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 cfs_spin_lock(&imp->imp_lock);
84 cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 cfs_list_del(&item->oic_item);
88 cfs_list_add(&item->oic_item,
90 item->oic_last_attempt = 0;
92 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
93 imp, imp->imp_obd->obd_name, uuid->uuid,
94 (priority ? ", moved to head" : ""));
95 cfs_spin_unlock(&imp->imp_lock);
96 GOTO(out_free, rc = 0);
101 imp_conn->oic_conn = ptlrpc_conn;
102 imp_conn->oic_uuid = *uuid;
103 imp_conn->oic_last_attempt = 0;
105 cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
107 cfs_list_add_tail(&imp_conn->oic_item,
108 &imp->imp_conn_list);
109 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
110 imp, imp->imp_obd->obd_name, uuid->uuid,
111 (priority ? "head" : "tail"));
113 cfs_spin_unlock(&imp->imp_lock);
114 GOTO(out_free, rc = -ENOENT);
117 cfs_spin_unlock(&imp->imp_lock);
121 OBD_FREE(imp_conn, sizeof(*imp_conn));
123 ptlrpc_connection_put(ptlrpc_conn);
127 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
129 return import_set_conn(imp, uuid, 1, 0);
132 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
135 return import_set_conn(imp, uuid, priority, 1);
138 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
140 struct obd_import_conn *imp_conn;
141 struct obd_export *dlmexp;
145 cfs_spin_lock(&imp->imp_lock);
146 if (cfs_list_empty(&imp->imp_conn_list)) {
147 LASSERT(!imp->imp_connection);
151 cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
152 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
154 LASSERT(imp_conn->oic_conn);
156 /* is current conn? */
157 if (imp_conn == imp->imp_conn_current) {
158 LASSERT(imp_conn->oic_conn == imp->imp_connection);
160 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
161 imp->imp_state != LUSTRE_IMP_DISCON) {
162 CERROR("can't remove current connection\n");
163 GOTO(out, rc = -EBUSY);
166 ptlrpc_connection_put(imp->imp_connection);
167 imp->imp_connection = NULL;
169 dlmexp = class_conn2export(&imp->imp_dlm_handle);
170 if (dlmexp && dlmexp->exp_connection) {
171 LASSERT(dlmexp->exp_connection ==
173 ptlrpc_connection_put(dlmexp->exp_connection);
174 dlmexp->exp_connection = NULL;
178 cfs_list_del(&imp_conn->oic_item);
179 ptlrpc_connection_put(imp_conn->oic_conn);
180 OBD_FREE(imp_conn, sizeof(*imp_conn));
181 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182 imp, imp->imp_obd->obd_name, uuid->uuid);
187 cfs_spin_unlock(&imp->imp_lock);
189 CERROR("connection %s not found\n", uuid->uuid);
193 void client_destroy_import(struct obd_import *imp)
195 /* drop security policy instance after all rpc finished/aborted
196 * to let all busy contexts be released. */
197 class_import_get(imp);
198 class_destroy_import(imp);
199 sptlrpc_import_sec_put(imp);
200 class_import_put(imp);
203 /* configure an RPC client OBD device
208 * 3 - inactive-on-startup
210 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
212 struct client_obd *cli = &obddev->u.cli;
213 struct obd_import *imp;
214 struct obd_uuid server_uuid;
215 int rq_portal, rp_portal, connect_op;
216 char *name = obddev->obd_type->typ_name;
220 /* In a more perfect world, we would hang a ptlrpc_client off of
221 * obd_type and just use the values from there. */
222 if (!strcmp(name, LUSTRE_OSC_NAME)) {
223 rq_portal = OST_REQUEST_PORTAL;
224 rp_portal = OSC_REPLY_PORTAL;
225 connect_op = OST_CONNECT;
226 cli->cl_sp_me = LUSTRE_SP_CLI;
227 cli->cl_sp_to = LUSTRE_SP_OST;
228 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
229 rq_portal = MDS_REQUEST_PORTAL;
230 rp_portal = MDC_REPLY_PORTAL;
231 connect_op = MDS_CONNECT;
232 cli->cl_sp_me = LUSTRE_SP_CLI;
233 cli->cl_sp_to = LUSTRE_SP_MDT;
234 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
235 rq_portal = MGS_REQUEST_PORTAL;
236 rp_portal = MGC_REPLY_PORTAL;
237 connect_op = MGS_CONNECT;
238 cli->cl_sp_me = LUSTRE_SP_MGC;
239 cli->cl_sp_to = LUSTRE_SP_MGS;
240 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
242 CERROR("unknown client OBD type \"%s\", can't setup\n",
247 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
248 CERROR("requires a TARGET UUID\n");
252 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
253 CERROR("client UUID must be less than 38 characters\n");
257 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
258 CERROR("setup requires a SERVER UUID\n");
262 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
263 CERROR("target UUID must be less than 38 characters\n");
267 cfs_init_rwsem(&cli->cl_sem);
268 cfs_sema_init(&cli->cl_mgc_sem, 1);
269 cli->cl_conn_count = 0;
270 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
271 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
272 sizeof(server_uuid)));
275 cli->cl_avail_grant = 0;
276 /* FIXME: should limit this for the sum of all cl_dirty_max */
277 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
278 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
279 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
280 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
281 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
282 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
283 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
284 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
285 client_obd_list_lock_init(&cli->cl_loi_list_lock);
286 cli->cl_r_in_flight = 0;
287 cli->cl_w_in_flight = 0;
289 cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
290 cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
291 cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
292 cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
293 cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
294 cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
295 cfs_waitq_init(&cli->cl_destroy_waitq);
296 cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
297 #ifdef ENABLE_CHECKSUM
298 /* Turn on checksumming by default. */
299 cli->cl_checksum = 1;
301 * The supported checksum types will be worked out at connect time
302 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
305 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
307 cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
308 cfs_atomic_set(&cli->cl_quota_resends, CLIENT_QUOTA_DEFAULT_RESENDS);
310 /* This value may be changed at connect time in
311 ptlrpc_connect_interpret. */
312 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
313 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
315 if (!strcmp(name, LUSTRE_MDC_NAME)) {
316 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
317 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
318 cli->cl_max_rpcs_in_flight = 2;
319 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
320 cli->cl_max_rpcs_in_flight = 3;
321 } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
322 cli->cl_max_rpcs_in_flight = 4;
324 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
329 CERROR("ldlm_get_ref failed: %d\n", rc);
333 ptlrpc_init_client(rq_portal, rp_portal, name,
334 &obddev->obd_ldlm_client);
336 imp = class_new_import(obddev);
338 GOTO(err_ldlm, rc = -ENOENT);
339 imp->imp_client = &obddev->obd_ldlm_client;
340 imp->imp_connect_op = connect_op;
341 imp->imp_initial_recov = 1;
342 imp->imp_initial_recov_bk = 0;
343 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
344 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
345 LUSTRE_CFG_BUFLEN(lcfg, 1));
346 class_import_put(imp);
348 rc = client_import_add_conn(imp, &server_uuid, 1);
350 CERROR("can't add initial connection\n");
351 GOTO(err_import, rc);
354 cli->cl_import = imp;
355 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
356 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
357 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
359 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
360 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
361 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
362 name, obddev->obd_name,
363 cli->cl_target_uuid.uuid);
364 cfs_spin_lock(&imp->imp_lock);
365 imp->imp_deactive = 1;
366 cfs_spin_unlock(&imp->imp_lock);
370 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
371 LDLM_NAMESPACE_CLIENT,
372 LDLM_NAMESPACE_GREEDY);
373 if (obddev->obd_namespace == NULL) {
374 CERROR("Unable to create client namespace - %s\n",
376 GOTO(err_import, rc = -ENOMEM);
379 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
384 class_destroy_import(imp);
392 int client_obd_cleanup(struct obd_device *obddev)
396 ldlm_namespace_free_post(obddev->obd_namespace);
397 obddev->obd_namespace = NULL;
403 /* ->o_connect() method for client side (OSC and MDC and MGC) */
404 int client_connect_import(const struct lu_env *env,
405 struct obd_export **exp,
406 struct obd_device *obd, struct obd_uuid *cluuid,
407 struct obd_connect_data *data, void *localdata)
409 struct client_obd *cli = &obd->u.cli;
410 struct obd_import *imp = cli->cl_import;
411 struct obd_connect_data *ocd;
412 struct lustre_handle conn = { 0 };
417 cfs_down_write(&cli->cl_sem);
418 if (cli->cl_conn_count > 0 )
419 GOTO(out_sem, rc = -EALREADY);
421 rc = class_connect(&conn, obd, cluuid);
425 cli->cl_conn_count++;
426 *exp = class_conn2export(&conn);
428 LASSERT(obd->obd_namespace);
430 imp->imp_dlm_handle = conn;
431 rc = ptlrpc_init_import(imp);
435 ocd = &imp->imp_connect_data;
438 imp->imp_connect_flags_orig = data->ocd_connect_flags;
441 rc = ptlrpc_connect_import(imp, NULL);
443 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
446 LASSERT((*exp)->exp_connection);
449 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
450 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
451 data->ocd_connect_flags, ocd->ocd_connect_flags);
452 data->ocd_connect_flags = ocd->ocd_connect_flags;
455 ptlrpc_pinger_add_import(imp);
461 cli->cl_conn_count--;
462 class_disconnect(*exp);
466 cfs_up_write(&cli->cl_sem);
471 int client_disconnect_export(struct obd_export *exp)
473 struct obd_device *obd = class_exp2obd(exp);
474 struct client_obd *cli;
475 struct obd_import *imp;
480 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
481 exp, exp ? exp->exp_handle.h_cookie : -1);
486 imp = cli->cl_import;
488 cfs_down_write(&cli->cl_sem);
489 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
492 if (!cli->cl_conn_count) {
493 CERROR("disconnecting disconnected device (%s)\n",
495 GOTO(out_disconnect, rc = -EINVAL);
498 cli->cl_conn_count--;
499 if (cli->cl_conn_count)
500 GOTO(out_disconnect, rc = 0);
502 /* Mark import deactivated now, so we don't try to reconnect if any
503 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
504 * fully deactivate the import, or that would drop all requests. */
505 cfs_spin_lock(&imp->imp_lock);
506 imp->imp_deactive = 1;
507 cfs_spin_unlock(&imp->imp_lock);
509 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
510 * delete it regardless. (It's safe to delete an import that was
512 (void)ptlrpc_pinger_del_import(imp);
514 if (obd->obd_namespace != NULL) {
515 /* obd_force == local only */
516 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
517 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
519 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
523 * there's no need to hold sem during disconnecting an import,
524 * and actually it may cause deadlock in gss.
526 cfs_up_write(&cli->cl_sem);
527 rc = ptlrpc_disconnect_import(imp, 0);
528 cfs_down_write(&cli->cl_sem);
530 ptlrpc_invalidate_import(imp);
532 if (imp->imp_rq_pool) {
533 ptlrpc_free_rq_pool(imp->imp_rq_pool);
534 imp->imp_rq_pool = NULL;
536 client_destroy_import(imp);
537 cli->cl_import = NULL;
542 /* use server style - class_disconnect should be always called for
544 err = class_disconnect(exp);
548 cfs_up_write(&cli->cl_sem);
553 int server_disconnect_export(struct obd_export *exp)
558 /* Disconnect early so that clients can't keep using export */
559 rc = class_disconnect(exp);
560 /* close import for avoid sending any requests */
561 if (exp->exp_imp_reverse)
562 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
564 if (exp->exp_obd->obd_namespace != NULL)
565 ldlm_cancel_locks_for_export(exp);
567 /* complete all outstanding replies */
568 cfs_spin_lock(&exp->exp_lock);
569 while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
570 struct ptlrpc_reply_state *rs =
571 cfs_list_entry(exp->exp_outstanding_replies.next,
572 struct ptlrpc_reply_state, rs_exp_list);
573 struct ptlrpc_service *svc = rs->rs_service;
575 cfs_spin_lock(&svc->srv_lock);
576 cfs_list_del_init(&rs->rs_exp_list);
577 cfs_spin_lock(&rs->rs_lock);
578 ptlrpc_schedule_difficult_reply(rs);
579 cfs_spin_unlock(&rs->rs_lock);
580 cfs_spin_unlock(&svc->srv_lock);
582 cfs_spin_unlock(&exp->exp_lock);
584 /* release nid stat refererence */
585 lprocfs_exp_cleanup(exp);
590 /* --------------------------------------------------------------------------
591 * from old lib/target.c
592 * -------------------------------------------------------------------------- */
594 static int target_handle_reconnect(struct lustre_handle *conn,
595 struct obd_export *exp,
596 struct obd_uuid *cluuid)
599 if (exp->exp_connection && exp->exp_imp_reverse) {
600 struct lustre_handle *hdl;
601 hdl = &exp->exp_imp_reverse->imp_remote_handle;
602 /* Might be a re-connect after a partition. */
603 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
604 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
606 conn->cookie = exp->exp_handle.h_cookie;
607 /* target_handle_connect() treats EALREADY and
608 * -EALREADY differently. EALREADY means we are
609 * doing a valid reconnect from the same client. */
612 CERROR("%s reconnecting from %s, "
613 "handle mismatch (ours "LPX64", theirs "
614 LPX64")\n", cluuid->uuid,
615 exp->exp_connection->c_remote_uuid.uuid,
616 hdl->cookie, conn->cookie);
617 memset(conn, 0, sizeof *conn);
618 /* target_handle_connect() treats EALREADY and
619 * -EALREADY differently. -EALREADY is an error
620 * (same UUID, different handle). */
625 conn->cookie = exp->exp_handle.h_cookie;
626 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
627 cluuid->uuid, exp, conn->cookie);
631 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
634 struct obd_export *exp = cb_data;
636 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
637 obd->obd_name, exp->exp_client_uuid.uuid);
639 cfs_spin_lock(&exp->exp_lock);
640 exp->exp_need_sync = 0;
641 cfs_spin_unlock(&exp->exp_lock);
642 class_export_cb_put(exp);
644 EXPORT_SYMBOL(target_client_add_cb);
647 target_start_and_reset_recovery_timer(struct obd_device *obd,
648 struct ptlrpc_request *req,
651 int target_handle_connect(struct ptlrpc_request *req)
653 struct obd_device *target, *targref = NULL;
654 struct obd_export *export = NULL;
655 struct obd_import *revimp;
656 struct lustre_handle conn;
657 struct lustre_handle *tmp;
658 struct obd_uuid tgtuuid;
659 struct obd_uuid cluuid;
660 struct obd_uuid remote_uuid;
664 struct obd_connect_data *data, *tmpdata;
665 lnet_nid_t *client_nid = NULL;
668 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
670 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
672 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
673 GOTO(out, rc = -EINVAL);
676 obd_str2uuid(&tgtuuid, str);
677 target = class_uuid2obd(&tgtuuid);
679 target = class_name2obd(str);
681 if (!target || target->obd_stopping || !target->obd_set_up) {
682 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
683 " for connect (%s)\n", str,
684 !target ? "no target" :
685 (target->obd_stopping ? "stopping" :
687 GOTO(out, rc = -ENODEV);
690 if (target->obd_no_conn) {
691 LCONSOLE_WARN("%s: temporarily refusing client connection "
692 "from %s\n", target->obd_name,
693 libcfs_nid2str(req->rq_peer.nid));
694 GOTO(out, rc = -EAGAIN);
697 /* Make sure the target isn't cleaned up while we're here. Yes,
698 there's still a race between the above check and our incref here.
699 Really, class_uuid2obd should take the ref. */
700 targref = class_incref(target, __FUNCTION__, cfs_current());
703 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
705 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
706 GOTO(out, rc = -EINVAL);
709 obd_str2uuid(&cluuid, str);
711 /* XXX extract a nettype and format accordingly */
712 switch (sizeof(lnet_nid_t)) {
713 /* NB the casts only avoid compiler warnings */
715 snprintf(remote_uuid.uuid, sizeof remote_uuid,
716 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
719 snprintf(remote_uuid.uuid, sizeof remote_uuid,
720 "NET_%x_UUID", (__u32)req->rq_peer.nid);
726 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
728 GOTO(out, rc = -EPROTO);
732 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
734 GOTO(out, rc = -EPROTO);
736 rc = req_capsule_server_pack(&req->rq_pill);
740 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
742 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
743 "libclient connection attempt");
744 GOTO(out, rc = -EPROTO);
745 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
746 LUSTRE_VERSION_ALLOWED_OFFSET ||
747 data->ocd_version > LUSTRE_VERSION_CODE +
748 LUSTRE_VERSION_ALLOWED_OFFSET) {
749 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
750 "libclient connection attempt",
751 data->ocd_version < LUSTRE_VERSION_CODE ?
753 OBD_OCD_VERSION_MAJOR(data->ocd_version),
754 OBD_OCD_VERSION_MINOR(data->ocd_version),
755 OBD_OCD_VERSION_PATCH(data->ocd_version),
756 OBD_OCD_VERSION_FIX(data->ocd_version));
757 data = req_capsule_server_sized_get(&req->rq_pill,
759 offsetof(typeof(*data), ocd_version) +
760 sizeof(data->ocd_version));
762 data->ocd_connect_flags = OBD_CONNECT_VERSION;
763 data->ocd_version = LUSTRE_VERSION_CODE;
765 GOTO(out, rc = -EPROTO);
769 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
770 (data->ocd_connect_flags & OBD_CONNECT_MDS))
773 /* lctl gets a backstage, all-access pass. */
774 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
775 goto dont_check_exports;
777 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
781 /* we've found an export in the hash */
782 if (export->exp_connecting) { /* bug 9635, et. al. */
783 CWARN("%s: exp %p already connecting\n",
784 export->exp_obd->obd_name, export);
785 class_export_put(export);
788 } else if (mds_conn && export->exp_connection) {
789 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
790 /* mds reconnected after failover */
791 CWARN("%s: received MDS connection from NID %s,"
792 " removing former export from NID %s\n",
793 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
794 libcfs_nid2str(export->exp_connection->c_peer.nid));
796 /* new mds connection from the same nid */
797 CWARN("%s: received new MDS connection from NID %s,"
798 " removing former export from same NID\n",
799 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
800 class_fail_export(export);
801 class_export_put(export);
804 } else if (export->exp_connection != NULL &&
805 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
806 (lustre_msg_get_op_flags(req->rq_reqmsg) &
807 MSG_CONNECT_INITIAL)) {
808 /* in mds failover we have static uuid but nid can be
810 CWARN("%s: cookie %s seen on new NID %s when "
811 "existing NID %s is already connected\n",
812 target->obd_name, cluuid.uuid,
813 libcfs_nid2str(req->rq_peer.nid),
814 libcfs_nid2str(export->exp_connection->c_peer.nid));
816 class_export_put(export);
819 cfs_spin_lock(&export->exp_lock);
820 export->exp_connecting = 1;
821 cfs_spin_unlock(&export->exp_lock);
822 class_export_put(export);
823 LASSERT(export->exp_obd == target);
825 rc = target_handle_reconnect(&conn, export, &cluuid);
828 /* If we found an export, we already unlocked. */
831 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
832 } else if (req->rq_export == NULL &&
833 cfs_atomic_read(&export->exp_rpc_count) > 0) {
834 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
835 target->obd_name, cluuid.uuid,
836 libcfs_nid2str(req->rq_peer.nid),
837 export, cfs_atomic_read(&export->exp_refcount));
838 GOTO(out, rc = -EBUSY);
839 } else if (req->rq_export != NULL &&
840 (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
841 /* the current connect rpc has increased exp_rpc_count */
842 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
843 target->obd_name, cluuid.uuid,
844 libcfs_nid2str(req->rq_peer.nid),
845 export, cfs_atomic_read(&export->exp_rpc_count) - 1);
846 cfs_spin_lock(&export->exp_lock);
847 if (req->rq_export->exp_conn_cnt <
848 lustre_msg_get_conn_cnt(req->rq_reqmsg))
849 /* try to abort active requests */
850 req->rq_export->exp_abort_active_req = 1;
851 cfs_spin_unlock(&export->exp_lock);
852 GOTO(out, rc = -EBUSY);
853 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
854 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
855 "cookies not random?\n", target->obd_name,
856 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
857 GOTO(out, rc = -EALREADY);
859 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
866 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
867 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
868 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
869 export, (long)cfs_time_current_sec(),
870 export ? (long)export->exp_last_request_time : 0);
872 /* Tell the client if we're in recovery. */
873 if (target->obd_recovering) {
874 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
875 /* If this is the first time a client connects,
876 reset the recovery timer */
878 target_start_and_reset_recovery_timer(target, req,
882 /* We want to handle EALREADY but *not* -EALREADY from
883 * target_handle_reconnect(), return reconnection state in a flag */
884 if (rc == EALREADY) {
885 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
891 /* Tell the client if we support replayable requests */
892 if (target->obd_replayable)
893 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
894 client_nid = &req->rq_peer.nid;
896 if (export == NULL) {
897 if (target->obd_recovering) {
900 t = cfs_timer_deadline(&target->obd_recovery_timer);
901 t = cfs_time_sub(t, cfs_time_current());
902 CERROR("%s: denying connection for new client %s (%s): "
903 "%d clients in recovery for "CFS_TIME_T"s\n",
905 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
906 cfs_atomic_read(&target-> \
907 obd_lock_replay_clients),
908 cfs_duration_sec(t));
912 rc = obd_connect(req->rq_svc_thread->t_env,
913 &export, target, &cluuid, data,
916 conn.cookie = export->exp_handle.h_cookie;
919 rc = obd_reconnect(req->rq_svc_thread->t_env,
920 export, target, &cluuid, data, client_nid);
922 /* prevous done via class_conn2export */
923 class_export_get(export);
927 /* Return only the parts of obd_connect_data that we understand, so the
928 * client knows that we don't understand the rest. */
930 tmpdata = req_capsule_server_get(&req->rq_pill,
932 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
936 /* If all else goes well, this is our RPC return code. */
939 lustre_msg_set_handle(req->rq_repmsg, &conn);
941 /* If the client and the server are the same node, we will already
942 * have an export that really points to the client's DLM export,
943 * because we have a shared handles table.
945 * XXX this will go away when shaver stops sending the "connect" handle
946 * in the real "remote handle" field of the request --phik 24 Apr 2003
948 if (req->rq_export != NULL)
949 class_export_put(req->rq_export);
951 req->rq_export = export;
953 cfs_spin_lock(&export->exp_lock);
954 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
955 cfs_spin_unlock(&export->exp_lock);
956 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
957 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
958 export->exp_conn_cnt,
959 lustre_msg_get_conn_cnt(req->rq_reqmsg));
961 GOTO(out, rc = -EALREADY);
963 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
964 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
965 export->exp_abort_active_req = 0;
967 /* request from liblustre? Don't evict it for not pinging. */
968 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
969 export->exp_libclient = 1;
970 cfs_spin_unlock(&export->exp_lock);
972 cfs_spin_lock(&target->obd_dev_lock);
973 cfs_list_del_init(&export->exp_obd_chain_timed);
974 cfs_spin_unlock(&target->obd_dev_lock);
976 cfs_spin_unlock(&export->exp_lock);
979 if (export->exp_connection != NULL) {
980 /* Check to see if connection came from another NID */
981 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
982 !cfs_hlist_unhashed(&export->exp_nid_hash))
983 cfs_hash_del(export->exp_obd->obd_nid_hash,
984 &export->exp_connection->c_peer.nid,
985 &export->exp_nid_hash);
987 ptlrpc_connection_put(export->exp_connection);
990 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
993 if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
994 cfs_hash_add_unique(export->exp_obd->obd_nid_hash,
995 &export->exp_connection->c_peer.nid,
996 &export->exp_nid_hash);
999 cfs_spin_lock_bh(&target->obd_processing_task_lock);
1000 if (target->obd_recovering && !export->exp_in_recovery) {
1001 cfs_spin_lock(&export->exp_lock);
1002 export->exp_in_recovery = 1;
1003 export->exp_req_replay_needed = 1;
1004 export->exp_lock_replay_needed = 1;
1005 cfs_spin_unlock(&export->exp_lock);
1006 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1007 && (data->ocd_transno == 0))
1008 CWARN("Connect with zero transno!\n");
1010 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1011 && data->ocd_transno < target->obd_next_recovery_transno)
1012 target->obd_next_recovery_transno = data->ocd_transno;
1013 target->obd_connected_clients++;
1014 cfs_atomic_inc(&target->obd_req_replay_clients);
1015 cfs_atomic_inc(&target->obd_lock_replay_clients);
1016 if (target->obd_connected_clients ==
1017 target->obd_max_recoverable_clients)
1018 cfs_waitq_signal(&target->obd_next_transno_waitq);
1020 cfs_spin_unlock_bh(&target->obd_processing_task_lock);
1021 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1024 if (export->exp_imp_reverse != NULL) {
1025 /* destroyed import can be still referenced in ctxt */
1026 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1027 KEY_REVIMP_UPD, 0, NULL, NULL);
1029 client_destroy_import(export->exp_imp_reverse);
1032 /* for the rest part, we return -ENOTCONN in case of errors
1033 * in order to let client initialize connection again.
1035 revimp = export->exp_imp_reverse = class_new_import(target);
1037 CERROR("fail to alloc new reverse import.\n");
1038 GOTO(out, rc = -ENOTCONN);
1041 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1042 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1043 revimp->imp_remote_handle = conn;
1044 revimp->imp_dlm_fake = 1;
1045 revimp->imp_state = LUSTRE_IMP_FULL;
1047 /* unknown versions will be caught in
1048 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1049 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1051 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1052 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1053 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1055 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1057 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1059 CERROR("Failed to get sec for reverse import: %d\n", rc);
1060 export->exp_imp_reverse = NULL;
1061 class_destroy_import(revimp);
1064 class_import_put(revimp);
1067 cfs_spin_lock(&export->exp_lock);
1068 export->exp_connecting = 0;
1069 cfs_spin_unlock(&export->exp_lock);
1072 class_decref(targref, __FUNCTION__, cfs_current());
1074 req->rq_status = rc;
1078 int target_handle_disconnect(struct ptlrpc_request *req)
1083 rc = req_capsule_server_pack(&req->rq_pill);
1087 /* keep the rq_export around so we can send the reply */
1088 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1093 void target_destroy_export(struct obd_export *exp)
1095 /* exports created from last_rcvd data, and "fake"
1096 exports created by lctl don't have an import */
1097 if (exp->exp_imp_reverse != NULL)
1098 client_destroy_import(exp->exp_imp_reverse);
1100 LASSERT(cfs_atomic_read(&exp->exp_locks_count) == 0);
1101 LASSERT(cfs_atomic_read(&exp->exp_rpc_count) == 0);
1102 LASSERT(cfs_atomic_read(&exp->exp_cb_count) == 0);
1103 LASSERT(cfs_atomic_read(&exp->exp_replay_count) == 0);
1107 * Recovery functions
1109 static void target_request_copy_get(struct ptlrpc_request *req)
1111 class_export_rpc_get(req->rq_export);
1112 LASSERT(cfs_list_empty(&req->rq_list));
1113 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1114 /* increase refcount to keep request in queue */
1115 LASSERT(cfs_atomic_read(&req->rq_refcount));
1116 cfs_atomic_inc(&req->rq_refcount);
1117 /** let export know it has replays to be handled */
1118 cfs_atomic_inc(&req->rq_export->exp_replay_count);
1119 /* release service thread while request is queued
1120 * we are moving the request from active processing
1121 * to waiting on the replay queue */
1122 ptlrpc_server_active_request_dec(req);
1125 static void target_request_copy_put(struct ptlrpc_request *req)
1127 LASSERT(cfs_list_empty(&req->rq_replay_list));
1128 LASSERT(cfs_atomic_read(&req->rq_export->exp_replay_count) > 0);
1129 cfs_atomic_dec(&req->rq_export->exp_replay_count);
1130 class_export_rpc_put(req->rq_export);
1131 /* ptlrpc_server_drop_request() assumes the request is active */
1132 ptlrpc_server_active_request_inc(req);
1133 ptlrpc_server_drop_request(req);
1136 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1138 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1139 struct obd_export *exp = req->rq_export;
1140 struct ptlrpc_request *reqiter;
1145 cfs_spin_lock(&exp->exp_lock);
1146 cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1148 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1155 /* we expect it with RESENT and REPLAY flags */
1156 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1157 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1158 CERROR("invalid flags %x of resent replay\n",
1159 lustre_msg_get_flags(req->rq_reqmsg));
1161 cfs_list_add_tail(&req->rq_replay_list,
1162 &exp->exp_req_replay_queue);
1165 cfs_spin_unlock(&exp->exp_lock);
1169 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1171 LASSERT(!cfs_list_empty(&req->rq_replay_list));
1172 LASSERT(req->rq_export);
1174 cfs_spin_lock(&req->rq_export->exp_lock);
1175 cfs_list_del_init(&req->rq_replay_list);
1176 cfs_spin_unlock(&req->rq_export->exp_lock);
1180 static void target_finish_recovery(struct obd_device *obd)
1183 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1186 ldlm_reprocess_all_ns(obd->obd_namespace);
1187 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1188 if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1189 !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1190 !cfs_list_empty(&obd->obd_final_req_queue)) {
1191 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1193 cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1194 cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1196 cfs_list_empty(&obd->obd_final_req_queue) ? \
1198 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1201 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1203 obd->obd_recovery_end = cfs_time_current_sec();
1205 /* when recovery finished, cleanup orphans on mds and ost */
1206 if (OBT(obd) && OBP(obd, postrecov)) {
1207 int rc = OBP(obd, postrecov)(obd);
1209 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1215 static void abort_req_replay_queue(struct obd_device *obd)
1217 struct ptlrpc_request *req, *n;
1218 cfs_list_t abort_list;
1220 CFS_INIT_LIST_HEAD(&abort_list);
1221 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1222 cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1223 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1224 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1225 DEBUG_REQ(D_WARNING, req, "aborted:");
1226 req->rq_status = -ENOTCONN;
1227 if (ptlrpc_error(req)) {
1228 DEBUG_REQ(D_ERROR, req,
1229 "failed abort_req_reply; skipping");
1231 target_exp_dequeue_req_replay(req);
1232 target_request_copy_put(req);
1236 static void abort_lock_replay_queue(struct obd_device *obd)
1238 struct ptlrpc_request *req, *n;
1239 cfs_list_t abort_list;
1241 CFS_INIT_LIST_HEAD(&abort_list);
1242 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1243 cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1244 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1245 cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1246 DEBUG_REQ(D_ERROR, req, "aborted:");
1247 req->rq_status = -ENOTCONN;
1248 if (ptlrpc_error(req)) {
1249 DEBUG_REQ(D_ERROR, req,
1250 "failed abort_lock_reply; skipping");
1252 target_request_copy_put(req);
1257 /* Called from a cleanup function if the device is being cleaned up
1258 forcefully. The exports should all have been disconnected already,
1259 the only thing left to do is
1260 - clear the recovery flags
1262 - free queued requests and replies, but don't send replies
1263 Because the obd_stopping flag is set, no new requests should be received.
1266 void target_cleanup_recovery(struct obd_device *obd)
1268 struct ptlrpc_request *req, *n;
1269 cfs_list_t clean_list;
1272 CFS_INIT_LIST_HEAD(&clean_list);
1273 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1274 if (!obd->obd_recovering) {
1275 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1279 obd->obd_recovering = obd->obd_abort_recovery = 0;
1280 target_cancel_recovery_timer(obd);
1282 cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1283 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1285 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1286 LASSERT(req->rq_reply_state == 0);
1287 target_exp_dequeue_req_replay(req);
1288 target_request_copy_put(req);
1291 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1292 cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1293 cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1294 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1296 cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1297 LASSERT(req->rq_reply_state == 0);
1298 target_request_copy_put(req);
1304 /* obd_processing_task_lock should be held */
1305 void target_cancel_recovery_timer(struct obd_device *obd)
1307 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1308 cfs_timer_disarm(&obd->obd_recovery_timer);
1311 /* extend = 1 means require at least "duration" seconds left in the timer,
1312 extend = 0 means set the total duration (start_recovery_timer) */
1313 static void reset_recovery_timer(struct obd_device *obd, int duration,
1316 cfs_time_t now = cfs_time_current_sec();
1317 cfs_duration_t left;
1319 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1320 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1321 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1325 left = cfs_time_sub(obd->obd_recovery_end, now);
1327 if (extend && (duration > left))
1328 obd->obd_recovery_timeout += duration - left;
1329 else if (!extend && (duration > obd->obd_recovery_timeout))
1330 /* Track the client's largest expected replay time */
1331 obd->obd_recovery_timeout = duration;
1334 * If total recovery time already exceed the
1335 * obd_recovery_max_time, then CRAY XT3 will
1336 * abort the recovery
1338 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1339 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1341 obd->obd_recovery_end = obd->obd_recovery_start +
1342 obd->obd_recovery_timeout;
1343 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1344 cfs_time_before(now, obd->obd_recovery_end)) {
1345 left = cfs_time_sub(obd->obd_recovery_end, now);
1346 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1348 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1349 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1350 obd->obd_name, (unsigned)left);
1353 static void check_and_start_recovery_timer(struct obd_device *obd)
1355 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1356 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1357 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1360 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1361 obd->obd_recovery_start = cfs_time_current_sec();
1363 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1364 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1366 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1369 /* Reset the timer with each new client connection */
1371 * This timer is actually reconnect_timer, which is for making sure
1372 * the total recovery window is at least as big as my reconnect
1373 * attempt timing. So the initial recovery time_out will be set to
1374 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1375 * from client is bigger than this, then the recovery time_out will
1376 * be extended to make sure the client could be reconnected, in the
1377 * process, the timeout from the new client should be ignored.
1381 target_start_and_reset_recovery_timer(struct obd_device *obd,
1382 struct ptlrpc_request *req,
1385 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1387 if (!new_client && service_time)
1388 /* Teach server about old server's estimates, as first guess
1389 * at how long new requests will take. */
1390 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1393 check_and_start_recovery_timer(obd);
1395 /* convert the service time to rpc timeout,
1396 * reuse service_time to limit stack usage */
1397 service_time = at_est2timeout(service_time);
1399 /* We expect other clients to timeout within service_time, then try
1400 * to reconnect, then try the failover server. The max delay between
1401 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1402 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1403 INITIAL_CONNECT_TIMEOUT);
1404 if (service_time > obd->obd_recovery_timeout && !new_client)
1405 reset_recovery_timer(obd, service_time, 0);
1410 /** Health checking routines */
1411 static inline int exp_connect_healthy(struct obd_export *exp)
1413 return (exp->exp_in_recovery);
1416 /** if export done req_replay or has replay in queue */
1417 static inline int exp_req_replay_healthy(struct obd_export *exp)
1419 return (!exp->exp_req_replay_needed ||
1420 cfs_atomic_read(&exp->exp_replay_count) > 0);
1422 /** if export done lock_replay or has replay in queue */
1423 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1425 return (!exp->exp_lock_replay_needed ||
1426 cfs_atomic_read(&exp->exp_replay_count) > 0);
1429 static inline int exp_vbr_healthy(struct obd_export *exp)
1431 return (!exp->exp_vbr_failed);
1434 static inline int exp_finished(struct obd_export *exp)
1436 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1439 /** Checking routines for recovery */
1440 static int check_for_clients(struct obd_device *obd)
1442 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1444 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1445 if (obd->obd_no_conn == 0 &&
1446 obd->obd_connected_clients + obd->obd_stale_clients ==
1447 obd->obd_max_recoverable_clients)
1452 static int check_for_next_transno(struct obd_device *obd)
1454 struct ptlrpc_request *req = NULL;
1455 int wake_up = 0, connected, completed, queue_len;
1456 __u64 next_transno, req_transno;
1458 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1460 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1461 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1462 struct ptlrpc_request, rq_list);
1463 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1468 connected = obd->obd_connected_clients;
1469 completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1470 queue_len = obd->obd_requests_queued_for_recovery;
1471 next_transno = obd->obd_next_recovery_transno;
1473 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1474 "req_transno: "LPU64", next_transno: "LPU64"\n",
1475 obd->obd_max_recoverable_clients, connected, completed,
1476 queue_len, req_transno, next_transno);
1478 if (obd->obd_abort_recovery) {
1479 CDEBUG(D_HA, "waking for aborted recovery\n");
1481 } else if (obd->obd_recovery_expired) {
1482 CDEBUG(D_HA, "waking for expired recovery\n");
1484 } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1485 CDEBUG(D_HA, "waking for completed recovery\n");
1487 } else if (req_transno == next_transno) {
1488 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1490 } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1492 /** handle gaps occured due to lost reply or VBR */
1493 LASSERTF(req_transno >= next_transno,
1494 "req_transno: "LPU64", next_transno: "LPU64"\n",
1495 req_transno, next_transno);
1496 if (req_transno > obd->obd_last_committed &&
1497 !obd->obd_version_recov)
1500 "%s: waking for gap in transno, VBR is %s (skip: "
1501 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1502 ", last_committed: "LPD64")\n",
1503 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1504 next_transno, queue_len, completed, connected,
1505 req_transno, obd->obd_last_committed);
1506 obd->obd_next_recovery_transno = req_transno;
1508 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1509 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1510 " by fail_lock, waking up ("LPD64")\n", next_transno);
1511 obd->obd_next_recovery_transno = req_transno;
1514 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1518 static int check_for_next_lock(struct obd_device *obd)
1522 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1523 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1524 CDEBUG(D_HA, "waking for next lock\n");
1526 } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1527 CDEBUG(D_HA, "waking for completed lock replay\n");
1529 } else if (obd->obd_abort_recovery) {
1530 CDEBUG(D_HA, "waking for aborted recovery\n");
1532 } else if (obd->obd_recovery_expired) {
1533 CDEBUG(D_HA, "waking for expired recovery\n");
1536 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1542 * wait for recovery events,
1543 * check its status with help of check_routine
1544 * evict dead clients via health_check
1546 static int target_recovery_overseer(struct obd_device *obd,
1547 int (*check_routine)(struct obd_device *),
1548 int (*health_check)(struct obd_export *))
1550 int abort = 0, expired = 1;
1553 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1554 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1555 abort = obd->obd_abort_recovery;
1556 expired = obd->obd_recovery_expired;
1557 obd->obd_recovery_expired = 0;
1558 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1560 CWARN("recovery is aborted, evict exports in recovery\n");
1561 /** evict exports which didn't finish recovery yet */
1562 class_disconnect_stale_exports(obd, exp_finished);
1563 } else if (expired) {
1564 /** If some clients died being recovered, evict them */
1565 CDEBUG(D_WARNING, "recovery is timed out, evict stale exports\n");
1566 /** evict cexports with no replay in queue, they are stalled */
1567 class_disconnect_stale_exports(obd, health_check);
1568 /** continue with VBR */
1569 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1570 obd->obd_version_recov = 1;
1571 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1573 * reset timer, recovery will proceed with versions now,
1574 * timeout is set just to handle reconnection delays
1576 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1577 /** Wait for recovery events again, after evicting bad clients */
1579 } while (!abort && expired);
1584 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1586 struct ptlrpc_request *req = NULL;
1589 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1590 obd->obd_next_recovery_transno);
1592 if (target_recovery_overseer(obd, check_for_next_transno,
1593 exp_req_replay_healthy)) {
1594 abort_req_replay_queue(obd);
1595 abort_lock_replay_queue(obd);
1598 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1599 if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1600 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1601 struct ptlrpc_request, rq_list);
1602 cfs_list_del_init(&req->rq_list);
1603 obd->obd_requests_queued_for_recovery--;
1604 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1606 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1607 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1608 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1609 /** evict exports failed VBR */
1610 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1615 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1617 struct ptlrpc_request *req = NULL;
1619 CDEBUG(D_HA, "Waiting for lock\n");
1620 if (target_recovery_overseer(obd, check_for_next_lock,
1621 exp_lock_replay_healthy))
1622 abort_lock_replay_queue(obd);
1624 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1625 if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1626 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1627 struct ptlrpc_request, rq_list);
1628 cfs_list_del_init(&req->rq_list);
1629 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1631 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1632 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1633 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1634 /** evict exports failed VBR */
1635 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1640 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1642 struct ptlrpc_request *req = NULL;
1644 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1645 if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1646 req = cfs_list_entry(obd->obd_final_req_queue.next,
1647 struct ptlrpc_request, rq_list);
1648 cfs_list_del_init(&req->rq_list);
1649 if (req->rq_export->exp_in_recovery) {
1650 cfs_spin_lock(&req->rq_export->exp_lock);
1651 req->rq_export->exp_in_recovery = 0;
1652 cfs_spin_unlock(&req->rq_export->exp_lock);
1655 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1659 static int handle_recovery_req(struct ptlrpc_thread *thread,
1660 struct ptlrpc_request *req,
1661 svc_handler_t handler)
1666 rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1668 CERROR("Failure to initialize session: %d\n", rc);
1669 GOTO(reqcopy_put, rc);
1672 * export can be evicted during recovery, no need to handle replays for
1673 * it after that, discard such request silently
1675 if (req->rq_export->exp_disconnected)
1676 GOTO(reqcopy_put, rc);
1678 req->rq_recov_session.lc_thread = thread;
1679 lu_context_enter(&req->rq_recov_session);
1680 req->rq_svc_thread = thread;
1681 req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1683 /* thread context */
1684 lu_context_enter(&thread->t_env->le_ctx);
1686 lu_context_exit(&thread->t_env->le_ctx);
1688 lu_context_exit(&req->rq_recov_session);
1689 lu_context_fini(&req->rq_recov_session);
1690 /* don't reset timer for final stage */
1691 if (!exp_finished(req->rq_export)) {
1693 * Add request timeout to the recovery time so next request from
1694 * this client may come in recovery time
1696 reset_recovery_timer(class_exp2obd(req->rq_export),
1697 AT_OFF ? obd_timeout :
1698 lustre_msg_get_timeout(req->rq_reqmsg), 1);
1704 static int target_recovery_thread(void *arg)
1706 struct lu_target *lut = arg;
1707 struct obd_device *obd = lut->lut_obd;
1708 struct ptlrpc_request *req;
1709 struct target_recovery_data *trd = &obd->obd_recovery_data;
1710 unsigned long delta;
1711 unsigned long flags;
1713 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1717 cfs_daemonize_ctxt("tgt_recov");
1719 SIGNAL_MASK_LOCK(current, flags);
1720 sigfillset(¤t->blocked);
1722 SIGNAL_MASK_UNLOCK(current, flags);
1724 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1728 thread->t_env = &env;
1729 env.le_ctx.lc_thread = thread;
1731 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1733 trd->trd_processing_task = cfs_curproc_pid();
1735 obd->obd_recovering = 1;
1736 cfs_complete(&trd->trd_starting);
1738 /* first of all, we have to know the first transno to replay */
1739 if (target_recovery_overseer(obd, check_for_clients,
1740 exp_connect_healthy)) {
1741 abort_req_replay_queue(obd);
1742 abort_lock_replay_queue(obd);
1745 /* next stage: replay requests */
1747 obd->obd_req_replaying = 1;
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_bh(&obd->obd_processing_task_lock);
1763 obd->obd_next_recovery_transno++;
1764 cfs_spin_unlock_bh(&obd->obd_processing_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_bh(&obd->obd_processing_task_lock);
1795 obd->obd_recovering = obd->obd_abort_recovery = 0;
1796 target_cancel_recovery_timer(obd);
1797 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1798 while ((req = target_next_final_ping(obd))) {
1799 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1800 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1801 libcfs_nid2str(req->rq_peer.nid));
1802 handle_recovery_req(thread, req,
1803 trd->trd_recovery_handler);
1804 target_request_copy_put(req);
1807 delta = (jiffies - delta) / CFS_HZ;
1808 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1809 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1810 if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1811 CWARN("too long recovery - read logs\n");
1812 libcfs_debug_dumplog();
1815 target_finish_recovery(obd);
1817 lu_context_fini(&env.le_ctx);
1818 trd->trd_processing_task = 0;
1819 cfs_complete(&trd->trd_finishing);
1823 static int target_start_recovery_thread(struct lu_target *lut,
1824 svc_handler_t handler)
1826 struct obd_device *obd = lut->lut_obd;
1828 struct target_recovery_data *trd = &obd->obd_recovery_data;
1830 memset(trd, 0, sizeof(*trd));
1831 cfs_init_completion(&trd->trd_starting);
1832 cfs_init_completion(&trd->trd_finishing);
1833 trd->trd_recovery_handler = handler;
1835 if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) {
1836 cfs_wait_for_completion(&trd->trd_starting);
1837 LASSERT(obd->obd_recovering != 0);
1844 void target_stop_recovery_thread(struct obd_device *obd)
1846 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1847 if (obd->obd_recovery_data.trd_processing_task > 0) {
1848 struct target_recovery_data *trd = &obd->obd_recovery_data;
1849 /** recovery can be done but postrecovery is not yet */
1850 if (obd->obd_recovering) {
1851 CERROR("%s: Aborting recovery\n", obd->obd_name);
1852 obd->obd_abort_recovery = 1;
1853 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1855 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1856 cfs_wait_for_completion(&trd->trd_finishing);
1858 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
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 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1880 obd->obd_recovery_expired = 1;
1881 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1882 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1885 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1887 struct obd_device *obd = lut->lut_obd;
1888 if (obd->obd_max_recoverable_clients == 0) {
1889 /** Update server last boot epoch */
1890 lut_boot_epoch_update(lut);
1894 CWARN("RECOVERY: service %s, %d recoverable clients, "
1895 "last_transno "LPU64"\n", obd->obd_name,
1896 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1897 LASSERT(obd->obd_stopping == 0);
1898 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1899 obd->obd_recovery_start = 0;
1900 obd->obd_recovery_end = 0;
1901 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1902 /* bz13079: this should be set to desired value for ost but not for mds */
1903 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1904 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1905 target_start_recovery_thread(lut, handler);
1907 EXPORT_SYMBOL(target_recovery_init);
1911 static int target_process_req_flags(struct obd_device *obd,
1912 struct ptlrpc_request *req)
1914 struct obd_export *exp = req->rq_export;
1915 LASSERT(exp != NULL);
1916 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1917 /* client declares he's ready to replay locks */
1918 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1919 if (exp->exp_req_replay_needed) {
1920 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) >
1922 cfs_spin_lock(&exp->exp_lock);
1923 exp->exp_req_replay_needed = 0;
1924 cfs_spin_unlock(&exp->exp_lock);
1925 cfs_atomic_dec(&obd->obd_req_replay_clients);
1927 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1929 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1930 /* client declares he's ready to complete recovery
1931 * so, we put the request on th final queue */
1932 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
1933 if (exp->exp_lock_replay_needed) {
1934 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) >
1936 cfs_spin_lock(&exp->exp_lock);
1937 exp->exp_lock_replay_needed = 0;
1938 cfs_spin_unlock(&exp->exp_lock);
1939 cfs_atomic_dec(&obd->obd_lock_replay_clients);
1941 cfs_spin_unlock_bh(&obd->obd_processing_task_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_spin_lock_bh(&obd->obd_processing_task_lock);
1968 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1969 if (obd->obd_recovering) {
1970 cfs_list_add_tail(&req->rq_list,
1971 &obd->obd_final_req_queue);
1973 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
1974 target_request_copy_put(req);
1975 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1977 cfs_spin_unlock_bh(&obd->obd_processing_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_spin_lock_bh(&obd->obd_processing_task_lock);
1985 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1986 LASSERT(obd->obd_recovering);
1987 /* usually due to recovery abort */
1988 if (!req->rq_export->exp_in_recovery) {
1989 cfs_spin_unlock_bh(&obd->obd_processing_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_bh(&obd->obd_processing_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 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2011 /* If we're processing the queue, we want don't want to queue this
2014 * Also, if this request has a transno less than the one we're waiting
2015 * for, we should process it now. It could (and currently always will)
2016 * be an open request for a descriptor that was opened some time ago.
2018 * Also, a resent, replayed request that has already been
2019 * handled will pass through here and be processed immediately.
2021 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2022 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2023 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2024 /* Processing the queue right now, don't re-add. */
2025 LASSERT(cfs_list_empty(&req->rq_list));
2026 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2029 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2031 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2034 target_request_copy_get(req);
2035 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2036 LASSERT(obd->obd_recovering);
2037 if (!req->rq_export->exp_in_recovery) {
2038 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2039 target_request_copy_put(req);
2042 LASSERT(req->rq_export->exp_req_replay_needed);
2044 if (target_exp_enqueue_req_replay(req)) {
2045 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2046 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2047 target_request_copy_put(req);
2052 cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2053 struct ptlrpc_request *reqiter =
2054 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2056 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2057 cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2062 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2064 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2065 "has been claimed by another client");
2066 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2067 target_exp_dequeue_req_replay(req);
2068 target_request_copy_put(req);
2074 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2076 obd->obd_requests_queued_for_recovery++;
2077 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2078 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2083 * Packs current SLV and Limit into \a req.
2085 int target_pack_pool_reply(struct ptlrpc_request *req)
2087 struct obd_device *obd;
2091 * Check that we still have all structures alive as this may
2092 * be some late rpc in shutdown time.
2094 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2095 !exp_connect_lru_resize(req->rq_export))) {
2096 lustre_msg_set_slv(req->rq_repmsg, 0);
2097 lustre_msg_set_limit(req->rq_repmsg, 0);
2102 * OBD is alive here as export is alive, which we checked above.
2104 obd = req->rq_export->exp_obd;
2106 cfs_read_lock(&obd->obd_pool_lock);
2107 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2108 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2109 cfs_read_unlock(&obd->obd_pool_lock);
2114 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2116 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2117 DEBUG_REQ(D_ERROR, req, "dropping reply");
2122 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2123 req->rq_status = rc;
2124 return (ptlrpc_send_error(req, 1));
2126 DEBUG_REQ(D_NET, req, "sending reply");
2129 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2132 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2135 struct ptlrpc_reply_state *rs;
2136 struct obd_device *obd;
2137 struct obd_export *exp;
2138 struct ptlrpc_service *svc;
2141 if (req->rq_no_reply) {
2146 svc = req->rq_rqbd->rqbd_service;
2147 rs = req->rq_reply_state;
2148 if (rs == NULL || !rs->rs_difficult) {
2150 target_send_reply_msg (req, rc, fail_id);
2155 /* must be an export if locks saved */
2156 LASSERT (req->rq_export != NULL);
2157 /* req/reply consistent */
2158 LASSERT (rs->rs_service == svc);
2161 LASSERT (!rs->rs_scheduled);
2162 LASSERT (!rs->rs_scheduled_ever);
2163 LASSERT (!rs->rs_handled);
2164 LASSERT (!rs->rs_on_net);
2165 LASSERT (rs->rs_export == NULL);
2166 LASSERT (cfs_list_empty(&rs->rs_obd_list));
2167 LASSERT (cfs_list_empty(&rs->rs_exp_list));
2169 exp = class_export_get (req->rq_export);
2172 /* disable reply scheduling while I'm setting up */
2173 rs->rs_scheduled = 1;
2175 rs->rs_xid = req->rq_xid;
2176 rs->rs_transno = req->rq_transno;
2177 rs->rs_export = exp;
2178 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2180 cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2181 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2182 rs->rs_transno, exp->exp_last_committed);
2183 if (rs->rs_transno > exp->exp_last_committed) {
2184 /* not committed already */
2185 cfs_list_add_tail(&rs->rs_obd_list,
2186 &exp->exp_uncommitted_replies);
2188 cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2190 cfs_spin_lock(&exp->exp_lock);
2191 cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2192 cfs_spin_unlock(&exp->exp_lock);
2194 netrc = target_send_reply_msg (req, rc, fail_id);
2196 cfs_spin_lock(&svc->srv_lock);
2198 cfs_atomic_inc(&svc->srv_n_difficult_replies);
2201 /* error sending: reply is off the net. Also we need +1
2202 * reply ref until ptlrpc_handle_rs() is done
2203 * with the reply state (if the send was successful, there
2204 * would have been +1 ref for the net, which
2205 * reply_out_callback leaves alone) */
2207 ptlrpc_rs_addref(rs);
2208 cfs_atomic_inc (&svc->srv_outstanding_replies);
2211 cfs_spin_lock(&rs->rs_lock);
2212 if (rs->rs_transno <= exp->exp_last_committed ||
2213 (!rs->rs_on_net && !rs->rs_no_ack) ||
2214 cfs_list_empty(&rs->rs_exp_list) || /* completed already */
2215 cfs_list_empty(&rs->rs_obd_list)) {
2216 CDEBUG(D_HA, "Schedule reply immediately\n");
2217 ptlrpc_dispatch_difficult_reply(rs);
2219 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2220 rs->rs_scheduled = 0; /* allow notifier to schedule */
2222 cfs_spin_unlock(&rs->rs_lock);
2223 cfs_spin_unlock(&svc->srv_lock);
2227 int target_handle_ping(struct ptlrpc_request *req)
2229 obd_ping(req->rq_export);
2230 return req_capsule_server_pack(&req->rq_pill);
2233 void target_committed_to_req(struct ptlrpc_request *req)
2235 struct obd_export *exp = req->rq_export;
2237 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2238 lustre_msg_set_last_committed(req->rq_repmsg,
2239 exp->exp_last_committed);
2241 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2242 "%d)", exp->exp_obd->obd_no_transno,
2243 req->rq_repmsg == NULL);
2245 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2246 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2248 EXPORT_SYMBOL(target_committed_to_req);
2250 int target_handle_qc_callback(struct ptlrpc_request *req)
2252 struct obd_quotactl *oqctl;
2253 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2255 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2256 if (oqctl == NULL) {
2257 CERROR("Can't unpack obd_quotactl\n");
2261 cli->cl_qchk_stat = oqctl->qc_stat;
2266 #ifdef HAVE_QUOTA_SUPPORT
2267 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2270 struct obd_device *obd = req->rq_export->exp_obd;
2271 struct obd_device *master_obd = NULL, *lov_obd = NULL;
2272 struct obd_device_target *obt;
2273 struct lustre_quota_ctxt *qctxt;
2274 struct qunit_data *qdata = NULL;
2278 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2281 rc = req_capsule_server_pack(&req->rq_pill);
2283 CERROR("packing reply failed!: rc = %d\n", rc);
2287 LASSERT(req->rq_export);
2289 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2290 if (IS_ERR(qdata)) {
2291 rc = PTR_ERR(qdata);
2292 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2293 req->rq_status = rc;
2297 /* we use the observer */
2298 if (obd_pin_observer(obd, &lov_obd) ||
2299 obd_pin_observer(lov_obd, &master_obd)) {
2300 CERROR("Can't find the observer, it is recovering\n");
2301 req->rq_status = -EAGAIN;
2305 obt = &master_obd->u.obt;
2306 qctxt = &obt->obt_qctxt;
2308 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2309 /* quota_type has not been processed yet, return EAGAIN
2310 * until we know whether or not quotas are supposed to
2312 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2314 req->rq_status = -EAGAIN;
2318 cfs_down_read(&obt->obt_rwsem);
2319 if (qctxt->lqc_lqs_hash == NULL) {
2320 cfs_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;
2330 LASSERT(qctxt->lqc_handler);
2331 rc = qctxt->lqc_handler(master_obd, qdata,
2332 lustre_msg_get_opc(req->rq_reqmsg));
2333 cfs_up_read(&obt->obt_rwsem);
2334 if (rc && rc != -EDQUOT)
2335 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2336 "dqacq/dqrel failed! (rc:%d)\n", rc);
2337 req->rq_status = rc;
2339 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2341 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2345 /* Block the quota req. b=14840 */
2346 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2351 obd_unpin_observer(lov_obd);
2353 obd_unpin_observer(obd);
2355 rc = ptlrpc_reply(req);
2359 #endif /* !__KERNEL__ */
2361 #endif /* HAVE_QUOTA_SUPPORT */
2363 /* Send a remote set_info_async.
2364 * This may go from client to server or server to client
2366 int target_set_info_rpc(struct obd_import *imp, int opcode,
2367 obd_count keylen, void *key,
2368 obd_count vallen, void *val,
2369 struct ptlrpc_request_set *set)
2371 struct ptlrpc_request *req;
2376 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
2380 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2381 RCL_CLIENT, keylen);
2382 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2383 RCL_CLIENT, vallen);
2384 rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, opcode);
2386 ptlrpc_request_free(req);
2390 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2391 memcpy(tmp, key, keylen);
2392 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2393 memcpy(tmp, val, vallen);
2395 ptlrpc_request_set_replen(req);
2398 ptlrpc_set_add_req(set, req);
2399 ptlrpc_check_set(NULL, set);
2401 rc = ptlrpc_queue_wait(req);
2402 ptlrpc_req_finished(req);
2407 EXPORT_SYMBOL(target_set_info_rpc);
2410 ldlm_mode_t lck_compat_array[] = {
2411 [LCK_EX] LCK_COMPAT_EX,
2412 [LCK_PW] LCK_COMPAT_PW,
2413 [LCK_PR] LCK_COMPAT_PR,
2414 [LCK_CW] LCK_COMPAT_CW,
2415 [LCK_CR] LCK_COMPAT_CR,
2416 [LCK_NL] LCK_COMPAT_NL,
2417 [LCK_GROUP] LCK_COMPAT_GROUP,
2418 [LCK_COS] LCK_COMPAT_COS,
2422 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2423 * not escape to the user level.
2425 int ldlm_error2errno(ldlm_error_t error)
2433 case ELDLM_LOCK_CHANGED:
2436 case ELDLM_LOCK_ABORTED:
2439 case ELDLM_LOCK_REPLACED:
2442 case ELDLM_NO_LOCK_DATA:
2445 case ELDLM_NAMESPACE_EXISTS:
2448 case ELDLM_BAD_NAMESPACE:
2452 if (((int)error) < 0) /* cast to signed type */
2453 result = error; /* as ldlm_error_t can be unsigned */
2455 CERROR("Invalid DLM result code: %i\n", error);
2461 EXPORT_SYMBOL(ldlm_error2errno);
2464 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2466 ldlm_error_t ldlm_errno2error(int err_no)
2475 error = ELDLM_LOCK_CHANGED;
2478 error = ELDLM_LOCK_ABORTED;
2481 error = ELDLM_LOCK_REPLACED;
2484 error = ELDLM_NO_LOCK_DATA;
2487 error = ELDLM_NAMESPACE_EXISTS;
2490 error = ELDLM_BAD_NAMESPACE;
2497 EXPORT_SYMBOL(ldlm_errno2error);
2499 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2500 void ldlm_dump_export_locks(struct obd_export *exp)
2502 cfs_spin_lock(&exp->exp_locks_list_guard);
2503 if (!cfs_list_empty(&exp->exp_locks_list)) {
2504 struct ldlm_lock *lock;
2506 CERROR("dumping locks for export %p,"
2507 "ignore if the unmount doesn't hang\n", exp);
2508 cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2509 ldlm_lock_dump(D_ERROR, lock, 0);
2511 cfs_spin_unlock(&exp->exp_locks_list_guard);