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 lustre_handle *dlm_handle,
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_export *exp;
395 struct obd_connect_data *ocd;
396 struct ldlm_namespace *to_be_freed = NULL;
400 down_write(&cli->cl_sem);
401 rc = class_connect(dlm_handle, obd, cluuid);
405 cli->cl_conn_count++;
406 if (cli->cl_conn_count > 1)
408 exp = class_conn2export(dlm_handle);
410 if (obd->obd_namespace != NULL)
411 CERROR("already have namespace!\n");
412 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
413 LDLM_NAMESPACE_CLIENT,
414 LDLM_NAMESPACE_GREEDY);
415 if (obd->obd_namespace == NULL)
416 GOTO(out_disco, rc = -ENOMEM);
418 imp->imp_dlm_handle = *dlm_handle;
419 rc = ptlrpc_init_import(imp);
423 ocd = &imp->imp_connect_data;
426 imp->imp_connect_flags_orig = data->ocd_connect_flags;
429 rc = ptlrpc_connect_import(imp, NULL);
431 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
434 LASSERT(exp->exp_connection);
437 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
438 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
439 data->ocd_connect_flags, ocd->ocd_connect_flags);
440 data->ocd_connect_flags = ocd->ocd_connect_flags;
443 ptlrpc_pinger_add_import(imp);
449 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
450 to_be_freed = obd->obd_namespace;
451 obd->obd_namespace = NULL;
453 cli->cl_conn_count--;
454 class_disconnect(exp);
456 class_export_put(exp);
459 up_write(&cli->cl_sem);
461 ldlm_namespace_free_post(to_be_freed);
466 int client_disconnect_export(struct obd_export *exp)
468 struct obd_device *obd = class_exp2obd(exp);
469 struct client_obd *cli;
470 struct obd_import *imp;
472 struct ldlm_namespace *to_be_freed = NULL;
476 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
477 exp, exp ? exp->exp_handle.h_cookie : -1);
482 imp = cli->cl_import;
484 down_write(&cli->cl_sem);
485 if (!cli->cl_conn_count) {
486 CERROR("disconnecting disconnected device (%s)\n",
488 GOTO(out_sem, rc = -EINVAL);
491 cli->cl_conn_count--;
492 if (cli->cl_conn_count)
493 GOTO(out_no_disconnect, rc = 0);
495 /* Mark import deactivated now, so we don't try to reconnect if any
496 * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't
497 * fully deactivate the import, or that would drop all requests. */
498 spin_lock(&imp->imp_lock);
499 imp->imp_deactive = 1;
500 spin_unlock(&imp->imp_lock);
502 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
503 * delete it regardless. (It's safe to delete an import that was
505 (void)ptlrpc_pinger_del_import(imp);
507 if (obd->obd_namespace != NULL) {
508 /* obd_force == local only */
509 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
510 obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
512 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
513 to_be_freed = obd->obd_namespace;
516 rc = ptlrpc_disconnect_import(imp, 0);
518 ptlrpc_invalidate_import(imp);
519 /* set obd_namespace to NULL only after invalidate, because we can have
520 * some connect requests in flight, and his need store a connect flags
521 * in obd_namespace. bug 14260 */
522 obd->obd_namespace = NULL;
524 if (imp->imp_rq_pool) {
525 ptlrpc_free_rq_pool(imp->imp_rq_pool);
526 imp->imp_rq_pool = NULL;
529 cli->cl_import = NULL;
533 err = class_disconnect(exp);
537 up_write(&cli->cl_sem);
539 ldlm_namespace_free_post(to_be_freed);
544 /* --------------------------------------------------------------------------
545 * from old lib/target.c
546 * -------------------------------------------------------------------------- */
548 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
549 struct obd_uuid *cluuid, int mds_conn)
552 if (exp->exp_connection && exp->exp_imp_reverse && !mds_conn) {
553 struct lustre_handle *hdl;
554 hdl = &exp->exp_imp_reverse->imp_remote_handle;
555 /* Might be a re-connect after a partition. */
556 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
557 CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
559 conn->cookie = exp->exp_handle.h_cookie;
560 /* target_handle_connect() treats EALREADY and
561 * -EALREADY differently. EALREADY means we are
562 * doing a valid reconnect from the same client. */
565 CERROR("%s reconnecting from %s, "
566 "handle mismatch (ours "LPX64", theirs "
567 LPX64")\n", cluuid->uuid,
568 exp->exp_connection->c_remote_uuid.uuid,
569 hdl->cookie, conn->cookie);
570 memset(conn, 0, sizeof *conn);
571 /* target_handle_connect() treats EALREADY and
572 * -EALREADY differently. -EALREADY is an error
573 * (same UUID, different handle). */
578 conn->cookie = exp->exp_handle.h_cookie;
579 CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
580 cluuid->uuid, exp, conn->cookie);
584 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
587 struct obd_export *exp = cb_data;
589 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
590 obd->obd_name, exp->exp_client_uuid.uuid);
592 spin_lock(&exp->exp_lock);
593 exp->exp_need_sync = 0;
594 spin_unlock(&exp->exp_lock);
596 EXPORT_SYMBOL(target_client_add_cb);
598 target_start_and_reset_recovery_timer(struct obd_device *obd,
599 struct ptlrpc_request *req,
602 int target_handle_connect(struct ptlrpc_request *req)
604 struct obd_device *target, *targref = NULL;
605 struct obd_export *export = NULL;
606 struct obd_import *revimp;
607 struct lustre_handle conn;
608 struct lustre_handle *tmp;
609 struct obd_uuid tgtuuid;
610 struct obd_uuid cluuid;
611 struct obd_uuid remote_uuid;
615 struct obd_connect_data *data, *tmpdata;
616 lnet_nid_t *client_nid = NULL;
619 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
621 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
623 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
624 GOTO(out, rc = -EINVAL);
627 obd_str2uuid(&tgtuuid, str);
628 target = class_uuid2obd(&tgtuuid);
630 target = class_name2obd(str);
632 if (!target || target->obd_stopping || !target->obd_set_up) {
633 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
634 " for connect (%s)\n", str,
635 !target ? "no target" :
636 (target->obd_stopping ? "stopping" :
638 GOTO(out, rc = -ENODEV);
641 if (target->obd_no_conn) {
642 LCONSOLE_WARN("%s: temporarily refusing client connection "
643 "from %s\n", target->obd_name,
644 libcfs_nid2str(req->rq_peer.nid));
645 GOTO(out, rc = -EAGAIN);
648 /* Make sure the target isn't cleaned up while we're here. Yes,
649 there's still a race between the above check and our incref here.
650 Really, class_uuid2obd should take the ref. */
651 targref = class_incref(target, __FUNCTION__, cfs_current());
654 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
656 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
657 GOTO(out, rc = -EINVAL);
660 obd_str2uuid(&cluuid, str);
662 /* XXX extract a nettype and format accordingly */
663 switch (sizeof(lnet_nid_t)) {
664 /* NB the casts only avoid compiler warnings */
666 snprintf(remote_uuid.uuid, sizeof remote_uuid,
667 "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
670 snprintf(remote_uuid.uuid, sizeof remote_uuid,
671 "NET_%x_UUID", (__u32)req->rq_peer.nid);
677 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
679 GOTO(out, rc = -EPROTO);
683 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
685 GOTO(out, rc = -EPROTO);
687 rc = req_capsule_server_pack(&req->rq_pill);
691 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
693 DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
694 "libclient connection attempt");
695 GOTO(out, rc = -EPROTO);
696 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
697 LUSTRE_VERSION_ALLOWED_OFFSET ||
698 data->ocd_version > LUSTRE_VERSION_CODE +
699 LUSTRE_VERSION_ALLOWED_OFFSET) {
700 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
701 "libclient connection attempt",
702 data->ocd_version < LUSTRE_VERSION_CODE ?
704 OBD_OCD_VERSION_MAJOR(data->ocd_version),
705 OBD_OCD_VERSION_MINOR(data->ocd_version),
706 OBD_OCD_VERSION_PATCH(data->ocd_version),
707 OBD_OCD_VERSION_FIX(data->ocd_version));
708 data = req_capsule_server_sized_get(&req->rq_pill,
710 offsetof(typeof(*data), ocd_version) +
711 sizeof(data->ocd_version));
713 data->ocd_connect_flags = OBD_CONNECT_VERSION;
714 data->ocd_version = LUSTRE_VERSION_CODE;
716 GOTO(out, rc = -EPROTO);
720 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
721 (data->ocd_connect_flags & OBD_CONNECT_MDS))
724 /* lctl gets a backstage, all-access pass. */
725 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
726 goto dont_check_exports;
728 spin_lock(&target->obd_dev_lock);
729 export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
731 if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
732 CWARN("%s: exp %p already connecting\n",
733 export->exp_obd->obd_name, export);
734 class_export_put(export);
737 } else if (export != NULL && export->exp_connection != NULL &&
738 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
740 /* in mds failover we have static uuid but nid can be
742 CWARN("%s: cookie %s seen on new NID %s when "
743 "existing NID %s is already connected\n",
744 target->obd_name, cluuid.uuid,
745 libcfs_nid2str(req->rq_peer.nid),
746 libcfs_nid2str(export->exp_connection->c_peer.nid));
748 class_export_put(export);
750 } else if (export != NULL) {
751 spin_lock(&export->exp_lock);
752 export->exp_connecting = 1;
753 spin_unlock(&export->exp_lock);
754 class_export_put(export);
755 spin_unlock(&target->obd_dev_lock);
756 LASSERT(export->exp_obd == target);
758 rc = target_handle_reconnect(&conn, export, &cluuid, mds_conn);
761 /* If we found an export, we already unlocked. */
763 spin_unlock(&target->obd_dev_lock);
764 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
765 } else if (req->rq_export == NULL &&
766 atomic_read(&export->exp_rpc_count) > 0) {
767 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
768 target->obd_name, cluuid.uuid,
769 libcfs_nid2str(req->rq_peer.nid),
770 export, atomic_read(&export->exp_refcount));
771 GOTO(out, rc = -EBUSY);
772 } else if (req->rq_export != NULL &&
773 (atomic_read(&export->exp_rpc_count) > 1)) {
774 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
775 target->obd_name, cluuid.uuid,
776 libcfs_nid2str(req->rq_peer.nid),
777 export, atomic_read(&export->exp_rpc_count));
778 GOTO(out, rc = -EBUSY);
779 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
781 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
782 "cookies not random?\n", target->obd_name,
783 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
784 GOTO(out, rc = -EALREADY);
786 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
787 if (req->rq_export == NULL && mds_conn)
788 export->exp_last_request_time =
789 max(export->exp_last_request_time,
790 (time_t)cfs_time_current_sec());
797 CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
798 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
799 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
800 export, (long)cfs_time_current_sec(),
801 export ? (long)export->exp_last_request_time : 0);
803 /* Tell the client if we're in recovery. */
804 if (target->obd_recovering) {
805 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
806 /* If this is the first time a client connects,
807 reset the recovery timer */
809 target_start_and_reset_recovery_timer(target, req,
813 /* We want to handle EALREADY but *not* -EALREADY from
814 * target_handle_reconnect(), return reconnection state in a flag */
815 if (rc == EALREADY) {
816 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
822 /* Tell the client if we support replayable requests */
823 if (target->obd_replayable)
824 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
825 client_nid = &req->rq_peer.nid;
827 if (export == NULL) {
828 if (target->obd_recovering) {
831 t = cfs_timer_deadline(&target->obd_recovery_timer);
832 t = cfs_time_sub(t, cfs_time_current());
833 CERROR("%s: denying connection for new client %s (%s): "
834 "%d clients in recovery for "CFS_TIME_T"s\n",
836 libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
837 target->obd_recoverable_clients,
838 cfs_duration_sec(t));
842 rc = obd_connect(req->rq_svc_thread->t_env,
843 &conn, target, &cluuid, data,
847 rc = obd_reconnect(req->rq_svc_thread->t_env,
848 export, target, &cluuid, data, client_nid);
852 /* Return only the parts of obd_connect_data that we understand, so the
853 * client knows that we don't understand the rest. */
855 tmpdata = req_capsule_server_get(&req->rq_pill,
857 //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
861 /* If all else goes well, this is our RPC return code. */
864 lustre_msg_set_handle(req->rq_repmsg, &conn);
866 /* ownership of this export ref transfers to the request AFTER we
867 * drop any previous reference the request had, but we don't want
868 * that to go to zero before we get our new export reference. */
869 export = class_conn2export(&conn);
871 DEBUG_REQ(D_ERROR, req, "Missing export!");
872 GOTO(out, rc = -ENODEV);
875 /* If the client and the server are the same node, we will already
876 * have an export that really points to the client's DLM export,
877 * because we have a shared handles table.
879 * XXX this will go away when shaver stops sending the "connect" handle
880 * in the real "remote handle" field of the request --phik 24 Apr 2003
882 if (req->rq_export != NULL)
883 class_export_put(req->rq_export);
885 req->rq_export = export;
887 spin_lock(&export->exp_lock);
889 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
890 } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
891 spin_unlock(&export->exp_lock);
892 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
893 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
894 export->exp_conn_cnt,
895 lustre_msg_get_conn_cnt(req->rq_reqmsg));
897 GOTO(out, rc = -EALREADY);
899 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
901 /* request from liblustre? Don't evict it for not pinging. */
902 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
903 export->exp_libclient = 1;
904 spin_unlock(&export->exp_lock);
906 spin_lock(&target->obd_dev_lock);
907 list_del_init(&export->exp_obd_chain_timed);
908 spin_unlock(&target->obd_dev_lock);
910 spin_unlock(&export->exp_lock);
913 if (export->exp_connection != NULL)
914 ptlrpc_connection_put(export->exp_connection);
915 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
918 if (hlist_unhashed(&export->exp_nid_hash)) {
919 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
920 &export->exp_connection->c_peer.nid,
921 &export->exp_nid_hash);
924 spin_lock_bh(&target->obd_processing_task_lock);
925 if (target->obd_recovering && !export->exp_in_recovery) {
926 spin_lock(&export->exp_lock);
927 export->exp_in_recovery = 1;
928 export->exp_req_replay_needed = 1;
929 export->exp_lock_replay_needed = 1;
930 spin_unlock(&export->exp_lock);
931 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
932 && (data->ocd_transno == 0))
933 CWARN("Connect with zero transno!\n");
935 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
936 && data->ocd_transno < target->obd_next_recovery_transno)
937 target->obd_next_recovery_transno = data->ocd_transno;
938 target->obd_connected_clients++;
939 /* each connected client is counted as recoverable */
940 target->obd_recoverable_clients++;
941 atomic_inc(&target->obd_req_replay_clients);
942 atomic_inc(&target->obd_lock_replay_clients);
943 if (target->obd_connected_clients ==
944 target->obd_max_recoverable_clients)
945 wake_up(&target->obd_next_transno_waitq);
947 spin_unlock_bh(&target->obd_processing_task_lock);
948 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
951 if (export->exp_imp_reverse != NULL) {
952 /* destroyed import can be still referenced in ctxt */
953 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
954 KEY_REVIMP_UPD, 0, NULL, NULL);
956 /* in some recovery senarios, previous ctx init rpc handled
957 * in sptlrpc_target_export_check() might be used to install
958 * a reverse ctx in this reverse import, and later OBD_CONNECT
959 * using the same gss ctx could reach here and following new
960 * reverse import. note all reverse ctx in new/old import are
961 * actually based on the same gss ctx. so we invalidate ctx
962 * here before destroy import, otherwise flush old import will
963 * lead to remote reverse ctx be destroied, thus the reverse
964 * ctx of new import will lost its peer.
965 * there might be a better way to deal with this???
967 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
969 destroy_import(export->exp_imp_reverse);
972 /* for the rest part, we return -ENOTCONN in case of errors
973 * in order to let client initialize connection again.
975 revimp = export->exp_imp_reverse = class_new_import(target);
977 CERROR("fail to alloc new reverse import.\n");
978 GOTO(out, rc = -ENOTCONN);
981 revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
982 revimp->imp_client = &export->exp_obd->obd_ldlm_client;
983 revimp->imp_remote_handle = conn;
984 revimp->imp_dlm_fake = 1;
985 revimp->imp_state = LUSTRE_IMP_FULL;
987 /* unknown versions will be caught in
988 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
989 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
991 if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
992 (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
993 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
995 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
997 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx,
998 req->rq_flvr.sf_rpc);
1000 CERROR("Failed to get sec for reverse import: %d\n", rc);
1001 export->exp_imp_reverse = NULL;
1002 class_destroy_import(revimp);
1005 class_import_put(revimp);
1008 spin_lock(&export->exp_lock);
1009 export->exp_connecting = 0;
1010 spin_unlock(&export->exp_lock);
1013 class_decref(targref, __FUNCTION__, cfs_current());
1015 req->rq_status = rc;
1019 int target_handle_disconnect(struct ptlrpc_request *req)
1024 rc = req_capsule_server_pack(&req->rq_pill);
1028 /* keep the rq_export around so we can send the reply */
1029 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1034 void target_destroy_export(struct obd_export *exp)
1036 /* exports created from last_rcvd data, and "fake"
1037 exports created by lctl don't have an import */
1038 if (exp->exp_imp_reverse != NULL)
1039 destroy_import(exp->exp_imp_reverse);
1041 /* We cancel locks at disconnect time, but this will catch any locks
1042 * granted in a race with recovery-induced disconnect. */
1043 if (exp->exp_obd->obd_namespace != NULL)
1044 ldlm_cancel_locks_for_export(exp);
1048 * Recovery functions
1051 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1053 struct ptlrpc_request *copy_req;
1054 struct lustre_msg *copy_reqmsg;
1055 struct ptlrpc_user_desc *udesc = NULL;
1057 OBD_ALLOC_PTR(copy_req);
1060 OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1062 OBD_FREE_PTR(copy_req);
1066 if (orig_req->rq_user_desc) {
1067 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1069 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1071 OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1072 OBD_FREE_PTR(copy_req);
1075 memcpy(udesc, orig_req->rq_user_desc,
1076 sptlrpc_user_desc_size(ngroups));
1079 *copy_req = *orig_req;
1080 memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1081 copy_req->rq_reqmsg = copy_reqmsg;
1082 copy_req->rq_user_desc = udesc;
1084 class_export_rpc_get(copy_req->rq_export);
1085 CFS_INIT_LIST_HEAD(©_req->rq_list);
1086 CFS_INIT_LIST_HEAD(©_req->rq_replay_list);
1087 sptlrpc_svc_ctx_addref(copy_req);
1089 if (copy_req->rq_reply_state) {
1090 /* the copied req takes over the reply state */
1091 orig_req->rq_reply_state = NULL;
1092 /* to catch further access */
1093 orig_req->rq_repmsg = NULL;
1094 orig_req->rq_replen = 0;
1100 void ptlrpc_free_clone(struct ptlrpc_request *req)
1102 LASSERT(list_empty(&req->rq_replay_list));
1104 ptlrpc_req_drop_rs(req);
1105 sptlrpc_svc_ctx_decref(req);
1106 class_export_rpc_put(req->rq_export);
1107 list_del_init(&req->rq_list);
1109 if (req->rq_user_desc) {
1110 int ngroups = req->rq_user_desc->pud_ngroups;
1111 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1113 OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1117 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1119 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1120 struct obd_export *exp = req->rq_export;
1121 struct ptlrpc_request *reqiter;
1126 spin_lock(&exp->exp_lock);
1127 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1129 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1136 /* we expect it with RESENT and REPLAY flags */
1137 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1138 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1139 CERROR("invalid flags %x of resent replay\n",
1140 lustre_msg_get_flags(req->rq_reqmsg));
1142 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1145 spin_unlock(&exp->exp_lock);
1149 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1151 LASSERT(!list_empty(&req->rq_replay_list));
1152 LASSERT(req->rq_export);
1154 spin_lock(&req->rq_export->exp_lock);
1155 list_del_init(&req->rq_replay_list);
1156 spin_unlock(&req->rq_export->exp_lock);
1160 static void target_finish_recovery(struct obd_device *obd)
1163 LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1166 ldlm_reprocess_all_ns(obd->obd_namespace);
1167 spin_lock_bh(&obd->obd_processing_task_lock);
1168 if (list_empty(&obd->obd_req_replay_queue) &&
1169 list_empty(&obd->obd_lock_replay_queue) &&
1170 list_empty(&obd->obd_final_req_queue)) {
1171 obd->obd_processing_task = 0;
1173 CERROR("%s: Recovery queues ( %s%s%s) are empty\n",
1175 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1176 list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1177 list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1178 spin_unlock_bh(&obd->obd_processing_task_lock);
1181 spin_unlock_bh(&obd->obd_processing_task_lock);
1183 /* when recovery finished, cleanup orphans on mds and ost */
1184 if (OBT(obd) && OBP(obd, postrecov)) {
1185 int rc = OBP(obd, postrecov)(obd);
1186 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1187 rc < 0 ? "failed" : "complete", rc);
1190 obd->obd_recovery_end = cfs_time_current_sec();
1194 static void abort_req_replay_queue(struct obd_device *obd)
1196 struct ptlrpc_request *req, *n;
1197 struct list_head abort_list;
1199 CFS_INIT_LIST_HEAD(&abort_list);
1200 spin_lock_bh(&obd->obd_processing_task_lock);
1201 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1202 spin_unlock_bh(&obd->obd_processing_task_lock);
1203 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1204 DEBUG_REQ(D_WARNING, req, "aborted:");
1205 req->rq_status = -ENOTCONN;
1206 if (ptlrpc_error(req)) {
1207 DEBUG_REQ(D_ERROR, req,
1208 "failed abort_req_reply; skipping");
1210 target_exp_dequeue_req_replay(req);
1211 ptlrpc_free_clone(req);
1215 static void abort_lock_replay_queue(struct obd_device *obd)
1217 struct ptlrpc_request *req, *n;
1218 struct list_head abort_list;
1220 CFS_INIT_LIST_HEAD(&abort_list);
1221 spin_lock_bh(&obd->obd_processing_task_lock);
1222 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1223 spin_unlock_bh(&obd->obd_processing_task_lock);
1224 list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1225 DEBUG_REQ(D_ERROR, req, "aborted:");
1226 req->rq_status = -ENOTCONN;
1227 if (ptlrpc_error(req)) {
1228 DEBUG_REQ(D_ERROR, req,
1229 "failed abort_lock_reply; skipping");
1231 ptlrpc_free_clone(req);
1236 /* Called from a cleanup function if the device is being cleaned up
1237 forcefully. The exports should all have been disconnected already,
1238 the only thing left to do is
1239 - clear the recovery flags
1241 - free queued requests and replies, but don't send replies
1242 Because the obd_stopping flag is set, no new requests should be received.
1245 void target_cleanup_recovery(struct obd_device *obd)
1247 struct ptlrpc_request *req, *n;
1248 struct list_head clean_list;
1251 LASSERT(obd->obd_stopping);
1253 CFS_INIT_LIST_HEAD(&clean_list);
1254 spin_lock_bh(&obd->obd_processing_task_lock);
1255 if (!obd->obd_recovering) {
1256 spin_unlock_bh(&obd->obd_processing_task_lock);
1260 obd->obd_recovering = obd->obd_abort_recovery = 0;
1261 target_cancel_recovery_timer(obd);
1263 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1264 spin_unlock_bh(&obd->obd_processing_task_lock);
1266 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1267 LASSERT(req->rq_reply_state == 0);
1268 target_exp_dequeue_req_replay(req);
1269 ptlrpc_free_clone(req);
1272 spin_lock_bh(&obd->obd_processing_task_lock);
1273 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1274 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1275 spin_unlock_bh(&obd->obd_processing_task_lock);
1277 list_for_each_entry_safe(req, n, &clean_list, rq_list){
1278 LASSERT(req->rq_reply_state == 0);
1279 ptlrpc_free_clone(req);
1285 /* obd_processing_task_lock should be held */
1286 void target_cancel_recovery_timer(struct obd_device *obd)
1288 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1289 cfs_timer_disarm(&obd->obd_recovery_timer);
1292 /* extend = 1 means require at least "duration" seconds left in the timer,
1293 extend = 0 means set the total duration (start_recovery_timer) */
1294 static void reset_recovery_timer(struct obd_device *obd, int duration,
1297 cfs_time_t now = cfs_time_current_sec();
1298 cfs_duration_t left;
1300 spin_lock_bh(&obd->obd_processing_task_lock);
1301 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1302 spin_unlock_bh(&obd->obd_processing_task_lock);
1306 left = cfs_time_sub(obd->obd_recovery_end, now);
1308 if (extend && (duration > left))
1309 obd->obd_recovery_timeout += duration - left;
1310 else if (!extend && (duration > obd->obd_recovery_timeout))
1311 /* Track the client's largest expected replay time */
1312 obd->obd_recovery_timeout = duration;
1315 * If total recovery time already exceed the
1316 * obd_recovery_max_time, then CRAY XT3 will
1317 * abort the recovery
1319 if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1320 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1322 obd->obd_recovery_end = obd->obd_recovery_start +
1323 obd->obd_recovery_timeout;
1324 if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1325 cfs_time_before(now, obd->obd_recovery_end)) {
1326 left = cfs_time_sub(obd->obd_recovery_end, now);
1327 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1329 spin_unlock_bh(&obd->obd_processing_task_lock);
1330 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1331 obd->obd_name, (unsigned)left);
1334 static void check_and_start_recovery_timer(struct obd_device *obd)
1336 spin_lock_bh(&obd->obd_processing_task_lock);
1337 if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1338 spin_unlock_bh(&obd->obd_processing_task_lock);
1341 CWARN("%s: starting recovery timer\n", obd->obd_name);
1342 obd->obd_recovery_start = cfs_time_current_sec();
1344 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1345 spin_unlock_bh(&obd->obd_processing_task_lock);
1347 reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1350 /* Reset the timer with each new client connection */
1352 * This timer is actually reconnect_timer, which is for making sure
1353 * the total recovery window is at least as big as my reconnect
1354 * attempt timing. So the initial recovery time_out will be set to
1355 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1356 * from client is bigger than this, then the recovery time_out will
1357 * be extend to make sure the client could be reconnected, in the
1358 * process, the timeout from the new client should be ignored.
1362 target_start_and_reset_recovery_timer(struct obd_device *obd,
1363 struct ptlrpc_request *req,
1366 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1368 if (!new_client && service_time)
1369 /* Teach server about old server's estimates, as first guess
1370 * at how long new requests will take. */
1371 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1374 check_and_start_recovery_timer(obd);
1376 /* convert the service time to rpc timeout,
1377 * reuse service_time to limit stack usage */
1378 service_time = at_est2timeout(service_time);
1380 /* We expect other clients to timeout within service_time, then try
1381 * to reconnect, then try the failover server. The max delay between
1382 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1383 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1384 INITIAL_CONNECT_TIMEOUT);
1385 if (service_time > obd->obd_recovery_timeout && !new_client)
1386 reset_recovery_timer(obd, service_time, 0);
1390 static int check_for_next_transno(struct obd_device *obd)
1392 struct ptlrpc_request *req = NULL;
1393 int wake_up = 0, connected, completed, queue_len, max;
1394 __u64 next_transno, req_transno;
1396 spin_lock_bh(&obd->obd_processing_task_lock);
1398 if (!list_empty(&obd->obd_req_replay_queue)) {
1399 req = list_entry(obd->obd_req_replay_queue.next,
1400 struct ptlrpc_request, rq_list);
1401 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1406 max = obd->obd_max_recoverable_clients;
1407 connected = obd->obd_connected_clients;
1408 completed = connected - obd->obd_recoverable_clients;
1409 queue_len = obd->obd_requests_queued_for_recovery;
1410 next_transno = obd->obd_next_recovery_transno;
1412 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1413 "req_transno: "LPU64", next_transno: "LPU64"\n",
1414 max, connected, completed, queue_len, req_transno, next_transno);
1416 if (obd->obd_abort_recovery) {
1417 CDEBUG(D_HA, "waking for aborted recovery\n");
1419 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1420 CDEBUG(D_HA, "waking for completed recovery\n");
1422 } else if (req_transno == next_transno) {
1423 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1425 } else if (queue_len + completed == max) {
1426 /* handle gaps occured due to lost reply. It is allowed gaps
1427 * because all clients are connected and there will be resend
1428 * for missed transaction */
1429 LASSERTF(req_transno >= next_transno,
1430 "req_transno: "LPU64", next_transno: "LPU64"\n",
1431 req_transno, next_transno);
1433 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1434 "waking for skipped transno (skip: "LPD64
1435 ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1436 next_transno, queue_len, completed, connected, req_transno);
1437 obd->obd_next_recovery_transno = req_transno;
1439 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1440 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1441 " by fail_lock, waking up ("LPD64")\n", next_transno);
1442 obd->obd_next_recovery_transno = req_transno;
1444 } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1445 /* some clients haven't connected in time, but we can try
1446 * to replay requests that demand on already committed ones
1447 * also, we can replay first non-committed transation */
1448 LASSERT(req_transno != 0);
1449 if (req_transno == obd->obd_last_committed + 1) {
1450 obd->obd_next_recovery_transno = req_transno;
1451 } else if (req_transno > obd->obd_last_committed) {
1452 /* can't continue recovery: have no needed transno */
1453 obd->obd_abort_recovery = 1;
1454 CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1455 "connected: %d, completed: %d, queue_len: %d, "
1456 "req_transno: "LPU64", next_transno: "LPU64"\n",
1457 max, connected, completed, queue_len,
1458 req_transno, next_transno);
1463 spin_unlock_bh(&obd->obd_processing_task_lock);
1467 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1469 struct l_wait_info lwi = { 0 };
1470 struct ptlrpc_request *req;
1472 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1473 obd->obd_next_recovery_transno);
1474 l_wait_event(obd->obd_next_transno_waitq,
1475 check_for_next_transno(obd), &lwi);
1477 spin_lock_bh(&obd->obd_processing_task_lock);
1478 if (obd->obd_abort_recovery) {
1480 } else if (!list_empty(&obd->obd_req_replay_queue)) {
1481 req = list_entry(obd->obd_req_replay_queue.next,
1482 struct ptlrpc_request, rq_list);
1483 target_exp_dequeue_req_replay(req);
1484 list_del_init(&req->rq_list);
1485 obd->obd_requests_queued_for_recovery--;
1489 spin_unlock_bh(&obd->obd_processing_task_lock);
1493 static int check_for_next_lock(struct obd_device *obd)
1495 struct ptlrpc_request *req = NULL;
1498 spin_lock_bh(&obd->obd_processing_task_lock);
1499 if (!list_empty(&obd->obd_lock_replay_queue)) {
1500 req = list_entry(obd->obd_lock_replay_queue.next,
1501 struct ptlrpc_request, rq_list);
1502 CDEBUG(D_HA, "waking for next lock\n");
1504 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1505 CDEBUG(D_HA, "waking for completed lock replay\n");
1507 } else if (obd->obd_abort_recovery) {
1508 CDEBUG(D_HA, "waking for aborted recovery\n");
1511 spin_unlock_bh(&obd->obd_processing_task_lock);
1516 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1518 struct l_wait_info lwi = { 0 };
1519 struct ptlrpc_request *req;
1521 CDEBUG(D_HA, "Waiting for lock\n");
1522 l_wait_event(obd->obd_next_transno_waitq,
1523 check_for_next_lock(obd), &lwi);
1525 spin_lock_bh(&obd->obd_processing_task_lock);
1526 if (obd->obd_abort_recovery) {
1528 } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1529 req = list_entry(obd->obd_lock_replay_queue.next,
1530 struct ptlrpc_request, rq_list);
1531 list_del_init(&req->rq_list);
1535 spin_unlock_bh(&obd->obd_processing_task_lock);
1539 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1541 struct ptlrpc_request *req;
1543 spin_lock_bh(&obd->obd_processing_task_lock);
1544 if (!list_empty(&obd->obd_final_req_queue)) {
1545 req = list_entry(obd->obd_final_req_queue.next,
1546 struct ptlrpc_request, rq_list);
1547 list_del_init(&req->rq_list);
1551 spin_unlock_bh(&obd->obd_processing_task_lock);
1555 static inline int req_replay_done(struct obd_export *exp)
1557 return (exp->exp_req_replay_needed == 0);
1560 static inline int lock_replay_done(struct obd_export *exp)
1562 return (exp->exp_lock_replay_needed == 0);
1565 static inline int connect_done(struct obd_export *exp)
1567 return (exp->exp_in_recovery != 0);
1570 static int check_for_clients(struct obd_device *obd)
1572 if (obd->obd_abort_recovery)
1574 LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1575 if (obd->obd_no_conn == 0 &&
1576 obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1581 static int handle_recovery_req(struct ptlrpc_thread *thread,
1582 struct ptlrpc_request *req,
1583 svc_handler_t handler)
1588 rc = lu_context_init(&req->rq_session, LCT_SESSION);
1590 CERROR("Failure to initialize session: %d\n", rc);
1593 req->rq_session.lc_thread = thread;
1594 lu_context_enter(&req->rq_session);
1595 req->rq_svc_thread = thread;
1596 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1598 /* thread context */
1599 lu_context_enter(&thread->t_env->le_ctx);
1601 lu_context_exit(&thread->t_env->le_ctx);
1603 lu_context_exit(&req->rq_session);
1604 lu_context_fini(&req->rq_session);
1605 /* don't reset timer for final stage */
1606 if (!req_replay_done(req->rq_export) ||
1607 !lock_replay_done(req->rq_export))
1608 reset_recovery_timer(class_exp2obd(req->rq_export),
1609 AT_OFF ? obd_timeout :
1610 at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1611 ptlrpc_free_clone(req);
1615 static void resume_recovery_timer(struct obd_device *obd)
1617 /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1618 reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1621 static int target_recovery_thread(void *arg)
1623 struct obd_device *obd = arg;
1624 struct ptlrpc_request *req;
1625 struct target_recovery_data *trd = &obd->obd_recovery_data;
1626 struct l_wait_info lwi = { 0 };
1627 unsigned long delta;
1628 unsigned long flags;
1630 struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1634 cfs_daemonize("tgt_recov");
1636 SIGNAL_MASK_LOCK(current, flags);
1637 sigfillset(¤t->blocked);
1639 SIGNAL_MASK_UNLOCK(current, flags);
1641 rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1645 thread->t_env = &env;
1646 env.le_ctx.lc_thread = thread;
1648 CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1650 trd->trd_processing_task = current->pid;
1652 obd->obd_recovering = 1;
1653 complete(&trd->trd_starting);
1655 /* first of all, we have to know the first transno to replay */
1656 obd->obd_abort_recovery = 0;
1657 l_wait_event(obd->obd_next_transno_waitq,
1658 check_for_clients(obd), &lwi);
1660 spin_lock_bh(&obd->obd_processing_task_lock);
1661 target_cancel_recovery_timer(obd);
1662 spin_unlock_bh(&obd->obd_processing_task_lock);
1664 /* If some clients haven't connected in time, evict them */
1665 if (obd->obd_abort_recovery) {
1666 CWARN("Some clients haven't connect in time (%d/%d),"
1667 "evict them\n", obd->obd_connected_clients,
1668 obd->obd_max_recoverable_clients);
1669 obd->obd_abort_recovery = obd->obd_stopping;
1670 class_disconnect_stale_exports(obd, connect_done);
1672 /* next stage: replay requests */
1674 obd->obd_req_replaying = 1;
1675 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1676 atomic_read(&obd->obd_req_replay_clients),
1677 obd->obd_next_recovery_transno);
1678 resume_recovery_timer(obd);
1679 while ((req = target_next_replay_req(obd))) {
1680 LASSERT(trd->trd_processing_task == current->pid);
1681 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1682 lustre_msg_get_transno(req->rq_reqmsg),
1683 libcfs_nid2str(req->rq_peer.nid));
1684 handle_recovery_req(thread, req,
1685 trd->trd_recovery_handler);
1686 obd->obd_replayed_requests++;
1687 spin_lock_bh(&obd->obd_processing_task_lock);
1688 obd->obd_next_recovery_transno++;
1689 spin_unlock_bh(&obd->obd_processing_task_lock);
1692 spin_lock_bh(&obd->obd_processing_task_lock);
1693 target_cancel_recovery_timer(obd);
1694 spin_unlock_bh(&obd->obd_processing_task_lock);
1696 /* If some clients haven't replayed requests in time, evict them */
1697 if (obd->obd_abort_recovery) {
1698 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1699 obd->obd_abort_recovery = obd->obd_stopping;
1700 class_disconnect_stale_exports(obd, req_replay_done);
1701 abort_req_replay_queue(obd);
1704 /* The second stage: replay locks */
1705 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1706 atomic_read(&obd->obd_lock_replay_clients));
1707 resume_recovery_timer(obd);
1708 while ((req = target_next_replay_lock(obd))) {
1709 LASSERT(trd->trd_processing_task == current->pid);
1710 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1711 libcfs_nid2str(req->rq_peer.nid));
1712 handle_recovery_req(thread, req,
1713 trd->trd_recovery_handler);
1714 obd->obd_replayed_locks++;
1717 spin_lock_bh(&obd->obd_processing_task_lock);
1718 target_cancel_recovery_timer(obd);
1719 spin_unlock_bh(&obd->obd_processing_task_lock);
1720 /* If some clients haven't replayed requests in time, evict them */
1721 if (obd->obd_abort_recovery) {
1723 CERROR("lock replay timed out, aborting ...\n");
1724 obd->obd_abort_recovery = obd->obd_stopping;
1725 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1726 abort_lock_replay_queue(obd);
1729 /* We drop recoverying flag to forward all new requests
1730 * to regular mds_handle() since now */
1731 spin_lock_bh(&obd->obd_processing_task_lock);
1732 obd->obd_recovering = obd->obd_abort_recovery = 0;
1733 spin_unlock_bh(&obd->obd_processing_task_lock);
1734 /* The third stage: reply on final pings */
1735 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1736 while ((req = target_next_final_ping(obd))) {
1737 LASSERT(trd->trd_processing_task == current->pid);
1738 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1739 libcfs_nid2str(req->rq_peer.nid));
1740 handle_recovery_req(thread, req,
1741 trd->trd_recovery_handler);
1744 delta = (jiffies - delta) / HZ;
1745 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1746 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1747 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1748 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1749 if (delta > obd_timeout * 2) {
1750 CWARN("too long recovery - read logs\n");
1751 libcfs_debug_dumplog();
1754 target_finish_recovery(obd);
1756 lu_context_fini(&env.le_ctx);
1757 trd->trd_processing_task = 0;
1758 complete(&trd->trd_finishing);
1762 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1765 struct target_recovery_data *trd = &obd->obd_recovery_data;
1767 memset(trd, 0, sizeof(*trd));
1768 init_completion(&trd->trd_starting);
1769 init_completion(&trd->trd_finishing);
1770 trd->trd_recovery_handler = handler;
1772 if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1773 wait_for_completion(&trd->trd_starting);
1774 LASSERT(obd->obd_recovering != 0);
1781 void target_stop_recovery_thread(struct obd_device *obd)
1783 spin_lock_bh(&obd->obd_processing_task_lock);
1784 if (obd->obd_recovery_data.trd_processing_task > 0) {
1785 struct target_recovery_data *trd = &obd->obd_recovery_data;
1786 CERROR("%s: Aborting recovery\n", obd->obd_name);
1787 obd->obd_abort_recovery = 1;
1788 wake_up(&obd->obd_next_transno_waitq);
1789 spin_unlock_bh(&obd->obd_processing_task_lock);
1790 wait_for_completion(&trd->trd_finishing);
1792 spin_unlock_bh(&obd->obd_processing_task_lock);
1796 void target_recovery_fini(struct obd_device *obd)
1798 class_disconnect_exports(obd);
1799 target_stop_recovery_thread(obd);
1800 target_cleanup_recovery(obd);
1802 EXPORT_SYMBOL(target_recovery_fini);
1804 static void target_recovery_expired(unsigned long castmeharder)
1806 struct obd_device *obd = (struct obd_device *)castmeharder;
1807 LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1808 "after %lds (%d clients did)\n",
1809 obd->obd_name, obd->obd_recoverable_clients,
1810 cfs_time_current_sec()- obd->obd_recovery_start,
1811 obd->obd_connected_clients);
1812 spin_lock_bh(&obd->obd_processing_task_lock);
1813 if (obd->obd_recovering)
1814 obd->obd_abort_recovery = 1;
1815 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1816 spin_unlock_bh(&obd->obd_processing_task_lock);
1819 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1821 if (obd->obd_max_recoverable_clients == 0)
1824 CWARN("RECOVERY: service %s, %d recoverable clients, "
1825 "last_transno "LPU64"\n", obd->obd_name,
1826 obd->obd_max_recoverable_clients, obd->obd_last_committed);
1827 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1828 obd->obd_recovery_start = 0;
1829 obd->obd_recovery_end = 0;
1830 obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1831 /* bz13079: this should be set to desired value for ost but not for mds */
1832 obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1833 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1834 target_start_recovery_thread(obd, handler);
1836 EXPORT_SYMBOL(target_recovery_init);
1840 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1842 struct obd_export *exp = req->rq_export;
1843 LASSERT(exp != NULL);
1844 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1845 /* client declares he's ready to replay locks */
1846 spin_lock_bh(&obd->obd_processing_task_lock);
1847 if (exp->exp_req_replay_needed) {
1848 LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1849 spin_lock(&exp->exp_lock);
1850 exp->exp_req_replay_needed = 0;
1851 spin_unlock(&exp->exp_lock);
1852 atomic_dec(&obd->obd_req_replay_clients);
1853 LASSERT(obd->obd_recoverable_clients > 0);
1854 obd->obd_recoverable_clients--;
1855 if (atomic_read(&obd->obd_req_replay_clients) == 0)
1856 CDEBUG(D_HA, "all clients have replayed reqs\n");
1857 wake_up(&obd->obd_next_transno_waitq);
1859 spin_unlock_bh(&obd->obd_processing_task_lock);
1861 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1862 /* client declares he's ready to complete recovery
1863 * so, we put the request on th final queue */
1864 spin_lock_bh(&obd->obd_processing_task_lock);
1865 if (exp->exp_lock_replay_needed) {
1866 LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1867 spin_lock(&exp->exp_lock);
1868 exp->exp_lock_replay_needed = 0;
1869 spin_unlock(&exp->exp_lock);
1870 atomic_dec(&obd->obd_lock_replay_clients);
1871 if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1872 CDEBUG(D_HA, "all clients have replayed locks\n");
1873 wake_up(&obd->obd_next_transno_waitq);
1875 spin_unlock_bh(&obd->obd_processing_task_lock);
1881 int target_queue_recovery_request(struct ptlrpc_request *req,
1882 struct obd_device *obd)
1884 struct list_head *tmp;
1886 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1890 if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1891 /* Processing the queue right now, don't re-add. */
1895 target_process_req_flags(obd, req);
1897 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1898 /* client declares he's ready to complete recovery
1899 * so, we put the request on th final queue */
1900 req = ptlrpc_clone_req(req);
1903 DEBUG_REQ(D_HA, req, "queue final req");
1904 spin_lock_bh(&obd->obd_processing_task_lock);
1905 if (obd->obd_recovering)
1906 list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1908 spin_unlock_bh(&obd->obd_processing_task_lock);
1909 ptlrpc_free_clone(req);
1910 if (obd->obd_stopping) {
1916 spin_unlock_bh(&obd->obd_processing_task_lock);
1919 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1920 /* client declares he's ready to replay locks */
1921 req = ptlrpc_clone_req(req);
1924 DEBUG_REQ(D_HA, req, "queue lock replay req");
1925 spin_lock_bh(&obd->obd_processing_task_lock);
1926 LASSERT(obd->obd_recovering);
1927 /* usually due to recovery abort */
1928 if (!req->rq_export->exp_in_recovery) {
1929 spin_unlock_bh(&obd->obd_processing_task_lock);
1930 ptlrpc_free_clone(req);
1933 LASSERT(req->rq_export->exp_lock_replay_needed);
1934 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1935 spin_unlock_bh(&obd->obd_processing_task_lock);
1936 wake_up(&obd->obd_next_transno_waitq);
1940 /* CAVEAT EMPTOR: The incoming request message has been swabbed
1941 * (i.e. buflens etc are in my own byte order), but type-dependent
1942 * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1945 CFS_INIT_LIST_HEAD(&req->rq_list);
1946 DEBUG_REQ(D_HA, req, "not queueing");
1950 spin_lock_bh(&obd->obd_processing_task_lock);
1952 /* If we're processing the queue, we want don't want to queue this
1955 * Also, if this request has a transno less than the one we're waiting
1956 * for, we should process it now. It could (and currently always will)
1957 * be an open request for a descriptor that was opened some time ago.
1959 * Also, a resent, replayed request that has already been
1960 * handled will pass through here and be processed immediately.
1962 CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1963 obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1964 if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1965 /* Processing the queue right now, don't re-add. */
1966 LASSERT(list_empty(&req->rq_list));
1967 spin_unlock_bh(&obd->obd_processing_task_lock);
1970 spin_unlock_bh(&obd->obd_processing_task_lock);
1972 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
1975 req = ptlrpc_clone_req(req);
1979 spin_lock_bh(&obd->obd_processing_task_lock);
1980 LASSERT(obd->obd_recovering);
1981 if (!req->rq_export->exp_in_recovery) {
1982 spin_unlock_bh(&obd->obd_processing_task_lock);
1983 ptlrpc_free_clone(req);
1986 LASSERT(req->rq_export->exp_req_replay_needed);
1988 if (target_exp_enqueue_req_replay(req)) {
1989 spin_unlock_bh(&obd->obd_processing_task_lock);
1990 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1991 ptlrpc_free_clone(req);
1996 list_for_each(tmp, &obd->obd_req_replay_queue) {
1997 struct ptlrpc_request *reqiter =
1998 list_entry(tmp, struct ptlrpc_request, rq_list);
2000 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2001 list_add_tail(&req->rq_list, &reqiter->rq_list);
2006 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2008 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2009 "has been claimed by another client");
2010 spin_unlock_bh(&obd->obd_processing_task_lock);
2011 target_exp_dequeue_req_replay(req);
2012 ptlrpc_free_clone(req);
2018 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2020 obd->obd_requests_queued_for_recovery++;
2021 wake_up(&obd->obd_next_transno_waitq);
2022 spin_unlock_bh(&obd->obd_processing_task_lock);
2026 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2028 return req->rq_export->exp_obd;
2031 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2033 LASSERT(exp != NULL);
2034 return &exp->exp_obd->obd_namespace->ns_pool;
2038 * Packs current SLV and Limit into \a req.
2040 int target_pack_pool_reply(struct ptlrpc_request *req)
2042 struct obd_device *obd;
2046 * Check that we still have all structures alive as this may
2047 * be some late rpc in shutdown time.
2049 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2050 !exp_connect_lru_resize(req->rq_export))) {
2051 lustre_msg_set_slv(req->rq_repmsg, 0);
2052 lustre_msg_set_limit(req->rq_repmsg, 0);
2057 * OBD is alive here as export is alive, which we checked above.
2059 obd = req->rq_export->exp_obd;
2061 read_lock(&obd->obd_pool_lock);
2062 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2063 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2064 read_unlock(&obd->obd_pool_lock);
2069 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2071 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2072 DEBUG_REQ(D_ERROR, req, "dropping reply");
2077 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2078 req->rq_status = rc;
2079 return (ptlrpc_send_error(req, 1));
2081 DEBUG_REQ(D_NET, req, "sending reply");
2084 return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2087 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2090 struct ptlrpc_reply_state *rs;
2091 struct obd_device *obd;
2092 struct obd_export *exp;
2093 struct ptlrpc_service *svc;
2096 if (req->rq_no_reply) {
2101 svc = req->rq_rqbd->rqbd_service;
2102 rs = req->rq_reply_state;
2103 if (rs == NULL || !rs->rs_difficult) {
2105 target_send_reply_msg (req, rc, fail_id);
2110 /* must be an export if locks saved */
2111 LASSERT (req->rq_export != NULL);
2112 /* req/reply consistent */
2113 LASSERT (rs->rs_service == svc);
2116 LASSERT (!rs->rs_scheduled);
2117 LASSERT (!rs->rs_scheduled_ever);
2118 LASSERT (!rs->rs_handled);
2119 LASSERT (!rs->rs_on_net);
2120 LASSERT (rs->rs_export == NULL);
2121 LASSERT (list_empty(&rs->rs_obd_list));
2122 LASSERT (list_empty(&rs->rs_exp_list));
2124 exp = class_export_get (req->rq_export);
2127 /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2128 rs->rs_scheduled = 1;
2130 rs->rs_xid = req->rq_xid;
2131 rs->rs_transno = req->rq_transno;
2132 rs->rs_export = exp;
2134 spin_lock(&obd->obd_uncommitted_replies_lock);
2136 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2137 rs->rs_transno, obd->obd_last_committed);
2138 if (rs->rs_transno > obd->obd_last_committed) {
2139 /* not committed already */
2140 list_add_tail (&rs->rs_obd_list,
2141 &obd->obd_uncommitted_replies);
2144 spin_unlock (&obd->obd_uncommitted_replies_lock);
2145 spin_lock (&exp->exp_lock);
2147 list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
2149 spin_unlock(&exp->exp_lock);
2151 netrc = target_send_reply_msg (req, rc, fail_id);
2153 spin_lock(&svc->srv_lock);
2155 atomic_inc(&svc->srv_n_difficult_replies);
2158 /* error sending: reply is off the net. Also we need +1
2159 * reply ref until ptlrpc_server_handle_reply() is done
2160 * with the reply state (if the send was successful, there
2161 * would have been +1 ref for the net, which
2162 * reply_out_callback leaves alone) */
2164 ptlrpc_rs_addref(rs);
2165 atomic_inc (&svc->srv_outstanding_replies);
2168 spin_lock(&rs->rs_lock);
2169 if (rs->rs_transno <= obd->obd_last_committed ||
2170 (!rs->rs_on_net && !rs->rs_no_ack) ||
2171 list_empty(&rs->rs_exp_list) || /* completed already */
2172 list_empty(&rs->rs_obd_list)) {
2173 CDEBUG(D_HA, "Schedule reply immediately\n");
2174 ptlrpc_dispatch_difficult_reply(rs);
2176 list_add (&rs->rs_list, &svc->srv_active_replies);
2177 rs->rs_scheduled = 0; /* allow notifier to schedule */
2179 spin_unlock(&rs->rs_lock);
2180 spin_unlock(&svc->srv_lock);
2184 int target_handle_ping(struct ptlrpc_request *req)
2186 obd_ping(req->rq_export);
2187 return req_capsule_server_pack(&req->rq_pill);
2190 void target_committed_to_req(struct ptlrpc_request *req)
2192 struct obd_device *obd;
2194 if (req == NULL || req->rq_export == NULL)
2197 obd = req->rq_export->exp_obd;
2201 if (!obd->obd_no_transno && req->rq_repmsg != NULL)
2202 lustre_msg_set_last_committed(req->rq_repmsg,
2203 obd->obd_last_committed);
2205 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2206 "%d)", obd->obd_no_transno, req->rq_repmsg == NULL);
2208 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2209 obd->obd_last_committed, req->rq_transno, req->rq_xid);
2212 EXPORT_SYMBOL(target_committed_to_req);
2214 int target_handle_qc_callback(struct ptlrpc_request *req)
2216 struct obd_quotactl *oqctl;
2217 struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2219 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2220 if (oqctl == NULL) {
2221 CERROR("Can't unpack obd_quotactl\n");
2225 cli->cl_qchk_stat = oqctl->qc_stat;
2230 #ifdef HAVE_QUOTA_SUPPORT
2231 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2234 struct obd_device *obd = req->rq_export->exp_obd;
2235 struct obd_device *master_obd;
2236 struct obd_device_target *obt;
2237 struct lustre_quota_ctxt *qctxt;
2238 struct qunit_data *qdata = NULL;
2242 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2245 rc = req_capsule_server_pack(&req->rq_pill);
2247 CERROR("packing reply failed!: rc = %d\n", rc);
2251 LASSERT(req->rq_export);
2253 OBD_ALLOC(qdata, sizeof(struct qunit_data));
2256 rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2258 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2262 /* we use the observer */
2263 if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2264 CERROR("Can't find the observer, it is recovering\n");
2265 req->rq_status = -EAGAIN;
2266 GOTO(send_reply, rc = -EAGAIN);
2269 master_obd = obd->obd_observer->obd_observer;
2270 obt = &master_obd->u.obt;
2271 qctxt = &obt->obt_qctxt;
2273 if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2274 /* quota_type has not been processed yet, return EAGAIN
2275 * until we know whether or not quotas are supposed to
2277 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2279 req->rq_status = -EAGAIN;
2280 rc = ptlrpc_reply(req);
2284 down_read(&obt->obt_rwsem);
2285 if (qctxt->lqc_lqs_hash == NULL) {
2286 up_read(&obt->obt_rwsem);
2287 /* quota_type has not been processed yet, return EAGAIN
2288 * until we know whether or not quotas are supposed to
2290 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2292 req->rq_status = -EAGAIN;
2293 rc = ptlrpc_reply(req);
2297 LASSERT(qctxt->lqc_handler);
2298 rc = qctxt->lqc_handler(master_obd, qdata,
2299 lustre_msg_get_opc(req->rq_reqmsg));
2300 up_read(&obt->obt_rwsem);
2301 if (rc && rc != -EDQUOT)
2302 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
2303 "dqacq failed! (rc:%d)\n", rc);
2304 req->rq_status = rc;
2306 /* there are three forms of qunit(historic causes), so we need to
2307 * adjust the same form to different forms slaves needed */
2308 rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2310 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2314 /* Block the quota req. b=14840 */
2315 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2317 rc = ptlrpc_reply(req);
2319 OBD_FREE(qdata, sizeof(struct qunit_data));
2323 #endif /* !__KERNEL__ */
2325 #endif /* HAVE_QUOTA_SUPPORT */
2327 ldlm_mode_t lck_compat_array[] = {
2328 [LCK_EX] LCK_COMPAT_EX,
2329 [LCK_PW] LCK_COMPAT_PW,
2330 [LCK_PR] LCK_COMPAT_PR,
2331 [LCK_CW] LCK_COMPAT_CW,
2332 [LCK_CR] LCK_COMPAT_CR,
2333 [LCK_NL] LCK_COMPAT_NL,
2334 [LCK_GROUP] LCK_COMPAT_GROUP,
2335 [LCK_COS] LCK_COMPAT_COS,
2339 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2340 * not escape to the user level.
2342 int ldlm_error2errno(ldlm_error_t error)
2350 case ELDLM_LOCK_CHANGED:
2353 case ELDLM_LOCK_ABORTED:
2356 case ELDLM_LOCK_REPLACED:
2359 case ELDLM_NO_LOCK_DATA:
2362 case ELDLM_NAMESPACE_EXISTS:
2365 case ELDLM_BAD_NAMESPACE:
2369 if (((int)error) < 0) /* cast to signed type */
2370 result = error; /* as ldlm_error_t can be unsigned */
2372 CERROR("Invalid DLM result code: %i\n", error);
2378 EXPORT_SYMBOL(ldlm_error2errno);
2381 * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2383 ldlm_error_t ldlm_errno2error(int err_no)
2392 error = ELDLM_LOCK_CHANGED;
2395 error = ELDLM_LOCK_ABORTED;
2398 error = ELDLM_LOCK_REPLACED;
2401 error = ELDLM_NO_LOCK_DATA;
2404 error = ELDLM_NAMESPACE_EXISTS;
2407 error = ELDLM_BAD_NAMESPACE;
2414 EXPORT_SYMBOL(ldlm_errno2error);