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 list_del(&imp_conn->oic_item);
180 ptlrpc_connection_put(imp_conn->oic_conn);
181 OBD_FREE(imp_conn, sizeof(*imp_conn));
182 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
183 imp, imp->imp_obd->obd_name, uuid->uuid);
188 spin_unlock(&imp->imp_lock);
190 CERROR("connection %s not found\n", uuid->uuid);
193 EXPORT_SYMBOL(client_import_del_conn);
196 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
197 * to find a conn uuid of \a imp which can reach \a peer.
199 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
200 struct obd_uuid *uuid)
202 struct obd_import_conn *conn;
206 spin_lock(&imp->imp_lock);
207 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
208 /* Check if conn UUID does have this peer NID. */
209 if (class_check_uuid(&conn->oic_uuid, peer)) {
210 *uuid = conn->oic_uuid;
215 spin_unlock(&imp->imp_lock);
218 EXPORT_SYMBOL(client_import_find_conn);
220 void client_destroy_import(struct obd_import *imp)
222 /* Drop security policy instance after all RPCs have finished/aborted
223 * to let all busy contexts be released. */
224 class_import_get(imp);
225 class_destroy_import(imp);
226 sptlrpc_import_sec_put(imp);
227 class_import_put(imp);
229 EXPORT_SYMBOL(client_destroy_import);
232 * Check whether or not the OSC is on MDT.
235 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
237 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
240 static int osc_on_mdt(char *obdname)
244 ptr = strrchr(obdname, '-');
248 if (strncmp(ptr + 1, "MDT", 3) == 0)
254 /* Configure an RPC client OBD device.
259 * 3 - inactive-on-startup
261 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
263 struct client_obd *cli = &obddev->u.cli;
264 struct obd_import *imp;
265 struct obd_uuid server_uuid;
266 int rq_portal, rp_portal, connect_op;
267 char *name = obddev->obd_type->typ_name;
268 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
269 char *cli_name = lustre_cfg_buf(lcfg, 0);
273 /* In a more perfect world, we would hang a ptlrpc_client off of
274 * obd_type and just use the values from there. */
275 if (!strcmp(name, LUSTRE_OSC_NAME)) {
276 rq_portal = OST_REQUEST_PORTAL;
277 rp_portal = OSC_REPLY_PORTAL;
278 connect_op = OST_CONNECT;
279 cli->cl_sp_me = LUSTRE_SP_CLI;
280 cli->cl_sp_to = LUSTRE_SP_OST;
281 ns_type = LDLM_NS_TYPE_OSC;
282 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
283 !strcmp(name, LUSTRE_LWP_NAME)) {
284 rq_portal = MDS_REQUEST_PORTAL;
285 rp_portal = MDC_REPLY_PORTAL;
286 connect_op = MDS_CONNECT;
287 if (is_lwp_on_ost(cli_name))
288 cli->cl_sp_me = LUSTRE_SP_OST;
289 else if (is_lwp_on_mdt(cli_name))
290 cli->cl_sp_me = LUSTRE_SP_MDT;
292 cli->cl_sp_me = LUSTRE_SP_CLI;
293 cli->cl_sp_to = LUSTRE_SP_MDT;
294 ns_type = LDLM_NS_TYPE_MDC;
295 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
296 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
297 /* OSP_on_MDT for other MDTs */
298 connect_op = MDS_CONNECT;
299 cli->cl_sp_to = LUSTRE_SP_MDT;
300 ns_type = LDLM_NS_TYPE_MDC;
301 rq_portal = OUT_PORTAL;
303 /* OSP on MDT for OST */
304 connect_op = OST_CONNECT;
305 cli->cl_sp_to = LUSTRE_SP_OST;
306 ns_type = LDLM_NS_TYPE_OSC;
307 rq_portal = OST_REQUEST_PORTAL;
309 rp_portal = OSC_REPLY_PORTAL;
310 cli->cl_sp_me = LUSTRE_SP_MDT;
311 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
312 rq_portal = MGS_REQUEST_PORTAL;
313 rp_portal = MGC_REPLY_PORTAL;
314 connect_op = MGS_CONNECT;
315 cli->cl_sp_me = LUSTRE_SP_MGC;
316 cli->cl_sp_to = LUSTRE_SP_MGS;
317 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
318 ns_type = LDLM_NS_TYPE_MGC;
320 CERROR("unknown client OBD type \"%s\", can't setup\n",
325 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
326 CERROR("requires a TARGET UUID\n");
330 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
331 CERROR("client UUID must be less than 38 characters\n");
335 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
336 CERROR("setup requires a SERVER UUID\n");
340 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
341 CERROR("target UUID must be less than 38 characters\n");
345 init_rwsem(&cli->cl_sem);
346 mutex_init(&cli->cl_mgc_mutex);
347 cli->cl_conn_count = 0;
348 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
349 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
350 sizeof(server_uuid)));
352 cli->cl_dirty_pages = 0;
353 cli->cl_avail_grant = 0;
354 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
355 /* cl_dirty_max_pages may be changed at connect time in
356 * ptlrpc_connect_interpret(). */
357 client_adjust_max_dirty(cli);
358 INIT_LIST_HEAD(&cli->cl_cache_waiters);
359 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
360 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
361 INIT_LIST_HEAD(&cli->cl_loi_write_list);
362 INIT_LIST_HEAD(&cli->cl_loi_read_list);
363 spin_lock_init(&cli->cl_loi_list_lock);
364 atomic_set(&cli->cl_pending_w_pages, 0);
365 atomic_set(&cli->cl_pending_r_pages, 0);
366 cli->cl_r_in_flight = 0;
367 cli->cl_w_in_flight = 0;
369 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
370 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
371 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
372 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
373 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
374 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
377 INIT_LIST_HEAD(&cli->cl_lru_osc);
378 atomic_set(&cli->cl_lru_shrinkers, 0);
379 atomic_long_set(&cli->cl_lru_busy, 0);
380 atomic_long_set(&cli->cl_lru_in_list, 0);
381 INIT_LIST_HEAD(&cli->cl_lru_list);
382 spin_lock_init(&cli->cl_lru_list_lock);
383 atomic_long_set(&cli->cl_unstable_count, 0);
384 INIT_LIST_HEAD(&cli->cl_shrink_list);
386 init_waitqueue_head(&cli->cl_destroy_waitq);
387 atomic_set(&cli->cl_destroy_in_flight, 0);
388 #ifdef ENABLE_CHECKSUM
389 /* Turn on checksumming by default. */
390 cli->cl_checksum = 1;
392 * The supported checksum types will be worked out at connect time
393 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
396 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
398 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
400 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
401 * from OFD after connecting. */
402 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
404 /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
405 * it will be updated at OSC connection time. */
406 cli->cl_chunkbits = PAGE_CACHE_SHIFT;
408 if (!strcmp(name, LUSTRE_MDC_NAME)) {
409 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
410 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
411 cli->cl_max_rpcs_in_flight = 2;
412 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
413 cli->cl_max_rpcs_in_flight = 3;
414 } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
415 cli->cl_max_rpcs_in_flight = 4;
417 if (osc_on_mdt(obddev->obd_name))
418 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
420 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
423 spin_lock_init(&cli->cl_mod_rpcs_lock);
424 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
425 cli->cl_max_mod_rpcs_in_flight = 0;
426 cli->cl_mod_rpcs_in_flight = 0;
427 cli->cl_close_rpcs_in_flight = 0;
428 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
429 cli->cl_mod_tag_bitmap = NULL;
431 if (connect_op == MDS_CONNECT) {
432 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
433 OBD_ALLOC(cli->cl_mod_tag_bitmap,
434 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
435 if (cli->cl_mod_tag_bitmap == NULL)
436 GOTO(err, rc = -ENOMEM);
441 CERROR("ldlm_get_ref failed: %d\n", rc);
445 ptlrpc_init_client(rq_portal, rp_portal, name,
446 &obddev->obd_ldlm_client);
448 imp = class_new_import(obddev);
450 GOTO(err_ldlm, rc = -ENOENT);
451 imp->imp_client = &obddev->obd_ldlm_client;
452 imp->imp_connect_op = connect_op;
453 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
454 LUSTRE_CFG_BUFLEN(lcfg, 1));
455 class_import_put(imp);
457 rc = client_import_add_conn(imp, &server_uuid, 1);
459 CERROR("can't add initial connection\n");
460 GOTO(err_import, rc);
463 cli->cl_import = imp;
464 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
465 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
467 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
468 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
469 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
470 name, obddev->obd_name,
471 cli->cl_target_uuid.uuid);
472 spin_lock(&imp->imp_lock);
473 imp->imp_deactive = 1;
474 spin_unlock(&imp->imp_lock);
478 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
479 LDLM_NAMESPACE_CLIENT,
480 LDLM_NAMESPACE_GREEDY,
482 if (obddev->obd_namespace == NULL) {
483 CERROR("Unable to create client namespace - %s\n",
485 GOTO(err_import, rc = -ENOMEM);
491 class_destroy_import(imp);
495 if (cli->cl_mod_tag_bitmap != NULL)
496 OBD_FREE(cli->cl_mod_tag_bitmap,
497 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
498 cli->cl_mod_tag_bitmap = NULL;
502 EXPORT_SYMBOL(client_obd_setup);
504 int client_obd_cleanup(struct obd_device *obddev)
506 struct client_obd *cli = &obddev->u.cli;
509 ldlm_namespace_free_post(obddev->obd_namespace);
510 obddev->obd_namespace = NULL;
512 obd_cleanup_client_import(obddev);
513 LASSERT(obddev->u.cli.cl_import == NULL);
517 if (cli->cl_mod_tag_bitmap != NULL)
518 OBD_FREE(cli->cl_mod_tag_bitmap,
519 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
520 cli->cl_mod_tag_bitmap = NULL;
524 EXPORT_SYMBOL(client_obd_cleanup);
526 /* ->o_connect() method for client side (OSC and MDC and MGC) */
527 int client_connect_import(const struct lu_env *env,
528 struct obd_export **exp,
529 struct obd_device *obd, struct obd_uuid *cluuid,
530 struct obd_connect_data *data, void *localdata)
532 struct client_obd *cli = &obd->u.cli;
533 struct obd_import *imp = cli->cl_import;
534 struct obd_connect_data *ocd;
535 struct lustre_handle conn = { 0 };
541 down_write(&cli->cl_sem);
542 if (cli->cl_conn_count > 0)
543 GOTO(out_sem, rc = -EALREADY);
545 rc = class_connect(&conn, obd, cluuid);
549 cli->cl_conn_count++;
550 *exp = class_conn2export(&conn);
552 LASSERT(obd->obd_namespace);
554 imp->imp_dlm_handle = conn;
555 rc = ptlrpc_init_import(imp);
559 ocd = &imp->imp_connect_data;
562 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
563 LUSTRE_MDC_NAME, 3) == 0;
565 data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
566 imp->imp_connect_flags_orig = data->ocd_connect_flags;
569 rc = ptlrpc_connect_import(imp);
571 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
574 LASSERT(*exp != NULL && (*exp)->exp_connection);
577 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
578 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
579 data->ocd_connect_flags, ocd->ocd_connect_flags);
580 data->ocd_connect_flags = ocd->ocd_connect_flags;
581 /* clear the flag as it was not set and is not known
584 data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
587 ptlrpc_pinger_add_import(imp);
593 cli->cl_conn_count--;
594 class_disconnect(*exp);
598 up_write(&cli->cl_sem);
602 EXPORT_SYMBOL(client_connect_import);
604 int client_disconnect_export(struct obd_export *exp)
606 struct obd_device *obd = class_exp2obd(exp);
607 struct client_obd *cli;
608 struct obd_import *imp;
613 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
614 exp, exp ? exp->exp_handle.h_cookie : -1);
619 imp = cli->cl_import;
621 down_write(&cli->cl_sem);
622 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
625 if (cli->cl_conn_count == 0) {
626 CERROR("disconnecting disconnected device (%s)\n",
628 GOTO(out_disconnect, rc = -EINVAL);
631 cli->cl_conn_count--;
632 if (cli->cl_conn_count != 0)
633 GOTO(out_disconnect, rc = 0);
635 /* Mark import deactivated now, so we don't try to reconnect if any
636 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
637 * fully deactivate the import, or that would drop all requests. */
638 spin_lock(&imp->imp_lock);
639 imp->imp_deactive = 1;
640 spin_unlock(&imp->imp_lock);
642 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
643 * delete it regardless. (It's safe to delete an import that was
645 (void)ptlrpc_pinger_del_import(imp);
647 if (obd->obd_namespace != NULL) {
648 /* obd_force == local only */
649 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
650 obd->obd_force ? LCF_LOCAL : 0, NULL);
651 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
654 /* There's no need to hold sem while disconnecting an import,
655 * and it may actually cause deadlock in GSS. */
656 up_write(&cli->cl_sem);
657 rc = ptlrpc_disconnect_import(imp, 0);
658 down_write(&cli->cl_sem);
660 ptlrpc_invalidate_import(imp);
665 /* Use server style - class_disconnect should be always called for
667 err = class_disconnect(exp);
671 up_write(&cli->cl_sem);
675 EXPORT_SYMBOL(client_disconnect_export);
677 #ifdef HAVE_SERVER_SUPPORT
678 int server_disconnect_export(struct obd_export *exp)
683 /* Disconnect early so that clients can't keep using export. */
684 rc = class_disconnect(exp);
685 /* Close import to avoid sending any requests. */
686 if (exp->exp_imp_reverse)
687 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
689 ldlm_bl_thread_wakeup();
691 /* complete all outstanding replies */
692 spin_lock(&exp->exp_lock);
693 while (!list_empty(&exp->exp_outstanding_replies)) {
694 struct ptlrpc_reply_state *rs =
695 list_entry(exp->exp_outstanding_replies.next,
696 struct ptlrpc_reply_state, rs_exp_list);
697 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
699 spin_lock(&svcpt->scp_rep_lock);
701 list_del_init(&rs->rs_exp_list);
702 spin_lock(&rs->rs_lock);
703 ptlrpc_schedule_difficult_reply(rs);
704 spin_unlock(&rs->rs_lock);
706 spin_unlock(&svcpt->scp_rep_lock);
708 spin_unlock(&exp->exp_lock);
712 EXPORT_SYMBOL(server_disconnect_export);
714 /* --------------------------------------------------------------------------
715 * from old lib/target.c
716 * -------------------------------------------------------------------------- */
718 static int target_handle_reconnect(struct lustre_handle *conn,
719 struct obd_export *exp,
720 struct obd_uuid *cluuid)
722 struct obd_device *target;
723 struct lustre_handle *hdl;
730 hdl = &exp->exp_imp_reverse->imp_remote_handle;
731 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
732 conn->cookie = exp->exp_handle.h_cookie;
733 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
734 " cookie "LPX64"\n", cluuid->uuid, exp, conn->cookie);
738 target = exp->exp_obd;
740 /* Might be a re-connect after a partition. */
741 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
742 LCONSOLE_WARN("%s: already connected client %s (at %s) "
743 "with handle "LPX64". Rejecting client "
744 "with the same UUID trying to reconnect "
745 "with handle "LPX64"\n", target->obd_name,
746 obd_uuid2str(&exp->exp_client_uuid),
747 obd_export_nid2str(exp),
748 hdl->cookie, conn->cookie);
749 memset(conn, 0, sizeof *conn);
750 /* target_handle_connect() treats EALREADY and
751 * -EALREADY differently. -EALREADY is an error
752 * (same UUID, different handle). */
756 if (!target->obd_recovering) {
757 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
758 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
759 obd_export_nid2str(exp));
760 GOTO(out_already, rc);
763 now = cfs_time_current();
764 deadline = cfs_timer_deadline(&target->obd_recovery_timer);
765 if (cfs_time_before(now, deadline)) {
766 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
767 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting,"
768 " waiting for %d clients in recovery for"
769 " %d:%.02d\n", target->obd_name,
770 obd_uuid2str(&exp->exp_client_uuid),
771 obd_export_nid2str(exp),
772 target->obd_max_recoverable_clients,
773 timeout / 60, timeout % 60);
775 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
776 LCONSOLE_WARN("%s: Recovery already passed deadline"
777 " %d:%.02d, It is most likely due to DNE"
778 " recovery is failed or stuck, please wait a"
779 " few more minutes or abort the recovery.\n",
780 target->obd_name, timeout / 60, timeout % 60);
784 conn->cookie = exp->exp_handle.h_cookie;
785 /* target_handle_connect() treats EALREADY and
786 * -EALREADY differently. EALREADY means we are
787 * doing a valid reconnect from the same client. */
791 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
794 struct obd_export *exp = cb_data;
796 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
797 obd->obd_name, exp->exp_client_uuid.uuid);
799 spin_lock(&exp->exp_lock);
800 exp->exp_need_sync = 0;
801 spin_unlock(&exp->exp_lock);
802 class_export_cb_put(exp);
806 check_and_start_recovery_timer(struct obd_device *obd,
807 struct ptlrpc_request *req, int new_client);
810 * update flags for import during reconnect process
812 static int rev_import_flags_update(struct obd_import *revimp,
813 struct ptlrpc_request *req)
816 struct obd_connect_data *data;
818 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
820 if (data->ocd_connect_flags & OBD_CONNECT_AT)
821 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
823 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
825 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
827 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
829 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
830 revimp->imp_client->cli_name,
831 libcfs_id2str(req->rq_peer), rc);
839 * Allocate a new reverse import for an export.
841 * \retval -errno in case error hit
842 * \retval 0 if reverse import correctly init
844 int rev_import_init(struct obd_export *export)
846 struct obd_device *obd = export->exp_obd;
847 struct obd_import *revimp;
849 LASSERT(export->exp_imp_reverse == NULL);
851 revimp = class_new_import(obd);
855 revimp->imp_remote_handle.cookie = 0ULL;
856 revimp->imp_client = &obd->obd_ldlm_client;
857 revimp->imp_dlm_fake = 1;
859 /* it is safe to connect import in new state as no sends possible */
860 spin_lock(&export->exp_lock);
861 export->exp_imp_reverse = revimp;
862 spin_unlock(&export->exp_lock);
863 class_import_put(revimp);
867 EXPORT_SYMBOL(rev_import_init);
870 * Handle reconnect for an export.
872 * \param exp export to handle reconnect process
873 * \param req client reconnect request
875 * \retval -rc in case securitfy flavor can't be changed
876 * \retval 0 in case none problems
878 static int rev_import_reconnect(struct obd_export *exp,
879 struct ptlrpc_request *req)
881 struct obd_import *revimp = exp->exp_imp_reverse;
882 struct lustre_handle *lh;
885 /* avoid sending a request until import flags are changed */
886 ptlrpc_import_enter_resend(revimp);
888 if (revimp->imp_connection != NULL)
889 ptlrpc_connection_put(revimp->imp_connection);
892 * client from recovery don't have a handle so we need to take from
893 * request. it may produce situation when wrong client connected
894 * to recovery as we trust a client uuid
896 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
897 revimp->imp_remote_handle = *lh;
899 /* unknown versions will be caught in
900 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
901 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
903 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
905 rc = rev_import_flags_update(revimp, req);
907 /* it is safe to still be in RECOVERY phase as we are not able
908 * to setup correct security flavor so requests are not able to
909 * be delivered correctly */
913 /* resend all rpc's via new connection */
914 return ptlrpc_import_recovery_state_machine(revimp);
917 int target_handle_connect(struct ptlrpc_request *req)
919 struct obd_device *target = NULL;
920 struct obd_export *export = NULL;
921 /* connect handle - filled from target_handle_reconnect in
923 struct lustre_handle conn;
924 struct lustre_handle *tmp;
925 struct obd_uuid tgtuuid;
926 struct obd_uuid cluuid;
931 bool mds_conn = false, lw_client = false;
932 bool mds_mds_conn = false;
933 bool new_mds_mds_conn = false;
934 bool target_referenced = false;
935 struct obd_connect_data *data, *tmpdata;
937 lnet_nid_t *client_nid = NULL;
940 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
942 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
944 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
945 GOTO(out, rc = -EINVAL);
948 obd_str2uuid(&tgtuuid, str);
949 target = class_uuid2obd(&tgtuuid);
951 target = class_name2obd(str);
954 deuuidify(str, NULL, &target_start, &target_len);
955 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
956 "from %s (no target). If you are running "
957 "an HA pair check that the target is "
958 "mounted on the other server.\n", str,
959 libcfs_nid2str(req->rq_peer.nid));
960 GOTO(out, rc = -ENODEV);
963 spin_lock(&target->obd_dev_lock);
964 if (target->obd_stopping || !target->obd_set_up) {
965 spin_unlock(&target->obd_dev_lock);
967 deuuidify(str, NULL, &target_start, &target_len);
968 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
969 target_len, target_start,
970 libcfs_nid2str(req->rq_peer.nid),
971 (target->obd_stopping ?
972 "stopping" : "not set up"));
973 GOTO(out, rc = -ENODEV);
976 if (target->obd_no_conn) {
977 spin_unlock(&target->obd_dev_lock);
979 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
980 "from %s\n", target->obd_name,
981 libcfs_nid2str(req->rq_peer.nid));
982 GOTO(out, rc = -EAGAIN);
985 /* Make sure the target isn't cleaned up while we're here. Yes,
986 * there's still a race between the above check and our incref here.
987 * Really, class_uuid2obd should take the ref. */
988 class_incref(target, __func__, current);
989 target_referenced = true;
991 target->obd_conn_inprogress++;
992 spin_unlock(&target->obd_dev_lock);
994 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
996 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
997 GOTO(out, rc = -EINVAL);
1000 obd_str2uuid(&cluuid, str);
1002 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1004 GOTO(out, rc = -EPROTO);
1008 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1010 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1012 GOTO(out, rc = -EPROTO);
1014 rc = req_capsule_server_pack(&req->rq_pill);
1018 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1019 /* Don't allow clients to connect that are using old 1.8 format
1020 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1021 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1022 * FULL20 flag should be set on all connections since 2.0, but no
1023 * longer affects behaviour.
1025 * Later this check will be disabled and the flag can be retired
1026 * completely once interop with 3.0 is no longer needed.
1028 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1029 GOTO(out, rc = -EPROTO);
1032 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1033 if (data->ocd_version < LUSTRE_VERSION_CODE -
1034 LUSTRE_VERSION_ALLOWED_OFFSET ||
1035 data->ocd_version > LUSTRE_VERSION_CODE +
1036 LUSTRE_VERSION_ALLOWED_OFFSET) {
1037 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1038 "libclient connection attempt",
1039 data->ocd_version < LUSTRE_VERSION_CODE ?
1041 OBD_OCD_VERSION_MAJOR(data->ocd_version),
1042 OBD_OCD_VERSION_MINOR(data->ocd_version),
1043 OBD_OCD_VERSION_PATCH(data->ocd_version),
1044 OBD_OCD_VERSION_FIX(data->ocd_version));
1045 data = req_capsule_server_sized_get(&req->rq_pill,
1047 offsetof(typeof(*data), ocd_version) +
1048 sizeof(data->ocd_version));
1050 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1051 data->ocd_version = LUSTRE_VERSION_CODE;
1053 GOTO(out, rc = -EPROTO);
1057 /* Note: lw_client is needed in MDS-MDS failover during update log
1058 * processing, so we needs to allow lw_client to be connected at
1059 * anytime, instead of only the initial connection */
1060 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1062 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1063 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1064 mds_mds_conn = (data->ocd_connect_flags &
1065 OBD_CONNECT_MDS_MDS) != 0;
1067 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1068 * for Imperative Recovery connection from MGC to MGS.
1070 * Via check OBD_CONNECT_FID, we can distinguish whether
1071 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1074 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1075 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1076 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1077 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1078 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1079 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1081 /* We do not support the MDT-MDT interoperations with
1082 * different version MDT because of protocol changes. */
1083 if (unlikely(major != LUSTRE_MAJOR ||
1084 minor != LUSTRE_MINOR ||
1085 abs(patch - LUSTRE_PATCH) > 3)) {
1086 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1087 "connection from different version MDT "
1088 "(%d.%d.%d.%d) %s %s\n",
1089 target->obd_name, LUSTRE_MAJOR,
1090 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1091 major, minor, patch,
1092 OBD_OCD_VERSION_FIX(data->ocd_version),
1093 libcfs_nid2str(req->rq_peer.nid), str);
1095 GOTO(out, rc = -EPROTO);
1100 /* lctl gets a backstage, all-access pass. */
1101 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1102 goto dont_check_exports;
1104 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1108 /* We've found an export in the hash. */
1110 spin_lock(&export->exp_lock);
1112 if (export->exp_connecting) { /* bug 9635, et. al. */
1113 spin_unlock(&export->exp_lock);
1114 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1115 export->exp_obd->obd_name, export,
1116 libcfs_nid2str(req->rq_peer.nid));
1117 class_export_put(export);
1120 } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
1121 spin_unlock(&export->exp_lock);
1122 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
1123 /* MDS or LWP reconnected after failover. */
1124 LCONSOLE_WARN("%s: Received %s connection from "
1125 "%s, removing former export from %s\n",
1126 target->obd_name, mds_conn ? "MDS" : "LWP",
1127 libcfs_nid2str(req->rq_peer.nid),
1128 libcfs_nid2str(export->exp_connection->c_peer.nid));
1130 /* New MDS connection from the same NID. */
1131 LCONSOLE_WARN("%s: Received new %s connection from "
1132 "%s, removing former export from same NID\n",
1133 target->obd_name, mds_conn ? "MDS" : "LWP",
1134 libcfs_nid2str(req->rq_peer.nid));
1135 class_fail_export(export);
1136 class_export_put(export);
1139 } else if (export->exp_connection != NULL &&
1140 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1141 (lustre_msg_get_op_flags(req->rq_reqmsg) &
1142 MSG_CONNECT_INITIAL)) {
1143 spin_unlock(&export->exp_lock);
1144 /* In MDS failover we have static UUID but NID can change. */
1145 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1146 "existing nid %s is already connected\n",
1147 target->obd_name, cluuid.uuid,
1148 libcfs_nid2str(req->rq_peer.nid),
1150 export->exp_connection->c_peer.nid));
1152 class_export_put(export);
1155 export->exp_connecting = 1;
1156 spin_unlock(&export->exp_lock);
1157 LASSERT(export->exp_obd == target);
1159 rc = target_handle_reconnect(&conn, export, &cluuid);
1162 /* If we found an export, we already unlocked. */
1165 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1166 } else if (req->rq_export == NULL &&
1167 atomic_read(&export->exp_rpc_count) > 0) {
1168 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1169 "still busy with %d references\n",
1170 target->obd_name, cluuid.uuid,
1171 libcfs_nid2str(req->rq_peer.nid),
1172 atomic_read(&export->exp_refcount));
1173 GOTO(out, rc = -EBUSY);
1174 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1175 if (!strstr(cluuid.uuid, "mdt"))
1176 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1177 "known client %s (at %s) because it "
1178 "is indicating it is a new client",
1179 target->obd_name, cluuid.uuid,
1180 libcfs_nid2str(req->rq_peer.nid));
1181 GOTO(out, rc = -EALREADY);
1183 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1190 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1191 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1192 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1193 export, (long)cfs_time_current_sec(),
1194 export ? (long)export->exp_last_request_time : 0);
1196 /* If this is the first time a client connects, reset the recovery
1197 * timer. Discard lightweight connections which might be local. */
1198 if (!lw_client && rc == 0 && target->obd_recovering)
1199 check_and_start_recovery_timer(target, req, export == NULL);
1201 /* We want to handle EALREADY but *not* -EALREADY from
1202 * target_handle_reconnect(), return reconnection state in a flag. */
1203 if (rc == EALREADY) {
1204 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1210 /* Tell the client if we support replayable requests. */
1211 if (target->obd_replayable)
1212 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1213 client_nid = &req->rq_peer.nid;
1215 if (export == NULL) {
1216 /* allow lightweight connections during recovery */
1217 /* allow "new" MDT to be connected during recovery, since we
1218 * need retrieve recovery update records from it */
1219 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1221 int c; /* connected */
1222 int i; /* in progress */
1224 int s; /* stale/evicted */
1226 c = atomic_read(&target->obd_connected_clients);
1227 i = atomic_read(&target->obd_lock_replay_clients);
1228 k = target->obd_max_recoverable_clients;
1229 s = target->obd_stale_clients;
1230 t = cfs_timer_deadline(&target->obd_recovery_timer);
1231 t = cfs_time_sub(t, cfs_time_current());
1232 t = cfs_duration_sec(t);
1233 LCONSOLE_WARN("%s: Denying connection for new client %s"
1234 "(at %s), waiting for %d known clients "
1235 "(%d recovered, %d in progress, and %d "
1236 "evicted) to recover in %d:%.02d\n",
1237 target->obd_name, cluuid.uuid,
1238 libcfs_nid2str(req->rq_peer.nid), k,
1239 c - i, i, s, (int)t / 60,
1244 rc = obd_connect(req->rq_svc_thread->t_env,
1245 &export, target, &cluuid, data,
1247 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1248 lustre_msg_add_op_flags(req->rq_repmsg,
1249 MSG_CONNECT_RECOVERING);
1251 conn.cookie = export->exp_handle.h_cookie;
1252 rc = rev_import_init(export);
1256 new_mds_mds_conn = true;
1259 rc = obd_reconnect(req->rq_svc_thread->t_env,
1260 export, target, &cluuid, data, client_nid);
1265 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1266 data->ocd_instance = target->u.obt.obt_instance;
1268 /* Return only the parts of obd_connect_data that we understand, so the
1269 * client knows that we don't understand the rest. */
1271 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1273 tmpdata = req_capsule_server_get(&req->rq_pill,
1275 /* Don't use struct assignment here, because the client reply
1276 * buffer may be smaller/larger than the local struct
1277 * obd_connect_data. */
1278 memcpy(tmpdata, data, min(tmpsize, size));
1281 /* If the client and the server are the same node, we will already
1282 * have an export that really points to the client's DLM export,
1283 * because we have a shared handles table.
1285 * XXX this will go away when shaver stops sending the "connect" handle
1286 * in the real "remote handle" field of the request --phik 24 Apr 2003
1288 ptlrpc_request_change_export(req, export);
1290 spin_lock(&export->exp_lock);
1291 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1292 spin_unlock(&export->exp_lock);
1293 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1294 "or equal conn_cnt: %d >= %d\n",
1295 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1296 export->exp_conn_cnt,
1297 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1299 GOTO(out, rc = -EALREADY);
1301 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1302 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1304 /* Don't evict liblustre clients for not pinging. */
1305 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1306 export->exp_libclient = 1;
1307 spin_unlock(&export->exp_lock);
1309 spin_lock(&target->obd_dev_lock);
1310 list_del_init(&export->exp_obd_chain_timed);
1311 spin_unlock(&target->obd_dev_lock);
1313 spin_unlock(&export->exp_lock);
1316 if (export->exp_connection != NULL) {
1317 /* Check to see if connection came from another NID. */
1318 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1319 !hlist_unhashed(&export->exp_nid_hash))
1320 cfs_hash_del(export->exp_obd->obd_nid_hash,
1321 &export->exp_connection->c_peer.nid,
1322 &export->exp_nid_hash);
1324 ptlrpc_connection_put(export->exp_connection);
1327 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1330 if (hlist_unhashed(&export->exp_nid_hash))
1331 cfs_hash_add(export->exp_obd->obd_nid_hash,
1332 &export->exp_connection->c_peer.nid,
1333 &export->exp_nid_hash);
1335 lustre_msg_set_handle(req->rq_repmsg, &conn);
1337 rc = rev_import_reconnect(export, req);
1341 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1343 __u64 transno = data->ocd_transno;
1345 spin_lock(&export->exp_lock);
1346 /* possible race with class_disconnect_stale_exports,
1347 * export may be already in the eviction process */
1348 if (export->exp_failed) {
1349 spin_unlock(&export->exp_lock);
1350 GOTO(out, rc = -ENODEV);
1352 export->exp_in_recovery = 1;
1353 export->exp_req_replay_needed = 1;
1354 export->exp_lock_replay_needed = 1;
1355 spin_unlock(&export->exp_lock);
1357 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1358 MSG_CONNECT_TRANSNO);
1359 if (has_transno && transno == 0)
1360 CWARN("Connect with zero transno!\n");
1362 if (has_transno && transno > 0 &&
1363 transno < target->obd_next_recovery_transno &&
1364 transno > target->obd_last_committed) {
1365 /* Another way is to use cmpxchg() to be lock-free. */
1366 spin_lock(&target->obd_recovery_task_lock);
1367 if (transno < target->obd_next_recovery_transno)
1368 target->obd_next_recovery_transno = transno;
1369 spin_unlock(&target->obd_recovery_task_lock);
1372 atomic_inc(&target->obd_req_replay_clients);
1373 atomic_inc(&target->obd_lock_replay_clients);
1374 /* Note: MDS-MDS connection is allowed to be connected during
1375 * recovery, no matter if the exports needs to be recoveried.
1376 * Because we need retrieve updates logs from all other MDTs.
1377 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1378 * also needs to be increased to match other recovery checking
1380 if (new_mds_mds_conn)
1381 target->obd_max_recoverable_clients++;
1382 if (atomic_inc_return(&target->obd_connected_clients) ==
1383 target->obd_max_recoverable_clients)
1384 wake_up(&target->obd_next_transno_waitq);
1387 /* Tell the client we're in recovery, when client is involved in it. */
1388 if (target->obd_recovering && !lw_client)
1389 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1393 spin_lock(&export->exp_lock);
1394 export->exp_connecting = 0;
1395 spin_unlock(&export->exp_lock);
1397 class_export_put(export);
1399 if (target_referenced == true && target != NULL) {
1400 spin_lock(&target->obd_dev_lock);
1401 target->obd_conn_inprogress--;
1402 spin_unlock(&target->obd_dev_lock);
1404 class_decref(target, __func__, current);
1406 req->rq_status = rc;
1410 int target_handle_disconnect(struct ptlrpc_request *req)
1415 rc = req_capsule_server_pack(&req->rq_pill);
1419 /* Keep the rq_export around so we can send the reply. */
1420 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1425 void target_destroy_export(struct obd_export *exp)
1427 struct obd_import *imp = NULL;
1428 /* exports created from last_rcvd data, and "fake"
1429 exports created by lctl don't have an import */
1430 spin_lock(&exp->exp_lock);
1431 if (exp->exp_imp_reverse != NULL) {
1432 imp = exp->exp_imp_reverse;
1433 exp->exp_imp_reverse = NULL;
1435 spin_unlock(&exp->exp_lock);
1437 client_destroy_import(imp);
1439 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1440 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1441 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1442 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1444 EXPORT_SYMBOL(target_destroy_export);
1447 * Recovery functions
1449 static void target_request_copy_get(struct ptlrpc_request *req)
1451 class_export_rpc_inc(req->rq_export);
1452 LASSERT(list_empty(&req->rq_list));
1453 INIT_LIST_HEAD(&req->rq_replay_list);
1455 /* Increase refcount to keep request in queue. */
1456 atomic_inc(&req->rq_refcount);
1457 /* Let export know it has replays to be handled. */
1458 atomic_inc(&req->rq_export->exp_replay_count);
1461 static void target_request_copy_put(struct ptlrpc_request *req)
1463 LASSERT(list_empty(&req->rq_replay_list));
1464 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1466 atomic_dec(&req->rq_export->exp_replay_count);
1467 class_export_rpc_dec(req->rq_export);
1468 ptlrpc_server_drop_request(req);
1471 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1473 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1474 struct obd_export *exp = req->rq_export;
1475 struct ptlrpc_request *reqiter;
1476 struct ptlrpc_request *dup_req = NULL;
1481 spin_lock(&exp->exp_lock);
1482 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1484 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1492 /* We expect it with RESENT and REPLAY flags. */
1493 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1494 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1495 CERROR("invalid flags %x of resent replay\n",
1496 lustre_msg_get_flags(req->rq_reqmsg));
1498 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1501 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1503 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1504 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1508 list_add_tail(&req->rq_replay_list,
1509 &exp->exp_req_replay_queue);
1512 spin_unlock(&exp->exp_lock);
1516 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1518 LASSERT(!list_empty(&req->rq_replay_list));
1519 LASSERT(req->rq_export);
1521 spin_lock(&req->rq_export->exp_lock);
1522 list_del_init(&req->rq_replay_list);
1523 spin_unlock(&req->rq_export->exp_lock);
1526 static void target_finish_recovery(struct lu_target *lut)
1528 struct obd_device *obd = lut->lut_obd;
1531 /* Only log a recovery message when recovery has occurred. */
1532 if (obd->obd_recovery_start) {
1533 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1534 obd->obd_recovery_start);
1535 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1536 "%d recovered and %d %s evicted.\n", obd->obd_name,
1537 (int)elapsed_time / 60, (int)elapsed_time % 60,
1538 obd->obd_max_recoverable_clients,
1539 atomic_read(&obd->obd_connected_clients),
1540 obd->obd_stale_clients,
1541 obd->obd_stale_clients == 1 ? "was" : "were");
1544 ldlm_reprocess_all_ns(obd->obd_namespace);
1545 spin_lock(&obd->obd_recovery_task_lock);
1546 if (!list_empty(&obd->obd_req_replay_queue) ||
1547 !list_empty(&obd->obd_lock_replay_queue) ||
1548 !list_empty(&obd->obd_final_req_queue)) {
1549 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1551 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1552 list_empty(&obd->obd_lock_replay_queue) ? \
1554 list_empty(&obd->obd_final_req_queue) ? \
1556 spin_unlock(&obd->obd_recovery_task_lock);
1559 spin_unlock(&obd->obd_recovery_task_lock);
1561 obd->obd_recovery_end = cfs_time_current_sec();
1563 /* When recovery finished, cleanup orphans on MDS and OST. */
1564 if (OBT(obd) && OBP(obd, postrecov)) {
1565 int rc = OBP(obd, postrecov)(obd);
1567 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1573 static void abort_req_replay_queue(struct obd_device *obd)
1575 struct ptlrpc_request *req, *n;
1576 struct list_head abort_list;
1578 INIT_LIST_HEAD(&abort_list);
1579 spin_lock(&obd->obd_recovery_task_lock);
1580 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1581 spin_unlock(&obd->obd_recovery_task_lock);
1582 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1583 DEBUG_REQ(D_WARNING, req, "aborted:");
1584 req->rq_status = -ENOTCONN;
1585 if (ptlrpc_error(req)) {
1586 DEBUG_REQ(D_ERROR, req,
1587 "failed abort_req_reply; skipping");
1589 target_exp_dequeue_req_replay(req);
1590 target_request_copy_put(req);
1594 static void abort_lock_replay_queue(struct obd_device *obd)
1596 struct ptlrpc_request *req, *n;
1597 struct list_head abort_list;
1599 INIT_LIST_HEAD(&abort_list);
1600 spin_lock(&obd->obd_recovery_task_lock);
1601 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1602 spin_unlock(&obd->obd_recovery_task_lock);
1603 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1604 DEBUG_REQ(D_ERROR, req, "aborted:");
1605 req->rq_status = -ENOTCONN;
1606 if (ptlrpc_error(req)) {
1607 DEBUG_REQ(D_ERROR, req,
1608 "failed abort_lock_reply; skipping");
1610 target_request_copy_put(req);
1614 /* Called from a cleanup function if the device is being cleaned up
1615 forcefully. The exports should all have been disconnected already,
1616 the only thing left to do is
1617 - clear the recovery flags
1619 - free queued requests and replies, but don't send replies
1620 Because the obd_stopping flag is set, no new requests should be received.
1623 void target_cleanup_recovery(struct obd_device *obd)
1625 struct ptlrpc_request *req, *n;
1626 struct list_head clean_list;
1628 INIT_LIST_HEAD(&clean_list);
1629 spin_lock(&obd->obd_dev_lock);
1630 if (!obd->obd_recovering) {
1631 spin_unlock(&obd->obd_dev_lock);
1635 obd->obd_recovering = obd->obd_abort_recovery = 0;
1636 spin_unlock(&obd->obd_dev_lock);
1638 spin_lock(&obd->obd_recovery_task_lock);
1639 target_cancel_recovery_timer(obd);
1640 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1641 spin_unlock(&obd->obd_recovery_task_lock);
1643 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1644 LASSERT(req->rq_reply_state == NULL);
1645 target_exp_dequeue_req_replay(req);
1646 target_request_copy_put(req);
1649 spin_lock(&obd->obd_recovery_task_lock);
1650 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1651 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1652 spin_unlock(&obd->obd_recovery_task_lock);
1654 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1655 LASSERT(req->rq_reply_state == NULL);
1656 target_request_copy_put(req);
1661 EXPORT_SYMBOL(target_cleanup_recovery);
1663 /* obd_recovery_task_lock should be held */
1664 void target_cancel_recovery_timer(struct obd_device *obd)
1666 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1667 cfs_timer_disarm(&obd->obd_recovery_timer);
1670 static void target_start_recovery_timer(struct obd_device *obd)
1672 if (obd->obd_recovery_start != 0)
1675 spin_lock(&obd->obd_dev_lock);
1676 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1677 spin_unlock(&obd->obd_dev_lock);
1681 LASSERT(obd->obd_recovery_timeout != 0);
1683 if (obd->obd_recovery_start != 0) {
1684 spin_unlock(&obd->obd_dev_lock);
1688 cfs_timer_arm(&obd->obd_recovery_timer,
1689 cfs_time_shift(obd->obd_recovery_timeout));
1690 obd->obd_recovery_start = cfs_time_current_sec();
1691 spin_unlock(&obd->obd_dev_lock);
1693 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1694 "or until %d client%s reconnect%s\n",
1696 obd->obd_recovery_timeout / 60,
1697 obd->obd_recovery_timeout % 60,
1698 obd->obd_max_recoverable_clients,
1699 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1700 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1704 * extend recovery window.
1706 * if @extend is true, extend recovery window to have @drt remaining at least;
1707 * otherwise, make sure the recovery timeout value is not less than @drt.
1709 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1713 cfs_duration_t left;
1716 spin_lock(&obd->obd_dev_lock);
1717 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1718 spin_unlock(&obd->obd_dev_lock);
1721 LASSERT(obd->obd_recovery_start != 0);
1723 now = cfs_time_current_sec();
1724 to = obd->obd_recovery_timeout;
1725 end = obd->obd_recovery_start + to;
1726 left = cfs_time_sub(end, now);
1728 if (extend && (drt > left)) {
1730 } else if (!extend && (drt > to)) {
1734 if (to > obd->obd_recovery_time_hard)
1735 to = obd->obd_recovery_time_hard;
1736 if (obd->obd_recovery_timeout < to) {
1737 obd->obd_recovery_timeout = to;
1738 end = obd->obd_recovery_start + to;
1739 cfs_timer_arm(&obd->obd_recovery_timer,
1740 cfs_time_shift(end - now));
1742 spin_unlock(&obd->obd_dev_lock);
1744 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1745 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1748 /* Reset the timer with each new client connection */
1750 * This timer is actually reconnect_timer, which is for making sure
1751 * the total recovery window is at least as big as my reconnect
1752 * attempt timing. So the initial recovery time_out will be set to
1753 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1754 * from client is bigger than this, then the recovery time_out will
1755 * be extended to make sure the client could be reconnected, in the
1756 * process, the timeout from the new client should be ignored.
1760 check_and_start_recovery_timer(struct obd_device *obd,
1761 struct ptlrpc_request *req,
1764 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1765 struct obd_device_target *obt = &obd->u.obt;
1767 if (!new_client && service_time)
1768 /* Teach server about old server's estimates, as first guess
1769 * at how long new requests will take. */
1770 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1773 target_start_recovery_timer(obd);
1775 /* Convert the service time to RPC timeout,
1776 * and reuse service_time to limit stack usage. */
1777 service_time = at_est2timeout(service_time);
1779 /* We expect other clients to timeout within service_time, then try
1780 * to reconnect, then try the failover server. The max delay between
1781 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1782 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1784 LASSERT(obt->obt_magic == OBT_MAGIC);
1785 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1786 if (service_time > obd->obd_recovery_timeout && !new_client)
1787 extend_recovery_timer(obd, service_time, false);
1790 /** Health checking routines */
1791 static inline int exp_connect_healthy(struct obd_export *exp)
1793 return (exp->exp_in_recovery);
1796 /** if export done req_replay or has replay in queue */
1797 static inline int exp_req_replay_healthy(struct obd_export *exp)
1799 return (!exp->exp_req_replay_needed ||
1800 atomic_read(&exp->exp_replay_count) > 0);
1804 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1806 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1807 exp_req_replay_healthy(exp);
1810 /** if export done lock_replay or has replay in queue */
1811 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1813 return (!exp->exp_lock_replay_needed ||
1814 atomic_read(&exp->exp_replay_count) > 0);
1817 static inline int exp_vbr_healthy(struct obd_export *exp)
1819 return (!exp->exp_vbr_failed);
1822 static inline int exp_finished(struct obd_export *exp)
1824 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1827 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1829 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1833 static int check_for_next_transno(struct lu_target *lut)
1835 struct ptlrpc_request *req = NULL;
1836 struct obd_device *obd = lut->lut_obd;
1837 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1838 int wake_up = 0, connected, completed, queue_len;
1839 __u64 req_transno = 0;
1840 __u64 update_transno = 0;
1841 __u64 next_transno = 0;
1844 spin_lock(&obd->obd_recovery_task_lock);
1845 if (!list_empty(&obd->obd_req_replay_queue)) {
1846 req = list_entry(obd->obd_req_replay_queue.next,
1847 struct ptlrpc_request, rq_list);
1848 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1852 update_transno = distribute_txn_get_next_transno(tdtd);
1854 connected = atomic_read(&obd->obd_connected_clients);
1855 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1856 queue_len = obd->obd_requests_queued_for_recovery;
1857 next_transno = obd->obd_next_recovery_transno;
1859 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1860 "req_transno: "LPU64", next_transno: "LPU64"\n",
1861 obd->obd_max_recoverable_clients, connected, completed,
1862 queue_len, req_transno, next_transno);
1864 if (obd->obd_abort_recovery) {
1865 CDEBUG(D_HA, "waking for aborted recovery\n");
1867 } else if (obd->obd_recovery_expired) {
1868 CDEBUG(D_HA, "waking for expired recovery\n");
1870 } else if (tdtd != NULL && req != NULL &&
1871 is_req_replayed_by_update(req)) {
1872 LASSERTF(req_transno < next_transno, "req_transno "LPU64
1873 "next_transno"LPU64"\n", req_transno, next_transno);
1874 CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n",
1877 } else if (req_transno == next_transno ||
1878 (update_transno != 0 && update_transno <= next_transno)) {
1879 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1881 } else if (queue_len > 0 &&
1882 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1884 /** handle gaps occured due to lost reply or VBR */
1885 LASSERTF(req_transno >= next_transno,
1886 "req_transno: "LPU64", next_transno: "LPU64"\n",
1887 req_transno, next_transno);
1888 if (req_transno > obd->obd_last_committed &&
1889 !obd->obd_version_recov)
1892 "%s: waking for gap in transno, VBR is %s (skip: "
1893 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1894 ", next_update "LPD64" last_committed: "LPD64")\n",
1895 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1896 next_transno, queue_len, completed, connected,
1897 req_transno, update_transno, obd->obd_last_committed);
1898 obd->obd_next_recovery_transno = req_transno;
1900 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1901 CDEBUG(D_HA, "waking for completed recovery\n");
1903 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1904 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1905 " by fail_lock, waking up ("LPD64")\n", next_transno);
1906 obd->obd_next_recovery_transno = req_transno;
1909 spin_unlock(&obd->obd_recovery_task_lock);
1913 static int check_for_next_lock(struct lu_target *lut)
1915 struct obd_device *obd = lut->lut_obd;
1918 spin_lock(&obd->obd_recovery_task_lock);
1919 if (!list_empty(&obd->obd_lock_replay_queue)) {
1920 CDEBUG(D_HA, "waking for next lock\n");
1922 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1923 CDEBUG(D_HA, "waking for completed lock replay\n");
1925 } else if (obd->obd_abort_recovery) {
1926 CDEBUG(D_HA, "waking for aborted recovery\n");
1928 } else if (obd->obd_recovery_expired) {
1929 CDEBUG(D_HA, "waking for expired recovery\n");
1932 spin_unlock(&obd->obd_recovery_task_lock);
1938 * wait for recovery events,
1939 * check its status with help of check_routine
1940 * evict dead clients via health_check
1942 static int target_recovery_overseer(struct lu_target *lut,
1943 int (*check_routine)(struct lu_target *),
1944 int (*health_check)(struct obd_export *))
1946 struct obd_device *obd = lut->lut_obd;
1947 struct target_distribute_txn_data *tdtd;
1949 if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1950 (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1951 __u64 next_update_transno = 0;
1953 /* Only abort the recovery if there are no update recovery
1954 * left in the queue */
1955 spin_lock(&obd->obd_recovery_task_lock);
1956 if (lut->lut_tdtd != NULL) {
1957 next_update_transno =
1958 distribute_txn_get_next_transno(lut->lut_tdtd);
1960 tdtd = lut->lut_tdtd;
1961 /* If next_update_transno == 0, it probably because
1962 * updatelog retrieve threads did not get any records
1963 * yet, let's wait those threads stopped */
1964 if (next_update_transno == 0) {
1965 struct l_wait_info lwi = { 0 };
1967 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1969 &tdtd->tdtd_recovery_threads_count) == 0,
1972 next_update_transno =
1973 distribute_txn_get_next_transno(
1978 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
1979 obd->obd_next_recovery_transno = next_update_transno;
1980 spin_unlock(&obd->obd_recovery_task_lock);
1981 /* Disconnect unfinished exports from clients, and
1982 * keep connection from MDT to make sure the update
1983 * recovery will still keep trying until some one
1984 * manually abort the recovery */
1985 class_disconnect_stale_exports(obd,
1986 exp_finished_or_from_mdt);
1987 /* Abort all of replay and replay lock req from
1989 abort_req_replay_queue(obd);
1990 abort_lock_replay_queue(obd);
1991 CDEBUG(D_HA, "%s: there are still update replay ("LPX64
1992 ")in the queue.\n", obd->obd_name,
1993 next_update_transno);
1995 obd->obd_abort_recovery = 1;
1996 spin_unlock(&obd->obd_recovery_task_lock);
1997 CWARN("%s recovery is aborted by hard timeout\n",
2002 while (wait_event_timeout(obd->obd_next_transno_waitq,
2004 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2005 /* wait indefinitely for event, but don't trigger watchdog */;
2007 if (obd->obd_abort_recovery) {
2008 CWARN("recovery is aborted, evict exports in recovery\n");
2009 if (lut->lut_tdtd != NULL) {
2010 struct l_wait_info lwi = { 0 };
2012 tdtd = lut->lut_tdtd;
2013 /* Let's wait all of the update log recovery thread
2015 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2016 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2018 /* Then abort the update recovery list */
2019 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2020 dtrq_list_destroy(lut->lut_tdtd);
2023 /** evict exports which didn't finish recovery yet */
2024 class_disconnect_stale_exports(obd, exp_finished);
2026 } else if (obd->obd_recovery_expired) {
2027 obd->obd_recovery_expired = 0;
2028 /** If some clients died being recovered, evict them */
2029 LCONSOLE_WARN("%s: recovery is timed out, "
2030 "evict stale exports\n", obd->obd_name);
2031 /** evict cexports with no replay in queue, they are stalled */
2032 class_disconnect_stale_exports(obd, health_check);
2034 /** continue with VBR */
2035 spin_lock(&obd->obd_dev_lock);
2036 obd->obd_version_recov = 1;
2037 spin_unlock(&obd->obd_dev_lock);
2039 * reset timer, recovery will proceed with versions now,
2040 * timeout is set just to handle reconnection delays
2042 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2043 /** Wait for recovery events again, after evicting bad clients */
2049 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2051 struct obd_device *obd = lut->lut_obd;
2052 struct ptlrpc_request *req = NULL;
2054 CDEBUG(D_HA, "Waiting for lock\n");
2055 if (target_recovery_overseer(lut, check_for_next_lock,
2056 exp_lock_replay_healthy))
2057 abort_lock_replay_queue(obd);
2059 spin_lock(&obd->obd_recovery_task_lock);
2060 if (!list_empty(&obd->obd_lock_replay_queue)) {
2061 req = list_entry(obd->obd_lock_replay_queue.next,
2062 struct ptlrpc_request, rq_list);
2063 list_del_init(&req->rq_list);
2064 spin_unlock(&obd->obd_recovery_task_lock);
2066 spin_unlock(&obd->obd_recovery_task_lock);
2067 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2068 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2069 /** evict exports failed VBR */
2070 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2075 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2077 struct ptlrpc_request *req = NULL;
2079 spin_lock(&obd->obd_recovery_task_lock);
2080 if (!list_empty(&obd->obd_final_req_queue)) {
2081 req = list_entry(obd->obd_final_req_queue.next,
2082 struct ptlrpc_request, rq_list);
2083 list_del_init(&req->rq_list);
2084 spin_unlock(&obd->obd_recovery_task_lock);
2085 if (req->rq_export->exp_in_recovery) {
2086 spin_lock(&req->rq_export->exp_lock);
2087 req->rq_export->exp_in_recovery = 0;
2088 spin_unlock(&req->rq_export->exp_lock);
2091 spin_unlock(&obd->obd_recovery_task_lock);
2096 static void handle_recovery_req(struct ptlrpc_thread *thread,
2097 struct ptlrpc_request *req,
2098 svc_handler_t handler)
2103 * export can be evicted during recovery, no need to handle replays for
2104 * it after that, discard such request silently
2106 if (req->rq_export->exp_disconnected)
2109 req->rq_session.lc_thread = thread;
2110 req->rq_svc_thread = thread;
2111 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2113 /* thread context */
2114 lu_context_enter(&thread->t_env->le_ctx);
2116 lu_context_exit(&thread->t_env->le_ctx);
2118 /* don't reset timer for final stage */
2119 if (!exp_finished(req->rq_export)) {
2120 int to = obd_timeout;
2123 * Add request timeout to the recovery time so next request from
2124 * this client may come in recovery time
2127 struct ptlrpc_service_part *svcpt;
2129 svcpt = req->rq_rqbd->rqbd_svcpt;
2130 /* If the server sent early reply for this request,
2131 * the client will recalculate the timeout according to
2132 * current server estimate service time, so we will
2133 * use the maxium timeout here for waiting the client
2134 * sending the next req */
2135 to = max((int)at_est2timeout(
2136 at_get(&svcpt->scp_at_estimate)),
2137 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2138 /* Add 2 net_latency, one for balance rq_deadline
2139 * (see ptl_send_rpc), one for resend the req to server,
2140 * Note: client will pack net_latency in replay req
2141 * (see ptlrpc_replay_req) */
2142 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2144 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2149 /** Checking routines for recovery */
2150 static int check_for_recovery_ready(struct lu_target *lut)
2152 struct obd_device *obd = lut->lut_obd;
2153 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2155 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2156 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2157 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2158 obd->obd_recovery_expired);
2160 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2161 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2162 if (clnts + obd->obd_stale_clients <
2163 obd->obd_max_recoverable_clients)
2167 if (lut->lut_tdtd != NULL) {
2168 if (!lut->lut_tdtd->tdtd_replay_ready &&
2169 !obd->obd_abort_recovery) {
2170 /* Let's extend recovery timer, in case the recovery
2171 * timer expired, and some clients got evicted */
2172 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2174 CDEBUG(D_HA, "%s update recovery is not ready,"
2175 " extend recovery %d\n", obd->obd_name,
2176 obd->obd_recovery_timeout);
2179 dtrq_list_dump(lut->lut_tdtd, D_HA);
2187 REQUEST_RECOVERY = 1,
2188 UPDATE_RECOVERY = 2,
2191 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2195 if (!list_empty(&obd->obd_req_replay_queue)) {
2196 struct ptlrpc_request *req;
2198 req = list_entry(obd->obd_req_replay_queue.next,
2199 struct ptlrpc_request, rq_list);
2200 transno = lustre_msg_get_transno(req->rq_reqmsg);
2206 static __u64 get_next_transno(struct lu_target *lut, int *type)
2208 struct obd_device *obd = lut->lut_obd;
2209 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2211 __u64 update_transno;
2214 transno = get_next_replay_req_transno(obd);
2216 *type = REQUEST_RECOVERY;
2221 update_transno = distribute_txn_get_next_transno(tdtd);
2222 if (transno == 0 || (transno >= update_transno &&
2223 update_transno != 0)) {
2224 transno = update_transno;
2226 *type = UPDATE_RECOVERY;
2233 * drop duplicate replay request
2235 * Because the operation has been replayed by update recovery, the request
2236 * with the same transno will be dropped and also notify the client to send
2237 * next replay request.
2239 * \param[in] env execution environment
2240 * \param[in] obd failover obd device
2241 * \param[in] req request to be dropped
2243 static void drop_duplicate_replay_req(struct lu_env *env,
2244 struct obd_device *obd,
2245 struct ptlrpc_request *req)
2247 DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2248 " update records are found.\n",
2249 lustre_msg_get_transno(req->rq_reqmsg),
2250 libcfs_nid2str(req->rq_peer.nid));
2252 /* Right now, only for MDS reint operation update replay and
2253 * normal request replay can have the same transno */
2254 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2255 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2256 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2257 if (likely(req->rq_export))
2258 target_committed_to_req(req);
2259 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2260 target_send_reply(req, req->rq_status, 0);
2262 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2263 libcfs_nid2str(req->rq_peer.nid));
2265 target_exp_dequeue_req_replay(req);
2266 target_request_copy_put(req);
2267 obd->obd_replayed_requests++;
2270 static void replay_request_or_update(struct lu_env *env,
2271 struct lu_target *lut,
2272 struct target_recovery_data *trd,
2273 struct ptlrpc_thread *thread)
2275 struct obd_device *obd = lut->lut_obd;
2276 struct ptlrpc_request *req = NULL;
2281 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2282 obd->obd_next_recovery_transno);
2284 /* Replay all of request and update by transno */
2286 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2288 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2290 /** It is needed to extend recovery window above
2291 * recovery_time_soft. Extending is possible only in the
2292 * end of recovery window (see more details in
2293 * handle_recovery_req()).
2295 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2297 if (target_recovery_overseer(lut, check_for_next_transno,
2298 exp_req_replay_healthy_or_from_mdt)) {
2299 abort_req_replay_queue(obd);
2300 abort_lock_replay_queue(obd);
2304 spin_lock(&obd->obd_recovery_task_lock);
2305 transno = get_next_transno(lut, &type);
2306 if (type == REQUEST_RECOVERY && transno != 0) {
2307 /* Drop replay request from client side, if the
2308 * replay has been executed by update with the
2310 req = list_entry(obd->obd_req_replay_queue.next,
2311 struct ptlrpc_request, rq_list);
2313 list_del_init(&req->rq_list);
2314 obd->obd_requests_queued_for_recovery--;
2315 spin_unlock(&obd->obd_recovery_task_lock);
2317 /* Let's check if the request has been redone by
2319 if (is_req_replayed_by_update(req)) {
2320 struct distribute_txn_replay_req *dtrq;
2322 dtrq = distribute_txn_lookup_finish_list(tdtd,
2324 LASSERT(dtrq != NULL);
2325 spin_lock(&tdtd->tdtd_replay_list_lock);
2326 list_del_init(&dtrq->dtrq_list);
2327 spin_unlock(&tdtd->tdtd_replay_list_lock);
2330 drop_duplicate_replay_req(env, obd, req);
2335 LASSERT(trd->trd_processing_task == current_pid());
2336 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2337 lustre_msg_get_transno(req->rq_reqmsg),
2338 libcfs_nid2str(req->rq_peer.nid));
2340 handle_recovery_req(thread, req,
2341 trd->trd_recovery_handler);
2343 * bz18031: increase next_recovery_transno before
2344 * target_request_copy_put() will drop exp_rpc reference
2346 spin_lock(&obd->obd_recovery_task_lock);
2347 obd->obd_next_recovery_transno++;
2348 spin_unlock(&obd->obd_recovery_task_lock);
2349 target_exp_dequeue_req_replay(req);
2350 target_request_copy_put(req);
2351 obd->obd_replayed_requests++;
2352 } else if (type == UPDATE_RECOVERY && transno != 0) {
2353 struct distribute_txn_replay_req *dtrq;
2356 spin_unlock(&obd->obd_recovery_task_lock);
2358 LASSERT(tdtd != NULL);
2359 dtrq = distribute_txn_get_next_req(tdtd);
2360 lu_context_enter(&thread->t_env->le_ctx);
2361 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2362 lu_context_exit(&thread->t_env->le_ctx);
2363 extend_recovery_timer(obd, obd_timeout, true);
2365 if (rc == 0 && dtrq->dtrq_xid != 0) {
2366 CDEBUG(D_HA, "Move x"LPU64" t"LPU64
2367 " to finish list\n", dtrq->dtrq_xid,
2368 dtrq->dtrq_master_transno);
2370 /* Add it to the replay finish list */
2371 spin_lock(&tdtd->tdtd_replay_list_lock);
2372 list_add(&dtrq->dtrq_list,
2373 &tdtd->tdtd_replay_finish_list);
2374 spin_unlock(&tdtd->tdtd_replay_list_lock);
2376 spin_lock(&obd->obd_recovery_task_lock);
2377 if (transno == obd->obd_next_recovery_transno)
2378 obd->obd_next_recovery_transno++;
2380 obd->obd_next_recovery_transno)
2381 obd->obd_next_recovery_transno =
2383 spin_unlock(&obd->obd_recovery_task_lock);
2388 spin_unlock(&obd->obd_recovery_task_lock);
2390 LASSERT(list_empty(&obd->obd_req_replay_queue));
2391 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2392 /** evict exports failed VBR */
2393 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2399 static int target_recovery_thread(void *arg)
2401 struct lu_target *lut = arg;
2402 struct obd_device *obd = lut->lut_obd;
2403 struct ptlrpc_request *req;
2404 struct target_recovery_data *trd = &obd->obd_recovery_data;
2405 unsigned long delta;
2407 struct ptlrpc_thread *thread = NULL;
2411 unshare_fs_struct();
2412 OBD_ALLOC_PTR(thread);
2418 OBD_FREE_PTR(thread);
2422 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2424 OBD_FREE_PTR(thread);
2429 thread->t_env = env;
2430 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2431 env->le_ctx.lc_thread = thread;
2432 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2433 thread->t_watchdog = NULL;
2435 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2437 trd->trd_processing_task = current_pid();
2439 spin_lock(&obd->obd_dev_lock);
2440 obd->obd_recovering = 1;
2441 spin_unlock(&obd->obd_dev_lock);
2442 complete(&trd->trd_starting);
2444 /* first of all, we have to know the first transno to replay */
2445 if (target_recovery_overseer(lut, check_for_recovery_ready,
2446 exp_connect_healthy)) {
2447 abort_req_replay_queue(obd);
2448 abort_lock_replay_queue(obd);
2449 if (lut->lut_tdtd != NULL)
2450 dtrq_list_destroy(lut->lut_tdtd);
2453 /* next stage: replay requests or update */
2455 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2456 atomic_read(&obd->obd_req_replay_clients),
2457 obd->obd_next_recovery_transno);
2458 replay_request_or_update(env, lut, trd, thread);
2461 * The second stage: replay locks
2463 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2464 atomic_read(&obd->obd_lock_replay_clients));
2465 while ((req = target_next_replay_lock(lut))) {
2466 LASSERT(trd->trd_processing_task == current_pid());
2467 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2468 libcfs_nid2str(req->rq_peer.nid));
2469 handle_recovery_req(thread, req,
2470 trd->trd_recovery_handler);
2471 target_request_copy_put(req);
2472 obd->obd_replayed_locks++;
2476 * The third stage: reply on final pings, at this moment all clients
2477 * must have request in final queue
2479 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2480 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2481 /** Update server last boot epoch */
2482 tgt_boot_epoch_update(lut);
2483 /* We drop recoverying flag to forward all new requests
2484 * to regular mds_handle() since now */
2485 spin_lock(&obd->obd_dev_lock);
2486 obd->obd_recovering = obd->obd_abort_recovery = 0;
2487 spin_unlock(&obd->obd_dev_lock);
2488 spin_lock(&obd->obd_recovery_task_lock);
2489 target_cancel_recovery_timer(obd);
2490 spin_unlock(&obd->obd_recovery_task_lock);
2491 while ((req = target_next_final_ping(obd))) {
2492 LASSERT(trd->trd_processing_task == current_pid());
2493 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2494 libcfs_nid2str(req->rq_peer.nid));
2495 handle_recovery_req(thread, req,
2496 trd->trd_recovery_handler);
2497 /* Because the waiting client can not send ping to server,
2498 * so we need refresh the last_request_time, to avoid the
2499 * export is being evicted */
2500 ptlrpc_update_export_timer(req->rq_export, 0);
2501 target_request_copy_put(req);
2504 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2505 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2506 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2507 if (delta > OBD_RECOVERY_TIME_SOFT) {
2508 CWARN("too long recovery - read logs\n");
2509 libcfs_debug_dumplog();
2512 target_finish_recovery(lut);
2514 lu_context_fini(&env->le_ctx);
2515 trd->trd_processing_task = 0;
2516 complete(&trd->trd_finishing);
2518 tgt_io_thread_done(thread);
2519 OBD_FREE_PTR(thread);
2524 static int target_start_recovery_thread(struct lu_target *lut,
2525 svc_handler_t handler)
2527 struct obd_device *obd = lut->lut_obd;
2529 struct target_recovery_data *trd = &obd->obd_recovery_data;
2532 memset(trd, 0, sizeof(*trd));
2533 init_completion(&trd->trd_starting);
2534 init_completion(&trd->trd_finishing);
2535 trd->trd_recovery_handler = handler;
2537 rc = server_name2index(obd->obd_name, &index, NULL);
2541 if (!IS_ERR(kthread_run(target_recovery_thread,
2542 lut, "tgt_recover_%d", index))) {
2543 wait_for_completion(&trd->trd_starting);
2544 LASSERT(obd->obd_recovering != 0);
2552 void target_stop_recovery_thread(struct obd_device *obd)
2554 if (obd->obd_recovery_data.trd_processing_task > 0) {
2555 struct target_recovery_data *trd = &obd->obd_recovery_data;
2556 /** recovery can be done but postrecovery is not yet */
2557 spin_lock(&obd->obd_dev_lock);
2558 if (obd->obd_recovering) {
2559 CERROR("%s: Aborting recovery\n", obd->obd_name);
2560 obd->obd_abort_recovery = 1;
2561 wake_up(&obd->obd_next_transno_waitq);
2563 spin_unlock(&obd->obd_dev_lock);
2564 wait_for_completion(&trd->trd_finishing);
2567 EXPORT_SYMBOL(target_stop_recovery_thread);
2569 void target_recovery_fini(struct obd_device *obd)
2571 class_disconnect_exports(obd);
2572 target_stop_recovery_thread(obd);
2573 target_cleanup_recovery(obd);
2575 EXPORT_SYMBOL(target_recovery_fini);
2577 static void target_recovery_expired(unsigned long castmeharder)
2579 struct obd_device *obd = (struct obd_device *)castmeharder;
2580 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2581 " after %lds (%d clients connected)\n",
2582 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2583 cfs_time_current_sec()- obd->obd_recovery_start,
2584 atomic_read(&obd->obd_connected_clients));
2586 obd->obd_recovery_expired = 1;
2587 wake_up(&obd->obd_next_transno_waitq);
2590 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2592 struct obd_device *obd = lut->lut_obd;
2594 if (obd->obd_max_recoverable_clients == 0) {
2595 /** Update server last boot epoch */
2596 tgt_boot_epoch_update(lut);
2600 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2601 "last_transno "LPU64"\n", obd->obd_name,
2602 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2603 LASSERT(obd->obd_stopping == 0);
2604 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2605 obd->obd_recovery_start = 0;
2606 obd->obd_recovery_end = 0;
2608 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2609 target_start_recovery_thread(lut, handler);
2611 EXPORT_SYMBOL(target_recovery_init);
2613 static int target_process_req_flags(struct obd_device *obd,
2614 struct ptlrpc_request *req)
2616 struct obd_export *exp = req->rq_export;
2617 LASSERT(exp != NULL);
2618 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2619 /* client declares he's ready to replay locks */
2620 spin_lock(&exp->exp_lock);
2621 if (exp->exp_req_replay_needed) {
2622 exp->exp_req_replay_needed = 0;
2623 spin_unlock(&exp->exp_lock);
2625 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2626 atomic_dec(&obd->obd_req_replay_clients);
2628 spin_unlock(&exp->exp_lock);
2631 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2632 /* client declares he's ready to complete recovery
2633 * so, we put the request on th final queue */
2634 spin_lock(&exp->exp_lock);
2635 if (exp->exp_lock_replay_needed) {
2636 exp->exp_lock_replay_needed = 0;
2637 spin_unlock(&exp->exp_lock);
2639 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2640 atomic_dec(&obd->obd_lock_replay_clients);
2642 spin_unlock(&exp->exp_lock);
2648 int target_queue_recovery_request(struct ptlrpc_request *req,
2649 struct obd_device *obd)
2651 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2652 struct ptlrpc_request *reqiter;
2656 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2657 /* Processing the queue right now, don't re-add. */
2661 target_process_req_flags(obd, req);
2663 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2664 /* client declares he's ready to complete recovery
2665 * so, we put the request on th final queue */
2666 target_request_copy_get(req);
2667 DEBUG_REQ(D_HA, req, "queue final req");
2668 wake_up(&obd->obd_next_transno_waitq);
2669 spin_lock(&obd->obd_recovery_task_lock);
2670 if (obd->obd_recovering) {
2671 list_add_tail(&req->rq_list,
2672 &obd->obd_final_req_queue);
2674 spin_unlock(&obd->obd_recovery_task_lock);
2675 target_request_copy_put(req);
2676 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2678 spin_unlock(&obd->obd_recovery_task_lock);
2681 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2682 /* client declares he's ready to replay locks */
2683 target_request_copy_get(req);
2684 DEBUG_REQ(D_HA, req, "queue lock replay req");
2685 wake_up(&obd->obd_next_transno_waitq);
2686 spin_lock(&obd->obd_recovery_task_lock);
2687 LASSERT(obd->obd_recovering);
2688 /* usually due to recovery abort */
2689 if (!req->rq_export->exp_in_recovery) {
2690 spin_unlock(&obd->obd_recovery_task_lock);
2691 target_request_copy_put(req);
2694 LASSERT(req->rq_export->exp_lock_replay_needed);
2695 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2696 spin_unlock(&obd->obd_recovery_task_lock);
2700 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2701 * (i.e. buflens etc are in my own byte order), but type-dependent
2702 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2705 INIT_LIST_HEAD(&req->rq_list);
2706 DEBUG_REQ(D_HA, req, "not queueing");
2710 /* If we're processing the queue, we want don't want to queue this
2713 * Also, if this request has a transno less than the one we're waiting
2714 * for, we should process it now. It could (and currently always will)
2715 * be an open request for a descriptor that was opened some time ago.
2717 * Also, a resent, replayed request that has already been
2718 * handled will pass through here and be processed immediately.
2720 CDEBUG(D_HA, "Next recovery transno: "LPU64
2721 ", current: "LPU64", replaying\n",
2722 obd->obd_next_recovery_transno, transno);
2724 /* If the request has been replayed by update replay, then sends this
2725 * request to the recovery thread (replay_request_or_update()), where
2726 * it will be handled */
2727 spin_lock(&obd->obd_recovery_task_lock);
2728 if (transno < obd->obd_next_recovery_transno &&
2729 !is_req_replayed_by_update(req)) {
2730 /* Processing the queue right now, don't re-add. */
2731 LASSERT(list_empty(&req->rq_list));
2732 spin_unlock(&obd->obd_recovery_task_lock);
2735 spin_unlock(&obd->obd_recovery_task_lock);
2737 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2740 target_request_copy_get(req);
2741 if (!req->rq_export->exp_in_recovery) {
2742 target_request_copy_put(req);
2745 LASSERT(req->rq_export->exp_req_replay_needed);
2747 if (target_exp_enqueue_req_replay(req)) {
2748 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2749 target_request_copy_put(req);
2754 spin_lock(&obd->obd_recovery_task_lock);
2755 LASSERT(obd->obd_recovering);
2756 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2757 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2758 list_add_tail(&req->rq_list, &reqiter->rq_list);
2763 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2765 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2766 "has been claimed by another client");
2767 spin_unlock(&obd->obd_recovery_task_lock);
2768 target_exp_dequeue_req_replay(req);
2769 target_request_copy_put(req);
2775 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2777 obd->obd_requests_queued_for_recovery++;
2778 spin_unlock(&obd->obd_recovery_task_lock);
2779 wake_up(&obd->obd_next_transno_waitq);
2783 int target_handle_ping(struct ptlrpc_request *req)
2785 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2786 return req_capsule_server_pack(&req->rq_pill);
2789 void target_committed_to_req(struct ptlrpc_request *req)
2791 struct obd_export *exp = req->rq_export;
2793 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2794 lustre_msg_set_last_committed(req->rq_repmsg,
2795 exp->exp_last_committed);
2797 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2798 "%d)", exp->exp_obd->obd_no_transno,
2799 req->rq_repmsg == NULL);
2801 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2802 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2805 #endif /* HAVE_SERVER_SUPPORT */
2808 * Packs current SLV and Limit into \a req.
2810 int target_pack_pool_reply(struct ptlrpc_request *req)
2812 struct obd_device *obd;
2815 /* Check that we still have all structures alive as this may
2816 * be some late RPC at shutdown time. */
2817 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2818 !exp_connect_lru_resize(req->rq_export))) {
2819 lustre_msg_set_slv(req->rq_repmsg, 0);
2820 lustre_msg_set_limit(req->rq_repmsg, 0);
2824 /* OBD is alive here as export is alive, which we checked above. */
2825 obd = req->rq_export->exp_obd;
2827 read_lock(&obd->obd_pool_lock);
2828 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2829 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2830 read_unlock(&obd->obd_pool_lock);
2835 static int target_send_reply_msg(struct ptlrpc_request *req,
2836 int rc, int fail_id)
2838 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2839 DEBUG_REQ(D_ERROR, req, "dropping reply");
2842 /* We can have a null rq_reqmsg in the event of bad signature or
2843 * no context when unwrapping */
2844 if (req->rq_reqmsg &&
2845 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2846 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2850 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2851 req->rq_status = rc;
2852 return ptlrpc_send_error(req, 1);
2854 DEBUG_REQ(D_NET, req, "sending reply");
2857 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2860 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2862 struct ptlrpc_service_part *svcpt;
2864 struct ptlrpc_reply_state *rs;
2865 struct obd_export *exp;
2868 if (req->rq_no_reply) {
2873 svcpt = req->rq_rqbd->rqbd_svcpt;
2874 rs = req->rq_reply_state;
2875 if (rs == NULL || !rs->rs_difficult) {
2877 target_send_reply_msg (req, rc, fail_id);
2882 /* must be an export if locks saved */
2883 LASSERT(req->rq_export != NULL);
2884 /* req/reply consistent */
2885 LASSERT(rs->rs_svcpt == svcpt);
2888 LASSERT(!rs->rs_scheduled);
2889 LASSERT(!rs->rs_scheduled_ever);
2890 LASSERT(!rs->rs_handled);
2891 LASSERT(!rs->rs_on_net);
2892 LASSERT(rs->rs_export == NULL);
2893 LASSERT(list_empty(&rs->rs_obd_list));
2894 LASSERT(list_empty(&rs->rs_exp_list));
2896 exp = class_export_get(req->rq_export);
2898 /* disable reply scheduling while I'm setting up */
2899 rs->rs_scheduled = 1;
2901 rs->rs_xid = req->rq_xid;
2902 rs->rs_transno = req->rq_transno;
2903 rs->rs_export = exp;
2904 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2906 spin_lock(&exp->exp_uncommitted_replies_lock);
2907 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2908 rs->rs_transno, exp->exp_last_committed);
2909 if (rs->rs_transno > exp->exp_last_committed) {
2910 /* not committed already */
2911 list_add_tail(&rs->rs_obd_list,
2912 &exp->exp_uncommitted_replies);
2914 spin_unlock(&exp->exp_uncommitted_replies_lock);
2916 spin_lock(&exp->exp_lock);
2917 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2918 spin_unlock(&exp->exp_lock);
2920 netrc = target_send_reply_msg(req, rc, fail_id);
2922 spin_lock(&svcpt->scp_rep_lock);
2924 atomic_inc(&svcpt->scp_nreps_difficult);
2927 /* error sending: reply is off the net. Also we need +1
2928 * reply ref until ptlrpc_handle_rs() is done
2929 * with the reply state (if the send was successful, there
2930 * would have been +1 ref for the net, which
2931 * reply_out_callback leaves alone) */
2933 ptlrpc_rs_addref(rs);
2936 spin_lock(&rs->rs_lock);
2937 if (rs->rs_transno <= exp->exp_last_committed ||
2938 (!rs->rs_on_net && !rs->rs_no_ack) ||
2939 list_empty(&rs->rs_exp_list) || /* completed already */
2940 list_empty(&rs->rs_obd_list)) {
2941 CDEBUG(D_HA, "Schedule reply immediately\n");
2942 ptlrpc_dispatch_difficult_reply(rs);
2944 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2945 rs->rs_scheduled = 0; /* allow notifier to schedule */
2947 spin_unlock(&rs->rs_lock);
2948 spin_unlock(&svcpt->scp_rep_lock);
2952 enum ldlm_mode lck_compat_array[] = {
2953 [LCK_EX] = LCK_COMPAT_EX,
2954 [LCK_PW] = LCK_COMPAT_PW,
2955 [LCK_PR] = LCK_COMPAT_PR,
2956 [LCK_CW] = LCK_COMPAT_CW,
2957 [LCK_CR] = LCK_COMPAT_CR,
2958 [LCK_NL] = LCK_COMPAT_NL,
2959 [LCK_GROUP] = LCK_COMPAT_GROUP,
2960 [LCK_COS] = LCK_COMPAT_COS,
2964 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2965 * not escape to the user level.
2967 int ldlm_error2errno(enum ldlm_error error)
2973 case ELDLM_LOCK_MATCHED:
2976 case ELDLM_LOCK_CHANGED:
2979 case ELDLM_LOCK_ABORTED:
2982 case ELDLM_LOCK_REPLACED:
2985 case ELDLM_NO_LOCK_DATA:
2988 case ELDLM_NAMESPACE_EXISTS:
2991 case ELDLM_BAD_NAMESPACE:
2995 if (((int)error) < 0) { /* cast to signed type */
2996 result = error; /* as ldlm_error can be unsigned */
2998 CERROR("Invalid DLM result code: %d\n", error);
3004 EXPORT_SYMBOL(ldlm_error2errno);
3007 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3009 enum ldlm_error ldlm_errno2error(int err_no)
3018 error = ELDLM_LOCK_CHANGED;
3021 error = ELDLM_LOCK_ABORTED;
3024 error = ELDLM_LOCK_REPLACED;
3027 error = ELDLM_NO_LOCK_DATA;
3030 error = ELDLM_NAMESPACE_EXISTS;
3033 error = ELDLM_BAD_NAMESPACE;
3041 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3042 void ldlm_dump_export_locks(struct obd_export *exp)
3044 spin_lock(&exp->exp_locks_list_guard);
3045 if (!list_empty(&exp->exp_locks_list)) {
3046 struct ldlm_lock *lock;
3048 CERROR("dumping locks for export %p,"
3049 "ignore if the unmount doesn't hang\n", exp);
3050 list_for_each_entry(lock, &exp->exp_locks_list,
3052 LDLM_ERROR(lock, "lock:");
3054 spin_unlock(&exp->exp_locks_list_guard);
3058 #ifdef HAVE_SERVER_SUPPORT
3059 static int target_bulk_timeout(void *data)
3062 /* We don't fail the connection here, because having the export
3063 * killed makes the (vital) call to commitrw very sad.
3068 static inline const char *bulk2type(struct ptlrpc_request *req)
3070 if (req->rq_bulk_read)
3072 if (req->rq_bulk_write)
3077 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3078 struct l_wait_info *lwi)
3080 struct ptlrpc_request *req = desc->bd_req;
3081 time_t start = cfs_time_current_sec();
3087 /* If there is eviction in progress, wait for it to finish. */
3088 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3089 *lwi = LWI_INTR(NULL, NULL);
3090 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3091 !atomic_read(&exp->exp_obd->
3092 obd_evict_inprogress),
3096 /* Check if client was evicted or reconnected already. */
3097 if (exp->exp_failed ||
3098 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3101 if (req->rq_bulk_read)
3102 rc = sptlrpc_svc_wrap_bulk(req, desc);
3104 if ((exp->exp_connect_data.ocd_connect_flags &
3105 OBD_CONNECT_BULK_MBITS) != 0)
3106 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3107 else /* old version, bulk matchbits is rq_xid */
3108 req->rq_mbits = req->rq_xid;
3111 rc = ptlrpc_start_bulk_transfer(desc);
3115 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3116 bulk2type(req), rc);
3120 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3121 ptlrpc_abort_bulk(desc);
3125 /* limit actual bulk transfer to bulk_timeout seconds */
3126 deadline = start + bulk_timeout;
3127 if (deadline > req->rq_deadline)
3128 deadline = req->rq_deadline;
3131 long timeoutl = deadline - cfs_time_current_sec();
3132 cfs_duration_t timeout = timeoutl <= 0 ?
3133 CFS_TICK : cfs_time_seconds(timeoutl);
3136 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3137 target_bulk_timeout, desc);
3138 rc = l_wait_event(desc->bd_waitq,
3139 !ptlrpc_server_bulk_active(desc) ||
3142 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3144 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3145 /* Wait again if we changed rq_deadline. */
3146 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3147 deadline = start + bulk_timeout;
3148 if (deadline > rq_deadline)
3149 deadline = rq_deadline;
3150 } while ((rc == -ETIMEDOUT) &&
3151 (deadline > cfs_time_current_sec()));
3153 if (rc == -ETIMEDOUT) {
3154 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3155 bulk2type(req), deadline - start,
3156 cfs_time_current_sec() - deadline);
3157 ptlrpc_abort_bulk(desc);
3158 } else if (exp->exp_failed) {
3159 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3162 ptlrpc_abort_bulk(desc);
3163 } else if (exp->exp_conn_cnt >
3164 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3165 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3167 /* We don't reply anyway. */
3169 ptlrpc_abort_bulk(desc);
3170 } else if (desc->bd_failure) {
3171 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3173 /* XXX should this be a different errno? */
3176 if (req->rq_bulk_write)
3177 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3178 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3179 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3180 bulk2type(req), desc->bd_nob_transferred,
3182 /* XXX should this be a different errno? */
3189 EXPORT_SYMBOL(target_bulk_io);
3191 #endif /* HAVE_SERVER_SUPPORT */