1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <libcfs/libcfs.h>
45 # include <liblustre.h>
48 #include <lustre_mds.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
54 /* @priority: if non-zero, move the selected to the list head
55 * @create: if zero, only search in existed connections
57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
58 int priority, int create)
60 struct ptlrpc_connection *ptlrpc_conn;
61 struct obd_import_conn *imp_conn = NULL, *item;
65 if (!create && !priority) {
66 CDEBUG(D_HA, "Nothing to do\n");
70 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
72 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
77 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
79 GOTO(out_put, rc = -ENOMEM);
83 spin_lock(&imp->imp_lock);
84 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
85 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
87 list_del(&item->oic_item);
88 list_add(&item->oic_item, &imp->imp_conn_list);
89 item->oic_last_attempt = 0;
91 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
92 imp, imp->imp_obd->obd_name, uuid->uuid,
93 (priority ? ", moved to head" : ""));
94 spin_unlock(&imp->imp_lock);
95 GOTO(out_free, rc = 0);
100 imp_conn->oic_conn = ptlrpc_conn;
101 imp_conn->oic_uuid = *uuid;
102 imp_conn->oic_last_attempt = 0;
104 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
106 list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
107 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
108 imp, imp->imp_obd->obd_name, uuid->uuid,
109 (priority ? "head" : "tail"));
111 spin_unlock(&imp->imp_lock);
112 GOTO(out_free, rc = -ENOENT);
115 spin_unlock(&imp->imp_lock);
119 OBD_FREE(imp_conn, sizeof(*imp_conn));
121 ptlrpc_connection_put(ptlrpc_conn);
125 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
127 return import_set_conn(imp, uuid, 1, 0);
130 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
133 return import_set_conn(imp, uuid, priority, 1);
136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
138 struct obd_import_conn *imp_conn;
139 struct obd_export *dlmexp;
143 spin_lock(&imp->imp_lock);
144 if (list_empty(&imp->imp_conn_list)) {
145 LASSERT(!imp->imp_connection);
149 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
150 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
152 LASSERT(imp_conn->oic_conn);
154 /* is current conn? */
155 if (imp_conn == imp->imp_conn_current) {
156 LASSERT(imp_conn->oic_conn == imp->imp_connection);
158 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
159 imp->imp_state != LUSTRE_IMP_DISCON) {
160 CERROR("can't remove current connection\n");
161 GOTO(out, rc = -EBUSY);
164 ptlrpc_connection_put(imp->imp_connection);
165 imp->imp_connection = NULL;
167 dlmexp = class_conn2export(&imp->imp_dlm_handle);
168 if (dlmexp && dlmexp->exp_connection) {
169 LASSERT(dlmexp->exp_connection ==
171 ptlrpc_connection_put(dlmexp->exp_connection);
172 dlmexp->exp_connection = NULL;
176 list_del(&imp_conn->oic_item);
177 ptlrpc_connection_put(imp_conn->oic_conn);
178 OBD_FREE(imp_conn, sizeof(*imp_conn));
179 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
180 imp, imp->imp_obd->obd_name, uuid->uuid);
185 spin_unlock(&imp->imp_lock);
187 CERROR("connection %s not found\n", uuid->uuid);
191 void client_destroy_import(struct obd_import *imp)
193 /* drop security policy instance after all rpc finished/aborted
194 * to let all busy contexts be released. */
195 class_import_get(imp);
196 class_destroy_import(imp);
197 sptlrpc_import_sec_put(imp);
198 class_import_put(imp);
201 /* configure an RPC client OBD device
206 * 3 - inactive-on-startup
208 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
210 struct client_obd *cli = &obddev->u.cli;
211 struct obd_import *imp;
212 struct obd_uuid server_uuid;
213 int rq_portal, rp_portal, connect_op;
214 char *name = obddev->obd_type->typ_name;
218 /* In a more perfect world, we would hang a ptlrpc_client off of
219 * obd_type and just use the values from there. */
220 if (!strcmp(name, LUSTRE_OSC_NAME)) {
221 rq_portal = OST_REQUEST_PORTAL;
222 rp_portal = OSC_REPLY_PORTAL;
223 connect_op = OST_CONNECT;
224 cli->cl_sp_me = LUSTRE_SP_CLI;
225 cli->cl_sp_to = LUSTRE_SP_OST;
226 } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
227 rq_portal = MDS_REQUEST_PORTAL;
228 rp_portal = MDC_REPLY_PORTAL;
229 connect_op = MDS_CONNECT;
230 cli->cl_sp_me = LUSTRE_SP_CLI;
231 cli->cl_sp_to = LUSTRE_SP_MDT;
232 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
233 rq_portal = MGS_REQUEST_PORTAL;
234 rp_portal = MGC_REPLY_PORTAL;
235 connect_op = MGS_CONNECT;
236 cli->cl_sp_me = LUSTRE_SP_MGC;
237 cli->cl_sp_to = LUSTRE_SP_MGS;
238 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
240 CERROR("unknown client OBD type \"%s\", can't setup\n",
245 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
246 CERROR("requires a TARGET UUID\n");
250 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
251 CERROR("client UUID must be less than 38 characters\n");
255 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
256 CERROR("setup requires a SERVER UUID\n");
260 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
261 CERROR("target UUID must be less than 38 characters\n");
265 init_rwsem(&cli->cl_sem);
266 sema_init(&cli->cl_mgc_sem, 1);
267 cli->cl_conn_count = 0;
268 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
269 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
270 sizeof(server_uuid)));
273 cli->cl_avail_grant = 0;
274 /* FIXME: should limit this for the sum of all cl_dirty_max */
275 cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
276 if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
277 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
278 CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
279 CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
280 CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
281 CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
282 CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
283 client_obd_list_lock_init(&cli->cl_loi_list_lock);
284 cli->cl_r_in_flight = 0;
285 cli->cl_w_in_flight = 0;
287 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
288 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
289 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
290 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
291 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
292 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
293 cfs_waitq_init(&cli->cl_destroy_waitq);
294 atomic_set(&cli->cl_destroy_in_flight, 0);
295 #ifdef ENABLE_CHECKSUM
296 /* Turn on checksumming by default. */
297 cli->cl_checksum = 1;
299 * The supported checksum types will be worked out at connect time
300 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
303 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
305 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
307 /* This value may be changed at connect time in
308 ptlrpc_connect_interpret. */
309 cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
310 (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
312 if (!strcmp(name, LUSTRE_MDC_NAME)) {
313 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
314 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
315 cli->cl_max_rpcs_in_flight = 2;
316 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
317 cli->cl_max_rpcs_in_flight = 3;
318 } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
319 cli->cl_max_rpcs_in_flight = 4;
321 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
326 CERROR("ldlm_get_ref failed: %d\n", rc);
330 ptlrpc_init_client(rq_portal, rp_portal, name,
331 &obddev->obd_ldlm_client);
333 imp = class_new_import(obddev);
335 GOTO(err_ldlm, rc = -ENOENT);
336 imp->imp_client = &obddev->obd_ldlm_client;
337 imp->imp_connect_op = connect_op;
338 imp->imp_initial_recov = 1;
339 imp->imp_initial_recov_bk = 0;
340 CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
341 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
342 LUSTRE_CFG_BUFLEN(lcfg, 1));
343 class_import_put(imp);
345 rc = client_import_add_conn(imp, &server_uuid, 1);
347 CERROR("can't add initial connection\n");
348 GOTO(err_import, rc);
351 cli->cl_import = imp;
352 /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
353 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
354 cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
356 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
357 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
358 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
359 name, obddev->obd_name,
360 cli->cl_target_uuid.uuid);
361 spin_lock(&imp->imp_lock);
362 imp->imp_invalid = 1;
363 spin_unlock(&imp->imp_lock);
367 cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
372 class_destroy_import(imp);
380 int client_obd_cleanup(struct obd_device *obddev)
387 /* ->o_connect() method for client side (OSC and MDC and MGC) */
388 int client_connect_import(const struct lu_env *env,
389 struct obd_export **exp,
390 struct obd_device *obd, struct obd_uuid *cluuid,
391 struct obd_connect_data *data, void *localdata)
393 struct client_obd *cli = &obd->u.cli;
394 struct obd_import *imp = cli->cl_import;
395 struct obd_connect_data *ocd;
396 struct ldlm_namespace *to_be_freed = NULL;
397 struct lustre_handle conn = { 0 };
402 down_write(&cli->cl_sem);
403 if (cli->cl_conn_count > 0 )
404 GOTO(out_sem, rc = -EALREADY);
406 rc = class_connect(&conn, obd, cluuid);
410 cli->cl_conn_count++;
411 *exp = class_conn2export(&conn);
413 if (obd->obd_namespace != NULL)
414 CERROR("already have namespace!\n");
417 * Deadlock case - bug 18380
419 up_write(&cli->cl_sem);
420 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
421 LDLM_NAMESPACE_CLIENT,
422 LDLM_NAMESPACE_GREEDY);
423 down_write(&cli->cl_sem);
424 if (obd->obd_namespace == NULL)
425 GOTO(out_disco, rc = -ENOMEM);
427 imp->imp_dlm_handle = conn;
428 rc = ptlrpc_init_import(imp);
432 ocd = &imp->imp_connect_data;
435 imp->imp_connect_flags_orig = data->ocd_connect_flags;
438 rc = ptlrpc_connect_import(imp, NULL);
440 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
443 LASSERT((*exp)->exp_connection);
446 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
447 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
448 data->ocd_connect_flags, ocd->ocd_connect_flags);
449 data->ocd_connect_flags = ocd->ocd_connect_flags;
452 ptlrpc_pinger_add_import(imp);
458 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
459 to_be_freed = obd->obd_namespace;
460 obd->obd_namespace = NULL;
462 cli->cl_conn_count--;
463 class_disconnect(*exp);
467 up_write(&cli->cl_sem);
469 ldlm_namespace_free_post(to_be_freed);
474 int client_disconnect_export(struct obd_export *exp)
476 struct obd_device *obd = class_exp2obd(exp);
477 struct client_obd *cli;
478 struct obd_import *imp;
480 struct ldlm_namespace *to_be_freed = NULL;
484 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
485 exp, exp ? exp->exp_handle.h_cookie : -1);
490 imp = cli->cl_import;
492 down_write(&cli->cl_sem);
493 if (!cli->cl_conn_count) {
494 CERROR("disconnecting disconnected device (%s)\n",
496 GOTO(out_disconnect, rc = -EINVAL);
499 cli->cl_conn_count--;
500 if (cli->cl_conn_count)
501 GOTO(out_disconnect, rc = 0);
503 /* Mark import deactivated now, so we don't try to reconnect if any
504 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
505 * fully deactivate the import, or that would drop all requests. */
506 spin_lock(&imp->imp_lock);
507 imp->imp_deactive = 1;
508 spin_unlock(&imp->imp_lock);
510 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
511 * delete it regardless. (It's safe to delete an import that was
513 (void)ptlrpc_pinger_del_import(imp);
515 if (obd->obd_namespace != NULL) {
516 /* obd_force == local only */
517 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
518 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
520 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
521 to_be_freed = obd->obd_namespace;
525 * there's no necessary to hold sem during diconnecting an import,
526 * and actually it may cause deadlock in gss.
528 up_write(&cli->cl_sem);
529 rc = ptlrpc_disconnect_import(imp, 0);
530 down_write(&cli->cl_sem);
532 ptlrpc_invalidate_import(imp);
533 /* set obd_namespace to NULL only after invalidate, because we can have
534 * some connect requests in flight, and his need store a connect flags
535 * in obd_namespace. bug 14260 */
536 obd->obd_namespace = NULL;
538 if (imp->imp_rq_pool) {
539 ptlrpc_free_rq_pool(imp->imp_rq_pool);
540 imp->imp_rq_pool = NULL;
542 client_destroy_import(imp);
543 cli->cl_import = NULL;
548 /* use server style - class_disconnect should be always called for
550 err = class_disconnect(exp);
554 up_write(&cli->cl_sem);
556 ldlm_namespace_free_post(to_be_freed);
561 /* --------------------------------------------------------------------------
562 * from old lib/target.c
563 * -------------------------------------------------------------------------- */
565 static int target_handle_reconnect(struct lustre_handle *conn,
566 struct obd_export *exp,
567 struct obd_uuid *cluuid)
570 if (exp->exp_connection && exp->exp_imp_reverse) {
571 struct lustre_handle *hdl;
572 hdl = &exp->exp_imp_reverse->imp_remote_handle;
573 /* Might be a re-connect after a partition. */
574 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
575 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
577 conn->cookie = exp->exp_handle.h_cookie;
578 /* target_handle_connect() treats EALREADY and
579 * -EALREADY differently. EALREADY means we are
580 * doing a valid reconnect from the same client. */
583 CERROR("%s reconnecting from %s, "
584 "handle mismatch (ours "LPX64", theirs "
585 LPX64")\n", cluuid->uuid,
586 exp->exp_connection->c_remote_uuid.uuid,
587 hdl->cookie, conn->cookie);
588 memset(conn, 0, sizeof *conn);
589 /* target_handle_connect() treats EALREADY and
590 * -EALREADY differently. -EALREADY is an error
591 * (same UUID, different handle). */
596 conn->cookie = exp->exp_handle.h_cookie;
597 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
598 cluuid->uuid, exp, conn->cookie);
602 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
605 struct obd_export *exp = cb_data;
607 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
608 obd->obd_name, exp->exp_client_uuid.uuid);
610 spin_lock(&exp->exp_lock);
611 exp->exp_need_sync = 0;
612 spin_unlock(&exp->exp_lock);
614 EXPORT_SYMBOL(target_client_add_cb);
616 target_start_and_reset_recovery_timer(struct obd_device *obd,
617 struct ptlrpc_request *req,
620 int target_handle_connect(struct ptlrpc_request *req)
622 struct obd_device *target, *targref = NULL;
623 struct obd_export *export = NULL;
624 struct obd_import *revimp;
625 struct lustre_handle conn;
626 struct lustre_handle *tmp;
627 struct obd_uuid tgtuuid;
628 struct obd_uuid cluuid;
629 struct obd_uuid remote_uuid;
633 struct obd_connect_data *data, *tmpdata;
634 lnet_nid_t *client_nid = NULL;
637 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
639 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
641 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
642 GOTO(out, rc = -EINVAL);
645 obd_str2uuid(&tgtuuid, str);
646 target = class_uuid2obd(&tgtuuid);
648 target = class_name2obd(str);
650 if (!target || target->obd_stopping || !target->obd_set_up) {
651 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
652 " for connect (%s)\n", str,
653 !target ? "no target" :
654 (target->obd_stopping ? "stopping" :
656 GOTO(out, rc = -ENODEV);
659 if (target->obd_no_conn) {
660 LCONSOLE_WARN("%s: temporarily refusing client connection "
661 "from %s\n", target->obd_name,
662 libcfs_nid2str(req->rq_peer.nid));
663 GOTO(out, rc = -EAGAIN);
666 /* Make sure the target isn't cleaned up while we're here. Yes,
667 there's still a race between the above check and our incref here.
668 Really, class_uuid2obd should take the ref. */
669 targref = class_incref(target, __FUNCTION__, cfs_current());
672 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
674 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
675 GOTO(out, rc = -EINVAL);
678 obd_str2uuid(&cluuid, str);
680 /* XXX extract a nettype and format accordingly */
681 switch (sizeof(lnet_nid_t)) {
682 /* NB the casts only avoid compiler warnings */
684 snprintf(remote_uuid.uuid, sizeof remote_uuid,
685 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
688 snprintf(remote_uuid.uuid, sizeof remote_uuid,
689 "NET_%x_UUID", (__u32)req->rq_peer.nid);
695 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
697 GOTO(out, rc = -EPROTO);
701 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
703 GOTO(out, rc = -EPROTO);
705 rc = req_capsule_server_pack(&req->rq_pill);
709 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
711 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
712 "libclient connection attempt");
713 GOTO(out, rc = -EPROTO);
714 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
715 LUSTRE_VERSION_ALLOWED_OFFSET ||
716 data->ocd_version > LUSTRE_VERSION_CODE +
717 LUSTRE_VERSION_ALLOWED_OFFSET) {
718 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
719 "libclient connection attempt",
720 data->ocd_version < LUSTRE_VERSION_CODE ?
722 OBD_OCD_VERSION_MAJOR(data->ocd_version),
723 OBD_OCD_VERSION_MINOR(data->ocd_version),
724 OBD_OCD_VERSION_PATCH(data->ocd_version),
725 OBD_OCD_VERSION_FIX(data->ocd_version));
726 data = req_capsule_server_sized_get(&req->rq_pill,
728 offsetof(typeof(*data), ocd_version) +
729 sizeof(data->ocd_version));
731 data->ocd_connect_flags = OBD_CONNECT_VERSION;
732 data->ocd_version = LUSTRE_VERSION_CODE;
734 GOTO(out, rc = -EPROTO);
738 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
739 (data->ocd_connect_flags & OBD_CONNECT_MDS))
742 /* lctl gets a backstage, all-access pass. */
743 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
744 goto dont_check_exports;
746 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
750 /* we've found an export in the hash */
751 if (export->exp_connecting) { /* bug 9635, et. al. */
752 CWARN("%s: exp %p already connecting\n",
753 export->exp_obd->obd_name, export);
754 class_export_put(export);
757 } else if (mds_conn && export->exp_connection) {
758 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
759 /* mds reconnected after failover */
760 CWARN("%s: received MDS connection from NID %s,"
761 " removing former export from NID %s\n",
762 target->obd_name, libcfs_nid2str(req->rq_peer.nid),
763 libcfs_nid2str(export->exp_connection->c_peer.nid));
765 /* new mds connection from the same nid */
766 CWARN("%s: received new MDS connection from NID %s,"
767 " removing former export from same NID\n",
768 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
769 class_fail_export(export);
770 class_export_put(export);
773 } else if (export->exp_connection != NULL &&
774 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
775 /* in mds failover we have static uuid but nid can be
777 CWARN("%s: cookie %s seen on new NID %s when "
778 "existing NID %s is already connected\n",
779 target->obd_name, cluuid.uuid,
780 libcfs_nid2str(req->rq_peer.nid),
781 libcfs_nid2str(export->exp_connection->c_peer.nid));
783 class_export_put(export);
786 spin_lock(&export->exp_lock);
787 export->exp_connecting = 1;
788 spin_unlock(&export->exp_lock);
789 class_export_put(export);
790 LASSERT(export->exp_obd == target);
792 rc = target_handle_reconnect(&conn, export, &cluuid);
795 /* If we found an export, we already unlocked. */
798 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
799 } else if (req->rq_export == NULL &&
800 atomic_read(&export->exp_rpc_count) > 0) {
801 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
802 target->obd_name, cluuid.uuid,
803 libcfs_nid2str(req->rq_peer.nid),
804 export, atomic_read(&export->exp_refcount));
805 GOTO(out, rc = -EBUSY);
806 } else if (req->rq_export != NULL &&
807 (atomic_read(&export->exp_rpc_count) > 1)) {
808 /* the current connect rpc has increased exp_rpc_count */
809 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
810 target->obd_name, cluuid.uuid,
811 libcfs_nid2str(req->rq_peer.nid),
812 export, atomic_read(&export->exp_rpc_count) - 1);
813 spin_lock(&export->exp_lock);
814 if (req->rq_export->exp_conn_cnt <
815 lustre_msg_get_conn_cnt(req->rq_reqmsg))
816 /* try to abort active requests */
817 req->rq_export->exp_abort_active_req = 1;
818 spin_unlock(&export->exp_lock);
819 GOTO(out, rc = -EBUSY);
820 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
821 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
822 "cookies not random?\n", target->obd_name,
823 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
824 GOTO(out, rc = -EALREADY);
826 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
833 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
834 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
835 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
836 export, (long)cfs_time_current_sec(),
837 export ? (long)export->exp_last_request_time : 0);
839 /* Tell the client if we're in recovery. */
840 if (target->obd_recovering) {
841 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
842 /* If this is the first time a client connects,
843 reset the recovery timer */
845 target_start_and_reset_recovery_timer(target, req,
849 /* We want to handle EALREADY but *not* -EALREADY from
850 * target_handle_reconnect(), return reconnection state in a flag */
851 if (rc == EALREADY) {
852 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
858 /* Tell the client if we support replayable requests */
859 if (target->obd_replayable)
860 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
861 client_nid = &req->rq_peer.nid;
863 if (export == NULL) {
864 if (target->obd_recovering) {
867 t = cfs_timer_deadline(&target->obd_recovery_timer);
868 t = cfs_time_sub(t, cfs_time_current());
869 CERROR("%s: denying connection for new client %s (%s): "
870 "%d clients in recovery for "CFS_TIME_T"s\n",
872 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
873 target->obd_recoverable_clients,
874 cfs_duration_sec(t));
878 rc = obd_connect(req->rq_svc_thread->t_env,
879 &export, target, &cluuid, data,
882 conn.cookie = export->exp_handle.h_cookie;
885 rc = obd_reconnect(req->rq_svc_thread->t_env,
886 export, target, &cluuid, data, client_nid);
888 /* prevous done via class_conn2export */
889 class_export_get(export);
893 /* Return only the parts of obd_connect_data that we understand, so the
894 * client knows that we don't understand the rest. */
896 tmpdata = req_capsule_server_get(&req->rq_pill,
898 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
902 /* If all else goes well, this is our RPC return code. */
905 lustre_msg_set_handle(req->rq_repmsg, &conn);
907 /* If the client and the server are the same node, we will already
908 * have an export that really points to the client's DLM export,
909 * because we have a shared handles table.
911 * XXX this will go away when shaver stops sending the "connect" handle
912 * in the real "remote handle" field of the request --phik 24 Apr 2003
914 if (req->rq_export != NULL)
915 class_export_put(req->rq_export);
917 req->rq_export = export;
919 spin_lock(&export->exp_lock);
920 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
921 spin_unlock(&export->exp_lock);
922 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
923 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
924 export->exp_conn_cnt,
925 lustre_msg_get_conn_cnt(req->rq_reqmsg));
927 GOTO(out, rc = -EALREADY);
929 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
930 export->exp_abort_active_req = 0;
932 /* request from liblustre? Don't evict it for not pinging. */
933 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
934 export->exp_libclient = 1;
935 spin_unlock(&export->exp_lock);
937 spin_lock(&target->obd_dev_lock);
938 list_del_init(&export->exp_obd_chain_timed);
939 spin_unlock(&target->obd_dev_lock);
941 spin_unlock(&export->exp_lock);
944 if (export->exp_connection != NULL)
945 ptlrpc_connection_put(export->exp_connection);
946 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
949 if (hlist_unhashed(&export->exp_nid_hash)) {
950 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
951 &export->exp_connection->c_peer.nid,
952 &export->exp_nid_hash);
955 spin_lock_bh(&target->obd_processing_task_lock);
956 if (target->obd_recovering && !export->exp_in_recovery) {
957 spin_lock(&export->exp_lock);
958 export->exp_in_recovery = 1;
959 export->exp_req_replay_needed = 1;
960 export->exp_lock_replay_needed = 1;
961 spin_unlock(&export->exp_lock);
962 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
963 && (data->ocd_transno == 0))
964 CWARN("Connect with zero transno!\n");
966 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
967 && data->ocd_transno < target->obd_next_recovery_transno)
968 target->obd_next_recovery_transno = data->ocd_transno;
969 target->obd_connected_clients++;
970 /* each connected client is counted as recoverable */
971 target->obd_recoverable_clients++;
972 atomic_inc(&target->obd_req_replay_clients);
973 atomic_inc(&target->obd_lock_replay_clients);
974 if (target->obd_connected_clients ==
975 target->obd_max_recoverable_clients)
976 cfs_waitq_signal(&target->obd_next_transno_waitq);
978 spin_unlock_bh(&target->obd_processing_task_lock);
979 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
982 if (export->exp_imp_reverse != NULL) {
983 /* destroyed import can be still referenced in ctxt */
984 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
985 KEY_REVIMP_UPD, 0, NULL, NULL);
987 /* in some recovery senarios, previous ctx init rpc handled
988 * in sptlrpc_target_export_check() might be used to install
989 * a reverse ctx in this reverse import, and later OBD_CONNECT
990 * using the same gss ctx could reach here and following new
991 * reverse import. note all reverse ctx in new/old import are
992 * actually based on the same gss ctx. so we invalidate ctx
993 * here before destroy import, otherwise flush old import will
994 * lead to remote reverse ctx be destroied, thus the reverse
995 * ctx of new import will lost its peer.
996 * there might be a better way to deal with this???
998 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
1000 client_destroy_import(export->exp_imp_reverse);
1003 /* for the rest part, we return -ENOTCONN in case of errors
1004 * in order to let client initialize connection again.
1006 revimp = export->exp_imp_reverse = class_new_import(target);
1008 CERROR("fail to alloc new reverse import.\n");
1009 GOTO(out, rc = -ENOTCONN);
1012 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1013 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1014 revimp->imp_remote_handle = conn;
1015 revimp->imp_dlm_fake = 1;
1016 revimp->imp_state = LUSTRE_IMP_FULL;
1018 /* unknown versions will be caught in
1019 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1020 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1022 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1023 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1024 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1026 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1028 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1030 CERROR("Failed to get sec for reverse import: %d\n", rc);
1031 export->exp_imp_reverse = NULL;
1032 class_destroy_import(revimp);
1035 class_import_put(revimp);
1038 spin_lock(&export->exp_lock);
1039 export->exp_connecting = 0;
1040 spin_unlock(&export->exp_lock);
1043 class_decref(targref, __FUNCTION__, cfs_current());
1045 req->rq_status = rc;
1049 int target_handle_disconnect(struct ptlrpc_request *req)
1054 rc = req_capsule_server_pack(&req->rq_pill);
1058 /* keep the rq_export around so we can send the reply */
1059 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1064 void target_destroy_export(struct obd_export *exp)
1066 /* exports created from last_rcvd data, and "fake"
1067 exports created by lctl don't have an import */
1068 if (exp->exp_imp_reverse != NULL)
1069 client_destroy_import(exp->exp_imp_reverse);
1071 /* We cancel locks at disconnect time, but this will catch any locks
1072 * granted in a race with recovery-induced disconnect. */
1073 if (exp->exp_obd->obd_namespace != NULL)
1074 ldlm_cancel_locks_for_export(exp);
1078 * Recovery functions
1081 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1083 struct ptlrpc_request *copy_req;
1084 struct lustre_msg *copy_reqmsg;
1085 struct ptlrpc_user_desc *udesc = NULL;
1087 OBD_ALLOC_PTR(copy_req);
1090 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1092 OBD_FREE_PTR(copy_req);
1096 if (orig_req->rq_user_desc) {
1097 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1099 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1101 OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1102 OBD_FREE_PTR(copy_req);
1105 memcpy(udesc, orig_req->rq_user_desc,
1106 sptlrpc_user_desc_size(ngroups));
1109 *copy_req = *orig_req;
1110 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1111 copy_req->rq_reqmsg = copy_reqmsg;
1112 copy_req->rq_user_desc = udesc;
1114 class_export_rpc_get(copy_req->rq_export);
1115 CFS_INIT_LIST_HEAD(©_req->rq_list);
1116 CFS_INIT_LIST_HEAD(©_req->rq_replay_list);
1117 sptlrpc_svc_ctx_addref(copy_req);
1119 if (copy_req->rq_reply_state) {
1120 /* the copied req takes over the reply state */
1121 orig_req->rq_reply_state = NULL;
1122 /* to catch further access */
1123 orig_req->rq_repmsg = NULL;
1124 orig_req->rq_replen = 0;
1130 void ptlrpc_free_clone(struct ptlrpc_request *req)
1132 LASSERT(list_empty(&req->rq_replay_list));
1134 ptlrpc_req_drop_rs(req);
1135 sptlrpc_svc_ctx_decref(req);
1136 class_export_rpc_put(req->rq_export);
1137 list_del_init(&req->rq_list);
1139 if (req->rq_user_desc) {
1140 int ngroups = req->rq_user_desc->pud_ngroups;
1141 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1143 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1147 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1149 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1150 struct obd_export *exp = req->rq_export;
1151 struct ptlrpc_request *reqiter;
1156 spin_lock(&exp->exp_lock);
1157 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1159 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1166 /* we expect it with RESENT and REPLAY flags */
1167 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1168 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1169 CERROR("invalid flags %x of resent replay\n",
1170 lustre_msg_get_flags(req->rq_reqmsg));
1172 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1175 spin_unlock(&exp->exp_lock);
1179 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1181 LASSERT(!list_empty(&req->rq_replay_list));
1182 LASSERT(req->rq_export);
1184 spin_lock(&req->rq_export->exp_lock);
1185 list_del_init(&req->rq_replay_list);
1186 spin_unlock(&req->rq_export->exp_lock);
1190 static void target_finish_recovery(struct obd_device *obd)
1193 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1196 ldlm_reprocess_all_ns(obd->obd_namespace);
1197 spin_lock_bh(&obd->obd_processing_task_lock);
1198 if (list_empty(&obd->obd_req_replay_queue) &&
1199 list_empty(&obd->obd_lock_replay_queue) &&
1200 list_empty(&obd->obd_final_req_queue)) {
1201 obd->obd_processing_task = 0;
1203 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1205 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1206 list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1207 list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1208 spin_unlock_bh(&obd->obd_processing_task_lock);
1211 spin_unlock_bh(&obd->obd_processing_task_lock);
1213 /* when recovery finished, cleanup orphans on mds and ost */
1214 if (OBT(obd) && OBP(obd, postrecov)) {
1215 int rc = OBP(obd, postrecov)(obd);
1217 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1221 obd->obd_recovery_end = cfs_time_current_sec();
1225 static void abort_req_replay_queue(struct obd_device *obd)
1227 struct ptlrpc_request *req, *n;
1228 struct list_head abort_list;
1230 CFS_INIT_LIST_HEAD(&abort_list);
1231 spin_lock_bh(&obd->obd_processing_task_lock);
1232 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1233 spin_unlock_bh(&obd->obd_processing_task_lock);
1234 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1235 DEBUG_REQ(D_WARNING, req, "aborted:");
1236 req->rq_status = -ENOTCONN;
1237 if (ptlrpc_error(req)) {
1238 DEBUG_REQ(D_ERROR, req,
1239 "failed abort_req_reply; skipping");
1241 target_exp_dequeue_req_replay(req);
1242 ptlrpc_free_clone(req);
1246 static void abort_lock_replay_queue(struct obd_device *obd)
1248 struct ptlrpc_request *req, *n;
1249 struct list_head abort_list;
1251 CFS_INIT_LIST_HEAD(&abort_list);
1252 spin_lock_bh(&obd->obd_processing_task_lock);
1253 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1254 spin_unlock_bh(&obd->obd_processing_task_lock);
1255 list_for_each_entry_safe(req, n, &abort_list, rq_list){
1256 DEBUG_REQ(D_ERROR, req, "aborted:");
1257 req->rq_status = -ENOTCONN;
1258 if (ptlrpc_error(req)) {
1259 DEBUG_REQ(D_ERROR, req,
1260 "failed abort_lock_reply; skipping");
1262 ptlrpc_free_clone(req);
1267 /* Called from a cleanup function if the device is being cleaned up
1268 forcefully. The exports should all have been disconnected already,
1269 the only thing left to do is
1270 - clear the recovery flags
1272 - free queued requests and replies, but don't send replies
1273 Because the obd_stopping flag is set, no new requests should be received.
1276 void target_cleanup_recovery(struct obd_device *obd)
1278 struct ptlrpc_request *req, *n;
1279 struct list_head clean_list;
1282 LASSERT(obd->obd_stopping);
1284 CFS_INIT_LIST_HEAD(&clean_list);
1285 spin_lock_bh(&obd->obd_processing_task_lock);
1286 if (!obd->obd_recovering) {
1287 spin_unlock_bh(&obd->obd_processing_task_lock);
1291 obd->obd_recovering = obd->obd_abort_recovery = 0;
1292 target_cancel_recovery_timer(obd);
1294 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1295 spin_unlock_bh(&obd->obd_processing_task_lock);
1297 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1298 LASSERT(req->rq_reply_state == 0);
1299 target_exp_dequeue_req_replay(req);
1300 ptlrpc_free_clone(req);
1303 spin_lock_bh(&obd->obd_processing_task_lock);
1304 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1305 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1306 spin_unlock_bh(&obd->obd_processing_task_lock);
1308 list_for_each_entry_safe(req, n, &clean_list, rq_list){
1309 LASSERT(req->rq_reply_state == 0);
1310 ptlrpc_free_clone(req);
1316 /* obd_processing_task_lock should be held */
1317 void target_cancel_recovery_timer(struct obd_device *obd)
1319 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1320 cfs_timer_disarm(&obd->obd_recovery_timer);
1323 /* extend = 1 means require at least "duration" seconds left in the timer,
1324 extend = 0 means set the total duration (start_recovery_timer) */
1325 static void reset_recovery_timer(struct obd_device *obd, int duration,
1328 cfs_time_t now = cfs_time_current_sec();
1329 cfs_duration_t left;
1331 spin_lock_bh(&obd->obd_processing_task_lock);
1332 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1333 spin_unlock_bh(&obd->obd_processing_task_lock);
1337 left = cfs_time_sub(obd->obd_recovery_end, now);
1339 if (extend && (duration > left))
1340 obd->obd_recovery_timeout += duration - left;
1341 else if (!extend && (duration > obd->obd_recovery_timeout))
1342 /* Track the client's largest expected replay time */
1343 obd->obd_recovery_timeout = duration;
1346 * If total recovery time already exceed the
1347 * obd_recovery_max_time, then CRAY XT3 will
1348 * abort the recovery
1350 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1351 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1353 obd->obd_recovery_end = obd->obd_recovery_start +
1354 obd->obd_recovery_timeout;
1355 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1356 cfs_time_before(now, obd->obd_recovery_end)) {
1357 left = cfs_time_sub(obd->obd_recovery_end, now);
1358 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1360 spin_unlock_bh(&obd->obd_processing_task_lock);
1361 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1362 obd->obd_name, (unsigned)left);
1365 static void check_and_start_recovery_timer(struct obd_device *obd)
1367 spin_lock_bh(&obd->obd_processing_task_lock);
1368 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1369 spin_unlock_bh(&obd->obd_processing_task_lock);
1372 CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1373 obd->obd_recovery_start = cfs_time_current_sec();
1375 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1376 spin_unlock_bh(&obd->obd_processing_task_lock);
1378 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1381 /* Reset the timer with each new client connection */
1383 * This timer is actually reconnect_timer, which is for making sure
1384 * the total recovery window is at least as big as my reconnect
1385 * attempt timing. So the initial recovery time_out will be set to
1386 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1387 * from client is bigger than this, then the recovery time_out will
1388 * be extend to make sure the client could be reconnected, in the
1389 * process, the timeout from the new client should be ignored.
1393 target_start_and_reset_recovery_timer(struct obd_device *obd,
1394 struct ptlrpc_request *req,
1397 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1399 if (!new_client && service_time)
1400 /* Teach server about old server's estimates, as first guess
1401 * at how long new requests will take. */
1402 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1405 check_and_start_recovery_timer(obd);
1407 /* convert the service time to rpc timeout,
1408 * reuse service_time to limit stack usage */
1409 service_time = at_est2timeout(service_time);
1411 /* We expect other clients to timeout within service_time, then try
1412 * to reconnect, then try the failover server. The max delay between
1413 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1414 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1415 INITIAL_CONNECT_TIMEOUT);
1416 if (service_time > obd->obd_recovery_timeout && !new_client)
1417 reset_recovery_timer(obd, service_time, 0);
1421 static int check_for_next_transno(struct obd_device *obd)
1423 struct ptlrpc_request *req = NULL;
1424 int wake_up = 0, connected, completed, queue_len, max;
1425 __u64 next_transno, req_transno;
1427 spin_lock_bh(&obd->obd_processing_task_lock);
1429 if (!list_empty(&obd->obd_req_replay_queue)) {
1430 req = list_entry(obd->obd_req_replay_queue.next,
1431 struct ptlrpc_request, rq_list);
1432 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1437 max = obd->obd_max_recoverable_clients;
1438 connected = obd->obd_connected_clients;
1439 completed = connected - obd->obd_recoverable_clients;
1440 queue_len = obd->obd_requests_queued_for_recovery;
1441 next_transno = obd->obd_next_recovery_transno;
1443 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1444 "req_transno: "LPU64", next_transno: "LPU64"\n",
1445 max, connected, completed, queue_len, req_transno, next_transno);
1447 if (obd->obd_abort_recovery) {
1448 CDEBUG(D_HA, "waking for aborted recovery\n");
1450 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1451 CDEBUG(D_HA, "waking for completed recovery\n");
1453 } else if (req_transno == next_transno) {
1454 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1456 } else if (queue_len + completed == max) {
1457 /* handle gaps occured due to lost reply. It is allowed gaps
1458 * because all clients are connected and there will be resend
1459 * for missed transaction */
1460 LASSERTF(req_transno >= next_transno,
1461 "req_transno: "LPU64", next_transno: "LPU64"\n",
1462 req_transno, next_transno);
1464 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1465 "waking for skipped transno (skip: "LPD64
1466 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1467 next_transno, queue_len, completed, connected, req_transno);
1468 obd->obd_next_recovery_transno = req_transno;
1470 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1471 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1472 " by fail_lock, waking up ("LPD64")\n", next_transno);
1473 obd->obd_next_recovery_transno = req_transno;
1475 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1476 /* some clients haven't connected in time, but we can try
1477 * to replay requests that demand on already committed ones
1478 * also, we can replay first non-committed transation */
1479 LASSERT(req_transno != 0);
1480 if (obd->obd_version_recov ||
1481 req_transno == obd->obd_last_committed + 1) {
1482 obd->obd_next_recovery_transno = req_transno;
1483 } else if (req_transno > obd->obd_last_committed) {
1484 /* can't continue recovery: have no needed transno */
1485 obd->obd_abort_recovery = 1;
1486 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1487 "connected: %d, completed: %d, queue_len: %d, "
1488 "req_transno: "LPU64", next_transno: "LPU64"\n",
1489 max, connected, completed, queue_len,
1490 req_transno, next_transno);
1495 spin_unlock_bh(&obd->obd_processing_task_lock);
1499 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1501 struct l_wait_info lwi = { 0 };
1502 struct ptlrpc_request *req;
1504 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1505 obd->obd_next_recovery_transno);
1506 l_wait_event(obd->obd_next_transno_waitq,
1507 check_for_next_transno(obd), &lwi);
1509 spin_lock_bh(&obd->obd_processing_task_lock);
1510 if (obd->obd_abort_recovery) {
1512 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1513 req = list_entry(obd->obd_req_replay_queue.next,
1514 struct ptlrpc_request, rq_list);
1515 list_del_init(&req->rq_list);
1516 obd->obd_requests_queued_for_recovery--;
1520 spin_unlock_bh(&obd->obd_processing_task_lock);
1524 static int check_for_next_lock(struct obd_device *obd)
1526 struct ptlrpc_request *req = NULL;
1529 spin_lock_bh(&obd->obd_processing_task_lock);
1530 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 CDEBUG(D_HA, "waking for next lock\n");
1535 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1536 CDEBUG(D_HA, "waking for completed lock replay\n");
1538 } else if (obd->obd_abort_recovery) {
1539 CDEBUG(D_HA, "waking for aborted recovery\n");
1542 spin_unlock_bh(&obd->obd_processing_task_lock);
1547 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1549 struct l_wait_info lwi = { 0 };
1550 struct ptlrpc_request *req;
1552 CDEBUG(D_HA, "Waiting for lock\n");
1553 l_wait_event(obd->obd_next_transno_waitq,
1554 check_for_next_lock(obd), &lwi);
1556 spin_lock_bh(&obd->obd_processing_task_lock);
1557 if (obd->obd_abort_recovery) {
1559 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1560 req = list_entry(obd->obd_lock_replay_queue.next,
1561 struct ptlrpc_request, rq_list);
1562 list_del_init(&req->rq_list);
1566 spin_unlock_bh(&obd->obd_processing_task_lock);
1570 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1572 struct ptlrpc_request *req;
1574 spin_lock_bh(&obd->obd_processing_task_lock);
1575 if (!list_empty(&obd->obd_final_req_queue)) {
1576 req = list_entry(obd->obd_final_req_queue.next,
1577 struct ptlrpc_request, rq_list);
1578 list_del_init(&req->rq_list);
1579 if (req->rq_export->exp_in_recovery) {
1580 spin_lock(&req->rq_export->exp_lock);
1581 req->rq_export->exp_in_recovery = 0;
1582 spin_unlock(&req->rq_export->exp_lock);
1587 spin_unlock_bh(&obd->obd_processing_task_lock);
1591 static inline int req_vbr_done(struct obd_export *exp)
1593 return (exp->exp_vbr_failed == 0);
1596 static inline int req_replay_done(struct obd_export *exp)
1598 return (exp->exp_req_replay_needed == 0);
1601 static inline int lock_replay_done(struct obd_export *exp)
1603 return (exp->exp_lock_replay_needed == 0);
1606 static inline int connect_done(struct obd_export *exp)
1608 return (exp->exp_in_recovery != 0);
1611 static int check_for_clients(struct obd_device *obd)
1613 if (obd->obd_abort_recovery || obd->obd_version_recov)
1615 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1616 if (obd->obd_no_conn == 0 &&
1617 obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1622 static int handle_recovery_req(struct ptlrpc_thread *thread,
1623 struct ptlrpc_request *req,
1624 svc_handler_t handler)
1629 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1631 CERROR("Failure to initialize session: %d\n", rc);
1634 req->rq_session.lc_thread = thread;
1635 lu_context_enter(&req->rq_session);
1636 req->rq_svc_thread = thread;
1637 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1639 /* thread context */
1640 lu_context_enter(&thread->t_env->le_ctx);
1642 lu_context_exit(&thread->t_env->le_ctx);
1644 lu_context_exit(&req->rq_session);
1645 lu_context_fini(&req->rq_session);
1646 /* don't reset timer for final stage */
1647 if (!req_replay_done(req->rq_export) ||
1648 !lock_replay_done(req->rq_export))
1649 reset_recovery_timer(class_exp2obd(req->rq_export),
1650 AT_OFF ? obd_timeout :
1651 at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1654 * bz18031: increase next_recovery_transno before ptlrpc_free_clone()
1655 * will drop exp_rpc reference
1657 if (!req_replay_done(req->rq_export)) {
1658 spin_lock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1659 req->rq_export->exp_obd->obd_next_recovery_transno++;
1660 spin_unlock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1661 target_exp_dequeue_req_replay(req);
1663 ptlrpc_free_clone(req);
1667 static int target_recovery_thread(void *arg)
1669 struct lu_target *lut = arg;
1670 struct obd_device *obd = lut->lut_obd;
1671 struct ptlrpc_request *req;
1672 struct target_recovery_data *trd = &obd->obd_recovery_data;
1673 struct l_wait_info lwi = { 0 };
1674 unsigned long delta;
1675 unsigned long flags;
1677 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1681 cfs_daemonize("tgt_recov");
1683 SIGNAL_MASK_LOCK(current, flags);
1684 sigfillset(¤t->blocked);
1686 SIGNAL_MASK_UNLOCK(current, flags);
1688 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1692 thread->t_env = &env;
1693 env.le_ctx.lc_thread = thread;
1695 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1697 trd->trd_processing_task = cfs_curproc_pid();
1699 obd->obd_recovering = 1;
1700 complete(&trd->trd_starting);
1702 /* first of all, we have to know the first transno to replay */
1703 obd->obd_abort_recovery = 0;
1704 l_wait_event(obd->obd_next_transno_waitq,
1705 check_for_clients(obd), &lwi);
1707 /* If some clients haven't connected in time, evict them */
1708 if (obd->obd_connected_clients < obd->obd_max_recoverable_clients) {
1709 CWARN("Some clients haven't connect in time (%d/%d),"
1710 "evict them\n", obd->obd_connected_clients,
1711 obd->obd_max_recoverable_clients);
1712 class_disconnect_stale_exports(obd, connect_done,
1713 exp_flags_from_obd(obd) |
1714 OBD_OPT_ABORT_RECOV);
1716 * if recovery proceeds with versions then some clients may be
1717 * timed out waiting for others and trying to reconnect.
1718 * Extend timer for such reconnect cases.
1720 if (obd->obd_version_recov)
1721 reset_recovery_timer(obd, RECONNECT_DELAY_MAX * 2, 1);
1724 /* next stage: replay requests */
1726 obd->obd_req_replaying = 1;
1727 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1728 atomic_read(&obd->obd_req_replay_clients),
1729 obd->obd_next_recovery_transno);
1730 while ((req = target_next_replay_req(obd))) {
1731 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1732 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1733 lustre_msg_get_transno(req->rq_reqmsg),
1734 libcfs_nid2str(req->rq_peer.nid));
1735 handle_recovery_req(thread, req,
1736 trd->trd_recovery_handler);
1737 obd->obd_replayed_requests++;
1740 /* If some clients haven't replayed requests in time, evict them */
1741 if (obd->obd_abort_recovery) {
1742 CDEBUG(D_WARNING, "req replay is aborted\n");
1743 class_disconnect_stale_exports(obd, req_replay_done,
1744 exp_flags_from_obd(obd) |
1745 OBD_OPT_ABORT_RECOV);
1746 abort_req_replay_queue(obd);
1748 LASSERT(list_empty(&obd->obd_req_replay_queue));
1750 /* The second stage: replay locks */
1751 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1752 atomic_read(&obd->obd_lock_replay_clients));
1753 while ((req = target_next_replay_lock(obd))) {
1754 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1755 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1756 libcfs_nid2str(req->rq_peer.nid));
1757 handle_recovery_req(thread, req,
1758 trd->trd_recovery_handler);
1759 obd->obd_replayed_locks++;
1762 /* If some clients haven't replayed requests in time, evict them */
1763 if (obd->obd_abort_recovery) {
1764 CERROR("lock replay is aborted\n");
1765 class_disconnect_stale_exports(obd, lock_replay_done,
1766 exp_flags_from_obd(obd) |
1767 OBD_OPT_ABORT_RECOV);
1768 abort_lock_replay_queue(obd);
1770 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1772 /* The third stage: reply on final pings */
1773 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1774 /** evict exports failed VBR */
1775 class_disconnect_stale_exports(obd, req_vbr_done,
1776 exp_flags_from_obd(obd) |
1777 OBD_OPT_ABORT_RECOV);
1778 /** Update server last boot epoch */
1779 lut_boot_epoch_update(lut);
1780 /* We drop recoverying flag to forward all new requests
1781 * to regular mds_handle() since now */
1782 spin_lock_bh(&obd->obd_processing_task_lock);
1783 obd->obd_recovering = obd->obd_abort_recovery = 0;
1784 target_cancel_recovery_timer(obd);
1785 spin_unlock_bh(&obd->obd_processing_task_lock);
1786 while ((req = target_next_final_ping(obd))) {
1787 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1788 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1789 libcfs_nid2str(req->rq_peer.nid));
1790 handle_recovery_req(thread, req,
1791 trd->trd_recovery_handler);
1794 delta = (jiffies - delta) / HZ;
1795 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1796 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1797 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1798 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1799 if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1800 CWARN("too long recovery - read logs\n");
1801 libcfs_debug_dumplog();
1804 target_finish_recovery(obd);
1806 lu_context_fini(&env.le_ctx);
1807 trd->trd_processing_task = 0;
1808 complete(&trd->trd_finishing);
1812 static int target_start_recovery_thread(struct lu_target *lut,
1813 svc_handler_t handler)
1815 struct obd_device *obd = lut->lut_obd;
1817 struct target_recovery_data *trd = &obd->obd_recovery_data;
1819 memset(trd, 0, sizeof(*trd));
1820 init_completion(&trd->trd_starting);
1821 init_completion(&trd->trd_finishing);
1822 trd->trd_recovery_handler = handler;
1824 if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1825 wait_for_completion(&trd->trd_starting);
1826 LASSERT(obd->obd_recovering != 0);
1833 void target_stop_recovery_thread(struct obd_device *obd)
1835 spin_lock_bh(&obd->obd_processing_task_lock);
1836 if (obd->obd_recovery_data.trd_processing_task > 0) {
1837 struct target_recovery_data *trd = &obd->obd_recovery_data;
1838 CERROR("%s: Aborting recovery\n", obd->obd_name);
1839 obd->obd_abort_recovery = 1;
1840 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1841 spin_unlock_bh(&obd->obd_processing_task_lock);
1842 wait_for_completion(&trd->trd_finishing);
1844 spin_unlock_bh(&obd->obd_processing_task_lock);
1848 void target_recovery_fini(struct obd_device *obd)
1850 class_disconnect_exports(obd);
1851 target_stop_recovery_thread(obd);
1852 target_cleanup_recovery(obd);
1854 EXPORT_SYMBOL(target_recovery_fini);
1856 static void target_recovery_expired(unsigned long castmeharder)
1858 struct obd_device *obd = (struct obd_device *)castmeharder;
1859 CDEBUG(D_HA, "%s: recovery timed out; %d clients never reconnected "
1860 "after %lds (%d clients did)\n",
1861 obd->obd_name, obd->obd_recoverable_clients,
1862 cfs_time_current_sec()- obd->obd_recovery_start,
1863 obd->obd_connected_clients);
1865 spin_lock_bh(&obd->obd_processing_task_lock);
1866 obd->obd_version_recov = 1;
1867 CDEBUG(D_INFO, "VBR is used for %d clients from t"LPU64"\n",
1868 atomic_read(&obd->obd_req_replay_clients),
1869 obd->obd_next_recovery_transno);
1870 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1871 spin_unlock_bh(&obd->obd_processing_task_lock);
1874 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1876 struct obd_device *obd = lut->lut_obd;
1877 if (obd->obd_max_recoverable_clients == 0)
1880 CWARN("RECOVERY: service %s, %d recoverable clients, "
1881 "last_transno "LPU64"\n", obd->obd_name,
1882 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1883 LASSERT(obd->obd_stopping == 0);
1884 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1885 obd->obd_recovery_start = 0;
1886 obd->obd_recovery_end = 0;
1887 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1888 /* bz13079: this should be set to desired value for ost but not for mds */
1889 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1890 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1891 target_start_recovery_thread(lut, handler);
1893 EXPORT_SYMBOL(target_recovery_init);
1897 static int target_process_req_flags(struct obd_device *obd,
1898 struct ptlrpc_request *req)
1900 struct obd_export *exp = req->rq_export;
1901 LASSERT(exp != NULL);
1902 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1903 /* client declares he's ready to replay locks */
1904 spin_lock_bh(&obd->obd_processing_task_lock);
1905 if (exp->exp_req_replay_needed) {
1906 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1907 spin_lock(&exp->exp_lock);
1908 exp->exp_req_replay_needed = 0;
1909 spin_unlock(&exp->exp_lock);
1910 atomic_dec(&obd->obd_req_replay_clients);
1911 LASSERT(obd->obd_recoverable_clients > 0);
1912 obd->obd_recoverable_clients--;
1913 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1914 CDEBUG(D_HA, "all clients have replayed reqs\n");
1916 spin_unlock_bh(&obd->obd_processing_task_lock);
1918 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1919 /* client declares he's ready to complete recovery
1920 * so, we put the request on th final queue */
1921 spin_lock_bh(&obd->obd_processing_task_lock);
1922 if (exp->exp_lock_replay_needed) {
1923 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1924 spin_lock(&exp->exp_lock);
1925 exp->exp_lock_replay_needed = 0;
1926 spin_unlock(&exp->exp_lock);
1927 atomic_dec(&obd->obd_lock_replay_clients);
1928 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1929 CDEBUG(D_HA, "all clients have replayed locks\n");
1931 spin_unlock_bh(&obd->obd_processing_task_lock);
1937 int target_queue_recovery_request(struct ptlrpc_request *req,
1938 struct obd_device *obd)
1940 struct list_head *tmp;
1942 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1945 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1946 /* Processing the queue right now, don't re-add. */
1950 target_process_req_flags(obd, req);
1952 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1953 /* client declares he's ready to complete recovery
1954 * so, we put the request on th final queue */
1955 req = ptlrpc_clone_req(req);
1958 DEBUG_REQ(D_HA, req, "queue final req");
1959 spin_lock_bh(&obd->obd_processing_task_lock);
1960 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1961 if (obd->obd_recovering)
1962 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1964 spin_unlock_bh(&obd->obd_processing_task_lock);
1965 ptlrpc_free_clone(req);
1966 if (obd->obd_stopping) {
1972 spin_unlock_bh(&obd->obd_processing_task_lock);
1975 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1976 /* client declares he's ready to replay locks */
1977 req = ptlrpc_clone_req(req);
1980 DEBUG_REQ(D_HA, req, "queue lock replay req");
1981 spin_lock_bh(&obd->obd_processing_task_lock);
1982 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1983 LASSERT(obd->obd_recovering);
1984 /* usually due to recovery abort */
1985 if (!req->rq_export->exp_in_recovery) {
1986 spin_unlock_bh(&obd->obd_processing_task_lock);
1987 ptlrpc_free_clone(req);
1990 LASSERT(req->rq_export->exp_lock_replay_needed);
1991 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1992 spin_unlock_bh(&obd->obd_processing_task_lock);
1996 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1997 * (i.e. buflens etc are in my own byte order), but type-dependent
1998 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2001 CFS_INIT_LIST_HEAD(&req->rq_list);
2002 DEBUG_REQ(D_HA, req, "not queueing");
2006 spin_lock_bh(&obd->obd_processing_task_lock);
2008 /* If we're processing the queue, we want don't want to queue this
2011 * Also, if this request has a transno less than the one we're waiting
2012 * for, we should process it now. It could (and currently always will)
2013 * be an open request for a descriptor that was opened some time ago.
2015 * Also, a resent, replayed request that has already been
2016 * handled will pass through here and be processed immediately.
2018 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2019 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2020 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2021 /* Processing the queue right now, don't re-add. */
2022 LASSERT(list_empty(&req->rq_list));
2023 spin_unlock_bh(&obd->obd_processing_task_lock);
2026 spin_unlock_bh(&obd->obd_processing_task_lock);
2028 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2031 req = ptlrpc_clone_req(req);
2035 spin_lock_bh(&obd->obd_processing_task_lock);
2036 LASSERT(obd->obd_recovering);
2037 if (!req->rq_export->exp_in_recovery) {
2038 spin_unlock_bh(&obd->obd_processing_task_lock);
2039 ptlrpc_free_clone(req);
2042 LASSERT(req->rq_export->exp_req_replay_needed);
2044 if (target_exp_enqueue_req_replay(req)) {
2045 spin_unlock_bh(&obd->obd_processing_task_lock);
2046 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2047 ptlrpc_free_clone(req);
2052 list_for_each(tmp, &obd->obd_req_replay_queue) {
2053 struct ptlrpc_request *reqiter =
2054 list_entry(tmp, struct ptlrpc_request, rq_list);
2056 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2057 list_add_tail(&req->rq_list, &reqiter->rq_list);
2062 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2064 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2065 "has been claimed by another client");
2066 spin_unlock_bh(&obd->obd_processing_task_lock);
2067 target_exp_dequeue_req_replay(req);
2068 ptlrpc_free_clone(req);
2074 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2076 obd->obd_requests_queued_for_recovery++;
2077 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2078 spin_unlock_bh(&obd->obd_processing_task_lock);
2082 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2084 return req->rq_export->exp_obd;
2087 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2089 LASSERT(exp != NULL);
2090 return &exp->exp_obd->obd_namespace->ns_pool;
2094 * Packs current SLV and Limit into \a req.
2096 int target_pack_pool_reply(struct ptlrpc_request *req)
2098 struct obd_device *obd;
2102 * Check that we still have all structures alive as this may
2103 * be some late rpc in shutdown time.
2105 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2106 !exp_connect_lru_resize(req->rq_export))) {
2107 lustre_msg_set_slv(req->rq_repmsg, 0);
2108 lustre_msg_set_limit(req->rq_repmsg, 0);
2113 * OBD is alive here as export is alive, which we checked above.
2115 obd = req->rq_export->exp_obd;
2117 read_lock(&obd->obd_pool_lock);
2118 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2119 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2120 read_unlock(&obd->obd_pool_lock);
2125 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2127 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2128 DEBUG_REQ(D_ERROR, req, "dropping reply");
2133 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2134 req->rq_status = rc;
2135 return (ptlrpc_send_error(req, 1));
2137 DEBUG_REQ(D_NET, req, "sending reply");
2140 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2143 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2146 struct ptlrpc_reply_state *rs;
2147 struct obd_device *obd;
2148 struct obd_export *exp;
2149 struct ptlrpc_service *svc;
2152 if (req->rq_no_reply) {
2157 svc = req->rq_rqbd->rqbd_service;
2158 rs = req->rq_reply_state;
2159 if (rs == NULL || !rs->rs_difficult) {
2161 target_send_reply_msg (req, rc, fail_id);
2166 /* must be an export if locks saved */
2167 LASSERT (req->rq_export != NULL);
2168 /* req/reply consistent */
2169 LASSERT (rs->rs_service == svc);
2172 LASSERT (!rs->rs_scheduled);
2173 LASSERT (!rs->rs_scheduled_ever);
2174 LASSERT (!rs->rs_handled);
2175 LASSERT (!rs->rs_on_net);
2176 LASSERT (rs->rs_export == NULL);
2177 LASSERT (list_empty(&rs->rs_obd_list));
2178 LASSERT (list_empty(&rs->rs_exp_list));
2180 exp = class_export_get (req->rq_export);
2183 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2184 rs->rs_scheduled = 1;
2186 rs->rs_xid = req->rq_xid;
2187 rs->rs_transno = req->rq_transno;
2188 rs->rs_export = exp;
2190 spin_lock(&exp->exp_uncommitted_replies_lock);
2191 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2192 rs->rs_transno, exp->exp_last_committed);
2193 if (rs->rs_transno > exp->exp_last_committed) {
2194 /* not committed already */
2195 list_add_tail(&rs->rs_obd_list,
2196 &exp->exp_uncommitted_replies);
2198 spin_unlock (&exp->exp_uncommitted_replies_lock);
2200 spin_lock(&exp->exp_lock);
2201 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2202 spin_unlock(&exp->exp_lock);
2204 netrc = target_send_reply_msg (req, rc, fail_id);
2206 spin_lock(&svc->srv_lock);
2208 atomic_inc(&svc->srv_n_difficult_replies);
2211 /* error sending: reply is off the net. Also we need +1
2212 * reply ref until ptlrpc_server_handle_reply() is done
2213 * with the reply state (if the send was successful, there
2214 * would have been +1 ref for the net, which
2215 * reply_out_callback leaves alone) */
2217 ptlrpc_rs_addref(rs);
2218 atomic_inc (&svc->srv_outstanding_replies);
2221 spin_lock(&rs->rs_lock);
2222 if (rs->rs_transno <= exp->exp_last_committed ||
2223 (!rs->rs_on_net && !rs->rs_no_ack) ||
2224 list_empty(&rs->rs_exp_list) || /* completed already */
2225 list_empty(&rs->rs_obd_list)) {
2226 CDEBUG(D_HA, "Schedule reply immediately\n");
2227 ptlrpc_dispatch_difficult_reply(rs);
2229 list_add (&rs->rs_list, &svc->srv_active_replies);
2230 rs->rs_scheduled = 0; /* allow notifier to schedule */
2232 spin_unlock(&rs->rs_lock);
2233 spin_unlock(&svc->srv_lock);
2237 int target_handle_ping(struct ptlrpc_request *req)
2239 obd_ping(req->rq_export);
2240 return req_capsule_server_pack(&req->rq_pill);
2243 void target_committed_to_req(struct ptlrpc_request *req)
2245 struct obd_export *exp = req->rq_export;
2247 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2248 lustre_msg_set_last_committed(req->rq_repmsg,
2249 exp->exp_last_committed);
2251 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2252 "%d)", exp->exp_obd->obd_no_transno,
2253 req->rq_repmsg == NULL);
2255 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2256 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2258 EXPORT_SYMBOL(target_committed_to_req);
2260 int target_handle_qc_callback(struct ptlrpc_request *req)
2262 struct obd_quotactl *oqctl;
2263 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2265 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2266 if (oqctl == NULL) {
2267 CERROR("Can't unpack obd_quotactl\n");
2271 cli->cl_qchk_stat = oqctl->qc_stat;
2276 #ifdef HAVE_QUOTA_SUPPORT
2277 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2280 struct obd_device *obd = req->rq_export->exp_obd;
2281 struct obd_device *master_obd;
2282 struct obd_device_target *obt;
2283 struct lustre_quota_ctxt *qctxt;
2284 struct qunit_data *qdata = NULL;
2288 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2291 rc = req_capsule_server_pack(&req->rq_pill);
2293 CERROR("packing reply failed!: rc = %d\n", rc);
2297 LASSERT(req->rq_export);
2299 qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2300 if (IS_ERR(qdata)) {
2301 rc = PTR_ERR(qdata);
2302 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2303 req->rq_status = rc;
2307 /* we use the observer */
2308 if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2309 CERROR("Can't find the observer, it is recovering\n");
2310 req->rq_status = -EAGAIN;
2314 master_obd = obd->obd_observer->obd_observer;
2315 obt = &master_obd->u.obt;
2316 qctxt = &obt->obt_qctxt;
2318 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2319 /* quota_type has not been processed yet, return EAGAIN
2320 * until we know whether or not quotas are supposed to
2322 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2324 req->rq_status = -EAGAIN;
2328 down_read(&obt->obt_rwsem);
2329 if (qctxt->lqc_lqs_hash == NULL) {
2330 up_read(&obt->obt_rwsem);
2331 /* quota_type has not been processed yet, return EAGAIN
2332 * until we know whether or not quotas are supposed to
2334 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2336 req->rq_status = -EAGAIN;
2340 LASSERT(qctxt->lqc_handler);
2341 rc = qctxt->lqc_handler(master_obd, qdata,
2342 lustre_msg_get_opc(req->rq_reqmsg));
2343 up_read(&obt->obt_rwsem);
2344 if (rc && rc != -EDQUOT)
2345 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2346 "dqacq/dqrel failed! (rc:%d)\n", rc);
2347 req->rq_status = rc;
2349 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2351 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2355 /* Block the quota req. b=14840 */
2356 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2360 rc = ptlrpc_reply(req);
2364 #endif /* !__KERNEL__ */
2366 #endif /* HAVE_QUOTA_SUPPORT */
2368 ldlm_mode_t lck_compat_array[] = {
2369 [LCK_EX] LCK_COMPAT_EX,
2370 [LCK_PW] LCK_COMPAT_PW,
2371 [LCK_PR] LCK_COMPAT_PR,
2372 [LCK_CW] LCK_COMPAT_CW,
2373 [LCK_CR] LCK_COMPAT_CR,
2374 [LCK_NL] LCK_COMPAT_NL,
2375 [LCK_GROUP] LCK_COMPAT_GROUP,
2376 [LCK_COS] LCK_COMPAT_COS,
2380 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2381 * not escape to the user level.
2383 int ldlm_error2errno(ldlm_error_t error)
2391 case ELDLM_LOCK_CHANGED:
2394 case ELDLM_LOCK_ABORTED:
2397 case ELDLM_LOCK_REPLACED:
2400 case ELDLM_NO_LOCK_DATA:
2403 case ELDLM_NAMESPACE_EXISTS:
2406 case ELDLM_BAD_NAMESPACE:
2410 if (((int)error) < 0) /* cast to signed type */
2411 result = error; /* as ldlm_error_t can be unsigned */
2413 CERROR("Invalid DLM result code: %i\n", error);
2419 EXPORT_SYMBOL(ldlm_error2errno);
2422 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2424 ldlm_error_t ldlm_errno2error(int err_no)
2433 error = ELDLM_LOCK_CHANGED;
2436 error = ELDLM_LOCK_ABORTED;
2439 error = ELDLM_LOCK_REPLACED;
2442 error = ELDLM_NO_LOCK_DATA;
2445 error = ELDLM_NAMESPACE_EXISTS;
2448 error = ELDLM_BAD_NAMESPACE;
2455 EXPORT_SYMBOL(ldlm_errno2error);