4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * This file deals with various client/target related logic including recovery.
40 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
44 #define DEBUG_SUBSYSTEM S_LDLM
46 #include <linux/kthread.h>
47 #include <libcfs/libcfs.h>
49 #include <obd_class.h>
50 #include <lustre_dlm.h>
51 #include <lustre_net.h>
52 #include <lustre_sec.h>
53 #include "ldlm_internal.h"
55 /* @priority: If non-zero, move the selected connection to the list head.
56 * @create: If zero, only search in existing connections.
58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
59 int priority, int create)
61 struct ptlrpc_connection *ptlrpc_conn;
62 struct obd_import_conn *imp_conn = NULL, *item;
66 if (!create && !priority) {
67 CDEBUG(D_HA, "Nothing to do\n");
71 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
73 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
78 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
80 GOTO(out_put, rc = -ENOMEM);
84 spin_lock(&imp->imp_lock);
85 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
86 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
88 list_del(&item->oic_item);
89 list_add(&item->oic_item,
91 item->oic_last_attempt = 0;
93 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
94 imp, imp->imp_obd->obd_name, uuid->uuid,
95 (priority ? ", moved to head" : ""));
96 spin_unlock(&imp->imp_lock);
97 GOTO(out_free, rc = 0);
100 /* No existing import connection found for \a uuid. */
102 imp_conn->oic_conn = ptlrpc_conn;
103 imp_conn->oic_uuid = *uuid;
104 imp_conn->oic_last_attempt = 0;
106 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
108 list_add_tail(&imp_conn->oic_item,
109 &imp->imp_conn_list);
110 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
111 imp, imp->imp_obd->obd_name, uuid->uuid,
112 (priority ? "head" : "tail"));
114 spin_unlock(&imp->imp_lock);
115 GOTO(out_free, rc = -ENOENT);
118 spin_unlock(&imp->imp_lock);
122 OBD_FREE(imp_conn, sizeof(*imp_conn));
124 ptlrpc_connection_put(ptlrpc_conn);
128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
130 return import_set_conn(imp, uuid, 1, 0);
133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
136 return import_set_conn(imp, uuid, priority, 1);
138 EXPORT_SYMBOL(client_import_add_conn);
140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
142 struct obd_import_conn *imp_conn;
143 struct obd_export *dlmexp;
147 spin_lock(&imp->imp_lock);
148 if (list_empty(&imp->imp_conn_list)) {
149 LASSERT(!imp->imp_connection);
153 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
154 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
156 LASSERT(imp_conn->oic_conn);
158 if (imp_conn == imp->imp_conn_current) {
159 LASSERT(imp_conn->oic_conn == imp->imp_connection);
161 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
162 imp->imp_state != LUSTRE_IMP_DISCON) {
163 CERROR("can't remove current connection\n");
164 GOTO(out, rc = -EBUSY);
167 ptlrpc_connection_put(imp->imp_connection);
168 imp->imp_connection = NULL;
170 dlmexp = class_conn2export(&imp->imp_dlm_handle);
171 if (dlmexp && dlmexp->exp_connection) {
172 LASSERT(dlmexp->exp_connection ==
174 ptlrpc_connection_put(dlmexp->exp_connection);
175 dlmexp->exp_connection = NULL;
179 class_export_put(dlmexp);
182 list_del(&imp_conn->oic_item);
183 ptlrpc_connection_put(imp_conn->oic_conn);
184 OBD_FREE(imp_conn, sizeof(*imp_conn));
185 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
186 imp, imp->imp_obd->obd_name, uuid->uuid);
191 spin_unlock(&imp->imp_lock);
193 CERROR("connection %s not found\n", uuid->uuid);
196 EXPORT_SYMBOL(client_import_del_conn);
199 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
200 * to find a conn uuid of \a imp which can reach \a peer.
202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
203 struct obd_uuid *uuid)
205 struct obd_import_conn *conn;
209 spin_lock(&imp->imp_lock);
210 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
211 /* Check if conn UUID does have this peer NID. */
212 if (class_check_uuid(&conn->oic_uuid, peer)) {
213 *uuid = conn->oic_uuid;
218 spin_unlock(&imp->imp_lock);
221 EXPORT_SYMBOL(client_import_find_conn);
223 void client_destroy_import(struct obd_import *imp)
225 /* Drop security policy instance after all RPCs have finished/aborted
226 * to let all busy contexts be released. */
227 class_import_get(imp);
228 class_destroy_import(imp);
229 sptlrpc_import_sec_put(imp);
230 class_import_put(imp);
232 EXPORT_SYMBOL(client_destroy_import);
235 * Check whether or not the OSC is on MDT.
238 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
240 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
243 static int osc_on_mdt(char *obdname)
247 ptr = strrchr(obdname, '-');
251 if (strncmp(ptr + 1, "MDT", 3) == 0)
257 /* Configure an RPC client OBD device.
262 * 3 - inactive-on-startup
264 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
266 struct client_obd *cli = &obddev->u.cli;
267 struct obd_import *imp;
268 struct obd_uuid server_uuid;
269 int rq_portal, rp_portal, connect_op;
270 char *name = obddev->obd_type->typ_name;
271 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
272 char *cli_name = lustre_cfg_buf(lcfg, 0);
276 /* In a more perfect world, we would hang a ptlrpc_client off of
277 * obd_type and just use the values from there. */
278 if (!strcmp(name, LUSTRE_OSC_NAME)) {
279 rq_portal = OST_REQUEST_PORTAL;
280 rp_portal = OSC_REPLY_PORTAL;
281 connect_op = OST_CONNECT;
282 cli->cl_sp_me = LUSTRE_SP_CLI;
283 cli->cl_sp_to = LUSTRE_SP_OST;
284 ns_type = LDLM_NS_TYPE_OSC;
285 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
286 !strcmp(name, LUSTRE_LWP_NAME)) {
287 rq_portal = MDS_REQUEST_PORTAL;
288 rp_portal = MDC_REPLY_PORTAL;
289 connect_op = MDS_CONNECT;
290 if (is_lwp_on_ost(cli_name))
291 cli->cl_sp_me = LUSTRE_SP_OST;
292 else if (is_lwp_on_mdt(cli_name))
293 cli->cl_sp_me = LUSTRE_SP_MDT;
295 cli->cl_sp_me = LUSTRE_SP_CLI;
296 cli->cl_sp_to = LUSTRE_SP_MDT;
297 ns_type = LDLM_NS_TYPE_MDC;
298 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
299 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
300 /* OSP_on_MDT for other MDTs */
301 connect_op = MDS_CONNECT;
302 cli->cl_sp_to = LUSTRE_SP_MDT;
303 ns_type = LDLM_NS_TYPE_MDC;
304 rq_portal = OUT_PORTAL;
306 /* OSP on MDT for OST */
307 connect_op = OST_CONNECT;
308 cli->cl_sp_to = LUSTRE_SP_OST;
309 ns_type = LDLM_NS_TYPE_OSC;
310 rq_portal = OST_REQUEST_PORTAL;
312 rp_portal = OSC_REPLY_PORTAL;
313 cli->cl_sp_me = LUSTRE_SP_MDT;
314 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
315 rq_portal = MGS_REQUEST_PORTAL;
316 rp_portal = MGC_REPLY_PORTAL;
317 connect_op = MGS_CONNECT;
318 cli->cl_sp_me = LUSTRE_SP_MGC;
319 cli->cl_sp_to = LUSTRE_SP_MGS;
320 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
321 ns_type = LDLM_NS_TYPE_MGC;
323 CERROR("unknown client OBD type \"%s\", can't setup\n",
328 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
329 CERROR("requires a TARGET UUID\n");
333 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
334 CERROR("client UUID must be less than 38 characters\n");
338 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
339 CERROR("setup requires a SERVER UUID\n");
343 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
344 CERROR("target UUID must be less than 38 characters\n");
348 init_rwsem(&cli->cl_sem);
349 mutex_init(&cli->cl_mgc_mutex);
350 cli->cl_conn_count = 0;
351 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
352 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
353 sizeof(server_uuid)));
355 cli->cl_dirty_pages = 0;
356 cli->cl_avail_grant = 0;
357 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
358 /* cl_dirty_max_pages may be changed at connect time in
359 * ptlrpc_connect_interpret(). */
360 client_adjust_max_dirty(cli);
361 INIT_LIST_HEAD(&cli->cl_cache_waiters);
362 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
363 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
364 INIT_LIST_HEAD(&cli->cl_loi_write_list);
365 INIT_LIST_HEAD(&cli->cl_loi_read_list);
366 spin_lock_init(&cli->cl_loi_list_lock);
367 atomic_set(&cli->cl_pending_w_pages, 0);
368 atomic_set(&cli->cl_pending_r_pages, 0);
369 cli->cl_r_in_flight = 0;
370 cli->cl_w_in_flight = 0;
372 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
373 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
374 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
375 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
376 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
377 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
380 INIT_LIST_HEAD(&cli->cl_lru_osc);
381 atomic_set(&cli->cl_lru_shrinkers, 0);
382 atomic_long_set(&cli->cl_lru_busy, 0);
383 atomic_long_set(&cli->cl_lru_in_list, 0);
384 INIT_LIST_HEAD(&cli->cl_lru_list);
385 spin_lock_init(&cli->cl_lru_list_lock);
386 atomic_long_set(&cli->cl_unstable_count, 0);
387 INIT_LIST_HEAD(&cli->cl_shrink_list);
389 init_waitqueue_head(&cli->cl_destroy_waitq);
390 atomic_set(&cli->cl_destroy_in_flight, 0);
391 #ifdef ENABLE_CHECKSUM
392 /* Turn on checksumming by default. */
393 cli->cl_checksum = 1;
395 * The supported checksum types will be worked out at connect time
396 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
399 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
401 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
403 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
404 * from OFD after connecting. */
405 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
407 /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
408 * it will be updated at OSC connection time. */
409 cli->cl_chunkbits = PAGE_CACHE_SHIFT;
411 if (!strcmp(name, LUSTRE_MDC_NAME)) {
412 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
413 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
414 cli->cl_max_rpcs_in_flight = 2;
415 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
416 cli->cl_max_rpcs_in_flight = 3;
417 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
418 cli->cl_max_rpcs_in_flight = 4;
420 if (osc_on_mdt(obddev->obd_name))
421 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
423 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
426 spin_lock_init(&cli->cl_mod_rpcs_lock);
427 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
428 cli->cl_max_mod_rpcs_in_flight = 0;
429 cli->cl_mod_rpcs_in_flight = 0;
430 cli->cl_close_rpcs_in_flight = 0;
431 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
432 cli->cl_mod_tag_bitmap = NULL;
434 if (connect_op == MDS_CONNECT) {
435 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
436 OBD_ALLOC(cli->cl_mod_tag_bitmap,
437 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
438 if (cli->cl_mod_tag_bitmap == NULL)
439 GOTO(err, rc = -ENOMEM);
444 CERROR("ldlm_get_ref failed: %d\n", rc);
448 ptlrpc_init_client(rq_portal, rp_portal, name,
449 &obddev->obd_ldlm_client);
451 imp = class_new_import(obddev);
453 GOTO(err_ldlm, rc = -ENOENT);
454 imp->imp_client = &obddev->obd_ldlm_client;
455 imp->imp_connect_op = connect_op;
456 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
457 LUSTRE_CFG_BUFLEN(lcfg, 1));
458 class_import_put(imp);
460 rc = client_import_add_conn(imp, &server_uuid, 1);
462 CERROR("can't add initial connection\n");
463 GOTO(err_import, rc);
466 cli->cl_import = imp;
467 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
468 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
470 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
471 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
472 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
473 name, obddev->obd_name,
474 cli->cl_target_uuid.uuid);
475 spin_lock(&imp->imp_lock);
476 imp->imp_deactive = 1;
477 spin_unlock(&imp->imp_lock);
481 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
482 LDLM_NAMESPACE_CLIENT,
483 LDLM_NAMESPACE_GREEDY,
485 if (obddev->obd_namespace == NULL) {
486 CERROR("Unable to create client namespace - %s\n",
488 GOTO(err_import, rc = -ENOMEM);
494 class_destroy_import(imp);
498 if (cli->cl_mod_tag_bitmap != NULL)
499 OBD_FREE(cli->cl_mod_tag_bitmap,
500 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
501 cli->cl_mod_tag_bitmap = NULL;
505 EXPORT_SYMBOL(client_obd_setup);
507 int client_obd_cleanup(struct obd_device *obddev)
509 struct client_obd *cli = &obddev->u.cli;
512 ldlm_namespace_free_post(obddev->obd_namespace);
513 obddev->obd_namespace = NULL;
515 obd_cleanup_client_import(obddev);
516 LASSERT(obddev->u.cli.cl_import == NULL);
520 if (cli->cl_mod_tag_bitmap != NULL)
521 OBD_FREE(cli->cl_mod_tag_bitmap,
522 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
523 cli->cl_mod_tag_bitmap = NULL;
527 EXPORT_SYMBOL(client_obd_cleanup);
529 /* ->o_connect() method for client side (OSC and MDC and MGC) */
530 int client_connect_import(const struct lu_env *env,
531 struct obd_export **exp,
532 struct obd_device *obd, struct obd_uuid *cluuid,
533 struct obd_connect_data *data, void *localdata)
535 struct client_obd *cli = &obd->u.cli;
536 struct obd_import *imp = cli->cl_import;
537 struct obd_connect_data *ocd;
538 struct lustre_handle conn = { 0 };
544 down_write(&cli->cl_sem);
545 if (cli->cl_conn_count > 0)
546 GOTO(out_sem, rc = -EALREADY);
548 rc = class_connect(&conn, obd, cluuid);
552 cli->cl_conn_count++;
553 *exp = class_conn2export(&conn);
555 LASSERT(obd->obd_namespace);
557 imp->imp_dlm_handle = conn;
558 rc = ptlrpc_init_import(imp);
562 ocd = &imp->imp_connect_data;
565 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
566 LUSTRE_MDC_NAME, 3) == 0;
568 data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
569 imp->imp_connect_flags_orig = data->ocd_connect_flags;
570 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
573 rc = ptlrpc_connect_import(imp);
576 data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
577 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
580 LASSERT(*exp != NULL && (*exp)->exp_connection);
583 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
584 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
585 data->ocd_connect_flags, ocd->ocd_connect_flags);
586 data->ocd_connect_flags = ocd->ocd_connect_flags;
587 /* clear the flag as it was not set and is not known
590 data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
593 ptlrpc_pinger_add_import(imp);
599 cli->cl_conn_count--;
600 class_disconnect(*exp);
604 up_write(&cli->cl_sem);
608 EXPORT_SYMBOL(client_connect_import);
610 int client_disconnect_export(struct obd_export *exp)
612 struct obd_device *obd = class_exp2obd(exp);
613 struct client_obd *cli;
614 struct obd_import *imp;
619 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
620 exp, exp ? exp->exp_handle.h_cookie : -1);
625 imp = cli->cl_import;
627 down_write(&cli->cl_sem);
628 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
631 if (cli->cl_conn_count == 0) {
632 CERROR("disconnecting disconnected device (%s)\n",
634 GOTO(out_disconnect, rc = -EINVAL);
637 cli->cl_conn_count--;
638 if (cli->cl_conn_count != 0)
639 GOTO(out_disconnect, rc = 0);
641 /* Mark import deactivated now, so we don't try to reconnect if any
642 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
643 * fully deactivate the import, or that would drop all requests. */
644 spin_lock(&imp->imp_lock);
645 imp->imp_deactive = 1;
646 spin_unlock(&imp->imp_lock);
648 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
649 * delete it regardless. (It's safe to delete an import that was
651 (void)ptlrpc_pinger_del_import(imp);
653 if (obd->obd_namespace != NULL) {
654 /* obd_force == local only */
655 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
656 obd->obd_force ? LCF_LOCAL : 0, NULL);
657 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
660 /* There's no need to hold sem while disconnecting an import,
661 * and it may actually cause deadlock in GSS. */
662 up_write(&cli->cl_sem);
663 rc = ptlrpc_disconnect_import(imp, 0);
664 down_write(&cli->cl_sem);
666 ptlrpc_invalidate_import(imp);
671 /* Use server style - class_disconnect should be always called for
673 err = class_disconnect(exp);
677 up_write(&cli->cl_sem);
681 EXPORT_SYMBOL(client_disconnect_export);
683 #ifdef HAVE_SERVER_SUPPORT
684 int server_disconnect_export(struct obd_export *exp)
689 /* Disconnect early so that clients can't keep using export. */
690 rc = class_disconnect(exp);
691 /* Close import to avoid sending any requests. */
692 if (exp->exp_imp_reverse)
693 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
695 ldlm_bl_thread_wakeup();
697 /* complete all outstanding replies */
698 spin_lock(&exp->exp_lock);
699 while (!list_empty(&exp->exp_outstanding_replies)) {
700 struct ptlrpc_reply_state *rs =
701 list_entry(exp->exp_outstanding_replies.next,
702 struct ptlrpc_reply_state, rs_exp_list);
703 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
705 spin_lock(&svcpt->scp_rep_lock);
707 list_del_init(&rs->rs_exp_list);
708 spin_lock(&rs->rs_lock);
709 ptlrpc_schedule_difficult_reply(rs);
710 spin_unlock(&rs->rs_lock);
712 spin_unlock(&svcpt->scp_rep_lock);
714 spin_unlock(&exp->exp_lock);
718 EXPORT_SYMBOL(server_disconnect_export);
720 /* --------------------------------------------------------------------------
721 * from old lib/target.c
722 * -------------------------------------------------------------------------- */
724 static int target_handle_reconnect(struct lustre_handle *conn,
725 struct obd_export *exp,
726 struct obd_uuid *cluuid)
728 struct obd_device *target;
729 struct lustre_handle *hdl;
736 hdl = &exp->exp_imp_reverse->imp_remote_handle;
737 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
738 conn->cookie = exp->exp_handle.h_cookie;
739 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
740 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
744 target = exp->exp_obd;
746 /* Might be a re-connect after a partition. */
747 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
748 LCONSOLE_WARN("%s: already connected client %s (at %s) "
749 "with handle %#llx. Rejecting client "
750 "with the same UUID trying to reconnect "
751 "with handle %#llx\n", target->obd_name,
752 obd_uuid2str(&exp->exp_client_uuid),
753 obd_export_nid2str(exp),
754 hdl->cookie, conn->cookie);
755 memset(conn, 0, sizeof *conn);
756 /* target_handle_connect() treats EALREADY and
757 * -EALREADY differently. -EALREADY is an error
758 * (same UUID, different handle). */
762 if (!target->obd_recovering) {
763 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
764 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
765 obd_export_nid2str(exp));
766 GOTO(out_already, rc);
769 now = cfs_time_current();
770 deadline = cfs_timer_deadline(&target->obd_recovery_timer);
771 if (cfs_time_before(now, deadline)) {
772 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
773 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting,"
774 " waiting for %d clients in recovery for"
775 " %d:%.02d\n", target->obd_name,
776 obd_uuid2str(&exp->exp_client_uuid),
777 obd_export_nid2str(exp),
778 target->obd_max_recoverable_clients,
779 timeout / 60, timeout % 60);
781 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
782 LCONSOLE_WARN("%s: Recovery already passed deadline"
783 " %d:%.02d, It is most likely due to DNE"
784 " recovery is failed or stuck, please wait a"
785 " few more minutes or abort the recovery.\n",
786 target->obd_name, timeout / 60, timeout % 60);
790 conn->cookie = exp->exp_handle.h_cookie;
791 /* target_handle_connect() treats EALREADY and
792 * -EALREADY differently. EALREADY means we are
793 * doing a valid reconnect from the same client. */
797 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
800 struct obd_export *exp = cb_data;
802 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
803 obd->obd_name, exp->exp_client_uuid.uuid);
805 spin_lock(&exp->exp_lock);
806 exp->exp_need_sync = 0;
807 spin_unlock(&exp->exp_lock);
808 class_export_cb_put(exp);
812 check_and_start_recovery_timer(struct obd_device *obd,
813 struct ptlrpc_request *req, int new_client);
816 * update flags for import during reconnect process
818 static int rev_import_flags_update(struct obd_import *revimp,
819 struct ptlrpc_request *req)
822 struct obd_connect_data *data;
824 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
826 if (data->ocd_connect_flags & OBD_CONNECT_AT)
827 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
829 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
831 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
833 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
835 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
836 revimp->imp_client->cli_name,
837 libcfs_id2str(req->rq_peer), rc);
845 * Allocate a new reverse import for an export.
847 * \retval -errno in case error hit
848 * \retval 0 if reverse import correctly init
850 int rev_import_init(struct obd_export *export)
852 struct obd_device *obd = export->exp_obd;
853 struct obd_import *revimp;
855 LASSERT(export->exp_imp_reverse == NULL);
857 revimp = class_new_import(obd);
861 revimp->imp_remote_handle.cookie = 0ULL;
862 revimp->imp_client = &obd->obd_ldlm_client;
863 revimp->imp_dlm_fake = 1;
865 /* it is safe to connect import in new state as no sends possible */
866 spin_lock(&export->exp_lock);
867 export->exp_imp_reverse = revimp;
868 spin_unlock(&export->exp_lock);
869 class_import_put(revimp);
873 EXPORT_SYMBOL(rev_import_init);
876 * Handle reconnect for an export.
878 * \param exp export to handle reconnect process
879 * \param req client reconnect request
881 * \retval -rc in case securitfy flavor can't be changed
882 * \retval 0 in case none problems
884 static int rev_import_reconnect(struct obd_export *exp,
885 struct ptlrpc_request *req)
887 struct obd_import *revimp = exp->exp_imp_reverse;
888 struct lustre_handle *lh;
891 /* avoid sending a request until import flags are changed */
892 ptlrpc_import_enter_resend(revimp);
894 if (revimp->imp_connection != NULL)
895 ptlrpc_connection_put(revimp->imp_connection);
898 * client from recovery don't have a handle so we need to take from
899 * request. it may produce situation when wrong client connected
900 * to recovery as we trust a client uuid
902 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
903 revimp->imp_remote_handle = *lh;
905 /* unknown versions will be caught in
906 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
907 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
909 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
911 rc = rev_import_flags_update(revimp, req);
913 /* it is safe to still be in RECOVERY phase as we are not able
914 * to setup correct security flavor so requests are not able to
915 * be delivered correctly */
919 /* resend all rpc's via new connection */
920 return ptlrpc_import_recovery_state_machine(revimp);
923 int target_handle_connect(struct ptlrpc_request *req)
925 struct obd_device *target = NULL;
926 struct obd_export *export = NULL;
927 /* connect handle - filled from target_handle_reconnect in
929 struct lustre_handle conn;
930 struct lustre_handle *tmp;
931 struct obd_uuid tgtuuid;
932 struct obd_uuid cluuid;
937 bool mds_conn = false, lw_client = false;
938 bool mds_mds_conn = false;
939 bool new_mds_mds_conn = false;
940 bool target_referenced = false;
941 struct obd_connect_data *data, *tmpdata;
943 lnet_nid_t *client_nid = NULL;
946 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
948 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
950 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
951 GOTO(out, rc = -EINVAL);
954 obd_str2uuid(&tgtuuid, str);
955 target = class_uuid2obd(&tgtuuid);
957 target = class_name2obd(str);
960 deuuidify(str, NULL, &target_start, &target_len);
961 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
962 "from %s (no target). If you are running "
963 "an HA pair check that the target is "
964 "mounted on the other server.\n", str,
965 libcfs_nid2str(req->rq_peer.nid));
966 GOTO(out, rc = -ENODEV);
969 spin_lock(&target->obd_dev_lock);
970 if (target->obd_stopping || !target->obd_set_up) {
971 spin_unlock(&target->obd_dev_lock);
973 deuuidify(str, NULL, &target_start, &target_len);
974 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
975 target_len, target_start,
976 libcfs_nid2str(req->rq_peer.nid),
977 (target->obd_stopping ?
978 "stopping" : "not set up"));
979 GOTO(out, rc = -ENODEV);
982 if (target->obd_no_conn) {
983 spin_unlock(&target->obd_dev_lock);
985 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
986 "from %s\n", target->obd_name,
987 libcfs_nid2str(req->rq_peer.nid));
988 GOTO(out, rc = -EAGAIN);
991 /* Make sure the target isn't cleaned up while we're here. Yes,
992 * there's still a race between the above check and our incref here.
993 * Really, class_uuid2obd should take the ref. */
994 class_incref(target, __func__, current);
995 target_referenced = true;
997 target->obd_conn_inprogress++;
998 spin_unlock(&target->obd_dev_lock);
1000 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1002 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1003 GOTO(out, rc = -EINVAL);
1006 obd_str2uuid(&cluuid, str);
1008 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1010 GOTO(out, rc = -EPROTO);
1014 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1016 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1018 GOTO(out, rc = -EPROTO);
1020 rc = req_capsule_server_pack(&req->rq_pill);
1024 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1025 /* Don't allow clients to connect that are using old 1.8 format
1026 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1027 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1028 * FULL20 flag should be set on all connections since 2.0, but no
1029 * longer affects behaviour.
1031 * Later this check will be disabled and the flag can be retired
1032 * completely once interop with 3.0 is no longer needed.
1034 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1035 GOTO(out, rc = -EPROTO);
1038 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1039 if (data->ocd_version < LUSTRE_VERSION_CODE -
1040 LUSTRE_VERSION_ALLOWED_OFFSET ||
1041 data->ocd_version > LUSTRE_VERSION_CODE +
1042 LUSTRE_VERSION_ALLOWED_OFFSET) {
1043 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1044 "libclient connection attempt",
1045 data->ocd_version < LUSTRE_VERSION_CODE ?
1047 OBD_OCD_VERSION_MAJOR(data->ocd_version),
1048 OBD_OCD_VERSION_MINOR(data->ocd_version),
1049 OBD_OCD_VERSION_PATCH(data->ocd_version),
1050 OBD_OCD_VERSION_FIX(data->ocd_version));
1051 data = req_capsule_server_sized_get(&req->rq_pill,
1053 offsetof(typeof(*data), ocd_version) +
1054 sizeof(data->ocd_version));
1056 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1057 data->ocd_version = LUSTRE_VERSION_CODE;
1059 GOTO(out, rc = -EPROTO);
1063 /* Note: lw_client is needed in MDS-MDS failover during update log
1064 * processing, so we needs to allow lw_client to be connected at
1065 * anytime, instead of only the initial connection */
1066 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1068 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1069 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1070 mds_mds_conn = (data->ocd_connect_flags &
1071 OBD_CONNECT_MDS_MDS) != 0;
1073 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1074 * for Imperative Recovery connection from MGC to MGS.
1076 * Via check OBD_CONNECT_FID, we can distinguish whether
1077 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1080 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1081 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1082 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1083 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1084 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1085 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1087 /* We do not support the MDT-MDT interoperations with
1088 * different version MDT because of protocol changes. */
1089 if (unlikely(major != LUSTRE_MAJOR ||
1090 minor != LUSTRE_MINOR ||
1091 abs(patch - LUSTRE_PATCH) > 3)) {
1092 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1093 "connection from different version MDT "
1094 "(%d.%d.%d.%d) %s %s\n",
1095 target->obd_name, LUSTRE_MAJOR,
1096 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1097 major, minor, patch,
1098 OBD_OCD_VERSION_FIX(data->ocd_version),
1099 libcfs_nid2str(req->rq_peer.nid), str);
1101 GOTO(out, rc = -EPROTO);
1106 /* lctl gets a backstage, all-access pass. */
1107 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1108 goto dont_check_exports;
1110 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1114 /* We've found an export in the hash. */
1116 spin_lock(&export->exp_lock);
1118 if (export->exp_connecting) { /* bug 9635, et. al. */
1119 spin_unlock(&export->exp_lock);
1120 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1121 export->exp_obd->obd_name, export,
1122 libcfs_nid2str(req->rq_peer.nid));
1123 class_export_put(export);
1126 } else if ((mds_conn || lw_client ||
1127 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1128 export->exp_connection != NULL) {
1129 spin_unlock(&export->exp_lock);
1130 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1131 /* MDS or LWP reconnected after failover. */
1132 LCONSOLE_WARN("%s: Received %s connection from "
1133 "%s, removing former export from %s\n",
1134 target->obd_name, mds_conn ? "MDS" : "LWP",
1135 libcfs_nid2str(req->rq_peer.nid),
1136 libcfs_nid2str(export->exp_connection->c_peer.nid));
1138 /* New MDS connection from the same NID. */
1139 LCONSOLE_WARN("%s: Received new %s connection from "
1140 "%s, removing former export from same NID\n",
1141 target->obd_name, mds_conn ? "MDS" : "LWP",
1142 libcfs_nid2str(req->rq_peer.nid));
1145 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1146 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1147 /* Because exports between MDTs will always be
1148 * kept, let's do not fail such export if they
1149 * come from the same NID, otherwise it might
1150 * cause eviction between MDTs, which might
1151 * cause namespace inconsistency */
1152 spin_lock(&export->exp_lock);
1153 export->exp_connecting = 1;
1154 spin_unlock(&export->exp_lock);
1155 conn.cookie = export->exp_handle.h_cookie;
1158 class_fail_export(export);
1159 class_export_put(export);
1163 } else if (export->exp_connection != NULL &&
1164 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1165 (lustre_msg_get_op_flags(req->rq_reqmsg) &
1166 MSG_CONNECT_INITIAL)) {
1167 spin_unlock(&export->exp_lock);
1168 /* In MDS failover we have static UUID but NID can change. */
1169 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1170 "existing nid %s is already connected\n",
1171 target->obd_name, cluuid.uuid,
1172 libcfs_nid2str(req->rq_peer.nid),
1174 export->exp_connection->c_peer.nid));
1176 class_export_put(export);
1179 export->exp_connecting = 1;
1180 spin_unlock(&export->exp_lock);
1181 LASSERT(export->exp_obd == target);
1183 rc = target_handle_reconnect(&conn, export, &cluuid);
1186 /* If we found an export, we already unlocked. */
1189 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1190 } else if (req->rq_export == NULL &&
1191 atomic_read(&export->exp_rpc_count) > 0) {
1192 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1193 "still busy with %d references\n",
1194 target->obd_name, cluuid.uuid,
1195 libcfs_nid2str(req->rq_peer.nid),
1196 atomic_read(&export->exp_refcount));
1197 GOTO(out, rc = -EBUSY);
1198 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1199 if (!strstr(cluuid.uuid, "mdt"))
1200 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1201 "known client %s (at %s) because it "
1202 "is indicating it is a new client",
1203 target->obd_name, cluuid.uuid,
1204 libcfs_nid2str(req->rq_peer.nid));
1205 GOTO(out, rc = -EALREADY);
1207 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1214 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1215 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1216 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1217 export, (long)cfs_time_current_sec(),
1218 export ? (long)export->exp_last_request_time : 0);
1220 /* If this is the first time a client connects, reset the recovery
1221 * timer. Discard lightweight connections which might be local. */
1222 if (!lw_client && rc == 0 && target->obd_recovering)
1223 check_and_start_recovery_timer(target, req, export == NULL);
1225 /* We want to handle EALREADY but *not* -EALREADY from
1226 * target_handle_reconnect(), return reconnection state in a flag. */
1227 if (rc == EALREADY) {
1228 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1234 /* Tell the client if we support replayable requests. */
1235 if (target->obd_replayable)
1236 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1237 client_nid = &req->rq_peer.nid;
1239 if (export == NULL) {
1240 /* allow lightweight connections during recovery */
1241 /* allow "new" MDT to be connected during recovery, since we
1242 * need retrieve recovery update records from it */
1243 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1245 int c; /* connected */
1246 int i; /* in progress */
1248 int s; /* stale/evicted */
1250 c = atomic_read(&target->obd_connected_clients);
1251 i = atomic_read(&target->obd_lock_replay_clients);
1252 k = target->obd_max_recoverable_clients;
1253 s = target->obd_stale_clients;
1254 t = cfs_timer_deadline(&target->obd_recovery_timer);
1255 t = cfs_time_sub(t, cfs_time_current());
1256 t = cfs_duration_sec(t);
1257 LCONSOLE_WARN("%s: Denying connection for new client %s"
1258 "(at %s), waiting for %d known clients "
1259 "(%d recovered, %d in progress, and %d "
1260 "evicted) to recover in %d:%.02d\n",
1261 target->obd_name, cluuid.uuid,
1262 libcfs_nid2str(req->rq_peer.nid), k,
1263 c - i, i, s, (int)t / 60,
1268 rc = obd_connect(req->rq_svc_thread->t_env,
1269 &export, target, &cluuid, data,
1271 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1272 lustre_msg_add_op_flags(req->rq_repmsg,
1273 MSG_CONNECT_RECOVERING);
1275 conn.cookie = export->exp_handle.h_cookie;
1276 rc = rev_import_init(export);
1280 new_mds_mds_conn = true;
1283 rc = obd_reconnect(req->rq_svc_thread->t_env,
1284 export, target, &cluuid, data, client_nid);
1289 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1290 data->ocd_instance = target->u.obt.obt_instance;
1292 /* Return only the parts of obd_connect_data that we understand, so the
1293 * client knows that we don't understand the rest. */
1295 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1297 tmpdata = req_capsule_server_get(&req->rq_pill,
1299 /* Don't use struct assignment here, because the client reply
1300 * buffer may be smaller/larger than the local struct
1301 * obd_connect_data. */
1302 memcpy(tmpdata, data, min(tmpsize, size));
1305 /* If the client and the server are the same node, we will already
1306 * have an export that really points to the client's DLM export,
1307 * because we have a shared handles table.
1309 * XXX this will go away when shaver stops sending the "connect" handle
1310 * in the real "remote handle" field of the request --phik 24 Apr 2003
1312 ptlrpc_request_change_export(req, export);
1314 spin_lock(&export->exp_lock);
1315 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1316 spin_unlock(&export->exp_lock);
1317 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1318 "or equal conn_cnt: %d >= %d\n",
1319 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1320 export->exp_conn_cnt,
1321 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1323 GOTO(out, rc = -EALREADY);
1325 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1326 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1328 /* Don't evict liblustre clients for not pinging. */
1329 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1330 export->exp_libclient = 1;
1331 spin_unlock(&export->exp_lock);
1333 spin_lock(&target->obd_dev_lock);
1334 list_del_init(&export->exp_obd_chain_timed);
1335 spin_unlock(&target->obd_dev_lock);
1337 spin_unlock(&export->exp_lock);
1340 if (export->exp_connection != NULL) {
1341 /* Check to see if connection came from another NID. */
1342 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1343 !hlist_unhashed(&export->exp_nid_hash))
1344 cfs_hash_del(export->exp_obd->obd_nid_hash,
1345 &export->exp_connection->c_peer.nid,
1346 &export->exp_nid_hash);
1348 ptlrpc_connection_put(export->exp_connection);
1351 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1354 if (hlist_unhashed(&export->exp_nid_hash))
1355 cfs_hash_add(export->exp_obd->obd_nid_hash,
1356 &export->exp_connection->c_peer.nid,
1357 &export->exp_nid_hash);
1359 lustre_msg_set_handle(req->rq_repmsg, &conn);
1361 rc = rev_import_reconnect(export, req);
1365 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1367 __u64 transno = data->ocd_transno;
1369 spin_lock(&export->exp_lock);
1370 /* possible race with class_disconnect_stale_exports,
1371 * export may be already in the eviction process */
1372 if (export->exp_failed) {
1373 spin_unlock(&export->exp_lock);
1374 GOTO(out, rc = -ENODEV);
1376 export->exp_in_recovery = 1;
1377 export->exp_req_replay_needed = 1;
1378 export->exp_lock_replay_needed = 1;
1379 spin_unlock(&export->exp_lock);
1381 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1382 MSG_CONNECT_TRANSNO);
1383 if (has_transno && transno == 0)
1384 CWARN("Connect with zero transno!\n");
1386 if (has_transno && transno > 0 &&
1387 transno < target->obd_next_recovery_transno &&
1388 transno > target->obd_last_committed) {
1389 /* Another way is to use cmpxchg() to be lock-free. */
1390 spin_lock(&target->obd_recovery_task_lock);
1391 if (transno < target->obd_next_recovery_transno)
1392 target->obd_next_recovery_transno = transno;
1393 spin_unlock(&target->obd_recovery_task_lock);
1396 atomic_inc(&target->obd_req_replay_clients);
1397 atomic_inc(&target->obd_lock_replay_clients);
1398 /* Note: MDS-MDS connection is allowed to be connected during
1399 * recovery, no matter if the exports needs to be recoveried.
1400 * Because we need retrieve updates logs from all other MDTs.
1401 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1402 * also needs to be increased to match other recovery checking
1404 if (new_mds_mds_conn)
1405 target->obd_max_recoverable_clients++;
1406 if (atomic_inc_return(&target->obd_connected_clients) ==
1407 target->obd_max_recoverable_clients)
1408 wake_up(&target->obd_next_transno_waitq);
1411 /* Tell the client we're in recovery, when client is involved in it. */
1412 if (target->obd_recovering && !lw_client)
1413 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1417 spin_lock(&export->exp_lock);
1418 export->exp_connecting = 0;
1419 spin_unlock(&export->exp_lock);
1421 class_export_put(export);
1423 if (target_referenced == true && target != NULL) {
1424 spin_lock(&target->obd_dev_lock);
1425 target->obd_conn_inprogress--;
1426 spin_unlock(&target->obd_dev_lock);
1428 class_decref(target, __func__, current);
1430 req->rq_status = rc;
1434 int target_handle_disconnect(struct ptlrpc_request *req)
1439 rc = req_capsule_server_pack(&req->rq_pill);
1443 /* Keep the rq_export around so we can send the reply. */
1444 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1449 void target_destroy_export(struct obd_export *exp)
1451 struct obd_import *imp = NULL;
1452 /* exports created from last_rcvd data, and "fake"
1453 exports created by lctl don't have an import */
1454 spin_lock(&exp->exp_lock);
1455 if (exp->exp_imp_reverse != NULL) {
1456 imp = exp->exp_imp_reverse;
1457 exp->exp_imp_reverse = NULL;
1459 spin_unlock(&exp->exp_lock);
1461 client_destroy_import(imp);
1463 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1464 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1465 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1466 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1468 EXPORT_SYMBOL(target_destroy_export);
1471 * Recovery functions
1473 static void target_request_copy_get(struct ptlrpc_request *req)
1475 class_export_rpc_inc(req->rq_export);
1476 LASSERT(list_empty(&req->rq_list));
1477 INIT_LIST_HEAD(&req->rq_replay_list);
1479 /* Increase refcount to keep request in queue. */
1480 atomic_inc(&req->rq_refcount);
1481 /* Let export know it has replays to be handled. */
1482 atomic_inc(&req->rq_export->exp_replay_count);
1485 static void target_request_copy_put(struct ptlrpc_request *req)
1487 LASSERT(list_empty(&req->rq_replay_list));
1488 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1490 atomic_dec(&req->rq_export->exp_replay_count);
1491 class_export_rpc_dec(req->rq_export);
1492 ptlrpc_server_drop_request(req);
1495 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1497 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1498 struct obd_export *exp = req->rq_export;
1499 struct ptlrpc_request *reqiter;
1500 struct ptlrpc_request *dup_req = NULL;
1505 spin_lock(&exp->exp_lock);
1506 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1508 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1516 /* We expect it with RESENT and REPLAY flags. */
1517 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1518 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1519 CERROR("invalid flags %x of resent replay\n",
1520 lustre_msg_get_flags(req->rq_reqmsg));
1522 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1525 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1527 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1528 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1532 list_add_tail(&req->rq_replay_list,
1533 &exp->exp_req_replay_queue);
1536 spin_unlock(&exp->exp_lock);
1540 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1542 LASSERT(!list_empty(&req->rq_replay_list));
1543 LASSERT(req->rq_export);
1545 spin_lock(&req->rq_export->exp_lock);
1546 list_del_init(&req->rq_replay_list);
1547 spin_unlock(&req->rq_export->exp_lock);
1550 static void target_finish_recovery(struct lu_target *lut)
1552 struct obd_device *obd = lut->lut_obd;
1555 /* Only log a recovery message when recovery has occurred. */
1556 if (obd->obd_recovery_start) {
1557 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1558 obd->obd_recovery_start);
1559 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1560 "%d recovered and %d %s evicted.\n", obd->obd_name,
1561 (int)elapsed_time / 60, (int)elapsed_time % 60,
1562 obd->obd_max_recoverable_clients,
1563 atomic_read(&obd->obd_connected_clients),
1564 obd->obd_stale_clients,
1565 obd->obd_stale_clients == 1 ? "was" : "were");
1568 ldlm_reprocess_all_ns(obd->obd_namespace);
1569 spin_lock(&obd->obd_recovery_task_lock);
1570 if (!list_empty(&obd->obd_req_replay_queue) ||
1571 !list_empty(&obd->obd_lock_replay_queue) ||
1572 !list_empty(&obd->obd_final_req_queue)) {
1573 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1575 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1576 list_empty(&obd->obd_lock_replay_queue) ? \
1578 list_empty(&obd->obd_final_req_queue) ? \
1580 spin_unlock(&obd->obd_recovery_task_lock);
1583 spin_unlock(&obd->obd_recovery_task_lock);
1585 obd->obd_recovery_end = cfs_time_current_sec();
1587 /* When recovery finished, cleanup orphans on MDS and OST. */
1588 if (OBT(obd) && OBP(obd, postrecov)) {
1589 int rc = OBP(obd, postrecov)(obd);
1591 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1597 static void abort_req_replay_queue(struct obd_device *obd)
1599 struct ptlrpc_request *req, *n;
1600 struct list_head abort_list;
1602 INIT_LIST_HEAD(&abort_list);
1603 spin_lock(&obd->obd_recovery_task_lock);
1604 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1605 spin_unlock(&obd->obd_recovery_task_lock);
1606 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1607 DEBUG_REQ(D_WARNING, req, "aborted:");
1608 req->rq_status = -ENOTCONN;
1609 if (ptlrpc_error(req)) {
1610 DEBUG_REQ(D_ERROR, req,
1611 "failed abort_req_reply; skipping");
1613 target_exp_dequeue_req_replay(req);
1614 target_request_copy_put(req);
1618 static void abort_lock_replay_queue(struct obd_device *obd)
1620 struct ptlrpc_request *req, *n;
1621 struct list_head abort_list;
1623 INIT_LIST_HEAD(&abort_list);
1624 spin_lock(&obd->obd_recovery_task_lock);
1625 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1626 spin_unlock(&obd->obd_recovery_task_lock);
1627 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1628 DEBUG_REQ(D_ERROR, req, "aborted:");
1629 req->rq_status = -ENOTCONN;
1630 if (ptlrpc_error(req)) {
1631 DEBUG_REQ(D_ERROR, req,
1632 "failed abort_lock_reply; skipping");
1634 target_request_copy_put(req);
1638 /* Called from a cleanup function if the device is being cleaned up
1639 forcefully. The exports should all have been disconnected already,
1640 the only thing left to do is
1641 - clear the recovery flags
1643 - free queued requests and replies, but don't send replies
1644 Because the obd_stopping flag is set, no new requests should be received.
1647 void target_cleanup_recovery(struct obd_device *obd)
1649 struct ptlrpc_request *req, *n;
1650 struct list_head clean_list;
1652 INIT_LIST_HEAD(&clean_list);
1653 spin_lock(&obd->obd_dev_lock);
1654 if (!obd->obd_recovering) {
1655 spin_unlock(&obd->obd_dev_lock);
1659 obd->obd_recovering = obd->obd_abort_recovery = 0;
1660 spin_unlock(&obd->obd_dev_lock);
1662 spin_lock(&obd->obd_recovery_task_lock);
1663 target_cancel_recovery_timer(obd);
1664 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1665 spin_unlock(&obd->obd_recovery_task_lock);
1667 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1668 LASSERT(req->rq_reply_state == NULL);
1669 target_exp_dequeue_req_replay(req);
1670 target_request_copy_put(req);
1673 spin_lock(&obd->obd_recovery_task_lock);
1674 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1675 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1676 spin_unlock(&obd->obd_recovery_task_lock);
1678 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1679 LASSERT(req->rq_reply_state == NULL);
1680 target_request_copy_put(req);
1685 EXPORT_SYMBOL(target_cleanup_recovery);
1687 /* obd_recovery_task_lock should be held */
1688 void target_cancel_recovery_timer(struct obd_device *obd)
1690 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1691 cfs_timer_disarm(&obd->obd_recovery_timer);
1694 static void target_start_recovery_timer(struct obd_device *obd)
1696 if (obd->obd_recovery_start != 0)
1699 spin_lock(&obd->obd_dev_lock);
1700 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1701 spin_unlock(&obd->obd_dev_lock);
1705 LASSERT(obd->obd_recovery_timeout != 0);
1707 if (obd->obd_recovery_start != 0) {
1708 spin_unlock(&obd->obd_dev_lock);
1712 cfs_timer_arm(&obd->obd_recovery_timer,
1713 cfs_time_shift(obd->obd_recovery_timeout));
1714 obd->obd_recovery_start = cfs_time_current_sec();
1715 spin_unlock(&obd->obd_dev_lock);
1717 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1718 "or until %d client%s reconnect%s\n",
1720 obd->obd_recovery_timeout / 60,
1721 obd->obd_recovery_timeout % 60,
1722 obd->obd_max_recoverable_clients,
1723 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1724 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1728 * extend recovery window.
1730 * if @extend is true, extend recovery window to have @drt remaining at least;
1731 * otherwise, make sure the recovery timeout value is not less than @drt.
1733 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1737 cfs_duration_t left;
1740 spin_lock(&obd->obd_dev_lock);
1741 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1742 spin_unlock(&obd->obd_dev_lock);
1745 LASSERT(obd->obd_recovery_start != 0);
1747 now = cfs_time_current_sec();
1748 to = obd->obd_recovery_timeout;
1749 end = obd->obd_recovery_start + to;
1750 left = cfs_time_sub(end, now);
1752 if (extend && (drt > left)) {
1754 } else if (!extend && (drt > to)) {
1758 if (to > obd->obd_recovery_time_hard)
1759 to = obd->obd_recovery_time_hard;
1760 if (obd->obd_recovery_timeout < to) {
1761 obd->obd_recovery_timeout = to;
1762 end = obd->obd_recovery_start + to;
1763 cfs_timer_arm(&obd->obd_recovery_timer,
1764 cfs_time_shift(end - now));
1766 spin_unlock(&obd->obd_dev_lock);
1768 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1769 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1772 /* Reset the timer with each new client connection */
1774 * This timer is actually reconnect_timer, which is for making sure
1775 * the total recovery window is at least as big as my reconnect
1776 * attempt timing. So the initial recovery time_out will be set to
1777 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1778 * from client is bigger than this, then the recovery time_out will
1779 * be extended to make sure the client could be reconnected, in the
1780 * process, the timeout from the new client should be ignored.
1784 check_and_start_recovery_timer(struct obd_device *obd,
1785 struct ptlrpc_request *req,
1788 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1789 struct obd_device_target *obt = &obd->u.obt;
1791 if (!new_client && service_time)
1792 /* Teach server about old server's estimates, as first guess
1793 * at how long new requests will take. */
1794 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1797 target_start_recovery_timer(obd);
1799 /* Convert the service time to RPC timeout,
1800 * and reuse service_time to limit stack usage. */
1801 service_time = at_est2timeout(service_time);
1803 /* We expect other clients to timeout within service_time, then try
1804 * to reconnect, then try the failover server. The max delay between
1805 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1806 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1808 LASSERT(obt->obt_magic == OBT_MAGIC);
1809 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1810 if (service_time > obd->obd_recovery_timeout && !new_client)
1811 extend_recovery_timer(obd, service_time, false);
1814 /** Health checking routines */
1815 static inline int exp_connect_healthy(struct obd_export *exp)
1817 return (exp->exp_in_recovery);
1820 /** if export done req_replay or has replay in queue */
1821 static inline int exp_req_replay_healthy(struct obd_export *exp)
1823 return (!exp->exp_req_replay_needed ||
1824 atomic_read(&exp->exp_replay_count) > 0);
1828 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1830 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1831 exp_req_replay_healthy(exp);
1834 /** if export done lock_replay or has replay in queue */
1835 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1837 return (!exp->exp_lock_replay_needed ||
1838 atomic_read(&exp->exp_replay_count) > 0);
1841 static inline int exp_vbr_healthy(struct obd_export *exp)
1843 return (!exp->exp_vbr_failed);
1846 static inline int exp_finished(struct obd_export *exp)
1848 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1851 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1853 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1857 static int check_for_next_transno(struct lu_target *lut)
1859 struct ptlrpc_request *req = NULL;
1860 struct obd_device *obd = lut->lut_obd;
1861 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1862 int wake_up = 0, connected, completed, queue_len;
1863 __u64 req_transno = 0;
1864 __u64 update_transno = 0;
1865 __u64 next_transno = 0;
1868 spin_lock(&obd->obd_recovery_task_lock);
1869 if (!list_empty(&obd->obd_req_replay_queue)) {
1870 req = list_entry(obd->obd_req_replay_queue.next,
1871 struct ptlrpc_request, rq_list);
1872 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1876 update_transno = distribute_txn_get_next_transno(tdtd);
1878 connected = atomic_read(&obd->obd_connected_clients);
1879 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1880 queue_len = obd->obd_requests_queued_for_recovery;
1881 next_transno = obd->obd_next_recovery_transno;
1883 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1884 "req_transno: %llu, next_transno: %llu\n",
1885 obd->obd_max_recoverable_clients, connected, completed,
1886 queue_len, req_transno, next_transno);
1888 if (obd->obd_abort_recovery) {
1889 CDEBUG(D_HA, "waking for aborted recovery\n");
1891 } else if (obd->obd_recovery_expired) {
1892 CDEBUG(D_HA, "waking for expired recovery\n");
1894 } else if (tdtd != NULL && req != NULL &&
1895 is_req_replayed_by_update(req)) {
1896 LASSERTF(req_transno < next_transno, "req_transno %llu"
1897 "next_transno%llu\n", req_transno, next_transno);
1898 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1901 } else if (req_transno == next_transno ||
1902 (update_transno != 0 && update_transno <= next_transno)) {
1903 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1905 } else if (queue_len > 0 &&
1906 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1907 /** handle gaps occured due to lost reply or VBR */
1908 LASSERTF(req_transno >= next_transno,
1909 "req_transno: %llu, next_transno: %llu\n",
1910 req_transno, next_transno);
1912 "%s: waking for gap in transno, VBR is %s (skip: "
1913 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1914 ", next_update %lld last_committed: %lld)\n",
1915 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1916 next_transno, queue_len, completed, connected,
1917 req_transno, update_transno, obd->obd_last_committed);
1918 obd->obd_next_recovery_transno = req_transno;
1920 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1921 CDEBUG(D_HA, "waking for completed recovery\n");
1923 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1924 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1925 " by fail_lock, waking up (%lld)\n", next_transno);
1926 obd->obd_next_recovery_transno = req_transno;
1929 spin_unlock(&obd->obd_recovery_task_lock);
1933 static int check_for_next_lock(struct lu_target *lut)
1935 struct obd_device *obd = lut->lut_obd;
1938 spin_lock(&obd->obd_recovery_task_lock);
1939 if (!list_empty(&obd->obd_lock_replay_queue)) {
1940 CDEBUG(D_HA, "waking for next lock\n");
1942 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1943 CDEBUG(D_HA, "waking for completed lock replay\n");
1945 } else if (obd->obd_abort_recovery) {
1946 CDEBUG(D_HA, "waking for aborted recovery\n");
1948 } else if (obd->obd_recovery_expired) {
1949 CDEBUG(D_HA, "waking for expired recovery\n");
1952 spin_unlock(&obd->obd_recovery_task_lock);
1958 * wait for recovery events,
1959 * check its status with help of check_routine
1960 * evict dead clients via health_check
1962 static int target_recovery_overseer(struct lu_target *lut,
1963 int (*check_routine)(struct lu_target *),
1964 int (*health_check)(struct obd_export *))
1966 struct obd_device *obd = lut->lut_obd;
1967 struct target_distribute_txn_data *tdtd;
1969 if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1970 (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1971 __u64 next_update_transno = 0;
1973 /* Only abort the recovery if there are no update recovery
1974 * left in the queue */
1975 spin_lock(&obd->obd_recovery_task_lock);
1976 if (lut->lut_tdtd != NULL) {
1977 next_update_transno =
1978 distribute_txn_get_next_transno(lut->lut_tdtd);
1980 tdtd = lut->lut_tdtd;
1981 /* If next_update_transno == 0, it probably because
1982 * updatelog retrieve threads did not get any records
1983 * yet, let's wait those threads stopped */
1984 if (next_update_transno == 0) {
1985 struct l_wait_info lwi = { 0 };
1987 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1989 &tdtd->tdtd_recovery_threads_count) == 0,
1992 next_update_transno =
1993 distribute_txn_get_next_transno(
1998 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
1999 obd->obd_next_recovery_transno = next_update_transno;
2000 spin_unlock(&obd->obd_recovery_task_lock);
2001 /* Disconnect unfinished exports from clients, and
2002 * keep connection from MDT to make sure the update
2003 * recovery will still keep trying until some one
2004 * manually abort the recovery */
2005 class_disconnect_stale_exports(obd,
2006 exp_finished_or_from_mdt);
2007 /* Abort all of replay and replay lock req from
2009 abort_req_replay_queue(obd);
2010 abort_lock_replay_queue(obd);
2011 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2012 ")in the queue.\n", obd->obd_name,
2013 next_update_transno);
2015 obd->obd_abort_recovery = 1;
2016 spin_unlock(&obd->obd_recovery_task_lock);
2017 CWARN("%s recovery is aborted by hard timeout\n",
2022 while (wait_event_timeout(obd->obd_next_transno_waitq,
2024 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2025 /* wait indefinitely for event, but don't trigger watchdog */;
2027 if (obd->obd_abort_recovery) {
2028 CWARN("recovery is aborted, evict exports in recovery\n");
2029 if (lut->lut_tdtd != NULL) {
2030 struct l_wait_info lwi = { 0 };
2032 tdtd = lut->lut_tdtd;
2033 /* Let's wait all of the update log recovery thread
2035 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2036 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2038 /* Then abort the update recovery list */
2039 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2040 dtrq_list_destroy(lut->lut_tdtd);
2043 /** evict exports which didn't finish recovery yet */
2044 class_disconnect_stale_exports(obd, exp_finished);
2046 } else if (obd->obd_recovery_expired) {
2047 obd->obd_recovery_expired = 0;
2048 /** If some clients died being recovered, evict them */
2049 LCONSOLE_WARN("%s: recovery is timed out, "
2050 "evict stale exports\n", obd->obd_name);
2051 /** evict cexports with no replay in queue, they are stalled */
2052 class_disconnect_stale_exports(obd, health_check);
2054 /** continue with VBR */
2055 spin_lock(&obd->obd_dev_lock);
2056 obd->obd_version_recov = 1;
2057 spin_unlock(&obd->obd_dev_lock);
2059 * reset timer, recovery will proceed with versions now,
2060 * timeout is set just to handle reconnection delays
2062 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2063 /** Wait for recovery events again, after evicting bad clients */
2069 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2071 struct obd_device *obd = lut->lut_obd;
2072 struct ptlrpc_request *req = NULL;
2074 CDEBUG(D_HA, "Waiting for lock\n");
2075 if (target_recovery_overseer(lut, check_for_next_lock,
2076 exp_lock_replay_healthy))
2077 abort_lock_replay_queue(obd);
2079 spin_lock(&obd->obd_recovery_task_lock);
2080 if (!list_empty(&obd->obd_lock_replay_queue)) {
2081 req = list_entry(obd->obd_lock_replay_queue.next,
2082 struct ptlrpc_request, rq_list);
2083 list_del_init(&req->rq_list);
2084 spin_unlock(&obd->obd_recovery_task_lock);
2086 spin_unlock(&obd->obd_recovery_task_lock);
2087 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2088 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2089 /** evict exports failed VBR */
2090 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2095 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2097 struct ptlrpc_request *req = NULL;
2099 spin_lock(&obd->obd_recovery_task_lock);
2100 if (!list_empty(&obd->obd_final_req_queue)) {
2101 req = list_entry(obd->obd_final_req_queue.next,
2102 struct ptlrpc_request, rq_list);
2103 list_del_init(&req->rq_list);
2104 spin_unlock(&obd->obd_recovery_task_lock);
2105 if (req->rq_export->exp_in_recovery) {
2106 spin_lock(&req->rq_export->exp_lock);
2107 req->rq_export->exp_in_recovery = 0;
2108 spin_unlock(&req->rq_export->exp_lock);
2111 spin_unlock(&obd->obd_recovery_task_lock);
2116 static void handle_recovery_req(struct ptlrpc_thread *thread,
2117 struct ptlrpc_request *req,
2118 svc_handler_t handler)
2123 * export can be evicted during recovery, no need to handle replays for
2124 * it after that, discard such request silently
2126 if (req->rq_export->exp_disconnected)
2129 req->rq_session.lc_thread = thread;
2130 req->rq_svc_thread = thread;
2131 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2133 /* thread context */
2134 lu_context_enter(&thread->t_env->le_ctx);
2136 lu_context_exit(&thread->t_env->le_ctx);
2138 /* don't reset timer for final stage */
2139 if (!exp_finished(req->rq_export)) {
2140 int to = obd_timeout;
2143 * Add request timeout to the recovery time so next request from
2144 * this client may come in recovery time
2147 struct ptlrpc_service_part *svcpt;
2149 svcpt = req->rq_rqbd->rqbd_svcpt;
2150 /* If the server sent early reply for this request,
2151 * the client will recalculate the timeout according to
2152 * current server estimate service time, so we will
2153 * use the maxium timeout here for waiting the client
2154 * sending the next req */
2155 to = max((int)at_est2timeout(
2156 at_get(&svcpt->scp_at_estimate)),
2157 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2158 /* Add 2 net_latency, one for balance rq_deadline
2159 * (see ptl_send_rpc), one for resend the req to server,
2160 * Note: client will pack net_latency in replay req
2161 * (see ptlrpc_replay_req) */
2162 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2164 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2169 /** Checking routines for recovery */
2170 static int check_for_recovery_ready(struct lu_target *lut)
2172 struct obd_device *obd = lut->lut_obd;
2173 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2175 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2176 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2177 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2178 obd->obd_recovery_expired);
2180 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2181 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2182 if (clnts + obd->obd_stale_clients <
2183 obd->obd_max_recoverable_clients)
2187 if (lut->lut_tdtd != NULL) {
2188 if (!lut->lut_tdtd->tdtd_replay_ready &&
2189 !obd->obd_abort_recovery) {
2190 /* Let's extend recovery timer, in case the recovery
2191 * timer expired, and some clients got evicted */
2192 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2194 CDEBUG(D_HA, "%s update recovery is not ready,"
2195 " extend recovery %d\n", obd->obd_name,
2196 obd->obd_recovery_timeout);
2199 dtrq_list_dump(lut->lut_tdtd, D_HA);
2207 REQUEST_RECOVERY = 1,
2208 UPDATE_RECOVERY = 2,
2211 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2215 if (!list_empty(&obd->obd_req_replay_queue)) {
2216 struct ptlrpc_request *req;
2218 req = list_entry(obd->obd_req_replay_queue.next,
2219 struct ptlrpc_request, rq_list);
2220 transno = lustre_msg_get_transno(req->rq_reqmsg);
2226 static __u64 get_next_transno(struct lu_target *lut, int *type)
2228 struct obd_device *obd = lut->lut_obd;
2229 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2231 __u64 update_transno;
2234 transno = get_next_replay_req_transno(obd);
2236 *type = REQUEST_RECOVERY;
2241 update_transno = distribute_txn_get_next_transno(tdtd);
2242 if (transno == 0 || (transno >= update_transno &&
2243 update_transno != 0)) {
2244 transno = update_transno;
2246 *type = UPDATE_RECOVERY;
2253 * drop duplicate replay request
2255 * Because the operation has been replayed by update recovery, the request
2256 * with the same transno will be dropped and also notify the client to send
2257 * next replay request.
2259 * \param[in] env execution environment
2260 * \param[in] obd failover obd device
2261 * \param[in] req request to be dropped
2263 static void drop_duplicate_replay_req(struct lu_env *env,
2264 struct obd_device *obd,
2265 struct ptlrpc_request *req)
2267 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2268 " update records are found.\n",
2269 lustre_msg_get_transno(req->rq_reqmsg),
2270 libcfs_nid2str(req->rq_peer.nid));
2272 /* Right now, only for MDS reint operation update replay and
2273 * normal request replay can have the same transno */
2274 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2275 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2276 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2277 if (likely(req->rq_export))
2278 target_committed_to_req(req);
2279 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2280 target_send_reply(req, req->rq_status, 0);
2282 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2283 libcfs_nid2str(req->rq_peer.nid));
2285 target_exp_dequeue_req_replay(req);
2286 target_request_copy_put(req);
2287 obd->obd_replayed_requests++;
2290 static void replay_request_or_update(struct lu_env *env,
2291 struct lu_target *lut,
2292 struct target_recovery_data *trd,
2293 struct ptlrpc_thread *thread)
2295 struct obd_device *obd = lut->lut_obd;
2296 struct ptlrpc_request *req = NULL;
2301 CDEBUG(D_HA, "Waiting for transno %lld\n",
2302 obd->obd_next_recovery_transno);
2304 /* Replay all of request and update by transno */
2306 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2308 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2310 /** It is needed to extend recovery window above
2311 * recovery_time_soft. Extending is possible only in the
2312 * end of recovery window (see more details in
2313 * handle_recovery_req()).
2315 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2317 if (target_recovery_overseer(lut, check_for_next_transno,
2318 exp_req_replay_healthy_or_from_mdt)) {
2319 abort_req_replay_queue(obd);
2320 abort_lock_replay_queue(obd);
2324 spin_lock(&obd->obd_recovery_task_lock);
2325 transno = get_next_transno(lut, &type);
2326 if (type == REQUEST_RECOVERY && transno != 0) {
2327 /* Drop replay request from client side, if the
2328 * replay has been executed by update with the
2330 req = list_entry(obd->obd_req_replay_queue.next,
2331 struct ptlrpc_request, rq_list);
2333 list_del_init(&req->rq_list);
2334 obd->obd_requests_queued_for_recovery--;
2335 spin_unlock(&obd->obd_recovery_task_lock);
2337 /* Let's check if the request has been redone by
2339 if (is_req_replayed_by_update(req)) {
2340 struct distribute_txn_replay_req *dtrq;
2342 dtrq = distribute_txn_lookup_finish_list(tdtd,
2344 LASSERT(dtrq != NULL);
2345 spin_lock(&tdtd->tdtd_replay_list_lock);
2346 list_del_init(&dtrq->dtrq_list);
2347 spin_unlock(&tdtd->tdtd_replay_list_lock);
2350 drop_duplicate_replay_req(env, obd, req);
2355 LASSERT(trd->trd_processing_task == current_pid());
2356 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2357 lustre_msg_get_transno(req->rq_reqmsg),
2358 libcfs_nid2str(req->rq_peer.nid));
2360 handle_recovery_req(thread, req,
2361 trd->trd_recovery_handler);
2363 * bz18031: increase next_recovery_transno before
2364 * target_request_copy_put() will drop exp_rpc reference
2366 spin_lock(&obd->obd_recovery_task_lock);
2367 obd->obd_next_recovery_transno++;
2368 spin_unlock(&obd->obd_recovery_task_lock);
2369 target_exp_dequeue_req_replay(req);
2370 target_request_copy_put(req);
2371 obd->obd_replayed_requests++;
2372 } else if (type == UPDATE_RECOVERY && transno != 0) {
2373 struct distribute_txn_replay_req *dtrq;
2376 spin_unlock(&obd->obd_recovery_task_lock);
2378 LASSERT(tdtd != NULL);
2379 dtrq = distribute_txn_get_next_req(tdtd);
2380 lu_context_enter(&thread->t_env->le_ctx);
2381 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2382 lu_context_exit(&thread->t_env->le_ctx);
2383 extend_recovery_timer(obd, obd_timeout, true);
2385 if (rc == 0 && dtrq->dtrq_xid != 0) {
2386 CDEBUG(D_HA, "Move x%llu t%llu"
2387 " to finish list\n", dtrq->dtrq_xid,
2388 dtrq->dtrq_master_transno);
2390 /* Add it to the replay finish list */
2391 spin_lock(&tdtd->tdtd_replay_list_lock);
2392 list_add(&dtrq->dtrq_list,
2393 &tdtd->tdtd_replay_finish_list);
2394 spin_unlock(&tdtd->tdtd_replay_list_lock);
2396 spin_lock(&obd->obd_recovery_task_lock);
2397 if (transno == obd->obd_next_recovery_transno)
2398 obd->obd_next_recovery_transno++;
2400 obd->obd_next_recovery_transno)
2401 obd->obd_next_recovery_transno =
2403 spin_unlock(&obd->obd_recovery_task_lock);
2408 spin_unlock(&obd->obd_recovery_task_lock);
2410 LASSERT(list_empty(&obd->obd_req_replay_queue));
2411 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2412 /** evict exports failed VBR */
2413 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2419 static int target_recovery_thread(void *arg)
2421 struct lu_target *lut = arg;
2422 struct obd_device *obd = lut->lut_obd;
2423 struct ptlrpc_request *req;
2424 struct target_recovery_data *trd = &obd->obd_recovery_data;
2425 unsigned long delta;
2427 struct ptlrpc_thread *thread = NULL;
2431 unshare_fs_struct();
2432 OBD_ALLOC_PTR(thread);
2438 OBD_FREE_PTR(thread);
2442 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2444 OBD_FREE_PTR(thread);
2449 thread->t_env = env;
2450 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2451 env->le_ctx.lc_thread = thread;
2452 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2453 thread->t_watchdog = NULL;
2455 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2457 trd->trd_processing_task = current_pid();
2459 spin_lock(&obd->obd_dev_lock);
2460 obd->obd_recovering = 1;
2461 spin_unlock(&obd->obd_dev_lock);
2462 complete(&trd->trd_starting);
2464 /* first of all, we have to know the first transno to replay */
2465 if (target_recovery_overseer(lut, check_for_recovery_ready,
2466 exp_connect_healthy)) {
2467 abort_req_replay_queue(obd);
2468 abort_lock_replay_queue(obd);
2469 if (lut->lut_tdtd != NULL)
2470 dtrq_list_destroy(lut->lut_tdtd);
2473 /* next stage: replay requests or update */
2475 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2476 atomic_read(&obd->obd_req_replay_clients),
2477 obd->obd_next_recovery_transno);
2478 replay_request_or_update(env, lut, trd, thread);
2481 * The second stage: replay locks
2483 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2484 atomic_read(&obd->obd_lock_replay_clients));
2485 while ((req = target_next_replay_lock(lut))) {
2486 LASSERT(trd->trd_processing_task == current_pid());
2487 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2488 libcfs_nid2str(req->rq_peer.nid));
2489 handle_recovery_req(thread, req,
2490 trd->trd_recovery_handler);
2491 target_request_copy_put(req);
2492 obd->obd_replayed_locks++;
2496 * The third stage: reply on final pings, at this moment all clients
2497 * must have request in final queue
2499 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2500 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2501 /** Update server last boot epoch */
2502 tgt_boot_epoch_update(lut);
2503 /* We drop recoverying flag to forward all new requests
2504 * to regular mds_handle() since now */
2505 spin_lock(&obd->obd_dev_lock);
2506 obd->obd_recovering = obd->obd_abort_recovery = 0;
2507 spin_unlock(&obd->obd_dev_lock);
2508 spin_lock(&obd->obd_recovery_task_lock);
2509 target_cancel_recovery_timer(obd);
2510 spin_unlock(&obd->obd_recovery_task_lock);
2511 while ((req = target_next_final_ping(obd))) {
2512 LASSERT(trd->trd_processing_task == current_pid());
2513 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2514 libcfs_nid2str(req->rq_peer.nid));
2515 handle_recovery_req(thread, req,
2516 trd->trd_recovery_handler);
2517 /* Because the waiting client can not send ping to server,
2518 * so we need refresh the last_request_time, to avoid the
2519 * export is being evicted */
2520 ptlrpc_update_export_timer(req->rq_export, 0);
2521 target_request_copy_put(req);
2524 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2525 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2526 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2527 if (delta > OBD_RECOVERY_TIME_SOFT) {
2528 CWARN("too long recovery - read logs\n");
2529 libcfs_debug_dumplog();
2532 target_finish_recovery(lut);
2534 lu_context_fini(&env->le_ctx);
2535 trd->trd_processing_task = 0;
2536 complete(&trd->trd_finishing);
2538 tgt_io_thread_done(thread);
2539 OBD_FREE_PTR(thread);
2544 static int target_start_recovery_thread(struct lu_target *lut,
2545 svc_handler_t handler)
2547 struct obd_device *obd = lut->lut_obd;
2549 struct target_recovery_data *trd = &obd->obd_recovery_data;
2552 memset(trd, 0, sizeof(*trd));
2553 init_completion(&trd->trd_starting);
2554 init_completion(&trd->trd_finishing);
2555 trd->trd_recovery_handler = handler;
2557 rc = server_name2index(obd->obd_name, &index, NULL);
2561 if (!IS_ERR(kthread_run(target_recovery_thread,
2562 lut, "tgt_recover_%d", index))) {
2563 wait_for_completion(&trd->trd_starting);
2564 LASSERT(obd->obd_recovering != 0);
2572 void target_stop_recovery_thread(struct obd_device *obd)
2574 if (obd->obd_recovery_data.trd_processing_task > 0) {
2575 struct target_recovery_data *trd = &obd->obd_recovery_data;
2576 /** recovery can be done but postrecovery is not yet */
2577 spin_lock(&obd->obd_dev_lock);
2578 if (obd->obd_recovering) {
2579 CERROR("%s: Aborting recovery\n", obd->obd_name);
2580 obd->obd_abort_recovery = 1;
2581 wake_up(&obd->obd_next_transno_waitq);
2583 spin_unlock(&obd->obd_dev_lock);
2584 wait_for_completion(&trd->trd_finishing);
2587 EXPORT_SYMBOL(target_stop_recovery_thread);
2589 void target_recovery_fini(struct obd_device *obd)
2591 class_disconnect_exports(obd);
2592 target_stop_recovery_thread(obd);
2593 target_cleanup_recovery(obd);
2595 EXPORT_SYMBOL(target_recovery_fini);
2597 static void target_recovery_expired(unsigned long castmeharder)
2599 struct obd_device *obd = (struct obd_device *)castmeharder;
2600 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2601 " after %lds (%d clients connected)\n",
2602 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2603 cfs_time_current_sec()- obd->obd_recovery_start,
2604 atomic_read(&obd->obd_connected_clients));
2606 obd->obd_recovery_expired = 1;
2607 wake_up(&obd->obd_next_transno_waitq);
2610 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2612 struct obd_device *obd = lut->lut_obd;
2614 if (obd->obd_max_recoverable_clients == 0) {
2615 /** Update server last boot epoch */
2616 tgt_boot_epoch_update(lut);
2620 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2621 "last_transno %llu\n", obd->obd_name,
2622 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2623 LASSERT(obd->obd_stopping == 0);
2624 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2625 obd->obd_recovery_start = 0;
2626 obd->obd_recovery_end = 0;
2628 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2629 target_start_recovery_thread(lut, handler);
2631 EXPORT_SYMBOL(target_recovery_init);
2633 static int target_process_req_flags(struct obd_device *obd,
2634 struct ptlrpc_request *req)
2636 struct obd_export *exp = req->rq_export;
2637 LASSERT(exp != NULL);
2638 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2639 /* client declares he's ready to replay locks */
2640 spin_lock(&exp->exp_lock);
2641 if (exp->exp_req_replay_needed) {
2642 exp->exp_req_replay_needed = 0;
2643 spin_unlock(&exp->exp_lock);
2645 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2646 atomic_dec(&obd->obd_req_replay_clients);
2648 spin_unlock(&exp->exp_lock);
2651 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2652 /* client declares he's ready to complete recovery
2653 * so, we put the request on th final queue */
2654 spin_lock(&exp->exp_lock);
2655 if (exp->exp_lock_replay_needed) {
2656 exp->exp_lock_replay_needed = 0;
2657 spin_unlock(&exp->exp_lock);
2659 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2660 atomic_dec(&obd->obd_lock_replay_clients);
2662 spin_unlock(&exp->exp_lock);
2668 int target_queue_recovery_request(struct ptlrpc_request *req,
2669 struct obd_device *obd)
2671 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2672 struct ptlrpc_request *reqiter;
2676 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2677 /* Processing the queue right now, don't re-add. */
2681 target_process_req_flags(obd, req);
2683 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2684 /* client declares he's ready to complete recovery
2685 * so, we put the request on th final queue */
2686 target_request_copy_get(req);
2687 DEBUG_REQ(D_HA, req, "queue final req");
2688 wake_up(&obd->obd_next_transno_waitq);
2689 spin_lock(&obd->obd_recovery_task_lock);
2690 if (obd->obd_recovering) {
2691 struct ptlrpc_request *tmp;
2692 struct ptlrpc_request *duplicate = NULL;
2694 if (likely(!req->rq_export->exp_replay_done)) {
2695 req->rq_export->exp_replay_done = 1;
2696 list_add_tail(&req->rq_list,
2697 &obd->obd_final_req_queue);
2698 spin_unlock(&obd->obd_recovery_task_lock);
2702 /* XXX O(n), but only happens if final ping is
2703 * timed out, probably reorganize the list as
2704 * a hash list later */
2705 list_for_each_entry_safe(reqiter, tmp,
2706 &obd->obd_final_req_queue,
2708 if (reqiter->rq_export == req->rq_export) {
2709 list_del_init(&reqiter->rq_list);
2710 duplicate = reqiter;
2715 list_add_tail(&req->rq_list,
2716 &obd->obd_final_req_queue);
2717 req->rq_export->exp_replay_done = 1;
2718 spin_unlock(&obd->obd_recovery_task_lock);
2720 if (duplicate != NULL) {
2721 DEBUG_REQ(D_HA, duplicate,
2722 "put prev final req\n");
2723 target_request_copy_put(duplicate);
2727 spin_unlock(&obd->obd_recovery_task_lock);
2728 target_request_copy_put(req);
2729 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2732 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2733 /* client declares he's ready to replay locks */
2734 target_request_copy_get(req);
2735 DEBUG_REQ(D_HA, req, "queue lock replay req");
2736 wake_up(&obd->obd_next_transno_waitq);
2737 spin_lock(&obd->obd_recovery_task_lock);
2738 LASSERT(obd->obd_recovering);
2739 /* usually due to recovery abort */
2740 if (!req->rq_export->exp_in_recovery) {
2741 spin_unlock(&obd->obd_recovery_task_lock);
2742 target_request_copy_put(req);
2745 LASSERT(req->rq_export->exp_lock_replay_needed);
2746 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2747 spin_unlock(&obd->obd_recovery_task_lock);
2751 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2752 * (i.e. buflens etc are in my own byte order), but type-dependent
2753 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2756 INIT_LIST_HEAD(&req->rq_list);
2757 DEBUG_REQ(D_HA, req, "not queueing");
2761 /* If we're processing the queue, we want don't want to queue this
2764 * Also, if this request has a transno less than the one we're waiting
2765 * for, we should process it now. It could (and currently always will)
2766 * be an open request for a descriptor that was opened some time ago.
2768 * Also, a resent, replayed request that has already been
2769 * handled will pass through here and be processed immediately.
2771 CDEBUG(D_HA, "Next recovery transno: %llu"
2772 ", current: %llu, replaying\n",
2773 obd->obd_next_recovery_transno, transno);
2775 /* If the request has been replayed by update replay, then sends this
2776 * request to the recovery thread (replay_request_or_update()), where
2777 * it will be handled */
2778 spin_lock(&obd->obd_recovery_task_lock);
2779 if (transno < obd->obd_next_recovery_transno &&
2780 !is_req_replayed_by_update(req)) {
2781 /* Processing the queue right now, don't re-add. */
2782 LASSERT(list_empty(&req->rq_list));
2783 spin_unlock(&obd->obd_recovery_task_lock);
2786 spin_unlock(&obd->obd_recovery_task_lock);
2788 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2791 target_request_copy_get(req);
2792 if (!req->rq_export->exp_in_recovery) {
2793 target_request_copy_put(req);
2796 LASSERT(req->rq_export->exp_req_replay_needed);
2798 if (target_exp_enqueue_req_replay(req)) {
2799 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2800 target_request_copy_put(req);
2805 spin_lock(&obd->obd_recovery_task_lock);
2806 LASSERT(obd->obd_recovering);
2807 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2808 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2809 list_add_tail(&req->rq_list, &reqiter->rq_list);
2814 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2816 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2817 "has been claimed by another client");
2818 spin_unlock(&obd->obd_recovery_task_lock);
2819 target_exp_dequeue_req_replay(req);
2820 target_request_copy_put(req);
2826 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2828 obd->obd_requests_queued_for_recovery++;
2829 spin_unlock(&obd->obd_recovery_task_lock);
2830 wake_up(&obd->obd_next_transno_waitq);
2834 int target_handle_ping(struct ptlrpc_request *req)
2836 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2837 return req_capsule_server_pack(&req->rq_pill);
2840 void target_committed_to_req(struct ptlrpc_request *req)
2842 struct obd_export *exp = req->rq_export;
2844 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2845 lustre_msg_set_last_committed(req->rq_repmsg,
2846 exp->exp_last_committed);
2848 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2849 "%d)", exp->exp_obd->obd_no_transno,
2850 req->rq_repmsg == NULL);
2852 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2853 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2856 #endif /* HAVE_SERVER_SUPPORT */
2859 * Packs current SLV and Limit into \a req.
2861 int target_pack_pool_reply(struct ptlrpc_request *req)
2863 struct obd_device *obd;
2866 /* Check that we still have all structures alive as this may
2867 * be some late RPC at shutdown time. */
2868 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2869 !exp_connect_lru_resize(req->rq_export))) {
2870 lustre_msg_set_slv(req->rq_repmsg, 0);
2871 lustre_msg_set_limit(req->rq_repmsg, 0);
2875 /* OBD is alive here as export is alive, which we checked above. */
2876 obd = req->rq_export->exp_obd;
2878 read_lock(&obd->obd_pool_lock);
2879 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2880 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2881 read_unlock(&obd->obd_pool_lock);
2886 static int target_send_reply_msg(struct ptlrpc_request *req,
2887 int rc, int fail_id)
2889 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2890 DEBUG_REQ(D_ERROR, req, "dropping reply");
2893 /* We can have a null rq_reqmsg in the event of bad signature or
2894 * no context when unwrapping */
2895 if (req->rq_reqmsg &&
2896 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2897 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2901 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2902 req->rq_status = rc;
2903 return ptlrpc_send_error(req, 1);
2905 DEBUG_REQ(D_NET, req, "sending reply");
2908 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2911 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2913 struct ptlrpc_service_part *svcpt;
2915 struct ptlrpc_reply_state *rs;
2916 struct obd_export *exp;
2919 if (req->rq_no_reply) {
2924 svcpt = req->rq_rqbd->rqbd_svcpt;
2925 rs = req->rq_reply_state;
2926 if (rs == NULL || !rs->rs_difficult) {
2928 target_send_reply_msg (req, rc, fail_id);
2933 /* must be an export if locks saved */
2934 LASSERT(req->rq_export != NULL);
2935 /* req/reply consistent */
2936 LASSERT(rs->rs_svcpt == svcpt);
2939 LASSERT(!rs->rs_scheduled);
2940 LASSERT(!rs->rs_scheduled_ever);
2941 LASSERT(!rs->rs_handled);
2942 LASSERT(!rs->rs_on_net);
2943 LASSERT(rs->rs_export == NULL);
2944 LASSERT(list_empty(&rs->rs_obd_list));
2945 LASSERT(list_empty(&rs->rs_exp_list));
2947 exp = class_export_get(req->rq_export);
2949 /* disable reply scheduling while I'm setting up */
2950 rs->rs_scheduled = 1;
2952 rs->rs_xid = req->rq_xid;
2953 rs->rs_transno = req->rq_transno;
2954 rs->rs_export = exp;
2955 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2957 spin_lock(&exp->exp_uncommitted_replies_lock);
2958 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2959 rs->rs_transno, exp->exp_last_committed);
2960 if (rs->rs_transno > exp->exp_last_committed) {
2961 /* not committed already */
2962 list_add_tail(&rs->rs_obd_list,
2963 &exp->exp_uncommitted_replies);
2965 spin_unlock(&exp->exp_uncommitted_replies_lock);
2967 spin_lock(&exp->exp_lock);
2968 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2969 spin_unlock(&exp->exp_lock);
2971 netrc = target_send_reply_msg(req, rc, fail_id);
2973 spin_lock(&svcpt->scp_rep_lock);
2975 atomic_inc(&svcpt->scp_nreps_difficult);
2978 /* error sending: reply is off the net. Also we need +1
2979 * reply ref until ptlrpc_handle_rs() is done
2980 * with the reply state (if the send was successful, there
2981 * would have been +1 ref for the net, which
2982 * reply_out_callback leaves alone) */
2984 ptlrpc_rs_addref(rs);
2987 spin_lock(&rs->rs_lock);
2988 if (rs->rs_transno <= exp->exp_last_committed ||
2989 (!rs->rs_on_net && !rs->rs_no_ack) ||
2990 list_empty(&rs->rs_exp_list) || /* completed already */
2991 list_empty(&rs->rs_obd_list)) {
2992 CDEBUG(D_HA, "Schedule reply immediately\n");
2993 ptlrpc_dispatch_difficult_reply(rs);
2995 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2996 rs->rs_scheduled = 0; /* allow notifier to schedule */
2998 spin_unlock(&rs->rs_lock);
2999 spin_unlock(&svcpt->scp_rep_lock);
3003 enum ldlm_mode lck_compat_array[] = {
3004 [LCK_EX] = LCK_COMPAT_EX,
3005 [LCK_PW] = LCK_COMPAT_PW,
3006 [LCK_PR] = LCK_COMPAT_PR,
3007 [LCK_CW] = LCK_COMPAT_CW,
3008 [LCK_CR] = LCK_COMPAT_CR,
3009 [LCK_NL] = LCK_COMPAT_NL,
3010 [LCK_GROUP] = LCK_COMPAT_GROUP,
3011 [LCK_COS] = LCK_COMPAT_COS,
3015 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3016 * not escape to the user level.
3018 int ldlm_error2errno(enum ldlm_error error)
3024 case ELDLM_LOCK_MATCHED:
3027 case ELDLM_LOCK_CHANGED:
3030 case ELDLM_LOCK_ABORTED:
3033 case ELDLM_LOCK_REPLACED:
3036 case ELDLM_NO_LOCK_DATA:
3039 case ELDLM_NAMESPACE_EXISTS:
3042 case ELDLM_BAD_NAMESPACE:
3046 if (((int)error) < 0) { /* cast to signed type */
3047 result = error; /* as ldlm_error can be unsigned */
3049 CERROR("Invalid DLM result code: %d\n", error);
3055 EXPORT_SYMBOL(ldlm_error2errno);
3058 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3060 enum ldlm_error ldlm_errno2error(int err_no)
3069 error = ELDLM_LOCK_CHANGED;
3072 error = ELDLM_LOCK_ABORTED;
3075 error = ELDLM_LOCK_REPLACED;
3078 error = ELDLM_NO_LOCK_DATA;
3081 error = ELDLM_NAMESPACE_EXISTS;
3084 error = ELDLM_BAD_NAMESPACE;
3092 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3093 void ldlm_dump_export_locks(struct obd_export *exp)
3095 spin_lock(&exp->exp_locks_list_guard);
3096 if (!list_empty(&exp->exp_locks_list)) {
3097 struct ldlm_lock *lock;
3099 CERROR("dumping locks for export %p,"
3100 "ignore if the unmount doesn't hang\n", exp);
3101 list_for_each_entry(lock, &exp->exp_locks_list,
3103 LDLM_ERROR(lock, "lock:");
3105 spin_unlock(&exp->exp_locks_list_guard);
3109 #ifdef HAVE_SERVER_SUPPORT
3110 static int target_bulk_timeout(void *data)
3113 /* We don't fail the connection here, because having the export
3114 * killed makes the (vital) call to commitrw very sad.
3119 static inline const char *bulk2type(struct ptlrpc_request *req)
3121 if (req->rq_bulk_read)
3123 if (req->rq_bulk_write)
3128 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3129 struct l_wait_info *lwi)
3131 struct ptlrpc_request *req = desc->bd_req;
3132 time_t start = cfs_time_current_sec();
3138 /* If there is eviction in progress, wait for it to finish. */
3139 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3140 *lwi = LWI_INTR(NULL, NULL);
3141 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3142 !atomic_read(&exp->exp_obd->
3143 obd_evict_inprogress),
3147 /* Check if client was evicted or reconnected already. */
3148 if (exp->exp_failed ||
3149 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3152 if (req->rq_bulk_read)
3153 rc = sptlrpc_svc_wrap_bulk(req, desc);
3155 if ((exp->exp_connect_data.ocd_connect_flags &
3156 OBD_CONNECT_BULK_MBITS) != 0)
3157 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3158 else /* old version, bulk matchbits is rq_xid */
3159 req->rq_mbits = req->rq_xid;
3162 rc = ptlrpc_start_bulk_transfer(desc);
3166 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3167 bulk2type(req), rc);
3171 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3172 ptlrpc_abort_bulk(desc);
3176 /* limit actual bulk transfer to bulk_timeout seconds */
3177 deadline = start + bulk_timeout;
3178 if (deadline > req->rq_deadline)
3179 deadline = req->rq_deadline;
3182 long timeoutl = deadline - cfs_time_current_sec();
3183 cfs_duration_t timeout = timeoutl <= 0 ?
3184 CFS_TICK : cfs_time_seconds(timeoutl);
3187 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3188 target_bulk_timeout, desc);
3189 rc = l_wait_event(desc->bd_waitq,
3190 !ptlrpc_server_bulk_active(desc) ||
3193 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3195 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3196 /* Wait again if we changed rq_deadline. */
3197 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3198 deadline = start + bulk_timeout;
3199 if (deadline > rq_deadline)
3200 deadline = rq_deadline;
3201 } while ((rc == -ETIMEDOUT) &&
3202 (deadline > cfs_time_current_sec()));
3204 if (rc == -ETIMEDOUT) {
3205 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3206 bulk2type(req), deadline - start,
3207 cfs_time_current_sec() - deadline);
3208 ptlrpc_abort_bulk(desc);
3209 } else if (exp->exp_failed) {
3210 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3213 ptlrpc_abort_bulk(desc);
3214 } else if (exp->exp_conn_cnt >
3215 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3216 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3218 /* We don't reply anyway. */
3220 ptlrpc_abort_bulk(desc);
3221 } else if (desc->bd_failure) {
3222 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3224 /* XXX should this be a different errno? */
3227 if (req->rq_bulk_write)
3228 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3229 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3230 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3231 bulk2type(req), desc->bd_nob_transferred,
3233 /* XXX should this be a different errno? */
3240 EXPORT_SYMBOL(target_bulk_io);
3242 #endif /* HAVE_SERVER_SUPPORT */