1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_mds.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: if non-zero, move the selected to the list head
55 * @create: if zero, only search in existed connections
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 cfs_spin_lock(&imp->imp_lock);
84 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 cfs_list_del(&item->oic_item);
88 cfs_list_add(&item->oic_item,
90 item->oic_last_attempt = 0;
92 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? ", moved to head" : ""));
95 cfs_spin_unlock(&imp->imp_lock);
96 GOTO(out_free, rc = 0);
101 imp_conn->oic_conn = ptlrpc_conn;
102 imp_conn->oic_uuid = *uuid;
103 imp_conn->oic_last_attempt = 0;
105 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107 cfs_list_add_tail(&imp_conn->oic_item,
108 &imp->imp_conn_list);
109 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
110 imp, imp->imp_obd->obd_name, uuid->uuid,
111 (priority ? "head" : "tail"));
113 cfs_spin_unlock(&imp->imp_lock);
114 GOTO(out_free, rc = -ENOENT);
117 cfs_spin_unlock(&imp->imp_lock);
121 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 ptlrpc_connection_put(ptlrpc_conn);
127 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 return import_set_conn(imp, uuid, 1, 0);
132 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
135 return import_set_conn(imp, uuid, priority, 1);
138 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
140 struct obd_import_conn *imp_conn;
141 struct obd_export *dlmexp;
145 cfs_spin_lock(&imp->imp_lock);
146 if (cfs_list_empty(&imp->imp_conn_list)) {
147 LASSERT(!imp->imp_connection);
151 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
152 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
154 LASSERT(imp_conn->oic_conn);
156 /* is current conn? */
157 if (imp_conn == imp->imp_conn_current) {
158 LASSERT(imp_conn->oic_conn == imp->imp_connection);
160 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
161 imp->imp_state != LUSTRE_IMP_DISCON) {
162 CERROR("can't remove current connection\n");
163 GOTO(out, rc = -EBUSY);
166 ptlrpc_connection_put(imp->imp_connection);
167 imp->imp_connection = NULL;
169 dlmexp = class_conn2export(&imp->imp_dlm_handle);
170 if (dlmexp && dlmexp->exp_connection) {
171 LASSERT(dlmexp->exp_connection ==
173 ptlrpc_connection_put(dlmexp->exp_connection);
174 dlmexp->exp_connection = NULL;
178 cfs_list_del(&imp_conn->oic_item);
179 ptlrpc_connection_put(imp_conn->oic_conn);
180 OBD_FREE(imp_conn, sizeof(*imp_conn));
181 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182 imp, imp->imp_obd->obd_name, uuid->uuid);
187 cfs_spin_unlock(&imp->imp_lock);
189 CERROR("connection %s not found\n", uuid->uuid);
193 void client_destroy_import(struct obd_import *imp)
195 /* drop security policy instance after all rpc finished/aborted
196 * to let all busy contexts be released. */
197 class_import_get(imp);
198 class_destroy_import(imp);
199 sptlrpc_import_sec_put(imp);
200 class_import_put(imp);
203 /* configure an RPC client OBD device
208 * 3 - inactive-on-startup
210 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
212 struct client_obd *cli = &obddev->u.cli;
213 struct obd_import *imp;
214 struct obd_uuid server_uuid;
215 int rq_portal, rp_portal, connect_op;
216 char *name = obddev->obd_type->typ_name;
220 /* In a more perfect world, we would hang a ptlrpc_client off of
221 * obd_type and just use the values from there. */
222 if (!strcmp(name, LUSTRE_OSC_NAME)) {
223 rq_portal = OST_REQUEST_PORTAL;
224 rp_portal = OSC_REPLY_PORTAL;
225 connect_op = OST_CONNECT;
226 cli->cl_sp_me = LUSTRE_SP_CLI;
227 cli->cl_sp_to = LUSTRE_SP_OST;
228 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
229 rq_portal = MDS_REQUEST_PORTAL;
230 rp_portal = MDC_REPLY_PORTAL;
231 connect_op = MDS_CONNECT;
232 cli->cl_sp_me = LUSTRE_SP_CLI;
233 cli->cl_sp_to = LUSTRE_SP_MDT;
234 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
235 rq_portal = MGS_REQUEST_PORTAL;
236 rp_portal = MGC_REPLY_PORTAL;
237 connect_op = MGS_CONNECT;
238 cli->cl_sp_me = LUSTRE_SP_MGC;
239 cli->cl_sp_to = LUSTRE_SP_MGS;
240 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
242 CERROR("unknown client OBD type \"%s\", can't setup\n",
247 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
248 CERROR("requires a TARGET UUID\n");
252 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
253 CERROR("client UUID must be less than 38 characters\n");
257 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
258 CERROR("setup requires a SERVER UUID\n");
262 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
263 CERROR("target UUID must be less than 38 characters\n");
267 cfs_init_rwsem(&cli->cl_sem);
268 cfs_sema_init(&cli->cl_mgc_sem, 1);
269 cli->cl_conn_count = 0;
270 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
271 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
272 sizeof(server_uuid)));
275 cli->cl_avail_grant = 0;
276 /* FIXME: should limit this for the sum of all cl_dirty_max */
277 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
278 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
279 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
280 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
281 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
282 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
283 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
284 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
285 client_obd_list_lock_init(&cli->cl_loi_list_lock);
286 cli->cl_r_in_flight = 0;
287 cli->cl_w_in_flight = 0;
289 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
290 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
291 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
292 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
293 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
294 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
295 cfs_waitq_init(&cli->cl_destroy_waitq);
296 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
297 #ifdef ENABLE_CHECKSUM
298 /* Turn on checksumming by default. */
299 cli->cl_checksum = 1;
301 * The supported checksum types will be worked out at connect time
302 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
305 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
307 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
309 /* This value may be changed at connect time in
310 ptlrpc_connect_interpret. */
311 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
312 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
314 if (!strcmp(name, LUSTRE_MDC_NAME)) {
315 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
316 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
317 cli->cl_max_rpcs_in_flight = 2;
318 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
319 cli->cl_max_rpcs_in_flight = 3;
320 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
321 cli->cl_max_rpcs_in_flight = 4;
323 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
328 CERROR("ldlm_get_ref failed: %d\n", rc);
332 ptlrpc_init_client(rq_portal, rp_portal, name,
333 &obddev->obd_ldlm_client);
335 imp = class_new_import(obddev);
337 GOTO(err_ldlm, rc = -ENOENT);
338 imp->imp_client = &obddev->obd_ldlm_client;
339 imp->imp_connect_op = connect_op;
340 imp->imp_initial_recov = 1;
341 imp->imp_initial_recov_bk = 0;
342 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
343 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
344 LUSTRE_CFG_BUFLEN(lcfg, 1));
345 class_import_put(imp);
347 rc = client_import_add_conn(imp, &server_uuid, 1);
349 CERROR("can't add initial connection\n");
350 GOTO(err_import, rc);
353 cli->cl_import = imp;
354 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
355 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
356 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
358 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
359 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
360 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
361 name, obddev->obd_name,
362 cli->cl_target_uuid.uuid);
363 cfs_spin_lock(&imp->imp_lock);
364 imp->imp_deactive = 1;
365 cfs_spin_unlock(&imp->imp_lock);
369 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
370 LDLM_NAMESPACE_CLIENT,
371 LDLM_NAMESPACE_GREEDY);
372 if (obddev->obd_namespace == NULL) {
373 CERROR("Unable to create client namespace - %s\n",
375 GOTO(err_import, rc = -ENOMEM);
378 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
383 class_destroy_import(imp);
391 int client_obd_cleanup(struct obd_device *obddev)
395 ldlm_namespace_free_post(obddev->obd_namespace);
396 obddev->obd_namespace = NULL;
402 /* ->o_connect() method for client side (OSC and MDC and MGC) */
403 int client_connect_import(const struct lu_env *env,
404 struct obd_export **exp,
405 struct obd_device *obd, struct obd_uuid *cluuid,
406 struct obd_connect_data *data, void *localdata)
408 struct client_obd *cli = &obd->u.cli;
409 struct obd_import *imp = cli->cl_import;
410 struct obd_connect_data *ocd;
411 struct lustre_handle conn = { 0 };
416 cfs_down_write(&cli->cl_sem);
417 if (cli->cl_conn_count > 0 )
418 GOTO(out_sem, rc = -EALREADY);
420 rc = class_connect(&conn, obd, cluuid);
424 cli->cl_conn_count++;
425 *exp = class_conn2export(&conn);
427 LASSERT(obd->obd_namespace);
429 imp->imp_dlm_handle = conn;
430 rc = ptlrpc_init_import(imp);
434 ocd = &imp->imp_connect_data;
437 imp->imp_connect_flags_orig = data->ocd_connect_flags;
440 rc = ptlrpc_connect_import(imp, NULL);
442 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
445 LASSERT((*exp)->exp_connection);
448 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
449 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
450 data->ocd_connect_flags, ocd->ocd_connect_flags);
451 data->ocd_connect_flags = ocd->ocd_connect_flags;
454 ptlrpc_pinger_add_import(imp);
460 cli->cl_conn_count--;
461 class_disconnect(*exp);
465 cfs_up_write(&cli->cl_sem);
470 int client_disconnect_export(struct obd_export *exp)
472 struct obd_device *obd = class_exp2obd(exp);
473 struct client_obd *cli;
474 struct obd_import *imp;
479 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
480 exp, exp ? exp->exp_handle.h_cookie : -1);
485 imp = cli->cl_import;
487 cfs_down_write(&cli->cl_sem);
488 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
491 if (!cli->cl_conn_count) {
492 CERROR("disconnecting disconnected device (%s)\n",
494 GOTO(out_disconnect, rc = -EINVAL);
497 cli->cl_conn_count--;
498 if (cli->cl_conn_count)
499 GOTO(out_disconnect, rc = 0);
501 /* Mark import deactivated now, so we don't try to reconnect if any
502 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
503 * fully deactivate the import, or that would drop all requests. */
504 cfs_spin_lock(&imp->imp_lock);
505 imp->imp_deactive = 1;
506 cfs_spin_unlock(&imp->imp_lock);
508 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
509 * delete it regardless. (It's safe to delete an import that was
511 (void)ptlrpc_pinger_del_import(imp);
513 if (obd->obd_namespace != NULL) {
514 /* obd_force == local only */
515 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
516 obd->obd_force ? LCF_LOCAL : 0, NULL);
517 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
521 * there's no need to hold sem during disconnecting an import,
522 * and actually it may cause deadlock in gss.
524 cfs_up_write(&cli->cl_sem);
525 rc = ptlrpc_disconnect_import(imp, 0);
526 cfs_down_write(&cli->cl_sem);
528 ptlrpc_invalidate_import(imp);
530 if (imp->imp_rq_pool) {
531 ptlrpc_free_rq_pool(imp->imp_rq_pool);
532 imp->imp_rq_pool = NULL;
534 client_destroy_import(imp);
535 cli->cl_import = NULL;
540 /* use server style - class_disconnect should be always called for
542 err = class_disconnect(exp);
546 cfs_up_write(&cli->cl_sem);
551 int server_disconnect_export(struct obd_export *exp)
556 /* Disconnect early so that clients can't keep using export */
557 rc = class_disconnect(exp);
558 /* close import for avoid sending any requests */
559 if (exp->exp_imp_reverse)
560 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
562 if (exp->exp_obd->obd_namespace != NULL)
563 ldlm_cancel_locks_for_export(exp);
565 /* complete all outstanding replies */
566 cfs_spin_lock(&exp->exp_lock);
567 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
568 struct ptlrpc_reply_state *rs =
569 cfs_list_entry(exp->exp_outstanding_replies.next,
570 struct ptlrpc_reply_state, rs_exp_list);
571 struct ptlrpc_service *svc = rs->rs_service;
573 cfs_spin_lock(&svc->srv_lock);
574 cfs_list_del_init(&rs->rs_exp_list);
575 cfs_spin_lock(&rs->rs_lock);
576 ptlrpc_schedule_difficult_reply(rs);
577 cfs_spin_unlock(&rs->rs_lock);
578 cfs_spin_unlock(&svc->srv_lock);
580 cfs_spin_unlock(&exp->exp_lock);
582 /* release nid stat refererence */
583 lprocfs_exp_cleanup(exp);
588 /* --------------------------------------------------------------------------
589 * from old lib/target.c
590 * -------------------------------------------------------------------------- */
592 static int target_handle_reconnect(struct lustre_handle *conn,
593 struct obd_export *exp,
594 struct obd_uuid *cluuid)
597 if (exp->exp_connection && exp->exp_imp_reverse) {
598 struct lustre_handle *hdl;
599 hdl = &exp->exp_imp_reverse->imp_remote_handle;
600 /* Might be a re-connect after a partition. */
601 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
602 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
604 conn->cookie = exp->exp_handle.h_cookie;
605 /* target_handle_connect() treats EALREADY and
606 * -EALREADY differently. EALREADY means we are
607 * doing a valid reconnect from the same client. */
610 CERROR("%s reconnecting from %s, "
611 "handle mismatch (ours "LPX64", theirs "
612 LPX64")\n", cluuid->uuid,
613 exp->exp_connection->c_remote_uuid.uuid,
614 hdl->cookie, conn->cookie);
615 memset(conn, 0, sizeof *conn);
616 /* target_handle_connect() treats EALREADY and
617 * -EALREADY differently. -EALREADY is an error
618 * (same UUID, different handle). */
623 conn->cookie = exp->exp_handle.h_cookie;
624 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
625 cluuid->uuid, exp, conn->cookie);
629 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
632 struct obd_export *exp = cb_data;
634 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
635 obd->obd_name, exp->exp_client_uuid.uuid);
637 cfs_spin_lock(&exp->exp_lock);
638 exp->exp_need_sync = 0;
639 cfs_spin_unlock(&exp->exp_lock);
640 class_export_cb_put(exp);
642 EXPORT_SYMBOL(target_client_add_cb);
645 target_start_and_reset_recovery_timer(struct obd_device *obd,
646 struct ptlrpc_request *req,
649 int target_handle_connect(struct ptlrpc_request *req)
651 struct obd_device *target, *targref = NULL;
652 struct obd_export *export = NULL;
653 struct obd_import *revimp;
654 struct lustre_handle conn;
655 struct lustre_handle *tmp;
656 struct obd_uuid tgtuuid;
657 struct obd_uuid cluuid;
658 struct obd_uuid remote_uuid;
662 struct obd_connect_data *data, *tmpdata;
663 lnet_nid_t *client_nid = NULL;
666 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
668 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
670 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
671 GOTO(out, rc = -EINVAL);
674 obd_str2uuid(&tgtuuid, str);
675 target = class_uuid2obd(&tgtuuid);
677 target = class_name2obd(str);
679 if (!target || target->obd_stopping || !target->obd_set_up) {
680 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
681 " for connect (%s)\n", str,
682 !target ? "no target" :
683 (target->obd_stopping ? "stopping" :
685 GOTO(out, rc = -ENODEV);
688 if (target->obd_no_conn) {
689 LCONSOLE_WARN("%s: temporarily refusing client connection "
690 "from %s\n", target->obd_name,
691 libcfs_nid2str(req->rq_peer.nid));
692 GOTO(out, rc = -EAGAIN);
695 /* Make sure the target isn't cleaned up while we're here. Yes,
696 there's still a race between the above check and our incref here.
697 Really, class_uuid2obd should take the ref. */
698 targref = class_incref(target, __FUNCTION__, cfs_current());
701 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
703 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
704 GOTO(out, rc = -EINVAL);
707 obd_str2uuid(&cluuid, str);
709 /* XXX extract a nettype and format accordingly */
710 switch (sizeof(lnet_nid_t)) {
711 /* NB the casts only avoid compiler warnings */
713 snprintf(remote_uuid.uuid, sizeof remote_uuid,
714 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
717 snprintf(remote_uuid.uuid, sizeof remote_uuid,
718 "NET_%x_UUID", (__u32)req->rq_peer.nid);
724 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
726 GOTO(out, rc = -EPROTO);
730 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
732 GOTO(out, rc = -EPROTO);
734 rc = req_capsule_server_pack(&req->rq_pill);
738 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
740 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
741 "libclient connection attempt");
742 GOTO(out, rc = -EPROTO);
743 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
744 LUSTRE_VERSION_ALLOWED_OFFSET ||
745 data->ocd_version > LUSTRE_VERSION_CODE +
746 LUSTRE_VERSION_ALLOWED_OFFSET) {
747 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
748 "libclient connection attempt",
749 data->ocd_version < LUSTRE_VERSION_CODE ?
751 OBD_OCD_VERSION_MAJOR(data->ocd_version),
752 OBD_OCD_VERSION_MINOR(data->ocd_version),
753 OBD_OCD_VERSION_PATCH(data->ocd_version),
754 OBD_OCD_VERSION_FIX(data->ocd_version));
755 data = req_capsule_server_sized_get(&req->rq_pill,
757 offsetof(typeof(*data), ocd_version) +
758 sizeof(data->ocd_version));
760 data->ocd_connect_flags = OBD_CONNECT_VERSION;
761 data->ocd_version = LUSTRE_VERSION_CODE;
763 GOTO(out, rc = -EPROTO);
767 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
768 (data->ocd_connect_flags & OBD_CONNECT_MDS))
771 /* lctl gets a backstage, all-access pass. */
772 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
773 goto dont_check_exports;
775 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
779 /* we've found an export in the hash */
780 if (export->exp_connecting) { /* bug 9635, et. al. */
781 CWARN("%s: exp %p already connecting\n",
782 export->exp_obd->obd_name, export);
783 class_export_put(export);
786 } else if (mds_conn && export->exp_connection) {
787 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
788 /* mds reconnected after failover */
789 CWARN("%s: received MDS connection from NID %s,"
790 " removing former export from NID %s\n",
791 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
792 libcfs_nid2str(export->exp_connection->c_peer.nid));
794 /* new mds connection from the same nid */
795 CWARN("%s: received new MDS connection from NID %s,"
796 " removing former export from same NID\n",
797 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
798 class_fail_export(export);
799 class_export_put(export);
802 } else if (export->exp_connection != NULL &&
803 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
804 (lustre_msg_get_op_flags(req->rq_reqmsg) &
805 MSG_CONNECT_INITIAL)) {
806 /* in mds failover we have static uuid but nid can be
808 CWARN("%s: cookie %s seen on new NID %s when "
809 "existing NID %s is already connected\n",
810 target->obd_name, cluuid.uuid,
811 libcfs_nid2str(req->rq_peer.nid),
812 libcfs_nid2str(export->exp_connection->c_peer.nid));
814 class_export_put(export);
817 cfs_spin_lock(&export->exp_lock);
818 export->exp_connecting = 1;
819 cfs_spin_unlock(&export->exp_lock);
820 class_export_put(export);
821 LASSERT(export->exp_obd == target);
823 rc = target_handle_reconnect(&conn, export, &cluuid);
826 /* If we found an export, we already unlocked. */
829 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
830 } else if (req->rq_export == NULL &&
831 cfs_atomic_read(&export->exp_rpc_count) > 0) {
832 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
833 target->obd_name, cluuid.uuid,
834 libcfs_nid2str(req->rq_peer.nid),
835 export, cfs_atomic_read(&export->exp_refcount));
836 GOTO(out, rc = -EBUSY);
837 } else if (req->rq_export != NULL &&
838 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
839 /* the current connect rpc has increased exp_rpc_count */
840 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
841 target->obd_name, cluuid.uuid,
842 libcfs_nid2str(req->rq_peer.nid),
843 export, cfs_atomic_read(&export->exp_rpc_count) - 1);
844 cfs_spin_lock(&export->exp_lock);
845 if (req->rq_export->exp_conn_cnt <
846 lustre_msg_get_conn_cnt(req->rq_reqmsg))
847 /* try to abort active requests */
848 req->rq_export->exp_abort_active_req = 1;
849 cfs_spin_unlock(&export->exp_lock);
850 GOTO(out, rc = -EBUSY);
851 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
852 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
853 "cookies not random?\n", target->obd_name,
854 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
855 GOTO(out, rc = -EALREADY);
857 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
864 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
865 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
866 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
867 export, (long)cfs_time_current_sec(),
868 export ? (long)export->exp_last_request_time : 0);
870 /* Tell the client if we're in recovery. */
871 if (target->obd_recovering) {
872 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
873 /* If this is the first time a client connects,
874 reset the recovery timer */
876 target_start_and_reset_recovery_timer(target, req,
880 /* We want to handle EALREADY but *not* -EALREADY from
881 * target_handle_reconnect(), return reconnection state in a flag */
882 if (rc == EALREADY) {
883 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
889 /* Tell the client if we support replayable requests */
890 if (target->obd_replayable)
891 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
892 client_nid = &req->rq_peer.nid;
894 if (export == NULL) {
895 if (target->obd_recovering) {
898 t = cfs_timer_deadline(&target->obd_recovery_timer);
899 t = cfs_time_sub(t, cfs_time_current());
900 CERROR("%s: denying connection for new client %s (%s): "
901 "%d clients in recovery for "CFS_TIME_T"s\n",
903 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
904 cfs_atomic_read(&target-> \
905 obd_lock_replay_clients),
906 cfs_duration_sec(t));
910 rc = obd_connect(req->rq_svc_thread->t_env,
911 &export, target, &cluuid, data,
914 conn.cookie = export->exp_handle.h_cookie;
917 rc = obd_reconnect(req->rq_svc_thread->t_env,
918 export, target, &cluuid, data, client_nid);
920 /* prevous done via class_conn2export */
921 class_export_get(export);
925 /* Return only the parts of obd_connect_data that we understand, so the
926 * client knows that we don't understand the rest. */
928 tmpdata = req_capsule_server_get(&req->rq_pill,
930 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
934 /* If all else goes well, this is our RPC return code. */
937 lustre_msg_set_handle(req->rq_repmsg, &conn);
939 /* If the client and the server are the same node, we will already
940 * have an export that really points to the client's DLM export,
941 * because we have a shared handles table.
943 * XXX this will go away when shaver stops sending the "connect" handle
944 * in the real "remote handle" field of the request --phik 24 Apr 2003
946 if (req->rq_export != NULL)
947 class_export_put(req->rq_export);
949 req->rq_export = export;
951 cfs_spin_lock(&export->exp_lock);
952 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
953 cfs_spin_unlock(&export->exp_lock);
954 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
955 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
956 export->exp_conn_cnt,
957 lustre_msg_get_conn_cnt(req->rq_reqmsg));
959 GOTO(out, rc = -EALREADY);
961 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
962 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
963 export->exp_abort_active_req = 0;
965 /* request from liblustre? Don't evict it for not pinging. */
966 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
967 export->exp_libclient = 1;
968 cfs_spin_unlock(&export->exp_lock);
970 cfs_spin_lock(&target->obd_dev_lock);
971 cfs_list_del_init(&export->exp_obd_chain_timed);
972 cfs_spin_unlock(&target->obd_dev_lock);
974 cfs_spin_unlock(&export->exp_lock);
977 if (export->exp_connection != NULL) {
978 /* Check to see if connection came from another NID */
979 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
980 !cfs_hlist_unhashed(&export->exp_nid_hash))
981 cfs_hash_del(export->exp_obd->obd_nid_hash,
982 &export->exp_connection->c_peer.nid,
983 &export->exp_nid_hash);
985 ptlrpc_connection_put(export->exp_connection);
988 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
991 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
992 cfs_hash_add_unique(export->exp_obd->obd_nid_hash,
993 &export->exp_connection->c_peer.nid,
994 &export->exp_nid_hash);
997 cfs_spin_lock_bh(&target->obd_processing_task_lock);
998 if (target->obd_recovering && !export->exp_in_recovery) {
999 cfs_spin_lock(&export->exp_lock);
1000 export->exp_in_recovery = 1;
1001 export->exp_req_replay_needed = 1;
1002 export->exp_lock_replay_needed = 1;
1003 cfs_spin_unlock(&export->exp_lock);
1004 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1005 && (data->ocd_transno == 0))
1006 CWARN("Connect with zero transno!\n");
1008 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1009 && data->ocd_transno < target->obd_next_recovery_transno)
1010 target->obd_next_recovery_transno = data->ocd_transno;
1011 target->obd_connected_clients++;
1012 cfs_atomic_inc(&target->obd_req_replay_clients);
1013 cfs_atomic_inc(&target->obd_lock_replay_clients);
1014 if (target->obd_connected_clients ==
1015 target->obd_max_recoverable_clients)
1016 cfs_waitq_signal(&target->obd_next_transno_waitq);
1018 cfs_spin_unlock_bh(&target->obd_processing_task_lock);
1019 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1022 if (export->exp_imp_reverse != NULL) {
1023 /* destroyed import can be still referenced in ctxt */
1024 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1025 KEY_REVIMP_UPD, 0, NULL, NULL);
1027 client_destroy_import(export->exp_imp_reverse);
1030 /* for the rest part, we return -ENOTCONN in case of errors
1031 * in order to let client initialize connection again.
1033 revimp = export->exp_imp_reverse = class_new_import(target);
1035 CERROR("fail to alloc new reverse import.\n");
1036 GOTO(out, rc = -ENOTCONN);
1039 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1040 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1041 revimp->imp_remote_handle = conn;
1042 revimp->imp_dlm_fake = 1;
1043 revimp->imp_state = LUSTRE_IMP_FULL;
1045 /* unknown versions will be caught in
1046 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1047 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1049 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1050 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1051 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1053 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1055 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1056 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1057 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1059 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1061 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1063 CERROR("Failed to get sec for reverse import: %d\n", rc);
1064 export->exp_imp_reverse = NULL;
1065 class_destroy_import(revimp);
1068 class_import_put(revimp);
1071 cfs_spin_lock(&export->exp_lock);
1072 export->exp_connecting = 0;
1073 cfs_spin_unlock(&export->exp_lock);
1076 class_decref(targref, __FUNCTION__, cfs_current());
1078 req->rq_status = rc;
1082 int target_handle_disconnect(struct ptlrpc_request *req)
1087 rc = req_capsule_server_pack(&req->rq_pill);
1091 /* keep the rq_export around so we can send the reply */
1092 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1097 void target_destroy_export(struct obd_export *exp)
1099 /* exports created from last_rcvd data, and "fake"
1100 exports created by lctl don't have an import */
1101 if (exp->exp_imp_reverse != NULL)
1102 client_destroy_import(exp->exp_imp_reverse);
1104 LASSERT(cfs_atomic_read(&exp->exp_locks_count) == 0);
1105 LASSERT(cfs_atomic_read(&exp->exp_rpc_count) == 0);
1106 LASSERT(cfs_atomic_read(&exp->exp_cb_count) == 0);
1107 LASSERT(cfs_atomic_read(&exp->exp_replay_count) == 0);
1111 * Recovery functions
1113 static void target_request_copy_get(struct ptlrpc_request *req)
1115 class_export_rpc_get(req->rq_export);
1116 LASSERT(cfs_list_empty(&req->rq_list));
1117 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1118 /* increase refcount to keep request in queue */
1119 LASSERT(cfs_atomic_read(&req->rq_refcount));
1120 cfs_atomic_inc(&req->rq_refcount);
1121 /** let export know it has replays to be handled */
1122 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1123 /* release service thread while request is queued
1124 * we are moving the request from active processing
1125 * to waiting on the replay queue */
1126 ptlrpc_server_active_request_dec(req);
1129 static void target_request_copy_put(struct ptlrpc_request *req)
1131 LASSERT(cfs_list_empty(&req->rq_replay_list));
1132 LASSERT(cfs_atomic_read(&req->rq_export->exp_replay_count) > 0);
1133 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1134 class_export_rpc_put(req->rq_export);
1135 /* ptlrpc_server_drop_request() assumes the request is active */
1136 ptlrpc_server_active_request_inc(req);
1137 ptlrpc_server_drop_request(req);
1140 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1142 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1143 struct obd_export *exp = req->rq_export;
1144 struct ptlrpc_request *reqiter;
1149 cfs_spin_lock(&exp->exp_lock);
1150 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1152 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1159 /* we expect it with RESENT and REPLAY flags */
1160 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1161 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1162 CERROR("invalid flags %x of resent replay\n",
1163 lustre_msg_get_flags(req->rq_reqmsg));
1165 cfs_list_add_tail(&req->rq_replay_list,
1166 &exp->exp_req_replay_queue);
1169 cfs_spin_unlock(&exp->exp_lock);
1173 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1175 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1176 LASSERT(req->rq_export);
1178 cfs_spin_lock(&req->rq_export->exp_lock);
1179 cfs_list_del_init(&req->rq_replay_list);
1180 cfs_spin_unlock(&req->rq_export->exp_lock);
1184 static void target_finish_recovery(struct obd_device *obd)
1187 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1190 ldlm_reprocess_all_ns(obd->obd_namespace);
1191 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1192 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1193 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1194 !cfs_list_empty(&obd->obd_final_req_queue)) {
1195 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1197 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1198 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1200 cfs_list_empty(&obd->obd_final_req_queue) ? \
1202 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1205 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1207 obd->obd_recovery_end = cfs_time_current_sec();
1209 /* when recovery finished, cleanup orphans on mds and ost */
1210 if (OBT(obd) && OBP(obd, postrecov)) {
1211 int rc = OBP(obd, postrecov)(obd);
1213 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1219 static void abort_req_replay_queue(struct obd_device *obd)
1221 struct ptlrpc_request *req, *n;
1222 cfs_list_t abort_list;
1224 CFS_INIT_LIST_HEAD(&abort_list);
1225 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1226 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1227 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1228 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1229 DEBUG_REQ(D_WARNING, req, "aborted:");
1230 req->rq_status = -ENOTCONN;
1231 if (ptlrpc_error(req)) {
1232 DEBUG_REQ(D_ERROR, req,
1233 "failed abort_req_reply; skipping");
1235 target_exp_dequeue_req_replay(req);
1236 target_request_copy_put(req);
1240 static void abort_lock_replay_queue(struct obd_device *obd)
1242 struct ptlrpc_request *req, *n;
1243 cfs_list_t abort_list;
1245 CFS_INIT_LIST_HEAD(&abort_list);
1246 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1247 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1248 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1249 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1250 DEBUG_REQ(D_ERROR, req, "aborted:");
1251 req->rq_status = -ENOTCONN;
1252 if (ptlrpc_error(req)) {
1253 DEBUG_REQ(D_ERROR, req,
1254 "failed abort_lock_reply; skipping");
1256 target_request_copy_put(req);
1261 /* Called from a cleanup function if the device is being cleaned up
1262 forcefully. The exports should all have been disconnected already,
1263 the only thing left to do is
1264 - clear the recovery flags
1266 - free queued requests and replies, but don't send replies
1267 Because the obd_stopping flag is set, no new requests should be received.
1270 void target_cleanup_recovery(struct obd_device *obd)
1272 struct ptlrpc_request *req, *n;
1273 cfs_list_t clean_list;
1276 CFS_INIT_LIST_HEAD(&clean_list);
1277 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1278 if (!obd->obd_recovering) {
1279 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1283 obd->obd_recovering = obd->obd_abort_recovery = 0;
1284 target_cancel_recovery_timer(obd);
1286 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1287 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1289 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1290 LASSERT(req->rq_reply_state == 0);
1291 target_exp_dequeue_req_replay(req);
1292 target_request_copy_put(req);
1295 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1296 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1297 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1298 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1300 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1301 LASSERT(req->rq_reply_state == 0);
1302 target_request_copy_put(req);
1308 /* obd_processing_task_lock should be held */
1309 void target_cancel_recovery_timer(struct obd_device *obd)
1311 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1312 cfs_timer_disarm(&obd->obd_recovery_timer);
1315 /* extend = 1 means require at least "duration" seconds left in the timer,
1316 extend = 0 means set the total duration (start_recovery_timer) */
1317 static void reset_recovery_timer(struct obd_device *obd, int duration,
1320 cfs_time_t now = cfs_time_current_sec();
1321 cfs_duration_t left;
1323 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1324 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1325 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1329 left = cfs_time_sub(obd->obd_recovery_end, now);
1331 if (extend && (duration > left))
1332 obd->obd_recovery_timeout += duration - left;
1333 else if (!extend && (duration > obd->obd_recovery_timeout))
1334 /* Track the client's largest expected replay time */
1335 obd->obd_recovery_timeout = duration;
1337 /* Hard limit of obd_recovery_time_hard which should not happen */
1338 if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1339 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1341 obd->obd_recovery_end = obd->obd_recovery_start +
1342 obd->obd_recovery_timeout;
1343 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1344 cfs_time_before(now, obd->obd_recovery_end)) {
1345 left = cfs_time_sub(obd->obd_recovery_end, now);
1346 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1348 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1349 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1350 obd->obd_name, (unsigned)left);
1353 static void check_and_start_recovery_timer(struct obd_device *obd)
1355 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1356 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1357 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1360 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1361 obd->obd_recovery_start = cfs_time_current_sec();
1362 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1364 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1367 /* Reset the timer with each new client connection */
1369 * This timer is actually reconnect_timer, which is for making sure
1370 * the total recovery window is at least as big as my reconnect
1371 * attempt timing. So the initial recovery time_out will be set to
1372 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1373 * from client is bigger than this, then the recovery time_out will
1374 * be extended to make sure the client could be reconnected, in the
1375 * process, the timeout from the new client should be ignored.
1379 target_start_and_reset_recovery_timer(struct obd_device *obd,
1380 struct ptlrpc_request *req,
1383 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1385 if (!new_client && service_time)
1386 /* Teach server about old server's estimates, as first guess
1387 * at how long new requests will take. */
1388 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1391 check_and_start_recovery_timer(obd);
1393 /* convert the service time to rpc timeout,
1394 * reuse service_time to limit stack usage */
1395 service_time = at_est2timeout(service_time);
1397 /* We expect other clients to timeout within service_time, then try
1398 * to reconnect, then try the failover server. The max delay between
1399 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1400 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1401 INITIAL_CONNECT_TIMEOUT);
1402 if (service_time > obd->obd_recovery_timeout && !new_client)
1403 reset_recovery_timer(obd, service_time, 0);
1408 /** Health checking routines */
1409 static inline int exp_connect_healthy(struct obd_export *exp)
1411 return (exp->exp_in_recovery);
1414 /** if export done req_replay or has replay in queue */
1415 static inline int exp_req_replay_healthy(struct obd_export *exp)
1417 return (!exp->exp_req_replay_needed ||
1418 cfs_atomic_read(&exp->exp_replay_count) > 0);
1420 /** if export done lock_replay or has replay in queue */
1421 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1423 return (!exp->exp_lock_replay_needed ||
1424 cfs_atomic_read(&exp->exp_replay_count) > 0);
1427 static inline int exp_vbr_healthy(struct obd_export *exp)
1429 return (!exp->exp_vbr_failed);
1432 static inline int exp_finished(struct obd_export *exp)
1434 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1437 /** Checking routines for recovery */
1438 static int check_for_clients(struct obd_device *obd)
1440 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1442 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1443 if (obd->obd_no_conn == 0 &&
1444 obd->obd_connected_clients + obd->obd_stale_clients ==
1445 obd->obd_max_recoverable_clients)
1450 static int check_for_next_transno(struct obd_device *obd)
1452 struct ptlrpc_request *req = NULL;
1453 int wake_up = 0, connected, completed, queue_len;
1454 __u64 next_transno, req_transno;
1456 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1458 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1459 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1460 struct ptlrpc_request, rq_list);
1461 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1466 connected = obd->obd_connected_clients;
1467 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1468 queue_len = obd->obd_requests_queued_for_recovery;
1469 next_transno = obd->obd_next_recovery_transno;
1471 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1472 "req_transno: "LPU64", next_transno: "LPU64"\n",
1473 obd->obd_max_recoverable_clients, connected, completed,
1474 queue_len, req_transno, next_transno);
1476 if (obd->obd_abort_recovery) {
1477 CDEBUG(D_HA, "waking for aborted recovery\n");
1479 } else if (obd->obd_recovery_expired) {
1480 CDEBUG(D_HA, "waking for expired recovery\n");
1482 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1483 CDEBUG(D_HA, "waking for completed recovery\n");
1485 } else if (req_transno == next_transno) {
1486 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1488 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1490 /** handle gaps occured due to lost reply or VBR */
1491 LASSERTF(req_transno >= next_transno,
1492 "req_transno: "LPU64", next_transno: "LPU64"\n",
1493 req_transno, next_transno);
1494 if (req_transno > obd->obd_last_committed &&
1495 !obd->obd_version_recov)
1498 "%s: waking for gap in transno, VBR is %s (skip: "
1499 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1500 ", last_committed: "LPD64")\n",
1501 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1502 next_transno, queue_len, completed, connected,
1503 req_transno, obd->obd_last_committed);
1504 obd->obd_next_recovery_transno = req_transno;
1506 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1507 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1508 " by fail_lock, waking up ("LPD64")\n", next_transno);
1509 obd->obd_next_recovery_transno = req_transno;
1512 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1516 static int check_for_next_lock(struct obd_device *obd)
1520 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1521 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1522 CDEBUG(D_HA, "waking for next lock\n");
1524 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1525 CDEBUG(D_HA, "waking for completed lock replay\n");
1527 } else if (obd->obd_abort_recovery) {
1528 CDEBUG(D_HA, "waking for aborted recovery\n");
1530 } else if (obd->obd_recovery_expired) {
1531 CDEBUG(D_HA, "waking for expired recovery\n");
1534 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1540 * wait for recovery events,
1541 * check its status with help of check_routine
1542 * evict dead clients via health_check
1544 static int target_recovery_overseer(struct obd_device *obd,
1545 int (*check_routine)(struct obd_device *),
1546 int (*health_check)(struct obd_export *))
1548 int abort = 0, expired = 1;
1551 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1552 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1553 abort = obd->obd_abort_recovery;
1554 expired = obd->obd_recovery_expired;
1555 obd->obd_recovery_expired = 0;
1556 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1558 CWARN("recovery is aborted, evict exports in recovery\n");
1559 /** evict exports which didn't finish recovery yet */
1560 class_disconnect_stale_exports(obd, exp_finished);
1561 } else if (expired) {
1562 /** If some clients died being recovered, evict them */
1563 CDEBUG(D_WARNING, "recovery is timed out, evict stale exports\n");
1564 /** evict cexports with no replay in queue, they are stalled */
1565 class_disconnect_stale_exports(obd, health_check);
1566 /** continue with VBR */
1567 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1568 obd->obd_version_recov = 1;
1569 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1571 * reset timer, recovery will proceed with versions now,
1572 * timeout is set just to handle reconnection delays
1574 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1575 /** Wait for recovery events again, after evicting bad clients */
1577 } while (!abort && expired);
1582 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1584 struct ptlrpc_request *req = NULL;
1587 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1588 obd->obd_next_recovery_transno);
1590 if (target_recovery_overseer(obd, check_for_next_transno,
1591 exp_req_replay_healthy)) {
1592 abort_req_replay_queue(obd);
1593 abort_lock_replay_queue(obd);
1596 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1597 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1598 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1599 struct ptlrpc_request, rq_list);
1600 cfs_list_del_init(&req->rq_list);
1601 obd->obd_requests_queued_for_recovery--;
1602 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1604 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1605 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1606 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1607 /** evict exports failed VBR */
1608 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1613 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1615 struct ptlrpc_request *req = NULL;
1617 CDEBUG(D_HA, "Waiting for lock\n");
1618 if (target_recovery_overseer(obd, check_for_next_lock,
1619 exp_lock_replay_healthy))
1620 abort_lock_replay_queue(obd);
1622 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1623 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1624 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1625 struct ptlrpc_request, rq_list);
1626 cfs_list_del_init(&req->rq_list);
1627 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1629 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1630 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1631 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1632 /** evict exports failed VBR */
1633 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1638 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1640 struct ptlrpc_request *req = NULL;
1642 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1643 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1644 req = cfs_list_entry(obd->obd_final_req_queue.next,
1645 struct ptlrpc_request, rq_list);
1646 cfs_list_del_init(&req->rq_list);
1647 if (req->rq_export->exp_in_recovery) {
1648 cfs_spin_lock(&req->rq_export->exp_lock);
1649 req->rq_export->exp_in_recovery = 0;
1650 cfs_spin_unlock(&req->rq_export->exp_lock);
1653 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1657 static int handle_recovery_req(struct ptlrpc_thread *thread,
1658 struct ptlrpc_request *req,
1659 svc_handler_t handler)
1664 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1666 CERROR("Failure to initialize session: %d\n", rc);
1667 GOTO(reqcopy_put, rc);
1670 * export can be evicted during recovery, no need to handle replays for
1671 * it after that, discard such request silently
1673 if (req->rq_export->exp_disconnected)
1674 GOTO(reqcopy_put, rc);
1676 req->rq_recov_session.lc_thread = thread;
1677 lu_context_enter(&req->rq_recov_session);
1678 req->rq_svc_thread = thread;
1679 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1681 /* thread context */
1682 lu_context_enter(&thread->t_env->le_ctx);
1684 lu_context_exit(&thread->t_env->le_ctx);
1686 lu_context_exit(&req->rq_recov_session);
1687 lu_context_fini(&req->rq_recov_session);
1688 /* don't reset timer for final stage */
1689 if (!exp_finished(req->rq_export)) {
1691 * Add request timeout to the recovery time so next request from
1692 * this client may come in recovery time
1694 reset_recovery_timer(class_exp2obd(req->rq_export),
1695 AT_OFF ? obd_timeout :
1696 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1702 static int target_recovery_thread(void *arg)
1704 struct lu_target *lut = arg;
1705 struct obd_device *obd = lut->lut_obd;
1706 struct ptlrpc_request *req;
1707 struct target_recovery_data *trd = &obd->obd_recovery_data;
1708 unsigned long delta;
1709 unsigned long flags;
1711 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1715 cfs_daemonize_ctxt("tgt_recov");
1717 SIGNAL_MASK_LOCK(current, flags);
1718 sigfillset(¤t->blocked);
1720 SIGNAL_MASK_UNLOCK(current, flags);
1722 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1726 thread->t_env = &env;
1727 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1728 env.le_ctx.lc_thread = thread;
1729 thread->t_data = NULL;
1731 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1733 trd->trd_processing_task = cfs_curproc_pid();
1735 obd->obd_recovering = 1;
1736 cfs_complete(&trd->trd_starting);
1738 /* first of all, we have to know the first transno to replay */
1739 if (target_recovery_overseer(obd, check_for_clients,
1740 exp_connect_healthy)) {
1741 abort_req_replay_queue(obd);
1742 abort_lock_replay_queue(obd);
1745 /* next stage: replay requests */
1747 obd->obd_req_replaying = 1;
1748 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1749 cfs_atomic_read(&obd->obd_req_replay_clients),
1750 obd->obd_next_recovery_transno);
1751 while ((req = target_next_replay_req(obd))) {
1752 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1753 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1754 lustre_msg_get_transno(req->rq_reqmsg),
1755 libcfs_nid2str(req->rq_peer.nid));
1756 handle_recovery_req(thread, req,
1757 trd->trd_recovery_handler);
1759 * bz18031: increase next_recovery_transno before
1760 * target_request_copy_put() will drop exp_rpc reference
1762 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1763 obd->obd_next_recovery_transno++;
1764 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1765 target_exp_dequeue_req_replay(req);
1766 target_request_copy_put(req);
1767 obd->obd_replayed_requests++;
1771 * The second stage: replay locks
1773 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1774 cfs_atomic_read(&obd->obd_lock_replay_clients));
1775 while ((req = target_next_replay_lock(obd))) {
1776 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1777 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1778 libcfs_nid2str(req->rq_peer.nid));
1779 handle_recovery_req(thread, req,
1780 trd->trd_recovery_handler);
1781 target_request_copy_put(req);
1782 obd->obd_replayed_locks++;
1786 * The third stage: reply on final pings, at this moment all clients
1787 * must have request in final queue
1789 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1790 /** Update server last boot epoch */
1791 lut_boot_epoch_update(lut);
1792 /* We drop recoverying flag to forward all new requests
1793 * to regular mds_handle() since now */
1794 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1795 obd->obd_recovering = obd->obd_abort_recovery = 0;
1796 target_cancel_recovery_timer(obd);
1797 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1798 while ((req = target_next_final_ping(obd))) {
1799 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1800 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1801 libcfs_nid2str(req->rq_peer.nid));
1802 handle_recovery_req(thread, req,
1803 trd->trd_recovery_handler);
1804 target_request_copy_put(req);
1807 delta = (jiffies - delta) / CFS_HZ;
1808 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1809 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1810 if (delta > OBD_RECOVERY_TIME_SOFT) {
1811 CWARN("too long recovery - read logs\n");
1812 libcfs_debug_dumplog();
1815 target_finish_recovery(obd);
1817 lu_context_fini(&env.le_ctx);
1818 trd->trd_processing_task = 0;
1819 cfs_complete(&trd->trd_finishing);
1823 static int target_start_recovery_thread(struct lu_target *lut,
1824 svc_handler_t handler)
1826 struct obd_device *obd = lut->lut_obd;
1828 struct target_recovery_data *trd = &obd->obd_recovery_data;
1830 memset(trd, 0, sizeof(*trd));
1831 cfs_init_completion(&trd->trd_starting);
1832 cfs_init_completion(&trd->trd_finishing);
1833 trd->trd_recovery_handler = handler;
1835 if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
1836 cfs_wait_for_completion(&trd->trd_starting);
1837 LASSERT(obd->obd_recovering != 0);
1844 void target_stop_recovery_thread(struct obd_device *obd)
1846 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1847 if (obd->obd_recovery_data.trd_processing_task > 0) {
1848 struct target_recovery_data *trd = &obd->obd_recovery_data;
1849 /** recovery can be done but postrecovery is not yet */
1850 if (obd->obd_recovering) {
1851 CERROR("%s: Aborting recovery\n", obd->obd_name);
1852 obd->obd_abort_recovery = 1;
1853 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1855 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1856 cfs_wait_for_completion(&trd->trd_finishing);
1858 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1862 void target_recovery_fini(struct obd_device *obd)
1864 class_disconnect_exports(obd);
1865 target_stop_recovery_thread(obd);
1866 target_cleanup_recovery(obd);
1868 EXPORT_SYMBOL(target_recovery_fini);
1870 static void target_recovery_expired(unsigned long castmeharder)
1872 struct obd_device *obd = (struct obd_device *)castmeharder;
1873 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1874 " after %lds (%d clients connected)\n",
1875 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1876 cfs_time_current_sec()- obd->obd_recovery_start,
1877 obd->obd_connected_clients);
1879 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1880 obd->obd_recovery_expired = 1;
1881 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1882 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1885 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1887 struct obd_device *obd = lut->lut_obd;
1888 if (obd->obd_max_recoverable_clients == 0) {
1889 /** Update server last boot epoch */
1890 lut_boot_epoch_update(lut);
1894 CWARN("RECOVERY: service %s, %d recoverable clients, "
1895 "last_transno "LPU64"\n", obd->obd_name,
1896 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1897 LASSERT(obd->obd_stopping == 0);
1898 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1899 obd->obd_recovery_start = 0;
1900 obd->obd_recovery_end = 0;
1902 /* both values can be get from mount data already */
1903 if (obd->obd_recovery_timeout == 0)
1904 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
1905 if (obd->obd_recovery_time_hard == 0)
1906 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
1907 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1908 target_start_recovery_thread(lut, handler);
1910 EXPORT_SYMBOL(target_recovery_init);
1914 static int target_process_req_flags(struct obd_device *obd,
1915 struct ptlrpc_request *req)
1917 struct obd_export *exp = req->rq_export;
1918 LASSERT(exp != NULL);
1919 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1920 /* client declares he's ready to replay locks */
1921 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1922 if (exp->exp_req_replay_needed) {
1923 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) >
1925 cfs_spin_lock(&exp->exp_lock);
1926 exp->exp_req_replay_needed = 0;
1927 cfs_spin_unlock(&exp->exp_lock);
1928 cfs_atomic_dec(&obd->obd_req_replay_clients);
1930 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1932 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1933 /* client declares he's ready to complete recovery
1934 * so, we put the request on th final queue */
1935 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1936 if (exp->exp_lock_replay_needed) {
1937 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) >
1939 cfs_spin_lock(&exp->exp_lock);
1940 exp->exp_lock_replay_needed = 0;
1941 cfs_spin_unlock(&exp->exp_lock);
1942 cfs_atomic_dec(&obd->obd_lock_replay_clients);
1944 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1950 int target_queue_recovery_request(struct ptlrpc_request *req,
1951 struct obd_device *obd)
1955 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1958 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1959 /* Processing the queue right now, don't re-add. */
1963 target_process_req_flags(obd, req);
1965 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1966 /* client declares he's ready to complete recovery
1967 * so, we put the request on th final queue */
1968 target_request_copy_get(req);
1969 DEBUG_REQ(D_HA, req, "queue final req");
1970 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1971 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1972 if (obd->obd_recovering) {
1973 cfs_list_add_tail(&req->rq_list,
1974 &obd->obd_final_req_queue);
1976 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1977 target_request_copy_put(req);
1978 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1980 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1983 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1984 /* client declares he's ready to replay locks */
1985 target_request_copy_get(req);
1986 DEBUG_REQ(D_HA, req, "queue lock replay req");
1987 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1988 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1989 LASSERT(obd->obd_recovering);
1990 /* usually due to recovery abort */
1991 if (!req->rq_export->exp_in_recovery) {
1992 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1993 target_request_copy_put(req);
1996 LASSERT(req->rq_export->exp_lock_replay_needed);
1997 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1998 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2002 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2003 * (i.e. buflens etc are in my own byte order), but type-dependent
2004 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2007 CFS_INIT_LIST_HEAD(&req->rq_list);
2008 DEBUG_REQ(D_HA, req, "not queueing");
2012 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2014 /* If we're processing the queue, we want don't want to queue this
2017 * Also, if this request has a transno less than the one we're waiting
2018 * for, we should process it now. It could (and currently always will)
2019 * be an open request for a descriptor that was opened some time ago.
2021 * Also, a resent, replayed request that has already been
2022 * handled will pass through here and be processed immediately.
2024 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2025 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2026 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2027 /* Processing the queue right now, don't re-add. */
2028 LASSERT(cfs_list_empty(&req->rq_list));
2029 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2032 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2034 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2037 target_request_copy_get(req);
2038 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2039 LASSERT(obd->obd_recovering);
2040 if (!req->rq_export->exp_in_recovery) {
2041 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2042 target_request_copy_put(req);
2045 LASSERT(req->rq_export->exp_req_replay_needed);
2047 if (target_exp_enqueue_req_replay(req)) {
2048 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2049 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2050 target_request_copy_put(req);
2055 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2056 struct ptlrpc_request *reqiter =
2057 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2059 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2060 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2065 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2067 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2068 "has been claimed by another client");
2069 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2070 target_exp_dequeue_req_replay(req);
2071 target_request_copy_put(req);
2077 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2079 obd->obd_requests_queued_for_recovery++;
2080 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2081 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2086 * Packs current SLV and Limit into \a req.
2088 int target_pack_pool_reply(struct ptlrpc_request *req)
2090 struct obd_device *obd;
2094 * Check that we still have all structures alive as this may
2095 * be some late rpc in shutdown time.
2097 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2098 !exp_connect_lru_resize(req->rq_export))) {
2099 lustre_msg_set_slv(req->rq_repmsg, 0);
2100 lustre_msg_set_limit(req->rq_repmsg, 0);
2105 * OBD is alive here as export is alive, which we checked above.
2107 obd = req->rq_export->exp_obd;
2109 cfs_read_lock(&obd->obd_pool_lock);
2110 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2111 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2112 cfs_read_unlock(&obd->obd_pool_lock);
2117 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2119 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2120 DEBUG_REQ(D_ERROR, req, "dropping reply");
2125 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2126 req->rq_status = rc;
2127 return (ptlrpc_send_error(req, 1));
2129 DEBUG_REQ(D_NET, req, "sending reply");
2132 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2135 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2138 struct ptlrpc_reply_state *rs;
2139 struct obd_device *obd;
2140 struct obd_export *exp;
2141 struct ptlrpc_service *svc;
2144 if (req->rq_no_reply) {
2149 svc = req->rq_rqbd->rqbd_service;
2150 rs = req->rq_reply_state;
2151 if (rs == NULL || !rs->rs_difficult) {
2153 target_send_reply_msg (req, rc, fail_id);
2158 /* must be an export if locks saved */
2159 LASSERT (req->rq_export != NULL);
2160 /* req/reply consistent */
2161 LASSERT (rs->rs_service == svc);
2164 LASSERT (!rs->rs_scheduled);
2165 LASSERT (!rs->rs_scheduled_ever);
2166 LASSERT (!rs->rs_handled);
2167 LASSERT (!rs->rs_on_net);
2168 LASSERT (rs->rs_export == NULL);
2169 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2170 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2172 exp = class_export_get (req->rq_export);
2175 /* disable reply scheduling while I'm setting up */
2176 rs->rs_scheduled = 1;
2178 rs->rs_xid = req->rq_xid;
2179 rs->rs_transno = req->rq_transno;
2180 rs->rs_export = exp;
2181 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2183 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2184 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2185 rs->rs_transno, exp->exp_last_committed);
2186 if (rs->rs_transno > exp->exp_last_committed) {
2187 /* not committed already */
2188 cfs_list_add_tail(&rs->rs_obd_list,
2189 &exp->exp_uncommitted_replies);
2191 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2193 cfs_spin_lock(&exp->exp_lock);
2194 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2195 cfs_spin_unlock(&exp->exp_lock);
2197 netrc = target_send_reply_msg (req, rc, fail_id);
2199 cfs_spin_lock(&svc->srv_lock);
2201 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2204 /* error sending: reply is off the net. Also we need +1
2205 * reply ref until ptlrpc_handle_rs() is done
2206 * with the reply state (if the send was successful, there
2207 * would have been +1 ref for the net, which
2208 * reply_out_callback leaves alone) */
2210 ptlrpc_rs_addref(rs);
2211 cfs_atomic_inc (&svc->srv_outstanding_replies);
2214 cfs_spin_lock(&rs->rs_lock);
2215 if (rs->rs_transno <= exp->exp_last_committed ||
2216 (!rs->rs_on_net && !rs->rs_no_ack) ||
2217 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2218 cfs_list_empty(&rs->rs_obd_list)) {
2219 CDEBUG(D_HA, "Schedule reply immediately\n");
2220 ptlrpc_dispatch_difficult_reply(rs);
2222 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2223 rs->rs_scheduled = 0; /* allow notifier to schedule */
2225 cfs_spin_unlock(&rs->rs_lock);
2226 cfs_spin_unlock(&svc->srv_lock);
2230 int target_handle_ping(struct ptlrpc_request *req)
2232 obd_ping(req->rq_export);
2233 return req_capsule_server_pack(&req->rq_pill);
2236 void target_committed_to_req(struct ptlrpc_request *req)
2238 struct obd_export *exp = req->rq_export;
2240 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2241 lustre_msg_set_last_committed(req->rq_repmsg,
2242 exp->exp_last_committed);
2244 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2245 "%d)", exp->exp_obd->obd_no_transno,
2246 req->rq_repmsg == NULL);
2248 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2249 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2251 EXPORT_SYMBOL(target_committed_to_req);
2253 int target_handle_qc_callback(struct ptlrpc_request *req)
2255 struct obd_quotactl *oqctl;
2256 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2258 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2259 if (oqctl == NULL) {
2260 CERROR("Can't unpack obd_quotactl\n");
2264 cli->cl_qchk_stat = oqctl->qc_stat;
2269 #ifdef HAVE_QUOTA_SUPPORT
2270 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2273 struct obd_device *obd = req->rq_export->exp_obd;
2274 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2275 struct obd_device_target *obt;
2276 struct lustre_quota_ctxt *qctxt;
2277 struct qunit_data *qdata = NULL;
2281 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2284 rc = req_capsule_server_pack(&req->rq_pill);
2286 CERROR("packing reply failed!: rc = %d\n", rc);
2290 LASSERT(req->rq_export);
2292 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2293 if (IS_ERR(qdata)) {
2294 rc = PTR_ERR(qdata);
2295 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2296 req->rq_status = rc;
2300 /* we use the observer */
2301 if (obd_pin_observer(obd, &lov_obd) ||
2302 obd_pin_observer(lov_obd, &master_obd)) {
2303 CERROR("Can't find the observer, it is recovering\n");
2304 req->rq_status = -EAGAIN;
2308 obt = &master_obd->u.obt;
2309 qctxt = &obt->obt_qctxt;
2311 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2312 /* quota_type has not been processed yet, return EAGAIN
2313 * until we know whether or not quotas are supposed to
2315 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2317 req->rq_status = -EAGAIN;
2321 cfs_down_read(&obt->obt_rwsem);
2322 if (qctxt->lqc_lqs_hash == NULL) {
2323 cfs_up_read(&obt->obt_rwsem);
2324 /* quota_type has not been processed yet, return EAGAIN
2325 * until we know whether or not quotas are supposed to
2327 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2329 req->rq_status = -EAGAIN;
2333 LASSERT(qctxt->lqc_handler);
2334 rc = qctxt->lqc_handler(master_obd, qdata,
2335 lustre_msg_get_opc(req->rq_reqmsg));
2336 cfs_up_read(&obt->obt_rwsem);
2337 if (rc && rc != -EDQUOT)
2338 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2339 "dqacq/dqrel failed! (rc:%d)\n", rc);
2340 req->rq_status = rc;
2342 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2344 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2348 /* Block the quota req. b=14840 */
2349 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2354 obd_unpin_observer(lov_obd);
2356 obd_unpin_observer(obd);
2358 rc = ptlrpc_reply(req);
2362 #endif /* !__KERNEL__ */
2364 #endif /* HAVE_QUOTA_SUPPORT */
2366 ldlm_mode_t lck_compat_array[] = {
2367 [LCK_EX] LCK_COMPAT_EX,
2368 [LCK_PW] LCK_COMPAT_PW,
2369 [LCK_PR] LCK_COMPAT_PR,
2370 [LCK_CW] LCK_COMPAT_CW,
2371 [LCK_CR] LCK_COMPAT_CR,
2372 [LCK_NL] LCK_COMPAT_NL,
2373 [LCK_GROUP] LCK_COMPAT_GROUP,
2374 [LCK_COS] LCK_COMPAT_COS,
2378 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2379 * not escape to the user level.
2381 int ldlm_error2errno(ldlm_error_t error)
2389 case ELDLM_LOCK_CHANGED:
2392 case ELDLM_LOCK_ABORTED:
2395 case ELDLM_LOCK_REPLACED:
2398 case ELDLM_NO_LOCK_DATA:
2401 case ELDLM_NAMESPACE_EXISTS:
2404 case ELDLM_BAD_NAMESPACE:
2408 if (((int)error) < 0) /* cast to signed type */
2409 result = error; /* as ldlm_error_t can be unsigned */
2411 CERROR("Invalid DLM result code: %i\n", error);
2417 EXPORT_SYMBOL(ldlm_error2errno);
2420 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2422 ldlm_error_t ldlm_errno2error(int err_no)
2431 error = ELDLM_LOCK_CHANGED;
2434 error = ELDLM_LOCK_ABORTED;
2437 error = ELDLM_LOCK_REPLACED;
2440 error = ELDLM_NO_LOCK_DATA;
2443 error = ELDLM_NAMESPACE_EXISTS;
2446 error = ELDLM_BAD_NAMESPACE;
2453 EXPORT_SYMBOL(ldlm_errno2error);
2455 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2456 void ldlm_dump_export_locks(struct obd_export *exp)
2458 cfs_spin_lock(&exp->exp_locks_list_guard);
2459 if (!cfs_list_empty(&exp->exp_locks_list)) {
2460 struct ldlm_lock *lock;
2462 CERROR("dumping locks for export %p,"
2463 "ignore if the unmount doesn't hang\n", exp);
2464 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2465 ldlm_lock_dump(D_ERROR, lock, 0);
2467 cfs_spin_unlock(&exp->exp_locks_list_guard);