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;
295 CERROR("unknown client OBD type \"%s\", can't setup\n",
300 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
301 CERROR("requires a TARGET UUID\n");
305 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
306 CERROR("client UUID must be less than 38 characters\n");
310 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
311 CERROR("setup requires a SERVER UUID\n");
315 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
316 CERROR("target UUID must be less than 38 characters\n");
320 cfs_init_rwsem(&cli->cl_sem);
321 cfs_sema_init(&cli->cl_mgc_sem, 1);
322 cli->cl_conn_count = 0;
323 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
324 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
325 sizeof(server_uuid)));
328 cli->cl_avail_grant = 0;
329 /* FIXME: should limit this for the sum of all cl_dirty_max */
330 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
331 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
332 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
333 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
334 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
335 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
336 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
337 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
338 client_obd_list_lock_init(&cli->cl_loi_list_lock);
339 cfs_atomic_set(&cli->cl_pending_w_pages, 0);
340 cfs_atomic_set(&cli->cl_pending_r_pages, 0);
341 cli->cl_r_in_flight = 0;
342 cli->cl_w_in_flight = 0;
344 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
345 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
346 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
347 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
348 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
349 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
350 cfs_waitq_init(&cli->cl_destroy_waitq);
351 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
352 #ifdef ENABLE_CHECKSUM
353 /* Turn on checksumming by default. */
354 cli->cl_checksum = 1;
356 * The supported checksum types will be worked out at connect time
357 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
360 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
362 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
364 /* This value may be changed at connect time in
365 ptlrpc_connect_interpret. */
366 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
367 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
369 if (!strcmp(name, LUSTRE_MDC_NAME)) {
370 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
371 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
372 cli->cl_max_rpcs_in_flight = 2;
373 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
374 cli->cl_max_rpcs_in_flight = 3;
375 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
376 cli->cl_max_rpcs_in_flight = 4;
378 if (osc_on_mdt(obddev->obd_name))
379 cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
381 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
385 CERROR("ldlm_get_ref failed: %d\n", rc);
389 ptlrpc_init_client(rq_portal, rp_portal, name,
390 &obddev->obd_ldlm_client);
392 imp = class_new_import(obddev);
394 GOTO(err_ldlm, rc = -ENOENT);
395 imp->imp_client = &obddev->obd_ldlm_client;
396 imp->imp_connect_op = connect_op;
397 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
398 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
399 LUSTRE_CFG_BUFLEN(lcfg, 1));
400 class_import_put(imp);
402 rc = client_import_add_conn(imp, &server_uuid, 1);
404 CERROR("can't add initial connection\n");
405 GOTO(err_import, rc);
408 cli->cl_import = imp;
409 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
410 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
411 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
413 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
414 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
415 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
416 name, obddev->obd_name,
417 cli->cl_target_uuid.uuid);
418 cfs_spin_lock(&imp->imp_lock);
419 imp->imp_deactive = 1;
420 cfs_spin_unlock(&imp->imp_lock);
424 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
425 LDLM_NAMESPACE_CLIENT,
426 LDLM_NAMESPACE_GREEDY,
428 if (obddev->obd_namespace == NULL) {
429 CERROR("Unable to create client namespace - %s\n",
431 GOTO(err_import, rc = -ENOMEM);
434 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
439 class_destroy_import(imp);
447 int client_obd_cleanup(struct obd_device *obddev)
451 ldlm_namespace_free_post(obddev->obd_namespace);
452 obddev->obd_namespace = NULL;
454 LASSERT(obddev->u.cli.cl_import == NULL);
460 /* ->o_connect() method for client side (OSC and MDC and MGC) */
461 int client_connect_import(const struct lu_env *env,
462 struct obd_export **exp,
463 struct obd_device *obd, struct obd_uuid *cluuid,
464 struct obd_connect_data *data, void *localdata)
466 struct client_obd *cli = &obd->u.cli;
467 struct obd_import *imp = cli->cl_import;
468 struct obd_connect_data *ocd;
469 struct lustre_handle conn = { 0 };
474 cfs_down_write(&cli->cl_sem);
475 if (cli->cl_conn_count > 0 )
476 GOTO(out_sem, rc = -EALREADY);
478 rc = class_connect(&conn, obd, cluuid);
482 cli->cl_conn_count++;
483 *exp = class_conn2export(&conn);
485 LASSERT(obd->obd_namespace);
487 imp->imp_dlm_handle = conn;
488 rc = ptlrpc_init_import(imp);
492 ocd = &imp->imp_connect_data;
495 imp->imp_connect_flags_orig = data->ocd_connect_flags;
498 rc = ptlrpc_connect_import(imp);
500 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
503 LASSERT((*exp)->exp_connection);
506 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
507 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
508 data->ocd_connect_flags, ocd->ocd_connect_flags);
509 data->ocd_connect_flags = ocd->ocd_connect_flags;
512 ptlrpc_pinger_add_import(imp);
518 cli->cl_conn_count--;
519 class_disconnect(*exp);
523 cfs_up_write(&cli->cl_sem);
528 int client_disconnect_export(struct obd_export *exp)
530 struct obd_device *obd = class_exp2obd(exp);
531 struct client_obd *cli;
532 struct obd_import *imp;
537 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
538 exp, exp ? exp->exp_handle.h_cookie : -1);
543 imp = cli->cl_import;
545 cfs_down_write(&cli->cl_sem);
546 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
549 if (!cli->cl_conn_count) {
550 CERROR("disconnecting disconnected device (%s)\n",
552 GOTO(out_disconnect, rc = -EINVAL);
555 cli->cl_conn_count--;
556 if (cli->cl_conn_count)
557 GOTO(out_disconnect, rc = 0);
559 /* Mark import deactivated now, so we don't try to reconnect if any
560 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
561 * fully deactivate the import, or that would drop all requests. */
562 cfs_spin_lock(&imp->imp_lock);
563 imp->imp_deactive = 1;
564 cfs_spin_unlock(&imp->imp_lock);
566 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
567 * delete it regardless. (It's safe to delete an import that was
569 (void)ptlrpc_pinger_del_import(imp);
571 if (obd->obd_namespace != NULL) {
572 /* obd_force == local only */
573 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
574 obd->obd_force ? LCF_LOCAL : 0, NULL);
575 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
579 * there's no need to hold sem during disconnecting an import,
580 * and actually it may cause deadlock in gss.
582 cfs_up_write(&cli->cl_sem);
583 rc = ptlrpc_disconnect_import(imp, 0);
584 cfs_down_write(&cli->cl_sem);
586 ptlrpc_invalidate_import(imp);
591 /* use server style - class_disconnect should be always called for
593 err = class_disconnect(exp);
597 cfs_up_write(&cli->cl_sem);
602 #ifdef HAVE_SERVER_SUPPORT
603 int server_disconnect_export(struct obd_export *exp)
608 /* Disconnect early so that clients can't keep using export */
609 rc = class_disconnect(exp);
610 /* close import for avoid sending any requests */
611 if (exp->exp_imp_reverse)
612 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
614 if (exp->exp_obd->obd_namespace != NULL)
615 ldlm_cancel_locks_for_export(exp);
617 /* complete all outstanding replies */
618 cfs_spin_lock(&exp->exp_lock);
619 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
620 struct ptlrpc_reply_state *rs =
621 cfs_list_entry(exp->exp_outstanding_replies.next,
622 struct ptlrpc_reply_state, rs_exp_list);
623 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
625 cfs_spin_lock(&svcpt->scp_rep_lock);
627 cfs_list_del_init(&rs->rs_exp_list);
628 cfs_spin_lock(&rs->rs_lock);
629 ptlrpc_schedule_difficult_reply(rs);
630 cfs_spin_unlock(&rs->rs_lock);
632 cfs_spin_unlock(&svcpt->scp_rep_lock);
634 cfs_spin_unlock(&exp->exp_lock);
639 /* --------------------------------------------------------------------------
640 * from old lib/target.c
641 * -------------------------------------------------------------------------- */
643 static int target_handle_reconnect(struct lustre_handle *conn,
644 struct obd_export *exp,
645 struct obd_uuid *cluuid)
649 if (exp->exp_connection && exp->exp_imp_reverse) {
650 struct lustre_handle *hdl;
651 struct obd_device *target;
653 hdl = &exp->exp_imp_reverse->imp_remote_handle;
654 target = exp->exp_obd;
656 /* Might be a re-connect after a partition. */
657 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
658 if (target->obd_recovering) {
659 int timeout = cfs_duration_sec(cfs_time_sub(
661 &target->obd_recovery_timer),
662 cfs_time_current()));
664 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
665 "ing, waiting for %d clients in recov"
666 "ery for %d:%.02d\n", target->obd_name,
667 obd_uuid2str(&exp->exp_client_uuid),
668 obd_export_nid2str(exp),
669 target->obd_max_recoverable_clients,
670 timeout / 60, timeout % 60);
672 LCONSOLE_WARN("%s: Client %s (at %s) "
673 "reconnecting\n", target->obd_name,
674 obd_uuid2str(&exp->exp_client_uuid),
675 obd_export_nid2str(exp));
678 conn->cookie = exp->exp_handle.h_cookie;
679 /* target_handle_connect() treats EALREADY and
680 * -EALREADY differently. EALREADY means we are
681 * doing a valid reconnect from the same client. */
684 LCONSOLE_WARN("%s: already connected client %s (at %s) "
685 "with handle "LPX64". Rejecting client "
686 "with the same UUID trying to reconnect "
687 "with handle "LPX64"\n", target->obd_name,
688 obd_uuid2str(&exp->exp_client_uuid),
689 obd_export_nid2str(exp),
690 hdl->cookie, conn->cookie);
691 memset(conn, 0, sizeof *conn);
692 /* target_handle_connect() treats EALREADY and
693 * -EALREADY differently. -EALREADY is an error
694 * (same UUID, different handle). */
699 conn->cookie = exp->exp_handle.h_cookie;
700 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
701 cluuid->uuid, exp, conn->cookie);
705 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
708 struct obd_export *exp = cb_data;
710 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
711 obd->obd_name, exp->exp_client_uuid.uuid);
713 cfs_spin_lock(&exp->exp_lock);
714 exp->exp_need_sync = 0;
715 cfs_spin_unlock(&exp->exp_lock);
716 class_export_cb_put(exp);
718 EXPORT_SYMBOL(target_client_add_cb);
722 check_and_start_recovery_timer(struct obd_device *obd,
723 struct ptlrpc_request *req, int new_client);
726 check_and_start_recovery_timer(struct obd_device *obd,
727 struct ptlrpc_request *req, int new_client)
732 int target_handle_connect(struct ptlrpc_request *req)
734 struct obd_device *target = NULL, *targref = NULL;
735 struct obd_export *export = NULL;
736 struct obd_import *revimp;
737 struct obd_import *tmp_imp = NULL;
738 struct lustre_handle conn;
739 struct lustre_handle *tmp;
740 struct obd_uuid tgtuuid;
741 struct obd_uuid cluuid;
742 struct obd_uuid remote_uuid;
747 bool mds_conn = false, lw_client = false;
748 struct obd_connect_data *data, *tmpdata;
750 lnet_nid_t *client_nid = NULL;
753 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
755 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
757 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
758 GOTO(out, rc = -EINVAL);
761 obd_str2uuid(&tgtuuid, str);
762 target = class_uuid2obd(&tgtuuid);
764 target = class_name2obd(str);
767 deuuidify(str, NULL, &target_start, &target_len);
768 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available for "
769 "connect (no target)\n", str);
770 GOTO(out, rc = -ENODEV);
773 cfs_spin_lock(&target->obd_dev_lock);
774 if (target->obd_stopping || !target->obd_set_up) {
775 cfs_spin_unlock(&target->obd_dev_lock);
777 deuuidify(str, NULL, &target_start, &target_len);
778 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
779 "from %s (%s)\n", target_len, target_start,
780 libcfs_nid2str(req->rq_peer.nid),
781 (target->obd_stopping ?
782 "stopping" : "not set up"));
783 GOTO(out, rc = -ENODEV);
786 if (target->obd_no_conn) {
787 cfs_spin_unlock(&target->obd_dev_lock);
789 LCONSOLE_WARN("%s: Temporarily refusing client connection "
790 "from %s\n", target->obd_name,
791 libcfs_nid2str(req->rq_peer.nid));
792 GOTO(out, rc = -EAGAIN);
795 /* Make sure the target isn't cleaned up while we're here. Yes,
796 there's still a race between the above check and our incref here.
797 Really, class_uuid2obd should take the ref. */
798 targref = class_incref(target, __FUNCTION__, cfs_current());
800 target->obd_conn_inprogress++;
801 cfs_spin_unlock(&target->obd_dev_lock);
803 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
805 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
806 GOTO(out, rc = -EINVAL);
809 obd_str2uuid(&cluuid, str);
811 /* XXX extract a nettype and format accordingly */
812 switch (sizeof(lnet_nid_t)) {
813 /* NB the casts only avoid compiler warnings */
815 snprintf(remote_uuid.uuid, sizeof remote_uuid,
816 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
819 snprintf(remote_uuid.uuid, sizeof remote_uuid,
820 "NET_%x_UUID", (__u32)req->rq_peer.nid);
826 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
828 GOTO(out, rc = -EPROTO);
832 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
834 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
836 GOTO(out, rc = -EPROTO);
838 rc = req_capsule_server_pack(&req->rq_pill);
842 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
844 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
845 "libclient connection attempt");
846 GOTO(out, rc = -EPROTO);
847 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
848 LUSTRE_VERSION_ALLOWED_OFFSET ||
849 data->ocd_version > LUSTRE_VERSION_CODE +
850 LUSTRE_VERSION_ALLOWED_OFFSET) {
851 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
852 "libclient connection attempt",
853 data->ocd_version < LUSTRE_VERSION_CODE ?
855 OBD_OCD_VERSION_MAJOR(data->ocd_version),
856 OBD_OCD_VERSION_MINOR(data->ocd_version),
857 OBD_OCD_VERSION_PATCH(data->ocd_version),
858 OBD_OCD_VERSION_FIX(data->ocd_version));
859 data = req_capsule_server_sized_get(&req->rq_pill,
861 offsetof(typeof(*data), ocd_version) +
862 sizeof(data->ocd_version));
864 data->ocd_connect_flags = OBD_CONNECT_VERSION;
865 data->ocd_version = LUSTRE_VERSION_CODE;
867 GOTO(out, rc = -EPROTO);
871 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
872 (data->ocd_connect_flags & OBD_CONNECT_MDS))
875 if ((data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0)
878 /* lctl gets a backstage, all-access pass. */
879 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
880 goto dont_check_exports;
882 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
886 /* we've found an export in the hash */
888 cfs_spin_lock(&export->exp_lock);
890 if (export->exp_connecting) { /* bug 9635, et. al. */
891 cfs_spin_unlock(&export->exp_lock);
892 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
893 export->exp_obd->obd_name, export,
894 libcfs_nid2str(req->rq_peer.nid));
895 class_export_put(export);
898 } else if (mds_conn && export->exp_connection) {
899 cfs_spin_unlock(&export->exp_lock);
900 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
901 /* mds reconnected after failover */
902 LCONSOLE_WARN("%s: Received MDS connection from "
903 "%s, removing former export from %s\n",
904 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
905 libcfs_nid2str(export->exp_connection->c_peer.nid));
907 /* new mds connection from the same nid */
908 LCONSOLE_WARN("%s: Received new MDS connection from "
909 "%s, removing former export from same NID\n",
910 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
911 class_fail_export(export);
912 class_export_put(export);
915 } else if (export->exp_connection != NULL &&
916 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
917 (lustre_msg_get_op_flags(req->rq_reqmsg) &
918 MSG_CONNECT_INITIAL)) {
919 cfs_spin_unlock(&export->exp_lock);
920 /* in mds failover we have static uuid but nid can be
922 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
923 "existing nid %s is already connected\n",
924 target->obd_name, cluuid.uuid,
925 libcfs_nid2str(req->rq_peer.nid),
927 export->exp_connection->c_peer.nid));
929 class_export_put(export);
932 export->exp_connecting = 1;
933 cfs_spin_unlock(&export->exp_lock);
934 LASSERT(export->exp_obd == target);
936 rc = target_handle_reconnect(&conn, export, &cluuid);
939 /* If we found an export, we already unlocked. */
942 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
943 } else if (req->rq_export == NULL &&
944 cfs_atomic_read(&export->exp_rpc_count) > 0) {
945 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
946 "still busy with %d references\n",
947 target->obd_name, cluuid.uuid,
948 libcfs_nid2str(req->rq_peer.nid),
949 cfs_atomic_read(&export->exp_refcount));
950 GOTO(out, rc = -EBUSY);
951 } else if (req->rq_export != NULL &&
952 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
953 /* the current connect rpc has increased exp_rpc_count */
954 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
955 "still busy with %d active RPCs\n",
956 target->obd_name, cluuid.uuid,
957 libcfs_nid2str(req->rq_peer.nid),
958 cfs_atomic_read(&export->exp_rpc_count) - 1);
959 cfs_spin_lock(&export->exp_lock);
960 if (req->rq_export->exp_conn_cnt <
961 lustre_msg_get_conn_cnt(req->rq_reqmsg))
962 /* try to abort active requests */
963 req->rq_export->exp_abort_active_req = 1;
964 cfs_spin_unlock(&export->exp_lock);
965 GOTO(out, rc = -EBUSY);
966 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
967 if (!strstr(cluuid.uuid, "mdt"))
968 LCONSOLE_WARN("%s: Rejecting reconnect from the "
969 "known client %s (at %s) because it "
970 "is indicating it is a new client",
971 target->obd_name, cluuid.uuid,
972 libcfs_nid2str(req->rq_peer.nid));
973 GOTO(out, rc = -EALREADY);
975 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
982 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
983 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
984 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
985 export, (long)cfs_time_current_sec(),
986 export ? (long)export->exp_last_request_time : 0);
988 /* If this is the first time a client connects, reset the recovery
989 * timer. Discard lightweight connections which might be local */
990 if (!lw_client && rc == 0 && target->obd_recovering)
991 check_and_start_recovery_timer(target, req, export == NULL);
993 /* We want to handle EALREADY but *not* -EALREADY from
994 * target_handle_reconnect(), return reconnection state in a flag */
995 if (rc == EALREADY) {
996 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1002 /* Tell the client if we support replayable requests */
1003 if (target->obd_replayable)
1004 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1005 client_nid = &req->rq_peer.nid;
1007 if (export == NULL) {
1008 /* allow lightweight connections during recovery */
1009 if (target->obd_recovering && !lw_client) {
1011 int c; /* connected */
1012 int i; /* in progress */
1015 c = cfs_atomic_read(&target->obd_connected_clients);
1016 i = cfs_atomic_read(&target->obd_lock_replay_clients);
1017 k = target->obd_max_recoverable_clients;
1018 t = cfs_timer_deadline(&target->obd_recovery_timer);
1019 t = cfs_time_sub(t, cfs_time_current());
1020 t = cfs_duration_sec(t);
1021 LCONSOLE_WARN("%s: Denying connection for new client "
1022 "%s (at %s), waiting for all %d known "
1023 "clients (%d recovered, %d in progress, "
1024 "and %d unseen) to recover in %d:%.02d\n",
1025 target->obd_name, cluuid.uuid,
1026 libcfs_nid2str(req->rq_peer.nid), k,
1027 c - i, i, k - c, (int)t / 60,
1032 rc = obd_connect(req->rq_svc_thread->t_env,
1033 &export, target, &cluuid, data,
1036 conn.cookie = export->exp_handle.h_cookie;
1039 rc = obd_reconnect(req->rq_svc_thread->t_env,
1040 export, target, &cluuid, data, client_nid);
1045 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
1046 /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
1047 * will do the swabbing for if the client is using the same endianness.
1049 * This fixup is version-limited, because we don't want to carry the
1050 * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
1051 * interop with unpatched 2.2 clients. For newer clients, servers
1052 * will never do MNE swabbing, let the client handle that. LU-1644 */
1053 export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) &&
1054 !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB);
1056 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
1059 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1060 data->ocd_instance = target->u.obt.obt_instance;
1062 /* Return only the parts of obd_connect_data that we understand, so the
1063 * client knows that we don't understand the rest. */
1065 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1067 tmpdata = req_capsule_server_get(&req->rq_pill,
1069 /* Don't use struct assignment here, because the client reply
1070 * buffer may be smaller/larger than the local struct
1071 * obd_connect_data. */
1072 memcpy(tmpdata, data, min(tmpsize, size));
1075 /* If all else goes well, this is our RPC return code. */
1078 lustre_msg_set_handle(req->rq_repmsg, &conn);
1080 /* If the client and the server are the same node, we will already
1081 * have an export that really points to the client's DLM export,
1082 * because we have a shared handles table.
1084 * XXX this will go away when shaver stops sending the "connect" handle
1085 * in the real "remote handle" field of the request --phik 24 Apr 2003
1087 if (req->rq_export != NULL)
1088 class_export_put(req->rq_export);
1090 /* request takes one export refcount */
1091 req->rq_export = class_export_get(export);
1093 cfs_spin_lock(&export->exp_lock);
1094 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1095 cfs_spin_unlock(&export->exp_lock);
1096 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1097 "conn_cnt: %d > %d\n",
1098 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1099 export->exp_conn_cnt,
1100 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1102 GOTO(out, rc = -EALREADY);
1104 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1105 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1106 export->exp_abort_active_req = 0;
1108 /* request from liblustre? Don't evict it for not pinging. */
1109 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1110 export->exp_libclient = 1;
1111 cfs_spin_unlock(&export->exp_lock);
1113 cfs_spin_lock(&target->obd_dev_lock);
1114 cfs_list_del_init(&export->exp_obd_chain_timed);
1115 cfs_spin_unlock(&target->obd_dev_lock);
1117 cfs_spin_unlock(&export->exp_lock);
1120 if (export->exp_connection != NULL) {
1121 /* Check to see if connection came from another NID */
1122 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1123 !cfs_hlist_unhashed(&export->exp_nid_hash))
1124 cfs_hash_del(export->exp_obd->obd_nid_hash,
1125 &export->exp_connection->c_peer.nid,
1126 &export->exp_nid_hash);
1128 ptlrpc_connection_put(export->exp_connection);
1131 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1134 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1135 cfs_hash_add(export->exp_obd->obd_nid_hash,
1136 &export->exp_connection->c_peer.nid,
1137 &export->exp_nid_hash);
1140 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1142 __u64 transno = data->ocd_transno;
1144 cfs_spin_lock(&export->exp_lock);
1145 /* possible race with class_disconnect_stale_exports,
1146 * export may be already in the eviction process */
1147 if (export->exp_failed) {
1148 cfs_spin_unlock(&export->exp_lock);
1149 GOTO(out, rc = -ENODEV);
1151 export->exp_in_recovery = 1;
1152 export->exp_req_replay_needed = 1;
1153 export->exp_lock_replay_needed = 1;
1154 cfs_spin_unlock(&export->exp_lock);
1156 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1157 MSG_CONNECT_TRANSNO);
1158 if (has_transno && transno == 0)
1159 CWARN("Connect with zero transno!\n");
1161 if (has_transno && transno > 0 &&
1162 transno < target->obd_next_recovery_transno &&
1163 transno > target->obd_last_committed) {
1164 /* another way is to use cmpxchg() so it will be
1166 cfs_spin_lock(&target->obd_recovery_task_lock);
1167 if (transno < target->obd_next_recovery_transno)
1168 target->obd_next_recovery_transno = transno;
1169 cfs_spin_unlock(&target->obd_recovery_task_lock);
1172 cfs_atomic_inc(&target->obd_req_replay_clients);
1173 cfs_atomic_inc(&target->obd_lock_replay_clients);
1174 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1175 target->obd_max_recoverable_clients)
1176 cfs_waitq_signal(&target->obd_next_transno_waitq);
1179 /* Tell the client we're in recovery, when client is involved in it. */
1180 if (target->obd_recovering && !lw_client)
1181 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1183 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1186 /* for the rest part, we return -ENOTCONN in case of errors
1187 * in order to let client initialize connection again.
1189 revimp = class_new_import(target);
1190 if (revimp == NULL) {
1191 CERROR("fail to alloc new reverse import.\n");
1192 GOTO(out, rc = -ENOTCONN);
1195 cfs_spin_lock(&export->exp_lock);
1196 if (export->exp_imp_reverse != NULL) {
1197 /* destroyed import can be still referenced in ctxt */
1198 obd_set_info_async(req->rq_svc_thread->t_env, export,
1199 sizeof(KEY_REVIMP_UPD), KEY_REVIMP_UPD,
1201 tmp_imp = export->exp_imp_reverse;
1203 export->exp_imp_reverse = revimp;
1204 cfs_spin_unlock(&export->exp_lock);
1206 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1207 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1208 revimp->imp_remote_handle = conn;
1209 revimp->imp_dlm_fake = 1;
1210 revimp->imp_state = LUSTRE_IMP_FULL;
1212 /* unknown versions will be caught in
1213 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1214 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1216 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1217 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1218 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1220 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1222 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1223 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1224 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1226 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1228 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1230 CERROR("Failed to get sec for reverse import: %d\n", rc);
1231 cfs_spin_lock(&export->exp_lock);
1232 export->exp_imp_reverse = NULL;
1233 cfs_spin_unlock(&export->exp_lock);
1234 class_destroy_import(revimp);
1237 class_import_put(revimp);
1240 if (tmp_imp != NULL)
1241 client_destroy_import(tmp_imp);
1243 cfs_spin_lock(&export->exp_lock);
1244 export->exp_connecting = 0;
1245 cfs_spin_unlock(&export->exp_lock);
1247 class_export_put(export);
1250 cfs_spin_lock(&target->obd_dev_lock);
1251 target->obd_conn_inprogress--;
1252 cfs_spin_unlock(&target->obd_dev_lock);
1254 class_decref(targref, __FUNCTION__, cfs_current());
1257 req->rq_status = rc;
1261 int target_handle_disconnect(struct ptlrpc_request *req)
1266 rc = req_capsule_server_pack(&req->rq_pill);
1270 /* keep the rq_export around so we can send the reply */
1271 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1276 void target_destroy_export(struct obd_export *exp)
1278 struct obd_import *imp = NULL;
1279 /* exports created from last_rcvd data, and "fake"
1280 exports created by lctl don't have an import */
1281 cfs_spin_lock(&exp->exp_lock);
1282 if (exp->exp_imp_reverse != NULL) {
1283 imp = exp->exp_imp_reverse;
1284 exp->exp_imp_reverse = NULL;
1286 cfs_spin_unlock(&exp->exp_lock);
1288 client_destroy_import(imp);
1290 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1291 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1292 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1293 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1297 * Recovery functions
1299 static void target_request_copy_get(struct ptlrpc_request *req)
1301 class_export_rpc_get(req->rq_export);
1302 LASSERT(cfs_list_empty(&req->rq_list));
1303 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1305 /* increase refcount to keep request in queue */
1306 cfs_atomic_inc(&req->rq_refcount);
1307 /** let export know it has replays to be handled */
1308 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1311 static void target_request_copy_put(struct ptlrpc_request *req)
1313 LASSERT(cfs_list_empty(&req->rq_replay_list));
1314 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1316 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1317 class_export_rpc_put(req->rq_export);
1318 ptlrpc_server_drop_request(req);
1321 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1323 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1324 struct obd_export *exp = req->rq_export;
1325 struct ptlrpc_request *reqiter;
1330 cfs_spin_lock(&exp->exp_lock);
1331 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1333 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1340 /* we expect it with RESENT and REPLAY flags */
1341 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1342 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1343 CERROR("invalid flags %x of resent replay\n",
1344 lustre_msg_get_flags(req->rq_reqmsg));
1346 cfs_list_add_tail(&req->rq_replay_list,
1347 &exp->exp_req_replay_queue);
1350 cfs_spin_unlock(&exp->exp_lock);
1354 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1356 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1357 LASSERT(req->rq_export);
1359 cfs_spin_lock(&req->rq_export->exp_lock);
1360 cfs_list_del_init(&req->rq_replay_list);
1361 cfs_spin_unlock(&req->rq_export->exp_lock);
1365 static void target_finish_recovery(struct obd_device *obd)
1367 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1368 obd->obd_recovery_start);
1371 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1372 "%d recovered and %d %s evicted.\n", obd->obd_name,
1373 (int)elapsed_time / 60, (int)elapsed_time % 60,
1374 obd->obd_max_recoverable_clients,
1375 cfs_atomic_read(&obd->obd_connected_clients),
1376 obd->obd_stale_clients,
1377 obd->obd_stale_clients == 1 ? "was" : "were");
1379 ldlm_reprocess_all_ns(obd->obd_namespace);
1380 cfs_spin_lock(&obd->obd_recovery_task_lock);
1381 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1382 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1383 !cfs_list_empty(&obd->obd_final_req_queue)) {
1384 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1386 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1387 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1389 cfs_list_empty(&obd->obd_final_req_queue) ? \
1391 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1394 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1396 obd->obd_recovery_end = cfs_time_current_sec();
1398 /* when recovery finished, cleanup orphans on mds and ost */
1399 if (OBT(obd) && OBP(obd, postrecov)) {
1400 int rc = OBP(obd, postrecov)(obd);
1402 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1408 static void abort_req_replay_queue(struct obd_device *obd)
1410 struct ptlrpc_request *req, *n;
1411 cfs_list_t abort_list;
1413 CFS_INIT_LIST_HEAD(&abort_list);
1414 cfs_spin_lock(&obd->obd_recovery_task_lock);
1415 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1416 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1417 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1418 DEBUG_REQ(D_WARNING, req, "aborted:");
1419 req->rq_status = -ENOTCONN;
1420 if (ptlrpc_error(req)) {
1421 DEBUG_REQ(D_ERROR, req,
1422 "failed abort_req_reply; skipping");
1424 target_exp_dequeue_req_replay(req);
1425 target_request_copy_put(req);
1429 static void abort_lock_replay_queue(struct obd_device *obd)
1431 struct ptlrpc_request *req, *n;
1432 cfs_list_t abort_list;
1434 CFS_INIT_LIST_HEAD(&abort_list);
1435 cfs_spin_lock(&obd->obd_recovery_task_lock);
1436 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1437 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1438 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1439 DEBUG_REQ(D_ERROR, req, "aborted:");
1440 req->rq_status = -ENOTCONN;
1441 if (ptlrpc_error(req)) {
1442 DEBUG_REQ(D_ERROR, req,
1443 "failed abort_lock_reply; skipping");
1445 target_request_copy_put(req);
1449 /* Called from a cleanup function if the device is being cleaned up
1450 forcefully. The exports should all have been disconnected already,
1451 the only thing left to do is
1452 - clear the recovery flags
1454 - free queued requests and replies, but don't send replies
1455 Because the obd_stopping flag is set, no new requests should be received.
1458 void target_cleanup_recovery(struct obd_device *obd)
1460 struct ptlrpc_request *req, *n;
1461 cfs_list_t clean_list;
1464 CFS_INIT_LIST_HEAD(&clean_list);
1465 cfs_spin_lock(&obd->obd_dev_lock);
1466 if (!obd->obd_recovering) {
1467 cfs_spin_unlock(&obd->obd_dev_lock);
1471 obd->obd_recovering = obd->obd_abort_recovery = 0;
1472 cfs_spin_unlock(&obd->obd_dev_lock);
1474 cfs_spin_lock(&obd->obd_recovery_task_lock);
1475 target_cancel_recovery_timer(obd);
1476 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1477 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1479 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1480 LASSERT(req->rq_reply_state == 0);
1481 target_exp_dequeue_req_replay(req);
1482 target_request_copy_put(req);
1485 cfs_spin_lock(&obd->obd_recovery_task_lock);
1486 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1487 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1488 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1490 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1491 LASSERT(req->rq_reply_state == 0);
1492 target_request_copy_put(req);
1498 /* obd_recovery_task_lock should be held */
1499 void target_cancel_recovery_timer(struct obd_device *obd)
1501 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1502 cfs_timer_disarm(&obd->obd_recovery_timer);
1505 static void target_start_recovery_timer(struct obd_device *obd)
1507 if (obd->obd_recovery_start != 0)
1510 cfs_spin_lock(&obd->obd_dev_lock);
1511 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1512 cfs_spin_unlock(&obd->obd_dev_lock);
1516 LASSERT(obd->obd_recovery_timeout != 0);
1518 if (obd->obd_recovery_start != 0) {
1519 cfs_spin_unlock(&obd->obd_dev_lock);
1523 cfs_timer_arm(&obd->obd_recovery_timer,
1524 cfs_time_shift(obd->obd_recovery_timeout));
1525 obd->obd_recovery_start = cfs_time_current_sec();
1526 cfs_spin_unlock(&obd->obd_dev_lock);
1528 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1529 "or until %d client%s reconnect%s\n",
1531 obd->obd_recovery_timeout / 60,
1532 obd->obd_recovery_timeout % 60,
1533 obd->obd_max_recoverable_clients,
1534 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1535 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1539 * extend recovery window.
1541 * if @extend is true, extend recovery window to have @drt remaining at least;
1542 * otherwise, make sure the recovery timeout value is not less than @drt.
1544 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1548 cfs_duration_t left;
1551 cfs_spin_lock(&obd->obd_dev_lock);
1552 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1553 cfs_spin_unlock(&obd->obd_dev_lock);
1556 LASSERT(obd->obd_recovery_start != 0);
1558 now = cfs_time_current_sec();
1559 to = obd->obd_recovery_timeout;
1560 end = obd->obd_recovery_start + to;
1561 left = cfs_time_sub(end, now);
1563 if (extend && (drt > left)) {
1565 } else if (!extend && (drt > to)) {
1567 /* reduce drt by already passed time */
1568 drt -= obd->obd_recovery_timeout - left;
1571 if (to > obd->obd_recovery_time_hard)
1572 to = obd->obd_recovery_time_hard;
1573 if (obd->obd_recovery_timeout < to) {
1574 obd->obd_recovery_timeout = to;
1575 cfs_timer_arm(&obd->obd_recovery_timer,
1576 cfs_time_shift(drt));
1578 cfs_spin_unlock(&obd->obd_dev_lock);
1580 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1581 obd->obd_name, (unsigned)drt);
1584 /* Reset the timer with each new client connection */
1586 * This timer is actually reconnect_timer, which is for making sure
1587 * the total recovery window is at least as big as my reconnect
1588 * attempt timing. So the initial recovery time_out will be set to
1589 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1590 * from client is bigger than this, then the recovery time_out will
1591 * be extended to make sure the client could be reconnected, in the
1592 * process, the timeout from the new client should be ignored.
1596 check_and_start_recovery_timer(struct obd_device *obd,
1597 struct ptlrpc_request *req,
1600 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1601 struct obd_device_target *obt = &obd->u.obt;
1602 struct lustre_sb_info *lsi;
1604 if (!new_client && service_time)
1605 /* Teach server about old server's estimates, as first guess
1606 * at how long new requests will take. */
1607 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1610 target_start_recovery_timer(obd);
1612 /* convert the service time to rpc timeout,
1613 * reuse service_time to limit stack usage */
1614 service_time = at_est2timeout(service_time);
1616 /* We expect other clients to timeout within service_time, then try
1617 * to reconnect, then try the failover server. The max delay between
1618 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1619 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1621 LASSERT(obt->obt_magic == OBT_MAGIC);
1622 lsi = s2lsi(obt->obt_sb);
1623 if (!(lsi->lsi_flags | LDD_F_IR_CAPABLE))
1624 service_time += 2 * (CONNECTION_SWITCH_MAX +
1625 CONNECTION_SWITCH_INC);
1626 if (service_time > obd->obd_recovery_timeout && !new_client)
1627 extend_recovery_timer(obd, service_time, false);
1630 /** Health checking routines */
1631 static inline int exp_connect_healthy(struct obd_export *exp)
1633 return (exp->exp_in_recovery);
1636 /** if export done req_replay or has replay in queue */
1637 static inline int exp_req_replay_healthy(struct obd_export *exp)
1639 return (!exp->exp_req_replay_needed ||
1640 cfs_atomic_read(&exp->exp_replay_count) > 0);
1642 /** if export done lock_replay or has replay in queue */
1643 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1645 return (!exp->exp_lock_replay_needed ||
1646 cfs_atomic_read(&exp->exp_replay_count) > 0);
1649 static inline int exp_vbr_healthy(struct obd_export *exp)
1651 return (!exp->exp_vbr_failed);
1654 static inline int exp_finished(struct obd_export *exp)
1656 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1659 /** Checking routines for recovery */
1660 static int check_for_clients(struct obd_device *obd)
1662 unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1664 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1666 LASSERT(clnts <= obd->obd_max_recoverable_clients);
1667 if (obd->obd_no_conn == 0 &&
1668 clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1673 static int check_for_next_transno(struct obd_device *obd)
1675 struct ptlrpc_request *req = NULL;
1676 int wake_up = 0, connected, completed, queue_len;
1677 __u64 next_transno, req_transno;
1680 cfs_spin_lock(&obd->obd_recovery_task_lock);
1681 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1682 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1683 struct ptlrpc_request, rq_list);
1684 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1689 connected = cfs_atomic_read(&obd->obd_connected_clients);
1690 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1691 queue_len = obd->obd_requests_queued_for_recovery;
1692 next_transno = obd->obd_next_recovery_transno;
1694 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1695 "req_transno: "LPU64", next_transno: "LPU64"\n",
1696 obd->obd_max_recoverable_clients, connected, completed,
1697 queue_len, req_transno, next_transno);
1699 if (obd->obd_abort_recovery) {
1700 CDEBUG(D_HA, "waking for aborted recovery\n");
1702 } else if (obd->obd_recovery_expired) {
1703 CDEBUG(D_HA, "waking for expired recovery\n");
1705 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1706 CDEBUG(D_HA, "waking for completed recovery\n");
1708 } else if (req_transno == next_transno) {
1709 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1711 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1713 /** handle gaps occured due to lost reply or VBR */
1714 LASSERTF(req_transno >= next_transno,
1715 "req_transno: "LPU64", next_transno: "LPU64"\n",
1716 req_transno, next_transno);
1717 if (req_transno > obd->obd_last_committed &&
1718 !obd->obd_version_recov)
1721 "%s: waking for gap in transno, VBR is %s (skip: "
1722 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1723 ", last_committed: "LPD64")\n",
1724 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1725 next_transno, queue_len, completed, connected,
1726 req_transno, obd->obd_last_committed);
1727 obd->obd_next_recovery_transno = req_transno;
1729 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1730 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1731 " by fail_lock, waking up ("LPD64")\n", next_transno);
1732 obd->obd_next_recovery_transno = req_transno;
1735 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1739 static int check_for_next_lock(struct obd_device *obd)
1743 cfs_spin_lock(&obd->obd_recovery_task_lock);
1744 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1745 CDEBUG(D_HA, "waking for next lock\n");
1747 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1748 CDEBUG(D_HA, "waking for completed lock replay\n");
1750 } else if (obd->obd_abort_recovery) {
1751 CDEBUG(D_HA, "waking for aborted recovery\n");
1753 } else if (obd->obd_recovery_expired) {
1754 CDEBUG(D_HA, "waking for expired recovery\n");
1757 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1763 * wait for recovery events,
1764 * check its status with help of check_routine
1765 * evict dead clients via health_check
1767 static int target_recovery_overseer(struct obd_device *obd,
1768 int (*check_routine)(struct obd_device *),
1769 int (*health_check)(struct obd_export *))
1772 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1773 if (obd->obd_abort_recovery) {
1774 CWARN("recovery is aborted, evict exports in recovery\n");
1775 /** evict exports which didn't finish recovery yet */
1776 class_disconnect_stale_exports(obd, exp_finished);
1778 } else if (obd->obd_recovery_expired) {
1779 obd->obd_recovery_expired = 0;
1780 /** If some clients died being recovered, evict them */
1781 LCONSOLE_WARN("%s: recovery is timed out, "
1782 "evict stale exports\n", obd->obd_name);
1783 /** evict cexports with no replay in queue, they are stalled */
1784 class_disconnect_stale_exports(obd, health_check);
1785 /** continue with VBR */
1786 cfs_spin_lock(&obd->obd_dev_lock);
1787 obd->obd_version_recov = 1;
1788 cfs_spin_unlock(&obd->obd_dev_lock);
1790 * reset timer, recovery will proceed with versions now,
1791 * timeout is set just to handle reconnection delays
1793 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1794 /** Wait for recovery events again, after evicting bad clients */
1800 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1802 struct ptlrpc_request *req = NULL;
1805 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1806 obd->obd_next_recovery_transno);
1808 if (target_recovery_overseer(obd, check_for_next_transno,
1809 exp_req_replay_healthy)) {
1810 abort_req_replay_queue(obd);
1811 abort_lock_replay_queue(obd);
1814 cfs_spin_lock(&obd->obd_recovery_task_lock);
1815 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1816 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1817 struct ptlrpc_request, rq_list);
1818 cfs_list_del_init(&req->rq_list);
1819 obd->obd_requests_queued_for_recovery--;
1820 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1822 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1823 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1824 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1825 /** evict exports failed VBR */
1826 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1831 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1833 struct ptlrpc_request *req = NULL;
1835 CDEBUG(D_HA, "Waiting for lock\n");
1836 if (target_recovery_overseer(obd, check_for_next_lock,
1837 exp_lock_replay_healthy))
1838 abort_lock_replay_queue(obd);
1840 cfs_spin_lock(&obd->obd_recovery_task_lock);
1841 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1842 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1843 struct ptlrpc_request, rq_list);
1844 cfs_list_del_init(&req->rq_list);
1845 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1847 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1848 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1849 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1850 /** evict exports failed VBR */
1851 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1856 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1858 struct ptlrpc_request *req = NULL;
1860 cfs_spin_lock(&obd->obd_recovery_task_lock);
1861 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1862 req = cfs_list_entry(obd->obd_final_req_queue.next,
1863 struct ptlrpc_request, rq_list);
1864 cfs_list_del_init(&req->rq_list);
1865 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1866 if (req->rq_export->exp_in_recovery) {
1867 cfs_spin_lock(&req->rq_export->exp_lock);
1868 req->rq_export->exp_in_recovery = 0;
1869 cfs_spin_unlock(&req->rq_export->exp_lock);
1872 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1877 static int handle_recovery_req(struct ptlrpc_thread *thread,
1878 struct ptlrpc_request *req,
1879 svc_handler_t handler)
1885 * export can be evicted during recovery, no need to handle replays for
1886 * it after that, discard such request silently
1888 if (req->rq_export->exp_disconnected)
1889 GOTO(reqcopy_put, rc = 0);
1891 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1893 CERROR("Failure to initialize session: %d\n", rc);
1894 GOTO(reqcopy_put, rc);
1897 req->rq_recov_session.lc_thread = thread;
1898 lu_context_enter(&req->rq_recov_session);
1899 req->rq_svc_thread = thread;
1900 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1902 /* thread context */
1903 lu_context_enter(&thread->t_env->le_ctx);
1905 lu_context_exit(&thread->t_env->le_ctx);
1907 lu_context_exit(&req->rq_recov_session);
1908 lu_context_fini(&req->rq_recov_session);
1909 /* don't reset timer for final stage */
1910 if (!exp_finished(req->rq_export)) {
1911 int to = obd_timeout;
1914 * Add request timeout to the recovery time so next request from
1915 * this client may come in recovery time
1918 struct ptlrpc_service_part *svcpt;
1920 svcpt = req->rq_rqbd->rqbd_svcpt;
1921 /* If the server sent early reply for this request,
1922 * the client will recalculate the timeout according to
1923 * current server estimate service time, so we will
1924 * use the maxium timeout here for waiting the client
1925 * sending the next req */
1926 to = max((int)at_est2timeout(
1927 at_get(&svcpt->scp_at_estimate)),
1928 (int)lustre_msg_get_timeout(req->rq_reqmsg));
1929 /* Add net_latency (see ptlrpc_replay_req) */
1930 to += lustre_msg_get_service_time(req->rq_reqmsg);
1932 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1938 static int target_recovery_thread(void *arg)
1940 struct lu_target *lut = arg;
1941 struct obd_device *obd = lut->lut_obd;
1942 struct ptlrpc_request *req;
1943 struct target_recovery_data *trd = &obd->obd_recovery_data;
1944 unsigned long delta;
1945 unsigned long flags;
1947 struct ptlrpc_thread *thread = NULL;
1951 cfs_daemonize_ctxt("tgt_recov");
1953 SIGNAL_MASK_LOCK(current, flags);
1954 sigfillset(¤t->blocked);
1956 SIGNAL_MASK_UNLOCK(current, flags);
1958 OBD_ALLOC_PTR(thread);
1964 OBD_FREE_PTR(thread);
1968 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1970 OBD_FREE_PTR(thread);
1975 thread->t_env = env;
1976 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1977 env->le_ctx.lc_thread = thread;
1978 thread->t_data = NULL;
1979 thread->t_watchdog = NULL;
1981 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1983 trd->trd_processing_task = cfs_curproc_pid();
1985 cfs_spin_lock(&obd->obd_dev_lock);
1986 obd->obd_recovering = 1;
1987 cfs_spin_unlock(&obd->obd_dev_lock);
1988 cfs_complete(&trd->trd_starting);
1990 /* first of all, we have to know the first transno to replay */
1991 if (target_recovery_overseer(obd, check_for_clients,
1992 exp_connect_healthy)) {
1993 abort_req_replay_queue(obd);
1994 abort_lock_replay_queue(obd);
1997 /* next stage: replay requests */
1999 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2000 cfs_atomic_read(&obd->obd_req_replay_clients),
2001 obd->obd_next_recovery_transno);
2002 while ((req = target_next_replay_req(obd))) {
2003 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2004 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2005 lustre_msg_get_transno(req->rq_reqmsg),
2006 libcfs_nid2str(req->rq_peer.nid));
2007 handle_recovery_req(thread, req,
2008 trd->trd_recovery_handler);
2010 * bz18031: increase next_recovery_transno before
2011 * target_request_copy_put() will drop exp_rpc reference
2013 cfs_spin_lock(&obd->obd_recovery_task_lock);
2014 obd->obd_next_recovery_transno++;
2015 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2016 target_exp_dequeue_req_replay(req);
2017 target_request_copy_put(req);
2018 obd->obd_replayed_requests++;
2022 * The second stage: replay locks
2024 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2025 cfs_atomic_read(&obd->obd_lock_replay_clients));
2026 while ((req = target_next_replay_lock(obd))) {
2027 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2028 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2029 libcfs_nid2str(req->rq_peer.nid));
2030 handle_recovery_req(thread, req,
2031 trd->trd_recovery_handler);
2032 target_request_copy_put(req);
2033 obd->obd_replayed_locks++;
2037 * The third stage: reply on final pings, at this moment all clients
2038 * must have request in final queue
2040 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2041 /** Update server last boot epoch */
2042 lut_boot_epoch_update(lut);
2043 /* We drop recoverying flag to forward all new requests
2044 * to regular mds_handle() since now */
2045 cfs_spin_lock(&obd->obd_dev_lock);
2046 obd->obd_recovering = obd->obd_abort_recovery = 0;
2047 cfs_spin_unlock(&obd->obd_dev_lock);
2048 cfs_spin_lock(&obd->obd_recovery_task_lock);
2049 target_cancel_recovery_timer(obd);
2050 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2051 while ((req = target_next_final_ping(obd))) {
2052 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2053 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2054 libcfs_nid2str(req->rq_peer.nid));
2055 handle_recovery_req(thread, req,
2056 trd->trd_recovery_handler);
2057 target_request_copy_put(req);
2060 delta = (jiffies - delta) / CFS_HZ;
2061 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2062 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2063 if (delta > OBD_RECOVERY_TIME_SOFT) {
2064 CWARN("too long recovery - read logs\n");
2065 libcfs_debug_dumplog();
2068 target_finish_recovery(obd);
2070 lu_context_fini(&env->le_ctx);
2071 trd->trd_processing_task = 0;
2072 cfs_complete(&trd->trd_finishing);
2074 OBD_FREE_PTR(thread);
2079 static int target_start_recovery_thread(struct lu_target *lut,
2080 svc_handler_t handler)
2082 struct obd_device *obd = lut->lut_obd;
2084 struct target_recovery_data *trd = &obd->obd_recovery_data;
2086 memset(trd, 0, sizeof(*trd));
2087 cfs_init_completion(&trd->trd_starting);
2088 cfs_init_completion(&trd->trd_finishing);
2089 trd->trd_recovery_handler = handler;
2091 if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
2092 cfs_wait_for_completion(&trd->trd_starting);
2093 LASSERT(obd->obd_recovering != 0);
2100 void target_stop_recovery_thread(struct obd_device *obd)
2102 if (obd->obd_recovery_data.trd_processing_task > 0) {
2103 struct target_recovery_data *trd = &obd->obd_recovery_data;
2104 /** recovery can be done but postrecovery is not yet */
2105 cfs_spin_lock(&obd->obd_dev_lock);
2106 if (obd->obd_recovering) {
2107 CERROR("%s: Aborting recovery\n", obd->obd_name);
2108 obd->obd_abort_recovery = 1;
2109 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2111 cfs_spin_unlock(&obd->obd_dev_lock);
2112 cfs_wait_for_completion(&trd->trd_finishing);
2116 void target_recovery_fini(struct obd_device *obd)
2118 class_disconnect_exports(obd);
2119 target_stop_recovery_thread(obd);
2120 target_cleanup_recovery(obd);
2122 EXPORT_SYMBOL(target_recovery_fini);
2124 static void target_recovery_expired(unsigned long castmeharder)
2126 struct obd_device *obd = (struct obd_device *)castmeharder;
2127 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2128 " after %lds (%d clients connected)\n",
2129 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2130 cfs_time_current_sec()- obd->obd_recovery_start,
2131 cfs_atomic_read(&obd->obd_connected_clients));
2133 obd->obd_recovery_expired = 1;
2134 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2137 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2139 struct obd_device *obd = lut->lut_obd;
2140 if (obd->obd_max_recoverable_clients == 0) {
2141 /** Update server last boot epoch */
2142 lut_boot_epoch_update(lut);
2146 CWARN("RECOVERY: service %s, %d recoverable clients, "
2147 "last_transno "LPU64"\n", obd->obd_name,
2148 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2149 LASSERT(obd->obd_stopping == 0);
2150 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2151 obd->obd_recovery_start = 0;
2152 obd->obd_recovery_end = 0;
2154 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2155 target_start_recovery_thread(lut, handler);
2157 EXPORT_SYMBOL(target_recovery_init);
2159 #endif /* __KERNEL__ */
2161 static int target_process_req_flags(struct obd_device *obd,
2162 struct ptlrpc_request *req)
2164 struct obd_export *exp = req->rq_export;
2165 LASSERT(exp != NULL);
2166 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2167 /* client declares he's ready to replay locks */
2168 cfs_spin_lock(&exp->exp_lock);
2169 if (exp->exp_req_replay_needed) {
2170 exp->exp_req_replay_needed = 0;
2171 cfs_spin_unlock(&exp->exp_lock);
2173 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2174 cfs_atomic_dec(&obd->obd_req_replay_clients);
2176 cfs_spin_unlock(&exp->exp_lock);
2179 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2180 /* client declares he's ready to complete recovery
2181 * so, we put the request on th final queue */
2182 cfs_spin_lock(&exp->exp_lock);
2183 if (exp->exp_lock_replay_needed) {
2184 exp->exp_lock_replay_needed = 0;
2185 cfs_spin_unlock(&exp->exp_lock);
2187 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2188 cfs_atomic_dec(&obd->obd_lock_replay_clients);
2190 cfs_spin_unlock(&exp->exp_lock);
2196 int target_queue_recovery_request(struct ptlrpc_request *req,
2197 struct obd_device *obd)
2201 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2204 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2205 /* Processing the queue right now, don't re-add. */
2209 target_process_req_flags(obd, req);
2211 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2212 /* client declares he's ready to complete recovery
2213 * so, we put the request on th final queue */
2214 target_request_copy_get(req);
2215 DEBUG_REQ(D_HA, req, "queue final req");
2216 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2217 cfs_spin_lock(&obd->obd_recovery_task_lock);
2218 if (obd->obd_recovering) {
2219 cfs_list_add_tail(&req->rq_list,
2220 &obd->obd_final_req_queue);
2222 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2223 target_request_copy_put(req);
2224 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2226 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2229 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2230 /* client declares he's ready to replay locks */
2231 target_request_copy_get(req);
2232 DEBUG_REQ(D_HA, req, "queue lock replay req");
2233 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2234 cfs_spin_lock(&obd->obd_recovery_task_lock);
2235 LASSERT(obd->obd_recovering);
2236 /* usually due to recovery abort */
2237 if (!req->rq_export->exp_in_recovery) {
2238 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2239 target_request_copy_put(req);
2242 LASSERT(req->rq_export->exp_lock_replay_needed);
2243 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2244 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2248 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2249 * (i.e. buflens etc are in my own byte order), but type-dependent
2250 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2253 CFS_INIT_LIST_HEAD(&req->rq_list);
2254 DEBUG_REQ(D_HA, req, "not queueing");
2258 /* If we're processing the queue, we want don't want to queue this
2261 * Also, if this request has a transno less than the one we're waiting
2262 * for, we should process it now. It could (and currently always will)
2263 * be an open request for a descriptor that was opened some time ago.
2265 * Also, a resent, replayed request that has already been
2266 * handled will pass through here and be processed immediately.
2268 CDEBUG(D_HA, "Next recovery transno: "LPU64
2269 ", current: "LPU64", replaying\n",
2270 obd->obd_next_recovery_transno, transno);
2271 cfs_spin_lock(&obd->obd_recovery_task_lock);
2272 if (transno < obd->obd_next_recovery_transno) {
2273 /* Processing the queue right now, don't re-add. */
2274 LASSERT(cfs_list_empty(&req->rq_list));
2275 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2278 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2280 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2283 target_request_copy_get(req);
2284 if (!req->rq_export->exp_in_recovery) {
2285 target_request_copy_put(req);
2288 LASSERT(req->rq_export->exp_req_replay_needed);
2290 if (target_exp_enqueue_req_replay(req)) {
2291 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2292 target_request_copy_put(req);
2297 cfs_spin_lock(&obd->obd_recovery_task_lock);
2298 LASSERT(obd->obd_recovering);
2299 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2300 struct ptlrpc_request *reqiter =
2301 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2303 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2304 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2309 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2311 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2312 "has been claimed by another client");
2313 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2314 target_exp_dequeue_req_replay(req);
2315 target_request_copy_put(req);
2321 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2323 obd->obd_requests_queued_for_recovery++;
2324 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2325 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2329 int target_handle_ping(struct ptlrpc_request *req)
2331 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2332 return req_capsule_server_pack(&req->rq_pill);
2335 void target_committed_to_req(struct ptlrpc_request *req)
2337 struct obd_export *exp = req->rq_export;
2339 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2340 lustre_msg_set_last_committed(req->rq_repmsg,
2341 exp->exp_last_committed);
2343 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2344 "%d)", exp->exp_obd->obd_no_transno,
2345 req->rq_repmsg == NULL);
2347 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2348 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2350 EXPORT_SYMBOL(target_committed_to_req);
2352 #endif /* HAVE_SERVER_SUPPORT */
2355 * Packs current SLV and Limit into \a req.
2357 int target_pack_pool_reply(struct ptlrpc_request *req)
2359 struct obd_device *obd;
2363 * Check that we still have all structures alive as this may
2364 * be some late rpc in shutdown time.
2366 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2367 !exp_connect_lru_resize(req->rq_export))) {
2368 lustre_msg_set_slv(req->rq_repmsg, 0);
2369 lustre_msg_set_limit(req->rq_repmsg, 0);
2374 * OBD is alive here as export is alive, which we checked above.
2376 obd = req->rq_export->exp_obd;
2378 cfs_read_lock(&obd->obd_pool_lock);
2379 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2380 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2381 cfs_read_unlock(&obd->obd_pool_lock);
2386 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2388 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2389 DEBUG_REQ(D_ERROR, req, "dropping reply");
2394 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2395 req->rq_status = rc;
2396 return (ptlrpc_send_error(req, 1));
2398 DEBUG_REQ(D_NET, req, "sending reply");
2401 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2404 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2406 struct ptlrpc_service_part *svcpt;
2408 struct ptlrpc_reply_state *rs;
2409 struct obd_export *exp;
2412 if (req->rq_no_reply) {
2417 svcpt = req->rq_rqbd->rqbd_svcpt;
2418 rs = req->rq_reply_state;
2419 if (rs == NULL || !rs->rs_difficult) {
2421 target_send_reply_msg (req, rc, fail_id);
2426 /* must be an export if locks saved */
2427 LASSERT (req->rq_export != NULL);
2428 /* req/reply consistent */
2429 LASSERT(rs->rs_svcpt == svcpt);
2432 LASSERT (!rs->rs_scheduled);
2433 LASSERT (!rs->rs_scheduled_ever);
2434 LASSERT (!rs->rs_handled);
2435 LASSERT (!rs->rs_on_net);
2436 LASSERT (rs->rs_export == NULL);
2437 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2438 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2440 exp = class_export_get (req->rq_export);
2442 /* disable reply scheduling while I'm setting up */
2443 rs->rs_scheduled = 1;
2445 rs->rs_xid = req->rq_xid;
2446 rs->rs_transno = req->rq_transno;
2447 rs->rs_export = exp;
2448 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2450 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2451 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2452 rs->rs_transno, exp->exp_last_committed);
2453 if (rs->rs_transno > exp->exp_last_committed) {
2454 /* not committed already */
2455 cfs_list_add_tail(&rs->rs_obd_list,
2456 &exp->exp_uncommitted_replies);
2458 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2460 cfs_spin_lock(&exp->exp_lock);
2461 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2462 cfs_spin_unlock(&exp->exp_lock);
2464 netrc = target_send_reply_msg (req, rc, fail_id);
2466 cfs_spin_lock(&svcpt->scp_rep_lock);
2468 cfs_atomic_inc(&svcpt->scp_nreps_difficult);
2471 /* error sending: reply is off the net. Also we need +1
2472 * reply ref until ptlrpc_handle_rs() is done
2473 * with the reply state (if the send was successful, there
2474 * would have been +1 ref for the net, which
2475 * reply_out_callback leaves alone) */
2477 ptlrpc_rs_addref(rs);
2480 cfs_spin_lock(&rs->rs_lock);
2481 if (rs->rs_transno <= exp->exp_last_committed ||
2482 (!rs->rs_on_net && !rs->rs_no_ack) ||
2483 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2484 cfs_list_empty(&rs->rs_obd_list)) {
2485 CDEBUG(D_HA, "Schedule reply immediately\n");
2486 ptlrpc_dispatch_difficult_reply(rs);
2488 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
2489 rs->rs_scheduled = 0; /* allow notifier to schedule */
2491 cfs_spin_unlock(&rs->rs_lock);
2492 cfs_spin_unlock(&svcpt->scp_rep_lock);
2496 int target_handle_qc_callback(struct ptlrpc_request *req)
2498 struct obd_quotactl *oqctl;
2499 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2501 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2502 if (oqctl == NULL) {
2503 CERROR("Can't unpack obd_quotactl\n");
2507 cli->cl_qchk_stat = oqctl->qc_stat;
2512 #ifdef HAVE_QUOTA_SUPPORT
2513 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2516 struct obd_device *obd = req->rq_export->exp_obd;
2517 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2518 struct obd_device_target *obt;
2519 struct lustre_quota_ctxt *qctxt;
2520 struct qunit_data *qdata = NULL;
2524 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2527 rc = req_capsule_server_pack(&req->rq_pill);
2529 CERROR("packing reply failed!: rc = %d\n", rc);
2533 LASSERT(req->rq_export);
2535 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2536 if (IS_ERR(qdata)) {
2537 rc = PTR_ERR(qdata);
2538 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2539 req->rq_status = rc;
2543 /* we use the observer */
2544 if (obd_pin_observer(obd, &lov_obd) ||
2545 obd_pin_observer(lov_obd, &master_obd)) {
2546 CERROR("Can't find the observer, it is recovering\n");
2547 req->rq_status = -EAGAIN;
2551 obt = &master_obd->u.obt;
2552 qctxt = &obt->obt_qctxt;
2554 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2555 /* quota_type has not been processed yet, return EAGAIN
2556 * until we know whether or not quotas are supposed to
2558 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2560 req->rq_status = -EAGAIN;
2564 cfs_down_read(&obt->obt_rwsem);
2565 if (qctxt->lqc_lqs_hash == NULL) {
2566 cfs_up_read(&obt->obt_rwsem);
2567 /* quota_type has not been processed yet, return EAGAIN
2568 * until we know whether or not quotas are supposed to
2570 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2572 req->rq_status = -EAGAIN;
2576 LASSERT(qctxt->lqc_handler);
2577 rc = qctxt->lqc_handler(master_obd, qdata,
2578 lustre_msg_get_opc(req->rq_reqmsg));
2579 cfs_up_read(&obt->obt_rwsem);
2580 if (rc && rc != -EDQUOT)
2581 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2582 "dqacq/dqrel failed! (rc:%d)\n", rc);
2583 req->rq_status = rc;
2585 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2587 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2591 /* Block the quota req. b=14840 */
2592 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2597 obd_unpin_observer(lov_obd);
2599 obd_unpin_observer(obd);
2601 rc = ptlrpc_reply(req);
2605 #endif /* !__KERNEL__ */
2607 #endif /* HAVE_QUOTA_SUPPORT */
2609 ldlm_mode_t lck_compat_array[] = {
2610 [LCK_EX] LCK_COMPAT_EX,
2611 [LCK_PW] LCK_COMPAT_PW,
2612 [LCK_PR] LCK_COMPAT_PR,
2613 [LCK_CW] LCK_COMPAT_CW,
2614 [LCK_CR] LCK_COMPAT_CR,
2615 [LCK_NL] LCK_COMPAT_NL,
2616 [LCK_GROUP] LCK_COMPAT_GROUP,
2617 [LCK_COS] LCK_COMPAT_COS,
2621 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2622 * not escape to the user level.
2624 int ldlm_error2errno(ldlm_error_t error)
2632 case ELDLM_LOCK_CHANGED:
2635 case ELDLM_LOCK_ABORTED:
2638 case ELDLM_LOCK_REPLACED:
2641 case ELDLM_NO_LOCK_DATA:
2644 case ELDLM_NAMESPACE_EXISTS:
2647 case ELDLM_BAD_NAMESPACE:
2651 if (((int)error) < 0) /* cast to signed type */
2652 result = error; /* as ldlm_error_t can be unsigned */
2654 CERROR("Invalid DLM result code: %d\n", error);
2660 EXPORT_SYMBOL(ldlm_error2errno);
2663 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2665 ldlm_error_t ldlm_errno2error(int err_no)
2674 error = ELDLM_LOCK_CHANGED;
2677 error = ELDLM_LOCK_ABORTED;
2680 error = ELDLM_LOCK_REPLACED;
2683 error = ELDLM_NO_LOCK_DATA;
2686 error = ELDLM_NAMESPACE_EXISTS;
2689 error = ELDLM_BAD_NAMESPACE;
2696 EXPORT_SYMBOL(ldlm_errno2error);
2698 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2699 void ldlm_dump_export_locks(struct obd_export *exp)
2701 cfs_spin_lock(&exp->exp_locks_list_guard);
2702 if (!cfs_list_empty(&exp->exp_locks_list)) {
2703 struct ldlm_lock *lock;
2705 CERROR("dumping locks for export %p,"
2706 "ignore if the unmount doesn't hang\n", exp);
2707 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2708 LDLM_ERROR(lock, "lock:");
2710 cfs_spin_unlock(&exp->exp_locks_list_guard);
2714 #ifdef HAVE_SERVER_SUPPORT
2715 static int target_bulk_timeout(void *data)
2718 /* We don't fail the connection here, because having the export
2719 * killed makes the (vital) call to commitrw very sad.
2724 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2726 return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2729 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2730 struct l_wait_info *lwi)
2732 struct ptlrpc_request *req = desc->bd_req;
2736 /* Check if there is eviction in progress, and if so, wait for
2738 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2739 *lwi = LWI_INTR(NULL, NULL);
2740 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2741 !cfs_atomic_read(&exp->exp_obd->
2742 obd_evict_inprogress),
2746 /* Check if client was evicted or tried to reconnect already */
2747 if (exp->exp_failed || exp->exp_abort_active_req) {
2750 if (desc->bd_type == BULK_PUT_SINK)
2751 rc = sptlrpc_svc_wrap_bulk(req, desc);
2753 rc = ptlrpc_start_bulk_transfer(desc);
2756 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2757 ptlrpc_abort_bulk(desc);
2758 } else if (rc == 0) {
2759 time_t start = cfs_time_current_sec();
2761 long timeoutl = req->rq_deadline - cfs_time_current_sec();
2762 cfs_duration_t timeout = timeoutl <= 0 ?
2763 CFS_TICK : cfs_time_seconds(timeoutl);
2764 *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2765 cfs_time_seconds(1),
2766 target_bulk_timeout,
2768 rc = l_wait_event(desc->bd_waitq,
2769 !ptlrpc_server_bulk_active(desc) ||
2771 exp->exp_abort_active_req,
2773 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2774 /* Wait again if we changed deadline */
2775 } while ((rc == -ETIMEDOUT) &&
2776 (req->rq_deadline > cfs_time_current_sec()));
2778 if (rc == -ETIMEDOUT) {
2779 DEBUG_REQ(D_ERROR, req,
2780 "timeout on bulk %s after %ld%+lds",
2782 req->rq_deadline - start,
2783 cfs_time_current_sec() -
2785 ptlrpc_abort_bulk(desc);
2786 } else if (exp->exp_failed) {
2787 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2790 ptlrpc_abort_bulk(desc);
2791 } else if (exp->exp_abort_active_req) {
2792 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2794 /* we don't reply anyway */
2796 ptlrpc_abort_bulk(desc);
2797 } else if (!desc->bd_success ||
2798 desc->bd_nob_transferred != desc->bd_nob) {
2799 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2801 "truncated" : "network error on",
2803 desc->bd_nob_transferred,
2805 /* XXX should this be a different errno? */
2807 } else if (desc->bd_type == BULK_GET_SINK) {
2808 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2811 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2812 bulk2type(desc), rc);
2817 EXPORT_SYMBOL(target_bulk_io);
2819 #endif /* HAVE_SERVER_SUPPORT */