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 ||
1121 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1122 export->exp_connection != NULL) {
1123 spin_unlock(&export->exp_lock);
1124 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1125 /* MDS or LWP reconnected after failover. */
1126 LCONSOLE_WARN("%s: Received %s connection from "
1127 "%s, removing former export from %s\n",
1128 target->obd_name, mds_conn ? "MDS" : "LWP",
1129 libcfs_nid2str(req->rq_peer.nid),
1130 libcfs_nid2str(export->exp_connection->c_peer.nid));
1132 /* New MDS connection from the same NID. */
1133 LCONSOLE_WARN("%s: Received new %s connection from "
1134 "%s, removing former export from same NID\n",
1135 target->obd_name, mds_conn ? "MDS" : "LWP",
1136 libcfs_nid2str(req->rq_peer.nid));
1139 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1140 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1141 /* Because exports between MDTs will always be
1142 * kept, let's do not fail such export if they
1143 * come from the same NID, otherwise it might
1144 * cause eviction between MDTs, which might
1145 * cause namespace inconsistency */
1146 spin_lock(&export->exp_lock);
1147 export->exp_connecting = 1;
1148 spin_unlock(&export->exp_lock);
1149 conn.cookie = export->exp_handle.h_cookie;
1152 class_fail_export(export);
1153 class_export_put(export);
1157 } else if (export->exp_connection != NULL &&
1158 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1159 (lustre_msg_get_op_flags(req->rq_reqmsg) &
1160 MSG_CONNECT_INITIAL)) {
1161 spin_unlock(&export->exp_lock);
1162 /* In MDS failover we have static UUID but NID can change. */
1163 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1164 "existing nid %s is already connected\n",
1165 target->obd_name, cluuid.uuid,
1166 libcfs_nid2str(req->rq_peer.nid),
1168 export->exp_connection->c_peer.nid));
1170 class_export_put(export);
1173 export->exp_connecting = 1;
1174 spin_unlock(&export->exp_lock);
1175 LASSERT(export->exp_obd == target);
1177 rc = target_handle_reconnect(&conn, export, &cluuid);
1180 /* If we found an export, we already unlocked. */
1183 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1184 } else if (req->rq_export == NULL &&
1185 atomic_read(&export->exp_rpc_count) > 0) {
1186 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1187 "still busy with %d references\n",
1188 target->obd_name, cluuid.uuid,
1189 libcfs_nid2str(req->rq_peer.nid),
1190 atomic_read(&export->exp_refcount));
1191 GOTO(out, rc = -EBUSY);
1192 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1193 if (!strstr(cluuid.uuid, "mdt"))
1194 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1195 "known client %s (at %s) because it "
1196 "is indicating it is a new client",
1197 target->obd_name, cluuid.uuid,
1198 libcfs_nid2str(req->rq_peer.nid));
1199 GOTO(out, rc = -EALREADY);
1201 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1208 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1209 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1210 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1211 export, (long)cfs_time_current_sec(),
1212 export ? (long)export->exp_last_request_time : 0);
1214 /* If this is the first time a client connects, reset the recovery
1215 * timer. Discard lightweight connections which might be local. */
1216 if (!lw_client && rc == 0 && target->obd_recovering)
1217 check_and_start_recovery_timer(target, req, export == NULL);
1219 /* We want to handle EALREADY but *not* -EALREADY from
1220 * target_handle_reconnect(), return reconnection state in a flag. */
1221 if (rc == EALREADY) {
1222 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1228 /* Tell the client if we support replayable requests. */
1229 if (target->obd_replayable)
1230 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1231 client_nid = &req->rq_peer.nid;
1233 if (export == NULL) {
1234 /* allow lightweight connections during recovery */
1235 /* allow "new" MDT to be connected during recovery, since we
1236 * need retrieve recovery update records from it */
1237 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1239 int c; /* connected */
1240 int i; /* in progress */
1242 int s; /* stale/evicted */
1244 c = atomic_read(&target->obd_connected_clients);
1245 i = atomic_read(&target->obd_lock_replay_clients);
1246 k = target->obd_max_recoverable_clients;
1247 s = target->obd_stale_clients;
1248 t = cfs_timer_deadline(&target->obd_recovery_timer);
1249 t = cfs_time_sub(t, cfs_time_current());
1250 t = cfs_duration_sec(t);
1251 LCONSOLE_WARN("%s: Denying connection for new client %s"
1252 "(at %s), waiting for %d known clients "
1253 "(%d recovered, %d in progress, and %d "
1254 "evicted) to recover in %d:%.02d\n",
1255 target->obd_name, cluuid.uuid,
1256 libcfs_nid2str(req->rq_peer.nid), k,
1257 c - i, i, s, (int)t / 60,
1262 rc = obd_connect(req->rq_svc_thread->t_env,
1263 &export, target, &cluuid, data,
1265 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1266 lustre_msg_add_op_flags(req->rq_repmsg,
1267 MSG_CONNECT_RECOVERING);
1269 conn.cookie = export->exp_handle.h_cookie;
1270 rc = rev_import_init(export);
1274 new_mds_mds_conn = true;
1277 rc = obd_reconnect(req->rq_svc_thread->t_env,
1278 export, target, &cluuid, data, client_nid);
1283 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1284 data->ocd_instance = target->u.obt.obt_instance;
1286 /* Return only the parts of obd_connect_data that we understand, so the
1287 * client knows that we don't understand the rest. */
1289 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1291 tmpdata = req_capsule_server_get(&req->rq_pill,
1293 /* Don't use struct assignment here, because the client reply
1294 * buffer may be smaller/larger than the local struct
1295 * obd_connect_data. */
1296 memcpy(tmpdata, data, min(tmpsize, size));
1299 /* If the client and the server are the same node, we will already
1300 * have an export that really points to the client's DLM export,
1301 * because we have a shared handles table.
1303 * XXX this will go away when shaver stops sending the "connect" handle
1304 * in the real "remote handle" field of the request --phik 24 Apr 2003
1306 ptlrpc_request_change_export(req, export);
1308 spin_lock(&export->exp_lock);
1309 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1310 spin_unlock(&export->exp_lock);
1311 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1312 "or equal conn_cnt: %d >= %d\n",
1313 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1314 export->exp_conn_cnt,
1315 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1317 GOTO(out, rc = -EALREADY);
1319 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1320 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1322 /* Don't evict liblustre clients for not pinging. */
1323 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1324 export->exp_libclient = 1;
1325 spin_unlock(&export->exp_lock);
1327 spin_lock(&target->obd_dev_lock);
1328 list_del_init(&export->exp_obd_chain_timed);
1329 spin_unlock(&target->obd_dev_lock);
1331 spin_unlock(&export->exp_lock);
1334 if (export->exp_connection != NULL) {
1335 /* Check to see if connection came from another NID. */
1336 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1337 !hlist_unhashed(&export->exp_nid_hash))
1338 cfs_hash_del(export->exp_obd->obd_nid_hash,
1339 &export->exp_connection->c_peer.nid,
1340 &export->exp_nid_hash);
1342 ptlrpc_connection_put(export->exp_connection);
1345 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1348 if (hlist_unhashed(&export->exp_nid_hash))
1349 cfs_hash_add(export->exp_obd->obd_nid_hash,
1350 &export->exp_connection->c_peer.nid,
1351 &export->exp_nid_hash);
1353 lustre_msg_set_handle(req->rq_repmsg, &conn);
1355 rc = rev_import_reconnect(export, req);
1359 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1361 __u64 transno = data->ocd_transno;
1363 spin_lock(&export->exp_lock);
1364 /* possible race with class_disconnect_stale_exports,
1365 * export may be already in the eviction process */
1366 if (export->exp_failed) {
1367 spin_unlock(&export->exp_lock);
1368 GOTO(out, rc = -ENODEV);
1370 export->exp_in_recovery = 1;
1371 export->exp_req_replay_needed = 1;
1372 export->exp_lock_replay_needed = 1;
1373 spin_unlock(&export->exp_lock);
1375 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1376 MSG_CONNECT_TRANSNO);
1377 if (has_transno && transno == 0)
1378 CWARN("Connect with zero transno!\n");
1380 if (has_transno && transno > 0 &&
1381 transno < target->obd_next_recovery_transno &&
1382 transno > target->obd_last_committed) {
1383 /* Another way is to use cmpxchg() to be lock-free. */
1384 spin_lock(&target->obd_recovery_task_lock);
1385 if (transno < target->obd_next_recovery_transno)
1386 target->obd_next_recovery_transno = transno;
1387 spin_unlock(&target->obd_recovery_task_lock);
1390 atomic_inc(&target->obd_req_replay_clients);
1391 atomic_inc(&target->obd_lock_replay_clients);
1392 /* Note: MDS-MDS connection is allowed to be connected during
1393 * recovery, no matter if the exports needs to be recoveried.
1394 * Because we need retrieve updates logs from all other MDTs.
1395 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1396 * also needs to be increased to match other recovery checking
1398 if (new_mds_mds_conn)
1399 target->obd_max_recoverable_clients++;
1400 if (atomic_inc_return(&target->obd_connected_clients) ==
1401 target->obd_max_recoverable_clients)
1402 wake_up(&target->obd_next_transno_waitq);
1405 /* Tell the client we're in recovery, when client is involved in it. */
1406 if (target->obd_recovering && !lw_client)
1407 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1411 spin_lock(&export->exp_lock);
1412 export->exp_connecting = 0;
1413 spin_unlock(&export->exp_lock);
1415 class_export_put(export);
1417 if (target_referenced == true && target != NULL) {
1418 spin_lock(&target->obd_dev_lock);
1419 target->obd_conn_inprogress--;
1420 spin_unlock(&target->obd_dev_lock);
1422 class_decref(target, __func__, current);
1424 req->rq_status = rc;
1428 int target_handle_disconnect(struct ptlrpc_request *req)
1433 rc = req_capsule_server_pack(&req->rq_pill);
1437 /* Keep the rq_export around so we can send the reply. */
1438 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1443 void target_destroy_export(struct obd_export *exp)
1445 struct obd_import *imp = NULL;
1446 /* exports created from last_rcvd data, and "fake"
1447 exports created by lctl don't have an import */
1448 spin_lock(&exp->exp_lock);
1449 if (exp->exp_imp_reverse != NULL) {
1450 imp = exp->exp_imp_reverse;
1451 exp->exp_imp_reverse = NULL;
1453 spin_unlock(&exp->exp_lock);
1455 client_destroy_import(imp);
1457 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1458 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1459 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1460 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1462 EXPORT_SYMBOL(target_destroy_export);
1465 * Recovery functions
1467 static void target_request_copy_get(struct ptlrpc_request *req)
1469 class_export_rpc_inc(req->rq_export);
1470 LASSERT(list_empty(&req->rq_list));
1471 INIT_LIST_HEAD(&req->rq_replay_list);
1473 /* Increase refcount to keep request in queue. */
1474 atomic_inc(&req->rq_refcount);
1475 /* Let export know it has replays to be handled. */
1476 atomic_inc(&req->rq_export->exp_replay_count);
1479 static void target_request_copy_put(struct ptlrpc_request *req)
1481 LASSERT(list_empty(&req->rq_replay_list));
1482 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1484 atomic_dec(&req->rq_export->exp_replay_count);
1485 class_export_rpc_dec(req->rq_export);
1486 ptlrpc_server_drop_request(req);
1489 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1491 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1492 struct obd_export *exp = req->rq_export;
1493 struct ptlrpc_request *reqiter;
1494 struct ptlrpc_request *dup_req = NULL;
1499 spin_lock(&exp->exp_lock);
1500 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1502 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1510 /* We expect it with RESENT and REPLAY flags. */
1511 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1512 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1513 CERROR("invalid flags %x of resent replay\n",
1514 lustre_msg_get_flags(req->rq_reqmsg));
1516 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1519 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1521 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1522 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1526 list_add_tail(&req->rq_replay_list,
1527 &exp->exp_req_replay_queue);
1530 spin_unlock(&exp->exp_lock);
1534 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1536 LASSERT(!list_empty(&req->rq_replay_list));
1537 LASSERT(req->rq_export);
1539 spin_lock(&req->rq_export->exp_lock);
1540 list_del_init(&req->rq_replay_list);
1541 spin_unlock(&req->rq_export->exp_lock);
1544 static void target_finish_recovery(struct lu_target *lut)
1546 struct obd_device *obd = lut->lut_obd;
1549 /* Only log a recovery message when recovery has occurred. */
1550 if (obd->obd_recovery_start) {
1551 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1552 obd->obd_recovery_start);
1553 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1554 "%d recovered and %d %s evicted.\n", obd->obd_name,
1555 (int)elapsed_time / 60, (int)elapsed_time % 60,
1556 obd->obd_max_recoverable_clients,
1557 atomic_read(&obd->obd_connected_clients),
1558 obd->obd_stale_clients,
1559 obd->obd_stale_clients == 1 ? "was" : "were");
1562 ldlm_reprocess_all_ns(obd->obd_namespace);
1563 spin_lock(&obd->obd_recovery_task_lock);
1564 if (!list_empty(&obd->obd_req_replay_queue) ||
1565 !list_empty(&obd->obd_lock_replay_queue) ||
1566 !list_empty(&obd->obd_final_req_queue)) {
1567 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1569 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1570 list_empty(&obd->obd_lock_replay_queue) ? \
1572 list_empty(&obd->obd_final_req_queue) ? \
1574 spin_unlock(&obd->obd_recovery_task_lock);
1577 spin_unlock(&obd->obd_recovery_task_lock);
1579 obd->obd_recovery_end = cfs_time_current_sec();
1581 /* When recovery finished, cleanup orphans on MDS and OST. */
1582 if (OBT(obd) && OBP(obd, postrecov)) {
1583 int rc = OBP(obd, postrecov)(obd);
1585 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1591 static void abort_req_replay_queue(struct obd_device *obd)
1593 struct ptlrpc_request *req, *n;
1594 struct list_head abort_list;
1596 INIT_LIST_HEAD(&abort_list);
1597 spin_lock(&obd->obd_recovery_task_lock);
1598 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1599 spin_unlock(&obd->obd_recovery_task_lock);
1600 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1601 DEBUG_REQ(D_WARNING, req, "aborted:");
1602 req->rq_status = -ENOTCONN;
1603 if (ptlrpc_error(req)) {
1604 DEBUG_REQ(D_ERROR, req,
1605 "failed abort_req_reply; skipping");
1607 target_exp_dequeue_req_replay(req);
1608 target_request_copy_put(req);
1612 static void abort_lock_replay_queue(struct obd_device *obd)
1614 struct ptlrpc_request *req, *n;
1615 struct list_head abort_list;
1617 INIT_LIST_HEAD(&abort_list);
1618 spin_lock(&obd->obd_recovery_task_lock);
1619 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1620 spin_unlock(&obd->obd_recovery_task_lock);
1621 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1622 DEBUG_REQ(D_ERROR, req, "aborted:");
1623 req->rq_status = -ENOTCONN;
1624 if (ptlrpc_error(req)) {
1625 DEBUG_REQ(D_ERROR, req,
1626 "failed abort_lock_reply; skipping");
1628 target_request_copy_put(req);
1632 /* Called from a cleanup function if the device is being cleaned up
1633 forcefully. The exports should all have been disconnected already,
1634 the only thing left to do is
1635 - clear the recovery flags
1637 - free queued requests and replies, but don't send replies
1638 Because the obd_stopping flag is set, no new requests should be received.
1641 void target_cleanup_recovery(struct obd_device *obd)
1643 struct ptlrpc_request *req, *n;
1644 struct list_head clean_list;
1646 INIT_LIST_HEAD(&clean_list);
1647 spin_lock(&obd->obd_dev_lock);
1648 if (!obd->obd_recovering) {
1649 spin_unlock(&obd->obd_dev_lock);
1653 obd->obd_recovering = obd->obd_abort_recovery = 0;
1654 spin_unlock(&obd->obd_dev_lock);
1656 spin_lock(&obd->obd_recovery_task_lock);
1657 target_cancel_recovery_timer(obd);
1658 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1659 spin_unlock(&obd->obd_recovery_task_lock);
1661 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1662 LASSERT(req->rq_reply_state == NULL);
1663 target_exp_dequeue_req_replay(req);
1664 target_request_copy_put(req);
1667 spin_lock(&obd->obd_recovery_task_lock);
1668 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1669 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1670 spin_unlock(&obd->obd_recovery_task_lock);
1672 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1673 LASSERT(req->rq_reply_state == NULL);
1674 target_request_copy_put(req);
1679 EXPORT_SYMBOL(target_cleanup_recovery);
1681 /* obd_recovery_task_lock should be held */
1682 void target_cancel_recovery_timer(struct obd_device *obd)
1684 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1685 cfs_timer_disarm(&obd->obd_recovery_timer);
1688 static void target_start_recovery_timer(struct obd_device *obd)
1690 if (obd->obd_recovery_start != 0)
1693 spin_lock(&obd->obd_dev_lock);
1694 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1695 spin_unlock(&obd->obd_dev_lock);
1699 LASSERT(obd->obd_recovery_timeout != 0);
1701 if (obd->obd_recovery_start != 0) {
1702 spin_unlock(&obd->obd_dev_lock);
1706 cfs_timer_arm(&obd->obd_recovery_timer,
1707 cfs_time_shift(obd->obd_recovery_timeout));
1708 obd->obd_recovery_start = cfs_time_current_sec();
1709 spin_unlock(&obd->obd_dev_lock);
1711 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1712 "or until %d client%s reconnect%s\n",
1714 obd->obd_recovery_timeout / 60,
1715 obd->obd_recovery_timeout % 60,
1716 obd->obd_max_recoverable_clients,
1717 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1718 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1722 * extend recovery window.
1724 * if @extend is true, extend recovery window to have @drt remaining at least;
1725 * otherwise, make sure the recovery timeout value is not less than @drt.
1727 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1731 cfs_duration_t left;
1734 spin_lock(&obd->obd_dev_lock);
1735 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1736 spin_unlock(&obd->obd_dev_lock);
1739 LASSERT(obd->obd_recovery_start != 0);
1741 now = cfs_time_current_sec();
1742 to = obd->obd_recovery_timeout;
1743 end = obd->obd_recovery_start + to;
1744 left = cfs_time_sub(end, now);
1746 if (extend && (drt > left)) {
1748 } else if (!extend && (drt > to)) {
1752 if (to > obd->obd_recovery_time_hard)
1753 to = obd->obd_recovery_time_hard;
1754 if (obd->obd_recovery_timeout < to) {
1755 obd->obd_recovery_timeout = to;
1756 end = obd->obd_recovery_start + to;
1757 cfs_timer_arm(&obd->obd_recovery_timer,
1758 cfs_time_shift(end - now));
1760 spin_unlock(&obd->obd_dev_lock);
1762 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1763 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1766 /* Reset the timer with each new client connection */
1768 * This timer is actually reconnect_timer, which is for making sure
1769 * the total recovery window is at least as big as my reconnect
1770 * attempt timing. So the initial recovery time_out will be set to
1771 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1772 * from client is bigger than this, then the recovery time_out will
1773 * be extended to make sure the client could be reconnected, in the
1774 * process, the timeout from the new client should be ignored.
1778 check_and_start_recovery_timer(struct obd_device *obd,
1779 struct ptlrpc_request *req,
1782 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1783 struct obd_device_target *obt = &obd->u.obt;
1785 if (!new_client && service_time)
1786 /* Teach server about old server's estimates, as first guess
1787 * at how long new requests will take. */
1788 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1791 target_start_recovery_timer(obd);
1793 /* Convert the service time to RPC timeout,
1794 * and reuse service_time to limit stack usage. */
1795 service_time = at_est2timeout(service_time);
1797 /* We expect other clients to timeout within service_time, then try
1798 * to reconnect, then try the failover server. The max delay between
1799 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1800 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1802 LASSERT(obt->obt_magic == OBT_MAGIC);
1803 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1804 if (service_time > obd->obd_recovery_timeout && !new_client)
1805 extend_recovery_timer(obd, service_time, false);
1808 /** Health checking routines */
1809 static inline int exp_connect_healthy(struct obd_export *exp)
1811 return (exp->exp_in_recovery);
1814 /** if export done req_replay or has replay in queue */
1815 static inline int exp_req_replay_healthy(struct obd_export *exp)
1817 return (!exp->exp_req_replay_needed ||
1818 atomic_read(&exp->exp_replay_count) > 0);
1822 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1824 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1825 exp_req_replay_healthy(exp);
1828 /** if export done lock_replay or has replay in queue */
1829 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1831 return (!exp->exp_lock_replay_needed ||
1832 atomic_read(&exp->exp_replay_count) > 0);
1835 static inline int exp_vbr_healthy(struct obd_export *exp)
1837 return (!exp->exp_vbr_failed);
1840 static inline int exp_finished(struct obd_export *exp)
1842 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1845 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1847 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1851 static int check_for_next_transno(struct lu_target *lut)
1853 struct ptlrpc_request *req = NULL;
1854 struct obd_device *obd = lut->lut_obd;
1855 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1856 int wake_up = 0, connected, completed, queue_len;
1857 __u64 req_transno = 0;
1858 __u64 update_transno = 0;
1859 __u64 next_transno = 0;
1862 spin_lock(&obd->obd_recovery_task_lock);
1863 if (!list_empty(&obd->obd_req_replay_queue)) {
1864 req = list_entry(obd->obd_req_replay_queue.next,
1865 struct ptlrpc_request, rq_list);
1866 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1870 update_transno = distribute_txn_get_next_transno(tdtd);
1872 connected = atomic_read(&obd->obd_connected_clients);
1873 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1874 queue_len = obd->obd_requests_queued_for_recovery;
1875 next_transno = obd->obd_next_recovery_transno;
1877 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1878 "req_transno: "LPU64", next_transno: "LPU64"\n",
1879 obd->obd_max_recoverable_clients, connected, completed,
1880 queue_len, req_transno, next_transno);
1882 if (obd->obd_abort_recovery) {
1883 CDEBUG(D_HA, "waking for aborted recovery\n");
1885 } else if (obd->obd_recovery_expired) {
1886 CDEBUG(D_HA, "waking for expired recovery\n");
1888 } else if (tdtd != NULL && req != NULL &&
1889 is_req_replayed_by_update(req)) {
1890 LASSERTF(req_transno < next_transno, "req_transno "LPU64
1891 "next_transno"LPU64"\n", req_transno, next_transno);
1892 CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n",
1895 } else if (req_transno == next_transno ||
1896 (update_transno != 0 && update_transno <= next_transno)) {
1897 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1899 } else if (queue_len > 0 &&
1900 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1902 /** handle gaps occured due to lost reply or VBR */
1903 LASSERTF(req_transno >= next_transno,
1904 "req_transno: "LPU64", next_transno: "LPU64"\n",
1905 req_transno, next_transno);
1906 if (req_transno > obd->obd_last_committed &&
1907 !obd->obd_version_recov)
1910 "%s: waking for gap in transno, VBR is %s (skip: "
1911 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1912 ", next_update "LPD64" last_committed: "LPD64")\n",
1913 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1914 next_transno, queue_len, completed, connected,
1915 req_transno, update_transno, obd->obd_last_committed);
1916 obd->obd_next_recovery_transno = req_transno;
1918 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1919 CDEBUG(D_HA, "waking for completed recovery\n");
1921 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1922 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1923 " by fail_lock, waking up ("LPD64")\n", next_transno);
1924 obd->obd_next_recovery_transno = req_transno;
1927 spin_unlock(&obd->obd_recovery_task_lock);
1931 static int check_for_next_lock(struct lu_target *lut)
1933 struct obd_device *obd = lut->lut_obd;
1936 spin_lock(&obd->obd_recovery_task_lock);
1937 if (!list_empty(&obd->obd_lock_replay_queue)) {
1938 CDEBUG(D_HA, "waking for next lock\n");
1940 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1941 CDEBUG(D_HA, "waking for completed lock replay\n");
1943 } else if (obd->obd_abort_recovery) {
1944 CDEBUG(D_HA, "waking for aborted recovery\n");
1946 } else if (obd->obd_recovery_expired) {
1947 CDEBUG(D_HA, "waking for expired recovery\n");
1950 spin_unlock(&obd->obd_recovery_task_lock);
1956 * wait for recovery events,
1957 * check its status with help of check_routine
1958 * evict dead clients via health_check
1960 static int target_recovery_overseer(struct lu_target *lut,
1961 int (*check_routine)(struct lu_target *),
1962 int (*health_check)(struct obd_export *))
1964 struct obd_device *obd = lut->lut_obd;
1965 struct target_distribute_txn_data *tdtd;
1967 if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1968 (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1969 __u64 next_update_transno = 0;
1971 /* Only abort the recovery if there are no update recovery
1972 * left in the queue */
1973 spin_lock(&obd->obd_recovery_task_lock);
1974 if (lut->lut_tdtd != NULL) {
1975 next_update_transno =
1976 distribute_txn_get_next_transno(lut->lut_tdtd);
1978 tdtd = lut->lut_tdtd;
1979 /* If next_update_transno == 0, it probably because
1980 * updatelog retrieve threads did not get any records
1981 * yet, let's wait those threads stopped */
1982 if (next_update_transno == 0) {
1983 struct l_wait_info lwi = { 0 };
1985 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1987 &tdtd->tdtd_recovery_threads_count) == 0,
1990 next_update_transno =
1991 distribute_txn_get_next_transno(
1996 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
1997 obd->obd_next_recovery_transno = next_update_transno;
1998 spin_unlock(&obd->obd_recovery_task_lock);
1999 /* Disconnect unfinished exports from clients, and
2000 * keep connection from MDT to make sure the update
2001 * recovery will still keep trying until some one
2002 * manually abort the recovery */
2003 class_disconnect_stale_exports(obd,
2004 exp_finished_or_from_mdt);
2005 /* Abort all of replay and replay lock req from
2007 abort_req_replay_queue(obd);
2008 abort_lock_replay_queue(obd);
2009 CDEBUG(D_HA, "%s: there are still update replay ("LPX64
2010 ")in the queue.\n", obd->obd_name,
2011 next_update_transno);
2013 obd->obd_abort_recovery = 1;
2014 spin_unlock(&obd->obd_recovery_task_lock);
2015 CWARN("%s recovery is aborted by hard timeout\n",
2020 while (wait_event_timeout(obd->obd_next_transno_waitq,
2022 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2023 /* wait indefinitely for event, but don't trigger watchdog */;
2025 if (obd->obd_abort_recovery) {
2026 CWARN("recovery is aborted, evict exports in recovery\n");
2027 if (lut->lut_tdtd != NULL) {
2028 struct l_wait_info lwi = { 0 };
2030 tdtd = lut->lut_tdtd;
2031 /* Let's wait all of the update log recovery thread
2033 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2034 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2036 /* Then abort the update recovery list */
2037 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2038 dtrq_list_destroy(lut->lut_tdtd);
2041 /** evict exports which didn't finish recovery yet */
2042 class_disconnect_stale_exports(obd, exp_finished);
2044 } else if (obd->obd_recovery_expired) {
2045 obd->obd_recovery_expired = 0;
2046 /** If some clients died being recovered, evict them */
2047 LCONSOLE_WARN("%s: recovery is timed out, "
2048 "evict stale exports\n", obd->obd_name);
2049 /** evict cexports with no replay in queue, they are stalled */
2050 class_disconnect_stale_exports(obd, health_check);
2052 /** continue with VBR */
2053 spin_lock(&obd->obd_dev_lock);
2054 obd->obd_version_recov = 1;
2055 spin_unlock(&obd->obd_dev_lock);
2057 * reset timer, recovery will proceed with versions now,
2058 * timeout is set just to handle reconnection delays
2060 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2061 /** Wait for recovery events again, after evicting bad clients */
2067 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2069 struct obd_device *obd = lut->lut_obd;
2070 struct ptlrpc_request *req = NULL;
2072 CDEBUG(D_HA, "Waiting for lock\n");
2073 if (target_recovery_overseer(lut, check_for_next_lock,
2074 exp_lock_replay_healthy))
2075 abort_lock_replay_queue(obd);
2077 spin_lock(&obd->obd_recovery_task_lock);
2078 if (!list_empty(&obd->obd_lock_replay_queue)) {
2079 req = list_entry(obd->obd_lock_replay_queue.next,
2080 struct ptlrpc_request, rq_list);
2081 list_del_init(&req->rq_list);
2082 spin_unlock(&obd->obd_recovery_task_lock);
2084 spin_unlock(&obd->obd_recovery_task_lock);
2085 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2086 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2087 /** evict exports failed VBR */
2088 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2093 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2095 struct ptlrpc_request *req = NULL;
2097 spin_lock(&obd->obd_recovery_task_lock);
2098 if (!list_empty(&obd->obd_final_req_queue)) {
2099 req = list_entry(obd->obd_final_req_queue.next,
2100 struct ptlrpc_request, rq_list);
2101 list_del_init(&req->rq_list);
2102 spin_unlock(&obd->obd_recovery_task_lock);
2103 if (req->rq_export->exp_in_recovery) {
2104 spin_lock(&req->rq_export->exp_lock);
2105 req->rq_export->exp_in_recovery = 0;
2106 spin_unlock(&req->rq_export->exp_lock);
2109 spin_unlock(&obd->obd_recovery_task_lock);
2114 static void handle_recovery_req(struct ptlrpc_thread *thread,
2115 struct ptlrpc_request *req,
2116 svc_handler_t handler)
2121 * export can be evicted during recovery, no need to handle replays for
2122 * it after that, discard such request silently
2124 if (req->rq_export->exp_disconnected)
2127 req->rq_session.lc_thread = thread;
2128 req->rq_svc_thread = thread;
2129 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2131 /* thread context */
2132 lu_context_enter(&thread->t_env->le_ctx);
2134 lu_context_exit(&thread->t_env->le_ctx);
2136 /* don't reset timer for final stage */
2137 if (!exp_finished(req->rq_export)) {
2138 int to = obd_timeout;
2141 * Add request timeout to the recovery time so next request from
2142 * this client may come in recovery time
2145 struct ptlrpc_service_part *svcpt;
2147 svcpt = req->rq_rqbd->rqbd_svcpt;
2148 /* If the server sent early reply for this request,
2149 * the client will recalculate the timeout according to
2150 * current server estimate service time, so we will
2151 * use the maxium timeout here for waiting the client
2152 * sending the next req */
2153 to = max((int)at_est2timeout(
2154 at_get(&svcpt->scp_at_estimate)),
2155 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2156 /* Add 2 net_latency, one for balance rq_deadline
2157 * (see ptl_send_rpc), one for resend the req to server,
2158 * Note: client will pack net_latency in replay req
2159 * (see ptlrpc_replay_req) */
2160 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2162 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2167 /** Checking routines for recovery */
2168 static int check_for_recovery_ready(struct lu_target *lut)
2170 struct obd_device *obd = lut->lut_obd;
2171 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2173 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2174 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2175 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2176 obd->obd_recovery_expired);
2178 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2179 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2180 if (clnts + obd->obd_stale_clients <
2181 obd->obd_max_recoverable_clients)
2185 if (lut->lut_tdtd != NULL) {
2186 if (!lut->lut_tdtd->tdtd_replay_ready &&
2187 !obd->obd_abort_recovery) {
2188 /* Let's extend recovery timer, in case the recovery
2189 * timer expired, and some clients got evicted */
2190 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2192 CDEBUG(D_HA, "%s update recovery is not ready,"
2193 " extend recovery %d\n", obd->obd_name,
2194 obd->obd_recovery_timeout);
2197 dtrq_list_dump(lut->lut_tdtd, D_HA);
2205 REQUEST_RECOVERY = 1,
2206 UPDATE_RECOVERY = 2,
2209 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2213 if (!list_empty(&obd->obd_req_replay_queue)) {
2214 struct ptlrpc_request *req;
2216 req = list_entry(obd->obd_req_replay_queue.next,
2217 struct ptlrpc_request, rq_list);
2218 transno = lustre_msg_get_transno(req->rq_reqmsg);
2224 static __u64 get_next_transno(struct lu_target *lut, int *type)
2226 struct obd_device *obd = lut->lut_obd;
2227 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2229 __u64 update_transno;
2232 transno = get_next_replay_req_transno(obd);
2234 *type = REQUEST_RECOVERY;
2239 update_transno = distribute_txn_get_next_transno(tdtd);
2240 if (transno == 0 || (transno >= update_transno &&
2241 update_transno != 0)) {
2242 transno = update_transno;
2244 *type = UPDATE_RECOVERY;
2251 * drop duplicate replay request
2253 * Because the operation has been replayed by update recovery, the request
2254 * with the same transno will be dropped and also notify the client to send
2255 * next replay request.
2257 * \param[in] env execution environment
2258 * \param[in] obd failover obd device
2259 * \param[in] req request to be dropped
2261 static void drop_duplicate_replay_req(struct lu_env *env,
2262 struct obd_device *obd,
2263 struct ptlrpc_request *req)
2265 DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2266 " update records are found.\n",
2267 lustre_msg_get_transno(req->rq_reqmsg),
2268 libcfs_nid2str(req->rq_peer.nid));
2270 /* Right now, only for MDS reint operation update replay and
2271 * normal request replay can have the same transno */
2272 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2273 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2274 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2275 if (likely(req->rq_export))
2276 target_committed_to_req(req);
2277 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2278 target_send_reply(req, req->rq_status, 0);
2280 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2281 libcfs_nid2str(req->rq_peer.nid));
2283 target_exp_dequeue_req_replay(req);
2284 target_request_copy_put(req);
2285 obd->obd_replayed_requests++;
2288 static void replay_request_or_update(struct lu_env *env,
2289 struct lu_target *lut,
2290 struct target_recovery_data *trd,
2291 struct ptlrpc_thread *thread)
2293 struct obd_device *obd = lut->lut_obd;
2294 struct ptlrpc_request *req = NULL;
2299 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2300 obd->obd_next_recovery_transno);
2302 /* Replay all of request and update by transno */
2304 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2306 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2308 /** It is needed to extend recovery window above
2309 * recovery_time_soft. Extending is possible only in the
2310 * end of recovery window (see more details in
2311 * handle_recovery_req()).
2313 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2315 if (target_recovery_overseer(lut, check_for_next_transno,
2316 exp_req_replay_healthy_or_from_mdt)) {
2317 abort_req_replay_queue(obd);
2318 abort_lock_replay_queue(obd);
2322 spin_lock(&obd->obd_recovery_task_lock);
2323 transno = get_next_transno(lut, &type);
2324 if (type == REQUEST_RECOVERY && transno != 0) {
2325 /* Drop replay request from client side, if the
2326 * replay has been executed by update with the
2328 req = list_entry(obd->obd_req_replay_queue.next,
2329 struct ptlrpc_request, rq_list);
2331 list_del_init(&req->rq_list);
2332 obd->obd_requests_queued_for_recovery--;
2333 spin_unlock(&obd->obd_recovery_task_lock);
2335 /* Let's check if the request has been redone by
2337 if (is_req_replayed_by_update(req)) {
2338 struct distribute_txn_replay_req *dtrq;
2340 dtrq = distribute_txn_lookup_finish_list(tdtd,
2342 LASSERT(dtrq != NULL);
2343 spin_lock(&tdtd->tdtd_replay_list_lock);
2344 list_del_init(&dtrq->dtrq_list);
2345 spin_unlock(&tdtd->tdtd_replay_list_lock);
2348 drop_duplicate_replay_req(env, obd, req);
2353 LASSERT(trd->trd_processing_task == current_pid());
2354 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2355 lustre_msg_get_transno(req->rq_reqmsg),
2356 libcfs_nid2str(req->rq_peer.nid));
2358 handle_recovery_req(thread, req,
2359 trd->trd_recovery_handler);
2361 * bz18031: increase next_recovery_transno before
2362 * target_request_copy_put() will drop exp_rpc reference
2364 spin_lock(&obd->obd_recovery_task_lock);
2365 obd->obd_next_recovery_transno++;
2366 spin_unlock(&obd->obd_recovery_task_lock);
2367 target_exp_dequeue_req_replay(req);
2368 target_request_copy_put(req);
2369 obd->obd_replayed_requests++;
2370 } else if (type == UPDATE_RECOVERY && transno != 0) {
2371 struct distribute_txn_replay_req *dtrq;
2374 spin_unlock(&obd->obd_recovery_task_lock);
2376 LASSERT(tdtd != NULL);
2377 dtrq = distribute_txn_get_next_req(tdtd);
2378 lu_context_enter(&thread->t_env->le_ctx);
2379 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2380 lu_context_exit(&thread->t_env->le_ctx);
2381 extend_recovery_timer(obd, obd_timeout, true);
2383 if (rc == 0 && dtrq->dtrq_xid != 0) {
2384 CDEBUG(D_HA, "Move x"LPU64" t"LPU64
2385 " to finish list\n", dtrq->dtrq_xid,
2386 dtrq->dtrq_master_transno);
2388 /* Add it to the replay finish list */
2389 spin_lock(&tdtd->tdtd_replay_list_lock);
2390 list_add(&dtrq->dtrq_list,
2391 &tdtd->tdtd_replay_finish_list);
2392 spin_unlock(&tdtd->tdtd_replay_list_lock);
2394 spin_lock(&obd->obd_recovery_task_lock);
2395 if (transno == obd->obd_next_recovery_transno)
2396 obd->obd_next_recovery_transno++;
2398 obd->obd_next_recovery_transno)
2399 obd->obd_next_recovery_transno =
2401 spin_unlock(&obd->obd_recovery_task_lock);
2406 spin_unlock(&obd->obd_recovery_task_lock);
2408 LASSERT(list_empty(&obd->obd_req_replay_queue));
2409 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2410 /** evict exports failed VBR */
2411 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2417 static int target_recovery_thread(void *arg)
2419 struct lu_target *lut = arg;
2420 struct obd_device *obd = lut->lut_obd;
2421 struct ptlrpc_request *req;
2422 struct target_recovery_data *trd = &obd->obd_recovery_data;
2423 unsigned long delta;
2425 struct ptlrpc_thread *thread = NULL;
2429 unshare_fs_struct();
2430 OBD_ALLOC_PTR(thread);
2436 OBD_FREE_PTR(thread);
2440 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2442 OBD_FREE_PTR(thread);
2447 thread->t_env = env;
2448 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2449 env->le_ctx.lc_thread = thread;
2450 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2451 thread->t_watchdog = NULL;
2453 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2455 trd->trd_processing_task = current_pid();
2457 spin_lock(&obd->obd_dev_lock);
2458 obd->obd_recovering = 1;
2459 spin_unlock(&obd->obd_dev_lock);
2460 complete(&trd->trd_starting);
2462 /* first of all, we have to know the first transno to replay */
2463 if (target_recovery_overseer(lut, check_for_recovery_ready,
2464 exp_connect_healthy)) {
2465 abort_req_replay_queue(obd);
2466 abort_lock_replay_queue(obd);
2467 if (lut->lut_tdtd != NULL)
2468 dtrq_list_destroy(lut->lut_tdtd);
2471 /* next stage: replay requests or update */
2473 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2474 atomic_read(&obd->obd_req_replay_clients),
2475 obd->obd_next_recovery_transno);
2476 replay_request_or_update(env, lut, trd, thread);
2479 * The second stage: replay locks
2481 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2482 atomic_read(&obd->obd_lock_replay_clients));
2483 while ((req = target_next_replay_lock(lut))) {
2484 LASSERT(trd->trd_processing_task == current_pid());
2485 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2486 libcfs_nid2str(req->rq_peer.nid));
2487 handle_recovery_req(thread, req,
2488 trd->trd_recovery_handler);
2489 target_request_copy_put(req);
2490 obd->obd_replayed_locks++;
2494 * The third stage: reply on final pings, at this moment all clients
2495 * must have request in final queue
2497 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2498 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2499 /** Update server last boot epoch */
2500 tgt_boot_epoch_update(lut);
2501 /* We drop recoverying flag to forward all new requests
2502 * to regular mds_handle() since now */
2503 spin_lock(&obd->obd_dev_lock);
2504 obd->obd_recovering = obd->obd_abort_recovery = 0;
2505 spin_unlock(&obd->obd_dev_lock);
2506 spin_lock(&obd->obd_recovery_task_lock);
2507 target_cancel_recovery_timer(obd);
2508 spin_unlock(&obd->obd_recovery_task_lock);
2509 while ((req = target_next_final_ping(obd))) {
2510 LASSERT(trd->trd_processing_task == current_pid());
2511 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2512 libcfs_nid2str(req->rq_peer.nid));
2513 handle_recovery_req(thread, req,
2514 trd->trd_recovery_handler);
2515 /* Because the waiting client can not send ping to server,
2516 * so we need refresh the last_request_time, to avoid the
2517 * export is being evicted */
2518 ptlrpc_update_export_timer(req->rq_export, 0);
2519 target_request_copy_put(req);
2522 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2523 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2524 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2525 if (delta > OBD_RECOVERY_TIME_SOFT) {
2526 CWARN("too long recovery - read logs\n");
2527 libcfs_debug_dumplog();
2530 target_finish_recovery(lut);
2532 lu_context_fini(&env->le_ctx);
2533 trd->trd_processing_task = 0;
2534 complete(&trd->trd_finishing);
2536 tgt_io_thread_done(thread);
2537 OBD_FREE_PTR(thread);
2542 static int target_start_recovery_thread(struct lu_target *lut,
2543 svc_handler_t handler)
2545 struct obd_device *obd = lut->lut_obd;
2547 struct target_recovery_data *trd = &obd->obd_recovery_data;
2550 memset(trd, 0, sizeof(*trd));
2551 init_completion(&trd->trd_starting);
2552 init_completion(&trd->trd_finishing);
2553 trd->trd_recovery_handler = handler;
2555 rc = server_name2index(obd->obd_name, &index, NULL);
2559 if (!IS_ERR(kthread_run(target_recovery_thread,
2560 lut, "tgt_recover_%d", index))) {
2561 wait_for_completion(&trd->trd_starting);
2562 LASSERT(obd->obd_recovering != 0);
2570 void target_stop_recovery_thread(struct obd_device *obd)
2572 if (obd->obd_recovery_data.trd_processing_task > 0) {
2573 struct target_recovery_data *trd = &obd->obd_recovery_data;
2574 /** recovery can be done but postrecovery is not yet */
2575 spin_lock(&obd->obd_dev_lock);
2576 if (obd->obd_recovering) {
2577 CERROR("%s: Aborting recovery\n", obd->obd_name);
2578 obd->obd_abort_recovery = 1;
2579 wake_up(&obd->obd_next_transno_waitq);
2581 spin_unlock(&obd->obd_dev_lock);
2582 wait_for_completion(&trd->trd_finishing);
2585 EXPORT_SYMBOL(target_stop_recovery_thread);
2587 void target_recovery_fini(struct obd_device *obd)
2589 class_disconnect_exports(obd);
2590 target_stop_recovery_thread(obd);
2591 target_cleanup_recovery(obd);
2593 EXPORT_SYMBOL(target_recovery_fini);
2595 static void target_recovery_expired(unsigned long castmeharder)
2597 struct obd_device *obd = (struct obd_device *)castmeharder;
2598 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2599 " after %lds (%d clients connected)\n",
2600 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2601 cfs_time_current_sec()- obd->obd_recovery_start,
2602 atomic_read(&obd->obd_connected_clients));
2604 obd->obd_recovery_expired = 1;
2605 wake_up(&obd->obd_next_transno_waitq);
2608 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2610 struct obd_device *obd = lut->lut_obd;
2612 if (obd->obd_max_recoverable_clients == 0) {
2613 /** Update server last boot epoch */
2614 tgt_boot_epoch_update(lut);
2618 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2619 "last_transno "LPU64"\n", obd->obd_name,
2620 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2621 LASSERT(obd->obd_stopping == 0);
2622 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2623 obd->obd_recovery_start = 0;
2624 obd->obd_recovery_end = 0;
2626 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2627 target_start_recovery_thread(lut, handler);
2629 EXPORT_SYMBOL(target_recovery_init);
2631 static int target_process_req_flags(struct obd_device *obd,
2632 struct ptlrpc_request *req)
2634 struct obd_export *exp = req->rq_export;
2635 LASSERT(exp != NULL);
2636 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2637 /* client declares he's ready to replay locks */
2638 spin_lock(&exp->exp_lock);
2639 if (exp->exp_req_replay_needed) {
2640 exp->exp_req_replay_needed = 0;
2641 spin_unlock(&exp->exp_lock);
2643 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2644 atomic_dec(&obd->obd_req_replay_clients);
2646 spin_unlock(&exp->exp_lock);
2649 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2650 /* client declares he's ready to complete recovery
2651 * so, we put the request on th final queue */
2652 spin_lock(&exp->exp_lock);
2653 if (exp->exp_lock_replay_needed) {
2654 exp->exp_lock_replay_needed = 0;
2655 spin_unlock(&exp->exp_lock);
2657 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2658 atomic_dec(&obd->obd_lock_replay_clients);
2660 spin_unlock(&exp->exp_lock);
2666 int target_queue_recovery_request(struct ptlrpc_request *req,
2667 struct obd_device *obd)
2669 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2670 struct ptlrpc_request *reqiter;
2674 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2675 /* Processing the queue right now, don't re-add. */
2679 target_process_req_flags(obd, req);
2681 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2682 /* client declares he's ready to complete recovery
2683 * so, we put the request on th final queue */
2684 target_request_copy_get(req);
2685 DEBUG_REQ(D_HA, req, "queue final req");
2686 wake_up(&obd->obd_next_transno_waitq);
2687 spin_lock(&obd->obd_recovery_task_lock);
2688 if (obd->obd_recovering) {
2689 list_add_tail(&req->rq_list,
2690 &obd->obd_final_req_queue);
2692 spin_unlock(&obd->obd_recovery_task_lock);
2693 target_request_copy_put(req);
2694 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2696 spin_unlock(&obd->obd_recovery_task_lock);
2699 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2700 /* client declares he's ready to replay locks */
2701 target_request_copy_get(req);
2702 DEBUG_REQ(D_HA, req, "queue lock replay req");
2703 wake_up(&obd->obd_next_transno_waitq);
2704 spin_lock(&obd->obd_recovery_task_lock);
2705 LASSERT(obd->obd_recovering);
2706 /* usually due to recovery abort */
2707 if (!req->rq_export->exp_in_recovery) {
2708 spin_unlock(&obd->obd_recovery_task_lock);
2709 target_request_copy_put(req);
2712 LASSERT(req->rq_export->exp_lock_replay_needed);
2713 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2714 spin_unlock(&obd->obd_recovery_task_lock);
2718 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2719 * (i.e. buflens etc are in my own byte order), but type-dependent
2720 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2723 INIT_LIST_HEAD(&req->rq_list);
2724 DEBUG_REQ(D_HA, req, "not queueing");
2728 /* If we're processing the queue, we want don't want to queue this
2731 * Also, if this request has a transno less than the one we're waiting
2732 * for, we should process it now. It could (and currently always will)
2733 * be an open request for a descriptor that was opened some time ago.
2735 * Also, a resent, replayed request that has already been
2736 * handled will pass through here and be processed immediately.
2738 CDEBUG(D_HA, "Next recovery transno: "LPU64
2739 ", current: "LPU64", replaying\n",
2740 obd->obd_next_recovery_transno, transno);
2742 /* If the request has been replayed by update replay, then sends this
2743 * request to the recovery thread (replay_request_or_update()), where
2744 * it will be handled */
2745 spin_lock(&obd->obd_recovery_task_lock);
2746 if (transno < obd->obd_next_recovery_transno &&
2747 !is_req_replayed_by_update(req)) {
2748 /* Processing the queue right now, don't re-add. */
2749 LASSERT(list_empty(&req->rq_list));
2750 spin_unlock(&obd->obd_recovery_task_lock);
2753 spin_unlock(&obd->obd_recovery_task_lock);
2755 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2758 target_request_copy_get(req);
2759 if (!req->rq_export->exp_in_recovery) {
2760 target_request_copy_put(req);
2763 LASSERT(req->rq_export->exp_req_replay_needed);
2765 if (target_exp_enqueue_req_replay(req)) {
2766 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2767 target_request_copy_put(req);
2772 spin_lock(&obd->obd_recovery_task_lock);
2773 LASSERT(obd->obd_recovering);
2774 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2775 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2776 list_add_tail(&req->rq_list, &reqiter->rq_list);
2781 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2783 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2784 "has been claimed by another client");
2785 spin_unlock(&obd->obd_recovery_task_lock);
2786 target_exp_dequeue_req_replay(req);
2787 target_request_copy_put(req);
2793 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2795 obd->obd_requests_queued_for_recovery++;
2796 spin_unlock(&obd->obd_recovery_task_lock);
2797 wake_up(&obd->obd_next_transno_waitq);
2801 int target_handle_ping(struct ptlrpc_request *req)
2803 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2804 return req_capsule_server_pack(&req->rq_pill);
2807 void target_committed_to_req(struct ptlrpc_request *req)
2809 struct obd_export *exp = req->rq_export;
2811 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2812 lustre_msg_set_last_committed(req->rq_repmsg,
2813 exp->exp_last_committed);
2815 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2816 "%d)", exp->exp_obd->obd_no_transno,
2817 req->rq_repmsg == NULL);
2819 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2820 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2823 #endif /* HAVE_SERVER_SUPPORT */
2826 * Packs current SLV and Limit into \a req.
2828 int target_pack_pool_reply(struct ptlrpc_request *req)
2830 struct obd_device *obd;
2833 /* Check that we still have all structures alive as this may
2834 * be some late RPC at shutdown time. */
2835 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2836 !exp_connect_lru_resize(req->rq_export))) {
2837 lustre_msg_set_slv(req->rq_repmsg, 0);
2838 lustre_msg_set_limit(req->rq_repmsg, 0);
2842 /* OBD is alive here as export is alive, which we checked above. */
2843 obd = req->rq_export->exp_obd;
2845 read_lock(&obd->obd_pool_lock);
2846 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2847 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2848 read_unlock(&obd->obd_pool_lock);
2853 static int target_send_reply_msg(struct ptlrpc_request *req,
2854 int rc, int fail_id)
2856 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2857 DEBUG_REQ(D_ERROR, req, "dropping reply");
2860 /* We can have a null rq_reqmsg in the event of bad signature or
2861 * no context when unwrapping */
2862 if (req->rq_reqmsg &&
2863 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2864 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2868 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2869 req->rq_status = rc;
2870 return ptlrpc_send_error(req, 1);
2872 DEBUG_REQ(D_NET, req, "sending reply");
2875 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2878 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2880 struct ptlrpc_service_part *svcpt;
2882 struct ptlrpc_reply_state *rs;
2883 struct obd_export *exp;
2886 if (req->rq_no_reply) {
2891 svcpt = req->rq_rqbd->rqbd_svcpt;
2892 rs = req->rq_reply_state;
2893 if (rs == NULL || !rs->rs_difficult) {
2895 target_send_reply_msg (req, rc, fail_id);
2900 /* must be an export if locks saved */
2901 LASSERT(req->rq_export != NULL);
2902 /* req/reply consistent */
2903 LASSERT(rs->rs_svcpt == svcpt);
2906 LASSERT(!rs->rs_scheduled);
2907 LASSERT(!rs->rs_scheduled_ever);
2908 LASSERT(!rs->rs_handled);
2909 LASSERT(!rs->rs_on_net);
2910 LASSERT(rs->rs_export == NULL);
2911 LASSERT(list_empty(&rs->rs_obd_list));
2912 LASSERT(list_empty(&rs->rs_exp_list));
2914 exp = class_export_get(req->rq_export);
2916 /* disable reply scheduling while I'm setting up */
2917 rs->rs_scheduled = 1;
2919 rs->rs_xid = req->rq_xid;
2920 rs->rs_transno = req->rq_transno;
2921 rs->rs_export = exp;
2922 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2924 spin_lock(&exp->exp_uncommitted_replies_lock);
2925 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2926 rs->rs_transno, exp->exp_last_committed);
2927 if (rs->rs_transno > exp->exp_last_committed) {
2928 /* not committed already */
2929 list_add_tail(&rs->rs_obd_list,
2930 &exp->exp_uncommitted_replies);
2932 spin_unlock(&exp->exp_uncommitted_replies_lock);
2934 spin_lock(&exp->exp_lock);
2935 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2936 spin_unlock(&exp->exp_lock);
2938 netrc = target_send_reply_msg(req, rc, fail_id);
2940 spin_lock(&svcpt->scp_rep_lock);
2942 atomic_inc(&svcpt->scp_nreps_difficult);
2945 /* error sending: reply is off the net. Also we need +1
2946 * reply ref until ptlrpc_handle_rs() is done
2947 * with the reply state (if the send was successful, there
2948 * would have been +1 ref for the net, which
2949 * reply_out_callback leaves alone) */
2951 ptlrpc_rs_addref(rs);
2954 spin_lock(&rs->rs_lock);
2955 if (rs->rs_transno <= exp->exp_last_committed ||
2956 (!rs->rs_on_net && !rs->rs_no_ack) ||
2957 list_empty(&rs->rs_exp_list) || /* completed already */
2958 list_empty(&rs->rs_obd_list)) {
2959 CDEBUG(D_HA, "Schedule reply immediately\n");
2960 ptlrpc_dispatch_difficult_reply(rs);
2962 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2963 rs->rs_scheduled = 0; /* allow notifier to schedule */
2965 spin_unlock(&rs->rs_lock);
2966 spin_unlock(&svcpt->scp_rep_lock);
2970 enum ldlm_mode lck_compat_array[] = {
2971 [LCK_EX] = LCK_COMPAT_EX,
2972 [LCK_PW] = LCK_COMPAT_PW,
2973 [LCK_PR] = LCK_COMPAT_PR,
2974 [LCK_CW] = LCK_COMPAT_CW,
2975 [LCK_CR] = LCK_COMPAT_CR,
2976 [LCK_NL] = LCK_COMPAT_NL,
2977 [LCK_GROUP] = LCK_COMPAT_GROUP,
2978 [LCK_COS] = LCK_COMPAT_COS,
2982 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2983 * not escape to the user level.
2985 int ldlm_error2errno(enum ldlm_error error)
2991 case ELDLM_LOCK_MATCHED:
2994 case ELDLM_LOCK_CHANGED:
2997 case ELDLM_LOCK_ABORTED:
3000 case ELDLM_LOCK_REPLACED:
3003 case ELDLM_NO_LOCK_DATA:
3006 case ELDLM_NAMESPACE_EXISTS:
3009 case ELDLM_BAD_NAMESPACE:
3013 if (((int)error) < 0) { /* cast to signed type */
3014 result = error; /* as ldlm_error can be unsigned */
3016 CERROR("Invalid DLM result code: %d\n", error);
3022 EXPORT_SYMBOL(ldlm_error2errno);
3025 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3027 enum ldlm_error ldlm_errno2error(int err_no)
3036 error = ELDLM_LOCK_CHANGED;
3039 error = ELDLM_LOCK_ABORTED;
3042 error = ELDLM_LOCK_REPLACED;
3045 error = ELDLM_NO_LOCK_DATA;
3048 error = ELDLM_NAMESPACE_EXISTS;
3051 error = ELDLM_BAD_NAMESPACE;
3059 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3060 void ldlm_dump_export_locks(struct obd_export *exp)
3062 spin_lock(&exp->exp_locks_list_guard);
3063 if (!list_empty(&exp->exp_locks_list)) {
3064 struct ldlm_lock *lock;
3066 CERROR("dumping locks for export %p,"
3067 "ignore if the unmount doesn't hang\n", exp);
3068 list_for_each_entry(lock, &exp->exp_locks_list,
3070 LDLM_ERROR(lock, "lock:");
3072 spin_unlock(&exp->exp_locks_list_guard);
3076 #ifdef HAVE_SERVER_SUPPORT
3077 static int target_bulk_timeout(void *data)
3080 /* We don't fail the connection here, because having the export
3081 * killed makes the (vital) call to commitrw very sad.
3086 static inline const char *bulk2type(struct ptlrpc_request *req)
3088 if (req->rq_bulk_read)
3090 if (req->rq_bulk_write)
3095 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3096 struct l_wait_info *lwi)
3098 struct ptlrpc_request *req = desc->bd_req;
3099 time_t start = cfs_time_current_sec();
3105 /* If there is eviction in progress, wait for it to finish. */
3106 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3107 *lwi = LWI_INTR(NULL, NULL);
3108 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3109 !atomic_read(&exp->exp_obd->
3110 obd_evict_inprogress),
3114 /* Check if client was evicted or reconnected already. */
3115 if (exp->exp_failed ||
3116 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3119 if (req->rq_bulk_read)
3120 rc = sptlrpc_svc_wrap_bulk(req, desc);
3122 if ((exp->exp_connect_data.ocd_connect_flags &
3123 OBD_CONNECT_BULK_MBITS) != 0)
3124 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3125 else /* old version, bulk matchbits is rq_xid */
3126 req->rq_mbits = req->rq_xid;
3129 rc = ptlrpc_start_bulk_transfer(desc);
3133 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3134 bulk2type(req), rc);
3138 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3139 ptlrpc_abort_bulk(desc);
3143 /* limit actual bulk transfer to bulk_timeout seconds */
3144 deadline = start + bulk_timeout;
3145 if (deadline > req->rq_deadline)
3146 deadline = req->rq_deadline;
3149 long timeoutl = deadline - cfs_time_current_sec();
3150 cfs_duration_t timeout = timeoutl <= 0 ?
3151 CFS_TICK : cfs_time_seconds(timeoutl);
3154 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3155 target_bulk_timeout, desc);
3156 rc = l_wait_event(desc->bd_waitq,
3157 !ptlrpc_server_bulk_active(desc) ||
3160 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3162 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3163 /* Wait again if we changed rq_deadline. */
3164 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3165 deadline = start + bulk_timeout;
3166 if (deadline > rq_deadline)
3167 deadline = rq_deadline;
3168 } while ((rc == -ETIMEDOUT) &&
3169 (deadline > cfs_time_current_sec()));
3171 if (rc == -ETIMEDOUT) {
3172 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3173 bulk2type(req), deadline - start,
3174 cfs_time_current_sec() - deadline);
3175 ptlrpc_abort_bulk(desc);
3176 } else if (exp->exp_failed) {
3177 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3180 ptlrpc_abort_bulk(desc);
3181 } else if (exp->exp_conn_cnt >
3182 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3183 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3185 /* We don't reply anyway. */
3187 ptlrpc_abort_bulk(desc);
3188 } else if (desc->bd_failure) {
3189 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3191 /* XXX should this be a different errno? */
3194 if (req->rq_bulk_write)
3195 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3196 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3197 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3198 bulk2type(req), desc->bd_nob_transferred,
3200 /* XXX should this be a different errno? */
3207 EXPORT_SYMBOL(target_bulk_io);
3209 #endif /* HAVE_SERVER_SUPPORT */