1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_mds.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: if non-zero, move the selected to the list head
55 * @create: if zero, only search in existed connections
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 cfs_spin_lock(&imp->imp_lock);
84 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 cfs_list_del(&item->oic_item);
88 cfs_list_add(&item->oic_item,
90 item->oic_last_attempt = 0;
92 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? ", moved to head" : ""));
95 cfs_spin_unlock(&imp->imp_lock);
96 GOTO(out_free, rc = 0);
101 imp_conn->oic_conn = ptlrpc_conn;
102 imp_conn->oic_uuid = *uuid;
103 imp_conn->oic_last_attempt = 0;
105 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107 cfs_list_add_tail(&imp_conn->oic_item,
108 &imp->imp_conn_list);
109 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
110 imp, imp->imp_obd->obd_name, uuid->uuid,
111 (priority ? "head" : "tail"));
113 cfs_spin_unlock(&imp->imp_lock);
114 GOTO(out_free, rc = -ENOENT);
117 cfs_spin_unlock(&imp->imp_lock);
121 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 ptlrpc_connection_put(ptlrpc_conn);
127 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 return import_set_conn(imp, uuid, 1, 0);
132 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
135 return import_set_conn(imp, uuid, priority, 1);
138 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
140 struct obd_import_conn *imp_conn;
141 struct obd_export *dlmexp;
145 cfs_spin_lock(&imp->imp_lock);
146 if (cfs_list_empty(&imp->imp_conn_list)) {
147 LASSERT(!imp->imp_connection);
151 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
152 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
154 LASSERT(imp_conn->oic_conn);
156 /* is current conn? */
157 if (imp_conn == imp->imp_conn_current) {
158 LASSERT(imp_conn->oic_conn == imp->imp_connection);
160 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
161 imp->imp_state != LUSTRE_IMP_DISCON) {
162 CERROR("can't remove current connection\n");
163 GOTO(out, rc = -EBUSY);
166 ptlrpc_connection_put(imp->imp_connection);
167 imp->imp_connection = NULL;
169 dlmexp = class_conn2export(&imp->imp_dlm_handle);
170 if (dlmexp && dlmexp->exp_connection) {
171 LASSERT(dlmexp->exp_connection ==
173 ptlrpc_connection_put(dlmexp->exp_connection);
174 dlmexp->exp_connection = NULL;
178 cfs_list_del(&imp_conn->oic_item);
179 ptlrpc_connection_put(imp_conn->oic_conn);
180 OBD_FREE(imp_conn, sizeof(*imp_conn));
181 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182 imp, imp->imp_obd->obd_name, uuid->uuid);
187 cfs_spin_unlock(&imp->imp_lock);
189 CERROR("connection %s not found\n", uuid->uuid);
193 void client_destroy_import(struct obd_import *imp)
195 /* drop security policy instance after all rpc finished/aborted
196 * to let all busy contexts be released. */
197 class_import_get(imp);
198 class_destroy_import(imp);
199 sptlrpc_import_sec_put(imp);
200 class_import_put(imp);
203 /* configure an RPC client OBD device
208 * 3 - inactive-on-startup
210 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
212 struct client_obd *cli = &obddev->u.cli;
213 struct obd_import *imp;
214 struct obd_uuid server_uuid;
215 int rq_portal, rp_portal, connect_op;
216 char *name = obddev->obd_type->typ_name;
217 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
221 /* In a more perfect world, we would hang a ptlrpc_client off of
222 * obd_type and just use the values from there. */
223 if (!strcmp(name, LUSTRE_OSC_NAME)) {
224 rq_portal = OST_REQUEST_PORTAL;
225 rp_portal = OSC_REPLY_PORTAL;
226 connect_op = OST_CONNECT;
227 cli->cl_sp_me = LUSTRE_SP_CLI;
228 cli->cl_sp_to = LUSTRE_SP_OST;
229 ns_type = LDLM_NS_TYPE_OSC;
231 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
232 rq_portal = MDS_REQUEST_PORTAL;
233 rp_portal = MDC_REPLY_PORTAL;
234 connect_op = MDS_CONNECT;
235 cli->cl_sp_me = LUSTRE_SP_CLI;
236 cli->cl_sp_to = LUSTRE_SP_MDT;
237 ns_type = LDLM_NS_TYPE_MDC;
239 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
240 rq_portal = MGS_REQUEST_PORTAL;
241 rp_portal = MGC_REPLY_PORTAL;
242 connect_op = MGS_CONNECT;
243 cli->cl_sp_me = LUSTRE_SP_MGC;
244 cli->cl_sp_to = LUSTRE_SP_MGS;
245 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
246 ns_type = LDLM_NS_TYPE_MGC;
249 CERROR("unknown client OBD type \"%s\", can't setup\n",
254 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
255 CERROR("requires a TARGET UUID\n");
259 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
260 CERROR("client UUID must be less than 38 characters\n");
264 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
265 CERROR("setup requires a SERVER UUID\n");
269 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
270 CERROR("target UUID must be less than 38 characters\n");
274 cfs_init_rwsem(&cli->cl_sem);
275 cfs_sema_init(&cli->cl_mgc_sem, 1);
276 cli->cl_conn_count = 0;
277 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
278 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
279 sizeof(server_uuid)));
282 cli->cl_avail_grant = 0;
283 /* FIXME: should limit this for the sum of all cl_dirty_max */
284 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
285 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
286 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
287 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
288 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
289 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
290 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
291 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
292 client_obd_list_lock_init(&cli->cl_loi_list_lock);
293 cli->cl_r_in_flight = 0;
294 cli->cl_w_in_flight = 0;
296 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
297 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
298 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
299 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
300 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
301 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
302 cfs_waitq_init(&cli->cl_destroy_waitq);
303 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
304 #ifdef ENABLE_CHECKSUM
305 /* Turn on checksumming by default. */
306 cli->cl_checksum = 1;
308 * The supported checksum types will be worked out at connect time
309 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
312 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
314 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
316 /* This value may be changed at connect time in
317 ptlrpc_connect_interpret. */
318 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
319 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
321 if (!strcmp(name, LUSTRE_MDC_NAME)) {
322 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
323 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
324 cli->cl_max_rpcs_in_flight = 2;
325 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
326 cli->cl_max_rpcs_in_flight = 3;
327 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
328 cli->cl_max_rpcs_in_flight = 4;
330 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
335 CERROR("ldlm_get_ref failed: %d\n", rc);
339 ptlrpc_init_client(rq_portal, rp_portal, name,
340 &obddev->obd_ldlm_client);
342 imp = class_new_import(obddev);
344 GOTO(err_ldlm, rc = -ENOENT);
345 imp->imp_client = &obddev->obd_ldlm_client;
346 imp->imp_connect_op = connect_op;
347 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
348 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
349 LUSTRE_CFG_BUFLEN(lcfg, 1));
350 class_import_put(imp);
352 rc = client_import_add_conn(imp, &server_uuid, 1);
354 CERROR("can't add initial connection\n");
355 GOTO(err_import, rc);
358 cli->cl_import = imp;
359 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
360 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
361 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
363 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
364 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
365 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
366 name, obddev->obd_name,
367 cli->cl_target_uuid.uuid);
368 cfs_spin_lock(&imp->imp_lock);
369 imp->imp_deactive = 1;
370 cfs_spin_unlock(&imp->imp_lock);
374 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
375 LDLM_NAMESPACE_CLIENT,
376 LDLM_NAMESPACE_GREEDY,
378 if (obddev->obd_namespace == NULL) {
379 CERROR("Unable to create client namespace - %s\n",
381 GOTO(err_import, rc = -ENOMEM);
384 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
389 class_destroy_import(imp);
397 int client_obd_cleanup(struct obd_device *obddev)
401 ldlm_namespace_free_post(obddev->obd_namespace);
402 obddev->obd_namespace = NULL;
408 /* ->o_connect() method for client side (OSC and MDC and MGC) */
409 int client_connect_import(const struct lu_env *env,
410 struct obd_export **exp,
411 struct obd_device *obd, struct obd_uuid *cluuid,
412 struct obd_connect_data *data, void *localdata)
414 struct client_obd *cli = &obd->u.cli;
415 struct obd_import *imp = cli->cl_import;
416 struct obd_connect_data *ocd;
417 struct lustre_handle conn = { 0 };
422 cfs_down_write(&cli->cl_sem);
423 if (cli->cl_conn_count > 0 )
424 GOTO(out_sem, rc = -EALREADY);
426 rc = class_connect(&conn, obd, cluuid);
430 cli->cl_conn_count++;
431 *exp = class_conn2export(&conn);
433 LASSERT(obd->obd_namespace);
435 imp->imp_dlm_handle = conn;
436 rc = ptlrpc_init_import(imp);
440 ocd = &imp->imp_connect_data;
443 imp->imp_connect_flags_orig = data->ocd_connect_flags;
446 rc = ptlrpc_connect_import(imp, NULL);
448 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
451 LASSERT((*exp)->exp_connection);
454 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
455 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
456 data->ocd_connect_flags, ocd->ocd_connect_flags);
457 data->ocd_connect_flags = ocd->ocd_connect_flags;
460 ptlrpc_pinger_add_import(imp);
466 cli->cl_conn_count--;
467 class_disconnect(*exp);
471 cfs_up_write(&cli->cl_sem);
476 int client_disconnect_export(struct obd_export *exp)
478 struct obd_device *obd = class_exp2obd(exp);
479 struct client_obd *cli;
480 struct obd_import *imp;
485 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
486 exp, exp ? exp->exp_handle.h_cookie : -1);
491 imp = cli->cl_import;
493 cfs_down_write(&cli->cl_sem);
494 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
497 if (!cli->cl_conn_count) {
498 CERROR("disconnecting disconnected device (%s)\n",
500 GOTO(out_disconnect, rc = -EINVAL);
503 cli->cl_conn_count--;
504 if (cli->cl_conn_count)
505 GOTO(out_disconnect, rc = 0);
507 /* Mark import deactivated now, so we don't try to reconnect if any
508 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
509 * fully deactivate the import, or that would drop all requests. */
510 cfs_spin_lock(&imp->imp_lock);
511 imp->imp_deactive = 1;
512 cfs_spin_unlock(&imp->imp_lock);
514 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
515 * delete it regardless. (It's safe to delete an import that was
517 (void)ptlrpc_pinger_del_import(imp);
519 if (obd->obd_namespace != NULL) {
520 /* obd_force == local only */
521 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
522 obd->obd_force ? LCF_LOCAL : 0, NULL);
523 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
527 * there's no need to hold sem during disconnecting an import,
528 * and actually it may cause deadlock in gss.
530 cfs_up_write(&cli->cl_sem);
531 rc = ptlrpc_disconnect_import(imp, 0);
532 cfs_down_write(&cli->cl_sem);
534 ptlrpc_invalidate_import(imp);
536 if (imp->imp_rq_pool) {
537 ptlrpc_free_rq_pool(imp->imp_rq_pool);
538 imp->imp_rq_pool = NULL;
540 client_destroy_import(imp);
541 cli->cl_import = NULL;
546 /* use server style - class_disconnect should be always called for
548 err = class_disconnect(exp);
552 cfs_up_write(&cli->cl_sem);
557 int server_disconnect_export(struct obd_export *exp)
562 /* Disconnect early so that clients can't keep using export */
563 rc = class_disconnect(exp);
564 /* close import for avoid sending any requests */
565 if (exp->exp_imp_reverse)
566 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
568 if (exp->exp_obd->obd_namespace != NULL)
569 ldlm_cancel_locks_for_export(exp);
571 /* complete all outstanding replies */
572 cfs_spin_lock(&exp->exp_lock);
573 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
574 struct ptlrpc_reply_state *rs =
575 cfs_list_entry(exp->exp_outstanding_replies.next,
576 struct ptlrpc_reply_state, rs_exp_list);
577 struct ptlrpc_service *svc = rs->rs_service;
579 cfs_spin_lock(&svc->srv_rs_lock);
580 cfs_list_del_init(&rs->rs_exp_list);
581 cfs_spin_lock(&rs->rs_lock);
582 ptlrpc_schedule_difficult_reply(rs);
583 cfs_spin_unlock(&rs->rs_lock);
584 cfs_spin_unlock(&svc->srv_rs_lock);
586 cfs_spin_unlock(&exp->exp_lock);
591 /* --------------------------------------------------------------------------
592 * from old lib/target.c
593 * -------------------------------------------------------------------------- */
595 static int target_handle_reconnect(struct lustre_handle *conn,
596 struct obd_export *exp,
597 struct obd_uuid *cluuid)
600 if (exp->exp_connection && exp->exp_imp_reverse) {
601 struct lustre_handle *hdl;
602 hdl = &exp->exp_imp_reverse->imp_remote_handle;
603 /* Might be a re-connect after a partition. */
604 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
605 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
607 conn->cookie = exp->exp_handle.h_cookie;
608 /* target_handle_connect() treats EALREADY and
609 * -EALREADY differently. EALREADY means we are
610 * doing a valid reconnect from the same client. */
613 CERROR("%s reconnecting from %s, "
614 "handle mismatch (ours "LPX64", theirs "
615 LPX64")\n", cluuid->uuid,
616 exp->exp_connection->c_remote_uuid.uuid,
617 hdl->cookie, conn->cookie);
618 memset(conn, 0, sizeof *conn);
619 /* target_handle_connect() treats EALREADY and
620 * -EALREADY differently. -EALREADY is an error
621 * (same UUID, different handle). */
626 conn->cookie = exp->exp_handle.h_cookie;
627 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
628 cluuid->uuid, exp, conn->cookie);
632 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
635 struct obd_export *exp = cb_data;
637 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
638 obd->obd_name, exp->exp_client_uuid.uuid);
640 cfs_spin_lock(&exp->exp_lock);
641 exp->exp_need_sync = 0;
642 cfs_spin_unlock(&exp->exp_lock);
643 class_export_cb_put(exp);
645 EXPORT_SYMBOL(target_client_add_cb);
648 target_start_and_reset_recovery_timer(struct obd_device *obd,
649 struct ptlrpc_request *req,
652 int target_handle_connect(struct ptlrpc_request *req)
654 struct obd_device *target, *targref = NULL;
655 struct obd_export *export = NULL;
656 struct obd_import *revimp;
657 struct lustre_handle conn;
658 struct lustre_handle *tmp;
659 struct obd_uuid tgtuuid;
660 struct obd_uuid cluuid;
661 struct obd_uuid remote_uuid;
665 struct obd_connect_data *data, *tmpdata;
666 lnet_nid_t *client_nid = NULL;
669 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
671 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
673 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
674 GOTO(out, rc = -EINVAL);
677 obd_str2uuid(&tgtuuid, str);
678 target = class_uuid2obd(&tgtuuid);
680 target = class_name2obd(str);
682 if (!target || target->obd_stopping || !target->obd_set_up) {
683 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
684 " for connect (%s)\n", str,
685 !target ? "no target" :
686 (target->obd_stopping ? "stopping" :
688 GOTO(out, rc = -ENODEV);
691 if (target->obd_no_conn) {
692 LCONSOLE_WARN("%s: temporarily refusing client connection "
693 "from %s\n", target->obd_name,
694 libcfs_nid2str(req->rq_peer.nid));
695 GOTO(out, rc = -EAGAIN);
698 /* Make sure the target isn't cleaned up while we're here. Yes,
699 there's still a race between the above check and our incref here.
700 Really, class_uuid2obd should take the ref. */
701 targref = class_incref(target, __FUNCTION__, cfs_current());
704 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
706 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
707 GOTO(out, rc = -EINVAL);
710 obd_str2uuid(&cluuid, str);
712 /* XXX extract a nettype and format accordingly */
713 switch (sizeof(lnet_nid_t)) {
714 /* NB the casts only avoid compiler warnings */
716 snprintf(remote_uuid.uuid, sizeof remote_uuid,
717 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
720 snprintf(remote_uuid.uuid, sizeof remote_uuid,
721 "NET_%x_UUID", (__u32)req->rq_peer.nid);
727 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
729 GOTO(out, rc = -EPROTO);
733 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
735 GOTO(out, rc = -EPROTO);
737 rc = req_capsule_server_pack(&req->rq_pill);
741 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
743 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
744 "libclient connection attempt");
745 GOTO(out, rc = -EPROTO);
746 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
747 LUSTRE_VERSION_ALLOWED_OFFSET ||
748 data->ocd_version > LUSTRE_VERSION_CODE +
749 LUSTRE_VERSION_ALLOWED_OFFSET) {
750 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
751 "libclient connection attempt",
752 data->ocd_version < LUSTRE_VERSION_CODE ?
754 OBD_OCD_VERSION_MAJOR(data->ocd_version),
755 OBD_OCD_VERSION_MINOR(data->ocd_version),
756 OBD_OCD_VERSION_PATCH(data->ocd_version),
757 OBD_OCD_VERSION_FIX(data->ocd_version));
758 data = req_capsule_server_sized_get(&req->rq_pill,
760 offsetof(typeof(*data), ocd_version) +
761 sizeof(data->ocd_version));
763 data->ocd_connect_flags = OBD_CONNECT_VERSION;
764 data->ocd_version = LUSTRE_VERSION_CODE;
766 GOTO(out, rc = -EPROTO);
770 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
771 (data->ocd_connect_flags & OBD_CONNECT_MDS))
774 /* lctl gets a backstage, all-access pass. */
775 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
776 goto dont_check_exports;
778 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
782 /* we've found an export in the hash */
783 if (export->exp_connecting) { /* bug 9635, et. al. */
784 CWARN("%s: exp %p already connecting\n",
785 export->exp_obd->obd_name, export);
786 class_export_put(export);
789 } else if (mds_conn && export->exp_connection) {
790 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
791 /* mds reconnected after failover */
792 CWARN("%s: received MDS connection from NID %s,"
793 " removing former export from NID %s\n",
794 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
795 libcfs_nid2str(export->exp_connection->c_peer.nid));
797 /* new mds connection from the same nid */
798 CWARN("%s: received new MDS connection from NID %s,"
799 " removing former export from same NID\n",
800 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
801 class_fail_export(export);
802 class_export_put(export);
805 } else if (export->exp_connection != NULL &&
806 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
807 (lustre_msg_get_op_flags(req->rq_reqmsg) &
808 MSG_CONNECT_INITIAL)) {
809 /* in mds failover we have static uuid but nid can be
811 CWARN("%s: cookie %s seen on new NID %s when "
812 "existing NID %s is already connected\n",
813 target->obd_name, cluuid.uuid,
814 libcfs_nid2str(req->rq_peer.nid),
815 libcfs_nid2str(export->exp_connection->c_peer.nid));
817 class_export_put(export);
820 cfs_spin_lock(&export->exp_lock);
821 export->exp_connecting = 1;
822 cfs_spin_unlock(&export->exp_lock);
823 class_export_put(export);
824 LASSERT(export->exp_obd == target);
826 rc = target_handle_reconnect(&conn, export, &cluuid);
829 /* If we found an export, we already unlocked. */
832 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
833 } else if (req->rq_export == NULL &&
834 cfs_atomic_read(&export->exp_rpc_count) > 0) {
835 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
836 target->obd_name, cluuid.uuid,
837 libcfs_nid2str(req->rq_peer.nid),
838 export, cfs_atomic_read(&export->exp_refcount));
839 GOTO(out, rc = -EBUSY);
840 } else if (req->rq_export != NULL &&
841 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
842 /* the current connect rpc has increased exp_rpc_count */
843 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
844 target->obd_name, cluuid.uuid,
845 libcfs_nid2str(req->rq_peer.nid),
846 export, cfs_atomic_read(&export->exp_rpc_count) - 1);
847 cfs_spin_lock(&export->exp_lock);
848 if (req->rq_export->exp_conn_cnt <
849 lustre_msg_get_conn_cnt(req->rq_reqmsg))
850 /* try to abort active requests */
851 req->rq_export->exp_abort_active_req = 1;
852 cfs_spin_unlock(&export->exp_lock);
853 GOTO(out, rc = -EBUSY);
854 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
855 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
856 "cookies not random?\n", target->obd_name,
857 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
858 GOTO(out, rc = -EALREADY);
860 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
867 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
868 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
869 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
870 export, (long)cfs_time_current_sec(),
871 export ? (long)export->exp_last_request_time : 0);
873 /* Tell the client if we're in recovery. */
874 if (target->obd_recovering) {
875 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
876 /* If this is the first time a client connects,
877 reset the recovery timer */
879 target_start_and_reset_recovery_timer(target, req,
883 /* We want to handle EALREADY but *not* -EALREADY from
884 * target_handle_reconnect(), return reconnection state in a flag */
885 if (rc == EALREADY) {
886 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
892 /* Tell the client if we support replayable requests */
893 if (target->obd_replayable)
894 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
895 client_nid = &req->rq_peer.nid;
897 if (export == NULL) {
898 if (target->obd_recovering) {
901 t = cfs_timer_deadline(&target->obd_recovery_timer);
902 t = cfs_time_sub(t, cfs_time_current());
903 CERROR("%s: denying connection for new client %s (%s): "
904 "%d clients in recovery for "CFS_TIME_T"s\n",
906 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
907 cfs_atomic_read(&target-> \
908 obd_lock_replay_clients),
909 cfs_duration_sec(t));
913 rc = obd_connect(req->rq_svc_thread->t_env,
914 &export, target, &cluuid, data,
917 conn.cookie = export->exp_handle.h_cookie;
920 rc = obd_reconnect(req->rq_svc_thread->t_env,
921 export, target, &cluuid, data, client_nid);
923 /* prevous done via class_conn2export */
924 class_export_get(export);
928 /* Return only the parts of obd_connect_data that we understand, so the
929 * client knows that we don't understand the rest. */
931 tmpdata = req_capsule_server_get(&req->rq_pill,
933 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
937 /* If all else goes well, this is our RPC return code. */
940 lustre_msg_set_handle(req->rq_repmsg, &conn);
942 /* If the client and the server are the same node, we will already
943 * have an export that really points to the client's DLM export,
944 * because we have a shared handles table.
946 * XXX this will go away when shaver stops sending the "connect" handle
947 * in the real "remote handle" field of the request --phik 24 Apr 2003
949 if (req->rq_export != NULL)
950 class_export_put(req->rq_export);
952 req->rq_export = export;
954 cfs_spin_lock(&export->exp_lock);
955 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
956 cfs_spin_unlock(&export->exp_lock);
957 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
958 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
959 export->exp_conn_cnt,
960 lustre_msg_get_conn_cnt(req->rq_reqmsg));
962 GOTO(out, rc = -EALREADY);
964 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
965 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
966 export->exp_abort_active_req = 0;
968 /* request from liblustre? Don't evict it for not pinging. */
969 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
970 export->exp_libclient = 1;
971 cfs_spin_unlock(&export->exp_lock);
973 cfs_spin_lock(&target->obd_dev_lock);
974 cfs_list_del_init(&export->exp_obd_chain_timed);
975 cfs_spin_unlock(&target->obd_dev_lock);
977 cfs_spin_unlock(&export->exp_lock);
980 if (export->exp_connection != NULL) {
981 /* Check to see if connection came from another NID */
982 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
983 !cfs_hlist_unhashed(&export->exp_nid_hash))
984 cfs_hash_del(export->exp_obd->obd_nid_hash,
985 &export->exp_connection->c_peer.nid,
986 &export->exp_nid_hash);
988 ptlrpc_connection_put(export->exp_connection);
991 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
994 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
995 cfs_hash_add(export->exp_obd->obd_nid_hash,
996 &export->exp_connection->c_peer.nid,
997 &export->exp_nid_hash);
1000 cfs_spin_lock(&target->obd_recovery_task_lock);
1001 if (target->obd_recovering && !export->exp_in_recovery) {
1002 cfs_spin_lock(&export->exp_lock);
1003 export->exp_in_recovery = 1;
1004 export->exp_req_replay_needed = 1;
1005 export->exp_lock_replay_needed = 1;
1006 cfs_spin_unlock(&export->exp_lock);
1007 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1008 && (data->ocd_transno == 0))
1009 CWARN("Connect with zero transno!\n");
1011 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1012 && data->ocd_transno < target->obd_next_recovery_transno)
1013 target->obd_next_recovery_transno = data->ocd_transno;
1014 target->obd_connected_clients++;
1015 cfs_atomic_inc(&target->obd_req_replay_clients);
1016 cfs_atomic_inc(&target->obd_lock_replay_clients);
1017 if (target->obd_connected_clients ==
1018 target->obd_max_recoverable_clients)
1019 cfs_waitq_signal(&target->obd_next_transno_waitq);
1021 cfs_spin_unlock(&target->obd_recovery_task_lock);
1022 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1025 if (export->exp_imp_reverse != NULL) {
1026 /* destroyed import can be still referenced in ctxt */
1027 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1028 KEY_REVIMP_UPD, 0, NULL, NULL);
1030 client_destroy_import(export->exp_imp_reverse);
1033 /* for the rest part, we return -ENOTCONN in case of errors
1034 * in order to let client initialize connection again.
1036 revimp = export->exp_imp_reverse = class_new_import(target);
1038 CERROR("fail to alloc new reverse import.\n");
1039 GOTO(out, rc = -ENOTCONN);
1042 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1043 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1044 revimp->imp_remote_handle = conn;
1045 revimp->imp_dlm_fake = 1;
1046 revimp->imp_state = LUSTRE_IMP_FULL;
1048 /* unknown versions will be caught in
1049 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1050 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1052 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1053 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1054 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1056 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1058 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1059 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1060 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1062 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1064 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1066 CERROR("Failed to get sec for reverse import: %d\n", rc);
1067 export->exp_imp_reverse = NULL;
1068 class_destroy_import(revimp);
1071 class_import_put(revimp);
1074 cfs_spin_lock(&export->exp_lock);
1075 export->exp_connecting = 0;
1076 cfs_spin_unlock(&export->exp_lock);
1079 class_decref(targref, __FUNCTION__, cfs_current());
1081 req->rq_status = rc;
1085 int target_handle_disconnect(struct ptlrpc_request *req)
1090 rc = req_capsule_server_pack(&req->rq_pill);
1094 /* keep the rq_export around so we can send the reply */
1095 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1100 void target_destroy_export(struct obd_export *exp)
1102 /* exports created from last_rcvd data, and "fake"
1103 exports created by lctl don't have an import */
1104 if (exp->exp_imp_reverse != NULL)
1105 client_destroy_import(exp->exp_imp_reverse);
1107 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1108 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1109 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1110 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1114 * Recovery functions
1116 static void target_request_copy_get(struct ptlrpc_request *req)
1118 class_export_rpc_get(req->rq_export);
1119 LASSERT(cfs_list_empty(&req->rq_list));
1120 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1122 /* increase refcount to keep request in queue */
1123 cfs_atomic_inc(&req->rq_refcount);
1124 /** let export know it has replays to be handled */
1125 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1126 /* release service thread while request is queued
1127 * we are moving the request from active processing
1128 * to waiting on the replay queue */
1129 ptlrpc_server_active_request_dec(req);
1132 static void target_request_copy_put(struct ptlrpc_request *req)
1134 LASSERT(cfs_list_empty(&req->rq_replay_list));
1135 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1137 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1138 class_export_rpc_put(req->rq_export);
1139 /* ptlrpc_server_drop_request() assumes the request is active */
1140 ptlrpc_server_active_request_inc(req);
1141 ptlrpc_server_drop_request(req);
1144 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1146 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1147 struct obd_export *exp = req->rq_export;
1148 struct ptlrpc_request *reqiter;
1153 cfs_spin_lock(&exp->exp_lock);
1154 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1156 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1163 /* we expect it with RESENT and REPLAY flags */
1164 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1165 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1166 CERROR("invalid flags %x of resent replay\n",
1167 lustre_msg_get_flags(req->rq_reqmsg));
1169 cfs_list_add_tail(&req->rq_replay_list,
1170 &exp->exp_req_replay_queue);
1173 cfs_spin_unlock(&exp->exp_lock);
1177 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1179 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1180 LASSERT(req->rq_export);
1182 cfs_spin_lock(&req->rq_export->exp_lock);
1183 cfs_list_del_init(&req->rq_replay_list);
1184 cfs_spin_unlock(&req->rq_export->exp_lock);
1188 static void target_finish_recovery(struct obd_device *obd)
1191 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1194 ldlm_reprocess_all_ns(obd->obd_namespace);
1195 cfs_spin_lock(&obd->obd_recovery_task_lock);
1196 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1197 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1198 !cfs_list_empty(&obd->obd_final_req_queue)) {
1199 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1201 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1202 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1204 cfs_list_empty(&obd->obd_final_req_queue) ? \
1206 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1209 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1211 obd->obd_recovery_end = cfs_time_current_sec();
1213 /* when recovery finished, cleanup orphans on mds and ost */
1214 if (OBT(obd) && OBP(obd, postrecov)) {
1215 int rc = OBP(obd, postrecov)(obd);
1217 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1223 static void abort_req_replay_queue(struct obd_device *obd)
1225 struct ptlrpc_request *req, *n;
1226 cfs_list_t abort_list;
1228 CFS_INIT_LIST_HEAD(&abort_list);
1229 cfs_spin_lock(&obd->obd_recovery_task_lock);
1230 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1231 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1232 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1233 DEBUG_REQ(D_WARNING, req, "aborted:");
1234 req->rq_status = -ENOTCONN;
1235 if (ptlrpc_error(req)) {
1236 DEBUG_REQ(D_ERROR, req,
1237 "failed abort_req_reply; skipping");
1239 target_exp_dequeue_req_replay(req);
1240 target_request_copy_put(req);
1244 static void abort_lock_replay_queue(struct obd_device *obd)
1246 struct ptlrpc_request *req, *n;
1247 cfs_list_t abort_list;
1249 CFS_INIT_LIST_HEAD(&abort_list);
1250 cfs_spin_lock(&obd->obd_recovery_task_lock);
1251 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1252 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1253 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1254 DEBUG_REQ(D_ERROR, req, "aborted:");
1255 req->rq_status = -ENOTCONN;
1256 if (ptlrpc_error(req)) {
1257 DEBUG_REQ(D_ERROR, req,
1258 "failed abort_lock_reply; skipping");
1260 target_request_copy_put(req);
1265 /* Called from a cleanup function if the device is being cleaned up
1266 forcefully. The exports should all have been disconnected already,
1267 the only thing left to do is
1268 - clear the recovery flags
1270 - free queued requests and replies, but don't send replies
1271 Because the obd_stopping flag is set, no new requests should be received.
1274 void target_cleanup_recovery(struct obd_device *obd)
1276 struct ptlrpc_request *req, *n;
1277 cfs_list_t clean_list;
1280 CFS_INIT_LIST_HEAD(&clean_list);
1281 cfs_spin_lock(&obd->obd_dev_lock);
1282 if (!obd->obd_recovering) {
1283 cfs_spin_unlock(&obd->obd_dev_lock);
1287 obd->obd_recovering = obd->obd_abort_recovery = 0;
1288 cfs_spin_unlock(&obd->obd_dev_lock);
1290 cfs_spin_lock(&obd->obd_recovery_task_lock);
1291 target_cancel_recovery_timer(obd);
1292 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1293 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1295 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1296 LASSERT(req->rq_reply_state == 0);
1297 target_exp_dequeue_req_replay(req);
1298 target_request_copy_put(req);
1301 cfs_spin_lock(&obd->obd_recovery_task_lock);
1302 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1303 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1304 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1306 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1307 LASSERT(req->rq_reply_state == 0);
1308 target_request_copy_put(req);
1314 /* obd_recovery_task_lock should be held */
1315 void target_cancel_recovery_timer(struct obd_device *obd)
1317 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1318 cfs_timer_disarm(&obd->obd_recovery_timer);
1321 /* extend = 1 means require at least "duration" seconds left in the timer,
1322 extend = 0 means set the total duration (start_recovery_timer) */
1323 static void reset_recovery_timer(struct obd_device *obd, int duration,
1326 cfs_time_t now = cfs_time_current_sec();
1327 cfs_duration_t left;
1329 cfs_spin_lock(&obd->obd_recovery_task_lock);
1330 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1331 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1335 left = cfs_time_sub(obd->obd_recovery_end, now);
1337 if (extend && (duration > left))
1338 obd->obd_recovery_timeout += duration - left;
1339 else if (!extend && (duration > obd->obd_recovery_timeout))
1340 /* Track the client's largest expected replay time */
1341 obd->obd_recovery_timeout = duration;
1343 /* Hard limit of obd_recovery_time_hard which should not happen */
1344 if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1345 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1347 obd->obd_recovery_end = obd->obd_recovery_start +
1348 obd->obd_recovery_timeout;
1349 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1350 cfs_time_before(now, obd->obd_recovery_end)) {
1351 left = cfs_time_sub(obd->obd_recovery_end, now);
1352 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1354 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1355 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1356 obd->obd_name, (unsigned)left);
1359 static void check_and_start_recovery_timer(struct obd_device *obd)
1361 cfs_spin_lock(&obd->obd_recovery_task_lock);
1362 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1363 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1366 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1367 obd->obd_recovery_start = cfs_time_current_sec();
1368 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1370 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1373 /* Reset the timer with each new client connection */
1375 * This timer is actually reconnect_timer, which is for making sure
1376 * the total recovery window is at least as big as my reconnect
1377 * attempt timing. So the initial recovery time_out will be set to
1378 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1379 * from client is bigger than this, then the recovery time_out will
1380 * be extended to make sure the client could be reconnected, in the
1381 * process, the timeout from the new client should be ignored.
1385 target_start_and_reset_recovery_timer(struct obd_device *obd,
1386 struct ptlrpc_request *req,
1389 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1391 if (!new_client && service_time)
1392 /* Teach server about old server's estimates, as first guess
1393 * at how long new requests will take. */
1394 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1397 check_and_start_recovery_timer(obd);
1399 /* convert the service time to rpc timeout,
1400 * reuse service_time to limit stack usage */
1401 service_time = at_est2timeout(service_time);
1403 /* We expect other clients to timeout within service_time, then try
1404 * to reconnect, then try the failover server. The max delay between
1405 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1406 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1407 INITIAL_CONNECT_TIMEOUT);
1408 if (service_time > obd->obd_recovery_timeout && !new_client)
1409 reset_recovery_timer(obd, service_time, 0);
1414 /** Health checking routines */
1415 static inline int exp_connect_healthy(struct obd_export *exp)
1417 return (exp->exp_in_recovery);
1420 /** if export done req_replay or has replay in queue */
1421 static inline int exp_req_replay_healthy(struct obd_export *exp)
1423 return (!exp->exp_req_replay_needed ||
1424 cfs_atomic_read(&exp->exp_replay_count) > 0);
1426 /** if export done lock_replay or has replay in queue */
1427 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1429 return (!exp->exp_lock_replay_needed ||
1430 cfs_atomic_read(&exp->exp_replay_count) > 0);
1433 static inline int exp_vbr_healthy(struct obd_export *exp)
1435 return (!exp->exp_vbr_failed);
1438 static inline int exp_finished(struct obd_export *exp)
1440 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1443 /** Checking routines for recovery */
1444 static int check_for_clients(struct obd_device *obd)
1446 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1448 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1449 if (obd->obd_no_conn == 0 &&
1450 obd->obd_connected_clients + obd->obd_stale_clients ==
1451 obd->obd_max_recoverable_clients)
1456 static int check_for_next_transno(struct obd_device *obd)
1458 struct ptlrpc_request *req = NULL;
1459 int wake_up = 0, connected, completed, queue_len;
1460 __u64 next_transno, req_transno;
1463 cfs_spin_lock(&obd->obd_recovery_task_lock);
1464 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1465 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1466 struct ptlrpc_request, rq_list);
1467 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1472 connected = obd->obd_connected_clients;
1473 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1474 queue_len = obd->obd_requests_queued_for_recovery;
1475 next_transno = obd->obd_next_recovery_transno;
1477 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1478 "req_transno: "LPU64", next_transno: "LPU64"\n",
1479 obd->obd_max_recoverable_clients, connected, completed,
1480 queue_len, req_transno, next_transno);
1482 if (obd->obd_abort_recovery) {
1483 CDEBUG(D_HA, "waking for aborted recovery\n");
1485 } else if (obd->obd_recovery_expired) {
1486 CDEBUG(D_HA, "waking for expired recovery\n");
1488 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1489 CDEBUG(D_HA, "waking for completed recovery\n");
1491 } else if (req_transno == next_transno) {
1492 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1494 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1496 /** handle gaps occured due to lost reply or VBR */
1497 LASSERTF(req_transno >= next_transno,
1498 "req_transno: "LPU64", next_transno: "LPU64"\n",
1499 req_transno, next_transno);
1500 if (req_transno > obd->obd_last_committed &&
1501 !obd->obd_version_recov)
1504 "%s: waking for gap in transno, VBR is %s (skip: "
1505 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1506 ", last_committed: "LPD64")\n",
1507 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1508 next_transno, queue_len, completed, connected,
1509 req_transno, obd->obd_last_committed);
1510 obd->obd_next_recovery_transno = req_transno;
1512 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1513 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1514 " by fail_lock, waking up ("LPD64")\n", next_transno);
1515 obd->obd_next_recovery_transno = req_transno;
1518 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1522 static int check_for_next_lock(struct obd_device *obd)
1526 cfs_spin_lock(&obd->obd_recovery_task_lock);
1527 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1528 CDEBUG(D_HA, "waking for next lock\n");
1530 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1531 CDEBUG(D_HA, "waking for completed lock replay\n");
1533 } else if (obd->obd_abort_recovery) {
1534 CDEBUG(D_HA, "waking for aborted recovery\n");
1536 } else if (obd->obd_recovery_expired) {
1537 CDEBUG(D_HA, "waking for expired recovery\n");
1540 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1546 * wait for recovery events,
1547 * check its status with help of check_routine
1548 * evict dead clients via health_check
1550 static int target_recovery_overseer(struct obd_device *obd,
1551 int (*check_routine)(struct obd_device *),
1552 int (*health_check)(struct obd_export *))
1555 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1556 if (obd->obd_abort_recovery) {
1557 CWARN("recovery is aborted, evict exports in recovery\n");
1558 /** evict exports which didn't finish recovery yet */
1559 class_disconnect_stale_exports(obd, exp_finished);
1561 } else if (obd->obd_recovery_expired) {
1562 obd->obd_recovery_expired = 0;
1563 /** If some clients died being recovered, evict them */
1565 "recovery is timed out, evict stale exports\n");
1566 /** evict cexports with no replay in queue, they are stalled */
1567 class_disconnect_stale_exports(obd, health_check);
1568 /** continue with VBR */
1569 cfs_spin_lock(&obd->obd_dev_lock);
1570 obd->obd_version_recov = 1;
1571 cfs_spin_unlock(&obd->obd_dev_lock);
1573 * reset timer, recovery will proceed with versions now,
1574 * timeout is set just to handle reconnection delays
1576 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1577 /** Wait for recovery events again, after evicting bad clients */
1583 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1585 struct ptlrpc_request *req = NULL;
1588 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1589 obd->obd_next_recovery_transno);
1591 if (target_recovery_overseer(obd, check_for_next_transno,
1592 exp_req_replay_healthy)) {
1593 abort_req_replay_queue(obd);
1594 abort_lock_replay_queue(obd);
1597 cfs_spin_lock(&obd->obd_recovery_task_lock);
1598 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1599 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1600 struct ptlrpc_request, rq_list);
1601 cfs_list_del_init(&req->rq_list);
1602 obd->obd_requests_queued_for_recovery--;
1603 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1605 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1606 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1607 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1608 /** evict exports failed VBR */
1609 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1614 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1616 struct ptlrpc_request *req = NULL;
1618 CDEBUG(D_HA, "Waiting for lock\n");
1619 if (target_recovery_overseer(obd, check_for_next_lock,
1620 exp_lock_replay_healthy))
1621 abort_lock_replay_queue(obd);
1623 cfs_spin_lock(&obd->obd_recovery_task_lock);
1624 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1625 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1626 struct ptlrpc_request, rq_list);
1627 cfs_list_del_init(&req->rq_list);
1628 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1630 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1631 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1632 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1633 /** evict exports failed VBR */
1634 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1639 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1641 struct ptlrpc_request *req = NULL;
1643 cfs_spin_lock(&obd->obd_recovery_task_lock);
1644 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1645 req = cfs_list_entry(obd->obd_final_req_queue.next,
1646 struct ptlrpc_request, rq_list);
1647 cfs_list_del_init(&req->rq_list);
1648 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1649 if (req->rq_export->exp_in_recovery) {
1650 cfs_spin_lock(&req->rq_export->exp_lock);
1651 req->rq_export->exp_in_recovery = 0;
1652 cfs_spin_unlock(&req->rq_export->exp_lock);
1655 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1660 static int handle_recovery_req(struct ptlrpc_thread *thread,
1661 struct ptlrpc_request *req,
1662 svc_handler_t handler)
1668 * export can be evicted during recovery, no need to handle replays for
1669 * it after that, discard such request silently
1671 if (req->rq_export->exp_disconnected)
1672 GOTO(reqcopy_put, rc = 0);
1674 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1676 CERROR("Failure to initialize session: %d\n", rc);
1677 GOTO(reqcopy_put, rc);
1680 req->rq_recov_session.lc_thread = thread;
1681 lu_context_enter(&req->rq_recov_session);
1682 req->rq_svc_thread = thread;
1683 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1685 /* thread context */
1686 lu_context_enter(&thread->t_env->le_ctx);
1688 lu_context_exit(&thread->t_env->le_ctx);
1690 lu_context_exit(&req->rq_recov_session);
1691 lu_context_fini(&req->rq_recov_session);
1692 /* don't reset timer for final stage */
1693 if (!exp_finished(req->rq_export)) {
1695 * Add request timeout to the recovery time so next request from
1696 * this client may come in recovery time
1698 reset_recovery_timer(class_exp2obd(req->rq_export),
1699 AT_OFF ? obd_timeout :
1700 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1706 static int target_recovery_thread(void *arg)
1708 struct lu_target *lut = arg;
1709 struct obd_device *obd = lut->lut_obd;
1710 struct ptlrpc_request *req;
1711 struct target_recovery_data *trd = &obd->obd_recovery_data;
1712 unsigned long delta;
1713 unsigned long flags;
1715 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1719 cfs_daemonize_ctxt("tgt_recov");
1721 SIGNAL_MASK_LOCK(current, flags);
1722 sigfillset(¤t->blocked);
1724 SIGNAL_MASK_UNLOCK(current, flags);
1726 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1730 thread->t_env = &env;
1731 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1732 env.le_ctx.lc_thread = thread;
1733 thread->t_data = NULL;
1735 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1737 trd->trd_processing_task = cfs_curproc_pid();
1739 cfs_spin_lock(&obd->obd_dev_lock);
1740 obd->obd_recovering = 1;
1741 cfs_spin_unlock(&obd->obd_dev_lock);
1742 cfs_complete(&trd->trd_starting);
1744 /* first of all, we have to know the first transno to replay */
1745 if (target_recovery_overseer(obd, check_for_clients,
1746 exp_connect_healthy)) {
1747 abort_req_replay_queue(obd);
1748 abort_lock_replay_queue(obd);
1751 /* next stage: replay requests */
1753 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1754 cfs_atomic_read(&obd->obd_req_replay_clients),
1755 obd->obd_next_recovery_transno);
1756 while ((req = target_next_replay_req(obd))) {
1757 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1758 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1759 lustre_msg_get_transno(req->rq_reqmsg),
1760 libcfs_nid2str(req->rq_peer.nid));
1761 handle_recovery_req(thread, req,
1762 trd->trd_recovery_handler);
1764 * bz18031: increase next_recovery_transno before
1765 * target_request_copy_put() will drop exp_rpc reference
1767 cfs_spin_lock(&obd->obd_recovery_task_lock);
1768 obd->obd_next_recovery_transno++;
1769 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1770 target_exp_dequeue_req_replay(req);
1771 target_request_copy_put(req);
1772 obd->obd_replayed_requests++;
1776 * The second stage: replay locks
1778 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1779 cfs_atomic_read(&obd->obd_lock_replay_clients));
1780 while ((req = target_next_replay_lock(obd))) {
1781 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1782 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1783 libcfs_nid2str(req->rq_peer.nid));
1784 handle_recovery_req(thread, req,
1785 trd->trd_recovery_handler);
1786 target_request_copy_put(req);
1787 obd->obd_replayed_locks++;
1791 * The third stage: reply on final pings, at this moment all clients
1792 * must have request in final queue
1794 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1795 /** Update server last boot epoch */
1796 lut_boot_epoch_update(lut);
1797 /* We drop recoverying flag to forward all new requests
1798 * to regular mds_handle() since now */
1799 cfs_spin_lock(&obd->obd_dev_lock);
1800 obd->obd_recovering = obd->obd_abort_recovery = 0;
1801 cfs_spin_unlock(&obd->obd_dev_lock);
1802 cfs_spin_lock(&obd->obd_recovery_task_lock);
1803 target_cancel_recovery_timer(obd);
1804 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1805 while ((req = target_next_final_ping(obd))) {
1806 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1807 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1808 libcfs_nid2str(req->rq_peer.nid));
1809 handle_recovery_req(thread, req,
1810 trd->trd_recovery_handler);
1811 target_request_copy_put(req);
1814 delta = (jiffies - delta) / CFS_HZ;
1815 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1816 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1817 if (delta > OBD_RECOVERY_TIME_SOFT) {
1818 CWARN("too long recovery - read logs\n");
1819 libcfs_debug_dumplog();
1822 target_finish_recovery(obd);
1824 lu_context_fini(&env.le_ctx);
1825 trd->trd_processing_task = 0;
1826 cfs_complete(&trd->trd_finishing);
1830 static int target_start_recovery_thread(struct lu_target *lut,
1831 svc_handler_t handler)
1833 struct obd_device *obd = lut->lut_obd;
1835 struct target_recovery_data *trd = &obd->obd_recovery_data;
1837 memset(trd, 0, sizeof(*trd));
1838 cfs_init_completion(&trd->trd_starting);
1839 cfs_init_completion(&trd->trd_finishing);
1840 trd->trd_recovery_handler = handler;
1842 if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
1843 cfs_wait_for_completion(&trd->trd_starting);
1844 LASSERT(obd->obd_recovering != 0);
1851 void target_stop_recovery_thread(struct obd_device *obd)
1853 if (obd->obd_recovery_data.trd_processing_task > 0) {
1854 struct target_recovery_data *trd = &obd->obd_recovery_data;
1855 /** recovery can be done but postrecovery is not yet */
1856 cfs_spin_lock(&obd->obd_dev_lock);
1857 if (obd->obd_recovering) {
1858 CERROR("%s: Aborting recovery\n", obd->obd_name);
1859 obd->obd_abort_recovery = 1;
1860 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1862 cfs_spin_unlock(&obd->obd_dev_lock);
1863 cfs_wait_for_completion(&trd->trd_finishing);
1867 void target_recovery_fini(struct obd_device *obd)
1869 class_disconnect_exports(obd);
1870 target_stop_recovery_thread(obd);
1871 target_cleanup_recovery(obd);
1873 EXPORT_SYMBOL(target_recovery_fini);
1875 static void target_recovery_expired(unsigned long castmeharder)
1877 struct obd_device *obd = (struct obd_device *)castmeharder;
1878 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1879 " after %lds (%d clients connected)\n",
1880 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1881 cfs_time_current_sec()- obd->obd_recovery_start,
1882 obd->obd_connected_clients);
1884 obd->obd_recovery_expired = 1;
1885 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1888 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1890 struct obd_device *obd = lut->lut_obd;
1891 if (obd->obd_max_recoverable_clients == 0) {
1892 /** Update server last boot epoch */
1893 lut_boot_epoch_update(lut);
1897 CWARN("RECOVERY: service %s, %d recoverable clients, "
1898 "last_transno "LPU64"\n", obd->obd_name,
1899 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1900 LASSERT(obd->obd_stopping == 0);
1901 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1902 obd->obd_recovery_start = 0;
1903 obd->obd_recovery_end = 0;
1905 /* both values can be get from mount data already */
1906 if (obd->obd_recovery_timeout == 0)
1907 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
1908 if (obd->obd_recovery_time_hard == 0)
1909 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
1910 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1911 target_start_recovery_thread(lut, handler);
1913 EXPORT_SYMBOL(target_recovery_init);
1917 static int target_process_req_flags(struct obd_device *obd,
1918 struct ptlrpc_request *req)
1920 struct obd_export *exp = req->rq_export;
1921 LASSERT(exp != NULL);
1922 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1923 /* client declares he's ready to replay locks */
1924 cfs_spin_lock(&exp->exp_lock);
1925 if (exp->exp_req_replay_needed) {
1926 exp->exp_req_replay_needed = 0;
1927 cfs_spin_unlock(&exp->exp_lock);
1929 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
1930 cfs_atomic_dec(&obd->obd_req_replay_clients);
1932 cfs_spin_unlock(&exp->exp_lock);
1935 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1936 /* client declares he's ready to complete recovery
1937 * so, we put the request on th final queue */
1938 cfs_spin_lock(&exp->exp_lock);
1939 if (exp->exp_lock_replay_needed) {
1940 exp->exp_lock_replay_needed = 0;
1941 cfs_spin_unlock(&exp->exp_lock);
1943 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
1944 cfs_atomic_dec(&obd->obd_lock_replay_clients);
1946 cfs_spin_unlock(&exp->exp_lock);
1952 int target_queue_recovery_request(struct ptlrpc_request *req,
1953 struct obd_device *obd)
1957 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1960 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1961 /* Processing the queue right now, don't re-add. */
1965 target_process_req_flags(obd, req);
1967 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1968 /* client declares he's ready to complete recovery
1969 * so, we put the request on th final queue */
1970 target_request_copy_get(req);
1971 DEBUG_REQ(D_HA, req, "queue final req");
1972 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1973 cfs_spin_lock(&obd->obd_recovery_task_lock);
1974 if (obd->obd_recovering) {
1975 cfs_list_add_tail(&req->rq_list,
1976 &obd->obd_final_req_queue);
1978 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1979 target_request_copy_put(req);
1980 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1982 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1985 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1986 /* client declares he's ready to replay locks */
1987 target_request_copy_get(req);
1988 DEBUG_REQ(D_HA, req, "queue lock replay req");
1989 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1990 cfs_spin_lock(&obd->obd_recovery_task_lock);
1991 LASSERT(obd->obd_recovering);
1992 /* usually due to recovery abort */
1993 if (!req->rq_export->exp_in_recovery) {
1994 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1995 target_request_copy_put(req);
1998 LASSERT(req->rq_export->exp_lock_replay_needed);
1999 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2000 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2004 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2005 * (i.e. buflens etc are in my own byte order), but type-dependent
2006 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2009 CFS_INIT_LIST_HEAD(&req->rq_list);
2010 DEBUG_REQ(D_HA, req, "not queueing");
2014 /* If we're processing the queue, we want don't want to queue this
2017 * Also, if this request has a transno less than the one we're waiting
2018 * for, we should process it now. It could (and currently always will)
2019 * be an open request for a descriptor that was opened some time ago.
2021 * Also, a resent, replayed request that has already been
2022 * handled will pass through here and be processed immediately.
2024 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2025 obd->obd_next_recovery_transno, transno);
2026 cfs_spin_lock(&obd->obd_recovery_task_lock);
2027 if (transno < obd->obd_next_recovery_transno) {
2028 /* Processing the queue right now, don't re-add. */
2029 LASSERT(cfs_list_empty(&req->rq_list));
2030 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2033 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2035 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2038 target_request_copy_get(req);
2039 if (!req->rq_export->exp_in_recovery) {
2040 target_request_copy_put(req);
2043 LASSERT(req->rq_export->exp_req_replay_needed);
2045 if (target_exp_enqueue_req_replay(req)) {
2046 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2047 target_request_copy_put(req);
2052 cfs_spin_lock(&obd->obd_recovery_task_lock);
2053 LASSERT(obd->obd_recovering);
2054 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2055 struct ptlrpc_request *reqiter =
2056 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2058 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2059 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2064 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2066 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2067 "has been claimed by another client");
2068 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2069 target_exp_dequeue_req_replay(req);
2070 target_request_copy_put(req);
2076 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2078 obd->obd_requests_queued_for_recovery++;
2079 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2080 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2085 * Packs current SLV and Limit into \a req.
2087 int target_pack_pool_reply(struct ptlrpc_request *req)
2089 struct obd_device *obd;
2093 * Check that we still have all structures alive as this may
2094 * be some late rpc in shutdown time.
2096 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2097 !exp_connect_lru_resize(req->rq_export))) {
2098 lustre_msg_set_slv(req->rq_repmsg, 0);
2099 lustre_msg_set_limit(req->rq_repmsg, 0);
2104 * OBD is alive here as export is alive, which we checked above.
2106 obd = req->rq_export->exp_obd;
2108 cfs_read_lock(&obd->obd_pool_lock);
2109 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2110 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2111 cfs_read_unlock(&obd->obd_pool_lock);
2116 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2118 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2119 DEBUG_REQ(D_ERROR, req, "dropping reply");
2124 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2125 req->rq_status = rc;
2126 return (ptlrpc_send_error(req, 1));
2128 DEBUG_REQ(D_NET, req, "sending reply");
2131 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2134 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2137 struct ptlrpc_reply_state *rs;
2138 struct obd_device *obd;
2139 struct obd_export *exp;
2140 struct ptlrpc_service *svc;
2143 if (req->rq_no_reply) {
2148 svc = req->rq_rqbd->rqbd_service;
2149 rs = req->rq_reply_state;
2150 if (rs == NULL || !rs->rs_difficult) {
2152 target_send_reply_msg (req, rc, fail_id);
2157 /* must be an export if locks saved */
2158 LASSERT (req->rq_export != NULL);
2159 /* req/reply consistent */
2160 LASSERT (rs->rs_service == svc);
2163 LASSERT (!rs->rs_scheduled);
2164 LASSERT (!rs->rs_scheduled_ever);
2165 LASSERT (!rs->rs_handled);
2166 LASSERT (!rs->rs_on_net);
2167 LASSERT (rs->rs_export == NULL);
2168 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2169 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2171 exp = class_export_get (req->rq_export);
2174 /* disable reply scheduling while I'm setting up */
2175 rs->rs_scheduled = 1;
2177 rs->rs_xid = req->rq_xid;
2178 rs->rs_transno = req->rq_transno;
2179 rs->rs_export = exp;
2180 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2182 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2183 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2184 rs->rs_transno, exp->exp_last_committed);
2185 if (rs->rs_transno > exp->exp_last_committed) {
2186 /* not committed already */
2187 cfs_list_add_tail(&rs->rs_obd_list,
2188 &exp->exp_uncommitted_replies);
2190 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2192 cfs_spin_lock(&exp->exp_lock);
2193 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2194 cfs_spin_unlock(&exp->exp_lock);
2196 netrc = target_send_reply_msg (req, rc, fail_id);
2198 cfs_spin_lock(&svc->srv_rs_lock);
2200 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2203 /* error sending: reply is off the net. Also we need +1
2204 * reply ref until ptlrpc_handle_rs() is done
2205 * with the reply state (if the send was successful, there
2206 * would have been +1 ref for the net, which
2207 * reply_out_callback leaves alone) */
2209 ptlrpc_rs_addref(rs);
2212 cfs_spin_lock(&rs->rs_lock);
2213 if (rs->rs_transno <= exp->exp_last_committed ||
2214 (!rs->rs_on_net && !rs->rs_no_ack) ||
2215 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2216 cfs_list_empty(&rs->rs_obd_list)) {
2217 CDEBUG(D_HA, "Schedule reply immediately\n");
2218 ptlrpc_dispatch_difficult_reply(rs);
2220 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2221 rs->rs_scheduled = 0; /* allow notifier to schedule */
2223 cfs_spin_unlock(&rs->rs_lock);
2224 cfs_spin_unlock(&svc->srv_rs_lock);
2228 int target_handle_ping(struct ptlrpc_request *req)
2230 obd_ping(req->rq_export);
2231 return req_capsule_server_pack(&req->rq_pill);
2234 void target_committed_to_req(struct ptlrpc_request *req)
2236 struct obd_export *exp = req->rq_export;
2238 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2239 lustre_msg_set_last_committed(req->rq_repmsg,
2240 exp->exp_last_committed);
2242 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2243 "%d)", exp->exp_obd->obd_no_transno,
2244 req->rq_repmsg == NULL);
2246 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2247 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2249 EXPORT_SYMBOL(target_committed_to_req);
2251 int target_handle_qc_callback(struct ptlrpc_request *req)
2253 struct obd_quotactl *oqctl;
2254 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2256 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2257 if (oqctl == NULL) {
2258 CERROR("Can't unpack obd_quotactl\n");
2262 cli->cl_qchk_stat = oqctl->qc_stat;
2267 #ifdef HAVE_QUOTA_SUPPORT
2268 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2271 struct obd_device *obd = req->rq_export->exp_obd;
2272 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2273 struct obd_device_target *obt;
2274 struct lustre_quota_ctxt *qctxt;
2275 struct qunit_data *qdata = NULL;
2279 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2282 rc = req_capsule_server_pack(&req->rq_pill);
2284 CERROR("packing reply failed!: rc = %d\n", rc);
2288 LASSERT(req->rq_export);
2290 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2291 if (IS_ERR(qdata)) {
2292 rc = PTR_ERR(qdata);
2293 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2294 req->rq_status = rc;
2298 /* we use the observer */
2299 if (obd_pin_observer(obd, &lov_obd) ||
2300 obd_pin_observer(lov_obd, &master_obd)) {
2301 CERROR("Can't find the observer, it is recovering\n");
2302 req->rq_status = -EAGAIN;
2306 obt = &master_obd->u.obt;
2307 qctxt = &obt->obt_qctxt;
2309 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2310 /* quota_type has not been processed yet, return EAGAIN
2311 * until we know whether or not quotas are supposed to
2313 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2315 req->rq_status = -EAGAIN;
2319 cfs_down_read(&obt->obt_rwsem);
2320 if (qctxt->lqc_lqs_hash == NULL) {
2321 cfs_up_read(&obt->obt_rwsem);
2322 /* quota_type has not been processed yet, return EAGAIN
2323 * until we know whether or not quotas are supposed to
2325 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2327 req->rq_status = -EAGAIN;
2331 LASSERT(qctxt->lqc_handler);
2332 rc = qctxt->lqc_handler(master_obd, qdata,
2333 lustre_msg_get_opc(req->rq_reqmsg));
2334 cfs_up_read(&obt->obt_rwsem);
2335 if (rc && rc != -EDQUOT)
2336 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2337 "dqacq/dqrel failed! (rc:%d)\n", rc);
2338 req->rq_status = rc;
2340 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2342 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2346 /* Block the quota req. b=14840 */
2347 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2352 obd_unpin_observer(lov_obd);
2354 obd_unpin_observer(obd);
2356 rc = ptlrpc_reply(req);
2360 #endif /* !__KERNEL__ */
2362 #endif /* HAVE_QUOTA_SUPPORT */
2364 ldlm_mode_t lck_compat_array[] = {
2365 [LCK_EX] LCK_COMPAT_EX,
2366 [LCK_PW] LCK_COMPAT_PW,
2367 [LCK_PR] LCK_COMPAT_PR,
2368 [LCK_CW] LCK_COMPAT_CW,
2369 [LCK_CR] LCK_COMPAT_CR,
2370 [LCK_NL] LCK_COMPAT_NL,
2371 [LCK_GROUP] LCK_COMPAT_GROUP,
2372 [LCK_COS] LCK_COMPAT_COS,
2376 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2377 * not escape to the user level.
2379 int ldlm_error2errno(ldlm_error_t error)
2387 case ELDLM_LOCK_CHANGED:
2390 case ELDLM_LOCK_ABORTED:
2393 case ELDLM_LOCK_REPLACED:
2396 case ELDLM_NO_LOCK_DATA:
2399 case ELDLM_NAMESPACE_EXISTS:
2402 case ELDLM_BAD_NAMESPACE:
2406 if (((int)error) < 0) /* cast to signed type */
2407 result = error; /* as ldlm_error_t can be unsigned */
2409 CERROR("Invalid DLM result code: %d\n", error);
2415 EXPORT_SYMBOL(ldlm_error2errno);
2418 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2420 ldlm_error_t ldlm_errno2error(int err_no)
2429 error = ELDLM_LOCK_CHANGED;
2432 error = ELDLM_LOCK_ABORTED;
2435 error = ELDLM_LOCK_REPLACED;
2438 error = ELDLM_NO_LOCK_DATA;
2441 error = ELDLM_NAMESPACE_EXISTS;
2444 error = ELDLM_BAD_NAMESPACE;
2451 EXPORT_SYMBOL(ldlm_errno2error);
2453 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2454 void ldlm_dump_export_locks(struct obd_export *exp)
2456 cfs_spin_lock(&exp->exp_locks_list_guard);
2457 if (!cfs_list_empty(&exp->exp_locks_list)) {
2458 struct ldlm_lock *lock;
2460 CERROR("dumping locks for export %p,"
2461 "ignore if the unmount doesn't hang\n", exp);
2462 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2463 ldlm_lock_dump(D_ERROR, lock, 0);
2465 cfs_spin_unlock(&exp->exp_locks_list_guard);