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);
306 atomic_set(&cli->cl_quota_resends, CLIENT_QUOTA_DEFAULT_RESENDS);
308 /* This value may be changed at connect time in
309 ptlrpc_connect_interpret. */
310 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
311 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
313 if (!strcmp(name, LUSTRE_MDC_NAME)) {
314 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
315 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
316 cli->cl_max_rpcs_in_flight = 2;
317 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
318 cli->cl_max_rpcs_in_flight = 3;
319 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
320 cli->cl_max_rpcs_in_flight = 4;
322 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
327 CERROR("ldlm_get_ref failed: %d\n", rc);
331 ptlrpc_init_client(rq_portal, rp_portal, name,
332 &obddev->obd_ldlm_client);
334 imp = class_new_import(obddev);
336 GOTO(err_ldlm, rc = -ENOENT);
337 imp->imp_client = &obddev->obd_ldlm_client;
338 imp->imp_connect_op = connect_op;
339 imp->imp_initial_recov = 1;
340 imp->imp_initial_recov_bk = 0;
341 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
342 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
343 LUSTRE_CFG_BUFLEN(lcfg, 1));
344 class_import_put(imp);
346 rc = client_import_add_conn(imp, &server_uuid, 1);
348 CERROR("can't add initial connection\n");
349 GOTO(err_import, rc);
352 cli->cl_import = imp;
353 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
354 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
355 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
357 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
358 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
359 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
360 name, obddev->obd_name,
361 cli->cl_target_uuid.uuid);
362 spin_lock(&imp->imp_lock);
363 imp->imp_deactive = 1;
364 spin_unlock(&imp->imp_lock);
368 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
369 LDLM_NAMESPACE_CLIENT,
370 LDLM_NAMESPACE_GREEDY);
371 if (obddev->obd_namespace == NULL) {
372 CERROR("Unable to create client namespace - %s\n",
374 GOTO(err_import, rc = -ENOMEM);
377 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
382 class_destroy_import(imp);
390 int client_obd_cleanup(struct obd_device *obddev)
394 ldlm_namespace_free_post(obddev->obd_namespace);
395 obddev->obd_namespace = NULL;
401 /* ->o_connect() method for client side (OSC and MDC and MGC) */
402 int client_connect_import(const struct lu_env *env,
403 struct obd_export **exp,
404 struct obd_device *obd, struct obd_uuid *cluuid,
405 struct obd_connect_data *data, void *localdata)
407 struct client_obd *cli = &obd->u.cli;
408 struct obd_import *imp = cli->cl_import;
409 struct obd_connect_data *ocd;
410 struct lustre_handle conn = { 0 };
415 down_write(&cli->cl_sem);
416 if (cli->cl_conn_count > 0 )
417 GOTO(out_sem, rc = -EALREADY);
419 rc = class_connect(&conn, obd, cluuid);
423 cli->cl_conn_count++;
424 *exp = class_conn2export(&conn);
426 LASSERT(obd->obd_namespace);
428 imp->imp_dlm_handle = conn;
429 rc = ptlrpc_init_import(imp);
433 ocd = &imp->imp_connect_data;
436 imp->imp_connect_flags_orig = data->ocd_connect_flags;
439 rc = ptlrpc_connect_import(imp, NULL);
441 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
444 LASSERT((*exp)->exp_connection);
447 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
448 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
449 data->ocd_connect_flags, ocd->ocd_connect_flags);
450 data->ocd_connect_flags = ocd->ocd_connect_flags;
453 ptlrpc_pinger_add_import(imp);
459 cli->cl_conn_count--;
460 class_disconnect(*exp);
464 up_write(&cli->cl_sem);
469 int client_disconnect_export(struct obd_export *exp)
471 struct obd_device *obd = class_exp2obd(exp);
472 struct client_obd *cli;
473 struct obd_import *imp;
478 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
479 exp, exp ? exp->exp_handle.h_cookie : -1);
484 imp = cli->cl_import;
486 down_write(&cli->cl_sem);
487 CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
490 if (!cli->cl_conn_count) {
491 CERROR("disconnecting disconnected device (%s)\n",
493 GOTO(out_disconnect, rc = -EINVAL);
496 cli->cl_conn_count--;
497 if (cli->cl_conn_count)
498 GOTO(out_disconnect, rc = 0);
500 /* Mark import deactivated now, so we don't try to reconnect if any
501 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
502 * fully deactivate the import, or that would drop all requests. */
503 spin_lock(&imp->imp_lock);
504 imp->imp_deactive = 1;
505 spin_unlock(&imp->imp_lock);
507 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
508 * delete it regardless. (It's safe to delete an import that was
510 (void)ptlrpc_pinger_del_import(imp);
512 if (obd->obd_namespace != NULL) {
513 /* obd_force == local only */
514 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
515 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
517 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
521 * there's no necessary to hold sem during diconnecting an import,
522 * and actually it may cause deadlock in gss.
524 up_write(&cli->cl_sem);
525 rc = ptlrpc_disconnect_import(imp, 0);
526 down_write(&cli->cl_sem);
528 ptlrpc_invalidate_import(imp);
530 if (imp->imp_rq_pool) {
531 ptlrpc_free_rq_pool(imp->imp_rq_pool);
532 imp->imp_rq_pool = NULL;
534 client_destroy_import(imp);
535 cli->cl_import = NULL;
540 /* use server style - class_disconnect should be always called for
542 err = class_disconnect(exp);
546 up_write(&cli->cl_sem);
551 /* --------------------------------------------------------------------------
552 * from old lib/target.c
553 * -------------------------------------------------------------------------- */
555 static int target_handle_reconnect(struct lustre_handle *conn,
556 struct obd_export *exp,
557 struct obd_uuid *cluuid)
560 if (exp->exp_connection && exp->exp_imp_reverse) {
561 struct lustre_handle *hdl;
562 hdl = &exp->exp_imp_reverse->imp_remote_handle;
563 /* Might be a re-connect after a partition. */
564 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
565 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
567 conn->cookie = exp->exp_handle.h_cookie;
568 /* target_handle_connect() treats EALREADY and
569 * -EALREADY differently. EALREADY means we are
570 * doing a valid reconnect from the same client. */
573 CERROR("%s reconnecting from %s, "
574 "handle mismatch (ours "LPX64", theirs "
575 LPX64")\n", cluuid->uuid,
576 exp->exp_connection->c_remote_uuid.uuid,
577 hdl->cookie, conn->cookie);
578 memset(conn, 0, sizeof *conn);
579 /* target_handle_connect() treats EALREADY and
580 * -EALREADY differently. -EALREADY is an error
581 * (same UUID, different handle). */
586 conn->cookie = exp->exp_handle.h_cookie;
587 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
588 cluuid->uuid, exp, conn->cookie);
592 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
595 struct obd_export *exp = cb_data;
597 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
598 obd->obd_name, exp->exp_client_uuid.uuid);
600 spin_lock(&exp->exp_lock);
601 exp->exp_need_sync = 0;
602 spin_unlock(&exp->exp_lock);
604 EXPORT_SYMBOL(target_client_add_cb);
606 target_start_and_reset_recovery_timer(struct obd_device *obd,
607 struct ptlrpc_request *req,
610 int target_handle_connect(struct ptlrpc_request *req)
612 struct obd_device *target, *targref = NULL;
613 struct obd_export *export = NULL;
614 struct obd_import *revimp;
615 struct lustre_handle conn;
616 struct lustre_handle *tmp;
617 struct obd_uuid tgtuuid;
618 struct obd_uuid cluuid;
619 struct obd_uuid remote_uuid;
623 struct obd_connect_data *data, *tmpdata;
624 lnet_nid_t *client_nid = NULL;
627 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
629 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
631 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
632 GOTO(out, rc = -EINVAL);
635 obd_str2uuid(&tgtuuid, str);
636 target = class_uuid2obd(&tgtuuid);
638 target = class_name2obd(str);
640 if (!target || target->obd_stopping || !target->obd_set_up) {
641 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
642 " for connect (%s)\n", str,
643 !target ? "no target" :
644 (target->obd_stopping ? "stopping" :
646 GOTO(out, rc = -ENODEV);
649 if (target->obd_no_conn) {
650 LCONSOLE_WARN("%s: temporarily refusing client connection "
651 "from %s\n", target->obd_name,
652 libcfs_nid2str(req->rq_peer.nid));
653 GOTO(out, rc = -EAGAIN);
656 /* Make sure the target isn't cleaned up while we're here. Yes,
657 there's still a race between the above check and our incref here.
658 Really, class_uuid2obd should take the ref. */
659 targref = class_incref(target, __FUNCTION__, cfs_current());
662 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
664 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
665 GOTO(out, rc = -EINVAL);
668 obd_str2uuid(&cluuid, str);
670 /* XXX extract a nettype and format accordingly */
671 switch (sizeof(lnet_nid_t)) {
672 /* NB the casts only avoid compiler warnings */
674 snprintf(remote_uuid.uuid, sizeof remote_uuid,
675 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
678 snprintf(remote_uuid.uuid, sizeof remote_uuid,
679 "NET_%x_UUID", (__u32)req->rq_peer.nid);
685 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
687 GOTO(out, rc = -EPROTO);
691 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
693 GOTO(out, rc = -EPROTO);
695 rc = req_capsule_server_pack(&req->rq_pill);
699 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
701 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
702 "libclient connection attempt");
703 GOTO(out, rc = -EPROTO);
704 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
705 LUSTRE_VERSION_ALLOWED_OFFSET ||
706 data->ocd_version > LUSTRE_VERSION_CODE +
707 LUSTRE_VERSION_ALLOWED_OFFSET) {
708 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
709 "libclient connection attempt",
710 data->ocd_version < LUSTRE_VERSION_CODE ?
712 OBD_OCD_VERSION_MAJOR(data->ocd_version),
713 OBD_OCD_VERSION_MINOR(data->ocd_version),
714 OBD_OCD_VERSION_PATCH(data->ocd_version),
715 OBD_OCD_VERSION_FIX(data->ocd_version));
716 data = req_capsule_server_sized_get(&req->rq_pill,
718 offsetof(typeof(*data), ocd_version) +
719 sizeof(data->ocd_version));
721 data->ocd_connect_flags = OBD_CONNECT_VERSION;
722 data->ocd_version = LUSTRE_VERSION_CODE;
724 GOTO(out, rc = -EPROTO);
728 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
729 (data->ocd_connect_flags & OBD_CONNECT_MDS))
732 /* lctl gets a backstage, all-access pass. */
733 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
734 goto dont_check_exports;
736 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
740 /* we've found an export in the hash */
741 if (export->exp_connecting) { /* bug 9635, et. al. */
742 CWARN("%s: exp %p already connecting\n",
743 export->exp_obd->obd_name, export);
744 class_export_put(export);
747 } else if (mds_conn && export->exp_connection) {
748 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
749 /* mds reconnected after failover */
750 CWARN("%s: received MDS connection from NID %s,"
751 " removing former export from NID %s\n",
752 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
753 libcfs_nid2str(export->exp_connection->c_peer.nid));
755 /* new mds connection from the same nid */
756 CWARN("%s: received new MDS connection from NID %s,"
757 " removing former export from same NID\n",
758 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
759 class_fail_export(export);
760 class_export_put(export);
763 } else if (export->exp_connection != NULL &&
764 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
765 (lustre_msg_get_op_flags(req->rq_reqmsg) &
766 MSG_CONNECT_INITIAL)) {
767 /* in mds failover we have static uuid but nid can be
769 CWARN("%s: cookie %s seen on new NID %s when "
770 "existing NID %s is already connected\n",
771 target->obd_name, cluuid.uuid,
772 libcfs_nid2str(req->rq_peer.nid),
773 libcfs_nid2str(export->exp_connection->c_peer.nid));
775 class_export_put(export);
778 spin_lock(&export->exp_lock);
779 export->exp_connecting = 1;
780 spin_unlock(&export->exp_lock);
781 class_export_put(export);
782 LASSERT(export->exp_obd == target);
784 rc = target_handle_reconnect(&conn, export, &cluuid);
787 /* If we found an export, we already unlocked. */
790 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
791 } else if (req->rq_export == NULL &&
792 atomic_read(&export->exp_rpc_count) > 0) {
793 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
794 target->obd_name, cluuid.uuid,
795 libcfs_nid2str(req->rq_peer.nid),
796 export, atomic_read(&export->exp_refcount));
797 GOTO(out, rc = -EBUSY);
798 } else if (req->rq_export != NULL &&
799 (atomic_read(&export->exp_rpc_count) > 1)) {
800 /* the current connect rpc has increased exp_rpc_count */
801 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
802 target->obd_name, cluuid.uuid,
803 libcfs_nid2str(req->rq_peer.nid),
804 export, atomic_read(&export->exp_rpc_count) - 1);
805 spin_lock(&export->exp_lock);
806 if (req->rq_export->exp_conn_cnt <
807 lustre_msg_get_conn_cnt(req->rq_reqmsg))
808 /* try to abort active requests */
809 req->rq_export->exp_abort_active_req = 1;
810 spin_unlock(&export->exp_lock);
811 GOTO(out, rc = -EBUSY);
812 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
813 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
814 "cookies not random?\n", target->obd_name,
815 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
816 GOTO(out, rc = -EALREADY);
818 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
825 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
826 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
827 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
828 export, (long)cfs_time_current_sec(),
829 export ? (long)export->exp_last_request_time : 0);
831 /* Tell the client if we're in recovery. */
832 if (target->obd_recovering) {
833 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
834 /* If this is the first time a client connects,
835 reset the recovery timer */
837 target_start_and_reset_recovery_timer(target, req,
841 /* We want to handle EALREADY but *not* -EALREADY from
842 * target_handle_reconnect(), return reconnection state in a flag */
843 if (rc == EALREADY) {
844 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
850 /* Tell the client if we support replayable requests */
851 if (target->obd_replayable)
852 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
853 client_nid = &req->rq_peer.nid;
855 if (export == NULL) {
856 if (target->obd_recovering) {
859 t = cfs_timer_deadline(&target->obd_recovery_timer);
860 t = cfs_time_sub(t, cfs_time_current());
861 CERROR("%s: denying connection for new client %s (%s): "
862 "%d clients in recovery for "CFS_TIME_T"s\n",
864 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
865 atomic_read(&target->obd_lock_replay_clients),
866 cfs_duration_sec(t));
870 rc = obd_connect(req->rq_svc_thread->t_env,
871 &export, target, &cluuid, data,
874 conn.cookie = export->exp_handle.h_cookie;
877 rc = obd_reconnect(req->rq_svc_thread->t_env,
878 export, target, &cluuid, data, client_nid);
880 /* prevous done via class_conn2export */
881 class_export_get(export);
885 /* Return only the parts of obd_connect_data that we understand, so the
886 * client knows that we don't understand the rest. */
888 tmpdata = req_capsule_server_get(&req->rq_pill,
890 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
894 /* If all else goes well, this is our RPC return code. */
897 lustre_msg_set_handle(req->rq_repmsg, &conn);
899 /* If the client and the server are the same node, we will already
900 * have an export that really points to the client's DLM export,
901 * because we have a shared handles table.
903 * XXX this will go away when shaver stops sending the "connect" handle
904 * in the real "remote handle" field of the request --phik 24 Apr 2003
906 if (req->rq_export != NULL)
907 class_export_put(req->rq_export);
909 req->rq_export = export;
911 spin_lock(&export->exp_lock);
912 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
913 spin_unlock(&export->exp_lock);
914 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
915 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
916 export->exp_conn_cnt,
917 lustre_msg_get_conn_cnt(req->rq_reqmsg));
919 GOTO(out, rc = -EALREADY);
921 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
922 export->exp_abort_active_req = 0;
924 /* request from liblustre? Don't evict it for not pinging. */
925 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
926 export->exp_libclient = 1;
927 spin_unlock(&export->exp_lock);
929 spin_lock(&target->obd_dev_lock);
930 list_del_init(&export->exp_obd_chain_timed);
931 spin_unlock(&target->obd_dev_lock);
933 spin_unlock(&export->exp_lock);
936 if (export->exp_connection != NULL) {
937 /* Check to see if connection came from another NID */
938 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
939 !hlist_unhashed(&export->exp_nid_hash))
940 lustre_hash_del(export->exp_obd->obd_nid_hash,
941 &export->exp_connection->c_peer.nid,
942 &export->exp_nid_hash);
944 ptlrpc_connection_put(export->exp_connection);
947 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
950 if (hlist_unhashed(&export->exp_nid_hash)) {
951 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
952 &export->exp_connection->c_peer.nid,
953 &export->exp_nid_hash);
956 spin_lock_bh(&target->obd_processing_task_lock);
957 if (target->obd_recovering && !export->exp_in_recovery) {
958 spin_lock(&export->exp_lock);
959 export->exp_in_recovery = 1;
960 export->exp_req_replay_needed = 1;
961 export->exp_lock_replay_needed = 1;
962 spin_unlock(&export->exp_lock);
963 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
964 && (data->ocd_transno == 0))
965 CWARN("Connect with zero transno!\n");
967 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
968 && data->ocd_transno < target->obd_next_recovery_transno)
969 target->obd_next_recovery_transno = data->ocd_transno;
970 target->obd_connected_clients++;
971 atomic_inc(&target->obd_req_replay_clients);
972 atomic_inc(&target->obd_lock_replay_clients);
973 if (target->obd_connected_clients ==
974 target->obd_max_recoverable_clients)
975 cfs_waitq_signal(&target->obd_next_transno_waitq);
977 spin_unlock_bh(&target->obd_processing_task_lock);
978 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
981 if (export->exp_imp_reverse != NULL) {
982 /* destroyed import can be still referenced in ctxt */
983 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
984 KEY_REVIMP_UPD, 0, NULL, NULL);
986 client_destroy_import(export->exp_imp_reverse);
989 /* for the rest part, we return -ENOTCONN in case of errors
990 * in order to let client initialize connection again.
992 revimp = export->exp_imp_reverse = class_new_import(target);
994 CERROR("fail to alloc new reverse import.\n");
995 GOTO(out, rc = -ENOTCONN);
998 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
999 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1000 revimp->imp_remote_handle = conn;
1001 revimp->imp_dlm_fake = 1;
1002 revimp->imp_state = LUSTRE_IMP_FULL;
1004 /* unknown versions will be caught in
1005 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1006 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1008 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1009 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1010 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1012 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1014 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1016 CERROR("Failed to get sec for reverse import: %d\n", rc);
1017 export->exp_imp_reverse = NULL;
1018 class_destroy_import(revimp);
1021 class_import_put(revimp);
1024 spin_lock(&export->exp_lock);
1025 export->exp_connecting = 0;
1026 spin_unlock(&export->exp_lock);
1029 class_decref(targref, __FUNCTION__, cfs_current());
1031 req->rq_status = rc;
1035 int target_handle_disconnect(struct ptlrpc_request *req)
1040 rc = req_capsule_server_pack(&req->rq_pill);
1044 /* keep the rq_export around so we can send the reply */
1045 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1050 void target_destroy_export(struct obd_export *exp)
1052 /* exports created from last_rcvd data, and "fake"
1053 exports created by lctl don't have an import */
1054 if (exp->exp_imp_reverse != NULL)
1055 client_destroy_import(exp->exp_imp_reverse);
1057 LASSERT(atomic_read(&exp->exp_locks_count) == 0);
1058 LASSERT(atomic_read(&exp->exp_rpc_count) == 0);
1059 LASSERT(atomic_read(&exp->exp_cb_count) == 0);
1060 LASSERT(atomic_read(&exp->exp_replay_count) == 0);
1064 * Recovery functions
1067 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1069 struct ptlrpc_request *copy_req;
1070 struct lustre_msg *copy_reqmsg;
1071 struct ptlrpc_user_desc *udesc = NULL;
1073 OBD_ALLOC_PTR(copy_req);
1076 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1078 OBD_FREE_PTR(copy_req);
1082 if (orig_req->rq_user_desc) {
1083 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1085 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1087 OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1088 OBD_FREE_PTR(copy_req);
1091 memcpy(udesc, orig_req->rq_user_desc,
1092 sptlrpc_user_desc_size(ngroups));
1095 *copy_req = *orig_req;
1096 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1097 copy_req->rq_reqmsg = copy_reqmsg;
1098 copy_req->rq_user_desc = udesc;
1100 class_export_rpc_get(copy_req->rq_export);
1101 CFS_INIT_LIST_HEAD(©_req->rq_list);
1102 CFS_INIT_LIST_HEAD(©_req->rq_replay_list);
1103 sptlrpc_svc_ctx_addref(copy_req);
1105 if (copy_req->rq_reply_state) {
1106 /* the copied req takes over the reply state */
1107 orig_req->rq_reply_state = NULL;
1108 /* to catch further access */
1109 orig_req->rq_repmsg = NULL;
1110 orig_req->rq_replen = 0;
1112 /** let export know it has replays to be handled */
1113 atomic_inc(©_req->rq_export->exp_replay_count);
1117 void ptlrpc_free_clone(struct ptlrpc_request *req)
1119 LASSERT(list_empty(&req->rq_replay_list));
1121 ptlrpc_req_drop_rs(req);
1122 sptlrpc_svc_ctx_decref(req);
1123 LASSERT(atomic_read(&req->rq_export->exp_replay_count) > 0);
1124 atomic_dec(&req->rq_export->exp_replay_count);
1125 class_export_rpc_put(req->rq_export);
1126 list_del_init(&req->rq_list);
1128 if (req->rq_user_desc) {
1129 int ngroups = req->rq_user_desc->pud_ngroups;
1130 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1132 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
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 spin_lock(&exp->exp_lock);
1146 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 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1164 spin_unlock(&exp->exp_lock);
1168 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1170 LASSERT(!list_empty(&req->rq_replay_list));
1171 LASSERT(req->rq_export);
1173 spin_lock(&req->rq_export->exp_lock);
1174 list_del_init(&req->rq_replay_list);
1175 spin_unlock(&req->rq_export->exp_lock);
1179 static void target_finish_recovery(struct obd_device *obd)
1182 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1185 ldlm_reprocess_all_ns(obd->obd_namespace);
1186 spin_lock_bh(&obd->obd_processing_task_lock);
1187 if (!list_empty(&obd->obd_req_replay_queue) ||
1188 !list_empty(&obd->obd_lock_replay_queue) ||
1189 !list_empty(&obd->obd_final_req_queue)) {
1190 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1192 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1193 list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1194 list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1195 spin_unlock_bh(&obd->obd_processing_task_lock);
1198 spin_unlock_bh(&obd->obd_processing_task_lock);
1200 obd->obd_recovery_end = cfs_time_current_sec();
1202 /* when recovery finished, cleanup orphans on mds and ost */
1203 if (OBT(obd) && OBP(obd, postrecov)) {
1204 int rc = OBP(obd, postrecov)(obd);
1206 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1212 static void abort_req_replay_queue(struct obd_device *obd)
1214 struct ptlrpc_request *req, *n;
1215 struct list_head abort_list;
1217 CFS_INIT_LIST_HEAD(&abort_list);
1218 spin_lock_bh(&obd->obd_processing_task_lock);
1219 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1220 spin_unlock_bh(&obd->obd_processing_task_lock);
1221 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1222 DEBUG_REQ(D_WARNING, req, "aborted:");
1223 req->rq_status = -ENOTCONN;
1224 if (ptlrpc_error(req)) {
1225 DEBUG_REQ(D_ERROR, req,
1226 "failed abort_req_reply; skipping");
1228 target_exp_dequeue_req_replay(req);
1229 ptlrpc_free_clone(req);
1233 static void abort_lock_replay_queue(struct obd_device *obd)
1235 struct ptlrpc_request *req, *n;
1236 struct list_head abort_list;
1238 CFS_INIT_LIST_HEAD(&abort_list);
1239 spin_lock_bh(&obd->obd_processing_task_lock);
1240 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1241 spin_unlock_bh(&obd->obd_processing_task_lock);
1242 list_for_each_entry_safe(req, n, &abort_list, rq_list){
1243 DEBUG_REQ(D_ERROR, req, "aborted:");
1244 req->rq_status = -ENOTCONN;
1245 if (ptlrpc_error(req)) {
1246 DEBUG_REQ(D_ERROR, req,
1247 "failed abort_lock_reply; skipping");
1249 ptlrpc_free_clone(req);
1254 /* Called from a cleanup function if the device is being cleaned up
1255 forcefully. The exports should all have been disconnected already,
1256 the only thing left to do is
1257 - clear the recovery flags
1259 - free queued requests and replies, but don't send replies
1260 Because the obd_stopping flag is set, no new requests should be received.
1263 void target_cleanup_recovery(struct obd_device *obd)
1265 struct ptlrpc_request *req, *n;
1266 struct list_head clean_list;
1269 CFS_INIT_LIST_HEAD(&clean_list);
1270 spin_lock_bh(&obd->obd_processing_task_lock);
1271 if (!obd->obd_recovering) {
1272 spin_unlock_bh(&obd->obd_processing_task_lock);
1276 obd->obd_recovering = obd->obd_abort_recovery = 0;
1277 target_cancel_recovery_timer(obd);
1279 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1280 spin_unlock_bh(&obd->obd_processing_task_lock);
1282 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1283 LASSERT(req->rq_reply_state == 0);
1284 target_exp_dequeue_req_replay(req);
1285 ptlrpc_free_clone(req);
1288 spin_lock_bh(&obd->obd_processing_task_lock);
1289 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1290 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1291 spin_unlock_bh(&obd->obd_processing_task_lock);
1293 list_for_each_entry_safe(req, n, &clean_list, rq_list){
1294 LASSERT(req->rq_reply_state == 0);
1295 ptlrpc_free_clone(req);
1301 /* obd_processing_task_lock should be held */
1302 void target_cancel_recovery_timer(struct obd_device *obd)
1304 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1305 cfs_timer_disarm(&obd->obd_recovery_timer);
1308 /* extend = 1 means require at least "duration" seconds left in the timer,
1309 extend = 0 means set the total duration (start_recovery_timer) */
1310 static void reset_recovery_timer(struct obd_device *obd, int duration,
1313 cfs_time_t now = cfs_time_current_sec();
1314 cfs_duration_t left;
1316 spin_lock_bh(&obd->obd_processing_task_lock);
1317 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1318 spin_unlock_bh(&obd->obd_processing_task_lock);
1322 left = cfs_time_sub(obd->obd_recovery_end, now);
1324 if (extend && (duration > left))
1325 obd->obd_recovery_timeout += duration - left;
1326 else if (!extend && (duration > obd->obd_recovery_timeout))
1327 /* Track the client's largest expected replay time */
1328 obd->obd_recovery_timeout = duration;
1331 * If total recovery time already exceed the
1332 * obd_recovery_max_time, then CRAY XT3 will
1333 * abort the recovery
1335 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1336 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1338 obd->obd_recovery_end = obd->obd_recovery_start +
1339 obd->obd_recovery_timeout;
1340 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1341 cfs_time_before(now, obd->obd_recovery_end)) {
1342 left = cfs_time_sub(obd->obd_recovery_end, now);
1343 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1345 spin_unlock_bh(&obd->obd_processing_task_lock);
1346 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1347 obd->obd_name, (unsigned)left);
1350 static void check_and_start_recovery_timer(struct obd_device *obd)
1352 spin_lock_bh(&obd->obd_processing_task_lock);
1353 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1354 spin_unlock_bh(&obd->obd_processing_task_lock);
1357 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1358 obd->obd_recovery_start = cfs_time_current_sec();
1360 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1361 spin_unlock_bh(&obd->obd_processing_task_lock);
1363 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1366 /* Reset the timer with each new client connection */
1368 * This timer is actually reconnect_timer, which is for making sure
1369 * the total recovery window is at least as big as my reconnect
1370 * attempt timing. So the initial recovery time_out will be set to
1371 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1372 * from client is bigger than this, then the recovery time_out will
1373 * be extend to make sure the client could be reconnected, in the
1374 * process, the timeout from the new client should be ignored.
1378 target_start_and_reset_recovery_timer(struct obd_device *obd,
1379 struct ptlrpc_request *req,
1382 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1384 if (!new_client && service_time)
1385 /* Teach server about old server's estimates, as first guess
1386 * at how long new requests will take. */
1387 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1390 check_and_start_recovery_timer(obd);
1392 /* convert the service time to rpc timeout,
1393 * reuse service_time to limit stack usage */
1394 service_time = at_est2timeout(service_time);
1396 /* We expect other clients to timeout within service_time, then try
1397 * to reconnect, then try the failover server. The max delay between
1398 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1399 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1400 INITIAL_CONNECT_TIMEOUT);
1401 if (service_time > obd->obd_recovery_timeout && !new_client)
1402 reset_recovery_timer(obd, service_time, 0);
1407 /** Health checking routines */
1408 static inline int exp_connect_healthy(struct obd_export *exp)
1410 return (exp->exp_in_recovery);
1413 /** if export done req_replay or has replay in queue */
1414 static inline int exp_req_replay_healthy(struct obd_export *exp)
1416 return (!exp->exp_req_replay_needed ||
1417 atomic_read(&exp->exp_replay_count) > 0);
1419 /** if export done lock_replay or has replay in queue */
1420 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1422 return (!exp->exp_lock_replay_needed ||
1423 atomic_read(&exp->exp_replay_count) > 0);
1426 static inline int exp_vbr_healthy(struct obd_export *exp)
1428 return (!exp->exp_vbr_failed);
1431 static inline int exp_finished(struct obd_export *exp)
1433 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1436 /** Checking routines for recovery */
1437 static int check_for_clients(struct obd_device *obd)
1439 if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1441 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1442 if (obd->obd_no_conn == 0 &&
1443 obd->obd_connected_clients + obd->obd_stale_clients ==
1444 obd->obd_max_recoverable_clients)
1449 static int check_for_next_transno(struct obd_device *obd)
1451 struct ptlrpc_request *req = NULL;
1452 int wake_up = 0, connected, completed, queue_len;
1453 __u64 next_transno, req_transno;
1455 spin_lock_bh(&obd->obd_processing_task_lock);
1457 if (!list_empty(&obd->obd_req_replay_queue)) {
1458 req = list_entry(obd->obd_req_replay_queue.next,
1459 struct ptlrpc_request, rq_list);
1460 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1465 connected = obd->obd_connected_clients;
1466 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1467 queue_len = obd->obd_requests_queued_for_recovery;
1468 next_transno = obd->obd_next_recovery_transno;
1470 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1471 "req_transno: "LPU64", next_transno: "LPU64"\n",
1472 obd->obd_max_recoverable_clients, connected, completed,
1473 queue_len, req_transno, next_transno);
1475 if (obd->obd_abort_recovery) {
1476 CDEBUG(D_HA, "waking for aborted recovery\n");
1478 } else if (obd->obd_recovery_expired) {
1479 CDEBUG(D_HA, "waking for expired recovery\n");
1481 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1482 CDEBUG(D_HA, "waking for completed recovery\n");
1484 } else if (req_transno == next_transno) {
1485 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1487 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1489 /** handle gaps occured due to lost reply or VBR */
1490 LASSERTF(req_transno >= next_transno,
1491 "req_transno: "LPU64", next_transno: "LPU64"\n",
1492 req_transno, next_transno);
1493 if (req_transno > obd->obd_last_committed &&
1494 !obd->obd_version_recov)
1497 "%s: waking for gap in transno, VBR is %s (skip: "
1498 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1499 ", last_committed: "LPD64")\n",
1500 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1501 next_transno, queue_len, completed, connected,
1502 req_transno, obd->obd_last_committed);
1503 obd->obd_next_recovery_transno = req_transno;
1505 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1506 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1507 " by fail_lock, waking up ("LPD64")\n", next_transno);
1508 obd->obd_next_recovery_transno = req_transno;
1511 spin_unlock_bh(&obd->obd_processing_task_lock);
1515 static int check_for_next_lock(struct obd_device *obd)
1519 spin_lock_bh(&obd->obd_processing_task_lock);
1520 if (!list_empty(&obd->obd_lock_replay_queue)) {
1521 CDEBUG(D_HA, "waking for next lock\n");
1523 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1524 CDEBUG(D_HA, "waking for completed lock replay\n");
1526 } else if (obd->obd_abort_recovery) {
1527 CDEBUG(D_HA, "waking for aborted recovery\n");
1529 } else if (obd->obd_recovery_expired) {
1530 CDEBUG(D_HA, "waking for expired recovery\n");
1533 spin_unlock_bh(&obd->obd_processing_task_lock);
1539 * wait for recovery events,
1540 * check its status with help of check_routine
1541 * evict dead clients via health_check
1543 static int target_recovery_overseer(struct obd_device *obd,
1544 int (*check_routine)(struct obd_device *),
1545 int (*health_check)(struct obd_export *))
1547 int abort = 0, expired = 1;
1550 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1551 spin_lock_bh(&obd->obd_processing_task_lock);
1552 abort = obd->obd_abort_recovery;
1553 expired = obd->obd_recovery_expired;
1554 obd->obd_recovery_expired = 0;
1555 spin_unlock_bh(&obd->obd_processing_task_lock);
1557 CWARN("recovery is aborted, evict exports in recovery\n");
1558 /** evict exports which didn't finish recovery yet */
1559 class_disconnect_stale_exports(obd, exp_finished);
1560 } else if (expired) {
1561 /** If some clients died being recovered, evict them */
1562 CDEBUG(D_WARNING, "recovery is timed out, evict stale exports\n");
1563 /** evict cexports with no replay in queue, they are stalled */
1564 class_disconnect_stale_exports(obd, health_check);
1565 /** continue with VBR */
1566 spin_lock_bh(&obd->obd_processing_task_lock);
1567 obd->obd_version_recov = 1;
1568 spin_unlock_bh(&obd->obd_processing_task_lock);
1570 * reset timer, recovery will proceed with versions now,
1571 * timeout is set just to handle reconnection delays
1573 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1574 /** Wait for recovery events again, after evicting bad clients */
1576 } while (!abort && expired);
1581 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1583 struct ptlrpc_request *req = NULL;
1586 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1587 obd->obd_next_recovery_transno);
1589 if (target_recovery_overseer(obd, check_for_next_transno,
1590 exp_req_replay_healthy)) {
1591 abort_req_replay_queue(obd);
1592 abort_lock_replay_queue(obd);
1595 spin_lock_bh(&obd->obd_processing_task_lock);
1596 if (!list_empty(&obd->obd_req_replay_queue)) {
1597 req = list_entry(obd->obd_req_replay_queue.next,
1598 struct ptlrpc_request, rq_list);
1599 list_del_init(&req->rq_list);
1600 obd->obd_requests_queued_for_recovery--;
1601 spin_unlock_bh(&obd->obd_processing_task_lock);
1603 spin_unlock_bh(&obd->obd_processing_task_lock);
1604 LASSERT(list_empty(&obd->obd_req_replay_queue));
1605 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1606 /** evict exports failed VBR */
1607 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1612 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1614 struct ptlrpc_request *req = NULL;
1616 CDEBUG(D_HA, "Waiting for lock\n");
1617 if (target_recovery_overseer(obd, check_for_next_lock,
1618 exp_lock_replay_healthy))
1619 abort_lock_replay_queue(obd);
1621 spin_lock_bh(&obd->obd_processing_task_lock);
1622 if (!list_empty(&obd->obd_lock_replay_queue)) {
1623 req = list_entry(obd->obd_lock_replay_queue.next,
1624 struct ptlrpc_request, rq_list);
1625 list_del_init(&req->rq_list);
1626 spin_unlock_bh(&obd->obd_processing_task_lock);
1628 spin_unlock_bh(&obd->obd_processing_task_lock);
1629 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1630 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1631 /** evict exports failed VBR */
1632 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1637 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1639 struct ptlrpc_request *req = NULL;
1641 spin_lock_bh(&obd->obd_processing_task_lock);
1642 if (!list_empty(&obd->obd_final_req_queue)) {
1643 req = list_entry(obd->obd_final_req_queue.next,
1644 struct ptlrpc_request, rq_list);
1645 list_del_init(&req->rq_list);
1646 if (req->rq_export->exp_in_recovery) {
1647 spin_lock(&req->rq_export->exp_lock);
1648 req->rq_export->exp_in_recovery = 0;
1649 spin_unlock(&req->rq_export->exp_lock);
1652 spin_unlock_bh(&obd->obd_processing_task_lock);
1656 static int handle_recovery_req(struct ptlrpc_thread *thread,
1657 struct ptlrpc_request *req,
1658 svc_handler_t handler)
1663 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1665 CERROR("Failure to initialize session: %d\n", rc);
1666 GOTO(free_clone, rc);
1669 * export can be evicted during recovery, no need to handle replays for
1670 * it after that, discard such request silently
1672 if (req->rq_export->exp_disconnected)
1673 GOTO(free_clone, rc);
1675 req->rq_session.lc_thread = thread;
1676 lu_context_enter(&req->rq_session);
1677 req->rq_svc_thread = thread;
1678 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1680 /* thread context */
1681 lu_context_enter(&thread->t_env->le_ctx);
1683 lu_context_exit(&thread->t_env->le_ctx);
1685 lu_context_exit(&req->rq_session);
1686 lu_context_fini(&req->rq_session);
1687 /* don't reset timer for final stage */
1688 if (!exp_finished(req->rq_export)) {
1690 * XXX: until bug 18948 is fixed (enable AT for request copy)
1691 * the client may reconnect during recovery so we may need to
1692 * wait RECONNECT_DELAY_MAX after each replay instead of
1693 * at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate);
1695 reset_recovery_timer(class_exp2obd(req->rq_export), AT_OFF ?
1696 obd_timeout : RECONNECT_DELAY_MAX, 1);
1699 * bz18031: increase next_recovery_transno before ptlrpc_free_clone()
1700 * will drop exp_rpc reference
1702 if (req->rq_export->exp_req_replay_needed) {
1703 spin_lock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1704 req->rq_export->exp_obd->obd_next_recovery_transno++;
1705 spin_unlock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1706 target_exp_dequeue_req_replay(req);
1709 ptlrpc_free_clone(req);
1713 static int target_recovery_thread(void *arg)
1715 struct lu_target *lut = arg;
1716 struct obd_device *obd = lut->lut_obd;
1717 struct ptlrpc_request *req;
1718 struct target_recovery_data *trd = &obd->obd_recovery_data;
1719 unsigned long delta;
1720 unsigned long flags;
1722 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1726 cfs_daemonize("tgt_recov");
1728 SIGNAL_MASK_LOCK(current, flags);
1729 sigfillset(¤t->blocked);
1731 SIGNAL_MASK_UNLOCK(current, flags);
1733 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1737 thread->t_env = &env;
1738 env.le_ctx.lc_thread = thread;
1740 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1742 trd->trd_processing_task = cfs_curproc_pid();
1744 obd->obd_recovering = 1;
1745 complete(&trd->trd_starting);
1747 /* first of all, we have to know the first transno to replay */
1748 if (target_recovery_overseer(obd, check_for_clients,
1749 exp_connect_healthy)) {
1750 abort_req_replay_queue(obd);
1751 abort_lock_replay_queue(obd);
1754 /* next stage: replay requests */
1756 obd->obd_req_replaying = 1;
1757 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1758 atomic_read(&obd->obd_req_replay_clients),
1759 obd->obd_next_recovery_transno);
1760 while ((req = target_next_replay_req(obd))) {
1761 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1762 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1763 lustre_msg_get_transno(req->rq_reqmsg),
1764 libcfs_nid2str(req->rq_peer.nid));
1765 handle_recovery_req(thread, req,
1766 trd->trd_recovery_handler);
1767 obd->obd_replayed_requests++;
1771 * The second stage: replay locks
1773 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1774 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 obd->obd_replayed_locks++;
1785 * The third stage: reply on final pings, at this moment all clients
1786 * must have request in final queue
1788 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1789 /** Update server last boot epoch */
1790 lut_boot_epoch_update(lut);
1791 /* We drop recoverying flag to forward all new requests
1792 * to regular mds_handle() since now */
1793 spin_lock_bh(&obd->obd_processing_task_lock);
1794 obd->obd_recovering = obd->obd_abort_recovery = 0;
1795 target_cancel_recovery_timer(obd);
1796 spin_unlock_bh(&obd->obd_processing_task_lock);
1797 while ((req = target_next_final_ping(obd))) {
1798 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1799 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1800 libcfs_nid2str(req->rq_peer.nid));
1801 handle_recovery_req(thread, req,
1802 trd->trd_recovery_handler);
1805 delta = (jiffies - delta) / HZ;
1806 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1807 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1808 if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1809 CWARN("too long recovery - read logs\n");
1810 libcfs_debug_dumplog();
1813 target_finish_recovery(obd);
1815 lu_context_fini(&env.le_ctx);
1816 trd->trd_processing_task = 0;
1817 complete(&trd->trd_finishing);
1821 static int target_start_recovery_thread(struct lu_target *lut,
1822 svc_handler_t handler)
1824 struct obd_device *obd = lut->lut_obd;
1826 struct target_recovery_data *trd = &obd->obd_recovery_data;
1828 memset(trd, 0, sizeof(*trd));
1829 init_completion(&trd->trd_starting);
1830 init_completion(&trd->trd_finishing);
1831 trd->trd_recovery_handler = handler;
1833 if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1834 wait_for_completion(&trd->trd_starting);
1835 LASSERT(obd->obd_recovering != 0);
1842 void target_stop_recovery_thread(struct obd_device *obd)
1844 spin_lock_bh(&obd->obd_processing_task_lock);
1845 if (obd->obd_recovery_data.trd_processing_task > 0) {
1846 struct target_recovery_data *trd = &obd->obd_recovery_data;
1847 /** recovery can be done but postrecovery is not yet */
1848 if (obd->obd_recovering) {
1849 CERROR("%s: Aborting recovery\n", obd->obd_name);
1850 obd->obd_abort_recovery = 1;
1851 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1853 spin_unlock_bh(&obd->obd_processing_task_lock);
1854 wait_for_completion(&trd->trd_finishing);
1856 spin_unlock_bh(&obd->obd_processing_task_lock);
1860 void target_recovery_fini(struct obd_device *obd)
1862 class_disconnect_exports(obd);
1863 target_stop_recovery_thread(obd);
1864 target_cleanup_recovery(obd);
1866 EXPORT_SYMBOL(target_recovery_fini);
1868 static void target_recovery_expired(unsigned long castmeharder)
1870 struct obd_device *obd = (struct obd_device *)castmeharder;
1871 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1872 " after %lds (%d clients connected)\n",
1873 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
1874 cfs_time_current_sec()- obd->obd_recovery_start,
1875 obd->obd_connected_clients);
1877 spin_lock_bh(&obd->obd_processing_task_lock);
1878 obd->obd_recovery_expired = 1;
1879 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1880 spin_unlock_bh(&obd->obd_processing_task_lock);
1883 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1885 struct obd_device *obd = lut->lut_obd;
1886 if (obd->obd_max_recoverable_clients == 0) {
1887 /** Update server last boot epoch */
1888 lut_boot_epoch_update(lut);
1892 CWARN("RECOVERY: service %s, %d recoverable clients, "
1893 "last_transno "LPU64"\n", obd->obd_name,
1894 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1895 LASSERT(obd->obd_stopping == 0);
1896 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1897 obd->obd_recovery_start = 0;
1898 obd->obd_recovery_end = 0;
1899 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1900 /* bz13079: this should be set to desired value for ost but not for mds */
1901 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1902 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1903 target_start_recovery_thread(lut, handler);
1905 EXPORT_SYMBOL(target_recovery_init);
1909 static int target_process_req_flags(struct obd_device *obd,
1910 struct ptlrpc_request *req)
1912 struct obd_export *exp = req->rq_export;
1913 LASSERT(exp != NULL);
1914 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1915 /* client declares he's ready to replay locks */
1916 spin_lock_bh(&obd->obd_processing_task_lock);
1917 if (exp->exp_req_replay_needed) {
1918 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1919 spin_lock(&exp->exp_lock);
1920 exp->exp_req_replay_needed = 0;
1921 spin_unlock(&exp->exp_lock);
1922 atomic_dec(&obd->obd_req_replay_clients);
1924 spin_unlock_bh(&obd->obd_processing_task_lock);
1926 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1927 /* client declares he's ready to complete recovery
1928 * so, we put the request on th final queue */
1929 spin_lock_bh(&obd->obd_processing_task_lock);
1930 if (exp->exp_lock_replay_needed) {
1931 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1932 spin_lock(&exp->exp_lock);
1933 exp->exp_lock_replay_needed = 0;
1934 spin_unlock(&exp->exp_lock);
1935 atomic_dec(&obd->obd_lock_replay_clients);
1937 spin_unlock_bh(&obd->obd_processing_task_lock);
1943 int target_queue_recovery_request(struct ptlrpc_request *req,
1944 struct obd_device *obd)
1946 struct list_head *tmp;
1948 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1951 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1952 /* Processing the queue right now, don't re-add. */
1956 target_process_req_flags(obd, req);
1958 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1959 /* client declares he's ready to complete recovery
1960 * so, we put the request on th final queue */
1961 req = ptlrpc_clone_req(req);
1964 DEBUG_REQ(D_HA, req, "queue final req");
1965 spin_lock_bh(&obd->obd_processing_task_lock);
1966 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1967 if (obd->obd_recovering)
1968 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1970 spin_unlock_bh(&obd->obd_processing_task_lock);
1971 ptlrpc_free_clone(req);
1972 if (obd->obd_stopping) {
1978 spin_unlock_bh(&obd->obd_processing_task_lock);
1981 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1982 /* client declares he's ready to replay locks */
1983 req = ptlrpc_clone_req(req);
1986 DEBUG_REQ(D_HA, req, "queue lock replay req");
1987 spin_lock_bh(&obd->obd_processing_task_lock);
1988 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1989 LASSERT(obd->obd_recovering);
1990 /* usually due to recovery abort */
1991 if (!req->rq_export->exp_in_recovery) {
1992 spin_unlock_bh(&obd->obd_processing_task_lock);
1993 ptlrpc_free_clone(req);
1996 LASSERT(req->rq_export->exp_lock_replay_needed);
1997 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1998 spin_unlock_bh(&obd->obd_processing_task_lock);
2002 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2003 * (i.e. buflens etc are in my own byte order), but type-dependent
2004 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2007 CFS_INIT_LIST_HEAD(&req->rq_list);
2008 DEBUG_REQ(D_HA, req, "not queueing");
2012 spin_lock_bh(&obd->obd_processing_task_lock);
2014 /* If we're processing the queue, we want don't want to queue this
2017 * Also, if this request has a transno less than the one we're waiting
2018 * for, we should process it now. It could (and currently always will)
2019 * be an open request for a descriptor that was opened some time ago.
2021 * Also, a resent, replayed request that has already been
2022 * handled will pass through here and be processed immediately.
2024 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2025 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2026 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2027 /* Processing the queue right now, don't re-add. */
2028 LASSERT(list_empty(&req->rq_list));
2029 spin_unlock_bh(&obd->obd_processing_task_lock);
2032 spin_unlock_bh(&obd->obd_processing_task_lock);
2034 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2037 req = ptlrpc_clone_req(req);
2041 spin_lock_bh(&obd->obd_processing_task_lock);
2042 LASSERT(obd->obd_recovering);
2043 if (!req->rq_export->exp_in_recovery) {
2044 spin_unlock_bh(&obd->obd_processing_task_lock);
2045 ptlrpc_free_clone(req);
2048 LASSERT(req->rq_export->exp_req_replay_needed);
2050 if (target_exp_enqueue_req_replay(req)) {
2051 spin_unlock_bh(&obd->obd_processing_task_lock);
2052 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2053 ptlrpc_free_clone(req);
2058 list_for_each(tmp, &obd->obd_req_replay_queue) {
2059 struct ptlrpc_request *reqiter =
2060 list_entry(tmp, struct ptlrpc_request, rq_list);
2062 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2063 list_add_tail(&req->rq_list, &reqiter->rq_list);
2068 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2070 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2071 "has been claimed by another client");
2072 spin_unlock_bh(&obd->obd_processing_task_lock);
2073 target_exp_dequeue_req_replay(req);
2074 ptlrpc_free_clone(req);
2080 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2082 obd->obd_requests_queued_for_recovery++;
2083 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2084 spin_unlock_bh(&obd->obd_processing_task_lock);
2089 * Packs current SLV and Limit into \a req.
2091 int target_pack_pool_reply(struct ptlrpc_request *req)
2093 struct obd_device *obd;
2097 * Check that we still have all structures alive as this may
2098 * be some late rpc in shutdown time.
2100 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2101 !exp_connect_lru_resize(req->rq_export))) {
2102 lustre_msg_set_slv(req->rq_repmsg, 0);
2103 lustre_msg_set_limit(req->rq_repmsg, 0);
2108 * OBD is alive here as export is alive, which we checked above.
2110 obd = req->rq_export->exp_obd;
2112 read_lock(&obd->obd_pool_lock);
2113 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2114 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2115 read_unlock(&obd->obd_pool_lock);
2120 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2122 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2123 DEBUG_REQ(D_ERROR, req, "dropping reply");
2128 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2129 req->rq_status = rc;
2130 return (ptlrpc_send_error(req, 1));
2132 DEBUG_REQ(D_NET, req, "sending reply");
2135 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2138 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2141 struct ptlrpc_reply_state *rs;
2142 struct obd_device *obd;
2143 struct obd_export *exp;
2144 struct ptlrpc_service *svc;
2147 if (req->rq_no_reply) {
2152 svc = req->rq_rqbd->rqbd_service;
2153 rs = req->rq_reply_state;
2154 if (rs == NULL || !rs->rs_difficult) {
2156 target_send_reply_msg (req, rc, fail_id);
2161 /* must be an export if locks saved */
2162 LASSERT (req->rq_export != NULL);
2163 /* req/reply consistent */
2164 LASSERT (rs->rs_service == svc);
2167 LASSERT (!rs->rs_scheduled);
2168 LASSERT (!rs->rs_scheduled_ever);
2169 LASSERT (!rs->rs_handled);
2170 LASSERT (!rs->rs_on_net);
2171 LASSERT (rs->rs_export == NULL);
2172 LASSERT (list_empty(&rs->rs_obd_list));
2173 LASSERT (list_empty(&rs->rs_exp_list));
2175 exp = class_export_get (req->rq_export);
2178 /* disable reply scheduling while I'm setting up */
2179 rs->rs_scheduled = 1;
2181 rs->rs_xid = req->rq_xid;
2182 rs->rs_transno = req->rq_transno;
2183 rs->rs_export = exp;
2184 rs->rs_opc = lustre_msg_get_opc(rs->rs_msg);
2186 spin_lock(&exp->exp_uncommitted_replies_lock);
2187 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2188 rs->rs_transno, exp->exp_last_committed);
2189 if (rs->rs_transno > exp->exp_last_committed) {
2190 /* not committed already */
2191 list_add_tail(&rs->rs_obd_list,
2192 &exp->exp_uncommitted_replies);
2194 spin_unlock (&exp->exp_uncommitted_replies_lock);
2196 spin_lock(&exp->exp_lock);
2197 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2198 spin_unlock(&exp->exp_lock);
2200 netrc = target_send_reply_msg (req, rc, fail_id);
2202 spin_lock(&svc->srv_lock);
2204 atomic_inc(&svc->srv_n_difficult_replies);
2207 /* error sending: reply is off the net. Also we need +1
2208 * reply ref until ptlrpc_handle_rs() is done
2209 * with the reply state (if the send was successful, there
2210 * would have been +1 ref for the net, which
2211 * reply_out_callback leaves alone) */
2213 ptlrpc_rs_addref(rs);
2214 atomic_inc (&svc->srv_outstanding_replies);
2217 spin_lock(&rs->rs_lock);
2218 if (rs->rs_transno <= exp->exp_last_committed ||
2219 (!rs->rs_on_net && !rs->rs_no_ack) ||
2220 list_empty(&rs->rs_exp_list) || /* completed already */
2221 list_empty(&rs->rs_obd_list)) {
2222 CDEBUG(D_HA, "Schedule reply immediately\n");
2223 ptlrpc_dispatch_difficult_reply(rs);
2225 list_add (&rs->rs_list, &svc->srv_active_replies);
2226 rs->rs_scheduled = 0; /* allow notifier to schedule */
2228 spin_unlock(&rs->rs_lock);
2229 spin_unlock(&svc->srv_lock);
2233 int target_handle_ping(struct ptlrpc_request *req)
2235 obd_ping(req->rq_export);
2236 return req_capsule_server_pack(&req->rq_pill);
2239 void target_committed_to_req(struct ptlrpc_request *req)
2241 struct obd_export *exp = req->rq_export;
2243 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2244 lustre_msg_set_last_committed(req->rq_repmsg,
2245 exp->exp_last_committed);
2247 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2248 "%d)", exp->exp_obd->obd_no_transno,
2249 req->rq_repmsg == NULL);
2251 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2252 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2254 EXPORT_SYMBOL(target_committed_to_req);
2256 int target_handle_qc_callback(struct ptlrpc_request *req)
2258 struct obd_quotactl *oqctl;
2259 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2261 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2262 if (oqctl == NULL) {
2263 CERROR("Can't unpack obd_quotactl\n");
2267 cli->cl_qchk_stat = oqctl->qc_stat;
2272 #ifdef HAVE_QUOTA_SUPPORT
2273 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2276 struct obd_device *obd = req->rq_export->exp_obd;
2277 struct obd_device *master_obd;
2278 struct obd_device_target *obt;
2279 struct lustre_quota_ctxt *qctxt;
2280 struct qunit_data *qdata = NULL;
2284 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2287 rc = req_capsule_server_pack(&req->rq_pill);
2289 CERROR("packing reply failed!: rc = %d\n", rc);
2293 LASSERT(req->rq_export);
2295 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2296 if (IS_ERR(qdata)) {
2297 rc = PTR_ERR(qdata);
2298 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2299 req->rq_status = rc;
2303 /* we use the observer */
2304 if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2305 CERROR("Can't find the observer, it is recovering\n");
2306 req->rq_status = -EAGAIN;
2310 master_obd = obd->obd_observer->obd_observer;
2311 obt = &master_obd->u.obt;
2312 qctxt = &obt->obt_qctxt;
2314 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2315 /* quota_type has not been processed yet, return EAGAIN
2316 * until we know whether or not quotas are supposed to
2318 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2320 req->rq_status = -EAGAIN;
2324 down_read(&obt->obt_rwsem);
2325 if (qctxt->lqc_lqs_hash == NULL) {
2326 up_read(&obt->obt_rwsem);
2327 /* quota_type has not been processed yet, return EAGAIN
2328 * until we know whether or not quotas are supposed to
2330 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2332 req->rq_status = -EAGAIN;
2336 LASSERT(qctxt->lqc_handler);
2337 rc = qctxt->lqc_handler(master_obd, qdata,
2338 lustre_msg_get_opc(req->rq_reqmsg));
2339 up_read(&obt->obt_rwsem);
2340 if (rc && rc != -EDQUOT)
2341 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2342 "dqacq/dqrel failed! (rc:%d)\n", rc);
2343 req->rq_status = rc;
2345 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2347 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2351 /* Block the quota req. b=14840 */
2352 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2356 rc = ptlrpc_reply(req);
2360 #endif /* !__KERNEL__ */
2362 #endif /* HAVE_QUOTA_SUPPORT */
2364 /* Send a remote set_info_async.
2365 * This may go from client to server or server to client
2367 int target_set_info_rpc(struct obd_import *imp, int opcode,
2368 obd_count keylen, void *key,
2369 obd_count vallen, void *val,
2370 struct ptlrpc_request_set *set)
2372 struct ptlrpc_request *req;
2377 req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
2381 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2382 RCL_CLIENT, keylen);
2383 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2384 RCL_CLIENT, vallen);
2385 rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, opcode);
2387 ptlrpc_request_free(req);
2391 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2392 memcpy(tmp, key, keylen);
2393 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2394 memcpy(tmp, val, vallen);
2396 ptlrpc_request_set_replen(req);
2399 ptlrpc_set_add_req(set, req);
2400 ptlrpc_check_set(NULL, set);
2402 rc = ptlrpc_queue_wait(req);
2403 ptlrpc_req_finished(req);
2408 EXPORT_SYMBOL(target_set_info_rpc);
2411 ldlm_mode_t lck_compat_array[] = {
2412 [LCK_EX] LCK_COMPAT_EX,
2413 [LCK_PW] LCK_COMPAT_PW,
2414 [LCK_PR] LCK_COMPAT_PR,
2415 [LCK_CW] LCK_COMPAT_CW,
2416 [LCK_CR] LCK_COMPAT_CR,
2417 [LCK_NL] LCK_COMPAT_NL,
2418 [LCK_GROUP] LCK_COMPAT_GROUP,
2419 [LCK_COS] LCK_COMPAT_COS,
2423 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2424 * not escape to the user level.
2426 int ldlm_error2errno(ldlm_error_t error)
2434 case ELDLM_LOCK_CHANGED:
2437 case ELDLM_LOCK_ABORTED:
2440 case ELDLM_LOCK_REPLACED:
2443 case ELDLM_NO_LOCK_DATA:
2446 case ELDLM_NAMESPACE_EXISTS:
2449 case ELDLM_BAD_NAMESPACE:
2453 if (((int)error) < 0) /* cast to signed type */
2454 result = error; /* as ldlm_error_t can be unsigned */
2456 CERROR("Invalid DLM result code: %i\n", error);
2462 EXPORT_SYMBOL(ldlm_error2errno);
2465 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2467 ldlm_error_t ldlm_errno2error(int err_no)
2476 error = ELDLM_LOCK_CHANGED;
2479 error = ELDLM_LOCK_ABORTED;
2482 error = ELDLM_LOCK_REPLACED;
2485 error = ELDLM_NO_LOCK_DATA;
2488 error = ELDLM_NAMESPACE_EXISTS;
2491 error = ELDLM_BAD_NAMESPACE;
2498 EXPORT_SYMBOL(ldlm_errno2error);