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;
242 /* FIXME: should limit this for the sum of all cl_dirty_max */
243 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
244 if (cli->cl_dirty_max >> PAGE_SHIFT > num_physpages / 8)
245 cli->cl_dirty_max = num_physpages << (PAGE_SHIFT - 3);
247 INIT_LIST_HEAD(&cli->cl_cache_waiters);
248 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
249 INIT_LIST_HEAD(&cli->cl_loi_write_list);
250 INIT_LIST_HEAD(&cli->cl_loi_read_list);
251 spin_lock_init(&cli->cl_loi_list_lock);
252 cli->cl_r_in_flight = 0;
253 cli->cl_w_in_flight = 0;
254 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
255 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
256 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
257 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
259 memset(&cli->cl_last_write_time, 0,
260 sizeof(cli->cl_last_write_time));
262 cli->cl_cache_wait_num = 0;
263 cli->cl_cache_wait_sum = 0;
264 cli->cl_write_gap_sum = 0;
265 cli->cl_write_gaps = 0;
266 cli->cl_write_num = 0;
267 cli->cl_read_num = 0;
269 cli->cl_dirty_num = 0;
270 cli->cl_dirty_sum = 0;
271 cli->cl_dirty_av = 0;
272 cli->cl_sync_rpcs = 0;
273 cli->cl_dirty_dmax = 0;
274 cli->cl_dirty_dmin = 0;
276 if (num_physpages >> (20 - PAGE_SHIFT) <= 128) { /* <= 128 MB */
277 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 4;
278 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 4;
280 } else if (num_physpages >> (20 - PAGE_SHIFT) <= 512) { /* <= 512 MB */
281 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES / 2;
282 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT / 2;
285 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
286 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
291 CERROR("ldlm_get_ref failed: %d\n", rc);
295 ptlrpc_init_client(rq_portal, rp_portal, name,
296 &obddev->obd_ldlm_client);
298 imp = class_new_import();
300 GOTO(err_ldlm, rc = -ENOENT);
301 imp->imp_client = &obddev->obd_ldlm_client;
302 imp->imp_obd = obddev;
303 imp->imp_connect_op = connect_op;
304 imp->imp_generation = 0;
305 imp->imp_initial_recov = 1;
306 INIT_LIST_HEAD(&imp->imp_pinger_chain);
307 memcpy(imp->imp_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
308 LUSTRE_CFG_BUFLEN(lcfg, 1));
309 class_import_put(imp);
311 rc = client_import_add_conn(imp, &server_uuid, 1);
313 CERROR("can't add initial connection\n");
314 GOTO(err_import, rc);
317 cli->cl_import = imp;
318 cli->cl_max_mds_easize = sizeof(struct lov_mds_md);
319 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
320 cli->cl_sandev = to_kdev_t(0);
322 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
323 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
324 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
325 name, obddev->obd_name,
326 imp->imp_target_uuid.uuid);
327 imp->imp_invalid = 1;
329 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0)
330 mgmt_name = lustre_cfg_string(lcfg, 4);
332 mgmt_name = lustre_cfg_string(lcfg, 3);
336 if (mgmt_name != NULL) {
337 /* Register with management client if we need to. */
338 CDEBUG(D_HA, "%s registering with %s for events about %s\n",
339 obddev->obd_name, mgmt_name, server_uuid.uuid);
341 mgmt_obd = class_name2obd(mgmt_name);
343 CERROR("can't find mgmtcli %s to register\n",
345 GOTO(err_import, rc = -ENOSYS);
348 register_f = (mgmtcli_register_for_events_t)symbol_get("mgmtcli_register_for_events");
350 CERROR("can't i_m_g mgmtcli_register_for_events\n");
351 GOTO(err_import, rc = -ENOSYS);
354 rc = register_f(mgmt_obd, obddev, &imp->imp_target_uuid);
355 symbol_put("mgmtcli_register_for_events");
358 cli->cl_mgmtcli_obd = mgmt_obd;
364 class_destroy_import(imp);
372 int client_obd_cleanup(struct obd_device *obddev, int flags)
374 struct client_obd *cli = &obddev->u.cli;
380 if (cli->cl_mgmtcli_obd) {
381 mgmtcli_deregister_for_events_t dereg_f;
383 dereg_f = (mgmtcli_deregister_for_events_t)symbol_get("mgmtcli_deregister_for_events");
384 dereg_f(cli->cl_mgmtcli_obd, obddev);
385 symbol_put("mgmtcli_deregister_for_events");
388 /* Here we try to drop the security structure after destroy import,
389 * to avoid issue of "sleep in spinlock".
391 class_import_get(cli->cl_import);
392 class_destroy_import(cli->cl_import);
393 ptlrpcs_import_drop_sec(cli->cl_import);
394 class_import_put(cli->cl_import);
395 cli->cl_import = NULL;
397 ldlm_put_ref(flags & OBD_OPT_FORCE);
401 int client_connect_import(struct lustre_handle *dlm_handle,
402 struct obd_device *obd,
403 struct obd_uuid *cluuid,
404 struct obd_connect_data *conn_data,
405 unsigned long connect_flags)
407 struct client_obd *cli = &obd->u.cli;
408 struct obd_import *imp = cli->cl_import;
409 struct obd_export *exp;
414 rc = class_connect(dlm_handle, obd, cluuid);
418 cli->cl_conn_count++;
419 if (cli->cl_conn_count > 1)
421 exp = class_conn2export(dlm_handle);
423 if (obd->obd_namespace != NULL)
424 CERROR("already have namespace!\n");
425 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
426 LDLM_NAMESPACE_CLIENT);
427 if (obd->obd_namespace == NULL)
428 GOTO(out_disco, rc = -ENOMEM);
430 rc = ptlrpcs_import_get_sec(imp);
434 imp->imp_dlm_handle = *dlm_handle;
435 rc = ptlrpc_init_import(imp);
439 imp->imp_connect_flags = connect_flags;
441 memcpy(&imp->imp_connect_data, conn_data, sizeof(*conn_data));
443 rc = ptlrpc_connect_import(imp, NULL);
445 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
448 LASSERT(exp->exp_connection);
449 ptlrpc_pinger_add_import(imp);
454 ldlm_namespace_free(obd->obd_namespace, 0);
455 obd->obd_namespace = NULL;
457 cli->cl_conn_count--;
458 class_disconnect(exp, 0);
460 class_export_put(exp);
467 int client_disconnect_export(struct obd_export *exp, unsigned long flags)
469 struct obd_device *obd = class_exp2obd(exp);
470 struct client_obd *cli = &obd->u.cli;
471 struct obd_import *imp = cli->cl_import;
476 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
477 exp, exp ? exp->exp_handle.h_cookie : -1);
482 if (!cli->cl_conn_count) {
483 CERROR("disconnecting disconnected device (%s)\n",
485 GOTO(out_sem, rc = -EINVAL);
488 cli->cl_conn_count--;
489 if (cli->cl_conn_count)
490 GOTO(out_no_disconnect, rc = 0);
492 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
493 * delete it regardless. (It's safe to delete an import that was
495 (void)ptlrpc_pinger_del_import(imp);
497 if (obd->obd_namespace != NULL) {
498 /* obd_no_recov == local only */
499 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
500 obd->obd_no_recov, NULL);
501 ldlm_namespace_free(obd->obd_namespace, obd->obd_no_recov);
502 obd->obd_namespace = NULL;
506 * Yeah, obd_no_recov also (mainly) means "forced shutdown".
508 if (obd->obd_no_recov)
509 ptlrpc_invalidate_import(imp, 0);
511 rc = ptlrpc_disconnect_import(imp);
515 err = class_disconnect(exp, 0);
523 /* --------------------------------------------------------------------------
524 * from old lib/target.c
525 * -------------------------------------------------------------------------- */
527 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
528 struct obd_uuid *cluuid, int initial_conn)
530 if (exp->exp_connection && !initial_conn) {
531 struct lustre_handle *hdl;
532 hdl = &exp->exp_imp_reverse->imp_remote_handle;
533 /* Might be a re-connect after a partition. */
534 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
535 CERROR("%s reconnecting\n", cluuid->uuid);
536 conn->cookie = exp->exp_handle.h_cookie;
539 CERROR("%s reconnecting from %s, "
540 "handle mismatch (ours "LPX64", theirs "
541 LPX64")\n", cluuid->uuid,
542 exp->exp_connection->c_remote_uuid.uuid,
543 hdl->cookie, conn->cookie);
544 memset(conn, 0, sizeof *conn);
549 conn->cookie = exp->exp_handle.h_cookie;
550 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
552 CDEBUG(D_IOCTL,"connect: cookie "LPX64"\n", conn->cookie);
556 static inline int ptlrpc_peer_is_local(struct ptlrpc_peer *peer)
558 ptl_process_id_t myid;
560 PtlGetId(peer->peer_ni->pni_ni_h, &myid);
561 return (memcmp(&peer->peer_id, &myid, sizeof(myid)) == 0);
564 /* To check whether the p_flavor is in deny list or not
567 * EPERM found, refuse
570 static int check_deny_list(struct list_head *head, __u32 flavor)
572 deny_sec_t *p_deny_sec = NULL;
573 deny_sec_t *n_deny_sec = NULL;
575 list_for_each_entry_safe(p_deny_sec, n_deny_sec, head, list) {
576 if (p_deny_sec->flavor == flavor)
582 int target_check_deny_sec(struct obd_device *target, struct ptlrpc_request *req)
587 flavor = req->rq_req_secflvr;
589 if (!strcmp(target->obd_type->typ_name, OBD_MDS_DEVICENAME)) {
590 spin_lock(&target->u.mds.mds_denylist_lock);
591 rc = check_deny_list(&target->u.mds.mds_denylist, flavor);
592 spin_unlock(&target->u.mds.mds_denylist_lock);
593 } else if (!strcmp(target->obd_type->typ_name, OBD_FILTER_DEVICENAME)) {
594 spin_lock(&target->u.filter.fo_denylist_lock);
595 rc = check_deny_list(&target->u.filter.fo_denylist, flavor);
596 spin_unlock(&target->u.filter.fo_denylist_lock);
602 int target_handle_connect(struct ptlrpc_request *req)
604 unsigned long connect_flags = 0, *cfp;
605 struct obd_device *target;
606 struct obd_export *export = NULL;
607 struct obd_import *revimp;
608 struct lustre_handle conn;
609 struct obd_uuid tgtuuid;
610 struct obd_uuid cluuid;
611 struct obd_uuid remote_uuid;
613 struct obd_connect_data *conn_data;
614 int conn_data_size = sizeof(*conn_data);
618 int initial_conn = 0;
619 char peer_str[PTL_NALFMT_SIZE];
620 const int offset = 1;
623 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
625 LASSERT_REQSWAB (req, offset + 0);
626 str = lustre_msg_string(req->rq_reqmsg, offset + 0,
627 sizeof(tgtuuid) - 1);
629 CERROR("bad target UUID for connect\n");
630 GOTO(out, rc = -EINVAL);
633 obd_str2uuid (&tgtuuid, str);
634 target = class_uuid2obd(&tgtuuid);
636 target = class_name2obd(str);
638 if (!target || target->obd_stopping || !target->obd_set_up) {
639 CERROR("UUID '%s' is not available for connect from %s\n",
640 str, req->rq_peerstr);
641 GOTO(out, rc = -ENODEV);
644 /* check the secure deny list of mds/ost */
645 rc = target_check_deny_sec(target, req);
649 LASSERT_REQSWAB (req, offset + 1);
650 str = lustre_msg_string(req->rq_reqmsg, offset + 1, sizeof(cluuid) - 1);
652 CERROR("bad client UUID for connect\n");
653 GOTO(out, rc = -EINVAL);
656 obd_str2uuid (&cluuid, str);
658 /* XXX extract a nettype and format accordingly */
659 switch (sizeof(ptl_nid_t)) {
660 /* NB the casts only avoid compiler warnings */
662 snprintf((char *)remote_uuid.uuid, sizeof(remote_uuid),
663 "NET_"LPX64"_UUID", (__u64)req->rq_peer.peer_id.nid);
666 snprintf((char *)remote_uuid.uuid, sizeof(remote_uuid),
667 "NET_%x_UUID", (__u32)req->rq_peer.peer_id.nid);
673 tmp = lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn));
675 GOTO(out, rc = -EPROTO);
677 memcpy(&conn, tmp, sizeof conn);
679 cfp = lustre_msg_buf(req->rq_reqmsg, offset + 3, sizeof(unsigned long));
680 LASSERT(cfp != NULL);
681 connect_flags = *cfp;
683 conn_data = lustre_swab_reqbuf(req, offset + 4, sizeof(*conn_data),
684 lustre_swab_connect);
686 GOTO(out, rc = -EPROTO);
688 rc = lustre_pack_reply(req, 1, &conn_data_size, NULL);
692 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
695 /* lctl gets a backstage, all-access pass. */
696 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
697 goto dont_check_exports;
699 spin_lock(&target->obd_dev_lock);
700 list_for_each(p, &target->obd_exports) {
701 export = list_entry(p, struct obd_export, exp_obd_chain);
702 if (obd_uuid_equals(&cluuid, &export->exp_client_uuid)) {
703 spin_unlock(&target->obd_dev_lock);
704 LASSERT(export->exp_obd == target);
706 rc = target_handle_reconnect(&conn, export, &cluuid,
712 /* If we found an export, we already unlocked. */
714 spin_unlock(&target->obd_dev_lock);
715 } else if (req->rq_export == NULL &&
716 atomic_read(&export->exp_rpc_count) > 0) {
717 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
718 target->obd_name, cluuid.uuid,
719 ptlrpc_peernid2str(&req->rq_peer, peer_str),
720 export, atomic_read(&export->exp_refcount));
721 GOTO(out, rc = -EBUSY);
722 } else if (req->rq_export != NULL &&
723 atomic_read(&export->exp_rpc_count) > 1) {
724 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
725 target->obd_name, cluuid.uuid,
726 ptlrpc_peernid2str(&req->rq_peer, peer_str),
727 export, atomic_read(&export->exp_rpc_count));
728 GOTO(out, rc = -EBUSY);
729 } else if (req->rq_reqmsg->conn_cnt == 1 && !initial_conn) {
730 CERROR("%s reconnected with 1 conn_cnt; cookies not random?\n",
732 GOTO(out, rc = -EALREADY);
735 /* Tell the client if we're in recovery. */
736 /* If this is the first client, start the recovery timer */
737 CWARN("%s: connection from %s@%s/%lu %st"LPU64"\n", target->obd_name,
738 cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str), *cfp,
739 target->obd_recovering ? "recovering/" : "", conn_data->transno);
741 if (target->obd_recovering) {
742 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
743 target_start_recovery_timer(target);
747 /* Tell the client if we support replayable requests */
748 if (target->obd_replayable)
749 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
752 if (export == NULL) {
753 if (target->obd_recovering) {
754 CERROR("%s denying connection for new client %s@%s: "
755 "%d clients in recovery for %lds\n", target->obd_name,
757 ptlrpc_peernid2str(&req->rq_peer, peer_str),
758 target->obd_recoverable_clients,
759 (target->obd_recovery_timer.expires-jiffies)/HZ);
763 rc = obd_connect(&conn, target, &cluuid, conn_data,
768 /* Return only the parts of obd_connect_data that we understand, so the
769 * client knows that we don't understand the rest. */
770 conn_data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
771 memcpy(lustre_msg_buf(req->rq_repmsg, 0, sizeof(*conn_data)), conn_data,
774 /* Tell the client if we support replayable requests */
775 if (target->obd_replayable)
776 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
778 /* If all else goes well, this is our RPC return code. */
781 if (rc && rc != EALREADY)
784 req->rq_repmsg->handle = conn;
786 /* If the client and the server are the same node, we will already
787 * have an export that really points to the client's DLM export,
788 * because we have a shared handles table.
790 * XXX this will go away when shaver stops sending the "connect" handle
791 * in the real "remote handle" field of the request --phik 24 Apr 2003
793 if (req->rq_export != NULL)
794 class_export_put(req->rq_export);
796 /* ownership of this export ref transfers to the request */
797 export = req->rq_export = class_conn2export(&conn);
798 LASSERT(export != NULL);
800 spin_lock_irqsave(&export->exp_lock, flags);
802 req->rq_repmsg->conn_cnt = export->exp_conn_cnt + 1;
803 } else if (export->exp_conn_cnt >= req->rq_reqmsg->conn_cnt) {
804 CERROR("%s@%s: already connected at a higher conn_cnt: %d > %d\n",
805 cluuid.uuid, ptlrpc_peernid2str(&req->rq_peer, peer_str),
806 export->exp_conn_cnt,
807 req->rq_reqmsg->conn_cnt);
808 spin_unlock_irqrestore(&export->exp_lock, flags);
809 GOTO(out, rc = -EALREADY);
811 export->exp_conn_cnt = req->rq_reqmsg->conn_cnt;
812 spin_unlock_irqrestore(&export->exp_lock, flags);
814 /* request from liblustre? */
815 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT)
816 export->exp_libclient = 1;
818 if (!(lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_ASYNC) &&
819 ptlrpc_peer_is_local(&req->rq_peer)) {
820 CWARN("%s: exp %p set sync\n", target->obd_name, export);
821 export->exp_sync = 1;
823 CDEBUG(D_HA, "%s: exp %p set async\n",target->obd_name,export);
824 export->exp_sync = 0;
827 if (export->exp_connection != NULL)
828 ptlrpc_put_connection(export->exp_connection);
829 export->exp_connection = ptlrpc_get_connection(&req->rq_peer,
832 if (rc == EALREADY) {
833 /* We indicate the reconnection in a flag, not an error code. */
834 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
838 spin_lock_bh(&target->obd_processing_task_lock);
839 if (target->obd_recovering && export->exp_connected == 0) {
840 __u64 t = conn_data->transno;
841 export->exp_connected = 1;
842 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
843 && t < target->obd_next_recovery_transno)
844 target->obd_next_recovery_transno = t;
845 target->obd_connected_clients++;
846 if (target->obd_connected_clients == target->obd_max_recoverable_clients)
847 wake_up(&target->obd_next_transno_waitq);
849 spin_unlock_bh(&target->obd_processing_task_lock);
851 memcpy(&conn, lustre_msg_buf(req->rq_reqmsg, offset + 2, sizeof(conn)),
854 if (export->exp_imp_reverse != NULL) {
855 /* same logic as client_obd_cleanup */
856 class_import_get(export->exp_imp_reverse);
857 class_destroy_import(export->exp_imp_reverse);
858 ptlrpcs_import_drop_sec(export->exp_imp_reverse);
859 class_import_put(export->exp_imp_reverse);
862 /* for the rest part, we return -ENOTCONN in case of errors
863 * in order to let client initialize connection again.
865 revimp = export->exp_imp_reverse = class_new_import();
867 CERROR("fail to alloc new reverse import.\n");
868 GOTO(out, rc = -ENOTCONN);
871 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
872 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
873 revimp->imp_remote_handle = conn;
874 revimp->imp_obd = target;
875 revimp->imp_dlm_fake = 1;
876 revimp->imp_state = LUSTRE_IMP_FULL;
878 rc = ptlrpcs_import_get_sec(revimp);
880 CERROR("reverse import can not get sec: %d\n", rc);
881 class_destroy_import(revimp);
882 export->exp_imp_reverse = NULL;
883 GOTO(out, rc = -ENOTCONN);
886 class_import_put(revimp);
888 rc = obd_connect_post(export, initial_conn, connect_flags);
895 int target_handle_disconnect(struct ptlrpc_request *req)
897 struct obd_export *exp;
901 rc = lustre_pack_reply(req, 0, NULL, NULL);
905 /* keep the rq_export around so we can send the reply */
906 exp = class_export_get(req->rq_export);
907 req->rq_status = obd_disconnect(exp, 0);
911 void target_destroy_export(struct obd_export *exp)
913 /* exports created from last_rcvd data, and "fake"
914 exports created by lctl don't have an import */
915 if (exp->exp_imp_reverse != NULL) {
916 ptlrpcs_import_drop_sec(exp->exp_imp_reverse);
917 class_destroy_import(exp->exp_imp_reverse);
920 /* We cancel locks at disconnect time, but this will catch any locks
921 * granted in a race with recovery-induced disconnect. */
922 if (exp->exp_obd->obd_namespace != NULL)
923 ldlm_cancel_locks_for_export(exp);
930 struct ptlrpc_request *
931 ptlrpc_clone_req( struct ptlrpc_request *orig_req)
933 struct ptlrpc_request *copy_req;
934 struct lustre_msg *copy_reqmsg;
936 OBD_ALLOC(copy_req, sizeof *copy_req);
939 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
941 OBD_FREE(copy_req, sizeof *copy_req);
945 memcpy(copy_req, orig_req, sizeof *copy_req);
946 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
947 /* the copied req takes over the reply state and security data */
948 orig_req->rq_reply_state = NULL;
949 orig_req->rq_svcsec_data = NULL;
951 copy_req->rq_reqmsg = copy_reqmsg;
952 class_export_get(copy_req->rq_export);
953 INIT_LIST_HEAD(©_req->rq_list);
958 void ptlrpc_free_clone( struct ptlrpc_request *req)
961 svcsec_cleanup_req(req);
963 class_export_put(req->rq_export);
964 list_del(&req->rq_list);
965 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
966 OBD_FREE(req, sizeof *req);
969 static void target_release_saved_req(struct ptlrpc_request *req)
972 svcsec_cleanup_req(req);
974 class_export_put(req->rq_export);
975 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
976 OBD_FREE(req, sizeof *req);
979 static void target_finish_recovery(struct obd_device *obd)
983 ldlm_reprocess_all_ns(obd->obd_namespace);
985 /* when recovery finished, cleanup orphans on mds and ost */
986 if (OBT(obd) && OBP(obd, postrecov)) {
987 rc = OBP(obd, postrecov)(obd);
989 CWARN("%s: all clients recovered, %d MDS "
990 "orphans deleted\n", obd->obd_name, rc);
992 CERROR("postrecov failed %d\n", rc);
995 obd->obd_recovery_end = LTIME_S(CURRENT_TIME);
999 static void abort_req_replay_queue(struct obd_device *obd)
1001 struct ptlrpc_request *req;
1002 struct list_head *tmp, *n;
1005 list_for_each_safe(tmp, n, &obd->obd_req_replay_queue) {
1006 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1007 list_del(&req->rq_list);
1008 DEBUG_REQ(D_ERROR, req, "aborted:");
1009 req->rq_status = -ENOTCONN;
1010 req->rq_type = PTL_RPC_MSG_ERR;
1011 rc = lustre_pack_reply(req, 0, NULL, NULL);
1015 DEBUG_REQ(D_ERROR, req,
1016 "packing failed for abort-reply; skipping");
1018 target_release_saved_req(req);
1022 static void abort_lock_replay_queue(struct obd_device *obd)
1024 struct ptlrpc_request *req;
1025 struct list_head *tmp, *n;
1028 list_for_each_safe(tmp, n, &obd->obd_lock_replay_queue) {
1029 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1030 list_del(&req->rq_list);
1031 DEBUG_REQ(D_ERROR, req, "aborted:");
1032 req->rq_status = -ENOTCONN;
1033 req->rq_type = PTL_RPC_MSG_ERR;
1034 rc = lustre_pack_reply(req, 0, NULL, NULL);
1038 DEBUG_REQ(D_ERROR, req,
1039 "packing failed for abort-reply; skipping");
1041 target_release_saved_req(req);
1045 /* Called from a cleanup function if the device is being cleaned up
1046 forcefully. The exports should all have been disconnected already,
1047 the only thing left to do is
1048 - clear the recovery flags
1050 - free queued requests and replies, but don't send replies
1051 Because the obd_stopping flag is set, no new requests should be received.
1054 void target_cleanup_recovery(struct obd_device *obd)
1056 struct list_head *tmp, *n;
1057 struct ptlrpc_request *req;
1059 spin_lock_bh(&obd->obd_processing_task_lock);
1060 if (!obd->obd_recovering) {
1061 spin_unlock_bh(&obd->obd_processing_task_lock);
1065 obd->obd_recovering = obd->obd_abort_recovery = 0;
1066 target_cancel_recovery_timer(obd);
1067 spin_unlock_bh(&obd->obd_processing_task_lock);
1069 list_for_each_safe(tmp, n, &obd->obd_req_replay_queue) {
1070 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1071 list_del(&req->rq_list);
1072 LASSERT (req->rq_reply_state == 0);
1073 target_release_saved_req(req);
1075 list_for_each_safe(tmp, n, &obd->obd_lock_replay_queue) {
1076 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1077 list_del(&req->rq_list);
1078 LASSERT (req->rq_reply_state == 0);
1079 target_release_saved_req(req);
1081 list_for_each_safe(tmp, n, &obd->obd_final_req_queue) {
1082 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1083 list_del(&req->rq_list);
1084 LASSERT (req->rq_reply_state == 0);
1085 target_release_saved_req(req);
1090 static void target_abort_recovery(void *data)
1092 struct obd_device *obd = data;
1094 LASSERT(!obd->obd_recovering);
1096 class_disconnect_stale_exports(obd, 0);
1098 CERROR("%s: recovery period over; disconnecting unfinished clients.\n",
1101 abort_recovery_queue(obd);
1102 target_finish_recovery(obd);
1103 ptlrpc_run_recovery_over_upcall(obd);
1107 static void target_recovery_expired(unsigned long castmeharder)
1109 struct obd_device *obd = (struct obd_device *)castmeharder;
1110 spin_lock_bh(&obd->obd_processing_task_lock);
1111 if (obd->obd_recovering)
1112 obd->obd_abort_recovery = 1;
1114 wake_up(&obd->obd_next_transno_waitq);
1115 spin_unlock_bh(&obd->obd_processing_task_lock);
1119 /* obd_processing_task_lock should be held */
1120 void target_cancel_recovery_timer(struct obd_device *obd)
1122 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1123 del_timer(&obd->obd_recovery_timer);
1127 static void reset_recovery_timer(struct obd_device *obd)
1129 spin_lock_bh(&obd->obd_processing_task_lock);
1130 if (!obd->obd_recovering) {
1131 spin_unlock_bh(&obd->obd_processing_task_lock);
1134 CDEBUG(D_HA, "timer will expire in %u seconds\n",
1135 OBD_RECOVERY_TIMEOUT / HZ);
1136 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
1137 spin_unlock_bh(&obd->obd_processing_task_lock);
1141 /* Only start it the first time called */
1142 void target_start_recovery_timer(struct obd_device *obd)
1144 spin_lock_bh(&obd->obd_processing_task_lock);
1145 if (!obd->obd_recovering || timer_pending(&obd->obd_recovery_timer)) {
1146 spin_unlock_bh(&obd->obd_processing_task_lock);
1149 CWARN("%s: starting recovery timer (%us)\n", obd->obd_name,
1150 OBD_RECOVERY_TIMEOUT / HZ);
1151 obd->obd_recovery_timer.function = target_recovery_expired;
1152 obd->obd_recovery_timer.data = (unsigned long)obd;
1153 mod_timer(&obd->obd_recovery_timer, jiffies + OBD_RECOVERY_TIMEOUT);
1154 spin_unlock_bh(&obd->obd_processing_task_lock);
1158 static int check_for_next_transno(struct obd_device *obd)
1160 struct ptlrpc_request *req = NULL;
1161 int wake_up = 0, connected, completed, queue_len, max;
1162 __u64 next_transno, req_transno;
1164 spin_lock_bh(&obd->obd_processing_task_lock);
1165 if (!list_empty(&obd->obd_req_replay_queue)) {
1166 req = list_entry(obd->obd_req_replay_queue.next,
1167 struct ptlrpc_request, rq_list);
1168 req_transno = req->rq_reqmsg->transno;
1173 max = obd->obd_max_recoverable_clients;
1174 connected = obd->obd_connected_clients;
1175 completed = max - obd->obd_recoverable_clients;
1176 queue_len = obd->obd_requests_queued_for_recovery;
1177 next_transno = obd->obd_next_recovery_transno;
1179 CDEBUG(D_HA,"max: %d, connected: %d, completed: %d, queue_len: %d, "
1180 "req_transno: "LPU64", next_transno: "LPU64"\n",
1181 max, connected, completed, queue_len, req_transno, next_transno);
1182 if (obd->obd_abort_recovery) {
1183 CDEBUG(D_HA, "waking for aborted recovery\n");
1185 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1186 CDEBUG(D_HA, "waking for completed recovery\n");
1188 } else if (req_transno == next_transno) {
1189 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1191 } else if (queue_len + completed == max) {
1192 LASSERT(req->rq_reqmsg->transno >= next_transno);
1193 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1194 "waking for skipped transno (skip: "LPD64
1195 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1196 next_transno, queue_len, completed, max, req_transno);
1197 obd->obd_next_recovery_transno = req_transno;
1199 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1200 /* some clients haven't connected in time, but we can try
1201 * to replay requests that demand on already committed ones
1202 * also, we can replay first non-committed transation */
1203 LASSERT(req_transno != 0);
1204 if (req_transno == obd->obd_last_committed + 1) {
1205 obd->obd_next_recovery_transno = req_transno;
1206 } else if (req_transno > obd->obd_last_committed) {
1207 /* can't continue recovery: have no needed transno */
1208 obd->obd_abort_recovery = 1;
1209 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1210 "connected: %d, completed: %d, queue_len: %d, "
1211 "req_transno: "LPU64", next_transno: "LPU64"\n",
1212 max, connected, completed, queue_len,
1213 req_transno, next_transno);
1217 spin_unlock_bh(&obd->obd_processing_task_lock);
1222 static struct ptlrpc_request *
1223 target_next_replay_req(struct obd_device *obd)
1225 struct l_wait_info lwi = { 0 };
1226 struct ptlrpc_request *req;
1228 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1229 obd->obd_next_recovery_transno);
1230 l_wait_event(obd->obd_next_transno_waitq,
1231 check_for_next_transno(obd), &lwi);
1233 spin_lock_bh(&obd->obd_processing_task_lock);
1234 if (obd->obd_abort_recovery) {
1236 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1237 req = list_entry(obd->obd_req_replay_queue.next,
1238 struct ptlrpc_request, rq_list);
1239 list_del_init(&req->rq_list);
1240 obd->obd_requests_queued_for_recovery--;
1244 spin_unlock_bh(&obd->obd_processing_task_lock);
1248 static int check_for_next_lock(struct obd_device *obd)
1250 struct ptlrpc_request *req = NULL;
1253 spin_lock_bh(&obd->obd_processing_task_lock);
1254 if (!list_empty(&obd->obd_lock_replay_queue)) {
1255 req = list_entry(obd->obd_lock_replay_queue.next,
1256 struct ptlrpc_request, rq_list);
1257 CDEBUG(D_HA, "waking for next lock\n");
1259 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1260 CDEBUG(D_HA, "waking for completed lock replay\n");
1262 } else if (obd->obd_abort_recovery) {
1263 CDEBUG(D_HA, "waking for aborted recovery\n");
1266 spin_unlock_bh(&obd->obd_processing_task_lock);
1271 static struct ptlrpc_request *
1272 target_next_replay_lock(struct obd_device *obd)
1274 struct l_wait_info lwi = { 0 };
1275 struct ptlrpc_request *req;
1277 CDEBUG(D_HA, "Waiting for lock\n");
1278 l_wait_event(obd->obd_next_transno_waitq,
1279 check_for_next_lock(obd), &lwi);
1281 spin_lock_bh(&obd->obd_processing_task_lock);
1282 if (obd->obd_abort_recovery) {
1284 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1285 req = list_entry(obd->obd_lock_replay_queue.next,
1286 struct ptlrpc_request, rq_list);
1287 list_del_init(&req->rq_list);
1291 spin_unlock_bh(&obd->obd_processing_task_lock);
1295 static struct ptlrpc_request *
1296 target_next_final_ping(struct obd_device *obd)
1298 struct ptlrpc_request *req;
1300 spin_lock_bh(&obd->obd_processing_task_lock);
1301 if (!list_empty(&obd->obd_final_req_queue)) {
1302 req = list_entry(obd->obd_final_req_queue.next,
1303 struct ptlrpc_request, rq_list);
1304 list_del_init(&req->rq_list);
1308 spin_unlock_bh(&obd->obd_processing_task_lock);
1312 static int req_replay_done(struct obd_export *exp)
1314 if (exp->exp_req_replay_needed)
1319 static int lock_replay_done(struct obd_export *exp)
1321 if (exp->exp_lock_replay_needed)
1326 static int connect_done(struct obd_export *exp)
1328 if (exp->exp_connected)
1333 static int check_for_clients(struct obd_device *obd)
1335 if (obd->obd_abort_recovery)
1337 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1338 if (obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1343 static int target_recovery_thread(void *arg)
1345 struct obd_device *obd = arg;
1346 struct ptlrpc_request *req;
1347 struct target_recovery_data *trd = &obd->obd_recovery_data;
1348 char peer_str[PTL_NALFMT_SIZE];
1349 struct l_wait_info lwi = { 0 };
1350 unsigned long delta;
1351 unsigned long flags;
1354 kportal_daemonize("tgt-recov");
1356 SIGNAL_MASK_LOCK(current, flags);
1357 sigfillset(¤t->blocked);
1359 SIGNAL_MASK_UNLOCK(current, flags);
1361 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1363 trd->trd_processing_task = current->pid;
1365 obd->obd_recovering = 1;
1366 complete(&trd->trd_starting);
1368 /* first of all, we have to know the first transno to replay */
1369 obd->obd_abort_recovery = 0;
1370 l_wait_event(obd->obd_next_transno_waitq,
1371 check_for_clients(obd), &lwi);
1373 spin_lock_bh(&obd->obd_processing_task_lock);
1374 target_cancel_recovery_timer(obd);
1375 spin_unlock_bh(&obd->obd_processing_task_lock);
1377 /* If some clients haven't connected in time, evict them */
1378 if (obd->obd_abort_recovery) {
1380 CDEBUG(D_ERROR, "few clients haven't connect in time (%d/%d),"
1381 "evict them ...\n", obd->obd_connected_clients,
1382 obd->obd_max_recoverable_clients);
1383 obd->obd_abort_recovery = 0;
1384 stale = class_disconnect_stale_exports(obd, connect_done, 0);
1385 atomic_sub(stale, &obd->obd_req_replay_clients);
1386 atomic_sub(stale, &obd->obd_lock_replay_clients);
1389 /* next stage: replay requests */
1391 obd->obd_req_replaying = 1;
1392 CDEBUG(D_ERROR, "1: request replay stage - %d clients from t"LPU64"\n",
1393 atomic_read(&obd->obd_req_replay_clients),
1394 obd->obd_next_recovery_transno);
1395 while ((req = target_next_replay_req(obd))) {
1396 LASSERT(trd->trd_processing_task == current->pid);
1397 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1398 req->rq_reqmsg->transno,
1399 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1400 (void)trd->trd_recovery_handler(req);
1401 obd->obd_replayed_requests++;
1402 reset_recovery_timer(obd);
1403 /* bug 1580: decide how to properly sync() in recovery*/
1404 //mds_fsync_super(mds->mds_sb);
1405 ptlrpc_free_clone(req);
1406 spin_lock_bh(&obd->obd_processing_task_lock);
1407 obd->obd_next_recovery_transno++;
1408 spin_unlock_bh(&obd->obd_processing_task_lock);
1411 spin_lock_bh(&obd->obd_processing_task_lock);
1412 target_cancel_recovery_timer(obd);
1413 spin_unlock_bh(&obd->obd_processing_task_lock);
1415 /* If some clients haven't replayed requests in time, evict them */
1416 if (obd->obd_abort_recovery) {
1418 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1419 obd->obd_abort_recovery = 0;
1420 stale = class_disconnect_stale_exports(obd, req_replay_done, 0);
1421 atomic_sub(stale, &obd->obd_lock_replay_clients);
1422 abort_req_replay_queue(obd);
1423 /* XXX for debuggin tests 11 and 17 */
1427 /* The second stage: replay locks */
1428 CDEBUG(D_ERROR, "2: lock replay stage - %d clients\n",
1429 atomic_read(&obd->obd_lock_replay_clients));
1430 while ((req = target_next_replay_lock(obd))) {
1431 LASSERT(trd->trd_processing_task == current->pid);
1432 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1433 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1434 (void)trd->trd_recovery_handler(req);
1435 reset_recovery_timer(obd);
1436 ptlrpc_free_clone(req);
1437 obd->obd_replayed_locks++;
1440 spin_lock_bh(&obd->obd_processing_task_lock);
1441 target_cancel_recovery_timer(obd);
1442 spin_unlock_bh(&obd->obd_processing_task_lock);
1444 /* If some clients haven't replayed requests in time, evict them */
1445 if (obd->obd_abort_recovery) {
1447 CERROR("lock replay timed out, aborting ...\n");
1448 obd->obd_abort_recovery = 0;
1449 stale = class_disconnect_stale_exports(obd, lock_replay_done, 0);
1450 abort_lock_replay_queue(obd);
1453 /* We drop recoverying flag to forward all new requests
1454 * to regular mds_handle() since now */
1455 spin_lock_bh(&obd->obd_processing_task_lock);
1456 obd->obd_recovering = 0;
1457 spin_unlock_bh(&obd->obd_processing_task_lock);
1459 /* The third stage: reply on final pings */
1460 CDEBUG(D_ERROR, "3: final stage - process recovery completion pings\n");
1461 while ((req = target_next_final_ping(obd))) {
1462 LASSERT(trd->trd_processing_task == current->pid);
1463 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1464 ptlrpc_peernid2str(&req->rq_peer, peer_str));
1465 (void)trd->trd_recovery_handler(req);
1466 ptlrpc_free_clone(req);
1469 delta = (jiffies - delta) / HZ;
1470 CDEBUG(D_ERROR,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1471 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1472 if (delta > obd_timeout * 2) {
1473 CWARN("too long recovery - read logs\n");
1474 portals_debug_dumplog();
1476 target_finish_recovery(obd);
1478 trd->trd_processing_task = 0;
1479 complete(&trd->trd_finishing);
1483 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1486 struct target_recovery_data *trd = &obd->obd_recovery_data;
1488 memset(trd, 0, sizeof(*trd));
1489 init_completion(&trd->trd_starting);
1490 init_completion(&trd->trd_finishing);
1491 trd->trd_recovery_handler = handler;
1493 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1494 wait_for_completion(&trd->trd_starting);
1495 LASSERT(obd->obd_recovering != 0);
1502 void target_stop_recovery_thread(struct obd_device *obd)
1504 spin_lock_bh(&obd->obd_processing_task_lock);
1505 if (obd->obd_recovery_data.trd_processing_task > 0) {
1506 struct target_recovery_data *trd = &obd->obd_recovery_data;
1507 CERROR("%s: aborting recovery\n", obd->obd_name);
1508 obd->obd_abort_recovery = 1;
1509 wake_up(&obd->obd_next_transno_waitq);
1510 spin_unlock_bh(&obd->obd_processing_task_lock);
1511 wait_for_completion(&trd->trd_finishing);
1513 spin_unlock_bh(&obd->obd_processing_task_lock);
1518 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1520 struct obd_export *exp = req->rq_export;
1521 LASSERT(exp != NULL);
1522 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1523 /* client declares he's ready to replay locks */
1524 spin_lock_bh(&obd->obd_processing_task_lock);
1525 if (exp->exp_req_replay_needed) {
1526 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1527 exp->exp_req_replay_needed = 0;
1528 atomic_dec(&obd->obd_req_replay_clients);
1529 obd->obd_recoverable_clients--;
1530 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1531 CDEBUG(D_HA, "all clients have replayed reqs\n");
1532 wake_up(&obd->obd_next_transno_waitq);
1534 spin_unlock_bh(&obd->obd_processing_task_lock);
1536 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1537 /* client declares he's ready to complete recovery
1538 * so, we put the request on th final queue */
1539 spin_lock_bh(&obd->obd_processing_task_lock);
1540 if (exp->exp_lock_replay_needed) {
1541 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1542 exp->exp_lock_replay_needed = 0;
1543 atomic_dec(&obd->obd_lock_replay_clients);
1544 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1545 CDEBUG(D_HA, "all clients have replayed locks\n");
1546 wake_up(&obd->obd_next_transno_waitq);
1548 spin_unlock_bh(&obd->obd_processing_task_lock);
1554 int target_queue_recovery_request(struct ptlrpc_request *req,
1555 struct obd_device *obd)
1557 struct list_head *tmp;
1559 __u64 transno = req->rq_reqmsg->transno;
1561 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
1562 /* Processing the queue right now, don't re-add. */
1566 target_process_req_flags(obd, req);
1568 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1569 /* client declares he's ready to complete recovery
1570 * so, we put the request on th final queue */
1571 req = ptlrpc_clone_req(req);
1574 DEBUG_REQ(D_HA, req, "queue final req");
1575 spin_lock_bh(&obd->obd_processing_task_lock);
1576 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1577 spin_unlock_bh(&obd->obd_processing_task_lock);
1580 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1581 /* client declares he's ready to replay locks */
1582 req = ptlrpc_clone_req(req);
1585 DEBUG_REQ(D_HA, req, "queue lock replay req");
1586 spin_lock_bh(&obd->obd_processing_task_lock);
1587 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1588 spin_unlock_bh(&obd->obd_processing_task_lock);
1589 wake_up(&obd->obd_next_transno_waitq);
1594 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1595 * (i.e. buflens etc are in my own byte order), but type-dependent
1596 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1599 INIT_LIST_HEAD(&req->rq_list);
1600 DEBUG_REQ(D_HA, req, "not queueing");
1605 /* If we're processing the queue, we want don't want to queue this
1608 * Also, if this request has a transno less than the one we're waiting
1609 * for, we should process it now. It could (and currently always will)
1610 * be an open request for a descriptor that was opened some time ago.
1612 * Also, a resent, replayed request that has already been
1613 * handled will pass through here and be processed immediately.
1615 spin_lock_bh(&obd->obd_processing_task_lock);
1616 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1617 /* Processing the queue right now, don't re-add. */
1618 LASSERT(list_empty(&req->rq_list));
1619 spin_unlock_bh(&obd->obd_processing_task_lock);
1622 spin_unlock_bh(&obd->obd_processing_task_lock);
1624 /* A resent, replayed request that is still on the queue; just drop it.
1625 The queued request will handle this. */
1626 if ((lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY))
1627 == (MSG_RESENT | MSG_REPLAY)) {
1628 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1632 req = ptlrpc_clone_req(req);
1636 spin_lock_bh(&obd->obd_processing_task_lock);
1639 list_for_each(tmp, &obd->obd_req_replay_queue) {
1640 struct ptlrpc_request *reqiter =
1641 list_entry(tmp, struct ptlrpc_request, rq_list);
1643 if (reqiter->rq_reqmsg->transno > transno) {
1644 list_add_tail(&req->rq_list, &reqiter->rq_list);
1651 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
1653 obd->obd_requests_queued_for_recovery++;
1654 wake_up(&obd->obd_next_transno_waitq);
1655 spin_unlock_bh(&obd->obd_processing_task_lock);
1659 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1661 return req->rq_export->exp_obd;
1665 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1667 if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1668 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1669 DEBUG_REQ(D_ERROR, req, "dropping reply");
1670 /* NB this does _not_ send with ACK disabled, to simulate
1671 * sending OK, but timing out for the ACK */
1672 if (req->rq_reply_state != NULL) {
1673 if (!req->rq_reply_state->rs_difficult) {
1674 lustre_free_reply_state (req->rq_reply_state);
1675 req->rq_reply_state = NULL;
1677 struct ptlrpc_service *svc =
1678 req->rq_rqbd->rqbd_srv_ni->sni_service;
1679 atomic_inc(&svc->srv_outstanding_replies);
1686 req->rq_status = rc;
1687 return (ptlrpc_error(req));
1689 DEBUG_REQ(D_NET, req, "sending reply");
1692 return (ptlrpc_send_reply(req, 1));
1696 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1699 unsigned long flags;
1700 struct ptlrpc_reply_state *rs;
1701 struct obd_device *obd;
1702 struct obd_export *exp;
1703 struct ptlrpc_srv_ni *sni;
1704 struct ptlrpc_service *svc;
1706 sni = req->rq_rqbd->rqbd_srv_ni;
1707 svc = sni->sni_service;
1709 rs = req->rq_reply_state;
1710 if (rs == NULL || !rs->rs_difficult) {
1711 /* The easy case; no notifiers and reply_out_callback()
1712 * cleans up (i.e. we can't look inside rs after a
1713 * successful send) */
1714 netrc = target_send_reply_msg (req, rc, fail_id);
1716 LASSERT (netrc == 0 || req->rq_reply_state == NULL);
1720 /* must be an export if locks saved */
1721 LASSERT (req->rq_export != NULL);
1722 /* req/reply consistent */
1723 LASSERT (rs->rs_srv_ni == sni);
1726 LASSERT (!rs->rs_scheduled);
1727 LASSERT (!rs->rs_scheduled_ever);
1728 LASSERT (!rs->rs_handled);
1729 LASSERT (!rs->rs_on_net);
1730 LASSERT (rs->rs_export == NULL);
1731 LASSERT (list_empty(&rs->rs_obd_list));
1732 LASSERT (list_empty(&rs->rs_exp_list));
1734 exp = class_export_get (req->rq_export);
1737 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1738 rs->rs_scheduled = 1;
1740 rs->rs_xid = req->rq_xid;
1741 rs->rs_transno = req->rq_transno;
1742 rs->rs_export = exp;
1744 spin_lock_irqsave (&obd->obd_uncommitted_replies_lock, flags);
1746 if (rs->rs_transno > obd->obd_last_committed) {
1747 /* not committed already */
1748 list_add_tail (&rs->rs_obd_list,
1749 &obd->obd_uncommitted_replies);
1752 spin_unlock (&obd->obd_uncommitted_replies_lock);
1753 spin_lock (&exp->exp_lock);
1755 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1757 spin_unlock_irqrestore (&exp->exp_lock, flags);
1759 netrc = target_send_reply_msg (req, rc, fail_id);
1761 spin_lock_irqsave (&svc->srv_lock, flags);
1763 svc->srv_n_difficult_replies++;
1765 if (netrc != 0) /* error sending: reply is off the net */
1768 if (!rs->rs_on_net || /* some notifier */
1769 list_empty(&rs->rs_exp_list) || /* completed already */
1770 list_empty(&rs->rs_obd_list)) {
1771 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
1772 wake_up (&svc->srv_waitq);
1774 list_add (&rs->rs_list, &sni->sni_active_replies);
1775 rs->rs_scheduled = 0; /* allow notifier to schedule */
1778 spin_unlock_irqrestore (&svc->srv_lock, flags);
1781 int target_handle_ping(struct ptlrpc_request *req)
1783 return lustre_pack_reply(req, 0, NULL, NULL);