1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_mds.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: if non-zero, move the selected to the list head
55 * @create: if zero, only search in existed connections
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 spin_lock(&imp->imp_lock);
84 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 list_del(&item->oic_item);
88 list_add(&item->oic_item, &imp->imp_conn_list);
89 item->oic_last_attempt = 0;
91 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
92 imp, imp->imp_obd->obd_name, uuid->uuid,
93 (priority ? ", moved to head" : ""));
94 spin_unlock(&imp->imp_lock);
95 GOTO(out_free, rc = 0);
100 imp_conn->oic_conn = ptlrpc_conn;
101 imp_conn->oic_uuid = *uuid;
102 imp_conn->oic_last_attempt = 0;
104 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
106 list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
107 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
108 imp, imp->imp_obd->obd_name, uuid->uuid,
109 (priority ? "head" : "tail"));
111 spin_unlock(&imp->imp_lock);
112 GOTO(out_free, rc = -ENOENT);
115 spin_unlock(&imp->imp_lock);
119 OBD_FREE(imp_conn, sizeof(*imp_conn));
121 ptlrpc_connection_put(ptlrpc_conn);
125 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
127 return import_set_conn(imp, uuid, 1, 0);
130 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
133 return import_set_conn(imp, uuid, priority, 1);
136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
138 struct obd_import_conn *imp_conn;
139 struct obd_export *dlmexp;
143 spin_lock(&imp->imp_lock);
144 if (list_empty(&imp->imp_conn_list)) {
145 LASSERT(!imp->imp_connection);
149 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
150 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
152 LASSERT(imp_conn->oic_conn);
154 /* is current conn? */
155 if (imp_conn == imp->imp_conn_current) {
156 LASSERT(imp_conn->oic_conn == imp->imp_connection);
158 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
159 imp->imp_state != LUSTRE_IMP_DISCON) {
160 CERROR("can't remove current connection\n");
161 GOTO(out, rc = -EBUSY);
164 ptlrpc_connection_put(imp->imp_connection);
165 imp->imp_connection = NULL;
167 dlmexp = class_conn2export(&imp->imp_dlm_handle);
168 if (dlmexp && dlmexp->exp_connection) {
169 LASSERT(dlmexp->exp_connection ==
171 ptlrpc_connection_put(dlmexp->exp_connection);
172 dlmexp->exp_connection = NULL;
176 list_del(&imp_conn->oic_item);
177 ptlrpc_connection_put(imp_conn->oic_conn);
178 OBD_FREE(imp_conn, sizeof(*imp_conn));
179 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
180 imp, imp->imp_obd->obd_name, uuid->uuid);
185 spin_unlock(&imp->imp_lock);
187 CERROR("connection %s not found\n", uuid->uuid);
191 void client_destroy_import(struct obd_import *imp)
193 /* drop security policy instance after all rpc finished/aborted
194 * to let all busy contexts be released. */
195 class_import_get(imp);
196 class_destroy_import(imp);
197 sptlrpc_import_sec_put(imp);
198 class_import_put(imp);
201 /* configure an RPC client OBD device
206 * 3 - inactive-on-startup
208 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
210 struct client_obd *cli = &obddev->u.cli;
211 struct obd_import *imp;
212 struct obd_uuid server_uuid;
213 int rq_portal, rp_portal, connect_op;
214 char *name = obddev->obd_type->typ_name;
218 /* In a more perfect world, we would hang a ptlrpc_client off of
219 * obd_type and just use the values from there. */
220 if (!strcmp(name, LUSTRE_OSC_NAME)) {
221 rq_portal = OST_REQUEST_PORTAL;
222 rp_portal = OSC_REPLY_PORTAL;
223 connect_op = OST_CONNECT;
224 cli->cl_sp_me = LUSTRE_SP_CLI;
225 cli->cl_sp_to = LUSTRE_SP_OST;
226 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
227 rq_portal = MDS_REQUEST_PORTAL;
228 rp_portal = MDC_REPLY_PORTAL;
229 connect_op = MDS_CONNECT;
230 cli->cl_sp_me = LUSTRE_SP_CLI;
231 cli->cl_sp_to = LUSTRE_SP_MDT;
232 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
233 rq_portal = MGS_REQUEST_PORTAL;
234 rp_portal = MGC_REPLY_PORTAL;
235 connect_op = MGS_CONNECT;
236 cli->cl_sp_me = LUSTRE_SP_MGC;
237 cli->cl_sp_to = LUSTRE_SP_MGS;
238 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
240 CERROR("unknown client OBD type \"%s\", can't setup\n",
245 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
246 CERROR("requires a TARGET UUID\n");
250 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
251 CERROR("client UUID must be less than 38 characters\n");
255 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
256 CERROR("setup requires a SERVER UUID\n");
260 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
261 CERROR("target UUID must be less than 38 characters\n");
265 init_rwsem(&cli->cl_sem);
266 sema_init(&cli->cl_mgc_sem, 1);
267 cli->cl_conn_count = 0;
268 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
269 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
270 sizeof(server_uuid)));
273 cli->cl_avail_grant = 0;
274 /* FIXME: should limit this for the sum of all cl_dirty_max */
275 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
276 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
277 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
278 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
279 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
280 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
281 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
282 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
283 client_obd_list_lock_init(&cli->cl_loi_list_lock);
284 cli->cl_r_in_flight = 0;
285 cli->cl_w_in_flight = 0;
287 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
288 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
289 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
290 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
291 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
292 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
293 cfs_waitq_init(&cli->cl_destroy_waitq);
294 atomic_set(&cli->cl_destroy_in_flight, 0);
295 #ifdef ENABLE_CHECKSUM
296 /* Turn on checksumming by default. */
297 cli->cl_checksum = 1;
299 * The supported checksum types will be worked out at connect time
300 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
303 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
305 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
307 /* This value may be changed at connect time in
308 ptlrpc_connect_interpret. */
309 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
310 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
312 if (!strcmp(name, LUSTRE_MDC_NAME)) {
313 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
314 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
315 cli->cl_max_rpcs_in_flight = 2;
316 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
317 cli->cl_max_rpcs_in_flight = 3;
318 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
319 cli->cl_max_rpcs_in_flight = 4;
321 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
326 CERROR("ldlm_get_ref failed: %d\n", rc);
330 ptlrpc_init_client(rq_portal, rp_portal, name,
331 &obddev->obd_ldlm_client);
333 imp = class_new_import(obddev);
335 GOTO(err_ldlm, rc = -ENOENT);
336 imp->imp_client = &obddev->obd_ldlm_client;
337 imp->imp_connect_op = connect_op;
338 imp->imp_initial_recov = 1;
339 imp->imp_initial_recov_bk = 0;
340 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
341 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
342 LUSTRE_CFG_BUFLEN(lcfg, 1));
343 class_import_put(imp);
345 rc = client_import_add_conn(imp, &server_uuid, 1);
347 CERROR("can't add initial connection\n");
348 GOTO(err_import, rc);
351 cli->cl_import = imp;
352 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
353 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
354 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
356 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
357 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
358 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
359 name, obddev->obd_name,
360 cli->cl_target_uuid.uuid);
361 spin_lock(&imp->imp_lock);
362 imp->imp_invalid = 1;
363 spin_unlock(&imp->imp_lock);
367 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
372 class_destroy_import(imp);
380 int client_obd_cleanup(struct obd_device *obddev)
387 /* ->o_connect() method for client side (OSC and MDC and MGC) */
388 int client_connect_import(const struct lu_env *env,
389 struct obd_export **exp,
390 struct obd_device *obd, struct obd_uuid *cluuid,
391 struct obd_connect_data *data, void *localdata)
393 struct client_obd *cli = &obd->u.cli;
394 struct obd_import *imp = cli->cl_import;
395 struct obd_connect_data *ocd;
396 struct ldlm_namespace *to_be_freed = NULL;
397 struct lustre_handle conn = { 0 };
402 down_write(&cli->cl_sem);
403 if (cli->cl_conn_count > 0 )
404 GOTO(out_sem, rc = -EALREADY);
406 rc = class_connect(&conn, obd, cluuid);
410 cli->cl_conn_count++;
411 *exp = class_conn2export(&conn);
413 if (obd->obd_namespace != NULL)
414 CERROR("already have namespace!\n");
417 * Deadlock case - bug 18380
419 up_write(&cli->cl_sem);
420 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
421 LDLM_NAMESPACE_CLIENT,
422 LDLM_NAMESPACE_GREEDY);
423 down_write(&cli->cl_sem);
424 if (obd->obd_namespace == NULL)
425 GOTO(out_disco, rc = -ENOMEM);
427 imp->imp_dlm_handle = conn;
428 rc = ptlrpc_init_import(imp);
432 ocd = &imp->imp_connect_data;
435 imp->imp_connect_flags_orig = data->ocd_connect_flags;
438 rc = ptlrpc_connect_import(imp, NULL);
440 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
443 LASSERT((*exp)->exp_connection);
446 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
447 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
448 data->ocd_connect_flags, ocd->ocd_connect_flags);
449 data->ocd_connect_flags = ocd->ocd_connect_flags;
452 ptlrpc_pinger_add_import(imp);
458 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
459 to_be_freed = obd->obd_namespace;
460 obd->obd_namespace = NULL;
462 cli->cl_conn_count--;
463 class_disconnect(*exp);
467 up_write(&cli->cl_sem);
469 ldlm_namespace_free_post(to_be_freed);
474 int client_disconnect_export(struct obd_export *exp)
476 struct obd_device *obd = class_exp2obd(exp);
477 struct client_obd *cli;
478 struct obd_import *imp;
480 struct ldlm_namespace *to_be_freed = NULL;
484 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
485 exp, exp ? exp->exp_handle.h_cookie : -1);
490 imp = cli->cl_import;
492 down_write(&cli->cl_sem);
493 if (!cli->cl_conn_count) {
494 CERROR("disconnecting disconnected device (%s)\n",
496 GOTO(out_sem, rc = -EINVAL);
499 cli->cl_conn_count--;
500 if (cli->cl_conn_count)
501 GOTO(out_no_disconnect, rc = 0);
503 /* Mark import deactivated now, so we don't try to reconnect if any
504 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
505 * fully deactivate the import, or that would drop all requests. */
506 spin_lock(&imp->imp_lock);
507 imp->imp_deactive = 1;
508 spin_unlock(&imp->imp_lock);
510 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
511 * delete it regardless. (It's safe to delete an import that was
513 (void)ptlrpc_pinger_del_import(imp);
515 if (obd->obd_namespace != NULL) {
516 /* obd_force == local only */
517 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
518 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
520 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
521 to_be_freed = obd->obd_namespace;
525 * there's no necessary to hold sem during diconnecting an import,
526 * and actually it may cause deadlock in gss.
528 up_write(&cli->cl_sem);
529 rc = ptlrpc_disconnect_import(imp, 0);
530 down_write(&cli->cl_sem);
532 ptlrpc_invalidate_import(imp);
533 /* set obd_namespace to NULL only after invalidate, because we can have
534 * some connect requests in flight, and his need store a connect flags
535 * in obd_namespace. bug 14260 */
536 obd->obd_namespace = NULL;
538 if (imp->imp_rq_pool) {
539 ptlrpc_free_rq_pool(imp->imp_rq_pool);
540 imp->imp_rq_pool = NULL;
542 client_destroy_import(imp);
543 cli->cl_import = NULL;
547 err = class_disconnect(exp);
551 up_write(&cli->cl_sem);
553 ldlm_namespace_free_post(to_be_freed);
558 /* --------------------------------------------------------------------------
559 * from old lib/target.c
560 * -------------------------------------------------------------------------- */
562 static int target_handle_reconnect(struct lustre_handle *conn,
563 struct obd_export *exp,
564 struct obd_uuid *cluuid)
567 if (exp->exp_connection && exp->exp_imp_reverse) {
568 struct lustre_handle *hdl;
569 hdl = &exp->exp_imp_reverse->imp_remote_handle;
570 /* Might be a re-connect after a partition. */
571 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
572 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
574 conn->cookie = exp->exp_handle.h_cookie;
575 /* target_handle_connect() treats EALREADY and
576 * -EALREADY differently. EALREADY means we are
577 * doing a valid reconnect from the same client. */
580 CERROR("%s reconnecting from %s, "
581 "handle mismatch (ours "LPX64", theirs "
582 LPX64")\n", cluuid->uuid,
583 exp->exp_connection->c_remote_uuid.uuid,
584 hdl->cookie, conn->cookie);
585 memset(conn, 0, sizeof *conn);
586 /* target_handle_connect() treats EALREADY and
587 * -EALREADY differently. -EALREADY is an error
588 * (same UUID, different handle). */
593 conn->cookie = exp->exp_handle.h_cookie;
594 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
595 cluuid->uuid, exp, conn->cookie);
599 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
602 struct obd_export *exp = cb_data;
604 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
605 obd->obd_name, exp->exp_client_uuid.uuid);
607 spin_lock(&exp->exp_lock);
608 exp->exp_need_sync = 0;
609 spin_unlock(&exp->exp_lock);
611 EXPORT_SYMBOL(target_client_add_cb);
613 target_start_and_reset_recovery_timer(struct obd_device *obd,
614 struct ptlrpc_request *req,
617 int target_handle_connect(struct ptlrpc_request *req)
619 struct obd_device *target, *targref = NULL;
620 struct obd_export *export = NULL;
621 struct obd_import *revimp;
622 struct lustre_handle conn;
623 struct lustre_handle *tmp;
624 struct obd_uuid tgtuuid;
625 struct obd_uuid cluuid;
626 struct obd_uuid remote_uuid;
630 struct obd_connect_data *data, *tmpdata;
631 lnet_nid_t *client_nid = NULL;
634 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
636 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
638 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
639 GOTO(out, rc = -EINVAL);
642 obd_str2uuid(&tgtuuid, str);
643 target = class_uuid2obd(&tgtuuid);
645 target = class_name2obd(str);
647 if (!target || target->obd_stopping || !target->obd_set_up) {
648 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
649 " for connect (%s)\n", str,
650 !target ? "no target" :
651 (target->obd_stopping ? "stopping" :
653 GOTO(out, rc = -ENODEV);
656 if (target->obd_no_conn) {
657 LCONSOLE_WARN("%s: temporarily refusing client connection "
658 "from %s\n", target->obd_name,
659 libcfs_nid2str(req->rq_peer.nid));
660 GOTO(out, rc = -EAGAIN);
663 /* Make sure the target isn't cleaned up while we're here. Yes,
664 there's still a race between the above check and our incref here.
665 Really, class_uuid2obd should take the ref. */
666 targref = class_incref(target, __FUNCTION__, cfs_current());
669 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
671 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
672 GOTO(out, rc = -EINVAL);
675 obd_str2uuid(&cluuid, str);
677 /* XXX extract a nettype and format accordingly */
678 switch (sizeof(lnet_nid_t)) {
679 /* NB the casts only avoid compiler warnings */
681 snprintf(remote_uuid.uuid, sizeof remote_uuid,
682 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
685 snprintf(remote_uuid.uuid, sizeof remote_uuid,
686 "NET_%x_UUID", (__u32)req->rq_peer.nid);
692 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
694 GOTO(out, rc = -EPROTO);
698 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
700 GOTO(out, rc = -EPROTO);
702 rc = req_capsule_server_pack(&req->rq_pill);
706 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
708 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
709 "libclient connection attempt");
710 GOTO(out, rc = -EPROTO);
711 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
712 LUSTRE_VERSION_ALLOWED_OFFSET ||
713 data->ocd_version > LUSTRE_VERSION_CODE +
714 LUSTRE_VERSION_ALLOWED_OFFSET) {
715 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
716 "libclient connection attempt",
717 data->ocd_version < LUSTRE_VERSION_CODE ?
719 OBD_OCD_VERSION_MAJOR(data->ocd_version),
720 OBD_OCD_VERSION_MINOR(data->ocd_version),
721 OBD_OCD_VERSION_PATCH(data->ocd_version),
722 OBD_OCD_VERSION_FIX(data->ocd_version));
723 data = req_capsule_server_sized_get(&req->rq_pill,
725 offsetof(typeof(*data), ocd_version) +
726 sizeof(data->ocd_version));
728 data->ocd_connect_flags = OBD_CONNECT_VERSION;
729 data->ocd_version = LUSTRE_VERSION_CODE;
731 GOTO(out, rc = -EPROTO);
735 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
736 (data->ocd_connect_flags & OBD_CONNECT_MDS))
739 /* lctl gets a backstage, all-access pass. */
740 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
741 goto dont_check_exports;
743 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
747 /* we've found an export in the hash */
748 if (export->exp_connecting) { /* bug 9635, et. al. */
749 CWARN("%s: exp %p already connecting\n",
750 export->exp_obd->obd_name, export);
751 class_export_put(export);
754 } else if (mds_conn && export->exp_connection) {
755 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
756 /* mds reconnected after failover */
757 CWARN("%s: received MDS connection from NID %s,"
758 " removing former export from NID %s\n",
759 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
760 libcfs_nid2str(export->exp_connection->c_peer.nid));
762 /* new mds connection from the same nid */
763 CWARN("%s: received new MDS connection from NID %s,"
764 " removing former export from same NID\n",
765 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
766 class_fail_export(export);
767 class_export_put(export);
770 } else if (export->exp_connection != NULL &&
771 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
772 /* in mds failover we have static uuid but nid can be
774 CWARN("%s: cookie %s seen on new NID %s when "
775 "existing NID %s is already connected\n",
776 target->obd_name, cluuid.uuid,
777 libcfs_nid2str(req->rq_peer.nid),
778 libcfs_nid2str(export->exp_connection->c_peer.nid));
780 class_export_put(export);
783 spin_lock(&export->exp_lock);
784 export->exp_connecting = 1;
785 spin_unlock(&export->exp_lock);
786 class_export_put(export);
787 LASSERT(export->exp_obd == target);
789 rc = target_handle_reconnect(&conn, export, &cluuid);
792 /* If we found an export, we already unlocked. */
795 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
796 } else if (req->rq_export == NULL &&
797 atomic_read(&export->exp_rpc_count) > 0) {
798 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
799 target->obd_name, cluuid.uuid,
800 libcfs_nid2str(req->rq_peer.nid),
801 export, atomic_read(&export->exp_refcount));
802 GOTO(out, rc = -EBUSY);
803 } else if (req->rq_export != NULL &&
804 (atomic_read(&export->exp_rpc_count) > 1)) {
805 /* the current connect rpc has increased exp_rpc_count */
806 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
807 target->obd_name, cluuid.uuid,
808 libcfs_nid2str(req->rq_peer.nid),
809 export, atomic_read(&export->exp_rpc_count) - 1);
810 GOTO(out, rc = -EBUSY);
811 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
812 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
813 "cookies not random?\n", target->obd_name,
814 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
815 GOTO(out, rc = -EALREADY);
817 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
824 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
825 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
826 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
827 export, (long)cfs_time_current_sec(),
828 export ? (long)export->exp_last_request_time : 0);
830 /* Tell the client if we're in recovery. */
831 if (target->obd_recovering) {
832 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
833 /* If this is the first time a client connects,
834 reset the recovery timer */
836 target_start_and_reset_recovery_timer(target, req,
840 /* We want to handle EALREADY but *not* -EALREADY from
841 * target_handle_reconnect(), return reconnection state in a flag */
842 if (rc == EALREADY) {
843 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
849 /* Tell the client if we support replayable requests */
850 if (target->obd_replayable)
851 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
852 client_nid = &req->rq_peer.nid;
854 if (export == NULL) {
855 if (target->obd_recovering) {
858 t = cfs_timer_deadline(&target->obd_recovery_timer);
859 t = cfs_time_sub(t, cfs_time_current());
860 CERROR("%s: denying connection for new client %s (%s): "
861 "%d clients in recovery for "CFS_TIME_T"s\n",
863 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
864 target->obd_recoverable_clients,
865 cfs_duration_sec(t));
869 rc = obd_connect(req->rq_svc_thread->t_env,
870 &export, target, &cluuid, data,
873 conn.cookie = export->exp_handle.h_cookie;
876 rc = obd_reconnect(req->rq_svc_thread->t_env,
877 export, target, &cluuid, data, client_nid);
879 /* prevous done via class_conn2export */
880 class_export_get(export);
884 /* Return only the parts of obd_connect_data that we understand, so the
885 * client knows that we don't understand the rest. */
887 tmpdata = req_capsule_server_get(&req->rq_pill,
889 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
893 /* If all else goes well, this is our RPC return code. */
896 lustre_msg_set_handle(req->rq_repmsg, &conn);
898 /* If the client and the server are the same node, we will already
899 * have an export that really points to the client's DLM export,
900 * because we have a shared handles table.
902 * XXX this will go away when shaver stops sending the "connect" handle
903 * in the real "remote handle" field of the request --phik 24 Apr 2003
905 if (req->rq_export != NULL)
906 class_export_put(req->rq_export);
908 req->rq_export = export;
910 spin_lock(&export->exp_lock);
911 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
912 spin_unlock(&export->exp_lock);
913 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
914 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
915 export->exp_conn_cnt,
916 lustre_msg_get_conn_cnt(req->rq_reqmsg));
918 GOTO(out, rc = -EALREADY);
920 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
922 /* request from liblustre? Don't evict it for not pinging. */
923 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
924 export->exp_libclient = 1;
925 spin_unlock(&export->exp_lock);
927 spin_lock(&target->obd_dev_lock);
928 list_del_init(&export->exp_obd_chain_timed);
929 spin_unlock(&target->obd_dev_lock);
931 spin_unlock(&export->exp_lock);
934 if (export->exp_connection != NULL)
935 ptlrpc_connection_put(export->exp_connection);
936 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
939 if (hlist_unhashed(&export->exp_nid_hash)) {
940 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
941 &export->exp_connection->c_peer.nid,
942 &export->exp_nid_hash);
945 spin_lock_bh(&target->obd_processing_task_lock);
946 if (target->obd_recovering && !export->exp_in_recovery) {
947 spin_lock(&export->exp_lock);
948 export->exp_in_recovery = 1;
949 export->exp_req_replay_needed = 1;
950 export->exp_lock_replay_needed = 1;
951 spin_unlock(&export->exp_lock);
952 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
953 && (data->ocd_transno == 0))
954 CWARN("Connect with zero transno!\n");
956 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
957 && data->ocd_transno < target->obd_next_recovery_transno)
958 target->obd_next_recovery_transno = data->ocd_transno;
959 target->obd_connected_clients++;
960 /* each connected client is counted as recoverable */
961 target->obd_recoverable_clients++;
962 atomic_inc(&target->obd_req_replay_clients);
963 atomic_inc(&target->obd_lock_replay_clients);
964 if (target->obd_connected_clients ==
965 target->obd_max_recoverable_clients)
966 cfs_waitq_signal(&target->obd_next_transno_waitq);
968 spin_unlock_bh(&target->obd_processing_task_lock);
969 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
972 if (export->exp_imp_reverse != NULL) {
973 /* destroyed import can be still referenced in ctxt */
974 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
975 KEY_REVIMP_UPD, 0, NULL, NULL);
977 /* in some recovery senarios, previous ctx init rpc handled
978 * in sptlrpc_target_export_check() might be used to install
979 * a reverse ctx in this reverse import, and later OBD_CONNECT
980 * using the same gss ctx could reach here and following new
981 * reverse import. note all reverse ctx in new/old import are
982 * actually based on the same gss ctx. so we invalidate ctx
983 * here before destroy import, otherwise flush old import will
984 * lead to remote reverse ctx be destroied, thus the reverse
985 * ctx of new import will lost its peer.
986 * there might be a better way to deal with this???
988 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
990 client_destroy_import(export->exp_imp_reverse);
993 /* for the rest part, we return -ENOTCONN in case of errors
994 * in order to let client initialize connection again.
996 revimp = export->exp_imp_reverse = class_new_import(target);
998 CERROR("fail to alloc new reverse import.\n");
999 GOTO(out, rc = -ENOTCONN);
1002 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1003 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1004 revimp->imp_remote_handle = conn;
1005 revimp->imp_dlm_fake = 1;
1006 revimp->imp_state = LUSTRE_IMP_FULL;
1007 /* this is a bit of a layering violation, but much less risk than
1008 * changing this very complex and race-prone code. bug=16839 */
1009 if (data->ocd_connect_flags & OBD_CONNECT_MDS)
1010 obd_set_info_async(export, sizeof(KEY_MDS_CONN), KEY_MDS_CONN,
1013 /* unknown versions will be caught in
1014 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1015 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1017 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1018 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1019 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1021 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1023 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1025 CERROR("Failed to get sec for reverse import: %d\n", rc);
1026 export->exp_imp_reverse = NULL;
1027 class_destroy_import(revimp);
1030 class_import_put(revimp);
1033 spin_lock(&export->exp_lock);
1034 export->exp_connecting = 0;
1035 spin_unlock(&export->exp_lock);
1038 class_decref(targref, __FUNCTION__, cfs_current());
1040 req->rq_status = rc;
1044 int target_handle_disconnect(struct ptlrpc_request *req)
1049 rc = req_capsule_server_pack(&req->rq_pill);
1053 /* keep the rq_export around so we can send the reply */
1054 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1059 void target_destroy_export(struct obd_export *exp)
1061 /* exports created from last_rcvd data, and "fake"
1062 exports created by lctl don't have an import */
1063 if (exp->exp_imp_reverse != NULL)
1064 client_destroy_import(exp->exp_imp_reverse);
1066 /* We cancel locks at disconnect time, but this will catch any locks
1067 * granted in a race with recovery-induced disconnect. */
1068 if (exp->exp_obd->obd_namespace != NULL)
1069 ldlm_cancel_locks_for_export(exp);
1073 * Recovery functions
1076 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1078 struct ptlrpc_request *copy_req;
1079 struct lustre_msg *copy_reqmsg;
1080 struct ptlrpc_user_desc *udesc = NULL;
1082 OBD_ALLOC_PTR(copy_req);
1085 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1087 OBD_FREE_PTR(copy_req);
1091 if (orig_req->rq_user_desc) {
1092 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1094 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1096 OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1097 OBD_FREE_PTR(copy_req);
1100 memcpy(udesc, orig_req->rq_user_desc,
1101 sptlrpc_user_desc_size(ngroups));
1104 *copy_req = *orig_req;
1105 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1106 copy_req->rq_reqmsg = copy_reqmsg;
1107 copy_req->rq_user_desc = udesc;
1109 class_export_rpc_get(copy_req->rq_export);
1110 CFS_INIT_LIST_HEAD(©_req->rq_list);
1111 CFS_INIT_LIST_HEAD(©_req->rq_replay_list);
1112 sptlrpc_svc_ctx_addref(copy_req);
1114 if (copy_req->rq_reply_state) {
1115 /* the copied req takes over the reply state */
1116 orig_req->rq_reply_state = NULL;
1117 /* to catch further access */
1118 orig_req->rq_repmsg = NULL;
1119 orig_req->rq_replen = 0;
1125 void ptlrpc_free_clone(struct ptlrpc_request *req)
1127 LASSERT(list_empty(&req->rq_replay_list));
1129 ptlrpc_req_drop_rs(req);
1130 sptlrpc_svc_ctx_decref(req);
1131 class_export_rpc_put(req->rq_export);
1132 list_del_init(&req->rq_list);
1134 if (req->rq_user_desc) {
1135 int ngroups = req->rq_user_desc->pud_ngroups;
1136 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1138 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1142 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1144 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1145 struct obd_export *exp = req->rq_export;
1146 struct ptlrpc_request *reqiter;
1151 spin_lock(&exp->exp_lock);
1152 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1154 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1161 /* we expect it with RESENT and REPLAY flags */
1162 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1163 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1164 CERROR("invalid flags %x of resent replay\n",
1165 lustre_msg_get_flags(req->rq_reqmsg));
1167 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1170 spin_unlock(&exp->exp_lock);
1174 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1176 LASSERT(!list_empty(&req->rq_replay_list));
1177 LASSERT(req->rq_export);
1179 spin_lock(&req->rq_export->exp_lock);
1180 list_del_init(&req->rq_replay_list);
1181 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 spin_lock_bh(&obd->obd_processing_task_lock);
1193 if (list_empty(&obd->obd_req_replay_queue) &&
1194 list_empty(&obd->obd_lock_replay_queue) &&
1195 list_empty(&obd->obd_final_req_queue)) {
1196 obd->obd_processing_task = 0;
1198 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1200 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1201 list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1202 list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1203 spin_unlock_bh(&obd->obd_processing_task_lock);
1206 spin_unlock_bh(&obd->obd_processing_task_lock);
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",
1216 obd->obd_recovery_end = cfs_time_current_sec();
1220 static void abort_req_replay_queue(struct obd_device *obd)
1222 struct ptlrpc_request *req, *n;
1223 struct list_head abort_list;
1225 CFS_INIT_LIST_HEAD(&abort_list);
1226 spin_lock_bh(&obd->obd_processing_task_lock);
1227 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1228 spin_unlock_bh(&obd->obd_processing_task_lock);
1229 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 ptlrpc_free_clone(req);
1241 static void abort_lock_replay_queue(struct obd_device *obd)
1243 struct ptlrpc_request *req, *n;
1244 struct list_head abort_list;
1246 CFS_INIT_LIST_HEAD(&abort_list);
1247 spin_lock_bh(&obd->obd_processing_task_lock);
1248 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1249 spin_unlock_bh(&obd->obd_processing_task_lock);
1250 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 ptlrpc_free_clone(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 struct list_head clean_list;
1277 LASSERT(obd->obd_stopping);
1279 CFS_INIT_LIST_HEAD(&clean_list);
1280 spin_lock_bh(&obd->obd_processing_task_lock);
1281 if (!obd->obd_recovering) {
1282 spin_unlock_bh(&obd->obd_processing_task_lock);
1286 obd->obd_recovering = obd->obd_abort_recovery = 0;
1287 target_cancel_recovery_timer(obd);
1289 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1290 spin_unlock_bh(&obd->obd_processing_task_lock);
1292 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 ptlrpc_free_clone(req);
1298 spin_lock_bh(&obd->obd_processing_task_lock);
1299 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1300 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1301 spin_unlock_bh(&obd->obd_processing_task_lock);
1303 list_for_each_entry_safe(req, n, &clean_list, rq_list){
1304 LASSERT(req->rq_reply_state == 0);
1305 ptlrpc_free_clone(req);
1311 /* obd_processing_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 spin_lock_bh(&obd->obd_processing_task_lock);
1327 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1328 spin_unlock_bh(&obd->obd_processing_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;
1341 * If total recovery time already exceed the
1342 * obd_recovery_max_time, then CRAY XT3 will
1343 * abort the recovery
1345 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1346 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1348 obd->obd_recovery_end = obd->obd_recovery_start +
1349 obd->obd_recovery_timeout;
1350 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1351 cfs_time_before(now, obd->obd_recovery_end)) {
1352 left = cfs_time_sub(obd->obd_recovery_end, now);
1353 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1355 spin_unlock_bh(&obd->obd_processing_task_lock);
1356 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1357 obd->obd_name, (unsigned)left);
1360 static void check_and_start_recovery_timer(struct obd_device *obd)
1362 spin_lock_bh(&obd->obd_processing_task_lock);
1363 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1364 spin_unlock_bh(&obd->obd_processing_task_lock);
1367 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1368 obd->obd_recovery_start = cfs_time_current_sec();
1370 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1371 spin_unlock_bh(&obd->obd_processing_task_lock);
1373 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1376 /* Reset the timer with each new client connection */
1378 * This timer is actually reconnect_timer, which is for making sure
1379 * the total recovery window is at least as big as my reconnect
1380 * attempt timing. So the initial recovery time_out will be set to
1381 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1382 * from client is bigger than this, then the recovery time_out will
1383 * be extend to make sure the client could be reconnected, in the
1384 * process, the timeout from the new client should be ignored.
1388 target_start_and_reset_recovery_timer(struct obd_device *obd,
1389 struct ptlrpc_request *req,
1392 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1394 if (!new_client && service_time)
1395 /* Teach server about old server's estimates, as first guess
1396 * at how long new requests will take. */
1397 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1400 check_and_start_recovery_timer(obd);
1402 /* convert the service time to rpc timeout,
1403 * reuse service_time to limit stack usage */
1404 service_time = at_est2timeout(service_time);
1406 /* We expect other clients to timeout within service_time, then try
1407 * to reconnect, then try the failover server. The max delay between
1408 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1409 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1410 INITIAL_CONNECT_TIMEOUT);
1411 if (service_time > obd->obd_recovery_timeout && !new_client)
1412 reset_recovery_timer(obd, service_time, 0);
1416 static int check_for_next_transno(struct obd_device *obd)
1418 struct ptlrpc_request *req = NULL;
1419 int wake_up = 0, connected, completed, queue_len, max;
1420 __u64 next_transno, req_transno;
1422 spin_lock_bh(&obd->obd_processing_task_lock);
1424 if (!list_empty(&obd->obd_req_replay_queue)) {
1425 req = list_entry(obd->obd_req_replay_queue.next,
1426 struct ptlrpc_request, rq_list);
1427 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1432 max = obd->obd_max_recoverable_clients;
1433 connected = obd->obd_connected_clients;
1434 completed = connected - obd->obd_recoverable_clients;
1435 queue_len = obd->obd_requests_queued_for_recovery;
1436 next_transno = obd->obd_next_recovery_transno;
1438 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1439 "req_transno: "LPU64", next_transno: "LPU64"\n",
1440 max, connected, completed, queue_len, req_transno, next_transno);
1442 if (obd->obd_abort_recovery) {
1443 CDEBUG(D_HA, "waking for aborted recovery\n");
1445 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1446 CDEBUG(D_HA, "waking for completed recovery\n");
1448 } else if (req_transno == next_transno) {
1449 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1451 } else if (queue_len + completed == max) {
1452 /* handle gaps occured due to lost reply. It is allowed gaps
1453 * because all clients are connected and there will be resend
1454 * for missed transaction */
1455 LASSERTF(req_transno >= next_transno,
1456 "req_transno: "LPU64", next_transno: "LPU64"\n",
1457 req_transno, next_transno);
1459 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1460 "waking for skipped transno (skip: "LPD64
1461 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1462 next_transno, queue_len, completed, connected, req_transno);
1463 obd->obd_next_recovery_transno = req_transno;
1465 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1466 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1467 " by fail_lock, waking up ("LPD64")\n", next_transno);
1468 obd->obd_next_recovery_transno = req_transno;
1470 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1471 /* some clients haven't connected in time, but we can try
1472 * to replay requests that demand on already committed ones
1473 * also, we can replay first non-committed transation */
1474 LASSERT(req_transno != 0);
1475 if (obd->obd_version_recov ||
1476 req_transno == obd->obd_last_committed + 1) {
1477 obd->obd_next_recovery_transno = req_transno;
1478 } else if (req_transno > obd->obd_last_committed) {
1479 /* can't continue recovery: have no needed transno */
1480 obd->obd_abort_recovery = 1;
1481 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1482 "connected: %d, completed: %d, queue_len: %d, "
1483 "req_transno: "LPU64", next_transno: "LPU64"\n",
1484 max, connected, completed, queue_len,
1485 req_transno, next_transno);
1490 spin_unlock_bh(&obd->obd_processing_task_lock);
1494 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1496 struct l_wait_info lwi = { 0 };
1497 struct ptlrpc_request *req;
1499 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1500 obd->obd_next_recovery_transno);
1501 l_wait_event(obd->obd_next_transno_waitq,
1502 check_for_next_transno(obd), &lwi);
1504 spin_lock_bh(&obd->obd_processing_task_lock);
1505 if (obd->obd_abort_recovery) {
1507 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1508 req = list_entry(obd->obd_req_replay_queue.next,
1509 struct ptlrpc_request, rq_list);
1510 target_exp_dequeue_req_replay(req);
1511 list_del_init(&req->rq_list);
1512 obd->obd_requests_queued_for_recovery--;
1516 spin_unlock_bh(&obd->obd_processing_task_lock);
1520 static int check_for_next_lock(struct obd_device *obd)
1522 struct ptlrpc_request *req = NULL;
1525 spin_lock_bh(&obd->obd_processing_task_lock);
1526 if (!list_empty(&obd->obd_lock_replay_queue)) {
1527 req = list_entry(obd->obd_lock_replay_queue.next,
1528 struct ptlrpc_request, rq_list);
1529 CDEBUG(D_HA, "waking for next lock\n");
1531 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1532 CDEBUG(D_HA, "waking for completed lock replay\n");
1534 } else if (obd->obd_abort_recovery) {
1535 CDEBUG(D_HA, "waking for aborted recovery\n");
1538 spin_unlock_bh(&obd->obd_processing_task_lock);
1543 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1545 struct l_wait_info lwi = { 0 };
1546 struct ptlrpc_request *req;
1548 CDEBUG(D_HA, "Waiting for lock\n");
1549 l_wait_event(obd->obd_next_transno_waitq,
1550 check_for_next_lock(obd), &lwi);
1552 spin_lock_bh(&obd->obd_processing_task_lock);
1553 if (obd->obd_abort_recovery) {
1555 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1556 req = list_entry(obd->obd_lock_replay_queue.next,
1557 struct ptlrpc_request, rq_list);
1558 list_del_init(&req->rq_list);
1562 spin_unlock_bh(&obd->obd_processing_task_lock);
1566 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1568 struct ptlrpc_request *req;
1570 spin_lock_bh(&obd->obd_processing_task_lock);
1571 if (!list_empty(&obd->obd_final_req_queue)) {
1572 req = list_entry(obd->obd_final_req_queue.next,
1573 struct ptlrpc_request, rq_list);
1574 list_del_init(&req->rq_list);
1575 if (req->rq_export->exp_in_recovery) {
1576 spin_lock(&req->rq_export->exp_lock);
1577 req->rq_export->exp_in_recovery = 0;
1578 spin_unlock(&req->rq_export->exp_lock);
1583 spin_unlock_bh(&obd->obd_processing_task_lock);
1587 static inline int req_vbr_done(struct obd_export *exp)
1589 return (exp->exp_vbr_failed == 0);
1592 static inline int req_replay_done(struct obd_export *exp)
1594 return (exp->exp_req_replay_needed == 0);
1597 static inline int lock_replay_done(struct obd_export *exp)
1599 return (exp->exp_lock_replay_needed == 0);
1602 static inline int connect_done(struct obd_export *exp)
1604 return (exp->exp_in_recovery != 0);
1607 static int check_for_clients(struct obd_device *obd)
1609 if (obd->obd_abort_recovery || obd->obd_version_recov)
1611 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1612 if (obd->obd_no_conn == 0 &&
1613 obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1618 static int handle_recovery_req(struct ptlrpc_thread *thread,
1619 struct ptlrpc_request *req,
1620 svc_handler_t handler)
1625 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1627 CERROR("Failure to initialize session: %d\n", rc);
1630 req->rq_session.lc_thread = thread;
1631 lu_context_enter(&req->rq_session);
1632 req->rq_svc_thread = thread;
1633 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1635 /* thread context */
1636 lu_context_enter(&thread->t_env->le_ctx);
1638 lu_context_exit(&thread->t_env->le_ctx);
1640 lu_context_exit(&req->rq_session);
1641 lu_context_fini(&req->rq_session);
1642 /* don't reset timer for final stage */
1643 if (!req_replay_done(req->rq_export) ||
1644 !lock_replay_done(req->rq_export))
1645 reset_recovery_timer(class_exp2obd(req->rq_export),
1646 AT_OFF ? obd_timeout :
1647 at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1648 ptlrpc_free_clone(req);
1652 static int target_recovery_thread(void *arg)
1654 struct lu_target *lut = arg;
1655 struct obd_device *obd = lut->lut_obd;
1656 struct ptlrpc_request *req;
1657 struct target_recovery_data *trd = &obd->obd_recovery_data;
1658 struct l_wait_info lwi = { 0 };
1659 unsigned long delta;
1660 unsigned long flags;
1662 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1666 cfs_daemonize("tgt_recov");
1668 SIGNAL_MASK_LOCK(current, flags);
1669 sigfillset(¤t->blocked);
1671 SIGNAL_MASK_UNLOCK(current, flags);
1673 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1677 thread->t_env = &env;
1678 env.le_ctx.lc_thread = thread;
1680 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1682 trd->trd_processing_task = cfs_curproc_pid();
1684 obd->obd_recovering = 1;
1685 complete(&trd->trd_starting);
1687 /* first of all, we have to know the first transno to replay */
1688 obd->obd_abort_recovery = 0;
1689 l_wait_event(obd->obd_next_transno_waitq,
1690 check_for_clients(obd), &lwi);
1692 /* If some clients haven't connected in time, evict them */
1693 if (obd->obd_connected_clients < obd->obd_max_recoverable_clients) {
1694 CWARN("Some clients haven't connect in time (%d/%d),"
1695 "evict them\n", obd->obd_connected_clients,
1696 obd->obd_max_recoverable_clients);
1697 class_disconnect_stale_exports(obd, connect_done,
1698 exp_flags_from_obd(obd) |
1699 OBD_OPT_ABORT_RECOV);
1701 * if recovery proceeds with versions then some clients may be
1702 * timed out waiting for others and trying to reconnect.
1703 * Extend timer for such reconnect cases.
1705 if (obd->obd_version_recov)
1706 reset_recovery_timer(obd, RECONNECT_DELAY_MAX * 2, 1);
1709 /* next stage: replay requests */
1711 obd->obd_req_replaying = 1;
1712 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1713 atomic_read(&obd->obd_req_replay_clients),
1714 obd->obd_next_recovery_transno);
1715 while ((req = target_next_replay_req(obd))) {
1716 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1717 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1718 lustre_msg_get_transno(req->rq_reqmsg),
1719 libcfs_nid2str(req->rq_peer.nid));
1720 handle_recovery_req(thread, req,
1721 trd->trd_recovery_handler);
1722 obd->obd_replayed_requests++;
1723 spin_lock_bh(&obd->obd_processing_task_lock);
1724 obd->obd_next_recovery_transno++;
1725 spin_unlock_bh(&obd->obd_processing_task_lock);
1728 /* If some clients haven't replayed requests in time, evict them */
1729 if (obd->obd_abort_recovery) {
1730 CDEBUG(D_WARNING, "req replay is aborted\n");
1731 class_disconnect_stale_exports(obd, req_replay_done,
1732 exp_flags_from_obd(obd) |
1733 OBD_OPT_ABORT_RECOV);
1734 abort_req_replay_queue(obd);
1736 LASSERT(list_empty(&obd->obd_req_replay_queue));
1738 /* The second stage: replay locks */
1739 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1740 atomic_read(&obd->obd_lock_replay_clients));
1741 while ((req = target_next_replay_lock(obd))) {
1742 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1743 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1744 libcfs_nid2str(req->rq_peer.nid));
1745 handle_recovery_req(thread, req,
1746 trd->trd_recovery_handler);
1747 obd->obd_replayed_locks++;
1750 /* If some clients haven't replayed requests in time, evict them */
1751 if (obd->obd_abort_recovery) {
1752 CERROR("lock replay is aborted\n");
1753 class_disconnect_stale_exports(obd, lock_replay_done,
1754 exp_flags_from_obd(obd) |
1755 OBD_OPT_ABORT_RECOV);
1756 abort_lock_replay_queue(obd);
1758 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1760 /* The third stage: reply on final pings */
1761 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1762 /** evict exports failed VBR */
1763 class_disconnect_stale_exports(obd, req_vbr_done,
1764 exp_flags_from_obd(obd) |
1765 OBD_OPT_ABORT_RECOV);
1766 /** Update server last boot epoch */
1767 lut_boot_epoch_update(lut);
1768 /* We drop recoverying flag to forward all new requests
1769 * to regular mds_handle() since now */
1770 spin_lock_bh(&obd->obd_processing_task_lock);
1771 obd->obd_recovering = obd->obd_abort_recovery = 0;
1772 target_cancel_recovery_timer(obd);
1773 spin_unlock_bh(&obd->obd_processing_task_lock);
1774 while ((req = target_next_final_ping(obd))) {
1775 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1776 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1777 libcfs_nid2str(req->rq_peer.nid));
1778 handle_recovery_req(thread, req,
1779 trd->trd_recovery_handler);
1782 delta = (jiffies - delta) / HZ;
1783 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1784 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1785 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1786 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1787 if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1788 CWARN("too long recovery - read logs\n");
1789 libcfs_debug_dumplog();
1792 target_finish_recovery(obd);
1794 lu_context_fini(&env.le_ctx);
1795 trd->trd_processing_task = 0;
1796 complete(&trd->trd_finishing);
1800 static int target_start_recovery_thread(struct lu_target *lut,
1801 svc_handler_t handler)
1803 struct obd_device *obd = lut->lut_obd;
1805 struct target_recovery_data *trd = &obd->obd_recovery_data;
1807 memset(trd, 0, sizeof(*trd));
1808 init_completion(&trd->trd_starting);
1809 init_completion(&trd->trd_finishing);
1810 trd->trd_recovery_handler = handler;
1812 if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1813 wait_for_completion(&trd->trd_starting);
1814 LASSERT(obd->obd_recovering != 0);
1821 void target_stop_recovery_thread(struct obd_device *obd)
1823 spin_lock_bh(&obd->obd_processing_task_lock);
1824 if (obd->obd_recovery_data.trd_processing_task > 0) {
1825 struct target_recovery_data *trd = &obd->obd_recovery_data;
1826 CERROR("%s: Aborting recovery\n", obd->obd_name);
1827 obd->obd_abort_recovery = 1;
1828 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1829 spin_unlock_bh(&obd->obd_processing_task_lock);
1830 wait_for_completion(&trd->trd_finishing);
1832 spin_unlock_bh(&obd->obd_processing_task_lock);
1836 void target_recovery_fini(struct obd_device *obd)
1838 class_disconnect_exports(obd);
1839 target_stop_recovery_thread(obd);
1840 target_cleanup_recovery(obd);
1842 EXPORT_SYMBOL(target_recovery_fini);
1844 static void target_recovery_expired(unsigned long castmeharder)
1846 struct obd_device *obd = (struct obd_device *)castmeharder;
1847 CDEBUG(D_HA, "%s: recovery timed out; %d clients never reconnected "
1848 "after %lds (%d clients did)\n",
1849 obd->obd_name, obd->obd_recoverable_clients,
1850 cfs_time_current_sec()- obd->obd_recovery_start,
1851 obd->obd_connected_clients);
1853 spin_lock_bh(&obd->obd_processing_task_lock);
1854 obd->obd_version_recov = 1;
1855 CDEBUG(D_INFO, "VBR is used for %d clients from t"LPU64"\n",
1856 atomic_read(&obd->obd_req_replay_clients),
1857 obd->obd_next_recovery_transno);
1858 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1859 spin_unlock_bh(&obd->obd_processing_task_lock);
1862 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1864 struct obd_device *obd = lut->lut_obd;
1865 if (obd->obd_max_recoverable_clients == 0)
1868 CWARN("RECOVERY: service %s, %d recoverable clients, "
1869 "last_transno "LPU64"\n", obd->obd_name,
1870 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1871 LASSERT(obd->obd_stopping == 0);
1872 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1873 obd->obd_recovery_start = 0;
1874 obd->obd_recovery_end = 0;
1875 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1876 /* bz13079: this should be set to desired value for ost but not for mds */
1877 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1878 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1879 target_start_recovery_thread(lut, handler);
1881 EXPORT_SYMBOL(target_recovery_init);
1885 static int target_process_req_flags(struct obd_device *obd,
1886 struct ptlrpc_request *req)
1888 struct obd_export *exp = req->rq_export;
1889 LASSERT(exp != NULL);
1890 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1891 /* client declares he's ready to replay locks */
1892 spin_lock_bh(&obd->obd_processing_task_lock);
1893 if (exp->exp_req_replay_needed) {
1894 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1895 spin_lock(&exp->exp_lock);
1896 exp->exp_req_replay_needed = 0;
1897 spin_unlock(&exp->exp_lock);
1898 atomic_dec(&obd->obd_req_replay_clients);
1899 LASSERT(obd->obd_recoverable_clients > 0);
1900 obd->obd_recoverable_clients--;
1901 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1902 CDEBUG(D_HA, "all clients have replayed reqs\n");
1904 spin_unlock_bh(&obd->obd_processing_task_lock);
1906 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1907 /* client declares he's ready to complete recovery
1908 * so, we put the request on th final queue */
1909 spin_lock_bh(&obd->obd_processing_task_lock);
1910 if (exp->exp_lock_replay_needed) {
1911 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1912 spin_lock(&exp->exp_lock);
1913 exp->exp_lock_replay_needed = 0;
1914 spin_unlock(&exp->exp_lock);
1915 atomic_dec(&obd->obd_lock_replay_clients);
1916 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1917 CDEBUG(D_HA, "all clients have replayed locks\n");
1919 spin_unlock_bh(&obd->obd_processing_task_lock);
1925 int target_queue_recovery_request(struct ptlrpc_request *req,
1926 struct obd_device *obd)
1928 struct list_head *tmp;
1930 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1933 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1934 /* Processing the queue right now, don't re-add. */
1938 target_process_req_flags(obd, req);
1940 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1941 /* client declares he's ready to complete recovery
1942 * so, we put the request on th final queue */
1943 req = ptlrpc_clone_req(req);
1946 DEBUG_REQ(D_HA, req, "queue final req");
1947 spin_lock_bh(&obd->obd_processing_task_lock);
1948 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1949 if (obd->obd_recovering)
1950 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1952 spin_unlock_bh(&obd->obd_processing_task_lock);
1953 ptlrpc_free_clone(req);
1954 if (obd->obd_stopping) {
1960 spin_unlock_bh(&obd->obd_processing_task_lock);
1963 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1964 /* client declares he's ready to replay locks */
1965 req = ptlrpc_clone_req(req);
1968 DEBUG_REQ(D_HA, req, "queue lock replay req");
1969 spin_lock_bh(&obd->obd_processing_task_lock);
1970 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1971 LASSERT(obd->obd_recovering);
1972 /* usually due to recovery abort */
1973 if (!req->rq_export->exp_in_recovery) {
1974 spin_unlock_bh(&obd->obd_processing_task_lock);
1975 ptlrpc_free_clone(req);
1978 LASSERT(req->rq_export->exp_lock_replay_needed);
1979 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1980 spin_unlock_bh(&obd->obd_processing_task_lock);
1984 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1985 * (i.e. buflens etc are in my own byte order), but type-dependent
1986 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1989 CFS_INIT_LIST_HEAD(&req->rq_list);
1990 DEBUG_REQ(D_HA, req, "not queueing");
1994 spin_lock_bh(&obd->obd_processing_task_lock);
1996 /* If we're processing the queue, we want don't want to queue this
1999 * Also, if this request has a transno less than the one we're waiting
2000 * for, we should process it now. It could (and currently always will)
2001 * be an open request for a descriptor that was opened some time ago.
2003 * Also, a resent, replayed request that has already been
2004 * handled will pass through here and be processed immediately.
2006 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2007 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2008 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2009 /* Processing the queue right now, don't re-add. */
2010 LASSERT(list_empty(&req->rq_list));
2011 spin_unlock_bh(&obd->obd_processing_task_lock);
2014 spin_unlock_bh(&obd->obd_processing_task_lock);
2016 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2019 req = ptlrpc_clone_req(req);
2023 spin_lock_bh(&obd->obd_processing_task_lock);
2024 LASSERT(obd->obd_recovering);
2025 if (!req->rq_export->exp_in_recovery) {
2026 spin_unlock_bh(&obd->obd_processing_task_lock);
2027 ptlrpc_free_clone(req);
2030 LASSERT(req->rq_export->exp_req_replay_needed);
2032 if (target_exp_enqueue_req_replay(req)) {
2033 spin_unlock_bh(&obd->obd_processing_task_lock);
2034 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2035 ptlrpc_free_clone(req);
2040 list_for_each(tmp, &obd->obd_req_replay_queue) {
2041 struct ptlrpc_request *reqiter =
2042 list_entry(tmp, struct ptlrpc_request, rq_list);
2044 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2045 list_add_tail(&req->rq_list, &reqiter->rq_list);
2050 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2052 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2053 "has been claimed by another client");
2054 spin_unlock_bh(&obd->obd_processing_task_lock);
2055 target_exp_dequeue_req_replay(req);
2056 ptlrpc_free_clone(req);
2062 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2064 obd->obd_requests_queued_for_recovery++;
2065 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2066 spin_unlock_bh(&obd->obd_processing_task_lock);
2070 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2072 return req->rq_export->exp_obd;
2075 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2077 LASSERT(exp != NULL);
2078 return &exp->exp_obd->obd_namespace->ns_pool;
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 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 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 (list_empty(&rs->rs_obd_list));
2166 LASSERT (list_empty(&rs->rs_exp_list));
2168 exp = class_export_get (req->rq_export);
2171 /* disable reply scheduling onto srv_reply_queue 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;
2178 spin_lock(&exp->exp_uncommitted_replies_lock);
2179 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2180 rs->rs_transno, exp->exp_last_committed);
2181 if (rs->rs_transno > exp->exp_last_committed) {
2182 /* not committed already */
2183 list_add_tail(&rs->rs_obd_list,
2184 &exp->exp_uncommitted_replies);
2186 spin_unlock (&exp->exp_uncommitted_replies_lock);
2188 spin_lock(&exp->exp_lock);
2189 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2190 spin_unlock(&exp->exp_lock);
2192 netrc = target_send_reply_msg (req, rc, fail_id);
2194 spin_lock(&svc->srv_lock);
2196 atomic_inc(&svc->srv_n_difficult_replies);
2199 /* error sending: reply is off the net. Also we need +1
2200 * reply ref until ptlrpc_server_handle_reply() is done
2201 * with the reply state (if the send was successful, there
2202 * would have been +1 ref for the net, which
2203 * reply_out_callback leaves alone) */
2205 ptlrpc_rs_addref(rs);
2206 atomic_inc (&svc->srv_outstanding_replies);
2209 spin_lock(&rs->rs_lock);
2210 if (rs->rs_transno <= exp->exp_last_committed ||
2211 (!rs->rs_on_net && !rs->rs_no_ack) ||
2212 list_empty(&rs->rs_exp_list) || /* completed already */
2213 list_empty(&rs->rs_obd_list)) {
2214 CDEBUG(D_HA, "Schedule reply immediately\n");
2215 ptlrpc_dispatch_difficult_reply(rs);
2217 list_add (&rs->rs_list, &svc->srv_active_replies);
2218 rs->rs_scheduled = 0; /* allow notifier to schedule */
2220 spin_unlock(&rs->rs_lock);
2221 spin_unlock(&svc->srv_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;
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 OBD_ALLOC(qdata, sizeof(struct qunit_data));
2290 rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2292 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2296 /* we use the observer */
2297 if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2298 CERROR("Can't find the observer, it is recovering\n");
2299 req->rq_status = -EAGAIN;
2300 GOTO(send_reply, rc = -EAGAIN);
2303 master_obd = obd->obd_observer->obd_observer;
2304 obt = &master_obd->u.obt;
2305 qctxt = &obt->obt_qctxt;
2307 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2308 /* quota_type has not been processed yet, return EAGAIN
2309 * until we know whether or not quotas are supposed to
2311 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2313 req->rq_status = -EAGAIN;
2314 rc = ptlrpc_reply(req);
2318 down_read(&obt->obt_rwsem);
2319 if (qctxt->lqc_lqs_hash == NULL) {
2320 up_read(&obt->obt_rwsem);
2321 /* quota_type has not been processed yet, return EAGAIN
2322 * until we know whether or not quotas are supposed to
2324 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2326 req->rq_status = -EAGAIN;
2327 rc = ptlrpc_reply(req);
2331 LASSERT(qctxt->lqc_handler);
2332 rc = qctxt->lqc_handler(master_obd, qdata,
2333 lustre_msg_get_opc(req->rq_reqmsg));
2334 up_read(&obt->obt_rwsem);
2335 if (rc && rc != -EDQUOT)
2336 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2337 "dqacq failed! (rc:%d)\n", rc);
2338 req->rq_status = rc;
2340 /* there are three forms of qunit(historic causes), so we need to
2341 * adjust the same form to different forms slaves needed */
2342 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2344 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2348 /* Block the quota req. b=14840 */
2349 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2351 rc = ptlrpc_reply(req);
2353 OBD_FREE(qdata, sizeof(struct qunit_data));
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: %i\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);