exp_handle.h_cookie : -1);
474 if (!cli->cl_conn_count) {
475 CERROR("disconnecting disconnected device (%s)\n",
477 GOTO(out_sem, rc = -EINVAL);
480 cli->cl_conn_count--;
481 if (cli->cl_conn_count)
482 GOTO(out_no_disconnect, rc = 0);
484 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
485 * delete it regardless. (It's safe to delete an import that was
487 (void)ptlrpc_pinger_del_import(imp);
489 if (obd->obd_namespace != NULL) {
490 /* obd_no_recov == local only */
491 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
492 obd->obd_no_recov, NULL);
493 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
494 obd->obd_namespace = NULL;
498 * Yeah, obd_no_recov also (mainly) means "forced shutdown".
500 if (obd->obd_no_recov)
501 ptlrpc_invalidate_import(imp, 0);
503 rc = ptlrpc_disconnect_import(imp);
507 err = class_disconnect(exp, 0);
515 /* --------------------------------------------------------------------------
516 * from old lib/target.c
517 * -------------------------------------------------------------------------- */
519 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
520 struct obd_uuid *cluuid, int initial_conn)
522 if (exp->exp_connection && !initial_conn) {
523 struct lustre_handle *hdl;
524 hdl = &exp->exp_imp_reverse->imp_remote_handle;
525 /* Might be a re-connect after a partition. */
526 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
527 CERROR("%s reconnecting\n", cluuid->uuid);
528 conn->cookie = exp->exp_handle.h_cookie;
531 CERROR("%s reconnecting from %s, "
532 "handle mismatch (ours "LPX64", theirs "
533 LPX64")\n", cluuid->uuid,
534 exp->exp_connection->c_remote_uuid.uuid,
535 hdl->cookie, conn->cookie);
536 memset(conn, 0, sizeof *conn);
541 conn->cookie = exp->exp_handle.h_cookie;
542 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
544 CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
548 static inline int ptlrpc_peer_is_local(struct ptlrpc_peer *peer)
550 ptl_process_id_t myid;
552 PtlGetId(peer->peer_ni->pni_ni_h, &myid);
553 return (memcmp(&peer->peer_id, &myid, sizeof(myid)) == 0);
556 /* To check whether the p_flavor is in deny list or not
559 * EPERM found, refuse
562 static int check_deny_list(struct list_head *head, __u32 flavor)
564 deny_sec_t *p_deny_sec = NULL;
565 deny_sec_t *n_deny_sec = NULL;
567 list_for_each_entry_safe(p_deny_sec, n_deny_sec, head, list) {
568 if (p_deny_sec->flavor == flavor)
574 int target_check_deny_sec(struct obd_device *target, struct ptlrpc_request *req)
579 flavor = req->rq_req_secflvr;
581 if (!strcmp(target->obd_type->typ_name, OBD_MDS_DEVICENAME)) {
582 spin_lock(&target->u.mds.mds_denylist_lock);
583 rc = check_deny_list(&target->u.mds.mds_denylist, flavor);
584 spin_unlock(&target->u.mds.mds_denylist_lock);
585 } else if (!strcmp(target->obd_type->typ_name, OBD_FILTER_DEVICENAME)) {
586 spin_lock(&target->u.filter.fo_denylist_lock);
587 rc = check_deny_list(&target->u.filter.fo_denylist, flavor);
588 spin_unlock(&target->u.filter.fo_denylist_lock);
594 int target_handle_connect(struct ptlrpc_request *req)
596 unsigned long connect_flags = 0, *cfp;
597 struct obd_device *target;
598 struct obd_export *export = NULL;
599 struct obd_import *revimp;
600 struct lustre_handle conn;
601 struct obd_uuid tgtuuid;
602 struct obd_uuid cluuid;
603 struct obd_uuid remote_uuid;
605 struct obd_connect_data *conn_data;
606 int conn_data_size = sizeof(*conn_data);
610 int initial_conn = 0;
611 char peer_str[PTL_NALFMT_SIZE];
612 const int offset = 1;
615 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
617 LASSERT_REQSWAB (req, offset + 0);
618 str = lustre_msg_string(req->rq_reqmsg, offset + 0,
619 sizeof(tgtuuid) - 1);
621 CERROR("bad target UUID for connect\n");
622 GOTO(out, rc = -EINVAL);
625 obd_str2uuid (&tgtuuid, str);
626 target = class_uuid2obd(&tgtuuid);
628 target = class_name2obd(str);
630 if (!target || target->obd_stopping || !target->obd_set_up) {
631 CERROR("UUID '%s' is not available for connect from %s\n",
632 str, req->rq_peerstr);
633 GOTO(out, rc = -ENODEV);
636 /* check the secure deny list of mds/ost */
637 rc = target_check_deny_sec(target, req);
641 LASSERT_REQSWAB (req, offset + 1);
642 str = lustre_msg_string(req->rq_reqmsg, offset + 1, sizeof(cluuid) - 1);
644 CERROR("bad client UUID for connect\n");
645 GOTO(out, rc = -EINVAL);
648 obd_str2uuid (&cluuid, str);
650 /* XXX extract a nettype and format accordingly */
651 switch (sizeof(ptl_nid_t)) {
652 /* NB the casts only avoid compiler warnings */
654 snprintf((char *)remote_uuid.uuid, sizeof(remote_uuid),
655 "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_id.nid);
658 snprintf((char *)remote_uuid.uuid, sizeof(remote_uuid),
659 "NET_%x_UUID", (__u32)req->rq_peer.peer_id.nid);
665 tmp = lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn));
667 GOTO(out, rc = -EPROTO);
669 memcpy(&conn, tmp, sizeof conn);
671 cfp = lustre_msg_buf(req->rq_reqmsg, offset + 3, sizeof(unsigned long));
672 LASSERT(cfp != NULL);
673 connect_flags = *cfp;
675 conn_data = lustre_swab_reqbuf(req, offset + 4, sizeof(*conn_data),
676 lustre_swab_connect);
678 GOTO(out, rc = -EPROTO);
680 rc = lustre_pack_reply(req, 1, &conn_data_size, NULL);
684 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
687 /* lctl gets a backstage, all-access pass. */
688 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
689 goto dont_check_exports;
691 spin_lock(&target->obd_dev_lock);
692 list_for_each(p, &target->obd_exports) {
693 export = list_entry(p, struct obd_export, exp_obd_chain);
694 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
695 spin_unlock(&target->obd_dev_lock);
696 LASSERT(export->exp_obd == target);
698 rc = target_handle_reconnect(&conn, export, &cluuid,
704 /* If we found an export, we already unlocked. */
706 spin_unlock(&target->obd_dev_lock);
707 } else if (req->rq_export == NULL &&
708 atomic_read(&export->exp_rpc_count) > 0) {
709 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
710 target->obd_name, cluuid.uuid,
711 ptlrpc_peernid2str(&req->rq_peer, peer_str),
712 export, atomic_read(&export->exp_refcount));
713 GOTO(out, rc = -EBUSY);
714 } else if (req->rq_export != NULL &&
715 atomic_read(&export->exp_rpc_count) > 1) {
716 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
717 target->obd_name, cluuid.uuid,
718 ptlrpc_peernid2str(&req->rq_peer, peer_str),
719 export, atomic_read(&export->exp_rpc_count));
720 GOTO(out, rc = -EBUSY);
721 } else if (req->rq_reqmsg->conn_cnt == 1 && !initial_conn) {
722 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
724 GOTO(out, rc = -EALREADY);
727 /* Tell the client if we're in recovery. */
728 /* If this is the first client, start the recovery timer */
729 CWARN("%s: connection from %s@%s/%lu %st"LPU64"\n", target->obd_name,
730 cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str), *cfp,
731 target->obd_recovering ? "recovering/" : "", conn_data->transno);
733 if (target->obd_recovering) {
734 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
735 target_start_recovery_timer(target);
739 /* Tell the client if we support replayable requests */
740 if (target->obd_replayable)
741 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
744 if (export == NULL) {
745 if (target->obd_recovering) {
746 CERROR("%s denying connection for new client %s@%s: "
747 "%d clients in recovery for %lds\n", target->obd_name,
749 ptlrpc_peernid2str(&req->rq_peer, peer_str),
750 target->obd_recoverable_clients,
751 (target->obd_recovery_timer.expires-jiffies)/HZ);
755 rc = obd_connect(&conn, target, &cluuid, conn_data,
760 /* Return only the parts of obd_connect_data that we understand, so the
761 * client knows that we don't understand the rest. */
762 conn_data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
763 memcpy(lustre_msg_buf(req->rq_repmsg, 0, sizeof(*conn_data)), conn_data,
766 /* Tell the client if we support replayable requests */
767 if (target->obd_replayable)
768 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
770 /* If all else goes well, this is our RPC return code. */
773 if (rc && rc != EALREADY)
776 req->rq_repmsg->handle = conn;
778 /* If the client and the server are the same node, we will already
779 * have an export that really points to the client's DLM export,
780 * because we have a shared handles table.
782 * XXX this will go away when shaver stops sending the "connect" handle
783 * in the real "remote handle" field of the request --phik 24 Apr 2003
785 if (req->rq_export != NULL)
786 class_export_put(req->rq_export);
788 /* ownership of this export ref transfers to the request */
789 export = req->rq_export = class_conn2export(&conn);
790 LASSERT(export != NULL);
792 spin_lock_irqsave(&export->exp_lock, flags);
794 req->rq_repmsg->conn_cnt = export->exp_conn_cnt + 1;
795 } else if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
796 CERROR("%s@%s: already connected at a higher conn_cnt: %d > %d\n",
797 cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str),
798 export->exp_conn_cnt,
799 req->rq_reqmsg->conn_cnt);
800 spin_unlock_irqrestore(&export->exp_lock, flags);
801 GOTO(out, rc = -EALREADY);
803 export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
804 spin_unlock_irqrestore(&export->exp_lock, flags);
806 /* request from liblustre? */
807 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
808 export->exp_libclient = 1;
810 if (!(lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_ASYNC) &&
811 ptlrpc_peer_is_local(&req->rq_peer)) {
812 CWARN("%s: exp %p set sync\n", target->obd_name, export);
813 export->exp_sync = 1;
815 CDEBUG(D_HA, "%s: exp %p set async\n",target->obd_name,export);
816 export->exp_sync = 0;
819 if (export->exp_connection != NULL)
820 ptlrpc_put_connection(export->exp_connection);
821 export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
824 if (rc == EALREADY) {
825 /* We indicate the reconnection in a flag, not an error code. */
826 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
830 spin_lock_bh(&target->obd_processing_task_lock);
831 if (target->obd_recovering && export->exp_connected == 0) {
832 __u64 t = conn_data->transno;
833 export->exp_connected = 1;
834 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
835 && t < target->obd_next_recovery_transno)
836 target->obd_next_recovery_transno = t;
837 target->obd_connected_clients++;
838 if (target->obd_connected_clients == target->obd_max_recoverable_clients)
839 wake_up(&target->obd_next_transno_waitq);
841 spin_unlock_bh(&target->obd_processing_task_lock);
843 memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn)),
846 if (export->exp_imp_reverse != NULL) {
847 /* same logic as client_obd_cleanup */
848 class_import_get(export->exp_imp_reverse);
849 class_destroy_import(export->exp_imp_reverse);
850 ptlrpcs_import_drop_sec(export->exp_imp_reverse);
851 class_import_put(export->exp_imp_reverse);
854 /* for the rest part, we return -ENOTCONN in case of errors
855 * in order to let client initialize connection again.
857 revimp = export->exp_imp_reverse = class_new_import();
859 CERROR("fail to alloc new reverse import.\n");
860 GOTO(out, rc = -ENOTCONN);
863 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
864 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
865 revimp->imp_remote_handle = conn;
866 revimp->imp_obd = target;
867 revimp->imp_dlm_fake = 1;
868 revimp->imp_state = LUSTRE_IMP_FULL;
870 rc = ptlrpcs_import_get_sec(revimp);
872 CERROR("reverse import can not get sec: %d\n", rc);
873 class_destroy_import(revimp);
874 export->exp_imp_reverse = NULL;
875 GOTO(out, rc = -ENOTCONN);
878 class_import_put(revimp);
880 rc = obd_connect_post(export, initial_conn, connect_flags);
887 int target_handle_disconnect(struct ptlrpc_request *req)
889 struct obd_export *exp;
893 rc = lustre_pack_reply(req, 0, NULL, NULL);
897 /* keep the rq_export around so we can send the reply */
898 exp = class_export_get(req->rq_export);
899 req->rq_status = obd_disconnect(exp, 0);
903 void target_destroy_export(struct obd_export *exp)
905 /* exports created from last_rcvd data, and "fake"
906 exports created by lctl don't have an import */
907 if (exp->exp_imp_reverse != NULL) {
908 ptlrpcs_import_drop_sec(exp->exp_imp_reverse);
909 class_destroy_import(exp->exp_imp_reverse);
912 /* We cancel locks at disconnect time, but this will catch any locks
913 * granted in a race with recovery-induced disconnect. */
914 if (exp->exp_obd->obd_namespace != NULL)
915 ldlm_cancel_locks_for_export(exp);
922 struct ptlrpc_request *
923 ptlrpc_clone_req( struct ptlrpc_request *orig_req)
925 struct ptlrpc_request *copy_req;
926 struct lustre_msg *copy_reqmsg;
928 OBD_ALLOC(copy_req, sizeof *copy_req);
931 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
933 OBD_FREE(copy_req, sizeof *copy_req);
937 memcpy(copy_req, orig_req, sizeof *copy_req);
938 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
939 /* the copied req takes over the reply state and security data */
940 orig_req->rq_reply_state = NULL;
941 orig_req->rq_svcsec_data = NULL;
943 copy_req->rq_reqmsg = copy_reqmsg;
944 class_export_get(copy_req->rq_export);
945 INIT_LIST_HEAD(©_req->rq_list);
950 void ptlrpc_free_clone( struct ptlrpc_request *req)
953 svcsec_cleanup_req(req);
955 class_export_put(req->rq_export);
956 list_del(&req->rq_list);
957 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
958 OBD_FREE(req, sizeof *req);
961 static void target_release_saved_req(struct ptlrpc_request *req)
964 svcsec_cleanup_req(req);
966 class_export_put(req->rq_export);
967 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
968 OBD_FREE(req, sizeof *req);
971 static void target_finish_recovery(struct obd_device *obd)
975 ldlm_reprocess_all_ns(obd->obd_namespace);
977 /* when recovery finished, cleanup orphans on mds and ost */
978 if (OBT(obd) && OBP(obd, postrecov)) {
979 rc = OBP(obd, postrecov)(obd);
981 CWARN("%s: all clients recovered, %d MDS "
982 "orphans deleted\n", obd->obd_name, rc);
984 CERROR("postrecov failed %d\n", rc);
987 obd->obd_recovery_end = LTIME_S(CURRENT_TIME);
991 static void abort_req_replay_queue(struct obd_device *obd)
993 struct ptlrpc_request *req;
994 struct list_head *tmp, *n;
997 list_for_each_safe(tmp, n, &obd->obd_req_replay_queue) {
998 req = list_entry(tmp, struct ptlrpc_request, rq_list);
999 list_del(&req->rq_list);
1000 DEBUG_REQ(D_ERROR, req, "aborted:");
1001 req->rq_status = -ENOTCONN;
1002 req->rq_type = PTL_RPC_MSG_ERR;
1003 rc = lustre_pack_reply(req, 0, NULL, NULL);
1007 DEBUG_REQ(D_ERROR, req,
1008 "packing failed for abort-reply; skipping");
1010 target_release_saved_req(req);
1014 static void abort_lock_replay_queue(struct obd_device *obd)
1016 struct ptlrpc_request *req;
1017 struct list_head *tmp, *n;
1020 list_for_each_safe(tmp, n, &obd->obd_lock_replay_queue) {
1021 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1022 list_del(&req->rq_list);
1023 DEBUG_REQ(D_ERROR, req, "aborted:");
1024 req->rq_status = -ENOTCONN;
1025 req->rq_type = PTL_RPC_MSG_ERR;
1026 rc = lustre_pack_reply(req, 0, NULL, NULL);
1030 DEBUG_REQ(D_ERROR, req,
1031 "packing failed for abort-reply; skipping");
1033 target_release_saved_req(req);
1037 /* Called from a cleanup function if the device is being cleaned up
1038 forcefully. The exports should all have been disconnected already,
1039 the only thing left to do is
1040 - clear the recovery flags
1042 - free queued requests and replies, but don't send replies
1043 Because the obd_stopping flag is set, no new requests should be received.
1046 void target_cleanup_recovery(struct obd_device *obd)
1048 struct list_head *tmp, *n;
1049 struct ptlrpc_request *req;
1051 spin_lock_bh(&obd->obd_processing_task_lock);
1052 if (!obd->obd_recovering) {
1053 spin_unlock_bh(&obd->obd_processing_task_lock);
1057 obd->obd_recovering = obd->obd_abort_recovery = 0;
1058 target_cancel_recovery_timer(obd);
1059 spin_unlock_bh(&obd->obd_processing_task_lock);
1061 list_for_each_safe(tmp, n, &obd->obd_req_replay_queue) {
1062 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1063 list_del(&req->rq_list);
1064 LASSERT (req->rq_reply_state == 0);
1065 target_release_saved_req(req);
1067 list_for_each_safe(tmp, n, &obd->obd_lock_replay_queue) {
1068 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1069 list_del(&req->rq_list);
1070 LASSERT (req->rq_reply_state == 0);
1071 target_release_saved_req(req);
1073 list_for_each_safe(tmp, n, &obd->obd_final_req_queue) {
1074 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1075 list_del(&req->rq_list);
1076 LASSERT (req->rq_reply_state == 0);
1077 target_release_saved_req(req);
1082 static void target_abort_recovery(void *data)
1084 struct obd_device *obd = data;
1086 LASSERT(!obd->obd_recovering);
1088 class_disconnect_stale_exports(obd, 0);
1090 CERROR("%s: recovery period over; disconnecting unfinished clients.\n",
1093 abort_recovery_queue(obd);
1094 target_finish_recovery(obd);
1095 ptlrpc_run_recovery_over_upcall(obd);
1099 static void target_recovery_expired(unsigned long castmeharder)
1101 struct obd_device *obd = (struct obd_device *)castmeharder;
1102 spin_lock_bh(&obd->obd_processing_task_lock);
1103 if (obd->obd_recovering)
1104 obd->obd_abort_recovery = 1;
1106 wake_up(&obd->obd_next_transno_waitq);
1107 spin_unlock_bh(&obd->obd_processing_task_lock);
1111 /* obd_processing_task_lock should be held */
1112 void target_cancel_recovery_timer(struct obd_device *obd)
1114 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1115 del_timer(&obd->obd_recovery_timer);
1119 static void reset_recovery_timer(struct obd_device *obd)
1121 spin_lock_bh(&obd->obd_processing_task_lock);
1122 if (!obd->obd_recovering) {
1123 spin_unlock_bh(&obd->obd_processing_task_lock);
1126 CDEBUG(D_HA, "timer will expire in %u seconds\n",
1127 OBD_RECOVERY_TIMEOUT / HZ);
1128 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
1129 spin_unlock_bh(&obd->obd_processing_task_lock);
1133 /* Only start it the first time called */
1134 void target_start_recovery_timer(struct obd_device *obd)
1136 spin_lock_bh(&obd->obd_processing_task_lock);
1137 if (!obd->obd_recovering || timer_pending(&obd->obd_recovery_timer)) {
1138 spin_unlock_bh(&obd->obd_processing_task_lock);
1141 CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1142 OBD_RECOVERY_TIMEOUT / HZ);
1143 obd->obd_recovery_timer.function = target_recovery_expired;
1144 obd->obd_recovery_timer.data = (unsigned long)obd;
1145 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
1146 spin_unlock_bh(&obd->obd_processing_task_lock);
1150 static int check_for_next_transno(struct obd_device *obd)
1152 struct ptlrpc_request *req = NULL;
1153 int wake_up = 0, connected, completed, queue_len, max;
1154 __u64 next_transno, req_transno;
1156 spin_lock_bh(&obd->obd_processing_task_lock);
1157 if (!list_empty(&obd->obd_req_replay_queue)) {
1158 req = list_entry(obd->obd_req_replay_queue.next,
1159 struct ptlrpc_request, rq_list);
1160 req_transno = req->rq_reqmsg->transno;
1165 max = obd->obd_max_recoverable_clients;
1166 connected = obd->obd_connected_clients;
1167 completed = max - obd->obd_recoverable_clients;
1168 queue_len = obd->obd_requests_queued_for_recovery;
1169 next_transno = obd->obd_next_recovery_transno;
1171 CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
1172 "req_transno: "LPU64", next_transno: "LPU64"\n",
1173 max, connected, completed, queue_len, req_transno, next_transno);
1174 if (obd->obd_abort_recovery) {
1175 CDEBUG(D_HA, "waking for aborted recovery\n");
1177 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1178 CDEBUG(D_HA, "waking for completed recovery\n");
1180 } else if (req_transno == next_transno) {
1181 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1183 } else if (queue_len + completed == max) {
1184 LASSERT(req->rq_reqmsg->transno >= next_transno);
1185 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1186 "waking for skipped transno (skip: "LPD64
1187 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1188 next_transno, queue_len, completed, max, req_transno);
1189 obd->obd_next_recovery_transno = req_transno;
1191 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1192 /* some clients haven't connected in time, but we can try
1193 * to replay requests that demand on already committed ones
1194 * also, we can replay first non-committed transation */
1195 LASSERT(req_transno != 0);
1196 if (req_transno == obd->obd_last_committed + 1) {
1197 obd->obd_next_recovery_transno = req_transno;
1198 } else if (req_transno > obd->obd_last_committed) {
1199 /* can't continue recovery: have no needed transno */
1200 obd->obd_abort_recovery = 1;
1201 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1202 "connected: %d, completed: %d, queue_len: %d, "
1203 "req_transno: "LPU64", next_transno: "LPU64"\n",
1204 max, connected, completed, queue_len,
1205 req_transno, next_transno);
1209 spin_unlock_bh(&obd->obd_processing_task_lock);
1214 static struct ptlrpc_request *
1215 target_next_replay_req(struct obd_device *obd)
1217 struct l_wait_info lwi = { 0 };
1218 struct ptlrpc_request *req;
1220 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1221 obd->obd_next_recovery_transno);
1222 l_wait_event(obd->obd_next_transno_waitq,
1223 check_for_next_transno(obd), &lwi);
1225 spin_lock_bh(&obd->obd_processing_task_lock);
1226 if (obd->obd_abort_recovery) {
1228 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1229 req = list_entry(obd->obd_req_replay_queue.next,
1230 struct ptlrpc_request, rq_list);
1231 list_del_init(&req->rq_list);
1232 obd->obd_requests_queued_for_recovery--;
1236 spin_unlock_bh(&obd->obd_processing_task_lock);
1240 static int check_for_next_lock(struct obd_device *obd)
1242 struct ptlrpc_request *req = NULL;
1245 spin_lock_bh(&obd->obd_processing_task_lock);
1246 if (!list_empty(&obd->obd_lock_replay_queue)) {
1247 req = list_entry(obd->obd_lock_replay_queue.next,
1248 struct ptlrpc_request, rq_list);
1249 CDEBUG(D_HA, "waking for next lock\n");
1251 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1252 CDEBUG(D_HA, "waking for completed lock replay\n");
1254 } else if (obd->obd_abort_recovery) {
1255 CDEBUG(D_HA, "waking for aborted recovery\n");
1258 spin_unlock_bh(&obd->obd_processing_task_lock);
1263 static struct ptlrpc_request *
1264 target_next_replay_lock(struct obd_device *obd)
1266 struct l_wait_info lwi = { 0 };
1267 struct ptlrpc_request *req;
1269 CDEBUG(D_HA, "Waiting for lock\n");
1270 l_wait_event(obd->obd_next_transno_waitq,
1271 check_for_next_lock(obd), &lwi);
1273 spin_lock_bh(&obd->obd_processing_task_lock);
1274 if (obd->obd_abort_recovery) {
1276 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1277 req = list_entry(obd->obd_lock_replay_queue.next,
1278 struct ptlrpc_request, rq_list);
1279 list_del_init(&req->rq_list);
1283 spin_unlock_bh(&obd->obd_processing_task_lock);
1287 static struct ptlrpc_request *
1288 target_next_final_ping(struct obd_device *obd)
1290 struct ptlrpc_request *req;
1292 spin_lock_bh(&obd->obd_processing_task_lock);
1293 if (!list_empty(&obd->obd_final_req_queue)) {
1294 req = list_entry(obd->obd_final_req_queue.next,
1295 struct ptlrpc_request, rq_list);
1296 list_del_init(&req->rq_list);
1300 spin_unlock_bh(&obd->obd_processing_task_lock);
1304 static int req_replay_done(struct obd_export *exp)
1306 if (exp->exp_req_replay_needed)
1311 static int lock_replay_done(struct obd_export *exp)
1313 if (exp->exp_lock_replay_needed)
1318 static int connect_done(struct obd_export *exp)
1320 if (exp->exp_connected)
1325 static int check_for_clients(struct obd_device *obd)
1327 if (obd->obd_abort_recovery)
1329 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1330 if (obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1335 static int target_recovery_thread(void *arg)
1337 struct obd_device *obd = arg;
1338 struct ptlrpc_request *req;
1339 struct target_recovery_data *trd = &obd->obd_recovery_data;
1340 char peer_str[PTL_NALFMT_SIZE];
1341 struct l_wait_info lwi = { 0 };
1342 unsigned long delta;
1343 unsigned long flags;
1346 kportal_daemonize("tgt-recov");
1348 SIGNAL_MASK_LOCK(current, flags);
1349 sigfillset(¤t->blocked);
1351 SIGNAL_MASK_UNLOCK(current, flags);
1353 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1355 trd->trd_processing_task = current->pid;
1357 obd->obd_recovering = 1;
1358 complete(&trd->trd_starting);
1360 /* first of all, we have to know the first transno to replay */
1361 obd->obd_abort_recovery = 0;
1362 l_wait_event(obd->obd_next_transno_waitq,
1363 check_for_clients(obd), &lwi);
1365 spin_lock_bh(&obd->obd_processing_task_lock);
1366 target_cancel_recovery_timer(obd);
1367 spin_unlock_bh(&obd->obd_processing_task_lock);
1369 /* If some clients haven't connected in time, evict them */
1370 if (obd->obd_abort_recovery) {
1372 CDEBUG(D_ERROR, "few clients haven't connect in time (%d/%d),"
1373 "evict them ...\n", obd->obd_connected_clients,
1374 obd->obd_max_recoverable_clients);
1375 obd->obd_abort_recovery = 0;
1376 stale = class_disconnect_stale_exports(obd, connect_done, 0);
1377 atomic_sub(stale, &obd->obd_req_replay_clients);
1378 atomic_sub(stale, &obd->obd_lock_replay_clients);
1381 /* next stage: replay requests */
1383 CDEBUG(D_ERROR, "1: request replay stage - %d clients from t"LPU64"\n",
1384 atomic_read(&obd->obd_req_replay_clients),
1385 obd->obd_next_recovery_transno);
1386 while ((req = target_next_replay_req(obd))) {
1387 LASSERT(trd->trd_processing_task == current->pid);
1388 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s: ",
1389 req->rq_reqmsg->transno,
1390 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1391 (void)trd->trd_recovery_handler(req);
1392 obd->obd_replayed_requests++;
1393 reset_recovery_timer(obd);
1394 /* bug 1580: decide how to properly sync() in recovery*/
1395 //mds_fsync_super(mds->mds_sb);
1396 ptlrpc_free_clone(req);
1397 spin_lock_bh(&obd->obd_processing_task_lock);
1398 obd->obd_next_recovery_transno++;
1399 spin_unlock_bh(&obd->obd_processing_task_lock);
1402 spin_lock_bh(&obd->obd_processing_task_lock);
1403 target_cancel_recovery_timer(obd);
1404 spin_unlock_bh(&obd->obd_processing_task_lock);
1406 /* If some clients haven't replayed requests in time, evict them */
1407 if (obd->obd_abort_recovery) {
1409 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1410 obd->obd_abort_recovery = 0;
1411 stale = class_disconnect_stale_exports(obd, req_replay_done, 0);
1412 atomic_sub(stale, &obd->obd_lock_replay_clients);
1413 abort_req_replay_queue(obd);
1414 /* XXX for debuggin tests 11 and 17 */
1418 /* The second stage: replay locks */
1419 CDEBUG(D_ERROR, "2: lock replay stage - %d clients\n",
1420 atomic_read(&obd->obd_lock_replay_clients));
1421 while ((req = target_next_replay_lock(obd))) {
1422 LASSERT(trd->trd_processing_task == current->pid);
1423 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1424 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1425 (void)trd->trd_recovery_handler(req);
1426 reset_recovery_timer(obd);
1427 ptlrpc_free_clone(req);
1428 obd->obd_replayed_locks++;
1431 spin_lock_bh(&obd->obd_processing_task_lock);
1432 target_cancel_recovery_timer(obd);
1433 spin_unlock_bh(&obd->obd_processing_task_lock);
1435 /* If some clients haven't replayed requests in time, evict them */
1436 if (obd->obd_abort_recovery) {
1438 CERROR("lock replay timed out, aborting ...\n");
1439 obd->obd_abort_recovery = 0;
1440 stale = class_disconnect_stale_exports(obd, lock_replay_done, 0);
1441 abort_lock_replay_queue(obd);
1444 /* We drop recoverying flag to forward all new requests
1445 * to regular mds_handle() since now */
1446 spin_lock_bh(&obd->obd_processing_task_lock);
1447 obd->obd_recovering = 0;
1448 spin_unlock_bh(&obd->obd_processing_task_lock);
1450 /* The third stage: reply on final pings */
1451 CDEBUG(D_ERROR, "3: final stage - process recovery completion pings\n");
1452 while ((req = target_next_final_ping(obd))) {
1453 LASSERT(trd->trd_processing_task == current->pid);
1454 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1455 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1456 (void)trd->trd_recovery_handler(req);
1457 ptlrpc_free_clone(req);
1460 delta = (jiffies - delta) / HZ;
1461 CDEBUG(D_ERROR,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1462 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1463 if (delta > obd_timeout * 2) {
1464 CWARN("too long recovery - read logs\n");
1465 portals_debug_dumplog();
1467 target_finish_recovery(obd);
1469 trd->trd_processing_task = 0;
1470 complete(&trd->trd_finishing);
1474 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1477 struct target_recovery_data *trd = &obd->obd_recovery_data;
1479 memset(trd, 0, sizeof(*trd));
1480 init_completion(&trd->trd_starting);
1481 init_completion(&trd->trd_finishing);
1482 trd->trd_recovery_handler = handler;
1484 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1485 wait_for_completion(&trd->trd_starting);
1486 LASSERT(obd->obd_recovering != 0);
1493 void target_stop_recovery_thread(struct obd_device *obd)
1495 spin_lock_bh(&obd->obd_processing_task_lock);
1496 if (obd->obd_recovery_data.trd_processing_task > 0) {
1497 struct target_recovery_data *trd = &obd->obd_recovery_data;
1498 CERROR("%s: aborting recovery\n", obd->obd_name);
1499 obd->obd_abort_recovery = 1;
1500 wake_up(&obd->obd_next_transno_waitq);
1501 spin_unlock_bh(&obd->obd_processing_task_lock);
1502 wait_for_completion(&trd->trd_finishing);
1504 spin_unlock_bh(&obd->obd_processing_task_lock);
1509 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1511 struct obd_export *exp = req->rq_export;
1512 LASSERT(exp != NULL);
1513 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1514 /* client declares he's ready to replay locks */
1515 spin_lock_bh(&obd->obd_processing_task_lock);
1516 if (exp->exp_req_replay_needed) {
1517 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1518 exp->exp_req_replay_needed = 0;
1519 atomic_dec(&obd->obd_req_replay_clients);
1520 obd->obd_recoverable_clients--;
1521 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1522 CDEBUG(D_HA, "all clients have replayed reqs\n");
1523 wake_up(&obd->obd_next_transno_waitq);
1525 spin_unlock_bh(&obd->obd_processing_task_lock);
1527 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1528 /* client declares he's ready to complete recovery
1529 * so, we put the request on th final queue */
1530 spin_lock_bh(&obd->obd_processing_task_lock);
1531 if (exp->exp_lock_replay_needed) {
1532 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1533 exp->exp_lock_replay_needed = 0;
1534 atomic_dec(&obd->obd_lock_replay_clients);
1535 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1536 CDEBUG(D_HA, "all clients have replayed locks\n");
1537 wake_up(&obd->obd_next_transno_waitq);
1539 spin_unlock_bh(&obd->obd_processing_task_lock);
1545 int target_queue_recovery_request(struct ptlrpc_request *req,
1546 struct obd_device *obd)
1548 struct list_head *tmp;
1550 __u64 transno = req->rq_reqmsg->transno;
1552 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
1553 /* Processing the queue right now, don't re-add. */
1557 target_process_req_flags(obd, req);
1559 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1560 /* client declares he's ready to complete recovery
1561 * so, we put the request on th final queue */
1562 req = ptlrpc_clone_req(req);
1565 DEBUG_REQ(D_HA, req, "queue final req");
1566 spin_lock_bh(&obd->obd_processing_task_lock);
1567 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1568 spin_unlock_bh(&obd->obd_processing_task_lock);
1571 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1572 /* client declares he's ready to replay locks */
1573 req = ptlrpc_clone_req(req);
1576 DEBUG_REQ(D_HA, req, "queue lock replay req");
1577 spin_lock_bh(&obd->obd_processing_task_lock);
1578 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1579 spin_unlock_bh(&obd->obd_processing_task_lock);
1580 wake_up(&obd->obd_next_transno_waitq);
1585 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1586 * (i.e. buflens etc are in my own byte order), but type-dependent
1587 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1590 INIT_LIST_HEAD(&req->rq_list);
1591 DEBUG_REQ(D_HA, req, "not queueing");
1596 /* If we're processing the queue, we want don't want to queue this
1599 * Also, if this request has a transno less than the one we're waiting
1600 * for, we should process it now. It could (and currently always will)
1601 * be an open request for a descriptor that was opened some time ago.
1603 * Also, a resent, replayed request that has already been
1604 * handled will pass through here and be processed immediately.
1606 spin_lock_bh(&obd->obd_processing_task_lock);
1607 if (transno < obd->obd_next_recovery_transno && check_for_clients(obd)) {
1608 /* Processing the queue right now, don't re-add. */
1609 LASSERT(list_empty(&req->rq_list));
1610 spin_unlock_bh(&obd->obd_processing_task_lock);
1613 spin_unlock_bh(&obd->obd_processing_task_lock);
1615 /* A resent, replayed request that is still on the queue; just drop it.
1616 The queued request will handle this. */
1617 if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY))
1618 == (MSG_RESENT | MSG_REPLAY)) {
1619 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1623 req = ptlrpc_clone_req(req);
1627 spin_lock_bh(&obd->obd_processing_task_lock);
1630 list_for_each(tmp, &obd->obd_req_replay_queue) {
1631 struct ptlrpc_request *reqiter =
1632 list_entry(tmp, struct ptlrpc_request, rq_list);
1634 if (reqiter->rq_reqmsg->transno > transno) {
1635 list_add_tail(&req->rq_list, &reqiter->rq_list);
1642 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1644 obd->obd_requests_queued_for_recovery++;
1645 wake_up(&obd->obd_next_transno_waitq);
1646 spin_unlock_bh(&obd->obd_processing_task_lock);
1650 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1652 return req->rq_export->exp_obd;
1656 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1658 if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1659 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1660 DEBUG_REQ(D_ERROR, req, "dropping reply");
1661 /* NB this does _not_ send with ACK disabled, to simulate
1662 * sending OK, but timing out for the ACK */
1663 if (req->rq_reply_state != NULL) {
1664 if (!req->rq_reply_state->rs_difficult) {
1665 lustre_free_reply_state (req->rq_reply_state);
1666 req->rq_reply_state = NULL;
1668 struct ptlrpc_service *svc =
1669 req->rq_rqbd->rqbd_srv_ni->sni_service;
1670 atomic_inc(&svc->srv_outstanding_replies);
1677 req->rq_status = rc;
1678 return (ptlrpc_error(req));
1680 DEBUG_REQ(D_NET, req, "sending reply");
1683 return (ptlrpc_send_reply(req, 1));
1687 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1690 unsigned long flags;
1691 struct ptlrpc_reply_state *rs;
1692 struct obd_device *obd;
1693 struct obd_export *exp;
1694 struct ptlrpc_srv_ni *sni;
1695 struct ptlrpc_service *svc;
1697 sni = req->rq_rqbd->rqbd_srv_ni;
1698 svc = sni->sni_service;
1700 rs = req->rq_reply_state;
1701 if (rs == NULL || !rs->rs_difficult) {
1702 /* The easy case; no notifiers and reply_out_callback()
1703 * cleans up (i.e. we can't look inside rs after a
1704 * successful send) */
1705 netrc = target_send_reply_msg (req, rc, fail_id);
1707 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1711 /* must be an export if locks saved */
1712 LASSERT (req->rq_export != NULL);
1713 /* req/reply consistent */
1714 LASSERT (rs->rs_srv_ni == sni);
1717 LASSERT (!rs->rs_scheduled);
1718 LASSERT (!rs->rs_scheduled_ever);
1719 LASSERT (!rs->rs_handled);
1720 LASSERT (!rs->rs_on_net);
1721 LASSERT (rs->rs_export == NULL);
1722 LASSERT (list_empty(&rs->rs_obd_list));
1723 LASSERT (list_empty(&rs->rs_exp_list));
1725 exp = class_export_get (req->rq_export);
1728 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1729 rs->rs_scheduled = 1;
1731 rs->rs_xid = req->rq_xid;
1732 rs->rs_transno = req->rq_transno;
1733 rs->rs_export = exp;
1735 spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1737 if (rs->rs_transno > obd->obd_last_committed) {
1738 /* not committed already */
1739 list_add_tail (&rs->rs_obd_list,
1740 &obd->obd_uncommitted_replies);
1743 spin_unlock (&obd->obd_uncommitted_replies_lock);
1744 spin_lock (&exp->exp_lock);
1746 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1748 spin_unlock_irqrestore (&exp->exp_lock, flags);
1750 netrc = target_send_reply_msg (req, rc, fail_id);
1752 spin_lock_irqsave (&svc->srv_lock, flags);
1754 svc->srv_n_difficult_replies++;
1756 if (netrc != 0) /* error sending: reply is off the net */
1759 if (!rs->rs_on_net || /* some notifier */
1760 list_empty(&rs->rs_exp_list) || /* completed already */
1761 list_empty(&rs->rs_obd_list)) {
1762 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1763 wake_up (&svc->srv_waitq);
1765 list_add (&rs->rs_list, &sni->sni_active_replies);
1766 rs->rs_scheduled = 0; /* allow notifier to schedule */
1769 spin_unlock_irqrestore (&svc->srv_lock, flags);
1772 int target_handle_ping(struct ptlrpc_request *req)
1774 return lustre_pack_reply(req, 0, NULL, NULL);