1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_mds.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: if non-zero, move the selected to the list head
55 * @create: if zero, only search in existed connections
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 cfs_spin_lock(&imp->imp_lock);
84 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 cfs_list_del(&item->oic_item);
88 cfs_list_add(&item->oic_item,
90 item->oic_last_attempt = 0;
92 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? ", moved to head" : ""));
95 cfs_spin_unlock(&imp->imp_lock);
96 GOTO(out_free, rc = 0);
101 imp_conn->oic_conn = ptlrpc_conn;
102 imp_conn->oic_uuid = *uuid;
103 imp_conn->oic_last_attempt = 0;
105 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107 cfs_list_add_tail(&imp_conn->oic_item,
108 &imp->imp_conn_list);
109 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
110 imp, imp->imp_obd->obd_name, uuid->uuid,
111 (priority ? "head" : "tail"));
113 cfs_spin_unlock(&imp->imp_lock);
114 GOTO(out_free, rc = -ENOENT);
117 cfs_spin_unlock(&imp->imp_lock);
121 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 ptlrpc_connection_put(ptlrpc_conn);
127 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 return import_set_conn(imp, uuid, 1, 0);
132 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
135 return import_set_conn(imp, uuid, priority, 1);
138 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
140 struct obd_import_conn *imp_conn;
141 struct obd_export *dlmexp;
145 cfs_spin_lock(&imp->imp_lock);
146 if (cfs_list_empty(&imp->imp_conn_list)) {
147 LASSERT(!imp->imp_connection);
151 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
152 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
154 LASSERT(imp_conn->oic_conn);
156 /* is current conn? */
157 if (imp_conn == imp->imp_conn_current) {
158 LASSERT(imp_conn->oic_conn == imp->imp_connection);
160 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
161 imp->imp_state != LUSTRE_IMP_DISCON) {
162 CERROR("can't remove current connection\n");
163 GOTO(out, rc = -EBUSY);
166 ptlrpc_connection_put(imp->imp_connection);
167 imp->imp_connection = NULL;
169 dlmexp = class_conn2export(&imp->imp_dlm_handle);
170 if (dlmexp && dlmexp->exp_connection) {
171 LASSERT(dlmexp->exp_connection ==
173 ptlrpc_connection_put(dlmexp->exp_connection);
174 dlmexp->exp_connection = NULL;
178 cfs_list_del(&imp_conn->oic_item);
179 ptlrpc_connection_put(imp_conn->oic_conn);
180 OBD_FREE(imp_conn, sizeof(*imp_conn));
181 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182 imp, imp->imp_obd->obd_name, uuid->uuid);
187 cfs_spin_unlock(&imp->imp_lock);
189 CERROR("connection %s not found\n", uuid->uuid);
193 void client_destroy_import(struct obd_import *imp)
195 /* drop security policy instance after all rpc finished/aborted
196 * to let all busy contexts be released. */
197 class_import_get(imp);
198 class_destroy_import(imp);
199 sptlrpc_import_sec_put(imp);
200 class_import_put(imp);
203 /* configure an RPC client OBD device
208 * 3 - inactive-on-startup
210 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
212 struct client_obd *cli = &obddev->u.cli;
213 struct obd_import *imp;
214 struct obd_uuid server_uuid;
215 int rq_portal, rp_portal, connect_op;
216 char *name = obddev->obd_type->typ_name;
217 ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
221 /* In a more perfect world, we would hang a ptlrpc_client off of
222 * obd_type and just use the values from there. */
223 if (!strcmp(name, LUSTRE_OSC_NAME)) {
224 rq_portal = OST_REQUEST_PORTAL;
225 rp_portal = OSC_REPLY_PORTAL;
226 connect_op = OST_CONNECT;
227 cli->cl_sp_me = LUSTRE_SP_CLI;
228 cli->cl_sp_to = LUSTRE_SP_OST;
229 ns_type = LDLM_NS_TYPE_OSC;
231 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
232 rq_portal = MDS_REQUEST_PORTAL;
233 rp_portal = MDC_REPLY_PORTAL;
234 connect_op = MDS_CONNECT;
235 cli->cl_sp_me = LUSTRE_SP_CLI;
236 cli->cl_sp_to = LUSTRE_SP_MDT;
237 ns_type = LDLM_NS_TYPE_MDC;
239 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
240 rq_portal = MGS_REQUEST_PORTAL;
241 rp_portal = MGC_REPLY_PORTAL;
242 connect_op = MGS_CONNECT;
243 cli->cl_sp_me = LUSTRE_SP_MGC;
244 cli->cl_sp_to = LUSTRE_SP_MGS;
245 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
246 ns_type = LDLM_NS_TYPE_MGC;
249 CERROR("unknown client OBD type \"%s\", can't setup\n",
254 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
255 CERROR("requires a TARGET UUID\n");
259 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
260 CERROR("client UUID must be less than 38 characters\n");
264 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
265 CERROR("setup requires a SERVER UUID\n");
269 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
270 CERROR("target UUID must be less than 38 characters\n");
274 cfs_init_rwsem(&cli->cl_sem);
275 cfs_sema_init(&cli->cl_mgc_sem, 1);
276 cli->cl_conn_count = 0;
277 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
278 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
279 sizeof(server_uuid)));
282 cli->cl_avail_grant = 0;
283 /* FIXME: should limit this for the sum of all cl_dirty_max */
284 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
285 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
286 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
287 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
288 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
289 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
290 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
291 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
292 CFS_INIT_LIST_HEAD(&cli->cl_loi_sync_fs_list);
293 client_obd_list_lock_init(&cli->cl_loi_list_lock);
294 cli->cl_r_in_flight = 0;
295 cli->cl_w_in_flight = 0;
297 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
298 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
299 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
300 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
301 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
302 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
303 cfs_waitq_init(&cli->cl_destroy_waitq);
304 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
305 #ifdef ENABLE_CHECKSUM
306 /* Turn on checksumming by default. */
307 cli->cl_checksum = 1;
309 * The supported checksum types will be worked out at connect time
310 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
313 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
315 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
317 /* This value may be changed at connect time in
318 ptlrpc_connect_interpret. */
319 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
320 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
322 if (!strcmp(name, LUSTRE_MDC_NAME)) {
323 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
324 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
325 cli->cl_max_rpcs_in_flight = 2;
326 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
327 cli->cl_max_rpcs_in_flight = 3;
328 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
329 cli->cl_max_rpcs_in_flight = 4;
331 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
336 CERROR("ldlm_get_ref failed: %d\n", rc);
340 ptlrpc_init_client(rq_portal, rp_portal, name,
341 &obddev->obd_ldlm_client);
343 imp = class_new_import(obddev);
345 GOTO(err_ldlm, rc = -ENOENT);
346 imp->imp_client = &obddev->obd_ldlm_client;
347 imp->imp_connect_op = connect_op;
348 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
349 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
350 LUSTRE_CFG_BUFLEN(lcfg, 1));
351 class_import_put(imp);
353 rc = client_import_add_conn(imp, &server_uuid, 1);
355 CERROR("can't add initial connection\n");
356 GOTO(err_import, rc);
359 cli->cl_import = imp;
360 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
361 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
362 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
364 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
365 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
366 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
367 name, obddev->obd_name,
368 cli->cl_target_uuid.uuid);
369 cfs_spin_lock(&imp->imp_lock);
370 imp->imp_deactive = 1;
371 cfs_spin_unlock(&imp->imp_lock);
375 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
376 LDLM_NAMESPACE_CLIENT,
377 LDLM_NAMESPACE_GREEDY,
379 if (obddev->obd_namespace == NULL) {
380 CERROR("Unable to create client namespace - %s\n",
382 GOTO(err_import, rc = -ENOMEM);
385 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
390 class_destroy_import(imp);
398 int client_obd_cleanup(struct obd_device *obddev)
402 ldlm_namespace_free_post(obddev->obd_namespace);
403 obddev->obd_namespace = NULL;
409 /* ->o_connect() method for client side (OSC and MDC and MGC) */
410 int client_connect_import(const struct lu_env *env,
411 struct obd_export **exp,
412 struct obd_device *obd, struct obd_uuid *cluuid,
413 struct obd_connect_data *data, void *localdata)
415 struct client_obd *cli = &obd->u.cli;
416 struct obd_import *imp = cli->cl_import;
417 struct obd_connect_data *ocd;
418 struct lustre_handle conn = { 0 };
423 cfs_down_write(&cli->cl_sem);
424 if (cli->cl_conn_count > 0 )
425 GOTO(out_sem, rc = -EALREADY);
427 rc = class_connect(&conn, obd, cluuid);
431 cli->cl_conn_count++;
432 *exp = class_conn2export(&conn);
434 LASSERT(obd->obd_namespace);
436 imp->imp_dlm_handle = conn;
437 rc = ptlrpc_init_import(imp);
441 ocd = &imp->imp_connect_data;
444 imp->imp_connect_flags_orig = data->ocd_connect_flags;
447 rc = ptlrpc_connect_import(imp, NULL);
449 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
452 LASSERT((*exp)->exp_connection);
455 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
456 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
457 data->ocd_connect_flags, ocd->ocd_connect_flags);
458 data->ocd_connect_flags = ocd->ocd_connect_flags;
461 ptlrpc_pinger_add_import(imp);
467 cli->cl_conn_count--;
468 class_disconnect(*exp);
472 cfs_up_write(&cli->cl_sem);
477 int client_disconnect_export(struct obd_export *exp)
479 struct obd_device *obd = class_exp2obd(exp);
480 struct client_obd *cli;
481 struct obd_import *imp;
486 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
487 exp, exp ? exp->exp_handle.h_cookie : -1);
492 imp = cli->cl_import;
494 cfs_down_write(&cli->cl_sem);
495 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
498 if (!cli->cl_conn_count) {
499 CERROR("disconnecting disconnected device (%s)\n",
501 GOTO(out_disconnect, rc = -EINVAL);
504 cli->cl_conn_count--;
505 if (cli->cl_conn_count)
506 GOTO(out_disconnect, rc = 0);
508 /* Mark import deactivated now, so we don't try to reconnect if any
509 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
510 * fully deactivate the import, or that would drop all requests. */
511 cfs_spin_lock(&imp->imp_lock);
512 imp->imp_deactive = 1;
513 cfs_spin_unlock(&imp->imp_lock);
515 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
516 * delete it regardless. (It's safe to delete an import that was
518 (void)ptlrpc_pinger_del_import(imp);
520 if (obd->obd_namespace != NULL) {
521 /* obd_force == local only */
522 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
523 obd->obd_force ? LCF_LOCAL : 0, NULL);
524 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
528 * there's no need to hold sem during disconnecting an import,
529 * and actually it may cause deadlock in gss.
531 cfs_up_write(&cli->cl_sem);
532 rc = ptlrpc_disconnect_import(imp, 0);
533 cfs_down_write(&cli->cl_sem);
535 ptlrpc_invalidate_import(imp);
537 if (imp->imp_rq_pool) {
538 ptlrpc_free_rq_pool(imp->imp_rq_pool);
539 imp->imp_rq_pool = NULL;
541 client_destroy_import(imp);
542 cli->cl_import = NULL;
547 /* use server style - class_disconnect should be always called for
549 err = class_disconnect(exp);
553 cfs_up_write(&cli->cl_sem);
558 int server_disconnect_export(struct obd_export *exp)
563 /* Disconnect early so that clients can't keep using export */
564 rc = class_disconnect(exp);
565 /* close import for avoid sending any requests */
566 if (exp->exp_imp_reverse)
567 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
569 if (exp->exp_obd->obd_namespace != NULL)
570 ldlm_cancel_locks_for_export(exp);
572 /* complete all outstanding replies */
573 cfs_spin_lock(&exp->exp_lock);
574 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
575 struct ptlrpc_reply_state *rs =
576 cfs_list_entry(exp->exp_outstanding_replies.next,
577 struct ptlrpc_reply_state, rs_exp_list);
578 struct ptlrpc_service *svc = rs->rs_service;
580 cfs_spin_lock(&svc->srv_rs_lock);
581 cfs_list_del_init(&rs->rs_exp_list);
582 cfs_spin_lock(&rs->rs_lock);
583 ptlrpc_schedule_difficult_reply(rs);
584 cfs_spin_unlock(&rs->rs_lock);
585 cfs_spin_unlock(&svc->srv_rs_lock);
587 cfs_spin_unlock(&exp->exp_lock);
592 /* --------------------------------------------------------------------------
593 * from old lib/target.c
594 * -------------------------------------------------------------------------- */
596 static int target_handle_reconnect(struct lustre_handle *conn,
597 struct obd_export *exp,
598 struct obd_uuid *cluuid)
601 if (exp->exp_connection && exp->exp_imp_reverse) {
602 struct lustre_handle *hdl;
603 hdl = &exp->exp_imp_reverse->imp_remote_handle;
604 /* Might be a re-connect after a partition. */
605 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
606 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
608 conn->cookie = exp->exp_handle.h_cookie;
609 /* target_handle_connect() treats EALREADY and
610 * -EALREADY differently. EALREADY means we are
611 * doing a valid reconnect from the same client. */
614 CERROR("%s reconnecting from %s, "
615 "handle mismatch (ours "LPX64", theirs "
616 LPX64")\n", cluuid->uuid,
617 exp->exp_connection->c_remote_uuid.uuid,
618 hdl->cookie, conn->cookie);
619 memset(conn, 0, sizeof *conn);
620 /* target_handle_connect() treats EALREADY and
621 * -EALREADY differently. -EALREADY is an error
622 * (same UUID, different handle). */
627 conn->cookie = exp->exp_handle.h_cookie;
628 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
629 cluuid->uuid, exp, conn->cookie);
633 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
636 struct obd_export *exp = cb_data;
638 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
639 obd->obd_name, exp->exp_client_uuid.uuid);
641 cfs_spin_lock(&exp->exp_lock);
642 exp->exp_need_sync = 0;
643 cfs_spin_unlock(&exp->exp_lock);
644 class_export_cb_put(exp);
646 EXPORT_SYMBOL(target_client_add_cb);
649 target_start_and_reset_recovery_timer(struct obd_device *obd,
650 struct ptlrpc_request *req,
653 int target_handle_connect(struct ptlrpc_request *req)
655 struct obd_device *target, *targref = NULL;
656 struct obd_export *export = NULL;
657 struct obd_import *revimp;
658 struct lustre_handle conn;
659 struct lustre_handle *tmp;
660 struct obd_uuid tgtuuid;
661 struct obd_uuid cluuid;
662 struct obd_uuid remote_uuid;
666 struct obd_connect_data *data, *tmpdata;
667 lnet_nid_t *client_nid = NULL;
670 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
672 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
674 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
675 GOTO(out, rc = -EINVAL);
678 obd_str2uuid(&tgtuuid, str);
679 target = class_uuid2obd(&tgtuuid);
681 target = class_name2obd(str);
683 if (!target || target->obd_stopping || !target->obd_set_up) {
684 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
685 " for connect (%s)\n", str,
686 !target ? "no target" :
687 (target->obd_stopping ? "stopping" :
689 GOTO(out, rc = -ENODEV);
692 if (target->obd_no_conn) {
693 LCONSOLE_WARN("%s: temporarily refusing client connection "
694 "from %s\n", target->obd_name,
695 libcfs_nid2str(req->rq_peer.nid));
696 GOTO(out, rc = -EAGAIN);
699 /* Make sure the target isn't cleaned up while we're here. Yes,
700 there's still a race between the above check and our incref here.
701 Really, class_uuid2obd should take the ref. */
702 targref = class_incref(target, __FUNCTION__, cfs_current());
705 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
707 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
708 GOTO(out, rc = -EINVAL);
711 obd_str2uuid(&cluuid, str);
713 /* XXX extract a nettype and format accordingly */
714 switch (sizeof(lnet_nid_t)) {
715 /* NB the casts only avoid compiler warnings */
717 snprintf(remote_uuid.uuid, sizeof remote_uuid,
718 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
721 snprintf(remote_uuid.uuid, sizeof remote_uuid,
722 "NET_%x_UUID", (__u32)req->rq_peer.nid);
728 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
730 GOTO(out, rc = -EPROTO);
734 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
736 GOTO(out, rc = -EPROTO);
738 rc = req_capsule_server_pack(&req->rq_pill);
742 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
744 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
745 "libclient connection attempt");
746 GOTO(out, rc = -EPROTO);
747 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
748 LUSTRE_VERSION_ALLOWED_OFFSET ||
749 data->ocd_version > LUSTRE_VERSION_CODE +
750 LUSTRE_VERSION_ALLOWED_OFFSET) {
751 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
752 "libclient connection attempt",
753 data->ocd_version < LUSTRE_VERSION_CODE ?
755 OBD_OCD_VERSION_MAJOR(data->ocd_version),
756 OBD_OCD_VERSION_MINOR(data->ocd_version),
757 OBD_OCD_VERSION_PATCH(data->ocd_version),
758 OBD_OCD_VERSION_FIX(data->ocd_version));
759 data = req_capsule_server_sized_get(&req->rq_pill,
761 offsetof(typeof(*data), ocd_version) +
762 sizeof(data->ocd_version));
764 data->ocd_connect_flags = OBD_CONNECT_VERSION;
765 data->ocd_version = LUSTRE_VERSION_CODE;
767 GOTO(out, rc = -EPROTO);
771 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
772 (data->ocd_connect_flags & OBD_CONNECT_MDS))
775 /* lctl gets a backstage, all-access pass. */
776 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
777 goto dont_check_exports;
779 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
783 /* we've found an export in the hash */
784 if (export->exp_connecting) { /* bug 9635, et. al. */
785 CWARN("%s: exp %p already connecting\n",
786 export->exp_obd->obd_name, export);
787 class_export_put(export);
790 } else if (mds_conn && export->exp_connection) {
791 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
792 /* mds reconnected after failover */
793 CWARN("%s: received MDS connection from NID %s,"
794 " removing former export from NID %s\n",
795 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
796 libcfs_nid2str(export->exp_connection->c_peer.nid));
798 /* new mds connection from the same nid */
799 CWARN("%s: received new MDS connection from NID %s,"
800 " removing former export from same NID\n",
801 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
802 class_fail_export(export);
803 class_export_put(export);
806 } else if (export->exp_connection != NULL &&
807 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
808 (lustre_msg_get_op_flags(req->rq_reqmsg) &
809 MSG_CONNECT_INITIAL)) {
810 /* in mds failover we have static uuid but nid can be
812 CWARN("%s: cookie %s seen on new NID %s when "
813 "existing NID %s is already connected\n",
814 target->obd_name, cluuid.uuid,
815 libcfs_nid2str(req->rq_peer.nid),
816 libcfs_nid2str(export->exp_connection->c_peer.nid));
818 class_export_put(export);
821 cfs_spin_lock(&export->exp_lock);
822 export->exp_connecting = 1;
823 cfs_spin_unlock(&export->exp_lock);
824 class_export_put(export);
825 LASSERT(export->exp_obd == target);
827 rc = target_handle_reconnect(&conn, export, &cluuid);
830 /* If we found an export, we already unlocked. */
833 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
834 } else if (req->rq_export == NULL &&
835 cfs_atomic_read(&export->exp_rpc_count) > 0) {
836 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
837 target->obd_name, cluuid.uuid,
838 libcfs_nid2str(req->rq_peer.nid),
839 export, cfs_atomic_read(&export->exp_refcount));
840 GOTO(out, rc = -EBUSY);
841 } else if (req->rq_export != NULL &&
842 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
843 /* the current connect rpc has increased exp_rpc_count */
844 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
845 target->obd_name, cluuid.uuid,
846 libcfs_nid2str(req->rq_peer.nid),
847 export, cfs_atomic_read(&export->exp_rpc_count) - 1);
848 cfs_spin_lock(&export->exp_lock);
849 if (req->rq_export->exp_conn_cnt <
850 lustre_msg_get_conn_cnt(req->rq_reqmsg))
851 /* try to abort active requests */
852 req->rq_export->exp_abort_active_req = 1;
853 cfs_spin_unlock(&export->exp_lock);
854 GOTO(out, rc = -EBUSY);
855 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
856 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
857 "cookies not random?\n", target->obd_name,
858 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
859 GOTO(out, rc = -EALREADY);
861 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
868 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
869 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
870 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
871 export, (long)cfs_time_current_sec(),
872 export ? (long)export->exp_last_request_time : 0);
874 /* Tell the client if we're in recovery. */
875 if (target->obd_recovering) {
876 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
877 /* If this is the first time a client connects,
878 reset the recovery timer */
880 target_start_and_reset_recovery_timer(target, req,
884 /* We want to handle EALREADY but *not* -EALREADY from
885 * target_handle_reconnect(), return reconnection state in a flag */
886 if (rc == EALREADY) {
887 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
893 /* Tell the client if we support replayable requests */
894 if (target->obd_replayable)
895 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
896 client_nid = &req->rq_peer.nid;
898 if (export == NULL) {
899 if (target->obd_recovering) {
902 t = cfs_timer_deadline(&target->obd_recovery_timer);
903 t = cfs_time_sub(t, cfs_time_current());
904 CERROR("%s: denying connection for new client %s (%s): "
905 "%d clients in recovery for "CFS_TIME_T"s\n",
907 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
908 cfs_atomic_read(&target-> \
909 obd_lock_replay_clients),
910 cfs_duration_sec(t));
914 rc = obd_connect(req->rq_svc_thread->t_env,
915 &export, target, &cluuid, data,
918 conn.cookie = export->exp_handle.h_cookie;
921 rc = obd_reconnect(req->rq_svc_thread->t_env,
922 export, target, &cluuid, data, client_nid);
924 /* prevous done via class_conn2export */
925 class_export_get(export);
929 /* Return only the parts of obd_connect_data that we understand, so the
930 * client knows that we don't understand the rest. */
932 tmpdata = req_capsule_server_get(&req->rq_pill,
934 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
938 /* If all else goes well, this is our RPC return code. */
941 lustre_msg_set_handle(req->rq_repmsg, &conn);
943 /* If the client and the server are the same node, we will already
944 * have an export that really points to the client's DLM export,
945 * because we have a shared handles table.
947 * XXX this will go away when shaver stops sending the "connect" handle
948 * in the real "remote handle" field of the request --phik 24 Apr 2003
950 if (req->rq_export != NULL)
951 class_export_put(req->rq_export);
953 req->rq_export = export;
955 cfs_spin_lock(&export->exp_lock);
956 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
957 cfs_spin_unlock(&export->exp_lock);
958 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
959 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
960 export->exp_conn_cnt,
961 lustre_msg_get_conn_cnt(req->rq_reqmsg));
963 GOTO(out, rc = -EALREADY);
965 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
966 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
967 export->exp_abort_active_req = 0;
969 /* request from liblustre? Don't evict it for not pinging. */
970 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
971 export->exp_libclient = 1;
972 cfs_spin_unlock(&export->exp_lock);
974 cfs_spin_lock(&target->obd_dev_lock);
975 cfs_list_del_init(&export->exp_obd_chain_timed);
976 cfs_spin_unlock(&target->obd_dev_lock);
978 cfs_spin_unlock(&export->exp_lock);
981 if (export->exp_connection != NULL) {
982 /* Check to see if connection came from another NID */
983 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
984 !cfs_hlist_unhashed(&export->exp_nid_hash))
985 cfs_hash_del(export->exp_obd->obd_nid_hash,
986 &export->exp_connection->c_peer.nid,
987 &export->exp_nid_hash);
989 ptlrpc_connection_put(export->exp_connection);
992 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
995 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
996 cfs_hash_add(export->exp_obd->obd_nid_hash,
997 &export->exp_connection->c_peer.nid,
998 &export->exp_nid_hash);
1001 cfs_spin_lock(&target->obd_recovery_task_lock);
1002 if (target->obd_recovering && !export->exp_in_recovery) {
1003 cfs_spin_lock(&export->exp_lock);
1004 export->exp_in_recovery = 1;
1005 export->exp_req_replay_needed = 1;
1006 export->exp_lock_replay_needed = 1;
1007 cfs_spin_unlock(&export->exp_lock);
1008 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1009 && (data->ocd_transno == 0))
1010 CWARN("Connect with zero transno!\n");
1012 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1013 && data->ocd_transno < target->obd_next_recovery_transno)
1014 target->obd_next_recovery_transno = data->ocd_transno;
1015 target->obd_connected_clients++;
1016 cfs_atomic_inc(&target->obd_req_replay_clients);
1017 cfs_atomic_inc(&target->obd_lock_replay_clients);
1018 if (target->obd_connected_clients ==
1019 target->obd_max_recoverable_clients)
1020 cfs_waitq_signal(&target->obd_next_transno_waitq);
1022 cfs_spin_unlock(&target->obd_recovery_task_lock);
1023 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1026 if (export->exp_imp_reverse != NULL) {
1027 /* destroyed import can be still referenced in ctxt */
1028 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1029 KEY_REVIMP_UPD, 0, NULL, NULL);
1031 client_destroy_import(export->exp_imp_reverse);
1034 /* for the rest part, we return -ENOTCONN in case of errors
1035 * in order to let client initialize connection again.
1037 revimp = export->exp_imp_reverse = class_new_import(target);
1039 CERROR("fail to alloc new reverse import.\n");
1040 GOTO(out, rc = -ENOTCONN);
1043 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1044 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1045 revimp->imp_remote_handle = conn;
1046 revimp->imp_dlm_fake = 1;
1047 revimp->imp_state = LUSTRE_IMP_FULL;
1049 /* unknown versions will be caught in
1050 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1051 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1053 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1054 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1055 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1057 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1059 if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1060 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1061 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1063 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1065 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1067 CERROR("Failed to get sec for reverse import: %d\n", rc);
1068 export->exp_imp_reverse = NULL;
1069 class_destroy_import(revimp);
1072 class_import_put(revimp);
1075 cfs_spin_lock(&export->exp_lock);
1076 export->exp_connecting = 0;
1077 cfs_spin_unlock(&export->exp_lock);
1080 class_decref(targref, __FUNCTION__, cfs_current());
1082 req->rq_status = rc;
1086 int target_handle_disconnect(struct ptlrpc_request *req)
1091 rc = req_capsule_server_pack(&req->rq_pill);
1095 /* keep the rq_export around so we can send the reply */
1096 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1101 void target_destroy_export(struct obd_export *exp)
1103 /* exports created from last_rcvd data, and "fake"
1104 exports created by lctl don't have an import */
1105 if (exp->exp_imp_reverse != NULL)
1106 client_destroy_import(exp->exp_imp_reverse);
1108 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1109 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1110 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1111 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1115 * Recovery functions
1117 static void target_request_copy_get(struct ptlrpc_request *req)
1119 class_export_rpc_get(req->rq_export);
1120 LASSERT(cfs_list_empty(&req->rq_list));
1121 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1123 /* increase refcount to keep request in queue */
1124 cfs_atomic_inc(&req->rq_refcount);
1125 /** let export know it has replays to be handled */
1126 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1129 static void target_request_copy_put(struct ptlrpc_request *req)
1131 LASSERT(cfs_list_empty(&req->rq_replay_list));
1132 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1134 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1135 class_export_rpc_put(req->rq_export);
1136 ptlrpc_server_drop_request(req);
1139 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1141 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1142 struct obd_export *exp = req->rq_export;
1143 struct ptlrpc_request *reqiter;
1148 cfs_spin_lock(&exp->exp_lock);
1149 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1151 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1158 /* we expect it with RESENT and REPLAY flags */
1159 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1160 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1161 CERROR("invalid flags %x of resent replay\n",
1162 lustre_msg_get_flags(req->rq_reqmsg));
1164 cfs_list_add_tail(&req->rq_replay_list,
1165 &exp->exp_req_replay_queue);
1168 cfs_spin_unlock(&exp->exp_lock);
1172 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1174 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1175 LASSERT(req->rq_export);
1177 cfs_spin_lock(&req->rq_export->exp_lock);
1178 cfs_list_del_init(&req->rq_replay_list);
1179 cfs_spin_unlock(&req->rq_export->exp_lock);
1183 static void target_finish_recovery(struct obd_device *obd)
1186 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1189 ldlm_reprocess_all_ns(obd->obd_namespace);
1190 cfs_spin_lock(&obd->obd_recovery_task_lock);
1191 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1192 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1193 !cfs_list_empty(&obd->obd_final_req_queue)) {
1194 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1196 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1197 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1199 cfs_list_empty(&obd->obd_final_req_queue) ? \
1201 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1204 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1206 obd->obd_recovery_end = cfs_time_current_sec();
1208 /* when recovery finished, cleanup orphans on mds and ost */
1209 if (OBT(obd) && OBP(obd, postrecov)) {
1210 int rc = OBP(obd, postrecov)(obd);
1212 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1218 static void abort_req_replay_queue(struct obd_device *obd)
1220 struct ptlrpc_request *req, *n;
1221 cfs_list_t abort_list;
1223 CFS_INIT_LIST_HEAD(&abort_list);
1224 cfs_spin_lock(&obd->obd_recovery_task_lock);
1225 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1226 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1227 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1228 DEBUG_REQ(D_WARNING, req, "aborted:");
1229 req->rq_status = -ENOTCONN;
1230 if (ptlrpc_error(req)) {
1231 DEBUG_REQ(D_ERROR, req,
1232 "failed abort_req_reply; skipping");
1234 target_exp_dequeue_req_replay(req);
1235 target_request_copy_put(req);
1239 static void abort_lock_replay_queue(struct obd_device *obd)
1241 struct ptlrpc_request *req, *n;
1242 cfs_list_t abort_list;
1244 CFS_INIT_LIST_HEAD(&abort_list);
1245 cfs_spin_lock(&obd->obd_recovery_task_lock);
1246 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1247 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1248 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1249 DEBUG_REQ(D_ERROR, req, "aborted:");
1250 req->rq_status = -ENOTCONN;
1251 if (ptlrpc_error(req)) {
1252 DEBUG_REQ(D_ERROR, req,
1253 "failed abort_lock_reply; skipping");
1255 target_request_copy_put(req);
1260 /* Called from a cleanup function if the device is being cleaned up
1261 forcefully. The exports should all have been disconnected already,
1262 the only thing left to do is
1263 - clear the recovery flags
1265 - free queued requests and replies, but don't send replies
1266 Because the obd_stopping flag is set, no new requests should be received.
1269 void target_cleanup_recovery(struct obd_device *obd)
1271 struct ptlrpc_request *req, *n;
1272 cfs_list_t clean_list;
1275 CFS_INIT_LIST_HEAD(&clean_list);
1276 cfs_spin_lock(&obd->obd_dev_lock);
1277 if (!obd->obd_recovering) {
1278 cfs_spin_unlock(&obd->obd_dev_lock);
1282 obd->obd_recovering = obd->obd_abort_recovery = 0;
1283 cfs_spin_unlock(&obd->obd_dev_lock);
1285 cfs_spin_lock(&obd->obd_recovery_task_lock);
1286 target_cancel_recovery_timer(obd);
1287 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1288 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1290 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1291 LASSERT(req->rq_reply_state == 0);
1292 target_exp_dequeue_req_replay(req);
1293 target_request_copy_put(req);
1296 cfs_spin_lock(&obd->obd_recovery_task_lock);
1297 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1298 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1299 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1301 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1302 LASSERT(req->rq_reply_state == 0);
1303 target_request_copy_put(req);
1309 /* obd_recovery_task_lock should be held */
1310 void target_cancel_recovery_timer(struct obd_device *obd)
1312 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1313 cfs_timer_disarm(&obd->obd_recovery_timer);
1316 /* extend = 1 means require at least "duration" seconds left in the timer,
1317 extend = 0 means set the total duration (start_recovery_timer) */
1318 static void reset_recovery_timer(struct obd_device *obd, int duration,
1321 cfs_time_t now = cfs_time_current_sec();
1322 cfs_duration_t left;
1324 cfs_spin_lock(&obd->obd_recovery_task_lock);
1325 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1326 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1330 left = cfs_time_sub(obd->obd_recovery_end, now);
1332 if (extend && (duration > left))
1333 obd->obd_recovery_timeout += duration - left;
1334 else if (!extend && (duration > obd->obd_recovery_timeout))
1335 /* Track the client's largest expected replay time */
1336 obd->obd_recovery_timeout = duration;
1338 /* Hard limit of obd_recovery_time_hard which should not happen */
1339 if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1340 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1342 obd->obd_recovery_end = obd->obd_recovery_start +
1343 obd->obd_recovery_timeout;
1344 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1345 cfs_time_before(now, obd->obd_recovery_end)) {
1346 left = cfs_time_sub(obd->obd_recovery_end, now);
1347 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1349 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1350 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1351 obd->obd_name, (unsigned)left);
1354 static void check_and_start_recovery_timer(struct obd_device *obd)
1356 cfs_spin_lock(&obd->obd_recovery_task_lock);
1357 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1358 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1361 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1362 obd->obd_recovery_start = cfs_time_current_sec();
1363 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1365 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1368 /* Reset the timer with each new client connection */
1370 * This timer is actually reconnect_timer, which is for making sure
1371 * the total recovery window is at least as big as my reconnect
1372 * attempt timing. So the initial recovery time_out will be set to
1373 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1374 * from client is bigger than this, then the recovery time_out will
1375 * be extended to make sure the client could be reconnected, in the
1376 * process, the timeout from the new client should be ignored.
1380 target_start_and_reset_recovery_timer(struct obd_device *obd,
1381 struct ptlrpc_request *req,
1384 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1386 if (!new_client && service_time)
1387 /* Teach server about old server's estimates, as first guess
1388 * at how long new requests will take. */
1389 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1392 check_and_start_recovery_timer(obd);
1394 /* convert the service time to rpc timeout,
1395 * reuse service_time to limit stack usage */
1396 service_time = at_est2timeout(service_time);
1398 /* We expect other clients to timeout within service_time, then try
1399 * to reconnect, then try the failover server. The max delay between
1400 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1401 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1402 INITIAL_CONNECT_TIMEOUT);
1403 if (service_time > obd->obd_recovery_timeout && !new_client)
1404 reset_recovery_timer(obd, service_time, 0);
1409 /** Health checking routines */
1410 static inline int exp_connect_healthy(struct obd_export *exp)
1412 return (exp->exp_in_recovery);
1415 /** if export done req_replay or has replay in queue */
1416 static inline int exp_req_replay_healthy(struct obd_export *exp)
1418 return (!exp->exp_req_replay_needed ||
1419 cfs_atomic_read(&exp->exp_replay_count) > 0);
1421 /** if export done lock_replay or has replay in queue */
1422 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1424 return (!exp->exp_lock_replay_needed ||
1425 cfs_atomic_read(&exp->exp_replay_count) > 0);
1428 static inline int exp_vbr_healthy(struct obd_export *exp)
1430 return (!exp->exp_vbr_failed);
1433 static inline int exp_finished(struct obd_export *exp)
1435 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1438 /** Checking routines for recovery */
1439 static int check_for_clients(struct obd_device *obd)
1441 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1443 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1444 if (obd->obd_no_conn == 0 &&
1445 obd->obd_connected_clients + obd->obd_stale_clients ==
1446 obd->obd_max_recoverable_clients)
1451 static int check_for_next_transno(struct obd_device *obd)
1453 struct ptlrpc_request *req = NULL;
1454 int wake_up = 0, connected, completed, queue_len;
1455 __u64 next_transno, req_transno;
1458 cfs_spin_lock(&obd->obd_recovery_task_lock);
1459 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1460 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1461 struct ptlrpc_request, rq_list);
1462 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1467 connected = obd->obd_connected_clients;
1468 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1469 queue_len = obd->obd_requests_queued_for_recovery;
1470 next_transno = obd->obd_next_recovery_transno;
1472 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1473 "req_transno: "LPU64", next_transno: "LPU64"\n",
1474 obd->obd_max_recoverable_clients, connected, completed,
1475 queue_len, req_transno, next_transno);
1477 if (obd->obd_abort_recovery) {
1478 CDEBUG(D_HA, "waking for aborted recovery\n");
1480 } else if (obd->obd_recovery_expired) {
1481 CDEBUG(D_HA, "waking for expired recovery\n");
1483 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1484 CDEBUG(D_HA, "waking for completed recovery\n");
1486 } else if (req_transno == next_transno) {
1487 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1489 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1491 /** handle gaps occured due to lost reply or VBR */
1492 LASSERTF(req_transno >= next_transno,
1493 "req_transno: "LPU64", next_transno: "LPU64"\n",
1494 req_transno, next_transno);
1495 if (req_transno > obd->obd_last_committed &&
1496 !obd->obd_version_recov)
1499 "%s: waking for gap in transno, VBR is %s (skip: "
1500 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1501 ", last_committed: "LPD64")\n",
1502 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1503 next_transno, queue_len, completed, connected,
1504 req_transno, obd->obd_last_committed);
1505 obd->obd_next_recovery_transno = req_transno;
1507 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1508 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1509 " by fail_lock, waking up ("LPD64")\n", next_transno);
1510 obd->obd_next_recovery_transno = req_transno;
1513 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1517 static int check_for_next_lock(struct obd_device *obd)
1521 cfs_spin_lock(&obd->obd_recovery_task_lock);
1522 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1523 CDEBUG(D_HA, "waking for next lock\n");
1525 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1526 CDEBUG(D_HA, "waking for completed lock replay\n");
1528 } else if (obd->obd_abort_recovery) {
1529 CDEBUG(D_HA, "waking for aborted recovery\n");
1531 } else if (obd->obd_recovery_expired) {
1532 CDEBUG(D_HA, "waking for expired recovery\n");
1535 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1541 * wait for recovery events,
1542 * check its status with help of check_routine
1543 * evict dead clients via health_check
1545 static int target_recovery_overseer(struct obd_device *obd,
1546 int (*check_routine)(struct obd_device *),
1547 int (*health_check)(struct obd_export *))
1550 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1551 if (obd->obd_abort_recovery) {
1552 CWARN("recovery is aborted, evict exports in recovery\n");
1553 /** evict exports which didn't finish recovery yet */
1554 class_disconnect_stale_exports(obd, exp_finished);
1556 } else if (obd->obd_recovery_expired) {
1557 obd->obd_recovery_expired = 0;
1558 /** If some clients died being recovered, evict them */
1560 "recovery is timed out, evict stale exports\n");
1561 /** evict cexports with no replay in queue, they are stalled */
1562 class_disconnect_stale_exports(obd, health_check);
1563 /** continue with VBR */
1564 cfs_spin_lock(&obd->obd_dev_lock);
1565 obd->obd_version_recov = 1;
1566 cfs_spin_unlock(&obd->obd_dev_lock);
1568 * reset timer, recovery will proceed with versions now,
1569 * timeout is set just to handle reconnection delays
1571 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1572 /** Wait for recovery events again, after evicting bad clients */
1578 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1580 struct ptlrpc_request *req = NULL;
1583 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1584 obd->obd_next_recovery_transno);
1586 if (target_recovery_overseer(obd, check_for_next_transno,
1587 exp_req_replay_healthy)) {
1588 abort_req_replay_queue(obd);
1589 abort_lock_replay_queue(obd);
1592 cfs_spin_lock(&obd->obd_recovery_task_lock);
1593 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1594 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1595 struct ptlrpc_request, rq_list);
1596 cfs_list_del_init(&req->rq_list);
1597 obd->obd_requests_queued_for_recovery--;
1598 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1600 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1601 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1602 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1603 /** evict exports failed VBR */
1604 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1609 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1611 struct ptlrpc_request *req = NULL;
1613 CDEBUG(D_HA, "Waiting for lock\n");
1614 if (target_recovery_overseer(obd, check_for_next_lock,
1615 exp_lock_replay_healthy))
1616 abort_lock_replay_queue(obd);
1618 cfs_spin_lock(&obd->obd_recovery_task_lock);
1619 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1620 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1621 struct ptlrpc_request, rq_list);
1622 cfs_list_del_init(&req->rq_list);
1623 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1625 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1626 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1627 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1628 /** evict exports failed VBR */
1629 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1634 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1636 struct ptlrpc_request *req = NULL;
1638 cfs_spin_lock(&obd->obd_recovery_task_lock);
1639 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1640 req = cfs_list_entry(obd->obd_final_req_queue.next,
1641 struct ptlrpc_request, rq_list);
1642 cfs_list_del_init(&req->rq_list);
1643 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1644 if (req->rq_export->exp_in_recovery) {
1645 cfs_spin_lock(&req->rq_export->exp_lock);
1646 req->rq_export->exp_in_recovery = 0;
1647 cfs_spin_unlock(&req->rq_export->exp_lock);
1650 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1655 static int handle_recovery_req(struct ptlrpc_thread *thread,
1656 struct ptlrpc_request *req,
1657 svc_handler_t handler)
1663 * export can be evicted during recovery, no need to handle replays for
1664 * it after that, discard such request silently
1666 if (req->rq_export->exp_disconnected)
1667 GOTO(reqcopy_put, rc = 0);
1669 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1671 CERROR("Failure to initialize session: %d\n", rc);
1672 GOTO(reqcopy_put, rc);
1675 req->rq_recov_session.lc_thread = thread;
1676 lu_context_enter(&req->rq_recov_session);
1677 req->rq_svc_thread = thread;
1678 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1680 /* thread context */
1681 lu_context_enter(&thread->t_env->le_ctx);
1683 lu_context_exit(&thread->t_env->le_ctx);
1685 lu_context_exit(&req->rq_recov_session);
1686 lu_context_fini(&req->rq_recov_session);
1687 /* don't reset timer for final stage */
1688 if (!exp_finished(req->rq_export)) {
1690 * Add request timeout to the recovery time so next request from
1691 * this client may come in recovery time
1693 reset_recovery_timer(class_exp2obd(req->rq_export),
1694 AT_OFF ? obd_timeout :
1695 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1701 static int target_recovery_thread(void *arg)
1703 struct lu_target *lut = arg;
1704 struct obd_device *obd = lut->lut_obd;
1705 struct ptlrpc_request *req;
1706 struct target_recovery_data *trd = &obd->obd_recovery_data;
1707 unsigned long delta;
1708 unsigned long flags;
1710 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1714 cfs_daemonize_ctxt("tgt_recov");
1716 SIGNAL_MASK_LOCK(current, flags);
1717 sigfillset(¤t->blocked);
1719 SIGNAL_MASK_UNLOCK(current, flags);
1721 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1725 thread->t_env = &env;
1726 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1727 env.le_ctx.lc_thread = thread;
1728 thread->t_data = NULL;
1730 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1732 trd->trd_processing_task = cfs_curproc_pid();
1734 cfs_spin_lock(&obd->obd_dev_lock);
1735 obd->obd_recovering = 1;
1736 cfs_spin_unlock(&obd->obd_dev_lock);
1737 cfs_complete(&trd->trd_starting);
1739 /* first of all, we have to know the first transno to replay */
1740 if (target_recovery_overseer(obd, check_for_clients,
1741 exp_connect_healthy)) {
1742 abort_req_replay_queue(obd);
1743 abort_lock_replay_queue(obd);
1746 /* next stage: replay requests */
1748 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1749 cfs_atomic_read(&obd->obd_req_replay_clients),
1750 obd->obd_next_recovery_transno);
1751 while ((req = target_next_replay_req(obd))) {
1752 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1753 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1754 lustre_msg_get_transno(req->rq_reqmsg),
1755 libcfs_nid2str(req->rq_peer.nid));
1756 handle_recovery_req(thread, req,
1757 trd->trd_recovery_handler);
1759 * bz18031: increase next_recovery_transno before
1760 * target_request_copy_put() will drop exp_rpc reference
1762 cfs_spin_lock(&obd->obd_recovery_task_lock);
1763 obd->obd_next_recovery_transno++;
1764 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1765 target_exp_dequeue_req_replay(req);
1766 target_request_copy_put(req);
1767 obd->obd_replayed_requests++;
1771 * The second stage: replay locks
1773 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1774 cfs_atomic_read(&obd->obd_lock_replay_clients));
1775 while ((req = target_next_replay_lock(obd))) {
1776 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1777 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1778 libcfs_nid2str(req->rq_peer.nid));
1779 handle_recovery_req(thread, req,
1780 trd->trd_recovery_handler);
1781 target_request_copy_put(req);
1782 obd->obd_replayed_locks++;
1786 * The third stage: reply on final pings, at this moment all clients
1787 * must have request in final queue
1789 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1790 /** Update server last boot epoch */
1791 lut_boot_epoch_update(lut);
1792 /* We drop recoverying flag to forward all new requests
1793 * to regular mds_handle() since now */
1794 cfs_spin_lock(&obd->obd_dev_lock);
1795 obd->obd_recovering = obd->obd_abort_recovery = 0;
1796 cfs_spin_unlock(&obd->obd_dev_lock);
1797 cfs_spin_lock(&obd->obd_recovery_task_lock);
1798 target_cancel_recovery_timer(obd);
1799 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1800 while ((req = target_next_final_ping(obd))) {
1801 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1802 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1803 libcfs_nid2str(req->rq_peer.nid));
1804 handle_recovery_req(thread, req,
1805 trd->trd_recovery_handler);
1806 target_request_copy_put(req);
1809 delta = (jiffies - delta) / CFS_HZ;
1810 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1811 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1812 if (delta > OBD_RECOVERY_TIME_SOFT) {
1813 CWARN("too long recovery - read logs\n");
1814 libcfs_debug_dumplog();
1817 target_finish_recovery(obd);
1819 lu_context_fini(&env.le_ctx);
1820 trd->trd_processing_task = 0;
1821 cfs_complete(&trd->trd_finishing);
1825 static int target_start_recovery_thread(struct lu_target *lut,
1826 svc_handler_t handler)
1828 struct obd_device *obd = lut->lut_obd;
1830 struct target_recovery_data *trd = &obd->obd_recovery_data;
1832 memset(trd, 0, sizeof(*trd));
1833 cfs_init_completion(&trd->trd_starting);
1834 cfs_init_completion(&trd->trd_finishing);
1835 trd->trd_recovery_handler = handler;
1837 if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
1838 cfs_wait_for_completion(&trd->trd_starting);
1839 LASSERT(obd->obd_recovering != 0);
1846 void target_stop_recovery_thread(struct obd_device *obd)
1848 if (obd->obd_recovery_data.trd_processing_task > 0) {
1849 struct target_recovery_data *trd = &obd->obd_recovery_data;
1850 /** recovery can be done but postrecovery is not yet */
1851 cfs_spin_lock(&obd->obd_dev_lock);
1852 if (obd->obd_recovering) {
1853 CERROR("%s: Aborting recovery\n", obd->obd_name);
1854 obd->obd_abort_recovery = 1;
1855 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1857 cfs_spin_unlock(&obd->obd_dev_lock);
1858 cfs_wait_for_completion(&trd->trd_finishing);
1862 void target_recovery_fini(struct obd_device *obd)
1864 class_disconnect_exports(obd);
1865 target_stop_recovery_thread(obd);
1866 target_cleanup_recovery(obd);
1868 EXPORT_SYMBOL(target_recovery_fini);
1870 static void target_recovery_expired(unsigned long castmeharder)
1872 struct obd_device *obd = (struct obd_device *)castmeharder;
1873 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1874 " after %lds (%d clients connected)\n",
1875 obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1876 cfs_time_current_sec()- obd->obd_recovery_start,
1877 obd->obd_connected_clients);
1879 obd->obd_recovery_expired = 1;
1880 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1883 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1885 struct obd_device *obd = lut->lut_obd;
1886 if (obd->obd_max_recoverable_clients == 0) {
1887 /** Update server last boot epoch */
1888 lut_boot_epoch_update(lut);
1892 CWARN("RECOVERY: service %s, %d recoverable clients, "
1893 "last_transno "LPU64"\n", obd->obd_name,
1894 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1895 LASSERT(obd->obd_stopping == 0);
1896 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1897 obd->obd_recovery_start = 0;
1898 obd->obd_recovery_end = 0;
1900 /* both values can be get from mount data already */
1901 if (obd->obd_recovery_timeout == 0)
1902 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
1903 if (obd->obd_recovery_time_hard == 0)
1904 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
1905 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1906 target_start_recovery_thread(lut, handler);
1908 EXPORT_SYMBOL(target_recovery_init);
1912 static int target_process_req_flags(struct obd_device *obd,
1913 struct ptlrpc_request *req)
1915 struct obd_export *exp = req->rq_export;
1916 LASSERT(exp != NULL);
1917 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1918 /* client declares he's ready to replay locks */
1919 cfs_spin_lock(&exp->exp_lock);
1920 if (exp->exp_req_replay_needed) {
1921 exp->exp_req_replay_needed = 0;
1922 cfs_spin_unlock(&exp->exp_lock);
1924 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
1925 cfs_atomic_dec(&obd->obd_req_replay_clients);
1927 cfs_spin_unlock(&exp->exp_lock);
1930 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1931 /* client declares he's ready to complete recovery
1932 * so, we put the request on th final queue */
1933 cfs_spin_lock(&exp->exp_lock);
1934 if (exp->exp_lock_replay_needed) {
1935 exp->exp_lock_replay_needed = 0;
1936 cfs_spin_unlock(&exp->exp_lock);
1938 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
1939 cfs_atomic_dec(&obd->obd_lock_replay_clients);
1941 cfs_spin_unlock(&exp->exp_lock);
1947 int target_queue_recovery_request(struct ptlrpc_request *req,
1948 struct obd_device *obd)
1952 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1955 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1956 /* Processing the queue right now, don't re-add. */
1960 target_process_req_flags(obd, req);
1962 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1963 /* client declares he's ready to complete recovery
1964 * so, we put the request on th final queue */
1965 target_request_copy_get(req);
1966 DEBUG_REQ(D_HA, req, "queue final req");
1967 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1968 cfs_spin_lock(&obd->obd_recovery_task_lock);
1969 if (obd->obd_recovering) {
1970 cfs_list_add_tail(&req->rq_list,
1971 &obd->obd_final_req_queue);
1973 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1974 target_request_copy_put(req);
1975 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1977 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1980 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1981 /* client declares he's ready to replay locks */
1982 target_request_copy_get(req);
1983 DEBUG_REQ(D_HA, req, "queue lock replay req");
1984 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1985 cfs_spin_lock(&obd->obd_recovery_task_lock);
1986 LASSERT(obd->obd_recovering);
1987 /* usually due to recovery abort */
1988 if (!req->rq_export->exp_in_recovery) {
1989 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1990 target_request_copy_put(req);
1993 LASSERT(req->rq_export->exp_lock_replay_needed);
1994 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1995 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1999 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2000 * (i.e. buflens etc are in my own byte order), but type-dependent
2001 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2004 CFS_INIT_LIST_HEAD(&req->rq_list);
2005 DEBUG_REQ(D_HA, req, "not queueing");
2009 /* If we're processing the queue, we want don't want to queue this
2012 * Also, if this request has a transno less than the one we're waiting
2013 * for, we should process it now. It could (and currently always will)
2014 * be an open request for a descriptor that was opened some time ago.
2016 * Also, a resent, replayed request that has already been
2017 * handled will pass through here and be processed immediately.
2019 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2020 obd->obd_next_recovery_transno, transno);
2021 cfs_spin_lock(&obd->obd_recovery_task_lock);
2022 if (transno < obd->obd_next_recovery_transno) {
2023 /* Processing the queue right now, don't re-add. */
2024 LASSERT(cfs_list_empty(&req->rq_list));
2025 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2028 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2030 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2033 target_request_copy_get(req);
2034 if (!req->rq_export->exp_in_recovery) {
2035 target_request_copy_put(req);
2038 LASSERT(req->rq_export->exp_req_replay_needed);
2040 if (target_exp_enqueue_req_replay(req)) {
2041 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2042 target_request_copy_put(req);
2047 cfs_spin_lock(&obd->obd_recovery_task_lock);
2048 LASSERT(obd->obd_recovering);
2049 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2050 struct ptlrpc_request *reqiter =
2051 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2053 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2054 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2059 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2061 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2062 "has been claimed by another client");
2063 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2064 target_exp_dequeue_req_replay(req);
2065 target_request_copy_put(req);
2071 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2073 obd->obd_requests_queued_for_recovery++;
2074 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2075 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2080 * Packs current SLV and Limit into \a req.
2082 int target_pack_pool_reply(struct ptlrpc_request *req)
2084 struct obd_device *obd;
2088 * Check that we still have all structures alive as this may
2089 * be some late rpc in shutdown time.
2091 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2092 !exp_connect_lru_resize(req->rq_export))) {
2093 lustre_msg_set_slv(req->rq_repmsg, 0);
2094 lustre_msg_set_limit(req->rq_repmsg, 0);
2099 * OBD is alive here as export is alive, which we checked above.
2101 obd = req->rq_export->exp_obd;
2103 cfs_read_lock(&obd->obd_pool_lock);
2104 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2105 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2106 cfs_read_unlock(&obd->obd_pool_lock);
2111 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2113 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2114 DEBUG_REQ(D_ERROR, req, "dropping reply");
2119 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2120 req->rq_status = rc;
2121 return (ptlrpc_send_error(req, 1));
2123 DEBUG_REQ(D_NET, req, "sending reply");
2126 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2129 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2132 struct ptlrpc_reply_state *rs;
2133 struct obd_device *obd;
2134 struct obd_export *exp;
2135 struct ptlrpc_service *svc;
2138 if (req->rq_no_reply) {
2143 svc = req->rq_rqbd->rqbd_service;
2144 rs = req->rq_reply_state;
2145 if (rs == NULL || !rs->rs_difficult) {
2147 target_send_reply_msg (req, rc, fail_id);
2152 /* must be an export if locks saved */
2153 LASSERT (req->rq_export != NULL);
2154 /* req/reply consistent */
2155 LASSERT (rs->rs_service == svc);
2158 LASSERT (!rs->rs_scheduled);
2159 LASSERT (!rs->rs_scheduled_ever);
2160 LASSERT (!rs->rs_handled);
2161 LASSERT (!rs->rs_on_net);
2162 LASSERT (rs->rs_export == NULL);
2163 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2164 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2166 exp = class_export_get (req->rq_export);
2169 /* disable reply scheduling while I'm setting up */
2170 rs->rs_scheduled = 1;
2172 rs->rs_xid = req->rq_xid;
2173 rs->rs_transno = req->rq_transno;
2174 rs->rs_export = exp;
2175 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2177 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2178 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2179 rs->rs_transno, exp->exp_last_committed);
2180 if (rs->rs_transno > exp->exp_last_committed) {
2181 /* not committed already */
2182 cfs_list_add_tail(&rs->rs_obd_list,
2183 &exp->exp_uncommitted_replies);
2185 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2187 cfs_spin_lock(&exp->exp_lock);
2188 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2189 cfs_spin_unlock(&exp->exp_lock);
2191 netrc = target_send_reply_msg (req, rc, fail_id);
2193 cfs_spin_lock(&svc->srv_rs_lock);
2195 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2198 /* error sending: reply is off the net. Also we need +1
2199 * reply ref until ptlrpc_handle_rs() is done
2200 * with the reply state (if the send was successful, there
2201 * would have been +1 ref for the net, which
2202 * reply_out_callback leaves alone) */
2204 ptlrpc_rs_addref(rs);
2207 cfs_spin_lock(&rs->rs_lock);
2208 if (rs->rs_transno <= exp->exp_last_committed ||
2209 (!rs->rs_on_net && !rs->rs_no_ack) ||
2210 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2211 cfs_list_empty(&rs->rs_obd_list)) {
2212 CDEBUG(D_HA, "Schedule reply immediately\n");
2213 ptlrpc_dispatch_difficult_reply(rs);
2215 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2216 rs->rs_scheduled = 0; /* allow notifier to schedule */
2218 cfs_spin_unlock(&rs->rs_lock);
2219 cfs_spin_unlock(&svc->srv_rs_lock);
2223 int target_handle_ping(struct ptlrpc_request *req)
2225 obd_ping(req->rq_export);
2226 return req_capsule_server_pack(&req->rq_pill);
2229 void target_committed_to_req(struct ptlrpc_request *req)
2231 struct obd_export *exp = req->rq_export;
2233 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2234 lustre_msg_set_last_committed(req->rq_repmsg,
2235 exp->exp_last_committed);
2237 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2238 "%d)", exp->exp_obd->obd_no_transno,
2239 req->rq_repmsg == NULL);
2241 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2242 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2244 EXPORT_SYMBOL(target_committed_to_req);
2246 int target_handle_qc_callback(struct ptlrpc_request *req)
2248 struct obd_quotactl *oqctl;
2249 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2251 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2252 if (oqctl == NULL) {
2253 CERROR("Can't unpack obd_quotactl\n");
2257 cli->cl_qchk_stat = oqctl->qc_stat;
2262 #ifdef HAVE_QUOTA_SUPPORT
2263 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2266 struct obd_device *obd = req->rq_export->exp_obd;
2267 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2268 struct obd_device_target *obt;
2269 struct lustre_quota_ctxt *qctxt;
2270 struct qunit_data *qdata = NULL;
2274 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2277 rc = req_capsule_server_pack(&req->rq_pill);
2279 CERROR("packing reply failed!: rc = %d\n", rc);
2283 LASSERT(req->rq_export);
2285 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2286 if (IS_ERR(qdata)) {
2287 rc = PTR_ERR(qdata);
2288 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2289 req->rq_status = rc;
2293 /* we use the observer */
2294 if (obd_pin_observer(obd, &lov_obd) ||
2295 obd_pin_observer(lov_obd, &master_obd)) {
2296 CERROR("Can't find the observer, it is recovering\n");
2297 req->rq_status = -EAGAIN;
2301 obt = &master_obd->u.obt;
2302 qctxt = &obt->obt_qctxt;
2304 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2305 /* quota_type has not been processed yet, return EAGAIN
2306 * until we know whether or not quotas are supposed to
2308 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2310 req->rq_status = -EAGAIN;
2314 cfs_down_read(&obt->obt_rwsem);
2315 if (qctxt->lqc_lqs_hash == NULL) {
2316 cfs_up_read(&obt->obt_rwsem);
2317 /* quota_type has not been processed yet, return EAGAIN
2318 * until we know whether or not quotas are supposed to
2320 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2322 req->rq_status = -EAGAIN;
2326 LASSERT(qctxt->lqc_handler);
2327 rc = qctxt->lqc_handler(master_obd, qdata,
2328 lustre_msg_get_opc(req->rq_reqmsg));
2329 cfs_up_read(&obt->obt_rwsem);
2330 if (rc && rc != -EDQUOT)
2331 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2332 "dqacq/dqrel failed! (rc:%d)\n", rc);
2333 req->rq_status = rc;
2335 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2337 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2341 /* Block the quota req. b=14840 */
2342 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2347 obd_unpin_observer(lov_obd);
2349 obd_unpin_observer(obd);
2351 rc = ptlrpc_reply(req);
2355 #endif /* !__KERNEL__ */
2357 #endif /* HAVE_QUOTA_SUPPORT */
2359 ldlm_mode_t lck_compat_array[] = {
2360 [LCK_EX] LCK_COMPAT_EX,
2361 [LCK_PW] LCK_COMPAT_PW,
2362 [LCK_PR] LCK_COMPAT_PR,
2363 [LCK_CW] LCK_COMPAT_CW,
2364 [LCK_CR] LCK_COMPAT_CR,
2365 [LCK_NL] LCK_COMPAT_NL,
2366 [LCK_GROUP] LCK_COMPAT_GROUP,
2367 [LCK_COS] LCK_COMPAT_COS,
2371 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2372 * not escape to the user level.
2374 int ldlm_error2errno(ldlm_error_t error)
2382 case ELDLM_LOCK_CHANGED:
2385 case ELDLM_LOCK_ABORTED:
2388 case ELDLM_LOCK_REPLACED:
2391 case ELDLM_NO_LOCK_DATA:
2394 case ELDLM_NAMESPACE_EXISTS:
2397 case ELDLM_BAD_NAMESPACE:
2401 if (((int)error) < 0) /* cast to signed type */
2402 result = error; /* as ldlm_error_t can be unsigned */
2404 CERROR("Invalid DLM result code: %d\n", error);
2410 EXPORT_SYMBOL(ldlm_error2errno);
2413 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2415 ldlm_error_t ldlm_errno2error(int err_no)
2424 error = ELDLM_LOCK_CHANGED;
2427 error = ELDLM_LOCK_ABORTED;
2430 error = ELDLM_LOCK_REPLACED;
2433 error = ELDLM_NO_LOCK_DATA;
2436 error = ELDLM_NAMESPACE_EXISTS;
2439 error = ELDLM_BAD_NAMESPACE;
2446 EXPORT_SYMBOL(ldlm_errno2error);
2448 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2449 void ldlm_dump_export_locks(struct obd_export *exp)
2451 cfs_spin_lock(&exp->exp_locks_list_guard);
2452 if (!cfs_list_empty(&exp->exp_locks_list)) {
2453 struct ldlm_lock *lock;
2455 CERROR("dumping locks for export %p,"
2456 "ignore if the unmount doesn't hang\n", exp);
2457 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2458 ldlm_lock_dump(D_ERROR, lock, 0);
2460 cfs_spin_unlock(&exp->exp_locks_list_guard);