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, 2017, 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/jiffies.h>
43 #include <linux/kthread.h>
44 #include <libcfs/libcfs.h>
46 #include <obd_class.h>
47 #include <lustre_dlm.h>
48 #include <lustre_net.h>
49 #include <lustre_sec.h>
50 #include "ldlm_internal.h"
52 /* @priority: If non-zero, move the selected connection to the list head.
53 * @create: If zero, only search in existing connections.
55 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
56 int priority, int create)
58 struct ptlrpc_connection *ptlrpc_conn;
59 struct obd_import_conn *imp_conn = NULL, *item;
60 lnet_nid_t nid4refnet = LNET_NID_ANY;
64 if (!create && !priority) {
65 CDEBUG(D_HA, "Nothing to do\n");
69 if (imp->imp_connection &&
70 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
71 /* nid4refnet is used to restrict network connections */
72 nid4refnet = imp->imp_connection->c_self;
73 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
75 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
80 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
82 GOTO(out_put, rc = -ENOMEM);
85 spin_lock(&imp->imp_lock);
86 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
87 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
89 list_del(&item->oic_item);
90 list_add(&item->oic_item,
92 item->oic_last_attempt = 0;
94 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
95 imp, imp->imp_obd->obd_name, uuid->uuid,
96 (priority ? ", moved to head" : ""));
97 spin_unlock(&imp->imp_lock);
98 GOTO(out_free, rc = 0);
101 /* No existing import connection found for \a uuid. */
103 imp_conn->oic_conn = ptlrpc_conn;
104 imp_conn->oic_uuid = *uuid;
105 imp_conn->oic_last_attempt = 0;
107 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
109 list_add_tail(&imp_conn->oic_item,
110 &imp->imp_conn_list);
111 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
112 imp, imp->imp_obd->obd_name, uuid->uuid,
113 (priority ? "head" : "tail"));
115 spin_unlock(&imp->imp_lock);
116 GOTO(out_free, rc = -ENOENT);
119 spin_unlock(&imp->imp_lock);
123 OBD_FREE(imp_conn, sizeof(*imp_conn));
125 ptlrpc_connection_put(ptlrpc_conn);
129 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
131 return import_set_conn(imp, uuid, 1, 0);
134 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
137 return import_set_conn(imp, uuid, priority, 1);
139 EXPORT_SYMBOL(client_import_add_conn);
141 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
143 struct obd_import_conn *imp_conn;
144 struct obd_export *dlmexp;
148 spin_lock(&imp->imp_lock);
149 if (list_empty(&imp->imp_conn_list)) {
150 LASSERT(!imp->imp_connection);
154 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
155 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
157 LASSERT(imp_conn->oic_conn);
159 if (imp_conn == imp->imp_conn_current) {
160 LASSERT(imp_conn->oic_conn == imp->imp_connection);
162 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
163 imp->imp_state != LUSTRE_IMP_DISCON) {
164 CERROR("can't remove current connection\n");
165 GOTO(out, rc = -EBUSY);
168 ptlrpc_connection_put(imp->imp_connection);
169 imp->imp_connection = NULL;
171 dlmexp = class_conn2export(&imp->imp_dlm_handle);
172 if (dlmexp && dlmexp->exp_connection) {
173 LASSERT(dlmexp->exp_connection ==
175 ptlrpc_connection_put(dlmexp->exp_connection);
176 dlmexp->exp_connection = NULL;
180 class_export_put(dlmexp);
183 list_del(&imp_conn->oic_item);
184 ptlrpc_connection_put(imp_conn->oic_conn);
185 OBD_FREE(imp_conn, sizeof(*imp_conn));
186 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
187 imp, imp->imp_obd->obd_name, uuid->uuid);
192 spin_unlock(&imp->imp_lock);
194 CERROR("connection %s not found\n", uuid->uuid);
197 EXPORT_SYMBOL(client_import_del_conn);
200 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
201 * to find a conn uuid of \a imp which can reach \a peer.
203 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
204 struct obd_uuid *uuid)
206 struct obd_import_conn *conn;
210 spin_lock(&imp->imp_lock);
211 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
212 /* Check if conn UUID does have this peer NID. */
213 if (class_check_uuid(&conn->oic_uuid, peer)) {
214 *uuid = conn->oic_uuid;
219 spin_unlock(&imp->imp_lock);
222 EXPORT_SYMBOL(client_import_find_conn);
224 void client_destroy_import(struct obd_import *imp)
226 /* Drop security policy instance after all RPCs have finished/aborted
227 * to let all busy contexts be released. */
228 class_import_get(imp);
229 class_destroy_import(imp);
230 sptlrpc_import_sec_put(imp);
231 class_import_put(imp);
233 EXPORT_SYMBOL(client_destroy_import);
236 * Check whether or not the OSC is on MDT.
239 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
241 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
244 static int osc_on_mdt(char *obdname)
248 ptr = strrchr(obdname, '-');
252 if (strncmp(ptr + 1, "MDT", 3) == 0)
258 /* Configure an RPC client OBD device.
263 * 3 - inactive-on-startup
264 * 4 - restrictive net
266 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
268 struct client_obd *cli = &obddev->u.cli;
269 struct obd_import *imp;
270 struct obd_uuid server_uuid;
271 int rq_portal, rp_portal, connect_op;
272 char *name = obddev->obd_type->typ_name;
273 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
274 char *cli_name = lustre_cfg_buf(lcfg, 0);
275 struct ptlrpc_connection fake_conn = { .c_self = 0,
276 .c_remote_uuid.uuid[0] = 0 };
280 /* In a more perfect world, we would hang a ptlrpc_client off of
281 * obd_type and just use the values from there. */
282 if (!strcmp(name, LUSTRE_OSC_NAME)) {
283 rq_portal = OST_REQUEST_PORTAL;
284 rp_portal = OSC_REPLY_PORTAL;
285 connect_op = OST_CONNECT;
286 cli->cl_sp_me = LUSTRE_SP_CLI;
287 cli->cl_sp_to = LUSTRE_SP_OST;
288 ns_type = LDLM_NS_TYPE_OSC;
289 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
290 !strcmp(name, LUSTRE_LWP_NAME)) {
291 rq_portal = MDS_REQUEST_PORTAL;
292 rp_portal = MDC_REPLY_PORTAL;
293 connect_op = MDS_CONNECT;
294 if (is_lwp_on_ost(cli_name))
295 cli->cl_sp_me = LUSTRE_SP_OST;
296 else if (is_lwp_on_mdt(cli_name))
297 cli->cl_sp_me = LUSTRE_SP_MDT;
299 cli->cl_sp_me = LUSTRE_SP_CLI;
300 cli->cl_sp_to = LUSTRE_SP_MDT;
301 ns_type = LDLM_NS_TYPE_MDC;
302 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
303 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
304 /* OSP_on_MDT for other MDTs */
305 connect_op = MDS_CONNECT;
306 cli->cl_sp_to = LUSTRE_SP_MDT;
307 ns_type = LDLM_NS_TYPE_MDC;
308 rq_portal = OUT_PORTAL;
310 /* OSP on MDT for OST */
311 connect_op = OST_CONNECT;
312 cli->cl_sp_to = LUSTRE_SP_OST;
313 ns_type = LDLM_NS_TYPE_OSC;
314 rq_portal = OST_REQUEST_PORTAL;
316 rp_portal = OSC_REPLY_PORTAL;
317 cli->cl_sp_me = LUSTRE_SP_MDT;
318 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
319 rq_portal = MGS_REQUEST_PORTAL;
320 rp_portal = MGC_REPLY_PORTAL;
321 connect_op = MGS_CONNECT;
322 cli->cl_sp_me = LUSTRE_SP_MGC;
323 cli->cl_sp_to = LUSTRE_SP_MGS;
324 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
325 ns_type = LDLM_NS_TYPE_MGC;
327 CERROR("unknown client OBD type \"%s\", can't setup\n",
332 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
333 CERROR("requires a TARGET UUID\n");
337 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
338 CERROR("client UUID must be less than 38 characters\n");
342 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
343 CERROR("setup requires a SERVER UUID\n");
347 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
348 CERROR("target UUID must be less than 38 characters\n");
352 init_rwsem(&cli->cl_sem);
353 mutex_init(&cli->cl_mgc_mutex);
355 init_rwsem(&cli->cl_seq_rwsem);
356 cli->cl_conn_count = 0;
357 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
358 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
359 sizeof(server_uuid)));
361 cli->cl_dirty_pages = 0;
362 cli->cl_avail_grant = 0;
363 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
364 /* cl_dirty_max_pages may be changed at connect time in
365 * ptlrpc_connect_interpret(). */
366 client_adjust_max_dirty(cli);
367 INIT_LIST_HEAD(&cli->cl_cache_waiters);
368 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
369 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
370 INIT_LIST_HEAD(&cli->cl_loi_write_list);
371 INIT_LIST_HEAD(&cli->cl_loi_read_list);
372 spin_lock_init(&cli->cl_loi_list_lock);
373 atomic_set(&cli->cl_pending_w_pages, 0);
374 atomic_set(&cli->cl_pending_r_pages, 0);
375 cli->cl_r_in_flight = 0;
376 cli->cl_w_in_flight = 0;
378 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
379 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
380 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
381 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
382 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
383 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
386 INIT_LIST_HEAD(&cli->cl_lru_osc);
387 atomic_set(&cli->cl_lru_shrinkers, 0);
388 atomic_long_set(&cli->cl_lru_busy, 0);
389 atomic_long_set(&cli->cl_lru_in_list, 0);
390 INIT_LIST_HEAD(&cli->cl_lru_list);
391 spin_lock_init(&cli->cl_lru_list_lock);
392 atomic_long_set(&cli->cl_unstable_count, 0);
393 INIT_LIST_HEAD(&cli->cl_shrink_list);
395 INIT_LIST_HEAD(&cli->cl_flight_waiters);
396 cli->cl_rpcs_in_flight = 0;
398 init_waitqueue_head(&cli->cl_destroy_waitq);
399 atomic_set(&cli->cl_destroy_in_flight, 0);
401 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
402 #ifdef ENABLE_CHECKSUM
403 /* Turn on checksumming by default. */
404 cli->cl_checksum = 1;
406 * The supported checksum types will be worked out at connect time
407 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
410 cli->cl_cksum_type = cli->cl_supp_cksum_types;
412 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
414 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
415 * from OFD after connecting. */
416 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
418 cli->cl_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
420 /* set cl_chunkbits default value to PAGE_SHIFT,
421 * it will be updated at OSC connection time. */
422 cli->cl_chunkbits = PAGE_SHIFT;
424 if (!strcmp(name, LUSTRE_MDC_NAME)) {
425 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
426 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
427 cli->cl_max_rpcs_in_flight = 2;
428 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
429 cli->cl_max_rpcs_in_flight = 3;
430 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
431 cli->cl_max_rpcs_in_flight = 4;
433 if (osc_on_mdt(obddev->obd_name))
434 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
436 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
439 spin_lock_init(&cli->cl_mod_rpcs_lock);
440 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
441 cli->cl_max_mod_rpcs_in_flight = 0;
442 cli->cl_mod_rpcs_in_flight = 0;
443 cli->cl_close_rpcs_in_flight = 0;
444 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
445 cli->cl_mod_tag_bitmap = NULL;
447 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
449 if (connect_op == MDS_CONNECT) {
450 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
451 OBD_ALLOC(cli->cl_mod_tag_bitmap,
452 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
453 if (cli->cl_mod_tag_bitmap == NULL)
454 GOTO(err, rc = -ENOMEM);
459 CERROR("ldlm_get_ref failed: %d\n", rc);
463 ptlrpc_init_client(rq_portal, rp_portal, name,
464 &obddev->obd_ldlm_client);
466 imp = class_new_import(obddev);
468 GOTO(err_ldlm, rc = -ENOENT);
469 imp->imp_client = &obddev->obd_ldlm_client;
470 imp->imp_connect_op = connect_op;
471 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
472 LUSTRE_CFG_BUFLEN(lcfg, 1));
473 class_import_put(imp);
475 if (lustre_cfg_buf(lcfg, 4)) {
476 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
478 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
480 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
481 obddev->obd_name, lustre_cfg_string(lcfg, 4),
483 GOTO(err_import, rc);
485 fake_conn.c_self = LNET_MKNID(refnet, 0);
486 imp->imp_connection = &fake_conn;
489 rc = client_import_add_conn(imp, &server_uuid, 1);
491 CERROR("can't add initial connection\n");
492 GOTO(err_import, rc);
494 imp->imp_connection = NULL;
496 cli->cl_import = imp;
497 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
498 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
500 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
501 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
502 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
503 name, obddev->obd_name,
504 cli->cl_target_uuid.uuid);
505 spin_lock(&imp->imp_lock);
506 imp->imp_deactive = 1;
507 spin_unlock(&imp->imp_lock);
511 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
512 LDLM_NAMESPACE_CLIENT,
513 LDLM_NAMESPACE_GREEDY,
515 if (obddev->obd_namespace == NULL) {
516 CERROR("Unable to create client namespace - %s\n",
518 GOTO(err_import, rc = -ENOMEM);
524 class_destroy_import(imp);
528 if (cli->cl_mod_tag_bitmap != NULL)
529 OBD_FREE(cli->cl_mod_tag_bitmap,
530 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
531 cli->cl_mod_tag_bitmap = NULL;
535 EXPORT_SYMBOL(client_obd_setup);
537 int client_obd_cleanup(struct obd_device *obddev)
539 struct client_obd *cli = &obddev->u.cli;
542 ldlm_namespace_free_post(obddev->obd_namespace);
543 obddev->obd_namespace = NULL;
545 obd_cleanup_client_import(obddev);
546 LASSERT(obddev->u.cli.cl_import == NULL);
550 if (cli->cl_mod_tag_bitmap != NULL)
551 OBD_FREE(cli->cl_mod_tag_bitmap,
552 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
553 cli->cl_mod_tag_bitmap = NULL;
557 EXPORT_SYMBOL(client_obd_cleanup);
559 /* ->o_connect() method for client side (OSC and MDC and MGC) */
560 int client_connect_import(const struct lu_env *env,
561 struct obd_export **exp,
562 struct obd_device *obd, struct obd_uuid *cluuid,
563 struct obd_connect_data *data, void *localdata)
565 struct client_obd *cli = &obd->u.cli;
566 struct obd_import *imp = cli->cl_import;
567 struct obd_connect_data *ocd;
568 struct lustre_handle conn = { 0 };
573 down_write(&cli->cl_sem);
574 if (cli->cl_conn_count > 0)
575 GOTO(out_sem, rc = -EALREADY);
577 rc = class_connect(&conn, obd, cluuid);
581 cli->cl_conn_count++;
582 *exp = class_conn2export(&conn);
584 LASSERT(obd->obd_namespace);
586 imp->imp_dlm_handle = conn;
587 rc = ptlrpc_init_import(imp);
591 ocd = &imp->imp_connect_data;
594 imp->imp_connect_flags_orig = data->ocd_connect_flags;
595 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
598 rc = ptlrpc_connect_import(imp);
600 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
603 LASSERT(*exp != NULL && (*exp)->exp_connection);
606 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
607 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
608 data->ocd_connect_flags, ocd->ocd_connect_flags);
609 data->ocd_connect_flags = ocd->ocd_connect_flags;
610 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
613 ptlrpc_pinger_add_import(imp);
619 cli->cl_conn_count--;
620 class_disconnect(*exp);
624 up_write(&cli->cl_sem);
628 EXPORT_SYMBOL(client_connect_import);
630 int client_disconnect_export(struct obd_export *exp)
632 struct obd_device *obd = class_exp2obd(exp);
633 struct client_obd *cli;
634 struct obd_import *imp;
639 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
640 exp, exp ? exp->exp_handle.h_cookie : -1);
645 imp = cli->cl_import;
647 down_write(&cli->cl_sem);
648 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
651 if (cli->cl_conn_count == 0) {
652 CERROR("disconnecting disconnected device (%s)\n",
654 GOTO(out_disconnect, rc = -EINVAL);
657 cli->cl_conn_count--;
658 if (cli->cl_conn_count != 0)
659 GOTO(out_disconnect, rc = 0);
661 /* Mark import deactivated now, so we don't try to reconnect if any
662 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
663 * fully deactivate the import, or that would drop all requests. */
664 spin_lock(&imp->imp_lock);
665 imp->imp_deactive = 1;
666 spin_unlock(&imp->imp_lock);
668 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
669 * delete it regardless. (It's safe to delete an import that was
671 (void)ptlrpc_pinger_del_import(imp);
673 if (obd->obd_namespace != NULL) {
674 /* obd_force == local only */
675 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
676 obd->obd_force ? LCF_LOCAL : 0, NULL);
677 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
680 /* There's no need to hold sem while disconnecting an import,
681 * and it may actually cause deadlock in GSS. */
682 up_write(&cli->cl_sem);
683 rc = ptlrpc_disconnect_import(imp, 0);
684 down_write(&cli->cl_sem);
686 ptlrpc_invalidate_import(imp);
691 /* Use server style - class_disconnect should be always called for
693 err = class_disconnect(exp);
697 up_write(&cli->cl_sem);
701 EXPORT_SYMBOL(client_disconnect_export);
703 #ifdef HAVE_SERVER_SUPPORT
704 int server_disconnect_export(struct obd_export *exp)
709 /* Disconnect early so that clients can't keep using export. */
710 rc = class_disconnect(exp);
711 /* Close import to avoid sending any requests. */
712 if (exp->exp_imp_reverse)
713 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
715 ldlm_bl_thread_wakeup();
717 /* complete all outstanding replies */
718 spin_lock(&exp->exp_lock);
719 while (!list_empty(&exp->exp_outstanding_replies)) {
720 struct ptlrpc_reply_state *rs =
721 list_entry(exp->exp_outstanding_replies.next,
722 struct ptlrpc_reply_state, rs_exp_list);
723 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
725 spin_lock(&svcpt->scp_rep_lock);
727 list_del_init(&rs->rs_exp_list);
729 spin_lock(&rs->rs_lock);
730 /* clear rs_convert_lock to make sure rs is handled and put */
731 rs->rs_convert_lock = 0;
732 ptlrpc_schedule_difficult_reply(rs);
733 spin_unlock(&rs->rs_lock);
735 spin_unlock(&svcpt->scp_rep_lock);
737 spin_unlock(&exp->exp_lock);
741 EXPORT_SYMBOL(server_disconnect_export);
743 /* --------------------------------------------------------------------------
744 * from old lib/target.c
745 * -------------------------------------------------------------------------- */
747 static int target_handle_reconnect(struct lustre_handle *conn,
748 struct obd_export *exp,
749 struct obd_uuid *cluuid)
751 struct obd_device *target;
752 struct lustre_handle *hdl;
759 hdl = &exp->exp_imp_reverse->imp_remote_handle;
760 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
761 conn->cookie = exp->exp_handle.h_cookie;
762 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
763 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
767 target = exp->exp_obd;
769 /* Might be a re-connect after a partition. */
770 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
771 LCONSOLE_WARN("%s: already connected client %s (at %s) "
772 "with handle %#llx. Rejecting client "
773 "with the same UUID trying to reconnect "
774 "with handle %#llx\n", target->obd_name,
775 obd_uuid2str(&exp->exp_client_uuid),
776 obd_export_nid2str(exp),
777 hdl->cookie, conn->cookie);
778 memset(conn, 0, sizeof *conn);
779 /* target_handle_connect() treats EALREADY and
780 * -EALREADY differently. -EALREADY is an error
781 * (same UUID, different handle). */
785 if (!target->obd_recovering) {
786 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
787 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
788 obd_export_nid2str(exp));
789 GOTO(out_already, rc);
792 now = ktime_get_seconds();
793 deadline = jiffies_to_msecs(target->obd_recovery_timer.expires) /
795 if (now < deadline) {
796 struct target_distribute_txn_data *tdtd;
801 timeout = deadline - now;
802 tdtd = class_exp2tgt(exp)->lut_tdtd;
803 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
804 buf = tdtd->tdtd_show_update_logs_retrievers(
805 tdtd->tdtd_show_retrievers_cbdata,
809 LCONSOLE_WARN("%s: Recovery already passed deadline "
810 "%lld:%.02lld. It is due to DNE recovery "
811 "failed/stuck on the %d MDT(s):%s. "
812 "Please wait until all MDTs recovered "
813 "or abort the recovery by force.\n",
814 target->obd_name, timeout / 60,
816 buf ? buf : "unknown (not enough RAM)");
818 LCONSOLE_WARN("%s: Recovery already passed deadline "
819 "%lld:%.02lld. If you do not want to wait "
820 "more, please abort the recovery by "
821 "force.\n", target->obd_name,
822 timeout / 60, timeout % 60);
827 timeout = now - deadline;
828 LCONSOLE_WARN("%s: Recovery already passed deadline"
829 " %lld:%.02lld, It is most likely due to DNE"
830 " recovery is failed or stuck, please wait a"
831 " few more minutes or abort the recovery.\n",
832 target->obd_name, timeout / 60, timeout % 60);
836 conn->cookie = exp->exp_handle.h_cookie;
837 /* target_handle_connect() treats EALREADY and
838 * -EALREADY differently. EALREADY means we are
839 * doing a valid reconnect from the same client. */
844 check_and_start_recovery_timer(struct obd_device *obd,
845 struct ptlrpc_request *req, int new_client);
848 * update flags for import during reconnect process
850 static int rev_import_flags_update(struct obd_import *revimp,
851 struct ptlrpc_request *req)
854 struct obd_connect_data *data;
856 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
858 if (data->ocd_connect_flags & OBD_CONNECT_AT)
859 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
861 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
863 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
865 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
867 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
868 revimp->imp_client->cli_name,
869 libcfs_id2str(req->rq_peer), rc);
877 * Allocate a new reverse import for an export.
879 * \retval -errno in case error hit
880 * \retval 0 if reverse import correctly init
882 int rev_import_init(struct obd_export *export)
884 struct obd_device *obd = export->exp_obd;
885 struct obd_import *revimp;
887 LASSERT(export->exp_imp_reverse == NULL);
889 revimp = class_new_import(obd);
893 revimp->imp_remote_handle.cookie = 0ULL;
894 revimp->imp_client = &obd->obd_ldlm_client;
895 revimp->imp_dlm_fake = 1;
897 /* it is safe to connect import in new state as no sends possible */
898 spin_lock(&export->exp_lock);
899 export->exp_imp_reverse = revimp;
900 spin_unlock(&export->exp_lock);
901 class_import_put(revimp);
905 EXPORT_SYMBOL(rev_import_init);
908 * Handle reconnect for an export.
910 * \param exp export to handle reconnect process
911 * \param req client reconnect request
913 * \retval -rc in case securitfy flavor can't be changed
914 * \retval 0 in case none problems
916 static int rev_import_reconnect(struct obd_export *exp,
917 struct ptlrpc_request *req)
919 struct obd_import *revimp = exp->exp_imp_reverse;
920 struct lustre_handle *lh;
923 /* avoid sending a request until import flags are changed */
924 ptlrpc_import_enter_resend(revimp);
926 if (revimp->imp_connection != NULL)
927 ptlrpc_connection_put(revimp->imp_connection);
930 * client from recovery don't have a handle so we need to take from
931 * request. it may produce situation when wrong client connected
932 * to recovery as we trust a client uuid
934 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
935 revimp->imp_remote_handle = *lh;
937 /* unknown versions will be caught in
938 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
939 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
941 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
943 rc = rev_import_flags_update(revimp, req);
945 /* it is safe to still be in RECOVERY phase as we are not able
946 * to setup correct security flavor so requests are not able to
947 * be delivered correctly */
951 /* resend all rpc's via new connection */
952 return ptlrpc_import_recovery_state_machine(revimp);
955 int target_handle_connect(struct ptlrpc_request *req)
957 struct obd_device *target = NULL;
958 struct obd_export *export = NULL;
959 /* connect handle - filled from target_handle_reconnect in
961 struct lustre_handle conn;
962 struct lustre_handle *tmp;
963 struct obd_uuid cluuid;
968 bool mds_conn = false, lw_client = false, initial_conn = false;
969 bool mds_mds_conn = false;
970 bool new_mds_mds_conn = false;
971 struct obd_connect_data *data, *tmpdata;
973 lnet_nid_t *client_nid = NULL;
976 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
978 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
980 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
981 GOTO(out, rc = -EINVAL);
984 target = class_dev_by_str(str);
986 deuuidify(str, NULL, &target_start, &target_len);
987 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
988 "from %s (no target). If you are running "
989 "an HA pair check that the target is "
990 "mounted on the other server.\n", str,
991 libcfs_nid2str(req->rq_peer.nid));
992 GOTO(out, rc = -ENODEV);
995 spin_lock(&target->obd_dev_lock);
997 target->obd_conn_inprogress++;
999 if (target->obd_stopping || !target->obd_set_up) {
1000 spin_unlock(&target->obd_dev_lock);
1002 deuuidify(str, NULL, &target_start, &target_len);
1003 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1004 target_len, target_start,
1005 libcfs_nid2str(req->rq_peer.nid),
1006 (target->obd_stopping ?
1007 "stopping" : "not set up"));
1008 GOTO(out, rc = -ENODEV);
1011 if (target->obd_no_conn) {
1012 spin_unlock(&target->obd_dev_lock);
1014 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1015 "from %s\n", target->obd_name,
1016 libcfs_nid2str(req->rq_peer.nid));
1017 GOTO(out, rc = -EAGAIN);
1020 spin_unlock(&target->obd_dev_lock);
1022 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1024 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1025 GOTO(out, rc = -EINVAL);
1028 obd_str2uuid(&cluuid, str);
1030 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1032 GOTO(out, rc = -EPROTO);
1036 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1038 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1040 GOTO(out, rc = -EPROTO);
1042 rc = req_capsule_server_pack(&req->rq_pill);
1046 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1047 /* Don't allow clients to connect that are using old 1.8 format
1048 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1049 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1050 * FULL20 flag should be set on all connections since 2.0, but no
1051 * longer affects behaviour.
1053 * Later this check will be disabled and the flag can be retired
1054 * completely once interop with 3.0 is no longer needed.
1056 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1057 GOTO(out, rc = -EPROTO);
1059 /* Don't allow liblustre clients to connect.
1060 * - testing was disabled in v2_2_50_0-61-g6a75d65
1061 * - building was disabled in v2_5_58_0-28-g7277179
1062 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1063 * - clients were refused connect for version difference > 0.0.1.32 */
1064 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1065 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1066 GOTO(out, rc = -EPROTO);
1070 /* Note: lw_client is needed in MDS-MDS failover during update log
1071 * processing, so we needs to allow lw_client to be connected at
1072 * anytime, instead of only the initial connection */
1073 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1075 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1076 initial_conn = true;
1077 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1078 mds_mds_conn = (data->ocd_connect_flags &
1079 OBD_CONNECT_MDS_MDS) != 0;
1081 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1082 * for Imperative Recovery connection from MGC to MGS.
1084 * Via check OBD_CONNECT_FID, we can distinguish whether
1085 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1086 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1089 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1090 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1091 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1092 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1093 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1094 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1096 /* We do not support the MDT-MDT interoperations with
1097 * different version MDT because of protocol changes. */
1098 if (unlikely(major != LUSTRE_MAJOR ||
1099 minor != LUSTRE_MINOR ||
1100 abs(patch - LUSTRE_PATCH) > 3)) {
1101 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1102 "connection from different version MDT "
1103 "(%d.%d.%d.%d) %s %s\n",
1104 target->obd_name, LUSTRE_MAJOR,
1105 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1106 major, minor, patch,
1107 OBD_OCD_VERSION_FIX(data->ocd_version),
1108 libcfs_nid2str(req->rq_peer.nid), str);
1110 GOTO(out, rc = -EPROTO);
1115 /* lctl gets a backstage, all-access pass. */
1116 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1117 goto dont_check_exports;
1119 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1123 /* We've found an export in the hash. */
1125 spin_lock(&export->exp_lock);
1127 if (export->exp_connecting) { /* bug 9635, et. al. */
1128 spin_unlock(&export->exp_lock);
1129 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1130 export->exp_obd->obd_name, export,
1131 libcfs_nid2str(req->rq_peer.nid));
1132 class_export_put(export);
1135 } else if ((mds_conn || (lw_client && initial_conn) ||
1136 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1137 export->exp_connection != NULL) {
1138 spin_unlock(&export->exp_lock);
1139 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1140 /* MDS or LWP reconnected after failover. */
1141 LCONSOLE_WARN("%s: Received %s connection from "
1142 "%s, removing former export from %s\n",
1143 target->obd_name, mds_conn ? "MDS" : "LWP",
1144 libcfs_nid2str(req->rq_peer.nid),
1145 libcfs_nid2str(export->exp_connection->c_peer.nid));
1147 /* New MDS connection from the same NID. */
1148 LCONSOLE_WARN("%s: Received new %s connection from "
1149 "%s, removing former export from same NID\n",
1150 target->obd_name, mds_conn ? "MDS" : "LWP",
1151 libcfs_nid2str(req->rq_peer.nid));
1154 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1155 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1156 /* Because exports between MDTs will always be
1157 * kept, let's do not fail such export if they
1158 * come from the same NID, otherwise it might
1159 * cause eviction between MDTs, which might
1160 * cause namespace inconsistency */
1161 spin_lock(&export->exp_lock);
1162 export->exp_connecting = 1;
1163 spin_unlock(&export->exp_lock);
1164 conn.cookie = export->exp_handle.h_cookie;
1167 class_fail_export(export);
1168 class_export_put(export);
1172 } else if (export->exp_connection != NULL && initial_conn &&
1173 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1174 spin_unlock(&export->exp_lock);
1175 /* In MDS failover we have static UUID but NID can change. */
1176 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1177 "existing nid %s is already connected\n",
1178 target->obd_name, cluuid.uuid,
1179 libcfs_nid2str(req->rq_peer.nid),
1181 export->exp_connection->c_peer.nid));
1183 class_export_put(export);
1186 export->exp_connecting = 1;
1187 spin_unlock(&export->exp_lock);
1188 LASSERT(export->exp_obd == target);
1190 rc = target_handle_reconnect(&conn, export, &cluuid);
1193 /* If we found an export, we already unlocked. */
1196 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1197 } else if (req->rq_export == NULL &&
1198 atomic_read(&export->exp_rpc_count) > 0) {
1199 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1200 "still busy with %d references\n",
1201 target->obd_name, cluuid.uuid,
1202 libcfs_nid2str(req->rq_peer.nid),
1203 atomic_read(&export->exp_refcount));
1204 GOTO(out, rc = -EBUSY);
1205 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1206 if (!strstr(cluuid.uuid, "mdt"))
1207 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1208 "known client %s (at %s) because it "
1209 "is indicating it is a new client",
1210 target->obd_name, cluuid.uuid,
1211 libcfs_nid2str(req->rq_peer.nid));
1212 GOTO(out, rc = -EALREADY);
1214 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1221 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1222 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1223 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1224 export, ktime_get_real_seconds(),
1225 export ? export->exp_last_request_time : 0);
1227 /* If this is the first time a client connects, reset the recovery
1228 * timer. Discard lightweight connections which might be local. */
1229 if (!lw_client && rc == 0 && target->obd_recovering)
1230 check_and_start_recovery_timer(target, req, export == NULL);
1232 /* We want to handle EALREADY but *not* -EALREADY from
1233 * target_handle_reconnect(), return reconnection state in a flag. */
1234 if (rc == EALREADY) {
1235 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1241 /* Tell the client if we support replayable requests. */
1242 if (target->obd_replayable)
1243 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1244 client_nid = &req->rq_peer.nid;
1246 if (export == NULL) {
1247 /* allow lightweight connections during recovery */
1248 /* allow "new" MDT to be connected during recovery, since we
1249 * need retrieve recovery update records from it */
1250 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1252 int c; /* connected */
1253 int i; /* in progress */
1255 int s; /* stale/evicted */
1257 c = atomic_read(&target->obd_connected_clients);
1258 i = atomic_read(&target->obd_lock_replay_clients);
1259 k = target->obd_max_recoverable_clients;
1260 s = target->obd_stale_clients;
1261 t = jiffies_to_msecs(target->obd_recovery_timer.expires);
1263 t -= ktime_get_seconds();
1264 LCONSOLE_WARN("%s: Denying connection for new client %s"
1265 "(at %s), waiting for %d known clients "
1266 "(%d recovered, %d in progress, and %d "
1267 "evicted) to recover in %lld:%.02lld\n",
1268 target->obd_name, cluuid.uuid,
1269 libcfs_nid2str(req->rq_peer.nid), k,
1270 c - i, i, s, t / 60, t % 60);
1274 rc = obd_connect(req->rq_svc_thread->t_env,
1275 &export, target, &cluuid, data,
1277 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1278 lustre_msg_add_op_flags(req->rq_repmsg,
1279 MSG_CONNECT_RECOVERING);
1281 conn.cookie = export->exp_handle.h_cookie;
1282 rc = rev_import_init(export);
1286 new_mds_mds_conn = true;
1289 rc = obd_reconnect(req->rq_svc_thread->t_env,
1290 export, target, &cluuid, data, client_nid);
1295 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1296 data->ocd_instance = target->u.obt.obt_instance;
1298 /* Return only the parts of obd_connect_data that we understand, so the
1299 * client knows that we don't understand the rest. */
1301 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1303 tmpdata = req_capsule_server_get(&req->rq_pill,
1305 /* Don't use struct assignment here, because the client reply
1306 * buffer may be smaller/larger than the local struct
1307 * obd_connect_data. */
1308 memcpy(tmpdata, data, min(tmpsize, size));
1311 /* If the client and the server are the same node, we will already
1312 * have an export that really points to the client's DLM export,
1313 * because we have a shared handles table.
1315 * XXX this will go away when shaver stops sending the "connect" handle
1316 * in the real "remote handle" field of the request --phik 24 Apr 2003
1318 ptlrpc_request_change_export(req, export);
1320 spin_lock(&export->exp_lock);
1321 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1322 spin_unlock(&export->exp_lock);
1323 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1324 "or equal conn_cnt: %d >= %d\n",
1325 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1326 export->exp_conn_cnt,
1327 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1329 GOTO(out, rc = -EALREADY);
1331 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1332 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1333 spin_unlock(&export->exp_lock);
1335 if (export->exp_connection != NULL) {
1336 /* Check to see if connection came from another NID. */
1337 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1338 !hlist_unhashed(&export->exp_nid_hash))
1339 cfs_hash_del(export->exp_obd->obd_nid_hash,
1340 &export->exp_connection->c_peer.nid,
1341 &export->exp_nid_hash);
1343 ptlrpc_connection_put(export->exp_connection);
1346 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1349 if (hlist_unhashed(&export->exp_nid_hash))
1350 cfs_hash_add(export->exp_obd->obd_nid_hash,
1351 &export->exp_connection->c_peer.nid,
1352 &export->exp_nid_hash);
1354 lustre_msg_set_handle(req->rq_repmsg, &conn);
1356 rc = rev_import_reconnect(export, req);
1360 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1362 __u64 transno = data->ocd_transno;
1364 spin_lock(&export->exp_lock);
1365 /* possible race with class_disconnect_stale_exports,
1366 * export may be already in the eviction process */
1367 if (export->exp_failed) {
1368 spin_unlock(&export->exp_lock);
1369 GOTO(out, rc = -ENODEV);
1371 export->exp_in_recovery = 1;
1372 export->exp_req_replay_needed = 1;
1373 export->exp_lock_replay_needed = 1;
1374 spin_unlock(&export->exp_lock);
1376 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1377 MSG_CONNECT_TRANSNO);
1378 if (has_transno && transno == 0)
1379 CWARN("Connect with zero transno!\n");
1381 if (has_transno && transno > 0 &&
1382 transno < target->obd_next_recovery_transno &&
1383 transno > target->obd_last_committed) {
1384 /* Another way is to use cmpxchg() to be lock-free. */
1385 spin_lock(&target->obd_recovery_task_lock);
1386 if (transno < target->obd_next_recovery_transno)
1387 target->obd_next_recovery_transno = transno;
1388 spin_unlock(&target->obd_recovery_task_lock);
1391 atomic_inc(&target->obd_req_replay_clients);
1392 atomic_inc(&target->obd_lock_replay_clients);
1393 /* Note: MDS-MDS connection is allowed to be connected during
1394 * recovery, no matter if the exports needs to be recoveried.
1395 * Because we need retrieve updates logs from all other MDTs.
1396 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1397 * also needs to be increased to match other recovery checking
1399 if (new_mds_mds_conn)
1400 target->obd_max_recoverable_clients++;
1401 if (atomic_inc_return(&target->obd_connected_clients) ==
1402 target->obd_max_recoverable_clients)
1403 wake_up(&target->obd_next_transno_waitq);
1406 /* Tell the client we're in recovery, when client is involved in it. */
1407 if (target->obd_recovering && !lw_client)
1408 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1412 spin_lock(&export->exp_lock);
1413 export->exp_connecting = 0;
1414 spin_unlock(&export->exp_lock);
1416 class_export_put(export);
1418 if (target != NULL) {
1419 spin_lock(&target->obd_dev_lock);
1420 target->obd_conn_inprogress--;
1421 spin_unlock(&target->obd_dev_lock);
1422 class_decref(target, "find", current);
1424 req->rq_status = rc;
1428 int target_handle_disconnect(struct ptlrpc_request *req)
1433 rc = req_capsule_server_pack(&req->rq_pill);
1437 /* Keep the rq_export around so we can send the reply. */
1438 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1443 void target_destroy_export(struct obd_export *exp)
1445 struct obd_import *imp = NULL;
1446 /* exports created from last_rcvd data, and "fake"
1447 exports created by lctl don't have an import */
1448 spin_lock(&exp->exp_lock);
1449 if (exp->exp_imp_reverse != NULL) {
1450 imp = exp->exp_imp_reverse;
1451 exp->exp_imp_reverse = NULL;
1453 spin_unlock(&exp->exp_lock);
1455 client_destroy_import(imp);
1457 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1458 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1459 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1460 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1462 EXPORT_SYMBOL(target_destroy_export);
1465 * Recovery functions
1467 static void target_request_copy_get(struct ptlrpc_request *req)
1469 class_export_rpc_inc(req->rq_export);
1470 LASSERT(list_empty(&req->rq_list));
1471 INIT_LIST_HEAD(&req->rq_replay_list);
1473 /* Increase refcount to keep request in queue. */
1474 atomic_inc(&req->rq_refcount);
1475 /* Let export know it has replays to be handled. */
1476 atomic_inc(&req->rq_export->exp_replay_count);
1479 static void target_request_copy_put(struct ptlrpc_request *req)
1481 LASSERT(list_empty(&req->rq_replay_list));
1482 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1484 atomic_dec(&req->rq_export->exp_replay_count);
1485 class_export_rpc_dec(req->rq_export);
1486 ptlrpc_server_drop_request(req);
1489 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1491 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1492 struct obd_export *exp = req->rq_export;
1493 struct ptlrpc_request *reqiter;
1494 struct ptlrpc_request *dup_req = NULL;
1499 spin_lock(&exp->exp_lock);
1500 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1502 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1510 /* We expect it with RESENT and REPLAY flags. */
1511 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1512 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1513 CERROR("invalid flags %x of resent replay\n",
1514 lustre_msg_get_flags(req->rq_reqmsg));
1516 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1519 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1521 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1522 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1526 list_add_tail(&req->rq_replay_list,
1527 &exp->exp_req_replay_queue);
1530 spin_unlock(&exp->exp_lock);
1534 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1536 LASSERT(!list_empty(&req->rq_replay_list));
1537 LASSERT(req->rq_export);
1539 spin_lock(&req->rq_export->exp_lock);
1540 list_del_init(&req->rq_replay_list);
1541 spin_unlock(&req->rq_export->exp_lock);
1544 static void target_finish_recovery(struct lu_target *lut)
1546 struct obd_device *obd = lut->lut_obd;
1549 /* Only log a recovery message when recovery has occurred. */
1550 if (obd->obd_recovery_start) {
1551 time64_t now = ktime_get_real_seconds();
1552 time64_t elapsed_time;
1554 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1555 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1556 "%d recovered and %d %s evicted.\n", obd->obd_name,
1557 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1558 obd->obd_max_recoverable_clients,
1559 atomic_read(&obd->obd_connected_clients),
1560 obd->obd_stale_clients,
1561 obd->obd_stale_clients == 1 ? "was" : "were");
1564 ldlm_reprocess_recovery_done(obd->obd_namespace);
1565 spin_lock(&obd->obd_recovery_task_lock);
1566 if (!list_empty(&obd->obd_req_replay_queue) ||
1567 !list_empty(&obd->obd_lock_replay_queue) ||
1568 !list_empty(&obd->obd_final_req_queue)) {
1569 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1571 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1572 list_empty(&obd->obd_lock_replay_queue) ? \
1574 list_empty(&obd->obd_final_req_queue) ? \
1576 spin_unlock(&obd->obd_recovery_task_lock);
1579 spin_unlock(&obd->obd_recovery_task_lock);
1581 obd->obd_recovery_end = ktime_get_real_seconds();
1583 /* When recovery finished, cleanup orphans on MDS and OST. */
1584 if (obd->obd_type && OBP(obd, postrecov)) {
1585 int rc = OBP(obd, postrecov)(obd);
1588 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1594 static void abort_req_replay_queue(struct obd_device *obd)
1596 struct ptlrpc_request *req, *n;
1597 struct list_head abort_list;
1599 INIT_LIST_HEAD(&abort_list);
1600 spin_lock(&obd->obd_recovery_task_lock);
1601 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1602 spin_unlock(&obd->obd_recovery_task_lock);
1603 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1604 DEBUG_REQ(D_WARNING, req, "aborted:");
1605 req->rq_status = -ENOTCONN;
1606 if (ptlrpc_error(req)) {
1607 DEBUG_REQ(D_ERROR, req,
1608 "failed abort_req_reply; skipping");
1610 target_exp_dequeue_req_replay(req);
1611 target_request_copy_put(req);
1615 static void abort_lock_replay_queue(struct obd_device *obd)
1617 struct ptlrpc_request *req, *n;
1618 struct list_head abort_list;
1620 INIT_LIST_HEAD(&abort_list);
1621 spin_lock(&obd->obd_recovery_task_lock);
1622 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1623 spin_unlock(&obd->obd_recovery_task_lock);
1624 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1625 DEBUG_REQ(D_ERROR, req, "aborted:");
1626 req->rq_status = -ENOTCONN;
1627 if (ptlrpc_error(req)) {
1628 DEBUG_REQ(D_ERROR, req,
1629 "failed abort_lock_reply; skipping");
1631 target_request_copy_put(req);
1635 /* Called from a cleanup function if the device is being cleaned up
1636 forcefully. The exports should all have been disconnected already,
1637 the only thing left to do is
1638 - clear the recovery flags
1640 - free queued requests and replies, but don't send replies
1641 Because the obd_stopping flag is set, no new requests should be received.
1644 void target_cleanup_recovery(struct obd_device *obd)
1646 struct ptlrpc_request *req, *n;
1647 struct list_head clean_list;
1649 INIT_LIST_HEAD(&clean_list);
1650 spin_lock(&obd->obd_dev_lock);
1651 if (!obd->obd_recovering) {
1652 spin_unlock(&obd->obd_dev_lock);
1656 obd->obd_recovering = obd->obd_abort_recovery = 0;
1657 spin_unlock(&obd->obd_dev_lock);
1659 spin_lock(&obd->obd_recovery_task_lock);
1660 target_cancel_recovery_timer(obd);
1661 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1662 spin_unlock(&obd->obd_recovery_task_lock);
1664 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1665 LASSERT(req->rq_reply_state == NULL);
1666 target_exp_dequeue_req_replay(req);
1667 target_request_copy_put(req);
1670 spin_lock(&obd->obd_recovery_task_lock);
1671 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1672 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1673 spin_unlock(&obd->obd_recovery_task_lock);
1675 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1676 LASSERT(req->rq_reply_state == NULL);
1677 target_request_copy_put(req);
1682 EXPORT_SYMBOL(target_cleanup_recovery);
1684 /* obd_recovery_task_lock should be held */
1685 void target_cancel_recovery_timer(struct obd_device *obd)
1687 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1688 del_timer(&obd->obd_recovery_timer);
1691 static void target_start_recovery_timer(struct obd_device *obd)
1693 if (obd->obd_recovery_start != 0)
1696 spin_lock(&obd->obd_dev_lock);
1697 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1698 spin_unlock(&obd->obd_dev_lock);
1702 LASSERT(obd->obd_recovery_timeout != 0);
1704 if (obd->obd_recovery_start != 0) {
1705 spin_unlock(&obd->obd_dev_lock);
1709 mod_timer(&obd->obd_recovery_timer,
1710 jiffies + cfs_time_seconds(obd->obd_recovery_timeout));
1711 obd->obd_recovery_start = ktime_get_real_seconds();
1712 spin_unlock(&obd->obd_dev_lock);
1714 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1716 obd->obd_recovery_timeout / 60,
1717 obd->obd_recovery_timeout % 60,
1718 obd->obd_max_recoverable_clients,
1719 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1720 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1724 * extend recovery window.
1726 * if @extend is true, extend recovery window to have @drt remaining at least;
1727 * otherwise, make sure the recovery timeout value is not less than @drt.
1729 static void extend_recovery_timer(struct obd_device *obd, time64_t drt,
1737 spin_lock(&obd->obd_dev_lock);
1738 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1739 spin_unlock(&obd->obd_dev_lock);
1742 LASSERT(obd->obd_recovery_start != 0);
1744 now = ktime_get_real_seconds();
1745 to = obd->obd_recovery_timeout;
1746 end = obd->obd_recovery_start + to;
1749 if (extend && (drt > left)) {
1751 } else if (!extend && (drt > to)) {
1755 if (to > obd->obd_recovery_time_hard) {
1756 to = obd->obd_recovery_time_hard;
1757 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1758 obd->obd_name, to, extend);
1761 if (obd->obd_recovery_timeout < to) {
1762 obd->obd_recovery_timeout = to;
1763 end = obd->obd_recovery_start + to;
1764 mod_timer(&obd->obd_recovery_timer,
1765 jiffies + cfs_time_seconds(end - now));
1767 spin_unlock(&obd->obd_dev_lock);
1769 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1770 obd->obd_name, (s64)(end - now));
1773 /* Reset the timer with each new client connection */
1775 * This timer is actually reconnect_timer, which is for making sure
1776 * the total recovery window is at least as big as my reconnect
1777 * attempt timing. So the initial recovery time_out will be set to
1778 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1779 * from client is bigger than this, then the recovery time_out will
1780 * be extended to make sure the client could be reconnected, in the
1781 * process, the timeout from the new client should be ignored.
1785 check_and_start_recovery_timer(struct obd_device *obd,
1786 struct ptlrpc_request *req,
1789 time64_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1790 struct obd_device_target *obt = &obd->u.obt;
1792 if (!new_client && service_time)
1793 /* Teach server about old server's estimates, as first guess
1794 * at how long new requests will take. */
1795 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1798 target_start_recovery_timer(obd);
1800 /* Convert the service time to RPC timeout,
1801 * and reuse service_time to limit stack usage.
1803 service_time = at_est2timeout(service_time);
1805 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1806 service_time < at_extra)
1807 service_time = at_extra;
1809 /* We expect other clients to timeout within service_time, then try
1810 * to reconnect, then try the failover server. The max delay between
1811 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1812 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1814 LASSERT(obt->obt_magic == OBT_MAGIC);
1815 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1816 if (service_time > obd->obd_recovery_timeout && !new_client)
1817 extend_recovery_timer(obd, service_time, false);
1820 /** Health checking routines */
1821 static inline int exp_connect_healthy(struct obd_export *exp)
1823 return (exp->exp_in_recovery);
1826 /** if export done req_replay or has replay in queue */
1827 static inline int exp_req_replay_healthy(struct obd_export *exp)
1829 return (!exp->exp_req_replay_needed ||
1830 atomic_read(&exp->exp_replay_count) > 0);
1834 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1836 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1837 exp_req_replay_healthy(exp);
1840 /** if export done lock_replay or has replay in queue */
1841 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1843 return (!exp->exp_lock_replay_needed ||
1844 atomic_read(&exp->exp_replay_count) > 0);
1847 static inline int exp_vbr_healthy(struct obd_export *exp)
1849 return (!exp->exp_vbr_failed);
1852 static inline int exp_finished(struct obd_export *exp)
1854 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1857 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1859 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1863 static int check_for_next_transno(struct lu_target *lut)
1865 struct ptlrpc_request *req = NULL;
1866 struct obd_device *obd = lut->lut_obd;
1867 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1868 int wake_up = 0, connected, completed, queue_len;
1869 __u64 req_transno = 0;
1870 __u64 update_transno = 0;
1871 __u64 next_transno = 0;
1874 spin_lock(&obd->obd_recovery_task_lock);
1875 if (!list_empty(&obd->obd_req_replay_queue)) {
1876 req = list_entry(obd->obd_req_replay_queue.next,
1877 struct ptlrpc_request, rq_list);
1878 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1882 update_transno = distribute_txn_get_next_transno(tdtd);
1884 connected = atomic_read(&obd->obd_connected_clients);
1885 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1886 queue_len = obd->obd_requests_queued_for_recovery;
1887 next_transno = obd->obd_next_recovery_transno;
1889 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1890 "req_transno: %llu, next_transno: %llu\n",
1891 obd->obd_max_recoverable_clients, connected, completed,
1892 queue_len, req_transno, next_transno);
1894 if (obd->obd_abort_recovery) {
1895 CDEBUG(D_HA, "waking for aborted recovery\n");
1897 } else if (obd->obd_recovery_expired) {
1898 CDEBUG(D_HA, "waking for expired recovery\n");
1900 } else if (tdtd != NULL && req != NULL &&
1901 is_req_replayed_by_update(req)) {
1902 LASSERTF(req_transno < next_transno, "req_transno %llu"
1903 "next_transno%llu\n", req_transno, next_transno);
1904 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1907 } else if (req_transno == next_transno ||
1908 (update_transno != 0 && update_transno <= next_transno)) {
1909 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1911 } else if (queue_len > 0 &&
1912 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1913 /** handle gaps occured due to lost reply or VBR */
1914 LASSERTF(req_transno >= next_transno,
1915 "req_transno: %llu, next_transno: %llu\n",
1916 req_transno, next_transno);
1918 "%s: waking for gap in transno, VBR is %s (skip: "
1919 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1920 ", next_update %lld last_committed: %lld)\n",
1921 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1922 next_transno, queue_len, completed, connected,
1923 req_transno, update_transno, obd->obd_last_committed);
1924 obd->obd_next_recovery_transno = req_transno;
1926 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1927 CDEBUG(D_HA, "waking for completed recovery\n");
1929 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1930 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1931 " by fail_lock, waking up (%lld)\n", next_transno);
1932 obd->obd_next_recovery_transno = req_transno;
1935 spin_unlock(&obd->obd_recovery_task_lock);
1939 static int check_for_next_lock(struct lu_target *lut)
1941 struct obd_device *obd = lut->lut_obd;
1944 spin_lock(&obd->obd_recovery_task_lock);
1945 if (!list_empty(&obd->obd_lock_replay_queue)) {
1946 CDEBUG(D_HA, "waking for next lock\n");
1948 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1949 CDEBUG(D_HA, "waking for completed lock replay\n");
1951 } else if (obd->obd_abort_recovery) {
1952 CDEBUG(D_HA, "waking for aborted recovery\n");
1954 } else if (obd->obd_recovery_expired) {
1955 CDEBUG(D_HA, "waking for expired recovery\n");
1958 spin_unlock(&obd->obd_recovery_task_lock);
1964 * wait for recovery events,
1965 * check its status with help of check_routine
1966 * evict dead clients via health_check
1968 static int target_recovery_overseer(struct lu_target *lut,
1969 int (*check_routine)(struct lu_target *),
1970 int (*health_check)(struct obd_export *))
1972 struct obd_device *obd = lut->lut_obd;
1973 struct target_distribute_txn_data *tdtd;
1977 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
1978 now = ktime_get_seconds();
1979 if (now - last > 600) {
1980 LCONSOLE_INFO("%s: in recovery but waiting for "
1981 "the first client to connect\n",
1986 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
1987 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
1988 __u64 next_update_transno = 0;
1990 /* Only abort the recovery if there are no update recovery
1991 * left in the queue */
1992 spin_lock(&obd->obd_recovery_task_lock);
1993 if (lut->lut_tdtd != NULL) {
1994 next_update_transno =
1995 distribute_txn_get_next_transno(lut->lut_tdtd);
1997 tdtd = lut->lut_tdtd;
1998 /* If next_update_transno == 0, it probably because
1999 * updatelog retrieve threads did not get any records
2000 * yet, let's wait those threads stopped */
2001 if (next_update_transno == 0) {
2002 struct l_wait_info lwi = { 0 };
2004 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2006 &tdtd->tdtd_recovery_threads_count) == 0,
2009 next_update_transno =
2010 distribute_txn_get_next_transno(
2015 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2016 obd->obd_next_recovery_transno = next_update_transno;
2017 spin_unlock(&obd->obd_recovery_task_lock);
2018 /* Disconnect unfinished exports from clients, and
2019 * keep connection from MDT to make sure the update
2020 * recovery will still keep trying until some one
2021 * manually abort the recovery */
2022 class_disconnect_stale_exports(obd,
2023 exp_finished_or_from_mdt);
2024 /* Abort all of replay and replay lock req from
2026 abort_req_replay_queue(obd);
2027 abort_lock_replay_queue(obd);
2028 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2029 ")in the queue.\n", obd->obd_name,
2030 next_update_transno);
2032 obd->obd_abort_recovery = 1;
2033 spin_unlock(&obd->obd_recovery_task_lock);
2034 CWARN("%s recovery is aborted by hard timeout\n",
2039 while (wait_event_timeout(obd->obd_next_transno_waitq,
2041 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2042 /* wait indefinitely for event, but don't trigger watchdog */;
2044 if (obd->obd_abort_recovery) {
2045 CWARN("recovery is aborted, evict exports in recovery\n");
2046 if (lut->lut_tdtd != NULL) {
2047 struct l_wait_info lwi = { 0 };
2049 tdtd = lut->lut_tdtd;
2050 /* Let's wait all of the update log recovery thread
2052 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2053 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2055 /* Then abort the update recovery list */
2056 dtrq_list_destroy(lut->lut_tdtd);
2059 /** evict exports which didn't finish recovery yet */
2060 class_disconnect_stale_exports(obd, exp_finished);
2062 } else if (obd->obd_recovery_expired) {
2063 obd->obd_recovery_expired = 0;
2064 /** If some clients died being recovered, evict them */
2065 LCONSOLE_WARN("%s: recovery is timed out, "
2066 "evict stale exports\n", obd->obd_name);
2067 /** evict cexports with no replay in queue, they are stalled */
2068 class_disconnect_stale_exports(obd, health_check);
2070 /** continue with VBR */
2071 spin_lock(&obd->obd_dev_lock);
2072 obd->obd_version_recov = 1;
2073 spin_unlock(&obd->obd_dev_lock);
2075 * reset timer, recovery will proceed with versions now,
2076 * timeout is set just to handle reconnection delays
2078 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2079 /** Wait for recovery events again, after evicting bad clients */
2085 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2087 struct obd_device *obd = lut->lut_obd;
2088 struct ptlrpc_request *req = NULL;
2090 CDEBUG(D_HA, "Waiting for lock\n");
2091 if (target_recovery_overseer(lut, check_for_next_lock,
2092 exp_lock_replay_healthy))
2093 abort_lock_replay_queue(obd);
2095 spin_lock(&obd->obd_recovery_task_lock);
2096 if (!list_empty(&obd->obd_lock_replay_queue)) {
2097 req = list_entry(obd->obd_lock_replay_queue.next,
2098 struct ptlrpc_request, rq_list);
2099 list_del_init(&req->rq_list);
2100 spin_unlock(&obd->obd_recovery_task_lock);
2102 spin_unlock(&obd->obd_recovery_task_lock);
2103 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2104 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2105 /** evict exports failed VBR */
2106 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2111 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2113 struct ptlrpc_request *req = NULL;
2115 spin_lock(&obd->obd_recovery_task_lock);
2116 if (!list_empty(&obd->obd_final_req_queue)) {
2117 req = list_entry(obd->obd_final_req_queue.next,
2118 struct ptlrpc_request, rq_list);
2119 list_del_init(&req->rq_list);
2120 spin_unlock(&obd->obd_recovery_task_lock);
2121 if (req->rq_export->exp_in_recovery) {
2122 spin_lock(&req->rq_export->exp_lock);
2123 req->rq_export->exp_in_recovery = 0;
2124 spin_unlock(&req->rq_export->exp_lock);
2127 spin_unlock(&obd->obd_recovery_task_lock);
2132 static void handle_recovery_req(struct ptlrpc_thread *thread,
2133 struct ptlrpc_request *req,
2134 svc_handler_t handler)
2139 * export can be evicted during recovery, no need to handle replays for
2140 * it after that, discard such request silently
2142 if (req->rq_export->exp_disconnected)
2145 req->rq_session.lc_thread = thread;
2146 req->rq_svc_thread = thread;
2147 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2149 /* thread context */
2150 lu_context_enter(&thread->t_env->le_ctx);
2152 lu_context_exit(&thread->t_env->le_ctx);
2154 /* don't reset timer for final stage */
2155 if (!exp_finished(req->rq_export)) {
2156 time64_t to = obd_timeout;
2159 * Add request timeout to the recovery time so next request from
2160 * this client may come in recovery time
2163 struct ptlrpc_service_part *svcpt;
2165 svcpt = req->rq_rqbd->rqbd_svcpt;
2166 /* If the server sent early reply for this request,
2167 * the client will recalculate the timeout according to
2168 * current server estimate service time, so we will
2169 * use the maxium timeout here for waiting the client
2170 * sending the next req */
2171 to = max((int)at_est2timeout(
2172 at_get(&svcpt->scp_at_estimate)),
2173 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2174 /* Add 2 net_latency, one for balance rq_deadline
2175 * (see ptl_send_rpc), one for resend the req to server,
2176 * Note: client will pack net_latency in replay req
2177 * (see ptlrpc_replay_req) */
2178 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2180 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2185 /** Checking routines for recovery */
2186 static int check_for_recovery_ready(struct lu_target *lut)
2188 struct obd_device *obd = lut->lut_obd;
2189 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2191 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2192 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2193 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2194 obd->obd_recovery_expired);
2196 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2197 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2198 if (clnts + obd->obd_stale_clients <
2199 obd->obd_max_recoverable_clients)
2203 if (lut->lut_tdtd != NULL) {
2204 if (!lut->lut_tdtd->tdtd_replay_ready &&
2205 !obd->obd_abort_recovery) {
2206 /* Let's extend recovery timer, in case the recovery
2207 * timer expired, and some clients got evicted */
2208 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2210 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2211 obd->obd_name, obd->obd_recovery_timeout);
2220 REQUEST_RECOVERY = 1,
2221 UPDATE_RECOVERY = 2,
2224 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2228 if (!list_empty(&obd->obd_req_replay_queue)) {
2229 struct ptlrpc_request *req;
2231 req = list_entry(obd->obd_req_replay_queue.next,
2232 struct ptlrpc_request, rq_list);
2233 transno = lustre_msg_get_transno(req->rq_reqmsg);
2239 static __u64 get_next_transno(struct lu_target *lut, int *type)
2241 struct obd_device *obd = lut->lut_obd;
2242 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2244 __u64 update_transno;
2247 transno = get_next_replay_req_transno(obd);
2249 *type = REQUEST_RECOVERY;
2254 update_transno = distribute_txn_get_next_transno(tdtd);
2255 if (transno == 0 || (transno >= update_transno &&
2256 update_transno != 0)) {
2257 transno = update_transno;
2259 *type = UPDATE_RECOVERY;
2266 * drop duplicate replay request
2268 * Because the operation has been replayed by update recovery, the request
2269 * with the same transno will be dropped and also notify the client to send
2270 * next replay request.
2272 * \param[in] env execution environment
2273 * \param[in] obd failover obd device
2274 * \param[in] req request to be dropped
2276 static void drop_duplicate_replay_req(struct lu_env *env,
2277 struct obd_device *obd,
2278 struct ptlrpc_request *req)
2280 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2281 " update records are found.\n",
2282 lustre_msg_get_transno(req->rq_reqmsg),
2283 libcfs_nid2str(req->rq_peer.nid));
2285 /* Right now, only for MDS reint operation update replay and
2286 * normal request replay can have the same transno */
2287 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2288 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2289 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2290 if (likely(req->rq_export))
2291 target_committed_to_req(req);
2292 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2293 target_send_reply(req, req->rq_status, 0);
2295 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2296 libcfs_nid2str(req->rq_peer.nid));
2298 target_exp_dequeue_req_replay(req);
2299 target_request_copy_put(req);
2300 obd->obd_replayed_requests++;
2303 static void replay_request_or_update(struct lu_env *env,
2304 struct lu_target *lut,
2305 struct target_recovery_data *trd,
2306 struct ptlrpc_thread *thread)
2308 struct obd_device *obd = lut->lut_obd;
2309 struct ptlrpc_request *req = NULL;
2314 CDEBUG(D_HA, "Waiting for transno %lld\n",
2315 obd->obd_next_recovery_transno);
2317 /* Replay all of request and update by transno */
2319 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2321 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2323 /** It is needed to extend recovery window above
2324 * recovery_time_soft. Extending is possible only in the
2325 * end of recovery window (see more details in
2326 * handle_recovery_req()).
2328 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2330 if (target_recovery_overseer(lut, check_for_next_transno,
2331 exp_req_replay_healthy_or_from_mdt)) {
2332 abort_req_replay_queue(obd);
2333 abort_lock_replay_queue(obd);
2337 spin_lock(&obd->obd_recovery_task_lock);
2338 transno = get_next_transno(lut, &type);
2339 if (type == REQUEST_RECOVERY && transno != 0) {
2340 /* Drop replay request from client side, if the
2341 * replay has been executed by update with the
2343 req = list_entry(obd->obd_req_replay_queue.next,
2344 struct ptlrpc_request, rq_list);
2346 list_del_init(&req->rq_list);
2347 obd->obd_requests_queued_for_recovery--;
2348 spin_unlock(&obd->obd_recovery_task_lock);
2350 /* Let's check if the request has been redone by
2352 if (is_req_replayed_by_update(req)) {
2353 struct distribute_txn_replay_req *dtrq;
2355 dtrq = distribute_txn_lookup_finish_list(tdtd,
2357 LASSERT(dtrq != NULL);
2358 spin_lock(&tdtd->tdtd_replay_list_lock);
2359 list_del_init(&dtrq->dtrq_list);
2360 spin_unlock(&tdtd->tdtd_replay_list_lock);
2363 drop_duplicate_replay_req(env, obd, req);
2368 LASSERT(trd->trd_processing_task == current_pid());
2369 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2370 lustre_msg_get_transno(req->rq_reqmsg),
2371 libcfs_nid2str(req->rq_peer.nid));
2373 handle_recovery_req(thread, req,
2374 trd->trd_recovery_handler);
2376 * bz18031: increase next_recovery_transno before
2377 * target_request_copy_put() will drop exp_rpc reference
2379 spin_lock(&obd->obd_recovery_task_lock);
2380 obd->obd_next_recovery_transno++;
2381 spin_unlock(&obd->obd_recovery_task_lock);
2382 target_exp_dequeue_req_replay(req);
2383 target_request_copy_put(req);
2384 obd->obd_replayed_requests++;
2385 } else if (type == UPDATE_RECOVERY && transno != 0) {
2386 struct distribute_txn_replay_req *dtrq;
2389 spin_unlock(&obd->obd_recovery_task_lock);
2391 LASSERT(tdtd != NULL);
2392 dtrq = distribute_txn_get_next_req(tdtd);
2393 lu_context_enter(&thread->t_env->le_ctx);
2394 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2395 lu_context_exit(&thread->t_env->le_ctx);
2396 extend_recovery_timer(obd, obd_timeout, true);
2398 if (rc == 0 && dtrq->dtrq_xid != 0) {
2399 CDEBUG(D_HA, "Move x%llu t%llu"
2400 " to finish list\n", dtrq->dtrq_xid,
2401 dtrq->dtrq_master_transno);
2403 /* Add it to the replay finish list */
2404 spin_lock(&tdtd->tdtd_replay_list_lock);
2405 list_add(&dtrq->dtrq_list,
2406 &tdtd->tdtd_replay_finish_list);
2407 spin_unlock(&tdtd->tdtd_replay_list_lock);
2409 spin_lock(&obd->obd_recovery_task_lock);
2410 if (transno == obd->obd_next_recovery_transno)
2411 obd->obd_next_recovery_transno++;
2413 obd->obd_next_recovery_transno)
2414 obd->obd_next_recovery_transno =
2416 spin_unlock(&obd->obd_recovery_task_lock);
2421 spin_unlock(&obd->obd_recovery_task_lock);
2423 LASSERT(list_empty(&obd->obd_req_replay_queue));
2424 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2425 /** evict exports failed VBR */
2426 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2432 static int target_recovery_thread(void *arg)
2434 struct lu_target *lut = arg;
2435 struct obd_device *obd = lut->lut_obd;
2436 struct ptlrpc_request *req;
2437 struct target_recovery_data *trd = &obd->obd_recovery_data;
2438 unsigned long delta;
2440 struct ptlrpc_thread *thread = NULL;
2444 unshare_fs_struct();
2445 OBD_ALLOC_PTR(thread);
2451 OBD_FREE_PTR(thread);
2455 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2457 OBD_FREE_PTR(thread);
2462 thread->t_env = env;
2463 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2464 env->le_ctx.lc_thread = thread;
2465 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2466 thread->t_watchdog = NULL;
2468 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2470 trd->trd_processing_task = current_pid();
2472 spin_lock(&obd->obd_dev_lock);
2473 obd->obd_recovering = 1;
2474 spin_unlock(&obd->obd_dev_lock);
2475 complete(&trd->trd_starting);
2477 /* first of all, we have to know the first transno to replay */
2478 if (target_recovery_overseer(lut, check_for_recovery_ready,
2479 exp_connect_healthy)) {
2480 abort_req_replay_queue(obd);
2481 abort_lock_replay_queue(obd);
2482 if (lut->lut_tdtd != NULL)
2483 dtrq_list_destroy(lut->lut_tdtd);
2486 /* next stage: replay requests or update */
2488 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2489 atomic_read(&obd->obd_req_replay_clients),
2490 obd->obd_next_recovery_transno);
2491 replay_request_or_update(env, lut, trd, thread);
2494 * The second stage: replay locks
2496 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2497 atomic_read(&obd->obd_lock_replay_clients));
2498 while ((req = target_next_replay_lock(lut))) {
2499 LASSERT(trd->trd_processing_task == current_pid());
2500 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2501 libcfs_nid2str(req->rq_peer.nid));
2502 handle_recovery_req(thread, req,
2503 trd->trd_recovery_handler);
2504 target_request_copy_put(req);
2505 obd->obd_replayed_locks++;
2509 * The third stage: reply on final pings, at this moment all clients
2510 * must have request in final queue
2512 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2513 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2514 /** Update server last boot epoch */
2515 tgt_boot_epoch_update(lut);
2516 /* We drop recoverying flag to forward all new requests
2517 * to regular mds_handle() since now */
2518 spin_lock(&obd->obd_dev_lock);
2519 obd->obd_recovering = obd->obd_abort_recovery = 0;
2520 spin_unlock(&obd->obd_dev_lock);
2521 spin_lock(&obd->obd_recovery_task_lock);
2522 target_cancel_recovery_timer(obd);
2523 spin_unlock(&obd->obd_recovery_task_lock);
2524 while ((req = target_next_final_ping(obd))) {
2525 LASSERT(trd->trd_processing_task == current_pid());
2526 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2527 libcfs_nid2str(req->rq_peer.nid));
2528 handle_recovery_req(thread, req,
2529 trd->trd_recovery_handler);
2530 /* Because the waiting client can not send ping to server,
2531 * so we need refresh the last_request_time, to avoid the
2532 * export is being evicted */
2533 ptlrpc_update_export_timer(req->rq_export, 0);
2534 target_request_copy_put(req);
2537 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2538 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2539 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2540 if (delta > OBD_RECOVERY_TIME_SOFT) {
2541 CWARN("too long recovery - read logs\n");
2542 libcfs_debug_dumplog();
2545 target_finish_recovery(lut);
2547 lu_context_fini(&env->le_ctx);
2548 trd->trd_processing_task = 0;
2549 complete(&trd->trd_finishing);
2551 tgt_io_thread_done(thread);
2552 OBD_FREE_PTR(thread);
2557 static int target_start_recovery_thread(struct lu_target *lut,
2558 svc_handler_t handler)
2560 struct obd_device *obd = lut->lut_obd;
2562 struct target_recovery_data *trd = &obd->obd_recovery_data;
2565 memset(trd, 0, sizeof(*trd));
2566 init_completion(&trd->trd_starting);
2567 init_completion(&trd->trd_finishing);
2568 trd->trd_recovery_handler = handler;
2570 rc = server_name2index(obd->obd_name, &index, NULL);
2574 if (!IS_ERR(kthread_run(target_recovery_thread,
2575 lut, "tgt_recover_%d", index))) {
2576 wait_for_completion(&trd->trd_starting);
2577 LASSERT(obd->obd_recovering != 0);
2585 void target_stop_recovery_thread(struct obd_device *obd)
2587 if (obd->obd_recovery_data.trd_processing_task > 0) {
2588 struct target_recovery_data *trd = &obd->obd_recovery_data;
2589 /** recovery can be done but postrecovery is not yet */
2590 spin_lock(&obd->obd_dev_lock);
2591 if (obd->obd_recovering) {
2592 CERROR("%s: Aborting recovery\n", obd->obd_name);
2593 obd->obd_abort_recovery = 1;
2594 wake_up(&obd->obd_next_transno_waitq);
2596 spin_unlock(&obd->obd_dev_lock);
2597 wait_for_completion(&trd->trd_finishing);
2600 EXPORT_SYMBOL(target_stop_recovery_thread);
2602 void target_recovery_fini(struct obd_device *obd)
2604 class_disconnect_exports(obd);
2605 target_stop_recovery_thread(obd);
2606 target_cleanup_recovery(obd);
2608 EXPORT_SYMBOL(target_recovery_fini);
2610 static void target_recovery_expired(cfs_timer_cb_arg_t data)
2612 struct obd_device *obd = cfs_from_timer(obd, data, obd_recovery_timer);
2613 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2614 " after %llus (%d clients connected)\n",
2615 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2616 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2617 atomic_read(&obd->obd_connected_clients));
2619 obd->obd_recovery_expired = 1;
2620 wake_up(&obd->obd_next_transno_waitq);
2623 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2625 struct obd_device *obd = lut->lut_obd;
2627 if (lut->lut_bottom->dd_rdonly)
2630 if (obd->obd_max_recoverable_clients == 0) {
2631 /** Update server last boot epoch */
2632 tgt_boot_epoch_update(lut);
2636 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2637 "last_transno %llu\n", obd->obd_name,
2638 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2639 LASSERT(obd->obd_stopping == 0);
2640 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2641 obd->obd_recovery_start = 0;
2642 obd->obd_recovery_end = 0;
2644 cfs_timer_setup(&obd->obd_recovery_timer, target_recovery_expired,
2645 (unsigned long)obd, 0);
2646 target_start_recovery_thread(lut, handler);
2648 EXPORT_SYMBOL(target_recovery_init);
2650 static int target_process_req_flags(struct obd_device *obd,
2651 struct ptlrpc_request *req)
2653 struct obd_export *exp = req->rq_export;
2654 LASSERT(exp != NULL);
2655 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2656 /* client declares he's ready to replay locks */
2657 spin_lock(&exp->exp_lock);
2658 if (exp->exp_req_replay_needed) {
2659 exp->exp_req_replay_needed = 0;
2660 spin_unlock(&exp->exp_lock);
2662 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2663 atomic_dec(&obd->obd_req_replay_clients);
2665 spin_unlock(&exp->exp_lock);
2668 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2669 /* client declares he's ready to complete recovery
2670 * so, we put the request on th final queue */
2671 spin_lock(&exp->exp_lock);
2672 if (exp->exp_lock_replay_needed) {
2673 exp->exp_lock_replay_needed = 0;
2674 spin_unlock(&exp->exp_lock);
2676 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2677 atomic_dec(&obd->obd_lock_replay_clients);
2679 spin_unlock(&exp->exp_lock);
2685 int target_queue_recovery_request(struct ptlrpc_request *req,
2686 struct obd_device *obd)
2688 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2689 struct ptlrpc_request *reqiter;
2693 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2694 /* Processing the queue right now, don't re-add. */
2698 target_process_req_flags(obd, req);
2700 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2701 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2702 if (cfs_fail_val == 1) {
2705 wake_up(&cfs_race_waitq);
2707 set_current_state(TASK_INTERRUPTIBLE);
2708 schedule_timeout(cfs_time_seconds(1));
2712 /* client declares he's ready to complete recovery
2713 * so, we put the request on th final queue */
2714 target_request_copy_get(req);
2715 DEBUG_REQ(D_HA, req, "queue final req");
2716 wake_up(&obd->obd_next_transno_waitq);
2717 spin_lock(&obd->obd_recovery_task_lock);
2718 if (obd->obd_recovering) {
2719 struct ptlrpc_request *tmp;
2720 struct ptlrpc_request *duplicate = NULL;
2722 if (likely(!req->rq_export->exp_replay_done)) {
2723 req->rq_export->exp_replay_done = 1;
2724 list_add_tail(&req->rq_list,
2725 &obd->obd_final_req_queue);
2726 spin_unlock(&obd->obd_recovery_task_lock);
2730 /* XXX O(n), but only happens if final ping is
2731 * timed out, probably reorganize the list as
2732 * a hash list later */
2733 list_for_each_entry_safe(reqiter, tmp,
2734 &obd->obd_final_req_queue,
2736 if (reqiter->rq_export == req->rq_export) {
2737 list_del_init(&reqiter->rq_list);
2738 duplicate = reqiter;
2743 list_add_tail(&req->rq_list,
2744 &obd->obd_final_req_queue);
2745 req->rq_export->exp_replay_done = 1;
2746 spin_unlock(&obd->obd_recovery_task_lock);
2748 if (duplicate != NULL) {
2749 DEBUG_REQ(D_HA, duplicate,
2750 "put prev final req\n");
2751 target_request_copy_put(duplicate);
2755 spin_unlock(&obd->obd_recovery_task_lock);
2756 target_request_copy_put(req);
2757 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2760 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2761 /* client declares he's ready to replay locks */
2762 target_request_copy_get(req);
2763 DEBUG_REQ(D_HA, req, "queue lock replay req");
2764 wake_up(&obd->obd_next_transno_waitq);
2765 spin_lock(&obd->obd_recovery_task_lock);
2766 LASSERT(obd->obd_recovering);
2767 /* usually due to recovery abort */
2768 if (!req->rq_export->exp_in_recovery) {
2769 spin_unlock(&obd->obd_recovery_task_lock);
2770 target_request_copy_put(req);
2773 LASSERT(req->rq_export->exp_lock_replay_needed);
2774 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2775 spin_unlock(&obd->obd_recovery_task_lock);
2779 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2780 * (i.e. buflens etc are in my own byte order), but type-dependent
2781 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2784 INIT_LIST_HEAD(&req->rq_list);
2785 DEBUG_REQ(D_HA, req, "not queueing");
2789 /* If we're processing the queue, we want don't want to queue this
2792 * Also, if this request has a transno less than the one we're waiting
2793 * for, we should process it now. It could (and currently always will)
2794 * be an open request for a descriptor that was opened some time ago.
2796 * Also, a resent, replayed request that has already been
2797 * handled will pass through here and be processed immediately.
2799 CDEBUG(D_HA, "Next recovery transno: %llu"
2800 ", current: %llu, replaying\n",
2801 obd->obd_next_recovery_transno, transno);
2803 /* If the request has been replayed by update replay, then sends this
2804 * request to the recovery thread (replay_request_or_update()), where
2805 * it will be handled */
2806 spin_lock(&obd->obd_recovery_task_lock);
2807 if (transno < obd->obd_next_recovery_transno &&
2808 !is_req_replayed_by_update(req)) {
2809 /* Processing the queue right now, don't re-add. */
2810 LASSERT(list_empty(&req->rq_list));
2811 spin_unlock(&obd->obd_recovery_task_lock);
2814 spin_unlock(&obd->obd_recovery_task_lock);
2816 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2819 target_request_copy_get(req);
2820 if (!req->rq_export->exp_in_recovery) {
2821 target_request_copy_put(req);
2824 LASSERT(req->rq_export->exp_req_replay_needed);
2826 if (target_exp_enqueue_req_replay(req)) {
2827 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2828 target_request_copy_put(req);
2833 spin_lock(&obd->obd_recovery_task_lock);
2834 LASSERT(obd->obd_recovering);
2835 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2836 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2837 list_add_tail(&req->rq_list, &reqiter->rq_list);
2842 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2844 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2845 "has been claimed by another client");
2846 spin_unlock(&obd->obd_recovery_task_lock);
2847 target_exp_dequeue_req_replay(req);
2848 target_request_copy_put(req);
2854 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2856 obd->obd_requests_queued_for_recovery++;
2857 spin_unlock(&obd->obd_recovery_task_lock);
2858 wake_up(&obd->obd_next_transno_waitq);
2862 int target_handle_ping(struct ptlrpc_request *req)
2864 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2865 return req_capsule_server_pack(&req->rq_pill);
2868 void target_committed_to_req(struct ptlrpc_request *req)
2870 struct obd_export *exp = req->rq_export;
2872 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2873 lustre_msg_set_last_committed(req->rq_repmsg,
2874 exp->exp_last_committed);
2876 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2877 "%d)", exp->exp_obd->obd_no_transno,
2878 req->rq_repmsg == NULL);
2880 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2881 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2884 #endif /* HAVE_SERVER_SUPPORT */
2887 * Packs current SLV and Limit into \a req.
2889 int target_pack_pool_reply(struct ptlrpc_request *req)
2891 struct obd_device *obd;
2894 /* Check that we still have all structures alive as this may
2895 * be some late RPC at shutdown time. */
2896 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2897 !exp_connect_lru_resize(req->rq_export))) {
2898 lustre_msg_set_slv(req->rq_repmsg, 0);
2899 lustre_msg_set_limit(req->rq_repmsg, 0);
2903 /* OBD is alive here as export is alive, which we checked above. */
2904 obd = req->rq_export->exp_obd;
2906 read_lock(&obd->obd_pool_lock);
2907 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2908 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2909 read_unlock(&obd->obd_pool_lock);
2914 static int target_send_reply_msg(struct ptlrpc_request *req,
2915 int rc, int fail_id)
2917 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2918 DEBUG_REQ(D_ERROR, req, "dropping reply");
2921 /* We can have a null rq_reqmsg in the event of bad signature or
2922 * no context when unwrapping */
2923 if (req->rq_reqmsg &&
2924 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2925 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2929 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2930 req->rq_status = rc;
2931 return ptlrpc_send_error(req, 1);
2933 DEBUG_REQ(D_NET, req, "sending reply");
2936 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2939 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2941 struct ptlrpc_service_part *svcpt;
2943 struct ptlrpc_reply_state *rs;
2944 struct obd_export *exp;
2947 if (req->rq_no_reply) {
2952 svcpt = req->rq_rqbd->rqbd_svcpt;
2953 rs = req->rq_reply_state;
2954 if (rs == NULL || !rs->rs_difficult) {
2956 target_send_reply_msg (req, rc, fail_id);
2961 /* must be an export if locks saved */
2962 LASSERT(req->rq_export != NULL);
2963 /* req/reply consistent */
2964 LASSERT(rs->rs_svcpt == svcpt);
2967 LASSERT(!rs->rs_scheduled);
2968 LASSERT(!rs->rs_scheduled_ever);
2969 LASSERT(!rs->rs_handled);
2970 LASSERT(!rs->rs_on_net);
2971 LASSERT(rs->rs_export == NULL);
2972 LASSERT(list_empty(&rs->rs_obd_list));
2973 LASSERT(list_empty(&rs->rs_exp_list));
2975 exp = class_export_get(req->rq_export);
2977 /* disable reply scheduling while I'm setting up */
2978 rs->rs_scheduled = 1;
2980 rs->rs_xid = req->rq_xid;
2981 rs->rs_transno = req->rq_transno;
2982 rs->rs_export = exp;
2983 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2985 spin_lock(&exp->exp_uncommitted_replies_lock);
2986 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2987 rs->rs_transno, exp->exp_last_committed);
2988 if (rs->rs_transno > exp->exp_last_committed) {
2989 /* not committed already */
2990 list_add_tail(&rs->rs_obd_list,
2991 &exp->exp_uncommitted_replies);
2993 spin_unlock(&exp->exp_uncommitted_replies_lock);
2995 spin_lock(&exp->exp_lock);
2996 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2997 spin_unlock(&exp->exp_lock);
2999 netrc = target_send_reply_msg(req, rc, fail_id);
3001 spin_lock(&svcpt->scp_rep_lock);
3003 atomic_inc(&svcpt->scp_nreps_difficult);
3006 /* error sending: reply is off the net. Also we need +1
3007 * reply ref until ptlrpc_handle_rs() is done
3008 * with the reply state (if the send was successful, there
3009 * would have been +1 ref for the net, which
3010 * reply_out_callback leaves alone) */
3012 ptlrpc_rs_addref(rs);
3015 spin_lock(&rs->rs_lock);
3016 if (rs->rs_transno <= exp->exp_last_committed ||
3017 (!rs->rs_on_net && !rs->rs_no_ack) ||
3018 list_empty(&rs->rs_exp_list) || /* completed already */
3019 list_empty(&rs->rs_obd_list)) {
3020 CDEBUG(D_HA, "Schedule reply immediately\n");
3021 ptlrpc_dispatch_difficult_reply(rs);
3023 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3024 rs->rs_scheduled = 0; /* allow notifier to schedule */
3026 spin_unlock(&rs->rs_lock);
3027 spin_unlock(&svcpt->scp_rep_lock);
3031 enum ldlm_mode lck_compat_array[] = {
3032 [LCK_EX] = LCK_COMPAT_EX,
3033 [LCK_PW] = LCK_COMPAT_PW,
3034 [LCK_PR] = LCK_COMPAT_PR,
3035 [LCK_CW] = LCK_COMPAT_CW,
3036 [LCK_CR] = LCK_COMPAT_CR,
3037 [LCK_NL] = LCK_COMPAT_NL,
3038 [LCK_GROUP] = LCK_COMPAT_GROUP,
3039 [LCK_COS] = LCK_COMPAT_COS,
3043 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3044 * not escape to the user level.
3046 int ldlm_error2errno(enum ldlm_error error)
3052 case ELDLM_LOCK_MATCHED:
3055 case ELDLM_LOCK_CHANGED:
3058 case ELDLM_LOCK_ABORTED:
3061 case ELDLM_LOCK_REPLACED:
3064 case ELDLM_NO_LOCK_DATA:
3067 case ELDLM_NAMESPACE_EXISTS:
3070 case ELDLM_BAD_NAMESPACE:
3074 if (((int)error) < 0) { /* cast to signed type */
3075 result = error; /* as ldlm_error can be unsigned */
3077 CERROR("Invalid DLM result code: %d\n", error);
3083 EXPORT_SYMBOL(ldlm_error2errno);
3086 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3088 enum ldlm_error ldlm_errno2error(int err_no)
3097 error = ELDLM_LOCK_CHANGED;
3100 error = ELDLM_LOCK_ABORTED;
3103 error = ELDLM_LOCK_REPLACED;
3106 error = ELDLM_NO_LOCK_DATA;
3109 error = ELDLM_NAMESPACE_EXISTS;
3112 error = ELDLM_BAD_NAMESPACE;
3120 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3121 void ldlm_dump_export_locks(struct obd_export *exp)
3123 spin_lock(&exp->exp_locks_list_guard);
3124 if (!list_empty(&exp->exp_locks_list)) {
3125 struct ldlm_lock *lock;
3127 CERROR("dumping locks for export %p,"
3128 "ignore if the unmount doesn't hang\n", exp);
3129 list_for_each_entry(lock, &exp->exp_locks_list,
3131 LDLM_ERROR(lock, "lock:");
3133 spin_unlock(&exp->exp_locks_list_guard);
3137 #ifdef HAVE_SERVER_SUPPORT
3138 static int target_bulk_timeout(void *data)
3141 /* We don't fail the connection here, because having the export
3142 * killed makes the (vital) call to commitrw very sad.
3147 static inline const char *bulk2type(struct ptlrpc_request *req)
3149 if (req->rq_bulk_read)
3151 if (req->rq_bulk_write)
3156 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3157 struct l_wait_info *lwi)
3159 struct ptlrpc_request *req = desc->bd_req;
3160 time64_t start = ktime_get_real_seconds();
3166 /* If there is eviction in progress, wait for it to finish. */
3167 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3168 *lwi = LWI_INTR(NULL, NULL);
3169 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3170 !atomic_read(&exp->exp_obd->
3171 obd_evict_inprogress),
3175 /* Check if client was evicted or reconnected already. */
3176 if (exp->exp_failed ||
3177 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3180 if (req->rq_bulk_read)
3181 rc = sptlrpc_svc_wrap_bulk(req, desc);
3183 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3184 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3185 else /* old version, bulk matchbits is rq_xid */
3186 req->rq_mbits = req->rq_xid;
3189 rc = ptlrpc_start_bulk_transfer(desc);
3193 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3194 bulk2type(req), rc);
3198 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3199 ptlrpc_abort_bulk(desc);
3203 /* limit actual bulk transfer to bulk_timeout seconds */
3204 deadline = start + bulk_timeout;
3205 if (deadline > req->rq_deadline)
3206 deadline = req->rq_deadline;
3209 time64_t timeoutl = deadline - ktime_get_real_seconds();
3210 long timeout_jiffies = timeoutl <= 0 ?
3211 1 : cfs_time_seconds(timeoutl);
3212 time64_t rq_deadline;
3214 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3215 cfs_time_seconds(1),
3216 target_bulk_timeout, desc);
3217 rc = l_wait_event(desc->bd_waitq,
3218 !ptlrpc_server_bulk_active(desc) ||
3221 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3223 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3224 /* Wait again if we changed rq_deadline. */
3225 rq_deadline = READ_ONCE(req->rq_deadline);
3226 deadline = start + bulk_timeout;
3227 if (deadline > rq_deadline)
3228 deadline = rq_deadline;
3229 } while (rc == -ETIMEDOUT &&
3230 deadline > ktime_get_real_seconds());
3232 if (rc == -ETIMEDOUT) {
3233 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3234 bulk2type(req), deadline - start,
3235 ktime_get_real_seconds() - deadline);
3236 ptlrpc_abort_bulk(desc);
3237 } else if (exp->exp_failed) {
3238 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3241 ptlrpc_abort_bulk(desc);
3242 } else if (exp->exp_conn_cnt >
3243 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3244 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3246 /* We don't reply anyway. */
3248 ptlrpc_abort_bulk(desc);
3249 } else if (desc->bd_failure) {
3250 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3252 /* XXX should this be a different errno? */
3255 if (req->rq_bulk_write)
3256 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3257 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3258 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3259 bulk2type(req), desc->bd_nob_transferred,
3261 /* XXX should this be a different errno? */
3268 EXPORT_SYMBOL(target_bulk_io);
3270 #endif /* HAVE_SERVER_SUPPORT */