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 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
41 # define EXPORT_SYMTAB
43 #define DEBUG_SUBSYSTEM S_LDLM
46 # include <libcfs/libcfs.h>
48 # include <liblustre.h>
51 #include <lustre_mds.h>
52 #include <lustre_dlm.h>
53 #include <lustre_net.h>
54 #include <lustre_sec.h>
55 #include "ldlm_internal.h"
57 /* @priority: if non-zero, move the selected to the list head
58 * @create: if zero, only search in existed connections
60 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
61 int priority, int create)
63 struct ptlrpc_connection *ptlrpc_conn;
64 struct obd_import_conn *imp_conn = NULL, *item;
68 if (!create && !priority) {
69 CDEBUG(D_HA, "Nothing to do\n");
73 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
75 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
80 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
82 GOTO(out_put, rc = -ENOMEM);
86 cfs_spin_lock(&imp->imp_lock);
87 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
88 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
90 cfs_list_del(&item->oic_item);
91 cfs_list_add(&item->oic_item,
93 item->oic_last_attempt = 0;
95 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
96 imp, imp->imp_obd->obd_name, uuid->uuid,
97 (priority ? ", moved to head" : ""));
98 cfs_spin_unlock(&imp->imp_lock);
99 GOTO(out_free, rc = 0);
104 imp_conn->oic_conn = ptlrpc_conn;
105 imp_conn->oic_uuid = *uuid;
106 imp_conn->oic_last_attempt = 0;
108 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
110 cfs_list_add_tail(&imp_conn->oic_item,
111 &imp->imp_conn_list);
112 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
113 imp, imp->imp_obd->obd_name, uuid->uuid,
114 (priority ? "head" : "tail"));
116 cfs_spin_unlock(&imp->imp_lock);
117 GOTO(out_free, rc = -ENOENT);
120 cfs_spin_unlock(&imp->imp_lock);
124 OBD_FREE(imp_conn, sizeof(*imp_conn));
126 ptlrpc_connection_put(ptlrpc_conn);
130 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
132 return import_set_conn(imp, uuid, 1, 0);
135 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
138 return import_set_conn(imp, uuid, priority, 1);
141 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
143 struct obd_import_conn *imp_conn;
144 struct obd_export *dlmexp;
148 cfs_spin_lock(&imp->imp_lock);
149 if (cfs_list_empty(&imp->imp_conn_list)) {
150 LASSERT(!imp->imp_connection);
154 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
155 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
157 LASSERT(imp_conn->oic_conn);
159 /* is current conn? */
160 if (imp_conn == imp->imp_conn_current) {
161 LASSERT(imp_conn->oic_conn == imp->imp_connection);
163 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
164 imp->imp_state != LUSTRE_IMP_DISCON) {
165 CERROR("can't remove current connection\n");
166 GOTO(out, rc = -EBUSY);
169 ptlrpc_connection_put(imp->imp_connection);
170 imp->imp_connection = NULL;
172 dlmexp = class_conn2export(&imp->imp_dlm_handle);
173 if (dlmexp && dlmexp->exp_connection) {
174 LASSERT(dlmexp->exp_connection ==
176 ptlrpc_connection_put(dlmexp->exp_connection);
177 dlmexp->exp_connection = NULL;
181 cfs_list_del(&imp_conn->oic_item);
182 ptlrpc_connection_put(imp_conn->oic_conn);
183 OBD_FREE(imp_conn, sizeof(*imp_conn));
184 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
185 imp, imp->imp_obd->obd_name, uuid->uuid);
190 cfs_spin_unlock(&imp->imp_lock);
192 CERROR("connection %s not found\n", uuid->uuid);
197 * Find conn uuid by peer nid. @peer is a server nid. This function is used
198 * to find a conn uuid of @imp which can reach @peer.
200 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
201 struct obd_uuid *uuid)
203 struct obd_import_conn *conn;
207 cfs_spin_lock(&imp->imp_lock);
208 cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
209 /* check if conn uuid does have this peer nid */
210 if (class_check_uuid(&conn->oic_uuid, peer)) {
211 *uuid = conn->oic_uuid;
216 cfs_spin_unlock(&imp->imp_lock);
219 EXPORT_SYMBOL(client_import_find_conn);
221 void client_destroy_import(struct obd_import *imp)
223 /* drop security policy instance after all rpc finished/aborted
224 * to let all busy contexts be released. */
225 class_import_get(imp);
226 class_destroy_import(imp);
227 sptlrpc_import_sec_put(imp);
228 class_import_put(imp);
231 /* configure an RPC client OBD device
236 * 3 - inactive-on-startup
238 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
240 struct client_obd *cli = &obddev->u.cli;
241 struct obd_import *imp;
242 struct obd_uuid server_uuid;
243 int rq_portal, rp_portal, connect_op;
244 char *name = obddev->obd_type->typ_name;
245 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
249 /* In a more perfect world, we would hang a ptlrpc_client off of
250 * obd_type and just use the values from there. */
251 if (!strcmp(name, LUSTRE_OSC_NAME)) {
252 rq_portal = OST_REQUEST_PORTAL;
253 rp_portal = OSC_REPLY_PORTAL;
254 connect_op = OST_CONNECT;
255 cli->cl_sp_me = LUSTRE_SP_CLI;
256 cli->cl_sp_to = LUSTRE_SP_OST;
257 ns_type = LDLM_NS_TYPE_OSC;
259 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
260 rq_portal = MDS_REQUEST_PORTAL;
261 rp_portal = MDC_REPLY_PORTAL;
262 connect_op = MDS_CONNECT;
263 cli->cl_sp_me = LUSTRE_SP_CLI;
264 cli->cl_sp_to = LUSTRE_SP_MDT;
265 ns_type = LDLM_NS_TYPE_MDC;
267 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
268 rq_portal = MGS_REQUEST_PORTAL;
269 rp_portal = MGC_REPLY_PORTAL;
270 connect_op = MGS_CONNECT;
271 cli->cl_sp_me = LUSTRE_SP_MGC;
272 cli->cl_sp_to = LUSTRE_SP_MGS;
273 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
274 ns_type = LDLM_NS_TYPE_MGC;
277 CERROR("unknown client OBD type \"%s\", can't setup\n",
282 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
283 CERROR("requires a TARGET UUID\n");
287 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
288 CERROR("client UUID must be less than 38 characters\n");
292 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
293 CERROR("setup requires a SERVER UUID\n");
297 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
298 CERROR("target UUID must be less than 38 characters\n");
302 cfs_init_rwsem(&cli->cl_sem);
303 cfs_sema_init(&cli->cl_mgc_sem, 1);
304 cli->cl_conn_count = 0;
305 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
306 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
307 sizeof(server_uuid)));
310 cli->cl_avail_grant = 0;
311 /* FIXME: should limit this for the sum of all cl_dirty_max */
312 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
313 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
314 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
315 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
316 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
317 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
318 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
319 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
320 client_obd_list_lock_init(&cli->cl_loi_list_lock);
321 cli->cl_r_in_flight = 0;
322 cli->cl_w_in_flight = 0;
324 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
325 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
326 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
327 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
328 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
329 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
330 cfs_waitq_init(&cli->cl_destroy_waitq);
331 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
332 #ifdef ENABLE_CHECKSUM
333 /* Turn on checksumming by default. */
334 cli->cl_checksum = 1;
336 * The supported checksum types will be worked out at connect time
337 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
340 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
342 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
344 /* This value may be changed at connect time in
345 ptlrpc_connect_interpret. */
346 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
347 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
349 if (!strcmp(name, LUSTRE_MDC_NAME)) {
350 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
351 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
352 cli->cl_max_rpcs_in_flight = 2;
353 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
354 cli->cl_max_rpcs_in_flight = 3;
355 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
356 cli->cl_max_rpcs_in_flight = 4;
358 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
363 CERROR("ldlm_get_ref failed: %d\n", rc);
367 ptlrpc_init_client(rq_portal, rp_portal, name,
368 &obddev->obd_ldlm_client);
370 imp = class_new_import(obddev);
372 GOTO(err_ldlm, rc = -ENOENT);
373 imp->imp_client = &obddev->obd_ldlm_client;
374 imp->imp_connect_op = connect_op;
375 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
376 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
377 LUSTRE_CFG_BUFLEN(lcfg, 1));
378 class_import_put(imp);
380 rc = client_import_add_conn(imp, &server_uuid, 1);
382 CERROR("can't add initial connection\n");
383 GOTO(err_import, rc);
386 cli->cl_import = imp;
387 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
388 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
389 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
391 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
392 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
393 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
394 name, obddev->obd_name,
395 cli->cl_target_uuid.uuid);
396 cfs_spin_lock(&imp->imp_lock);
397 imp->imp_deactive = 1;
398 cfs_spin_unlock(&imp->imp_lock);
402 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
403 LDLM_NAMESPACE_CLIENT,
404 LDLM_NAMESPACE_GREEDY,
406 if (obddev->obd_namespace == NULL) {
407 CERROR("Unable to create client namespace - %s\n",
409 GOTO(err_import, rc = -ENOMEM);
412 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
417 class_destroy_import(imp);
425 int client_obd_cleanup(struct obd_device *obddev)
429 ldlm_namespace_free_post(obddev->obd_namespace);
430 obddev->obd_namespace = NULL;
436 /* ->o_connect() method for client side (OSC and MDC and MGC) */
437 int client_connect_import(const struct lu_env *env,
438 struct obd_export **exp,
439 struct obd_device *obd, struct obd_uuid *cluuid,
440 struct obd_connect_data *data, void *localdata)
442 struct client_obd *cli = &obd->u.cli;
443 struct obd_import *imp = cli->cl_import;
444 struct obd_connect_data *ocd;
445 struct lustre_handle conn = { 0 };
450 cfs_down_write(&cli->cl_sem);
451 if (cli->cl_conn_count > 0 )
452 GOTO(out_sem, rc = -EALREADY);
454 rc = class_connect(&conn, obd, cluuid);
458 cli->cl_conn_count++;
459 *exp = class_conn2export(&conn);
461 LASSERT(obd->obd_namespace);
463 imp->imp_dlm_handle = conn;
464 rc = ptlrpc_init_import(imp);
468 ocd = &imp->imp_connect_data;
471 imp->imp_connect_flags_orig = data->ocd_connect_flags;
474 rc = ptlrpc_connect_import(imp);
476 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
479 LASSERT((*exp)->exp_connection);
482 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
483 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
484 data->ocd_connect_flags, ocd->ocd_connect_flags);
485 data->ocd_connect_flags = ocd->ocd_connect_flags;
488 ptlrpc_pinger_add_import(imp);
494 cli->cl_conn_count--;
495 class_disconnect(*exp);
499 cfs_up_write(&cli->cl_sem);
504 int client_disconnect_export(struct obd_export *exp)
506 struct obd_device *obd = class_exp2obd(exp);
507 struct client_obd *cli;
508 struct obd_import *imp;
513 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
514 exp, exp ? exp->exp_handle.h_cookie : -1);
519 imp = cli->cl_import;
521 cfs_down_write(&cli->cl_sem);
522 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
525 if (!cli->cl_conn_count) {
526 CERROR("disconnecting disconnected device (%s)\n",
528 GOTO(out_disconnect, rc = -EINVAL);
531 cli->cl_conn_count--;
532 if (cli->cl_conn_count)
533 GOTO(out_disconnect, rc = 0);
535 /* Mark import deactivated now, so we don't try to reconnect if any
536 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
537 * fully deactivate the import, or that would drop all requests. */
538 cfs_spin_lock(&imp->imp_lock);
539 imp->imp_deactive = 1;
540 cfs_spin_unlock(&imp->imp_lock);
542 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
543 * delete it regardless. (It's safe to delete an import that was
545 (void)ptlrpc_pinger_del_import(imp);
547 if (obd->obd_namespace != NULL) {
548 /* obd_force == local only */
549 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
550 obd->obd_force ? LCF_LOCAL : 0, NULL);
551 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
555 * there's no need to hold sem during disconnecting an import,
556 * and actually it may cause deadlock in gss.
558 cfs_up_write(&cli->cl_sem);
559 rc = ptlrpc_disconnect_import(imp, 0);
560 cfs_down_write(&cli->cl_sem);
562 ptlrpc_invalidate_import(imp);
564 if (imp->imp_rq_pool) {
565 ptlrpc_free_rq_pool(imp->imp_rq_pool);
566 imp->imp_rq_pool = NULL;
568 client_destroy_import(imp);
569 cli->cl_import = NULL;
574 /* use server style - class_disconnect should be always called for
576 err = class_disconnect(exp);
580 cfs_up_write(&cli->cl_sem);
585 int server_disconnect_export(struct obd_export *exp)
590 /* Disconnect early so that clients can't keep using export */
591 rc = class_disconnect(exp);
592 /* close import for avoid sending any requests */
593 if (exp->exp_imp_reverse)
594 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
596 if (exp->exp_obd->obd_namespace != NULL)
597 ldlm_cancel_locks_for_export(exp);
599 /* complete all outstanding replies */
600 cfs_spin_lock(&exp->exp_lock);
601 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
602 struct ptlrpc_reply_state *rs =
603 cfs_list_entry(exp->exp_outstanding_replies.next,
604 struct ptlrpc_reply_state, rs_exp_list);
605 struct ptlrpc_service *svc = rs->rs_service;
607 cfs_spin_lock(&svc->srv_rs_lock);
608 cfs_list_del_init(&rs->rs_exp_list);
609 cfs_spin_lock(&rs->rs_lock);
610 ptlrpc_schedule_difficult_reply(rs);
611 cfs_spin_unlock(&rs->rs_lock);
612 cfs_spin_unlock(&svc->srv_rs_lock);
614 cfs_spin_unlock(&exp->exp_lock);
619 /* --------------------------------------------------------------------------
620 * from old lib/target.c
621 * -------------------------------------------------------------------------- */
623 static int target_handle_reconnect(struct lustre_handle *conn,
624 struct obd_export *exp,
625 struct obd_uuid *cluuid)
628 if (exp->exp_connection && exp->exp_imp_reverse) {
629 struct lustre_handle *hdl;
630 hdl = &exp->exp_imp_reverse->imp_remote_handle;
631 /* Might be a re-connect after a partition. */
632 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
633 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
635 conn->cookie = exp->exp_handle.h_cookie;
636 /* target_handle_connect() treats EALREADY and
637 * -EALREADY differently. EALREADY means we are
638 * doing a valid reconnect from the same client. */
641 CERROR("%s reconnecting from %s, "
642 "handle mismatch (ours "LPX64", theirs "
643 LPX64")\n", cluuid->uuid,
644 exp->exp_connection->c_remote_uuid.uuid,
645 hdl->cookie, conn->cookie);
646 memset(conn, 0, sizeof *conn);
647 /* target_handle_connect() treats EALREADY and
648 * -EALREADY differently. -EALREADY is an error
649 * (same UUID, different handle). */
654 conn->cookie = exp->exp_handle.h_cookie;
655 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
656 cluuid->uuid, exp, conn->cookie);
660 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
663 struct obd_export *exp = cb_data;
665 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
666 obd->obd_name, exp->exp_client_uuid.uuid);
668 cfs_spin_lock(&exp->exp_lock);
669 exp->exp_need_sync = 0;
670 cfs_spin_unlock(&exp->exp_lock);
671 class_export_cb_put(exp);
673 EXPORT_SYMBOL(target_client_add_cb);
677 check_and_start_recovery_timer(struct obd_device *obd,
678 struct ptlrpc_request *req, int new_client);
681 check_and_start_recovery_timer(struct obd_device *obd,
682 struct ptlrpc_request *req, int new_client)
687 int target_handle_connect(struct ptlrpc_request *req)
689 struct obd_device *target, *targref = NULL;
690 struct obd_export *export = NULL;
691 struct obd_import *revimp;
692 struct lustre_handle conn;
693 struct lustre_handle *tmp;
694 struct obd_uuid tgtuuid;
695 struct obd_uuid cluuid;
696 struct obd_uuid remote_uuid;
700 struct obd_connect_data *data, *tmpdata;
702 lnet_nid_t *client_nid = NULL;
705 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
707 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
709 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
710 GOTO(out, rc = -EINVAL);
713 obd_str2uuid(&tgtuuid, str);
714 target = class_uuid2obd(&tgtuuid);
716 target = class_name2obd(str);
718 if (!target || target->obd_stopping || !target->obd_set_up) {
719 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
720 "for connect (%s)\n", str,
721 !target ? "no target" :
722 (target->obd_stopping ? "stopping" :
724 GOTO(out, rc = -ENODEV);
727 if (target->obd_no_conn) {
728 LCONSOLE_WARN("%s: temporarily refusing client connection "
729 "from %s\n", target->obd_name,
730 libcfs_nid2str(req->rq_peer.nid));
731 GOTO(out, rc = -EAGAIN);
734 /* Make sure the target isn't cleaned up while we're here. Yes,
735 there's still a race between the above check and our incref here.
736 Really, class_uuid2obd should take the ref. */
737 targref = class_incref(target, __FUNCTION__, cfs_current());
740 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
742 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
743 GOTO(out, rc = -EINVAL);
746 obd_str2uuid(&cluuid, str);
748 /* XXX extract a nettype and format accordingly */
749 switch (sizeof(lnet_nid_t)) {
750 /* NB the casts only avoid compiler warnings */
752 snprintf(remote_uuid.uuid, sizeof remote_uuid,
753 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
756 snprintf(remote_uuid.uuid, sizeof remote_uuid,
757 "NET_%x_UUID", (__u32)req->rq_peer.nid);
763 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
765 GOTO(out, rc = -EPROTO);
769 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
771 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
773 GOTO(out, rc = -EPROTO);
775 rc = req_capsule_server_pack(&req->rq_pill);
779 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
781 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
782 "libclient connection attempt");
783 GOTO(out, rc = -EPROTO);
784 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
785 LUSTRE_VERSION_ALLOWED_OFFSET ||
786 data->ocd_version > LUSTRE_VERSION_CODE +
787 LUSTRE_VERSION_ALLOWED_OFFSET) {
788 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
789 "libclient connection attempt",
790 data->ocd_version < LUSTRE_VERSION_CODE ?
792 OBD_OCD_VERSION_MAJOR(data->ocd_version),
793 OBD_OCD_VERSION_MINOR(data->ocd_version),
794 OBD_OCD_VERSION_PATCH(data->ocd_version),
795 OBD_OCD_VERSION_FIX(data->ocd_version));
796 data = req_capsule_server_sized_get(&req->rq_pill,
798 offsetof(typeof(*data), ocd_version) +
799 sizeof(data->ocd_version));
801 data->ocd_connect_flags = OBD_CONNECT_VERSION;
802 data->ocd_version = LUSTRE_VERSION_CODE;
804 GOTO(out, rc = -EPROTO);
808 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
809 (data->ocd_connect_flags & OBD_CONNECT_MDS))
812 /* lctl gets a backstage, all-access pass. */
813 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
814 goto dont_check_exports;
816 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
820 /* we've found an export in the hash */
821 if (export->exp_connecting) { /* bug 9635, et. al. */
822 CWARN("%s: exp %p already connecting\n",
823 export->exp_obd->obd_name, export);
824 class_export_put(export);
827 } else if (mds_conn && export->exp_connection) {
828 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
829 /* mds reconnected after failover */
830 CWARN("%s: received MDS connection from NID %s,"
831 " removing former export from NID %s\n",
832 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
833 libcfs_nid2str(export->exp_connection->c_peer.nid));
835 /* new mds connection from the same nid */
836 CWARN("%s: received new MDS connection from NID %s,"
837 " removing former export from same NID\n",
838 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
839 class_fail_export(export);
840 class_export_put(export);
843 } else if (export->exp_connection != NULL &&
844 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
845 (lustre_msg_get_op_flags(req->rq_reqmsg) &
846 MSG_CONNECT_INITIAL)) {
847 /* in mds failover we have static uuid but nid can be
849 CWARN("%s: cookie %s seen on new NID %s when "
850 "existing NID %s is already connected\n",
851 target->obd_name, cluuid.uuid,
852 libcfs_nid2str(req->rq_peer.nid),
853 libcfs_nid2str(export->exp_connection->c_peer.nid));
855 class_export_put(export);
858 cfs_spin_lock(&export->exp_lock);
859 export->exp_connecting = 1;
860 cfs_spin_unlock(&export->exp_lock);
861 class_export_put(export);
862 LASSERT(export->exp_obd == target);
864 rc = target_handle_reconnect(&conn, export, &cluuid);
867 /* If we found an export, we already unlocked. */
870 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
871 } else if (req->rq_export == NULL &&
872 cfs_atomic_read(&export->exp_rpc_count) > 0) {
873 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
874 target->obd_name, cluuid.uuid,
875 libcfs_nid2str(req->rq_peer.nid),
876 export, cfs_atomic_read(&export->exp_refcount));
877 GOTO(out, rc = -EBUSY);
878 } else if (req->rq_export != NULL &&
879 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
880 /* the current connect rpc has increased exp_rpc_count */
881 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
882 target->obd_name, cluuid.uuid,
883 libcfs_nid2str(req->rq_peer.nid),
884 export, cfs_atomic_read(&export->exp_rpc_count) - 1);
885 cfs_spin_lock(&export->exp_lock);
886 if (req->rq_export->exp_conn_cnt <
887 lustre_msg_get_conn_cnt(req->rq_reqmsg))
888 /* try to abort active requests */
889 req->rq_export->exp_abort_active_req = 1;
890 cfs_spin_unlock(&export->exp_lock);
891 GOTO(out, rc = -EBUSY);
892 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
893 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
894 "cookies not random?\n", target->obd_name,
895 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
896 GOTO(out, rc = -EALREADY);
898 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
905 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
906 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
907 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
908 export, (long)cfs_time_current_sec(),
909 export ? (long)export->exp_last_request_time : 0);
911 /* If this is the first time a client connects, reset the recovery
913 if (rc == 0 && target->obd_recovering)
914 check_and_start_recovery_timer(target, req, export == NULL);
916 /* We want to handle EALREADY but *not* -EALREADY from
917 * target_handle_reconnect(), return reconnection state in a flag */
918 if (rc == EALREADY) {
919 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
925 /* Tell the client if we support replayable requests */
926 if (target->obd_replayable)
927 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
928 client_nid = &req->rq_peer.nid;
930 if (export == NULL) {
931 if (target->obd_recovering) {
934 t = cfs_timer_deadline(&target->obd_recovery_timer);
935 t = cfs_time_sub(t, cfs_time_current());
936 CERROR("%s: denying connection for new client %s (%s): "
937 "%d clients in recovery for "CFS_TIME_T"s\n",
939 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
940 cfs_atomic_read(&target-> \
941 obd_lock_replay_clients),
942 cfs_duration_sec(t));
946 rc = obd_connect(req->rq_svc_thread->t_env,
947 &export, target, &cluuid, data,
950 conn.cookie = export->exp_handle.h_cookie;
953 rc = obd_reconnect(req->rq_svc_thread->t_env,
954 export, target, &cluuid, data, client_nid);
956 /* prevous done via class_conn2export */
957 class_export_get(export);
962 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
963 data->ocd_instance = target->u.obt.obt_instance;
965 /* Return only the parts of obd_connect_data that we understand, so the
966 * client knows that we don't understand the rest. */
968 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
970 tmpdata = req_capsule_server_get(&req->rq_pill,
972 /* Don't use struct assignment here, because the client reply
973 * buffer may be smaller/larger than the local struct
974 * obd_connect_data. */
975 memcpy(tmpdata, data, min(tmpsize, size));
978 /* If all else goes well, this is our RPC return code. */
981 lustre_msg_set_handle(req->rq_repmsg, &conn);
983 /* If the client and the server are the same node, we will already
984 * have an export that really points to the client's DLM export,
985 * because we have a shared handles table.
987 * XXX this will go away when shaver stops sending the "connect" handle
988 * in the real "remote handle" field of the request --phik 24 Apr 2003
990 if (req->rq_export != NULL)
991 class_export_put(req->rq_export);
993 req->rq_export = export;
995 cfs_spin_lock(&export->exp_lock);
996 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
997 cfs_spin_unlock(&export->exp_lock);
998 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
999 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1000 export->exp_conn_cnt,
1001 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1003 GOTO(out, rc = -EALREADY);
1005 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1006 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1007 export->exp_abort_active_req = 0;
1009 /* request from liblustre? Don't evict it for not pinging. */
1010 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1011 export->exp_libclient = 1;
1012 cfs_spin_unlock(&export->exp_lock);
1014 cfs_spin_lock(&target->obd_dev_lock);
1015 cfs_list_del_init(&export->exp_obd_chain_timed);
1016 cfs_spin_unlock(&target->obd_dev_lock);
1018 cfs_spin_unlock(&export->exp_lock);
1021 if (export->exp_connection != NULL) {
1022 /* Check to see if connection came from another NID */
1023 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1024 !cfs_hlist_unhashed(&export->exp_nid_hash))
1025 cfs_hash_del(export->exp_obd->obd_nid_hash,
1026 &export->exp_connection->c_peer.nid,
1027 &export->exp_nid_hash);
1029 ptlrpc_connection_put(export->exp_connection);
1032 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1035 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1036 cfs_hash_add(export->exp_obd->obd_nid_hash,
1037 &export->exp_connection->c_peer.nid,
1038 &export->exp_nid_hash);
1041 if (target->obd_recovering && !export->exp_in_recovery) {
1043 __u64 transno = data->ocd_transno;
1045 cfs_spin_lock(&export->exp_lock);
1046 export->exp_in_recovery = 1;
1047 export->exp_req_replay_needed = 1;
1048 export->exp_lock_replay_needed = 1;
1049 cfs_spin_unlock(&export->exp_lock);
1051 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1052 MSG_CONNECT_TRANSNO);
1053 if (has_transno && transno == 0)
1054 CWARN("Connect with zero transno!\n");
1056 if (has_transno && transno > 0 &&
1057 transno < target->obd_next_recovery_transno &&
1058 transno > target->obd_last_committed) {
1059 /* another way is to use cmpxchg() so it will be
1061 cfs_spin_lock(&target->obd_recovery_task_lock);
1062 if (transno < target->obd_next_recovery_transno)
1063 target->obd_next_recovery_transno = transno;
1064 cfs_spin_unlock(&target->obd_recovery_task_lock);
1067 cfs_atomic_inc(&target->obd_req_replay_clients);
1068 cfs_atomic_inc(&target->obd_lock_replay_clients);
1069 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1070 target->obd_max_recoverable_clients)
1071 cfs_waitq_signal(&target->obd_next_transno_waitq);
1074 /* Tell the client we're in recovery, when client is involved in it. */
1075 if (target->obd_recovering)
1076 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1078 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1081 if (export->exp_imp_reverse != NULL) {
1082 /* destroyed import can be still referenced in ctxt */
1083 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1084 KEY_REVIMP_UPD, 0, NULL, NULL);
1086 client_destroy_import(export->exp_imp_reverse);
1089 /* for the rest part, we return -ENOTCONN in case of errors
1090 * in order to let client initialize connection again.
1092 revimp = export->exp_imp_reverse = class_new_import(target);
1094 CERROR("fail to alloc new reverse import.\n");
1095 GOTO(out, rc = -ENOTCONN);
1098 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1099 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1100 revimp->imp_remote_handle = conn;
1101 revimp->imp_dlm_fake = 1;
1102 revimp->imp_state = LUSTRE_IMP_FULL;
1104 /* unknown versions will be caught in
1105 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1106 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1108 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1109 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1110 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1112 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1114 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1115 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1116 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1118 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1120 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1122 CERROR("Failed to get sec for reverse import: %d\n", rc);
1123 export->exp_imp_reverse = NULL;
1124 class_destroy_import(revimp);
1127 class_import_put(revimp);
1130 cfs_spin_lock(&export->exp_lock);
1131 export->exp_connecting = 0;
1132 cfs_spin_unlock(&export->exp_lock);
1135 class_decref(targref, __FUNCTION__, cfs_current());
1137 req->rq_status = rc;
1141 int target_handle_disconnect(struct ptlrpc_request *req)
1146 rc = req_capsule_server_pack(&req->rq_pill);
1150 /* keep the rq_export around so we can send the reply */
1151 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1156 void target_destroy_export(struct obd_export *exp)
1158 /* exports created from last_rcvd data, and "fake"
1159 exports created by lctl don't have an import */
1160 if (exp->exp_imp_reverse != NULL)
1161 client_destroy_import(exp->exp_imp_reverse);
1163 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1164 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1165 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1166 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1170 * Recovery functions
1172 static void target_request_copy_get(struct ptlrpc_request *req)
1174 class_export_rpc_get(req->rq_export);
1175 LASSERT(cfs_list_empty(&req->rq_list));
1176 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1178 /* increase refcount to keep request in queue */
1179 cfs_atomic_inc(&req->rq_refcount);
1180 /** let export know it has replays to be handled */
1181 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1184 static void target_request_copy_put(struct ptlrpc_request *req)
1186 LASSERT(cfs_list_empty(&req->rq_replay_list));
1187 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1189 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1190 class_export_rpc_put(req->rq_export);
1191 ptlrpc_server_drop_request(req);
1194 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1196 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1197 struct obd_export *exp = req->rq_export;
1198 struct ptlrpc_request *reqiter;
1203 cfs_spin_lock(&exp->exp_lock);
1204 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1206 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1213 /* we expect it with RESENT and REPLAY flags */
1214 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1215 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1216 CERROR("invalid flags %x of resent replay\n",
1217 lustre_msg_get_flags(req->rq_reqmsg));
1219 cfs_list_add_tail(&req->rq_replay_list,
1220 &exp->exp_req_replay_queue);
1223 cfs_spin_unlock(&exp->exp_lock);
1227 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1229 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1230 LASSERT(req->rq_export);
1232 cfs_spin_lock(&req->rq_export->exp_lock);
1233 cfs_list_del_init(&req->rq_replay_list);
1234 cfs_spin_unlock(&req->rq_export->exp_lock);
1238 static void target_finish_recovery(struct obd_device *obd)
1240 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1241 obd->obd_recovery_start);
1244 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1245 "%d recovered and %d %s evicted.\n", obd->obd_name,
1246 (int)elapsed_time / 60, (int)elapsed_time % 60,
1247 obd->obd_max_recoverable_clients,
1248 cfs_atomic_read(&obd->obd_connected_clients),
1249 obd->obd_stale_clients,
1250 obd->obd_stale_clients == 1 ? "was" : "were");
1252 ldlm_reprocess_all_ns(obd->obd_namespace);
1253 cfs_spin_lock(&obd->obd_recovery_task_lock);
1254 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1255 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1256 !cfs_list_empty(&obd->obd_final_req_queue)) {
1257 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1259 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1260 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1262 cfs_list_empty(&obd->obd_final_req_queue) ? \
1264 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1267 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1269 obd->obd_recovery_end = cfs_time_current_sec();
1271 /* when recovery finished, cleanup orphans on mds and ost */
1272 if (OBT(obd) && OBP(obd, postrecov)) {
1273 int rc = OBP(obd, postrecov)(obd);
1275 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1281 static void abort_req_replay_queue(struct obd_device *obd)
1283 struct ptlrpc_request *req, *n;
1284 cfs_list_t abort_list;
1286 CFS_INIT_LIST_HEAD(&abort_list);
1287 cfs_spin_lock(&obd->obd_recovery_task_lock);
1288 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1289 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1290 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1291 DEBUG_REQ(D_WARNING, req, "aborted:");
1292 req->rq_status = -ENOTCONN;
1293 if (ptlrpc_error(req)) {
1294 DEBUG_REQ(D_ERROR, req,
1295 "failed abort_req_reply; skipping");
1297 target_exp_dequeue_req_replay(req);
1298 target_request_copy_put(req);
1302 static void abort_lock_replay_queue(struct obd_device *obd)
1304 struct ptlrpc_request *req, *n;
1305 cfs_list_t abort_list;
1307 CFS_INIT_LIST_HEAD(&abort_list);
1308 cfs_spin_lock(&obd->obd_recovery_task_lock);
1309 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1310 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1311 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1312 DEBUG_REQ(D_ERROR, req, "aborted:");
1313 req->rq_status = -ENOTCONN;
1314 if (ptlrpc_error(req)) {
1315 DEBUG_REQ(D_ERROR, req,
1316 "failed abort_lock_reply; skipping");
1318 target_request_copy_put(req);
1322 /* Called from a cleanup function if the device is being cleaned up
1323 forcefully. The exports should all have been disconnected already,
1324 the only thing left to do is
1325 - clear the recovery flags
1327 - free queued requests and replies, but don't send replies
1328 Because the obd_stopping flag is set, no new requests should be received.
1331 void target_cleanup_recovery(struct obd_device *obd)
1333 struct ptlrpc_request *req, *n;
1334 cfs_list_t clean_list;
1337 CFS_INIT_LIST_HEAD(&clean_list);
1338 cfs_spin_lock(&obd->obd_dev_lock);
1339 if (!obd->obd_recovering) {
1340 cfs_spin_unlock(&obd->obd_dev_lock);
1344 obd->obd_recovering = obd->obd_abort_recovery = 0;
1345 cfs_spin_unlock(&obd->obd_dev_lock);
1347 cfs_spin_lock(&obd->obd_recovery_task_lock);
1348 target_cancel_recovery_timer(obd);
1349 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1350 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1352 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1353 LASSERT(req->rq_reply_state == 0);
1354 target_exp_dequeue_req_replay(req);
1355 target_request_copy_put(req);
1358 cfs_spin_lock(&obd->obd_recovery_task_lock);
1359 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1360 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1361 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1363 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1364 LASSERT(req->rq_reply_state == 0);
1365 target_request_copy_put(req);
1371 /* obd_recovery_task_lock should be held */
1372 void target_cancel_recovery_timer(struct obd_device *obd)
1374 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1375 cfs_timer_disarm(&obd->obd_recovery_timer);
1378 static void target_start_recovery_timer(struct obd_device *obd)
1380 if (obd->obd_recovery_start != 0)
1383 cfs_spin_lock(&obd->obd_dev_lock);
1384 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1385 cfs_spin_unlock(&obd->obd_dev_lock);
1389 LASSERT(obd->obd_recovery_timeout != 0);
1391 if (obd->obd_recovery_start != 0) {
1392 cfs_spin_unlock(&obd->obd_dev_lock);
1396 cfs_timer_arm(&obd->obd_recovery_timer,
1397 cfs_time_shift(obd->obd_recovery_timeout));
1398 obd->obd_recovery_start = cfs_time_current_sec();
1399 cfs_spin_unlock(&obd->obd_dev_lock);
1401 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1402 "or until %d client%s reconnect%s\n",
1404 obd->obd_recovery_timeout / 60,
1405 obd->obd_recovery_timeout % 60,
1406 obd->obd_max_recoverable_clients,
1407 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1408 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1412 * extend recovery window.
1414 * if @extend is true, extend recovery window to have @drt remaining at least;
1415 * otherwise, make sure the recovery timeout value is not less than @drt.
1417 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1421 cfs_duration_t left;
1424 cfs_spin_lock(&obd->obd_dev_lock);
1425 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1426 cfs_spin_unlock(&obd->obd_dev_lock);
1430 LASSERT(obd->obd_recovery_start != 0);
1432 now = cfs_time_current_sec();
1433 to = obd->obd_recovery_timeout;
1434 end = obd->obd_recovery_start + to;
1435 left = cfs_time_sub(end, now);
1437 if (extend && (drt > left)) {
1439 } else if (!extend && (drt > to)) {
1441 /* reduce drt by already passed time */
1442 drt -= obd->obd_recovery_timeout - left;
1445 if (to > obd->obd_recovery_time_hard)
1446 to = obd->obd_recovery_time_hard;
1447 if (obd->obd_recovery_timeout < to) {
1448 obd->obd_recovery_timeout = to;
1449 cfs_timer_arm(&obd->obd_recovery_timer,
1450 cfs_time_shift(drt));
1452 cfs_spin_unlock(&obd->obd_dev_lock);
1454 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1455 obd->obd_name, (unsigned)drt);
1458 /* Reset the timer with each new client connection */
1460 * This timer is actually reconnect_timer, which is for making sure
1461 * the total recovery window is at least as big as my reconnect
1462 * attempt timing. So the initial recovery time_out will be set to
1463 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1464 * from client is bigger than this, then the recovery time_out will
1465 * be extended to make sure the client could be reconnected, in the
1466 * process, the timeout from the new client should be ignored.
1470 check_and_start_recovery_timer(struct obd_device *obd,
1471 struct ptlrpc_request *req,
1474 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1475 struct obd_device_target *obt = &obd->u.obt;
1476 struct lustre_sb_info *lsi;
1478 if (!new_client && service_time)
1479 /* Teach server about old server's estimates, as first guess
1480 * at how long new requests will take. */
1481 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1484 target_start_recovery_timer(obd);
1486 /* convert the service time to rpc timeout,
1487 * reuse service_time to limit stack usage */
1488 service_time = at_est2timeout(service_time);
1490 /* We expect other clients to timeout within service_time, then try
1491 * to reconnect, then try the failover server. The max delay between
1492 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1493 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1495 LASSERT(obt->obt_magic == OBT_MAGIC);
1496 lsi = s2lsi(obt->obt_sb);
1497 if (!(lsi->lsi_flags | LSI_IR_CAPABLE))
1498 service_time += 2 * (CONNECTION_SWITCH_MAX +
1499 CONNECTION_SWITCH_INC);
1500 if (service_time > obd->obd_recovery_timeout && !new_client)
1501 extend_recovery_timer(obd, service_time, false);
1504 /** Health checking routines */
1505 static inline int exp_connect_healthy(struct obd_export *exp)
1507 return (exp->exp_in_recovery);
1510 /** if export done req_replay or has replay in queue */
1511 static inline int exp_req_replay_healthy(struct obd_export *exp)
1513 return (!exp->exp_req_replay_needed ||
1514 cfs_atomic_read(&exp->exp_replay_count) > 0);
1516 /** if export done lock_replay or has replay in queue */
1517 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1519 return (!exp->exp_lock_replay_needed ||
1520 cfs_atomic_read(&exp->exp_replay_count) > 0);
1523 static inline int exp_vbr_healthy(struct obd_export *exp)
1525 return (!exp->exp_vbr_failed);
1528 static inline int exp_finished(struct obd_export *exp)
1530 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1533 /** Checking routines for recovery */
1534 static int check_for_clients(struct obd_device *obd)
1536 unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1538 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1540 LASSERT(clnts <= obd->obd_max_recoverable_clients);
1541 if (obd->obd_no_conn == 0 &&
1542 clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1547 static int check_for_next_transno(struct obd_device *obd)
1549 struct ptlrpc_request *req = NULL;
1550 int wake_up = 0, connected, completed, queue_len;
1551 __u64 next_transno, req_transno;
1554 cfs_spin_lock(&obd->obd_recovery_task_lock);
1555 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1556 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1557 struct ptlrpc_request, rq_list);
1558 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1563 connected = cfs_atomic_read(&obd->obd_connected_clients);
1564 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1565 queue_len = obd->obd_requests_queued_for_recovery;
1566 next_transno = obd->obd_next_recovery_transno;
1568 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1569 "req_transno: "LPU64", next_transno: "LPU64"\n",
1570 obd->obd_max_recoverable_clients, connected, completed,
1571 queue_len, req_transno, next_transno);
1573 if (obd->obd_abort_recovery) {
1574 CDEBUG(D_HA, "waking for aborted recovery\n");
1576 } else if (obd->obd_recovery_expired) {
1577 CDEBUG(D_HA, "waking for expired recovery\n");
1579 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1580 CDEBUG(D_HA, "waking for completed recovery\n");
1582 } else if (req_transno == next_transno) {
1583 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1585 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1587 /** handle gaps occured due to lost reply or VBR */
1588 LASSERTF(req_transno >= next_transno,
1589 "req_transno: "LPU64", next_transno: "LPU64"\n",
1590 req_transno, next_transno);
1591 if (req_transno > obd->obd_last_committed &&
1592 !obd->obd_version_recov)
1595 "%s: waking for gap in transno, VBR is %s (skip: "
1596 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1597 ", last_committed: "LPD64")\n",
1598 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1599 next_transno, queue_len, completed, connected,
1600 req_transno, obd->obd_last_committed);
1601 obd->obd_next_recovery_transno = req_transno;
1603 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1604 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1605 " by fail_lock, waking up ("LPD64")\n", next_transno);
1606 obd->obd_next_recovery_transno = req_transno;
1609 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1613 static int check_for_next_lock(struct obd_device *obd)
1617 cfs_spin_lock(&obd->obd_recovery_task_lock);
1618 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1619 CDEBUG(D_HA, "waking for next lock\n");
1621 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1622 CDEBUG(D_HA, "waking for completed lock replay\n");
1624 } else if (obd->obd_abort_recovery) {
1625 CDEBUG(D_HA, "waking for aborted recovery\n");
1627 } else if (obd->obd_recovery_expired) {
1628 CDEBUG(D_HA, "waking for expired recovery\n");
1631 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1637 * wait for recovery events,
1638 * check its status with help of check_routine
1639 * evict dead clients via health_check
1641 static int target_recovery_overseer(struct obd_device *obd,
1642 int (*check_routine)(struct obd_device *),
1643 int (*health_check)(struct obd_export *))
1646 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1647 if (obd->obd_abort_recovery) {
1648 CWARN("recovery is aborted, evict exports in recovery\n");
1649 /** evict exports which didn't finish recovery yet */
1650 class_disconnect_stale_exports(obd, exp_finished);
1652 } else if (obd->obd_recovery_expired) {
1653 obd->obd_recovery_expired = 0;
1654 /** If some clients died being recovered, evict them */
1656 "recovery is timed out, evict stale exports\n");
1657 /** evict cexports with no replay in queue, they are stalled */
1658 class_disconnect_stale_exports(obd, health_check);
1659 /** continue with VBR */
1660 cfs_spin_lock(&obd->obd_dev_lock);
1661 obd->obd_version_recov = 1;
1662 cfs_spin_unlock(&obd->obd_dev_lock);
1664 * reset timer, recovery will proceed with versions now,
1665 * timeout is set just to handle reconnection delays
1667 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1668 /** Wait for recovery events again, after evicting bad clients */
1674 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1676 struct ptlrpc_request *req = NULL;
1679 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1680 obd->obd_next_recovery_transno);
1682 if (target_recovery_overseer(obd, check_for_next_transno,
1683 exp_req_replay_healthy)) {
1684 abort_req_replay_queue(obd);
1685 abort_lock_replay_queue(obd);
1688 cfs_spin_lock(&obd->obd_recovery_task_lock);
1689 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1690 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1691 struct ptlrpc_request, rq_list);
1692 cfs_list_del_init(&req->rq_list);
1693 obd->obd_requests_queued_for_recovery--;
1694 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1696 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1697 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1698 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1699 /** evict exports failed VBR */
1700 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1705 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1707 struct ptlrpc_request *req = NULL;
1709 CDEBUG(D_HA, "Waiting for lock\n");
1710 if (target_recovery_overseer(obd, check_for_next_lock,
1711 exp_lock_replay_healthy))
1712 abort_lock_replay_queue(obd);
1714 cfs_spin_lock(&obd->obd_recovery_task_lock);
1715 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1716 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1717 struct ptlrpc_request, rq_list);
1718 cfs_list_del_init(&req->rq_list);
1719 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1721 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1722 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1723 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1724 /** evict exports failed VBR */
1725 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1730 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1732 struct ptlrpc_request *req = NULL;
1734 cfs_spin_lock(&obd->obd_recovery_task_lock);
1735 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1736 req = cfs_list_entry(obd->obd_final_req_queue.next,
1737 struct ptlrpc_request, rq_list);
1738 cfs_list_del_init(&req->rq_list);
1739 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1740 if (req->rq_export->exp_in_recovery) {
1741 cfs_spin_lock(&req->rq_export->exp_lock);
1742 req->rq_export->exp_in_recovery = 0;
1743 cfs_spin_unlock(&req->rq_export->exp_lock);
1746 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1751 static int handle_recovery_req(struct ptlrpc_thread *thread,
1752 struct ptlrpc_request *req,
1753 svc_handler_t handler)
1759 * export can be evicted during recovery, no need to handle replays for
1760 * it after that, discard such request silently
1762 if (req->rq_export->exp_disconnected)
1763 GOTO(reqcopy_put, rc = 0);
1765 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1767 CERROR("Failure to initialize session: %d\n", rc);
1768 GOTO(reqcopy_put, rc);
1771 req->rq_recov_session.lc_thread = thread;
1772 lu_context_enter(&req->rq_recov_session);
1773 req->rq_svc_thread = thread;
1774 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1776 /* thread context */
1777 lu_context_enter(&thread->t_env->le_ctx);
1779 lu_context_exit(&thread->t_env->le_ctx);
1781 lu_context_exit(&req->rq_recov_session);
1782 lu_context_fini(&req->rq_recov_session);
1783 /* don't reset timer for final stage */
1784 if (!exp_finished(req->rq_export)) {
1785 int to = obd_timeout;
1788 * Add request timeout to the recovery time so next request from
1789 * this client may come in recovery time
1792 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1793 /* If the server sent early reply for this request,
1794 * the client will recalculate the timeout according to
1795 * current server estimate service time, so we will
1796 * use the maxium timeout here for waiting the client
1797 * sending the next req */
1798 to = max((int)at_est2timeout(
1799 at_get(&svc->srv_at_estimate)),
1800 (int)lustre_msg_get_timeout(req->rq_reqmsg));
1801 /* Add net_latency (see ptlrpc_replay_req) */
1802 to += lustre_msg_get_service_time(req->rq_reqmsg);
1804 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1810 static int target_recovery_thread(void *arg)
1812 struct lu_target *lut = arg;
1813 struct obd_device *obd = lut->lut_obd;
1814 struct ptlrpc_request *req;
1815 struct target_recovery_data *trd = &obd->obd_recovery_data;
1816 unsigned long delta;
1817 unsigned long flags;
1819 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1823 cfs_daemonize_ctxt("tgt_recov");
1825 SIGNAL_MASK_LOCK(current, flags);
1826 sigfillset(¤t->blocked);
1828 SIGNAL_MASK_UNLOCK(current, flags);
1830 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1834 thread->t_env = &env;
1835 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1836 env.le_ctx.lc_thread = thread;
1837 thread->t_data = NULL;
1838 thread->t_watchdog = NULL;
1840 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1842 trd->trd_processing_task = cfs_curproc_pid();
1844 cfs_spin_lock(&obd->obd_dev_lock);
1845 obd->obd_recovering = 1;
1846 cfs_spin_unlock(&obd->obd_dev_lock);
1847 cfs_complete(&trd->trd_starting);
1849 /* first of all, we have to know the first transno to replay */
1850 if (target_recovery_overseer(obd, check_for_clients,
1851 exp_connect_healthy)) {
1852 abort_req_replay_queue(obd);
1853 abort_lock_replay_queue(obd);
1856 /* next stage: replay requests */
1858 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1859 cfs_atomic_read(&obd->obd_req_replay_clients),
1860 obd->obd_next_recovery_transno);
1861 while ((req = target_next_replay_req(obd))) {
1862 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1863 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1864 lustre_msg_get_transno(req->rq_reqmsg),
1865 libcfs_nid2str(req->rq_peer.nid));
1866 handle_recovery_req(thread, req,
1867 trd->trd_recovery_handler);
1869 * bz18031: increase next_recovery_transno before
1870 * target_request_copy_put() will drop exp_rpc reference
1872 cfs_spin_lock(&obd->obd_recovery_task_lock);
1873 obd->obd_next_recovery_transno++;
1874 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1875 target_exp_dequeue_req_replay(req);
1876 target_request_copy_put(req);
1877 obd->obd_replayed_requests++;
1881 * The second stage: replay locks
1883 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1884 cfs_atomic_read(&obd->obd_lock_replay_clients));
1885 while ((req = target_next_replay_lock(obd))) {
1886 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1887 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1888 libcfs_nid2str(req->rq_peer.nid));
1889 handle_recovery_req(thread, req,
1890 trd->trd_recovery_handler);
1891 target_request_copy_put(req);
1892 obd->obd_replayed_locks++;
1896 * The third stage: reply on final pings, at this moment all clients
1897 * must have request in final queue
1899 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1900 /** Update server last boot epoch */
1901 lut_boot_epoch_update(lut);
1902 /* We drop recoverying flag to forward all new requests
1903 * to regular mds_handle() since now */
1904 cfs_spin_lock(&obd->obd_dev_lock);
1905 obd->obd_recovering = obd->obd_abort_recovery = 0;
1906 cfs_spin_unlock(&obd->obd_dev_lock);
1907 cfs_spin_lock(&obd->obd_recovery_task_lock);
1908 target_cancel_recovery_timer(obd);
1909 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1910 while ((req = target_next_final_ping(obd))) {
1911 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1912 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1913 libcfs_nid2str(req->rq_peer.nid));
1914 handle_recovery_req(thread, req,
1915 trd->trd_recovery_handler);
1916 target_request_copy_put(req);
1919 delta = (jiffies - delta) / CFS_HZ;
1920 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1921 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1922 if (delta > OBD_RECOVERY_TIME_SOFT) {
1923 CWARN("too long recovery - read logs\n");
1924 libcfs_debug_dumplog();
1927 target_finish_recovery(obd);
1929 lu_context_fini(&env.le_ctx);
1930 trd->trd_processing_task = 0;
1931 cfs_complete(&trd->trd_finishing);
1935 static int target_start_recovery_thread(struct lu_target *lut,
1936 svc_handler_t handler)
1938 struct obd_device *obd = lut->lut_obd;
1940 struct target_recovery_data *trd = &obd->obd_recovery_data;
1942 memset(trd, 0, sizeof(*trd));
1943 cfs_init_completion(&trd->trd_starting);
1944 cfs_init_completion(&trd->trd_finishing);
1945 trd->trd_recovery_handler = handler;
1947 if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
1948 cfs_wait_for_completion(&trd->trd_starting);
1949 LASSERT(obd->obd_recovering != 0);
1956 void target_stop_recovery_thread(struct obd_device *obd)
1958 if (obd->obd_recovery_data.trd_processing_task > 0) {
1959 struct target_recovery_data *trd = &obd->obd_recovery_data;
1960 /** recovery can be done but postrecovery is not yet */
1961 cfs_spin_lock(&obd->obd_dev_lock);
1962 if (obd->obd_recovering) {
1963 CERROR("%s: Aborting recovery\n", obd->obd_name);
1964 obd->obd_abort_recovery = 1;
1965 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1967 cfs_spin_unlock(&obd->obd_dev_lock);
1968 cfs_wait_for_completion(&trd->trd_finishing);
1972 void target_recovery_fini(struct obd_device *obd)
1974 class_disconnect_exports(obd);
1975 target_stop_recovery_thread(obd);
1976 target_cleanup_recovery(obd);
1978 EXPORT_SYMBOL(target_recovery_fini);
1980 static void target_recovery_expired(unsigned long castmeharder)
1982 struct obd_device *obd = (struct obd_device *)castmeharder;
1983 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1984 " after %lds (%d clients connected)\n",
1985 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1986 cfs_time_current_sec()- obd->obd_recovery_start,
1987 cfs_atomic_read(&obd->obd_connected_clients));
1989 obd->obd_recovery_expired = 1;
1990 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1993 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1995 struct obd_device *obd = lut->lut_obd;
1996 if (obd->obd_max_recoverable_clients == 0) {
1997 /** Update server last boot epoch */
1998 lut_boot_epoch_update(lut);
2002 CWARN("RECOVERY: service %s, %d recoverable clients, "
2003 "last_transno "LPU64"\n", obd->obd_name,
2004 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2005 LASSERT(obd->obd_stopping == 0);
2006 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2007 obd->obd_recovery_start = 0;
2008 obd->obd_recovery_end = 0;
2010 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2011 target_start_recovery_thread(lut, handler);
2013 EXPORT_SYMBOL(target_recovery_init);
2017 static int target_process_req_flags(struct obd_device *obd,
2018 struct ptlrpc_request *req)
2020 struct obd_export *exp = req->rq_export;
2021 LASSERT(exp != NULL);
2022 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2023 /* client declares he's ready to replay locks */
2024 cfs_spin_lock(&exp->exp_lock);
2025 if (exp->exp_req_replay_needed) {
2026 exp->exp_req_replay_needed = 0;
2027 cfs_spin_unlock(&exp->exp_lock);
2029 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2030 cfs_atomic_dec(&obd->obd_req_replay_clients);
2032 cfs_spin_unlock(&exp->exp_lock);
2035 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2036 /* client declares he's ready to complete recovery
2037 * so, we put the request on th final queue */
2038 cfs_spin_lock(&exp->exp_lock);
2039 if (exp->exp_lock_replay_needed) {
2040 exp->exp_lock_replay_needed = 0;
2041 cfs_spin_unlock(&exp->exp_lock);
2043 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2044 cfs_atomic_dec(&obd->obd_lock_replay_clients);
2046 cfs_spin_unlock(&exp->exp_lock);
2052 int target_queue_recovery_request(struct ptlrpc_request *req,
2053 struct obd_device *obd)
2057 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2060 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2061 /* Processing the queue right now, don't re-add. */
2065 target_process_req_flags(obd, req);
2067 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2068 /* client declares he's ready to complete recovery
2069 * so, we put the request on th final queue */
2070 target_request_copy_get(req);
2071 DEBUG_REQ(D_HA, req, "queue final req");
2072 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2073 cfs_spin_lock(&obd->obd_recovery_task_lock);
2074 if (obd->obd_recovering) {
2075 cfs_list_add_tail(&req->rq_list,
2076 &obd->obd_final_req_queue);
2078 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2079 target_request_copy_put(req);
2080 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2082 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2085 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2086 /* client declares he's ready to replay locks */
2087 target_request_copy_get(req);
2088 DEBUG_REQ(D_HA, req, "queue lock replay req");
2089 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2090 cfs_spin_lock(&obd->obd_recovery_task_lock);
2091 LASSERT(obd->obd_recovering);
2092 /* usually due to recovery abort */
2093 if (!req->rq_export->exp_in_recovery) {
2094 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2095 target_request_copy_put(req);
2098 LASSERT(req->rq_export->exp_lock_replay_needed);
2099 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2100 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2104 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2105 * (i.e. buflens etc are in my own byte order), but type-dependent
2106 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2109 CFS_INIT_LIST_HEAD(&req->rq_list);
2110 DEBUG_REQ(D_HA, req, "not queueing");
2114 /* If we're processing the queue, we want don't want to queue this
2117 * Also, if this request has a transno less than the one we're waiting
2118 * for, we should process it now. It could (and currently always will)
2119 * be an open request for a descriptor that was opened some time ago.
2121 * Also, a resent, replayed request that has already been
2122 * handled will pass through here and be processed immediately.
2124 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2125 obd->obd_next_recovery_transno, transno);
2126 cfs_spin_lock(&obd->obd_recovery_task_lock);
2127 if (transno < obd->obd_next_recovery_transno) {
2128 /* Processing the queue right now, don't re-add. */
2129 LASSERT(cfs_list_empty(&req->rq_list));
2130 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2133 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2135 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2138 target_request_copy_get(req);
2139 if (!req->rq_export->exp_in_recovery) {
2140 target_request_copy_put(req);
2143 LASSERT(req->rq_export->exp_req_replay_needed);
2145 if (target_exp_enqueue_req_replay(req)) {
2146 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2147 target_request_copy_put(req);
2152 cfs_spin_lock(&obd->obd_recovery_task_lock);
2153 LASSERT(obd->obd_recovering);
2154 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2155 struct ptlrpc_request *reqiter =
2156 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2158 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2159 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2164 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2166 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2167 "has been claimed by another client");
2168 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2169 target_exp_dequeue_req_replay(req);
2170 target_request_copy_put(req);
2176 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2178 obd->obd_requests_queued_for_recovery++;
2179 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2180 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2185 * Packs current SLV and Limit into \a req.
2187 int target_pack_pool_reply(struct ptlrpc_request *req)
2189 struct obd_device *obd;
2193 * Check that we still have all structures alive as this may
2194 * be some late rpc in shutdown time.
2196 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2197 !exp_connect_lru_resize(req->rq_export))) {
2198 lustre_msg_set_slv(req->rq_repmsg, 0);
2199 lustre_msg_set_limit(req->rq_repmsg, 0);
2204 * OBD is alive here as export is alive, which we checked above.
2206 obd = req->rq_export->exp_obd;
2208 cfs_read_lock(&obd->obd_pool_lock);
2209 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2210 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2211 cfs_read_unlock(&obd->obd_pool_lock);
2216 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2218 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2219 DEBUG_REQ(D_ERROR, req, "dropping reply");
2224 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2225 req->rq_status = rc;
2226 return (ptlrpc_send_error(req, 1));
2228 DEBUG_REQ(D_NET, req, "sending reply");
2231 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2234 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2237 struct ptlrpc_reply_state *rs;
2238 struct obd_export *exp;
2239 struct ptlrpc_service *svc;
2242 if (req->rq_no_reply) {
2247 svc = req->rq_rqbd->rqbd_service;
2248 rs = req->rq_reply_state;
2249 if (rs == NULL || !rs->rs_difficult) {
2251 target_send_reply_msg (req, rc, fail_id);
2256 /* must be an export if locks saved */
2257 LASSERT (req->rq_export != NULL);
2258 /* req/reply consistent */
2259 LASSERT (rs->rs_service == svc);
2262 LASSERT (!rs->rs_scheduled);
2263 LASSERT (!rs->rs_scheduled_ever);
2264 LASSERT (!rs->rs_handled);
2265 LASSERT (!rs->rs_on_net);
2266 LASSERT (rs->rs_export == NULL);
2267 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2268 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2270 exp = class_export_get (req->rq_export);
2272 /* disable reply scheduling while I'm setting up */
2273 rs->rs_scheduled = 1;
2275 rs->rs_xid = req->rq_xid;
2276 rs->rs_transno = req->rq_transno;
2277 rs->rs_export = exp;
2278 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2280 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2281 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2282 rs->rs_transno, exp->exp_last_committed);
2283 if (rs->rs_transno > exp->exp_last_committed) {
2284 /* not committed already */
2285 cfs_list_add_tail(&rs->rs_obd_list,
2286 &exp->exp_uncommitted_replies);
2288 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2290 cfs_spin_lock(&exp->exp_lock);
2291 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2292 cfs_spin_unlock(&exp->exp_lock);
2294 netrc = target_send_reply_msg (req, rc, fail_id);
2296 cfs_spin_lock(&svc->srv_rs_lock);
2298 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2301 /* error sending: reply is off the net. Also we need +1
2302 * reply ref until ptlrpc_handle_rs() is done
2303 * with the reply state (if the send was successful, there
2304 * would have been +1 ref for the net, which
2305 * reply_out_callback leaves alone) */
2307 ptlrpc_rs_addref(rs);
2310 cfs_spin_lock(&rs->rs_lock);
2311 if (rs->rs_transno <= exp->exp_last_committed ||
2312 (!rs->rs_on_net && !rs->rs_no_ack) ||
2313 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2314 cfs_list_empty(&rs->rs_obd_list)) {
2315 CDEBUG(D_HA, "Schedule reply immediately\n");
2316 ptlrpc_dispatch_difficult_reply(rs);
2318 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2319 rs->rs_scheduled = 0; /* allow notifier to schedule */
2321 cfs_spin_unlock(&rs->rs_lock);
2322 cfs_spin_unlock(&svc->srv_rs_lock);
2326 int target_handle_ping(struct ptlrpc_request *req)
2328 obd_ping(req->rq_export);
2329 return req_capsule_server_pack(&req->rq_pill);
2332 void target_committed_to_req(struct ptlrpc_request *req)
2334 struct obd_export *exp = req->rq_export;
2336 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2337 lustre_msg_set_last_committed(req->rq_repmsg,
2338 exp->exp_last_committed);
2340 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2341 "%d)", exp->exp_obd->obd_no_transno,
2342 req->rq_repmsg == NULL);
2344 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2345 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2347 EXPORT_SYMBOL(target_committed_to_req);
2349 int target_handle_qc_callback(struct ptlrpc_request *req)
2351 struct obd_quotactl *oqctl;
2352 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2354 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2355 if (oqctl == NULL) {
2356 CERROR("Can't unpack obd_quotactl\n");
2360 cli->cl_qchk_stat = oqctl->qc_stat;
2365 #ifdef HAVE_QUOTA_SUPPORT
2366 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2369 struct obd_device *obd = req->rq_export->exp_obd;
2370 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2371 struct obd_device_target *obt;
2372 struct lustre_quota_ctxt *qctxt;
2373 struct qunit_data *qdata = NULL;
2377 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2380 rc = req_capsule_server_pack(&req->rq_pill);
2382 CERROR("packing reply failed!: rc = %d\n", rc);
2386 LASSERT(req->rq_export);
2388 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2389 if (IS_ERR(qdata)) {
2390 rc = PTR_ERR(qdata);
2391 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2392 req->rq_status = rc;
2396 /* we use the observer */
2397 if (obd_pin_observer(obd, &lov_obd) ||
2398 obd_pin_observer(lov_obd, &master_obd)) {
2399 CERROR("Can't find the observer, it is recovering\n");
2400 req->rq_status = -EAGAIN;
2404 obt = &master_obd->u.obt;
2405 qctxt = &obt->obt_qctxt;
2407 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2408 /* quota_type has not been processed yet, return EAGAIN
2409 * until we know whether or not quotas are supposed to
2411 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2413 req->rq_status = -EAGAIN;
2417 cfs_down_read(&obt->obt_rwsem);
2418 if (qctxt->lqc_lqs_hash == NULL) {
2419 cfs_up_read(&obt->obt_rwsem);
2420 /* quota_type has not been processed yet, return EAGAIN
2421 * until we know whether or not quotas are supposed to
2423 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2425 req->rq_status = -EAGAIN;
2429 LASSERT(qctxt->lqc_handler);
2430 rc = qctxt->lqc_handler(master_obd, qdata,
2431 lustre_msg_get_opc(req->rq_reqmsg));
2432 cfs_up_read(&obt->obt_rwsem);
2433 if (rc && rc != -EDQUOT)
2434 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2435 "dqacq/dqrel failed! (rc:%d)\n", rc);
2436 req->rq_status = rc;
2438 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2440 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2444 /* Block the quota req. b=14840 */
2445 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2450 obd_unpin_observer(lov_obd);
2452 obd_unpin_observer(obd);
2454 rc = ptlrpc_reply(req);
2458 #endif /* !__KERNEL__ */
2460 #endif /* HAVE_QUOTA_SUPPORT */
2462 ldlm_mode_t lck_compat_array[] = {
2463 [LCK_EX] LCK_COMPAT_EX,
2464 [LCK_PW] LCK_COMPAT_PW,
2465 [LCK_PR] LCK_COMPAT_PR,
2466 [LCK_CW] LCK_COMPAT_CW,
2467 [LCK_CR] LCK_COMPAT_CR,
2468 [LCK_NL] LCK_COMPAT_NL,
2469 [LCK_GROUP] LCK_COMPAT_GROUP,
2470 [LCK_COS] LCK_COMPAT_COS,
2474 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2475 * not escape to the user level.
2477 int ldlm_error2errno(ldlm_error_t error)
2485 case ELDLM_LOCK_CHANGED:
2488 case ELDLM_LOCK_ABORTED:
2491 case ELDLM_LOCK_REPLACED:
2494 case ELDLM_NO_LOCK_DATA:
2497 case ELDLM_NAMESPACE_EXISTS:
2500 case ELDLM_BAD_NAMESPACE:
2504 if (((int)error) < 0) /* cast to signed type */
2505 result = error; /* as ldlm_error_t can be unsigned */
2507 CERROR("Invalid DLM result code: %d\n", error);
2513 EXPORT_SYMBOL(ldlm_error2errno);
2516 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2518 ldlm_error_t ldlm_errno2error(int err_no)
2527 error = ELDLM_LOCK_CHANGED;
2530 error = ELDLM_LOCK_ABORTED;
2533 error = ELDLM_LOCK_REPLACED;
2536 error = ELDLM_NO_LOCK_DATA;
2539 error = ELDLM_NAMESPACE_EXISTS;
2542 error = ELDLM_BAD_NAMESPACE;
2549 EXPORT_SYMBOL(ldlm_errno2error);
2551 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2552 void ldlm_dump_export_locks(struct obd_export *exp)
2554 cfs_spin_lock(&exp->exp_locks_list_guard);
2555 if (!cfs_list_empty(&exp->exp_locks_list)) {
2556 struct ldlm_lock *lock;
2558 CERROR("dumping locks for export %p,"
2559 "ignore if the unmount doesn't hang\n", exp);
2560 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2561 ldlm_lock_dump(D_ERROR, lock, 0);
2563 cfs_spin_unlock(&exp->exp_locks_list_guard);
2567 static int target_bulk_timeout(void *data)
2570 /* We don't fail the connection here, because having the export
2571 * killed makes the (vital) call to commitrw very sad.
2576 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2578 return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2581 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2582 struct l_wait_info *lwi)
2584 struct ptlrpc_request *req = desc->bd_req;
2588 /* Check if there is eviction in progress, and if so, wait for
2590 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2591 *lwi = LWI_INTR(NULL, NULL);
2592 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2593 !cfs_atomic_read(&exp->exp_obd->
2594 obd_evict_inprogress),
2598 /* Check if client was evicted or tried to reconnect already */
2599 if (exp->exp_failed || exp->exp_abort_active_req) {
2602 if (desc->bd_type == BULK_PUT_SINK)
2603 rc = sptlrpc_svc_wrap_bulk(req, desc);
2605 rc = ptlrpc_start_bulk_transfer(desc);
2608 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2609 ptlrpc_abort_bulk(desc);
2610 } else if (rc == 0) {
2611 time_t start = cfs_time_current_sec();
2613 long timeoutl = req->rq_deadline - cfs_time_current_sec();
2614 cfs_duration_t timeout = timeoutl <= 0 ?
2615 CFS_TICK : cfs_time_seconds(timeoutl);
2616 *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2617 cfs_time_seconds(1),
2618 target_bulk_timeout,
2620 rc = l_wait_event(desc->bd_waitq,
2621 !ptlrpc_server_bulk_active(desc) ||
2623 exp->exp_abort_active_req,
2625 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2626 /* Wait again if we changed deadline */
2627 } while ((rc == -ETIMEDOUT) &&
2628 (req->rq_deadline > cfs_time_current_sec()));
2630 if (rc == -ETIMEDOUT) {
2631 DEBUG_REQ(D_ERROR, req,
2632 "timeout on bulk %s after %ld%+lds",
2634 req->rq_deadline - start,
2635 cfs_time_current_sec() -
2637 ptlrpc_abort_bulk(desc);
2638 } else if (exp->exp_failed) {
2639 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2642 ptlrpc_abort_bulk(desc);
2643 } else if (exp->exp_abort_active_req) {
2644 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2646 /* we don't reply anyway */
2648 ptlrpc_abort_bulk(desc);
2649 } else if (!desc->bd_success ||
2650 desc->bd_nob_transferred != desc->bd_nob) {
2651 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2653 "truncated" : "network error on",
2655 desc->bd_nob_transferred,
2657 /* XXX should this be a different errno? */
2659 } else if (desc->bd_type == BULK_GET_SINK) {
2660 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2663 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2664 bulk2type(desc), rc);
2669 EXPORT_SYMBOL(target_bulk_io);