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);
196 void client_destroy_import(struct obd_import *imp)
198 /* drop security policy instance after all rpc finished/aborted
199 * to let all busy contexts be released. */
200 class_import_get(imp);
201 class_destroy_import(imp);
202 sptlrpc_import_sec_put(imp);
203 class_import_put(imp);
206 /* configure an RPC client OBD device
211 * 3 - inactive-on-startup
213 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
215 struct client_obd *cli = &obddev->u.cli;
216 struct obd_import *imp;
217 struct obd_uuid server_uuid;
218 int rq_portal, rp_portal, connect_op;
219 char *name = obddev->obd_type->typ_name;
220 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
224 /* In a more perfect world, we would hang a ptlrpc_client off of
225 * obd_type and just use the values from there. */
226 if (!strcmp(name, LUSTRE_OSC_NAME)) {
227 rq_portal = OST_REQUEST_PORTAL;
228 rp_portal = OSC_REPLY_PORTAL;
229 connect_op = OST_CONNECT;
230 cli->cl_sp_me = LUSTRE_SP_CLI;
231 cli->cl_sp_to = LUSTRE_SP_OST;
232 ns_type = LDLM_NS_TYPE_OSC;
234 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
235 rq_portal = MDS_REQUEST_PORTAL;
236 rp_portal = MDC_REPLY_PORTAL;
237 connect_op = MDS_CONNECT;
238 cli->cl_sp_me = LUSTRE_SP_CLI;
239 cli->cl_sp_to = LUSTRE_SP_MDT;
240 ns_type = LDLM_NS_TYPE_MDC;
242 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
243 rq_portal = MGS_REQUEST_PORTAL;
244 rp_portal = MGC_REPLY_PORTAL;
245 connect_op = MGS_CONNECT;
246 cli->cl_sp_me = LUSTRE_SP_MGC;
247 cli->cl_sp_to = LUSTRE_SP_MGS;
248 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
249 ns_type = LDLM_NS_TYPE_MGC;
252 CERROR("unknown client OBD type \"%s\", can't setup\n",
257 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
258 CERROR("requires a TARGET UUID\n");
262 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
263 CERROR("client UUID must be less than 38 characters\n");
267 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
268 CERROR("setup requires a SERVER UUID\n");
272 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
273 CERROR("target UUID must be less than 38 characters\n");
277 cfs_init_rwsem(&cli->cl_sem);
278 cfs_sema_init(&cli->cl_mgc_sem, 1);
279 cli->cl_conn_count = 0;
280 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
281 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
282 sizeof(server_uuid)));
285 cli->cl_avail_grant = 0;
286 /* FIXME: should limit this for the sum of all cl_dirty_max */
287 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
288 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
289 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
290 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
291 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
292 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
293 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
294 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
295 client_obd_list_lock_init(&cli->cl_loi_list_lock);
296 cli->cl_r_in_flight = 0;
297 cli->cl_w_in_flight = 0;
299 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
300 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
301 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
302 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
303 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
304 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
305 cfs_waitq_init(&cli->cl_destroy_waitq);
306 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
307 #ifdef ENABLE_CHECKSUM
308 /* Turn on checksumming by default. */
309 cli->cl_checksum = 1;
311 * The supported checksum types will be worked out at connect time
312 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
315 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
317 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
319 /* This value may be changed at connect time in
320 ptlrpc_connect_interpret. */
321 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
322 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
324 if (!strcmp(name, LUSTRE_MDC_NAME)) {
325 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
326 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
327 cli->cl_max_rpcs_in_flight = 2;
328 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
329 cli->cl_max_rpcs_in_flight = 3;
330 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
331 cli->cl_max_rpcs_in_flight = 4;
333 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
338 CERROR("ldlm_get_ref failed: %d\n", rc);
342 ptlrpc_init_client(rq_portal, rp_portal, name,
343 &obddev->obd_ldlm_client);
345 imp = class_new_import(obddev);
347 GOTO(err_ldlm, rc = -ENOENT);
348 imp->imp_client = &obddev->obd_ldlm_client;
349 imp->imp_connect_op = connect_op;
350 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
351 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
352 LUSTRE_CFG_BUFLEN(lcfg, 1));
353 class_import_put(imp);
355 rc = client_import_add_conn(imp, &server_uuid, 1);
357 CERROR("can't add initial connection\n");
358 GOTO(err_import, rc);
361 cli->cl_import = imp;
362 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
363 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
364 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
366 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
367 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
368 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
369 name, obddev->obd_name,
370 cli->cl_target_uuid.uuid);
371 cfs_spin_lock(&imp->imp_lock);
372 imp->imp_deactive = 1;
373 cfs_spin_unlock(&imp->imp_lock);
377 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
378 LDLM_NAMESPACE_CLIENT,
379 LDLM_NAMESPACE_GREEDY,
381 if (obddev->obd_namespace == NULL) {
382 CERROR("Unable to create client namespace - %s\n",
384 GOTO(err_import, rc = -ENOMEM);
387 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
392 class_destroy_import(imp);
400 int client_obd_cleanup(struct obd_device *obddev)
404 ldlm_namespace_free_post(obddev->obd_namespace);
405 obddev->obd_namespace = NULL;
411 /* ->o_connect() method for client side (OSC and MDC and MGC) */
412 int client_connect_import(const struct lu_env *env,
413 struct obd_export **exp,
414 struct obd_device *obd, struct obd_uuid *cluuid,
415 struct obd_connect_data *data, void *localdata)
417 struct client_obd *cli = &obd->u.cli;
418 struct obd_import *imp = cli->cl_import;
419 struct obd_connect_data *ocd;
420 struct lustre_handle conn = { 0 };
425 cfs_down_write(&cli->cl_sem);
426 if (cli->cl_conn_count > 0 )
427 GOTO(out_sem, rc = -EALREADY);
429 rc = class_connect(&conn, obd, cluuid);
433 cli->cl_conn_count++;
434 *exp = class_conn2export(&conn);
436 LASSERT(obd->obd_namespace);
438 imp->imp_dlm_handle = conn;
439 rc = ptlrpc_init_import(imp);
443 ocd = &imp->imp_connect_data;
446 imp->imp_connect_flags_orig = data->ocd_connect_flags;
449 rc = ptlrpc_connect_import(imp, NULL);
451 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
454 LASSERT((*exp)->exp_connection);
457 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
458 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
459 data->ocd_connect_flags, ocd->ocd_connect_flags);
460 data->ocd_connect_flags = ocd->ocd_connect_flags;
463 ptlrpc_pinger_add_import(imp);
469 cli->cl_conn_count--;
470 class_disconnect(*exp);
474 cfs_up_write(&cli->cl_sem);
479 int client_disconnect_export(struct obd_export *exp)
481 struct obd_device *obd = class_exp2obd(exp);
482 struct client_obd *cli;
483 struct obd_import *imp;
488 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
489 exp, exp ? exp->exp_handle.h_cookie : -1);
494 imp = cli->cl_import;
496 cfs_down_write(&cli->cl_sem);
497 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
500 if (!cli->cl_conn_count) {
501 CERROR("disconnecting disconnected device (%s)\n",
503 GOTO(out_disconnect, rc = -EINVAL);
506 cli->cl_conn_count--;
507 if (cli->cl_conn_count)
508 GOTO(out_disconnect, rc = 0);
510 /* Mark import deactivated now, so we don't try to reconnect if any
511 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
512 * fully deactivate the import, or that would drop all requests. */
513 cfs_spin_lock(&imp->imp_lock);
514 imp->imp_deactive = 1;
515 cfs_spin_unlock(&imp->imp_lock);
517 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
518 * delete it regardless. (It's safe to delete an import that was
520 (void)ptlrpc_pinger_del_import(imp);
522 if (obd->obd_namespace != NULL) {
523 /* obd_force == local only */
524 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
525 obd->obd_force ? LCF_LOCAL : 0, NULL);
526 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
530 * there's no need to hold sem during disconnecting an import,
531 * and actually it may cause deadlock in gss.
533 cfs_up_write(&cli->cl_sem);
534 rc = ptlrpc_disconnect_import(imp, 0);
535 cfs_down_write(&cli->cl_sem);
537 ptlrpc_invalidate_import(imp);
539 if (imp->imp_rq_pool) {
540 ptlrpc_free_rq_pool(imp->imp_rq_pool);
541 imp->imp_rq_pool = NULL;
543 client_destroy_import(imp);
544 cli->cl_import = NULL;
549 /* use server style - class_disconnect should be always called for
551 err = class_disconnect(exp);
555 cfs_up_write(&cli->cl_sem);
560 int server_disconnect_export(struct obd_export *exp)
565 /* Disconnect early so that clients can't keep using export */
566 rc = class_disconnect(exp);
567 /* close import for avoid sending any requests */
568 if (exp->exp_imp_reverse)
569 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
571 if (exp->exp_obd->obd_namespace != NULL)
572 ldlm_cancel_locks_for_export(exp);
574 /* complete all outstanding replies */
575 cfs_spin_lock(&exp->exp_lock);
576 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
577 struct ptlrpc_reply_state *rs =
578 cfs_list_entry(exp->exp_outstanding_replies.next,
579 struct ptlrpc_reply_state, rs_exp_list);
580 struct ptlrpc_service *svc = rs->rs_service;
582 cfs_spin_lock(&svc->srv_rs_lock);
583 cfs_list_del_init(&rs->rs_exp_list);
584 cfs_spin_lock(&rs->rs_lock);
585 ptlrpc_schedule_difficult_reply(rs);
586 cfs_spin_unlock(&rs->rs_lock);
587 cfs_spin_unlock(&svc->srv_rs_lock);
589 cfs_spin_unlock(&exp->exp_lock);
594 /* --------------------------------------------------------------------------
595 * from old lib/target.c
596 * -------------------------------------------------------------------------- */
598 static int target_handle_reconnect(struct lustre_handle *conn,
599 struct obd_export *exp,
600 struct obd_uuid *cluuid)
603 if (exp->exp_connection && exp->exp_imp_reverse) {
604 struct lustre_handle *hdl;
605 hdl = &exp->exp_imp_reverse->imp_remote_handle;
606 /* Might be a re-connect after a partition. */
607 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
608 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
610 conn->cookie = exp->exp_handle.h_cookie;
611 /* target_handle_connect() treats EALREADY and
612 * -EALREADY differently. EALREADY means we are
613 * doing a valid reconnect from the same client. */
616 CERROR("%s reconnecting from %s, "
617 "handle mismatch (ours "LPX64", theirs "
618 LPX64")\n", cluuid->uuid,
619 exp->exp_connection->c_remote_uuid.uuid,
620 hdl->cookie, conn->cookie);
621 memset(conn, 0, sizeof *conn);
622 /* target_handle_connect() treats EALREADY and
623 * -EALREADY differently. -EALREADY is an error
624 * (same UUID, different handle). */
629 conn->cookie = exp->exp_handle.h_cookie;
630 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
631 cluuid->uuid, exp, conn->cookie);
635 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
638 struct obd_export *exp = cb_data;
640 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
641 obd->obd_name, exp->exp_client_uuid.uuid);
643 cfs_spin_lock(&exp->exp_lock);
644 exp->exp_need_sync = 0;
645 cfs_spin_unlock(&exp->exp_lock);
646 class_export_cb_put(exp);
648 EXPORT_SYMBOL(target_client_add_cb);
651 target_start_and_reset_recovery_timer(struct obd_device *obd,
652 struct ptlrpc_request *req,
655 int target_handle_connect(struct ptlrpc_request *req)
657 struct obd_device *target, *targref = NULL;
658 struct obd_export *export = NULL;
659 struct obd_import *revimp;
660 struct lustre_handle conn;
661 struct lustre_handle *tmp;
662 struct obd_uuid tgtuuid;
663 struct obd_uuid cluuid;
664 struct obd_uuid remote_uuid;
668 struct obd_connect_data *data, *tmpdata;
670 lnet_nid_t *client_nid = NULL;
673 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
675 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
677 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
678 GOTO(out, rc = -EINVAL);
681 obd_str2uuid(&tgtuuid, str);
682 target = class_uuid2obd(&tgtuuid);
684 target = class_name2obd(str);
686 if (!target || target->obd_stopping || !target->obd_set_up) {
687 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
688 " for connect (%s)\n", str,
689 !target ? "no target" :
690 (target->obd_stopping ? "stopping" :
692 GOTO(out, rc = -ENODEV);
695 if (target->obd_no_conn) {
696 LCONSOLE_WARN("%s: temporarily refusing client connection "
697 "from %s\n", target->obd_name,
698 libcfs_nid2str(req->rq_peer.nid));
699 GOTO(out, rc = -EAGAIN);
702 /* Make sure the target isn't cleaned up while we're here. Yes,
703 there's still a race between the above check and our incref here.
704 Really, class_uuid2obd should take the ref. */
705 targref = class_incref(target, __FUNCTION__, cfs_current());
708 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
710 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
711 GOTO(out, rc = -EINVAL);
714 obd_str2uuid(&cluuid, str);
716 /* XXX extract a nettype and format accordingly */
717 switch (sizeof(lnet_nid_t)) {
718 /* NB the casts only avoid compiler warnings */
720 snprintf(remote_uuid.uuid, sizeof remote_uuid,
721 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
724 snprintf(remote_uuid.uuid, sizeof remote_uuid,
725 "NET_%x_UUID", (__u32)req->rq_peer.nid);
731 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
733 GOTO(out, rc = -EPROTO);
737 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
739 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
741 GOTO(out, rc = -EPROTO);
743 rc = req_capsule_server_pack(&req->rq_pill);
747 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
749 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
750 "libclient connection attempt");
751 GOTO(out, rc = -EPROTO);
752 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
753 LUSTRE_VERSION_ALLOWED_OFFSET ||
754 data->ocd_version > LUSTRE_VERSION_CODE +
755 LUSTRE_VERSION_ALLOWED_OFFSET) {
756 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
757 "libclient connection attempt",
758 data->ocd_version < LUSTRE_VERSION_CODE ?
760 OBD_OCD_VERSION_MAJOR(data->ocd_version),
761 OBD_OCD_VERSION_MINOR(data->ocd_version),
762 OBD_OCD_VERSION_PATCH(data->ocd_version),
763 OBD_OCD_VERSION_FIX(data->ocd_version));
764 data = req_capsule_server_sized_get(&req->rq_pill,
766 offsetof(typeof(*data), ocd_version) +
767 sizeof(data->ocd_version));
769 data->ocd_connect_flags = OBD_CONNECT_VERSION;
770 data->ocd_version = LUSTRE_VERSION_CODE;
772 GOTO(out, rc = -EPROTO);
776 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
777 (data->ocd_connect_flags & OBD_CONNECT_MDS))
780 /* lctl gets a backstage, all-access pass. */
781 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
782 goto dont_check_exports;
784 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
788 /* we've found an export in the hash */
789 if (export->exp_connecting) { /* bug 9635, et. al. */
790 CWARN("%s: exp %p already connecting\n",
791 export->exp_obd->obd_name, export);
792 class_export_put(export);
795 } else if (mds_conn && export->exp_connection) {
796 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
797 /* mds reconnected after failover */
798 CWARN("%s: received MDS connection from NID %s,"
799 " removing former export from NID %s\n",
800 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
801 libcfs_nid2str(export->exp_connection->c_peer.nid));
803 /* new mds connection from the same nid */
804 CWARN("%s: received new MDS connection from NID %s,"
805 " removing former export from same NID\n",
806 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
807 class_fail_export(export);
808 class_export_put(export);
811 } else if (export->exp_connection != NULL &&
812 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
813 (lustre_msg_get_op_flags(req->rq_reqmsg) &
814 MSG_CONNECT_INITIAL)) {
815 /* in mds failover we have static uuid but nid can be
817 CWARN("%s: cookie %s seen on new NID %s when "
818 "existing NID %s is already connected\n",
819 target->obd_name, cluuid.uuid,
820 libcfs_nid2str(req->rq_peer.nid),
821 libcfs_nid2str(export->exp_connection->c_peer.nid));
823 class_export_put(export);
826 cfs_spin_lock(&export->exp_lock);
827 export->exp_connecting = 1;
828 cfs_spin_unlock(&export->exp_lock);
829 class_export_put(export);
830 LASSERT(export->exp_obd == target);
832 rc = target_handle_reconnect(&conn, export, &cluuid);
835 /* If we found an export, we already unlocked. */
838 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
839 } else if (req->rq_export == NULL &&
840 cfs_atomic_read(&export->exp_rpc_count) > 0) {
841 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
842 target->obd_name, cluuid.uuid,
843 libcfs_nid2str(req->rq_peer.nid),
844 export, cfs_atomic_read(&export->exp_refcount));
845 GOTO(out, rc = -EBUSY);
846 } else if (req->rq_export != NULL &&
847 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
848 /* the current connect rpc has increased exp_rpc_count */
849 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
850 target->obd_name, cluuid.uuid,
851 libcfs_nid2str(req->rq_peer.nid),
852 export, cfs_atomic_read(&export->exp_rpc_count) - 1);
853 cfs_spin_lock(&export->exp_lock);
854 if (req->rq_export->exp_conn_cnt <
855 lustre_msg_get_conn_cnt(req->rq_reqmsg))
856 /* try to abort active requests */
857 req->rq_export->exp_abort_active_req = 1;
858 cfs_spin_unlock(&export->exp_lock);
859 GOTO(out, rc = -EBUSY);
860 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
861 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
862 "cookies not random?\n", target->obd_name,
863 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
864 GOTO(out, rc = -EALREADY);
866 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
873 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
874 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
875 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
876 export, (long)cfs_time_current_sec(),
877 export ? (long)export->exp_last_request_time : 0);
879 /* If this is the first time a client connects,
880 * reset the recovery timer */
881 if (rc == 0 && target->obd_recovering)
882 target_start_and_reset_recovery_timer(target, req, !export);
884 /* We want to handle EALREADY but *not* -EALREADY from
885 * target_handle_reconnect(), return reconnection state in a flag */
886 if (rc == EALREADY) {
887 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
893 /* Tell the client if we support replayable requests */
894 if (target->obd_replayable)
895 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
896 client_nid = &req->rq_peer.nid;
898 if (export == NULL) {
899 if (target->obd_recovering) {
902 t = cfs_timer_deadline(&target->obd_recovery_timer);
903 t = cfs_time_sub(t, cfs_time_current());
904 CERROR("%s: denying connection for new client %s (%s): "
905 "%d clients in recovery for "CFS_TIME_T"s\n",
907 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
908 cfs_atomic_read(&target-> \
909 obd_lock_replay_clients),
910 cfs_duration_sec(t));
914 rc = obd_connect(req->rq_svc_thread->t_env,
915 &export, target, &cluuid, data,
918 conn.cookie = export->exp_handle.h_cookie;
921 rc = obd_reconnect(req->rq_svc_thread->t_env,
922 export, target, &cluuid, data, client_nid);
924 /* prevous done via class_conn2export */
925 class_export_get(export);
929 /* Return only the parts of obd_connect_data that we understand, so the
930 * client knows that we don't understand the rest. */
932 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
934 tmpdata = req_capsule_server_get(&req->rq_pill,
936 /* Don't use struct assignment here, because the client reply
937 * buffer may be smaller/larger than the local struct
938 * obd_connect_data. */
939 memcpy(tmpdata, data, min(tmpsize, size));
942 /* If all else goes well, this is our RPC return code. */
945 lustre_msg_set_handle(req->rq_repmsg, &conn);
947 /* If the client and the server are the same node, we will already
948 * have an export that really points to the client's DLM export,
949 * because we have a shared handles table.
951 * XXX this will go away when shaver stops sending the "connect" handle
952 * in the real "remote handle" field of the request --phik 24 Apr 2003
954 if (req->rq_export != NULL)
955 class_export_put(req->rq_export);
957 req->rq_export = export;
959 cfs_spin_lock(&export->exp_lock);
960 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
961 cfs_spin_unlock(&export->exp_lock);
962 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
963 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
964 export->exp_conn_cnt,
965 lustre_msg_get_conn_cnt(req->rq_reqmsg));
967 GOTO(out, rc = -EALREADY);
969 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
970 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
971 export->exp_abort_active_req = 0;
973 /* request from liblustre? Don't evict it for not pinging. */
974 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
975 export->exp_libclient = 1;
976 cfs_spin_unlock(&export->exp_lock);
978 cfs_spin_lock(&target->obd_dev_lock);
979 cfs_list_del_init(&export->exp_obd_chain_timed);
980 cfs_spin_unlock(&target->obd_dev_lock);
982 cfs_spin_unlock(&export->exp_lock);
985 if (export->exp_connection != NULL) {
986 /* Check to see if connection came from another NID */
987 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
988 !cfs_hlist_unhashed(&export->exp_nid_hash))
989 cfs_hash_del(export->exp_obd->obd_nid_hash,
990 &export->exp_connection->c_peer.nid,
991 &export->exp_nid_hash);
993 ptlrpc_connection_put(export->exp_connection);
996 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
999 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1000 cfs_hash_add(export->exp_obd->obd_nid_hash,
1001 &export->exp_connection->c_peer.nid,
1002 &export->exp_nid_hash);
1005 cfs_spin_lock(&target->obd_recovery_task_lock);
1006 if (target->obd_recovering && !export->exp_in_recovery) {
1007 cfs_spin_lock(&export->exp_lock);
1008 export->exp_in_recovery = 1;
1009 export->exp_req_replay_needed = 1;
1010 export->exp_lock_replay_needed = 1;
1011 cfs_spin_unlock(&export->exp_lock);
1012 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1013 && (data->ocd_transno == 0))
1014 CWARN("Connect with zero transno!\n");
1016 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1017 && data->ocd_transno < target->obd_next_recovery_transno)
1018 target->obd_next_recovery_transno = data->ocd_transno;
1019 target->obd_connected_clients++;
1020 cfs_atomic_inc(&target->obd_req_replay_clients);
1021 cfs_atomic_inc(&target->obd_lock_replay_clients);
1022 if (target->obd_connected_clients ==
1023 target->obd_max_recoverable_clients)
1024 cfs_waitq_signal(&target->obd_next_transno_waitq);
1026 cfs_spin_unlock(&target->obd_recovery_task_lock);
1028 /* Tell the client we're in recovery, when client is involved in it. */
1029 if (target->obd_recovering)
1030 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1032 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1035 if (export->exp_imp_reverse != NULL) {
1036 /* destroyed import can be still referenced in ctxt */
1037 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1038 KEY_REVIMP_UPD, 0, NULL, NULL);
1040 client_destroy_import(export->exp_imp_reverse);
1043 /* for the rest part, we return -ENOTCONN in case of errors
1044 * in order to let client initialize connection again.
1046 revimp = export->exp_imp_reverse = class_new_import(target);
1048 CERROR("fail to alloc new reverse import.\n");
1049 GOTO(out, rc = -ENOTCONN);
1052 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1053 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1054 revimp->imp_remote_handle = conn;
1055 revimp->imp_dlm_fake = 1;
1056 revimp->imp_state = LUSTRE_IMP_FULL;
1058 /* unknown versions will be caught in
1059 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1060 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1062 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1063 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1064 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1066 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1068 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1069 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1070 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1072 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1074 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1076 CERROR("Failed to get sec for reverse import: %d\n", rc);
1077 export->exp_imp_reverse = NULL;
1078 class_destroy_import(revimp);
1081 class_import_put(revimp);
1084 cfs_spin_lock(&export->exp_lock);
1085 export->exp_connecting = 0;
1086 cfs_spin_unlock(&export->exp_lock);
1089 class_decref(targref, __FUNCTION__, cfs_current());
1091 req->rq_status = rc;
1095 int target_handle_disconnect(struct ptlrpc_request *req)
1100 rc = req_capsule_server_pack(&req->rq_pill);
1104 /* keep the rq_export around so we can send the reply */
1105 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1110 void target_destroy_export(struct obd_export *exp)
1112 /* exports created from last_rcvd data, and "fake"
1113 exports created by lctl don't have an import */
1114 if (exp->exp_imp_reverse != NULL)
1115 client_destroy_import(exp->exp_imp_reverse);
1117 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1118 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1119 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1120 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1124 * Recovery functions
1126 static void target_request_copy_get(struct ptlrpc_request *req)
1128 class_export_rpc_get(req->rq_export);
1129 LASSERT(cfs_list_empty(&req->rq_list));
1130 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1132 /* increase refcount to keep request in queue */
1133 cfs_atomic_inc(&req->rq_refcount);
1134 /** let export know it has replays to be handled */
1135 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1138 static void target_request_copy_put(struct ptlrpc_request *req)
1140 LASSERT(cfs_list_empty(&req->rq_replay_list));
1141 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1143 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1144 class_export_rpc_put(req->rq_export);
1145 ptlrpc_server_drop_request(req);
1148 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1150 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1151 struct obd_export *exp = req->rq_export;
1152 struct ptlrpc_request *reqiter;
1157 cfs_spin_lock(&exp->exp_lock);
1158 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1160 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1167 /* we expect it with RESENT and REPLAY flags */
1168 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1169 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1170 CERROR("invalid flags %x of resent replay\n",
1171 lustre_msg_get_flags(req->rq_reqmsg));
1173 cfs_list_add_tail(&req->rq_replay_list,
1174 &exp->exp_req_replay_queue);
1177 cfs_spin_unlock(&exp->exp_lock);
1181 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1183 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1184 LASSERT(req->rq_export);
1186 cfs_spin_lock(&req->rq_export->exp_lock);
1187 cfs_list_del_init(&req->rq_replay_list);
1188 cfs_spin_unlock(&req->rq_export->exp_lock);
1192 static void target_finish_recovery(struct obd_device *obd)
1195 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1198 ldlm_reprocess_all_ns(obd->obd_namespace);
1199 cfs_spin_lock(&obd->obd_recovery_task_lock);
1200 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1201 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1202 !cfs_list_empty(&obd->obd_final_req_queue)) {
1203 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1205 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1206 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1208 cfs_list_empty(&obd->obd_final_req_queue) ? \
1210 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1213 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1215 obd->obd_recovery_end = cfs_time_current_sec();
1217 /* when recovery finished, cleanup orphans on mds and ost */
1218 if (OBT(obd) && OBP(obd, postrecov)) {
1219 int rc = OBP(obd, postrecov)(obd);
1221 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1227 static void abort_req_replay_queue(struct obd_device *obd)
1229 struct ptlrpc_request *req, *n;
1230 cfs_list_t abort_list;
1232 CFS_INIT_LIST_HEAD(&abort_list);
1233 cfs_spin_lock(&obd->obd_recovery_task_lock);
1234 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1235 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1236 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1237 DEBUG_REQ(D_WARNING, req, "aborted:");
1238 req->rq_status = -ENOTCONN;
1239 if (ptlrpc_error(req)) {
1240 DEBUG_REQ(D_ERROR, req,
1241 "failed abort_req_reply; skipping");
1243 target_exp_dequeue_req_replay(req);
1244 target_request_copy_put(req);
1248 static void abort_lock_replay_queue(struct obd_device *obd)
1250 struct ptlrpc_request *req, *n;
1251 cfs_list_t abort_list;
1253 CFS_INIT_LIST_HEAD(&abort_list);
1254 cfs_spin_lock(&obd->obd_recovery_task_lock);
1255 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1256 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1257 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1258 DEBUG_REQ(D_ERROR, req, "aborted:");
1259 req->rq_status = -ENOTCONN;
1260 if (ptlrpc_error(req)) {
1261 DEBUG_REQ(D_ERROR, req,
1262 "failed abort_lock_reply; skipping");
1264 target_request_copy_put(req);
1269 /* Called from a cleanup function if the device is being cleaned up
1270 forcefully. The exports should all have been disconnected already,
1271 the only thing left to do is
1272 - clear the recovery flags
1274 - free queued requests and replies, but don't send replies
1275 Because the obd_stopping flag is set, no new requests should be received.
1278 void target_cleanup_recovery(struct obd_device *obd)
1280 struct ptlrpc_request *req, *n;
1281 cfs_list_t clean_list;
1284 CFS_INIT_LIST_HEAD(&clean_list);
1285 cfs_spin_lock(&obd->obd_dev_lock);
1286 if (!obd->obd_recovering) {
1287 cfs_spin_unlock(&obd->obd_dev_lock);
1291 obd->obd_recovering = obd->obd_abort_recovery = 0;
1292 cfs_spin_unlock(&obd->obd_dev_lock);
1294 cfs_spin_lock(&obd->obd_recovery_task_lock);
1295 target_cancel_recovery_timer(obd);
1296 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1297 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1299 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1300 LASSERT(req->rq_reply_state == 0);
1301 target_exp_dequeue_req_replay(req);
1302 target_request_copy_put(req);
1305 cfs_spin_lock(&obd->obd_recovery_task_lock);
1306 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1307 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1308 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1310 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1311 LASSERT(req->rq_reply_state == 0);
1312 target_request_copy_put(req);
1318 /* obd_recovery_task_lock should be held */
1319 void target_cancel_recovery_timer(struct obd_device *obd)
1321 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1322 cfs_timer_disarm(&obd->obd_recovery_timer);
1325 /* extend = 1 means require at least "duration" seconds left in the timer,
1326 extend = 0 means set the total duration (start_recovery_timer) */
1327 static void reset_recovery_timer(struct obd_device *obd, int duration,
1330 cfs_time_t now = cfs_time_current_sec();
1331 cfs_duration_t left;
1333 cfs_spin_lock(&obd->obd_recovery_task_lock);
1334 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1335 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1339 left = cfs_time_sub(obd->obd_recovery_end, now);
1341 if (extend && (duration > left))
1342 obd->obd_recovery_timeout += duration - left;
1343 else if (!extend && (duration > obd->obd_recovery_timeout))
1344 /* Track the client's largest expected replay time */
1345 obd->obd_recovery_timeout = duration;
1347 /* Hard limit of obd_recovery_time_hard which should not happen */
1348 if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1349 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1351 obd->obd_recovery_end = obd->obd_recovery_start +
1352 obd->obd_recovery_timeout;
1353 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1354 cfs_time_before(now, obd->obd_recovery_end)) {
1355 left = cfs_time_sub(obd->obd_recovery_end, now);
1356 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1358 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1359 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1360 obd->obd_name, (unsigned)left);
1363 static void check_and_start_recovery_timer(struct obd_device *obd)
1365 cfs_spin_lock(&obd->obd_recovery_task_lock);
1366 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1367 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1370 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1371 obd->obd_recovery_start = cfs_time_current_sec();
1372 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1374 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1377 /* Reset the timer with each new client connection */
1379 * This timer is actually reconnect_timer, which is for making sure
1380 * the total recovery window is at least as big as my reconnect
1381 * attempt timing. So the initial recovery time_out will be set to
1382 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1383 * from client is bigger than this, then the recovery time_out will
1384 * be extended to make sure the client could be reconnected, in the
1385 * process, the timeout from the new client should be ignored.
1389 target_start_and_reset_recovery_timer(struct obd_device *obd,
1390 struct ptlrpc_request *req,
1393 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1395 if (!new_client && service_time)
1396 /* Teach server about old server's estimates, as first guess
1397 * at how long new requests will take. */
1398 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1401 check_and_start_recovery_timer(obd);
1403 /* convert the service time to rpc timeout,
1404 * reuse service_time to limit stack usage */
1405 service_time = at_est2timeout(service_time);
1407 /* We expect other clients to timeout within service_time, then try
1408 * to reconnect, then try the failover server. The max delay between
1409 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1410 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1411 INITIAL_CONNECT_TIMEOUT);
1412 if (service_time > obd->obd_recovery_timeout && !new_client)
1413 reset_recovery_timer(obd, service_time, 0);
1418 /** Health checking routines */
1419 static inline int exp_connect_healthy(struct obd_export *exp)
1421 return (exp->exp_in_recovery);
1424 /** if export done req_replay or has replay in queue */
1425 static inline int exp_req_replay_healthy(struct obd_export *exp)
1427 return (!exp->exp_req_replay_needed ||
1428 cfs_atomic_read(&exp->exp_replay_count) > 0);
1430 /** if export done lock_replay or has replay in queue */
1431 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1433 return (!exp->exp_lock_replay_needed ||
1434 cfs_atomic_read(&exp->exp_replay_count) > 0);
1437 static inline int exp_vbr_healthy(struct obd_export *exp)
1439 return (!exp->exp_vbr_failed);
1442 static inline int exp_finished(struct obd_export *exp)
1444 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1447 /** Checking routines for recovery */
1448 static int check_for_clients(struct obd_device *obd)
1450 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1452 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1453 if (obd->obd_no_conn == 0 &&
1454 obd->obd_connected_clients + obd->obd_stale_clients ==
1455 obd->obd_max_recoverable_clients)
1460 static int check_for_next_transno(struct obd_device *obd)
1462 struct ptlrpc_request *req = NULL;
1463 int wake_up = 0, connected, completed, queue_len;
1464 __u64 next_transno, req_transno;
1467 cfs_spin_lock(&obd->obd_recovery_task_lock);
1468 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1469 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1470 struct ptlrpc_request, rq_list);
1471 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1476 connected = obd->obd_connected_clients;
1477 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1478 queue_len = obd->obd_requests_queued_for_recovery;
1479 next_transno = obd->obd_next_recovery_transno;
1481 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1482 "req_transno: "LPU64", next_transno: "LPU64"\n",
1483 obd->obd_max_recoverable_clients, connected, completed,
1484 queue_len, req_transno, next_transno);
1486 if (obd->obd_abort_recovery) {
1487 CDEBUG(D_HA, "waking for aborted recovery\n");
1489 } else if (obd->obd_recovery_expired) {
1490 CDEBUG(D_HA, "waking for expired recovery\n");
1492 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1493 CDEBUG(D_HA, "waking for completed recovery\n");
1495 } else if (req_transno == next_transno) {
1496 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1498 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1500 /** handle gaps occured due to lost reply or VBR */
1501 LASSERTF(req_transno >= next_transno,
1502 "req_transno: "LPU64", next_transno: "LPU64"\n",
1503 req_transno, next_transno);
1504 if (req_transno > obd->obd_last_committed &&
1505 !obd->obd_version_recov)
1508 "%s: waking for gap in transno, VBR is %s (skip: "
1509 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1510 ", last_committed: "LPD64")\n",
1511 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1512 next_transno, queue_len, completed, connected,
1513 req_transno, obd->obd_last_committed);
1514 obd->obd_next_recovery_transno = req_transno;
1516 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1517 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1518 " by fail_lock, waking up ("LPD64")\n", next_transno);
1519 obd->obd_next_recovery_transno = req_transno;
1522 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1526 static int check_for_next_lock(struct obd_device *obd)
1530 cfs_spin_lock(&obd->obd_recovery_task_lock);
1531 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1532 CDEBUG(D_HA, "waking for next lock\n");
1534 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1535 CDEBUG(D_HA, "waking for completed lock replay\n");
1537 } else if (obd->obd_abort_recovery) {
1538 CDEBUG(D_HA, "waking for aborted recovery\n");
1540 } else if (obd->obd_recovery_expired) {
1541 CDEBUG(D_HA, "waking for expired recovery\n");
1544 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1550 * wait for recovery events,
1551 * check its status with help of check_routine
1552 * evict dead clients via health_check
1554 static int target_recovery_overseer(struct obd_device *obd,
1555 int (*check_routine)(struct obd_device *),
1556 int (*health_check)(struct obd_export *))
1559 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1560 if (obd->obd_abort_recovery) {
1561 CWARN("recovery is aborted, evict exports in recovery\n");
1562 /** evict exports which didn't finish recovery yet */
1563 class_disconnect_stale_exports(obd, exp_finished);
1565 } else if (obd->obd_recovery_expired) {
1566 obd->obd_recovery_expired = 0;
1567 /** If some clients died being recovered, evict them */
1569 "recovery is timed out, evict stale exports\n");
1570 /** evict cexports with no replay in queue, they are stalled */
1571 class_disconnect_stale_exports(obd, health_check);
1572 /** continue with VBR */
1573 cfs_spin_lock(&obd->obd_dev_lock);
1574 obd->obd_version_recov = 1;
1575 cfs_spin_unlock(&obd->obd_dev_lock);
1577 * reset timer, recovery will proceed with versions now,
1578 * timeout is set just to handle reconnection delays
1580 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1581 /** Wait for recovery events again, after evicting bad clients */
1587 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1589 struct ptlrpc_request *req = NULL;
1592 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1593 obd->obd_next_recovery_transno);
1595 if (target_recovery_overseer(obd, check_for_next_transno,
1596 exp_req_replay_healthy)) {
1597 abort_req_replay_queue(obd);
1598 abort_lock_replay_queue(obd);
1601 cfs_spin_lock(&obd->obd_recovery_task_lock);
1602 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1603 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1604 struct ptlrpc_request, rq_list);
1605 cfs_list_del_init(&req->rq_list);
1606 obd->obd_requests_queued_for_recovery--;
1607 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1609 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1610 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1611 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1612 /** evict exports failed VBR */
1613 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1618 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1620 struct ptlrpc_request *req = NULL;
1622 CDEBUG(D_HA, "Waiting for lock\n");
1623 if (target_recovery_overseer(obd, check_for_next_lock,
1624 exp_lock_replay_healthy))
1625 abort_lock_replay_queue(obd);
1627 cfs_spin_lock(&obd->obd_recovery_task_lock);
1628 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1629 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1630 struct ptlrpc_request, rq_list);
1631 cfs_list_del_init(&req->rq_list);
1632 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1634 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1635 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1636 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1637 /** evict exports failed VBR */
1638 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1643 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1645 struct ptlrpc_request *req = NULL;
1647 cfs_spin_lock(&obd->obd_recovery_task_lock);
1648 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1649 req = cfs_list_entry(obd->obd_final_req_queue.next,
1650 struct ptlrpc_request, rq_list);
1651 cfs_list_del_init(&req->rq_list);
1652 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1653 if (req->rq_export->exp_in_recovery) {
1654 cfs_spin_lock(&req->rq_export->exp_lock);
1655 req->rq_export->exp_in_recovery = 0;
1656 cfs_spin_unlock(&req->rq_export->exp_lock);
1659 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1664 static int handle_recovery_req(struct ptlrpc_thread *thread,
1665 struct ptlrpc_request *req,
1666 svc_handler_t handler)
1672 * export can be evicted during recovery, no need to handle replays for
1673 * it after that, discard such request silently
1675 if (req->rq_export->exp_disconnected)
1676 GOTO(reqcopy_put, rc = 0);
1678 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1680 CERROR("Failure to initialize session: %d\n", rc);
1681 GOTO(reqcopy_put, rc);
1684 req->rq_recov_session.lc_thread = thread;
1685 lu_context_enter(&req->rq_recov_session);
1686 req->rq_svc_thread = thread;
1687 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1689 /* thread context */
1690 lu_context_enter(&thread->t_env->le_ctx);
1692 lu_context_exit(&thread->t_env->le_ctx);
1694 lu_context_exit(&req->rq_recov_session);
1695 lu_context_fini(&req->rq_recov_session);
1696 /* don't reset timer for final stage */
1697 if (!exp_finished(req->rq_export)) {
1699 * Add request timeout to the recovery time so next request from
1700 * this client may come in recovery time
1702 reset_recovery_timer(class_exp2obd(req->rq_export),
1703 AT_OFF ? obd_timeout :
1704 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1710 static int target_recovery_thread(void *arg)
1712 struct lu_target *lut = arg;
1713 struct obd_device *obd = lut->lut_obd;
1714 struct ptlrpc_request *req;
1715 struct target_recovery_data *trd = &obd->obd_recovery_data;
1716 unsigned long delta;
1717 unsigned long flags;
1719 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1723 cfs_daemonize_ctxt("tgt_recov");
1725 SIGNAL_MASK_LOCK(current, flags);
1726 sigfillset(¤t->blocked);
1728 SIGNAL_MASK_UNLOCK(current, flags);
1730 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1734 thread->t_env = &env;
1735 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1736 env.le_ctx.lc_thread = thread;
1737 thread->t_data = NULL;
1739 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1741 trd->trd_processing_task = cfs_curproc_pid();
1743 cfs_spin_lock(&obd->obd_dev_lock);
1744 obd->obd_recovering = 1;
1745 cfs_spin_unlock(&obd->obd_dev_lock);
1746 cfs_complete(&trd->trd_starting);
1748 /* first of all, we have to know the first transno to replay */
1749 if (target_recovery_overseer(obd, check_for_clients,
1750 exp_connect_healthy)) {
1751 abort_req_replay_queue(obd);
1752 abort_lock_replay_queue(obd);
1755 /* next stage: replay requests */
1757 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1758 cfs_atomic_read(&obd->obd_req_replay_clients),
1759 obd->obd_next_recovery_transno);
1760 while ((req = target_next_replay_req(obd))) {
1761 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1762 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1763 lustre_msg_get_transno(req->rq_reqmsg),
1764 libcfs_nid2str(req->rq_peer.nid));
1765 handle_recovery_req(thread, req,
1766 trd->trd_recovery_handler);
1768 * bz18031: increase next_recovery_transno before
1769 * target_request_copy_put() will drop exp_rpc reference
1771 cfs_spin_lock(&obd->obd_recovery_task_lock);
1772 obd->obd_next_recovery_transno++;
1773 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1774 target_exp_dequeue_req_replay(req);
1775 target_request_copy_put(req);
1776 obd->obd_replayed_requests++;
1780 * The second stage: replay locks
1782 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1783 cfs_atomic_read(&obd->obd_lock_replay_clients));
1784 while ((req = target_next_replay_lock(obd))) {
1785 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1786 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1787 libcfs_nid2str(req->rq_peer.nid));
1788 handle_recovery_req(thread, req,
1789 trd->trd_recovery_handler);
1790 target_request_copy_put(req);
1791 obd->obd_replayed_locks++;
1795 * The third stage: reply on final pings, at this moment all clients
1796 * must have request in final queue
1798 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1799 /** Update server last boot epoch */
1800 lut_boot_epoch_update(lut);
1801 /* We drop recoverying flag to forward all new requests
1802 * to regular mds_handle() since now */
1803 cfs_spin_lock(&obd->obd_dev_lock);
1804 obd->obd_recovering = obd->obd_abort_recovery = 0;
1805 cfs_spin_unlock(&obd->obd_dev_lock);
1806 cfs_spin_lock(&obd->obd_recovery_task_lock);
1807 target_cancel_recovery_timer(obd);
1808 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1809 while ((req = target_next_final_ping(obd))) {
1810 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1811 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1812 libcfs_nid2str(req->rq_peer.nid));
1813 handle_recovery_req(thread, req,
1814 trd->trd_recovery_handler);
1815 target_request_copy_put(req);
1818 delta = (jiffies - delta) / CFS_HZ;
1819 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1820 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1821 if (delta > OBD_RECOVERY_TIME_SOFT) {
1822 CWARN("too long recovery - read logs\n");
1823 libcfs_debug_dumplog();
1826 target_finish_recovery(obd);
1828 lu_context_fini(&env.le_ctx);
1829 trd->trd_processing_task = 0;
1830 cfs_complete(&trd->trd_finishing);
1834 static int target_start_recovery_thread(struct lu_target *lut,
1835 svc_handler_t handler)
1837 struct obd_device *obd = lut->lut_obd;
1839 struct target_recovery_data *trd = &obd->obd_recovery_data;
1841 memset(trd, 0, sizeof(*trd));
1842 cfs_init_completion(&trd->trd_starting);
1843 cfs_init_completion(&trd->trd_finishing);
1844 trd->trd_recovery_handler = handler;
1846 if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
1847 cfs_wait_for_completion(&trd->trd_starting);
1848 LASSERT(obd->obd_recovering != 0);
1855 void target_stop_recovery_thread(struct obd_device *obd)
1857 if (obd->obd_recovery_data.trd_processing_task > 0) {
1858 struct target_recovery_data *trd = &obd->obd_recovery_data;
1859 /** recovery can be done but postrecovery is not yet */
1860 cfs_spin_lock(&obd->obd_dev_lock);
1861 if (obd->obd_recovering) {
1862 CERROR("%s: Aborting recovery\n", obd->obd_name);
1863 obd->obd_abort_recovery = 1;
1864 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1866 cfs_spin_unlock(&obd->obd_dev_lock);
1867 cfs_wait_for_completion(&trd->trd_finishing);
1871 void target_recovery_fini(struct obd_device *obd)
1873 class_disconnect_exports(obd);
1874 target_stop_recovery_thread(obd);
1875 target_cleanup_recovery(obd);
1877 EXPORT_SYMBOL(target_recovery_fini);
1879 static void target_recovery_expired(unsigned long castmeharder)
1881 struct obd_device *obd = (struct obd_device *)castmeharder;
1882 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1883 " after %lds (%d clients connected)\n",
1884 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1885 cfs_time_current_sec()- obd->obd_recovery_start,
1886 obd->obd_connected_clients);
1888 obd->obd_recovery_expired = 1;
1889 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1892 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1894 struct obd_device *obd = lut->lut_obd;
1895 if (obd->obd_max_recoverable_clients == 0) {
1896 /** Update server last boot epoch */
1897 lut_boot_epoch_update(lut);
1901 CWARN("RECOVERY: service %s, %d recoverable clients, "
1902 "last_transno "LPU64"\n", obd->obd_name,
1903 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1904 LASSERT(obd->obd_stopping == 0);
1905 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1906 obd->obd_recovery_start = 0;
1907 obd->obd_recovery_end = 0;
1909 /* both values can be get from mount data already */
1910 if (obd->obd_recovery_timeout == 0)
1911 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
1912 if (obd->obd_recovery_time_hard == 0)
1913 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
1914 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1915 target_start_recovery_thread(lut, handler);
1917 EXPORT_SYMBOL(target_recovery_init);
1921 static int target_process_req_flags(struct obd_device *obd,
1922 struct ptlrpc_request *req)
1924 struct obd_export *exp = req->rq_export;
1925 LASSERT(exp != NULL);
1926 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1927 /* client declares he's ready to replay locks */
1928 cfs_spin_lock(&exp->exp_lock);
1929 if (exp->exp_req_replay_needed) {
1930 exp->exp_req_replay_needed = 0;
1931 cfs_spin_unlock(&exp->exp_lock);
1933 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
1934 cfs_atomic_dec(&obd->obd_req_replay_clients);
1936 cfs_spin_unlock(&exp->exp_lock);
1939 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1940 /* client declares he's ready to complete recovery
1941 * so, we put the request on th final queue */
1942 cfs_spin_lock(&exp->exp_lock);
1943 if (exp->exp_lock_replay_needed) {
1944 exp->exp_lock_replay_needed = 0;
1945 cfs_spin_unlock(&exp->exp_lock);
1947 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
1948 cfs_atomic_dec(&obd->obd_lock_replay_clients);
1950 cfs_spin_unlock(&exp->exp_lock);
1956 int target_queue_recovery_request(struct ptlrpc_request *req,
1957 struct obd_device *obd)
1961 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1964 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1965 /* Processing the queue right now, don't re-add. */
1969 target_process_req_flags(obd, req);
1971 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1972 /* client declares he's ready to complete recovery
1973 * so, we put the request on th final queue */
1974 target_request_copy_get(req);
1975 DEBUG_REQ(D_HA, req, "queue final req");
1976 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1977 cfs_spin_lock(&obd->obd_recovery_task_lock);
1978 if (obd->obd_recovering) {
1979 cfs_list_add_tail(&req->rq_list,
1980 &obd->obd_final_req_queue);
1982 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1983 target_request_copy_put(req);
1984 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1986 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1989 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1990 /* client declares he's ready to replay locks */
1991 target_request_copy_get(req);
1992 DEBUG_REQ(D_HA, req, "queue lock replay req");
1993 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1994 cfs_spin_lock(&obd->obd_recovery_task_lock);
1995 LASSERT(obd->obd_recovering);
1996 /* usually due to recovery abort */
1997 if (!req->rq_export->exp_in_recovery) {
1998 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1999 target_request_copy_put(req);
2002 LASSERT(req->rq_export->exp_lock_replay_needed);
2003 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2004 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2008 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2009 * (i.e. buflens etc are in my own byte order), but type-dependent
2010 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2013 CFS_INIT_LIST_HEAD(&req->rq_list);
2014 DEBUG_REQ(D_HA, req, "not queueing");
2018 /* If we're processing the queue, we want don't want to queue this
2021 * Also, if this request has a transno less than the one we're waiting
2022 * for, we should process it now. It could (and currently always will)
2023 * be an open request for a descriptor that was opened some time ago.
2025 * Also, a resent, replayed request that has already been
2026 * handled will pass through here and be processed immediately.
2028 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2029 obd->obd_next_recovery_transno, transno);
2030 cfs_spin_lock(&obd->obd_recovery_task_lock);
2031 if (transno < obd->obd_next_recovery_transno) {
2032 /* Processing the queue right now, don't re-add. */
2033 LASSERT(cfs_list_empty(&req->rq_list));
2034 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2037 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2039 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2042 target_request_copy_get(req);
2043 if (!req->rq_export->exp_in_recovery) {
2044 target_request_copy_put(req);
2047 LASSERT(req->rq_export->exp_req_replay_needed);
2049 if (target_exp_enqueue_req_replay(req)) {
2050 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2051 target_request_copy_put(req);
2056 cfs_spin_lock(&obd->obd_recovery_task_lock);
2057 LASSERT(obd->obd_recovering);
2058 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2059 struct ptlrpc_request *reqiter =
2060 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2062 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2063 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2068 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2070 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2071 "has been claimed by another client");
2072 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2073 target_exp_dequeue_req_replay(req);
2074 target_request_copy_put(req);
2080 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2082 obd->obd_requests_queued_for_recovery++;
2083 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2084 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2089 * Packs current SLV and Limit into \a req.
2091 int target_pack_pool_reply(struct ptlrpc_request *req)
2093 struct obd_device *obd;
2097 * Check that we still have all structures alive as this may
2098 * be some late rpc in shutdown time.
2100 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2101 !exp_connect_lru_resize(req->rq_export))) {
2102 lustre_msg_set_slv(req->rq_repmsg, 0);
2103 lustre_msg_set_limit(req->rq_repmsg, 0);
2108 * OBD is alive here as export is alive, which we checked above.
2110 obd = req->rq_export->exp_obd;
2112 cfs_read_lock(&obd->obd_pool_lock);
2113 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2114 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2115 cfs_read_unlock(&obd->obd_pool_lock);
2120 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2122 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2123 DEBUG_REQ(D_ERROR, req, "dropping reply");
2128 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2129 req->rq_status = rc;
2130 return (ptlrpc_send_error(req, 1));
2132 DEBUG_REQ(D_NET, req, "sending reply");
2135 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2138 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2141 struct ptlrpc_reply_state *rs;
2142 struct obd_device *obd;
2143 struct obd_export *exp;
2144 struct ptlrpc_service *svc;
2147 if (req->rq_no_reply) {
2152 svc = req->rq_rqbd->rqbd_service;
2153 rs = req->rq_reply_state;
2154 if (rs == NULL || !rs->rs_difficult) {
2156 target_send_reply_msg (req, rc, fail_id);
2161 /* must be an export if locks saved */
2162 LASSERT (req->rq_export != NULL);
2163 /* req/reply consistent */
2164 LASSERT (rs->rs_service == svc);
2167 LASSERT (!rs->rs_scheduled);
2168 LASSERT (!rs->rs_scheduled_ever);
2169 LASSERT (!rs->rs_handled);
2170 LASSERT (!rs->rs_on_net);
2171 LASSERT (rs->rs_export == NULL);
2172 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2173 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2175 exp = class_export_get (req->rq_export);
2178 /* disable reply scheduling while I'm setting up */
2179 rs->rs_scheduled = 1;
2181 rs->rs_xid = req->rq_xid;
2182 rs->rs_transno = req->rq_transno;
2183 rs->rs_export = exp;
2184 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2186 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2187 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2188 rs->rs_transno, exp->exp_last_committed);
2189 if (rs->rs_transno > exp->exp_last_committed) {
2190 /* not committed already */
2191 cfs_list_add_tail(&rs->rs_obd_list,
2192 &exp->exp_uncommitted_replies);
2194 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2196 cfs_spin_lock(&exp->exp_lock);
2197 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2198 cfs_spin_unlock(&exp->exp_lock);
2200 netrc = target_send_reply_msg (req, rc, fail_id);
2202 cfs_spin_lock(&svc->srv_rs_lock);
2204 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2207 /* error sending: reply is off the net. Also we need +1
2208 * reply ref until ptlrpc_handle_rs() is done
2209 * with the reply state (if the send was successful, there
2210 * would have been +1 ref for the net, which
2211 * reply_out_callback leaves alone) */
2213 ptlrpc_rs_addref(rs);
2216 cfs_spin_lock(&rs->rs_lock);
2217 if (rs->rs_transno <= exp->exp_last_committed ||
2218 (!rs->rs_on_net && !rs->rs_no_ack) ||
2219 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2220 cfs_list_empty(&rs->rs_obd_list)) {
2221 CDEBUG(D_HA, "Schedule reply immediately\n");
2222 ptlrpc_dispatch_difficult_reply(rs);
2224 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2225 rs->rs_scheduled = 0; /* allow notifier to schedule */
2227 cfs_spin_unlock(&rs->rs_lock);
2228 cfs_spin_unlock(&svc->srv_rs_lock);
2232 int target_handle_ping(struct ptlrpc_request *req)
2234 obd_ping(req->rq_export);
2235 return req_capsule_server_pack(&req->rq_pill);
2238 void target_committed_to_req(struct ptlrpc_request *req)
2240 struct obd_export *exp = req->rq_export;
2242 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2243 lustre_msg_set_last_committed(req->rq_repmsg,
2244 exp->exp_last_committed);
2246 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2247 "%d)", exp->exp_obd->obd_no_transno,
2248 req->rq_repmsg == NULL);
2250 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2251 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2253 EXPORT_SYMBOL(target_committed_to_req);
2255 int target_handle_qc_callback(struct ptlrpc_request *req)
2257 struct obd_quotactl *oqctl;
2258 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2260 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2261 if (oqctl == NULL) {
2262 CERROR("Can't unpack obd_quotactl\n");
2266 cli->cl_qchk_stat = oqctl->qc_stat;
2271 #ifdef HAVE_QUOTA_SUPPORT
2272 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2275 struct obd_device *obd = req->rq_export->exp_obd;
2276 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2277 struct obd_device_target *obt;
2278 struct lustre_quota_ctxt *qctxt;
2279 struct qunit_data *qdata = NULL;
2283 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2286 rc = req_capsule_server_pack(&req->rq_pill);
2288 CERROR("packing reply failed!: rc = %d\n", rc);
2292 LASSERT(req->rq_export);
2294 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2295 if (IS_ERR(qdata)) {
2296 rc = PTR_ERR(qdata);
2297 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2298 req->rq_status = rc;
2302 /* we use the observer */
2303 if (obd_pin_observer(obd, &lov_obd) ||
2304 obd_pin_observer(lov_obd, &master_obd)) {
2305 CERROR("Can't find the observer, it is recovering\n");
2306 req->rq_status = -EAGAIN;
2310 obt = &master_obd->u.obt;
2311 qctxt = &obt->obt_qctxt;
2313 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2314 /* quota_type has not been processed yet, return EAGAIN
2315 * until we know whether or not quotas are supposed to
2317 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2319 req->rq_status = -EAGAIN;
2323 cfs_down_read(&obt->obt_rwsem);
2324 if (qctxt->lqc_lqs_hash == NULL) {
2325 cfs_up_read(&obt->obt_rwsem);
2326 /* quota_type has not been processed yet, return EAGAIN
2327 * until we know whether or not quotas are supposed to
2329 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2331 req->rq_status = -EAGAIN;
2335 LASSERT(qctxt->lqc_handler);
2336 rc = qctxt->lqc_handler(master_obd, qdata,
2337 lustre_msg_get_opc(req->rq_reqmsg));
2338 cfs_up_read(&obt->obt_rwsem);
2339 if (rc && rc != -EDQUOT)
2340 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2341 "dqacq/dqrel failed! (rc:%d)\n", rc);
2342 req->rq_status = rc;
2344 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2346 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2350 /* Block the quota req. b=14840 */
2351 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2356 obd_unpin_observer(lov_obd);
2358 obd_unpin_observer(obd);
2360 rc = ptlrpc_reply(req);
2364 #endif /* !__KERNEL__ */
2366 #endif /* HAVE_QUOTA_SUPPORT */
2368 ldlm_mode_t lck_compat_array[] = {
2369 [LCK_EX] LCK_COMPAT_EX,
2370 [LCK_PW] LCK_COMPAT_PW,
2371 [LCK_PR] LCK_COMPAT_PR,
2372 [LCK_CW] LCK_COMPAT_CW,
2373 [LCK_CR] LCK_COMPAT_CR,
2374 [LCK_NL] LCK_COMPAT_NL,
2375 [LCK_GROUP] LCK_COMPAT_GROUP,
2376 [LCK_COS] LCK_COMPAT_COS,
2380 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2381 * not escape to the user level.
2383 int ldlm_error2errno(ldlm_error_t error)
2391 case ELDLM_LOCK_CHANGED:
2394 case ELDLM_LOCK_ABORTED:
2397 case ELDLM_LOCK_REPLACED:
2400 case ELDLM_NO_LOCK_DATA:
2403 case ELDLM_NAMESPACE_EXISTS:
2406 case ELDLM_BAD_NAMESPACE:
2410 if (((int)error) < 0) /* cast to signed type */
2411 result = error; /* as ldlm_error_t can be unsigned */
2413 CERROR("Invalid DLM result code: %d\n", error);
2419 EXPORT_SYMBOL(ldlm_error2errno);
2422 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2424 ldlm_error_t ldlm_errno2error(int err_no)
2433 error = ELDLM_LOCK_CHANGED;
2436 error = ELDLM_LOCK_ABORTED;
2439 error = ELDLM_LOCK_REPLACED;
2442 error = ELDLM_NO_LOCK_DATA;
2445 error = ELDLM_NAMESPACE_EXISTS;
2448 error = ELDLM_BAD_NAMESPACE;
2455 EXPORT_SYMBOL(ldlm_errno2error);
2457 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2458 void ldlm_dump_export_locks(struct obd_export *exp)
2460 cfs_spin_lock(&exp->exp_locks_list_guard);
2461 if (!cfs_list_empty(&exp->exp_locks_list)) {
2462 struct ldlm_lock *lock;
2464 CERROR("dumping locks for export %p,"
2465 "ignore if the unmount doesn't hang\n", exp);
2466 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2467 ldlm_lock_dump(D_ERROR, lock, 0);
2469 cfs_spin_unlock(&exp->exp_locks_list_guard);
2473 static int target_bulk_timeout(void *data)
2476 /* We don't fail the connection here, because having the export
2477 * killed makes the (vital) call to commitrw very sad.
2482 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2484 return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2487 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2488 struct l_wait_info *lwi)
2490 struct ptlrpc_request *req = desc->bd_req;
2494 /* Check if there is eviction in progress, and if so, wait for
2496 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2497 *lwi = LWI_INTR(NULL, NULL);
2498 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2499 !cfs_atomic_read(&exp->exp_obd->
2500 obd_evict_inprogress),
2504 /* Check if client was evicted or tried to reconnect already */
2505 if (exp->exp_failed || exp->exp_abort_active_req) {
2508 if (desc->bd_type == BULK_PUT_SINK)
2509 rc = sptlrpc_svc_wrap_bulk(req, desc);
2511 rc = ptlrpc_start_bulk_transfer(desc);
2514 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2515 ptlrpc_abort_bulk(desc);
2516 } else if (rc == 0) {
2517 time_t start = cfs_time_current_sec();
2519 long timeoutl = req->rq_deadline - cfs_time_current_sec();
2520 cfs_duration_t timeout = timeoutl <= 0 ?
2521 CFS_TICK : cfs_time_seconds(timeoutl);
2522 *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2523 cfs_time_seconds(1),
2524 target_bulk_timeout,
2526 rc = l_wait_event(desc->bd_waitq,
2527 !ptlrpc_server_bulk_active(desc) ||
2529 exp->exp_abort_active_req,
2531 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2532 /* Wait again if we changed deadline */
2533 } while ((rc == -ETIMEDOUT) &&
2534 (req->rq_deadline > cfs_time_current_sec()));
2536 if (rc == -ETIMEDOUT) {
2537 DEBUG_REQ(D_ERROR, req,
2538 "timeout on bulk %s after %ld%+lds",
2540 req->rq_deadline - start,
2541 cfs_time_current_sec() -
2543 ptlrpc_abort_bulk(desc);
2544 } else if (exp->exp_failed) {
2545 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2548 ptlrpc_abort_bulk(desc);
2549 } else if (exp->exp_abort_active_req) {
2550 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2552 /* we don't reply anyway */
2554 ptlrpc_abort_bulk(desc);
2555 } else if (!desc->bd_success ||
2556 desc->bd_nob_transferred != desc->bd_nob) {
2557 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2559 "truncated" : "network error on",
2561 desc->bd_nob_transferred,
2563 /* XXX should this be a different errno? */
2565 } else if (desc->bd_type == BULK_GET_SINK) {
2566 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2569 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2570 bulk2type(desc), rc);
2575 EXPORT_SYMBOL(target_bulk_io);