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 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
41 # define EXPORT_SYMTAB
43 #define DEBUG_SUBSYSTEM S_LDLM
46 # include <libcfs/libcfs.h>
48 # include <liblustre.h>
51 #include <lustre_mds.h>
52 #include <lustre_dlm.h>
53 #include <lustre_net.h>
54 #include <lustre_sec.h>
55 #include "ldlm_internal.h"
57 /* @priority: if non-zero, move the selected to the list head
58 * @create: if zero, only search in existed connections
60 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
61 int priority, int create)
63 struct ptlrpc_connection *ptlrpc_conn;
64 struct obd_import_conn *imp_conn = NULL, *item;
68 if (!create && !priority) {
69 CDEBUG(D_HA, "Nothing to do\n");
73 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
75 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
80 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
82 GOTO(out_put, rc = -ENOMEM);
86 cfs_spin_lock(&imp->imp_lock);
87 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
88 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
90 cfs_list_del(&item->oic_item);
91 cfs_list_add(&item->oic_item,
93 item->oic_last_attempt = 0;
95 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
96 imp, imp->imp_obd->obd_name, uuid->uuid,
97 (priority ? ", moved to head" : ""));
98 cfs_spin_unlock(&imp->imp_lock);
99 GOTO(out_free, rc = 0);
104 imp_conn->oic_conn = ptlrpc_conn;
105 imp_conn->oic_uuid = *uuid;
106 imp_conn->oic_last_attempt = 0;
108 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
110 cfs_list_add_tail(&imp_conn->oic_item,
111 &imp->imp_conn_list);
112 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
113 imp, imp->imp_obd->obd_name, uuid->uuid,
114 (priority ? "head" : "tail"));
116 cfs_spin_unlock(&imp->imp_lock);
117 GOTO(out_free, rc = -ENOENT);
120 cfs_spin_unlock(&imp->imp_lock);
124 OBD_FREE(imp_conn, sizeof(*imp_conn));
126 ptlrpc_connection_put(ptlrpc_conn);
130 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
132 return import_set_conn(imp, uuid, 1, 0);
135 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
138 return import_set_conn(imp, uuid, priority, 1);
141 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
143 struct obd_import_conn *imp_conn;
144 struct obd_export *dlmexp;
148 cfs_spin_lock(&imp->imp_lock);
149 if (cfs_list_empty(&imp->imp_conn_list)) {
150 LASSERT(!imp->imp_connection);
154 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
155 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
157 LASSERT(imp_conn->oic_conn);
159 /* is current conn? */
160 if (imp_conn == imp->imp_conn_current) {
161 LASSERT(imp_conn->oic_conn == imp->imp_connection);
163 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
164 imp->imp_state != LUSTRE_IMP_DISCON) {
165 CERROR("can't remove current connection\n");
166 GOTO(out, rc = -EBUSY);
169 ptlrpc_connection_put(imp->imp_connection);
170 imp->imp_connection = NULL;
172 dlmexp = class_conn2export(&imp->imp_dlm_handle);
173 if (dlmexp && dlmexp->exp_connection) {
174 LASSERT(dlmexp->exp_connection ==
176 ptlrpc_connection_put(dlmexp->exp_connection);
177 dlmexp->exp_connection = NULL;
181 cfs_list_del(&imp_conn->oic_item);
182 ptlrpc_connection_put(imp_conn->oic_conn);
183 OBD_FREE(imp_conn, sizeof(*imp_conn));
184 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
185 imp, imp->imp_obd->obd_name, uuid->uuid);
190 cfs_spin_unlock(&imp->imp_lock);
192 CERROR("connection %s not found\n", uuid->uuid);
196 void client_destroy_import(struct obd_import *imp)
198 /* drop security policy instance after all rpc finished/aborted
199 * to let all busy contexts be released. */
200 class_import_get(imp);
201 class_destroy_import(imp);
202 sptlrpc_import_sec_put(imp);
203 class_import_put(imp);
206 /* configure an RPC client OBD device
211 * 3 - inactive-on-startup
213 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
215 struct client_obd *cli = &obddev->u.cli;
216 struct obd_import *imp;
217 struct obd_uuid server_uuid;
218 int rq_portal, rp_portal, connect_op;
219 char *name = obddev->obd_type->typ_name;
220 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
224 /* In a more perfect world, we would hang a ptlrpc_client off of
225 * obd_type and just use the values from there. */
226 if (!strcmp(name, LUSTRE_OSC_NAME)) {
227 rq_portal = OST_REQUEST_PORTAL;
228 rp_portal = OSC_REPLY_PORTAL;
229 connect_op = OST_CONNECT;
230 cli->cl_sp_me = LUSTRE_SP_CLI;
231 cli->cl_sp_to = LUSTRE_SP_OST;
232 ns_type = LDLM_NS_TYPE_OSC;
234 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
235 rq_portal = MDS_REQUEST_PORTAL;
236 rp_portal = MDC_REPLY_PORTAL;
237 connect_op = MDS_CONNECT;
238 cli->cl_sp_me = LUSTRE_SP_CLI;
239 cli->cl_sp_to = LUSTRE_SP_MDT;
240 ns_type = LDLM_NS_TYPE_MDC;
242 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
243 rq_portal = MGS_REQUEST_PORTAL;
244 rp_portal = MGC_REPLY_PORTAL;
245 connect_op = MGS_CONNECT;
246 cli->cl_sp_me = LUSTRE_SP_MGC;
247 cli->cl_sp_to = LUSTRE_SP_MGS;
248 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
249 ns_type = LDLM_NS_TYPE_MGC;
252 CERROR("unknown client OBD type \"%s\", can't setup\n",
257 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
258 CERROR("requires a TARGET UUID\n");
262 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
263 CERROR("client UUID must be less than 38 characters\n");
267 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
268 CERROR("setup requires a SERVER UUID\n");
272 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
273 CERROR("target UUID must be less than 38 characters\n");
277 cfs_init_rwsem(&cli->cl_sem);
278 cfs_sema_init(&cli->cl_mgc_sem, 1);
279 cli->cl_conn_count = 0;
280 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
281 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
282 sizeof(server_uuid)));
285 cli->cl_avail_grant = 0;
286 /* FIXME: should limit this for the sum of all cl_dirty_max */
287 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
288 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
289 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
290 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
291 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
292 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
293 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
294 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
295 client_obd_list_lock_init(&cli->cl_loi_list_lock);
296 cli->cl_r_in_flight = 0;
297 cli->cl_w_in_flight = 0;
299 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
300 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
301 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
302 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
303 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
304 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
305 cfs_waitq_init(&cli->cl_destroy_waitq);
306 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
307 #ifdef ENABLE_CHECKSUM
308 /* Turn on checksumming by default. */
309 cli->cl_checksum = 1;
311 * The supported checksum types will be worked out at connect time
312 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
315 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
317 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
319 /* This value may be changed at connect time in
320 ptlrpc_connect_interpret. */
321 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
322 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
324 if (!strcmp(name, LUSTRE_MDC_NAME)) {
325 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
326 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
327 cli->cl_max_rpcs_in_flight = 2;
328 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
329 cli->cl_max_rpcs_in_flight = 3;
330 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
331 cli->cl_max_rpcs_in_flight = 4;
333 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
338 CERROR("ldlm_get_ref failed: %d\n", rc);
342 ptlrpc_init_client(rq_portal, rp_portal, name,
343 &obddev->obd_ldlm_client);
345 imp = class_new_import(obddev);
347 GOTO(err_ldlm, rc = -ENOENT);
348 imp->imp_client = &obddev->obd_ldlm_client;
349 imp->imp_connect_op = connect_op;
350 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
351 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
352 LUSTRE_CFG_BUFLEN(lcfg, 1));
353 class_import_put(imp);
355 rc = client_import_add_conn(imp, &server_uuid, 1);
357 CERROR("can't add initial connection\n");
358 GOTO(err_import, rc);
361 cli->cl_import = imp;
362 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
363 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
364 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
366 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
367 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
368 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
369 name, obddev->obd_name,
370 cli->cl_target_uuid.uuid);
371 cfs_spin_lock(&imp->imp_lock);
372 imp->imp_deactive = 1;
373 cfs_spin_unlock(&imp->imp_lock);
377 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
378 LDLM_NAMESPACE_CLIENT,
379 LDLM_NAMESPACE_GREEDY,
381 if (obddev->obd_namespace == NULL) {
382 CERROR("Unable to create client namespace - %s\n",
384 GOTO(err_import, rc = -ENOMEM);
387 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
392 class_destroy_import(imp);
400 int client_obd_cleanup(struct obd_device *obddev)
404 ldlm_namespace_free_post(obddev->obd_namespace);
405 obddev->obd_namespace = NULL;
411 /* ->o_connect() method for client side (OSC and MDC and MGC) */
412 int client_connect_import(const struct lu_env *env,
413 struct obd_export **exp,
414 struct obd_device *obd, struct obd_uuid *cluuid,
415 struct obd_connect_data *data, void *localdata)
417 struct client_obd *cli = &obd->u.cli;
418 struct obd_import *imp = cli->cl_import;
419 struct obd_connect_data *ocd;
420 struct lustre_handle conn = { 0 };
425 cfs_down_write(&cli->cl_sem);
426 if (cli->cl_conn_count > 0 )
427 GOTO(out_sem, rc = -EALREADY);
429 rc = class_connect(&conn, obd, cluuid);
433 cli->cl_conn_count++;
434 *exp = class_conn2export(&conn);
436 LASSERT(obd->obd_namespace);
438 imp->imp_dlm_handle = conn;
439 rc = ptlrpc_init_import(imp);
443 ocd = &imp->imp_connect_data;
446 imp->imp_connect_flags_orig = data->ocd_connect_flags;
449 rc = ptlrpc_connect_import(imp, NULL);
451 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
454 LASSERT((*exp)->exp_connection);
457 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
458 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
459 data->ocd_connect_flags, ocd->ocd_connect_flags);
460 data->ocd_connect_flags = ocd->ocd_connect_flags;
463 ptlrpc_pinger_add_import(imp);
469 cli->cl_conn_count--;
470 class_disconnect(*exp);
474 cfs_up_write(&cli->cl_sem);
479 int client_disconnect_export(struct obd_export *exp)
481 struct obd_device *obd = class_exp2obd(exp);
482 struct client_obd *cli;
483 struct obd_import *imp;
488 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
489 exp, exp ? exp->exp_handle.h_cookie : -1);
494 imp = cli->cl_import;
496 cfs_down_write(&cli->cl_sem);
497 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
500 if (!cli->cl_conn_count) {
501 CERROR("disconnecting disconnected device (%s)\n",
503 GOTO(out_disconnect, rc = -EINVAL);
506 cli->cl_conn_count--;
507 if (cli->cl_conn_count)
508 GOTO(out_disconnect, rc = 0);
510 /* Mark import deactivated now, so we don't try to reconnect if any
511 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
512 * fully deactivate the import, or that would drop all requests. */
513 cfs_spin_lock(&imp->imp_lock);
514 imp->imp_deactive = 1;
515 cfs_spin_unlock(&imp->imp_lock);
517 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
518 * delete it regardless. (It's safe to delete an import that was
520 (void)ptlrpc_pinger_del_import(imp);
522 if (obd->obd_namespace != NULL) {
523 /* obd_force == local only */
524 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
525 obd->obd_force ? LCF_LOCAL : 0, NULL);
526 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
530 * there's no need to hold sem during disconnecting an import,
531 * and actually it may cause deadlock in gss.
533 cfs_up_write(&cli->cl_sem);
534 rc = ptlrpc_disconnect_import(imp, 0);
535 cfs_down_write(&cli->cl_sem);
537 ptlrpc_invalidate_import(imp);
539 if (imp->imp_rq_pool) {
540 ptlrpc_free_rq_pool(imp->imp_rq_pool);
541 imp->imp_rq_pool = NULL;
543 client_destroy_import(imp);
544 cli->cl_import = NULL;
549 /* use server style - class_disconnect should be always called for
551 err = class_disconnect(exp);
555 cfs_up_write(&cli->cl_sem);
560 int server_disconnect_export(struct obd_export *exp)
565 /* Disconnect early so that clients can't keep using export */
566 rc = class_disconnect(exp);
567 /* close import for avoid sending any requests */
568 if (exp->exp_imp_reverse)
569 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
571 if (exp->exp_obd->obd_namespace != NULL)
572 ldlm_cancel_locks_for_export(exp);
574 /* complete all outstanding replies */
575 cfs_spin_lock(&exp->exp_lock);
576 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
577 struct ptlrpc_reply_state *rs =
578 cfs_list_entry(exp->exp_outstanding_replies.next,
579 struct ptlrpc_reply_state, rs_exp_list);
580 struct ptlrpc_service *svc = rs->rs_service;
582 cfs_spin_lock(&svc->srv_rs_lock);
583 cfs_list_del_init(&rs->rs_exp_list);
584 cfs_spin_lock(&rs->rs_lock);
585 ptlrpc_schedule_difficult_reply(rs);
586 cfs_spin_unlock(&rs->rs_lock);
587 cfs_spin_unlock(&svc->srv_rs_lock);
589 cfs_spin_unlock(&exp->exp_lock);
594 /* --------------------------------------------------------------------------
595 * from old lib/target.c
596 * -------------------------------------------------------------------------- */
598 static int target_handle_reconnect(struct lustre_handle *conn,
599 struct obd_export *exp,
600 struct obd_uuid *cluuid)
603 if (exp->exp_connection && exp->exp_imp_reverse) {
604 struct lustre_handle *hdl;
605 hdl = &exp->exp_imp_reverse->imp_remote_handle;
606 /* Might be a re-connect after a partition. */
607 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
608 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
610 conn->cookie = exp->exp_handle.h_cookie;
611 /* target_handle_connect() treats EALREADY and
612 * -EALREADY differently. EALREADY means we are
613 * doing a valid reconnect from the same client. */
616 CERROR("%s reconnecting from %s, "
617 "handle mismatch (ours "LPX64", theirs "
618 LPX64")\n", cluuid->uuid,
619 exp->exp_connection->c_remote_uuid.uuid,
620 hdl->cookie, conn->cookie);
621 memset(conn, 0, sizeof *conn);
622 /* target_handle_connect() treats EALREADY and
623 * -EALREADY differently. -EALREADY is an error
624 * (same UUID, different handle). */
629 conn->cookie = exp->exp_handle.h_cookie;
630 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
631 cluuid->uuid, exp, conn->cookie);
635 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
638 struct obd_export *exp = cb_data;
640 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
641 obd->obd_name, exp->exp_client_uuid.uuid);
643 cfs_spin_lock(&exp->exp_lock);
644 exp->exp_need_sync = 0;
645 cfs_spin_unlock(&exp->exp_lock);
646 class_export_cb_put(exp);
648 EXPORT_SYMBOL(target_client_add_cb);
651 target_start_and_reset_recovery_timer(struct obd_device *obd,
652 struct ptlrpc_request *req,
655 int target_handle_connect(struct ptlrpc_request *req)
657 struct obd_device *target, *targref = NULL;
658 struct obd_export *export = NULL;
659 struct obd_import *revimp;
660 struct lustre_handle conn;
661 struct lustre_handle *tmp;
662 struct obd_uuid tgtuuid;
663 struct obd_uuid cluuid;
664 struct obd_uuid remote_uuid;
668 struct obd_connect_data *data, *tmpdata;
669 lnet_nid_t *client_nid = NULL;
672 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
674 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
676 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
677 GOTO(out, rc = -EINVAL);
680 obd_str2uuid(&tgtuuid, str);
681 target = class_uuid2obd(&tgtuuid);
683 target = class_name2obd(str);
685 if (!target || target->obd_stopping || !target->obd_set_up) {
686 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
687 " for connect (%s)\n", str,
688 !target ? "no target" :
689 (target->obd_stopping ? "stopping" :
691 GOTO(out, rc = -ENODEV);
694 if (target->obd_no_conn) {
695 LCONSOLE_WARN("%s: temporarily refusing client connection "
696 "from %s\n", target->obd_name,
697 libcfs_nid2str(req->rq_peer.nid));
698 GOTO(out, rc = -EAGAIN);
701 /* Make sure the target isn't cleaned up while we're here. Yes,
702 there's still a race between the above check and our incref here.
703 Really, class_uuid2obd should take the ref. */
704 targref = class_incref(target, __FUNCTION__, cfs_current());
707 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
709 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
710 GOTO(out, rc = -EINVAL);
713 obd_str2uuid(&cluuid, str);
715 /* XXX extract a nettype and format accordingly */
716 switch (sizeof(lnet_nid_t)) {
717 /* NB the casts only avoid compiler warnings */
719 snprintf(remote_uuid.uuid, sizeof remote_uuid,
720 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
723 snprintf(remote_uuid.uuid, sizeof remote_uuid,
724 "NET_%x_UUID", (__u32)req->rq_peer.nid);
730 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
732 GOTO(out, rc = -EPROTO);
736 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
738 GOTO(out, rc = -EPROTO);
740 rc = req_capsule_server_pack(&req->rq_pill);
744 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
746 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
747 "libclient connection attempt");
748 GOTO(out, rc = -EPROTO);
749 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
750 LUSTRE_VERSION_ALLOWED_OFFSET ||
751 data->ocd_version > LUSTRE_VERSION_CODE +
752 LUSTRE_VERSION_ALLOWED_OFFSET) {
753 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
754 "libclient connection attempt",
755 data->ocd_version < LUSTRE_VERSION_CODE ?
757 OBD_OCD_VERSION_MAJOR(data->ocd_version),
758 OBD_OCD_VERSION_MINOR(data->ocd_version),
759 OBD_OCD_VERSION_PATCH(data->ocd_version),
760 OBD_OCD_VERSION_FIX(data->ocd_version));
761 data = req_capsule_server_sized_get(&req->rq_pill,
763 offsetof(typeof(*data), ocd_version) +
764 sizeof(data->ocd_version));
766 data->ocd_connect_flags = OBD_CONNECT_VERSION;
767 data->ocd_version = LUSTRE_VERSION_CODE;
769 GOTO(out, rc = -EPROTO);
773 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
774 (data->ocd_connect_flags & OBD_CONNECT_MDS))
777 /* lctl gets a backstage, all-access pass. */
778 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
779 goto dont_check_exports;
781 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
785 /* we've found an export in the hash */
786 if (export->exp_connecting) { /* bug 9635, et. al. */
787 CWARN("%s: exp %p already connecting\n",
788 export->exp_obd->obd_name, export);
789 class_export_put(export);
792 } else if (mds_conn && export->exp_connection) {
793 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
794 /* mds reconnected after failover */
795 CWARN("%s: received MDS connection from NID %s,"
796 " removing former export from NID %s\n",
797 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
798 libcfs_nid2str(export->exp_connection->c_peer.nid));
800 /* new mds connection from the same nid */
801 CWARN("%s: received new MDS connection from NID %s,"
802 " removing former export from same NID\n",
803 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
804 class_fail_export(export);
805 class_export_put(export);
808 } else if (export->exp_connection != NULL &&
809 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
810 (lustre_msg_get_op_flags(req->rq_reqmsg) &
811 MSG_CONNECT_INITIAL)) {
812 /* in mds failover we have static uuid but nid can be
814 CWARN("%s: cookie %s seen on new NID %s when "
815 "existing NID %s is already connected\n",
816 target->obd_name, cluuid.uuid,
817 libcfs_nid2str(req->rq_peer.nid),
818 libcfs_nid2str(export->exp_connection->c_peer.nid));
820 class_export_put(export);
823 cfs_spin_lock(&export->exp_lock);
824 export->exp_connecting = 1;
825 cfs_spin_unlock(&export->exp_lock);
826 class_export_put(export);
827 LASSERT(export->exp_obd == target);
829 rc = target_handle_reconnect(&conn, export, &cluuid);
832 /* If we found an export, we already unlocked. */
835 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
836 } else if (req->rq_export == NULL &&
837 cfs_atomic_read(&export->exp_rpc_count) > 0) {
838 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
839 target->obd_name, cluuid.uuid,
840 libcfs_nid2str(req->rq_peer.nid),
841 export, cfs_atomic_read(&export->exp_refcount));
842 GOTO(out, rc = -EBUSY);
843 } else if (req->rq_export != NULL &&
844 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
845 /* the current connect rpc has increased exp_rpc_count */
846 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
847 target->obd_name, cluuid.uuid,
848 libcfs_nid2str(req->rq_peer.nid),
849 export, cfs_atomic_read(&export->exp_rpc_count) - 1);
850 cfs_spin_lock(&export->exp_lock);
851 if (req->rq_export->exp_conn_cnt <
852 lustre_msg_get_conn_cnt(req->rq_reqmsg))
853 /* try to abort active requests */
854 req->rq_export->exp_abort_active_req = 1;
855 cfs_spin_unlock(&export->exp_lock);
856 GOTO(out, rc = -EBUSY);
857 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
858 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
859 "cookies not random?\n", target->obd_name,
860 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
861 GOTO(out, rc = -EALREADY);
863 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
870 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
871 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
872 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
873 export, (long)cfs_time_current_sec(),
874 export ? (long)export->exp_last_request_time : 0);
876 /* If this is the first time a client connects,
877 * reset the recovery timer */
878 if (rc == 0 && target->obd_recovering)
879 target_start_and_reset_recovery_timer(target, req, !export);
881 /* We want to handle EALREADY but *not* -EALREADY from
882 * target_handle_reconnect(), return reconnection state in a flag */
883 if (rc == EALREADY) {
884 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
890 /* Tell the client if we support replayable requests */
891 if (target->obd_replayable)
892 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
893 client_nid = &req->rq_peer.nid;
895 if (export == NULL) {
896 if (target->obd_recovering) {
899 t = cfs_timer_deadline(&target->obd_recovery_timer);
900 t = cfs_time_sub(t, cfs_time_current());
901 CERROR("%s: denying connection for new client %s (%s): "
902 "%d clients in recovery for "CFS_TIME_T"s\n",
904 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
905 cfs_atomic_read(&target-> \
906 obd_lock_replay_clients),
907 cfs_duration_sec(t));
911 rc = obd_connect(req->rq_svc_thread->t_env,
912 &export, target, &cluuid, data,
915 conn.cookie = export->exp_handle.h_cookie;
918 rc = obd_reconnect(req->rq_svc_thread->t_env,
919 export, target, &cluuid, data, client_nid);
921 /* prevous done via class_conn2export */
922 class_export_get(export);
926 /* Return only the parts of obd_connect_data that we understand, so the
927 * client knows that we don't understand the rest. */
929 tmpdata = req_capsule_server_get(&req->rq_pill,
931 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
935 /* If all else goes well, this is our RPC return code. */
938 lustre_msg_set_handle(req->rq_repmsg, &conn);
940 /* If the client and the server are the same node, we will already
941 * have an export that really points to the client's DLM export,
942 * because we have a shared handles table.
944 * XXX this will go away when shaver stops sending the "connect" handle
945 * in the real "remote handle" field of the request --phik 24 Apr 2003
947 if (req->rq_export != NULL)
948 class_export_put(req->rq_export);
950 req->rq_export = export;
952 cfs_spin_lock(&export->exp_lock);
953 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
954 cfs_spin_unlock(&export->exp_lock);
955 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
956 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
957 export->exp_conn_cnt,
958 lustre_msg_get_conn_cnt(req->rq_reqmsg));
960 GOTO(out, rc = -EALREADY);
962 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
963 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
964 export->exp_abort_active_req = 0;
966 /* request from liblustre? Don't evict it for not pinging. */
967 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
968 export->exp_libclient = 1;
969 cfs_spin_unlock(&export->exp_lock);
971 cfs_spin_lock(&target->obd_dev_lock);
972 cfs_list_del_init(&export->exp_obd_chain_timed);
973 cfs_spin_unlock(&target->obd_dev_lock);
975 cfs_spin_unlock(&export->exp_lock);
978 if (export->exp_connection != NULL) {
979 /* Check to see if connection came from another NID */
980 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
981 !cfs_hlist_unhashed(&export->exp_nid_hash))
982 cfs_hash_del(export->exp_obd->obd_nid_hash,
983 &export->exp_connection->c_peer.nid,
984 &export->exp_nid_hash);
986 ptlrpc_connection_put(export->exp_connection);
989 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
992 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
993 cfs_hash_add(export->exp_obd->obd_nid_hash,
994 &export->exp_connection->c_peer.nid,
995 &export->exp_nid_hash);
998 cfs_spin_lock(&target->obd_recovery_task_lock);
999 if (target->obd_recovering && !export->exp_in_recovery) {
1000 cfs_spin_lock(&export->exp_lock);
1001 export->exp_in_recovery = 1;
1002 export->exp_req_replay_needed = 1;
1003 export->exp_lock_replay_needed = 1;
1004 cfs_spin_unlock(&export->exp_lock);
1005 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1006 && (data->ocd_transno == 0))
1007 CWARN("Connect with zero transno!\n");
1009 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1010 && data->ocd_transno < target->obd_next_recovery_transno)
1011 target->obd_next_recovery_transno = data->ocd_transno;
1012 target->obd_connected_clients++;
1013 cfs_atomic_inc(&target->obd_req_replay_clients);
1014 cfs_atomic_inc(&target->obd_lock_replay_clients);
1015 if (target->obd_connected_clients ==
1016 target->obd_max_recoverable_clients)
1017 cfs_waitq_signal(&target->obd_next_transno_waitq);
1019 cfs_spin_unlock(&target->obd_recovery_task_lock);
1021 /* Tell the client we're in recovery, when client is involved in it. */
1022 if (target->obd_recovering)
1023 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1025 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1028 if (export->exp_imp_reverse != NULL) {
1029 /* destroyed import can be still referenced in ctxt */
1030 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1031 KEY_REVIMP_UPD, 0, NULL, NULL);
1033 client_destroy_import(export->exp_imp_reverse);
1036 /* for the rest part, we return -ENOTCONN in case of errors
1037 * in order to let client initialize connection again.
1039 revimp = export->exp_imp_reverse = class_new_import(target);
1041 CERROR("fail to alloc new reverse import.\n");
1042 GOTO(out, rc = -ENOTCONN);
1045 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1046 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1047 revimp->imp_remote_handle = conn;
1048 revimp->imp_dlm_fake = 1;
1049 revimp->imp_state = LUSTRE_IMP_FULL;
1051 /* unknown versions will be caught in
1052 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1053 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1055 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1056 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1057 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1059 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1061 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1062 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1063 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1065 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1067 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1069 CERROR("Failed to get sec for reverse import: %d\n", rc);
1070 export->exp_imp_reverse = NULL;
1071 class_destroy_import(revimp);
1074 class_import_put(revimp);
1077 cfs_spin_lock(&export->exp_lock);
1078 export->exp_connecting = 0;
1079 cfs_spin_unlock(&export->exp_lock);
1082 class_decref(targref, __FUNCTION__, cfs_current());
1084 req->rq_status = rc;
1088 int target_handle_disconnect(struct ptlrpc_request *req)
1093 rc = req_capsule_server_pack(&req->rq_pill);
1097 /* keep the rq_export around so we can send the reply */
1098 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1103 void target_destroy_export(struct obd_export *exp)
1105 /* exports created from last_rcvd data, and "fake"
1106 exports created by lctl don't have an import */
1107 if (exp->exp_imp_reverse != NULL)
1108 client_destroy_import(exp->exp_imp_reverse);
1110 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1111 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1112 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1113 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1117 * Recovery functions
1119 static void target_request_copy_get(struct ptlrpc_request *req)
1121 class_export_rpc_get(req->rq_export);
1122 LASSERT(cfs_list_empty(&req->rq_list));
1123 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1125 /* increase refcount to keep request in queue */
1126 cfs_atomic_inc(&req->rq_refcount);
1127 /** let export know it has replays to be handled */
1128 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1131 static void target_request_copy_put(struct ptlrpc_request *req)
1133 LASSERT(cfs_list_empty(&req->rq_replay_list));
1134 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1136 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1137 class_export_rpc_put(req->rq_export);
1138 ptlrpc_server_drop_request(req);
1141 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1143 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1144 struct obd_export *exp = req->rq_export;
1145 struct ptlrpc_request *reqiter;
1150 cfs_spin_lock(&exp->exp_lock);
1151 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1153 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1160 /* we expect it with RESENT and REPLAY flags */
1161 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1162 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1163 CERROR("invalid flags %x of resent replay\n",
1164 lustre_msg_get_flags(req->rq_reqmsg));
1166 cfs_list_add_tail(&req->rq_replay_list,
1167 &exp->exp_req_replay_queue);
1170 cfs_spin_unlock(&exp->exp_lock);
1174 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1176 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1177 LASSERT(req->rq_export);
1179 cfs_spin_lock(&req->rq_export->exp_lock);
1180 cfs_list_del_init(&req->rq_replay_list);
1181 cfs_spin_unlock(&req->rq_export->exp_lock);
1185 static void target_finish_recovery(struct obd_device *obd)
1188 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1191 ldlm_reprocess_all_ns(obd->obd_namespace);
1192 cfs_spin_lock(&obd->obd_recovery_task_lock);
1193 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1194 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1195 !cfs_list_empty(&obd->obd_final_req_queue)) {
1196 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1198 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1199 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1201 cfs_list_empty(&obd->obd_final_req_queue) ? \
1203 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1206 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1208 obd->obd_recovery_end = cfs_time_current_sec();
1210 /* when recovery finished, cleanup orphans on mds and ost */
1211 if (OBT(obd) && OBP(obd, postrecov)) {
1212 int rc = OBP(obd, postrecov)(obd);
1214 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1220 static void abort_req_replay_queue(struct obd_device *obd)
1222 struct ptlrpc_request *req, *n;
1223 cfs_list_t abort_list;
1225 CFS_INIT_LIST_HEAD(&abort_list);
1226 cfs_spin_lock(&obd->obd_recovery_task_lock);
1227 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1228 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1229 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1230 DEBUG_REQ(D_WARNING, req, "aborted:");
1231 req->rq_status = -ENOTCONN;
1232 if (ptlrpc_error(req)) {
1233 DEBUG_REQ(D_ERROR, req,
1234 "failed abort_req_reply; skipping");
1236 target_exp_dequeue_req_replay(req);
1237 target_request_copy_put(req);
1241 static void abort_lock_replay_queue(struct obd_device *obd)
1243 struct ptlrpc_request *req, *n;
1244 cfs_list_t abort_list;
1246 CFS_INIT_LIST_HEAD(&abort_list);
1247 cfs_spin_lock(&obd->obd_recovery_task_lock);
1248 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1249 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1250 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1251 DEBUG_REQ(D_ERROR, req, "aborted:");
1252 req->rq_status = -ENOTCONN;
1253 if (ptlrpc_error(req)) {
1254 DEBUG_REQ(D_ERROR, req,
1255 "failed abort_lock_reply; skipping");
1257 target_request_copy_put(req);
1262 /* Called from a cleanup function if the device is being cleaned up
1263 forcefully. The exports should all have been disconnected already,
1264 the only thing left to do is
1265 - clear the recovery flags
1267 - free queued requests and replies, but don't send replies
1268 Because the obd_stopping flag is set, no new requests should be received.
1271 void target_cleanup_recovery(struct obd_device *obd)
1273 struct ptlrpc_request *req, *n;
1274 cfs_list_t clean_list;
1277 CFS_INIT_LIST_HEAD(&clean_list);
1278 cfs_spin_lock(&obd->obd_dev_lock);
1279 if (!obd->obd_recovering) {
1280 cfs_spin_unlock(&obd->obd_dev_lock);
1284 obd->obd_recovering = obd->obd_abort_recovery = 0;
1285 cfs_spin_unlock(&obd->obd_dev_lock);
1287 cfs_spin_lock(&obd->obd_recovery_task_lock);
1288 target_cancel_recovery_timer(obd);
1289 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1290 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1292 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1293 LASSERT(req->rq_reply_state == 0);
1294 target_exp_dequeue_req_replay(req);
1295 target_request_copy_put(req);
1298 cfs_spin_lock(&obd->obd_recovery_task_lock);
1299 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1300 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1301 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1303 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1304 LASSERT(req->rq_reply_state == 0);
1305 target_request_copy_put(req);
1311 /* obd_recovery_task_lock should be held */
1312 void target_cancel_recovery_timer(struct obd_device *obd)
1314 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1315 cfs_timer_disarm(&obd->obd_recovery_timer);
1318 /* extend = 1 means require at least "duration" seconds left in the timer,
1319 extend = 0 means set the total duration (start_recovery_timer) */
1320 static void reset_recovery_timer(struct obd_device *obd, int duration,
1323 cfs_time_t now = cfs_time_current_sec();
1324 cfs_duration_t left;
1326 cfs_spin_lock(&obd->obd_recovery_task_lock);
1327 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1328 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1332 left = cfs_time_sub(obd->obd_recovery_end, now);
1334 if (extend && (duration > left))
1335 obd->obd_recovery_timeout += duration - left;
1336 else if (!extend && (duration > obd->obd_recovery_timeout))
1337 /* Track the client's largest expected replay time */
1338 obd->obd_recovery_timeout = duration;
1340 /* Hard limit of obd_recovery_time_hard which should not happen */
1341 if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1342 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1344 obd->obd_recovery_end = obd->obd_recovery_start +
1345 obd->obd_recovery_timeout;
1346 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1347 cfs_time_before(now, obd->obd_recovery_end)) {
1348 left = cfs_time_sub(obd->obd_recovery_end, now);
1349 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1351 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1352 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1353 obd->obd_name, (unsigned)left);
1356 static void check_and_start_recovery_timer(struct obd_device *obd)
1358 cfs_spin_lock(&obd->obd_recovery_task_lock);
1359 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1360 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1363 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1364 obd->obd_recovery_start = cfs_time_current_sec();
1365 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1367 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1370 /* Reset the timer with each new client connection */
1372 * This timer is actually reconnect_timer, which is for making sure
1373 * the total recovery window is at least as big as my reconnect
1374 * attempt timing. So the initial recovery time_out will be set to
1375 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1376 * from client is bigger than this, then the recovery time_out will
1377 * be extended to make sure the client could be reconnected, in the
1378 * process, the timeout from the new client should be ignored.
1382 target_start_and_reset_recovery_timer(struct obd_device *obd,
1383 struct ptlrpc_request *req,
1386 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1388 if (!new_client && service_time)
1389 /* Teach server about old server's estimates, as first guess
1390 * at how long new requests will take. */
1391 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1394 check_and_start_recovery_timer(obd);
1396 /* convert the service time to rpc timeout,
1397 * reuse service_time to limit stack usage */
1398 service_time = at_est2timeout(service_time);
1400 /* We expect other clients to timeout within service_time, then try
1401 * to reconnect, then try the failover server. The max delay between
1402 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1403 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1404 INITIAL_CONNECT_TIMEOUT);
1405 if (service_time > obd->obd_recovery_timeout && !new_client)
1406 reset_recovery_timer(obd, service_time, 0);
1411 /** Health checking routines */
1412 static inline int exp_connect_healthy(struct obd_export *exp)
1414 return (exp->exp_in_recovery);
1417 /** if export done req_replay or has replay in queue */
1418 static inline int exp_req_replay_healthy(struct obd_export *exp)
1420 return (!exp->exp_req_replay_needed ||
1421 cfs_atomic_read(&exp->exp_replay_count) > 0);
1423 /** if export done lock_replay or has replay in queue */
1424 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1426 return (!exp->exp_lock_replay_needed ||
1427 cfs_atomic_read(&exp->exp_replay_count) > 0);
1430 static inline int exp_vbr_healthy(struct obd_export *exp)
1432 return (!exp->exp_vbr_failed);
1435 static inline int exp_finished(struct obd_export *exp)
1437 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1440 /** Checking routines for recovery */
1441 static int check_for_clients(struct obd_device *obd)
1443 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1445 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1446 if (obd->obd_no_conn == 0 &&
1447 obd->obd_connected_clients + obd->obd_stale_clients ==
1448 obd->obd_max_recoverable_clients)
1453 static int check_for_next_transno(struct obd_device *obd)
1455 struct ptlrpc_request *req = NULL;
1456 int wake_up = 0, connected, completed, queue_len;
1457 __u64 next_transno, req_transno;
1460 cfs_spin_lock(&obd->obd_recovery_task_lock);
1461 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1462 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1463 struct ptlrpc_request, rq_list);
1464 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1469 connected = obd->obd_connected_clients;
1470 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1471 queue_len = obd->obd_requests_queued_for_recovery;
1472 next_transno = obd->obd_next_recovery_transno;
1474 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1475 "req_transno: "LPU64", next_transno: "LPU64"\n",
1476 obd->obd_max_recoverable_clients, connected, completed,
1477 queue_len, req_transno, next_transno);
1479 if (obd->obd_abort_recovery) {
1480 CDEBUG(D_HA, "waking for aborted recovery\n");
1482 } else if (obd->obd_recovery_expired) {
1483 CDEBUG(D_HA, "waking for expired recovery\n");
1485 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1486 CDEBUG(D_HA, "waking for completed recovery\n");
1488 } else if (req_transno == next_transno) {
1489 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1491 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1493 /** handle gaps occured due to lost reply or VBR */
1494 LASSERTF(req_transno >= next_transno,
1495 "req_transno: "LPU64", next_transno: "LPU64"\n",
1496 req_transno, next_transno);
1497 if (req_transno > obd->obd_last_committed &&
1498 !obd->obd_version_recov)
1501 "%s: waking for gap in transno, VBR is %s (skip: "
1502 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1503 ", last_committed: "LPD64")\n",
1504 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1505 next_transno, queue_len, completed, connected,
1506 req_transno, obd->obd_last_committed);
1507 obd->obd_next_recovery_transno = req_transno;
1509 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1510 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1511 " by fail_lock, waking up ("LPD64")\n", next_transno);
1512 obd->obd_next_recovery_transno = req_transno;
1515 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1519 static int check_for_next_lock(struct obd_device *obd)
1523 cfs_spin_lock(&obd->obd_recovery_task_lock);
1524 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1525 CDEBUG(D_HA, "waking for next lock\n");
1527 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1528 CDEBUG(D_HA, "waking for completed lock replay\n");
1530 } else if (obd->obd_abort_recovery) {
1531 CDEBUG(D_HA, "waking for aborted recovery\n");
1533 } else if (obd->obd_recovery_expired) {
1534 CDEBUG(D_HA, "waking for expired recovery\n");
1537 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1543 * wait for recovery events,
1544 * check its status with help of check_routine
1545 * evict dead clients via health_check
1547 static int target_recovery_overseer(struct obd_device *obd,
1548 int (*check_routine)(struct obd_device *),
1549 int (*health_check)(struct obd_export *))
1552 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1553 if (obd->obd_abort_recovery) {
1554 CWARN("recovery is aborted, evict exports in recovery\n");
1555 /** evict exports which didn't finish recovery yet */
1556 class_disconnect_stale_exports(obd, exp_finished);
1558 } else if (obd->obd_recovery_expired) {
1559 obd->obd_recovery_expired = 0;
1560 /** If some clients died being recovered, evict them */
1562 "recovery is timed out, evict stale exports\n");
1563 /** evict cexports with no replay in queue, they are stalled */
1564 class_disconnect_stale_exports(obd, health_check);
1565 /** continue with VBR */
1566 cfs_spin_lock(&obd->obd_dev_lock);
1567 obd->obd_version_recov = 1;
1568 cfs_spin_unlock(&obd->obd_dev_lock);
1570 * reset timer, recovery will proceed with versions now,
1571 * timeout is set just to handle reconnection delays
1573 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1574 /** Wait for recovery events again, after evicting bad clients */
1580 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1582 struct ptlrpc_request *req = NULL;
1585 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1586 obd->obd_next_recovery_transno);
1588 if (target_recovery_overseer(obd, check_for_next_transno,
1589 exp_req_replay_healthy)) {
1590 abort_req_replay_queue(obd);
1591 abort_lock_replay_queue(obd);
1594 cfs_spin_lock(&obd->obd_recovery_task_lock);
1595 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1596 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1597 struct ptlrpc_request, rq_list);
1598 cfs_list_del_init(&req->rq_list);
1599 obd->obd_requests_queued_for_recovery--;
1600 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1602 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1603 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1604 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1605 /** evict exports failed VBR */
1606 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1611 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1613 struct ptlrpc_request *req = NULL;
1615 CDEBUG(D_HA, "Waiting for lock\n");
1616 if (target_recovery_overseer(obd, check_for_next_lock,
1617 exp_lock_replay_healthy))
1618 abort_lock_replay_queue(obd);
1620 cfs_spin_lock(&obd->obd_recovery_task_lock);
1621 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1622 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1623 struct ptlrpc_request, rq_list);
1624 cfs_list_del_init(&req->rq_list);
1625 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1627 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1628 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1629 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1630 /** evict exports failed VBR */
1631 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1636 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1638 struct ptlrpc_request *req = NULL;
1640 cfs_spin_lock(&obd->obd_recovery_task_lock);
1641 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1642 req = cfs_list_entry(obd->obd_final_req_queue.next,
1643 struct ptlrpc_request, rq_list);
1644 cfs_list_del_init(&req->rq_list);
1645 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1646 if (req->rq_export->exp_in_recovery) {
1647 cfs_spin_lock(&req->rq_export->exp_lock);
1648 req->rq_export->exp_in_recovery = 0;
1649 cfs_spin_unlock(&req->rq_export->exp_lock);
1652 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1657 static int handle_recovery_req(struct ptlrpc_thread *thread,
1658 struct ptlrpc_request *req,
1659 svc_handler_t handler)
1665 * export can be evicted during recovery, no need to handle replays for
1666 * it after that, discard such request silently
1668 if (req->rq_export->exp_disconnected)
1669 GOTO(reqcopy_put, rc = 0);
1671 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1673 CERROR("Failure to initialize session: %d\n", rc);
1674 GOTO(reqcopy_put, rc);
1677 req->rq_recov_session.lc_thread = thread;
1678 lu_context_enter(&req->rq_recov_session);
1679 req->rq_svc_thread = thread;
1680 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1682 /* thread context */
1683 lu_context_enter(&thread->t_env->le_ctx);
1685 lu_context_exit(&thread->t_env->le_ctx);
1687 lu_context_exit(&req->rq_recov_session);
1688 lu_context_fini(&req->rq_recov_session);
1689 /* don't reset timer for final stage */
1690 if (!exp_finished(req->rq_export)) {
1692 * Add request timeout to the recovery time so next request from
1693 * this client may come in recovery time
1695 reset_recovery_timer(class_exp2obd(req->rq_export),
1696 AT_OFF ? obd_timeout :
1697 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1703 static int target_recovery_thread(void *arg)
1705 struct lu_target *lut = arg;
1706 struct obd_device *obd = lut->lut_obd;
1707 struct ptlrpc_request *req;
1708 struct target_recovery_data *trd = &obd->obd_recovery_data;
1709 unsigned long delta;
1710 unsigned long flags;
1712 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1716 cfs_daemonize_ctxt("tgt_recov");
1718 SIGNAL_MASK_LOCK(current, flags);
1719 sigfillset(¤t->blocked);
1721 SIGNAL_MASK_UNLOCK(current, flags);
1723 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1727 thread->t_env = &env;
1728 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1729 env.le_ctx.lc_thread = thread;
1730 thread->t_data = NULL;
1732 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1734 trd->trd_processing_task = cfs_curproc_pid();
1736 cfs_spin_lock(&obd->obd_dev_lock);
1737 obd->obd_recovering = 1;
1738 cfs_spin_unlock(&obd->obd_dev_lock);
1739 cfs_complete(&trd->trd_starting);
1741 /* first of all, we have to know the first transno to replay */
1742 if (target_recovery_overseer(obd, check_for_clients,
1743 exp_connect_healthy)) {
1744 abort_req_replay_queue(obd);
1745 abort_lock_replay_queue(obd);
1748 /* next stage: replay requests */
1750 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1751 cfs_atomic_read(&obd->obd_req_replay_clients),
1752 obd->obd_next_recovery_transno);
1753 while ((req = target_next_replay_req(obd))) {
1754 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1755 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1756 lustre_msg_get_transno(req->rq_reqmsg),
1757 libcfs_nid2str(req->rq_peer.nid));
1758 handle_recovery_req(thread, req,
1759 trd->trd_recovery_handler);
1761 * bz18031: increase next_recovery_transno before
1762 * target_request_copy_put() will drop exp_rpc reference
1764 cfs_spin_lock(&obd->obd_recovery_task_lock);
1765 obd->obd_next_recovery_transno++;
1766 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1767 target_exp_dequeue_req_replay(req);
1768 target_request_copy_put(req);
1769 obd->obd_replayed_requests++;
1773 * The second stage: replay locks
1775 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1776 cfs_atomic_read(&obd->obd_lock_replay_clients));
1777 while ((req = target_next_replay_lock(obd))) {
1778 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1779 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1780 libcfs_nid2str(req->rq_peer.nid));
1781 handle_recovery_req(thread, req,
1782 trd->trd_recovery_handler);
1783 target_request_copy_put(req);
1784 obd->obd_replayed_locks++;
1788 * The third stage: reply on final pings, at this moment all clients
1789 * must have request in final queue
1791 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1792 /** Update server last boot epoch */
1793 lut_boot_epoch_update(lut);
1794 /* We drop recoverying flag to forward all new requests
1795 * to regular mds_handle() since now */
1796 cfs_spin_lock(&obd->obd_dev_lock);
1797 obd->obd_recovering = obd->obd_abort_recovery = 0;
1798 cfs_spin_unlock(&obd->obd_dev_lock);
1799 cfs_spin_lock(&obd->obd_recovery_task_lock);
1800 target_cancel_recovery_timer(obd);
1801 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1802 while ((req = target_next_final_ping(obd))) {
1803 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1804 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1805 libcfs_nid2str(req->rq_peer.nid));
1806 handle_recovery_req(thread, req,
1807 trd->trd_recovery_handler);
1808 target_request_copy_put(req);
1811 delta = (jiffies - delta) / CFS_HZ;
1812 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1813 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1814 if (delta > OBD_RECOVERY_TIME_SOFT) {
1815 CWARN("too long recovery - read logs\n");
1816 libcfs_debug_dumplog();
1819 target_finish_recovery(obd);
1821 lu_context_fini(&env.le_ctx);
1822 trd->trd_processing_task = 0;
1823 cfs_complete(&trd->trd_finishing);
1827 static int target_start_recovery_thread(struct lu_target *lut,
1828 svc_handler_t handler)
1830 struct obd_device *obd = lut->lut_obd;
1832 struct target_recovery_data *trd = &obd->obd_recovery_data;
1834 memset(trd, 0, sizeof(*trd));
1835 cfs_init_completion(&trd->trd_starting);
1836 cfs_init_completion(&trd->trd_finishing);
1837 trd->trd_recovery_handler = handler;
1839 if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
1840 cfs_wait_for_completion(&trd->trd_starting);
1841 LASSERT(obd->obd_recovering != 0);
1848 void target_stop_recovery_thread(struct obd_device *obd)
1850 if (obd->obd_recovery_data.trd_processing_task > 0) {
1851 struct target_recovery_data *trd = &obd->obd_recovery_data;
1852 /** recovery can be done but postrecovery is not yet */
1853 cfs_spin_lock(&obd->obd_dev_lock);
1854 if (obd->obd_recovering) {
1855 CERROR("%s: Aborting recovery\n", obd->obd_name);
1856 obd->obd_abort_recovery = 1;
1857 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1859 cfs_spin_unlock(&obd->obd_dev_lock);
1860 cfs_wait_for_completion(&trd->trd_finishing);
1864 void target_recovery_fini(struct obd_device *obd)
1866 class_disconnect_exports(obd);
1867 target_stop_recovery_thread(obd);
1868 target_cleanup_recovery(obd);
1870 EXPORT_SYMBOL(target_recovery_fini);
1872 static void target_recovery_expired(unsigned long castmeharder)
1874 struct obd_device *obd = (struct obd_device *)castmeharder;
1875 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1876 " after %lds (%d clients connected)\n",
1877 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1878 cfs_time_current_sec()- obd->obd_recovery_start,
1879 obd->obd_connected_clients);
1881 obd->obd_recovery_expired = 1;
1882 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1885 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1887 struct obd_device *obd = lut->lut_obd;
1888 if (obd->obd_max_recoverable_clients == 0) {
1889 /** Update server last boot epoch */
1890 lut_boot_epoch_update(lut);
1894 CWARN("RECOVERY: service %s, %d recoverable clients, "
1895 "last_transno "LPU64"\n", obd->obd_name,
1896 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1897 LASSERT(obd->obd_stopping == 0);
1898 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1899 obd->obd_recovery_start = 0;
1900 obd->obd_recovery_end = 0;
1902 /* both values can be get from mount data already */
1903 if (obd->obd_recovery_timeout == 0)
1904 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
1905 if (obd->obd_recovery_time_hard == 0)
1906 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
1907 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1908 target_start_recovery_thread(lut, handler);
1910 EXPORT_SYMBOL(target_recovery_init);
1914 static int target_process_req_flags(struct obd_device *obd,
1915 struct ptlrpc_request *req)
1917 struct obd_export *exp = req->rq_export;
1918 LASSERT(exp != NULL);
1919 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1920 /* client declares he's ready to replay locks */
1921 cfs_spin_lock(&exp->exp_lock);
1922 if (exp->exp_req_replay_needed) {
1923 exp->exp_req_replay_needed = 0;
1924 cfs_spin_unlock(&exp->exp_lock);
1926 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
1927 cfs_atomic_dec(&obd->obd_req_replay_clients);
1929 cfs_spin_unlock(&exp->exp_lock);
1932 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1933 /* client declares he's ready to complete recovery
1934 * so, we put the request on th final queue */
1935 cfs_spin_lock(&exp->exp_lock);
1936 if (exp->exp_lock_replay_needed) {
1937 exp->exp_lock_replay_needed = 0;
1938 cfs_spin_unlock(&exp->exp_lock);
1940 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
1941 cfs_atomic_dec(&obd->obd_lock_replay_clients);
1943 cfs_spin_unlock(&exp->exp_lock);
1949 int target_queue_recovery_request(struct ptlrpc_request *req,
1950 struct obd_device *obd)
1954 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1957 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1958 /* Processing the queue right now, don't re-add. */
1962 target_process_req_flags(obd, req);
1964 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1965 /* client declares he's ready to complete recovery
1966 * so, we put the request on th final queue */
1967 target_request_copy_get(req);
1968 DEBUG_REQ(D_HA, req, "queue final req");
1969 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1970 cfs_spin_lock(&obd->obd_recovery_task_lock);
1971 if (obd->obd_recovering) {
1972 cfs_list_add_tail(&req->rq_list,
1973 &obd->obd_final_req_queue);
1975 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1976 target_request_copy_put(req);
1977 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1979 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1982 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1983 /* client declares he's ready to replay locks */
1984 target_request_copy_get(req);
1985 DEBUG_REQ(D_HA, req, "queue lock replay req");
1986 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1987 cfs_spin_lock(&obd->obd_recovery_task_lock);
1988 LASSERT(obd->obd_recovering);
1989 /* usually due to recovery abort */
1990 if (!req->rq_export->exp_in_recovery) {
1991 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1992 target_request_copy_put(req);
1995 LASSERT(req->rq_export->exp_lock_replay_needed);
1996 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1997 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2001 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2002 * (i.e. buflens etc are in my own byte order), but type-dependent
2003 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2006 CFS_INIT_LIST_HEAD(&req->rq_list);
2007 DEBUG_REQ(D_HA, req, "not queueing");
2011 /* If we're processing the queue, we want don't want to queue this
2014 * Also, if this request has a transno less than the one we're waiting
2015 * for, we should process it now. It could (and currently always will)
2016 * be an open request for a descriptor that was opened some time ago.
2018 * Also, a resent, replayed request that has already been
2019 * handled will pass through here and be processed immediately.
2021 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2022 obd->obd_next_recovery_transno, transno);
2023 cfs_spin_lock(&obd->obd_recovery_task_lock);
2024 if (transno < obd->obd_next_recovery_transno) {
2025 /* Processing the queue right now, don't re-add. */
2026 LASSERT(cfs_list_empty(&req->rq_list));
2027 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2030 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2032 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2035 target_request_copy_get(req);
2036 if (!req->rq_export->exp_in_recovery) {
2037 target_request_copy_put(req);
2040 LASSERT(req->rq_export->exp_req_replay_needed);
2042 if (target_exp_enqueue_req_replay(req)) {
2043 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2044 target_request_copy_put(req);
2049 cfs_spin_lock(&obd->obd_recovery_task_lock);
2050 LASSERT(obd->obd_recovering);
2051 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2052 struct ptlrpc_request *reqiter =
2053 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2055 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2056 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2061 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2063 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2064 "has been claimed by another client");
2065 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2066 target_exp_dequeue_req_replay(req);
2067 target_request_copy_put(req);
2073 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2075 obd->obd_requests_queued_for_recovery++;
2076 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2077 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2082 * Packs current SLV and Limit into \a req.
2084 int target_pack_pool_reply(struct ptlrpc_request *req)
2086 struct obd_device *obd;
2090 * Check that we still have all structures alive as this may
2091 * be some late rpc in shutdown time.
2093 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2094 !exp_connect_lru_resize(req->rq_export))) {
2095 lustre_msg_set_slv(req->rq_repmsg, 0);
2096 lustre_msg_set_limit(req->rq_repmsg, 0);
2101 * OBD is alive here as export is alive, which we checked above.
2103 obd = req->rq_export->exp_obd;
2105 cfs_read_lock(&obd->obd_pool_lock);
2106 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2107 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2108 cfs_read_unlock(&obd->obd_pool_lock);
2113 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2115 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2116 DEBUG_REQ(D_ERROR, req, "dropping reply");
2121 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2122 req->rq_status = rc;
2123 return (ptlrpc_send_error(req, 1));
2125 DEBUG_REQ(D_NET, req, "sending reply");
2128 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2131 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2134 struct ptlrpc_reply_state *rs;
2135 struct obd_device *obd;
2136 struct obd_export *exp;
2137 struct ptlrpc_service *svc;
2140 if (req->rq_no_reply) {
2145 svc = req->rq_rqbd->rqbd_service;
2146 rs = req->rq_reply_state;
2147 if (rs == NULL || !rs->rs_difficult) {
2149 target_send_reply_msg (req, rc, fail_id);
2154 /* must be an export if locks saved */
2155 LASSERT (req->rq_export != NULL);
2156 /* req/reply consistent */
2157 LASSERT (rs->rs_service == svc);
2160 LASSERT (!rs->rs_scheduled);
2161 LASSERT (!rs->rs_scheduled_ever);
2162 LASSERT (!rs->rs_handled);
2163 LASSERT (!rs->rs_on_net);
2164 LASSERT (rs->rs_export == NULL);
2165 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2166 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2168 exp = class_export_get (req->rq_export);
2171 /* disable reply scheduling while I'm setting up */
2172 rs->rs_scheduled = 1;
2174 rs->rs_xid = req->rq_xid;
2175 rs->rs_transno = req->rq_transno;
2176 rs->rs_export = exp;
2177 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2179 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2180 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2181 rs->rs_transno, exp->exp_last_committed);
2182 if (rs->rs_transno > exp->exp_last_committed) {
2183 /* not committed already */
2184 cfs_list_add_tail(&rs->rs_obd_list,
2185 &exp->exp_uncommitted_replies);
2187 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2189 cfs_spin_lock(&exp->exp_lock);
2190 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2191 cfs_spin_unlock(&exp->exp_lock);
2193 netrc = target_send_reply_msg (req, rc, fail_id);
2195 cfs_spin_lock(&svc->srv_rs_lock);
2197 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2200 /* error sending: reply is off the net. Also we need +1
2201 * reply ref until ptlrpc_handle_rs() is done
2202 * with the reply state (if the send was successful, there
2203 * would have been +1 ref for the net, which
2204 * reply_out_callback leaves alone) */
2206 ptlrpc_rs_addref(rs);
2209 cfs_spin_lock(&rs->rs_lock);
2210 if (rs->rs_transno <= exp->exp_last_committed ||
2211 (!rs->rs_on_net && !rs->rs_no_ack) ||
2212 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2213 cfs_list_empty(&rs->rs_obd_list)) {
2214 CDEBUG(D_HA, "Schedule reply immediately\n");
2215 ptlrpc_dispatch_difficult_reply(rs);
2217 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2218 rs->rs_scheduled = 0; /* allow notifier to schedule */
2220 cfs_spin_unlock(&rs->rs_lock);
2221 cfs_spin_unlock(&svc->srv_rs_lock);
2225 int target_handle_ping(struct ptlrpc_request *req)
2227 obd_ping(req->rq_export);
2228 return req_capsule_server_pack(&req->rq_pill);
2231 void target_committed_to_req(struct ptlrpc_request *req)
2233 struct obd_export *exp = req->rq_export;
2235 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2236 lustre_msg_set_last_committed(req->rq_repmsg,
2237 exp->exp_last_committed);
2239 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2240 "%d)", exp->exp_obd->obd_no_transno,
2241 req->rq_repmsg == NULL);
2243 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2244 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2246 EXPORT_SYMBOL(target_committed_to_req);
2248 int target_handle_qc_callback(struct ptlrpc_request *req)
2250 struct obd_quotactl *oqctl;
2251 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2253 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2254 if (oqctl == NULL) {
2255 CERROR("Can't unpack obd_quotactl\n");
2259 cli->cl_qchk_stat = oqctl->qc_stat;
2264 #ifdef HAVE_QUOTA_SUPPORT
2265 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2268 struct obd_device *obd = req->rq_export->exp_obd;
2269 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2270 struct obd_device_target *obt;
2271 struct lustre_quota_ctxt *qctxt;
2272 struct qunit_data *qdata = NULL;
2276 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2279 rc = req_capsule_server_pack(&req->rq_pill);
2281 CERROR("packing reply failed!: rc = %d\n", rc);
2285 LASSERT(req->rq_export);
2287 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2288 if (IS_ERR(qdata)) {
2289 rc = PTR_ERR(qdata);
2290 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2291 req->rq_status = rc;
2295 /* we use the observer */
2296 if (obd_pin_observer(obd, &lov_obd) ||
2297 obd_pin_observer(lov_obd, &master_obd)) {
2298 CERROR("Can't find the observer, it is recovering\n");
2299 req->rq_status = -EAGAIN;
2303 obt = &master_obd->u.obt;
2304 qctxt = &obt->obt_qctxt;
2306 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2307 /* quota_type has not been processed yet, return EAGAIN
2308 * until we know whether or not quotas are supposed to
2310 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2312 req->rq_status = -EAGAIN;
2316 cfs_down_read(&obt->obt_rwsem);
2317 if (qctxt->lqc_lqs_hash == NULL) {
2318 cfs_up_read(&obt->obt_rwsem);
2319 /* quota_type has not been processed yet, return EAGAIN
2320 * until we know whether or not quotas are supposed to
2322 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2324 req->rq_status = -EAGAIN;
2328 LASSERT(qctxt->lqc_handler);
2329 rc = qctxt->lqc_handler(master_obd, qdata,
2330 lustre_msg_get_opc(req->rq_reqmsg));
2331 cfs_up_read(&obt->obt_rwsem);
2332 if (rc && rc != -EDQUOT)
2333 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2334 "dqacq/dqrel failed! (rc:%d)\n", rc);
2335 req->rq_status = rc;
2337 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2339 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2343 /* Block the quota req. b=14840 */
2344 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2349 obd_unpin_observer(lov_obd);
2351 obd_unpin_observer(obd);
2353 rc = ptlrpc_reply(req);
2357 #endif /* !__KERNEL__ */
2359 #endif /* HAVE_QUOTA_SUPPORT */
2361 ldlm_mode_t lck_compat_array[] = {
2362 [LCK_EX] LCK_COMPAT_EX,
2363 [LCK_PW] LCK_COMPAT_PW,
2364 [LCK_PR] LCK_COMPAT_PR,
2365 [LCK_CW] LCK_COMPAT_CW,
2366 [LCK_CR] LCK_COMPAT_CR,
2367 [LCK_NL] LCK_COMPAT_NL,
2368 [LCK_GROUP] LCK_COMPAT_GROUP,
2369 [LCK_COS] LCK_COMPAT_COS,
2373 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2374 * not escape to the user level.
2376 int ldlm_error2errno(ldlm_error_t error)
2384 case ELDLM_LOCK_CHANGED:
2387 case ELDLM_LOCK_ABORTED:
2390 case ELDLM_LOCK_REPLACED:
2393 case ELDLM_NO_LOCK_DATA:
2396 case ELDLM_NAMESPACE_EXISTS:
2399 case ELDLM_BAD_NAMESPACE:
2403 if (((int)error) < 0) /* cast to signed type */
2404 result = error; /* as ldlm_error_t can be unsigned */
2406 CERROR("Invalid DLM result code: %d\n", error);
2412 EXPORT_SYMBOL(ldlm_error2errno);
2415 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2417 ldlm_error_t ldlm_errno2error(int err_no)
2426 error = ELDLM_LOCK_CHANGED;
2429 error = ELDLM_LOCK_ABORTED;
2432 error = ELDLM_LOCK_REPLACED;
2435 error = ELDLM_NO_LOCK_DATA;
2438 error = ELDLM_NAMESPACE_EXISTS;
2441 error = ELDLM_BAD_NAMESPACE;
2448 EXPORT_SYMBOL(ldlm_errno2error);
2450 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2451 void ldlm_dump_export_locks(struct obd_export *exp)
2453 cfs_spin_lock(&exp->exp_locks_list_guard);
2454 if (!cfs_list_empty(&exp->exp_locks_list)) {
2455 struct ldlm_lock *lock;
2457 CERROR("dumping locks for export %p,"
2458 "ignore if the unmount doesn't hang\n", exp);
2459 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2460 ldlm_lock_dump(D_ERROR, lock, 0);
2462 cfs_spin_unlock(&exp->exp_locks_list_guard);
2466 static int target_bulk_timeout(void *data)
2469 /* We don't fail the connection here, because having the export
2470 * killed makes the (vital) call to commitrw very sad.
2475 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2477 return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2480 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2481 struct l_wait_info *lwi)
2483 struct ptlrpc_request *req = desc->bd_req;
2487 /* Check if there is eviction in progress, and if so, wait for
2489 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2490 *lwi = LWI_INTR(NULL, NULL);
2491 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2492 !cfs_atomic_read(&exp->exp_obd->
2493 obd_evict_inprogress),
2497 /* Check if client was evicted or tried to reconnect already */
2498 if (exp->exp_failed || exp->exp_abort_active_req) {
2501 if (desc->bd_type == BULK_PUT_SINK)
2502 rc = sptlrpc_svc_wrap_bulk(req, desc);
2504 rc = ptlrpc_start_bulk_transfer(desc);
2507 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2508 ptlrpc_abort_bulk(desc);
2509 } else if (rc == 0) {
2510 time_t start = cfs_time_current_sec();
2512 long timeoutl = req->rq_deadline - cfs_time_current_sec();
2513 cfs_duration_t timeout = timeoutl <= 0 ?
2514 CFS_TICK : cfs_time_seconds(timeoutl);
2515 *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2516 cfs_time_seconds(1),
2517 target_bulk_timeout,
2519 rc = l_wait_event(desc->bd_waitq,
2520 !ptlrpc_server_bulk_active(desc) ||
2522 exp->exp_abort_active_req,
2524 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2525 /* Wait again if we changed deadline */
2526 } while ((rc == -ETIMEDOUT) &&
2527 (req->rq_deadline > cfs_time_current_sec()));
2529 if (rc == -ETIMEDOUT) {
2530 DEBUG_REQ(D_ERROR, req,
2531 "timeout on bulk %s after %ld%+lds",
2533 req->rq_deadline - start,
2534 cfs_time_current_sec() -
2536 ptlrpc_abort_bulk(desc);
2537 } else if (exp->exp_failed) {
2538 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2541 ptlrpc_abort_bulk(desc);
2542 } else if (exp->exp_abort_active_req) {
2543 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2545 /* we don't reply anyway */
2547 ptlrpc_abort_bulk(desc);
2548 } else if (!desc->bd_success ||
2549 desc->bd_nob_transferred != desc->bd_nob) {
2550 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2552 "truncated" : "network error on",
2554 desc->bd_nob_transferred,
2556 /* XXX should this be a different errno? */
2558 } else if (desc->bd_type == BULK_GET_SINK) {
2559 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2562 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2563 bulk2type(desc), rc);
2568 EXPORT_SYMBOL(target_bulk_io);