1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_mds.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: if non-zero, move the selected to the list head
55 * @create: if zero, only search in existed connections
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 cfs_spin_lock(&imp->imp_lock);
84 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 cfs_list_del(&item->oic_item);
88 cfs_list_add(&item->oic_item,
90 item->oic_last_attempt = 0;
92 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? ", moved to head" : ""));
95 cfs_spin_unlock(&imp->imp_lock);
96 GOTO(out_free, rc = 0);
101 imp_conn->oic_conn = ptlrpc_conn;
102 imp_conn->oic_uuid = *uuid;
103 imp_conn->oic_last_attempt = 0;
105 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107 cfs_list_add_tail(&imp_conn->oic_item,
108 &imp->imp_conn_list);
109 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
110 imp, imp->imp_obd->obd_name, uuid->uuid,
111 (priority ? "head" : "tail"));
113 cfs_spin_unlock(&imp->imp_lock);
114 GOTO(out_free, rc = -ENOENT);
117 cfs_spin_unlock(&imp->imp_lock);
121 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 ptlrpc_connection_put(ptlrpc_conn);
127 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 return import_set_conn(imp, uuid, 1, 0);
132 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
135 return import_set_conn(imp, uuid, priority, 1);
138 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
140 struct obd_import_conn *imp_conn;
141 struct obd_export *dlmexp;
145 cfs_spin_lock(&imp->imp_lock);
146 if (cfs_list_empty(&imp->imp_conn_list)) {
147 LASSERT(!imp->imp_connection);
151 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
152 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
154 LASSERT(imp_conn->oic_conn);
156 /* is current conn? */
157 if (imp_conn == imp->imp_conn_current) {
158 LASSERT(imp_conn->oic_conn == imp->imp_connection);
160 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
161 imp->imp_state != LUSTRE_IMP_DISCON) {
162 CERROR("can't remove current connection\n");
163 GOTO(out, rc = -EBUSY);
166 ptlrpc_connection_put(imp->imp_connection);
167 imp->imp_connection = NULL;
169 dlmexp = class_conn2export(&imp->imp_dlm_handle);
170 if (dlmexp && dlmexp->exp_connection) {
171 LASSERT(dlmexp->exp_connection ==
173 ptlrpc_connection_put(dlmexp->exp_connection);
174 dlmexp->exp_connection = NULL;
178 cfs_list_del(&imp_conn->oic_item);
179 ptlrpc_connection_put(imp_conn->oic_conn);
180 OBD_FREE(imp_conn, sizeof(*imp_conn));
181 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182 imp, imp->imp_obd->obd_name, uuid->uuid);
187 cfs_spin_unlock(&imp->imp_lock);
189 CERROR("connection %s not found\n", uuid->uuid);
193 void client_destroy_import(struct obd_import *imp)
195 /* drop security policy instance after all rpc finished/aborted
196 * to let all busy contexts be released. */
197 class_import_get(imp);
198 class_destroy_import(imp);
199 sptlrpc_import_sec_put(imp);
200 class_import_put(imp);
203 /* configure an RPC client OBD device
208 * 3 - inactive-on-startup
210 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
212 struct client_obd *cli = &obddev->u.cli;
213 struct obd_import *imp;
214 struct obd_uuid server_uuid;
215 int rq_portal, rp_portal, connect_op;
216 char *name = obddev->obd_type->typ_name;
217 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
221 /* In a more perfect world, we would hang a ptlrpc_client off of
222 * obd_type and just use the values from there. */
223 if (!strcmp(name, LUSTRE_OSC_NAME)) {
224 rq_portal = OST_REQUEST_PORTAL;
225 rp_portal = OSC_REPLY_PORTAL;
226 connect_op = OST_CONNECT;
227 cli->cl_sp_me = LUSTRE_SP_CLI;
228 cli->cl_sp_to = LUSTRE_SP_OST;
229 ns_type = LDLM_NS_TYPE_OSC;
231 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
232 rq_portal = MDS_REQUEST_PORTAL;
233 rp_portal = MDC_REPLY_PORTAL;
234 connect_op = MDS_CONNECT;
235 cli->cl_sp_me = LUSTRE_SP_CLI;
236 cli->cl_sp_to = LUSTRE_SP_MDT;
237 ns_type = LDLM_NS_TYPE_MDC;
239 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
240 rq_portal = MGS_REQUEST_PORTAL;
241 rp_portal = MGC_REPLY_PORTAL;
242 connect_op = MGS_CONNECT;
243 cli->cl_sp_me = LUSTRE_SP_MGC;
244 cli->cl_sp_to = LUSTRE_SP_MGS;
245 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
246 ns_type = LDLM_NS_TYPE_MGC;
249 CERROR("unknown client OBD type \"%s\", can't setup\n",
254 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
255 CERROR("requires a TARGET UUID\n");
259 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
260 CERROR("client UUID must be less than 38 characters\n");
264 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
265 CERROR("setup requires a SERVER UUID\n");
269 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
270 CERROR("target UUID must be less than 38 characters\n");
274 cfs_init_rwsem(&cli->cl_sem);
275 cfs_sema_init(&cli->cl_mgc_sem, 1);
276 cli->cl_conn_count = 0;
277 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
278 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
279 sizeof(server_uuid)));
282 cli->cl_avail_grant = 0;
283 /* FIXME: should limit this for the sum of all cl_dirty_max */
284 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
285 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
286 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
287 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
288 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
289 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
290 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
291 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
292 CFS_INIT_LIST_HEAD(&cli->cl_loi_sync_fs_list);
293 client_obd_list_lock_init(&cli->cl_loi_list_lock);
294 cli->cl_r_in_flight = 0;
295 cli->cl_w_in_flight = 0;
297 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
298 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
299 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
300 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
301 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
302 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
303 cfs_waitq_init(&cli->cl_destroy_waitq);
304 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
305 #ifdef ENABLE_CHECKSUM
306 /* Turn on checksumming by default. */
307 cli->cl_checksum = 1;
309 * The supported checksum types will be worked out at connect time
310 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
313 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
315 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
317 /* This value may be changed at connect time in
318 ptlrpc_connect_interpret. */
319 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
320 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
322 if (!strcmp(name, LUSTRE_MDC_NAME)) {
323 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
324 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
325 cli->cl_max_rpcs_in_flight = 2;
326 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
327 cli->cl_max_rpcs_in_flight = 3;
328 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
329 cli->cl_max_rpcs_in_flight = 4;
331 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
336 CERROR("ldlm_get_ref failed: %d\n", rc);
340 ptlrpc_init_client(rq_portal, rp_portal, name,
341 &obddev->obd_ldlm_client);
343 imp = class_new_import(obddev);
345 GOTO(err_ldlm, rc = -ENOENT);
346 imp->imp_client = &obddev->obd_ldlm_client;
347 imp->imp_connect_op = connect_op;
348 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
349 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
350 LUSTRE_CFG_BUFLEN(lcfg, 1));
351 class_import_put(imp);
353 rc = client_import_add_conn(imp, &server_uuid, 1);
355 CERROR("can't add initial connection\n");
356 GOTO(err_import, rc);
359 cli->cl_import = imp;
360 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
361 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
362 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
364 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
365 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
366 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
367 name, obddev->obd_name,
368 cli->cl_target_uuid.uuid);
369 cfs_spin_lock(&imp->imp_lock);
370 imp->imp_deactive = 1;
371 cfs_spin_unlock(&imp->imp_lock);
375 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
376 LDLM_NAMESPACE_CLIENT,
377 LDLM_NAMESPACE_GREEDY,
379 if (obddev->obd_namespace == NULL) {
380 CERROR("Unable to create client namespace - %s\n",
382 GOTO(err_import, rc = -ENOMEM);
385 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
390 class_destroy_import(imp);
398 int client_obd_cleanup(struct obd_device *obddev)
402 ldlm_namespace_free_post(obddev->obd_namespace);
403 obddev->obd_namespace = NULL;
409 /* ->o_connect() method for client side (OSC and MDC and MGC) */
410 int client_connect_import(const struct lu_env *env,
411 struct obd_export **exp,
412 struct obd_device *obd, struct obd_uuid *cluuid,
413 struct obd_connect_data *data, void *localdata)
415 struct client_obd *cli = &obd->u.cli;
416 struct obd_import *imp = cli->cl_import;
417 struct obd_connect_data *ocd;
418 struct lustre_handle conn = { 0 };
423 cfs_down_write(&cli->cl_sem);
424 if (cli->cl_conn_count > 0 )
425 GOTO(out_sem, rc = -EALREADY);
427 rc = class_connect(&conn, obd, cluuid);
431 cli->cl_conn_count++;
432 *exp = class_conn2export(&conn);
434 LASSERT(obd->obd_namespace);
436 imp->imp_dlm_handle = conn;
437 rc = ptlrpc_init_import(imp);
441 ocd = &imp->imp_connect_data;
444 imp->imp_connect_flags_orig = data->ocd_connect_flags;
447 rc = ptlrpc_connect_import(imp, NULL);
449 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
452 LASSERT((*exp)->exp_connection);
455 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
456 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
457 data->ocd_connect_flags, ocd->ocd_connect_flags);
458 data->ocd_connect_flags = ocd->ocd_connect_flags;
461 ptlrpc_pinger_add_import(imp);
467 cli->cl_conn_count--;
468 class_disconnect(*exp);
472 cfs_up_write(&cli->cl_sem);
477 int client_disconnect_export(struct obd_export *exp)
479 struct obd_device *obd = class_exp2obd(exp);
480 struct client_obd *cli;
481 struct obd_import *imp;
486 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
487 exp, exp ? exp->exp_handle.h_cookie : -1);
492 imp = cli->cl_import;
494 cfs_down_write(&cli->cl_sem);
495 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
498 if (!cli->cl_conn_count) {
499 CERROR("disconnecting disconnected device (%s)\n",
501 GOTO(out_disconnect, rc = -EINVAL);
504 cli->cl_conn_count--;
505 if (cli->cl_conn_count)
506 GOTO(out_disconnect, rc = 0);
508 /* Mark import deactivated now, so we don't try to reconnect if any
509 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
510 * fully deactivate the import, or that would drop all requests. */
511 cfs_spin_lock(&imp->imp_lock);
512 imp->imp_deactive = 1;
513 cfs_spin_unlock(&imp->imp_lock);
515 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
516 * delete it regardless. (It's safe to delete an import that was
518 (void)ptlrpc_pinger_del_import(imp);
520 if (obd->obd_namespace != NULL) {
521 /* obd_force == local only */
522 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
523 obd->obd_force ? LCF_LOCAL : 0, NULL);
524 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
528 * there's no need to hold sem during disconnecting an import,
529 * and actually it may cause deadlock in gss.
531 cfs_up_write(&cli->cl_sem);
532 rc = ptlrpc_disconnect_import(imp, 0);
533 cfs_down_write(&cli->cl_sem);
535 ptlrpc_invalidate_import(imp);
537 if (imp->imp_rq_pool) {
538 ptlrpc_free_rq_pool(imp->imp_rq_pool);
539 imp->imp_rq_pool = NULL;
541 client_destroy_import(imp);
542 cli->cl_import = NULL;
547 /* use server style - class_disconnect should be always called for
549 err = class_disconnect(exp);
553 cfs_up_write(&cli->cl_sem);
558 int server_disconnect_export(struct obd_export *exp)
563 /* Disconnect early so that clients can't keep using export */
564 rc = class_disconnect(exp);
565 /* close import for avoid sending any requests */
566 if (exp->exp_imp_reverse)
567 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
569 if (exp->exp_obd->obd_namespace != NULL)
570 ldlm_cancel_locks_for_export(exp);
572 /* complete all outstanding replies */
573 cfs_spin_lock(&exp->exp_lock);
574 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
575 struct ptlrpc_reply_state *rs =
576 cfs_list_entry(exp->exp_outstanding_replies.next,
577 struct ptlrpc_reply_state, rs_exp_list);
578 struct ptlrpc_service *svc = rs->rs_service;
580 cfs_spin_lock(&svc->srv_rs_lock);
581 cfs_list_del_init(&rs->rs_exp_list);
582 cfs_spin_lock(&rs->rs_lock);
583 ptlrpc_schedule_difficult_reply(rs);
584 cfs_spin_unlock(&rs->rs_lock);
585 cfs_spin_unlock(&svc->srv_rs_lock);
587 cfs_spin_unlock(&exp->exp_lock);
592 /* --------------------------------------------------------------------------
593 * from old lib/target.c
594 * -------------------------------------------------------------------------- */
596 static int target_handle_reconnect(struct lustre_handle *conn,
597 struct obd_export *exp,
598 struct obd_uuid *cluuid)
601 if (exp->exp_connection && exp->exp_imp_reverse) {
602 struct lustre_handle *hdl;
603 hdl = &exp->exp_imp_reverse->imp_remote_handle;
604 /* Might be a re-connect after a partition. */
605 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
606 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
608 conn->cookie = exp->exp_handle.h_cookie;
609 /* target_handle_connect() treats EALREADY and
610 * -EALREADY differently. EALREADY means we are
611 * doing a valid reconnect from the same client. */
614 CERROR("%s reconnecting from %s, "
615 "handle mismatch (ours "LPX64", theirs "
616 LPX64")\n", cluuid->uuid,
617 exp->exp_connection->c_remote_uuid.uuid,
618 hdl->cookie, conn->cookie);
619 memset(conn, 0, sizeof *conn);
620 /* target_handle_connect() treats EALREADY and
621 * -EALREADY differently. -EALREADY is an error
622 * (same UUID, different handle). */
627 conn->cookie = exp->exp_handle.h_cookie;
628 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
629 cluuid->uuid, exp, conn->cookie);
633 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
636 struct obd_export *exp = cb_data;
638 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
639 obd->obd_name, exp->exp_client_uuid.uuid);
641 cfs_spin_lock(&exp->exp_lock);
642 exp->exp_need_sync = 0;
643 cfs_spin_unlock(&exp->exp_lock);
644 class_export_cb_put(exp);
646 EXPORT_SYMBOL(target_client_add_cb);
649 target_start_and_reset_recovery_timer(struct obd_device *obd,
650 struct ptlrpc_request *req,
653 int target_handle_connect(struct ptlrpc_request *req)
655 struct obd_device *target, *targref = NULL;
656 struct obd_export *export = NULL;
657 struct obd_import *revimp;
658 struct lustre_handle conn;
659 struct lustre_handle *tmp;
660 struct obd_uuid tgtuuid;
661 struct obd_uuid cluuid;
662 struct obd_uuid remote_uuid;
666 struct obd_connect_data *data, *tmpdata;
667 lnet_nid_t *client_nid = NULL;
670 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
672 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
674 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
675 GOTO(out, rc = -EINVAL);
678 obd_str2uuid(&tgtuuid, str);
679 target = class_uuid2obd(&tgtuuid);
681 target = class_name2obd(str);
683 if (!target || target->obd_stopping || !target->obd_set_up) {
684 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
685 " for connect (%s)\n", str,
686 !target ? "no target" :
687 (target->obd_stopping ? "stopping" :
689 GOTO(out, rc = -ENODEV);
692 if (target->obd_no_conn) {
693 LCONSOLE_WARN("%s: temporarily refusing client connection "
694 "from %s\n", target->obd_name,
695 libcfs_nid2str(req->rq_peer.nid));
696 GOTO(out, rc = -EAGAIN);
699 /* Make sure the target isn't cleaned up while we're here. Yes,
700 there's still a race between the above check and our incref here.
701 Really, class_uuid2obd should take the ref. */
702 targref = class_incref(target, __FUNCTION__, cfs_current());
705 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
707 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
708 GOTO(out, rc = -EINVAL);
711 obd_str2uuid(&cluuid, str);
713 /* XXX extract a nettype and format accordingly */
714 switch (sizeof(lnet_nid_t)) {
715 /* NB the casts only avoid compiler warnings */
717 snprintf(remote_uuid.uuid, sizeof remote_uuid,
718 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
721 snprintf(remote_uuid.uuid, sizeof remote_uuid,
722 "NET_%x_UUID", (__u32)req->rq_peer.nid);
728 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
730 GOTO(out, rc = -EPROTO);
734 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
736 GOTO(out, rc = -EPROTO);
738 rc = req_capsule_server_pack(&req->rq_pill);
742 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
744 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
745 "libclient connection attempt");
746 GOTO(out, rc = -EPROTO);
747 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
748 LUSTRE_VERSION_ALLOWED_OFFSET ||
749 data->ocd_version > LUSTRE_VERSION_CODE +
750 LUSTRE_VERSION_ALLOWED_OFFSET) {
751 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
752 "libclient connection attempt",
753 data->ocd_version < LUSTRE_VERSION_CODE ?
755 OBD_OCD_VERSION_MAJOR(data->ocd_version),
756 OBD_OCD_VERSION_MINOR(data->ocd_version),
757 OBD_OCD_VERSION_PATCH(data->ocd_version),
758 OBD_OCD_VERSION_FIX(data->ocd_version));
759 data = req_capsule_server_sized_get(&req->rq_pill,
761 offsetof(typeof(*data), ocd_version) +
762 sizeof(data->ocd_version));
764 data->ocd_connect_flags = OBD_CONNECT_VERSION;
765 data->ocd_version = LUSTRE_VERSION_CODE;
767 GOTO(out, rc = -EPROTO);
771 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
772 (data->ocd_connect_flags & OBD_CONNECT_MDS))
775 /* lctl gets a backstage, all-access pass. */
776 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
777 goto dont_check_exports;
779 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
783 /* we've found an export in the hash */
784 if (export->exp_connecting) { /* bug 9635, et. al. */
785 CWARN("%s: exp %p already connecting\n",
786 export->exp_obd->obd_name, export);
787 class_export_put(export);
790 } else if (mds_conn && export->exp_connection) {
791 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
792 /* mds reconnected after failover */
793 CWARN("%s: received MDS connection from NID %s,"
794 " removing former export from NID %s\n",
795 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
796 libcfs_nid2str(export->exp_connection->c_peer.nid));
798 /* new mds connection from the same nid */
799 CWARN("%s: received new MDS connection from NID %s,"
800 " removing former export from same NID\n",
801 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
802 class_fail_export(export);
803 class_export_put(export);
806 } else if (export->exp_connection != NULL &&
807 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
808 (lustre_msg_get_op_flags(req->rq_reqmsg) &
809 MSG_CONNECT_INITIAL)) {
810 /* in mds failover we have static uuid but nid can be
812 CWARN("%s: cookie %s seen on new NID %s when "
813 "existing NID %s is already connected\n",
814 target->obd_name, cluuid.uuid,
815 libcfs_nid2str(req->rq_peer.nid),
816 libcfs_nid2str(export->exp_connection->c_peer.nid));
818 class_export_put(export);
821 cfs_spin_lock(&export->exp_lock);
822 export->exp_connecting = 1;
823 cfs_spin_unlock(&export->exp_lock);
824 class_export_put(export);
825 LASSERT(export->exp_obd == target);
827 rc = target_handle_reconnect(&conn, export, &cluuid);
830 /* If we found an export, we already unlocked. */
833 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
834 } else if (req->rq_export == NULL &&
835 cfs_atomic_read(&export->exp_rpc_count) > 0) {
836 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
837 target->obd_name, cluuid.uuid,
838 libcfs_nid2str(req->rq_peer.nid),
839 export, cfs_atomic_read(&export->exp_refcount));
840 GOTO(out, rc = -EBUSY);
841 } else if (req->rq_export != NULL &&
842 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
843 /* the current connect rpc has increased exp_rpc_count */
844 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
845 target->obd_name, cluuid.uuid,
846 libcfs_nid2str(req->rq_peer.nid),
847 export, cfs_atomic_read(&export->exp_rpc_count) - 1);
848 cfs_spin_lock(&export->exp_lock);
849 if (req->rq_export->exp_conn_cnt <
850 lustre_msg_get_conn_cnt(req->rq_reqmsg))
851 /* try to abort active requests */
852 req->rq_export->exp_abort_active_req = 1;
853 cfs_spin_unlock(&export->exp_lock);
854 GOTO(out, rc = -EBUSY);
855 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
856 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
857 "cookies not random?\n", target->obd_name,
858 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
859 GOTO(out, rc = -EALREADY);
861 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
868 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
869 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
870 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
871 export, (long)cfs_time_current_sec(),
872 export ? (long)export->exp_last_request_time : 0);
874 /* Tell the client if we're in recovery. */
875 if (target->obd_recovering) {
876 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
877 /* If this is the first time a client connects,
878 reset the recovery timer */
880 target_start_and_reset_recovery_timer(target, req,
884 /* We want to handle EALREADY but *not* -EALREADY from
885 * target_handle_reconnect(), return reconnection state in a flag */
886 if (rc == EALREADY) {
887 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
893 /* Tell the client if we support replayable requests */
894 if (target->obd_replayable)
895 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
896 client_nid = &req->rq_peer.nid;
898 if (export == NULL) {
899 if (target->obd_recovering) {
902 t = cfs_timer_deadline(&target->obd_recovery_timer);
903 t = cfs_time_sub(t, cfs_time_current());
904 CERROR("%s: denying connection for new client %s (%s): "
905 "%d clients in recovery for "CFS_TIME_T"s\n",
907 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
908 cfs_atomic_read(&target-> \
909 obd_lock_replay_clients),
910 cfs_duration_sec(t));
914 rc = obd_connect(req->rq_svc_thread->t_env,
915 &export, target, &cluuid, data,
918 conn.cookie = export->exp_handle.h_cookie;
921 rc = obd_reconnect(req->rq_svc_thread->t_env,
922 export, target, &cluuid, data, client_nid);
924 /* prevous done via class_conn2export */
925 class_export_get(export);
929 /* Return only the parts of obd_connect_data that we understand, so the
930 * client knows that we don't understand the rest. */
932 tmpdata = req_capsule_server_get(&req->rq_pill,
934 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
938 /* If all else goes well, this is our RPC return code. */
941 lustre_msg_set_handle(req->rq_repmsg, &conn);
943 /* If the client and the server are the same node, we will already
944 * have an export that really points to the client's DLM export,
945 * because we have a shared handles table.
947 * XXX this will go away when shaver stops sending the "connect" handle
948 * in the real "remote handle" field of the request --phik 24 Apr 2003
950 if (req->rq_export != NULL)
951 class_export_put(req->rq_export);
953 req->rq_export = export;
955 cfs_spin_lock(&export->exp_lock);
956 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
957 cfs_spin_unlock(&export->exp_lock);
958 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
959 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
960 export->exp_conn_cnt,
961 lustre_msg_get_conn_cnt(req->rq_reqmsg));
963 GOTO(out, rc = -EALREADY);
965 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
966 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
967 export->exp_abort_active_req = 0;
969 /* request from liblustre? Don't evict it for not pinging. */
970 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
971 export->exp_libclient = 1;
972 cfs_spin_unlock(&export->exp_lock);
974 cfs_spin_lock(&target->obd_dev_lock);
975 cfs_list_del_init(&export->exp_obd_chain_timed);
976 cfs_spin_unlock(&target->obd_dev_lock);
978 cfs_spin_unlock(&export->exp_lock);
981 if (export->exp_connection != NULL) {
982 /* Check to see if connection came from another NID */
983 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
984 !cfs_hlist_unhashed(&export->exp_nid_hash))
985 cfs_hash_del(export->exp_obd->obd_nid_hash,
986 &export->exp_connection->c_peer.nid,
987 &export->exp_nid_hash);
989 ptlrpc_connection_put(export->exp_connection);
992 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
995 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
996 cfs_hash_add(export->exp_obd->obd_nid_hash,
997 &export->exp_connection->c_peer.nid,
998 &export->exp_nid_hash);
1001 cfs_spin_lock(&target->obd_recovery_task_lock);
1002 if (target->obd_recovering && !export->exp_in_recovery) {
1003 cfs_spin_lock(&export->exp_lock);
1004 export->exp_in_recovery = 1;
1005 export->exp_req_replay_needed = 1;
1006 export->exp_lock_replay_needed = 1;
1007 cfs_spin_unlock(&export->exp_lock);
1008 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1009 && (data->ocd_transno == 0))
1010 CWARN("Connect with zero transno!\n");
1012 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1013 && data->ocd_transno < target->obd_next_recovery_transno)
1014 target->obd_next_recovery_transno = data->ocd_transno;
1015 target->obd_connected_clients++;
1016 cfs_atomic_inc(&target->obd_req_replay_clients);
1017 cfs_atomic_inc(&target->obd_lock_replay_clients);
1018 if (target->obd_connected_clients ==
1019 target->obd_max_recoverable_clients)
1020 cfs_waitq_signal(&target->obd_next_transno_waitq);
1022 cfs_spin_unlock(&target->obd_recovery_task_lock);
1023 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1026 if (export->exp_imp_reverse != NULL) {
1027 /* destroyed import can be still referenced in ctxt */
1028 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1029 KEY_REVIMP_UPD, 0, NULL, NULL);
1031 client_destroy_import(export->exp_imp_reverse);
1034 /* for the rest part, we return -ENOTCONN in case of errors
1035 * in order to let client initialize connection again.
1037 revimp = export->exp_imp_reverse = class_new_import(target);
1039 CERROR("fail to alloc new reverse import.\n");
1040 GOTO(out, rc = -ENOTCONN);
1043 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1044 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1045 revimp->imp_remote_handle = conn;
1046 revimp->imp_dlm_fake = 1;
1047 revimp->imp_state = LUSTRE_IMP_FULL;
1049 /* unknown versions will be caught in
1050 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1051 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1053 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1054 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1055 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1057 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1059 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1060 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1061 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1063 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1065 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1067 CERROR("Failed to get sec for reverse import: %d\n", rc);
1068 export->exp_imp_reverse = NULL;
1069 class_destroy_import(revimp);
1072 class_import_put(revimp);
1075 cfs_spin_lock(&export->exp_lock);
1076 export->exp_connecting = 0;
1077 cfs_spin_unlock(&export->exp_lock);
1080 class_decref(targref, __FUNCTION__, cfs_current());
1082 req->rq_status = rc;
1086 int target_handle_disconnect(struct ptlrpc_request *req)
1091 rc = req_capsule_server_pack(&req->rq_pill);
1095 /* keep the rq_export around so we can send the reply */
1096 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1101 void target_destroy_export(struct obd_export *exp)
1103 /* exports created from last_rcvd data, and "fake"
1104 exports created by lctl don't have an import */
1105 if (exp->exp_imp_reverse != NULL)
1106 client_destroy_import(exp->exp_imp_reverse);
1108 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1109 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1110 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1111 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1115 * Recovery functions
1117 static void target_request_copy_get(struct ptlrpc_request *req)
1119 class_export_rpc_get(req->rq_export);
1120 LASSERT(cfs_list_empty(&req->rq_list));
1121 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1123 /* increase refcount to keep request in queue */
1124 cfs_atomic_inc(&req->rq_refcount);
1125 /** let export know it has replays to be handled */
1126 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1127 /* release service thread while request is queued
1128 * we are moving the request from active processing
1129 * to waiting on the replay queue */
1130 ptlrpc_server_active_request_dec(req);
1133 static void target_request_copy_put(struct ptlrpc_request *req)
1135 LASSERT(cfs_list_empty(&req->rq_replay_list));
1136 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1138 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1139 class_export_rpc_put(req->rq_export);
1140 /* ptlrpc_server_drop_request() assumes the request is active */
1141 ptlrpc_server_active_request_inc(req);
1142 ptlrpc_server_drop_request(req);
1145 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1147 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1148 struct obd_export *exp = req->rq_export;
1149 struct ptlrpc_request *reqiter;
1154 cfs_spin_lock(&exp->exp_lock);
1155 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1157 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1164 /* we expect it with RESENT and REPLAY flags */
1165 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1166 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1167 CERROR("invalid flags %x of resent replay\n",
1168 lustre_msg_get_flags(req->rq_reqmsg));
1170 cfs_list_add_tail(&req->rq_replay_list,
1171 &exp->exp_req_replay_queue);
1174 cfs_spin_unlock(&exp->exp_lock);
1178 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1180 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1181 LASSERT(req->rq_export);
1183 cfs_spin_lock(&req->rq_export->exp_lock);
1184 cfs_list_del_init(&req->rq_replay_list);
1185 cfs_spin_unlock(&req->rq_export->exp_lock);
1189 static void target_finish_recovery(struct obd_device *obd)
1192 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1195 ldlm_reprocess_all_ns(obd->obd_namespace);
1196 cfs_spin_lock(&obd->obd_recovery_task_lock);
1197 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1198 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1199 !cfs_list_empty(&obd->obd_final_req_queue)) {
1200 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1202 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1203 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1205 cfs_list_empty(&obd->obd_final_req_queue) ? \
1207 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1210 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1212 obd->obd_recovery_end = cfs_time_current_sec();
1214 /* when recovery finished, cleanup orphans on mds and ost */
1215 if (OBT(obd) && OBP(obd, postrecov)) {
1216 int rc = OBP(obd, postrecov)(obd);
1218 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1224 static void abort_req_replay_queue(struct obd_device *obd)
1226 struct ptlrpc_request *req, *n;
1227 cfs_list_t abort_list;
1229 CFS_INIT_LIST_HEAD(&abort_list);
1230 cfs_spin_lock(&obd->obd_recovery_task_lock);
1231 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1232 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1233 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1234 DEBUG_REQ(D_WARNING, req, "aborted:");
1235 req->rq_status = -ENOTCONN;
1236 if (ptlrpc_error(req)) {
1237 DEBUG_REQ(D_ERROR, req,
1238 "failed abort_req_reply; skipping");
1240 target_exp_dequeue_req_replay(req);
1241 target_request_copy_put(req);
1245 static void abort_lock_replay_queue(struct obd_device *obd)
1247 struct ptlrpc_request *req, *n;
1248 cfs_list_t abort_list;
1250 CFS_INIT_LIST_HEAD(&abort_list);
1251 cfs_spin_lock(&obd->obd_recovery_task_lock);
1252 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1253 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1254 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1255 DEBUG_REQ(D_ERROR, req, "aborted:");
1256 req->rq_status = -ENOTCONN;
1257 if (ptlrpc_error(req)) {
1258 DEBUG_REQ(D_ERROR, req,
1259 "failed abort_lock_reply; skipping");
1261 target_request_copy_put(req);
1266 /* Called from a cleanup function if the device is being cleaned up
1267 forcefully. The exports should all have been disconnected already,
1268 the only thing left to do is
1269 - clear the recovery flags
1271 - free queued requests and replies, but don't send replies
1272 Because the obd_stopping flag is set, no new requests should be received.
1275 void target_cleanup_recovery(struct obd_device *obd)
1277 struct ptlrpc_request *req, *n;
1278 cfs_list_t clean_list;
1281 CFS_INIT_LIST_HEAD(&clean_list);
1282 cfs_spin_lock(&obd->obd_dev_lock);
1283 if (!obd->obd_recovering) {
1284 cfs_spin_unlock(&obd->obd_dev_lock);
1288 obd->obd_recovering = obd->obd_abort_recovery = 0;
1289 cfs_spin_unlock(&obd->obd_dev_lock);
1291 cfs_spin_lock(&obd->obd_recovery_task_lock);
1292 target_cancel_recovery_timer(obd);
1293 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1294 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1296 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1297 LASSERT(req->rq_reply_state == 0);
1298 target_exp_dequeue_req_replay(req);
1299 target_request_copy_put(req);
1302 cfs_spin_lock(&obd->obd_recovery_task_lock);
1303 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1304 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1305 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1307 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1308 LASSERT(req->rq_reply_state == 0);
1309 target_request_copy_put(req);
1315 /* obd_recovery_task_lock should be held */
1316 void target_cancel_recovery_timer(struct obd_device *obd)
1318 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1319 cfs_timer_disarm(&obd->obd_recovery_timer);
1322 /* extend = 1 means require at least "duration" seconds left in the timer,
1323 extend = 0 means set the total duration (start_recovery_timer) */
1324 static void reset_recovery_timer(struct obd_device *obd, int duration,
1327 cfs_time_t now = cfs_time_current_sec();
1328 cfs_duration_t left;
1330 cfs_spin_lock(&obd->obd_recovery_task_lock);
1331 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1332 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1336 left = cfs_time_sub(obd->obd_recovery_end, now);
1338 if (extend && (duration > left))
1339 obd->obd_recovery_timeout += duration - left;
1340 else if (!extend && (duration > obd->obd_recovery_timeout))
1341 /* Track the client's largest expected replay time */
1342 obd->obd_recovery_timeout = duration;
1344 /* Hard limit of obd_recovery_time_hard which should not happen */
1345 if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1346 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1348 obd->obd_recovery_end = obd->obd_recovery_start +
1349 obd->obd_recovery_timeout;
1350 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1351 cfs_time_before(now, obd->obd_recovery_end)) {
1352 left = cfs_time_sub(obd->obd_recovery_end, now);
1353 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1355 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1356 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1357 obd->obd_name, (unsigned)left);
1360 static void check_and_start_recovery_timer(struct obd_device *obd)
1362 cfs_spin_lock(&obd->obd_recovery_task_lock);
1363 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1364 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1367 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1368 obd->obd_recovery_start = cfs_time_current_sec();
1369 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1371 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1374 /* Reset the timer with each new client connection */
1376 * This timer is actually reconnect_timer, which is for making sure
1377 * the total recovery window is at least as big as my reconnect
1378 * attempt timing. So the initial recovery time_out will be set to
1379 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1380 * from client is bigger than this, then the recovery time_out will
1381 * be extended to make sure the client could be reconnected, in the
1382 * process, the timeout from the new client should be ignored.
1386 target_start_and_reset_recovery_timer(struct obd_device *obd,
1387 struct ptlrpc_request *req,
1390 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1392 if (!new_client && service_time)
1393 /* Teach server about old server's estimates, as first guess
1394 * at how long new requests will take. */
1395 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1398 check_and_start_recovery_timer(obd);
1400 /* convert the service time to rpc timeout,
1401 * reuse service_time to limit stack usage */
1402 service_time = at_est2timeout(service_time);
1404 /* We expect other clients to timeout within service_time, then try
1405 * to reconnect, then try the failover server. The max delay between
1406 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1407 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1408 INITIAL_CONNECT_TIMEOUT);
1409 if (service_time > obd->obd_recovery_timeout && !new_client)
1410 reset_recovery_timer(obd, service_time, 0);
1415 /** Health checking routines */
1416 static inline int exp_connect_healthy(struct obd_export *exp)
1418 return (exp->exp_in_recovery);
1421 /** if export done req_replay or has replay in queue */
1422 static inline int exp_req_replay_healthy(struct obd_export *exp)
1424 return (!exp->exp_req_replay_needed ||
1425 cfs_atomic_read(&exp->exp_replay_count) > 0);
1427 /** if export done lock_replay or has replay in queue */
1428 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1430 return (!exp->exp_lock_replay_needed ||
1431 cfs_atomic_read(&exp->exp_replay_count) > 0);
1434 static inline int exp_vbr_healthy(struct obd_export *exp)
1436 return (!exp->exp_vbr_failed);
1439 static inline int exp_finished(struct obd_export *exp)
1441 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1444 /** Checking routines for recovery */
1445 static int check_for_clients(struct obd_device *obd)
1447 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1449 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1450 if (obd->obd_no_conn == 0 &&
1451 obd->obd_connected_clients + obd->obd_stale_clients ==
1452 obd->obd_max_recoverable_clients)
1457 static int check_for_next_transno(struct obd_device *obd)
1459 struct ptlrpc_request *req = NULL;
1460 int wake_up = 0, connected, completed, queue_len;
1461 __u64 next_transno, req_transno;
1464 cfs_spin_lock(&obd->obd_recovery_task_lock);
1465 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1466 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1467 struct ptlrpc_request, rq_list);
1468 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1473 connected = obd->obd_connected_clients;
1474 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1475 queue_len = obd->obd_requests_queued_for_recovery;
1476 next_transno = obd->obd_next_recovery_transno;
1478 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1479 "req_transno: "LPU64", next_transno: "LPU64"\n",
1480 obd->obd_max_recoverable_clients, connected, completed,
1481 queue_len, req_transno, next_transno);
1483 if (obd->obd_abort_recovery) {
1484 CDEBUG(D_HA, "waking for aborted recovery\n");
1486 } else if (obd->obd_recovery_expired) {
1487 CDEBUG(D_HA, "waking for expired recovery\n");
1489 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1490 CDEBUG(D_HA, "waking for completed recovery\n");
1492 } else if (req_transno == next_transno) {
1493 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1495 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1497 /** handle gaps occured due to lost reply or VBR */
1498 LASSERTF(req_transno >= next_transno,
1499 "req_transno: "LPU64", next_transno: "LPU64"\n",
1500 req_transno, next_transno);
1501 if (req_transno > obd->obd_last_committed &&
1502 !obd->obd_version_recov)
1505 "%s: waking for gap in transno, VBR is %s (skip: "
1506 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1507 ", last_committed: "LPD64")\n",
1508 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1509 next_transno, queue_len, completed, connected,
1510 req_transno, obd->obd_last_committed);
1511 obd->obd_next_recovery_transno = req_transno;
1513 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1514 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1515 " by fail_lock, waking up ("LPD64")\n", next_transno);
1516 obd->obd_next_recovery_transno = req_transno;
1519 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1523 static int check_for_next_lock(struct obd_device *obd)
1527 cfs_spin_lock(&obd->obd_recovery_task_lock);
1528 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1529 CDEBUG(D_HA, "waking for next lock\n");
1531 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1532 CDEBUG(D_HA, "waking for completed lock replay\n");
1534 } else if (obd->obd_abort_recovery) {
1535 CDEBUG(D_HA, "waking for aborted recovery\n");
1537 } else if (obd->obd_recovery_expired) {
1538 CDEBUG(D_HA, "waking for expired recovery\n");
1541 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1547 * wait for recovery events,
1548 * check its status with help of check_routine
1549 * evict dead clients via health_check
1551 static int target_recovery_overseer(struct obd_device *obd,
1552 int (*check_routine)(struct obd_device *),
1553 int (*health_check)(struct obd_export *))
1556 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1557 if (obd->obd_abort_recovery) {
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);
1562 } else if (obd->obd_recovery_expired) {
1563 obd->obd_recovery_expired = 0;
1564 /** If some clients died being recovered, evict them */
1566 "recovery is timed out, evict stale exports\n");
1567 /** evict cexports with no replay in queue, they are stalled */
1568 class_disconnect_stale_exports(obd, health_check);
1569 /** continue with VBR */
1570 cfs_spin_lock(&obd->obd_dev_lock);
1571 obd->obd_version_recov = 1;
1572 cfs_spin_unlock(&obd->obd_dev_lock);
1574 * reset timer, recovery will proceed with versions now,
1575 * timeout is set just to handle reconnection delays
1577 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1578 /** Wait for recovery events again, after evicting bad clients */
1584 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1586 struct ptlrpc_request *req = NULL;
1589 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1590 obd->obd_next_recovery_transno);
1592 if (target_recovery_overseer(obd, check_for_next_transno,
1593 exp_req_replay_healthy)) {
1594 abort_req_replay_queue(obd);
1595 abort_lock_replay_queue(obd);
1598 cfs_spin_lock(&obd->obd_recovery_task_lock);
1599 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1600 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1601 struct ptlrpc_request, rq_list);
1602 cfs_list_del_init(&req->rq_list);
1603 obd->obd_requests_queued_for_recovery--;
1604 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1606 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1607 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1608 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1609 /** evict exports failed VBR */
1610 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1615 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1617 struct ptlrpc_request *req = NULL;
1619 CDEBUG(D_HA, "Waiting for lock\n");
1620 if (target_recovery_overseer(obd, check_for_next_lock,
1621 exp_lock_replay_healthy))
1622 abort_lock_replay_queue(obd);
1624 cfs_spin_lock(&obd->obd_recovery_task_lock);
1625 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1626 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1627 struct ptlrpc_request, rq_list);
1628 cfs_list_del_init(&req->rq_list);
1629 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1631 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1632 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1633 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1634 /** evict exports failed VBR */
1635 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1640 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1642 struct ptlrpc_request *req = NULL;
1644 cfs_spin_lock(&obd->obd_recovery_task_lock);
1645 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1646 req = cfs_list_entry(obd->obd_final_req_queue.next,
1647 struct ptlrpc_request, rq_list);
1648 cfs_list_del_init(&req->rq_list);
1649 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1650 if (req->rq_export->exp_in_recovery) {
1651 cfs_spin_lock(&req->rq_export->exp_lock);
1652 req->rq_export->exp_in_recovery = 0;
1653 cfs_spin_unlock(&req->rq_export->exp_lock);
1656 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1661 static int handle_recovery_req(struct ptlrpc_thread *thread,
1662 struct ptlrpc_request *req,
1663 svc_handler_t handler)
1669 * export can be evicted during recovery, no need to handle replays for
1670 * it after that, discard such request silently
1672 if (req->rq_export->exp_disconnected)
1673 GOTO(reqcopy_put, rc = 0);
1675 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1677 CERROR("Failure to initialize session: %d\n", rc);
1678 GOTO(reqcopy_put, rc);
1681 req->rq_recov_session.lc_thread = thread;
1682 lu_context_enter(&req->rq_recov_session);
1683 req->rq_svc_thread = thread;
1684 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1686 /* thread context */
1687 lu_context_enter(&thread->t_env->le_ctx);
1689 lu_context_exit(&thread->t_env->le_ctx);
1691 lu_context_exit(&req->rq_recov_session);
1692 lu_context_fini(&req->rq_recov_session);
1693 /* don't reset timer for final stage */
1694 if (!exp_finished(req->rq_export)) {
1696 * Add request timeout to the recovery time so next request from
1697 * this client may come in recovery time
1699 reset_recovery_timer(class_exp2obd(req->rq_export),
1700 AT_OFF ? obd_timeout :
1701 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1707 static int target_recovery_thread(void *arg)
1709 struct lu_target *lut = arg;
1710 struct obd_device *obd = lut->lut_obd;
1711 struct ptlrpc_request *req;
1712 struct target_recovery_data *trd = &obd->obd_recovery_data;
1713 unsigned long delta;
1714 unsigned long flags;
1716 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1720 cfs_daemonize_ctxt("tgt_recov");
1722 SIGNAL_MASK_LOCK(current, flags);
1723 sigfillset(¤t->blocked);
1725 SIGNAL_MASK_UNLOCK(current, flags);
1727 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1731 thread->t_env = &env;
1732 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1733 env.le_ctx.lc_thread = thread;
1734 thread->t_data = NULL;
1736 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1738 trd->trd_processing_task = cfs_curproc_pid();
1740 cfs_spin_lock(&obd->obd_dev_lock);
1741 obd->obd_recovering = 1;
1742 cfs_spin_unlock(&obd->obd_dev_lock);
1743 cfs_complete(&trd->trd_starting);
1745 /* first of all, we have to know the first transno to replay */
1746 if (target_recovery_overseer(obd, check_for_clients,
1747 exp_connect_healthy)) {
1748 abort_req_replay_queue(obd);
1749 abort_lock_replay_queue(obd);
1752 /* next stage: replay requests */
1754 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1755 cfs_atomic_read(&obd->obd_req_replay_clients),
1756 obd->obd_next_recovery_transno);
1757 while ((req = target_next_replay_req(obd))) {
1758 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1759 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1760 lustre_msg_get_transno(req->rq_reqmsg),
1761 libcfs_nid2str(req->rq_peer.nid));
1762 handle_recovery_req(thread, req,
1763 trd->trd_recovery_handler);
1765 * bz18031: increase next_recovery_transno before
1766 * target_request_copy_put() will drop exp_rpc reference
1768 cfs_spin_lock(&obd->obd_recovery_task_lock);
1769 obd->obd_next_recovery_transno++;
1770 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1771 target_exp_dequeue_req_replay(req);
1772 target_request_copy_put(req);
1773 obd->obd_replayed_requests++;
1777 * The second stage: replay locks
1779 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1780 cfs_atomic_read(&obd->obd_lock_replay_clients));
1781 while ((req = target_next_replay_lock(obd))) {
1782 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1783 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1784 libcfs_nid2str(req->rq_peer.nid));
1785 handle_recovery_req(thread, req,
1786 trd->trd_recovery_handler);
1787 target_request_copy_put(req);
1788 obd->obd_replayed_locks++;
1792 * The third stage: reply on final pings, at this moment all clients
1793 * must have request in final queue
1795 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1796 /** Update server last boot epoch */
1797 lut_boot_epoch_update(lut);
1798 /* We drop recoverying flag to forward all new requests
1799 * to regular mds_handle() since now */
1800 cfs_spin_lock(&obd->obd_dev_lock);
1801 obd->obd_recovering = obd->obd_abort_recovery = 0;
1802 cfs_spin_unlock(&obd->obd_dev_lock);
1803 cfs_spin_lock(&obd->obd_recovery_task_lock);
1804 target_cancel_recovery_timer(obd);
1805 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1806 while ((req = target_next_final_ping(obd))) {
1807 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1808 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1809 libcfs_nid2str(req->rq_peer.nid));
1810 handle_recovery_req(thread, req,
1811 trd->trd_recovery_handler);
1812 target_request_copy_put(req);
1815 delta = (jiffies - delta) / CFS_HZ;
1816 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1817 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1818 if (delta > OBD_RECOVERY_TIME_SOFT) {
1819 CWARN("too long recovery - read logs\n");
1820 libcfs_debug_dumplog();
1823 target_finish_recovery(obd);
1825 lu_context_fini(&env.le_ctx);
1826 trd->trd_processing_task = 0;
1827 cfs_complete(&trd->trd_finishing);
1831 static int target_start_recovery_thread(struct lu_target *lut,
1832 svc_handler_t handler)
1834 struct obd_device *obd = lut->lut_obd;
1836 struct target_recovery_data *trd = &obd->obd_recovery_data;
1838 memset(trd, 0, sizeof(*trd));
1839 cfs_init_completion(&trd->trd_starting);
1840 cfs_init_completion(&trd->trd_finishing);
1841 trd->trd_recovery_handler = handler;
1843 if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
1844 cfs_wait_for_completion(&trd->trd_starting);
1845 LASSERT(obd->obd_recovering != 0);
1852 void target_stop_recovery_thread(struct obd_device *obd)
1854 if (obd->obd_recovery_data.trd_processing_task > 0) {
1855 struct target_recovery_data *trd = &obd->obd_recovery_data;
1856 /** recovery can be done but postrecovery is not yet */
1857 cfs_spin_lock(&obd->obd_dev_lock);
1858 if (obd->obd_recovering) {
1859 CERROR("%s: Aborting recovery\n", obd->obd_name);
1860 obd->obd_abort_recovery = 1;
1861 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1863 cfs_spin_unlock(&obd->obd_dev_lock);
1864 cfs_wait_for_completion(&trd->trd_finishing);
1868 void target_recovery_fini(struct obd_device *obd)
1870 class_disconnect_exports(obd);
1871 target_stop_recovery_thread(obd);
1872 target_cleanup_recovery(obd);
1874 EXPORT_SYMBOL(target_recovery_fini);
1876 static void target_recovery_expired(unsigned long castmeharder)
1878 struct obd_device *obd = (struct obd_device *)castmeharder;
1879 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1880 " after %lds (%d clients connected)\n",
1881 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1882 cfs_time_current_sec()- obd->obd_recovery_start,
1883 obd->obd_connected_clients);
1885 obd->obd_recovery_expired = 1;
1886 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1889 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1891 struct obd_device *obd = lut->lut_obd;
1892 if (obd->obd_max_recoverable_clients == 0) {
1893 /** Update server last boot epoch */
1894 lut_boot_epoch_update(lut);
1898 CWARN("RECOVERY: service %s, %d recoverable clients, "
1899 "last_transno "LPU64"\n", obd->obd_name,
1900 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1901 LASSERT(obd->obd_stopping == 0);
1902 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1903 obd->obd_recovery_start = 0;
1904 obd->obd_recovery_end = 0;
1906 /* both values can be get from mount data already */
1907 if (obd->obd_recovery_timeout == 0)
1908 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
1909 if (obd->obd_recovery_time_hard == 0)
1910 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
1911 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1912 target_start_recovery_thread(lut, handler);
1914 EXPORT_SYMBOL(target_recovery_init);
1918 static int target_process_req_flags(struct obd_device *obd,
1919 struct ptlrpc_request *req)
1921 struct obd_export *exp = req->rq_export;
1922 LASSERT(exp != NULL);
1923 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1924 /* client declares he's ready to replay locks */
1925 cfs_spin_lock(&exp->exp_lock);
1926 if (exp->exp_req_replay_needed) {
1927 exp->exp_req_replay_needed = 0;
1928 cfs_spin_unlock(&exp->exp_lock);
1930 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
1931 cfs_atomic_dec(&obd->obd_req_replay_clients);
1933 cfs_spin_unlock(&exp->exp_lock);
1936 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1937 /* client declares he's ready to complete recovery
1938 * so, we put the request on th final queue */
1939 cfs_spin_lock(&exp->exp_lock);
1940 if (exp->exp_lock_replay_needed) {
1941 exp->exp_lock_replay_needed = 0;
1942 cfs_spin_unlock(&exp->exp_lock);
1944 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
1945 cfs_atomic_dec(&obd->obd_lock_replay_clients);
1947 cfs_spin_unlock(&exp->exp_lock);
1953 int target_queue_recovery_request(struct ptlrpc_request *req,
1954 struct obd_device *obd)
1958 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1961 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1962 /* Processing the queue right now, don't re-add. */
1966 target_process_req_flags(obd, req);
1968 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1969 /* client declares he's ready to complete recovery
1970 * so, we put the request on th final queue */
1971 target_request_copy_get(req);
1972 DEBUG_REQ(D_HA, req, "queue final req");
1973 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1974 cfs_spin_lock(&obd->obd_recovery_task_lock);
1975 if (obd->obd_recovering) {
1976 cfs_list_add_tail(&req->rq_list,
1977 &obd->obd_final_req_queue);
1979 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1980 target_request_copy_put(req);
1981 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1983 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1986 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1987 /* client declares he's ready to replay locks */
1988 target_request_copy_get(req);
1989 DEBUG_REQ(D_HA, req, "queue lock replay req");
1990 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1991 cfs_spin_lock(&obd->obd_recovery_task_lock);
1992 LASSERT(obd->obd_recovering);
1993 /* usually due to recovery abort */
1994 if (!req->rq_export->exp_in_recovery) {
1995 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1996 target_request_copy_put(req);
1999 LASSERT(req->rq_export->exp_lock_replay_needed);
2000 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2001 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2005 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2006 * (i.e. buflens etc are in my own byte order), but type-dependent
2007 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2010 CFS_INIT_LIST_HEAD(&req->rq_list);
2011 DEBUG_REQ(D_HA, req, "not queueing");
2015 /* If we're processing the queue, we want don't want to queue this
2018 * Also, if this request has a transno less than the one we're waiting
2019 * for, we should process it now. It could (and currently always will)
2020 * be an open request for a descriptor that was opened some time ago.
2022 * Also, a resent, replayed request that has already been
2023 * handled will pass through here and be processed immediately.
2025 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2026 obd->obd_next_recovery_transno, transno);
2027 cfs_spin_lock(&obd->obd_recovery_task_lock);
2028 if (transno < obd->obd_next_recovery_transno) {
2029 /* Processing the queue right now, don't re-add. */
2030 LASSERT(cfs_list_empty(&req->rq_list));
2031 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2034 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2036 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2039 target_request_copy_get(req);
2040 if (!req->rq_export->exp_in_recovery) {
2041 target_request_copy_put(req);
2044 LASSERT(req->rq_export->exp_req_replay_needed);
2046 if (target_exp_enqueue_req_replay(req)) {
2047 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2048 target_request_copy_put(req);
2053 cfs_spin_lock(&obd->obd_recovery_task_lock);
2054 LASSERT(obd->obd_recovering);
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(&obd->obd_recovery_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_spin_unlock(&obd->obd_recovery_task_lock);
2081 cfs_waitq_signal(&obd->obd_next_transno_waitq);
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_rs_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);
2213 cfs_spin_lock(&rs->rs_lock);
2214 if (rs->rs_transno <= exp->exp_last_committed ||
2215 (!rs->rs_on_net && !rs->rs_no_ack) ||
2216 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2217 cfs_list_empty(&rs->rs_obd_list)) {
2218 CDEBUG(D_HA, "Schedule reply immediately\n");
2219 ptlrpc_dispatch_difficult_reply(rs);
2221 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2222 rs->rs_scheduled = 0; /* allow notifier to schedule */
2224 cfs_spin_unlock(&rs->rs_lock);
2225 cfs_spin_unlock(&svc->srv_rs_lock);
2229 int target_handle_ping(struct ptlrpc_request *req)
2231 obd_ping(req->rq_export);
2232 return req_capsule_server_pack(&req->rq_pill);
2235 void target_committed_to_req(struct ptlrpc_request *req)
2237 struct obd_export *exp = req->rq_export;
2239 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2240 lustre_msg_set_last_committed(req->rq_repmsg,
2241 exp->exp_last_committed);
2243 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2244 "%d)", exp->exp_obd->obd_no_transno,
2245 req->rq_repmsg == NULL);
2247 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2248 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2250 EXPORT_SYMBOL(target_committed_to_req);
2252 int target_handle_qc_callback(struct ptlrpc_request *req)
2254 struct obd_quotactl *oqctl;
2255 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2257 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2258 if (oqctl == NULL) {
2259 CERROR("Can't unpack obd_quotactl\n");
2263 cli->cl_qchk_stat = oqctl->qc_stat;
2268 #ifdef HAVE_QUOTA_SUPPORT
2269 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2272 struct obd_device *obd = req->rq_export->exp_obd;
2273 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2274 struct obd_device_target *obt;
2275 struct lustre_quota_ctxt *qctxt;
2276 struct qunit_data *qdata = NULL;
2280 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2283 rc = req_capsule_server_pack(&req->rq_pill);
2285 CERROR("packing reply failed!: rc = %d\n", rc);
2289 LASSERT(req->rq_export);
2291 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2292 if (IS_ERR(qdata)) {
2293 rc = PTR_ERR(qdata);
2294 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2295 req->rq_status = rc;
2299 /* we use the observer */
2300 if (obd_pin_observer(obd, &lov_obd) ||
2301 obd_pin_observer(lov_obd, &master_obd)) {
2302 CERROR("Can't find the observer, it is recovering\n");
2303 req->rq_status = -EAGAIN;
2307 obt = &master_obd->u.obt;
2308 qctxt = &obt->obt_qctxt;
2310 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2311 /* quota_type has not been processed yet, return EAGAIN
2312 * until we know whether or not quotas are supposed to
2314 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2316 req->rq_status = -EAGAIN;
2320 cfs_down_read(&obt->obt_rwsem);
2321 if (qctxt->lqc_lqs_hash == NULL) {
2322 cfs_up_read(&obt->obt_rwsem);
2323 /* quota_type has not been processed yet, return EAGAIN
2324 * until we know whether or not quotas are supposed to
2326 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2328 req->rq_status = -EAGAIN;
2332 LASSERT(qctxt->lqc_handler);
2333 rc = qctxt->lqc_handler(master_obd, qdata,
2334 lustre_msg_get_opc(req->rq_reqmsg));
2335 cfs_up_read(&obt->obt_rwsem);
2336 if (rc && rc != -EDQUOT)
2337 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2338 "dqacq/dqrel failed! (rc:%d)\n", rc);
2339 req->rq_status = rc;
2341 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2343 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2347 /* Block the quota req. b=14840 */
2348 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2353 obd_unpin_observer(lov_obd);
2355 obd_unpin_observer(obd);
2357 rc = ptlrpc_reply(req);
2361 #endif /* !__KERNEL__ */
2363 #endif /* HAVE_QUOTA_SUPPORT */
2365 ldlm_mode_t lck_compat_array[] = {
2366 [LCK_EX] LCK_COMPAT_EX,
2367 [LCK_PW] LCK_COMPAT_PW,
2368 [LCK_PR] LCK_COMPAT_PR,
2369 [LCK_CW] LCK_COMPAT_CW,
2370 [LCK_CR] LCK_COMPAT_CR,
2371 [LCK_NL] LCK_COMPAT_NL,
2372 [LCK_GROUP] LCK_COMPAT_GROUP,
2373 [LCK_COS] LCK_COMPAT_COS,
2377 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2378 * not escape to the user level.
2380 int ldlm_error2errno(ldlm_error_t error)
2388 case ELDLM_LOCK_CHANGED:
2391 case ELDLM_LOCK_ABORTED:
2394 case ELDLM_LOCK_REPLACED:
2397 case ELDLM_NO_LOCK_DATA:
2400 case ELDLM_NAMESPACE_EXISTS:
2403 case ELDLM_BAD_NAMESPACE:
2407 if (((int)error) < 0) /* cast to signed type */
2408 result = error; /* as ldlm_error_t can be unsigned */
2410 CERROR("Invalid DLM result code: %d\n", error);
2416 EXPORT_SYMBOL(ldlm_error2errno);
2419 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2421 ldlm_error_t ldlm_errno2error(int err_no)
2430 error = ELDLM_LOCK_CHANGED;
2433 error = ELDLM_LOCK_ABORTED;
2436 error = ELDLM_LOCK_REPLACED;
2439 error = ELDLM_NO_LOCK_DATA;
2442 error = ELDLM_NAMESPACE_EXISTS;
2445 error = ELDLM_BAD_NAMESPACE;
2452 EXPORT_SYMBOL(ldlm_errno2error);
2454 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2455 void ldlm_dump_export_locks(struct obd_export *exp)
2457 cfs_spin_lock(&exp->exp_locks_list_guard);
2458 if (!cfs_list_empty(&exp->exp_locks_list)) {
2459 struct ldlm_lock *lock;
2461 CERROR("dumping locks for export %p,"
2462 "ignore if the unmount doesn't hang\n", exp);
2463 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2464 ldlm_lock_dump(D_ERROR, lock, 0);
2466 cfs_spin_unlock(&exp->exp_locks_list_guard);