1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_mds.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: if non-zero, move the selected to the list head
55 * @create: if zero, only search in existed connections
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 cfs_spin_lock(&imp->imp_lock);
84 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 cfs_list_del(&item->oic_item);
88 cfs_list_add(&item->oic_item,
90 item->oic_last_attempt = 0;
92 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? ", moved to head" : ""));
95 cfs_spin_unlock(&imp->imp_lock);
96 GOTO(out_free, rc = 0);
101 imp_conn->oic_conn = ptlrpc_conn;
102 imp_conn->oic_uuid = *uuid;
103 imp_conn->oic_last_attempt = 0;
105 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107 cfs_list_add_tail(&imp_conn->oic_item,
108 &imp->imp_conn_list);
109 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
110 imp, imp->imp_obd->obd_name, uuid->uuid,
111 (priority ? "head" : "tail"));
113 cfs_spin_unlock(&imp->imp_lock);
114 GOTO(out_free, rc = -ENOENT);
117 cfs_spin_unlock(&imp->imp_lock);
121 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 ptlrpc_connection_put(ptlrpc_conn);
127 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 return import_set_conn(imp, uuid, 1, 0);
132 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
135 return import_set_conn(imp, uuid, priority, 1);
138 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
140 struct obd_import_conn *imp_conn;
141 struct obd_export *dlmexp;
145 cfs_spin_lock(&imp->imp_lock);
146 if (cfs_list_empty(&imp->imp_conn_list)) {
147 LASSERT(!imp->imp_connection);
151 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
152 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
154 LASSERT(imp_conn->oic_conn);
156 /* is current conn? */
157 if (imp_conn == imp->imp_conn_current) {
158 LASSERT(imp_conn->oic_conn == imp->imp_connection);
160 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
161 imp->imp_state != LUSTRE_IMP_DISCON) {
162 CERROR("can't remove current connection\n");
163 GOTO(out, rc = -EBUSY);
166 ptlrpc_connection_put(imp->imp_connection);
167 imp->imp_connection = NULL;
169 dlmexp = class_conn2export(&imp->imp_dlm_handle);
170 if (dlmexp && dlmexp->exp_connection) {
171 LASSERT(dlmexp->exp_connection ==
173 ptlrpc_connection_put(dlmexp->exp_connection);
174 dlmexp->exp_connection = NULL;
178 cfs_list_del(&imp_conn->oic_item);
179 ptlrpc_connection_put(imp_conn->oic_conn);
180 OBD_FREE(imp_conn, sizeof(*imp_conn));
181 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182 imp, imp->imp_obd->obd_name, uuid->uuid);
187 cfs_spin_unlock(&imp->imp_lock);
189 CERROR("connection %s not found\n", uuid->uuid);
193 void client_destroy_import(struct obd_import *imp)
195 /* drop security policy instance after all rpc finished/aborted
196 * to let all busy contexts be released. */
197 class_import_get(imp);
198 class_destroy_import(imp);
199 sptlrpc_import_sec_put(imp);
200 class_import_put(imp);
203 /* configure an RPC client OBD device
208 * 3 - inactive-on-startup
210 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
212 struct client_obd *cli = &obddev->u.cli;
213 struct obd_import *imp;
214 struct obd_uuid server_uuid;
215 int rq_portal, rp_portal, connect_op;
216 char *name = obddev->obd_type->typ_name;
220 /* In a more perfect world, we would hang a ptlrpc_client off of
221 * obd_type and just use the values from there. */
222 if (!strcmp(name, LUSTRE_OSC_NAME)) {
223 rq_portal = OST_REQUEST_PORTAL;
224 rp_portal = OSC_REPLY_PORTAL;
225 connect_op = OST_CONNECT;
226 cli->cl_sp_me = LUSTRE_SP_CLI;
227 cli->cl_sp_to = LUSTRE_SP_OST;
228 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
229 rq_portal = MDS_REQUEST_PORTAL;
230 rp_portal = MDC_REPLY_PORTAL;
231 connect_op = MDS_CONNECT;
232 cli->cl_sp_me = LUSTRE_SP_CLI;
233 cli->cl_sp_to = LUSTRE_SP_MDT;
234 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
235 rq_portal = MGS_REQUEST_PORTAL;
236 rp_portal = MGC_REPLY_PORTAL;
237 connect_op = MGS_CONNECT;
238 cli->cl_sp_me = LUSTRE_SP_MGC;
239 cli->cl_sp_to = LUSTRE_SP_MGS;
240 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
242 CERROR("unknown client OBD type \"%s\", can't setup\n",
247 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
248 CERROR("requires a TARGET UUID\n");
252 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
253 CERROR("client UUID must be less than 38 characters\n");
257 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
258 CERROR("setup requires a SERVER UUID\n");
262 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
263 CERROR("target UUID must be less than 38 characters\n");
267 cfs_init_rwsem(&cli->cl_sem);
268 cfs_sema_init(&cli->cl_mgc_sem, 1);
269 cli->cl_conn_count = 0;
270 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
271 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
272 sizeof(server_uuid)));
275 cli->cl_avail_grant = 0;
276 /* FIXME: should limit this for the sum of all cl_dirty_max */
277 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
278 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
279 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
280 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
281 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
282 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
283 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
284 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
285 client_obd_list_lock_init(&cli->cl_loi_list_lock);
286 cli->cl_r_in_flight = 0;
287 cli->cl_w_in_flight = 0;
289 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
290 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
291 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
292 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
293 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
294 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
295 cfs_waitq_init(&cli->cl_destroy_waitq);
296 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
297 #ifdef ENABLE_CHECKSUM
298 /* Turn on checksumming by default. */
299 cli->cl_checksum = 1;
301 * The supported checksum types will be worked out at connect time
302 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
305 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
307 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
309 /* This value may be changed at connect time in
310 ptlrpc_connect_interpret. */
311 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
312 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
314 if (!strcmp(name, LUSTRE_MDC_NAME)) {
315 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
316 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
317 cli->cl_max_rpcs_in_flight = 2;
318 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
319 cli->cl_max_rpcs_in_flight = 3;
320 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
321 cli->cl_max_rpcs_in_flight = 4;
323 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
328 CERROR("ldlm_get_ref failed: %d\n", rc);
332 ptlrpc_init_client(rq_portal, rp_portal, name,
333 &obddev->obd_ldlm_client);
335 imp = class_new_import(obddev);
337 GOTO(err_ldlm, rc = -ENOENT);
338 imp->imp_client = &obddev->obd_ldlm_client;
339 imp->imp_connect_op = connect_op;
340 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
341 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
342 LUSTRE_CFG_BUFLEN(lcfg, 1));
343 class_import_put(imp);
345 rc = client_import_add_conn(imp, &server_uuid, 1);
347 CERROR("can't add initial connection\n");
348 GOTO(err_import, rc);
351 cli->cl_import = imp;
352 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
353 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
354 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
356 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
357 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
358 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
359 name, obddev->obd_name,
360 cli->cl_target_uuid.uuid);
361 cfs_spin_lock(&imp->imp_lock);
362 imp->imp_deactive = 1;
363 cfs_spin_unlock(&imp->imp_lock);
367 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
368 LDLM_NAMESPACE_CLIENT,
369 LDLM_NAMESPACE_GREEDY);
370 if (obddev->obd_namespace == NULL) {
371 CERROR("Unable to create client namespace - %s\n",
373 GOTO(err_import, rc = -ENOMEM);
376 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
381 class_destroy_import(imp);
389 int client_obd_cleanup(struct obd_device *obddev)
393 ldlm_namespace_free_post(obddev->obd_namespace);
394 obddev->obd_namespace = NULL;
400 /* ->o_connect() method for client side (OSC and MDC and MGC) */
401 int client_connect_import(const struct lu_env *env,
402 struct obd_export **exp,
403 struct obd_device *obd, struct obd_uuid *cluuid,
404 struct obd_connect_data *data, void *localdata)
406 struct client_obd *cli = &obd->u.cli;
407 struct obd_import *imp = cli->cl_import;
408 struct obd_connect_data *ocd;
409 struct lustre_handle conn = { 0 };
414 cfs_down_write(&cli->cl_sem);
415 if (cli->cl_conn_count > 0 )
416 GOTO(out_sem, rc = -EALREADY);
418 rc = class_connect(&conn, obd, cluuid);
422 cli->cl_conn_count++;
423 *exp = class_conn2export(&conn);
425 LASSERT(obd->obd_namespace);
427 imp->imp_dlm_handle = conn;
428 rc = ptlrpc_init_import(imp);
432 ocd = &imp->imp_connect_data;
435 imp->imp_connect_flags_orig = data->ocd_connect_flags;
438 rc = ptlrpc_connect_import(imp, NULL);
440 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
443 LASSERT((*exp)->exp_connection);
446 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
447 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
448 data->ocd_connect_flags, ocd->ocd_connect_flags);
449 data->ocd_connect_flags = ocd->ocd_connect_flags;
452 ptlrpc_pinger_add_import(imp);
458 cli->cl_conn_count--;
459 class_disconnect(*exp);
463 cfs_up_write(&cli->cl_sem);
468 int client_disconnect_export(struct obd_export *exp)
470 struct obd_device *obd = class_exp2obd(exp);
471 struct client_obd *cli;
472 struct obd_import *imp;
477 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
478 exp, exp ? exp->exp_handle.h_cookie : -1);
483 imp = cli->cl_import;
485 cfs_down_write(&cli->cl_sem);
486 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
489 if (!cli->cl_conn_count) {
490 CERROR("disconnecting disconnected device (%s)\n",
492 GOTO(out_disconnect, rc = -EINVAL);
495 cli->cl_conn_count--;
496 if (cli->cl_conn_count)
497 GOTO(out_disconnect, rc = 0);
499 /* Mark import deactivated now, so we don't try to reconnect if any
500 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
501 * fully deactivate the import, or that would drop all requests. */
502 cfs_spin_lock(&imp->imp_lock);
503 imp->imp_deactive = 1;
504 cfs_spin_unlock(&imp->imp_lock);
506 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
507 * delete it regardless. (It's safe to delete an import that was
509 (void)ptlrpc_pinger_del_import(imp);
511 if (obd->obd_namespace != NULL) {
512 /* obd_force == local only */
513 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
514 obd->obd_force ? LCF_LOCAL : 0, NULL);
515 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
519 * there's no need to hold sem during disconnecting an import,
520 * and actually it may cause deadlock in gss.
522 cfs_up_write(&cli->cl_sem);
523 rc = ptlrpc_disconnect_import(imp, 0);
524 cfs_down_write(&cli->cl_sem);
526 ptlrpc_invalidate_import(imp);
528 if (imp->imp_rq_pool) {
529 ptlrpc_free_rq_pool(imp->imp_rq_pool);
530 imp->imp_rq_pool = NULL;
532 client_destroy_import(imp);
533 cli->cl_import = NULL;
538 /* use server style - class_disconnect should be always called for
540 err = class_disconnect(exp);
544 cfs_up_write(&cli->cl_sem);
549 int server_disconnect_export(struct obd_export *exp)
554 /* Disconnect early so that clients can't keep using export */
555 rc = class_disconnect(exp);
556 /* close import for avoid sending any requests */
557 if (exp->exp_imp_reverse)
558 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
560 if (exp->exp_obd->obd_namespace != NULL)
561 ldlm_cancel_locks_for_export(exp);
563 /* complete all outstanding replies */
564 cfs_spin_lock(&exp->exp_lock);
565 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
566 struct ptlrpc_reply_state *rs =
567 cfs_list_entry(exp->exp_outstanding_replies.next,
568 struct ptlrpc_reply_state, rs_exp_list);
569 struct ptlrpc_service *svc = rs->rs_service;
571 cfs_spin_lock(&svc->srv_rs_lock);
572 cfs_list_del_init(&rs->rs_exp_list);
573 cfs_spin_lock(&rs->rs_lock);
574 ptlrpc_schedule_difficult_reply(rs);
575 cfs_spin_unlock(&rs->rs_lock);
576 cfs_spin_unlock(&svc->srv_rs_lock);
578 cfs_spin_unlock(&exp->exp_lock);
583 /* --------------------------------------------------------------------------
584 * from old lib/target.c
585 * -------------------------------------------------------------------------- */
587 static int target_handle_reconnect(struct lustre_handle *conn,
588 struct obd_export *exp,
589 struct obd_uuid *cluuid)
592 if (exp->exp_connection && exp->exp_imp_reverse) {
593 struct lustre_handle *hdl;
594 hdl = &exp->exp_imp_reverse->imp_remote_handle;
595 /* Might be a re-connect after a partition. */
596 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
597 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
599 conn->cookie = exp->exp_handle.h_cookie;
600 /* target_handle_connect() treats EALREADY and
601 * -EALREADY differently. EALREADY means we are
602 * doing a valid reconnect from the same client. */
605 CERROR("%s reconnecting from %s, "
606 "handle mismatch (ours "LPX64", theirs "
607 LPX64")\n", cluuid->uuid,
608 exp->exp_connection->c_remote_uuid.uuid,
609 hdl->cookie, conn->cookie);
610 memset(conn, 0, sizeof *conn);
611 /* target_handle_connect() treats EALREADY and
612 * -EALREADY differently. -EALREADY is an error
613 * (same UUID, different handle). */
618 conn->cookie = exp->exp_handle.h_cookie;
619 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
620 cluuid->uuid, exp, conn->cookie);
624 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
627 struct obd_export *exp = cb_data;
629 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
630 obd->obd_name, exp->exp_client_uuid.uuid);
632 cfs_spin_lock(&exp->exp_lock);
633 exp->exp_need_sync = 0;
634 cfs_spin_unlock(&exp->exp_lock);
635 class_export_cb_put(exp);
637 EXPORT_SYMBOL(target_client_add_cb);
640 target_start_and_reset_recovery_timer(struct obd_device *obd,
641 struct ptlrpc_request *req,
644 int target_handle_connect(struct ptlrpc_request *req)
646 struct obd_device *target, *targref = NULL;
647 struct obd_export *export = NULL;
648 struct obd_import *revimp;
649 struct lustre_handle conn;
650 struct lustre_handle *tmp;
651 struct obd_uuid tgtuuid;
652 struct obd_uuid cluuid;
653 struct obd_uuid remote_uuid;
657 struct obd_connect_data *data, *tmpdata;
658 lnet_nid_t *client_nid = NULL;
661 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
663 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
665 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
666 GOTO(out, rc = -EINVAL);
669 obd_str2uuid(&tgtuuid, str);
670 target = class_uuid2obd(&tgtuuid);
672 target = class_name2obd(str);
674 if (!target || target->obd_stopping || !target->obd_set_up) {
675 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
676 " for connect (%s)\n", str,
677 !target ? "no target" :
678 (target->obd_stopping ? "stopping" :
680 GOTO(out, rc = -ENODEV);
683 if (target->obd_no_conn) {
684 LCONSOLE_WARN("%s: temporarily refusing client connection "
685 "from %s\n", target->obd_name,
686 libcfs_nid2str(req->rq_peer.nid));
687 GOTO(out, rc = -EAGAIN);
690 /* Make sure the target isn't cleaned up while we're here. Yes,
691 there's still a race between the above check and our incref here.
692 Really, class_uuid2obd should take the ref. */
693 targref = class_incref(target, __FUNCTION__, cfs_current());
696 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
698 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
699 GOTO(out, rc = -EINVAL);
702 obd_str2uuid(&cluuid, str);
704 /* XXX extract a nettype and format accordingly */
705 switch (sizeof(lnet_nid_t)) {
706 /* NB the casts only avoid compiler warnings */
708 snprintf(remote_uuid.uuid, sizeof remote_uuid,
709 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
712 snprintf(remote_uuid.uuid, sizeof remote_uuid,
713 "NET_%x_UUID", (__u32)req->rq_peer.nid);
719 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
721 GOTO(out, rc = -EPROTO);
725 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
727 GOTO(out, rc = -EPROTO);
729 rc = req_capsule_server_pack(&req->rq_pill);
733 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
735 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
736 "libclient connection attempt");
737 GOTO(out, rc = -EPROTO);
738 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
739 LUSTRE_VERSION_ALLOWED_OFFSET ||
740 data->ocd_version > LUSTRE_VERSION_CODE +
741 LUSTRE_VERSION_ALLOWED_OFFSET) {
742 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
743 "libclient connection attempt",
744 data->ocd_version < LUSTRE_VERSION_CODE ?
746 OBD_OCD_VERSION_MAJOR(data->ocd_version),
747 OBD_OCD_VERSION_MINOR(data->ocd_version),
748 OBD_OCD_VERSION_PATCH(data->ocd_version),
749 OBD_OCD_VERSION_FIX(data->ocd_version));
750 data = req_capsule_server_sized_get(&req->rq_pill,
752 offsetof(typeof(*data), ocd_version) +
753 sizeof(data->ocd_version));
755 data->ocd_connect_flags = OBD_CONNECT_VERSION;
756 data->ocd_version = LUSTRE_VERSION_CODE;
758 GOTO(out, rc = -EPROTO);
762 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
763 (data->ocd_connect_flags & OBD_CONNECT_MDS))
766 /* lctl gets a backstage, all-access pass. */
767 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
768 goto dont_check_exports;
770 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
774 /* we've found an export in the hash */
775 if (export->exp_connecting) { /* bug 9635, et. al. */
776 CWARN("%s: exp %p already connecting\n",
777 export->exp_obd->obd_name, export);
778 class_export_put(export);
781 } else if (mds_conn && export->exp_connection) {
782 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
783 /* mds reconnected after failover */
784 CWARN("%s: received MDS connection from NID %s,"
785 " removing former export from NID %s\n",
786 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
787 libcfs_nid2str(export->exp_connection->c_peer.nid));
789 /* new mds connection from the same nid */
790 CWARN("%s: received new MDS connection from NID %s,"
791 " removing former export from same NID\n",
792 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
793 class_fail_export(export);
794 class_export_put(export);
797 } else if (export->exp_connection != NULL &&
798 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
799 (lustre_msg_get_op_flags(req->rq_reqmsg) &
800 MSG_CONNECT_INITIAL)) {
801 /* in mds failover we have static uuid but nid can be
803 CWARN("%s: cookie %s seen on new NID %s when "
804 "existing NID %s is already connected\n",
805 target->obd_name, cluuid.uuid,
806 libcfs_nid2str(req->rq_peer.nid),
807 libcfs_nid2str(export->exp_connection->c_peer.nid));
809 class_export_put(export);
812 cfs_spin_lock(&export->exp_lock);
813 export->exp_connecting = 1;
814 cfs_spin_unlock(&export->exp_lock);
815 class_export_put(export);
816 LASSERT(export->exp_obd == target);
818 rc = target_handle_reconnect(&conn, export, &cluuid);
821 /* If we found an export, we already unlocked. */
824 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
825 } else if (req->rq_export == NULL &&
826 cfs_atomic_read(&export->exp_rpc_count) > 0) {
827 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
828 target->obd_name, cluuid.uuid,
829 libcfs_nid2str(req->rq_peer.nid),
830 export, cfs_atomic_read(&export->exp_refcount));
831 GOTO(out, rc = -EBUSY);
832 } else if (req->rq_export != NULL &&
833 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
834 /* the current connect rpc has increased exp_rpc_count */
835 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
836 target->obd_name, cluuid.uuid,
837 libcfs_nid2str(req->rq_peer.nid),
838 export, cfs_atomic_read(&export->exp_rpc_count) - 1);
839 cfs_spin_lock(&export->exp_lock);
840 if (req->rq_export->exp_conn_cnt <
841 lustre_msg_get_conn_cnt(req->rq_reqmsg))
842 /* try to abort active requests */
843 req->rq_export->exp_abort_active_req = 1;
844 cfs_spin_unlock(&export->exp_lock);
845 GOTO(out, rc = -EBUSY);
846 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
847 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
848 "cookies not random?\n", target->obd_name,
849 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
850 GOTO(out, rc = -EALREADY);
852 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
859 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
860 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
861 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
862 export, (long)cfs_time_current_sec(),
863 export ? (long)export->exp_last_request_time : 0);
865 /* Tell the client if we're in recovery. */
866 if (target->obd_recovering) {
867 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
868 /* If this is the first time a client connects,
869 reset the recovery timer */
871 target_start_and_reset_recovery_timer(target, req,
875 /* We want to handle EALREADY but *not* -EALREADY from
876 * target_handle_reconnect(), return reconnection state in a flag */
877 if (rc == EALREADY) {
878 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
884 /* Tell the client if we support replayable requests */
885 if (target->obd_replayable)
886 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
887 client_nid = &req->rq_peer.nid;
889 if (export == NULL) {
890 if (target->obd_recovering) {
893 t = cfs_timer_deadline(&target->obd_recovery_timer);
894 t = cfs_time_sub(t, cfs_time_current());
895 CERROR("%s: denying connection for new client %s (%s): "
896 "%d clients in recovery for "CFS_TIME_T"s\n",
898 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
899 cfs_atomic_read(&target-> \
900 obd_lock_replay_clients),
901 cfs_duration_sec(t));
905 rc = obd_connect(req->rq_svc_thread->t_env,
906 &export, target, &cluuid, data,
909 conn.cookie = export->exp_handle.h_cookie;
912 rc = obd_reconnect(req->rq_svc_thread->t_env,
913 export, target, &cluuid, data, client_nid);
915 /* prevous done via class_conn2export */
916 class_export_get(export);
920 /* Return only the parts of obd_connect_data that we understand, so the
921 * client knows that we don't understand the rest. */
923 tmpdata = req_capsule_server_get(&req->rq_pill,
925 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
929 /* If all else goes well, this is our RPC return code. */
932 lustre_msg_set_handle(req->rq_repmsg, &conn);
934 /* If the client and the server are the same node, we will already
935 * have an export that really points to the client's DLM export,
936 * because we have a shared handles table.
938 * XXX this will go away when shaver stops sending the "connect" handle
939 * in the real "remote handle" field of the request --phik 24 Apr 2003
941 if (req->rq_export != NULL)
942 class_export_put(req->rq_export);
944 req->rq_export = export;
946 cfs_spin_lock(&export->exp_lock);
947 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
948 cfs_spin_unlock(&export->exp_lock);
949 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
950 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
951 export->exp_conn_cnt,
952 lustre_msg_get_conn_cnt(req->rq_reqmsg));
954 GOTO(out, rc = -EALREADY);
956 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
957 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
958 export->exp_abort_active_req = 0;
960 /* request from liblustre? Don't evict it for not pinging. */
961 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
962 export->exp_libclient = 1;
963 cfs_spin_unlock(&export->exp_lock);
965 cfs_spin_lock(&target->obd_dev_lock);
966 cfs_list_del_init(&export->exp_obd_chain_timed);
967 cfs_spin_unlock(&target->obd_dev_lock);
969 cfs_spin_unlock(&export->exp_lock);
972 if (export->exp_connection != NULL) {
973 /* Check to see if connection came from another NID */
974 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
975 !cfs_hlist_unhashed(&export->exp_nid_hash))
976 cfs_hash_del(export->exp_obd->obd_nid_hash,
977 &export->exp_connection->c_peer.nid,
978 &export->exp_nid_hash);
980 ptlrpc_connection_put(export->exp_connection);
983 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
986 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
987 cfs_hash_add(export->exp_obd->obd_nid_hash,
988 &export->exp_connection->c_peer.nid,
989 &export->exp_nid_hash);
992 cfs_spin_lock(&target->obd_recovery_task_lock);
993 if (target->obd_recovering && !export->exp_in_recovery) {
994 cfs_spin_lock(&export->exp_lock);
995 export->exp_in_recovery = 1;
996 export->exp_req_replay_needed = 1;
997 export->exp_lock_replay_needed = 1;
998 cfs_spin_unlock(&export->exp_lock);
999 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1000 && (data->ocd_transno == 0))
1001 CWARN("Connect with zero transno!\n");
1003 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1004 && data->ocd_transno < target->obd_next_recovery_transno)
1005 target->obd_next_recovery_transno = data->ocd_transno;
1006 target->obd_connected_clients++;
1007 cfs_atomic_inc(&target->obd_req_replay_clients);
1008 cfs_atomic_inc(&target->obd_lock_replay_clients);
1009 if (target->obd_connected_clients ==
1010 target->obd_max_recoverable_clients)
1011 cfs_waitq_signal(&target->obd_next_transno_waitq);
1013 cfs_spin_unlock(&target->obd_recovery_task_lock);
1014 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1017 if (export->exp_imp_reverse != NULL) {
1018 /* destroyed import can be still referenced in ctxt */
1019 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1020 KEY_REVIMP_UPD, 0, NULL, NULL);
1022 client_destroy_import(export->exp_imp_reverse);
1025 /* for the rest part, we return -ENOTCONN in case of errors
1026 * in order to let client initialize connection again.
1028 revimp = export->exp_imp_reverse = class_new_import(target);
1030 CERROR("fail to alloc new reverse import.\n");
1031 GOTO(out, rc = -ENOTCONN);
1034 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1035 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1036 revimp->imp_remote_handle = conn;
1037 revimp->imp_dlm_fake = 1;
1038 revimp->imp_state = LUSTRE_IMP_FULL;
1040 /* unknown versions will be caught in
1041 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1042 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1044 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1045 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1046 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1048 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1050 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1051 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1052 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1054 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1056 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1058 CERROR("Failed to get sec for reverse import: %d\n", rc);
1059 export->exp_imp_reverse = NULL;
1060 class_destroy_import(revimp);
1063 class_import_put(revimp);
1066 cfs_spin_lock(&export->exp_lock);
1067 export->exp_connecting = 0;
1068 cfs_spin_unlock(&export->exp_lock);
1071 class_decref(targref, __FUNCTION__, cfs_current());
1073 req->rq_status = rc;
1077 int target_handle_disconnect(struct ptlrpc_request *req)
1082 rc = req_capsule_server_pack(&req->rq_pill);
1086 /* keep the rq_export around so we can send the reply */
1087 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1092 void target_destroy_export(struct obd_export *exp)
1094 /* exports created from last_rcvd data, and "fake"
1095 exports created by lctl don't have an import */
1096 if (exp->exp_imp_reverse != NULL)
1097 client_destroy_import(exp->exp_imp_reverse);
1099 LASSERT(cfs_atomic_read(&exp->exp_locks_count) == 0);
1100 LASSERT(cfs_atomic_read(&exp->exp_rpc_count) == 0);
1101 LASSERT(cfs_atomic_read(&exp->exp_cb_count) == 0);
1102 LASSERT(cfs_atomic_read(&exp->exp_replay_count) == 0);
1106 * Recovery functions
1108 static void target_request_copy_get(struct ptlrpc_request *req)
1110 class_export_rpc_get(req->rq_export);
1111 LASSERT(cfs_list_empty(&req->rq_list));
1112 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1113 /* increase refcount to keep request in queue */
1114 LASSERT(cfs_atomic_read(&req->rq_refcount));
1115 cfs_atomic_inc(&req->rq_refcount);
1116 /** let export know it has replays to be handled */
1117 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1118 /* release service thread while request is queued
1119 * we are moving the request from active processing
1120 * to waiting on the replay queue */
1121 ptlrpc_server_active_request_dec(req);
1124 static void target_request_copy_put(struct ptlrpc_request *req)
1126 LASSERT(cfs_list_empty(&req->rq_replay_list));
1127 LASSERT(cfs_atomic_read(&req->rq_export->exp_replay_count) > 0);
1128 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1129 class_export_rpc_put(req->rq_export);
1130 /* ptlrpc_server_drop_request() assumes the request is active */
1131 ptlrpc_server_active_request_inc(req);
1132 ptlrpc_server_drop_request(req);
1135 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1137 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1138 struct obd_export *exp = req->rq_export;
1139 struct ptlrpc_request *reqiter;
1144 cfs_spin_lock(&exp->exp_lock);
1145 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1147 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1154 /* we expect it with RESENT and REPLAY flags */
1155 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1156 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1157 CERROR("invalid flags %x of resent replay\n",
1158 lustre_msg_get_flags(req->rq_reqmsg));
1160 cfs_list_add_tail(&req->rq_replay_list,
1161 &exp->exp_req_replay_queue);
1164 cfs_spin_unlock(&exp->exp_lock);
1168 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1170 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1171 LASSERT(req->rq_export);
1173 cfs_spin_lock(&req->rq_export->exp_lock);
1174 cfs_list_del_init(&req->rq_replay_list);
1175 cfs_spin_unlock(&req->rq_export->exp_lock);
1179 static void target_finish_recovery(struct obd_device *obd)
1182 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1185 ldlm_reprocess_all_ns(obd->obd_namespace);
1186 cfs_spin_lock(&obd->obd_recovery_task_lock);
1187 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1188 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1189 !cfs_list_empty(&obd->obd_final_req_queue)) {
1190 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1192 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1193 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1195 cfs_list_empty(&obd->obd_final_req_queue) ? \
1197 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1200 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1202 obd->obd_recovery_end = cfs_time_current_sec();
1204 /* when recovery finished, cleanup orphans on mds and ost */
1205 if (OBT(obd) && OBP(obd, postrecov)) {
1206 int rc = OBP(obd, postrecov)(obd);
1208 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1214 static void abort_req_replay_queue(struct obd_device *obd)
1216 struct ptlrpc_request *req, *n;
1217 cfs_list_t abort_list;
1219 CFS_INIT_LIST_HEAD(&abort_list);
1220 cfs_spin_lock(&obd->obd_recovery_task_lock);
1221 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1222 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1223 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1224 DEBUG_REQ(D_WARNING, req, "aborted:");
1225 req->rq_status = -ENOTCONN;
1226 if (ptlrpc_error(req)) {
1227 DEBUG_REQ(D_ERROR, req,
1228 "failed abort_req_reply; skipping");
1230 target_exp_dequeue_req_replay(req);
1231 target_request_copy_put(req);
1235 static void abort_lock_replay_queue(struct obd_device *obd)
1237 struct ptlrpc_request *req, *n;
1238 cfs_list_t abort_list;
1240 CFS_INIT_LIST_HEAD(&abort_list);
1241 cfs_spin_lock(&obd->obd_recovery_task_lock);
1242 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1243 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1244 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1245 DEBUG_REQ(D_ERROR, req, "aborted:");
1246 req->rq_status = -ENOTCONN;
1247 if (ptlrpc_error(req)) {
1248 DEBUG_REQ(D_ERROR, req,
1249 "failed abort_lock_reply; skipping");
1251 target_request_copy_put(req);
1256 /* Called from a cleanup function if the device is being cleaned up
1257 forcefully. The exports should all have been disconnected already,
1258 the only thing left to do is
1259 - clear the recovery flags
1261 - free queued requests and replies, but don't send replies
1262 Because the obd_stopping flag is set, no new requests should be received.
1265 void target_cleanup_recovery(struct obd_device *obd)
1267 struct ptlrpc_request *req, *n;
1268 cfs_list_t clean_list;
1271 CFS_INIT_LIST_HEAD(&clean_list);
1272 cfs_spin_lock(&obd->obd_dev_lock);
1273 if (!obd->obd_recovering) {
1274 cfs_spin_unlock(&obd->obd_dev_lock);
1278 obd->obd_recovering = obd->obd_abort_recovery = 0;
1279 cfs_spin_unlock(&obd->obd_dev_lock);
1281 cfs_spin_lock(&obd->obd_recovery_task_lock);
1282 target_cancel_recovery_timer(obd);
1283 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1284 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1286 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1287 LASSERT(req->rq_reply_state == 0);
1288 target_exp_dequeue_req_replay(req);
1289 target_request_copy_put(req);
1292 cfs_spin_lock(&obd->obd_recovery_task_lock);
1293 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1294 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1295 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1297 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1298 LASSERT(req->rq_reply_state == 0);
1299 target_request_copy_put(req);
1305 /* obd_recovery_task_lock should be held */
1306 void target_cancel_recovery_timer(struct obd_device *obd)
1308 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1309 cfs_timer_disarm(&obd->obd_recovery_timer);
1312 /* extend = 1 means require at least "duration" seconds left in the timer,
1313 extend = 0 means set the total duration (start_recovery_timer) */
1314 static void reset_recovery_timer(struct obd_device *obd, int duration,
1317 cfs_time_t now = cfs_time_current_sec();
1318 cfs_duration_t left;
1320 cfs_spin_lock(&obd->obd_recovery_task_lock);
1321 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1322 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1326 left = cfs_time_sub(obd->obd_recovery_end, now);
1328 if (extend && (duration > left))
1329 obd->obd_recovery_timeout += duration - left;
1330 else if (!extend && (duration > obd->obd_recovery_timeout))
1331 /* Track the client's largest expected replay time */
1332 obd->obd_recovery_timeout = duration;
1334 /* Hard limit of obd_recovery_time_hard which should not happen */
1335 if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1336 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1338 obd->obd_recovery_end = obd->obd_recovery_start +
1339 obd->obd_recovery_timeout;
1340 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1341 cfs_time_before(now, obd->obd_recovery_end)) {
1342 left = cfs_time_sub(obd->obd_recovery_end, now);
1343 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1345 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1346 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1347 obd->obd_name, (unsigned)left);
1350 static void check_and_start_recovery_timer(struct obd_device *obd)
1352 cfs_spin_lock(&obd->obd_recovery_task_lock);
1353 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1354 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1357 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1358 obd->obd_recovery_start = cfs_time_current_sec();
1359 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1361 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1364 /* Reset the timer with each new client connection */
1366 * This timer is actually reconnect_timer, which is for making sure
1367 * the total recovery window is at least as big as my reconnect
1368 * attempt timing. So the initial recovery time_out will be set to
1369 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1370 * from client is bigger than this, then the recovery time_out will
1371 * be extended to make sure the client could be reconnected, in the
1372 * process, the timeout from the new client should be ignored.
1376 target_start_and_reset_recovery_timer(struct obd_device *obd,
1377 struct ptlrpc_request *req,
1380 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1382 if (!new_client && service_time)
1383 /* Teach server about old server's estimates, as first guess
1384 * at how long new requests will take. */
1385 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1388 check_and_start_recovery_timer(obd);
1390 /* convert the service time to rpc timeout,
1391 * reuse service_time to limit stack usage */
1392 service_time = at_est2timeout(service_time);
1394 /* We expect other clients to timeout within service_time, then try
1395 * to reconnect, then try the failover server. The max delay between
1396 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1397 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1398 INITIAL_CONNECT_TIMEOUT);
1399 if (service_time > obd->obd_recovery_timeout && !new_client)
1400 reset_recovery_timer(obd, service_time, 0);
1405 /** Health checking routines */
1406 static inline int exp_connect_healthy(struct obd_export *exp)
1408 return (exp->exp_in_recovery);
1411 /** if export done req_replay or has replay in queue */
1412 static inline int exp_req_replay_healthy(struct obd_export *exp)
1414 return (!exp->exp_req_replay_needed ||
1415 cfs_atomic_read(&exp->exp_replay_count) > 0);
1417 /** if export done lock_replay or has replay in queue */
1418 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1420 return (!exp->exp_lock_replay_needed ||
1421 cfs_atomic_read(&exp->exp_replay_count) > 0);
1424 static inline int exp_vbr_healthy(struct obd_export *exp)
1426 return (!exp->exp_vbr_failed);
1429 static inline int exp_finished(struct obd_export *exp)
1431 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1434 /** Checking routines for recovery */
1435 static int check_for_clients(struct obd_device *obd)
1437 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1439 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1440 if (obd->obd_no_conn == 0 &&
1441 obd->obd_connected_clients + obd->obd_stale_clients ==
1442 obd->obd_max_recoverable_clients)
1447 static int check_for_next_transno(struct obd_device *obd)
1449 struct ptlrpc_request *req = NULL;
1450 int wake_up = 0, connected, completed, queue_len;
1451 __u64 next_transno, req_transno;
1454 cfs_spin_lock(&obd->obd_recovery_task_lock);
1455 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1456 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1457 struct ptlrpc_request, rq_list);
1458 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1463 connected = obd->obd_connected_clients;
1464 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1465 queue_len = obd->obd_requests_queued_for_recovery;
1466 next_transno = obd->obd_next_recovery_transno;
1468 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1469 "req_transno: "LPU64", next_transno: "LPU64"\n",
1470 obd->obd_max_recoverable_clients, connected, completed,
1471 queue_len, req_transno, next_transno);
1473 if (obd->obd_abort_recovery) {
1474 CDEBUG(D_HA, "waking for aborted recovery\n");
1476 } else if (obd->obd_recovery_expired) {
1477 CDEBUG(D_HA, "waking for expired recovery\n");
1479 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1480 CDEBUG(D_HA, "waking for completed recovery\n");
1482 } else if (req_transno == next_transno) {
1483 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1485 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1487 /** handle gaps occured due to lost reply or VBR */
1488 LASSERTF(req_transno >= next_transno,
1489 "req_transno: "LPU64", next_transno: "LPU64"\n",
1490 req_transno, next_transno);
1491 if (req_transno > obd->obd_last_committed &&
1492 !obd->obd_version_recov)
1495 "%s: waking for gap in transno, VBR is %s (skip: "
1496 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1497 ", last_committed: "LPD64")\n",
1498 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1499 next_transno, queue_len, completed, connected,
1500 req_transno, obd->obd_last_committed);
1501 obd->obd_next_recovery_transno = req_transno;
1503 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1504 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1505 " by fail_lock, waking up ("LPD64")\n", next_transno);
1506 obd->obd_next_recovery_transno = req_transno;
1509 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1513 static int check_for_next_lock(struct obd_device *obd)
1517 cfs_spin_lock(&obd->obd_recovery_task_lock);
1518 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1519 CDEBUG(D_HA, "waking for next lock\n");
1521 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1522 CDEBUG(D_HA, "waking for completed lock replay\n");
1524 } else if (obd->obd_abort_recovery) {
1525 CDEBUG(D_HA, "waking for aborted recovery\n");
1527 } else if (obd->obd_recovery_expired) {
1528 CDEBUG(D_HA, "waking for expired recovery\n");
1531 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1537 * wait for recovery events,
1538 * check its status with help of check_routine
1539 * evict dead clients via health_check
1541 static int target_recovery_overseer(struct obd_device *obd,
1542 int (*check_routine)(struct obd_device *),
1543 int (*health_check)(struct obd_export *))
1546 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1547 if (obd->obd_abort_recovery) {
1548 CWARN("recovery is aborted, evict exports in recovery\n");
1549 /** evict exports which didn't finish recovery yet */
1550 class_disconnect_stale_exports(obd, exp_finished);
1552 } else if (obd->obd_recovery_expired) {
1553 obd->obd_recovery_expired = 0;
1554 /** If some clients died being recovered, evict them */
1556 "recovery is timed out, evict stale exports\n");
1557 /** evict cexports with no replay in queue, they are stalled */
1558 class_disconnect_stale_exports(obd, health_check);
1559 /** continue with VBR */
1560 cfs_spin_lock(&obd->obd_dev_lock);
1561 obd->obd_version_recov = 1;
1562 cfs_spin_unlock(&obd->obd_dev_lock);
1564 * reset timer, recovery will proceed with versions now,
1565 * timeout is set just to handle reconnection delays
1567 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1568 /** Wait for recovery events again, after evicting bad clients */
1574 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1576 struct ptlrpc_request *req = NULL;
1579 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1580 obd->obd_next_recovery_transno);
1582 if (target_recovery_overseer(obd, check_for_next_transno,
1583 exp_req_replay_healthy)) {
1584 abort_req_replay_queue(obd);
1585 abort_lock_replay_queue(obd);
1588 cfs_spin_lock(&obd->obd_recovery_task_lock);
1589 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1590 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1591 struct ptlrpc_request, rq_list);
1592 cfs_list_del_init(&req->rq_list);
1593 obd->obd_requests_queued_for_recovery--;
1594 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1596 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1597 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1598 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1599 /** evict exports failed VBR */
1600 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1605 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1607 struct ptlrpc_request *req = NULL;
1609 CDEBUG(D_HA, "Waiting for lock\n");
1610 if (target_recovery_overseer(obd, check_for_next_lock,
1611 exp_lock_replay_healthy))
1612 abort_lock_replay_queue(obd);
1614 cfs_spin_lock(&obd->obd_recovery_task_lock);
1615 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1616 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1617 struct ptlrpc_request, rq_list);
1618 cfs_list_del_init(&req->rq_list);
1619 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1621 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1622 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1623 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1624 /** evict exports failed VBR */
1625 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1630 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1632 struct ptlrpc_request *req = NULL;
1634 cfs_spin_lock(&obd->obd_recovery_task_lock);
1635 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1636 req = cfs_list_entry(obd->obd_final_req_queue.next,
1637 struct ptlrpc_request, rq_list);
1638 cfs_list_del_init(&req->rq_list);
1639 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1640 if (req->rq_export->exp_in_recovery) {
1641 cfs_spin_lock(&req->rq_export->exp_lock);
1642 req->rq_export->exp_in_recovery = 0;
1643 cfs_spin_unlock(&req->rq_export->exp_lock);
1646 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1651 static int handle_recovery_req(struct ptlrpc_thread *thread,
1652 struct ptlrpc_request *req,
1653 svc_handler_t handler)
1659 * export can be evicted during recovery, no need to handle replays for
1660 * it after that, discard such request silently
1662 if (req->rq_export->exp_disconnected)
1663 GOTO(reqcopy_put, rc = 0);
1665 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1667 CERROR("Failure to initialize session: %d\n", rc);
1668 GOTO(reqcopy_put, rc);
1671 req->rq_recov_session.lc_thread = thread;
1672 lu_context_enter(&req->rq_recov_session);
1673 req->rq_svc_thread = thread;
1674 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1676 /* thread context */
1677 lu_context_enter(&thread->t_env->le_ctx);
1679 lu_context_exit(&thread->t_env->le_ctx);
1681 lu_context_exit(&req->rq_recov_session);
1682 lu_context_fini(&req->rq_recov_session);
1683 /* don't reset timer for final stage */
1684 if (!exp_finished(req->rq_export)) {
1686 * Add request timeout to the recovery time so next request from
1687 * this client may come in recovery time
1689 reset_recovery_timer(class_exp2obd(req->rq_export),
1690 AT_OFF ? obd_timeout :
1691 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1697 static int target_recovery_thread(void *arg)
1699 struct lu_target *lut = arg;
1700 struct obd_device *obd = lut->lut_obd;
1701 struct ptlrpc_request *req;
1702 struct target_recovery_data *trd = &obd->obd_recovery_data;
1703 unsigned long delta;
1704 unsigned long flags;
1706 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1710 cfs_daemonize_ctxt("tgt_recov");
1712 SIGNAL_MASK_LOCK(current, flags);
1713 sigfillset(¤t->blocked);
1715 SIGNAL_MASK_UNLOCK(current, flags);
1717 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1721 thread->t_env = &env;
1722 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1723 env.le_ctx.lc_thread = thread;
1724 thread->t_data = NULL;
1726 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1728 trd->trd_processing_task = cfs_curproc_pid();
1730 cfs_spin_lock(&obd->obd_dev_lock);
1731 obd->obd_recovering = 1;
1732 cfs_spin_unlock(&obd->obd_dev_lock);
1733 cfs_complete(&trd->trd_starting);
1735 /* first of all, we have to know the first transno to replay */
1736 if (target_recovery_overseer(obd, check_for_clients,
1737 exp_connect_healthy)) {
1738 abort_req_replay_queue(obd);
1739 abort_lock_replay_queue(obd);
1742 /* next stage: replay requests */
1744 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1745 cfs_atomic_read(&obd->obd_req_replay_clients),
1746 obd->obd_next_recovery_transno);
1747 while ((req = target_next_replay_req(obd))) {
1748 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1749 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1750 lustre_msg_get_transno(req->rq_reqmsg),
1751 libcfs_nid2str(req->rq_peer.nid));
1752 handle_recovery_req(thread, req,
1753 trd->trd_recovery_handler);
1755 * bz18031: increase next_recovery_transno before
1756 * target_request_copy_put() will drop exp_rpc reference
1758 cfs_spin_lock(&obd->obd_recovery_task_lock);
1759 obd->obd_next_recovery_transno++;
1760 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1761 target_exp_dequeue_req_replay(req);
1762 target_request_copy_put(req);
1763 obd->obd_replayed_requests++;
1767 * The second stage: replay locks
1769 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1770 cfs_atomic_read(&obd->obd_lock_replay_clients));
1771 while ((req = target_next_replay_lock(obd))) {
1772 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1773 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1774 libcfs_nid2str(req->rq_peer.nid));
1775 handle_recovery_req(thread, req,
1776 trd->trd_recovery_handler);
1777 target_request_copy_put(req);
1778 obd->obd_replayed_locks++;
1782 * The third stage: reply on final pings, at this moment all clients
1783 * must have request in final queue
1785 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1786 /** Update server last boot epoch */
1787 lut_boot_epoch_update(lut);
1788 /* We drop recoverying flag to forward all new requests
1789 * to regular mds_handle() since now */
1790 cfs_spin_lock(&obd->obd_dev_lock);
1791 obd->obd_recovering = obd->obd_abort_recovery = 0;
1792 cfs_spin_unlock(&obd->obd_dev_lock);
1793 cfs_spin_lock(&obd->obd_recovery_task_lock);
1794 target_cancel_recovery_timer(obd);
1795 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1796 while ((req = target_next_final_ping(obd))) {
1797 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1798 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1799 libcfs_nid2str(req->rq_peer.nid));
1800 handle_recovery_req(thread, req,
1801 trd->trd_recovery_handler);
1802 target_request_copy_put(req);
1805 delta = (jiffies - delta) / CFS_HZ;
1806 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1807 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1808 if (delta > OBD_RECOVERY_TIME_SOFT) {
1809 CWARN("too long recovery - read logs\n");
1810 libcfs_debug_dumplog();
1813 target_finish_recovery(obd);
1815 lu_context_fini(&env.le_ctx);
1816 trd->trd_processing_task = 0;
1817 cfs_complete(&trd->trd_finishing);
1821 static int target_start_recovery_thread(struct lu_target *lut,
1822 svc_handler_t handler)
1824 struct obd_device *obd = lut->lut_obd;
1826 struct target_recovery_data *trd = &obd->obd_recovery_data;
1828 memset(trd, 0, sizeof(*trd));
1829 cfs_init_completion(&trd->trd_starting);
1830 cfs_init_completion(&trd->trd_finishing);
1831 trd->trd_recovery_handler = handler;
1833 if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
1834 cfs_wait_for_completion(&trd->trd_starting);
1835 LASSERT(obd->obd_recovering != 0);
1842 void target_stop_recovery_thread(struct obd_device *obd)
1844 if (obd->obd_recovery_data.trd_processing_task > 0) {
1845 struct target_recovery_data *trd = &obd->obd_recovery_data;
1846 /** recovery can be done but postrecovery is not yet */
1847 cfs_spin_lock(&obd->obd_dev_lock);
1848 if (obd->obd_recovering) {
1849 CERROR("%s: Aborting recovery\n", obd->obd_name);
1850 obd->obd_abort_recovery = 1;
1851 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1853 cfs_spin_unlock(&obd->obd_dev_lock);
1854 cfs_wait_for_completion(&trd->trd_finishing);
1858 void target_recovery_fini(struct obd_device *obd)
1860 class_disconnect_exports(obd);
1861 target_stop_recovery_thread(obd);
1862 target_cleanup_recovery(obd);
1864 EXPORT_SYMBOL(target_recovery_fini);
1866 static void target_recovery_expired(unsigned long castmeharder)
1868 struct obd_device *obd = (struct obd_device *)castmeharder;
1869 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1870 " after %lds (%d clients connected)\n",
1871 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1872 cfs_time_current_sec()- obd->obd_recovery_start,
1873 obd->obd_connected_clients);
1875 obd->obd_recovery_expired = 1;
1876 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1879 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1881 struct obd_device *obd = lut->lut_obd;
1882 if (obd->obd_max_recoverable_clients == 0) {
1883 /** Update server last boot epoch */
1884 lut_boot_epoch_update(lut);
1888 CWARN("RECOVERY: service %s, %d recoverable clients, "
1889 "last_transno "LPU64"\n", obd->obd_name,
1890 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1891 LASSERT(obd->obd_stopping == 0);
1892 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1893 obd->obd_recovery_start = 0;
1894 obd->obd_recovery_end = 0;
1896 /* both values can be get from mount data already */
1897 if (obd->obd_recovery_timeout == 0)
1898 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
1899 if (obd->obd_recovery_time_hard == 0)
1900 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
1901 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1902 target_start_recovery_thread(lut, handler);
1904 EXPORT_SYMBOL(target_recovery_init);
1908 static int target_process_req_flags(struct obd_device *obd,
1909 struct ptlrpc_request *req)
1911 struct obd_export *exp = req->rq_export;
1912 LASSERT(exp != NULL);
1913 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1914 /* client declares he's ready to replay locks */
1915 cfs_spin_lock(&exp->exp_lock);
1916 if (exp->exp_req_replay_needed) {
1917 exp->exp_req_replay_needed = 0;
1918 cfs_spin_unlock(&exp->exp_lock);
1919 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients));
1920 cfs_atomic_dec(&obd->obd_req_replay_clients);
1922 cfs_spin_unlock(&exp->exp_lock);
1925 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1926 /* client declares he's ready to complete recovery
1927 * so, we put the request on th final queue */
1928 cfs_spin_lock(&exp->exp_lock);
1929 if (exp->exp_lock_replay_needed) {
1930 exp->exp_lock_replay_needed = 0;
1931 cfs_spin_unlock(&exp->exp_lock);
1932 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients));
1933 cfs_atomic_dec(&obd->obd_lock_replay_clients);
1935 cfs_spin_unlock(&exp->exp_lock);
1941 int target_queue_recovery_request(struct ptlrpc_request *req,
1942 struct obd_device *obd)
1946 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1949 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1950 /* Processing the queue right now, don't re-add. */
1954 target_process_req_flags(obd, req);
1956 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1957 /* client declares he's ready to complete recovery
1958 * so, we put the request on th final queue */
1959 target_request_copy_get(req);
1960 DEBUG_REQ(D_HA, req, "queue final req");
1961 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1962 cfs_spin_lock(&obd->obd_recovery_task_lock);
1963 if (obd->obd_recovering) {
1964 cfs_list_add_tail(&req->rq_list,
1965 &obd->obd_final_req_queue);
1967 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1968 target_request_copy_put(req);
1969 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1971 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1974 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1975 /* client declares he's ready to replay locks */
1976 target_request_copy_get(req);
1977 DEBUG_REQ(D_HA, req, "queue lock replay req");
1978 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1979 cfs_spin_lock(&obd->obd_recovery_task_lock);
1980 LASSERT(obd->obd_recovering);
1981 /* usually due to recovery abort */
1982 if (!req->rq_export->exp_in_recovery) {
1983 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1984 target_request_copy_put(req);
1987 LASSERT(req->rq_export->exp_lock_replay_needed);
1988 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1989 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1993 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1994 * (i.e. buflens etc are in my own byte order), but type-dependent
1995 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1998 CFS_INIT_LIST_HEAD(&req->rq_list);
1999 DEBUG_REQ(D_HA, req, "not queueing");
2003 /* If we're processing the queue, we want don't want to queue this
2006 * Also, if this request has a transno less than the one we're waiting
2007 * for, we should process it now. It could (and currently always will)
2008 * be an open request for a descriptor that was opened some time ago.
2010 * Also, a resent, replayed request that has already been
2011 * handled will pass through here and be processed immediately.
2013 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2014 obd->obd_next_recovery_transno, transno);
2015 cfs_spin_lock(&obd->obd_recovery_task_lock);
2016 if (transno < obd->obd_next_recovery_transno) {
2017 /* Processing the queue right now, don't re-add. */
2018 LASSERT(cfs_list_empty(&req->rq_list));
2019 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2022 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2024 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2027 target_request_copy_get(req);
2028 if (!req->rq_export->exp_in_recovery) {
2029 target_request_copy_put(req);
2032 LASSERT(req->rq_export->exp_req_replay_needed);
2034 if (target_exp_enqueue_req_replay(req)) {
2035 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2036 target_request_copy_put(req);
2041 cfs_spin_lock(&obd->obd_recovery_task_lock);
2042 LASSERT(obd->obd_recovering);
2043 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2044 struct ptlrpc_request *reqiter =
2045 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2047 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2048 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2053 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2055 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2056 "has been claimed by another client");
2057 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2058 target_exp_dequeue_req_replay(req);
2059 target_request_copy_put(req);
2065 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2067 obd->obd_requests_queued_for_recovery++;
2068 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2069 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2074 * Packs current SLV and Limit into \a req.
2076 int target_pack_pool_reply(struct ptlrpc_request *req)
2078 struct obd_device *obd;
2082 * Check that we still have all structures alive as this may
2083 * be some late rpc in shutdown time.
2085 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2086 !exp_connect_lru_resize(req->rq_export))) {
2087 lustre_msg_set_slv(req->rq_repmsg, 0);
2088 lustre_msg_set_limit(req->rq_repmsg, 0);
2093 * OBD is alive here as export is alive, which we checked above.
2095 obd = req->rq_export->exp_obd;
2097 cfs_read_lock(&obd->obd_pool_lock);
2098 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2099 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2100 cfs_read_unlock(&obd->obd_pool_lock);
2105 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2107 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2108 DEBUG_REQ(D_ERROR, req, "dropping reply");
2113 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2114 req->rq_status = rc;
2115 return (ptlrpc_send_error(req, 1));
2117 DEBUG_REQ(D_NET, req, "sending reply");
2120 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2123 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2126 struct ptlrpc_reply_state *rs;
2127 struct obd_device *obd;
2128 struct obd_export *exp;
2129 struct ptlrpc_service *svc;
2132 if (req->rq_no_reply) {
2137 svc = req->rq_rqbd->rqbd_service;
2138 rs = req->rq_reply_state;
2139 if (rs == NULL || !rs->rs_difficult) {
2141 target_send_reply_msg (req, rc, fail_id);
2146 /* must be an export if locks saved */
2147 LASSERT (req->rq_export != NULL);
2148 /* req/reply consistent */
2149 LASSERT (rs->rs_service == svc);
2152 LASSERT (!rs->rs_scheduled);
2153 LASSERT (!rs->rs_scheduled_ever);
2154 LASSERT (!rs->rs_handled);
2155 LASSERT (!rs->rs_on_net);
2156 LASSERT (rs->rs_export == NULL);
2157 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2158 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2160 exp = class_export_get (req->rq_export);
2163 /* disable reply scheduling while I'm setting up */
2164 rs->rs_scheduled = 1;
2166 rs->rs_xid = req->rq_xid;
2167 rs->rs_transno = req->rq_transno;
2168 rs->rs_export = exp;
2169 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2171 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2172 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2173 rs->rs_transno, exp->exp_last_committed);
2174 if (rs->rs_transno > exp->exp_last_committed) {
2175 /* not committed already */
2176 cfs_list_add_tail(&rs->rs_obd_list,
2177 &exp->exp_uncommitted_replies);
2179 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2181 cfs_spin_lock(&exp->exp_lock);
2182 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2183 cfs_spin_unlock(&exp->exp_lock);
2185 netrc = target_send_reply_msg (req, rc, fail_id);
2187 cfs_spin_lock(&svc->srv_rs_lock);
2189 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2192 /* error sending: reply is off the net. Also we need +1
2193 * reply ref until ptlrpc_handle_rs() is done
2194 * with the reply state (if the send was successful, there
2195 * would have been +1 ref for the net, which
2196 * reply_out_callback leaves alone) */
2198 ptlrpc_rs_addref(rs);
2201 cfs_spin_lock(&rs->rs_lock);
2202 if (rs->rs_transno <= exp->exp_last_committed ||
2203 (!rs->rs_on_net && !rs->rs_no_ack) ||
2204 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2205 cfs_list_empty(&rs->rs_obd_list)) {
2206 CDEBUG(D_HA, "Schedule reply immediately\n");
2207 ptlrpc_dispatch_difficult_reply(rs);
2209 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2210 rs->rs_scheduled = 0; /* allow notifier to schedule */
2212 cfs_spin_unlock(&rs->rs_lock);
2213 cfs_spin_unlock(&svc->srv_rs_lock);
2217 int target_handle_ping(struct ptlrpc_request *req)
2219 obd_ping(req->rq_export);
2220 return req_capsule_server_pack(&req->rq_pill);
2223 void target_committed_to_req(struct ptlrpc_request *req)
2225 struct obd_export *exp = req->rq_export;
2227 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2228 lustre_msg_set_last_committed(req->rq_repmsg,
2229 exp->exp_last_committed);
2231 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2232 "%d)", exp->exp_obd->obd_no_transno,
2233 req->rq_repmsg == NULL);
2235 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2236 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2238 EXPORT_SYMBOL(target_committed_to_req);
2240 int target_handle_qc_callback(struct ptlrpc_request *req)
2242 struct obd_quotactl *oqctl;
2243 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2245 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2246 if (oqctl == NULL) {
2247 CERROR("Can't unpack obd_quotactl\n");
2251 cli->cl_qchk_stat = oqctl->qc_stat;
2256 #ifdef HAVE_QUOTA_SUPPORT
2257 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2260 struct obd_device *obd = req->rq_export->exp_obd;
2261 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2262 struct obd_device_target *obt;
2263 struct lustre_quota_ctxt *qctxt;
2264 struct qunit_data *qdata = NULL;
2268 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2271 rc = req_capsule_server_pack(&req->rq_pill);
2273 CERROR("packing reply failed!: rc = %d\n", rc);
2277 LASSERT(req->rq_export);
2279 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2280 if (IS_ERR(qdata)) {
2281 rc = PTR_ERR(qdata);
2282 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2283 req->rq_status = rc;
2287 /* we use the observer */
2288 if (obd_pin_observer(obd, &lov_obd) ||
2289 obd_pin_observer(lov_obd, &master_obd)) {
2290 CERROR("Can't find the observer, it is recovering\n");
2291 req->rq_status = -EAGAIN;
2295 obt = &master_obd->u.obt;
2296 qctxt = &obt->obt_qctxt;
2298 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2299 /* quota_type has not been processed yet, return EAGAIN
2300 * until we know whether or not quotas are supposed to
2302 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2304 req->rq_status = -EAGAIN;
2308 cfs_down_read(&obt->obt_rwsem);
2309 if (qctxt->lqc_lqs_hash == NULL) {
2310 cfs_up_read(&obt->obt_rwsem);
2311 /* quota_type has not been processed yet, return EAGAIN
2312 * until we know whether or not quotas are supposed to
2314 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2316 req->rq_status = -EAGAIN;
2320 LASSERT(qctxt->lqc_handler);
2321 rc = qctxt->lqc_handler(master_obd, qdata,
2322 lustre_msg_get_opc(req->rq_reqmsg));
2323 cfs_up_read(&obt->obt_rwsem);
2324 if (rc && rc != -EDQUOT)
2325 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2326 "dqacq/dqrel failed! (rc:%d)\n", rc);
2327 req->rq_status = rc;
2329 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2331 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2335 /* Block the quota req. b=14840 */
2336 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2341 obd_unpin_observer(lov_obd);
2343 obd_unpin_observer(obd);
2345 rc = ptlrpc_reply(req);
2349 #endif /* !__KERNEL__ */
2351 #endif /* HAVE_QUOTA_SUPPORT */
2353 ldlm_mode_t lck_compat_array[] = {
2354 [LCK_EX] LCK_COMPAT_EX,
2355 [LCK_PW] LCK_COMPAT_PW,
2356 [LCK_PR] LCK_COMPAT_PR,
2357 [LCK_CW] LCK_COMPAT_CW,
2358 [LCK_CR] LCK_COMPAT_CR,
2359 [LCK_NL] LCK_COMPAT_NL,
2360 [LCK_GROUP] LCK_COMPAT_GROUP,
2361 [LCK_COS] LCK_COMPAT_COS,
2365 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2366 * not escape to the user level.
2368 int ldlm_error2errno(ldlm_error_t error)
2376 case ELDLM_LOCK_CHANGED:
2379 case ELDLM_LOCK_ABORTED:
2382 case ELDLM_LOCK_REPLACED:
2385 case ELDLM_NO_LOCK_DATA:
2388 case ELDLM_NAMESPACE_EXISTS:
2391 case ELDLM_BAD_NAMESPACE:
2395 if (((int)error) < 0) /* cast to signed type */
2396 result = error; /* as ldlm_error_t can be unsigned */
2398 CERROR("Invalid DLM result code: %d\n", error);
2404 EXPORT_SYMBOL(ldlm_error2errno);
2407 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2409 ldlm_error_t ldlm_errno2error(int err_no)
2418 error = ELDLM_LOCK_CHANGED;
2421 error = ELDLM_LOCK_ABORTED;
2424 error = ELDLM_LOCK_REPLACED;
2427 error = ELDLM_NO_LOCK_DATA;
2430 error = ELDLM_NAMESPACE_EXISTS;
2433 error = ELDLM_BAD_NAMESPACE;
2440 EXPORT_SYMBOL(ldlm_errno2error);
2442 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2443 void ldlm_dump_export_locks(struct obd_export *exp)
2445 cfs_spin_lock(&exp->exp_locks_list_guard);
2446 if (!cfs_list_empty(&exp->exp_locks_list)) {
2447 struct ldlm_lock *lock;
2449 CERROR("dumping locks for export %p,"
2450 "ignore if the unmount doesn't hang\n", exp);
2451 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2452 ldlm_lock_dump(D_ERROR, lock, 0);
2454 cfs_spin_unlock(&exp->exp_locks_list_guard);