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");
416 * Deadlock case - bug 18380
418 up_write(&cli->cl_sem);
419 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
420 LDLM_NAMESPACE_CLIENT,
421 LDLM_NAMESPACE_GREEDY);
422 down_write(&cli->cl_sem);
423 if (obd->obd_namespace == NULL)
424 GOTO(out_disco, rc = -ENOMEM);
426 imp->imp_dlm_handle = conn;
427 rc = ptlrpc_init_import(imp);
431 ocd = &imp->imp_connect_data;
434 imp->imp_connect_flags_orig = data->ocd_connect_flags;
437 rc = ptlrpc_connect_import(imp, NULL);
439 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
442 LASSERT((*exp)->exp_connection);
445 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
446 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
447 data->ocd_connect_flags, ocd->ocd_connect_flags);
448 data->ocd_connect_flags = ocd->ocd_connect_flags;
451 ptlrpc_pinger_add_import(imp);
457 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
458 to_be_freed = obd->obd_namespace;
459 obd->obd_namespace = NULL;
461 cli->cl_conn_count--;
462 class_disconnect(*exp);
466 up_write(&cli->cl_sem);
468 ldlm_namespace_free_post(to_be_freed);
473 int client_disconnect_export(struct obd_export *exp)
475 struct obd_device *obd = class_exp2obd(exp);
476 struct client_obd *cli;
477 struct obd_import *imp;
479 struct ldlm_namespace *to_be_freed = NULL;
483 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
484 exp, exp ? exp->exp_handle.h_cookie : -1);
489 imp = cli->cl_import;
491 down_write(&cli->cl_sem);
492 if (!cli->cl_conn_count) {
493 CERROR("disconnecting disconnected device (%s)\n",
495 GOTO(out_sem, rc = -EINVAL);
498 cli->cl_conn_count--;
499 if (cli->cl_conn_count)
500 GOTO(out_no_disconnect, rc = 0);
502 /* Mark import deactivated now, so we don't try to reconnect if any
503 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
504 * fully deactivate the import, or that would drop all requests. */
505 spin_lock(&imp->imp_lock);
506 imp->imp_deactive = 1;
507 spin_unlock(&imp->imp_lock);
509 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
510 * delete it regardless. (It's safe to delete an import that was
512 (void)ptlrpc_pinger_del_import(imp);
514 if (obd->obd_namespace != NULL) {
515 /* obd_force == local only */
516 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
517 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
519 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
520 to_be_freed = obd->obd_namespace;
524 * there's no necessary to hold sem during diconnecting an import,
525 * and actually it may cause deadlock in gss.
527 up_write(&cli->cl_sem);
528 rc = ptlrpc_disconnect_import(imp, 0);
529 down_write(&cli->cl_sem);
531 ptlrpc_invalidate_import(imp);
532 /* set obd_namespace to NULL only after invalidate, because we can have
533 * some connect requests in flight, and his need store a connect flags
534 * in obd_namespace. bug 14260 */
535 obd->obd_namespace = NULL;
537 if (imp->imp_rq_pool) {
538 ptlrpc_free_rq_pool(imp->imp_rq_pool);
539 imp->imp_rq_pool = NULL;
542 cli->cl_import = NULL;
546 err = class_disconnect(exp);
550 up_write(&cli->cl_sem);
552 ldlm_namespace_free_post(to_be_freed);
557 /* --------------------------------------------------------------------------
558 * from old lib/target.c
559 * -------------------------------------------------------------------------- */
561 static int target_handle_reconnect(struct lustre_handle *conn,
562 struct obd_export *exp,
563 struct obd_uuid *cluuid)
566 if (exp->exp_connection && exp->exp_imp_reverse) {
567 struct lustre_handle *hdl;
568 hdl = &exp->exp_imp_reverse->imp_remote_handle;
569 /* Might be a re-connect after a partition. */
570 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
571 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
573 conn->cookie = exp->exp_handle.h_cookie;
574 /* target_handle_connect() treats EALREADY and
575 * -EALREADY differently. EALREADY means we are
576 * doing a valid reconnect from the same client. */
579 CERROR("%s reconnecting from %s, "
580 "handle mismatch (ours "LPX64", theirs "
581 LPX64")\n", cluuid->uuid,
582 exp->exp_connection->c_remote_uuid.uuid,
583 hdl->cookie, conn->cookie);
584 memset(conn, 0, sizeof *conn);
585 /* target_handle_connect() treats EALREADY and
586 * -EALREADY differently. -EALREADY is an error
587 * (same UUID, different handle). */
592 conn->cookie = exp->exp_handle.h_cookie;
593 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
594 cluuid->uuid, exp, conn->cookie);
598 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
601 struct obd_export *exp = cb_data;
603 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
604 obd->obd_name, exp->exp_client_uuid.uuid);
606 spin_lock(&exp->exp_lock);
607 exp->exp_need_sync = 0;
608 spin_unlock(&exp->exp_lock);
610 EXPORT_SYMBOL(target_client_add_cb);
612 target_start_and_reset_recovery_timer(struct obd_device *obd,
613 struct ptlrpc_request *req,
616 int target_handle_connect(struct ptlrpc_request *req)
618 struct obd_device *target, *targref = NULL;
619 struct obd_export *export = NULL;
620 struct obd_import *revimp;
621 struct lustre_handle conn;
622 struct lustre_handle *tmp;
623 struct obd_uuid tgtuuid;
624 struct obd_uuid cluuid;
625 struct obd_uuid remote_uuid;
629 struct obd_connect_data *data, *tmpdata;
630 lnet_nid_t *client_nid = NULL;
633 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
635 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
637 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
638 GOTO(out, rc = -EINVAL);
641 obd_str2uuid(&tgtuuid, str);
642 target = class_uuid2obd(&tgtuuid);
644 target = class_name2obd(str);
646 if (!target || target->obd_stopping || !target->obd_set_up) {
647 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
648 " for connect (%s)\n", str,
649 !target ? "no target" :
650 (target->obd_stopping ? "stopping" :
652 GOTO(out, rc = -ENODEV);
655 if (target->obd_no_conn) {
656 LCONSOLE_WARN("%s: temporarily refusing client connection "
657 "from %s\n", target->obd_name,
658 libcfs_nid2str(req->rq_peer.nid));
659 GOTO(out, rc = -EAGAIN);
662 /* Make sure the target isn't cleaned up while we're here. Yes,
663 there's still a race between the above check and our incref here.
664 Really, class_uuid2obd should take the ref. */
665 targref = class_incref(target, __FUNCTION__, cfs_current());
668 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
670 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
671 GOTO(out, rc = -EINVAL);
674 obd_str2uuid(&cluuid, str);
676 /* XXX extract a nettype and format accordingly */
677 switch (sizeof(lnet_nid_t)) {
678 /* NB the casts only avoid compiler warnings */
680 snprintf(remote_uuid.uuid, sizeof remote_uuid,
681 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
684 snprintf(remote_uuid.uuid, sizeof remote_uuid,
685 "NET_%x_UUID", (__u32)req->rq_peer.nid);
691 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
693 GOTO(out, rc = -EPROTO);
697 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
699 GOTO(out, rc = -EPROTO);
701 rc = req_capsule_server_pack(&req->rq_pill);
705 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
707 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
708 "libclient connection attempt");
709 GOTO(out, rc = -EPROTO);
710 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
711 LUSTRE_VERSION_ALLOWED_OFFSET ||
712 data->ocd_version > LUSTRE_VERSION_CODE +
713 LUSTRE_VERSION_ALLOWED_OFFSET) {
714 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
715 "libclient connection attempt",
716 data->ocd_version < LUSTRE_VERSION_CODE ?
718 OBD_OCD_VERSION_MAJOR(data->ocd_version),
719 OBD_OCD_VERSION_MINOR(data->ocd_version),
720 OBD_OCD_VERSION_PATCH(data->ocd_version),
721 OBD_OCD_VERSION_FIX(data->ocd_version));
722 data = req_capsule_server_sized_get(&req->rq_pill,
724 offsetof(typeof(*data), ocd_version) +
725 sizeof(data->ocd_version));
727 data->ocd_connect_flags = OBD_CONNECT_VERSION;
728 data->ocd_version = LUSTRE_VERSION_CODE;
730 GOTO(out, rc = -EPROTO);
734 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
735 (data->ocd_connect_flags & OBD_CONNECT_MDS))
738 /* lctl gets a backstage, all-access pass. */
739 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
740 goto dont_check_exports;
742 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
746 /* we've found an export in the hash */
747 if (export->exp_connecting) { /* bug 9635, et. al. */
748 CWARN("%s: exp %p already connecting\n",
749 export->exp_obd->obd_name, export);
750 class_export_put(export);
753 } else if (mds_conn && export->exp_connection) {
754 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
755 /* mds reconnected after failover */
756 CWARN("%s: received MDS connection from NID %s,"
757 " removing former export from NID %s\n",
758 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
759 libcfs_nid2str(export->exp_connection->c_peer.nid));
761 /* new mds connection from the same nid */
762 CWARN("%s: received new MDS connection from NID %s,"
763 " removing former export from same NID\n",
764 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
765 class_fail_export(export);
766 class_export_put(export);
769 } else if (export->exp_connection != NULL &&
770 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
771 /* in mds failover we have static uuid but nid can be
773 CWARN("%s: cookie %s seen on new NID %s when "
774 "existing NID %s is already connected\n",
775 target->obd_name, cluuid.uuid,
776 libcfs_nid2str(req->rq_peer.nid),
777 libcfs_nid2str(export->exp_connection->c_peer.nid));
779 class_export_put(export);
782 spin_lock(&export->exp_lock);
783 export->exp_connecting = 1;
784 spin_unlock(&export->exp_lock);
785 class_export_put(export);
786 LASSERT(export->exp_obd == target);
788 rc = target_handle_reconnect(&conn, export, &cluuid);
791 /* If we found an export, we already unlocked. */
794 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
795 } else if (req->rq_export == NULL &&
796 atomic_read(&export->exp_rpc_count) > 0) {
797 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
798 target->obd_name, cluuid.uuid,
799 libcfs_nid2str(req->rq_peer.nid),
800 export, atomic_read(&export->exp_refcount));
801 GOTO(out, rc = -EBUSY);
802 } else if (req->rq_export != NULL &&
803 (atomic_read(&export->exp_rpc_count) > 1)) {
804 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
805 target->obd_name, cluuid.uuid,
806 libcfs_nid2str(req->rq_peer.nid),
807 export, atomic_read(&export->exp_rpc_count));
808 GOTO(out, rc = -EBUSY);
809 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
810 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
811 "cookies not random?\n", target->obd_name,
812 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
813 GOTO(out, rc = -EALREADY);
815 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
822 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
823 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
824 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
825 export, (long)cfs_time_current_sec(),
826 export ? (long)export->exp_last_request_time : 0);
828 /* Tell the client if we're in recovery. */
829 if (target->obd_recovering) {
830 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
831 /* If this is the first time a client connects,
832 reset the recovery timer */
834 target_start_and_reset_recovery_timer(target, req,
838 /* We want to handle EALREADY but *not* -EALREADY from
839 * target_handle_reconnect(), return reconnection state in a flag */
840 if (rc == EALREADY) {
841 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
847 /* Tell the client if we support replayable requests */
848 if (target->obd_replayable)
849 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
850 client_nid = &req->rq_peer.nid;
852 if (export == NULL) {
853 if (target->obd_recovering) {
856 t = cfs_timer_deadline(&target->obd_recovery_timer);
857 t = cfs_time_sub(t, cfs_time_current());
858 CERROR("%s: denying connection for new client %s (%s): "
859 "%d clients in recovery for "CFS_TIME_T"s\n",
861 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
862 target->obd_recoverable_clients,
863 cfs_duration_sec(t));
867 rc = obd_connect(req->rq_svc_thread->t_env,
868 &export, target, &cluuid, data,
871 conn.cookie = export->exp_handle.h_cookie;
874 rc = obd_reconnect(req->rq_svc_thread->t_env,
875 export, target, &cluuid, data, client_nid);
877 /* prevous done via class_conn2export */
878 class_export_get(export);
882 /* Return only the parts of obd_connect_data that we understand, so the
883 * client knows that we don't understand the rest. */
885 tmpdata = req_capsule_server_get(&req->rq_pill,
887 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
891 /* If all else goes well, this is our RPC return code. */
894 lustre_msg_set_handle(req->rq_repmsg, &conn);
896 /* If the client and the server are the same node, we will already
897 * have an export that really points to the client's DLM export,
898 * because we have a shared handles table.
900 * XXX this will go away when shaver stops sending the "connect" handle
901 * in the real "remote handle" field of the request --phik 24 Apr 2003
903 if (req->rq_export != NULL)
904 class_export_put(req->rq_export);
906 req->rq_export = export;
908 spin_lock(&export->exp_lock);
909 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
910 spin_unlock(&export->exp_lock);
911 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
912 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
913 export->exp_conn_cnt,
914 lustre_msg_get_conn_cnt(req->rq_reqmsg));
916 GOTO(out, rc = -EALREADY);
918 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
920 /* request from liblustre? Don't evict it for not pinging. */
921 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
922 export->exp_libclient = 1;
923 spin_unlock(&export->exp_lock);
925 spin_lock(&target->obd_dev_lock);
926 list_del_init(&export->exp_obd_chain_timed);
927 spin_unlock(&target->obd_dev_lock);
929 spin_unlock(&export->exp_lock);
932 if (export->exp_connection != NULL)
933 ptlrpc_connection_put(export->exp_connection);
934 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
937 if (hlist_unhashed(&export->exp_nid_hash)) {
938 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
939 &export->exp_connection->c_peer.nid,
940 &export->exp_nid_hash);
943 spin_lock_bh(&target->obd_processing_task_lock);
944 if (target->obd_recovering && !export->exp_in_recovery) {
945 spin_lock(&export->exp_lock);
946 export->exp_in_recovery = 1;
947 export->exp_req_replay_needed = 1;
948 export->exp_lock_replay_needed = 1;
949 spin_unlock(&export->exp_lock);
950 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
951 && (data->ocd_transno == 0))
952 CWARN("Connect with zero transno!\n");
954 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
955 && data->ocd_transno < target->obd_next_recovery_transno)
956 target->obd_next_recovery_transno = data->ocd_transno;
957 target->obd_connected_clients++;
958 /* each connected client is counted as recoverable */
959 target->obd_recoverable_clients++;
960 atomic_inc(&target->obd_req_replay_clients);
961 atomic_inc(&target->obd_lock_replay_clients);
962 if (target->obd_connected_clients ==
963 target->obd_max_recoverable_clients)
964 wake_up(&target->obd_next_transno_waitq);
966 spin_unlock_bh(&target->obd_processing_task_lock);
967 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
970 if (export->exp_imp_reverse != NULL) {
971 /* destroyed import can be still referenced in ctxt */
972 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
973 KEY_REVIMP_UPD, 0, NULL, NULL);
975 /* in some recovery senarios, previous ctx init rpc handled
976 * in sptlrpc_target_export_check() might be used to install
977 * a reverse ctx in this reverse import, and later OBD_CONNECT
978 * using the same gss ctx could reach here and following new
979 * reverse import. note all reverse ctx in new/old import are
980 * actually based on the same gss ctx. so we invalidate ctx
981 * here before destroy import, otherwise flush old import will
982 * lead to remote reverse ctx be destroied, thus the reverse
983 * ctx of new import will lost its peer.
984 * there might be a better way to deal with this???
986 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
988 destroy_import(export->exp_imp_reverse);
991 /* for the rest part, we return -ENOTCONN in case of errors
992 * in order to let client initialize connection again.
994 revimp = export->exp_imp_reverse = class_new_import(target);
996 CERROR("fail to alloc new reverse import.\n");
997 GOTO(out, rc = -ENOTCONN);
1000 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1001 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1002 revimp->imp_remote_handle = conn;
1003 revimp->imp_dlm_fake = 1;
1004 revimp->imp_state = LUSTRE_IMP_FULL;
1006 /* unknown versions will be caught in
1007 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1008 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1010 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1011 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1012 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1014 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1016 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1018 CERROR("Failed to get sec for reverse import: %d\n", rc);
1019 export->exp_imp_reverse = NULL;
1020 class_destroy_import(revimp);
1023 class_import_put(revimp);
1026 spin_lock(&export->exp_lock);
1027 export->exp_connecting = 0;
1028 spin_unlock(&export->exp_lock);
1031 class_decref(targref, __FUNCTION__, cfs_current());
1033 req->rq_status = rc;
1037 int target_handle_disconnect(struct ptlrpc_request *req)
1042 rc = req_capsule_server_pack(&req->rq_pill);
1046 /* keep the rq_export around so we can send the reply */
1047 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1052 void target_destroy_export(struct obd_export *exp)
1054 /* exports created from last_rcvd data, and "fake"
1055 exports created by lctl don't have an import */
1056 if (exp->exp_imp_reverse != NULL)
1057 destroy_import(exp->exp_imp_reverse);
1059 /* We cancel locks at disconnect time, but this will catch any locks
1060 * granted in a race with recovery-induced disconnect. */
1061 if (exp->exp_obd->obd_namespace != NULL)
1062 ldlm_cancel_locks_for_export(exp);
1066 * Recovery functions
1069 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1071 struct ptlrpc_request *copy_req;
1072 struct lustre_msg *copy_reqmsg;
1073 struct ptlrpc_user_desc *udesc = NULL;
1075 OBD_ALLOC_PTR(copy_req);
1078 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1080 OBD_FREE_PTR(copy_req);
1084 if (orig_req->rq_user_desc) {
1085 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1087 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1089 OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1090 OBD_FREE_PTR(copy_req);
1093 memcpy(udesc, orig_req->rq_user_desc,
1094 sptlrpc_user_desc_size(ngroups));
1097 *copy_req = *orig_req;
1098 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1099 copy_req->rq_reqmsg = copy_reqmsg;
1100 copy_req->rq_user_desc = udesc;
1102 class_export_rpc_get(copy_req->rq_export);
1103 CFS_INIT_LIST_HEAD(©_req->rq_list);
1104 CFS_INIT_LIST_HEAD(©_req->rq_replay_list);
1105 sptlrpc_svc_ctx_addref(copy_req);
1107 if (copy_req->rq_reply_state) {
1108 /* the copied req takes over the reply state */
1109 orig_req->rq_reply_state = NULL;
1110 /* to catch further access */
1111 orig_req->rq_repmsg = NULL;
1112 orig_req->rq_replen = 0;
1118 void ptlrpc_free_clone(struct ptlrpc_request *req)
1120 LASSERT(list_empty(&req->rq_replay_list));
1122 ptlrpc_req_drop_rs(req);
1123 sptlrpc_svc_ctx_decref(req);
1124 class_export_rpc_put(req->rq_export);
1125 list_del_init(&req->rq_list);
1127 if (req->rq_user_desc) {
1128 int ngroups = req->rq_user_desc->pud_ngroups;
1129 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1131 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1135 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1137 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1138 struct obd_export *exp = req->rq_export;
1139 struct ptlrpc_request *reqiter;
1144 spin_lock(&exp->exp_lock);
1145 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1147 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1154 /* we expect it with RESENT and REPLAY flags */
1155 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1156 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1157 CERROR("invalid flags %x of resent replay\n",
1158 lustre_msg_get_flags(req->rq_reqmsg));
1160 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1163 spin_unlock(&exp->exp_lock);
1167 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1169 LASSERT(!list_empty(&req->rq_replay_list));
1170 LASSERT(req->rq_export);
1172 spin_lock(&req->rq_export->exp_lock);
1173 list_del_init(&req->rq_replay_list);
1174 spin_unlock(&req->rq_export->exp_lock);
1178 static void target_finish_recovery(struct obd_device *obd)
1181 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1184 ldlm_reprocess_all_ns(obd->obd_namespace);
1185 spin_lock_bh(&obd->obd_processing_task_lock);
1186 if (list_empty(&obd->obd_req_replay_queue) &&
1187 list_empty(&obd->obd_lock_replay_queue) &&
1188 list_empty(&obd->obd_final_req_queue)) {
1189 obd->obd_processing_task = 0;
1191 CERROR("%s: Recovery queues ( %s%s%s) are empty\n",
1193 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1194 list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1195 list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1196 spin_unlock_bh(&obd->obd_processing_task_lock);
1199 spin_unlock_bh(&obd->obd_processing_task_lock);
1201 /* when recovery finished, cleanup orphans on mds and ost */
1202 if (OBT(obd) && OBP(obd, postrecov)) {
1203 int rc = OBP(obd, postrecov)(obd);
1204 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1205 rc < 0 ? "failed" : "complete", rc);
1208 obd->obd_recovery_end = cfs_time_current_sec();
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, &obd->obd_lock_replay_queue, 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 LASSERT(obd->obd_stopping);
1271 CFS_INIT_LIST_HEAD(&clean_list);
1272 spin_lock_bh(&obd->obd_processing_task_lock);
1273 if (!obd->obd_recovering) {
1274 spin_unlock_bh(&obd->obd_processing_task_lock);
1278 obd->obd_recovering = obd->obd_abort_recovery = 0;
1279 target_cancel_recovery_timer(obd);
1281 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1282 spin_unlock_bh(&obd->obd_processing_task_lock);
1284 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1285 LASSERT(req->rq_reply_state == 0);
1286 target_exp_dequeue_req_replay(req);
1287 ptlrpc_free_clone(req);
1290 spin_lock_bh(&obd->obd_processing_task_lock);
1291 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1292 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1293 spin_unlock_bh(&obd->obd_processing_task_lock);
1295 list_for_each_entry_safe(req, n, &clean_list, rq_list){
1296 LASSERT(req->rq_reply_state == 0);
1297 ptlrpc_free_clone(req);
1303 /* obd_processing_task_lock should be held */
1304 void target_cancel_recovery_timer(struct obd_device *obd)
1306 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1307 cfs_timer_disarm(&obd->obd_recovery_timer);
1310 /* extend = 1 means require at least "duration" seconds left in the timer,
1311 extend = 0 means set the total duration (start_recovery_timer) */
1312 static void reset_recovery_timer(struct obd_device *obd, int duration,
1315 cfs_time_t now = cfs_time_current_sec();
1316 cfs_duration_t left;
1318 spin_lock_bh(&obd->obd_processing_task_lock);
1319 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1320 spin_unlock_bh(&obd->obd_processing_task_lock);
1324 left = cfs_time_sub(obd->obd_recovery_end, now);
1326 if (extend && (duration > left))
1327 obd->obd_recovery_timeout += duration - left;
1328 else if (!extend && (duration > obd->obd_recovery_timeout))
1329 /* Track the client's largest expected replay time */
1330 obd->obd_recovery_timeout = duration;
1333 * If total recovery time already exceed the
1334 * obd_recovery_max_time, then CRAY XT3 will
1335 * abort the recovery
1337 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1338 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1340 obd->obd_recovery_end = obd->obd_recovery_start +
1341 obd->obd_recovery_timeout;
1342 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1343 cfs_time_before(now, obd->obd_recovery_end)) {
1344 left = cfs_time_sub(obd->obd_recovery_end, now);
1345 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1347 spin_unlock_bh(&obd->obd_processing_task_lock);
1348 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1349 obd->obd_name, (unsigned)left);
1352 static void check_and_start_recovery_timer(struct obd_device *obd)
1354 spin_lock_bh(&obd->obd_processing_task_lock);
1355 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1356 spin_unlock_bh(&obd->obd_processing_task_lock);
1359 CWARN("%s: starting recovery timer\n", obd->obd_name);
1360 obd->obd_recovery_start = cfs_time_current_sec();
1362 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1363 spin_unlock_bh(&obd->obd_processing_task_lock);
1365 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1368 /* Reset the timer with each new client connection */
1370 * This timer is actually reconnect_timer, which is for making sure
1371 * the total recovery window is at least as big as my reconnect
1372 * attempt timing. So the initial recovery time_out will be set to
1373 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1374 * from client is bigger than this, then the recovery time_out will
1375 * be extend to make sure the client could be reconnected, in the
1376 * process, the timeout from the new client should be ignored.
1380 target_start_and_reset_recovery_timer(struct obd_device *obd,
1381 struct ptlrpc_request *req,
1384 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1386 if (!new_client && service_time)
1387 /* Teach server about old server's estimates, as first guess
1388 * at how long new requests will take. */
1389 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1392 check_and_start_recovery_timer(obd);
1394 /* convert the service time to rpc timeout,
1395 * reuse service_time to limit stack usage */
1396 service_time = at_est2timeout(service_time);
1398 /* We expect other clients to timeout within service_time, then try
1399 * to reconnect, then try the failover server. The max delay between
1400 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1401 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1402 INITIAL_CONNECT_TIMEOUT);
1403 if (service_time > obd->obd_recovery_timeout && !new_client)
1404 reset_recovery_timer(obd, service_time, 0);
1408 static int check_for_next_transno(struct obd_device *obd)
1410 struct ptlrpc_request *req = NULL;
1411 int wake_up = 0, connected, completed, queue_len, max;
1412 __u64 next_transno, req_transno;
1414 spin_lock_bh(&obd->obd_processing_task_lock);
1416 if (!list_empty(&obd->obd_req_replay_queue)) {
1417 req = list_entry(obd->obd_req_replay_queue.next,
1418 struct ptlrpc_request, rq_list);
1419 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1424 max = obd->obd_max_recoverable_clients;
1425 connected = obd->obd_connected_clients;
1426 completed = connected - obd->obd_recoverable_clients;
1427 queue_len = obd->obd_requests_queued_for_recovery;
1428 next_transno = obd->obd_next_recovery_transno;
1430 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1431 "req_transno: "LPU64", next_transno: "LPU64"\n",
1432 max, connected, completed, queue_len, req_transno, next_transno);
1434 if (obd->obd_abort_recovery) {
1435 CDEBUG(D_HA, "waking for aborted recovery\n");
1437 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1438 CDEBUG(D_HA, "waking for completed recovery\n");
1440 } else if (req_transno == next_transno) {
1441 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1443 } else if (queue_len + completed == max) {
1444 /* handle gaps occured due to lost reply. It is allowed gaps
1445 * because all clients are connected and there will be resend
1446 * for missed transaction */
1447 LASSERTF(req_transno >= next_transno,
1448 "req_transno: "LPU64", next_transno: "LPU64"\n",
1449 req_transno, next_transno);
1451 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1452 "waking for skipped transno (skip: "LPD64
1453 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1454 next_transno, queue_len, completed, connected, req_transno);
1455 obd->obd_next_recovery_transno = req_transno;
1457 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1458 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1459 " by fail_lock, waking up ("LPD64")\n", next_transno);
1460 obd->obd_next_recovery_transno = req_transno;
1462 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1463 /* some clients haven't connected in time, but we can try
1464 * to replay requests that demand on already committed ones
1465 * also, we can replay first non-committed transation */
1466 LASSERT(req_transno != 0);
1467 if (req_transno == obd->obd_last_committed + 1) {
1468 obd->obd_next_recovery_transno = req_transno;
1469 } else if (req_transno > obd->obd_last_committed) {
1470 /* can't continue recovery: have no needed transno */
1471 obd->obd_abort_recovery = 1;
1472 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1473 "connected: %d, completed: %d, queue_len: %d, "
1474 "req_transno: "LPU64", next_transno: "LPU64"\n",
1475 max, connected, completed, queue_len,
1476 req_transno, next_transno);
1481 spin_unlock_bh(&obd->obd_processing_task_lock);
1485 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1487 struct l_wait_info lwi = { 0 };
1488 struct ptlrpc_request *req;
1490 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1491 obd->obd_next_recovery_transno);
1492 l_wait_event(obd->obd_next_transno_waitq,
1493 check_for_next_transno(obd), &lwi);
1495 spin_lock_bh(&obd->obd_processing_task_lock);
1496 if (obd->obd_abort_recovery) {
1498 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1499 req = list_entry(obd->obd_req_replay_queue.next,
1500 struct ptlrpc_request, rq_list);
1501 target_exp_dequeue_req_replay(req);
1502 list_del_init(&req->rq_list);
1503 obd->obd_requests_queued_for_recovery--;
1507 spin_unlock_bh(&obd->obd_processing_task_lock);
1511 static int check_for_next_lock(struct obd_device *obd)
1513 struct ptlrpc_request *req = NULL;
1516 spin_lock_bh(&obd->obd_processing_task_lock);
1517 if (!list_empty(&obd->obd_lock_replay_queue)) {
1518 req = list_entry(obd->obd_lock_replay_queue.next,
1519 struct ptlrpc_request, rq_list);
1520 CDEBUG(D_HA, "waking for next lock\n");
1522 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1523 CDEBUG(D_HA, "waking for completed lock replay\n");
1525 } else if (obd->obd_abort_recovery) {
1526 CDEBUG(D_HA, "waking for aborted recovery\n");
1529 spin_unlock_bh(&obd->obd_processing_task_lock);
1534 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1536 struct l_wait_info lwi = { 0 };
1537 struct ptlrpc_request *req;
1539 CDEBUG(D_HA, "Waiting for lock\n");
1540 l_wait_event(obd->obd_next_transno_waitq,
1541 check_for_next_lock(obd), &lwi);
1543 spin_lock_bh(&obd->obd_processing_task_lock);
1544 if (obd->obd_abort_recovery) {
1546 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1547 req = list_entry(obd->obd_lock_replay_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 struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1559 struct ptlrpc_request *req;
1561 spin_lock_bh(&obd->obd_processing_task_lock);
1562 if (!list_empty(&obd->obd_final_req_queue)) {
1563 req = list_entry(obd->obd_final_req_queue.next,
1564 struct ptlrpc_request, rq_list);
1565 list_del_init(&req->rq_list);
1569 spin_unlock_bh(&obd->obd_processing_task_lock);
1573 static inline int req_replay_done(struct obd_export *exp)
1575 return (exp->exp_req_replay_needed == 0);
1578 static inline int lock_replay_done(struct obd_export *exp)
1580 return (exp->exp_lock_replay_needed == 0);
1583 static inline int connect_done(struct obd_export *exp)
1585 return (exp->exp_in_recovery != 0);
1588 static int check_for_clients(struct obd_device *obd)
1590 if (obd->obd_abort_recovery)
1592 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1593 if (obd->obd_no_conn == 0 &&
1594 obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1599 static int handle_recovery_req(struct ptlrpc_thread *thread,
1600 struct ptlrpc_request *req,
1601 svc_handler_t handler)
1606 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1608 CERROR("Failure to initialize session: %d\n", rc);
1611 req->rq_session.lc_thread = thread;
1612 lu_context_enter(&req->rq_session);
1613 req->rq_svc_thread = thread;
1614 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1616 /* thread context */
1617 lu_context_enter(&thread->t_env->le_ctx);
1619 lu_context_exit(&thread->t_env->le_ctx);
1621 lu_context_exit(&req->rq_session);
1622 lu_context_fini(&req->rq_session);
1623 /* don't reset timer for final stage */
1624 if (!req_replay_done(req->rq_export) ||
1625 !lock_replay_done(req->rq_export))
1626 reset_recovery_timer(class_exp2obd(req->rq_export),
1627 AT_OFF ? obd_timeout :
1628 at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1629 ptlrpc_free_clone(req);
1633 static void resume_recovery_timer(struct obd_device *obd)
1635 /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1636 reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1639 static int target_recovery_thread(void *arg)
1641 struct obd_device *obd = arg;
1642 struct ptlrpc_request *req;
1643 struct target_recovery_data *trd = &obd->obd_recovery_data;
1644 struct l_wait_info lwi = { 0 };
1645 unsigned long delta;
1646 unsigned long flags;
1648 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1652 cfs_daemonize("tgt_recov");
1654 SIGNAL_MASK_LOCK(current, flags);
1655 sigfillset(¤t->blocked);
1657 SIGNAL_MASK_UNLOCK(current, flags);
1659 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1663 thread->t_env = &env;
1664 env.le_ctx.lc_thread = thread;
1666 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1668 trd->trd_processing_task = current->pid;
1670 obd->obd_recovering = 1;
1671 complete(&trd->trd_starting);
1673 /* first of all, we have to know the first transno to replay */
1674 obd->obd_abort_recovery = 0;
1675 l_wait_event(obd->obd_next_transno_waitq,
1676 check_for_clients(obd), &lwi);
1678 spin_lock_bh(&obd->obd_processing_task_lock);
1679 target_cancel_recovery_timer(obd);
1680 spin_unlock_bh(&obd->obd_processing_task_lock);
1682 /* If some clients haven't connected in time, evict them */
1683 if (obd->obd_abort_recovery) {
1684 CWARN("Some clients haven't connect in time (%d/%d),"
1685 "evict them\n", obd->obd_connected_clients,
1686 obd->obd_max_recoverable_clients);
1687 obd->obd_abort_recovery = obd->obd_stopping;
1688 class_disconnect_stale_exports(obd, connect_done,
1689 exp_flags_from_obd(obd) |
1690 OBD_OPT_ABORT_RECOV);
1692 /* next stage: replay requests */
1694 obd->obd_req_replaying = 1;
1695 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1696 atomic_read(&obd->obd_req_replay_clients),
1697 obd->obd_next_recovery_transno);
1698 resume_recovery_timer(obd);
1699 while ((req = target_next_replay_req(obd))) {
1700 LASSERT(trd->trd_processing_task == current->pid);
1701 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1702 lustre_msg_get_transno(req->rq_reqmsg),
1703 libcfs_nid2str(req->rq_peer.nid));
1704 handle_recovery_req(thread, req,
1705 trd->trd_recovery_handler);
1706 obd->obd_replayed_requests++;
1707 spin_lock_bh(&obd->obd_processing_task_lock);
1708 obd->obd_next_recovery_transno++;
1709 spin_unlock_bh(&obd->obd_processing_task_lock);
1712 spin_lock_bh(&obd->obd_processing_task_lock);
1713 target_cancel_recovery_timer(obd);
1714 spin_unlock_bh(&obd->obd_processing_task_lock);
1716 /* If some clients haven't replayed requests in time, evict them */
1717 if (obd->obd_abort_recovery) {
1718 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1719 obd->obd_abort_recovery = obd->obd_stopping;
1720 class_disconnect_stale_exports(obd, req_replay_done,
1721 exp_flags_from_obd(obd) |
1722 OBD_OPT_ABORT_RECOV);
1723 abort_req_replay_queue(obd);
1726 /* The second stage: replay locks */
1727 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1728 atomic_read(&obd->obd_lock_replay_clients));
1729 resume_recovery_timer(obd);
1730 while ((req = target_next_replay_lock(obd))) {
1731 LASSERT(trd->trd_processing_task == current->pid);
1732 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1733 libcfs_nid2str(req->rq_peer.nid));
1734 handle_recovery_req(thread, req,
1735 trd->trd_recovery_handler);
1736 obd->obd_replayed_locks++;
1739 spin_lock_bh(&obd->obd_processing_task_lock);
1740 target_cancel_recovery_timer(obd);
1741 spin_unlock_bh(&obd->obd_processing_task_lock);
1742 /* If some clients haven't replayed requests in time, evict them */
1743 if (obd->obd_abort_recovery) {
1745 CERROR("lock replay timed out, aborting ...\n");
1746 obd->obd_abort_recovery = obd->obd_stopping;
1747 stale = class_disconnect_stale_exports(obd, lock_replay_done,
1748 exp_flags_from_obd(obd) |
1749 OBD_OPT_ABORT_RECOV);
1750 abort_lock_replay_queue(obd);
1753 /* We drop recoverying flag to forward all new requests
1754 * to regular mds_handle() since now */
1755 spin_lock_bh(&obd->obd_processing_task_lock);
1756 obd->obd_recovering = obd->obd_abort_recovery = 0;
1757 spin_unlock_bh(&obd->obd_processing_task_lock);
1758 /* The third stage: reply on final pings */
1759 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1760 while ((req = target_next_final_ping(obd))) {
1761 LASSERT(trd->trd_processing_task == current->pid);
1762 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1763 libcfs_nid2str(req->rq_peer.nid));
1764 handle_recovery_req(thread, req,
1765 trd->trd_recovery_handler);
1768 delta = (jiffies - delta) / HZ;
1769 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1770 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1771 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1772 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1773 if (delta > obd_timeout * 2) {
1774 CWARN("too long recovery - read logs\n");
1775 libcfs_debug_dumplog();
1778 target_finish_recovery(obd);
1780 lu_context_fini(&env.le_ctx);
1781 trd->trd_processing_task = 0;
1782 complete(&trd->trd_finishing);
1786 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1789 struct target_recovery_data *trd = &obd->obd_recovery_data;
1791 memset(trd, 0, sizeof(*trd));
1792 init_completion(&trd->trd_starting);
1793 init_completion(&trd->trd_finishing);
1794 trd->trd_recovery_handler = handler;
1796 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1797 wait_for_completion(&trd->trd_starting);
1798 LASSERT(obd->obd_recovering != 0);
1805 void target_stop_recovery_thread(struct obd_device *obd)
1807 spin_lock_bh(&obd->obd_processing_task_lock);
1808 if (obd->obd_recovery_data.trd_processing_task > 0) {
1809 struct target_recovery_data *trd = &obd->obd_recovery_data;
1810 CERROR("%s: Aborting recovery\n", obd->obd_name);
1811 obd->obd_abort_recovery = 1;
1812 wake_up(&obd->obd_next_transno_waitq);
1813 spin_unlock_bh(&obd->obd_processing_task_lock);
1814 wait_for_completion(&trd->trd_finishing);
1816 spin_unlock_bh(&obd->obd_processing_task_lock);
1820 void target_recovery_fini(struct obd_device *obd)
1822 class_disconnect_exports(obd);
1823 target_stop_recovery_thread(obd);
1824 target_cleanup_recovery(obd);
1826 EXPORT_SYMBOL(target_recovery_fini);
1828 static void target_recovery_expired(unsigned long castmeharder)
1830 struct obd_device *obd = (struct obd_device *)castmeharder;
1831 LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1832 "after %lds (%d clients did)\n",
1833 obd->obd_name, obd->obd_recoverable_clients,
1834 cfs_time_current_sec()- obd->obd_recovery_start,
1835 obd->obd_connected_clients);
1836 spin_lock_bh(&obd->obd_processing_task_lock);
1837 if (obd->obd_recovering)
1838 obd->obd_abort_recovery = 1;
1839 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1840 spin_unlock_bh(&obd->obd_processing_task_lock);
1843 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1845 if (obd->obd_max_recoverable_clients == 0)
1848 CWARN("RECOVERY: service %s, %d recoverable clients, "
1849 "last_transno "LPU64"\n", obd->obd_name,
1850 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1851 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1852 obd->obd_recovery_start = 0;
1853 obd->obd_recovery_end = 0;
1854 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1855 /* bz13079: this should be set to desired value for ost but not for mds */
1856 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1857 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1858 target_start_recovery_thread(obd, handler);
1860 EXPORT_SYMBOL(target_recovery_init);
1864 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1866 struct obd_export *exp = req->rq_export;
1867 LASSERT(exp != NULL);
1868 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1869 /* client declares he's ready to replay locks */
1870 spin_lock_bh(&obd->obd_processing_task_lock);
1871 if (exp->exp_req_replay_needed) {
1872 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1873 spin_lock(&exp->exp_lock);
1874 exp->exp_req_replay_needed = 0;
1875 spin_unlock(&exp->exp_lock);
1876 atomic_dec(&obd->obd_req_replay_clients);
1877 LASSERT(obd->obd_recoverable_clients > 0);
1878 obd->obd_recoverable_clients--;
1879 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1880 CDEBUG(D_HA, "all clients have replayed reqs\n");
1881 wake_up(&obd->obd_next_transno_waitq);
1883 spin_unlock_bh(&obd->obd_processing_task_lock);
1885 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1886 /* client declares he's ready to complete recovery
1887 * so, we put the request on th final queue */
1888 spin_lock_bh(&obd->obd_processing_task_lock);
1889 if (exp->exp_lock_replay_needed) {
1890 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1891 spin_lock(&exp->exp_lock);
1892 exp->exp_lock_replay_needed = 0;
1893 spin_unlock(&exp->exp_lock);
1894 atomic_dec(&obd->obd_lock_replay_clients);
1895 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1896 CDEBUG(D_HA, "all clients have replayed locks\n");
1897 wake_up(&obd->obd_next_transno_waitq);
1899 spin_unlock_bh(&obd->obd_processing_task_lock);
1905 int target_queue_recovery_request(struct ptlrpc_request *req,
1906 struct obd_device *obd)
1908 struct list_head *tmp;
1910 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1914 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1915 /* Processing the queue right now, don't re-add. */
1919 target_process_req_flags(obd, req);
1921 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1922 /* client declares he's ready to complete recovery
1923 * so, we put the request on th final queue */
1924 req = ptlrpc_clone_req(req);
1927 DEBUG_REQ(D_HA, req, "queue final req");
1928 spin_lock_bh(&obd->obd_processing_task_lock);
1929 if (obd->obd_recovering)
1930 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1932 spin_unlock_bh(&obd->obd_processing_task_lock);
1933 ptlrpc_free_clone(req);
1934 if (obd->obd_stopping) {
1940 spin_unlock_bh(&obd->obd_processing_task_lock);
1943 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1944 /* client declares he's ready to replay locks */
1945 req = ptlrpc_clone_req(req);
1948 DEBUG_REQ(D_HA, req, "queue lock replay req");
1949 spin_lock_bh(&obd->obd_processing_task_lock);
1950 LASSERT(obd->obd_recovering);
1951 /* usually due to recovery abort */
1952 if (!req->rq_export->exp_in_recovery) {
1953 spin_unlock_bh(&obd->obd_processing_task_lock);
1954 ptlrpc_free_clone(req);
1957 LASSERT(req->rq_export->exp_lock_replay_needed);
1958 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1959 spin_unlock_bh(&obd->obd_processing_task_lock);
1960 wake_up(&obd->obd_next_transno_waitq);
1964 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1965 * (i.e. buflens etc are in my own byte order), but type-dependent
1966 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1969 CFS_INIT_LIST_HEAD(&req->rq_list);
1970 DEBUG_REQ(D_HA, req, "not queueing");
1974 spin_lock_bh(&obd->obd_processing_task_lock);
1976 /* If we're processing the queue, we want don't want to queue this
1979 * Also, if this request has a transno less than the one we're waiting
1980 * for, we should process it now. It could (and currently always will)
1981 * be an open request for a descriptor that was opened some time ago.
1983 * Also, a resent, replayed request that has already been
1984 * handled will pass through here and be processed immediately.
1986 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1987 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1988 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1989 /* Processing the queue right now, don't re-add. */
1990 LASSERT(list_empty(&req->rq_list));
1991 spin_unlock_bh(&obd->obd_processing_task_lock);
1994 spin_unlock_bh(&obd->obd_processing_task_lock);
1996 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
1999 req = ptlrpc_clone_req(req);
2003 spin_lock_bh(&obd->obd_processing_task_lock);
2004 LASSERT(obd->obd_recovering);
2005 if (!req->rq_export->exp_in_recovery) {
2006 spin_unlock_bh(&obd->obd_processing_task_lock);
2007 ptlrpc_free_clone(req);
2010 LASSERT(req->rq_export->exp_req_replay_needed);
2012 if (target_exp_enqueue_req_replay(req)) {
2013 spin_unlock_bh(&obd->obd_processing_task_lock);
2014 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2015 ptlrpc_free_clone(req);
2020 list_for_each(tmp, &obd->obd_req_replay_queue) {
2021 struct ptlrpc_request *reqiter =
2022 list_entry(tmp, struct ptlrpc_request, rq_list);
2024 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2025 list_add_tail(&req->rq_list, &reqiter->rq_list);
2030 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2032 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2033 "has been claimed by another client");
2034 spin_unlock_bh(&obd->obd_processing_task_lock);
2035 target_exp_dequeue_req_replay(req);
2036 ptlrpc_free_clone(req);
2042 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2044 obd->obd_requests_queued_for_recovery++;
2045 wake_up(&obd->obd_next_transno_waitq);
2046 spin_unlock_bh(&obd->obd_processing_task_lock);
2050 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2052 return req->rq_export->exp_obd;
2055 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2057 LASSERT(exp != NULL);
2058 return &exp->exp_obd->obd_namespace->ns_pool;
2062 * Packs current SLV and Limit into \a req.
2064 int target_pack_pool_reply(struct ptlrpc_request *req)
2066 struct obd_device *obd;
2070 * Check that we still have all structures alive as this may
2071 * be some late rpc in shutdown time.
2073 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2074 !exp_connect_lru_resize(req->rq_export))) {
2075 lustre_msg_set_slv(req->rq_repmsg, 0);
2076 lustre_msg_set_limit(req->rq_repmsg, 0);
2081 * OBD is alive here as export is alive, which we checked above.
2083 obd = req->rq_export->exp_obd;
2085 read_lock(&obd->obd_pool_lock);
2086 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2087 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2088 read_unlock(&obd->obd_pool_lock);
2093 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2095 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2096 DEBUG_REQ(D_ERROR, req, "dropping reply");
2101 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2102 req->rq_status = rc;
2103 return (ptlrpc_send_error(req, 1));
2105 DEBUG_REQ(D_NET, req, "sending reply");
2108 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2111 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2114 struct ptlrpc_reply_state *rs;
2115 struct obd_device *obd;
2116 struct obd_export *exp;
2117 struct ptlrpc_service *svc;
2120 if (req->rq_no_reply) {
2125 svc = req->rq_rqbd->rqbd_service;
2126 rs = req->rq_reply_state;
2127 if (rs == NULL || !rs->rs_difficult) {
2129 target_send_reply_msg (req, rc, fail_id);
2134 /* must be an export if locks saved */
2135 LASSERT (req->rq_export != NULL);
2136 /* req/reply consistent */
2137 LASSERT (rs->rs_service == svc);
2140 LASSERT (!rs->rs_scheduled);
2141 LASSERT (!rs->rs_scheduled_ever);
2142 LASSERT (!rs->rs_handled);
2143 LASSERT (!rs->rs_on_net);
2144 LASSERT (rs->rs_export == NULL);
2145 LASSERT (list_empty(&rs->rs_obd_list));
2146 LASSERT (list_empty(&rs->rs_exp_list));
2148 exp = class_export_get (req->rq_export);
2151 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2152 rs->rs_scheduled = 1;
2154 rs->rs_xid = req->rq_xid;
2155 rs->rs_transno = req->rq_transno;
2156 rs->rs_export = exp;
2158 spin_lock(&obd->obd_uncommitted_replies_lock);
2160 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2161 rs->rs_transno, obd->obd_last_committed);
2162 if (rs->rs_transno > obd->obd_last_committed) {
2163 /* not committed already */
2164 list_add_tail (&rs->rs_obd_list,
2165 &obd->obd_uncommitted_replies);
2168 spin_unlock (&obd->obd_uncommitted_replies_lock);
2169 spin_lock (&exp->exp_lock);
2171 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
2173 spin_unlock(&exp->exp_lock);
2175 netrc = target_send_reply_msg (req, rc, fail_id);
2177 spin_lock(&svc->srv_lock);
2179 atomic_inc(&svc->srv_n_difficult_replies);
2182 /* error sending: reply is off the net. Also we need +1
2183 * reply ref until ptlrpc_server_handle_reply() is done
2184 * with the reply state (if the send was successful, there
2185 * would have been +1 ref for the net, which
2186 * reply_out_callback leaves alone) */
2188 ptlrpc_rs_addref(rs);
2189 atomic_inc (&svc->srv_outstanding_replies);
2192 spin_lock(&rs->rs_lock);
2193 if (rs->rs_transno <= obd->obd_last_committed ||
2194 (!rs->rs_on_net && !rs->rs_no_ack) ||
2195 list_empty(&rs->rs_exp_list) || /* completed already */
2196 list_empty(&rs->rs_obd_list)) {
2197 CDEBUG(D_HA, "Schedule reply immediately\n");
2198 ptlrpc_dispatch_difficult_reply(rs);
2200 list_add (&rs->rs_list, &svc->srv_active_replies);
2201 rs->rs_scheduled = 0; /* allow notifier to schedule */
2203 spin_unlock(&rs->rs_lock);
2204 spin_unlock(&svc->srv_lock);
2208 int target_handle_ping(struct ptlrpc_request *req)
2210 obd_ping(req->rq_export);
2211 return req_capsule_server_pack(&req->rq_pill);
2214 void target_committed_to_req(struct ptlrpc_request *req)
2216 struct obd_device *obd;
2218 if (req == NULL || req->rq_export == NULL)
2221 obd = req->rq_export->exp_obd;
2225 if (!obd->obd_no_transno && req->rq_repmsg != NULL)
2226 lustre_msg_set_last_committed(req->rq_repmsg,
2227 obd->obd_last_committed);
2229 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2230 "%d)", obd->obd_no_transno, req->rq_repmsg == NULL);
2232 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2233 obd->obd_last_committed, req->rq_transno, req->rq_xid);
2236 EXPORT_SYMBOL(target_committed_to_req);
2238 int target_handle_qc_callback(struct ptlrpc_request *req)
2240 struct obd_quotactl *oqctl;
2241 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2243 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2244 if (oqctl == NULL) {
2245 CERROR("Can't unpack obd_quotactl\n");
2249 cli->cl_qchk_stat = oqctl->qc_stat;
2254 #ifdef HAVE_QUOTA_SUPPORT
2255 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2258 struct obd_device *obd = req->rq_export->exp_obd;
2259 struct obd_device *master_obd;
2260 struct obd_device_target *obt;
2261 struct lustre_quota_ctxt *qctxt;
2262 struct qunit_data *qdata = NULL;
2266 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2269 rc = req_capsule_server_pack(&req->rq_pill);
2271 CERROR("packing reply failed!: rc = %d\n", rc);
2275 LASSERT(req->rq_export);
2277 OBD_ALLOC(qdata, sizeof(struct qunit_data));
2280 rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2282 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2286 /* we use the observer */
2287 if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2288 CERROR("Can't find the observer, it is recovering\n");
2289 req->rq_status = -EAGAIN;
2290 GOTO(send_reply, rc = -EAGAIN);
2293 master_obd = obd->obd_observer->obd_observer;
2294 obt = &master_obd->u.obt;
2295 qctxt = &obt->obt_qctxt;
2297 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2298 /* quota_type has not been processed yet, return EAGAIN
2299 * until we know whether or not quotas are supposed to
2301 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2303 req->rq_status = -EAGAIN;
2304 rc = ptlrpc_reply(req);
2308 down_read(&obt->obt_rwsem);
2309 if (qctxt->lqc_lqs_hash == NULL) {
2310 up_read(&obt->obt_rwsem);
2311 /* quota_type has not been processed yet, return EAGAIN
2312 * until we know whether or not quotas are supposed to
2314 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2316 req->rq_status = -EAGAIN;
2317 rc = ptlrpc_reply(req);
2321 LASSERT(qctxt->lqc_handler);
2322 rc = qctxt->lqc_handler(master_obd, qdata,
2323 lustre_msg_get_opc(req->rq_reqmsg));
2324 up_read(&obt->obt_rwsem);
2325 if (rc && rc != -EDQUOT)
2326 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2327 "dqacq failed! (rc:%d)\n", rc);
2328 req->rq_status = rc;
2330 /* there are three forms of qunit(historic causes), so we need to
2331 * adjust the same form to different forms slaves needed */
2332 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2334 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2338 /* Block the quota req. b=14840 */
2339 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2341 rc = ptlrpc_reply(req);
2343 OBD_FREE(qdata, sizeof(struct qunit_data));
2347 #endif /* !__KERNEL__ */
2349 #endif /* HAVE_QUOTA_SUPPORT */
2351 ldlm_mode_t lck_compat_array[] = {
2352 [LCK_EX] LCK_COMPAT_EX,
2353 [LCK_PW] LCK_COMPAT_PW,
2354 [LCK_PR] LCK_COMPAT_PR,
2355 [LCK_CW] LCK_COMPAT_CW,
2356 [LCK_CR] LCK_COMPAT_CR,
2357 [LCK_NL] LCK_COMPAT_NL,
2358 [LCK_GROUP] LCK_COMPAT_GROUP,
2359 [LCK_COS] LCK_COMPAT_COS,
2363 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2364 * not escape to the user level.
2366 int ldlm_error2errno(ldlm_error_t error)
2374 case ELDLM_LOCK_CHANGED:
2377 case ELDLM_LOCK_ABORTED:
2380 case ELDLM_LOCK_REPLACED:
2383 case ELDLM_NO_LOCK_DATA:
2386 case ELDLM_NAMESPACE_EXISTS:
2389 case ELDLM_BAD_NAMESPACE:
2393 if (((int)error) < 0) /* cast to signed type */
2394 result = error; /* as ldlm_error_t can be unsigned */
2396 CERROR("Invalid DLM result code: %i\n", error);
2402 EXPORT_SYMBOL(ldlm_error2errno);
2405 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2407 ldlm_error_t ldlm_errno2error(int err_no)
2416 error = ELDLM_LOCK_CHANGED;
2419 error = ELDLM_LOCK_ABORTED;
2422 error = ELDLM_LOCK_REPLACED;
2425 error = ELDLM_NO_LOCK_DATA;
2428 error = ELDLM_NAMESPACE_EXISTS;
2431 error = ELDLM_BAD_NAMESPACE;
2438 EXPORT_SYMBOL(ldlm_errno2error);