4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_LDLM
40 # include <libcfs/libcfs.h>
42 # include <liblustre.h>
45 #include <obd_class.h>
46 #include <lustre_dlm.h>
47 #include <lustre_net.h>
48 #include <lustre_sec.h>
49 #include "ldlm_internal.h"
51 /* @priority: if non-zero, move the selected to the list head
52 * @create: if zero, only search in existed connections
54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
55 int priority, int create)
57 struct ptlrpc_connection *ptlrpc_conn;
58 struct obd_import_conn *imp_conn = NULL, *item;
62 if (!create && !priority) {
63 CDEBUG(D_HA, "Nothing to do\n");
67 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
69 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
74 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
76 GOTO(out_put, rc = -ENOMEM);
80 cfs_spin_lock(&imp->imp_lock);
81 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
82 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
84 cfs_list_del(&item->oic_item);
85 cfs_list_add(&item->oic_item,
87 item->oic_last_attempt = 0;
89 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
90 imp, imp->imp_obd->obd_name, uuid->uuid,
91 (priority ? ", moved to head" : ""));
92 cfs_spin_unlock(&imp->imp_lock);
93 GOTO(out_free, rc = 0);
98 imp_conn->oic_conn = ptlrpc_conn;
99 imp_conn->oic_uuid = *uuid;
100 imp_conn->oic_last_attempt = 0;
102 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
104 cfs_list_add_tail(&imp_conn->oic_item,
105 &imp->imp_conn_list);
106 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
107 imp, imp->imp_obd->obd_name, uuid->uuid,
108 (priority ? "head" : "tail"));
110 cfs_spin_unlock(&imp->imp_lock);
111 GOTO(out_free, rc = -ENOENT);
114 cfs_spin_unlock(&imp->imp_lock);
118 OBD_FREE(imp_conn, sizeof(*imp_conn));
120 ptlrpc_connection_put(ptlrpc_conn);
124 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
126 return import_set_conn(imp, uuid, 1, 0);
129 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
132 return import_set_conn(imp, uuid, priority, 1);
135 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
137 struct obd_import_conn *imp_conn;
138 struct obd_export *dlmexp;
142 cfs_spin_lock(&imp->imp_lock);
143 if (cfs_list_empty(&imp->imp_conn_list)) {
144 LASSERT(!imp->imp_connection);
148 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
149 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
151 LASSERT(imp_conn->oic_conn);
153 /* is current conn? */
154 if (imp_conn == imp->imp_conn_current) {
155 LASSERT(imp_conn->oic_conn == imp->imp_connection);
157 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
158 imp->imp_state != LUSTRE_IMP_DISCON) {
159 CERROR("can't remove current connection\n");
160 GOTO(out, rc = -EBUSY);
163 ptlrpc_connection_put(imp->imp_connection);
164 imp->imp_connection = NULL;
166 dlmexp = class_conn2export(&imp->imp_dlm_handle);
167 if (dlmexp && dlmexp->exp_connection) {
168 LASSERT(dlmexp->exp_connection ==
170 ptlrpc_connection_put(dlmexp->exp_connection);
171 dlmexp->exp_connection = NULL;
175 cfs_list_del(&imp_conn->oic_item);
176 ptlrpc_connection_put(imp_conn->oic_conn);
177 OBD_FREE(imp_conn, sizeof(*imp_conn));
178 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
179 imp, imp->imp_obd->obd_name, uuid->uuid);
184 cfs_spin_unlock(&imp->imp_lock);
186 CERROR("connection %s not found\n", uuid->uuid);
191 * Find conn uuid by peer nid. @peer is a server nid. This function is used
192 * to find a conn uuid of @imp which can reach @peer.
194 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
195 struct obd_uuid *uuid)
197 struct obd_import_conn *conn;
201 cfs_spin_lock(&imp->imp_lock);
202 cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
203 /* check if conn uuid does have this peer nid */
204 if (class_check_uuid(&conn->oic_uuid, peer)) {
205 *uuid = conn->oic_uuid;
210 cfs_spin_unlock(&imp->imp_lock);
213 EXPORT_SYMBOL(client_import_find_conn);
215 void client_destroy_import(struct obd_import *imp)
217 /* drop security policy instance after all rpc finished/aborted
218 * to let all busy contexts be released. */
219 class_import_get(imp);
220 class_destroy_import(imp);
221 sptlrpc_import_sec_put(imp);
222 class_import_put(imp);
224 EXPORT_SYMBOL(client_destroy_import);
227 * check whether the osc is on MDT or not
230 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
232 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
235 static int osc_on_mdt(char *obdname)
239 ptr = strrchr(obdname, '-');
243 if (strncmp(ptr + 1, "MDT", 3) == 0)
249 /* configure an RPC client OBD device
254 * 3 - inactive-on-startup
256 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
258 struct client_obd *cli = &obddev->u.cli;
259 struct obd_import *imp;
260 struct obd_uuid server_uuid;
261 int rq_portal, rp_portal, connect_op;
262 char *name = obddev->obd_type->typ_name;
263 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
267 /* In a more perfect world, we would hang a ptlrpc_client off of
268 * obd_type and just use the values from there. */
269 if (!strcmp(name, LUSTRE_OSC_NAME)) {
270 rq_portal = OST_REQUEST_PORTAL;
271 rp_portal = OSC_REPLY_PORTAL;
272 connect_op = OST_CONNECT;
273 cli->cl_sp_me = LUSTRE_SP_CLI;
274 cli->cl_sp_to = LUSTRE_SP_OST;
275 ns_type = LDLM_NS_TYPE_OSC;
277 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
278 rq_portal = MDS_REQUEST_PORTAL;
279 rp_portal = MDC_REPLY_PORTAL;
280 connect_op = MDS_CONNECT;
281 cli->cl_sp_me = LUSTRE_SP_CLI;
282 cli->cl_sp_to = LUSTRE_SP_MDT;
283 ns_type = LDLM_NS_TYPE_MDC;
285 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
286 rq_portal = MGS_REQUEST_PORTAL;
287 rp_portal = MGC_REPLY_PORTAL;
288 connect_op = MGS_CONNECT;
289 cli->cl_sp_me = LUSTRE_SP_MGC;
290 cli->cl_sp_to = LUSTRE_SP_MGS;
291 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
292 ns_type = LDLM_NS_TYPE_MGC;
294 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
295 rq_portal = OST_REQUEST_PORTAL;
296 rp_portal = OSC_REPLY_PORTAL;
297 connect_op = OST_CONNECT;
298 cli->cl_sp_me = LUSTRE_SP_CLI;
299 cli->cl_sp_to = LUSTRE_SP_OST;
300 ns_type = LDLM_NS_TYPE_OSC;
303 CERROR("unknown client OBD type \"%s\", can't setup\n",
308 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
309 CERROR("requires a TARGET UUID\n");
313 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
314 CERROR("client UUID must be less than 38 characters\n");
318 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
319 CERROR("setup requires a SERVER UUID\n");
323 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
324 CERROR("target UUID must be less than 38 characters\n");
328 cfs_init_rwsem(&cli->cl_sem);
329 cfs_sema_init(&cli->cl_mgc_sem, 1);
330 cli->cl_conn_count = 0;
331 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
332 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
333 sizeof(server_uuid)));
336 cli->cl_avail_grant = 0;
337 /* FIXME: should limit this for the sum of all cl_dirty_max */
338 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
339 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
340 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
341 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
342 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
343 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
344 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
345 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
346 client_obd_list_lock_init(&cli->cl_loi_list_lock);
347 cfs_atomic_set(&cli->cl_pending_w_pages, 0);
348 cfs_atomic_set(&cli->cl_pending_r_pages, 0);
349 cli->cl_r_in_flight = 0;
350 cli->cl_w_in_flight = 0;
352 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
353 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
354 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
355 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
356 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
357 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
358 cfs_waitq_init(&cli->cl_destroy_waitq);
359 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
360 #ifdef ENABLE_CHECKSUM
361 /* Turn on checksumming by default. */
362 cli->cl_checksum = 1;
364 * The supported checksum types will be worked out at connect time
365 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
368 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
370 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
372 /* This value may be changed at connect time in
373 ptlrpc_connect_interpret. */
374 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
375 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
377 if (!strcmp(name, LUSTRE_MDC_NAME)) {
378 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
379 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
380 cli->cl_max_rpcs_in_flight = 2;
381 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
382 cli->cl_max_rpcs_in_flight = 3;
383 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
384 cli->cl_max_rpcs_in_flight = 4;
386 if (osc_on_mdt(obddev->obd_name))
387 cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
389 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
393 CERROR("ldlm_get_ref failed: %d\n", rc);
397 ptlrpc_init_client(rq_portal, rp_portal, name,
398 &obddev->obd_ldlm_client);
400 imp = class_new_import(obddev);
402 GOTO(err_ldlm, rc = -ENOENT);
403 imp->imp_client = &obddev->obd_ldlm_client;
404 imp->imp_connect_op = connect_op;
405 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
406 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
407 LUSTRE_CFG_BUFLEN(lcfg, 1));
408 class_import_put(imp);
410 rc = client_import_add_conn(imp, &server_uuid, 1);
412 CERROR("can't add initial connection\n");
413 GOTO(err_import, rc);
416 cli->cl_import = imp;
417 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
418 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
419 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
421 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
422 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
423 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
424 name, obddev->obd_name,
425 cli->cl_target_uuid.uuid);
426 cfs_spin_lock(&imp->imp_lock);
427 imp->imp_deactive = 1;
428 cfs_spin_unlock(&imp->imp_lock);
432 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
433 LDLM_NAMESPACE_CLIENT,
434 LDLM_NAMESPACE_GREEDY,
436 if (obddev->obd_namespace == NULL) {
437 CERROR("Unable to create client namespace - %s\n",
439 GOTO(err_import, rc = -ENOMEM);
442 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
447 class_destroy_import(imp);
455 int client_obd_cleanup(struct obd_device *obddev)
459 ldlm_namespace_free_post(obddev->obd_namespace);
460 obddev->obd_namespace = NULL;
462 LASSERT(obddev->u.cli.cl_import == NULL);
468 /* ->o_connect() method for client side (OSC and MDC and MGC) */
469 int client_connect_import(const struct lu_env *env,
470 struct obd_export **exp,
471 struct obd_device *obd, struct obd_uuid *cluuid,
472 struct obd_connect_data *data, void *localdata)
474 struct client_obd *cli = &obd->u.cli;
475 struct obd_import *imp = cli->cl_import;
476 struct obd_connect_data *ocd;
477 struct lustre_handle conn = { 0 };
482 cfs_down_write(&cli->cl_sem);
483 if (cli->cl_conn_count > 0 )
484 GOTO(out_sem, rc = -EALREADY);
486 rc = class_connect(&conn, obd, cluuid);
490 cli->cl_conn_count++;
491 *exp = class_conn2export(&conn);
493 LASSERT(obd->obd_namespace);
495 imp->imp_dlm_handle = conn;
496 rc = ptlrpc_init_import(imp);
500 ocd = &imp->imp_connect_data;
503 imp->imp_connect_flags_orig = data->ocd_connect_flags;
506 rc = ptlrpc_connect_import(imp);
508 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
511 LASSERT((*exp)->exp_connection);
514 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
515 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
516 data->ocd_connect_flags, ocd->ocd_connect_flags);
517 data->ocd_connect_flags = ocd->ocd_connect_flags;
520 ptlrpc_pinger_add_import(imp);
526 cli->cl_conn_count--;
527 class_disconnect(*exp);
531 cfs_up_write(&cli->cl_sem);
536 int client_disconnect_export(struct obd_export *exp)
538 struct obd_device *obd = class_exp2obd(exp);
539 struct client_obd *cli;
540 struct obd_import *imp;
545 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
546 exp, exp ? exp->exp_handle.h_cookie : -1);
551 imp = cli->cl_import;
553 cfs_down_write(&cli->cl_sem);
554 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
557 if (!cli->cl_conn_count) {
558 CERROR("disconnecting disconnected device (%s)\n",
560 GOTO(out_disconnect, rc = -EINVAL);
563 cli->cl_conn_count--;
564 if (cli->cl_conn_count)
565 GOTO(out_disconnect, rc = 0);
567 /* Mark import deactivated now, so we don't try to reconnect if any
568 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
569 * fully deactivate the import, or that would drop all requests. */
570 cfs_spin_lock(&imp->imp_lock);
571 imp->imp_deactive = 1;
572 cfs_spin_unlock(&imp->imp_lock);
574 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
575 * delete it regardless. (It's safe to delete an import that was
577 (void)ptlrpc_pinger_del_import(imp);
579 if (obd->obd_namespace != NULL) {
580 /* obd_force == local only */
581 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
582 obd->obd_force ? LCF_LOCAL : 0, NULL);
583 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
587 * there's no need to hold sem during disconnecting an import,
588 * and actually it may cause deadlock in gss.
590 cfs_up_write(&cli->cl_sem);
591 rc = ptlrpc_disconnect_import(imp, 0);
592 cfs_down_write(&cli->cl_sem);
594 ptlrpc_invalidate_import(imp);
599 /* use server style - class_disconnect should be always called for
601 err = class_disconnect(exp);
605 cfs_up_write(&cli->cl_sem);
610 #ifdef HAVE_SERVER_SUPPORT
611 int server_disconnect_export(struct obd_export *exp)
616 /* Disconnect early so that clients can't keep using export */
617 rc = class_disconnect(exp);
618 /* close import for avoid sending any requests */
619 if (exp->exp_imp_reverse)
620 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
622 if (exp->exp_obd->obd_namespace != NULL)
623 ldlm_cancel_locks_for_export(exp);
625 /* complete all outstanding replies */
626 cfs_spin_lock(&exp->exp_lock);
627 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
628 struct ptlrpc_reply_state *rs =
629 cfs_list_entry(exp->exp_outstanding_replies.next,
630 struct ptlrpc_reply_state, rs_exp_list);
631 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
633 cfs_spin_lock(&svcpt->scp_rep_lock);
635 cfs_list_del_init(&rs->rs_exp_list);
636 cfs_spin_lock(&rs->rs_lock);
637 ptlrpc_schedule_difficult_reply(rs);
638 cfs_spin_unlock(&rs->rs_lock);
640 cfs_spin_unlock(&svcpt->scp_rep_lock);
642 cfs_spin_unlock(&exp->exp_lock);
647 /* --------------------------------------------------------------------------
648 * from old lib/target.c
649 * -------------------------------------------------------------------------- */
651 static int target_handle_reconnect(struct lustre_handle *conn,
652 struct obd_export *exp,
653 struct obd_uuid *cluuid)
657 if (exp->exp_connection && exp->exp_imp_reverse) {
658 struct lustre_handle *hdl;
659 struct obd_device *target;
661 hdl = &exp->exp_imp_reverse->imp_remote_handle;
662 target = exp->exp_obd;
664 /* Might be a re-connect after a partition. */
665 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
666 if (target->obd_recovering) {
667 int timeout = cfs_duration_sec(cfs_time_sub(
669 &target->obd_recovery_timer),
670 cfs_time_current()));
672 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
673 "ing, waiting for %d clients in recov"
674 "ery for %d:%.02d\n", target->obd_name,
675 obd_uuid2str(&exp->exp_client_uuid),
676 obd_export_nid2str(exp),
677 target->obd_max_recoverable_clients,
678 timeout / 60, timeout % 60);
680 LCONSOLE_WARN("%s: Client %s (at %s) "
681 "reconnecting\n", target->obd_name,
682 obd_uuid2str(&exp->exp_client_uuid),
683 obd_export_nid2str(exp));
686 conn->cookie = exp->exp_handle.h_cookie;
687 /* target_handle_connect() treats EALREADY and
688 * -EALREADY differently. EALREADY means we are
689 * doing a valid reconnect from the same client. */
692 LCONSOLE_WARN("%s: already connected client %s (at %s) "
693 "with handle "LPX64". Rejecting client "
694 "with the same UUID trying to reconnect "
695 "with handle "LPX64"\n", target->obd_name,
696 obd_uuid2str(&exp->exp_client_uuid),
697 obd_export_nid2str(exp),
698 hdl->cookie, conn->cookie);
699 memset(conn, 0, sizeof *conn);
700 /* target_handle_connect() treats EALREADY and
701 * -EALREADY differently. -EALREADY is an error
702 * (same UUID, different handle). */
707 conn->cookie = exp->exp_handle.h_cookie;
708 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
709 cluuid->uuid, exp, conn->cookie);
713 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
716 struct obd_export *exp = cb_data;
718 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
719 obd->obd_name, exp->exp_client_uuid.uuid);
721 cfs_spin_lock(&exp->exp_lock);
722 exp->exp_need_sync = 0;
723 cfs_spin_unlock(&exp->exp_lock);
724 class_export_cb_put(exp);
726 EXPORT_SYMBOL(target_client_add_cb);
730 check_and_start_recovery_timer(struct obd_device *obd,
731 struct ptlrpc_request *req, int new_client);
734 check_and_start_recovery_timer(struct obd_device *obd,
735 struct ptlrpc_request *req, int new_client)
740 int target_handle_connect(struct ptlrpc_request *req)
742 struct obd_device *target = NULL, *targref = NULL;
743 struct obd_export *export = NULL;
744 struct obd_import *revimp;
745 struct obd_import *tmp_imp = NULL;
746 struct lustre_handle conn;
747 struct lustre_handle *tmp;
748 struct obd_uuid tgtuuid;
749 struct obd_uuid cluuid;
750 struct obd_uuid remote_uuid;
755 bool mds_conn = false, lw_client = false;
756 struct obd_connect_data *data, *tmpdata;
758 lnet_nid_t *client_nid = NULL;
761 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
763 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
765 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
766 GOTO(out, rc = -EINVAL);
769 obd_str2uuid(&tgtuuid, str);
770 target = class_uuid2obd(&tgtuuid);
772 target = class_name2obd(str);
775 deuuidify(str, NULL, &target_start, &target_len);
776 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available for "
777 "connect (no target)\n", str);
778 GOTO(out, rc = -ENODEV);
781 cfs_spin_lock(&target->obd_dev_lock);
782 if (target->obd_stopping || !target->obd_set_up) {
783 cfs_spin_unlock(&target->obd_dev_lock);
785 deuuidify(str, NULL, &target_start, &target_len);
786 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
787 "from %s (%s)\n", target_len, target_start,
788 libcfs_nid2str(req->rq_peer.nid),
789 (target->obd_stopping ?
790 "stopping" : "not set up"));
791 GOTO(out, rc = -ENODEV);
794 if (target->obd_no_conn) {
795 cfs_spin_unlock(&target->obd_dev_lock);
797 LCONSOLE_WARN("%s: Temporarily refusing client connection "
798 "from %s\n", target->obd_name,
799 libcfs_nid2str(req->rq_peer.nid));
800 GOTO(out, rc = -EAGAIN);
803 /* Make sure the target isn't cleaned up while we're here. Yes,
804 there's still a race between the above check and our incref here.
805 Really, class_uuid2obd should take the ref. */
806 targref = class_incref(target, __FUNCTION__, cfs_current());
808 target->obd_conn_inprogress++;
809 cfs_spin_unlock(&target->obd_dev_lock);
811 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
813 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
814 GOTO(out, rc = -EINVAL);
817 obd_str2uuid(&cluuid, str);
819 /* XXX extract a nettype and format accordingly */
820 switch (sizeof(lnet_nid_t)) {
821 /* NB the casts only avoid compiler warnings */
823 snprintf(remote_uuid.uuid, sizeof remote_uuid,
824 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
827 snprintf(remote_uuid.uuid, sizeof remote_uuid,
828 "NET_%x_UUID", (__u32)req->rq_peer.nid);
834 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
836 GOTO(out, rc = -EPROTO);
840 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
842 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
844 GOTO(out, rc = -EPROTO);
846 rc = req_capsule_server_pack(&req->rq_pill);
850 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
852 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
853 "libclient connection attempt");
854 GOTO(out, rc = -EPROTO);
855 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
856 LUSTRE_VERSION_ALLOWED_OFFSET ||
857 data->ocd_version > LUSTRE_VERSION_CODE +
858 LUSTRE_VERSION_ALLOWED_OFFSET) {
859 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
860 "libclient connection attempt",
861 data->ocd_version < LUSTRE_VERSION_CODE ?
863 OBD_OCD_VERSION_MAJOR(data->ocd_version),
864 OBD_OCD_VERSION_MINOR(data->ocd_version),
865 OBD_OCD_VERSION_PATCH(data->ocd_version),
866 OBD_OCD_VERSION_FIX(data->ocd_version));
867 data = req_capsule_server_sized_get(&req->rq_pill,
869 offsetof(typeof(*data), ocd_version) +
870 sizeof(data->ocd_version));
872 data->ocd_connect_flags = OBD_CONNECT_VERSION;
873 data->ocd_version = LUSTRE_VERSION_CODE;
875 GOTO(out, rc = -EPROTO);
879 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
880 (data->ocd_connect_flags & OBD_CONNECT_MDS))
883 if ((data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0)
886 /* lctl gets a backstage, all-access pass. */
887 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
888 goto dont_check_exports;
890 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
894 /* we've found an export in the hash */
896 cfs_spin_lock(&export->exp_lock);
898 if (export->exp_connecting) { /* bug 9635, et. al. */
899 cfs_spin_unlock(&export->exp_lock);
900 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
901 export->exp_obd->obd_name, export,
902 libcfs_nid2str(req->rq_peer.nid));
903 class_export_put(export);
906 } else if (mds_conn && export->exp_connection) {
907 cfs_spin_unlock(&export->exp_lock);
908 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
909 /* mds reconnected after failover */
910 LCONSOLE_WARN("%s: Received MDS connection from "
911 "%s, removing former export from %s\n",
912 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
913 libcfs_nid2str(export->exp_connection->c_peer.nid));
915 /* new mds connection from the same nid */
916 LCONSOLE_WARN("%s: Received new MDS connection from "
917 "%s, removing former export from same NID\n",
918 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
919 class_fail_export(export);
920 class_export_put(export);
923 } else if (export->exp_connection != NULL &&
924 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
925 (lustre_msg_get_op_flags(req->rq_reqmsg) &
926 MSG_CONNECT_INITIAL)) {
927 cfs_spin_unlock(&export->exp_lock);
928 /* in mds failover we have static uuid but nid can be
930 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
931 "existing nid %s is already connected\n",
932 target->obd_name, cluuid.uuid,
933 libcfs_nid2str(req->rq_peer.nid),
935 export->exp_connection->c_peer.nid));
937 class_export_put(export);
940 export->exp_connecting = 1;
941 cfs_spin_unlock(&export->exp_lock);
942 LASSERT(export->exp_obd == target);
944 rc = target_handle_reconnect(&conn, export, &cluuid);
947 /* If we found an export, we already unlocked. */
950 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
951 } else if (req->rq_export == NULL &&
952 cfs_atomic_read(&export->exp_rpc_count) > 0) {
953 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
954 "still busy with %d references\n",
955 target->obd_name, cluuid.uuid,
956 libcfs_nid2str(req->rq_peer.nid),
957 cfs_atomic_read(&export->exp_refcount));
958 GOTO(out, rc = -EBUSY);
959 } else if (req->rq_export != NULL &&
960 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
961 /* the current connect rpc has increased exp_rpc_count */
962 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
963 "still busy with %d active RPCs\n",
964 target->obd_name, cluuid.uuid,
965 libcfs_nid2str(req->rq_peer.nid),
966 cfs_atomic_read(&export->exp_rpc_count) - 1);
967 cfs_spin_lock(&export->exp_lock);
968 if (req->rq_export->exp_conn_cnt <
969 lustre_msg_get_conn_cnt(req->rq_reqmsg))
970 /* try to abort active requests */
971 req->rq_export->exp_abort_active_req = 1;
972 cfs_spin_unlock(&export->exp_lock);
973 GOTO(out, rc = -EBUSY);
974 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
975 if (!strstr(cluuid.uuid, "mdt"))
976 LCONSOLE_WARN("%s: Rejecting reconnect from the "
977 "known client %s (at %s) because it "
978 "is indicating it is a new client",
979 target->obd_name, cluuid.uuid,
980 libcfs_nid2str(req->rq_peer.nid));
981 GOTO(out, rc = -EALREADY);
983 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
990 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
991 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
992 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
993 export, (long)cfs_time_current_sec(),
994 export ? (long)export->exp_last_request_time : 0);
996 /* If this is the first time a client connects, reset the recovery
997 * timer. Discard lightweight connections which might be local */
998 if (!lw_client && rc == 0 && target->obd_recovering)
999 check_and_start_recovery_timer(target, req, export == NULL);
1001 /* We want to handle EALREADY but *not* -EALREADY from
1002 * target_handle_reconnect(), return reconnection state in a flag */
1003 if (rc == EALREADY) {
1004 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1010 /* Tell the client if we support replayable requests */
1011 if (target->obd_replayable)
1012 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1013 client_nid = &req->rq_peer.nid;
1015 if (export == NULL) {
1016 /* allow lightweight connections during recovery */
1017 if (target->obd_recovering && !lw_client) {
1019 int c; /* connected */
1020 int i; /* in progress */
1023 c = cfs_atomic_read(&target->obd_connected_clients);
1024 i = cfs_atomic_read(&target->obd_lock_replay_clients);
1025 k = target->obd_max_recoverable_clients;
1026 t = cfs_timer_deadline(&target->obd_recovery_timer);
1027 t = cfs_time_sub(t, cfs_time_current());
1028 t = cfs_duration_sec(t);
1029 LCONSOLE_WARN("%s: Denying connection for new client "
1030 "%s (at %s), waiting for all %d known "
1031 "clients (%d recovered, %d in progress, "
1032 "and %d unseen) to recover in %d:%.02d\n",
1033 target->obd_name, cluuid.uuid,
1034 libcfs_nid2str(req->rq_peer.nid), k,
1035 c - i, i, k - c, (int)t / 60,
1040 rc = obd_connect(req->rq_svc_thread->t_env,
1041 &export, target, &cluuid, data,
1044 conn.cookie = export->exp_handle.h_cookie;
1047 rc = obd_reconnect(req->rq_svc_thread->t_env,
1048 export, target, &cluuid, data, client_nid);
1053 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
1054 /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
1055 * will do the swabbing for if the client is using the same endianness.
1057 * This fixup is version-limited, because we don't want to carry the
1058 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
1059 * interop with unpatched 2.2 clients. For newer clients, servers
1060 * will never do MNE swabbing, let the client handle that. LU-1644 */
1061 export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) &&
1062 !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB);
1064 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
1067 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1068 data->ocd_instance = target->u.obt.obt_instance;
1070 /* Return only the parts of obd_connect_data that we understand, so the
1071 * client knows that we don't understand the rest. */
1073 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1075 tmpdata = req_capsule_server_get(&req->rq_pill,
1077 /* Don't use struct assignment here, because the client reply
1078 * buffer may be smaller/larger than the local struct
1079 * obd_connect_data. */
1080 memcpy(tmpdata, data, min(tmpsize, size));
1083 /* If all else goes well, this is our RPC return code. */
1086 lustre_msg_set_handle(req->rq_repmsg, &conn);
1088 /* If the client and the server are the same node, we will already
1089 * have an export that really points to the client's DLM export,
1090 * because we have a shared handles table.
1092 * XXX this will go away when shaver stops sending the "connect" handle
1093 * in the real "remote handle" field of the request --phik 24 Apr 2003
1095 if (req->rq_export != NULL)
1096 class_export_put(req->rq_export);
1098 /* request takes one export refcount */
1099 req->rq_export = class_export_get(export);
1101 cfs_spin_lock(&export->exp_lock);
1102 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1103 cfs_spin_unlock(&export->exp_lock);
1104 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1105 "conn_cnt: %d > %d\n",
1106 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1107 export->exp_conn_cnt,
1108 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1110 GOTO(out, rc = -EALREADY);
1112 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1113 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1114 export->exp_abort_active_req = 0;
1116 /* request from liblustre? Don't evict it for not pinging. */
1117 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1118 export->exp_libclient = 1;
1119 cfs_spin_unlock(&export->exp_lock);
1121 cfs_spin_lock(&target->obd_dev_lock);
1122 cfs_list_del_init(&export->exp_obd_chain_timed);
1123 cfs_spin_unlock(&target->obd_dev_lock);
1125 cfs_spin_unlock(&export->exp_lock);
1128 if (export->exp_connection != NULL) {
1129 /* Check to see if connection came from another NID */
1130 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1131 !cfs_hlist_unhashed(&export->exp_nid_hash))
1132 cfs_hash_del(export->exp_obd->obd_nid_hash,
1133 &export->exp_connection->c_peer.nid,
1134 &export->exp_nid_hash);
1136 ptlrpc_connection_put(export->exp_connection);
1139 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1142 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1143 cfs_hash_add(export->exp_obd->obd_nid_hash,
1144 &export->exp_connection->c_peer.nid,
1145 &export->exp_nid_hash);
1148 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1150 __u64 transno = data->ocd_transno;
1152 cfs_spin_lock(&export->exp_lock);
1153 /* possible race with class_disconnect_stale_exports,
1154 * export may be already in the eviction process */
1155 if (export->exp_failed) {
1156 cfs_spin_unlock(&export->exp_lock);
1157 GOTO(out, rc = -ENODEV);
1159 export->exp_in_recovery = 1;
1160 export->exp_req_replay_needed = 1;
1161 export->exp_lock_replay_needed = 1;
1162 cfs_spin_unlock(&export->exp_lock);
1164 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1165 MSG_CONNECT_TRANSNO);
1166 if (has_transno && transno == 0)
1167 CWARN("Connect with zero transno!\n");
1169 if (has_transno && transno > 0 &&
1170 transno < target->obd_next_recovery_transno &&
1171 transno > target->obd_last_committed) {
1172 /* another way is to use cmpxchg() so it will be
1174 cfs_spin_lock(&target->obd_recovery_task_lock);
1175 if (transno < target->obd_next_recovery_transno)
1176 target->obd_next_recovery_transno = transno;
1177 cfs_spin_unlock(&target->obd_recovery_task_lock);
1180 cfs_atomic_inc(&target->obd_req_replay_clients);
1181 cfs_atomic_inc(&target->obd_lock_replay_clients);
1182 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1183 target->obd_max_recoverable_clients)
1184 cfs_waitq_signal(&target->obd_next_transno_waitq);
1187 /* Tell the client we're in recovery, when client is involved in it. */
1188 if (target->obd_recovering && !lw_client)
1189 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1191 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1194 /* for the rest part, we return -ENOTCONN in case of errors
1195 * in order to let client initialize connection again.
1197 revimp = class_new_import(target);
1198 if (revimp == NULL) {
1199 CERROR("fail to alloc new reverse import.\n");
1200 GOTO(out, rc = -ENOTCONN);
1203 cfs_spin_lock(&export->exp_lock);
1204 if (export->exp_imp_reverse != NULL) {
1205 /* destroyed import can be still referenced in ctxt */
1206 obd_set_info_async(req->rq_svc_thread->t_env, export,
1207 sizeof(KEY_REVIMP_UPD), KEY_REVIMP_UPD,
1209 tmp_imp = export->exp_imp_reverse;
1211 export->exp_imp_reverse = revimp;
1212 cfs_spin_unlock(&export->exp_lock);
1214 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1215 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1216 revimp->imp_remote_handle = conn;
1217 revimp->imp_dlm_fake = 1;
1218 revimp->imp_state = LUSTRE_IMP_FULL;
1220 /* unknown versions will be caught in
1221 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1222 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1224 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1225 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1226 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1228 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1230 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1231 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1232 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1234 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1236 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1238 CERROR("Failed to get sec for reverse import: %d\n", rc);
1239 cfs_spin_lock(&export->exp_lock);
1240 export->exp_imp_reverse = NULL;
1241 cfs_spin_unlock(&export->exp_lock);
1242 class_destroy_import(revimp);
1245 class_import_put(revimp);
1248 if (tmp_imp != NULL)
1249 client_destroy_import(tmp_imp);
1251 cfs_spin_lock(&export->exp_lock);
1252 export->exp_connecting = 0;
1253 cfs_spin_unlock(&export->exp_lock);
1255 class_export_put(export);
1258 cfs_spin_lock(&target->obd_dev_lock);
1259 target->obd_conn_inprogress--;
1260 cfs_spin_unlock(&target->obd_dev_lock);
1262 class_decref(targref, __FUNCTION__, cfs_current());
1265 req->rq_status = rc;
1269 int target_handle_disconnect(struct ptlrpc_request *req)
1274 rc = req_capsule_server_pack(&req->rq_pill);
1278 /* keep the rq_export around so we can send the reply */
1279 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1284 void target_destroy_export(struct obd_export *exp)
1286 struct obd_import *imp = NULL;
1287 /* exports created from last_rcvd data, and "fake"
1288 exports created by lctl don't have an import */
1289 cfs_spin_lock(&exp->exp_lock);
1290 if (exp->exp_imp_reverse != NULL) {
1291 imp = exp->exp_imp_reverse;
1292 exp->exp_imp_reverse = NULL;
1294 cfs_spin_unlock(&exp->exp_lock);
1296 client_destroy_import(imp);
1298 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1299 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1300 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1301 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1305 * Recovery functions
1307 static void target_request_copy_get(struct ptlrpc_request *req)
1309 class_export_rpc_get(req->rq_export);
1310 LASSERT(cfs_list_empty(&req->rq_list));
1311 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1313 /* increase refcount to keep request in queue */
1314 cfs_atomic_inc(&req->rq_refcount);
1315 /** let export know it has replays to be handled */
1316 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1319 static void target_request_copy_put(struct ptlrpc_request *req)
1321 LASSERT(cfs_list_empty(&req->rq_replay_list));
1322 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1324 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1325 class_export_rpc_put(req->rq_export);
1326 ptlrpc_server_drop_request(req);
1329 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1331 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1332 struct obd_export *exp = req->rq_export;
1333 struct ptlrpc_request *reqiter;
1338 cfs_spin_lock(&exp->exp_lock);
1339 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1341 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1348 /* we expect it with RESENT and REPLAY flags */
1349 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1350 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1351 CERROR("invalid flags %x of resent replay\n",
1352 lustre_msg_get_flags(req->rq_reqmsg));
1354 cfs_list_add_tail(&req->rq_replay_list,
1355 &exp->exp_req_replay_queue);
1358 cfs_spin_unlock(&exp->exp_lock);
1362 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1364 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1365 LASSERT(req->rq_export);
1367 cfs_spin_lock(&req->rq_export->exp_lock);
1368 cfs_list_del_init(&req->rq_replay_list);
1369 cfs_spin_unlock(&req->rq_export->exp_lock);
1373 static void target_finish_recovery(struct obd_device *obd)
1375 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1376 obd->obd_recovery_start);
1379 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1380 "%d recovered and %d %s evicted.\n", obd->obd_name,
1381 (int)elapsed_time / 60, (int)elapsed_time % 60,
1382 obd->obd_max_recoverable_clients,
1383 cfs_atomic_read(&obd->obd_connected_clients),
1384 obd->obd_stale_clients,
1385 obd->obd_stale_clients == 1 ? "was" : "were");
1387 ldlm_reprocess_all_ns(obd->obd_namespace);
1388 cfs_spin_lock(&obd->obd_recovery_task_lock);
1389 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1390 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1391 !cfs_list_empty(&obd->obd_final_req_queue)) {
1392 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1394 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1395 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1397 cfs_list_empty(&obd->obd_final_req_queue) ? \
1399 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1402 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1404 obd->obd_recovery_end = cfs_time_current_sec();
1406 /* when recovery finished, cleanup orphans on mds and ost */
1407 if (OBT(obd) && OBP(obd, postrecov)) {
1408 int rc = OBP(obd, postrecov)(obd);
1410 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1416 static void abort_req_replay_queue(struct obd_device *obd)
1418 struct ptlrpc_request *req, *n;
1419 cfs_list_t abort_list;
1421 CFS_INIT_LIST_HEAD(&abort_list);
1422 cfs_spin_lock(&obd->obd_recovery_task_lock);
1423 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1424 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1425 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1426 DEBUG_REQ(D_WARNING, req, "aborted:");
1427 req->rq_status = -ENOTCONN;
1428 if (ptlrpc_error(req)) {
1429 DEBUG_REQ(D_ERROR, req,
1430 "failed abort_req_reply; skipping");
1432 target_exp_dequeue_req_replay(req);
1433 target_request_copy_put(req);
1437 static void abort_lock_replay_queue(struct obd_device *obd)
1439 struct ptlrpc_request *req, *n;
1440 cfs_list_t abort_list;
1442 CFS_INIT_LIST_HEAD(&abort_list);
1443 cfs_spin_lock(&obd->obd_recovery_task_lock);
1444 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1445 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1446 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1447 DEBUG_REQ(D_ERROR, req, "aborted:");
1448 req->rq_status = -ENOTCONN;
1449 if (ptlrpc_error(req)) {
1450 DEBUG_REQ(D_ERROR, req,
1451 "failed abort_lock_reply; skipping");
1453 target_request_copy_put(req);
1457 /* Called from a cleanup function if the device is being cleaned up
1458 forcefully. The exports should all have been disconnected already,
1459 the only thing left to do is
1460 - clear the recovery flags
1462 - free queued requests and replies, but don't send replies
1463 Because the obd_stopping flag is set, no new requests should be received.
1466 void target_cleanup_recovery(struct obd_device *obd)
1468 struct ptlrpc_request *req, *n;
1469 cfs_list_t clean_list;
1472 CFS_INIT_LIST_HEAD(&clean_list);
1473 cfs_spin_lock(&obd->obd_dev_lock);
1474 if (!obd->obd_recovering) {
1475 cfs_spin_unlock(&obd->obd_dev_lock);
1479 obd->obd_recovering = obd->obd_abort_recovery = 0;
1480 cfs_spin_unlock(&obd->obd_dev_lock);
1482 cfs_spin_lock(&obd->obd_recovery_task_lock);
1483 target_cancel_recovery_timer(obd);
1484 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1485 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1487 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1488 LASSERT(req->rq_reply_state == 0);
1489 target_exp_dequeue_req_replay(req);
1490 target_request_copy_put(req);
1493 cfs_spin_lock(&obd->obd_recovery_task_lock);
1494 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1495 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1496 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1498 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1499 LASSERT(req->rq_reply_state == 0);
1500 target_request_copy_put(req);
1506 /* obd_recovery_task_lock should be held */
1507 void target_cancel_recovery_timer(struct obd_device *obd)
1509 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1510 cfs_timer_disarm(&obd->obd_recovery_timer);
1513 static void target_start_recovery_timer(struct obd_device *obd)
1515 if (obd->obd_recovery_start != 0)
1518 cfs_spin_lock(&obd->obd_dev_lock);
1519 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1520 cfs_spin_unlock(&obd->obd_dev_lock);
1524 LASSERT(obd->obd_recovery_timeout != 0);
1526 if (obd->obd_recovery_start != 0) {
1527 cfs_spin_unlock(&obd->obd_dev_lock);
1531 cfs_timer_arm(&obd->obd_recovery_timer,
1532 cfs_time_shift(obd->obd_recovery_timeout));
1533 obd->obd_recovery_start = cfs_time_current_sec();
1534 cfs_spin_unlock(&obd->obd_dev_lock);
1536 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1537 "or until %d client%s reconnect%s\n",
1539 obd->obd_recovery_timeout / 60,
1540 obd->obd_recovery_timeout % 60,
1541 obd->obd_max_recoverable_clients,
1542 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1543 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1547 * extend recovery window.
1549 * if @extend is true, extend recovery window to have @drt remaining at least;
1550 * otherwise, make sure the recovery timeout value is not less than @drt.
1552 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1556 cfs_duration_t left;
1559 cfs_spin_lock(&obd->obd_dev_lock);
1560 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1561 cfs_spin_unlock(&obd->obd_dev_lock);
1564 LASSERT(obd->obd_recovery_start != 0);
1566 now = cfs_time_current_sec();
1567 to = obd->obd_recovery_timeout;
1568 end = obd->obd_recovery_start + to;
1569 left = cfs_time_sub(end, now);
1571 if (extend && (drt > left)) {
1573 } else if (!extend && (drt > to)) {
1575 /* reduce drt by already passed time */
1576 drt -= obd->obd_recovery_timeout - left;
1579 if (to > obd->obd_recovery_time_hard)
1580 to = obd->obd_recovery_time_hard;
1581 if (obd->obd_recovery_timeout < to) {
1582 obd->obd_recovery_timeout = to;
1583 cfs_timer_arm(&obd->obd_recovery_timer,
1584 cfs_time_shift(drt));
1586 cfs_spin_unlock(&obd->obd_dev_lock);
1588 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1589 obd->obd_name, (unsigned)drt);
1592 /* Reset the timer with each new client connection */
1594 * This timer is actually reconnect_timer, which is for making sure
1595 * the total recovery window is at least as big as my reconnect
1596 * attempt timing. So the initial recovery time_out will be set to
1597 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1598 * from client is bigger than this, then the recovery time_out will
1599 * be extended to make sure the client could be reconnected, in the
1600 * process, the timeout from the new client should be ignored.
1604 check_and_start_recovery_timer(struct obd_device *obd,
1605 struct ptlrpc_request *req,
1608 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1609 struct obd_device_target *obt = &obd->u.obt;
1610 struct lustre_sb_info *lsi;
1612 if (!new_client && service_time)
1613 /* Teach server about old server's estimates, as first guess
1614 * at how long new requests will take. */
1615 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1618 target_start_recovery_timer(obd);
1620 /* convert the service time to rpc timeout,
1621 * reuse service_time to limit stack usage */
1622 service_time = at_est2timeout(service_time);
1624 /* We expect other clients to timeout within service_time, then try
1625 * to reconnect, then try the failover server. The max delay between
1626 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1627 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1629 LASSERT(obt->obt_magic == OBT_MAGIC);
1630 lsi = s2lsi(obt->obt_sb);
1631 if (!(lsi->lsi_flags | LDD_F_IR_CAPABLE))
1632 service_time += 2 * (CONNECTION_SWITCH_MAX +
1633 CONNECTION_SWITCH_INC);
1634 if (service_time > obd->obd_recovery_timeout && !new_client)
1635 extend_recovery_timer(obd, service_time, false);
1638 /** Health checking routines */
1639 static inline int exp_connect_healthy(struct obd_export *exp)
1641 return (exp->exp_in_recovery);
1644 /** if export done req_replay or has replay in queue */
1645 static inline int exp_req_replay_healthy(struct obd_export *exp)
1647 return (!exp->exp_req_replay_needed ||
1648 cfs_atomic_read(&exp->exp_replay_count) > 0);
1650 /** if export done lock_replay or has replay in queue */
1651 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1653 return (!exp->exp_lock_replay_needed ||
1654 cfs_atomic_read(&exp->exp_replay_count) > 0);
1657 static inline int exp_vbr_healthy(struct obd_export *exp)
1659 return (!exp->exp_vbr_failed);
1662 static inline int exp_finished(struct obd_export *exp)
1664 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1667 /** Checking routines for recovery */
1668 static int check_for_clients(struct obd_device *obd)
1670 unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1672 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1674 LASSERT(clnts <= obd->obd_max_recoverable_clients);
1675 if (obd->obd_no_conn == 0 &&
1676 clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1681 static int check_for_next_transno(struct obd_device *obd)
1683 struct ptlrpc_request *req = NULL;
1684 int wake_up = 0, connected, completed, queue_len;
1685 __u64 next_transno, req_transno;
1688 cfs_spin_lock(&obd->obd_recovery_task_lock);
1689 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1690 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1691 struct ptlrpc_request, rq_list);
1692 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1697 connected = cfs_atomic_read(&obd->obd_connected_clients);
1698 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1699 queue_len = obd->obd_requests_queued_for_recovery;
1700 next_transno = obd->obd_next_recovery_transno;
1702 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1703 "req_transno: "LPU64", next_transno: "LPU64"\n",
1704 obd->obd_max_recoverable_clients, connected, completed,
1705 queue_len, req_transno, next_transno);
1707 if (obd->obd_abort_recovery) {
1708 CDEBUG(D_HA, "waking for aborted recovery\n");
1710 } else if (obd->obd_recovery_expired) {
1711 CDEBUG(D_HA, "waking for expired recovery\n");
1713 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1714 CDEBUG(D_HA, "waking for completed recovery\n");
1716 } else if (req_transno == next_transno) {
1717 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1719 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1721 /** handle gaps occured due to lost reply or VBR */
1722 LASSERTF(req_transno >= next_transno,
1723 "req_transno: "LPU64", next_transno: "LPU64"\n",
1724 req_transno, next_transno);
1725 if (req_transno > obd->obd_last_committed &&
1726 !obd->obd_version_recov)
1729 "%s: waking for gap in transno, VBR is %s (skip: "
1730 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1731 ", last_committed: "LPD64")\n",
1732 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1733 next_transno, queue_len, completed, connected,
1734 req_transno, obd->obd_last_committed);
1735 obd->obd_next_recovery_transno = req_transno;
1737 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1738 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1739 " by fail_lock, waking up ("LPD64")\n", next_transno);
1740 obd->obd_next_recovery_transno = req_transno;
1743 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1747 static int check_for_next_lock(struct obd_device *obd)
1751 cfs_spin_lock(&obd->obd_recovery_task_lock);
1752 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1753 CDEBUG(D_HA, "waking for next lock\n");
1755 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1756 CDEBUG(D_HA, "waking for completed lock replay\n");
1758 } else if (obd->obd_abort_recovery) {
1759 CDEBUG(D_HA, "waking for aborted recovery\n");
1761 } else if (obd->obd_recovery_expired) {
1762 CDEBUG(D_HA, "waking for expired recovery\n");
1765 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1771 * wait for recovery events,
1772 * check its status with help of check_routine
1773 * evict dead clients via health_check
1775 static int target_recovery_overseer(struct obd_device *obd,
1776 int (*check_routine)(struct obd_device *),
1777 int (*health_check)(struct obd_export *))
1780 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1781 if (obd->obd_abort_recovery) {
1782 CWARN("recovery is aborted, evict exports in recovery\n");
1783 /** evict exports which didn't finish recovery yet */
1784 class_disconnect_stale_exports(obd, exp_finished);
1786 } else if (obd->obd_recovery_expired) {
1787 obd->obd_recovery_expired = 0;
1788 /** If some clients died being recovered, evict them */
1789 LCONSOLE_WARN("%s: recovery is timed out, "
1790 "evict stale exports\n", obd->obd_name);
1791 /** evict cexports with no replay in queue, they are stalled */
1792 class_disconnect_stale_exports(obd, health_check);
1793 /** continue with VBR */
1794 cfs_spin_lock(&obd->obd_dev_lock);
1795 obd->obd_version_recov = 1;
1796 cfs_spin_unlock(&obd->obd_dev_lock);
1798 * reset timer, recovery will proceed with versions now,
1799 * timeout is set just to handle reconnection delays
1801 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1802 /** Wait for recovery events again, after evicting bad clients */
1808 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1810 struct ptlrpc_request *req = NULL;
1813 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1814 obd->obd_next_recovery_transno);
1816 if (target_recovery_overseer(obd, check_for_next_transno,
1817 exp_req_replay_healthy)) {
1818 abort_req_replay_queue(obd);
1819 abort_lock_replay_queue(obd);
1822 cfs_spin_lock(&obd->obd_recovery_task_lock);
1823 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1824 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1825 struct ptlrpc_request, rq_list);
1826 cfs_list_del_init(&req->rq_list);
1827 obd->obd_requests_queued_for_recovery--;
1828 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1830 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1831 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1832 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1833 /** evict exports failed VBR */
1834 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1839 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1841 struct ptlrpc_request *req = NULL;
1843 CDEBUG(D_HA, "Waiting for lock\n");
1844 if (target_recovery_overseer(obd, check_for_next_lock,
1845 exp_lock_replay_healthy))
1846 abort_lock_replay_queue(obd);
1848 cfs_spin_lock(&obd->obd_recovery_task_lock);
1849 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1850 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1851 struct ptlrpc_request, rq_list);
1852 cfs_list_del_init(&req->rq_list);
1853 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1855 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1856 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1857 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1858 /** evict exports failed VBR */
1859 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1864 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1866 struct ptlrpc_request *req = NULL;
1868 cfs_spin_lock(&obd->obd_recovery_task_lock);
1869 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1870 req = cfs_list_entry(obd->obd_final_req_queue.next,
1871 struct ptlrpc_request, rq_list);
1872 cfs_list_del_init(&req->rq_list);
1873 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1874 if (req->rq_export->exp_in_recovery) {
1875 cfs_spin_lock(&req->rq_export->exp_lock);
1876 req->rq_export->exp_in_recovery = 0;
1877 cfs_spin_unlock(&req->rq_export->exp_lock);
1880 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1885 static int handle_recovery_req(struct ptlrpc_thread *thread,
1886 struct ptlrpc_request *req,
1887 svc_handler_t handler)
1893 * export can be evicted during recovery, no need to handle replays for
1894 * it after that, discard such request silently
1896 if (req->rq_export->exp_disconnected)
1897 GOTO(reqcopy_put, rc = 0);
1899 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1901 CERROR("Failure to initialize session: %d\n", rc);
1902 GOTO(reqcopy_put, rc);
1905 req->rq_recov_session.lc_thread = thread;
1906 lu_context_enter(&req->rq_recov_session);
1907 req->rq_svc_thread = thread;
1908 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1910 /* thread context */
1911 lu_context_enter(&thread->t_env->le_ctx);
1913 lu_context_exit(&thread->t_env->le_ctx);
1915 lu_context_exit(&req->rq_recov_session);
1916 lu_context_fini(&req->rq_recov_session);
1917 /* don't reset timer for final stage */
1918 if (!exp_finished(req->rq_export)) {
1919 int to = obd_timeout;
1922 * Add request timeout to the recovery time so next request from
1923 * this client may come in recovery time
1926 struct ptlrpc_service_part *svcpt;
1928 svcpt = req->rq_rqbd->rqbd_svcpt;
1929 /* If the server sent early reply for this request,
1930 * the client will recalculate the timeout according to
1931 * current server estimate service time, so we will
1932 * use the maxium timeout here for waiting the client
1933 * sending the next req */
1934 to = max((int)at_est2timeout(
1935 at_get(&svcpt->scp_at_estimate)),
1936 (int)lustre_msg_get_timeout(req->rq_reqmsg));
1937 /* Add net_latency (see ptlrpc_replay_req) */
1938 to += lustre_msg_get_service_time(req->rq_reqmsg);
1940 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1946 static int target_recovery_thread(void *arg)
1948 struct lu_target *lut = arg;
1949 struct obd_device *obd = lut->lut_obd;
1950 struct ptlrpc_request *req;
1951 struct target_recovery_data *trd = &obd->obd_recovery_data;
1952 unsigned long delta;
1953 unsigned long flags;
1955 struct ptlrpc_thread *thread = NULL;
1959 cfs_daemonize_ctxt("tgt_recov");
1961 SIGNAL_MASK_LOCK(current, flags);
1962 sigfillset(¤t->blocked);
1964 SIGNAL_MASK_UNLOCK(current, flags);
1966 OBD_ALLOC_PTR(thread);
1972 OBD_FREE_PTR(thread);
1976 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1978 OBD_FREE_PTR(thread);
1983 thread->t_env = env;
1984 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1985 env->le_ctx.lc_thread = thread;
1986 thread->t_data = NULL;
1987 thread->t_watchdog = NULL;
1989 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1991 trd->trd_processing_task = cfs_curproc_pid();
1993 cfs_spin_lock(&obd->obd_dev_lock);
1994 obd->obd_recovering = 1;
1995 cfs_spin_unlock(&obd->obd_dev_lock);
1996 cfs_complete(&trd->trd_starting);
1998 /* first of all, we have to know the first transno to replay */
1999 if (target_recovery_overseer(obd, check_for_clients,
2000 exp_connect_healthy)) {
2001 abort_req_replay_queue(obd);
2002 abort_lock_replay_queue(obd);
2005 /* next stage: replay requests */
2007 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2008 cfs_atomic_read(&obd->obd_req_replay_clients),
2009 obd->obd_next_recovery_transno);
2010 while ((req = target_next_replay_req(obd))) {
2011 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2012 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2013 lustre_msg_get_transno(req->rq_reqmsg),
2014 libcfs_nid2str(req->rq_peer.nid));
2015 handle_recovery_req(thread, req,
2016 trd->trd_recovery_handler);
2018 * bz18031: increase next_recovery_transno before
2019 * target_request_copy_put() will drop exp_rpc reference
2021 cfs_spin_lock(&obd->obd_recovery_task_lock);
2022 obd->obd_next_recovery_transno++;
2023 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2024 target_exp_dequeue_req_replay(req);
2025 target_request_copy_put(req);
2026 obd->obd_replayed_requests++;
2030 * The second stage: replay locks
2032 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2033 cfs_atomic_read(&obd->obd_lock_replay_clients));
2034 while ((req = target_next_replay_lock(obd))) {
2035 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2036 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2037 libcfs_nid2str(req->rq_peer.nid));
2038 handle_recovery_req(thread, req,
2039 trd->trd_recovery_handler);
2040 target_request_copy_put(req);
2041 obd->obd_replayed_locks++;
2045 * The third stage: reply on final pings, at this moment all clients
2046 * must have request in final queue
2048 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2049 /** Update server last boot epoch */
2050 lut_boot_epoch_update(lut);
2051 /* We drop recoverying flag to forward all new requests
2052 * to regular mds_handle() since now */
2053 cfs_spin_lock(&obd->obd_dev_lock);
2054 obd->obd_recovering = obd->obd_abort_recovery = 0;
2055 cfs_spin_unlock(&obd->obd_dev_lock);
2056 cfs_spin_lock(&obd->obd_recovery_task_lock);
2057 target_cancel_recovery_timer(obd);
2058 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2059 while ((req = target_next_final_ping(obd))) {
2060 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2061 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2062 libcfs_nid2str(req->rq_peer.nid));
2063 handle_recovery_req(thread, req,
2064 trd->trd_recovery_handler);
2065 target_request_copy_put(req);
2068 delta = (jiffies - delta) / CFS_HZ;
2069 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2070 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2071 if (delta > OBD_RECOVERY_TIME_SOFT) {
2072 CWARN("too long recovery - read logs\n");
2073 libcfs_debug_dumplog();
2076 target_finish_recovery(obd);
2078 lu_context_fini(&env->le_ctx);
2079 trd->trd_processing_task = 0;
2080 cfs_complete(&trd->trd_finishing);
2082 OBD_FREE_PTR(thread);
2087 static int target_start_recovery_thread(struct lu_target *lut,
2088 svc_handler_t handler)
2090 struct obd_device *obd = lut->lut_obd;
2092 struct target_recovery_data *trd = &obd->obd_recovery_data;
2094 memset(trd, 0, sizeof(*trd));
2095 cfs_init_completion(&trd->trd_starting);
2096 cfs_init_completion(&trd->trd_finishing);
2097 trd->trd_recovery_handler = handler;
2099 if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
2100 cfs_wait_for_completion(&trd->trd_starting);
2101 LASSERT(obd->obd_recovering != 0);
2108 void target_stop_recovery_thread(struct obd_device *obd)
2110 if (obd->obd_recovery_data.trd_processing_task > 0) {
2111 struct target_recovery_data *trd = &obd->obd_recovery_data;
2112 /** recovery can be done but postrecovery is not yet */
2113 cfs_spin_lock(&obd->obd_dev_lock);
2114 if (obd->obd_recovering) {
2115 CERROR("%s: Aborting recovery\n", obd->obd_name);
2116 obd->obd_abort_recovery = 1;
2117 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2119 cfs_spin_unlock(&obd->obd_dev_lock);
2120 cfs_wait_for_completion(&trd->trd_finishing);
2124 void target_recovery_fini(struct obd_device *obd)
2126 class_disconnect_exports(obd);
2127 target_stop_recovery_thread(obd);
2128 target_cleanup_recovery(obd);
2130 EXPORT_SYMBOL(target_recovery_fini);
2132 static void target_recovery_expired(unsigned long castmeharder)
2134 struct obd_device *obd = (struct obd_device *)castmeharder;
2135 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2136 " after %lds (%d clients connected)\n",
2137 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2138 cfs_time_current_sec()- obd->obd_recovery_start,
2139 cfs_atomic_read(&obd->obd_connected_clients));
2141 obd->obd_recovery_expired = 1;
2142 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2145 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2147 struct obd_device *obd = lut->lut_obd;
2148 if (obd->obd_max_recoverable_clients == 0) {
2149 /** Update server last boot epoch */
2150 lut_boot_epoch_update(lut);
2154 CWARN("RECOVERY: service %s, %d recoverable clients, "
2155 "last_transno "LPU64"\n", obd->obd_name,
2156 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2157 LASSERT(obd->obd_stopping == 0);
2158 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2159 obd->obd_recovery_start = 0;
2160 obd->obd_recovery_end = 0;
2162 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2163 target_start_recovery_thread(lut, handler);
2165 EXPORT_SYMBOL(target_recovery_init);
2167 #endif /* __KERNEL__ */
2169 static int target_process_req_flags(struct obd_device *obd,
2170 struct ptlrpc_request *req)
2172 struct obd_export *exp = req->rq_export;
2173 LASSERT(exp != NULL);
2174 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2175 /* client declares he's ready to replay locks */
2176 cfs_spin_lock(&exp->exp_lock);
2177 if (exp->exp_req_replay_needed) {
2178 exp->exp_req_replay_needed = 0;
2179 cfs_spin_unlock(&exp->exp_lock);
2181 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2182 cfs_atomic_dec(&obd->obd_req_replay_clients);
2184 cfs_spin_unlock(&exp->exp_lock);
2187 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2188 /* client declares he's ready to complete recovery
2189 * so, we put the request on th final queue */
2190 cfs_spin_lock(&exp->exp_lock);
2191 if (exp->exp_lock_replay_needed) {
2192 exp->exp_lock_replay_needed = 0;
2193 cfs_spin_unlock(&exp->exp_lock);
2195 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2196 cfs_atomic_dec(&obd->obd_lock_replay_clients);
2198 cfs_spin_unlock(&exp->exp_lock);
2204 int target_queue_recovery_request(struct ptlrpc_request *req,
2205 struct obd_device *obd)
2209 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2212 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2213 /* Processing the queue right now, don't re-add. */
2217 target_process_req_flags(obd, req);
2219 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2220 /* client declares he's ready to complete recovery
2221 * so, we put the request on th final queue */
2222 target_request_copy_get(req);
2223 DEBUG_REQ(D_HA, req, "queue final req");
2224 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2225 cfs_spin_lock(&obd->obd_recovery_task_lock);
2226 if (obd->obd_recovering) {
2227 cfs_list_add_tail(&req->rq_list,
2228 &obd->obd_final_req_queue);
2230 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2231 target_request_copy_put(req);
2232 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2234 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2237 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2238 /* client declares he's ready to replay locks */
2239 target_request_copy_get(req);
2240 DEBUG_REQ(D_HA, req, "queue lock replay req");
2241 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2242 cfs_spin_lock(&obd->obd_recovery_task_lock);
2243 LASSERT(obd->obd_recovering);
2244 /* usually due to recovery abort */
2245 if (!req->rq_export->exp_in_recovery) {
2246 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2247 target_request_copy_put(req);
2250 LASSERT(req->rq_export->exp_lock_replay_needed);
2251 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2252 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2256 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2257 * (i.e. buflens etc are in my own byte order), but type-dependent
2258 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2261 CFS_INIT_LIST_HEAD(&req->rq_list);
2262 DEBUG_REQ(D_HA, req, "not queueing");
2266 /* If we're processing the queue, we want don't want to queue this
2269 * Also, if this request has a transno less than the one we're waiting
2270 * for, we should process it now. It could (and currently always will)
2271 * be an open request for a descriptor that was opened some time ago.
2273 * Also, a resent, replayed request that has already been
2274 * handled will pass through here and be processed immediately.
2276 CDEBUG(D_HA, "Next recovery transno: "LPU64
2277 ", current: "LPU64", replaying\n",
2278 obd->obd_next_recovery_transno, transno);
2279 cfs_spin_lock(&obd->obd_recovery_task_lock);
2280 if (transno < obd->obd_next_recovery_transno) {
2281 /* Processing the queue right now, don't re-add. */
2282 LASSERT(cfs_list_empty(&req->rq_list));
2283 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2286 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2288 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2291 target_request_copy_get(req);
2292 if (!req->rq_export->exp_in_recovery) {
2293 target_request_copy_put(req);
2296 LASSERT(req->rq_export->exp_req_replay_needed);
2298 if (target_exp_enqueue_req_replay(req)) {
2299 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2300 target_request_copy_put(req);
2305 cfs_spin_lock(&obd->obd_recovery_task_lock);
2306 LASSERT(obd->obd_recovering);
2307 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2308 struct ptlrpc_request *reqiter =
2309 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2311 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2312 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2317 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2319 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2320 "has been claimed by another client");
2321 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2322 target_exp_dequeue_req_replay(req);
2323 target_request_copy_put(req);
2329 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2331 obd->obd_requests_queued_for_recovery++;
2332 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2333 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2337 int target_handle_ping(struct ptlrpc_request *req)
2339 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2340 return req_capsule_server_pack(&req->rq_pill);
2343 void target_committed_to_req(struct ptlrpc_request *req)
2345 struct obd_export *exp = req->rq_export;
2347 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2348 lustre_msg_set_last_committed(req->rq_repmsg,
2349 exp->exp_last_committed);
2351 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2352 "%d)", exp->exp_obd->obd_no_transno,
2353 req->rq_repmsg == NULL);
2355 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2356 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2358 EXPORT_SYMBOL(target_committed_to_req);
2360 #endif /* HAVE_SERVER_SUPPORT */
2363 * Packs current SLV and Limit into \a req.
2365 int target_pack_pool_reply(struct ptlrpc_request *req)
2367 struct obd_device *obd;
2371 * Check that we still have all structures alive as this may
2372 * be some late rpc in shutdown time.
2374 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2375 !exp_connect_lru_resize(req->rq_export))) {
2376 lustre_msg_set_slv(req->rq_repmsg, 0);
2377 lustre_msg_set_limit(req->rq_repmsg, 0);
2382 * OBD is alive here as export is alive, which we checked above.
2384 obd = req->rq_export->exp_obd;
2386 cfs_read_lock(&obd->obd_pool_lock);
2387 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2388 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2389 cfs_read_unlock(&obd->obd_pool_lock);
2394 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2396 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2397 DEBUG_REQ(D_ERROR, req, "dropping reply");
2402 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2403 req->rq_status = rc;
2404 return (ptlrpc_send_error(req, 1));
2406 DEBUG_REQ(D_NET, req, "sending reply");
2409 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2412 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2414 struct ptlrpc_service_part *svcpt;
2416 struct ptlrpc_reply_state *rs;
2417 struct obd_export *exp;
2420 if (req->rq_no_reply) {
2425 svcpt = req->rq_rqbd->rqbd_svcpt;
2426 rs = req->rq_reply_state;
2427 if (rs == NULL || !rs->rs_difficult) {
2429 target_send_reply_msg (req, rc, fail_id);
2434 /* must be an export if locks saved */
2435 LASSERT (req->rq_export != NULL);
2436 /* req/reply consistent */
2437 LASSERT(rs->rs_svcpt == svcpt);
2440 LASSERT (!rs->rs_scheduled);
2441 LASSERT (!rs->rs_scheduled_ever);
2442 LASSERT (!rs->rs_handled);
2443 LASSERT (!rs->rs_on_net);
2444 LASSERT (rs->rs_export == NULL);
2445 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2446 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2448 exp = class_export_get (req->rq_export);
2450 /* disable reply scheduling while I'm setting up */
2451 rs->rs_scheduled = 1;
2453 rs->rs_xid = req->rq_xid;
2454 rs->rs_transno = req->rq_transno;
2455 rs->rs_export = exp;
2456 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2458 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2459 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2460 rs->rs_transno, exp->exp_last_committed);
2461 if (rs->rs_transno > exp->exp_last_committed) {
2462 /* not committed already */
2463 cfs_list_add_tail(&rs->rs_obd_list,
2464 &exp->exp_uncommitted_replies);
2466 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2468 cfs_spin_lock(&exp->exp_lock);
2469 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2470 cfs_spin_unlock(&exp->exp_lock);
2472 netrc = target_send_reply_msg (req, rc, fail_id);
2474 cfs_spin_lock(&svcpt->scp_rep_lock);
2476 cfs_atomic_inc(&svcpt->scp_nreps_difficult);
2479 /* error sending: reply is off the net. Also we need +1
2480 * reply ref until ptlrpc_handle_rs() is done
2481 * with the reply state (if the send was successful, there
2482 * would have been +1 ref for the net, which
2483 * reply_out_callback leaves alone) */
2485 ptlrpc_rs_addref(rs);
2488 cfs_spin_lock(&rs->rs_lock);
2489 if (rs->rs_transno <= exp->exp_last_committed ||
2490 (!rs->rs_on_net && !rs->rs_no_ack) ||
2491 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2492 cfs_list_empty(&rs->rs_obd_list)) {
2493 CDEBUG(D_HA, "Schedule reply immediately\n");
2494 ptlrpc_dispatch_difficult_reply(rs);
2496 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
2497 rs->rs_scheduled = 0; /* allow notifier to schedule */
2499 cfs_spin_unlock(&rs->rs_lock);
2500 cfs_spin_unlock(&svcpt->scp_rep_lock);
2504 ldlm_mode_t lck_compat_array[] = {
2505 [LCK_EX] LCK_COMPAT_EX,
2506 [LCK_PW] LCK_COMPAT_PW,
2507 [LCK_PR] LCK_COMPAT_PR,
2508 [LCK_CW] LCK_COMPAT_CW,
2509 [LCK_CR] LCK_COMPAT_CR,
2510 [LCK_NL] LCK_COMPAT_NL,
2511 [LCK_GROUP] LCK_COMPAT_GROUP,
2512 [LCK_COS] LCK_COMPAT_COS,
2516 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2517 * not escape to the user level.
2519 int ldlm_error2errno(ldlm_error_t error)
2527 case ELDLM_LOCK_CHANGED:
2530 case ELDLM_LOCK_ABORTED:
2533 case ELDLM_LOCK_REPLACED:
2536 case ELDLM_NO_LOCK_DATA:
2539 case ELDLM_NAMESPACE_EXISTS:
2542 case ELDLM_BAD_NAMESPACE:
2546 if (((int)error) < 0) /* cast to signed type */
2547 result = error; /* as ldlm_error_t can be unsigned */
2549 CERROR("Invalid DLM result code: %d\n", error);
2555 EXPORT_SYMBOL(ldlm_error2errno);
2558 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2560 ldlm_error_t ldlm_errno2error(int err_no)
2569 error = ELDLM_LOCK_CHANGED;
2572 error = ELDLM_LOCK_ABORTED;
2575 error = ELDLM_LOCK_REPLACED;
2578 error = ELDLM_NO_LOCK_DATA;
2581 error = ELDLM_NAMESPACE_EXISTS;
2584 error = ELDLM_BAD_NAMESPACE;
2591 EXPORT_SYMBOL(ldlm_errno2error);
2593 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2594 void ldlm_dump_export_locks(struct obd_export *exp)
2596 cfs_spin_lock(&exp->exp_locks_list_guard);
2597 if (!cfs_list_empty(&exp->exp_locks_list)) {
2598 struct ldlm_lock *lock;
2600 CERROR("dumping locks for export %p,"
2601 "ignore if the unmount doesn't hang\n", exp);
2602 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2603 LDLM_ERROR(lock, "lock:");
2605 cfs_spin_unlock(&exp->exp_locks_list_guard);
2609 #ifdef HAVE_SERVER_SUPPORT
2610 static int target_bulk_timeout(void *data)
2613 /* We don't fail the connection here, because having the export
2614 * killed makes the (vital) call to commitrw very sad.
2619 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2621 return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2624 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2625 struct l_wait_info *lwi)
2627 struct ptlrpc_request *req = desc->bd_req;
2631 /* Check if there is eviction in progress, and if so, wait for
2633 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2634 *lwi = LWI_INTR(NULL, NULL);
2635 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2636 !cfs_atomic_read(&exp->exp_obd->
2637 obd_evict_inprogress),
2641 /* Check if client was evicted or tried to reconnect already */
2642 if (exp->exp_failed || exp->exp_abort_active_req) {
2645 if (desc->bd_type == BULK_PUT_SINK)
2646 rc = sptlrpc_svc_wrap_bulk(req, desc);
2648 rc = ptlrpc_start_bulk_transfer(desc);
2651 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2652 ptlrpc_abort_bulk(desc);
2653 } else if (rc == 0) {
2654 time_t start = cfs_time_current_sec();
2656 long timeoutl = req->rq_deadline - cfs_time_current_sec();
2657 cfs_duration_t timeout = timeoutl <= 0 ?
2658 CFS_TICK : cfs_time_seconds(timeoutl);
2659 *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2660 cfs_time_seconds(1),
2661 target_bulk_timeout,
2663 rc = l_wait_event(desc->bd_waitq,
2664 !ptlrpc_server_bulk_active(desc) ||
2666 exp->exp_abort_active_req,
2668 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2669 /* Wait again if we changed deadline */
2670 } while ((rc == -ETIMEDOUT) &&
2671 (req->rq_deadline > cfs_time_current_sec()));
2673 if (rc == -ETIMEDOUT) {
2674 DEBUG_REQ(D_ERROR, req,
2675 "timeout on bulk %s after %ld%+lds",
2677 req->rq_deadline - start,
2678 cfs_time_current_sec() -
2680 ptlrpc_abort_bulk(desc);
2681 } else if (exp->exp_failed) {
2682 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2685 ptlrpc_abort_bulk(desc);
2686 } else if (exp->exp_abort_active_req) {
2687 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2689 /* we don't reply anyway */
2691 ptlrpc_abort_bulk(desc);
2692 } else if (!desc->bd_success ||
2693 desc->bd_nob_transferred != desc->bd_nob) {
2694 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2696 "truncated" : "network error on",
2698 desc->bd_nob_transferred,
2700 /* XXX should this be a different errno? */
2702 } else if (desc->bd_type == BULK_GET_SINK) {
2703 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2706 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2707 bulk2type(desc), rc);
2712 EXPORT_SYMBOL(target_bulk_io);
2714 #endif /* HAVE_SERVER_SUPPORT */