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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * This file deals with various client/target related logic including recovery.
36 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
40 #define DEBUG_SUBSYSTEM S_LDLM
42 #include <linux/kthread.h>
43 #include <libcfs/libcfs.h>
45 #include <obd_class.h>
46 #include <lustre_dlm.h>
47 #include <lustre_net.h>
48 #include <lustre_sec.h>
49 #include "ldlm_internal.h"
51 /* @priority: If non-zero, move the selected connection to the list head.
52 * @create: If zero, only search in existing connections.
54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
55 int priority, int create)
57 struct ptlrpc_connection *ptlrpc_conn;
58 struct obd_import_conn *imp_conn = NULL, *item;
62 if (!create && !priority) {
63 CDEBUG(D_HA, "Nothing to do\n");
67 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
69 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
74 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
76 GOTO(out_put, rc = -ENOMEM);
80 spin_lock(&imp->imp_lock);
81 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
82 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
84 list_del(&item->oic_item);
85 list_add(&item->oic_item,
87 item->oic_last_attempt = 0;
89 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
90 imp, imp->imp_obd->obd_name, uuid->uuid,
91 (priority ? ", moved to head" : ""));
92 spin_unlock(&imp->imp_lock);
93 GOTO(out_free, rc = 0);
96 /* No existing import connection found for \a uuid. */
98 imp_conn->oic_conn = ptlrpc_conn;
99 imp_conn->oic_uuid = *uuid;
100 imp_conn->oic_last_attempt = 0;
102 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
104 list_add_tail(&imp_conn->oic_item,
105 &imp->imp_conn_list);
106 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
107 imp, imp->imp_obd->obd_name, uuid->uuid,
108 (priority ? "head" : "tail"));
110 spin_unlock(&imp->imp_lock);
111 GOTO(out_free, rc = -ENOENT);
114 spin_unlock(&imp->imp_lock);
118 OBD_FREE(imp_conn, sizeof(*imp_conn));
120 ptlrpc_connection_put(ptlrpc_conn);
124 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
126 return import_set_conn(imp, uuid, 1, 0);
129 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
132 return import_set_conn(imp, uuid, priority, 1);
134 EXPORT_SYMBOL(client_import_add_conn);
136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
138 struct obd_import_conn *imp_conn;
139 struct obd_export *dlmexp;
143 spin_lock(&imp->imp_lock);
144 if (list_empty(&imp->imp_conn_list)) {
145 LASSERT(!imp->imp_connection);
149 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
150 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
152 LASSERT(imp_conn->oic_conn);
154 if (imp_conn == imp->imp_conn_current) {
155 LASSERT(imp_conn->oic_conn == imp->imp_connection);
157 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
158 imp->imp_state != LUSTRE_IMP_DISCON) {
159 CERROR("can't remove current connection\n");
160 GOTO(out, rc = -EBUSY);
163 ptlrpc_connection_put(imp->imp_connection);
164 imp->imp_connection = NULL;
166 dlmexp = class_conn2export(&imp->imp_dlm_handle);
167 if (dlmexp && dlmexp->exp_connection) {
168 LASSERT(dlmexp->exp_connection ==
170 ptlrpc_connection_put(dlmexp->exp_connection);
171 dlmexp->exp_connection = NULL;
175 class_export_put(dlmexp);
178 list_del(&imp_conn->oic_item);
179 ptlrpc_connection_put(imp_conn->oic_conn);
180 OBD_FREE(imp_conn, sizeof(*imp_conn));
181 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
182 imp, imp->imp_obd->obd_name, uuid->uuid);
187 spin_unlock(&imp->imp_lock);
189 CERROR("connection %s not found\n", uuid->uuid);
192 EXPORT_SYMBOL(client_import_del_conn);
195 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
196 * to find a conn uuid of \a imp which can reach \a peer.
198 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
199 struct obd_uuid *uuid)
201 struct obd_import_conn *conn;
205 spin_lock(&imp->imp_lock);
206 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
207 /* Check if conn UUID does have this peer NID. */
208 if (class_check_uuid(&conn->oic_uuid, peer)) {
209 *uuid = conn->oic_uuid;
214 spin_unlock(&imp->imp_lock);
217 EXPORT_SYMBOL(client_import_find_conn);
219 void client_destroy_import(struct obd_import *imp)
221 /* Drop security policy instance after all RPCs have finished/aborted
222 * to let all busy contexts be released. */
223 class_import_get(imp);
224 class_destroy_import(imp);
225 sptlrpc_import_sec_put(imp);
226 class_import_put(imp);
228 EXPORT_SYMBOL(client_destroy_import);
231 * Check whether or not the OSC is on MDT.
234 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
236 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
239 static int osc_on_mdt(char *obdname)
243 ptr = strrchr(obdname, '-');
247 if (strncmp(ptr + 1, "MDT", 3) == 0)
253 /* Configure an RPC client OBD device.
258 * 3 - inactive-on-startup
260 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
262 struct client_obd *cli = &obddev->u.cli;
263 struct obd_import *imp;
264 struct obd_uuid server_uuid;
265 int rq_portal, rp_portal, connect_op;
266 char *name = obddev->obd_type->typ_name;
267 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
268 char *cli_name = lustre_cfg_buf(lcfg, 0);
272 /* In a more perfect world, we would hang a ptlrpc_client off of
273 * obd_type and just use the values from there. */
274 if (!strcmp(name, LUSTRE_OSC_NAME)) {
275 rq_portal = OST_REQUEST_PORTAL;
276 rp_portal = OSC_REPLY_PORTAL;
277 connect_op = OST_CONNECT;
278 cli->cl_sp_me = LUSTRE_SP_CLI;
279 cli->cl_sp_to = LUSTRE_SP_OST;
280 ns_type = LDLM_NS_TYPE_OSC;
281 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
282 !strcmp(name, LUSTRE_LWP_NAME)) {
283 rq_portal = MDS_REQUEST_PORTAL;
284 rp_portal = MDC_REPLY_PORTAL;
285 connect_op = MDS_CONNECT;
286 if (is_lwp_on_ost(cli_name))
287 cli->cl_sp_me = LUSTRE_SP_OST;
288 else if (is_lwp_on_mdt(cli_name))
289 cli->cl_sp_me = LUSTRE_SP_MDT;
291 cli->cl_sp_me = LUSTRE_SP_CLI;
292 cli->cl_sp_to = LUSTRE_SP_MDT;
293 ns_type = LDLM_NS_TYPE_MDC;
294 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
295 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
296 /* OSP_on_MDT for other MDTs */
297 connect_op = MDS_CONNECT;
298 cli->cl_sp_to = LUSTRE_SP_MDT;
299 ns_type = LDLM_NS_TYPE_MDC;
300 rq_portal = OUT_PORTAL;
302 /* OSP on MDT for OST */
303 connect_op = OST_CONNECT;
304 cli->cl_sp_to = LUSTRE_SP_OST;
305 ns_type = LDLM_NS_TYPE_OSC;
306 rq_portal = OST_REQUEST_PORTAL;
308 rp_portal = OSC_REPLY_PORTAL;
309 cli->cl_sp_me = LUSTRE_SP_MDT;
310 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
311 rq_portal = MGS_REQUEST_PORTAL;
312 rp_portal = MGC_REPLY_PORTAL;
313 connect_op = MGS_CONNECT;
314 cli->cl_sp_me = LUSTRE_SP_MGC;
315 cli->cl_sp_to = LUSTRE_SP_MGS;
316 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
317 ns_type = LDLM_NS_TYPE_MGC;
319 CERROR("unknown client OBD type \"%s\", can't setup\n",
324 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
325 CERROR("requires a TARGET UUID\n");
329 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
330 CERROR("client UUID must be less than 38 characters\n");
334 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
335 CERROR("setup requires a SERVER UUID\n");
339 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
340 CERROR("target UUID must be less than 38 characters\n");
344 init_rwsem(&cli->cl_sem);
345 mutex_init(&cli->cl_mgc_mutex);
346 cli->cl_conn_count = 0;
347 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
348 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
349 sizeof(server_uuid)));
351 cli->cl_dirty_pages = 0;
352 cli->cl_avail_grant = 0;
353 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
354 /* cl_dirty_max_pages may be changed at connect time in
355 * ptlrpc_connect_interpret(). */
356 client_adjust_max_dirty(cli);
357 INIT_LIST_HEAD(&cli->cl_cache_waiters);
358 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
359 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
360 INIT_LIST_HEAD(&cli->cl_loi_write_list);
361 INIT_LIST_HEAD(&cli->cl_loi_read_list);
362 spin_lock_init(&cli->cl_loi_list_lock);
363 atomic_set(&cli->cl_pending_w_pages, 0);
364 atomic_set(&cli->cl_pending_r_pages, 0);
365 cli->cl_r_in_flight = 0;
366 cli->cl_w_in_flight = 0;
368 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
369 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
370 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
371 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
372 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
373 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
376 INIT_LIST_HEAD(&cli->cl_lru_osc);
377 atomic_set(&cli->cl_lru_shrinkers, 0);
378 atomic_long_set(&cli->cl_lru_busy, 0);
379 atomic_long_set(&cli->cl_lru_in_list, 0);
380 INIT_LIST_HEAD(&cli->cl_lru_list);
381 spin_lock_init(&cli->cl_lru_list_lock);
382 atomic_long_set(&cli->cl_unstable_count, 0);
383 INIT_LIST_HEAD(&cli->cl_shrink_list);
385 init_waitqueue_head(&cli->cl_destroy_waitq);
386 atomic_set(&cli->cl_destroy_in_flight, 0);
387 #ifdef ENABLE_CHECKSUM
388 /* Turn on checksumming by default. */
389 cli->cl_checksum = 1;
391 * The supported checksum types will be worked out at connect time
392 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
395 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
397 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
399 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
400 * from OFD after connecting. */
401 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
403 /* set cl_chunkbits default value to PAGE_SHIFT,
404 * it will be updated at OSC connection time. */
405 cli->cl_chunkbits = PAGE_SHIFT;
407 if (!strcmp(name, LUSTRE_MDC_NAME)) {
408 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
409 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
410 cli->cl_max_rpcs_in_flight = 2;
411 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
412 cli->cl_max_rpcs_in_flight = 3;
413 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
414 cli->cl_max_rpcs_in_flight = 4;
416 if (osc_on_mdt(obddev->obd_name))
417 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
419 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
422 spin_lock_init(&cli->cl_mod_rpcs_lock);
423 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
424 cli->cl_max_mod_rpcs_in_flight = 0;
425 cli->cl_mod_rpcs_in_flight = 0;
426 cli->cl_close_rpcs_in_flight = 0;
427 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
428 cli->cl_mod_tag_bitmap = NULL;
430 if (connect_op == MDS_CONNECT) {
431 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
432 OBD_ALLOC(cli->cl_mod_tag_bitmap,
433 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
434 if (cli->cl_mod_tag_bitmap == NULL)
435 GOTO(err, rc = -ENOMEM);
440 CERROR("ldlm_get_ref failed: %d\n", rc);
444 ptlrpc_init_client(rq_portal, rp_portal, name,
445 &obddev->obd_ldlm_client);
447 imp = class_new_import(obddev);
449 GOTO(err_ldlm, rc = -ENOENT);
450 imp->imp_client = &obddev->obd_ldlm_client;
451 imp->imp_connect_op = connect_op;
452 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
453 LUSTRE_CFG_BUFLEN(lcfg, 1));
454 class_import_put(imp);
456 rc = client_import_add_conn(imp, &server_uuid, 1);
458 CERROR("can't add initial connection\n");
459 GOTO(err_import, rc);
462 cli->cl_import = imp;
463 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
464 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
466 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
467 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
468 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
469 name, obddev->obd_name,
470 cli->cl_target_uuid.uuid);
471 spin_lock(&imp->imp_lock);
472 imp->imp_deactive = 1;
473 spin_unlock(&imp->imp_lock);
477 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
478 LDLM_NAMESPACE_CLIENT,
479 LDLM_NAMESPACE_GREEDY,
481 if (obddev->obd_namespace == NULL) {
482 CERROR("Unable to create client namespace - %s\n",
484 GOTO(err_import, rc = -ENOMEM);
490 class_destroy_import(imp);
494 if (cli->cl_mod_tag_bitmap != NULL)
495 OBD_FREE(cli->cl_mod_tag_bitmap,
496 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
497 cli->cl_mod_tag_bitmap = NULL;
501 EXPORT_SYMBOL(client_obd_setup);
503 int client_obd_cleanup(struct obd_device *obddev)
505 struct client_obd *cli = &obddev->u.cli;
508 ldlm_namespace_free_post(obddev->obd_namespace);
509 obddev->obd_namespace = NULL;
511 obd_cleanup_client_import(obddev);
512 LASSERT(obddev->u.cli.cl_import == NULL);
516 if (cli->cl_mod_tag_bitmap != NULL)
517 OBD_FREE(cli->cl_mod_tag_bitmap,
518 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
519 cli->cl_mod_tag_bitmap = NULL;
523 EXPORT_SYMBOL(client_obd_cleanup);
525 /* ->o_connect() method for client side (OSC and MDC and MGC) */
526 int client_connect_import(const struct lu_env *env,
527 struct obd_export **exp,
528 struct obd_device *obd, struct obd_uuid *cluuid,
529 struct obd_connect_data *data, void *localdata)
531 struct client_obd *cli = &obd->u.cli;
532 struct obd_import *imp = cli->cl_import;
533 struct obd_connect_data *ocd;
534 struct lustre_handle conn = { 0 };
540 down_write(&cli->cl_sem);
541 if (cli->cl_conn_count > 0)
542 GOTO(out_sem, rc = -EALREADY);
544 rc = class_connect(&conn, obd, cluuid);
548 cli->cl_conn_count++;
549 *exp = class_conn2export(&conn);
551 LASSERT(obd->obd_namespace);
553 imp->imp_dlm_handle = conn;
554 rc = ptlrpc_init_import(imp);
558 ocd = &imp->imp_connect_data;
561 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
562 LUSTRE_MDC_NAME, 3) == 0;
564 data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
565 imp->imp_connect_flags_orig = data->ocd_connect_flags;
566 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
569 rc = ptlrpc_connect_import(imp);
572 data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
573 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
576 LASSERT(*exp != NULL && (*exp)->exp_connection);
579 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
580 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
581 data->ocd_connect_flags, ocd->ocd_connect_flags);
582 data->ocd_connect_flags = ocd->ocd_connect_flags;
583 /* clear the flag as it was not set and is not known
586 data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
589 ptlrpc_pinger_add_import(imp);
595 cli->cl_conn_count--;
596 class_disconnect(*exp);
600 up_write(&cli->cl_sem);
604 EXPORT_SYMBOL(client_connect_import);
606 int client_disconnect_export(struct obd_export *exp)
608 struct obd_device *obd = class_exp2obd(exp);
609 struct client_obd *cli;
610 struct obd_import *imp;
615 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
616 exp, exp ? exp->exp_handle.h_cookie : -1);
621 imp = cli->cl_import;
623 down_write(&cli->cl_sem);
624 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
627 if (cli->cl_conn_count == 0) {
628 CERROR("disconnecting disconnected device (%s)\n",
630 GOTO(out_disconnect, rc = -EINVAL);
633 cli->cl_conn_count--;
634 if (cli->cl_conn_count != 0)
635 GOTO(out_disconnect, rc = 0);
637 /* Mark import deactivated now, so we don't try to reconnect if any
638 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
639 * fully deactivate the import, or that would drop all requests. */
640 spin_lock(&imp->imp_lock);
641 imp->imp_deactive = 1;
642 spin_unlock(&imp->imp_lock);
644 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
645 * delete it regardless. (It's safe to delete an import that was
647 (void)ptlrpc_pinger_del_import(imp);
649 if (obd->obd_namespace != NULL) {
650 /* obd_force == local only */
651 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
652 obd->obd_force ? LCF_LOCAL : 0, NULL);
653 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
656 /* There's no need to hold sem while disconnecting an import,
657 * and it may actually cause deadlock in GSS. */
658 up_write(&cli->cl_sem);
659 rc = ptlrpc_disconnect_import(imp, 0);
660 down_write(&cli->cl_sem);
662 ptlrpc_invalidate_import(imp);
667 /* Use server style - class_disconnect should be always called for
669 err = class_disconnect(exp);
673 up_write(&cli->cl_sem);
677 EXPORT_SYMBOL(client_disconnect_export);
679 #ifdef HAVE_SERVER_SUPPORT
680 int server_disconnect_export(struct obd_export *exp)
685 /* Disconnect early so that clients can't keep using export. */
686 rc = class_disconnect(exp);
687 /* Close import to avoid sending any requests. */
688 if (exp->exp_imp_reverse)
689 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
691 ldlm_bl_thread_wakeup();
693 /* complete all outstanding replies */
694 spin_lock(&exp->exp_lock);
695 while (!list_empty(&exp->exp_outstanding_replies)) {
696 struct ptlrpc_reply_state *rs =
697 list_entry(exp->exp_outstanding_replies.next,
698 struct ptlrpc_reply_state, rs_exp_list);
699 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
701 spin_lock(&svcpt->scp_rep_lock);
703 list_del_init(&rs->rs_exp_list);
704 spin_lock(&rs->rs_lock);
705 ptlrpc_schedule_difficult_reply(rs);
706 spin_unlock(&rs->rs_lock);
708 spin_unlock(&svcpt->scp_rep_lock);
710 spin_unlock(&exp->exp_lock);
714 EXPORT_SYMBOL(server_disconnect_export);
716 /* --------------------------------------------------------------------------
717 * from old lib/target.c
718 * -------------------------------------------------------------------------- */
720 static int target_handle_reconnect(struct lustre_handle *conn,
721 struct obd_export *exp,
722 struct obd_uuid *cluuid)
724 struct obd_device *target;
725 struct lustre_handle *hdl;
732 hdl = &exp->exp_imp_reverse->imp_remote_handle;
733 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
734 conn->cookie = exp->exp_handle.h_cookie;
735 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
736 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
740 target = exp->exp_obd;
742 /* Might be a re-connect after a partition. */
743 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
744 LCONSOLE_WARN("%s: already connected client %s (at %s) "
745 "with handle %#llx. Rejecting client "
746 "with the same UUID trying to reconnect "
747 "with handle %#llx\n", target->obd_name,
748 obd_uuid2str(&exp->exp_client_uuid),
749 obd_export_nid2str(exp),
750 hdl->cookie, conn->cookie);
751 memset(conn, 0, sizeof *conn);
752 /* target_handle_connect() treats EALREADY and
753 * -EALREADY differently. -EALREADY is an error
754 * (same UUID, different handle). */
758 if (!target->obd_recovering) {
759 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
760 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
761 obd_export_nid2str(exp));
762 GOTO(out_already, rc);
765 now = cfs_time_current();
766 deadline = cfs_timer_deadline(&target->obd_recovery_timer);
767 if (cfs_time_before(now, deadline)) {
768 struct target_distribute_txn_data *tdtd =
769 class_exp2tgt(exp)->lut_tdtd;
774 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
775 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
776 buf = tdtd->tdtd_show_update_logs_retrievers(
777 tdtd->tdtd_show_retrievers_cbdata,
781 LCONSOLE_WARN("%s: Recovery already passed deadline "
782 "%d:%.02d. It is due to DNE recovery "
783 "failed/stuck on the %d MDT(s):%s. "
784 "Please wait until all MDTs recovered "
785 "or abort the recovery by force.\n",
786 target->obd_name, timeout / 60,
788 buf ? buf : "unknown (not enough RAM)");
790 LCONSOLE_WARN("%s: Recovery already passed deadline "
791 "%d:%.02d. If you do not want to wait "
792 "more, please abort the recovery by "
793 "force.\n", target->obd_name,
794 timeout / 60, timeout % 60);
799 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
800 LCONSOLE_WARN("%s: Recovery already passed deadline"
801 " %d:%.02d, It is most likely due to DNE"
802 " recovery is failed or stuck, please wait a"
803 " few more minutes or abort the recovery.\n",
804 target->obd_name, timeout / 60, timeout % 60);
808 conn->cookie = exp->exp_handle.h_cookie;
809 /* target_handle_connect() treats EALREADY and
810 * -EALREADY differently. EALREADY means we are
811 * doing a valid reconnect from the same client. */
815 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
818 struct obd_export *exp = cb_data;
820 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
821 obd->obd_name, exp->exp_client_uuid.uuid);
823 spin_lock(&exp->exp_lock);
824 exp->exp_need_sync = 0;
825 spin_unlock(&exp->exp_lock);
826 class_export_cb_put(exp);
830 check_and_start_recovery_timer(struct obd_device *obd,
831 struct ptlrpc_request *req, int new_client);
834 * update flags for import during reconnect process
836 static int rev_import_flags_update(struct obd_import *revimp,
837 struct ptlrpc_request *req)
840 struct obd_connect_data *data;
842 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
844 if (data->ocd_connect_flags & OBD_CONNECT_AT)
845 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
847 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
849 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
851 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
853 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
854 revimp->imp_client->cli_name,
855 libcfs_id2str(req->rq_peer), rc);
863 * Allocate a new reverse import for an export.
865 * \retval -errno in case error hit
866 * \retval 0 if reverse import correctly init
868 int rev_import_init(struct obd_export *export)
870 struct obd_device *obd = export->exp_obd;
871 struct obd_import *revimp;
873 LASSERT(export->exp_imp_reverse == NULL);
875 revimp = class_new_import(obd);
879 revimp->imp_remote_handle.cookie = 0ULL;
880 revimp->imp_client = &obd->obd_ldlm_client;
881 revimp->imp_dlm_fake = 1;
883 /* it is safe to connect import in new state as no sends possible */
884 spin_lock(&export->exp_lock);
885 export->exp_imp_reverse = revimp;
886 spin_unlock(&export->exp_lock);
887 class_import_put(revimp);
891 EXPORT_SYMBOL(rev_import_init);
894 * Handle reconnect for an export.
896 * \param exp export to handle reconnect process
897 * \param req client reconnect request
899 * \retval -rc in case securitfy flavor can't be changed
900 * \retval 0 in case none problems
902 static int rev_import_reconnect(struct obd_export *exp,
903 struct ptlrpc_request *req)
905 struct obd_import *revimp = exp->exp_imp_reverse;
906 struct lustre_handle *lh;
909 /* avoid sending a request until import flags are changed */
910 ptlrpc_import_enter_resend(revimp);
912 if (revimp->imp_connection != NULL)
913 ptlrpc_connection_put(revimp->imp_connection);
916 * client from recovery don't have a handle so we need to take from
917 * request. it may produce situation when wrong client connected
918 * to recovery as we trust a client uuid
920 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
921 revimp->imp_remote_handle = *lh;
923 /* unknown versions will be caught in
924 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
925 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
927 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
929 rc = rev_import_flags_update(revimp, req);
931 /* it is safe to still be in RECOVERY phase as we are not able
932 * to setup correct security flavor so requests are not able to
933 * be delivered correctly */
937 /* resend all rpc's via new connection */
938 return ptlrpc_import_recovery_state_machine(revimp);
941 int target_handle_connect(struct ptlrpc_request *req)
943 struct obd_device *target = NULL;
944 struct obd_export *export = NULL;
945 /* connect handle - filled from target_handle_reconnect in
947 struct lustre_handle conn;
948 struct lustre_handle *tmp;
949 struct obd_uuid tgtuuid;
950 struct obd_uuid cluuid;
955 bool mds_conn = false, lw_client = false;
956 bool mds_mds_conn = false;
957 bool new_mds_mds_conn = false;
958 bool target_referenced = false;
959 struct obd_connect_data *data, *tmpdata;
961 lnet_nid_t *client_nid = NULL;
964 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
966 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
968 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
969 GOTO(out, rc = -EINVAL);
972 obd_str2uuid(&tgtuuid, str);
973 target = class_uuid2obd(&tgtuuid);
975 target = class_name2obd(str);
978 deuuidify(str, NULL, &target_start, &target_len);
979 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
980 "from %s (no target). If you are running "
981 "an HA pair check that the target is "
982 "mounted on the other server.\n", str,
983 libcfs_nid2str(req->rq_peer.nid));
984 GOTO(out, rc = -ENODEV);
987 spin_lock(&target->obd_dev_lock);
988 if (target->obd_stopping || !target->obd_set_up) {
989 spin_unlock(&target->obd_dev_lock);
991 deuuidify(str, NULL, &target_start, &target_len);
992 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
993 target_len, target_start,
994 libcfs_nid2str(req->rq_peer.nid),
995 (target->obd_stopping ?
996 "stopping" : "not set up"));
997 GOTO(out, rc = -ENODEV);
1000 if (target->obd_no_conn) {
1001 spin_unlock(&target->obd_dev_lock);
1003 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1004 "from %s\n", target->obd_name,
1005 libcfs_nid2str(req->rq_peer.nid));
1006 GOTO(out, rc = -EAGAIN);
1009 /* Make sure the target isn't cleaned up while we're here. Yes,
1010 * there's still a race between the above check and our incref here.
1011 * Really, class_uuid2obd should take the ref. */
1012 class_incref(target, __func__, current);
1013 target_referenced = true;
1015 target->obd_conn_inprogress++;
1016 spin_unlock(&target->obd_dev_lock);
1018 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1020 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1021 GOTO(out, rc = -EINVAL);
1024 obd_str2uuid(&cluuid, str);
1026 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1028 GOTO(out, rc = -EPROTO);
1032 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1034 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1036 GOTO(out, rc = -EPROTO);
1038 rc = req_capsule_server_pack(&req->rq_pill);
1042 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1043 /* Don't allow clients to connect that are using old 1.8 format
1044 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1045 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1046 * FULL20 flag should be set on all connections since 2.0, but no
1047 * longer affects behaviour.
1049 * Later this check will be disabled and the flag can be retired
1050 * completely once interop with 3.0 is no longer needed.
1052 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1053 GOTO(out, rc = -EPROTO);
1056 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1057 if (data->ocd_version < LUSTRE_VERSION_CODE -
1058 LUSTRE_VERSION_ALLOWED_OFFSET ||
1059 data->ocd_version > LUSTRE_VERSION_CODE +
1060 LUSTRE_VERSION_ALLOWED_OFFSET) {
1061 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1062 "libclient connection attempt",
1063 data->ocd_version < LUSTRE_VERSION_CODE ?
1065 OBD_OCD_VERSION_MAJOR(data->ocd_version),
1066 OBD_OCD_VERSION_MINOR(data->ocd_version),
1067 OBD_OCD_VERSION_PATCH(data->ocd_version),
1068 OBD_OCD_VERSION_FIX(data->ocd_version));
1069 data = req_capsule_server_sized_get(&req->rq_pill,
1071 offsetof(typeof(*data), ocd_version) +
1072 sizeof(data->ocd_version));
1074 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1075 data->ocd_version = LUSTRE_VERSION_CODE;
1077 GOTO(out, rc = -EPROTO);
1081 /* Note: lw_client is needed in MDS-MDS failover during update log
1082 * processing, so we needs to allow lw_client to be connected at
1083 * anytime, instead of only the initial connection */
1084 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1086 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1087 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1088 mds_mds_conn = (data->ocd_connect_flags &
1089 OBD_CONNECT_MDS_MDS) != 0;
1091 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1092 * for Imperative Recovery connection from MGC to MGS.
1094 * Via check OBD_CONNECT_FID, we can distinguish whether
1095 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1098 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1099 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1100 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1101 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1102 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1103 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1105 /* We do not support the MDT-MDT interoperations with
1106 * different version MDT because of protocol changes. */
1107 if (unlikely(major != LUSTRE_MAJOR ||
1108 minor != LUSTRE_MINOR ||
1109 abs(patch - LUSTRE_PATCH) > 3)) {
1110 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1111 "connection from different version MDT "
1112 "(%d.%d.%d.%d) %s %s\n",
1113 target->obd_name, LUSTRE_MAJOR,
1114 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1115 major, minor, patch,
1116 OBD_OCD_VERSION_FIX(data->ocd_version),
1117 libcfs_nid2str(req->rq_peer.nid), str);
1119 GOTO(out, rc = -EPROTO);
1124 /* lctl gets a backstage, all-access pass. */
1125 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1126 goto dont_check_exports;
1128 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1132 /* We've found an export in the hash. */
1134 spin_lock(&export->exp_lock);
1136 if (export->exp_connecting) { /* bug 9635, et. al. */
1137 spin_unlock(&export->exp_lock);
1138 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1139 export->exp_obd->obd_name, export,
1140 libcfs_nid2str(req->rq_peer.nid));
1141 class_export_put(export);
1144 } else if ((mds_conn || lw_client ||
1145 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1146 export->exp_connection != NULL) {
1147 spin_unlock(&export->exp_lock);
1148 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1149 /* MDS or LWP reconnected after failover. */
1150 LCONSOLE_WARN("%s: Received %s connection from "
1151 "%s, removing former export from %s\n",
1152 target->obd_name, mds_conn ? "MDS" : "LWP",
1153 libcfs_nid2str(req->rq_peer.nid),
1154 libcfs_nid2str(export->exp_connection->c_peer.nid));
1156 /* New MDS connection from the same NID. */
1157 LCONSOLE_WARN("%s: Received new %s connection from "
1158 "%s, removing former export from same NID\n",
1159 target->obd_name, mds_conn ? "MDS" : "LWP",
1160 libcfs_nid2str(req->rq_peer.nid));
1163 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1164 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1165 /* Because exports between MDTs will always be
1166 * kept, let's do not fail such export if they
1167 * come from the same NID, otherwise it might
1168 * cause eviction between MDTs, which might
1169 * cause namespace inconsistency */
1170 spin_lock(&export->exp_lock);
1171 export->exp_connecting = 1;
1172 spin_unlock(&export->exp_lock);
1173 conn.cookie = export->exp_handle.h_cookie;
1176 class_fail_export(export);
1177 class_export_put(export);
1181 } else if (export->exp_connection != NULL &&
1182 req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1183 (lustre_msg_get_op_flags(req->rq_reqmsg) &
1184 MSG_CONNECT_INITIAL)) {
1185 spin_unlock(&export->exp_lock);
1186 /* In MDS failover we have static UUID but NID can change. */
1187 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1188 "existing nid %s is already connected\n",
1189 target->obd_name, cluuid.uuid,
1190 libcfs_nid2str(req->rq_peer.nid),
1192 export->exp_connection->c_peer.nid));
1194 class_export_put(export);
1197 export->exp_connecting = 1;
1198 spin_unlock(&export->exp_lock);
1199 LASSERT(export->exp_obd == target);
1201 rc = target_handle_reconnect(&conn, export, &cluuid);
1204 /* If we found an export, we already unlocked. */
1207 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1208 } else if (req->rq_export == NULL &&
1209 atomic_read(&export->exp_rpc_count) > 0) {
1210 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1211 "still busy with %d references\n",
1212 target->obd_name, cluuid.uuid,
1213 libcfs_nid2str(req->rq_peer.nid),
1214 atomic_read(&export->exp_refcount));
1215 GOTO(out, rc = -EBUSY);
1216 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1217 if (!strstr(cluuid.uuid, "mdt"))
1218 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1219 "known client %s (at %s) because it "
1220 "is indicating it is a new client",
1221 target->obd_name, cluuid.uuid,
1222 libcfs_nid2str(req->rq_peer.nid));
1223 GOTO(out, rc = -EALREADY);
1225 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1232 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1233 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1234 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1235 export, (long)cfs_time_current_sec(),
1236 export ? (long)export->exp_last_request_time : 0);
1238 /* If this is the first time a client connects, reset the recovery
1239 * timer. Discard lightweight connections which might be local. */
1240 if (!lw_client && rc == 0 && target->obd_recovering)
1241 check_and_start_recovery_timer(target, req, export == NULL);
1243 /* We want to handle EALREADY but *not* -EALREADY from
1244 * target_handle_reconnect(), return reconnection state in a flag. */
1245 if (rc == EALREADY) {
1246 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1252 /* Tell the client if we support replayable requests. */
1253 if (target->obd_replayable)
1254 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1255 client_nid = &req->rq_peer.nid;
1257 if (export == NULL) {
1258 /* allow lightweight connections during recovery */
1259 /* allow "new" MDT to be connected during recovery, since we
1260 * need retrieve recovery update records from it */
1261 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1263 int c; /* connected */
1264 int i; /* in progress */
1266 int s; /* stale/evicted */
1268 c = atomic_read(&target->obd_connected_clients);
1269 i = atomic_read(&target->obd_lock_replay_clients);
1270 k = target->obd_max_recoverable_clients;
1271 s = target->obd_stale_clients;
1272 t = cfs_timer_deadline(&target->obd_recovery_timer);
1273 t = cfs_time_sub(t, cfs_time_current());
1274 t = cfs_duration_sec(t);
1275 LCONSOLE_WARN("%s: Denying connection for new client %s"
1276 "(at %s), waiting for %d known clients "
1277 "(%d recovered, %d in progress, and %d "
1278 "evicted) to recover in %d:%.02d\n",
1279 target->obd_name, cluuid.uuid,
1280 libcfs_nid2str(req->rq_peer.nid), k,
1281 c - i, i, s, (int)t / 60,
1286 rc = obd_connect(req->rq_svc_thread->t_env,
1287 &export, target, &cluuid, data,
1289 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1290 lustre_msg_add_op_flags(req->rq_repmsg,
1291 MSG_CONNECT_RECOVERING);
1293 conn.cookie = export->exp_handle.h_cookie;
1294 rc = rev_import_init(export);
1298 new_mds_mds_conn = true;
1301 rc = obd_reconnect(req->rq_svc_thread->t_env,
1302 export, target, &cluuid, data, client_nid);
1307 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1308 data->ocd_instance = target->u.obt.obt_instance;
1310 /* Return only the parts of obd_connect_data that we understand, so the
1311 * client knows that we don't understand the rest. */
1313 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1315 tmpdata = req_capsule_server_get(&req->rq_pill,
1317 /* Don't use struct assignment here, because the client reply
1318 * buffer may be smaller/larger than the local struct
1319 * obd_connect_data. */
1320 memcpy(tmpdata, data, min(tmpsize, size));
1323 /* If the client and the server are the same node, we will already
1324 * have an export that really points to the client's DLM export,
1325 * because we have a shared handles table.
1327 * XXX this will go away when shaver stops sending the "connect" handle
1328 * in the real "remote handle" field of the request --phik 24 Apr 2003
1330 ptlrpc_request_change_export(req, export);
1332 spin_lock(&export->exp_lock);
1333 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1334 spin_unlock(&export->exp_lock);
1335 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1336 "or equal conn_cnt: %d >= %d\n",
1337 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1338 export->exp_conn_cnt,
1339 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1341 GOTO(out, rc = -EALREADY);
1343 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1344 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1346 /* Don't evict liblustre clients for not pinging. */
1347 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1348 export->exp_libclient = 1;
1349 spin_unlock(&export->exp_lock);
1351 spin_lock(&target->obd_dev_lock);
1352 list_del_init(&export->exp_obd_chain_timed);
1353 spin_unlock(&target->obd_dev_lock);
1355 spin_unlock(&export->exp_lock);
1358 if (export->exp_connection != NULL) {
1359 /* Check to see if connection came from another NID. */
1360 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1361 !hlist_unhashed(&export->exp_nid_hash))
1362 cfs_hash_del(export->exp_obd->obd_nid_hash,
1363 &export->exp_connection->c_peer.nid,
1364 &export->exp_nid_hash);
1366 ptlrpc_connection_put(export->exp_connection);
1369 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1372 if (hlist_unhashed(&export->exp_nid_hash))
1373 cfs_hash_add(export->exp_obd->obd_nid_hash,
1374 &export->exp_connection->c_peer.nid,
1375 &export->exp_nid_hash);
1377 lustre_msg_set_handle(req->rq_repmsg, &conn);
1379 rc = rev_import_reconnect(export, req);
1383 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1385 __u64 transno = data->ocd_transno;
1387 spin_lock(&export->exp_lock);
1388 /* possible race with class_disconnect_stale_exports,
1389 * export may be already in the eviction process */
1390 if (export->exp_failed) {
1391 spin_unlock(&export->exp_lock);
1392 GOTO(out, rc = -ENODEV);
1394 export->exp_in_recovery = 1;
1395 export->exp_req_replay_needed = 1;
1396 export->exp_lock_replay_needed = 1;
1397 spin_unlock(&export->exp_lock);
1399 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1400 MSG_CONNECT_TRANSNO);
1401 if (has_transno && transno == 0)
1402 CWARN("Connect with zero transno!\n");
1404 if (has_transno && transno > 0 &&
1405 transno < target->obd_next_recovery_transno &&
1406 transno > target->obd_last_committed) {
1407 /* Another way is to use cmpxchg() to be lock-free. */
1408 spin_lock(&target->obd_recovery_task_lock);
1409 if (transno < target->obd_next_recovery_transno)
1410 target->obd_next_recovery_transno = transno;
1411 spin_unlock(&target->obd_recovery_task_lock);
1414 atomic_inc(&target->obd_req_replay_clients);
1415 atomic_inc(&target->obd_lock_replay_clients);
1416 /* Note: MDS-MDS connection is allowed to be connected during
1417 * recovery, no matter if the exports needs to be recoveried.
1418 * Because we need retrieve updates logs from all other MDTs.
1419 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1420 * also needs to be increased to match other recovery checking
1422 if (new_mds_mds_conn)
1423 target->obd_max_recoverable_clients++;
1424 if (atomic_inc_return(&target->obd_connected_clients) ==
1425 target->obd_max_recoverable_clients)
1426 wake_up(&target->obd_next_transno_waitq);
1429 /* Tell the client we're in recovery, when client is involved in it. */
1430 if (target->obd_recovering && !lw_client)
1431 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1435 spin_lock(&export->exp_lock);
1436 export->exp_connecting = 0;
1437 spin_unlock(&export->exp_lock);
1439 class_export_put(export);
1441 if (target_referenced == true && target != NULL) {
1442 spin_lock(&target->obd_dev_lock);
1443 target->obd_conn_inprogress--;
1444 spin_unlock(&target->obd_dev_lock);
1446 class_decref(target, __func__, current);
1448 req->rq_status = rc;
1452 int target_handle_disconnect(struct ptlrpc_request *req)
1457 rc = req_capsule_server_pack(&req->rq_pill);
1461 /* Keep the rq_export around so we can send the reply. */
1462 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1467 void target_destroy_export(struct obd_export *exp)
1469 struct obd_import *imp = NULL;
1470 /* exports created from last_rcvd data, and "fake"
1471 exports created by lctl don't have an import */
1472 spin_lock(&exp->exp_lock);
1473 if (exp->exp_imp_reverse != NULL) {
1474 imp = exp->exp_imp_reverse;
1475 exp->exp_imp_reverse = NULL;
1477 spin_unlock(&exp->exp_lock);
1479 client_destroy_import(imp);
1481 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1482 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1483 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1484 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1486 EXPORT_SYMBOL(target_destroy_export);
1489 * Recovery functions
1491 static void target_request_copy_get(struct ptlrpc_request *req)
1493 class_export_rpc_inc(req->rq_export);
1494 LASSERT(list_empty(&req->rq_list));
1495 INIT_LIST_HEAD(&req->rq_replay_list);
1497 /* Increase refcount to keep request in queue. */
1498 atomic_inc(&req->rq_refcount);
1499 /* Let export know it has replays to be handled. */
1500 atomic_inc(&req->rq_export->exp_replay_count);
1503 static void target_request_copy_put(struct ptlrpc_request *req)
1505 LASSERT(list_empty(&req->rq_replay_list));
1506 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1508 atomic_dec(&req->rq_export->exp_replay_count);
1509 class_export_rpc_dec(req->rq_export);
1510 ptlrpc_server_drop_request(req);
1513 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1515 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1516 struct obd_export *exp = req->rq_export;
1517 struct ptlrpc_request *reqiter;
1518 struct ptlrpc_request *dup_req = NULL;
1523 spin_lock(&exp->exp_lock);
1524 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1526 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1534 /* We expect it with RESENT and REPLAY flags. */
1535 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1536 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1537 CERROR("invalid flags %x of resent replay\n",
1538 lustre_msg_get_flags(req->rq_reqmsg));
1540 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1543 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1545 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1546 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1550 list_add_tail(&req->rq_replay_list,
1551 &exp->exp_req_replay_queue);
1554 spin_unlock(&exp->exp_lock);
1558 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1560 LASSERT(!list_empty(&req->rq_replay_list));
1561 LASSERT(req->rq_export);
1563 spin_lock(&req->rq_export->exp_lock);
1564 list_del_init(&req->rq_replay_list);
1565 spin_unlock(&req->rq_export->exp_lock);
1568 static void target_finish_recovery(struct lu_target *lut)
1570 struct obd_device *obd = lut->lut_obd;
1573 /* Only log a recovery message when recovery has occurred. */
1574 if (obd->obd_recovery_start) {
1575 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1576 obd->obd_recovery_start);
1577 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1578 "%d recovered and %d %s evicted.\n", obd->obd_name,
1579 (int)elapsed_time / 60, (int)elapsed_time % 60,
1580 obd->obd_max_recoverable_clients,
1581 atomic_read(&obd->obd_connected_clients),
1582 obd->obd_stale_clients,
1583 obd->obd_stale_clients == 1 ? "was" : "were");
1586 ldlm_reprocess_all_ns(obd->obd_namespace);
1587 spin_lock(&obd->obd_recovery_task_lock);
1588 if (!list_empty(&obd->obd_req_replay_queue) ||
1589 !list_empty(&obd->obd_lock_replay_queue) ||
1590 !list_empty(&obd->obd_final_req_queue)) {
1591 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1593 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1594 list_empty(&obd->obd_lock_replay_queue) ? \
1596 list_empty(&obd->obd_final_req_queue) ? \
1598 spin_unlock(&obd->obd_recovery_task_lock);
1601 spin_unlock(&obd->obd_recovery_task_lock);
1603 obd->obd_recovery_end = cfs_time_current_sec();
1605 /* When recovery finished, cleanup orphans on MDS and OST. */
1606 if (OBT(obd) && OBP(obd, postrecov)) {
1607 int rc = OBP(obd, postrecov)(obd);
1609 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1615 static void abort_req_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_req_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_WARNING, req, "aborted:");
1626 req->rq_status = -ENOTCONN;
1627 if (ptlrpc_error(req)) {
1628 DEBUG_REQ(D_ERROR, req,
1629 "failed abort_req_reply; skipping");
1631 target_exp_dequeue_req_replay(req);
1632 target_request_copy_put(req);
1636 static void abort_lock_replay_queue(struct obd_device *obd)
1638 struct ptlrpc_request *req, *n;
1639 struct list_head abort_list;
1641 INIT_LIST_HEAD(&abort_list);
1642 spin_lock(&obd->obd_recovery_task_lock);
1643 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1644 spin_unlock(&obd->obd_recovery_task_lock);
1645 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1646 DEBUG_REQ(D_ERROR, req, "aborted:");
1647 req->rq_status = -ENOTCONN;
1648 if (ptlrpc_error(req)) {
1649 DEBUG_REQ(D_ERROR, req,
1650 "failed abort_lock_reply; skipping");
1652 target_request_copy_put(req);
1656 /* Called from a cleanup function if the device is being cleaned up
1657 forcefully. The exports should all have been disconnected already,
1658 the only thing left to do is
1659 - clear the recovery flags
1661 - free queued requests and replies, but don't send replies
1662 Because the obd_stopping flag is set, no new requests should be received.
1665 void target_cleanup_recovery(struct obd_device *obd)
1667 struct ptlrpc_request *req, *n;
1668 struct list_head clean_list;
1670 INIT_LIST_HEAD(&clean_list);
1671 spin_lock(&obd->obd_dev_lock);
1672 if (!obd->obd_recovering) {
1673 spin_unlock(&obd->obd_dev_lock);
1677 obd->obd_recovering = obd->obd_abort_recovery = 0;
1678 spin_unlock(&obd->obd_dev_lock);
1680 spin_lock(&obd->obd_recovery_task_lock);
1681 target_cancel_recovery_timer(obd);
1682 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1683 spin_unlock(&obd->obd_recovery_task_lock);
1685 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1686 LASSERT(req->rq_reply_state == NULL);
1687 target_exp_dequeue_req_replay(req);
1688 target_request_copy_put(req);
1691 spin_lock(&obd->obd_recovery_task_lock);
1692 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1693 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1694 spin_unlock(&obd->obd_recovery_task_lock);
1696 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1697 LASSERT(req->rq_reply_state == NULL);
1698 target_request_copy_put(req);
1703 EXPORT_SYMBOL(target_cleanup_recovery);
1705 /* obd_recovery_task_lock should be held */
1706 void target_cancel_recovery_timer(struct obd_device *obd)
1708 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1709 cfs_timer_disarm(&obd->obd_recovery_timer);
1712 static void target_start_recovery_timer(struct obd_device *obd)
1714 if (obd->obd_recovery_start != 0)
1717 spin_lock(&obd->obd_dev_lock);
1718 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1719 spin_unlock(&obd->obd_dev_lock);
1723 LASSERT(obd->obd_recovery_timeout != 0);
1725 if (obd->obd_recovery_start != 0) {
1726 spin_unlock(&obd->obd_dev_lock);
1730 cfs_timer_arm(&obd->obd_recovery_timer,
1731 cfs_time_shift(obd->obd_recovery_timeout));
1732 obd->obd_recovery_start = cfs_time_current_sec();
1733 spin_unlock(&obd->obd_dev_lock);
1735 LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1736 "or until %d client%s reconnect%s\n",
1738 obd->obd_recovery_timeout / 60,
1739 obd->obd_recovery_timeout % 60,
1740 obd->obd_max_recoverable_clients,
1741 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1742 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1746 * extend recovery window.
1748 * if @extend is true, extend recovery window to have @drt remaining at least;
1749 * otherwise, make sure the recovery timeout value is not less than @drt.
1751 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1755 cfs_duration_t left;
1758 spin_lock(&obd->obd_dev_lock);
1759 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1760 spin_unlock(&obd->obd_dev_lock);
1763 LASSERT(obd->obd_recovery_start != 0);
1765 now = cfs_time_current_sec();
1766 to = obd->obd_recovery_timeout;
1767 end = obd->obd_recovery_start + to;
1768 left = cfs_time_sub(end, now);
1770 if (extend && (drt > left)) {
1772 } else if (!extend && (drt > to)) {
1776 if (to > obd->obd_recovery_time_hard)
1777 to = obd->obd_recovery_time_hard;
1778 if (obd->obd_recovery_timeout < to) {
1779 obd->obd_recovery_timeout = to;
1780 end = obd->obd_recovery_start + to;
1781 cfs_timer_arm(&obd->obd_recovery_timer,
1782 cfs_time_shift(end - now));
1784 spin_unlock(&obd->obd_dev_lock);
1786 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1787 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1790 /* Reset the timer with each new client connection */
1792 * This timer is actually reconnect_timer, which is for making sure
1793 * the total recovery window is at least as big as my reconnect
1794 * attempt timing. So the initial recovery time_out will be set to
1795 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1796 * from client is bigger than this, then the recovery time_out will
1797 * be extended to make sure the client could be reconnected, in the
1798 * process, the timeout from the new client should be ignored.
1802 check_and_start_recovery_timer(struct obd_device *obd,
1803 struct ptlrpc_request *req,
1806 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1807 struct obd_device_target *obt = &obd->u.obt;
1809 if (!new_client && service_time)
1810 /* Teach server about old server's estimates, as first guess
1811 * at how long new requests will take. */
1812 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1815 target_start_recovery_timer(obd);
1817 /* Convert the service time to RPC timeout,
1818 * and reuse service_time to limit stack usage. */
1819 service_time = at_est2timeout(service_time);
1821 /* We expect other clients to timeout within service_time, then try
1822 * to reconnect, then try the failover server. The max delay between
1823 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1824 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1826 LASSERT(obt->obt_magic == OBT_MAGIC);
1827 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1828 if (service_time > obd->obd_recovery_timeout && !new_client)
1829 extend_recovery_timer(obd, service_time, false);
1832 /** Health checking routines */
1833 static inline int exp_connect_healthy(struct obd_export *exp)
1835 return (exp->exp_in_recovery);
1838 /** if export done req_replay or has replay in queue */
1839 static inline int exp_req_replay_healthy(struct obd_export *exp)
1841 return (!exp->exp_req_replay_needed ||
1842 atomic_read(&exp->exp_replay_count) > 0);
1846 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1848 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1849 exp_req_replay_healthy(exp);
1852 /** if export done lock_replay or has replay in queue */
1853 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1855 return (!exp->exp_lock_replay_needed ||
1856 atomic_read(&exp->exp_replay_count) > 0);
1859 static inline int exp_vbr_healthy(struct obd_export *exp)
1861 return (!exp->exp_vbr_failed);
1864 static inline int exp_finished(struct obd_export *exp)
1866 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1869 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1871 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1875 static int check_for_next_transno(struct lu_target *lut)
1877 struct ptlrpc_request *req = NULL;
1878 struct obd_device *obd = lut->lut_obd;
1879 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1880 int wake_up = 0, connected, completed, queue_len;
1881 __u64 req_transno = 0;
1882 __u64 update_transno = 0;
1883 __u64 next_transno = 0;
1886 spin_lock(&obd->obd_recovery_task_lock);
1887 if (!list_empty(&obd->obd_req_replay_queue)) {
1888 req = list_entry(obd->obd_req_replay_queue.next,
1889 struct ptlrpc_request, rq_list);
1890 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1894 update_transno = distribute_txn_get_next_transno(tdtd);
1896 connected = atomic_read(&obd->obd_connected_clients);
1897 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1898 queue_len = obd->obd_requests_queued_for_recovery;
1899 next_transno = obd->obd_next_recovery_transno;
1901 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1902 "req_transno: %llu, next_transno: %llu\n",
1903 obd->obd_max_recoverable_clients, connected, completed,
1904 queue_len, req_transno, next_transno);
1906 if (obd->obd_abort_recovery) {
1907 CDEBUG(D_HA, "waking for aborted recovery\n");
1909 } else if (obd->obd_recovery_expired) {
1910 CDEBUG(D_HA, "waking for expired recovery\n");
1912 } else if (tdtd != NULL && req != NULL &&
1913 is_req_replayed_by_update(req)) {
1914 LASSERTF(req_transno < next_transno, "req_transno %llu"
1915 "next_transno%llu\n", req_transno, next_transno);
1916 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1919 } else if (req_transno == next_transno ||
1920 (update_transno != 0 && update_transno <= next_transno)) {
1921 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1923 } else if (queue_len > 0 &&
1924 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1925 /** handle gaps occured due to lost reply or VBR */
1926 LASSERTF(req_transno >= next_transno,
1927 "req_transno: %llu, next_transno: %llu\n",
1928 req_transno, next_transno);
1930 "%s: waking for gap in transno, VBR is %s (skip: "
1931 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1932 ", next_update %lld last_committed: %lld)\n",
1933 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1934 next_transno, queue_len, completed, connected,
1935 req_transno, update_transno, obd->obd_last_committed);
1936 obd->obd_next_recovery_transno = req_transno;
1938 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1939 CDEBUG(D_HA, "waking for completed recovery\n");
1941 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1942 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1943 " by fail_lock, waking up (%lld)\n", next_transno);
1944 obd->obd_next_recovery_transno = req_transno;
1947 spin_unlock(&obd->obd_recovery_task_lock);
1951 static int check_for_next_lock(struct lu_target *lut)
1953 struct obd_device *obd = lut->lut_obd;
1956 spin_lock(&obd->obd_recovery_task_lock);
1957 if (!list_empty(&obd->obd_lock_replay_queue)) {
1958 CDEBUG(D_HA, "waking for next lock\n");
1960 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1961 CDEBUG(D_HA, "waking for completed lock replay\n");
1963 } else if (obd->obd_abort_recovery) {
1964 CDEBUG(D_HA, "waking for aborted recovery\n");
1966 } else if (obd->obd_recovery_expired) {
1967 CDEBUG(D_HA, "waking for expired recovery\n");
1970 spin_unlock(&obd->obd_recovery_task_lock);
1976 * wait for recovery events,
1977 * check its status with help of check_routine
1978 * evict dead clients via health_check
1980 static int target_recovery_overseer(struct lu_target *lut,
1981 int (*check_routine)(struct lu_target *),
1982 int (*health_check)(struct obd_export *))
1984 struct obd_device *obd = lut->lut_obd;
1985 struct target_distribute_txn_data *tdtd;
1987 if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1988 (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1989 __u64 next_update_transno = 0;
1991 /* Only abort the recovery if there are no update recovery
1992 * left in the queue */
1993 spin_lock(&obd->obd_recovery_task_lock);
1994 if (lut->lut_tdtd != NULL) {
1995 next_update_transno =
1996 distribute_txn_get_next_transno(lut->lut_tdtd);
1998 tdtd = lut->lut_tdtd;
1999 /* If next_update_transno == 0, it probably because
2000 * updatelog retrieve threads did not get any records
2001 * yet, let's wait those threads stopped */
2002 if (next_update_transno == 0) {
2003 struct l_wait_info lwi = { 0 };
2005 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2007 &tdtd->tdtd_recovery_threads_count) == 0,
2010 next_update_transno =
2011 distribute_txn_get_next_transno(
2016 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2017 obd->obd_next_recovery_transno = next_update_transno;
2018 spin_unlock(&obd->obd_recovery_task_lock);
2019 /* Disconnect unfinished exports from clients, and
2020 * keep connection from MDT to make sure the update
2021 * recovery will still keep trying until some one
2022 * manually abort the recovery */
2023 class_disconnect_stale_exports(obd,
2024 exp_finished_or_from_mdt);
2025 /* Abort all of replay and replay lock req from
2027 abort_req_replay_queue(obd);
2028 abort_lock_replay_queue(obd);
2029 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2030 ")in the queue.\n", obd->obd_name,
2031 next_update_transno);
2033 obd->obd_abort_recovery = 1;
2034 spin_unlock(&obd->obd_recovery_task_lock);
2035 CWARN("%s recovery is aborted by hard timeout\n",
2040 while (wait_event_timeout(obd->obd_next_transno_waitq,
2042 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2043 /* wait indefinitely for event, but don't trigger watchdog */;
2045 if (obd->obd_abort_recovery) {
2046 CWARN("recovery is aborted, evict exports in recovery\n");
2047 if (lut->lut_tdtd != NULL) {
2048 struct l_wait_info lwi = { 0 };
2050 tdtd = lut->lut_tdtd;
2051 /* Let's wait all of the update log recovery thread
2053 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2054 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2056 /* Then abort the update recovery list */
2057 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2058 dtrq_list_destroy(lut->lut_tdtd);
2061 /** evict exports which didn't finish recovery yet */
2062 class_disconnect_stale_exports(obd, exp_finished);
2064 } else if (obd->obd_recovery_expired) {
2065 obd->obd_recovery_expired = 0;
2066 /** If some clients died being recovered, evict them */
2067 LCONSOLE_WARN("%s: recovery is timed out, "
2068 "evict stale exports\n", obd->obd_name);
2069 /** evict cexports with no replay in queue, they are stalled */
2070 class_disconnect_stale_exports(obd, health_check);
2072 /** continue with VBR */
2073 spin_lock(&obd->obd_dev_lock);
2074 obd->obd_version_recov = 1;
2075 spin_unlock(&obd->obd_dev_lock);
2077 * reset timer, recovery will proceed with versions now,
2078 * timeout is set just to handle reconnection delays
2080 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2081 /** Wait for recovery events again, after evicting bad clients */
2087 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2089 struct obd_device *obd = lut->lut_obd;
2090 struct ptlrpc_request *req = NULL;
2092 CDEBUG(D_HA, "Waiting for lock\n");
2093 if (target_recovery_overseer(lut, check_for_next_lock,
2094 exp_lock_replay_healthy))
2095 abort_lock_replay_queue(obd);
2097 spin_lock(&obd->obd_recovery_task_lock);
2098 if (!list_empty(&obd->obd_lock_replay_queue)) {
2099 req = list_entry(obd->obd_lock_replay_queue.next,
2100 struct ptlrpc_request, rq_list);
2101 list_del_init(&req->rq_list);
2102 spin_unlock(&obd->obd_recovery_task_lock);
2104 spin_unlock(&obd->obd_recovery_task_lock);
2105 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2106 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2107 /** evict exports failed VBR */
2108 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2113 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2115 struct ptlrpc_request *req = NULL;
2117 spin_lock(&obd->obd_recovery_task_lock);
2118 if (!list_empty(&obd->obd_final_req_queue)) {
2119 req = list_entry(obd->obd_final_req_queue.next,
2120 struct ptlrpc_request, rq_list);
2121 list_del_init(&req->rq_list);
2122 spin_unlock(&obd->obd_recovery_task_lock);
2123 if (req->rq_export->exp_in_recovery) {
2124 spin_lock(&req->rq_export->exp_lock);
2125 req->rq_export->exp_in_recovery = 0;
2126 spin_unlock(&req->rq_export->exp_lock);
2129 spin_unlock(&obd->obd_recovery_task_lock);
2134 static void handle_recovery_req(struct ptlrpc_thread *thread,
2135 struct ptlrpc_request *req,
2136 svc_handler_t handler)
2141 * export can be evicted during recovery, no need to handle replays for
2142 * it after that, discard such request silently
2144 if (req->rq_export->exp_disconnected)
2147 req->rq_session.lc_thread = thread;
2148 req->rq_svc_thread = thread;
2149 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2151 /* thread context */
2152 lu_context_enter(&thread->t_env->le_ctx);
2154 lu_context_exit(&thread->t_env->le_ctx);
2156 /* don't reset timer for final stage */
2157 if (!exp_finished(req->rq_export)) {
2158 int to = obd_timeout;
2161 * Add request timeout to the recovery time so next request from
2162 * this client may come in recovery time
2165 struct ptlrpc_service_part *svcpt;
2167 svcpt = req->rq_rqbd->rqbd_svcpt;
2168 /* If the server sent early reply for this request,
2169 * the client will recalculate the timeout according to
2170 * current server estimate service time, so we will
2171 * use the maxium timeout here for waiting the client
2172 * sending the next req */
2173 to = max((int)at_est2timeout(
2174 at_get(&svcpt->scp_at_estimate)),
2175 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2176 /* Add 2 net_latency, one for balance rq_deadline
2177 * (see ptl_send_rpc), one for resend the req to server,
2178 * Note: client will pack net_latency in replay req
2179 * (see ptlrpc_replay_req) */
2180 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2182 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2187 /** Checking routines for recovery */
2188 static int check_for_recovery_ready(struct lu_target *lut)
2190 struct obd_device *obd = lut->lut_obd;
2191 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2193 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2194 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2195 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2196 obd->obd_recovery_expired);
2198 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2199 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2200 if (clnts + obd->obd_stale_clients <
2201 obd->obd_max_recoverable_clients)
2205 if (lut->lut_tdtd != NULL) {
2206 if (!lut->lut_tdtd->tdtd_replay_ready &&
2207 !obd->obd_abort_recovery) {
2208 /* Let's extend recovery timer, in case the recovery
2209 * timer expired, and some clients got evicted */
2210 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2212 CDEBUG(D_HA, "%s update recovery is not ready,"
2213 " extend recovery %d\n", obd->obd_name,
2214 obd->obd_recovery_timeout);
2217 dtrq_list_dump(lut->lut_tdtd, D_HA);
2225 REQUEST_RECOVERY = 1,
2226 UPDATE_RECOVERY = 2,
2229 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2233 if (!list_empty(&obd->obd_req_replay_queue)) {
2234 struct ptlrpc_request *req;
2236 req = list_entry(obd->obd_req_replay_queue.next,
2237 struct ptlrpc_request, rq_list);
2238 transno = lustre_msg_get_transno(req->rq_reqmsg);
2244 static __u64 get_next_transno(struct lu_target *lut, int *type)
2246 struct obd_device *obd = lut->lut_obd;
2247 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2249 __u64 update_transno;
2252 transno = get_next_replay_req_transno(obd);
2254 *type = REQUEST_RECOVERY;
2259 update_transno = distribute_txn_get_next_transno(tdtd);
2260 if (transno == 0 || (transno >= update_transno &&
2261 update_transno != 0)) {
2262 transno = update_transno;
2264 *type = UPDATE_RECOVERY;
2271 * drop duplicate replay request
2273 * Because the operation has been replayed by update recovery, the request
2274 * with the same transno will be dropped and also notify the client to send
2275 * next replay request.
2277 * \param[in] env execution environment
2278 * \param[in] obd failover obd device
2279 * \param[in] req request to be dropped
2281 static void drop_duplicate_replay_req(struct lu_env *env,
2282 struct obd_device *obd,
2283 struct ptlrpc_request *req)
2285 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2286 " update records are found.\n",
2287 lustre_msg_get_transno(req->rq_reqmsg),
2288 libcfs_nid2str(req->rq_peer.nid));
2290 /* Right now, only for MDS reint operation update replay and
2291 * normal request replay can have the same transno */
2292 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2293 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2294 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2295 if (likely(req->rq_export))
2296 target_committed_to_req(req);
2297 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2298 target_send_reply(req, req->rq_status, 0);
2300 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2301 libcfs_nid2str(req->rq_peer.nid));
2303 target_exp_dequeue_req_replay(req);
2304 target_request_copy_put(req);
2305 obd->obd_replayed_requests++;
2308 static void replay_request_or_update(struct lu_env *env,
2309 struct lu_target *lut,
2310 struct target_recovery_data *trd,
2311 struct ptlrpc_thread *thread)
2313 struct obd_device *obd = lut->lut_obd;
2314 struct ptlrpc_request *req = NULL;
2319 CDEBUG(D_HA, "Waiting for transno %lld\n",
2320 obd->obd_next_recovery_transno);
2322 /* Replay all of request and update by transno */
2324 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2326 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2328 /** It is needed to extend recovery window above
2329 * recovery_time_soft. Extending is possible only in the
2330 * end of recovery window (see more details in
2331 * handle_recovery_req()).
2333 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2335 if (target_recovery_overseer(lut, check_for_next_transno,
2336 exp_req_replay_healthy_or_from_mdt)) {
2337 abort_req_replay_queue(obd);
2338 abort_lock_replay_queue(obd);
2342 spin_lock(&obd->obd_recovery_task_lock);
2343 transno = get_next_transno(lut, &type);
2344 if (type == REQUEST_RECOVERY && transno != 0) {
2345 /* Drop replay request from client side, if the
2346 * replay has been executed by update with the
2348 req = list_entry(obd->obd_req_replay_queue.next,
2349 struct ptlrpc_request, rq_list);
2351 list_del_init(&req->rq_list);
2352 obd->obd_requests_queued_for_recovery--;
2353 spin_unlock(&obd->obd_recovery_task_lock);
2355 /* Let's check if the request has been redone by
2357 if (is_req_replayed_by_update(req)) {
2358 struct distribute_txn_replay_req *dtrq;
2360 dtrq = distribute_txn_lookup_finish_list(tdtd,
2362 LASSERT(dtrq != NULL);
2363 spin_lock(&tdtd->tdtd_replay_list_lock);
2364 list_del_init(&dtrq->dtrq_list);
2365 spin_unlock(&tdtd->tdtd_replay_list_lock);
2368 drop_duplicate_replay_req(env, obd, req);
2373 LASSERT(trd->trd_processing_task == current_pid());
2374 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2375 lustre_msg_get_transno(req->rq_reqmsg),
2376 libcfs_nid2str(req->rq_peer.nid));
2378 handle_recovery_req(thread, req,
2379 trd->trd_recovery_handler);
2381 * bz18031: increase next_recovery_transno before
2382 * target_request_copy_put() will drop exp_rpc reference
2384 spin_lock(&obd->obd_recovery_task_lock);
2385 obd->obd_next_recovery_transno++;
2386 spin_unlock(&obd->obd_recovery_task_lock);
2387 target_exp_dequeue_req_replay(req);
2388 target_request_copy_put(req);
2389 obd->obd_replayed_requests++;
2390 } else if (type == UPDATE_RECOVERY && transno != 0) {
2391 struct distribute_txn_replay_req *dtrq;
2394 spin_unlock(&obd->obd_recovery_task_lock);
2396 LASSERT(tdtd != NULL);
2397 dtrq = distribute_txn_get_next_req(tdtd);
2398 lu_context_enter(&thread->t_env->le_ctx);
2399 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2400 lu_context_exit(&thread->t_env->le_ctx);
2401 extend_recovery_timer(obd, obd_timeout, true);
2403 if (rc == 0 && dtrq->dtrq_xid != 0) {
2404 CDEBUG(D_HA, "Move x%llu t%llu"
2405 " to finish list\n", dtrq->dtrq_xid,
2406 dtrq->dtrq_master_transno);
2408 /* Add it to the replay finish list */
2409 spin_lock(&tdtd->tdtd_replay_list_lock);
2410 list_add(&dtrq->dtrq_list,
2411 &tdtd->tdtd_replay_finish_list);
2412 spin_unlock(&tdtd->tdtd_replay_list_lock);
2414 spin_lock(&obd->obd_recovery_task_lock);
2415 if (transno == obd->obd_next_recovery_transno)
2416 obd->obd_next_recovery_transno++;
2418 obd->obd_next_recovery_transno)
2419 obd->obd_next_recovery_transno =
2421 spin_unlock(&obd->obd_recovery_task_lock);
2426 spin_unlock(&obd->obd_recovery_task_lock);
2428 LASSERT(list_empty(&obd->obd_req_replay_queue));
2429 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2430 /** evict exports failed VBR */
2431 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2437 static int target_recovery_thread(void *arg)
2439 struct lu_target *lut = arg;
2440 struct obd_device *obd = lut->lut_obd;
2441 struct ptlrpc_request *req;
2442 struct target_recovery_data *trd = &obd->obd_recovery_data;
2443 unsigned long delta;
2445 struct ptlrpc_thread *thread = NULL;
2449 unshare_fs_struct();
2450 OBD_ALLOC_PTR(thread);
2456 OBD_FREE_PTR(thread);
2460 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2462 OBD_FREE_PTR(thread);
2467 thread->t_env = env;
2468 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2469 env->le_ctx.lc_thread = thread;
2470 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2471 thread->t_watchdog = NULL;
2473 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2475 trd->trd_processing_task = current_pid();
2477 spin_lock(&obd->obd_dev_lock);
2478 obd->obd_recovering = 1;
2479 spin_unlock(&obd->obd_dev_lock);
2480 complete(&trd->trd_starting);
2482 /* first of all, we have to know the first transno to replay */
2483 if (target_recovery_overseer(lut, check_for_recovery_ready,
2484 exp_connect_healthy)) {
2485 abort_req_replay_queue(obd);
2486 abort_lock_replay_queue(obd);
2487 if (lut->lut_tdtd != NULL)
2488 dtrq_list_destroy(lut->lut_tdtd);
2491 /* next stage: replay requests or update */
2493 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2494 atomic_read(&obd->obd_req_replay_clients),
2495 obd->obd_next_recovery_transno);
2496 replay_request_or_update(env, lut, trd, thread);
2499 * The second stage: replay locks
2501 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2502 atomic_read(&obd->obd_lock_replay_clients));
2503 while ((req = target_next_replay_lock(lut))) {
2504 LASSERT(trd->trd_processing_task == current_pid());
2505 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2506 libcfs_nid2str(req->rq_peer.nid));
2507 handle_recovery_req(thread, req,
2508 trd->trd_recovery_handler);
2509 target_request_copy_put(req);
2510 obd->obd_replayed_locks++;
2514 * The third stage: reply on final pings, at this moment all clients
2515 * must have request in final queue
2517 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2518 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2519 /** Update server last boot epoch */
2520 tgt_boot_epoch_update(lut);
2521 /* We drop recoverying flag to forward all new requests
2522 * to regular mds_handle() since now */
2523 spin_lock(&obd->obd_dev_lock);
2524 obd->obd_recovering = obd->obd_abort_recovery = 0;
2525 spin_unlock(&obd->obd_dev_lock);
2526 spin_lock(&obd->obd_recovery_task_lock);
2527 target_cancel_recovery_timer(obd);
2528 spin_unlock(&obd->obd_recovery_task_lock);
2529 while ((req = target_next_final_ping(obd))) {
2530 LASSERT(trd->trd_processing_task == current_pid());
2531 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2532 libcfs_nid2str(req->rq_peer.nid));
2533 handle_recovery_req(thread, req,
2534 trd->trd_recovery_handler);
2535 /* Because the waiting client can not send ping to server,
2536 * so we need refresh the last_request_time, to avoid the
2537 * export is being evicted */
2538 ptlrpc_update_export_timer(req->rq_export, 0);
2539 target_request_copy_put(req);
2542 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2543 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2544 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2545 if (delta > OBD_RECOVERY_TIME_SOFT) {
2546 CWARN("too long recovery - read logs\n");
2547 libcfs_debug_dumplog();
2550 target_finish_recovery(lut);
2552 lu_context_fini(&env->le_ctx);
2553 trd->trd_processing_task = 0;
2554 complete(&trd->trd_finishing);
2556 tgt_io_thread_done(thread);
2557 OBD_FREE_PTR(thread);
2562 static int target_start_recovery_thread(struct lu_target *lut,
2563 svc_handler_t handler)
2565 struct obd_device *obd = lut->lut_obd;
2567 struct target_recovery_data *trd = &obd->obd_recovery_data;
2570 memset(trd, 0, sizeof(*trd));
2571 init_completion(&trd->trd_starting);
2572 init_completion(&trd->trd_finishing);
2573 trd->trd_recovery_handler = handler;
2575 rc = server_name2index(obd->obd_name, &index, NULL);
2579 if (!IS_ERR(kthread_run(target_recovery_thread,
2580 lut, "tgt_recover_%d", index))) {
2581 wait_for_completion(&trd->trd_starting);
2582 LASSERT(obd->obd_recovering != 0);
2590 void target_stop_recovery_thread(struct obd_device *obd)
2592 if (obd->obd_recovery_data.trd_processing_task > 0) {
2593 struct target_recovery_data *trd = &obd->obd_recovery_data;
2594 /** recovery can be done but postrecovery is not yet */
2595 spin_lock(&obd->obd_dev_lock);
2596 if (obd->obd_recovering) {
2597 CERROR("%s: Aborting recovery\n", obd->obd_name);
2598 obd->obd_abort_recovery = 1;
2599 wake_up(&obd->obd_next_transno_waitq);
2601 spin_unlock(&obd->obd_dev_lock);
2602 wait_for_completion(&trd->trd_finishing);
2605 EXPORT_SYMBOL(target_stop_recovery_thread);
2607 void target_recovery_fini(struct obd_device *obd)
2609 class_disconnect_exports(obd);
2610 target_stop_recovery_thread(obd);
2611 target_cleanup_recovery(obd);
2613 EXPORT_SYMBOL(target_recovery_fini);
2615 static void target_recovery_expired(unsigned long castmeharder)
2617 struct obd_device *obd = (struct obd_device *)castmeharder;
2618 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2619 " after %lds (%d clients connected)\n",
2620 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2621 cfs_time_current_sec()- obd->obd_recovery_start,
2622 atomic_read(&obd->obd_connected_clients));
2624 obd->obd_recovery_expired = 1;
2625 wake_up(&obd->obd_next_transno_waitq);
2628 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2630 struct obd_device *obd = lut->lut_obd;
2632 if (obd->obd_max_recoverable_clients == 0) {
2633 /** Update server last boot epoch */
2634 tgt_boot_epoch_update(lut);
2638 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2639 "last_transno %llu\n", obd->obd_name,
2640 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2641 LASSERT(obd->obd_stopping == 0);
2642 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2643 obd->obd_recovery_start = 0;
2644 obd->obd_recovery_end = 0;
2646 cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2647 target_start_recovery_thread(lut, handler);
2649 EXPORT_SYMBOL(target_recovery_init);
2651 static int target_process_req_flags(struct obd_device *obd,
2652 struct ptlrpc_request *req)
2654 struct obd_export *exp = req->rq_export;
2655 LASSERT(exp != NULL);
2656 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2657 /* client declares he's ready to replay locks */
2658 spin_lock(&exp->exp_lock);
2659 if (exp->exp_req_replay_needed) {
2660 exp->exp_req_replay_needed = 0;
2661 spin_unlock(&exp->exp_lock);
2663 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2664 atomic_dec(&obd->obd_req_replay_clients);
2666 spin_unlock(&exp->exp_lock);
2669 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2670 /* client declares he's ready to complete recovery
2671 * so, we put the request on th final queue */
2672 spin_lock(&exp->exp_lock);
2673 if (exp->exp_lock_replay_needed) {
2674 exp->exp_lock_replay_needed = 0;
2675 spin_unlock(&exp->exp_lock);
2677 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2678 atomic_dec(&obd->obd_lock_replay_clients);
2680 spin_unlock(&exp->exp_lock);
2686 int target_queue_recovery_request(struct ptlrpc_request *req,
2687 struct obd_device *obd)
2689 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2690 struct ptlrpc_request *reqiter;
2694 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2695 /* Processing the queue right now, don't re-add. */
2699 target_process_req_flags(obd, req);
2701 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2702 /* client declares he's ready to complete recovery
2703 * so, we put the request on th final queue */
2704 target_request_copy_get(req);
2705 DEBUG_REQ(D_HA, req, "queue final req");
2706 wake_up(&obd->obd_next_transno_waitq);
2707 spin_lock(&obd->obd_recovery_task_lock);
2708 if (obd->obd_recovering) {
2709 struct ptlrpc_request *tmp;
2710 struct ptlrpc_request *duplicate = NULL;
2712 if (likely(!req->rq_export->exp_replay_done)) {
2713 req->rq_export->exp_replay_done = 1;
2714 list_add_tail(&req->rq_list,
2715 &obd->obd_final_req_queue);
2716 spin_unlock(&obd->obd_recovery_task_lock);
2720 /* XXX O(n), but only happens if final ping is
2721 * timed out, probably reorganize the list as
2722 * a hash list later */
2723 list_for_each_entry_safe(reqiter, tmp,
2724 &obd->obd_final_req_queue,
2726 if (reqiter->rq_export == req->rq_export) {
2727 list_del_init(&reqiter->rq_list);
2728 duplicate = reqiter;
2733 list_add_tail(&req->rq_list,
2734 &obd->obd_final_req_queue);
2735 req->rq_export->exp_replay_done = 1;
2736 spin_unlock(&obd->obd_recovery_task_lock);
2738 if (duplicate != NULL) {
2739 DEBUG_REQ(D_HA, duplicate,
2740 "put prev final req\n");
2741 target_request_copy_put(duplicate);
2745 spin_unlock(&obd->obd_recovery_task_lock);
2746 target_request_copy_put(req);
2747 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2750 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2751 /* client declares he's ready to replay locks */
2752 target_request_copy_get(req);
2753 DEBUG_REQ(D_HA, req, "queue lock replay req");
2754 wake_up(&obd->obd_next_transno_waitq);
2755 spin_lock(&obd->obd_recovery_task_lock);
2756 LASSERT(obd->obd_recovering);
2757 /* usually due to recovery abort */
2758 if (!req->rq_export->exp_in_recovery) {
2759 spin_unlock(&obd->obd_recovery_task_lock);
2760 target_request_copy_put(req);
2763 LASSERT(req->rq_export->exp_lock_replay_needed);
2764 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2765 spin_unlock(&obd->obd_recovery_task_lock);
2769 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2770 * (i.e. buflens etc are in my own byte order), but type-dependent
2771 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2774 INIT_LIST_HEAD(&req->rq_list);
2775 DEBUG_REQ(D_HA, req, "not queueing");
2779 /* If we're processing the queue, we want don't want to queue this
2782 * Also, if this request has a transno less than the one we're waiting
2783 * for, we should process it now. It could (and currently always will)
2784 * be an open request for a descriptor that was opened some time ago.
2786 * Also, a resent, replayed request that has already been
2787 * handled will pass through here and be processed immediately.
2789 CDEBUG(D_HA, "Next recovery transno: %llu"
2790 ", current: %llu, replaying\n",
2791 obd->obd_next_recovery_transno, transno);
2793 /* If the request has been replayed by update replay, then sends this
2794 * request to the recovery thread (replay_request_or_update()), where
2795 * it will be handled */
2796 spin_lock(&obd->obd_recovery_task_lock);
2797 if (transno < obd->obd_next_recovery_transno &&
2798 !is_req_replayed_by_update(req)) {
2799 /* Processing the queue right now, don't re-add. */
2800 LASSERT(list_empty(&req->rq_list));
2801 spin_unlock(&obd->obd_recovery_task_lock);
2804 spin_unlock(&obd->obd_recovery_task_lock);
2806 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2809 target_request_copy_get(req);
2810 if (!req->rq_export->exp_in_recovery) {
2811 target_request_copy_put(req);
2814 LASSERT(req->rq_export->exp_req_replay_needed);
2816 if (target_exp_enqueue_req_replay(req)) {
2817 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2818 target_request_copy_put(req);
2823 spin_lock(&obd->obd_recovery_task_lock);
2824 LASSERT(obd->obd_recovering);
2825 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2826 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2827 list_add_tail(&req->rq_list, &reqiter->rq_list);
2832 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2834 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2835 "has been claimed by another client");
2836 spin_unlock(&obd->obd_recovery_task_lock);
2837 target_exp_dequeue_req_replay(req);
2838 target_request_copy_put(req);
2844 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2846 obd->obd_requests_queued_for_recovery++;
2847 spin_unlock(&obd->obd_recovery_task_lock);
2848 wake_up(&obd->obd_next_transno_waitq);
2852 int target_handle_ping(struct ptlrpc_request *req)
2854 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2855 return req_capsule_server_pack(&req->rq_pill);
2858 void target_committed_to_req(struct ptlrpc_request *req)
2860 struct obd_export *exp = req->rq_export;
2862 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2863 lustre_msg_set_last_committed(req->rq_repmsg,
2864 exp->exp_last_committed);
2866 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2867 "%d)", exp->exp_obd->obd_no_transno,
2868 req->rq_repmsg == NULL);
2870 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2871 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2874 #endif /* HAVE_SERVER_SUPPORT */
2877 * Packs current SLV and Limit into \a req.
2879 int target_pack_pool_reply(struct ptlrpc_request *req)
2881 struct obd_device *obd;
2884 /* Check that we still have all structures alive as this may
2885 * be some late RPC at shutdown time. */
2886 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2887 !exp_connect_lru_resize(req->rq_export))) {
2888 lustre_msg_set_slv(req->rq_repmsg, 0);
2889 lustre_msg_set_limit(req->rq_repmsg, 0);
2893 /* OBD is alive here as export is alive, which we checked above. */
2894 obd = req->rq_export->exp_obd;
2896 read_lock(&obd->obd_pool_lock);
2897 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2898 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2899 read_unlock(&obd->obd_pool_lock);
2904 static int target_send_reply_msg(struct ptlrpc_request *req,
2905 int rc, int fail_id)
2907 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2908 DEBUG_REQ(D_ERROR, req, "dropping reply");
2911 /* We can have a null rq_reqmsg in the event of bad signature or
2912 * no context when unwrapping */
2913 if (req->rq_reqmsg &&
2914 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2915 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2919 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2920 req->rq_status = rc;
2921 return ptlrpc_send_error(req, 1);
2923 DEBUG_REQ(D_NET, req, "sending reply");
2926 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2929 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2931 struct ptlrpc_service_part *svcpt;
2933 struct ptlrpc_reply_state *rs;
2934 struct obd_export *exp;
2937 if (req->rq_no_reply) {
2942 svcpt = req->rq_rqbd->rqbd_svcpt;
2943 rs = req->rq_reply_state;
2944 if (rs == NULL || !rs->rs_difficult) {
2946 target_send_reply_msg (req, rc, fail_id);
2951 /* must be an export if locks saved */
2952 LASSERT(req->rq_export != NULL);
2953 /* req/reply consistent */
2954 LASSERT(rs->rs_svcpt == svcpt);
2957 LASSERT(!rs->rs_scheduled);
2958 LASSERT(!rs->rs_scheduled_ever);
2959 LASSERT(!rs->rs_handled);
2960 LASSERT(!rs->rs_on_net);
2961 LASSERT(rs->rs_export == NULL);
2962 LASSERT(list_empty(&rs->rs_obd_list));
2963 LASSERT(list_empty(&rs->rs_exp_list));
2965 exp = class_export_get(req->rq_export);
2967 /* disable reply scheduling while I'm setting up */
2968 rs->rs_scheduled = 1;
2970 rs->rs_xid = req->rq_xid;
2971 rs->rs_transno = req->rq_transno;
2972 rs->rs_export = exp;
2973 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2975 spin_lock(&exp->exp_uncommitted_replies_lock);
2976 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2977 rs->rs_transno, exp->exp_last_committed);
2978 if (rs->rs_transno > exp->exp_last_committed) {
2979 /* not committed already */
2980 list_add_tail(&rs->rs_obd_list,
2981 &exp->exp_uncommitted_replies);
2983 spin_unlock(&exp->exp_uncommitted_replies_lock);
2985 spin_lock(&exp->exp_lock);
2986 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2987 spin_unlock(&exp->exp_lock);
2989 netrc = target_send_reply_msg(req, rc, fail_id);
2991 spin_lock(&svcpt->scp_rep_lock);
2993 atomic_inc(&svcpt->scp_nreps_difficult);
2996 /* error sending: reply is off the net. Also we need +1
2997 * reply ref until ptlrpc_handle_rs() is done
2998 * with the reply state (if the send was successful, there
2999 * would have been +1 ref for the net, which
3000 * reply_out_callback leaves alone) */
3002 ptlrpc_rs_addref(rs);
3005 spin_lock(&rs->rs_lock);
3006 if (rs->rs_transno <= exp->exp_last_committed ||
3007 (!rs->rs_on_net && !rs->rs_no_ack) ||
3008 list_empty(&rs->rs_exp_list) || /* completed already */
3009 list_empty(&rs->rs_obd_list)) {
3010 CDEBUG(D_HA, "Schedule reply immediately\n");
3011 ptlrpc_dispatch_difficult_reply(rs);
3013 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3014 rs->rs_scheduled = 0; /* allow notifier to schedule */
3016 spin_unlock(&rs->rs_lock);
3017 spin_unlock(&svcpt->scp_rep_lock);
3021 enum ldlm_mode lck_compat_array[] = {
3022 [LCK_EX] = LCK_COMPAT_EX,
3023 [LCK_PW] = LCK_COMPAT_PW,
3024 [LCK_PR] = LCK_COMPAT_PR,
3025 [LCK_CW] = LCK_COMPAT_CW,
3026 [LCK_CR] = LCK_COMPAT_CR,
3027 [LCK_NL] = LCK_COMPAT_NL,
3028 [LCK_GROUP] = LCK_COMPAT_GROUP,
3029 [LCK_COS] = LCK_COMPAT_COS,
3033 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3034 * not escape to the user level.
3036 int ldlm_error2errno(enum ldlm_error error)
3042 case ELDLM_LOCK_MATCHED:
3045 case ELDLM_LOCK_CHANGED:
3048 case ELDLM_LOCK_ABORTED:
3051 case ELDLM_LOCK_REPLACED:
3054 case ELDLM_NO_LOCK_DATA:
3057 case ELDLM_NAMESPACE_EXISTS:
3060 case ELDLM_BAD_NAMESPACE:
3064 if (((int)error) < 0) { /* cast to signed type */
3065 result = error; /* as ldlm_error can be unsigned */
3067 CERROR("Invalid DLM result code: %d\n", error);
3073 EXPORT_SYMBOL(ldlm_error2errno);
3076 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3078 enum ldlm_error ldlm_errno2error(int err_no)
3087 error = ELDLM_LOCK_CHANGED;
3090 error = ELDLM_LOCK_ABORTED;
3093 error = ELDLM_LOCK_REPLACED;
3096 error = ELDLM_NO_LOCK_DATA;
3099 error = ELDLM_NAMESPACE_EXISTS;
3102 error = ELDLM_BAD_NAMESPACE;
3110 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3111 void ldlm_dump_export_locks(struct obd_export *exp)
3113 spin_lock(&exp->exp_locks_list_guard);
3114 if (!list_empty(&exp->exp_locks_list)) {
3115 struct ldlm_lock *lock;
3117 CERROR("dumping locks for export %p,"
3118 "ignore if the unmount doesn't hang\n", exp);
3119 list_for_each_entry(lock, &exp->exp_locks_list,
3121 LDLM_ERROR(lock, "lock:");
3123 spin_unlock(&exp->exp_locks_list_guard);
3127 #ifdef HAVE_SERVER_SUPPORT
3128 static int target_bulk_timeout(void *data)
3131 /* We don't fail the connection here, because having the export
3132 * killed makes the (vital) call to commitrw very sad.
3137 static inline const char *bulk2type(struct ptlrpc_request *req)
3139 if (req->rq_bulk_read)
3141 if (req->rq_bulk_write)
3146 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3147 struct l_wait_info *lwi)
3149 struct ptlrpc_request *req = desc->bd_req;
3150 time_t start = cfs_time_current_sec();
3156 /* If there is eviction in progress, wait for it to finish. */
3157 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3158 *lwi = LWI_INTR(NULL, NULL);
3159 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3160 !atomic_read(&exp->exp_obd->
3161 obd_evict_inprogress),
3165 /* Check if client was evicted or reconnected already. */
3166 if (exp->exp_failed ||
3167 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3170 if (req->rq_bulk_read)
3171 rc = sptlrpc_svc_wrap_bulk(req, desc);
3173 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3174 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3175 else /* old version, bulk matchbits is rq_xid */
3176 req->rq_mbits = req->rq_xid;
3179 rc = ptlrpc_start_bulk_transfer(desc);
3183 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3184 bulk2type(req), rc);
3188 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3189 ptlrpc_abort_bulk(desc);
3193 /* limit actual bulk transfer to bulk_timeout seconds */
3194 deadline = start + bulk_timeout;
3195 if (deadline > req->rq_deadline)
3196 deadline = req->rq_deadline;
3199 long timeoutl = deadline - cfs_time_current_sec();
3200 cfs_duration_t timeout = timeoutl <= 0 ?
3201 CFS_TICK : cfs_time_seconds(timeoutl);
3204 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3205 target_bulk_timeout, desc);
3206 rc = l_wait_event(desc->bd_waitq,
3207 !ptlrpc_server_bulk_active(desc) ||
3210 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3212 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3213 /* Wait again if we changed rq_deadline. */
3214 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3215 deadline = start + bulk_timeout;
3216 if (deadline > rq_deadline)
3217 deadline = rq_deadline;
3218 } while ((rc == -ETIMEDOUT) &&
3219 (deadline > cfs_time_current_sec()));
3221 if (rc == -ETIMEDOUT) {
3222 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3223 bulk2type(req), deadline - start,
3224 cfs_time_current_sec() - deadline);
3225 ptlrpc_abort_bulk(desc);
3226 } else if (exp->exp_failed) {
3227 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3230 ptlrpc_abort_bulk(desc);
3231 } else if (exp->exp_conn_cnt >
3232 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3233 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3235 /* We don't reply anyway. */
3237 ptlrpc_abort_bulk(desc);
3238 } else if (desc->bd_failure) {
3239 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3241 /* XXX should this be a different errno? */
3244 if (req->rq_bulk_write)
3245 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3246 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3247 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3248 bulk2type(req), desc->bd_nob_transferred,
3250 /* XXX should this be a different errno? */
3257 EXPORT_SYMBOL(target_bulk_io);
3259 #endif /* HAVE_SERVER_SUPPORT */