1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_mds.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: if non-zero, move the selected to the list head
55 * @create: if zero, only search in existed connections
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 cfs_spin_lock(&imp->imp_lock);
84 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 cfs_list_del(&item->oic_item);
88 cfs_list_add(&item->oic_item,
90 item->oic_last_attempt = 0;
92 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? ", moved to head" : ""));
95 cfs_spin_unlock(&imp->imp_lock);
96 GOTO(out_free, rc = 0);
101 imp_conn->oic_conn = ptlrpc_conn;
102 imp_conn->oic_uuid = *uuid;
103 imp_conn->oic_last_attempt = 0;
105 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107 cfs_list_add_tail(&imp_conn->oic_item,
108 &imp->imp_conn_list);
109 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
110 imp, imp->imp_obd->obd_name, uuid->uuid,
111 (priority ? "head" : "tail"));
113 cfs_spin_unlock(&imp->imp_lock);
114 GOTO(out_free, rc = -ENOENT);
117 cfs_spin_unlock(&imp->imp_lock);
121 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 ptlrpc_connection_put(ptlrpc_conn);
127 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 return import_set_conn(imp, uuid, 1, 0);
132 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
135 return import_set_conn(imp, uuid, priority, 1);
138 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
140 struct obd_import_conn *imp_conn;
141 struct obd_export *dlmexp;
145 cfs_spin_lock(&imp->imp_lock);
146 if (cfs_list_empty(&imp->imp_conn_list)) {
147 LASSERT(!imp->imp_connection);
151 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
152 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
154 LASSERT(imp_conn->oic_conn);
156 /* is current conn? */
157 if (imp_conn == imp->imp_conn_current) {
158 LASSERT(imp_conn->oic_conn == imp->imp_connection);
160 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
161 imp->imp_state != LUSTRE_IMP_DISCON) {
162 CERROR("can't remove current connection\n");
163 GOTO(out, rc = -EBUSY);
166 ptlrpc_connection_put(imp->imp_connection);
167 imp->imp_connection = NULL;
169 dlmexp = class_conn2export(&imp->imp_dlm_handle);
170 if (dlmexp && dlmexp->exp_connection) {
171 LASSERT(dlmexp->exp_connection ==
173 ptlrpc_connection_put(dlmexp->exp_connection);
174 dlmexp->exp_connection = NULL;
178 cfs_list_del(&imp_conn->oic_item);
179 ptlrpc_connection_put(imp_conn->oic_conn);
180 OBD_FREE(imp_conn, sizeof(*imp_conn));
181 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182 imp, imp->imp_obd->obd_name, uuid->uuid);
187 cfs_spin_unlock(&imp->imp_lock);
189 CERROR("connection %s not found\n", uuid->uuid);
193 void client_destroy_import(struct obd_import *imp)
195 /* drop security policy instance after all rpc finished/aborted
196 * to let all busy contexts be released. */
197 class_import_get(imp);
198 class_destroy_import(imp);
199 sptlrpc_import_sec_put(imp);
200 class_import_put(imp);
203 /* configure an RPC client OBD device
208 * 3 - inactive-on-startup
210 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
212 struct client_obd *cli = &obddev->u.cli;
213 struct obd_import *imp;
214 struct obd_uuid server_uuid;
215 int rq_portal, rp_portal, connect_op;
216 char *name = obddev->obd_type->typ_name;
220 /* In a more perfect world, we would hang a ptlrpc_client off of
221 * obd_type and just use the values from there. */
222 if (!strcmp(name, LUSTRE_OSC_NAME)) {
223 rq_portal = OST_REQUEST_PORTAL;
224 rp_portal = OSC_REPLY_PORTAL;
225 connect_op = OST_CONNECT;
226 cli->cl_sp_me = LUSTRE_SP_CLI;
227 cli->cl_sp_to = LUSTRE_SP_OST;
228 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
229 rq_portal = MDS_REQUEST_PORTAL;
230 rp_portal = MDC_REPLY_PORTAL;
231 connect_op = MDS_CONNECT;
232 cli->cl_sp_me = LUSTRE_SP_CLI;
233 cli->cl_sp_to = LUSTRE_SP_MDT;
234 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
235 rq_portal = MGS_REQUEST_PORTAL;
236 rp_portal = MGC_REPLY_PORTAL;
237 connect_op = MGS_CONNECT;
238 cli->cl_sp_me = LUSTRE_SP_MGC;
239 cli->cl_sp_to = LUSTRE_SP_MGS;
240 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
242 CERROR("unknown client OBD type \"%s\", can't setup\n",
247 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
248 CERROR("requires a TARGET UUID\n");
252 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
253 CERROR("client UUID must be less than 38 characters\n");
257 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
258 CERROR("setup requires a SERVER UUID\n");
262 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
263 CERROR("target UUID must be less than 38 characters\n");
267 cfs_init_rwsem(&cli->cl_sem);
268 cfs_sema_init(&cli->cl_mgc_sem, 1);
269 cli->cl_conn_count = 0;
270 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
271 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
272 sizeof(server_uuid)));
275 cli->cl_avail_grant = 0;
276 /* FIXME: should limit this for the sum of all cl_dirty_max */
277 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
278 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
279 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
280 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
281 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
282 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
283 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
284 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
285 CFS_INIT_LIST_HEAD(&cli->cl_loi_sync_fs_list);
286 client_obd_list_lock_init(&cli->cl_loi_list_lock);
287 cli->cl_r_in_flight = 0;
288 cli->cl_w_in_flight = 0;
290 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
291 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
292 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
293 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
294 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
295 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
296 cfs_waitq_init(&cli->cl_destroy_waitq);
297 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
298 #ifdef ENABLE_CHECKSUM
299 /* Turn on checksumming by default. */
300 cli->cl_checksum = 1;
302 * The supported checksum types will be worked out at connect time
303 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
306 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
308 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
310 /* This value may be changed at connect time in
311 ptlrpc_connect_interpret. */
312 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
313 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
315 if (!strcmp(name, LUSTRE_MDC_NAME)) {
316 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
317 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
318 cli->cl_max_rpcs_in_flight = 2;
319 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
320 cli->cl_max_rpcs_in_flight = 3;
321 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
322 cli->cl_max_rpcs_in_flight = 4;
324 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
329 CERROR("ldlm_get_ref failed: %d\n", rc);
333 ptlrpc_init_client(rq_portal, rp_portal, name,
334 &obddev->obd_ldlm_client);
336 imp = class_new_import(obddev);
338 GOTO(err_ldlm, rc = -ENOENT);
339 imp->imp_client = &obddev->obd_ldlm_client;
340 imp->imp_connect_op = connect_op;
341 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
342 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
343 LUSTRE_CFG_BUFLEN(lcfg, 1));
344 class_import_put(imp);
346 rc = client_import_add_conn(imp, &server_uuid, 1);
348 CERROR("can't add initial connection\n");
349 GOTO(err_import, rc);
352 cli->cl_import = imp;
353 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
354 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
355 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
357 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
358 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
359 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
360 name, obddev->obd_name,
361 cli->cl_target_uuid.uuid);
362 cfs_spin_lock(&imp->imp_lock);
363 imp->imp_deactive = 1;
364 cfs_spin_unlock(&imp->imp_lock);
368 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
369 LDLM_NAMESPACE_CLIENT,
370 LDLM_NAMESPACE_GREEDY);
371 if (obddev->obd_namespace == NULL) {
372 CERROR("Unable to create client namespace - %s\n",
374 GOTO(err_import, rc = -ENOMEM);
377 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
382 class_destroy_import(imp);
390 int client_obd_cleanup(struct obd_device *obddev)
394 ldlm_namespace_free_post(obddev->obd_namespace);
395 obddev->obd_namespace = NULL;
401 /* ->o_connect() method for client side (OSC and MDC and MGC) */
402 int client_connect_import(const struct lu_env *env,
403 struct obd_export **exp,
404 struct obd_device *obd, struct obd_uuid *cluuid,
405 struct obd_connect_data *data, void *localdata)
407 struct client_obd *cli = &obd->u.cli;
408 struct obd_import *imp = cli->cl_import;
409 struct obd_connect_data *ocd;
410 struct lustre_handle conn = { 0 };
415 cfs_down_write(&cli->cl_sem);
416 if (cli->cl_conn_count > 0 )
417 GOTO(out_sem, rc = -EALREADY);
419 rc = class_connect(&conn, obd, cluuid);
423 cli->cl_conn_count++;
424 *exp = class_conn2export(&conn);
426 LASSERT(obd->obd_namespace);
428 imp->imp_dlm_handle = conn;
429 rc = ptlrpc_init_import(imp);
433 ocd = &imp->imp_connect_data;
436 imp->imp_connect_flags_orig = data->ocd_connect_flags;
439 rc = ptlrpc_connect_import(imp, NULL);
441 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
444 LASSERT((*exp)->exp_connection);
447 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
448 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
449 data->ocd_connect_flags, ocd->ocd_connect_flags);
450 data->ocd_connect_flags = ocd->ocd_connect_flags;
453 ptlrpc_pinger_add_import(imp);
459 cli->cl_conn_count--;
460 class_disconnect(*exp);
464 cfs_up_write(&cli->cl_sem);
469 int client_disconnect_export(struct obd_export *exp)
471 struct obd_device *obd = class_exp2obd(exp);
472 struct client_obd *cli;
473 struct obd_import *imp;
478 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
479 exp, exp ? exp->exp_handle.h_cookie : -1);
484 imp = cli->cl_import;
486 cfs_down_write(&cli->cl_sem);
487 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
490 if (!cli->cl_conn_count) {
491 CERROR("disconnecting disconnected device (%s)\n",
493 GOTO(out_disconnect, rc = -EINVAL);
496 cli->cl_conn_count--;
497 if (cli->cl_conn_count)
498 GOTO(out_disconnect, rc = 0);
500 /* Mark import deactivated now, so we don't try to reconnect if any
501 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
502 * fully deactivate the import, or that would drop all requests. */
503 cfs_spin_lock(&imp->imp_lock);
504 imp->imp_deactive = 1;
505 cfs_spin_unlock(&imp->imp_lock);
507 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
508 * delete it regardless. (It's safe to delete an import that was
510 (void)ptlrpc_pinger_del_import(imp);
512 if (obd->obd_namespace != NULL) {
513 /* obd_force == local only */
514 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
515 obd->obd_force ? LCF_LOCAL : 0, NULL);
516 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
520 * there's no need to hold sem during disconnecting an import,
521 * and actually it may cause deadlock in gss.
523 cfs_up_write(&cli->cl_sem);
524 rc = ptlrpc_disconnect_import(imp, 0);
525 cfs_down_write(&cli->cl_sem);
527 ptlrpc_invalidate_import(imp);
529 if (imp->imp_rq_pool) {
530 ptlrpc_free_rq_pool(imp->imp_rq_pool);
531 imp->imp_rq_pool = NULL;
533 client_destroy_import(imp);
534 cli->cl_import = NULL;
539 /* use server style - class_disconnect should be always called for
541 err = class_disconnect(exp);
545 cfs_up_write(&cli->cl_sem);
550 int server_disconnect_export(struct obd_export *exp)
555 /* Disconnect early so that clients can't keep using export */
556 rc = class_disconnect(exp);
557 /* close import for avoid sending any requests */
558 if (exp->exp_imp_reverse)
559 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
561 if (exp->exp_obd->obd_namespace != NULL)
562 ldlm_cancel_locks_for_export(exp);
564 /* complete all outstanding replies */
565 cfs_spin_lock(&exp->exp_lock);
566 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
567 struct ptlrpc_reply_state *rs =
568 cfs_list_entry(exp->exp_outstanding_replies.next,
569 struct ptlrpc_reply_state, rs_exp_list);
570 struct ptlrpc_service *svc = rs->rs_service;
572 cfs_spin_lock(&svc->srv_lock);
573 cfs_list_del_init(&rs->rs_exp_list);
574 cfs_spin_lock(&rs->rs_lock);
575 ptlrpc_schedule_difficult_reply(rs);
576 cfs_spin_unlock(&rs->rs_lock);
577 cfs_spin_unlock(&svc->srv_lock);
579 cfs_spin_unlock(&exp->exp_lock);
584 /* --------------------------------------------------------------------------
585 * from old lib/target.c
586 * -------------------------------------------------------------------------- */
588 static int target_handle_reconnect(struct lustre_handle *conn,
589 struct obd_export *exp,
590 struct obd_uuid *cluuid)
593 if (exp->exp_connection && exp->exp_imp_reverse) {
594 struct lustre_handle *hdl;
595 hdl = &exp->exp_imp_reverse->imp_remote_handle;
596 /* Might be a re-connect after a partition. */
597 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
598 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
600 conn->cookie = exp->exp_handle.h_cookie;
601 /* target_handle_connect() treats EALREADY and
602 * -EALREADY differently. EALREADY means we are
603 * doing a valid reconnect from the same client. */
606 CERROR("%s reconnecting from %s, "
607 "handle mismatch (ours "LPX64", theirs "
608 LPX64")\n", cluuid->uuid,
609 exp->exp_connection->c_remote_uuid.uuid,
610 hdl->cookie, conn->cookie);
611 memset(conn, 0, sizeof *conn);
612 /* target_handle_connect() treats EALREADY and
613 * -EALREADY differently. -EALREADY is an error
614 * (same UUID, different handle). */
619 conn->cookie = exp->exp_handle.h_cookie;
620 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
621 cluuid->uuid, exp, conn->cookie);
625 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
628 struct obd_export *exp = cb_data;
630 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
631 obd->obd_name, exp->exp_client_uuid.uuid);
633 cfs_spin_lock(&exp->exp_lock);
634 exp->exp_need_sync = 0;
635 cfs_spin_unlock(&exp->exp_lock);
636 class_export_cb_put(exp);
638 EXPORT_SYMBOL(target_client_add_cb);
641 target_start_and_reset_recovery_timer(struct obd_device *obd,
642 struct ptlrpc_request *req,
645 int target_handle_connect(struct ptlrpc_request *req)
647 struct obd_device *target, *targref = NULL;
648 struct obd_export *export = NULL;
649 struct obd_import *revimp;
650 struct lustre_handle conn;
651 struct lustre_handle *tmp;
652 struct obd_uuid tgtuuid;
653 struct obd_uuid cluuid;
654 struct obd_uuid remote_uuid;
658 struct obd_connect_data *data, *tmpdata;
659 lnet_nid_t *client_nid = NULL;
662 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
664 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
666 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
667 GOTO(out, rc = -EINVAL);
670 obd_str2uuid(&tgtuuid, str);
671 target = class_uuid2obd(&tgtuuid);
673 target = class_name2obd(str);
675 if (!target || target->obd_stopping || !target->obd_set_up) {
676 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
677 " for connect (%s)\n", str,
678 !target ? "no target" :
679 (target->obd_stopping ? "stopping" :
681 GOTO(out, rc = -ENODEV);
684 if (target->obd_no_conn) {
685 LCONSOLE_WARN("%s: temporarily refusing client connection "
686 "from %s\n", target->obd_name,
687 libcfs_nid2str(req->rq_peer.nid));
688 GOTO(out, rc = -EAGAIN);
691 /* Make sure the target isn't cleaned up while we're here. Yes,
692 there's still a race between the above check and our incref here.
693 Really, class_uuid2obd should take the ref. */
694 targref = class_incref(target, __FUNCTION__, cfs_current());
697 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
699 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
700 GOTO(out, rc = -EINVAL);
703 obd_str2uuid(&cluuid, str);
705 /* XXX extract a nettype and format accordingly */
706 switch (sizeof(lnet_nid_t)) {
707 /* NB the casts only avoid compiler warnings */
709 snprintf(remote_uuid.uuid, sizeof remote_uuid,
710 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
713 snprintf(remote_uuid.uuid, sizeof remote_uuid,
714 "NET_%x_UUID", (__u32)req->rq_peer.nid);
720 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
722 GOTO(out, rc = -EPROTO);
726 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
728 GOTO(out, rc = -EPROTO);
730 rc = req_capsule_server_pack(&req->rq_pill);
734 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
736 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
737 "libclient connection attempt");
738 GOTO(out, rc = -EPROTO);
739 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
740 LUSTRE_VERSION_ALLOWED_OFFSET ||
741 data->ocd_version > LUSTRE_VERSION_CODE +
742 LUSTRE_VERSION_ALLOWED_OFFSET) {
743 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
744 "libclient connection attempt",
745 data->ocd_version < LUSTRE_VERSION_CODE ?
747 OBD_OCD_VERSION_MAJOR(data->ocd_version),
748 OBD_OCD_VERSION_MINOR(data->ocd_version),
749 OBD_OCD_VERSION_PATCH(data->ocd_version),
750 OBD_OCD_VERSION_FIX(data->ocd_version));
751 data = req_capsule_server_sized_get(&req->rq_pill,
753 offsetof(typeof(*data), ocd_version) +
754 sizeof(data->ocd_version));
756 data->ocd_connect_flags = OBD_CONNECT_VERSION;
757 data->ocd_version = LUSTRE_VERSION_CODE;
759 GOTO(out, rc = -EPROTO);
763 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
764 (data->ocd_connect_flags & OBD_CONNECT_MDS))
767 /* lctl gets a backstage, all-access pass. */
768 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
769 goto dont_check_exports;
771 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
775 /* we've found an export in the hash */
776 if (export->exp_connecting) { /* bug 9635, et. al. */
777 CWARN("%s: exp %p already connecting\n",
778 export->exp_obd->obd_name, export);
779 class_export_put(export);
782 } else if (mds_conn && export->exp_connection) {
783 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
784 /* mds reconnected after failover */
785 CWARN("%s: received MDS connection from NID %s,"
786 " removing former export from NID %s\n",
787 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
788 libcfs_nid2str(export->exp_connection->c_peer.nid));
790 /* new mds connection from the same nid */
791 CWARN("%s: received new MDS connection from NID %s,"
792 " removing former export from same NID\n",
793 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
794 class_fail_export(export);
795 class_export_put(export);
798 } else if (export->exp_connection != NULL &&
799 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
800 (lustre_msg_get_op_flags(req->rq_reqmsg) &
801 MSG_CONNECT_INITIAL)) {
802 /* in mds failover we have static uuid but nid can be
804 CWARN("%s: cookie %s seen on new NID %s when "
805 "existing NID %s is already connected\n",
806 target->obd_name, cluuid.uuid,
807 libcfs_nid2str(req->rq_peer.nid),
808 libcfs_nid2str(export->exp_connection->c_peer.nid));
810 class_export_put(export);
813 cfs_spin_lock(&export->exp_lock);
814 export->exp_connecting = 1;
815 cfs_spin_unlock(&export->exp_lock);
816 class_export_put(export);
817 LASSERT(export->exp_obd == target);
819 rc = target_handle_reconnect(&conn, export, &cluuid);
822 /* If we found an export, we already unlocked. */
825 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
826 } else if (req->rq_export == NULL &&
827 cfs_atomic_read(&export->exp_rpc_count) > 0) {
828 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
829 target->obd_name, cluuid.uuid,
830 libcfs_nid2str(req->rq_peer.nid),
831 export, cfs_atomic_read(&export->exp_refcount));
832 GOTO(out, rc = -EBUSY);
833 } else if (req->rq_export != NULL &&
834 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
835 /* the current connect rpc has increased exp_rpc_count */
836 CWARN("%s: refuse reconnection 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_rpc_count) - 1);
840 cfs_spin_lock(&export->exp_lock);
841 if (req->rq_export->exp_conn_cnt <
842 lustre_msg_get_conn_cnt(req->rq_reqmsg))
843 /* try to abort active requests */
844 req->rq_export->exp_abort_active_req = 1;
845 cfs_spin_unlock(&export->exp_lock);
846 GOTO(out, rc = -EBUSY);
847 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
848 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
849 "cookies not random?\n", target->obd_name,
850 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
851 GOTO(out, rc = -EALREADY);
853 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
860 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
861 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
862 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
863 export, (long)cfs_time_current_sec(),
864 export ? (long)export->exp_last_request_time : 0);
866 /* Tell the client if we're in recovery. */
867 if (target->obd_recovering) {
868 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
869 /* If this is the first time a client connects,
870 reset the recovery timer */
872 target_start_and_reset_recovery_timer(target, req,
876 /* We want to handle EALREADY but *not* -EALREADY from
877 * target_handle_reconnect(), return reconnection state in a flag */
878 if (rc == EALREADY) {
879 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
885 /* Tell the client if we support replayable requests */
886 if (target->obd_replayable)
887 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
888 client_nid = &req->rq_peer.nid;
890 if (export == NULL) {
891 if (target->obd_recovering) {
894 t = cfs_timer_deadline(&target->obd_recovery_timer);
895 t = cfs_time_sub(t, cfs_time_current());
896 CERROR("%s: denying connection for new client %s (%s): "
897 "%d clients in recovery for "CFS_TIME_T"s\n",
899 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
900 cfs_atomic_read(&target-> \
901 obd_lock_replay_clients),
902 cfs_duration_sec(t));
906 rc = obd_connect(req->rq_svc_thread->t_env,
907 &export, target, &cluuid, data,
910 conn.cookie = export->exp_handle.h_cookie;
913 rc = obd_reconnect(req->rq_svc_thread->t_env,
914 export, target, &cluuid, data, client_nid);
916 /* prevous done via class_conn2export */
917 class_export_get(export);
921 /* Return only the parts of obd_connect_data that we understand, so the
922 * client knows that we don't understand the rest. */
924 tmpdata = req_capsule_server_get(&req->rq_pill,
926 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
930 /* If all else goes well, this is our RPC return code. */
933 lustre_msg_set_handle(req->rq_repmsg, &conn);
935 /* If the client and the server are the same node, we will already
936 * have an export that really points to the client's DLM export,
937 * because we have a shared handles table.
939 * XXX this will go away when shaver stops sending the "connect" handle
940 * in the real "remote handle" field of the request --phik 24 Apr 2003
942 if (req->rq_export != NULL)
943 class_export_put(req->rq_export);
945 req->rq_export = export;
947 cfs_spin_lock(&export->exp_lock);
948 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
949 cfs_spin_unlock(&export->exp_lock);
950 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
951 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
952 export->exp_conn_cnt,
953 lustre_msg_get_conn_cnt(req->rq_reqmsg));
955 GOTO(out, rc = -EALREADY);
957 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
958 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
959 export->exp_abort_active_req = 0;
961 /* request from liblustre? Don't evict it for not pinging. */
962 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
963 export->exp_libclient = 1;
964 cfs_spin_unlock(&export->exp_lock);
966 cfs_spin_lock(&target->obd_dev_lock);
967 cfs_list_del_init(&export->exp_obd_chain_timed);
968 cfs_spin_unlock(&target->obd_dev_lock);
970 cfs_spin_unlock(&export->exp_lock);
973 if (export->exp_connection != NULL) {
974 /* Check to see if connection came from another NID */
975 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
976 !cfs_hlist_unhashed(&export->exp_nid_hash))
977 cfs_hash_del(export->exp_obd->obd_nid_hash,
978 &export->exp_connection->c_peer.nid,
979 &export->exp_nid_hash);
981 ptlrpc_connection_put(export->exp_connection);
984 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
987 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
988 cfs_hash_add(export->exp_obd->obd_nid_hash,
989 &export->exp_connection->c_peer.nid,
990 &export->exp_nid_hash);
993 cfs_spin_lock(&target->obd_recovery_task_lock);
994 if (target->obd_recovering && !export->exp_in_recovery) {
995 cfs_spin_lock(&export->exp_lock);
996 export->exp_in_recovery = 1;
997 export->exp_req_replay_needed = 1;
998 export->exp_lock_replay_needed = 1;
999 cfs_spin_unlock(&export->exp_lock);
1000 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1001 && (data->ocd_transno == 0))
1002 CWARN("Connect with zero transno!\n");
1004 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1005 && data->ocd_transno < target->obd_next_recovery_transno)
1006 target->obd_next_recovery_transno = data->ocd_transno;
1007 target->obd_connected_clients++;
1008 cfs_atomic_inc(&target->obd_req_replay_clients);
1009 cfs_atomic_inc(&target->obd_lock_replay_clients);
1010 if (target->obd_connected_clients ==
1011 target->obd_max_recoverable_clients)
1012 cfs_waitq_signal(&target->obd_next_transno_waitq);
1014 cfs_spin_unlock(&target->obd_recovery_task_lock);
1015 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1018 if (export->exp_imp_reverse != NULL) {
1019 /* destroyed import can be still referenced in ctxt */
1020 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1021 KEY_REVIMP_UPD, 0, NULL, NULL);
1023 client_destroy_import(export->exp_imp_reverse);
1026 /* for the rest part, we return -ENOTCONN in case of errors
1027 * in order to let client initialize connection again.
1029 revimp = export->exp_imp_reverse = class_new_import(target);
1031 CERROR("fail to alloc new reverse import.\n");
1032 GOTO(out, rc = -ENOTCONN);
1035 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1036 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1037 revimp->imp_remote_handle = conn;
1038 revimp->imp_dlm_fake = 1;
1039 revimp->imp_state = LUSTRE_IMP_FULL;
1041 /* unknown versions will be caught in
1042 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1043 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1045 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1046 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1047 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1049 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1051 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1052 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1053 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1055 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1057 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1059 CERROR("Failed to get sec for reverse import: %d\n", rc);
1060 export->exp_imp_reverse = NULL;
1061 class_destroy_import(revimp);
1064 class_import_put(revimp);
1067 cfs_spin_lock(&export->exp_lock);
1068 export->exp_connecting = 0;
1069 cfs_spin_unlock(&export->exp_lock);
1072 class_decref(targref, __FUNCTION__, cfs_current());
1074 req->rq_status = rc;
1078 int target_handle_disconnect(struct ptlrpc_request *req)
1083 rc = req_capsule_server_pack(&req->rq_pill);
1087 /* keep the rq_export around so we can send the reply */
1088 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1093 void target_destroy_export(struct obd_export *exp)
1095 /* exports created from last_rcvd data, and "fake"
1096 exports created by lctl don't have an import */
1097 if (exp->exp_imp_reverse != NULL)
1098 client_destroy_import(exp->exp_imp_reverse);
1100 LASSERT(cfs_atomic_read(&exp->exp_locks_count) == 0);
1101 LASSERT(cfs_atomic_read(&exp->exp_rpc_count) == 0);
1102 LASSERT(cfs_atomic_read(&exp->exp_cb_count) == 0);
1103 LASSERT(cfs_atomic_read(&exp->exp_replay_count) == 0);
1107 * Recovery functions
1109 static void target_request_copy_get(struct ptlrpc_request *req)
1111 class_export_rpc_get(req->rq_export);
1112 LASSERT(cfs_list_empty(&req->rq_list));
1113 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1114 /* increase refcount to keep request in queue */
1115 LASSERT(cfs_atomic_read(&req->rq_refcount));
1116 cfs_atomic_inc(&req->rq_refcount);
1117 /** let export know it has replays to be handled */
1118 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1119 /* release service thread while request is queued
1120 * we are moving the request from active processing
1121 * to waiting on the replay queue */
1122 ptlrpc_server_active_request_dec(req);
1125 static void target_request_copy_put(struct ptlrpc_request *req)
1127 LASSERT(cfs_list_empty(&req->rq_replay_list));
1128 LASSERT(cfs_atomic_read(&req->rq_export->exp_replay_count) > 0);
1129 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1130 class_export_rpc_put(req->rq_export);
1131 /* ptlrpc_server_drop_request() assumes the request is active */
1132 ptlrpc_server_active_request_inc(req);
1133 ptlrpc_server_drop_request(req);
1136 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1138 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1139 struct obd_export *exp = req->rq_export;
1140 struct ptlrpc_request *reqiter;
1145 cfs_spin_lock(&exp->exp_lock);
1146 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1148 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1155 /* we expect it with RESENT and REPLAY flags */
1156 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1157 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1158 CERROR("invalid flags %x of resent replay\n",
1159 lustre_msg_get_flags(req->rq_reqmsg));
1161 cfs_list_add_tail(&req->rq_replay_list,
1162 &exp->exp_req_replay_queue);
1165 cfs_spin_unlock(&exp->exp_lock);
1169 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1171 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1172 LASSERT(req->rq_export);
1174 cfs_spin_lock(&req->rq_export->exp_lock);
1175 cfs_list_del_init(&req->rq_replay_list);
1176 cfs_spin_unlock(&req->rq_export->exp_lock);
1180 static void target_finish_recovery(struct obd_device *obd)
1183 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1186 ldlm_reprocess_all_ns(obd->obd_namespace);
1187 cfs_spin_lock(&obd->obd_recovery_task_lock);
1188 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1189 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1190 !cfs_list_empty(&obd->obd_final_req_queue)) {
1191 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1193 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1194 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1196 cfs_list_empty(&obd->obd_final_req_queue) ? \
1198 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1201 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1203 obd->obd_recovery_end = cfs_time_current_sec();
1205 /* when recovery finished, cleanup orphans on mds and ost */
1206 if (OBT(obd) && OBP(obd, postrecov)) {
1207 int rc = OBP(obd, postrecov)(obd);
1209 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1215 static void abort_req_replay_queue(struct obd_device *obd)
1217 struct ptlrpc_request *req, *n;
1218 cfs_list_t abort_list;
1220 CFS_INIT_LIST_HEAD(&abort_list);
1221 cfs_spin_lock(&obd->obd_recovery_task_lock);
1222 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1223 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1224 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1225 DEBUG_REQ(D_WARNING, req, "aborted:");
1226 req->rq_status = -ENOTCONN;
1227 if (ptlrpc_error(req)) {
1228 DEBUG_REQ(D_ERROR, req,
1229 "failed abort_req_reply; skipping");
1231 target_exp_dequeue_req_replay(req);
1232 target_request_copy_put(req);
1236 static void abort_lock_replay_queue(struct obd_device *obd)
1238 struct ptlrpc_request *req, *n;
1239 cfs_list_t abort_list;
1241 CFS_INIT_LIST_HEAD(&abort_list);
1242 cfs_spin_lock(&obd->obd_recovery_task_lock);
1243 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1244 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1245 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1246 DEBUG_REQ(D_ERROR, req, "aborted:");
1247 req->rq_status = -ENOTCONN;
1248 if (ptlrpc_error(req)) {
1249 DEBUG_REQ(D_ERROR, req,
1250 "failed abort_lock_reply; skipping");
1252 target_request_copy_put(req);
1257 /* Called from a cleanup function if the device is being cleaned up
1258 forcefully. The exports should all have been disconnected already,
1259 the only thing left to do is
1260 - clear the recovery flags
1262 - free queued requests and replies, but don't send replies
1263 Because the obd_stopping flag is set, no new requests should be received.
1266 void target_cleanup_recovery(struct obd_device *obd)
1268 struct ptlrpc_request *req, *n;
1269 cfs_list_t clean_list;
1272 CFS_INIT_LIST_HEAD(&clean_list);
1273 cfs_spin_lock(&obd->obd_dev_lock);
1274 if (!obd->obd_recovering) {
1275 cfs_spin_unlock(&obd->obd_dev_lock);
1279 obd->obd_recovering = obd->obd_abort_recovery = 0;
1280 cfs_spin_unlock(&obd->obd_dev_lock);
1282 cfs_spin_lock(&obd->obd_recovery_task_lock);
1283 target_cancel_recovery_timer(obd);
1284 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1285 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1287 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1288 LASSERT(req->rq_reply_state == 0);
1289 target_exp_dequeue_req_replay(req);
1290 target_request_copy_put(req);
1293 cfs_spin_lock(&obd->obd_recovery_task_lock);
1294 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1295 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1296 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1298 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1299 LASSERT(req->rq_reply_state == 0);
1300 target_request_copy_put(req);
1306 /* obd_recovery_task_lock should be held */
1307 void target_cancel_recovery_timer(struct obd_device *obd)
1309 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1310 cfs_timer_disarm(&obd->obd_recovery_timer);
1313 /* extend = 1 means require at least "duration" seconds left in the timer,
1314 extend = 0 means set the total duration (start_recovery_timer) */
1315 static void reset_recovery_timer(struct obd_device *obd, int duration,
1318 cfs_time_t now = cfs_time_current_sec();
1319 cfs_duration_t left;
1321 cfs_spin_lock(&obd->obd_recovery_task_lock);
1322 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1323 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1327 left = cfs_time_sub(obd->obd_recovery_end, now);
1329 if (extend && (duration > left))
1330 obd->obd_recovery_timeout += duration - left;
1331 else if (!extend && (duration > obd->obd_recovery_timeout))
1332 /* Track the client's largest expected replay time */
1333 obd->obd_recovery_timeout = duration;
1335 /* Hard limit of obd_recovery_time_hard which should not happen */
1336 if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1337 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1339 obd->obd_recovery_end = obd->obd_recovery_start +
1340 obd->obd_recovery_timeout;
1341 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1342 cfs_time_before(now, obd->obd_recovery_end)) {
1343 left = cfs_time_sub(obd->obd_recovery_end, now);
1344 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1346 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1347 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1348 obd->obd_name, (unsigned)left);
1351 static void check_and_start_recovery_timer(struct obd_device *obd)
1353 cfs_spin_lock(&obd->obd_recovery_task_lock);
1354 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1355 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1358 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1359 obd->obd_recovery_start = cfs_time_current_sec();
1360 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1362 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1365 /* Reset the timer with each new client connection */
1367 * This timer is actually reconnect_timer, which is for making sure
1368 * the total recovery window is at least as big as my reconnect
1369 * attempt timing. So the initial recovery time_out will be set to
1370 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1371 * from client is bigger than this, then the recovery time_out will
1372 * be extended to make sure the client could be reconnected, in the
1373 * process, the timeout from the new client should be ignored.
1377 target_start_and_reset_recovery_timer(struct obd_device *obd,
1378 struct ptlrpc_request *req,
1381 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1383 if (!new_client && service_time)
1384 /* Teach server about old server's estimates, as first guess
1385 * at how long new requests will take. */
1386 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1389 check_and_start_recovery_timer(obd);
1391 /* convert the service time to rpc timeout,
1392 * reuse service_time to limit stack usage */
1393 service_time = at_est2timeout(service_time);
1395 /* We expect other clients to timeout within service_time, then try
1396 * to reconnect, then try the failover server. The max delay between
1397 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1398 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1399 INITIAL_CONNECT_TIMEOUT);
1400 if (service_time > obd->obd_recovery_timeout && !new_client)
1401 reset_recovery_timer(obd, service_time, 0);
1406 /** Health checking routines */
1407 static inline int exp_connect_healthy(struct obd_export *exp)
1409 return (exp->exp_in_recovery);
1412 /** if export done req_replay or has replay in queue */
1413 static inline int exp_req_replay_healthy(struct obd_export *exp)
1415 return (!exp->exp_req_replay_needed ||
1416 cfs_atomic_read(&exp->exp_replay_count) > 0);
1418 /** if export done lock_replay or has replay in queue */
1419 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1421 return (!exp->exp_lock_replay_needed ||
1422 cfs_atomic_read(&exp->exp_replay_count) > 0);
1425 static inline int exp_vbr_healthy(struct obd_export *exp)
1427 return (!exp->exp_vbr_failed);
1430 static inline int exp_finished(struct obd_export *exp)
1432 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1435 /** Checking routines for recovery */
1436 static int check_for_clients(struct obd_device *obd)
1438 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1440 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1441 if (obd->obd_no_conn == 0 &&
1442 obd->obd_connected_clients + obd->obd_stale_clients ==
1443 obd->obd_max_recoverable_clients)
1448 static int check_for_next_transno(struct obd_device *obd)
1450 struct ptlrpc_request *req = NULL;
1451 int wake_up = 0, connected, completed, queue_len;
1452 __u64 next_transno, req_transno;
1455 cfs_spin_lock(&obd->obd_recovery_task_lock);
1456 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1457 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1458 struct ptlrpc_request, rq_list);
1459 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1464 connected = obd->obd_connected_clients;
1465 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1466 queue_len = obd->obd_requests_queued_for_recovery;
1467 next_transno = obd->obd_next_recovery_transno;
1469 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1470 "req_transno: "LPU64", next_transno: "LPU64"\n",
1471 obd->obd_max_recoverable_clients, connected, completed,
1472 queue_len, req_transno, next_transno);
1474 if (obd->obd_abort_recovery) {
1475 CDEBUG(D_HA, "waking for aborted recovery\n");
1477 } else if (obd->obd_recovery_expired) {
1478 CDEBUG(D_HA, "waking for expired recovery\n");
1480 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1481 CDEBUG(D_HA, "waking for completed recovery\n");
1483 } else if (req_transno == next_transno) {
1484 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1486 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1488 /** handle gaps occured due to lost reply or VBR */
1489 LASSERTF(req_transno >= next_transno,
1490 "req_transno: "LPU64", next_transno: "LPU64"\n",
1491 req_transno, next_transno);
1492 if (req_transno > obd->obd_last_committed &&
1493 !obd->obd_version_recov)
1496 "%s: waking for gap in transno, VBR is %s (skip: "
1497 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1498 ", last_committed: "LPD64")\n",
1499 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1500 next_transno, queue_len, completed, connected,
1501 req_transno, obd->obd_last_committed);
1502 obd->obd_next_recovery_transno = req_transno;
1504 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1505 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1506 " by fail_lock, waking up ("LPD64")\n", next_transno);
1507 obd->obd_next_recovery_transno = req_transno;
1510 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1514 static int check_for_next_lock(struct obd_device *obd)
1518 cfs_spin_lock(&obd->obd_recovery_task_lock);
1519 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1520 CDEBUG(D_HA, "waking for next lock\n");
1522 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1523 CDEBUG(D_HA, "waking for completed lock replay\n");
1525 } else if (obd->obd_abort_recovery) {
1526 CDEBUG(D_HA, "waking for aborted recovery\n");
1528 } else if (obd->obd_recovery_expired) {
1529 CDEBUG(D_HA, "waking for expired recovery\n");
1532 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1538 * wait for recovery events,
1539 * check its status with help of check_routine
1540 * evict dead clients via health_check
1542 static int target_recovery_overseer(struct obd_device *obd,
1543 int (*check_routine)(struct obd_device *),
1544 int (*health_check)(struct obd_export *))
1547 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1548 if (obd->obd_abort_recovery) {
1549 CWARN("recovery is aborted, evict exports in recovery\n");
1550 /** evict exports which didn't finish recovery yet */
1551 class_disconnect_stale_exports(obd, exp_finished);
1553 } else if (obd->obd_recovery_expired) {
1554 obd->obd_recovery_expired = 0;
1555 /** If some clients died being recovered, evict them */
1557 "recovery is timed out, evict stale exports\n");
1558 /** evict cexports with no replay in queue, they are stalled */
1559 class_disconnect_stale_exports(obd, health_check);
1560 /** continue with VBR */
1561 cfs_spin_lock(&obd->obd_dev_lock);
1562 obd->obd_version_recov = 1;
1563 cfs_spin_unlock(&obd->obd_dev_lock);
1565 * reset timer, recovery will proceed with versions now,
1566 * timeout is set just to handle reconnection delays
1568 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1569 /** Wait for recovery events again, after evicting bad clients */
1575 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1577 struct ptlrpc_request *req = NULL;
1580 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1581 obd->obd_next_recovery_transno);
1583 if (target_recovery_overseer(obd, check_for_next_transno,
1584 exp_req_replay_healthy)) {
1585 abort_req_replay_queue(obd);
1586 abort_lock_replay_queue(obd);
1589 cfs_spin_lock(&obd->obd_recovery_task_lock);
1590 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1591 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1592 struct ptlrpc_request, rq_list);
1593 cfs_list_del_init(&req->rq_list);
1594 obd->obd_requests_queued_for_recovery--;
1595 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1597 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1598 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1599 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1600 /** evict exports failed VBR */
1601 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1606 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1608 struct ptlrpc_request *req = NULL;
1610 CDEBUG(D_HA, "Waiting for lock\n");
1611 if (target_recovery_overseer(obd, check_for_next_lock,
1612 exp_lock_replay_healthy))
1613 abort_lock_replay_queue(obd);
1615 cfs_spin_lock(&obd->obd_recovery_task_lock);
1616 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1617 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1618 struct ptlrpc_request, rq_list);
1619 cfs_list_del_init(&req->rq_list);
1620 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1622 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1623 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1624 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1625 /** evict exports failed VBR */
1626 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1631 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1633 struct ptlrpc_request *req = NULL;
1635 cfs_spin_lock(&obd->obd_recovery_task_lock);
1636 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1637 req = cfs_list_entry(obd->obd_final_req_queue.next,
1638 struct ptlrpc_request, rq_list);
1639 cfs_list_del_init(&req->rq_list);
1640 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1641 if (req->rq_export->exp_in_recovery) {
1642 cfs_spin_lock(&req->rq_export->exp_lock);
1643 req->rq_export->exp_in_recovery = 0;
1644 cfs_spin_unlock(&req->rq_export->exp_lock);
1647 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1652 static int handle_recovery_req(struct ptlrpc_thread *thread,
1653 struct ptlrpc_request *req,
1654 svc_handler_t handler)
1660 * export can be evicted during recovery, no need to handle replays for
1661 * it after that, discard such request silently
1663 if (req->rq_export->exp_disconnected)
1664 GOTO(reqcopy_put, rc = 0);
1666 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1668 CERROR("Failure to initialize session: %d\n", rc);
1669 GOTO(reqcopy_put, rc);
1672 req->rq_recov_session.lc_thread = thread;
1673 lu_context_enter(&req->rq_recov_session);
1674 req->rq_svc_thread = thread;
1675 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1677 /* thread context */
1678 lu_context_enter(&thread->t_env->le_ctx);
1680 lu_context_exit(&thread->t_env->le_ctx);
1682 lu_context_exit(&req->rq_recov_session);
1683 lu_context_fini(&req->rq_recov_session);
1684 /* don't reset timer for final stage */
1685 if (!exp_finished(req->rq_export)) {
1687 * Add request timeout to the recovery time so next request from
1688 * this client may come in recovery time
1690 reset_recovery_timer(class_exp2obd(req->rq_export),
1691 AT_OFF ? obd_timeout :
1692 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1698 static int target_recovery_thread(void *arg)
1700 struct lu_target *lut = arg;
1701 struct obd_device *obd = lut->lut_obd;
1702 struct ptlrpc_request *req;
1703 struct target_recovery_data *trd = &obd->obd_recovery_data;
1704 unsigned long delta;
1705 unsigned long flags;
1707 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1711 cfs_daemonize_ctxt("tgt_recov");
1713 SIGNAL_MASK_LOCK(current, flags);
1714 sigfillset(¤t->blocked);
1716 SIGNAL_MASK_UNLOCK(current, flags);
1718 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1722 thread->t_env = &env;
1723 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1724 env.le_ctx.lc_thread = thread;
1725 thread->t_data = NULL;
1727 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1729 trd->trd_processing_task = cfs_curproc_pid();
1731 cfs_spin_lock(&obd->obd_dev_lock);
1732 obd->obd_recovering = 1;
1733 cfs_spin_unlock(&obd->obd_dev_lock);
1734 cfs_complete(&trd->trd_starting);
1736 /* first of all, we have to know the first transno to replay */
1737 if (target_recovery_overseer(obd, check_for_clients,
1738 exp_connect_healthy)) {
1739 abort_req_replay_queue(obd);
1740 abort_lock_replay_queue(obd);
1743 /* next stage: replay requests */
1745 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1746 cfs_atomic_read(&obd->obd_req_replay_clients),
1747 obd->obd_next_recovery_transno);
1748 while ((req = target_next_replay_req(obd))) {
1749 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1750 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1751 lustre_msg_get_transno(req->rq_reqmsg),
1752 libcfs_nid2str(req->rq_peer.nid));
1753 handle_recovery_req(thread, req,
1754 trd->trd_recovery_handler);
1756 * bz18031: increase next_recovery_transno before
1757 * target_request_copy_put() will drop exp_rpc reference
1759 cfs_spin_lock(&obd->obd_recovery_task_lock);
1760 obd->obd_next_recovery_transno++;
1761 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1762 target_exp_dequeue_req_replay(req);
1763 target_request_copy_put(req);
1764 obd->obd_replayed_requests++;
1768 * The second stage: replay locks
1770 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1771 cfs_atomic_read(&obd->obd_lock_replay_clients));
1772 while ((req = target_next_replay_lock(obd))) {
1773 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1774 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1775 libcfs_nid2str(req->rq_peer.nid));
1776 handle_recovery_req(thread, req,
1777 trd->trd_recovery_handler);
1778 target_request_copy_put(req);
1779 obd->obd_replayed_locks++;
1783 * The third stage: reply on final pings, at this moment all clients
1784 * must have request in final queue
1786 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1787 /** Update server last boot epoch */
1788 lut_boot_epoch_update(lut);
1789 /* We drop recoverying flag to forward all new requests
1790 * to regular mds_handle() since now */
1791 cfs_spin_lock(&obd->obd_dev_lock);
1792 obd->obd_recovering = obd->obd_abort_recovery = 0;
1793 cfs_spin_unlock(&obd->obd_dev_lock);
1794 cfs_spin_lock(&obd->obd_recovery_task_lock);
1795 target_cancel_recovery_timer(obd);
1796 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1797 while ((req = target_next_final_ping(obd))) {
1798 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1799 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1800 libcfs_nid2str(req->rq_peer.nid));
1801 handle_recovery_req(thread, req,
1802 trd->trd_recovery_handler);
1803 target_request_copy_put(req);
1806 delta = (jiffies - delta) / CFS_HZ;
1807 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1808 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1809 if (delta > OBD_RECOVERY_TIME_SOFT) {
1810 CWARN("too long recovery - read logs\n");
1811 libcfs_debug_dumplog();
1814 target_finish_recovery(obd);
1816 lu_context_fini(&env.le_ctx);
1817 trd->trd_processing_task = 0;
1818 cfs_complete(&trd->trd_finishing);
1822 static int target_start_recovery_thread(struct lu_target *lut,
1823 svc_handler_t handler)
1825 struct obd_device *obd = lut->lut_obd;
1827 struct target_recovery_data *trd = &obd->obd_recovery_data;
1829 memset(trd, 0, sizeof(*trd));
1830 cfs_init_completion(&trd->trd_starting);
1831 cfs_init_completion(&trd->trd_finishing);
1832 trd->trd_recovery_handler = handler;
1834 if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
1835 cfs_wait_for_completion(&trd->trd_starting);
1836 LASSERT(obd->obd_recovering != 0);
1843 void target_stop_recovery_thread(struct obd_device *obd)
1845 if (obd->obd_recovery_data.trd_processing_task > 0) {
1846 struct target_recovery_data *trd = &obd->obd_recovery_data;
1847 /** recovery can be done but postrecovery is not yet */
1848 cfs_spin_lock(&obd->obd_dev_lock);
1849 if (obd->obd_recovering) {
1850 CERROR("%s: Aborting recovery\n", obd->obd_name);
1851 obd->obd_abort_recovery = 1;
1852 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1854 cfs_spin_unlock(&obd->obd_dev_lock);
1855 cfs_wait_for_completion(&trd->trd_finishing);
1859 void target_recovery_fini(struct obd_device *obd)
1861 class_disconnect_exports(obd);
1862 target_stop_recovery_thread(obd);
1863 target_cleanup_recovery(obd);
1865 EXPORT_SYMBOL(target_recovery_fini);
1867 static void target_recovery_expired(unsigned long castmeharder)
1869 struct obd_device *obd = (struct obd_device *)castmeharder;
1870 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1871 " after %lds (%d clients connected)\n",
1872 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1873 cfs_time_current_sec()- obd->obd_recovery_start,
1874 obd->obd_connected_clients);
1876 obd->obd_recovery_expired = 1;
1877 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1880 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1882 struct obd_device *obd = lut->lut_obd;
1883 if (obd->obd_max_recoverable_clients == 0) {
1884 /** Update server last boot epoch */
1885 lut_boot_epoch_update(lut);
1889 CWARN("RECOVERY: service %s, %d recoverable clients, "
1890 "last_transno "LPU64"\n", obd->obd_name,
1891 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1892 LASSERT(obd->obd_stopping == 0);
1893 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1894 obd->obd_recovery_start = 0;
1895 obd->obd_recovery_end = 0;
1897 /* both values can be get from mount data already */
1898 if (obd->obd_recovery_timeout == 0)
1899 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
1900 if (obd->obd_recovery_time_hard == 0)
1901 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
1902 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1903 target_start_recovery_thread(lut, handler);
1905 EXPORT_SYMBOL(target_recovery_init);
1909 static int target_process_req_flags(struct obd_device *obd,
1910 struct ptlrpc_request *req)
1912 struct obd_export *exp = req->rq_export;
1913 LASSERT(exp != NULL);
1914 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1915 /* client declares he's ready to replay locks */
1916 cfs_spin_lock(&exp->exp_lock);
1917 if (exp->exp_req_replay_needed) {
1918 exp->exp_req_replay_needed = 0;
1919 cfs_spin_unlock(&exp->exp_lock);
1920 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients));
1921 cfs_atomic_dec(&obd->obd_req_replay_clients);
1923 cfs_spin_unlock(&exp->exp_lock);
1926 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1927 /* client declares he's ready to complete recovery
1928 * so, we put the request on th final queue */
1929 cfs_spin_lock(&exp->exp_lock);
1930 if (exp->exp_lock_replay_needed) {
1931 exp->exp_lock_replay_needed = 0;
1932 cfs_spin_unlock(&exp->exp_lock);
1933 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients));
1934 cfs_atomic_dec(&obd->obd_lock_replay_clients);
1936 cfs_spin_unlock(&exp->exp_lock);
1942 int target_queue_recovery_request(struct ptlrpc_request *req,
1943 struct obd_device *obd)
1947 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1950 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1951 /* Processing the queue right now, don't re-add. */
1955 target_process_req_flags(obd, req);
1957 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1958 /* client declares he's ready to complete recovery
1959 * so, we put the request on th final queue */
1960 target_request_copy_get(req);
1961 DEBUG_REQ(D_HA, req, "queue final req");
1962 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1963 cfs_spin_lock(&obd->obd_recovery_task_lock);
1964 if (obd->obd_recovering) {
1965 cfs_list_add_tail(&req->rq_list,
1966 &obd->obd_final_req_queue);
1968 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1969 target_request_copy_put(req);
1970 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1972 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1975 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1976 /* client declares he's ready to replay locks */
1977 target_request_copy_get(req);
1978 DEBUG_REQ(D_HA, req, "queue lock replay req");
1979 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1980 cfs_spin_lock(&obd->obd_recovery_task_lock);
1981 LASSERT(obd->obd_recovering);
1982 /* usually due to recovery abort */
1983 if (!req->rq_export->exp_in_recovery) {
1984 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1985 target_request_copy_put(req);
1988 LASSERT(req->rq_export->exp_lock_replay_needed);
1989 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1990 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1994 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1995 * (i.e. buflens etc are in my own byte order), but type-dependent
1996 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1999 CFS_INIT_LIST_HEAD(&req->rq_list);
2000 DEBUG_REQ(D_HA, req, "not queueing");
2004 /* If we're processing the queue, we want don't want to queue this
2007 * Also, if this request has a transno less than the one we're waiting
2008 * for, we should process it now. It could (and currently always will)
2009 * be an open request for a descriptor that was opened some time ago.
2011 * Also, a resent, replayed request that has already been
2012 * handled will pass through here and be processed immediately.
2014 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2015 obd->obd_next_recovery_transno, transno);
2016 cfs_spin_lock(&obd->obd_recovery_task_lock);
2017 if (transno < obd->obd_next_recovery_transno) {
2018 /* Processing the queue right now, don't re-add. */
2019 LASSERT(cfs_list_empty(&req->rq_list));
2020 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2023 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2025 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2028 target_request_copy_get(req);
2029 if (!req->rq_export->exp_in_recovery) {
2030 target_request_copy_put(req);
2033 LASSERT(req->rq_export->exp_req_replay_needed);
2035 if (target_exp_enqueue_req_replay(req)) {
2036 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2037 target_request_copy_put(req);
2042 cfs_spin_lock(&obd->obd_recovery_task_lock);
2043 LASSERT(obd->obd_recovering);
2044 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2045 struct ptlrpc_request *reqiter =
2046 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2048 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2049 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2054 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2056 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2057 "has been claimed by another client");
2058 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2059 target_exp_dequeue_req_replay(req);
2060 target_request_copy_put(req);
2066 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2068 obd->obd_requests_queued_for_recovery++;
2069 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2070 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2075 * Packs current SLV and Limit into \a req.
2077 int target_pack_pool_reply(struct ptlrpc_request *req)
2079 struct obd_device *obd;
2083 * Check that we still have all structures alive as this may
2084 * be some late rpc in shutdown time.
2086 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2087 !exp_connect_lru_resize(req->rq_export))) {
2088 lustre_msg_set_slv(req->rq_repmsg, 0);
2089 lustre_msg_set_limit(req->rq_repmsg, 0);
2094 * OBD is alive here as export is alive, which we checked above.
2096 obd = req->rq_export->exp_obd;
2098 cfs_read_lock(&obd->obd_pool_lock);
2099 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2100 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2101 cfs_read_unlock(&obd->obd_pool_lock);
2106 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2108 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2109 DEBUG_REQ(D_ERROR, req, "dropping reply");
2114 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2115 req->rq_status = rc;
2116 return (ptlrpc_send_error(req, 1));
2118 DEBUG_REQ(D_NET, req, "sending reply");
2121 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2124 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2127 struct ptlrpc_reply_state *rs;
2128 struct obd_device *obd;
2129 struct obd_export *exp;
2130 struct ptlrpc_service *svc;
2133 if (req->rq_no_reply) {
2138 svc = req->rq_rqbd->rqbd_service;
2139 rs = req->rq_reply_state;
2140 if (rs == NULL || !rs->rs_difficult) {
2142 target_send_reply_msg (req, rc, fail_id);
2147 /* must be an export if locks saved */
2148 LASSERT (req->rq_export != NULL);
2149 /* req/reply consistent */
2150 LASSERT (rs->rs_service == svc);
2153 LASSERT (!rs->rs_scheduled);
2154 LASSERT (!rs->rs_scheduled_ever);
2155 LASSERT (!rs->rs_handled);
2156 LASSERT (!rs->rs_on_net);
2157 LASSERT (rs->rs_export == NULL);
2158 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2159 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2161 exp = class_export_get (req->rq_export);
2164 /* disable reply scheduling while I'm setting up */
2165 rs->rs_scheduled = 1;
2167 rs->rs_xid = req->rq_xid;
2168 rs->rs_transno = req->rq_transno;
2169 rs->rs_export = exp;
2170 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2172 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2173 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2174 rs->rs_transno, exp->exp_last_committed);
2175 if (rs->rs_transno > exp->exp_last_committed) {
2176 /* not committed already */
2177 cfs_list_add_tail(&rs->rs_obd_list,
2178 &exp->exp_uncommitted_replies);
2180 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2182 cfs_spin_lock(&exp->exp_lock);
2183 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2184 cfs_spin_unlock(&exp->exp_lock);
2186 netrc = target_send_reply_msg (req, rc, fail_id);
2188 cfs_spin_lock(&svc->srv_lock);
2190 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2193 /* error sending: reply is off the net. Also we need +1
2194 * reply ref until ptlrpc_handle_rs() is done
2195 * with the reply state (if the send was successful, there
2196 * would have been +1 ref for the net, which
2197 * reply_out_callback leaves alone) */
2199 ptlrpc_rs_addref(rs);
2200 cfs_atomic_inc (&svc->srv_outstanding_replies);
2203 cfs_spin_lock(&rs->rs_lock);
2204 if (rs->rs_transno <= exp->exp_last_committed ||
2205 (!rs->rs_on_net && !rs->rs_no_ack) ||
2206 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2207 cfs_list_empty(&rs->rs_obd_list)) {
2208 CDEBUG(D_HA, "Schedule reply immediately\n");
2209 ptlrpc_dispatch_difficult_reply(rs);
2211 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2212 rs->rs_scheduled = 0; /* allow notifier to schedule */
2214 cfs_spin_unlock(&rs->rs_lock);
2215 cfs_spin_unlock(&svc->srv_lock);
2219 int target_handle_ping(struct ptlrpc_request *req)
2221 obd_ping(req->rq_export);
2222 return req_capsule_server_pack(&req->rq_pill);
2225 void target_committed_to_req(struct ptlrpc_request *req)
2227 struct obd_export *exp = req->rq_export;
2229 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2230 lustre_msg_set_last_committed(req->rq_repmsg,
2231 exp->exp_last_committed);
2233 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2234 "%d)", exp->exp_obd->obd_no_transno,
2235 req->rq_repmsg == NULL);
2237 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2238 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2240 EXPORT_SYMBOL(target_committed_to_req);
2242 int target_handle_qc_callback(struct ptlrpc_request *req)
2244 struct obd_quotactl *oqctl;
2245 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2247 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2248 if (oqctl == NULL) {
2249 CERROR("Can't unpack obd_quotactl\n");
2253 cli->cl_qchk_stat = oqctl->qc_stat;
2258 #ifdef HAVE_QUOTA_SUPPORT
2259 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2262 struct obd_device *obd = req->rq_export->exp_obd;
2263 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2264 struct obd_device_target *obt;
2265 struct lustre_quota_ctxt *qctxt;
2266 struct qunit_data *qdata = NULL;
2270 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2273 rc = req_capsule_server_pack(&req->rq_pill);
2275 CERROR("packing reply failed!: rc = %d\n", rc);
2279 LASSERT(req->rq_export);
2281 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2282 if (IS_ERR(qdata)) {
2283 rc = PTR_ERR(qdata);
2284 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2285 req->rq_status = rc;
2289 /* we use the observer */
2290 if (obd_pin_observer(obd, &lov_obd) ||
2291 obd_pin_observer(lov_obd, &master_obd)) {
2292 CERROR("Can't find the observer, it is recovering\n");
2293 req->rq_status = -EAGAIN;
2297 obt = &master_obd->u.obt;
2298 qctxt = &obt->obt_qctxt;
2300 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2301 /* quota_type has not been processed yet, return EAGAIN
2302 * until we know whether or not quotas are supposed to
2304 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2306 req->rq_status = -EAGAIN;
2310 cfs_down_read(&obt->obt_rwsem);
2311 if (qctxt->lqc_lqs_hash == NULL) {
2312 cfs_up_read(&obt->obt_rwsem);
2313 /* quota_type has not been processed yet, return EAGAIN
2314 * until we know whether or not quotas are supposed to
2316 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2318 req->rq_status = -EAGAIN;
2322 LASSERT(qctxt->lqc_handler);
2323 rc = qctxt->lqc_handler(master_obd, qdata,
2324 lustre_msg_get_opc(req->rq_reqmsg));
2325 cfs_up_read(&obt->obt_rwsem);
2326 if (rc && rc != -EDQUOT)
2327 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2328 "dqacq/dqrel failed! (rc:%d)\n", rc);
2329 req->rq_status = rc;
2331 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2333 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2337 /* Block the quota req. b=14840 */
2338 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2343 obd_unpin_observer(lov_obd);
2345 obd_unpin_observer(obd);
2347 rc = ptlrpc_reply(req);
2351 #endif /* !__KERNEL__ */
2353 #endif /* HAVE_QUOTA_SUPPORT */
2355 ldlm_mode_t lck_compat_array[] = {
2356 [LCK_EX] LCK_COMPAT_EX,
2357 [LCK_PW] LCK_COMPAT_PW,
2358 [LCK_PR] LCK_COMPAT_PR,
2359 [LCK_CW] LCK_COMPAT_CW,
2360 [LCK_CR] LCK_COMPAT_CR,
2361 [LCK_NL] LCK_COMPAT_NL,
2362 [LCK_GROUP] LCK_COMPAT_GROUP,
2363 [LCK_COS] LCK_COMPAT_COS,
2367 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2368 * not escape to the user level.
2370 int ldlm_error2errno(ldlm_error_t error)
2378 case ELDLM_LOCK_CHANGED:
2381 case ELDLM_LOCK_ABORTED:
2384 case ELDLM_LOCK_REPLACED:
2387 case ELDLM_NO_LOCK_DATA:
2390 case ELDLM_NAMESPACE_EXISTS:
2393 case ELDLM_BAD_NAMESPACE:
2397 if (((int)error) < 0) /* cast to signed type */
2398 result = error; /* as ldlm_error_t can be unsigned */
2400 CERROR("Invalid DLM result code: %d\n", error);
2406 EXPORT_SYMBOL(ldlm_error2errno);
2409 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2411 ldlm_error_t ldlm_errno2error(int err_no)
2420 error = ELDLM_LOCK_CHANGED;
2423 error = ELDLM_LOCK_ABORTED;
2426 error = ELDLM_LOCK_REPLACED;
2429 error = ELDLM_NO_LOCK_DATA;
2432 error = ELDLM_NAMESPACE_EXISTS;
2435 error = ELDLM_BAD_NAMESPACE;
2442 EXPORT_SYMBOL(ldlm_errno2error);
2444 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2445 void ldlm_dump_export_locks(struct obd_export *exp)
2447 cfs_spin_lock(&exp->exp_locks_list_guard);
2448 if (!cfs_list_empty(&exp->exp_locks_list)) {
2449 struct ldlm_lock *lock;
2451 CERROR("dumping locks for export %p,"
2452 "ignore if the unmount doesn't hang\n", exp);
2453 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2454 ldlm_lock_dump(D_ERROR, lock, 0);
2456 cfs_spin_unlock(&exp->exp_locks_list_guard);