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 item->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 static void 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_write_list);
281 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
282 client_obd_list_lock_init(&cli->cl_loi_list_lock);
283 cli->cl_r_in_flight = 0;
284 cli->cl_w_in_flight = 0;
286 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
287 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
288 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
289 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
290 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
291 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
292 cfs_waitq_init(&cli->cl_destroy_waitq);
293 atomic_set(&cli->cl_destroy_in_flight, 0);
294 #ifdef ENABLE_CHECKSUM
295 /* Turn on checksumming by default. */
296 cli->cl_checksum = 1;
298 * The supported checksum types will be worked out at connect time
299 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
302 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
304 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
306 /* This value may be changed at connect time in
307 ptlrpc_connect_interpret. */
308 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
309 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
311 if (!strcmp(name, LUSTRE_MDC_NAME)) {
312 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
313 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
314 cli->cl_max_rpcs_in_flight = 2;
315 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
316 cli->cl_max_rpcs_in_flight = 3;
317 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
318 cli->cl_max_rpcs_in_flight = 4;
320 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
325 CERROR("ldlm_get_ref failed: %d\n", rc);
329 ptlrpc_init_client(rq_portal, rp_portal, name,
330 &obddev->obd_ldlm_client);
332 imp = class_new_import(obddev);
334 GOTO(err_ldlm, rc = -ENOENT);
335 imp->imp_client = &obddev->obd_ldlm_client;
336 imp->imp_connect_op = connect_op;
337 imp->imp_initial_recov = 1;
338 imp->imp_initial_recov_bk = 0;
339 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
340 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
341 LUSTRE_CFG_BUFLEN(lcfg, 1));
342 class_import_put(imp);
344 rc = client_import_add_conn(imp, &server_uuid, 1);
346 CERROR("can't add initial connection\n");
347 GOTO(err_import, rc);
350 cli->cl_import = imp;
351 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
352 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
353 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
355 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
356 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
357 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
358 name, obddev->obd_name,
359 cli->cl_target_uuid.uuid);
360 spin_lock(&imp->imp_lock);
361 imp->imp_invalid = 1;
362 spin_unlock(&imp->imp_lock);
366 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
371 class_destroy_import(imp);
379 int client_obd_cleanup(struct obd_device *obddev)
386 /* ->o_connect() method for client side (OSC and MDC and MGC) */
387 int client_connect_import(const struct lu_env *env,
388 struct obd_export **exp,
389 struct obd_device *obd, struct obd_uuid *cluuid,
390 struct obd_connect_data *data, void *localdata)
392 struct client_obd *cli = &obd->u.cli;
393 struct obd_import *imp = cli->cl_import;
394 struct obd_connect_data *ocd;
395 struct ldlm_namespace *to_be_freed = NULL;
396 struct lustre_handle conn = { 0 };
401 down_write(&cli->cl_sem);
402 rc = class_connect(&conn, obd, cluuid);
406 *exp = class_conn2export(&conn);
408 cli->cl_conn_count++;
409 if (cli->cl_conn_count > 1)
412 if (obd->obd_namespace != NULL)
413 CERROR("already have namespace!\n");
414 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
415 LDLM_NAMESPACE_CLIENT,
416 LDLM_NAMESPACE_GREEDY);
417 if (obd->obd_namespace == NULL)
418 GOTO(out_disco, rc = -ENOMEM);
420 imp->imp_dlm_handle = conn;
421 rc = ptlrpc_init_import(imp);
425 ocd = &imp->imp_connect_data;
428 imp->imp_connect_flags_orig = data->ocd_connect_flags;
431 rc = ptlrpc_connect_import(imp, NULL);
433 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
436 LASSERT((*exp)->exp_connection);
439 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
440 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
441 data->ocd_connect_flags, ocd->ocd_connect_flags);
442 data->ocd_connect_flags = ocd->ocd_connect_flags;
445 ptlrpc_pinger_add_import(imp);
451 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
452 to_be_freed = obd->obd_namespace;
453 obd->obd_namespace = NULL;
455 cli->cl_conn_count--;
456 class_disconnect(*exp);
460 up_write(&cli->cl_sem);
462 ldlm_namespace_free_post(to_be_freed);
467 int client_disconnect_export(struct obd_export *exp)
469 struct obd_device *obd = class_exp2obd(exp);
470 struct client_obd *cli;
471 struct obd_import *imp;
473 struct ldlm_namespace *to_be_freed = NULL;
477 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
478 exp, exp ? exp->exp_handle.h_cookie : -1);
483 imp = cli->cl_import;
485 down_write(&cli->cl_sem);
486 if (!cli->cl_conn_count) {
487 CERROR("disconnecting disconnected device (%s)\n",
489 GOTO(out_sem, rc = -EINVAL);
492 cli->cl_conn_count--;
493 if (cli->cl_conn_count)
494 GOTO(out_no_disconnect, rc = 0);
496 /* Mark import deactivated now, so we don't try to reconnect if any
497 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
498 * fully deactivate the import, or that would drop all requests. */
499 spin_lock(&imp->imp_lock);
500 imp->imp_deactive = 1;
501 spin_unlock(&imp->imp_lock);
503 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
504 * delete it regardless. (It's safe to delete an import that was
506 (void)ptlrpc_pinger_del_import(imp);
508 if (obd->obd_namespace != NULL) {
509 /* obd_force == local only */
510 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
511 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
513 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
514 to_be_freed = obd->obd_namespace;
518 * there's no necessary to hold sem during diconnecting an import,
519 * and actually it may cause deadlock in gss.
521 up_write(&cli->cl_sem);
522 rc = ptlrpc_disconnect_import(imp, 0);
523 down_write(&cli->cl_sem);
525 ptlrpc_invalidate_import(imp);
526 /* set obd_namespace to NULL only after invalidate, because we can have
527 * some connect requests in flight, and his need store a connect flags
528 * in obd_namespace. bug 14260 */
529 obd->obd_namespace = NULL;
531 if (imp->imp_rq_pool) {
532 ptlrpc_free_rq_pool(imp->imp_rq_pool);
533 imp->imp_rq_pool = NULL;
536 cli->cl_import = NULL;
540 err = class_disconnect(exp);
544 up_write(&cli->cl_sem);
546 ldlm_namespace_free_post(to_be_freed);
551 /* --------------------------------------------------------------------------
552 * from old lib/target.c
553 * -------------------------------------------------------------------------- */
555 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
556 struct obd_uuid *cluuid, int mds_conn)
559 if (exp->exp_connection && exp->exp_imp_reverse && !mds_conn) {
560 struct lustre_handle *hdl;
561 hdl = &exp->exp_imp_reverse->imp_remote_handle;
562 /* Might be a re-connect after a partition. */
563 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
564 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
566 conn->cookie = exp->exp_handle.h_cookie;
567 /* target_handle_connect() treats EALREADY and
568 * -EALREADY differently. EALREADY means we are
569 * doing a valid reconnect from the same client. */
572 CERROR("%s reconnecting from %s, "
573 "handle mismatch (ours "LPX64", theirs "
574 LPX64")\n", cluuid->uuid,
575 exp->exp_connection->c_remote_uuid.uuid,
576 hdl->cookie, conn->cookie);
577 memset(conn, 0, sizeof *conn);
578 /* target_handle_connect() treats EALREADY and
579 * -EALREADY differently. -EALREADY is an error
580 * (same UUID, different handle). */
585 conn->cookie = exp->exp_handle.h_cookie;
586 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
587 cluuid->uuid, exp, conn->cookie);
591 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
594 struct obd_export *exp = cb_data;
596 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
597 obd->obd_name, exp->exp_client_uuid.uuid);
599 spin_lock(&exp->exp_lock);
600 exp->exp_need_sync = 0;
601 spin_unlock(&exp->exp_lock);
603 EXPORT_SYMBOL(target_client_add_cb);
605 target_start_and_reset_recovery_timer(struct obd_device *obd,
606 struct ptlrpc_request *req,
609 int target_handle_connect(struct ptlrpc_request *req)
611 struct obd_device *target, *targref = NULL;
612 struct obd_export *export = NULL;
613 struct obd_import *revimp;
614 struct lustre_handle conn;
615 struct lustre_handle *tmp;
616 struct obd_uuid tgtuuid;
617 struct obd_uuid cluuid;
618 struct obd_uuid remote_uuid;
622 struct obd_connect_data *data, *tmpdata;
623 lnet_nid_t *client_nid = NULL;
626 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
628 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
630 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
631 GOTO(out, rc = -EINVAL);
634 obd_str2uuid(&tgtuuid, str);
635 target = class_uuid2obd(&tgtuuid);
637 target = class_name2obd(str);
639 if (!target || target->obd_stopping || !target->obd_set_up) {
640 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
641 " for connect (%s)\n", str,
642 !target ? "no target" :
643 (target->obd_stopping ? "stopping" :
645 GOTO(out, rc = -ENODEV);
648 if (target->obd_no_conn) {
649 LCONSOLE_WARN("%s: temporarily refusing client connection "
650 "from %s\n", target->obd_name,
651 libcfs_nid2str(req->rq_peer.nid));
652 GOTO(out, rc = -EAGAIN);
655 /* Make sure the target isn't cleaned up while we're here. Yes,
656 there's still a race between the above check and our incref here.
657 Really, class_uuid2obd should take the ref. */
658 targref = class_incref(target, __FUNCTION__, cfs_current());
661 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
663 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
664 GOTO(out, rc = -EINVAL);
667 obd_str2uuid(&cluuid, str);
669 /* XXX extract a nettype and format accordingly */
670 switch (sizeof(lnet_nid_t)) {
671 /* NB the casts only avoid compiler warnings */
673 snprintf(remote_uuid.uuid, sizeof remote_uuid,
674 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
677 snprintf(remote_uuid.uuid, sizeof remote_uuid,
678 "NET_%x_UUID", (__u32)req->rq_peer.nid);
684 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
686 GOTO(out, rc = -EPROTO);
690 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
692 GOTO(out, rc = -EPROTO);
694 rc = req_capsule_server_pack(&req->rq_pill);
698 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
700 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
701 "libclient connection attempt");
702 GOTO(out, rc = -EPROTO);
703 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
704 LUSTRE_VERSION_ALLOWED_OFFSET ||
705 data->ocd_version > LUSTRE_VERSION_CODE +
706 LUSTRE_VERSION_ALLOWED_OFFSET) {
707 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
708 "libclient connection attempt",
709 data->ocd_version < LUSTRE_VERSION_CODE ?
711 OBD_OCD_VERSION_MAJOR(data->ocd_version),
712 OBD_OCD_VERSION_MINOR(data->ocd_version),
713 OBD_OCD_VERSION_PATCH(data->ocd_version),
714 OBD_OCD_VERSION_FIX(data->ocd_version));
715 data = req_capsule_server_sized_get(&req->rq_pill,
717 offsetof(typeof(*data), ocd_version) +
718 sizeof(data->ocd_version));
720 data->ocd_connect_flags = OBD_CONNECT_VERSION;
721 data->ocd_version = LUSTRE_VERSION_CODE;
723 GOTO(out, rc = -EPROTO);
727 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
728 (data->ocd_connect_flags & OBD_CONNECT_MDS))
731 /* lctl gets a backstage, all-access pass. */
732 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
733 goto dont_check_exports;
735 spin_lock(&target->obd_dev_lock);
736 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
738 if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
739 CWARN("%s: exp %p already connecting\n",
740 export->exp_obd->obd_name, export);
741 class_export_put(export);
744 } else if (export != NULL && export->exp_connection != NULL &&
745 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
747 /* in mds failover we have static uuid but nid can be
749 CWARN("%s: cookie %s seen on new NID %s when "
750 "existing NID %s is already connected\n",
751 target->obd_name, cluuid.uuid,
752 libcfs_nid2str(req->rq_peer.nid),
753 libcfs_nid2str(export->exp_connection->c_peer.nid));
755 class_export_put(export);
757 } else if (export != NULL) {
758 spin_lock(&export->exp_lock);
759 export->exp_connecting = 1;
760 spin_unlock(&export->exp_lock);
761 class_export_put(export);
762 spin_unlock(&target->obd_dev_lock);
763 LASSERT(export->exp_obd == target);
765 rc = target_handle_reconnect(&conn, export, &cluuid, mds_conn);
768 /* If we found an export, we already unlocked. */
770 spin_unlock(&target->obd_dev_lock);
771 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
772 } else if (req->rq_export == NULL &&
773 atomic_read(&export->exp_rpc_count) > 0) {
774 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
775 target->obd_name, cluuid.uuid,
776 libcfs_nid2str(req->rq_peer.nid),
777 export, atomic_read(&export->exp_refcount));
778 GOTO(out, rc = -EBUSY);
779 } else if (req->rq_export != NULL &&
780 (atomic_read(&export->exp_rpc_count) > 1)) {
781 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
782 target->obd_name, cluuid.uuid,
783 libcfs_nid2str(req->rq_peer.nid),
784 export, atomic_read(&export->exp_rpc_count));
785 GOTO(out, rc = -EBUSY);
786 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
788 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
789 "cookies not random?\n", target->obd_name,
790 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
791 GOTO(out, rc = -EALREADY);
793 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
794 if (req->rq_export == NULL && mds_conn)
795 export->exp_last_request_time =
796 max(export->exp_last_request_time,
797 (time_t)cfs_time_current_sec());
804 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
805 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
806 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
807 export, (long)cfs_time_current_sec(),
808 export ? (long)export->exp_last_request_time : 0);
810 /* Tell the client if we're in recovery. */
811 if (target->obd_recovering) {
812 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
813 /* If this is the first time a client connects,
814 reset the recovery timer */
816 target_start_and_reset_recovery_timer(target, req,
820 /* We want to handle EALREADY but *not* -EALREADY from
821 * target_handle_reconnect(), return reconnection state in a flag */
822 if (rc == EALREADY) {
823 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
829 /* Tell the client if we support replayable requests */
830 if (target->obd_replayable)
831 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
832 client_nid = &req->rq_peer.nid;
834 if (export == NULL) {
835 if (target->obd_recovering) {
838 t = cfs_timer_deadline(&target->obd_recovery_timer);
839 t = cfs_time_sub(t, cfs_time_current());
840 CERROR("%s: denying connection for new client %s (%s): "
841 "%d clients in recovery for "CFS_TIME_T"s\n",
843 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
844 target->obd_recoverable_clients,
845 cfs_duration_sec(t));
849 rc = obd_connect(req->rq_svc_thread->t_env,
850 &export, target, &cluuid, data,
853 conn.cookie = export->exp_handle.h_cookie;
856 rc = obd_reconnect(req->rq_svc_thread->t_env,
857 export, target, &cluuid, data, client_nid);
859 /* prevous done via class_conn2export */
860 class_export_get(export);
864 /* Return only the parts of obd_connect_data that we understand, so the
865 * client knows that we don't understand the rest. */
867 tmpdata = req_capsule_server_get(&req->rq_pill,
869 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
873 /* If all else goes well, this is our RPC return code. */
876 lustre_msg_set_handle(req->rq_repmsg, &conn);
878 /* If the client and the server are the same node, we will already
879 * have an export that really points to the client's DLM export,
880 * because we have a shared handles table.
882 * XXX this will go away when shaver stops sending the "connect" handle
883 * in the real "remote handle" field of the request --phik 24 Apr 2003
885 if (req->rq_export != NULL)
886 class_export_put(req->rq_export);
888 req->rq_export = export;
890 spin_lock(&export->exp_lock);
892 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
893 } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
894 spin_unlock(&export->exp_lock);
895 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
896 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
897 export->exp_conn_cnt,
898 lustre_msg_get_conn_cnt(req->rq_reqmsg));
900 GOTO(out, rc = -EALREADY);
902 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
904 /* request from liblustre? Don't evict it for not pinging. */
905 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
906 export->exp_libclient = 1;
907 spin_unlock(&export->exp_lock);
909 spin_lock(&target->obd_dev_lock);
910 list_del_init(&export->exp_obd_chain_timed);
911 spin_unlock(&target->obd_dev_lock);
913 spin_unlock(&export->exp_lock);
916 if (export->exp_connection != NULL)
917 ptlrpc_connection_put(export->exp_connection);
918 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
921 if (hlist_unhashed(&export->exp_nid_hash)) {
922 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
923 &export->exp_connection->c_peer.nid,
924 &export->exp_nid_hash);
927 spin_lock_bh(&target->obd_processing_task_lock);
928 if (target->obd_recovering && !export->exp_in_recovery) {
929 spin_lock(&export->exp_lock);
930 export->exp_in_recovery = 1;
931 export->exp_req_replay_needed = 1;
932 export->exp_lock_replay_needed = 1;
933 spin_unlock(&export->exp_lock);
934 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
935 && (data->ocd_transno == 0))
936 CWARN("Connect with zero transno!\n");
938 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
939 && data->ocd_transno < target->obd_next_recovery_transno)
940 target->obd_next_recovery_transno = data->ocd_transno;
941 target->obd_connected_clients++;
942 /* each connected client is counted as recoverable */
943 target->obd_recoverable_clients++;
944 atomic_inc(&target->obd_req_replay_clients);
945 atomic_inc(&target->obd_lock_replay_clients);
946 if (target->obd_connected_clients ==
947 target->obd_max_recoverable_clients)
948 wake_up(&target->obd_next_transno_waitq);
950 spin_unlock_bh(&target->obd_processing_task_lock);
951 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
954 if (export->exp_imp_reverse != NULL) {
955 /* destroyed import can be still referenced in ctxt */
956 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
957 KEY_REVIMP_UPD, 0, NULL, NULL);
959 /* in some recovery senarios, previous ctx init rpc handled
960 * in sptlrpc_target_export_check() might be used to install
961 * a reverse ctx in this reverse import, and later OBD_CONNECT
962 * using the same gss ctx could reach here and following new
963 * reverse import. note all reverse ctx in new/old import are
964 * actually based on the same gss ctx. so we invalidate ctx
965 * here before destroy import, otherwise flush old import will
966 * lead to remote reverse ctx be destroied, thus the reverse
967 * ctx of new import will lost its peer.
968 * there might be a better way to deal with this???
970 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
972 destroy_import(export->exp_imp_reverse);
975 /* for the rest part, we return -ENOTCONN in case of errors
976 * in order to let client initialize connection again.
978 revimp = export->exp_imp_reverse = class_new_import(target);
980 CERROR("fail to alloc new reverse import.\n");
981 GOTO(out, rc = -ENOTCONN);
984 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
985 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
986 revimp->imp_remote_handle = conn;
987 revimp->imp_dlm_fake = 1;
988 revimp->imp_state = LUSTRE_IMP_FULL;
990 /* unknown versions will be caught in
991 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
992 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
994 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
995 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
996 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
998 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1000 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1002 CERROR("Failed to get sec for reverse import: %d\n", rc);
1003 export->exp_imp_reverse = NULL;
1004 class_destroy_import(revimp);
1007 class_import_put(revimp);
1010 spin_lock(&export->exp_lock);
1011 export->exp_connecting = 0;
1012 spin_unlock(&export->exp_lock);
1015 class_decref(targref, __FUNCTION__, cfs_current());
1017 req->rq_status = rc;
1021 int target_handle_disconnect(struct ptlrpc_request *req)
1026 rc = req_capsule_server_pack(&req->rq_pill);
1030 /* keep the rq_export around so we can send the reply */
1031 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1036 void target_destroy_export(struct obd_export *exp)
1038 /* exports created from last_rcvd data, and "fake"
1039 exports created by lctl don't have an import */
1040 if (exp->exp_imp_reverse != NULL)
1041 destroy_import(exp->exp_imp_reverse);
1043 /* We cancel locks at disconnect time, but this will catch any locks
1044 * granted in a race with recovery-induced disconnect. */
1045 if (exp->exp_obd->obd_namespace != NULL)
1046 ldlm_cancel_locks_for_export(exp);
1050 * Recovery functions
1053 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1055 struct ptlrpc_request *copy_req;
1056 struct lustre_msg *copy_reqmsg;
1057 struct ptlrpc_user_desc *udesc = NULL;
1059 OBD_ALLOC_PTR(copy_req);
1062 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1064 OBD_FREE_PTR(copy_req);
1068 if (orig_req->rq_user_desc) {
1069 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1071 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1073 OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1074 OBD_FREE_PTR(copy_req);
1077 memcpy(udesc, orig_req->rq_user_desc,
1078 sptlrpc_user_desc_size(ngroups));
1081 *copy_req = *orig_req;
1082 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1083 copy_req->rq_reqmsg = copy_reqmsg;
1084 copy_req->rq_user_desc = udesc;
1086 class_export_rpc_get(copy_req->rq_export);
1087 CFS_INIT_LIST_HEAD(©_req->rq_list);
1088 CFS_INIT_LIST_HEAD(©_req->rq_replay_list);
1089 sptlrpc_svc_ctx_addref(copy_req);
1091 if (copy_req->rq_reply_state) {
1092 /* the copied req takes over the reply state */
1093 orig_req->rq_reply_state = NULL;
1094 /* to catch further access */
1095 orig_req->rq_repmsg = NULL;
1096 orig_req->rq_replen = 0;
1102 void ptlrpc_free_clone(struct ptlrpc_request *req)
1104 LASSERT(list_empty(&req->rq_replay_list));
1106 ptlrpc_req_drop_rs(req);
1107 sptlrpc_svc_ctx_decref(req);
1108 class_export_rpc_put(req->rq_export);
1109 list_del_init(&req->rq_list);
1111 if (req->rq_user_desc) {
1112 int ngroups = req->rq_user_desc->pud_ngroups;
1113 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1115 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1119 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1121 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1122 struct obd_export *exp = req->rq_export;
1123 struct ptlrpc_request *reqiter;
1128 spin_lock(&exp->exp_lock);
1129 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1131 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1138 /* we expect it with RESENT and REPLAY flags */
1139 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1140 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1141 CERROR("invalid flags %x of resent replay\n",
1142 lustre_msg_get_flags(req->rq_reqmsg));
1144 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1147 spin_unlock(&exp->exp_lock);
1151 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1153 LASSERT(!list_empty(&req->rq_replay_list));
1154 LASSERT(req->rq_export);
1156 spin_lock(&req->rq_export->exp_lock);
1157 list_del_init(&req->rq_replay_list);
1158 spin_unlock(&req->rq_export->exp_lock);
1162 static void target_finish_recovery(struct obd_device *obd)
1165 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1168 ldlm_reprocess_all_ns(obd->obd_namespace);
1169 spin_lock_bh(&obd->obd_processing_task_lock);
1170 if (list_empty(&obd->obd_req_replay_queue) &&
1171 list_empty(&obd->obd_lock_replay_queue) &&
1172 list_empty(&obd->obd_final_req_queue)) {
1173 obd->obd_processing_task = 0;
1175 CERROR("%s: Recovery queues ( %s%s%s) are empty\n",
1177 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1178 list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1179 list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1180 spin_unlock_bh(&obd->obd_processing_task_lock);
1183 spin_unlock_bh(&obd->obd_processing_task_lock);
1185 /* when recovery finished, cleanup orphans on mds and ost */
1186 if (OBT(obd) && OBP(obd, postrecov)) {
1187 int rc = OBP(obd, postrecov)(obd);
1188 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1189 rc < 0 ? "failed" : "complete", rc);
1192 obd->obd_recovery_end = cfs_time_current_sec();
1196 static void abort_req_replay_queue(struct obd_device *obd)
1198 struct ptlrpc_request *req, *n;
1199 struct list_head abort_list;
1201 CFS_INIT_LIST_HEAD(&abort_list);
1202 spin_lock_bh(&obd->obd_processing_task_lock);
1203 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1204 spin_unlock_bh(&obd->obd_processing_task_lock);
1205 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1206 DEBUG_REQ(D_WARNING, req, "aborted:");
1207 req->rq_status = -ENOTCONN;
1208 if (ptlrpc_error(req)) {
1209 DEBUG_REQ(D_ERROR, req,
1210 "failed abort_req_reply; skipping");
1212 target_exp_dequeue_req_replay(req);
1213 ptlrpc_free_clone(req);
1217 static void abort_lock_replay_queue(struct obd_device *obd)
1219 struct ptlrpc_request *req, *n;
1220 struct list_head abort_list;
1222 CFS_INIT_LIST_HEAD(&abort_list);
1223 spin_lock_bh(&obd->obd_processing_task_lock);
1224 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1225 spin_unlock_bh(&obd->obd_processing_task_lock);
1226 list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1227 DEBUG_REQ(D_ERROR, req, "aborted:");
1228 req->rq_status = -ENOTCONN;
1229 if (ptlrpc_error(req)) {
1230 DEBUG_REQ(D_ERROR, req,
1231 "failed abort_lock_reply; skipping");
1233 ptlrpc_free_clone(req);
1238 /* Called from a cleanup function if the device is being cleaned up
1239 forcefully. The exports should all have been disconnected already,
1240 the only thing left to do is
1241 - clear the recovery flags
1243 - free queued requests and replies, but don't send replies
1244 Because the obd_stopping flag is set, no new requests should be received.
1247 void target_cleanup_recovery(struct obd_device *obd)
1249 struct ptlrpc_request *req, *n;
1250 struct list_head clean_list;
1253 LASSERT(obd->obd_stopping);
1255 CFS_INIT_LIST_HEAD(&clean_list);
1256 spin_lock_bh(&obd->obd_processing_task_lock);
1257 if (!obd->obd_recovering) {
1258 spin_unlock_bh(&obd->obd_processing_task_lock);
1262 obd->obd_recovering = obd->obd_abort_recovery = 0;
1263 target_cancel_recovery_timer(obd);
1265 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1266 spin_unlock_bh(&obd->obd_processing_task_lock);
1268 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1269 LASSERT(req->rq_reply_state == 0);
1270 target_exp_dequeue_req_replay(req);
1271 ptlrpc_free_clone(req);
1274 spin_lock_bh(&obd->obd_processing_task_lock);
1275 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1276 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1277 spin_unlock_bh(&obd->obd_processing_task_lock);
1279 list_for_each_entry_safe(req, n, &clean_list, rq_list){
1280 LASSERT(req->rq_reply_state == 0);
1281 ptlrpc_free_clone(req);
1287 /* obd_processing_task_lock should be held */
1288 void target_cancel_recovery_timer(struct obd_device *obd)
1290 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1291 cfs_timer_disarm(&obd->obd_recovery_timer);
1294 /* extend = 1 means require at least "duration" seconds left in the timer,
1295 extend = 0 means set the total duration (start_recovery_timer) */
1296 static void reset_recovery_timer(struct obd_device *obd, int duration,
1299 cfs_time_t now = cfs_time_current_sec();
1300 cfs_duration_t left;
1302 spin_lock_bh(&obd->obd_processing_task_lock);
1303 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1304 spin_unlock_bh(&obd->obd_processing_task_lock);
1308 left = cfs_time_sub(obd->obd_recovery_end, now);
1310 if (extend && (duration > left))
1311 obd->obd_recovery_timeout += duration - left;
1312 else if (!extend && (duration > obd->obd_recovery_timeout))
1313 /* Track the client's largest expected replay time */
1314 obd->obd_recovery_timeout = duration;
1317 * If total recovery time already exceed the
1318 * obd_recovery_max_time, then CRAY XT3 will
1319 * abort the recovery
1321 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1322 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1324 obd->obd_recovery_end = obd->obd_recovery_start +
1325 obd->obd_recovery_timeout;
1326 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1327 cfs_time_before(now, obd->obd_recovery_end)) {
1328 left = cfs_time_sub(obd->obd_recovery_end, now);
1329 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1331 spin_unlock_bh(&obd->obd_processing_task_lock);
1332 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1333 obd->obd_name, (unsigned)left);
1336 static void check_and_start_recovery_timer(struct obd_device *obd)
1338 spin_lock_bh(&obd->obd_processing_task_lock);
1339 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1340 spin_unlock_bh(&obd->obd_processing_task_lock);
1343 CWARN("%s: starting recovery timer\n", obd->obd_name);
1344 obd->obd_recovery_start = cfs_time_current_sec();
1346 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1347 spin_unlock_bh(&obd->obd_processing_task_lock);
1349 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1352 /* Reset the timer with each new client connection */
1354 * This timer is actually reconnect_timer, which is for making sure
1355 * the total recovery window is at least as big as my reconnect
1356 * attempt timing. So the initial recovery time_out will be set to
1357 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1358 * from client is bigger than this, then the recovery time_out will
1359 * be extend to make sure the client could be reconnected, in the
1360 * process, the timeout from the new client should be ignored.
1364 target_start_and_reset_recovery_timer(struct obd_device *obd,
1365 struct ptlrpc_request *req,
1368 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1370 if (!new_client && service_time)
1371 /* Teach server about old server's estimates, as first guess
1372 * at how long new requests will take. */
1373 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1376 check_and_start_recovery_timer(obd);
1378 /* convert the service time to rpc timeout,
1379 * reuse service_time to limit stack usage */
1380 service_time = at_est2timeout(service_time);
1382 /* We expect other clients to timeout within service_time, then try
1383 * to reconnect, then try the failover server. The max delay between
1384 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1385 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1386 INITIAL_CONNECT_TIMEOUT);
1387 if (service_time > obd->obd_recovery_timeout && !new_client)
1388 reset_recovery_timer(obd, service_time, 0);
1392 static int check_for_next_transno(struct obd_device *obd)
1394 struct ptlrpc_request *req = NULL;
1395 int wake_up = 0, connected, completed, queue_len, max;
1396 __u64 next_transno, req_transno;
1398 spin_lock_bh(&obd->obd_processing_task_lock);
1400 if (!list_empty(&obd->obd_req_replay_queue)) {
1401 req = list_entry(obd->obd_req_replay_queue.next,
1402 struct ptlrpc_request, rq_list);
1403 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1408 max = obd->obd_max_recoverable_clients;
1409 connected = obd->obd_connected_clients;
1410 completed = connected - obd->obd_recoverable_clients;
1411 queue_len = obd->obd_requests_queued_for_recovery;
1412 next_transno = obd->obd_next_recovery_transno;
1414 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1415 "req_transno: "LPU64", next_transno: "LPU64"\n",
1416 max, connected, completed, queue_len, req_transno, next_transno);
1418 if (obd->obd_abort_recovery) {
1419 CDEBUG(D_HA, "waking for aborted recovery\n");
1421 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1422 CDEBUG(D_HA, "waking for completed recovery\n");
1424 } else if (req_transno == next_transno) {
1425 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1427 } else if (queue_len + completed == max) {
1428 /* handle gaps occured due to lost reply. It is allowed gaps
1429 * because all clients are connected and there will be resend
1430 * for missed transaction */
1431 LASSERTF(req_transno >= next_transno,
1432 "req_transno: "LPU64", next_transno: "LPU64"\n",
1433 req_transno, next_transno);
1435 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1436 "waking for skipped transno (skip: "LPD64
1437 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1438 next_transno, queue_len, completed, connected, req_transno);
1439 obd->obd_next_recovery_transno = req_transno;
1441 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1442 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1443 " by fail_lock, waking up ("LPD64")\n", next_transno);
1444 obd->obd_next_recovery_transno = req_transno;
1446 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1447 /* some clients haven't connected in time, but we can try
1448 * to replay requests that demand on already committed ones
1449 * also, we can replay first non-committed transation */
1450 LASSERT(req_transno != 0);
1451 if (req_transno == obd->obd_last_committed + 1) {
1452 obd->obd_next_recovery_transno = req_transno;
1453 } else if (req_transno > obd->obd_last_committed) {
1454 /* can't continue recovery: have no needed transno */
1455 obd->obd_abort_recovery = 1;
1456 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1457 "connected: %d, completed: %d, queue_len: %d, "
1458 "req_transno: "LPU64", next_transno: "LPU64"\n",
1459 max, connected, completed, queue_len,
1460 req_transno, next_transno);
1465 spin_unlock_bh(&obd->obd_processing_task_lock);
1469 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1471 struct l_wait_info lwi = { 0 };
1472 struct ptlrpc_request *req;
1474 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1475 obd->obd_next_recovery_transno);
1476 l_wait_event(obd->obd_next_transno_waitq,
1477 check_for_next_transno(obd), &lwi);
1479 spin_lock_bh(&obd->obd_processing_task_lock);
1480 if (obd->obd_abort_recovery) {
1482 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1483 req = list_entry(obd->obd_req_replay_queue.next,
1484 struct ptlrpc_request, rq_list);
1485 target_exp_dequeue_req_replay(req);
1486 list_del_init(&req->rq_list);
1487 obd->obd_requests_queued_for_recovery--;
1491 spin_unlock_bh(&obd->obd_processing_task_lock);
1495 static int check_for_next_lock(struct obd_device *obd)
1497 struct ptlrpc_request *req = NULL;
1500 spin_lock_bh(&obd->obd_processing_task_lock);
1501 if (!list_empty(&obd->obd_lock_replay_queue)) {
1502 req = list_entry(obd->obd_lock_replay_queue.next,
1503 struct ptlrpc_request, rq_list);
1504 CDEBUG(D_HA, "waking for next lock\n");
1506 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1507 CDEBUG(D_HA, "waking for completed lock replay\n");
1509 } else if (obd->obd_abort_recovery) {
1510 CDEBUG(D_HA, "waking for aborted recovery\n");
1513 spin_unlock_bh(&obd->obd_processing_task_lock);
1518 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1520 struct l_wait_info lwi = { 0 };
1521 struct ptlrpc_request *req;
1523 CDEBUG(D_HA, "Waiting for lock\n");
1524 l_wait_event(obd->obd_next_transno_waitq,
1525 check_for_next_lock(obd), &lwi);
1527 spin_lock_bh(&obd->obd_processing_task_lock);
1528 if (obd->obd_abort_recovery) {
1530 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1531 req = list_entry(obd->obd_lock_replay_queue.next,
1532 struct ptlrpc_request, rq_list);
1533 list_del_init(&req->rq_list);
1537 spin_unlock_bh(&obd->obd_processing_task_lock);
1541 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1543 struct ptlrpc_request *req;
1545 spin_lock_bh(&obd->obd_processing_task_lock);
1546 if (!list_empty(&obd->obd_final_req_queue)) {
1547 req = list_entry(obd->obd_final_req_queue.next,
1548 struct ptlrpc_request, rq_list);
1549 list_del_init(&req->rq_list);
1553 spin_unlock_bh(&obd->obd_processing_task_lock);
1557 static inline int req_replay_done(struct obd_export *exp)
1559 return (exp->exp_req_replay_needed == 0);
1562 static inline int lock_replay_done(struct obd_export *exp)
1564 return (exp->exp_lock_replay_needed == 0);
1567 static inline int connect_done(struct obd_export *exp)
1569 return (exp->exp_in_recovery != 0);
1572 static int check_for_clients(struct obd_device *obd)
1574 if (obd->obd_abort_recovery)
1576 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1577 if (obd->obd_no_conn == 0 &&
1578 obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1583 static int handle_recovery_req(struct ptlrpc_thread *thread,
1584 struct ptlrpc_request *req,
1585 svc_handler_t handler)
1590 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1592 CERROR("Failure to initialize session: %d\n", rc);
1595 req->rq_session.lc_thread = thread;
1596 lu_context_enter(&req->rq_session);
1597 req->rq_svc_thread = thread;
1598 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1600 /* thread context */
1601 lu_context_enter(&thread->t_env->le_ctx);
1603 lu_context_exit(&thread->t_env->le_ctx);
1605 lu_context_exit(&req->rq_session);
1606 lu_context_fini(&req->rq_session);
1607 /* don't reset timer for final stage */
1608 if (!req_replay_done(req->rq_export) ||
1609 !lock_replay_done(req->rq_export))
1610 reset_recovery_timer(class_exp2obd(req->rq_export),
1611 AT_OFF ? obd_timeout :
1612 at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1613 ptlrpc_free_clone(req);
1617 static void resume_recovery_timer(struct obd_device *obd)
1619 /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1620 reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1623 static int target_recovery_thread(void *arg)
1625 struct obd_device *obd = arg;
1626 struct ptlrpc_request *req;
1627 struct target_recovery_data *trd = &obd->obd_recovery_data;
1628 struct l_wait_info lwi = { 0 };
1629 unsigned long delta;
1630 unsigned long flags;
1632 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1636 cfs_daemonize("tgt_recov");
1638 SIGNAL_MASK_LOCK(current, flags);
1639 sigfillset(¤t->blocked);
1641 SIGNAL_MASK_UNLOCK(current, flags);
1643 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1647 thread->t_env = &env;
1648 env.le_ctx.lc_thread = thread;
1650 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1652 trd->trd_processing_task = current->pid;
1654 obd->obd_recovering = 1;
1655 complete(&trd->trd_starting);
1657 /* first of all, we have to know the first transno to replay */
1658 obd->obd_abort_recovery = 0;
1659 l_wait_event(obd->obd_next_transno_waitq,
1660 check_for_clients(obd), &lwi);
1662 spin_lock_bh(&obd->obd_processing_task_lock);
1663 target_cancel_recovery_timer(obd);
1664 spin_unlock_bh(&obd->obd_processing_task_lock);
1666 /* If some clients haven't connected in time, evict them */
1667 if (obd->obd_abort_recovery) {
1668 CWARN("Some clients haven't connect in time (%d/%d),"
1669 "evict them\n", obd->obd_connected_clients,
1670 obd->obd_max_recoverable_clients);
1671 obd->obd_abort_recovery = obd->obd_stopping;
1672 class_disconnect_stale_exports(obd, connect_done,
1673 exp_flags_from_obd(obd) |
1674 OBD_OPT_ABORT_RECOV);
1676 /* next stage: replay requests */
1678 obd->obd_req_replaying = 1;
1679 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1680 atomic_read(&obd->obd_req_replay_clients),
1681 obd->obd_next_recovery_transno);
1682 resume_recovery_timer(obd);
1683 while ((req = target_next_replay_req(obd))) {
1684 LASSERT(trd->trd_processing_task == current->pid);
1685 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1686 lustre_msg_get_transno(req->rq_reqmsg),
1687 libcfs_nid2str(req->rq_peer.nid));
1688 handle_recovery_req(thread, req,
1689 trd->trd_recovery_handler);
1690 obd->obd_replayed_requests++;
1691 spin_lock_bh(&obd->obd_processing_task_lock);
1692 obd->obd_next_recovery_transno++;
1693 spin_unlock_bh(&obd->obd_processing_task_lock);
1696 spin_lock_bh(&obd->obd_processing_task_lock);
1697 target_cancel_recovery_timer(obd);
1698 spin_unlock_bh(&obd->obd_processing_task_lock);
1700 /* If some clients haven't replayed requests in time, evict them */
1701 if (obd->obd_abort_recovery) {
1702 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1703 obd->obd_abort_recovery = obd->obd_stopping;
1704 class_disconnect_stale_exports(obd, req_replay_done,
1705 exp_flags_from_obd(obd) |
1706 OBD_OPT_ABORT_RECOV);
1707 abort_req_replay_queue(obd);
1710 /* The second stage: replay locks */
1711 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1712 atomic_read(&obd->obd_lock_replay_clients));
1713 resume_recovery_timer(obd);
1714 while ((req = target_next_replay_lock(obd))) {
1715 LASSERT(trd->trd_processing_task == current->pid);
1716 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1717 libcfs_nid2str(req->rq_peer.nid));
1718 handle_recovery_req(thread, req,
1719 trd->trd_recovery_handler);
1720 obd->obd_replayed_locks++;
1723 spin_lock_bh(&obd->obd_processing_task_lock);
1724 target_cancel_recovery_timer(obd);
1725 spin_unlock_bh(&obd->obd_processing_task_lock);
1726 /* If some clients haven't replayed requests in time, evict them */
1727 if (obd->obd_abort_recovery) {
1729 CERROR("lock replay timed out, aborting ...\n");
1730 obd->obd_abort_recovery = obd->obd_stopping;
1731 stale = class_disconnect_stale_exports(obd, lock_replay_done,
1732 exp_flags_from_obd(obd) |
1733 OBD_OPT_ABORT_RECOV);
1734 abort_lock_replay_queue(obd);
1737 /* We drop recoverying flag to forward all new requests
1738 * to regular mds_handle() since now */
1739 spin_lock_bh(&obd->obd_processing_task_lock);
1740 obd->obd_recovering = obd->obd_abort_recovery = 0;
1741 spin_unlock_bh(&obd->obd_processing_task_lock);
1742 /* The third stage: reply on final pings */
1743 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1744 while ((req = target_next_final_ping(obd))) {
1745 LASSERT(trd->trd_processing_task == current->pid);
1746 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1747 libcfs_nid2str(req->rq_peer.nid));
1748 handle_recovery_req(thread, req,
1749 trd->trd_recovery_handler);
1752 delta = (jiffies - delta) / HZ;
1753 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1754 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1755 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1756 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1757 if (delta > obd_timeout * 2) {
1758 CWARN("too long recovery - read logs\n");
1759 libcfs_debug_dumplog();
1762 target_finish_recovery(obd);
1764 lu_context_fini(&env.le_ctx);
1765 trd->trd_processing_task = 0;
1766 complete(&trd->trd_finishing);
1770 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1773 struct target_recovery_data *trd = &obd->obd_recovery_data;
1775 memset(trd, 0, sizeof(*trd));
1776 init_completion(&trd->trd_starting);
1777 init_completion(&trd->trd_finishing);
1778 trd->trd_recovery_handler = handler;
1780 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1781 wait_for_completion(&trd->trd_starting);
1782 LASSERT(obd->obd_recovering != 0);
1789 void target_stop_recovery_thread(struct obd_device *obd)
1791 spin_lock_bh(&obd->obd_processing_task_lock);
1792 if (obd->obd_recovery_data.trd_processing_task > 0) {
1793 struct target_recovery_data *trd = &obd->obd_recovery_data;
1794 CERROR("%s: Aborting recovery\n", obd->obd_name);
1795 obd->obd_abort_recovery = 1;
1796 wake_up(&obd->obd_next_transno_waitq);
1797 spin_unlock_bh(&obd->obd_processing_task_lock);
1798 wait_for_completion(&trd->trd_finishing);
1800 spin_unlock_bh(&obd->obd_processing_task_lock);
1804 void target_recovery_fini(struct obd_device *obd)
1806 class_disconnect_exports(obd);
1807 target_stop_recovery_thread(obd);
1808 target_cleanup_recovery(obd);
1810 EXPORT_SYMBOL(target_recovery_fini);
1812 static void target_recovery_expired(unsigned long castmeharder)
1814 struct obd_device *obd = (struct obd_device *)castmeharder;
1815 LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1816 "after %lds (%d clients did)\n",
1817 obd->obd_name, obd->obd_recoverable_clients,
1818 cfs_time_current_sec()- obd->obd_recovery_start,
1819 obd->obd_connected_clients);
1820 spin_lock_bh(&obd->obd_processing_task_lock);
1821 if (obd->obd_recovering)
1822 obd->obd_abort_recovery = 1;
1823 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1824 spin_unlock_bh(&obd->obd_processing_task_lock);
1827 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1829 if (obd->obd_max_recoverable_clients == 0)
1832 CWARN("RECOVERY: service %s, %d recoverable clients, "
1833 "last_transno "LPU64"\n", obd->obd_name,
1834 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1835 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1836 obd->obd_recovery_start = 0;
1837 obd->obd_recovery_end = 0;
1838 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1839 /* bz13079: this should be set to desired value for ost but not for mds */
1840 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1841 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1842 target_start_recovery_thread(obd, handler);
1844 EXPORT_SYMBOL(target_recovery_init);
1848 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1850 struct obd_export *exp = req->rq_export;
1851 LASSERT(exp != NULL);
1852 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1853 /* client declares he's ready to replay locks */
1854 spin_lock_bh(&obd->obd_processing_task_lock);
1855 if (exp->exp_req_replay_needed) {
1856 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1857 spin_lock(&exp->exp_lock);
1858 exp->exp_req_replay_needed = 0;
1859 spin_unlock(&exp->exp_lock);
1860 atomic_dec(&obd->obd_req_replay_clients);
1861 LASSERT(obd->obd_recoverable_clients > 0);
1862 obd->obd_recoverable_clients--;
1863 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1864 CDEBUG(D_HA, "all clients have replayed reqs\n");
1865 wake_up(&obd->obd_next_transno_waitq);
1867 spin_unlock_bh(&obd->obd_processing_task_lock);
1869 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1870 /* client declares he's ready to complete recovery
1871 * so, we put the request on th final queue */
1872 spin_lock_bh(&obd->obd_processing_task_lock);
1873 if (exp->exp_lock_replay_needed) {
1874 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1875 spin_lock(&exp->exp_lock);
1876 exp->exp_lock_replay_needed = 0;
1877 spin_unlock(&exp->exp_lock);
1878 atomic_dec(&obd->obd_lock_replay_clients);
1879 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1880 CDEBUG(D_HA, "all clients have replayed locks\n");
1881 wake_up(&obd->obd_next_transno_waitq);
1883 spin_unlock_bh(&obd->obd_processing_task_lock);
1889 int target_queue_recovery_request(struct ptlrpc_request *req,
1890 struct obd_device *obd)
1892 struct list_head *tmp;
1894 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1898 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1899 /* Processing the queue right now, don't re-add. */
1903 target_process_req_flags(obd, req);
1905 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1906 /* client declares he's ready to complete recovery
1907 * so, we put the request on th final queue */
1908 req = ptlrpc_clone_req(req);
1911 DEBUG_REQ(D_HA, req, "queue final req");
1912 spin_lock_bh(&obd->obd_processing_task_lock);
1913 if (obd->obd_recovering)
1914 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1916 spin_unlock_bh(&obd->obd_processing_task_lock);
1917 ptlrpc_free_clone(req);
1918 if (obd->obd_stopping) {
1924 spin_unlock_bh(&obd->obd_processing_task_lock);
1927 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1928 /* client declares he's ready to replay locks */
1929 req = ptlrpc_clone_req(req);
1932 DEBUG_REQ(D_HA, req, "queue lock replay req");
1933 spin_lock_bh(&obd->obd_processing_task_lock);
1934 LASSERT(obd->obd_recovering);
1935 /* usually due to recovery abort */
1936 if (!req->rq_export->exp_in_recovery) {
1937 spin_unlock_bh(&obd->obd_processing_task_lock);
1938 ptlrpc_free_clone(req);
1941 LASSERT(req->rq_export->exp_lock_replay_needed);
1942 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1943 spin_unlock_bh(&obd->obd_processing_task_lock);
1944 wake_up(&obd->obd_next_transno_waitq);
1948 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1949 * (i.e. buflens etc are in my own byte order), but type-dependent
1950 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1953 CFS_INIT_LIST_HEAD(&req->rq_list);
1954 DEBUG_REQ(D_HA, req, "not queueing");
1958 spin_lock_bh(&obd->obd_processing_task_lock);
1960 /* If we're processing the queue, we want don't want to queue this
1963 * Also, if this request has a transno less than the one we're waiting
1964 * for, we should process it now. It could (and currently always will)
1965 * be an open request for a descriptor that was opened some time ago.
1967 * Also, a resent, replayed request that has already been
1968 * handled will pass through here and be processed immediately.
1970 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1971 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1972 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1973 /* Processing the queue right now, don't re-add. */
1974 LASSERT(list_empty(&req->rq_list));
1975 spin_unlock_bh(&obd->obd_processing_task_lock);
1978 spin_unlock_bh(&obd->obd_processing_task_lock);
1980 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
1983 req = ptlrpc_clone_req(req);
1987 spin_lock_bh(&obd->obd_processing_task_lock);
1988 LASSERT(obd->obd_recovering);
1989 if (!req->rq_export->exp_in_recovery) {
1990 spin_unlock_bh(&obd->obd_processing_task_lock);
1991 ptlrpc_free_clone(req);
1994 LASSERT(req->rq_export->exp_req_replay_needed);
1996 if (target_exp_enqueue_req_replay(req)) {
1997 spin_unlock_bh(&obd->obd_processing_task_lock);
1998 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1999 ptlrpc_free_clone(req);
2004 list_for_each(tmp, &obd->obd_req_replay_queue) {
2005 struct ptlrpc_request *reqiter =
2006 list_entry(tmp, struct ptlrpc_request, rq_list);
2008 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2009 list_add_tail(&req->rq_list, &reqiter->rq_list);
2014 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2016 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2017 "has been claimed by another client");
2018 spin_unlock_bh(&obd->obd_processing_task_lock);
2019 target_exp_dequeue_req_replay(req);
2020 ptlrpc_free_clone(req);
2026 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2028 obd->obd_requests_queued_for_recovery++;
2029 wake_up(&obd->obd_next_transno_waitq);
2030 spin_unlock_bh(&obd->obd_processing_task_lock);
2034 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2036 return req->rq_export->exp_obd;
2039 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2041 LASSERT(exp != NULL);
2042 return &exp->exp_obd->obd_namespace->ns_pool;
2046 * Packs current SLV and Limit into \a req.
2048 int target_pack_pool_reply(struct ptlrpc_request *req)
2050 struct obd_device *obd;
2054 * Check that we still have all structures alive as this may
2055 * be some late rpc in shutdown time.
2057 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2058 !exp_connect_lru_resize(req->rq_export))) {
2059 lustre_msg_set_slv(req->rq_repmsg, 0);
2060 lustre_msg_set_limit(req->rq_repmsg, 0);
2065 * OBD is alive here as export is alive, which we checked above.
2067 obd = req->rq_export->exp_obd;
2069 read_lock(&obd->obd_pool_lock);
2070 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2071 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2072 read_unlock(&obd->obd_pool_lock);
2077 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2079 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2080 DEBUG_REQ(D_ERROR, req, "dropping reply");
2085 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2086 req->rq_status = rc;
2087 return (ptlrpc_send_error(req, 1));
2089 DEBUG_REQ(D_NET, req, "sending reply");
2092 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2095 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2098 struct ptlrpc_reply_state *rs;
2099 struct obd_device *obd;
2100 struct obd_export *exp;
2101 struct ptlrpc_service *svc;
2104 if (req->rq_no_reply) {
2109 svc = req->rq_rqbd->rqbd_service;
2110 rs = req->rq_reply_state;
2111 if (rs == NULL || !rs->rs_difficult) {
2113 target_send_reply_msg (req, rc, fail_id);
2118 /* must be an export if locks saved */
2119 LASSERT (req->rq_export != NULL);
2120 /* req/reply consistent */
2121 LASSERT (rs->rs_service == svc);
2124 LASSERT (!rs->rs_scheduled);
2125 LASSERT (!rs->rs_scheduled_ever);
2126 LASSERT (!rs->rs_handled);
2127 LASSERT (!rs->rs_on_net);
2128 LASSERT (rs->rs_export == NULL);
2129 LASSERT (list_empty(&rs->rs_obd_list));
2130 LASSERT (list_empty(&rs->rs_exp_list));
2132 exp = class_export_get (req->rq_export);
2135 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2136 rs->rs_scheduled = 1;
2138 rs->rs_xid = req->rq_xid;
2139 rs->rs_transno = req->rq_transno;
2140 rs->rs_export = exp;
2142 spin_lock(&obd->obd_uncommitted_replies_lock);
2144 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2145 rs->rs_transno, obd->obd_last_committed);
2146 if (rs->rs_transno > obd->obd_last_committed) {
2147 /* not committed already */
2148 list_add_tail (&rs->rs_obd_list,
2149 &obd->obd_uncommitted_replies);
2152 spin_unlock (&obd->obd_uncommitted_replies_lock);
2153 spin_lock (&exp->exp_lock);
2155 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
2157 spin_unlock(&exp->exp_lock);
2159 netrc = target_send_reply_msg (req, rc, fail_id);
2161 spin_lock(&svc->srv_lock);
2163 atomic_inc(&svc->srv_n_difficult_replies);
2166 /* error sending: reply is off the net. Also we need +1
2167 * reply ref until ptlrpc_server_handle_reply() is done
2168 * with the reply state (if the send was successful, there
2169 * would have been +1 ref for the net, which
2170 * reply_out_callback leaves alone) */
2172 ptlrpc_rs_addref(rs);
2173 atomic_inc (&svc->srv_outstanding_replies);
2176 spin_lock(&rs->rs_lock);
2177 if (rs->rs_transno <= obd->obd_last_committed ||
2178 (!rs->rs_on_net && !rs->rs_no_ack) ||
2179 list_empty(&rs->rs_exp_list) || /* completed already */
2180 list_empty(&rs->rs_obd_list)) {
2181 CDEBUG(D_HA, "Schedule reply immediately\n");
2182 ptlrpc_dispatch_difficult_reply(rs);
2184 list_add (&rs->rs_list, &svc->srv_active_replies);
2185 rs->rs_scheduled = 0; /* allow notifier to schedule */
2187 spin_unlock(&rs->rs_lock);
2188 spin_unlock(&svc->srv_lock);
2192 int target_handle_ping(struct ptlrpc_request *req)
2194 obd_ping(req->rq_export);
2195 return req_capsule_server_pack(&req->rq_pill);
2198 void target_committed_to_req(struct ptlrpc_request *req)
2200 struct obd_device *obd;
2202 if (req == NULL || req->rq_export == NULL)
2205 obd = req->rq_export->exp_obd;
2209 if (!obd->obd_no_transno && req->rq_repmsg != NULL)
2210 lustre_msg_set_last_committed(req->rq_repmsg,
2211 obd->obd_last_committed);
2213 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2214 "%d)", obd->obd_no_transno, req->rq_repmsg == NULL);
2216 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2217 obd->obd_last_committed, req->rq_transno, req->rq_xid);
2220 EXPORT_SYMBOL(target_committed_to_req);
2222 int target_handle_qc_callback(struct ptlrpc_request *req)
2224 struct obd_quotactl *oqctl;
2225 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2227 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2228 if (oqctl == NULL) {
2229 CERROR("Can't unpack obd_quotactl\n");
2233 cli->cl_qchk_stat = oqctl->qc_stat;
2238 #ifdef HAVE_QUOTA_SUPPORT
2239 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2242 struct obd_device *obd = req->rq_export->exp_obd;
2243 struct obd_device *master_obd;
2244 struct obd_device_target *obt;
2245 struct lustre_quota_ctxt *qctxt;
2246 struct qunit_data *qdata = NULL;
2250 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2253 rc = req_capsule_server_pack(&req->rq_pill);
2255 CERROR("packing reply failed!: rc = %d\n", rc);
2259 LASSERT(req->rq_export);
2261 OBD_ALLOC(qdata, sizeof(struct qunit_data));
2264 rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2266 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2270 /* we use the observer */
2271 if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2272 CERROR("Can't find the observer, it is recovering\n");
2273 req->rq_status = -EAGAIN;
2274 GOTO(send_reply, rc = -EAGAIN);
2277 master_obd = obd->obd_observer->obd_observer;
2278 obt = &master_obd->u.obt;
2279 qctxt = &obt->obt_qctxt;
2281 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2282 /* quota_type has not been processed yet, return EAGAIN
2283 * until we know whether or not quotas are supposed to
2285 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2287 req->rq_status = -EAGAIN;
2288 rc = ptlrpc_reply(req);
2292 down_read(&obt->obt_rwsem);
2293 if (qctxt->lqc_lqs_hash == NULL) {
2294 up_read(&obt->obt_rwsem);
2295 /* quota_type has not been processed yet, return EAGAIN
2296 * until we know whether or not quotas are supposed to
2298 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2300 req->rq_status = -EAGAIN;
2301 rc = ptlrpc_reply(req);
2305 LASSERT(qctxt->lqc_handler);
2306 rc = qctxt->lqc_handler(master_obd, qdata,
2307 lustre_msg_get_opc(req->rq_reqmsg));
2308 up_read(&obt->obt_rwsem);
2309 if (rc && rc != -EDQUOT)
2310 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2311 "dqacq failed! (rc:%d)\n", rc);
2312 req->rq_status = rc;
2314 /* there are three forms of qunit(historic causes), so we need to
2315 * adjust the same form to different forms slaves needed */
2316 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2318 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2322 /* Block the quota req. b=14840 */
2323 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2325 rc = ptlrpc_reply(req);
2327 OBD_FREE(qdata, sizeof(struct qunit_data));
2331 #endif /* !__KERNEL__ */
2333 #endif /* HAVE_QUOTA_SUPPORT */
2335 ldlm_mode_t lck_compat_array[] = {
2336 [LCK_EX] LCK_COMPAT_EX,
2337 [LCK_PW] LCK_COMPAT_PW,
2338 [LCK_PR] LCK_COMPAT_PR,
2339 [LCK_CW] LCK_COMPAT_CW,
2340 [LCK_CR] LCK_COMPAT_CR,
2341 [LCK_NL] LCK_COMPAT_NL,
2342 [LCK_GROUP] LCK_COMPAT_GROUP,
2343 [LCK_COS] LCK_COMPAT_COS,
2347 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2348 * not escape to the user level.
2350 int ldlm_error2errno(ldlm_error_t error)
2358 case ELDLM_LOCK_CHANGED:
2361 case ELDLM_LOCK_ABORTED:
2364 case ELDLM_LOCK_REPLACED:
2367 case ELDLM_NO_LOCK_DATA:
2370 case ELDLM_NAMESPACE_EXISTS:
2373 case ELDLM_BAD_NAMESPACE:
2377 if (((int)error) < 0) /* cast to signed type */
2378 result = error; /* as ldlm_error_t can be unsigned */
2380 CERROR("Invalid DLM result code: %i\n", error);
2386 EXPORT_SYMBOL(ldlm_error2errno);
2389 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2391 ldlm_error_t ldlm_errno2error(int err_no)
2400 error = ELDLM_LOCK_CHANGED;
2403 error = ELDLM_LOCK_ABORTED;
2406 error = ELDLM_LOCK_REPLACED;
2409 error = ELDLM_NO_LOCK_DATA;
2412 error = ELDLM_NAMESPACE_EXISTS;
2415 error = ELDLM_BAD_NAMESPACE;
2422 EXPORT_SYMBOL(ldlm_errno2error);