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;
567 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
570 rc = ptlrpc_connect_import(imp);
573 data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
574 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
577 LASSERT(*exp != NULL && (*exp)->exp_connection);
580 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
581 ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
582 data->ocd_connect_flags, ocd->ocd_connect_flags);
583 data->ocd_connect_flags = ocd->ocd_connect_flags;
584 /* clear the flag as it was not set and is not known
587 data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
590 ptlrpc_pinger_add_import(imp);
596 cli->cl_conn_count--;
597 class_disconnect(*exp);
601 up_write(&cli->cl_sem);
605 EXPORT_SYMBOL(client_connect_import);
607 int client_disconnect_export(struct obd_export *exp)
609 struct obd_device *obd = class_exp2obd(exp);
610 struct client_obd *cli;
611 struct obd_import *imp;
616 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
617 exp, exp ? exp->exp_handle.h_cookie : -1);
622 imp = cli->cl_import;
624 down_write(&cli->cl_sem);
625 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
628 if (cli->cl_conn_count == 0) {
629 CERROR("disconnecting disconnected device (%s)\n",
631 GOTO(out_disconnect, rc = -EINVAL);
634 cli->cl_conn_count--;
635 if (cli->cl_conn_count != 0)
636 GOTO(out_disconnect, rc = 0);
638 /* Mark import deactivated now, so we don't try to reconnect if any
639 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
640 * fully deactivate the import, or that would drop all requests. */
641 spin_lock(&imp->imp_lock);
642 imp->imp_deactive = 1;
643 spin_unlock(&imp->imp_lock);
645 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
646 * delete it regardless. (It's safe to delete an import that was
648 (void)ptlrpc_pinger_del_import(imp);
650 if (obd->obd_namespace != NULL) {
651 /* obd_force == local only */
652 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
653 obd->obd_force ? LCF_LOCAL : 0, NULL);
654 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
657 /* There's no need to hold sem while disconnecting an import,
658 * and it may actually cause deadlock in GSS. */
659 up_write(&cli->cl_sem);
660 rc = ptlrpc_disconnect_import(imp, 0);
661 down_write(&cli->cl_sem);
663 ptlrpc_invalidate_import(imp);
668 /* Use server style - class_disconnect should be always called for
670 err = class_disconnect(exp);
674 up_write(&cli->cl_sem);
678 EXPORT_SYMBOL(client_disconnect_export);
680 #ifdef HAVE_SERVER_SUPPORT
681 int server_disconnect_export(struct obd_export *exp)
686 /* Disconnect early so that clients can't keep using export. */
687 rc = class_disconnect(exp);
688 /* Close import to avoid sending any requests. */
689 if (exp->exp_imp_reverse)
690 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
692 ldlm_bl_thread_wakeup();
694 /* complete all outstanding replies */
695 spin_lock(&exp->exp_lock);
696 while (!list_empty(&exp->exp_outstanding_replies)) {
697 struct ptlrpc_reply_state *rs =
698 list_entry(exp->exp_outstanding_replies.next,
699 struct ptlrpc_reply_state, rs_exp_list);
700 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
702 spin_lock(&svcpt->scp_rep_lock);
704 list_del_init(&rs->rs_exp_list);
705 spin_lock(&rs->rs_lock);
706 ptlrpc_schedule_difficult_reply(rs);
707 spin_unlock(&rs->rs_lock);
709 spin_unlock(&svcpt->scp_rep_lock);
711 spin_unlock(&exp->exp_lock);
715 EXPORT_SYMBOL(server_disconnect_export);
717 /* --------------------------------------------------------------------------
718 * from old lib/target.c
719 * -------------------------------------------------------------------------- */
721 static int target_handle_reconnect(struct lustre_handle *conn,
722 struct obd_export *exp,
723 struct obd_uuid *cluuid)
725 struct obd_device *target;
726 struct lustre_handle *hdl;
733 hdl = &exp->exp_imp_reverse->imp_remote_handle;
734 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
735 conn->cookie = exp->exp_handle.h_cookie;
736 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
737 " cookie "LPX64"\n", cluuid->uuid, exp, conn->cookie);
741 target = exp->exp_obd;
743 /* Might be a re-connect after a partition. */
744 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
745 LCONSOLE_WARN("%s: already connected client %s (at %s) "
746 "with handle "LPX64". Rejecting client "
747 "with the same UUID trying to reconnect "
748 "with handle "LPX64"\n", target->obd_name,
749 obd_uuid2str(&exp->exp_client_uuid),
750 obd_export_nid2str(exp),
751 hdl->cookie, conn->cookie);
752 memset(conn, 0, sizeof *conn);
753 /* target_handle_connect() treats EALREADY and
754 * -EALREADY differently. -EALREADY is an error
755 * (same UUID, different handle). */
759 if (!target->obd_recovering) {
760 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
761 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
762 obd_export_nid2str(exp));
763 GOTO(out_already, rc);
766 now = cfs_time_current();
767 deadline = cfs_timer_deadline(&target->obd_recovery_timer);
768 if (cfs_time_before(now, deadline)) {
769 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
770 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting,"
771 " waiting for %d clients in recovery for"
772 " %d:%.02d\n", target->obd_name,
773 obd_uuid2str(&exp->exp_client_uuid),
774 obd_export_nid2str(exp),
775 target->obd_max_recoverable_clients,
776 timeout / 60, timeout % 60);
778 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
779 LCONSOLE_WARN("%s: Recovery already passed deadline"
780 " %d:%.02d, It is most likely due to DNE"
781 " recovery is failed or stuck, please wait a"
782 " few more minutes or abort the recovery.\n",
783 target->obd_name, timeout / 60, timeout % 60);
787 conn->cookie = exp->exp_handle.h_cookie;
788 /* target_handle_connect() treats EALREADY and
789 * -EALREADY differently. EALREADY means we are
790 * doing a valid reconnect from the same client. */
794 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
797 struct obd_export *exp = cb_data;
799 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
800 obd->obd_name, exp->exp_client_uuid.uuid);
802 spin_lock(&exp->exp_lock);
803 exp->exp_need_sync = 0;
804 spin_unlock(&exp->exp_lock);
805 class_export_cb_put(exp);
809 check_and_start_recovery_timer(struct obd_device *obd,
810 struct ptlrpc_request *req, int new_client);
813 * update flags for import during reconnect process
815 static int rev_import_flags_update(struct obd_import *revimp,
816 struct ptlrpc_request *req)
819 struct obd_connect_data *data;
821 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
823 if (data->ocd_connect_flags & OBD_CONNECT_AT)
824 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
826 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
828 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
830 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
832 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
833 revimp->imp_client->cli_name,
834 libcfs_id2str(req->rq_peer), rc);
842 * Allocate a new reverse import for an export.
844 * \retval -errno in case error hit
845 * \retval 0 if reverse import correctly init
847 int rev_import_init(struct obd_export *export)
849 struct obd_device *obd = export->exp_obd;
850 struct obd_import *revimp;
852 LASSERT(export->exp_imp_reverse == NULL);
854 revimp = class_new_import(obd);
858 revimp->imp_remote_handle.cookie = 0ULL;
859 revimp->imp_client = &obd->obd_ldlm_client;
860 revimp->imp_dlm_fake = 1;
862 /* it is safe to connect import in new state as no sends possible */
863 spin_lock(&export->exp_lock);
864 export->exp_imp_reverse = revimp;
865 spin_unlock(&export->exp_lock);
866 class_import_put(revimp);
870 EXPORT_SYMBOL(rev_import_init);
873 * Handle reconnect for an export.
875 * \param exp export to handle reconnect process
876 * \param req client reconnect request
878 * \retval -rc in case securitfy flavor can't be changed
879 * \retval 0 in case none problems
881 static int rev_import_reconnect(struct obd_export *exp,
882 struct ptlrpc_request *req)
884 struct obd_import *revimp = exp->exp_imp_reverse;
885 struct lustre_handle *lh;
888 /* avoid sending a request until import flags are changed */
889 ptlrpc_import_enter_resend(revimp);
891 if (revimp->imp_connection != NULL)
892 ptlrpc_connection_put(revimp->imp_connection);
895 * client from recovery don't have a handle so we need to take from
896 * request. it may produce situation when wrong client connected
897 * to recovery as we trust a client uuid
899 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
900 revimp->imp_remote_handle = *lh;
902 /* unknown versions will be caught in
903 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
904 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
906 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
908 rc = rev_import_flags_update(revimp, req);
910 /* it is safe to still be in RECOVERY phase as we are not able
911 * to setup correct security flavor so requests are not able to
912 * be delivered correctly */
916 /* resend all rpc's via new connection */
917 return ptlrpc_import_recovery_state_machine(revimp);
920 int target_handle_connect(struct ptlrpc_request *req)
922 struct obd_device *target = NULL;
923 struct obd_export *export = NULL;
924 /* connect handle - filled from target_handle_reconnect in
926 struct lustre_handle conn;
927 struct lustre_handle *tmp;
928 struct obd_uuid tgtuuid;
929 struct obd_uuid cluuid;
934 bool mds_conn = false, lw_client = false;
935 bool mds_mds_conn = false;
936 bool new_mds_mds_conn = false;
937 bool target_referenced = false;
938 struct obd_connect_data *data, *tmpdata;
940 lnet_nid_t *client_nid = NULL;
943 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
945 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
947 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
948 GOTO(out, rc = -EINVAL);
951 obd_str2uuid(&tgtuuid, str);
952 target = class_uuid2obd(&tgtuuid);
954 target = class_name2obd(str);
957 deuuidify(str, NULL, &target_start, &target_len);
958 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
959 "from %s (no target). If you are running "
960 "an HA pair check that the target is "
961 "mounted on the other server.\n", str,
962 libcfs_nid2str(req->rq_peer.nid));
963 GOTO(out, rc = -ENODEV);
966 spin_lock(&target->obd_dev_lock);
967 if (target->obd_stopping || !target->obd_set_up) {
968 spin_unlock(&target->obd_dev_lock);
970 deuuidify(str, NULL, &target_start, &target_len);
971 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
972 target_len, target_start,
973 libcfs_nid2str(req->rq_peer.nid),
974 (target->obd_stopping ?
975 "stopping" : "not set up"));
976 GOTO(out, rc = -ENODEV);
979 if (target->obd_no_conn) {
980 spin_unlock(&target->obd_dev_lock);
982 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
983 "from %s\n", target->obd_name,
984 libcfs_nid2str(req->rq_peer.nid));
985 GOTO(out, rc = -EAGAIN);
988 /* Make sure the target isn't cleaned up while we're here. Yes,
989 * there's still a race between the above check and our incref here.
990 * Really, class_uuid2obd should take the ref. */
991 class_incref(target, __func__, current);
992 target_referenced = true;
994 target->obd_conn_inprogress++;
995 spin_unlock(&target->obd_dev_lock);
997 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
999 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1000 GOTO(out, rc = -EINVAL);
1003 obd_str2uuid(&cluuid, str);
1005 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1007 GOTO(out, rc = -EPROTO);
1011 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1013 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1015 GOTO(out, rc = -EPROTO);
1017 rc = req_capsule_server_pack(&req->rq_pill);
1021 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1022 /* Don't allow clients to connect that are using old 1.8 format
1023 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1024 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1025 * FULL20 flag should be set on all connections since 2.0, but no
1026 * longer affects behaviour.
1028 * Later this check will be disabled and the flag can be retired
1029 * completely once interop with 3.0 is no longer needed.
1031 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1032 GOTO(out, rc = -EPROTO);
1035 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1036 if (data->ocd_version < LUSTRE_VERSION_CODE -
1037 LUSTRE_VERSION_ALLOWED_OFFSET ||
1038 data->ocd_version > LUSTRE_VERSION_CODE +
1039 LUSTRE_VERSION_ALLOWED_OFFSET) {
1040 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1041 "libclient connection attempt",
1042 data->ocd_version < LUSTRE_VERSION_CODE ?
1044 OBD_OCD_VERSION_MAJOR(data->ocd_version),
1045 OBD_OCD_VERSION_MINOR(data->ocd_version),
1046 OBD_OCD_VERSION_PATCH(data->ocd_version),
1047 OBD_OCD_VERSION_FIX(data->ocd_version));
1048 data = req_capsule_server_sized_get(&req->rq_pill,
1050 offsetof(typeof(*data), ocd_version) +
1051 sizeof(data->ocd_version));
1053 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1054 data->ocd_version = LUSTRE_VERSION_CODE;
1056 GOTO(out, rc = -EPROTO);
1060 /* Note: lw_client is needed in MDS-MDS failover during update log
1061 * processing, so we needs to allow lw_client to be connected at
1062 * anytime, instead of only the initial connection */
1063 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1065 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1066 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1067 mds_mds_conn = (data->ocd_connect_flags &
1068 OBD_CONNECT_MDS_MDS) != 0;
1070 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1071 * for Imperative Recovery connection from MGC to MGS.
1073 * Via check OBD_CONNECT_FID, we can distinguish whether
1074 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1077 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1078 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1079 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1080 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1081 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1082 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1084 /* We do not support the MDT-MDT interoperations with
1085 * different version MDT because of protocol changes. */
1086 if (unlikely(major != LUSTRE_MAJOR ||
1087 minor != LUSTRE_MINOR ||
1088 abs(patch - LUSTRE_PATCH) > 3)) {
1089 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1090 "connection from different version MDT "
1091 "(%d.%d.%d.%d) %s %s\n",
1092 target->obd_name, LUSTRE_MAJOR,
1093 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1094 major, minor, patch,
1095 OBD_OCD_VERSION_FIX(data->ocd_version),
1096 libcfs_nid2str(req->rq_peer.nid), str);
1098 GOTO(out, rc = -EPROTO);
1103 /* lctl gets a backstage, all-access pass. */
1104 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1105 goto dont_check_exports;
1107 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1111 /* We've found an export in the hash. */
1113 spin_lock(&export->exp_lock);
1115 if (export->exp_connecting) { /* bug 9635, et. al. */
1116 spin_unlock(&export->exp_lock);
1117 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1118 export->exp_obd->obd_name, export,
1119 libcfs_nid2str(req->rq_peer.nid));
1120 class_export_put(export);
1123 } else if ((mds_conn || lw_client ||
1124 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1125 export->exp_connection != NULL) {
1126 spin_unlock(&export->exp_lock);
1127 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1128 /* MDS or LWP reconnected after failover. */
1129 LCONSOLE_WARN("%s: Received %s connection from "
1130 "%s, removing former export from %s\n",
1131 target->obd_name, mds_conn ? "MDS" : "LWP",
1132 libcfs_nid2str(req->rq_peer.nid),
1133 libcfs_nid2str(export->exp_connection->c_peer.nid));
1135 /* New MDS connection from the same NID. */
1136 LCONSOLE_WARN("%s: Received new %s connection from "
1137 "%s, removing former export from same NID\n",
1138 target->obd_name, mds_conn ? "MDS" : "LWP",
1139 libcfs_nid2str(req->rq_peer.nid));
1142 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1143 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1144 /* Because exports between MDTs will always be
1145 * kept, let's do not fail such export if they
1146 * come from the same NID, otherwise it might
1147 * cause eviction between MDTs, which might
1148 * cause namespace inconsistency */
1149 spin_lock(&export->exp_lock);
1150 export->exp_connecting = 1;
1151 spin_unlock(&export->exp_lock);
1152 conn.cookie = export->exp_handle.h_cookie;
1155 class_fail_export(export);
1156 class_export_put(export);
1160 } else if (export->exp_connection != NULL &&
1161 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1162 (lustre_msg_get_op_flags(req->rq_reqmsg) &
1163 MSG_CONNECT_INITIAL)) {
1164 spin_unlock(&export->exp_lock);
1165 /* In MDS failover we have static UUID but NID can change. */
1166 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1167 "existing nid %s is already connected\n",
1168 target->obd_name, cluuid.uuid,
1169 libcfs_nid2str(req->rq_peer.nid),
1171 export->exp_connection->c_peer.nid));
1173 class_export_put(export);
1176 export->exp_connecting = 1;
1177 spin_unlock(&export->exp_lock);
1178 LASSERT(export->exp_obd == target);
1180 rc = target_handle_reconnect(&conn, export, &cluuid);
1183 /* If we found an export, we already unlocked. */
1186 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1187 } else if (req->rq_export == NULL &&
1188 atomic_read(&export->exp_rpc_count) > 0) {
1189 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1190 "still busy with %d references\n",
1191 target->obd_name, cluuid.uuid,
1192 libcfs_nid2str(req->rq_peer.nid),
1193 atomic_read(&export->exp_refcount));
1194 GOTO(out, rc = -EBUSY);
1195 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1196 if (!strstr(cluuid.uuid, "mdt"))
1197 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1198 "known client %s (at %s) because it "
1199 "is indicating it is a new client",
1200 target->obd_name, cluuid.uuid,
1201 libcfs_nid2str(req->rq_peer.nid));
1202 GOTO(out, rc = -EALREADY);
1204 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1211 CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1212 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1213 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1214 export, (long)cfs_time_current_sec(),
1215 export ? (long)export->exp_last_request_time : 0);
1217 /* If this is the first time a client connects, reset the recovery
1218 * timer. Discard lightweight connections which might be local. */
1219 if (!lw_client && rc == 0 && target->obd_recovering)
1220 check_and_start_recovery_timer(target, req, export == NULL);
1222 /* We want to handle EALREADY but *not* -EALREADY from
1223 * target_handle_reconnect(), return reconnection state in a flag. */
1224 if (rc == EALREADY) {
1225 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1231 /* Tell the client if we support replayable requests. */
1232 if (target->obd_replayable)
1233 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1234 client_nid = &req->rq_peer.nid;
1236 if (export == NULL) {
1237 /* allow lightweight connections during recovery */
1238 /* allow "new" MDT to be connected during recovery, since we
1239 * need retrieve recovery update records from it */
1240 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1242 int c; /* connected */
1243 int i; /* in progress */
1245 int s; /* stale/evicted */
1247 c = atomic_read(&target->obd_connected_clients);
1248 i = atomic_read(&target->obd_lock_replay_clients);
1249 k = target->obd_max_recoverable_clients;
1250 s = target->obd_stale_clients;
1251 t = cfs_timer_deadline(&target->obd_recovery_timer);
1252 t = cfs_time_sub(t, cfs_time_current());
1253 t = cfs_duration_sec(t);
1254 LCONSOLE_WARN("%s: Denying connection for new client %s"
1255 "(at %s), waiting for %d known clients "
1256 "(%d recovered, %d in progress, and %d "
1257 "evicted) to recover in %d:%.02d\n",
1258 target->obd_name, cluuid.uuid,
1259 libcfs_nid2str(req->rq_peer.nid), k,
1260 c - i, i, s, (int)t / 60,
1265 rc = obd_connect(req->rq_svc_thread->t_env,
1266 &export, target, &cluuid, data,
1268 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1269 lustre_msg_add_op_flags(req->rq_repmsg,
1270 MSG_CONNECT_RECOVERING);
1272 conn.cookie = export->exp_handle.h_cookie;
1273 rc = rev_import_init(export);
1277 new_mds_mds_conn = true;
1280 rc = obd_reconnect(req->rq_svc_thread->t_env,
1281 export, target, &cluuid, data, client_nid);
1286 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1287 data->ocd_instance = target->u.obt.obt_instance;
1289 /* Return only the parts of obd_connect_data that we understand, so the
1290 * client knows that we don't understand the rest. */
1292 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1294 tmpdata = req_capsule_server_get(&req->rq_pill,
1296 /* Don't use struct assignment here, because the client reply
1297 * buffer may be smaller/larger than the local struct
1298 * obd_connect_data. */
1299 memcpy(tmpdata, data, min(tmpsize, size));
1302 /* If the client and the server are the same node, we will already
1303 * have an export that really points to the client's DLM export,
1304 * because we have a shared handles table.
1306 * XXX this will go away when shaver stops sending the "connect" handle
1307 * in the real "remote handle" field of the request --phik 24 Apr 2003
1309 ptlrpc_request_change_export(req, export);
1311 spin_lock(&export->exp_lock);
1312 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1313 spin_unlock(&export->exp_lock);
1314 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1315 "or equal conn_cnt: %d >= %d\n",
1316 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1317 export->exp_conn_cnt,
1318 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1320 GOTO(out, rc = -EALREADY);
1322 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1323 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1325 /* Don't evict liblustre clients for not pinging. */
1326 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1327 export->exp_libclient = 1;
1328 spin_unlock(&export->exp_lock);
1330 spin_lock(&target->obd_dev_lock);
1331 list_del_init(&export->exp_obd_chain_timed);
1332 spin_unlock(&target->obd_dev_lock);
1334 spin_unlock(&export->exp_lock);
1337 if (export->exp_connection != NULL) {
1338 /* Check to see if connection came from another NID. */
1339 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1340 !hlist_unhashed(&export->exp_nid_hash))
1341 cfs_hash_del(export->exp_obd->obd_nid_hash,
1342 &export->exp_connection->c_peer.nid,
1343 &export->exp_nid_hash);
1345 ptlrpc_connection_put(export->exp_connection);
1348 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1351 if (hlist_unhashed(&export->exp_nid_hash))
1352 cfs_hash_add(export->exp_obd->obd_nid_hash,
1353 &export->exp_connection->c_peer.nid,
1354 &export->exp_nid_hash);
1356 lustre_msg_set_handle(req->rq_repmsg, &conn);
1358 rc = rev_import_reconnect(export, req);
1362 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1364 __u64 transno = data->ocd_transno;
1366 spin_lock(&export->exp_lock);
1367 /* possible race with class_disconnect_stale_exports,
1368 * export may be already in the eviction process */
1369 if (export->exp_failed) {
1370 spin_unlock(&export->exp_lock);
1371 GOTO(out, rc = -ENODEV);
1373 export->exp_in_recovery = 1;
1374 export->exp_req_replay_needed = 1;
1375 export->exp_lock_replay_needed = 1;
1376 spin_unlock(&export->exp_lock);
1378 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1379 MSG_CONNECT_TRANSNO);
1380 if (has_transno && transno == 0)
1381 CWARN("Connect with zero transno!\n");
1383 if (has_transno && transno > 0 &&
1384 transno < target->obd_next_recovery_transno &&
1385 transno > target->obd_last_committed) {
1386 /* Another way is to use cmpxchg() to be lock-free. */
1387 spin_lock(&target->obd_recovery_task_lock);
1388 if (transno < target->obd_next_recovery_transno)
1389 target->obd_next_recovery_transno = transno;
1390 spin_unlock(&target->obd_recovery_task_lock);
1393 atomic_inc(&target->obd_req_replay_clients);
1394 atomic_inc(&target->obd_lock_replay_clients);
1395 /* Note: MDS-MDS connection is allowed to be connected during
1396 * recovery, no matter if the exports needs to be recoveried.
1397 * Because we need retrieve updates logs from all other MDTs.
1398 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1399 * also needs to be increased to match other recovery checking
1401 if (new_mds_mds_conn)
1402 target->obd_max_recoverable_clients++;
1403 if (atomic_inc_return(&target->obd_connected_clients) ==
1404 target->obd_max_recoverable_clients)
1405 wake_up(&target->obd_next_transno_waitq);
1408 /* Tell the client we're in recovery, when client is involved in it. */
1409 if (target->obd_recovering && !lw_client)
1410 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1414 spin_lock(&export->exp_lock);
1415 export->exp_connecting = 0;
1416 spin_unlock(&export->exp_lock);
1418 class_export_put(export);
1420 if (target_referenced == true && target != NULL) {
1421 spin_lock(&target->obd_dev_lock);
1422 target->obd_conn_inprogress--;
1423 spin_unlock(&target->obd_dev_lock);
1425 class_decref(target, __func__, current);
1427 req->rq_status = rc;
1431 int target_handle_disconnect(struct ptlrpc_request *req)
1436 rc = req_capsule_server_pack(&req->rq_pill);
1440 /* Keep the rq_export around so we can send the reply. */
1441 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1446 void target_destroy_export(struct obd_export *exp)
1448 struct obd_import *imp = NULL;
1449 /* exports created from last_rcvd data, and "fake"
1450 exports created by lctl don't have an import */
1451 spin_lock(&exp->exp_lock);
1452 if (exp->exp_imp_reverse != NULL) {
1453 imp = exp->exp_imp_reverse;
1454 exp->exp_imp_reverse = NULL;
1456 spin_unlock(&exp->exp_lock);
1458 client_destroy_import(imp);
1460 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1461 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1462 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1463 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1465 EXPORT_SYMBOL(target_destroy_export);
1468 * Recovery functions
1470 static void target_request_copy_get(struct ptlrpc_request *req)
1472 class_export_rpc_inc(req->rq_export);
1473 LASSERT(list_empty(&req->rq_list));
1474 INIT_LIST_HEAD(&req->rq_replay_list);
1476 /* Increase refcount to keep request in queue. */
1477 atomic_inc(&req->rq_refcount);
1478 /* Let export know it has replays to be handled. */
1479 atomic_inc(&req->rq_export->exp_replay_count);
1482 static void target_request_copy_put(struct ptlrpc_request *req)
1484 LASSERT(list_empty(&req->rq_replay_list));
1485 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1487 atomic_dec(&req->rq_export->exp_replay_count);
1488 class_export_rpc_dec(req->rq_export);
1489 ptlrpc_server_drop_request(req);
1492 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1494 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1495 struct obd_export *exp = req->rq_export;
1496 struct ptlrpc_request *reqiter;
1497 struct ptlrpc_request *dup_req = NULL;
1502 spin_lock(&exp->exp_lock);
1503 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1505 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1513 /* We expect it with RESENT and REPLAY flags. */
1514 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1515 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1516 CERROR("invalid flags %x of resent replay\n",
1517 lustre_msg_get_flags(req->rq_reqmsg));
1519 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1522 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1524 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1525 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1529 list_add_tail(&req->rq_replay_list,
1530 &exp->exp_req_replay_queue);
1533 spin_unlock(&exp->exp_lock);
1537 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1539 LASSERT(!list_empty(&req->rq_replay_list));
1540 LASSERT(req->rq_export);
1542 spin_lock(&req->rq_export->exp_lock);
1543 list_del_init(&req->rq_replay_list);
1544 spin_unlock(&req->rq_export->exp_lock);
1547 static void target_finish_recovery(struct lu_target *lut)
1549 struct obd_device *obd = lut->lut_obd;
1552 /* Only log a recovery message when recovery has occurred. */
1553 if (obd->obd_recovery_start) {
1554 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1555 obd->obd_recovery_start);
1556 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1557 "%d recovered and %d %s evicted.\n", obd->obd_name,
1558 (int)elapsed_time / 60, (int)elapsed_time % 60,
1559 obd->obd_max_recoverable_clients,
1560 atomic_read(&obd->obd_connected_clients),
1561 obd->obd_stale_clients,
1562 obd->obd_stale_clients == 1 ? "was" : "were");
1565 ldlm_reprocess_all_ns(obd->obd_namespace);
1566 spin_lock(&obd->obd_recovery_task_lock);
1567 if (!list_empty(&obd->obd_req_replay_queue) ||
1568 !list_empty(&obd->obd_lock_replay_queue) ||
1569 !list_empty(&obd->obd_final_req_queue)) {
1570 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1572 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1573 list_empty(&obd->obd_lock_replay_queue) ? \
1575 list_empty(&obd->obd_final_req_queue) ? \
1577 spin_unlock(&obd->obd_recovery_task_lock);
1580 spin_unlock(&obd->obd_recovery_task_lock);
1582 obd->obd_recovery_end = cfs_time_current_sec();
1584 /* When recovery finished, cleanup orphans on MDS and OST. */
1585 if (OBT(obd) && OBP(obd, postrecov)) {
1586 int rc = OBP(obd, postrecov)(obd);
1588 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1594 static void abort_req_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_req_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_WARNING, req, "aborted:");
1605 req->rq_status = -ENOTCONN;
1606 if (ptlrpc_error(req)) {
1607 DEBUG_REQ(D_ERROR, req,
1608 "failed abort_req_reply; skipping");
1610 target_exp_dequeue_req_replay(req);
1611 target_request_copy_put(req);
1615 static void abort_lock_replay_queue(struct obd_device *obd)
1617 struct ptlrpc_request *req, *n;
1618 struct list_head abort_list;
1620 INIT_LIST_HEAD(&abort_list);
1621 spin_lock(&obd->obd_recovery_task_lock);
1622 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1623 spin_unlock(&obd->obd_recovery_task_lock);
1624 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1625 DEBUG_REQ(D_ERROR, req, "aborted:");
1626 req->rq_status = -ENOTCONN;
1627 if (ptlrpc_error(req)) {
1628 DEBUG_REQ(D_ERROR, req,
1629 "failed abort_lock_reply; skipping");
1631 target_request_copy_put(req);
1635 /* Called from a cleanup function if the device is being cleaned up
1636 forcefully. The exports should all have been disconnected already,
1637 the only thing left to do is
1638 - clear the recovery flags
1640 - free queued requests and replies, but don't send replies
1641 Because the obd_stopping flag is set, no new requests should be received.
1644 void target_cleanup_recovery(struct obd_device *obd)
1646 struct ptlrpc_request *req, *n;
1647 struct list_head clean_list;
1649 INIT_LIST_HEAD(&clean_list);
1650 spin_lock(&obd->obd_dev_lock);
1651 if (!obd->obd_recovering) {
1652 spin_unlock(&obd->obd_dev_lock);
1656 obd->obd_recovering = obd->obd_abort_recovery = 0;
1657 spin_unlock(&obd->obd_dev_lock);
1659 spin_lock(&obd->obd_recovery_task_lock);
1660 target_cancel_recovery_timer(obd);
1661 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1662 spin_unlock(&obd->obd_recovery_task_lock);
1664 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1665 LASSERT(req->rq_reply_state == NULL);
1666 target_exp_dequeue_req_replay(req);
1667 target_request_copy_put(req);
1670 spin_lock(&obd->obd_recovery_task_lock);
1671 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1672 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1673 spin_unlock(&obd->obd_recovery_task_lock);
1675 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1676 LASSERT(req->rq_reply_state == NULL);
1677 target_request_copy_put(req);
1682 EXPORT_SYMBOL(target_cleanup_recovery);
1684 /* obd_recovery_task_lock should be held */
1685 void target_cancel_recovery_timer(struct obd_device *obd)
1687 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1688 cfs_timer_disarm(&obd->obd_recovery_timer);
1691 static void target_start_recovery_timer(struct obd_device *obd)
1693 if (obd->obd_recovery_start != 0)
1696 spin_lock(&obd->obd_dev_lock);
1697 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1698 spin_unlock(&obd->obd_dev_lock);
1702 LASSERT(obd->obd_recovery_timeout != 0);
1704 if (obd->obd_recovery_start != 0) {
1705 spin_unlock(&obd->obd_dev_lock);
1709 cfs_timer_arm(&obd->obd_recovery_timer,
1710 cfs_time_shift(obd->obd_recovery_timeout));
1711 obd->obd_recovery_start = cfs_time_current_sec();
1712 spin_unlock(&obd->obd_dev_lock);
1714 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1715 "or until %d client%s reconnect%s\n",
1717 obd->obd_recovery_timeout / 60,
1718 obd->obd_recovery_timeout % 60,
1719 obd->obd_max_recoverable_clients,
1720 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1721 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1725 * extend recovery window.
1727 * if @extend is true, extend recovery window to have @drt remaining at least;
1728 * otherwise, make sure the recovery timeout value is not less than @drt.
1730 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1734 cfs_duration_t left;
1737 spin_lock(&obd->obd_dev_lock);
1738 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1739 spin_unlock(&obd->obd_dev_lock);
1742 LASSERT(obd->obd_recovery_start != 0);
1744 now = cfs_time_current_sec();
1745 to = obd->obd_recovery_timeout;
1746 end = obd->obd_recovery_start + to;
1747 left = cfs_time_sub(end, now);
1749 if (extend && (drt > left)) {
1751 } else if (!extend && (drt > to)) {
1755 if (to > obd->obd_recovery_time_hard)
1756 to = obd->obd_recovery_time_hard;
1757 if (obd->obd_recovery_timeout < to) {
1758 obd->obd_recovery_timeout = to;
1759 end = obd->obd_recovery_start + to;
1760 cfs_timer_arm(&obd->obd_recovery_timer,
1761 cfs_time_shift(end - now));
1763 spin_unlock(&obd->obd_dev_lock);
1765 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1766 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1769 /* Reset the timer with each new client connection */
1771 * This timer is actually reconnect_timer, which is for making sure
1772 * the total recovery window is at least as big as my reconnect
1773 * attempt timing. So the initial recovery time_out will be set to
1774 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1775 * from client is bigger than this, then the recovery time_out will
1776 * be extended to make sure the client could be reconnected, in the
1777 * process, the timeout from the new client should be ignored.
1781 check_and_start_recovery_timer(struct obd_device *obd,
1782 struct ptlrpc_request *req,
1785 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1786 struct obd_device_target *obt = &obd->u.obt;
1788 if (!new_client && service_time)
1789 /* Teach server about old server's estimates, as first guess
1790 * at how long new requests will take. */
1791 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1794 target_start_recovery_timer(obd);
1796 /* Convert the service time to RPC timeout,
1797 * and reuse service_time to limit stack usage. */
1798 service_time = at_est2timeout(service_time);
1800 /* We expect other clients to timeout within service_time, then try
1801 * to reconnect, then try the failover server. The max delay between
1802 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1803 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1805 LASSERT(obt->obt_magic == OBT_MAGIC);
1806 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1807 if (service_time > obd->obd_recovery_timeout && !new_client)
1808 extend_recovery_timer(obd, service_time, false);
1811 /** Health checking routines */
1812 static inline int exp_connect_healthy(struct obd_export *exp)
1814 return (exp->exp_in_recovery);
1817 /** if export done req_replay or has replay in queue */
1818 static inline int exp_req_replay_healthy(struct obd_export *exp)
1820 return (!exp->exp_req_replay_needed ||
1821 atomic_read(&exp->exp_replay_count) > 0);
1825 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1827 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1828 exp_req_replay_healthy(exp);
1831 /** if export done lock_replay or has replay in queue */
1832 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1834 return (!exp->exp_lock_replay_needed ||
1835 atomic_read(&exp->exp_replay_count) > 0);
1838 static inline int exp_vbr_healthy(struct obd_export *exp)
1840 return (!exp->exp_vbr_failed);
1843 static inline int exp_finished(struct obd_export *exp)
1845 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1848 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1850 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1854 static int check_for_next_transno(struct lu_target *lut)
1856 struct ptlrpc_request *req = NULL;
1857 struct obd_device *obd = lut->lut_obd;
1858 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1859 int wake_up = 0, connected, completed, queue_len;
1860 __u64 req_transno = 0;
1861 __u64 update_transno = 0;
1862 __u64 next_transno = 0;
1865 spin_lock(&obd->obd_recovery_task_lock);
1866 if (!list_empty(&obd->obd_req_replay_queue)) {
1867 req = list_entry(obd->obd_req_replay_queue.next,
1868 struct ptlrpc_request, rq_list);
1869 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1873 update_transno = distribute_txn_get_next_transno(tdtd);
1875 connected = atomic_read(&obd->obd_connected_clients);
1876 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1877 queue_len = obd->obd_requests_queued_for_recovery;
1878 next_transno = obd->obd_next_recovery_transno;
1880 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1881 "req_transno: "LPU64", next_transno: "LPU64"\n",
1882 obd->obd_max_recoverable_clients, connected, completed,
1883 queue_len, req_transno, next_transno);
1885 if (obd->obd_abort_recovery) {
1886 CDEBUG(D_HA, "waking for aborted recovery\n");
1888 } else if (obd->obd_recovery_expired) {
1889 CDEBUG(D_HA, "waking for expired recovery\n");
1891 } else if (tdtd != NULL && req != NULL &&
1892 is_req_replayed_by_update(req)) {
1893 LASSERTF(req_transno < next_transno, "req_transno "LPU64
1894 "next_transno"LPU64"\n", req_transno, next_transno);
1895 CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n",
1898 } else if (req_transno == next_transno ||
1899 (update_transno != 0 && update_transno <= next_transno)) {
1900 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1902 } else if (queue_len > 0 &&
1903 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1905 /** handle gaps occured due to lost reply or VBR */
1906 LASSERTF(req_transno >= next_transno,
1907 "req_transno: "LPU64", next_transno: "LPU64"\n",
1908 req_transno, next_transno);
1909 if (req_transno > obd->obd_last_committed &&
1910 !obd->obd_version_recov)
1913 "%s: waking for gap in transno, VBR is %s (skip: "
1914 LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1915 ", next_update "LPD64" last_committed: "LPD64")\n",
1916 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1917 next_transno, queue_len, completed, connected,
1918 req_transno, update_transno, obd->obd_last_committed);
1919 obd->obd_next_recovery_transno = req_transno;
1921 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1922 CDEBUG(D_HA, "waking for completed recovery\n");
1924 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1925 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1926 " by fail_lock, waking up ("LPD64")\n", next_transno);
1927 obd->obd_next_recovery_transno = req_transno;
1930 spin_unlock(&obd->obd_recovery_task_lock);
1934 static int check_for_next_lock(struct lu_target *lut)
1936 struct obd_device *obd = lut->lut_obd;
1939 spin_lock(&obd->obd_recovery_task_lock);
1940 if (!list_empty(&obd->obd_lock_replay_queue)) {
1941 CDEBUG(D_HA, "waking for next lock\n");
1943 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1944 CDEBUG(D_HA, "waking for completed lock replay\n");
1946 } else if (obd->obd_abort_recovery) {
1947 CDEBUG(D_HA, "waking for aborted recovery\n");
1949 } else if (obd->obd_recovery_expired) {
1950 CDEBUG(D_HA, "waking for expired recovery\n");
1953 spin_unlock(&obd->obd_recovery_task_lock);
1959 * wait for recovery events,
1960 * check its status with help of check_routine
1961 * evict dead clients via health_check
1963 static int target_recovery_overseer(struct lu_target *lut,
1964 int (*check_routine)(struct lu_target *),
1965 int (*health_check)(struct obd_export *))
1967 struct obd_device *obd = lut->lut_obd;
1968 struct target_distribute_txn_data *tdtd;
1970 if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1971 (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1972 __u64 next_update_transno = 0;
1974 /* Only abort the recovery if there are no update recovery
1975 * left in the queue */
1976 spin_lock(&obd->obd_recovery_task_lock);
1977 if (lut->lut_tdtd != NULL) {
1978 next_update_transno =
1979 distribute_txn_get_next_transno(lut->lut_tdtd);
1981 tdtd = lut->lut_tdtd;
1982 /* If next_update_transno == 0, it probably because
1983 * updatelog retrieve threads did not get any records
1984 * yet, let's wait those threads stopped */
1985 if (next_update_transno == 0) {
1986 struct l_wait_info lwi = { 0 };
1988 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1990 &tdtd->tdtd_recovery_threads_count) == 0,
1993 next_update_transno =
1994 distribute_txn_get_next_transno(
1999 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2000 obd->obd_next_recovery_transno = next_update_transno;
2001 spin_unlock(&obd->obd_recovery_task_lock);
2002 /* Disconnect unfinished exports from clients, and
2003 * keep connection from MDT to make sure the update
2004 * recovery will still keep trying until some one
2005 * manually abort the recovery */
2006 class_disconnect_stale_exports(obd,
2007 exp_finished_or_from_mdt);
2008 /* Abort all of replay and replay lock req from
2010 abort_req_replay_queue(obd);
2011 abort_lock_replay_queue(obd);
2012 CDEBUG(D_HA, "%s: there are still update replay ("LPX64
2013 ")in the queue.\n", obd->obd_name,
2014 next_update_transno);
2016 obd->obd_abort_recovery = 1;
2017 spin_unlock(&obd->obd_recovery_task_lock);
2018 CWARN("%s recovery is aborted by hard timeout\n",
2023 while (wait_event_timeout(obd->obd_next_transno_waitq,
2025 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2026 /* wait indefinitely for event, but don't trigger watchdog */;
2028 if (obd->obd_abort_recovery) {
2029 CWARN("recovery is aborted, evict exports in recovery\n");
2030 if (lut->lut_tdtd != NULL) {
2031 struct l_wait_info lwi = { 0 };
2033 tdtd = lut->lut_tdtd;
2034 /* Let's wait all of the update log recovery thread
2036 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2037 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2039 /* Then abort the update recovery list */
2040 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2041 dtrq_list_destroy(lut->lut_tdtd);
2044 /** evict exports which didn't finish recovery yet */
2045 class_disconnect_stale_exports(obd, exp_finished);
2047 } else if (obd->obd_recovery_expired) {
2048 obd->obd_recovery_expired = 0;
2049 /** If some clients died being recovered, evict them */
2050 LCONSOLE_WARN("%s: recovery is timed out, "
2051 "evict stale exports\n", obd->obd_name);
2052 /** evict cexports with no replay in queue, they are stalled */
2053 class_disconnect_stale_exports(obd, health_check);
2055 /** continue with VBR */
2056 spin_lock(&obd->obd_dev_lock);
2057 obd->obd_version_recov = 1;
2058 spin_unlock(&obd->obd_dev_lock);
2060 * reset timer, recovery will proceed with versions now,
2061 * timeout is set just to handle reconnection delays
2063 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2064 /** Wait for recovery events again, after evicting bad clients */
2070 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2072 struct obd_device *obd = lut->lut_obd;
2073 struct ptlrpc_request *req = NULL;
2075 CDEBUG(D_HA, "Waiting for lock\n");
2076 if (target_recovery_overseer(lut, check_for_next_lock,
2077 exp_lock_replay_healthy))
2078 abort_lock_replay_queue(obd);
2080 spin_lock(&obd->obd_recovery_task_lock);
2081 if (!list_empty(&obd->obd_lock_replay_queue)) {
2082 req = list_entry(obd->obd_lock_replay_queue.next,
2083 struct ptlrpc_request, rq_list);
2084 list_del_init(&req->rq_list);
2085 spin_unlock(&obd->obd_recovery_task_lock);
2087 spin_unlock(&obd->obd_recovery_task_lock);
2088 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2089 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2090 /** evict exports failed VBR */
2091 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2096 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2098 struct ptlrpc_request *req = NULL;
2100 spin_lock(&obd->obd_recovery_task_lock);
2101 if (!list_empty(&obd->obd_final_req_queue)) {
2102 req = list_entry(obd->obd_final_req_queue.next,
2103 struct ptlrpc_request, rq_list);
2104 list_del_init(&req->rq_list);
2105 spin_unlock(&obd->obd_recovery_task_lock);
2106 if (req->rq_export->exp_in_recovery) {
2107 spin_lock(&req->rq_export->exp_lock);
2108 req->rq_export->exp_in_recovery = 0;
2109 spin_unlock(&req->rq_export->exp_lock);
2112 spin_unlock(&obd->obd_recovery_task_lock);
2117 static void handle_recovery_req(struct ptlrpc_thread *thread,
2118 struct ptlrpc_request *req,
2119 svc_handler_t handler)
2124 * export can be evicted during recovery, no need to handle replays for
2125 * it after that, discard such request silently
2127 if (req->rq_export->exp_disconnected)
2130 req->rq_session.lc_thread = thread;
2131 req->rq_svc_thread = thread;
2132 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2134 /* thread context */
2135 lu_context_enter(&thread->t_env->le_ctx);
2137 lu_context_exit(&thread->t_env->le_ctx);
2139 /* don't reset timer for final stage */
2140 if (!exp_finished(req->rq_export)) {
2141 int to = obd_timeout;
2144 * Add request timeout to the recovery time so next request from
2145 * this client may come in recovery time
2148 struct ptlrpc_service_part *svcpt;
2150 svcpt = req->rq_rqbd->rqbd_svcpt;
2151 /* If the server sent early reply for this request,
2152 * the client will recalculate the timeout according to
2153 * current server estimate service time, so we will
2154 * use the maxium timeout here for waiting the client
2155 * sending the next req */
2156 to = max((int)at_est2timeout(
2157 at_get(&svcpt->scp_at_estimate)),
2158 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2159 /* Add 2 net_latency, one for balance rq_deadline
2160 * (see ptl_send_rpc), one for resend the req to server,
2161 * Note: client will pack net_latency in replay req
2162 * (see ptlrpc_replay_req) */
2163 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2165 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2170 /** Checking routines for recovery */
2171 static int check_for_recovery_ready(struct lu_target *lut)
2173 struct obd_device *obd = lut->lut_obd;
2174 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2176 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2177 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2178 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2179 obd->obd_recovery_expired);
2181 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2182 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2183 if (clnts + obd->obd_stale_clients <
2184 obd->obd_max_recoverable_clients)
2188 if (lut->lut_tdtd != NULL) {
2189 if (!lut->lut_tdtd->tdtd_replay_ready &&
2190 !obd->obd_abort_recovery) {
2191 /* Let's extend recovery timer, in case the recovery
2192 * timer expired, and some clients got evicted */
2193 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2195 CDEBUG(D_HA, "%s update recovery is not ready,"
2196 " extend recovery %d\n", obd->obd_name,
2197 obd->obd_recovery_timeout);
2200 dtrq_list_dump(lut->lut_tdtd, D_HA);
2208 REQUEST_RECOVERY = 1,
2209 UPDATE_RECOVERY = 2,
2212 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2216 if (!list_empty(&obd->obd_req_replay_queue)) {
2217 struct ptlrpc_request *req;
2219 req = list_entry(obd->obd_req_replay_queue.next,
2220 struct ptlrpc_request, rq_list);
2221 transno = lustre_msg_get_transno(req->rq_reqmsg);
2227 static __u64 get_next_transno(struct lu_target *lut, int *type)
2229 struct obd_device *obd = lut->lut_obd;
2230 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2232 __u64 update_transno;
2235 transno = get_next_replay_req_transno(obd);
2237 *type = REQUEST_RECOVERY;
2242 update_transno = distribute_txn_get_next_transno(tdtd);
2243 if (transno == 0 || (transno >= update_transno &&
2244 update_transno != 0)) {
2245 transno = update_transno;
2247 *type = UPDATE_RECOVERY;
2254 * drop duplicate replay request
2256 * Because the operation has been replayed by update recovery, the request
2257 * with the same transno will be dropped and also notify the client to send
2258 * next replay request.
2260 * \param[in] env execution environment
2261 * \param[in] obd failover obd device
2262 * \param[in] req request to be dropped
2264 static void drop_duplicate_replay_req(struct lu_env *env,
2265 struct obd_device *obd,
2266 struct ptlrpc_request *req)
2268 DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2269 " update records are found.\n",
2270 lustre_msg_get_transno(req->rq_reqmsg),
2271 libcfs_nid2str(req->rq_peer.nid));
2273 /* Right now, only for MDS reint operation update replay and
2274 * normal request replay can have the same transno */
2275 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2276 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2277 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2278 if (likely(req->rq_export))
2279 target_committed_to_req(req);
2280 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2281 target_send_reply(req, req->rq_status, 0);
2283 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2284 libcfs_nid2str(req->rq_peer.nid));
2286 target_exp_dequeue_req_replay(req);
2287 target_request_copy_put(req);
2288 obd->obd_replayed_requests++;
2291 static void replay_request_or_update(struct lu_env *env,
2292 struct lu_target *lut,
2293 struct target_recovery_data *trd,
2294 struct ptlrpc_thread *thread)
2296 struct obd_device *obd = lut->lut_obd;
2297 struct ptlrpc_request *req = NULL;
2302 CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2303 obd->obd_next_recovery_transno);
2305 /* Replay all of request and update by transno */
2307 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2309 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2311 /** It is needed to extend recovery window above
2312 * recovery_time_soft. Extending is possible only in the
2313 * end of recovery window (see more details in
2314 * handle_recovery_req()).
2316 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2318 if (target_recovery_overseer(lut, check_for_next_transno,
2319 exp_req_replay_healthy_or_from_mdt)) {
2320 abort_req_replay_queue(obd);
2321 abort_lock_replay_queue(obd);
2325 spin_lock(&obd->obd_recovery_task_lock);
2326 transno = get_next_transno(lut, &type);
2327 if (type == REQUEST_RECOVERY && transno != 0) {
2328 /* Drop replay request from client side, if the
2329 * replay has been executed by update with the
2331 req = list_entry(obd->obd_req_replay_queue.next,
2332 struct ptlrpc_request, rq_list);
2334 list_del_init(&req->rq_list);
2335 obd->obd_requests_queued_for_recovery--;
2336 spin_unlock(&obd->obd_recovery_task_lock);
2338 /* Let's check if the request has been redone by
2340 if (is_req_replayed_by_update(req)) {
2341 struct distribute_txn_replay_req *dtrq;
2343 dtrq = distribute_txn_lookup_finish_list(tdtd,
2345 LASSERT(dtrq != NULL);
2346 spin_lock(&tdtd->tdtd_replay_list_lock);
2347 list_del_init(&dtrq->dtrq_list);
2348 spin_unlock(&tdtd->tdtd_replay_list_lock);
2351 drop_duplicate_replay_req(env, obd, req);
2356 LASSERT(trd->trd_processing_task == current_pid());
2357 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2358 lustre_msg_get_transno(req->rq_reqmsg),
2359 libcfs_nid2str(req->rq_peer.nid));
2361 handle_recovery_req(thread, req,
2362 trd->trd_recovery_handler);
2364 * bz18031: increase next_recovery_transno before
2365 * target_request_copy_put() will drop exp_rpc reference
2367 spin_lock(&obd->obd_recovery_task_lock);
2368 obd->obd_next_recovery_transno++;
2369 spin_unlock(&obd->obd_recovery_task_lock);
2370 target_exp_dequeue_req_replay(req);
2371 target_request_copy_put(req);
2372 obd->obd_replayed_requests++;
2373 } else if (type == UPDATE_RECOVERY && transno != 0) {
2374 struct distribute_txn_replay_req *dtrq;
2377 spin_unlock(&obd->obd_recovery_task_lock);
2379 LASSERT(tdtd != NULL);
2380 dtrq = distribute_txn_get_next_req(tdtd);
2381 lu_context_enter(&thread->t_env->le_ctx);
2382 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2383 lu_context_exit(&thread->t_env->le_ctx);
2384 extend_recovery_timer(obd, obd_timeout, true);
2386 if (rc == 0 && dtrq->dtrq_xid != 0) {
2387 CDEBUG(D_HA, "Move x"LPU64" t"LPU64
2388 " to finish list\n", dtrq->dtrq_xid,
2389 dtrq->dtrq_master_transno);
2391 /* Add it to the replay finish list */
2392 spin_lock(&tdtd->tdtd_replay_list_lock);
2393 list_add(&dtrq->dtrq_list,
2394 &tdtd->tdtd_replay_finish_list);
2395 spin_unlock(&tdtd->tdtd_replay_list_lock);
2397 spin_lock(&obd->obd_recovery_task_lock);
2398 if (transno == obd->obd_next_recovery_transno)
2399 obd->obd_next_recovery_transno++;
2401 obd->obd_next_recovery_transno)
2402 obd->obd_next_recovery_transno =
2404 spin_unlock(&obd->obd_recovery_task_lock);
2409 spin_unlock(&obd->obd_recovery_task_lock);
2411 LASSERT(list_empty(&obd->obd_req_replay_queue));
2412 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2413 /** evict exports failed VBR */
2414 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2420 static int target_recovery_thread(void *arg)
2422 struct lu_target *lut = arg;
2423 struct obd_device *obd = lut->lut_obd;
2424 struct ptlrpc_request *req;
2425 struct target_recovery_data *trd = &obd->obd_recovery_data;
2426 unsigned long delta;
2428 struct ptlrpc_thread *thread = NULL;
2432 unshare_fs_struct();
2433 OBD_ALLOC_PTR(thread);
2439 OBD_FREE_PTR(thread);
2443 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2445 OBD_FREE_PTR(thread);
2450 thread->t_env = env;
2451 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2452 env->le_ctx.lc_thread = thread;
2453 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2454 thread->t_watchdog = NULL;
2456 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2458 trd->trd_processing_task = current_pid();
2460 spin_lock(&obd->obd_dev_lock);
2461 obd->obd_recovering = 1;
2462 spin_unlock(&obd->obd_dev_lock);
2463 complete(&trd->trd_starting);
2465 /* first of all, we have to know the first transno to replay */
2466 if (target_recovery_overseer(lut, check_for_recovery_ready,
2467 exp_connect_healthy)) {
2468 abort_req_replay_queue(obd);
2469 abort_lock_replay_queue(obd);
2470 if (lut->lut_tdtd != NULL)
2471 dtrq_list_destroy(lut->lut_tdtd);
2474 /* next stage: replay requests or update */
2476 CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2477 atomic_read(&obd->obd_req_replay_clients),
2478 obd->obd_next_recovery_transno);
2479 replay_request_or_update(env, lut, trd, thread);
2482 * The second stage: replay locks
2484 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2485 atomic_read(&obd->obd_lock_replay_clients));
2486 while ((req = target_next_replay_lock(lut))) {
2487 LASSERT(trd->trd_processing_task == current_pid());
2488 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2489 libcfs_nid2str(req->rq_peer.nid));
2490 handle_recovery_req(thread, req,
2491 trd->trd_recovery_handler);
2492 target_request_copy_put(req);
2493 obd->obd_replayed_locks++;
2497 * The third stage: reply on final pings, at this moment all clients
2498 * must have request in final queue
2500 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2501 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2502 /** Update server last boot epoch */
2503 tgt_boot_epoch_update(lut);
2504 /* We drop recoverying flag to forward all new requests
2505 * to regular mds_handle() since now */
2506 spin_lock(&obd->obd_dev_lock);
2507 obd->obd_recovering = obd->obd_abort_recovery = 0;
2508 spin_unlock(&obd->obd_dev_lock);
2509 spin_lock(&obd->obd_recovery_task_lock);
2510 target_cancel_recovery_timer(obd);
2511 spin_unlock(&obd->obd_recovery_task_lock);
2512 while ((req = target_next_final_ping(obd))) {
2513 LASSERT(trd->trd_processing_task == current_pid());
2514 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2515 libcfs_nid2str(req->rq_peer.nid));
2516 handle_recovery_req(thread, req,
2517 trd->trd_recovery_handler);
2518 /* Because the waiting client can not send ping to server,
2519 * so we need refresh the last_request_time, to avoid the
2520 * export is being evicted */
2521 ptlrpc_update_export_timer(req->rq_export, 0);
2522 target_request_copy_put(req);
2525 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2526 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2527 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2528 if (delta > OBD_RECOVERY_TIME_SOFT) {
2529 CWARN("too long recovery - read logs\n");
2530 libcfs_debug_dumplog();
2533 target_finish_recovery(lut);
2535 lu_context_fini(&env->le_ctx);
2536 trd->trd_processing_task = 0;
2537 complete(&trd->trd_finishing);
2539 tgt_io_thread_done(thread);
2540 OBD_FREE_PTR(thread);
2545 static int target_start_recovery_thread(struct lu_target *lut,
2546 svc_handler_t handler)
2548 struct obd_device *obd = lut->lut_obd;
2550 struct target_recovery_data *trd = &obd->obd_recovery_data;
2553 memset(trd, 0, sizeof(*trd));
2554 init_completion(&trd->trd_starting);
2555 init_completion(&trd->trd_finishing);
2556 trd->trd_recovery_handler = handler;
2558 rc = server_name2index(obd->obd_name, &index, NULL);
2562 if (!IS_ERR(kthread_run(target_recovery_thread,
2563 lut, "tgt_recover_%d", index))) {
2564 wait_for_completion(&trd->trd_starting);
2565 LASSERT(obd->obd_recovering != 0);
2573 void target_stop_recovery_thread(struct obd_device *obd)
2575 if (obd->obd_recovery_data.trd_processing_task > 0) {
2576 struct target_recovery_data *trd = &obd->obd_recovery_data;
2577 /** recovery can be done but postrecovery is not yet */
2578 spin_lock(&obd->obd_dev_lock);
2579 if (obd->obd_recovering) {
2580 CERROR("%s: Aborting recovery\n", obd->obd_name);
2581 obd->obd_abort_recovery = 1;
2582 wake_up(&obd->obd_next_transno_waitq);
2584 spin_unlock(&obd->obd_dev_lock);
2585 wait_for_completion(&trd->trd_finishing);
2588 EXPORT_SYMBOL(target_stop_recovery_thread);
2590 void target_recovery_fini(struct obd_device *obd)
2592 class_disconnect_exports(obd);
2593 target_stop_recovery_thread(obd);
2594 target_cleanup_recovery(obd);
2596 EXPORT_SYMBOL(target_recovery_fini);
2598 static void target_recovery_expired(unsigned long castmeharder)
2600 struct obd_device *obd = (struct obd_device *)castmeharder;
2601 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2602 " after %lds (%d clients connected)\n",
2603 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2604 cfs_time_current_sec()- obd->obd_recovery_start,
2605 atomic_read(&obd->obd_connected_clients));
2607 obd->obd_recovery_expired = 1;
2608 wake_up(&obd->obd_next_transno_waitq);
2611 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2613 struct obd_device *obd = lut->lut_obd;
2615 if (obd->obd_max_recoverable_clients == 0) {
2616 /** Update server last boot epoch */
2617 tgt_boot_epoch_update(lut);
2621 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2622 "last_transno "LPU64"\n", obd->obd_name,
2623 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2624 LASSERT(obd->obd_stopping == 0);
2625 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2626 obd->obd_recovery_start = 0;
2627 obd->obd_recovery_end = 0;
2629 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2630 target_start_recovery_thread(lut, handler);
2632 EXPORT_SYMBOL(target_recovery_init);
2634 static int target_process_req_flags(struct obd_device *obd,
2635 struct ptlrpc_request *req)
2637 struct obd_export *exp = req->rq_export;
2638 LASSERT(exp != NULL);
2639 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2640 /* client declares he's ready to replay locks */
2641 spin_lock(&exp->exp_lock);
2642 if (exp->exp_req_replay_needed) {
2643 exp->exp_req_replay_needed = 0;
2644 spin_unlock(&exp->exp_lock);
2646 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2647 atomic_dec(&obd->obd_req_replay_clients);
2649 spin_unlock(&exp->exp_lock);
2652 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2653 /* client declares he's ready to complete recovery
2654 * so, we put the request on th final queue */
2655 spin_lock(&exp->exp_lock);
2656 if (exp->exp_lock_replay_needed) {
2657 exp->exp_lock_replay_needed = 0;
2658 spin_unlock(&exp->exp_lock);
2660 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2661 atomic_dec(&obd->obd_lock_replay_clients);
2663 spin_unlock(&exp->exp_lock);
2669 int target_queue_recovery_request(struct ptlrpc_request *req,
2670 struct obd_device *obd)
2672 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2673 struct ptlrpc_request *reqiter;
2677 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2678 /* Processing the queue right now, don't re-add. */
2682 target_process_req_flags(obd, req);
2684 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2685 /* client declares he's ready to complete recovery
2686 * so, we put the request on th final queue */
2687 target_request_copy_get(req);
2688 DEBUG_REQ(D_HA, req, "queue final req");
2689 wake_up(&obd->obd_next_transno_waitq);
2690 spin_lock(&obd->obd_recovery_task_lock);
2691 if (obd->obd_recovering) {
2692 list_add_tail(&req->rq_list,
2693 &obd->obd_final_req_queue);
2695 spin_unlock(&obd->obd_recovery_task_lock);
2696 target_request_copy_put(req);
2697 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2699 spin_unlock(&obd->obd_recovery_task_lock);
2702 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2703 /* client declares he's ready to replay locks */
2704 target_request_copy_get(req);
2705 DEBUG_REQ(D_HA, req, "queue lock replay req");
2706 wake_up(&obd->obd_next_transno_waitq);
2707 spin_lock(&obd->obd_recovery_task_lock);
2708 LASSERT(obd->obd_recovering);
2709 /* usually due to recovery abort */
2710 if (!req->rq_export->exp_in_recovery) {
2711 spin_unlock(&obd->obd_recovery_task_lock);
2712 target_request_copy_put(req);
2715 LASSERT(req->rq_export->exp_lock_replay_needed);
2716 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2717 spin_unlock(&obd->obd_recovery_task_lock);
2721 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2722 * (i.e. buflens etc are in my own byte order), but type-dependent
2723 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2726 INIT_LIST_HEAD(&req->rq_list);
2727 DEBUG_REQ(D_HA, req, "not queueing");
2731 /* If we're processing the queue, we want don't want to queue this
2734 * Also, if this request has a transno less than the one we're waiting
2735 * for, we should process it now. It could (and currently always will)
2736 * be an open request for a descriptor that was opened some time ago.
2738 * Also, a resent, replayed request that has already been
2739 * handled will pass through here and be processed immediately.
2741 CDEBUG(D_HA, "Next recovery transno: "LPU64
2742 ", current: "LPU64", replaying\n",
2743 obd->obd_next_recovery_transno, transno);
2745 /* If the request has been replayed by update replay, then sends this
2746 * request to the recovery thread (replay_request_or_update()), where
2747 * it will be handled */
2748 spin_lock(&obd->obd_recovery_task_lock);
2749 if (transno < obd->obd_next_recovery_transno &&
2750 !is_req_replayed_by_update(req)) {
2751 /* Processing the queue right now, don't re-add. */
2752 LASSERT(list_empty(&req->rq_list));
2753 spin_unlock(&obd->obd_recovery_task_lock);
2756 spin_unlock(&obd->obd_recovery_task_lock);
2758 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2761 target_request_copy_get(req);
2762 if (!req->rq_export->exp_in_recovery) {
2763 target_request_copy_put(req);
2766 LASSERT(req->rq_export->exp_req_replay_needed);
2768 if (target_exp_enqueue_req_replay(req)) {
2769 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2770 target_request_copy_put(req);
2775 spin_lock(&obd->obd_recovery_task_lock);
2776 LASSERT(obd->obd_recovering);
2777 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2778 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2779 list_add_tail(&req->rq_list, &reqiter->rq_list);
2784 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2786 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2787 "has been claimed by another client");
2788 spin_unlock(&obd->obd_recovery_task_lock);
2789 target_exp_dequeue_req_replay(req);
2790 target_request_copy_put(req);
2796 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2798 obd->obd_requests_queued_for_recovery++;
2799 spin_unlock(&obd->obd_recovery_task_lock);
2800 wake_up(&obd->obd_next_transno_waitq);
2804 int target_handle_ping(struct ptlrpc_request *req)
2806 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2807 return req_capsule_server_pack(&req->rq_pill);
2810 void target_committed_to_req(struct ptlrpc_request *req)
2812 struct obd_export *exp = req->rq_export;
2814 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2815 lustre_msg_set_last_committed(req->rq_repmsg,
2816 exp->exp_last_committed);
2818 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2819 "%d)", exp->exp_obd->obd_no_transno,
2820 req->rq_repmsg == NULL);
2822 CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2823 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2826 #endif /* HAVE_SERVER_SUPPORT */
2829 * Packs current SLV and Limit into \a req.
2831 int target_pack_pool_reply(struct ptlrpc_request *req)
2833 struct obd_device *obd;
2836 /* Check that we still have all structures alive as this may
2837 * be some late RPC at shutdown time. */
2838 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2839 !exp_connect_lru_resize(req->rq_export))) {
2840 lustre_msg_set_slv(req->rq_repmsg, 0);
2841 lustre_msg_set_limit(req->rq_repmsg, 0);
2845 /* OBD is alive here as export is alive, which we checked above. */
2846 obd = req->rq_export->exp_obd;
2848 read_lock(&obd->obd_pool_lock);
2849 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2850 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2851 read_unlock(&obd->obd_pool_lock);
2856 static int target_send_reply_msg(struct ptlrpc_request *req,
2857 int rc, int fail_id)
2859 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2860 DEBUG_REQ(D_ERROR, req, "dropping reply");
2863 /* We can have a null rq_reqmsg in the event of bad signature or
2864 * no context when unwrapping */
2865 if (req->rq_reqmsg &&
2866 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2867 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2871 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2872 req->rq_status = rc;
2873 return ptlrpc_send_error(req, 1);
2875 DEBUG_REQ(D_NET, req, "sending reply");
2878 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2881 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2883 struct ptlrpc_service_part *svcpt;
2885 struct ptlrpc_reply_state *rs;
2886 struct obd_export *exp;
2889 if (req->rq_no_reply) {
2894 svcpt = req->rq_rqbd->rqbd_svcpt;
2895 rs = req->rq_reply_state;
2896 if (rs == NULL || !rs->rs_difficult) {
2898 target_send_reply_msg (req, rc, fail_id);
2903 /* must be an export if locks saved */
2904 LASSERT(req->rq_export != NULL);
2905 /* req/reply consistent */
2906 LASSERT(rs->rs_svcpt == svcpt);
2909 LASSERT(!rs->rs_scheduled);
2910 LASSERT(!rs->rs_scheduled_ever);
2911 LASSERT(!rs->rs_handled);
2912 LASSERT(!rs->rs_on_net);
2913 LASSERT(rs->rs_export == NULL);
2914 LASSERT(list_empty(&rs->rs_obd_list));
2915 LASSERT(list_empty(&rs->rs_exp_list));
2917 exp = class_export_get(req->rq_export);
2919 /* disable reply scheduling while I'm setting up */
2920 rs->rs_scheduled = 1;
2922 rs->rs_xid = req->rq_xid;
2923 rs->rs_transno = req->rq_transno;
2924 rs->rs_export = exp;
2925 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2927 spin_lock(&exp->exp_uncommitted_replies_lock);
2928 CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2929 rs->rs_transno, exp->exp_last_committed);
2930 if (rs->rs_transno > exp->exp_last_committed) {
2931 /* not committed already */
2932 list_add_tail(&rs->rs_obd_list,
2933 &exp->exp_uncommitted_replies);
2935 spin_unlock(&exp->exp_uncommitted_replies_lock);
2937 spin_lock(&exp->exp_lock);
2938 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2939 spin_unlock(&exp->exp_lock);
2941 netrc = target_send_reply_msg(req, rc, fail_id);
2943 spin_lock(&svcpt->scp_rep_lock);
2945 atomic_inc(&svcpt->scp_nreps_difficult);
2948 /* error sending: reply is off the net. Also we need +1
2949 * reply ref until ptlrpc_handle_rs() is done
2950 * with the reply state (if the send was successful, there
2951 * would have been +1 ref for the net, which
2952 * reply_out_callback leaves alone) */
2954 ptlrpc_rs_addref(rs);
2957 spin_lock(&rs->rs_lock);
2958 if (rs->rs_transno <= exp->exp_last_committed ||
2959 (!rs->rs_on_net && !rs->rs_no_ack) ||
2960 list_empty(&rs->rs_exp_list) || /* completed already */
2961 list_empty(&rs->rs_obd_list)) {
2962 CDEBUG(D_HA, "Schedule reply immediately\n");
2963 ptlrpc_dispatch_difficult_reply(rs);
2965 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2966 rs->rs_scheduled = 0; /* allow notifier to schedule */
2968 spin_unlock(&rs->rs_lock);
2969 spin_unlock(&svcpt->scp_rep_lock);
2973 enum ldlm_mode lck_compat_array[] = {
2974 [LCK_EX] = LCK_COMPAT_EX,
2975 [LCK_PW] = LCK_COMPAT_PW,
2976 [LCK_PR] = LCK_COMPAT_PR,
2977 [LCK_CW] = LCK_COMPAT_CW,
2978 [LCK_CR] = LCK_COMPAT_CR,
2979 [LCK_NL] = LCK_COMPAT_NL,
2980 [LCK_GROUP] = LCK_COMPAT_GROUP,
2981 [LCK_COS] = LCK_COMPAT_COS,
2985 * Rather arbitrary mapping from LDLM error codes to errno values. This should
2986 * not escape to the user level.
2988 int ldlm_error2errno(enum ldlm_error error)
2994 case ELDLM_LOCK_MATCHED:
2997 case ELDLM_LOCK_CHANGED:
3000 case ELDLM_LOCK_ABORTED:
3003 case ELDLM_LOCK_REPLACED:
3006 case ELDLM_NO_LOCK_DATA:
3009 case ELDLM_NAMESPACE_EXISTS:
3012 case ELDLM_BAD_NAMESPACE:
3016 if (((int)error) < 0) { /* cast to signed type */
3017 result = error; /* as ldlm_error can be unsigned */
3019 CERROR("Invalid DLM result code: %d\n", error);
3025 EXPORT_SYMBOL(ldlm_error2errno);
3028 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3030 enum ldlm_error ldlm_errno2error(int err_no)
3039 error = ELDLM_LOCK_CHANGED;
3042 error = ELDLM_LOCK_ABORTED;
3045 error = ELDLM_LOCK_REPLACED;
3048 error = ELDLM_NO_LOCK_DATA;
3051 error = ELDLM_NAMESPACE_EXISTS;
3054 error = ELDLM_BAD_NAMESPACE;
3062 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3063 void ldlm_dump_export_locks(struct obd_export *exp)
3065 spin_lock(&exp->exp_locks_list_guard);
3066 if (!list_empty(&exp->exp_locks_list)) {
3067 struct ldlm_lock *lock;
3069 CERROR("dumping locks for export %p,"
3070 "ignore if the unmount doesn't hang\n", exp);
3071 list_for_each_entry(lock, &exp->exp_locks_list,
3073 LDLM_ERROR(lock, "lock:");
3075 spin_unlock(&exp->exp_locks_list_guard);
3079 #ifdef HAVE_SERVER_SUPPORT
3080 static int target_bulk_timeout(void *data)
3083 /* We don't fail the connection here, because having the export
3084 * killed makes the (vital) call to commitrw very sad.
3089 static inline const char *bulk2type(struct ptlrpc_request *req)
3091 if (req->rq_bulk_read)
3093 if (req->rq_bulk_write)
3098 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3099 struct l_wait_info *lwi)
3101 struct ptlrpc_request *req = desc->bd_req;
3102 time_t start = cfs_time_current_sec();
3108 /* If there is eviction in progress, wait for it to finish. */
3109 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3110 *lwi = LWI_INTR(NULL, NULL);
3111 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3112 !atomic_read(&exp->exp_obd->
3113 obd_evict_inprogress),
3117 /* Check if client was evicted or reconnected already. */
3118 if (exp->exp_failed ||
3119 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3122 if (req->rq_bulk_read)
3123 rc = sptlrpc_svc_wrap_bulk(req, desc);
3125 if ((exp->exp_connect_data.ocd_connect_flags &
3126 OBD_CONNECT_BULK_MBITS) != 0)
3127 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3128 else /* old version, bulk matchbits is rq_xid */
3129 req->rq_mbits = req->rq_xid;
3132 rc = ptlrpc_start_bulk_transfer(desc);
3136 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3137 bulk2type(req), rc);
3141 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3142 ptlrpc_abort_bulk(desc);
3146 /* limit actual bulk transfer to bulk_timeout seconds */
3147 deadline = start + bulk_timeout;
3148 if (deadline > req->rq_deadline)
3149 deadline = req->rq_deadline;
3152 long timeoutl = deadline - cfs_time_current_sec();
3153 cfs_duration_t timeout = timeoutl <= 0 ?
3154 CFS_TICK : cfs_time_seconds(timeoutl);
3157 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3158 target_bulk_timeout, desc);
3159 rc = l_wait_event(desc->bd_waitq,
3160 !ptlrpc_server_bulk_active(desc) ||
3163 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3165 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3166 /* Wait again if we changed rq_deadline. */
3167 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3168 deadline = start + bulk_timeout;
3169 if (deadline > rq_deadline)
3170 deadline = rq_deadline;
3171 } while ((rc == -ETIMEDOUT) &&
3172 (deadline > cfs_time_current_sec()));
3174 if (rc == -ETIMEDOUT) {
3175 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3176 bulk2type(req), deadline - start,
3177 cfs_time_current_sec() - deadline);
3178 ptlrpc_abort_bulk(desc);
3179 } else if (exp->exp_failed) {
3180 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3183 ptlrpc_abort_bulk(desc);
3184 } else if (exp->exp_conn_cnt >
3185 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3186 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3188 /* We don't reply anyway. */
3190 ptlrpc_abort_bulk(desc);
3191 } else if (desc->bd_failure) {
3192 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3194 /* XXX should this be a different errno? */
3197 if (req->rq_bulk_write)
3198 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3199 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3200 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3201 bulk2type(req), desc->bd_nob_transferred,
3203 /* XXX should this be a different errno? */
3210 EXPORT_SYMBOL(target_bulk_io);
3212 #endif /* HAVE_SERVER_SUPPORT */