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/
33 * This file deals with various client/target related logic including recovery.
35 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
39 #define DEBUG_SUBSYSTEM S_LDLM
41 #include <cl_object.h>
42 #include <linux/fs_struct.h>
43 #include <linux/jiffies.h>
44 #include <linux/kernel.h>
45 #include <linux/kthread.h>
46 #include <libcfs/libcfs.h>
48 #include <obd_class.h>
49 #include <lustre_dlm.h>
50 #include <lustre_net.h>
51 #include <lustre_sec.h>
52 #include "ldlm_internal.h"
55 * @priority: If non-zero, move the selected connection to the list head.
56 * @create: If zero, only search in existing connections.
58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
59 int priority, int create)
61 struct ptlrpc_connection *ptlrpc_conn;
62 struct obd_import_conn *imp_conn = NULL, *item;
63 u32 refnet = LNET_NET_ANY;
68 if (!create && !priority) {
69 CDEBUG(D_HA, "Nothing to do\n");
73 if (imp->imp_connection &&
74 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
75 /* refnet is used to restrict network connections */
76 refnet = LNET_NID_NET(&imp->imp_connection->c_self);
77 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, refnet);
79 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
84 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
86 GOTO(out_put, rc = -ENOMEM);
89 spin_lock(&imp->imp_lock);
90 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
91 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
93 list_move(&item->oic_item,
95 item->oic_last_attempt = 0;
97 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
98 imp, imp->imp_obd->obd_name, uuid->uuid,
99 (priority ? ", moved to head" : ""));
100 spin_unlock(&imp->imp_lock);
101 GOTO(out_free, rc = 0);
104 /* No existing import connection found for \a uuid. */
106 imp_conn->oic_conn = ptlrpc_conn;
107 imp_conn->oic_uuid = *uuid;
108 imp_conn->oic_last_attempt = 0;
110 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
112 list_add_tail(&imp_conn->oic_item,
113 &imp->imp_conn_list);
114 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
115 imp, imp->imp_obd->obd_name, uuid->uuid,
116 (priority ? "head" : "tail"));
118 spin_unlock(&imp->imp_lock);
119 GOTO(out_free, rc = -ENOENT);
122 spin_unlock(&imp->imp_lock);
126 OBD_FREE(imp_conn, sizeof(*imp_conn));
128 ptlrpc_connection_put(ptlrpc_conn);
132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
134 return import_set_conn(imp, uuid, 1, 0);
137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
140 return import_set_conn(imp, uuid, priority, 1);
142 EXPORT_SYMBOL(client_import_add_conn);
144 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
145 lnet_nid_t prim_nid, int priority)
147 struct ptlrpc_connection *ptlrpc_conn;
150 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, LNET_NIDNET(prim_nid));
152 const char *str_uuid = obd_uuid2str(uuid);
154 rc = class_add_uuid(str_uuid, prim_nid);
156 CERROR("%s: failed to add UUID '%s': rc = %d\n",
157 imp->imp_obd->obd_name, str_uuid, rc);
161 return import_set_conn(imp, uuid, priority, 1);
163 EXPORT_SYMBOL(client_import_dyn_add_conn);
165 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
166 int nid_count, struct obd_uuid *uuid)
168 struct obd_import_conn *conn;
172 if (nid_count <= 0 || !nids)
175 spin_lock(&imp->imp_lock);
176 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
177 if (class_check_uuid(&conn->oic_uuid, nids[0])) {
178 *uuid = conn->oic_uuid;
179 spin_unlock(&imp->imp_lock);
180 rc = class_add_nids_to_uuid(&conn->oic_uuid, nids,
185 spin_unlock(&imp->imp_lock);
188 EXPORT_SYMBOL(client_import_add_nids_to_conn);
190 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
192 struct obd_import_conn *imp_conn;
193 struct obd_export *dlmexp;
198 spin_lock(&imp->imp_lock);
199 if (list_empty(&imp->imp_conn_list)) {
200 LASSERT(!imp->imp_connection);
204 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
205 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
207 LASSERT(imp_conn->oic_conn);
209 if (imp_conn == imp->imp_conn_current) {
210 LASSERT(imp_conn->oic_conn == imp->imp_connection);
212 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
213 imp->imp_state != LUSTRE_IMP_DISCON) {
214 CERROR("can't remove current connection\n");
215 GOTO(out, rc = -EBUSY);
218 ptlrpc_connection_put(imp->imp_connection);
219 imp->imp_connection = NULL;
221 dlmexp = class_conn2export(&imp->imp_dlm_handle);
222 if (dlmexp && dlmexp->exp_connection) {
223 LASSERT(dlmexp->exp_connection ==
225 ptlrpc_connection_put(dlmexp->exp_connection);
226 dlmexp->exp_connection = NULL;
230 class_export_put(dlmexp);
233 list_del(&imp_conn->oic_item);
234 ptlrpc_connection_put(imp_conn->oic_conn);
235 OBD_FREE(imp_conn, sizeof(*imp_conn));
236 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
237 imp, imp->imp_obd->obd_name, uuid->uuid);
242 spin_unlock(&imp->imp_lock);
244 CERROR("connection %s not found\n", uuid->uuid);
247 EXPORT_SYMBOL(client_import_del_conn);
250 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
251 * to find a conn uuid of \a imp which can reach \a peer.
253 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
254 struct obd_uuid *uuid)
256 struct obd_import_conn *conn;
261 spin_lock(&imp->imp_lock);
262 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
263 /* Check if conn UUID does have this peer NID. */
264 if (class_check_uuid(&conn->oic_uuid, peer)) {
265 *uuid = conn->oic_uuid;
270 spin_unlock(&imp->imp_lock);
273 EXPORT_SYMBOL(client_import_find_conn);
275 void client_destroy_import(struct obd_import *imp)
278 * Drop security policy instance after all RPCs have finished/aborted
279 * to let all busy contexts be released.
281 class_import_get(imp);
282 class_destroy_import(imp);
283 sptlrpc_import_sec_put(imp);
284 class_import_put(imp);
286 EXPORT_SYMBOL(client_destroy_import);
289 * Check whether or not the OSC is on MDT.
292 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
294 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
297 static int osc_on_mdt(char *obdname)
301 ptr = strrchr(obdname, '-');
305 if (strncmp(ptr + 1, "MDT", 3) == 0)
312 * Configure an RPC client OBD device.
317 * 3 - inactive-on-startup
318 * 4 - restrictive net
320 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
322 struct client_obd *cli = &obd->u.cli;
323 struct obd_import *imp;
324 struct obd_uuid server_uuid;
325 int rq_portal, rp_portal, connect_op;
326 const char *name = obd->obd_type->typ_name;
327 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
328 char *cli_name = lustre_cfg_buf(lcfg, 0);
329 struct ptlrpc_connection fake_conn = { .c_self = {},
330 .c_remote_uuid.uuid[0] = 0 };
336 * In a more perfect world, we would hang a ptlrpc_client off of
337 * obd_type and just use the values from there.
339 if (!strcmp(name, LUSTRE_OSC_NAME)) {
340 rq_portal = OST_REQUEST_PORTAL;
341 rp_portal = OSC_REPLY_PORTAL;
342 connect_op = OST_CONNECT;
343 cli->cl_sp_me = LUSTRE_SP_CLI;
344 cli->cl_sp_to = LUSTRE_SP_OST;
345 ns_type = LDLM_NS_TYPE_OSC;
346 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
347 !strcmp(name, LUSTRE_LWP_NAME)) {
348 rq_portal = MDS_REQUEST_PORTAL;
349 rp_portal = MDC_REPLY_PORTAL;
350 connect_op = MDS_CONNECT;
351 if (is_lwp_on_ost(cli_name))
352 cli->cl_sp_me = LUSTRE_SP_OST;
353 else if (is_lwp_on_mdt(cli_name))
354 cli->cl_sp_me = LUSTRE_SP_MDT;
356 cli->cl_sp_me = LUSTRE_SP_CLI;
357 cli->cl_sp_to = LUSTRE_SP_MDT;
358 ns_type = LDLM_NS_TYPE_MDC;
359 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
360 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
361 /* OSP_on_MDT for other MDTs */
362 connect_op = MDS_CONNECT;
363 cli->cl_sp_to = LUSTRE_SP_MDT;
364 ns_type = LDLM_NS_TYPE_MDC;
365 rq_portal = OUT_PORTAL;
367 /* OSP on MDT for OST */
368 connect_op = OST_CONNECT;
369 cli->cl_sp_to = LUSTRE_SP_OST;
370 ns_type = LDLM_NS_TYPE_OSC;
371 rq_portal = OST_REQUEST_PORTAL;
373 rp_portal = OSC_REPLY_PORTAL;
374 cli->cl_sp_me = LUSTRE_SP_MDT;
375 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
376 rq_portal = MGS_REQUEST_PORTAL;
377 rp_portal = MGC_REPLY_PORTAL;
378 connect_op = MGS_CONNECT;
379 cli->cl_sp_me = LUSTRE_SP_MGC;
380 cli->cl_sp_to = LUSTRE_SP_MGS;
381 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
382 ns_type = LDLM_NS_TYPE_MGC;
384 CERROR("unknown client OBD type \"%s\", can't setup\n",
389 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
390 CERROR("requires a TARGET UUID\n");
394 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
395 CERROR("client UUID must be less than 38 characters\n");
399 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
400 CERROR("setup requires a SERVER UUID\n");
404 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
405 CERROR("target UUID must be less than 38 characters\n");
409 init_rwsem(&cli->cl_sem);
410 mutex_init(&cli->cl_mgc_mutex);
412 init_rwsem(&cli->cl_seq_rwsem);
413 cli->cl_conn_count = 0;
414 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
415 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
416 sizeof(server_uuid)));
418 cli->cl_dirty_pages = 0;
419 cli->cl_dirty_max_pages = 0;
420 cli->cl_avail_grant = 0;
421 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
423 * cl_dirty_max_pages may be changed at connect time in
424 * ptlrpc_connect_interpret().
426 client_adjust_max_dirty(cli);
427 init_waitqueue_head(&cli->cl_cache_waiters);
428 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
429 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
430 INIT_LIST_HEAD(&cli->cl_loi_write_list);
431 INIT_LIST_HEAD(&cli->cl_loi_read_list);
432 spin_lock_init(&cli->cl_loi_list_lock);
433 atomic_set(&cli->cl_pending_w_pages, 0);
434 atomic_set(&cli->cl_pending_r_pages, 0);
435 cli->cl_r_in_flight = 0;
436 cli->cl_w_in_flight = 0;
438 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
439 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
440 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
441 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
442 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
443 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
446 INIT_LIST_HEAD(&cli->cl_lru_osc);
447 atomic_set(&cli->cl_lru_shrinkers, 0);
448 atomic_long_set(&cli->cl_lru_busy, 0);
449 atomic_long_set(&cli->cl_lru_in_list, 0);
450 INIT_LIST_HEAD(&cli->cl_lru_list);
451 spin_lock_init(&cli->cl_lru_list_lock);
452 atomic_long_set(&cli->cl_unstable_count, 0);
453 INIT_LIST_HEAD(&cli->cl_shrink_list);
454 INIT_LIST_HEAD(&cli->cl_grant_chain);
456 INIT_LIST_HEAD(&cli->cl_flight_waiters);
457 cli->cl_rpcs_in_flight = 0;
459 init_waitqueue_head(&cli->cl_destroy_waitq);
460 atomic_set(&cli->cl_destroy_in_flight, 0);
463 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
464 cli->cl_preferred_cksum_type = 0;
465 #ifdef ENABLE_CHECKSUM
466 /* Turn on checksumming by default. */
467 cli->cl_checksum = 1;
469 * The supported checksum types will be worked out at connect time
470 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
473 cli->cl_cksum_type = cli->cl_supp_cksum_types;
475 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
478 * Set it to possible maximum size. It may be reduced by ocd_brw_size
479 * from OFD after connecting.
481 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
483 cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
486 * set cl_chunkbits default value to PAGE_SHIFT,
487 * it will be updated at OSC connection time.
489 cli->cl_chunkbits = PAGE_SHIFT;
491 if (!strcmp(name, LUSTRE_MDC_NAME)) {
492 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
493 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
494 cli->cl_max_rpcs_in_flight = 2;
495 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
496 cli->cl_max_rpcs_in_flight = 3;
497 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
498 cli->cl_max_rpcs_in_flight = 4;
500 if (osc_on_mdt(obd->obd_name))
501 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
503 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
506 spin_lock_init(&cli->cl_mod_rpcs_lock);
507 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
508 cli->cl_max_mod_rpcs_in_flight = 0;
509 cli->cl_mod_rpcs_in_flight = 0;
510 cli->cl_close_rpcs_in_flight = 0;
511 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
512 cli->cl_mod_tag_bitmap = NULL;
514 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
516 if (connect_op == MDS_CONNECT) {
517 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
518 OBD_ALLOC(cli->cl_mod_tag_bitmap,
519 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
520 if (cli->cl_mod_tag_bitmap == NULL)
521 GOTO(err, rc = -ENOMEM);
526 CERROR("ldlm_get_ref failed: %d\n", rc);
530 ptlrpc_init_client(rq_portal, rp_portal, name,
531 &obd->obd_ldlm_client);
533 imp = class_new_import(obd);
535 GOTO(err_ldlm, rc = -ENOENT);
536 imp->imp_client = &obd->obd_ldlm_client;
537 imp->imp_connect_op = connect_op;
538 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
539 LUSTRE_CFG_BUFLEN(lcfg, 1));
540 class_import_put(imp);
542 if (lustre_cfg_buf(lcfg, 4)) {
543 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
545 if (refnet == LNET_NET_ANY) {
547 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
548 obd->obd_name, lustre_cfg_string(lcfg, 4),
550 GOTO(err_import, rc);
552 lnet_nid4_to_nid(LNET_MKNID(refnet, 0), &fake_conn.c_self);
553 imp->imp_connection = &fake_conn;
556 rc = client_import_add_conn(imp, &server_uuid, 1);
558 CERROR("can't add initial connection\n");
559 GOTO(err_import, rc);
561 imp->imp_connection = NULL;
563 cli->cl_import = imp;
564 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
565 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
567 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
568 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
569 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
571 cli->cl_target_uuid.uuid);
572 spin_lock(&imp->imp_lock);
573 imp->imp_deactive = 1;
574 spin_unlock(&imp->imp_lock);
578 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
579 LDLM_NAMESPACE_CLIENT,
580 LDLM_NAMESPACE_GREEDY,
582 if (IS_ERR(obd->obd_namespace)) {
583 rc = PTR_ERR(obd->obd_namespace);
584 CERROR("%s: unable to create client namespace: rc = %d\n",
586 obd->obd_namespace = NULL;
587 GOTO(err_import, rc);
593 class_destroy_import(imp);
597 if (cli->cl_mod_tag_bitmap != NULL)
598 OBD_FREE(cli->cl_mod_tag_bitmap,
599 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
600 cli->cl_mod_tag_bitmap = NULL;
604 EXPORT_SYMBOL(client_obd_setup);
606 int client_obd_cleanup(struct obd_device *obd)
608 struct client_obd *cli = &obd->u.cli;
612 ldlm_namespace_free_post(obd->obd_namespace);
613 obd->obd_namespace = NULL;
615 obd_cleanup_client_import(obd);
616 LASSERT(obd->u.cli.cl_import == NULL);
620 if (cli->cl_mod_tag_bitmap != NULL)
621 OBD_FREE(cli->cl_mod_tag_bitmap,
622 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
623 cli->cl_mod_tag_bitmap = NULL;
627 EXPORT_SYMBOL(client_obd_cleanup);
629 /* ->o_connect() method for client side (OSC and MDC and MGC) */
630 int client_connect_import(const struct lu_env *env,
631 struct obd_export **exp,
632 struct obd_device *obd, struct obd_uuid *cluuid,
633 struct obd_connect_data *data, void *localdata)
635 struct client_obd *cli = &obd->u.cli;
636 struct obd_import *imp = cli->cl_import;
637 struct obd_connect_data *ocd;
638 struct lustre_handle conn = { 0 };
644 down_write(&cli->cl_sem);
645 if (cli->cl_conn_count > 0)
646 GOTO(out_sem, rc = -EALREADY);
648 rc = class_connect(&conn, obd, cluuid);
652 cli->cl_conn_count++;
653 *exp = class_conn2export(&conn);
655 LASSERT(obd->obd_namespace);
657 imp->imp_dlm_handle = conn;
658 rc = ptlrpc_init_import(imp);
662 ocd = &imp->imp_connect_data;
665 imp->imp_connect_flags_orig = data->ocd_connect_flags;
666 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
669 rc = ptlrpc_connect_import(imp);
671 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
674 LASSERT(*exp != NULL && (*exp)->exp_connection);
677 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
678 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
679 data->ocd_connect_flags, ocd->ocd_connect_flags);
680 data->ocd_connect_flags = ocd->ocd_connect_flags;
681 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
684 ptlrpc_pinger_add_import(imp);
690 cli->cl_conn_count--;
691 class_disconnect(*exp);
695 up_write(&cli->cl_sem);
697 if (!rc && localdata) {
698 LASSERT(cli->cl_cache == NULL); /* only once */
699 cli->cl_cache = (struct cl_client_cache *)localdata;
700 cl_cache_incref(cli->cl_cache);
701 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
703 /* add this osc into entity list */
704 LASSERT(list_empty(&cli->cl_lru_osc));
705 spin_lock(&cli->cl_cache->ccc_lru_lock);
706 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
707 spin_unlock(&cli->cl_cache->ccc_lru_lock);
712 EXPORT_SYMBOL(client_connect_import);
714 int client_disconnect_export(struct obd_export *exp)
716 struct obd_device *obd = class_exp2obd(exp);
717 struct client_obd *cli;
718 struct obd_import *imp;
724 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
725 exp, exp ? exp->exp_handle.h_cookie : -1);
730 imp = cli->cl_import;
732 down_write(&cli->cl_sem);
733 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
736 if (cli->cl_conn_count == 0) {
737 CERROR("disconnecting disconnected device (%s)\n",
739 GOTO(out_disconnect, rc = -EINVAL);
742 cli->cl_conn_count--;
743 if (cli->cl_conn_count != 0)
744 GOTO(out_disconnect, rc = 0);
747 * Mark import deactivated now, so we don't try to reconnect if any
748 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
749 * fully deactivate the import, or that would drop all requests.
751 spin_lock(&imp->imp_lock);
752 imp->imp_deactive = 1;
753 spin_unlock(&imp->imp_lock);
756 * Some non-replayable imports (MDS's OSCs) are pinged, so just
757 * delete it regardless. (It's safe to delete an import that was
760 (void)ptlrpc_pinger_del_import(imp);
762 if (obd->obd_namespace != NULL) {
763 /* obd_force == local only */
764 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
765 obd->obd_force ? LCF_LOCAL : 0, NULL);
766 ldlm_namespace_free_prior(obd->obd_namespace, imp,
771 * There's no need to hold sem while disconnecting an import,
772 * and it may actually cause deadlock in GSS.
774 up_write(&cli->cl_sem);
775 rc = ptlrpc_disconnect_import(imp, 0);
776 down_write(&cli->cl_sem);
778 ptlrpc_invalidate_import(imp);
784 * Use server style - class_disconnect should be always called for
787 err = class_disconnect(exp);
791 up_write(&cli->cl_sem);
795 EXPORT_SYMBOL(client_disconnect_export);
797 #ifdef HAVE_SERVER_SUPPORT
798 int server_disconnect_export(struct obd_export *exp)
804 /* Disconnect early so that clients can't keep using export. */
805 rc = class_disconnect(exp);
806 /* Close import to avoid sending any requests. */
807 if (exp->exp_imp_reverse)
808 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
810 ldlm_bl_thread_wakeup();
812 /* complete all outstanding replies */
813 spin_lock(&exp->exp_lock);
814 while (!list_empty(&exp->exp_outstanding_replies)) {
815 struct ptlrpc_reply_state *rs =
816 list_entry(exp->exp_outstanding_replies.next,
817 struct ptlrpc_reply_state, rs_exp_list);
818 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
820 spin_lock(&svcpt->scp_rep_lock);
822 list_del_init(&rs->rs_exp_list);
824 spin_lock(&rs->rs_lock);
825 /* clear rs_convert_lock to make sure rs is handled and put */
826 rs->rs_convert_lock = 0;
827 ptlrpc_schedule_difficult_reply(rs);
828 spin_unlock(&rs->rs_lock);
830 spin_unlock(&svcpt->scp_rep_lock);
832 spin_unlock(&exp->exp_lock);
836 EXPORT_SYMBOL(server_disconnect_export);
838 static inline int target_check_recovery_timer(struct obd_device *target)
843 if (!target->obd_recovering || target->obd_recovery_start == 0)
846 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
847 timeout = ktime_divns(remaining, NSEC_PER_SEC);
851 /* the recovery timer should expire, but it isn't triggered,
852 * it's better to abort the recovery of this target to speed up
853 * the recovery of the whole cluster. */
854 spin_lock(&target->obd_dev_lock);
855 if (target->obd_recovering) {
856 CERROR("%s: Aborting recovery\n", target->obd_name);
857 target->obd_abort_recovery = 1;
858 wake_up(&target->obd_next_transno_waitq);
860 spin_unlock(&target->obd_dev_lock);
865 * --------------------------------------------------------------------------
866 * from old lib/target.c
867 * --------------------------------------------------------------------------
869 static int target_handle_reconnect(struct lustre_handle *conn,
870 struct obd_export *exp,
871 struct obd_uuid *cluuid)
873 struct obd_device *target;
874 struct lustre_handle *hdl;
880 hdl = &exp->exp_imp_reverse->imp_remote_handle;
881 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
882 conn->cookie = exp->exp_handle.h_cookie;
884 "connect export for UUID '%s' at %p, cookie %#llx\n",
885 cluuid->uuid, exp, conn->cookie);
889 target = exp->exp_obd;
891 /* Might be a re-connect after a partition. */
892 if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
893 LCONSOLE_WARN("%s: already connected client %s (at %s) with handle %#llx. Rejecting client with the same UUID trying to reconnect with handle %#llx\n",
895 obd_uuid2str(&exp->exp_client_uuid),
896 obd_export_nid2str(exp),
897 hdl->cookie, conn->cookie);
898 memset(conn, 0, sizeof(*conn));
900 * target_handle_connect() treats EALREADY and
901 * -EALREADY differently. -EALREADY is an error
902 * (same UUID, different handle).
907 if (!target->obd_recovering) {
908 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
909 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
910 obd_export_nid2str(exp));
911 GOTO(out_already, rc);
914 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
915 timeout = ktime_divns(remaining, NSEC_PER_SEC);
917 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
919 obd_uuid2str(&exp->exp_client_uuid),
920 obd_export_nid2str(exp),
921 atomic_read(&target->obd_max_recoverable_clients),
922 timeout / 60, timeout % 60);
924 struct target_distribute_txn_data *tdtd;
929 target_check_recovery_timer(target);
931 tdtd = class_exp2tgt(exp)->lut_tdtd;
932 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
933 buf = tdtd->tdtd_show_update_logs_retrievers(
934 tdtd->tdtd_show_retrievers_cbdata,
938 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting, waiting for %d MDTs (%s) in recovery for %lld:%.02lld. Please wait until all MDTs recovered or you may force MDT evicition via 'lctl --device %s abort_recovery.\n",
940 obd_uuid2str(&exp->exp_client_uuid),
941 obd_export_nid2str(exp), count,
942 buf ? buf : "unknown (not enough RAM)",
943 (abs(timeout) + target->obd_recovery_timeout) / 60,
944 (abs(timeout) + target->obd_recovery_timeout) % 60,
947 LCONSOLE_WARN("%s: Recovery already passed deadline %lld:%.02lld. If you do not want to wait more, you may force taget eviction via 'lctl --device %s abort_recovery.\n",
948 target->obd_name, abs(timeout) / 60,
949 abs(timeout) % 60, target->obd_name);
956 conn->cookie = exp->exp_handle.h_cookie;
958 * target_handle_connect() treats EALREADY and
959 * -EALREADY differently. EALREADY means we are
960 * doing a valid reconnect from the same client.
966 check_and_start_recovery_timer(struct obd_device *obd,
967 struct ptlrpc_request *req, int new_client);
970 * update flags for import during reconnect process
972 static int rev_import_flags_update(struct obd_import *revimp,
973 struct ptlrpc_request *req)
976 struct obd_connect_data *data;
978 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
980 if (data->ocd_connect_flags & OBD_CONNECT_AT)
981 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
983 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
985 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
987 revimp->imp_connect_data = *data;
988 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
990 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
991 revimp->imp_client->cli_name,
992 libcfs_id2str(req->rq_peer), rc);
1000 * Allocate a new reverse import for an export.
1002 * \retval -errno in case error hit
1003 * \retval 0 if reverse import correctly init
1005 int rev_import_init(struct obd_export *export)
1007 struct obd_device *obd = export->exp_obd;
1008 struct obd_import *revimp;
1010 LASSERT(export->exp_imp_reverse == NULL);
1012 revimp = class_new_import(obd);
1016 revimp->imp_remote_handle.cookie = 0ULL;
1017 revimp->imp_client = &obd->obd_ldlm_client;
1018 revimp->imp_dlm_fake = 1;
1020 /* it is safe to connect import in new state as no sends possible */
1021 spin_lock(&export->exp_lock);
1022 export->exp_imp_reverse = revimp;
1023 spin_unlock(&export->exp_lock);
1024 class_import_put(revimp);
1028 EXPORT_SYMBOL(rev_import_init);
1031 * Handle reconnect for an export.
1033 * \param exp export to handle reconnect process
1034 * \param req client reconnect request
1036 * \retval -rc in case securitfy flavor can't be changed
1037 * \retval 0 in case none problems
1039 static int rev_import_reconnect(struct obd_export *exp,
1040 struct ptlrpc_request *req)
1042 struct obd_import *revimp = exp->exp_imp_reverse;
1043 struct lustre_handle *lh;
1046 /* avoid sending a request until import flags are changed */
1047 ptlrpc_import_enter_resend(revimp);
1049 ptlrpc_connection_put(revimp->imp_connection);
1052 * client from recovery don't have a handle so we need to take from
1053 * request. it may produce situation when wrong client connected
1054 * to recovery as we trust a client uuid
1056 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1057 revimp->imp_remote_handle = *lh;
1060 * unknown versions will be caught in
1061 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1063 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1065 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1067 rc = rev_import_flags_update(revimp, req);
1070 * it is safe to still be in RECOVERY phase as we are not able
1071 * to setup correct security flavor so requests are not able to
1072 * be delivered correctly
1077 /* resend all rpc's via new connection */
1078 return ptlrpc_import_recovery_state_machine(revimp);
1081 int target_handle_connect(struct ptlrpc_request *req)
1083 struct obd_device *target = NULL;
1084 struct obd_export *export = NULL;
1086 * connect handle - filled from target_handle_reconnect in
1089 struct lustre_handle conn;
1090 struct lustre_handle *tmp;
1091 struct obd_uuid cluuid;
1096 bool mds_conn = false, lw_client = false, initial_conn = false;
1097 bool mds_mds_conn = false;
1098 bool new_mds_mds_conn = false;
1099 struct obd_connect_data *data, *tmpdata;
1101 lnet_nid_t *client_nid = NULL;
1102 struct ptlrpc_connection *pcon = NULL;
1106 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1108 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1110 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1111 GOTO(out, rc = -EINVAL);
1114 target = class_dev_by_str(str);
1116 deuuidify(str, NULL, &target_start, &target_len);
1117 LCONSOLE_ERROR_MSG(0x137,
1118 "%s: not available for connect from %s (no target). If you are running an HA pair check that the target is mounted on the other server.\n",
1119 str, libcfs_nid2str(req->rq_peer.nid));
1120 GOTO(out, rc = -ENODEV);
1123 spin_lock(&target->obd_dev_lock);
1125 target->obd_conn_inprogress++;
1127 if (target->obd_stopping || !target->obd_set_up) {
1128 spin_unlock(&target->obd_dev_lock);
1130 deuuidify(str, NULL, &target_start, &target_len);
1131 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1132 target_len, target_start,
1133 libcfs_nid2str(req->rq_peer.nid),
1134 (target->obd_stopping ?
1135 "stopping" : "not set up"));
1136 GOTO(out, rc = -ENODEV);
1139 if (target->obd_no_conn) {
1140 spin_unlock(&target->obd_dev_lock);
1143 "%s: Temporarily refusing client connection from %s\n",
1144 target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1145 GOTO(out, rc = -EAGAIN);
1148 spin_unlock(&target->obd_dev_lock);
1150 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1152 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1153 GOTO(out, rc = -EINVAL);
1156 obd_str2uuid(&cluuid, str);
1158 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1160 GOTO(out, rc = -EPROTO);
1164 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1166 if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1167 GOTO(out, rc = -EPROTO);
1168 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1170 GOTO(out, rc = -EPROTO);
1172 rc = req_capsule_server_pack(&req->rq_pill);
1176 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1178 * Don't allow clients to connect that are using old 1.8 format
1179 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1180 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1181 * FULL20 flag should be set on all connections since 2.0, but no
1182 * longer affects behaviour.
1184 * Later this check will be disabled and the flag can be retired
1185 * completely once interop with 3.0 is no longer needed.
1187 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1188 GOTO(out, rc = -EPROTO);
1191 * Don't allow liblustre clients to connect.
1192 * - testing was disabled in v2_2_50_0-61-g6a75d65
1193 * - building was disabled in v2_5_58_0-28-g7277179
1194 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1195 * - clients were refused connect for version difference > 0.0.1.32
1197 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1198 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1199 GOTO(out, rc = -EPROTO);
1204 * Note: lw_client is needed in MDS-MDS failover during update log
1205 * processing, so we needs to allow lw_client to be connected at
1206 * anytime, instead of only the initial connection
1208 lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1210 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1211 initial_conn = true;
1212 mds_conn = OCD_HAS_FLAG(data, MDS);
1213 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1216 * OBD_CONNECT_MNE_SWAB is removed at 2.14
1217 * Checking OBD_CONNECT_FID can be removed in the future.
1219 * Via check OBD_CONNECT_FID, we can distinguish whether
1220 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1221 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1224 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1225 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1226 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1227 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1228 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1229 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1232 * We do not support the MDT-MDT interoperations with
1233 * different version MDT because of protocol changes.
1235 if (unlikely(major != LUSTRE_MAJOR ||
1236 minor != LUSTRE_MINOR ||
1237 abs(patch - LUSTRE_PATCH) > 3)) {
1238 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1239 target->obd_name, LUSTRE_MAJOR,
1240 LUSTRE_MINOR, LUSTRE_PATCH,
1241 LUSTRE_FIX, major, minor, patch,
1242 OBD_OCD_VERSION_FIX(data->ocd_version),
1243 libcfs_nid2str(req->rq_peer.nid),
1245 GOTO(out, rc = -EPROTO);
1250 /* lctl gets a backstage, all-access pass. */
1251 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1252 goto dont_check_exports;
1254 export = obd_uuid_lookup(target, &cluuid);
1258 /* We've found an export in the hash. */
1260 spin_lock(&export->exp_lock);
1262 if (export->exp_connecting) { /* b=9635, et. al. */
1263 spin_unlock(&export->exp_lock);
1264 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1265 export->exp_obd->obd_name, export,
1266 libcfs_nid2str(req->rq_peer.nid));
1267 class_export_put(export);
1270 } else if ((mds_conn || (lw_client && initial_conn) ||
1271 OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
1272 spin_unlock(&export->exp_lock);
1273 if (req->rq_peer.nid !=
1274 lnet_nid_to_nid4(&export->exp_connection->c_peer.nid)) {
1275 /* MDS or LWP reconnected after failover. */
1276 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1278 lw_client ? "LWP" : "MDS",
1279 libcfs_nid2str(req->rq_peer.nid),
1280 libcfs_nidstr(&export->exp_connection->c_peer.nid));
1282 /* New connection from the same NID. */
1283 LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1285 lw_client ? "LWP" : "MDS",
1286 libcfs_nid2str(req->rq_peer.nid),
1287 OCD_HAS_FLAG(data, MDS_MDS) ?
1291 if (req->rq_peer.nid ==
1292 lnet_nid_to_nid4(&export->exp_connection->c_peer.nid) &&
1293 OCD_HAS_FLAG(data, MDS_MDS)) {
1295 * Because exports between MDTs will always be
1296 * kept, let's do not fail such export if they
1297 * come from the same NID, otherwise it might
1298 * cause eviction between MDTs, which might
1299 * cause namespace inconsistency
1301 spin_lock(&export->exp_lock);
1302 export->exp_connecting = 1;
1303 export->exp_conn_cnt = 0;
1304 spin_unlock(&export->exp_lock);
1305 conn.cookie = export->exp_handle.h_cookie;
1308 class_fail_export(export);
1309 class_export_put(export);
1313 } else if (export->exp_connection != NULL && initial_conn &&
1314 req->rq_peer.nid != lnet_nid_to_nid4(&export->exp_connection->c_peer.nid)) {
1315 spin_unlock(&export->exp_lock);
1316 /* In MDS failover we have static UUID but NID can change. */
1317 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1318 target->obd_name, cluuid.uuid,
1319 libcfs_nid2str(req->rq_peer.nid),
1321 &export->exp_connection->c_peer.nid));
1323 class_export_put(export);
1325 } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1327 spin_unlock(&export->exp_lock);
1330 export->exp_connecting = 1;
1331 spin_unlock(&export->exp_lock);
1332 LASSERT(export->exp_obd == target);
1334 rc = target_handle_reconnect(&conn, export, &cluuid);
1337 /* If we found an export, we already unlocked. */
1340 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1341 } else if (req->rq_export == NULL &&
1342 atomic_read(&export->exp_rpc_count) > 0) {
1343 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1344 target->obd_name, cluuid.uuid,
1345 libcfs_nid2str(req->rq_peer.nid),
1346 refcount_read(&export->exp_handle.h_ref));
1347 GOTO(out, rc = -EBUSY);
1348 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1350 if (!strstr(cluuid.uuid, "mdt"))
1351 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1352 target->obd_name, cluuid.uuid,
1353 libcfs_nid2str(req->rq_peer.nid));
1354 GOTO(out, rc = -EALREADY);
1356 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1362 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1363 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1364 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1365 export, ktime_get_seconds(),
1366 export ? export->exp_last_request_time : 0);
1369 * If this is the first time a client connects, reset the recovery
1370 * timer. Discard lightweight connections which might be local.
1372 if (!lw_client && rc == 0 && target->obd_recovering)
1373 check_and_start_recovery_timer(target, req, export == NULL);
1376 * We want to handle EALREADY but *not* -EALREADY from
1377 * target_handle_reconnect(), return reconnection state in a flag.
1379 if (rc == EALREADY) {
1380 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1386 /* Tell the client if we support replayable requests. */
1387 if (target->obd_replayable)
1388 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1389 client_nid = &req->rq_peer.nid;
1391 if (export == NULL) {
1392 /* allow lightweight connections during recovery */
1394 * allow "new" MDT to be connected during recovery, since we
1395 * need retrieve recovery update records from it
1397 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1398 struct hrtimer *timer = &target->obd_recovery_timer;
1407 connected = atomic_read(&target->obd_connected_clients);
1408 in_progress = atomic_read(&target->obd_lock_replay_clients);
1410 atomic_read(&target->obd_max_recoverable_clients);
1411 stale = target->obd_stale_clients;
1412 remaining = hrtimer_get_remaining(timer);
1413 left = ktime_divns(remaining, NSEC_PER_SEC);
1415 if (ktime_to_ns(remaining) > 0) {
1416 msg = "to recover in";
1419 msg = "already passed deadline";
1422 target_check_recovery_timer(target);
1425 LCONSOLE_WARN("%s: Denying connection for new client %s (at %s), waiting for %d known clients (%d recovered, %d in progress, and %d evicted) %s %lld:%.02lld\n",
1426 target->obd_name, cluuid.uuid,
1427 libcfs_nid2str(req->rq_peer.nid), known,
1428 connected - in_progress, in_progress,
1429 stale, msg, timeout / 60, timeout % 60);
1433 rc = obd_connect(req->rq_svc_thread->t_env,
1434 &export, target, &cluuid, data,
1436 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1437 lustre_msg_add_op_flags(req->rq_repmsg,
1438 MSG_CONNECT_RECOVERING);
1440 conn.cookie = export->exp_handle.h_cookie;
1441 rc = rev_import_init(export);
1445 new_mds_mds_conn = true;
1448 rc = obd_reconnect(req->rq_svc_thread->t_env,
1449 export, target, &cluuid, data, client_nid);
1454 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1455 data->ocd_instance = target->u.obt.obt_instance;
1458 * Return only the parts of obd_connect_data that we understand, so the
1459 * client knows that we don't understand the rest.
1462 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1464 tmpdata = req_capsule_server_get(&req->rq_pill,
1467 * Don't use struct assignment here, because the client reply
1468 * buffer may be smaller/larger than the local struct
1471 memcpy(tmpdata, data, min(tmpsize, size));
1475 * If the client and the server are the same node, we will already
1476 * have an export that really points to the client's DLM export,
1477 * because we have a shared handles table.
1479 * XXX this will go away when shaver stops sending the "connect" handle
1480 * in the real "remote handle" field of the request --phik 24 Apr 2003
1482 ptlrpc_request_change_export(req, export);
1484 pcon = ptlrpc_connection_get(req->rq_peer,
1485 &req->rq_self, &cluuid);
1487 GOTO(out, rc = -ENOTCONN);
1489 spin_lock(&export->exp_lock);
1491 if (export->exp_disconnected) {
1492 spin_unlock(&export->exp_lock);
1493 GOTO(out, rc = -ENODEV);
1495 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1496 spin_unlock(&export->exp_lock);
1498 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1499 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1500 export->exp_conn_cnt,
1501 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1503 GOTO(out, rc = -EALREADY);
1505 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1506 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1508 /* Check to see if connection came from another NID. */
1509 if (export->exp_connection != NULL &&
1510 lnet_nid_to_nid4(&export->exp_connection->c_peer.nid) !=
1512 obd_nid_del(export->exp_obd, export);
1513 ptlrpc_connection_put(export->exp_connection);
1514 export->exp_connection = NULL;
1517 if (export->exp_connection == NULL) {
1518 export->exp_connection = pcon;
1521 obd_nid_add(export->exp_obd, export);
1523 spin_unlock(&export->exp_lock);
1525 lustre_msg_set_handle(req->rq_repmsg, &conn);
1527 rc = rev_import_reconnect(export, req);
1531 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1533 __u64 transno = data->ocd_transno;
1535 spin_lock(&export->exp_lock);
1537 * possible race with class_disconnect_stale_exports,
1538 * export may be already in the eviction process
1540 if (export->exp_failed) {
1541 spin_unlock(&export->exp_lock);
1542 GOTO(out, rc = -ENODEV);
1544 export->exp_in_recovery = 1;
1545 export->exp_req_replay_needed = 1;
1546 export->exp_lock_replay_needed = 1;
1547 spin_unlock(&export->exp_lock);
1549 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1550 MSG_CONNECT_TRANSNO);
1551 if (has_transno && transno == 0)
1552 CWARN("Connect with zero transno!\n");
1554 if (has_transno && transno > 0 &&
1555 transno < target->obd_next_recovery_transno &&
1556 transno > target->obd_last_committed) {
1557 /* Another way is to use cmpxchg() to be lock-free. */
1558 spin_lock(&target->obd_recovery_task_lock);
1559 if (transno < target->obd_next_recovery_transno)
1560 target->obd_next_recovery_transno = transno;
1561 spin_unlock(&target->obd_recovery_task_lock);
1564 atomic_inc(&target->obd_req_replay_clients);
1565 atomic_inc(&target->obd_lock_replay_clients);
1567 * Note: MDS-MDS connection is allowed to be connected during
1568 * recovery, no matter if the exports needs to be recoveried.
1569 * Because we need retrieve updates logs from all other MDTs.
1570 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1571 * also needs to be increased to match other recovery checking
1574 if (new_mds_mds_conn)
1575 atomic_inc(&target->obd_max_recoverable_clients);
1577 if (atomic_inc_return(&target->obd_connected_clients) ==
1578 atomic_read(&target->obd_max_recoverable_clients))
1579 wake_up(&target->obd_next_transno_waitq);
1582 /* Tell the client we're in recovery, when client is involved in it. */
1583 if (target->obd_recovering && !lw_client)
1584 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1588 spin_lock(&export->exp_lock);
1589 export->exp_connecting = 0;
1590 spin_unlock(&export->exp_lock);
1592 class_export_put(export);
1594 if (target != NULL) {
1595 spin_lock(&target->obd_dev_lock);
1596 target->obd_conn_inprogress--;
1597 spin_unlock(&target->obd_dev_lock);
1598 class_decref(target, "find", current);
1601 ptlrpc_connection_put(pcon);
1602 req->rq_status = rc;
1606 int target_handle_disconnect(struct ptlrpc_request *req)
1612 rc = req_capsule_server_pack(&req->rq_pill);
1616 /* In case of target disconnect, updating sec ctx immediately is
1617 * required in order to record latest sequence number used.
1618 * Sequence is normally updated on export destroy, but this event
1619 * can occur too late, ie after a new target connect request has
1621 * Maintaining correct sequence when client connection becomes idle
1622 * ensures that GSS does not erroneously consider requests as replays.
1624 rc = sptlrpc_export_update_ctx(req->rq_export);
1628 /* Keep the rq_export around so we can send the reply. */
1629 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1634 void target_destroy_export(struct obd_export *exp)
1636 struct obd_import *imp = NULL;
1638 * exports created from last_rcvd data, and "fake"
1639 * exports created by lctl don't have an import
1641 spin_lock(&exp->exp_lock);
1642 if (exp->exp_imp_reverse != NULL) {
1643 imp = exp->exp_imp_reverse;
1644 exp->exp_imp_reverse = NULL;
1646 spin_unlock(&exp->exp_lock);
1648 client_destroy_import(imp);
1650 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1651 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1652 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1653 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1655 EXPORT_SYMBOL(target_destroy_export);
1658 * Recovery functions
1660 static void target_request_copy_get(struct ptlrpc_request *req)
1662 class_export_rpc_inc(req->rq_export);
1663 LASSERT(list_empty(&req->rq_list));
1664 INIT_LIST_HEAD(&req->rq_replay_list);
1666 /* Increase refcount to keep request in queue. */
1667 atomic_inc(&req->rq_refcount);
1668 /* Let export know it has replays to be handled. */
1669 atomic_inc(&req->rq_export->exp_replay_count);
1672 static void target_request_copy_put(struct ptlrpc_request *req)
1674 LASSERT(list_empty(&req->rq_replay_list));
1675 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1677 atomic_dec(&req->rq_export->exp_replay_count);
1678 class_export_rpc_dec(req->rq_export);
1679 ptlrpc_server_drop_request(req);
1682 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1684 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1685 struct obd_export *exp = req->rq_export;
1686 struct ptlrpc_request *reqiter;
1687 struct ptlrpc_request *dup_req = NULL;
1692 spin_lock(&exp->exp_lock);
1693 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1695 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1703 /* We expect it with RESENT and REPLAY flags. */
1704 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1705 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1706 CERROR("invalid flags %x of resent replay\n",
1707 lustre_msg_get_flags(req->rq_reqmsg));
1709 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1712 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1714 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1715 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1719 list_add_tail(&req->rq_replay_list,
1720 &exp->exp_req_replay_queue);
1723 spin_unlock(&exp->exp_lock);
1727 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1729 LASSERT(!list_empty(&req->rq_replay_list));
1730 LASSERT(req->rq_export);
1732 spin_lock(&req->rq_export->exp_lock);
1733 list_del_init(&req->rq_replay_list);
1734 spin_unlock(&req->rq_export->exp_lock);
1737 static void target_finish_recovery(struct lu_target *lut)
1739 struct obd_device *obd = lut->lut_obd;
1743 /* Only log a recovery message when recovery has occurred. */
1744 if (obd->obd_recovery_start) {
1745 time64_t now = ktime_get_seconds();
1746 time64_t elapsed_time;
1748 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1750 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1751 obd->obd_name, elapsed_time / 60,
1753 atomic_read(&obd->obd_max_recoverable_clients),
1754 atomic_read(&obd->obd_connected_clients),
1755 obd->obd_stale_clients,
1756 obd->obd_stale_clients == 1 ? "was" : "were");
1757 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1758 libcfs_debug_dumplog();
1761 ldlm_reprocess_recovery_done(obd->obd_namespace);
1762 spin_lock(&obd->obd_recovery_task_lock);
1763 if (!list_empty(&obd->obd_req_replay_queue) ||
1764 !list_empty(&obd->obd_lock_replay_queue) ||
1765 !list_empty(&obd->obd_final_req_queue)) {
1766 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1768 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1769 list_empty(&obd->obd_lock_replay_queue) ? \
1771 list_empty(&obd->obd_final_req_queue) ? \
1773 spin_unlock(&obd->obd_recovery_task_lock);
1776 spin_unlock(&obd->obd_recovery_task_lock);
1778 obd->obd_recovery_end = ktime_get_seconds();
1780 /* When recovery finished, cleanup orphans on MDS and OST. */
1781 if (obd->obd_type && OBP(obd, postrecov)) {
1782 int rc = OBP(obd, postrecov)(obd);
1785 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1791 static void abort_req_replay_queue(struct obd_device *obd)
1793 struct ptlrpc_request *req, *n;
1794 LIST_HEAD(abort_list);
1796 spin_lock(&obd->obd_recovery_task_lock);
1797 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1798 spin_unlock(&obd->obd_recovery_task_lock);
1799 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1800 DEBUG_REQ(D_WARNING, req, "aborted:");
1801 req->rq_status = -ENOTCONN;
1802 if (ptlrpc_error(req)) {
1803 DEBUG_REQ(D_ERROR, req,
1804 "failed abort_req_reply; skipping");
1806 target_exp_dequeue_req_replay(req);
1807 target_request_copy_put(req);
1811 static void abort_lock_replay_queue(struct obd_device *obd)
1813 struct ptlrpc_request *req, *n;
1814 LIST_HEAD(abort_list);
1816 spin_lock(&obd->obd_recovery_task_lock);
1817 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1818 spin_unlock(&obd->obd_recovery_task_lock);
1819 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1820 DEBUG_REQ(D_ERROR, req, "aborted:");
1821 req->rq_status = -ENOTCONN;
1822 if (ptlrpc_error(req)) {
1823 DEBUG_REQ(D_ERROR, req,
1824 "failed abort_lock_reply; skipping");
1826 target_request_copy_put(req);
1831 * Called from a cleanup function if the device is being cleaned up
1832 * forcefully. The exports should all have been disconnected already,
1833 * the only thing left to do is
1834 * - clear the recovery flags
1835 * - cancel the timer
1836 * - free queued requests and replies, but don't send replies
1837 * Because the obd_stopping flag is set, no new requests should be received.
1839 void target_cleanup_recovery(struct obd_device *obd)
1841 struct ptlrpc_request *req, *n;
1842 LIST_HEAD(clean_list);
1844 spin_lock(&obd->obd_dev_lock);
1845 if (!obd->obd_recovering) {
1846 spin_unlock(&obd->obd_dev_lock);
1850 obd->obd_recovering = obd->obd_abort_recovery = 0;
1851 obd->obd_abort_recov_mdt = 0;
1852 spin_unlock(&obd->obd_dev_lock);
1854 spin_lock(&obd->obd_recovery_task_lock);
1855 target_cancel_recovery_timer(obd);
1856 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1857 spin_unlock(&obd->obd_recovery_task_lock);
1859 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1860 LASSERT(req->rq_reply_state == NULL);
1861 target_exp_dequeue_req_replay(req);
1862 target_request_copy_put(req);
1865 spin_lock(&obd->obd_recovery_task_lock);
1866 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1867 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1868 spin_unlock(&obd->obd_recovery_task_lock);
1870 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1871 LASSERT(req->rq_reply_state == NULL);
1872 target_request_copy_put(req);
1877 EXPORT_SYMBOL(target_cleanup_recovery);
1879 /* obd_recovery_task_lock should be held */
1880 void target_cancel_recovery_timer(struct obd_device *obd)
1882 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1883 hrtimer_cancel(&obd->obd_recovery_timer);
1886 static void target_start_recovery_timer(struct obd_device *obd)
1890 if (obd->obd_recovery_start != 0)
1893 spin_lock(&obd->obd_dev_lock);
1894 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1895 spin_unlock(&obd->obd_dev_lock);
1899 LASSERT(obd->obd_recovery_timeout != 0);
1901 if (obd->obd_recovery_start != 0) {
1902 spin_unlock(&obd->obd_dev_lock);
1906 obd->obd_recovery_start = ktime_get_seconds();
1907 delay = ktime_set(obd->obd_recovery_start +
1908 obd->obd_recovery_timeout, 0);
1909 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1910 spin_unlock(&obd->obd_dev_lock);
1912 LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1914 obd->obd_recovery_timeout / 60,
1915 obd->obd_recovery_timeout % 60,
1916 atomic_read(&obd->obd_max_recoverable_clients),
1917 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1919 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1924 * extend recovery window.
1926 * if @extend is true, extend recovery window to have @dr_timeout remaining
1927 * at least; otherwise, make sure the recovery timeout value is not less
1930 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1937 spin_lock(&obd->obd_dev_lock);
1938 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1939 obd->obd_stopping) {
1940 spin_unlock(&obd->obd_dev_lock);
1943 LASSERT(obd->obd_recovery_start != 0);
1945 left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1946 left = ktime_divns(left_ns, NSEC_PER_SEC);
1949 timeout = obd->obd_recovery_timeout;
1950 /* dr_timeout will happen after the hrtimer has expired.
1951 * Add the excess time to the soft recovery timeout without
1952 * exceeding the hard recovery timeout.
1954 if (dr_timeout > left) {
1955 timeout += dr_timeout - left;
1956 timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1960 timeout = clamp_t(timeout_t, dr_timeout,
1961 obd->obd_recovery_timeout,
1962 obd->obd_recovery_time_hard);
1965 if (timeout == obd->obd_recovery_time_hard)
1966 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1967 obd->obd_name, timeout, extend);
1969 if (obd->obd_recovery_timeout < timeout) {
1972 obd->obd_recovery_timeout = timeout;
1973 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1974 now = ktime_set(ktime_get_seconds(), 0);
1975 left_ns = ktime_sub(end, now);
1976 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1977 left = ktime_divns(left_ns, NSEC_PER_SEC);
1979 spin_unlock(&obd->obd_dev_lock);
1981 CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1982 obd->obd_name, left);
1985 /* Reset the timer with each new client connection */
1987 * This timer is actually reconnect_timer, which is for making sure
1988 * the total recovery window is at least as big as my reconnect
1989 * attempt timing. So the initial recovery time_out will be set to
1990 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1991 * from client is bigger than this, then the recovery time_out will
1992 * be extended to make sure the client could be reconnected, in the
1993 * process, the timeout from the new client should be ignored.
1996 check_and_start_recovery_timer(struct obd_device *obd,
1997 struct ptlrpc_request *req,
2000 timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
2001 struct obd_device_target *obt = &obd->u.obt;
2003 if (!new_client && service_timeout)
2005 * Teach server about old server's estimates, as first guess
2006 * at how long new requests will take.
2008 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2011 target_start_recovery_timer(obd);
2014 * Convert the service time to RPC timeout,
2015 * and reuse service_timeout to limit stack usage.
2017 service_timeout = at_est2timeout(service_timeout);
2019 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2020 service_timeout < at_extra)
2021 service_timeout = at_extra;
2024 * We expect other clients to timeout within service_timeout, then try
2025 * to reconnect, then try the failover server. The max delay between
2026 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2028 service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2030 LASSERT(obt->obt_magic == OBT_MAGIC);
2031 service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2032 if (service_timeout > obd->obd_recovery_timeout && !new_client)
2033 extend_recovery_timer(obd, service_timeout, false);
2036 /** Health checking routines */
2037 static inline int exp_connect_healthy(struct obd_export *exp)
2039 return exp->exp_in_recovery;
2042 /** if export done req_replay or has replay in queue */
2043 static inline int exp_req_replay_healthy(struct obd_export *exp)
2045 return (!exp->exp_req_replay_needed ||
2046 atomic_read(&exp->exp_replay_count) > 0);
2050 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2052 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2053 exp_req_replay_healthy(exp);
2056 /** if export done lock_replay or has replay in queue */
2057 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2059 return (!exp->exp_lock_replay_needed ||
2060 atomic_read(&exp->exp_replay_count) > 0);
2063 static inline int exp_vbr_healthy(struct obd_export *exp)
2065 return !exp->exp_vbr_failed;
2068 static inline int exp_finished(struct obd_export *exp)
2070 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2073 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2075 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2079 static int check_for_next_transno(struct lu_target *lut)
2081 struct ptlrpc_request *req = NULL;
2082 struct obd_device *obd = lut->lut_obd;
2083 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2084 int wake_up = 0, connected, completed, queue_len;
2085 __u64 req_transno = 0;
2086 __u64 update_transno = 0;
2087 __u64 next_transno = 0;
2091 spin_lock(&obd->obd_recovery_task_lock);
2092 if (!list_empty(&obd->obd_req_replay_queue)) {
2093 req = list_entry(obd->obd_req_replay_queue.next,
2094 struct ptlrpc_request, rq_list);
2095 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2098 if (!obd->obd_abort_recov_mdt && tdtd)
2099 update_transno = distribute_txn_get_next_transno(tdtd);
2101 connected = atomic_read(&obd->obd_connected_clients);
2102 completed = connected - atomic_read(&obd->obd_req_replay_clients);
2103 queue_len = obd->obd_requests_queued_for_recovery;
2104 next_transno = obd->obd_next_recovery_transno;
2107 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2108 atomic_read(&obd->obd_max_recoverable_clients),
2109 connected, completed,
2110 queue_len, req_transno, next_transno);
2112 if (obd->obd_abort_recovery) {
2113 CDEBUG(D_HA, "waking for aborted recovery\n");
2115 } else if (obd->obd_recovery_expired) {
2116 CDEBUG(D_HA, "waking for expired recovery\n");
2118 } else if (!obd->obd_abort_recov_mdt && tdtd && req &&
2119 is_req_replayed_by_update(req)) {
2120 LASSERTF(req_transno < next_transno,
2121 "req_transno %llu next_transno%llu\n", req_transno,
2123 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2126 } else if (req_transno == next_transno ||
2127 (update_transno != 0 && update_transno <= next_transno)) {
2128 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2130 } else if (queue_len > 0 &&
2131 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2132 /** handle gaps occured due to lost reply or VBR */
2133 LASSERTF(req_transno >= next_transno,
2134 "req_transno: %llu, next_transno: %llu\n",
2135 req_transno, next_transno);
2137 "%s: waking for gap in transno, VBR is %s (skip: %lld, ql: %d, comp: %d, conn: %d, next: %lld, next_update %lld last_committed: %lld)\n",
2138 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2139 next_transno, queue_len, completed, connected,
2140 req_transno, update_transno, obd->obd_last_committed);
2141 obd->obd_next_recovery_transno = req_transno;
2143 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2144 CDEBUG(D_HA, "waking for completed recovery\n");
2146 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2148 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2150 obd->obd_next_recovery_transno = req_transno;
2153 spin_unlock(&obd->obd_recovery_task_lock);
2157 static int check_for_next_lock(struct lu_target *lut)
2159 struct obd_device *obd = lut->lut_obd;
2162 spin_lock(&obd->obd_recovery_task_lock);
2163 if (!list_empty(&obd->obd_lock_replay_queue)) {
2164 CDEBUG(D_HA, "waking for next lock\n");
2166 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2167 CDEBUG(D_HA, "waking for completed lock replay\n");
2169 } else if (obd->obd_abort_recovery) {
2170 CDEBUG(D_HA, "waking for aborted recovery\n");
2172 } else if (obd->obd_recovery_expired) {
2173 CDEBUG(D_HA, "waking for expired recovery\n");
2176 spin_unlock(&obd->obd_recovery_task_lock);
2181 static int check_update_llog(struct lu_target *lut)
2183 struct obd_device *obd = lut->lut_obd;
2184 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2186 if (obd->obd_abort_recovery) {
2187 CDEBUG(D_HA, "waking for aborted recovery\n");
2191 if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2192 CDEBUG(D_HA, "waking for completion of reading update log\n");
2200 * wait for recovery events,
2201 * check its status with help of check_routine
2202 * evict dead clients via health_check
2204 static int target_recovery_overseer(struct lu_target *lut,
2205 int (*check_routine)(struct lu_target *),
2206 int (*health_check)(struct obd_export *))
2208 struct obd_device *obd = lut->lut_obd;
2209 struct target_distribute_txn_data *tdtd;
2213 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2214 now = ktime_get_seconds();
2215 if (now - last > 600) {
2216 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2221 if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2222 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2223 __u64 next_update_transno = 0;
2226 * Only abort the recovery if there are no update recovery
2229 spin_lock(&obd->obd_recovery_task_lock);
2230 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd) {
2231 next_update_transno =
2232 distribute_txn_get_next_transno(lut->lut_tdtd);
2234 tdtd = lut->lut_tdtd;
2236 * If next_update_transno == 0, it probably because
2237 * updatelog retrieve threads did not get any records
2238 * yet, let's wait those threads stopped
2240 if (next_update_transno == 0) {
2241 spin_unlock(&obd->obd_recovery_task_lock);
2243 while (wait_event_timeout(
2244 tdtd->tdtd_recovery_threads_waitq,
2245 check_update_llog(lut),
2246 cfs_time_seconds(60)) == 0);
2248 spin_lock(&obd->obd_recovery_task_lock);
2249 next_update_transno =
2250 distribute_txn_get_next_transno(tdtd);
2254 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2255 obd->obd_next_recovery_transno = next_update_transno;
2256 spin_unlock(&obd->obd_recovery_task_lock);
2258 * Disconnect unfinished exports from clients, and
2259 * keep connection from MDT to make sure the update
2260 * recovery will still keep trying until some one
2261 * manually abort the recovery
2263 class_disconnect_stale_exports(obd,
2264 exp_finished_or_from_mdt);
2265 /* Abort all of replay & replay lock req from clients */
2266 abort_req_replay_queue(obd);
2267 abort_lock_replay_queue(obd);
2269 "%s: there are still update replay (%#llx)in the queue.\n",
2270 obd->obd_name, next_update_transno);
2272 obd->obd_abort_recovery = 1;
2273 spin_unlock(&obd->obd_recovery_task_lock);
2274 CWARN("%s recovery is aborted by hard timeout\n",
2279 while (wait_event_timeout(obd->obd_next_transno_waitq,
2281 cfs_time_seconds(60)) == 0)
2282 ; /* wait indefinitely for event, but don't trigger watchdog */
2284 if (obd->obd_abort_recovery) {
2285 CWARN("recovery is aborted, evict exports in recovery\n");
2286 if (lut->lut_tdtd != NULL) {
2287 tdtd = lut->lut_tdtd;
2289 * Let's wait all of the update log recovery thread
2293 tdtd->tdtd_recovery_threads_waitq,
2294 atomic_read(&tdtd->tdtd_recovery_threads_count)
2296 /* Then abort the update recovery list */
2297 dtrq_list_destroy(lut->lut_tdtd);
2300 /** evict exports which didn't finish recovery yet */
2301 class_disconnect_stale_exports(obd, exp_finished);
2303 } else if (obd->obd_recovery_expired) {
2304 obd->obd_recovery_expired = 0;
2306 /** If some clients died being recovered, evict them */
2307 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2309 /** evict cexports with no replay in queue, they are stalled */
2310 class_disconnect_stale_exports(obd, health_check);
2312 /** continue with VBR */
2313 spin_lock(&obd->obd_dev_lock);
2314 obd->obd_version_recov = 1;
2315 spin_unlock(&obd->obd_dev_lock);
2317 * reset timer, recovery will proceed with versions now,
2318 * timeout is set just to handle reconnection delays
2320 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2322 * Wait for recovery events again, after evicting bad clients
2329 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2331 struct obd_device *obd = lut->lut_obd;
2332 struct ptlrpc_request *req = NULL;
2334 CDEBUG(D_HA, "Waiting for lock\n");
2335 if (target_recovery_overseer(lut, check_for_next_lock,
2336 exp_lock_replay_healthy))
2337 abort_lock_replay_queue(obd);
2339 spin_lock(&obd->obd_recovery_task_lock);
2340 if (!list_empty(&obd->obd_lock_replay_queue)) {
2341 req = list_entry(obd->obd_lock_replay_queue.next,
2342 struct ptlrpc_request, rq_list);
2343 list_del_init(&req->rq_list);
2344 spin_unlock(&obd->obd_recovery_task_lock);
2346 spin_unlock(&obd->obd_recovery_task_lock);
2347 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2348 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2349 /** evict exports failed VBR */
2350 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2355 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2357 struct ptlrpc_request *req = NULL;
2359 spin_lock(&obd->obd_recovery_task_lock);
2360 if (!list_empty(&obd->obd_final_req_queue)) {
2361 req = list_entry(obd->obd_final_req_queue.next,
2362 struct ptlrpc_request, rq_list);
2363 list_del_init(&req->rq_list);
2364 spin_unlock(&obd->obd_recovery_task_lock);
2365 if (req->rq_export->exp_in_recovery) {
2366 spin_lock(&req->rq_export->exp_lock);
2367 req->rq_export->exp_in_recovery = 0;
2368 spin_unlock(&req->rq_export->exp_lock);
2371 spin_unlock(&obd->obd_recovery_task_lock);
2376 static void handle_recovery_req(struct ptlrpc_thread *thread,
2377 struct ptlrpc_request *req,
2378 svc_handler_t handler)
2383 * export can be evicted during recovery, no need to handle replays for
2384 * it after that, discard such request silently
2386 if (req->rq_export->exp_disconnected)
2389 req->rq_session.lc_thread = thread;
2390 req->rq_svc_thread = thread;
2391 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2393 /* thread context */
2394 lu_context_enter(&thread->t_env->le_ctx);
2396 lu_context_exit(&thread->t_env->le_ctx);
2398 req->rq_svc_thread->t_env->le_ses = NULL;
2400 /* don't reset timer for final stage */
2401 if (!exp_finished(req->rq_export)) {
2402 timeout_t timeout = obd_timeout;
2405 * Add request @timeout to the recovery time so next request from
2406 * this client may come in recovery time
2409 struct ptlrpc_service_part *svcpt;
2410 timeout_t est_timeout;
2412 svcpt = req->rq_rqbd->rqbd_svcpt;
2414 * If the server sent early reply for this request,
2415 * the client will recalculate the timeout according to
2416 * current server estimate service time, so we will
2417 * use the maxium timeout here for waiting the client
2418 * sending the next req
2420 est_timeout = at_get(&svcpt->scp_at_estimate);
2421 timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2422 lustre_msg_get_timeout(req->rq_reqmsg));
2424 * Add 2 net_latency, one for balance rq_deadline
2425 * (see ptl_send_rpc), one for resend the req to server,
2426 * Note: client will pack net_latency in replay req
2427 * (see ptlrpc_replay_req)
2429 timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2431 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2437 /** Checking routines for recovery */
2438 static int check_for_recovery_ready(struct lu_target *lut)
2440 struct obd_device *obd = lut->lut_obd;
2441 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2444 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2445 clnts, obd->obd_stale_clients,
2446 atomic_read(&obd->obd_max_recoverable_clients),
2447 obd->obd_abort_recovery, obd->obd_recovery_expired);
2449 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2451 atomic_read(&obd->obd_max_recoverable_clients));
2452 if (clnts + obd->obd_stale_clients <
2453 atomic_read(&obd->obd_max_recoverable_clients))
2457 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd != NULL) {
2458 if (!lut->lut_tdtd->tdtd_replay_ready &&
2459 !obd->obd_abort_recovery && !obd->obd_stopping) {
2461 * Let's extend recovery timer, in case the recovery
2462 * timer expired, and some clients got evicted
2464 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2467 "%s update recovery is not ready, extend recovery %d\n",
2468 obd->obd_name, obd->obd_recovery_timeout);
2477 REQUEST_RECOVERY = 1,
2478 UPDATE_RECOVERY = 2,
2481 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2485 if (!list_empty(&obd->obd_req_replay_queue)) {
2486 struct ptlrpc_request *req;
2488 req = list_entry(obd->obd_req_replay_queue.next,
2489 struct ptlrpc_request, rq_list);
2490 transno = lustre_msg_get_transno(req->rq_reqmsg);
2496 static __u64 get_next_transno(struct lu_target *lut, int *type)
2498 struct obd_device *obd = lut->lut_obd;
2499 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2501 __u64 update_transno;
2505 transno = get_next_replay_req_transno(obd);
2507 *type = REQUEST_RECOVERY;
2509 if (!tdtd || obd->obd_abort_recov_mdt)
2512 update_transno = distribute_txn_get_next_transno(tdtd);
2513 if (transno == 0 || (transno >= update_transno &&
2514 update_transno != 0)) {
2515 transno = update_transno;
2517 *type = UPDATE_RECOVERY;
2524 * drop duplicate replay request
2526 * Because the operation has been replayed by update recovery, the request
2527 * with the same transno will be dropped and also notify the client to send
2528 * next replay request.
2530 * \param[in] env execution environment
2531 * \param[in] obd failover obd device
2532 * \param[in] req request to be dropped
2534 static void drop_duplicate_replay_req(struct lu_env *env,
2535 struct obd_device *obd,
2536 struct ptlrpc_request *req)
2538 DEBUG_REQ(D_HA, req,
2539 "remove t%lld from %s because duplicate update records found",
2540 lustre_msg_get_transno(req->rq_reqmsg),
2541 libcfs_nid2str(req->rq_peer.nid));
2544 * Right now, only for MDS reint operation update replay and
2545 * normal request replay can have the same transno
2547 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2548 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2549 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2550 if (likely(req->rq_export))
2551 target_committed_to_req(req);
2552 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2553 target_send_reply(req, req->rq_status, 0);
2555 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2556 libcfs_nid2str(req->rq_peer.nid));
2558 target_exp_dequeue_req_replay(req);
2559 target_request_copy_put(req);
2560 obd->obd_replayed_requests++;
2563 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2565 static void replay_request_or_update(struct lu_env *env,
2566 struct lu_target *lut,
2567 struct target_recovery_data *trd,
2568 struct ptlrpc_thread *thread)
2570 struct obd_device *obd = lut->lut_obd;
2571 struct ptlrpc_request *req = NULL;
2577 CDEBUG(D_HA, "Waiting for transno %lld\n",
2578 obd->obd_next_recovery_transno);
2580 /* Replay all of request and update by transno */
2582 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2584 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2587 * It is needed to extend recovery window above
2588 * recovery_time_soft. Extending is possible only in the
2589 * end of recovery window (see more details in
2590 * handle_recovery_req()).
2592 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2594 if (target_recovery_overseer(lut, check_for_next_transno,
2595 exp_req_replay_healthy_or_from_mdt)) {
2596 abort_req_replay_queue(obd);
2597 abort_lock_replay_queue(obd);
2601 spin_lock(&obd->obd_recovery_task_lock);
2602 transno = get_next_transno(lut, &type);
2603 if (type == REQUEST_RECOVERY && transno != 0) {
2605 * Drop replay request from client side, if the
2606 * replay has been executed by update with the
2609 req = list_entry(obd->obd_req_replay_queue.next,
2610 struct ptlrpc_request, rq_list);
2612 list_del_init(&req->rq_list);
2613 obd->obd_requests_queued_for_recovery--;
2614 spin_unlock(&obd->obd_recovery_task_lock);
2617 * Let's check if the request has been redone by
2620 if (is_req_replayed_by_update(req)) {
2621 struct distribute_txn_replay_req *dtrq;
2623 dtrq = distribute_txn_lookup_finish_list(tdtd,
2625 LASSERT(dtrq != NULL);
2626 spin_lock(&tdtd->tdtd_replay_list_lock);
2627 list_del_init(&dtrq->dtrq_list);
2628 spin_unlock(&tdtd->tdtd_replay_list_lock);
2631 drop_duplicate_replay_req(env, obd, req);
2636 LASSERT(trd->trd_processing_task == current->pid);
2637 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2639 lustre_msg_get_transno(req->rq_reqmsg),
2640 libcfs_nid2str(req->rq_peer.nid));
2642 ptlrpc_watchdog_init(&thread->t_watchdog,
2644 handle_recovery_req(thread, req,
2645 trd->trd_recovery_handler);
2646 ptlrpc_watchdog_disable(&thread->t_watchdog);
2649 * bz18031: increase next_recovery_transno before
2650 * target_request_copy_put() will drop exp_rpc reference
2652 spin_lock(&obd->obd_recovery_task_lock);
2653 obd->obd_next_recovery_transno++;
2654 spin_unlock(&obd->obd_recovery_task_lock);
2655 target_exp_dequeue_req_replay(req);
2656 target_request_copy_put(req);
2657 obd->obd_replayed_requests++;
2658 } else if (type == UPDATE_RECOVERY && transno != 0) {
2659 struct distribute_txn_replay_req *dtrq;
2662 spin_unlock(&obd->obd_recovery_task_lock);
2664 LASSERT(tdtd != NULL);
2665 dtrq = distribute_txn_get_next_req(tdtd);
2666 lu_context_enter(&thread->t_env->le_ctx);
2667 ptlrpc_watchdog_init(&thread->t_watchdog,
2669 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2670 ptlrpc_watchdog_disable(&thread->t_watchdog);
2671 lu_context_exit(&thread->t_env->le_ctx);
2672 extend_recovery_timer(obd, obd_timeout, true);
2674 if (rc == 0 && dtrq->dtrq_xid != 0) {
2676 "Move x%llu t%llu to finish list\n",
2678 dtrq->dtrq_master_transno);
2680 /* Add it to the replay finish list */
2681 spin_lock(&tdtd->tdtd_replay_list_lock);
2682 list_add(&dtrq->dtrq_list,
2683 &tdtd->tdtd_replay_finish_list);
2684 spin_unlock(&tdtd->tdtd_replay_list_lock);
2686 spin_lock(&obd->obd_recovery_task_lock);
2687 if (transno == obd->obd_next_recovery_transno)
2688 obd->obd_next_recovery_transno++;
2690 obd->obd_next_recovery_transno)
2691 obd->obd_next_recovery_transno =
2693 spin_unlock(&obd->obd_recovery_task_lock);
2698 spin_unlock(&obd->obd_recovery_task_lock);
2700 LASSERT(list_empty(&obd->obd_req_replay_queue));
2701 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2702 /** evict exports failed VBR */
2703 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2709 static int target_recovery_thread(void *arg)
2711 struct lu_target *lut = arg;
2712 struct obd_device *obd = lut->lut_obd;
2713 struct ptlrpc_request *req;
2714 struct target_recovery_data *trd = &obd->obd_recovery_data;
2715 unsigned long delta;
2717 struct ptlrpc_thread *thread = NULL;
2721 unshare_fs_struct();
2722 OBD_ALLOC_PTR(thread);
2728 GOTO(out_thread, rc = -ENOMEM);
2729 rc = lu_env_add(env);
2733 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2735 GOTO(out_env_remove, rc);
2737 thread->t_env = env;
2738 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2739 thread->t_task = current;
2740 env->le_ctx.lc_thread = thread;
2741 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2743 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2745 trd->trd_processing_task = current->pid;
2747 spin_lock(&obd->obd_dev_lock);
2748 obd->obd_recovering = 1;
2749 spin_unlock(&obd->obd_dev_lock);
2750 complete(&trd->trd_starting);
2752 /* first of all, we have to know the first transno to replay */
2753 if (target_recovery_overseer(lut, check_for_recovery_ready,
2754 exp_connect_healthy)) {
2755 abort_req_replay_queue(obd);
2756 abort_lock_replay_queue(obd);
2757 if (lut->lut_tdtd != NULL)
2758 dtrq_list_destroy(lut->lut_tdtd);
2761 /* next stage: replay requests or update */
2763 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2764 atomic_read(&obd->obd_req_replay_clients),
2765 obd->obd_next_recovery_transno);
2766 replay_request_or_update(env, lut, trd, thread);
2769 * The second stage: replay locks
2771 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2772 atomic_read(&obd->obd_lock_replay_clients));
2773 while ((req = target_next_replay_lock(lut))) {
2774 LASSERT(trd->trd_processing_task == current->pid);
2775 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2776 libcfs_nid2str(req->rq_peer.nid));
2777 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2778 req->rq_status = -ENODEV;
2779 target_request_copy_put(req);
2782 handle_recovery_req(thread, req,
2783 trd->trd_recovery_handler);
2784 target_request_copy_put(req);
2785 obd->obd_replayed_locks++;
2789 * The third stage: reply on final pings, at this moment all clients
2790 * must have request in final queue
2792 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2793 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2794 /** Update server last boot epoch */
2795 tgt_boot_epoch_update(lut);
2797 * We drop recoverying flag to forward all new requests
2798 * to regular mds_handle() since now
2800 spin_lock(&obd->obd_dev_lock);
2801 obd->obd_recovering = obd->obd_abort_recovery = 0;
2802 obd->obd_abort_recov_mdt = 0;
2803 spin_unlock(&obd->obd_dev_lock);
2804 spin_lock(&obd->obd_recovery_task_lock);
2805 target_cancel_recovery_timer(obd);
2806 spin_unlock(&obd->obd_recovery_task_lock);
2807 while ((req = target_next_final_ping(obd))) {
2808 LASSERT(trd->trd_processing_task == current->pid);
2809 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2810 libcfs_nid2str(req->rq_peer.nid));
2811 handle_recovery_req(thread, req,
2812 trd->trd_recovery_handler);
2814 * Because the waiting client can not send ping to server,
2815 * so we need refresh the last_request_time, to avoid the
2816 * export is being evicted
2818 ptlrpc_update_export_timer(req->rq_export, 0);
2819 target_request_copy_put(req);
2822 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2823 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2824 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2825 if (delta > OBD_RECOVERY_TIME_SOFT) {
2826 CWARN("too long recovery - read logs\n");
2827 libcfs_debug_dumplog();
2830 target_finish_recovery(lut);
2831 lu_context_fini(&env->le_ctx);
2832 trd->trd_processing_task = 0;
2833 complete_all(&trd->trd_finishing);
2834 tgt_io_thread_done(thread);
2840 OBD_FREE_PTR(thread);
2844 static int target_start_recovery_thread(struct lu_target *lut,
2845 svc_handler_t handler)
2847 struct obd_device *obd = lut->lut_obd;
2849 struct target_recovery_data *trd = &obd->obd_recovery_data;
2852 memset(trd, 0, sizeof(*trd));
2853 init_completion(&trd->trd_starting);
2854 init_completion(&trd->trd_finishing);
2855 trd->trd_recovery_handler = handler;
2857 rc = server_name2index(obd->obd_name, &index, NULL);
2861 if (!IS_ERR(kthread_run(target_recovery_thread,
2862 lut, "tgt_recover_%d", index))) {
2863 wait_for_completion(&trd->trd_starting);
2864 LASSERT(obd->obd_recovering != 0);
2872 void target_stop_recovery_thread(struct obd_device *obd)
2874 if (obd->obd_recovery_data.trd_processing_task > 0) {
2875 struct target_recovery_data *trd = &obd->obd_recovery_data;
2876 /** recovery can be done but postrecovery is not yet */
2877 spin_lock(&obd->obd_dev_lock);
2878 if (obd->obd_recovering) {
2879 CERROR("%s: Aborting recovery\n", obd->obd_name);
2880 obd->obd_abort_recovery = 1;
2881 wake_up(&obd->obd_next_transno_waitq);
2883 spin_unlock(&obd->obd_dev_lock);
2884 wait_for_completion(&trd->trd_finishing);
2887 EXPORT_SYMBOL(target_stop_recovery_thread);
2889 void target_recovery_fini(struct obd_device *obd)
2891 class_disconnect_exports(obd);
2892 target_stop_recovery_thread(obd);
2893 target_cleanup_recovery(obd);
2895 EXPORT_SYMBOL(target_recovery_fini);
2897 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2899 struct obd_device *obd = container_of(timer, struct obd_device,
2900 obd_recovery_timer);
2903 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2904 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2905 ktime_get_seconds() - obd->obd_recovery_start,
2906 atomic_read(&obd->obd_connected_clients));
2908 obd->obd_recovery_expired = 1;
2909 wake_up(&obd->obd_next_transno_waitq);
2910 return HRTIMER_NORESTART;
2913 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2915 struct obd_device *obd = lut->lut_obd;
2917 if (lut->lut_bottom->dd_rdonly)
2920 if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2921 /** Update server last boot epoch */
2922 tgt_boot_epoch_update(lut);
2926 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2927 "last_transno %llu\n", obd->obd_name,
2928 atomic_read(&obd->obd_max_recoverable_clients),
2929 obd->obd_last_committed);
2930 LASSERT(obd->obd_stopping == 0);
2931 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2932 obd->obd_recovery_start = 0;
2933 obd->obd_recovery_end = 0;
2935 hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2937 obd->obd_recovery_timer.function = &target_recovery_expired;
2938 target_start_recovery_thread(lut, handler);
2940 EXPORT_SYMBOL(target_recovery_init);
2942 static int target_process_req_flags(struct obd_device *obd,
2943 struct ptlrpc_request *req)
2945 struct obd_export *exp = req->rq_export;
2947 LASSERT(exp != NULL);
2948 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2949 /* client declares he's ready to replay locks */
2950 spin_lock(&exp->exp_lock);
2951 if (exp->exp_req_replay_needed) {
2952 exp->exp_req_replay_needed = 0;
2953 spin_unlock(&exp->exp_lock);
2955 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2956 atomic_dec(&obd->obd_req_replay_clients);
2958 spin_unlock(&exp->exp_lock);
2961 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2963 * client declares he's ready to complete recovery
2964 * so, we put the request on th final queue
2966 spin_lock(&exp->exp_lock);
2967 if (exp->exp_lock_replay_needed) {
2968 exp->exp_lock_replay_needed = 0;
2969 spin_unlock(&exp->exp_lock);
2971 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2972 atomic_dec(&obd->obd_lock_replay_clients);
2974 spin_unlock(&exp->exp_lock);
2980 int target_queue_recovery_request(struct ptlrpc_request *req,
2981 struct obd_device *obd)
2983 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2984 struct ptlrpc_request *reqiter;
2989 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2990 /* Processing the queue right now, don't re-add. */
2994 target_process_req_flags(obd, req);
2996 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2997 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2998 if (cfs_fail_val == 1) {
3001 wake_up(&cfs_race_waitq);
3003 schedule_timeout_interruptible(
3004 cfs_time_seconds(1));
3009 * client declares he's ready to complete recovery
3010 * so, we put the request on th final queue
3012 target_request_copy_get(req);
3013 DEBUG_REQ(D_HA, req, "queue final req");
3014 wake_up(&obd->obd_next_transno_waitq);
3015 spin_lock(&obd->obd_recovery_task_lock);
3016 if (obd->obd_recovering) {
3017 struct ptlrpc_request *tmp;
3018 struct ptlrpc_request *duplicate = NULL;
3020 if (likely(!req->rq_export->exp_replay_done)) {
3021 req->rq_export->exp_replay_done = 1;
3022 list_add_tail(&req->rq_list,
3023 &obd->obd_final_req_queue);
3024 spin_unlock(&obd->obd_recovery_task_lock);
3029 * XXX O(n), but only happens if final ping is
3030 * timed out, probably reorganize the list as
3033 list_for_each_entry_safe(reqiter, tmp,
3034 &obd->obd_final_req_queue,
3036 if (reqiter->rq_export == req->rq_export) {
3037 list_del_init(&reqiter->rq_list);
3038 duplicate = reqiter;
3043 list_add_tail(&req->rq_list,
3044 &obd->obd_final_req_queue);
3045 req->rq_export->exp_replay_done = 1;
3046 spin_unlock(&obd->obd_recovery_task_lock);
3048 if (duplicate != NULL) {
3049 DEBUG_REQ(D_HA, duplicate,
3050 "put prev final req");
3051 target_request_copy_put(duplicate);
3055 spin_unlock(&obd->obd_recovery_task_lock);
3056 target_request_copy_put(req);
3057 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3060 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3061 /* client declares he's ready to replay locks */
3062 target_request_copy_get(req);
3063 DEBUG_REQ(D_HA, req, "queue lock replay req");
3064 wake_up(&obd->obd_next_transno_waitq);
3065 spin_lock(&obd->obd_recovery_task_lock);
3066 LASSERT(obd->obd_recovering);
3067 /* usually due to recovery abort */
3068 if (!req->rq_export->exp_in_recovery) {
3069 spin_unlock(&obd->obd_recovery_task_lock);
3070 target_request_copy_put(req);
3073 LASSERT(req->rq_export->exp_lock_replay_needed);
3074 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3075 spin_unlock(&obd->obd_recovery_task_lock);
3080 * CAVEAT EMPTOR: The incoming request message has been swabbed
3081 * (i.e. buflens etc are in my own byte order), but type-dependent
3082 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3086 INIT_LIST_HEAD(&req->rq_list);
3087 DEBUG_REQ(D_HA, req, "not queueing");
3092 * If we're processing the queue, we want don't want to queue this
3095 * Also, if this request has a transno less than the one we're waiting
3096 * for, we should process it now. It could (and currently always will)
3097 * be an open request for a descriptor that was opened some time ago.
3099 * Also, a resent, replayed request that has already been
3100 * handled will pass through here and be processed immediately.
3103 "Next recovery transno: %llu, current: %llu, replaying\n",
3104 obd->obd_next_recovery_transno, transno);
3107 * If the request has been replayed by update replay, then sends this
3108 * request to the recovery thread (replay_request_or_update()), where
3109 * it will be handled
3111 spin_lock(&obd->obd_recovery_task_lock);
3112 if (transno < obd->obd_next_recovery_transno &&
3113 !is_req_replayed_by_update(req)) {
3114 /* Processing the queue right now, don't re-add. */
3115 LASSERT(list_empty(&req->rq_list));
3116 spin_unlock(&obd->obd_recovery_task_lock);
3119 spin_unlock(&obd->obd_recovery_task_lock);
3121 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3124 target_request_copy_get(req);
3125 if (!req->rq_export->exp_in_recovery) {
3126 target_request_copy_put(req);
3129 LASSERT(req->rq_export->exp_req_replay_needed);
3131 if (target_exp_enqueue_req_replay(req)) {
3132 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3133 target_request_copy_put(req);
3138 spin_lock(&obd->obd_recovery_task_lock);
3139 LASSERT(obd->obd_recovering);
3140 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3141 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3142 list_add_tail(&req->rq_list, &reqiter->rq_list);
3147 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3149 DEBUG_REQ(D_ERROR, req,
3150 "dropping replay: transno has been claimed by another client");
3151 spin_unlock(&obd->obd_recovery_task_lock);
3152 target_exp_dequeue_req_replay(req);
3153 target_request_copy_put(req);
3159 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3161 obd->obd_requests_queued_for_recovery++;
3162 spin_unlock(&obd->obd_recovery_task_lock);
3163 wake_up(&obd->obd_next_transno_waitq);
3167 void target_committed_to_req(struct ptlrpc_request *req)
3169 struct obd_export *exp = req->rq_export;
3171 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3172 lustre_msg_set_last_committed(req->rq_repmsg,
3173 exp->exp_last_committed);
3175 DEBUG_REQ(D_IOCTL, req,
3176 "not sending last_committed update (%d/%d)",
3177 exp->exp_obd->obd_no_transno,
3178 req->rq_repmsg == NULL);
3180 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3181 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3184 #endif /* HAVE_SERVER_SUPPORT */
3187 * Packs current SLV and Limit into \a req.
3189 int target_pack_pool_reply(struct ptlrpc_request *req)
3191 struct obd_device *obd;
3196 * Check that we still have all structures alive as this may
3197 * be some late RPC at shutdown time.
3199 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3200 !exp_connect_lru_resize(req->rq_export))) {
3201 lustre_msg_set_slv(req->rq_repmsg, 0);
3202 lustre_msg_set_limit(req->rq_repmsg, 0);
3206 /* OBD is alive here as export is alive, which we checked above. */
3207 obd = req->rq_export->exp_obd;
3209 read_lock(&obd->obd_pool_lock);
3210 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3211 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3212 read_unlock(&obd->obd_pool_lock);
3217 static int target_send_reply_msg(struct ptlrpc_request *req,
3218 int rc, int fail_id)
3220 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3221 DEBUG_REQ(D_ERROR, req, "dropping reply");
3225 * We can have a null rq_reqmsg in the event of bad signature or
3226 * no context when unwrapping
3228 if (req->rq_reqmsg &&
3229 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3230 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3234 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3235 req->rq_status = rc;
3236 return ptlrpc_send_error(req, 1);
3238 DEBUG_REQ(D_NET, req, "sending reply");
3240 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3243 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3245 struct ptlrpc_service_part *svcpt;
3247 struct ptlrpc_reply_state *rs;
3248 struct obd_export *exp;
3252 if (req->rq_no_reply) {
3257 svcpt = req->rq_rqbd->rqbd_svcpt;
3258 rs = req->rq_reply_state;
3259 if (rs == NULL || !rs->rs_difficult) {
3261 target_send_reply_msg(req, rc, fail_id);
3266 /* must be an export if locks saved */
3267 LASSERT(req->rq_export != NULL);
3268 /* req/reply consistent */
3269 LASSERT(rs->rs_svcpt == svcpt);
3272 LASSERT(!rs->rs_scheduled);
3273 LASSERT(!rs->rs_scheduled_ever);
3274 LASSERT(!rs->rs_handled);
3275 LASSERT(!rs->rs_sent);
3276 LASSERT(!rs->rs_unlinked);
3277 LASSERT(rs->rs_export == NULL);
3278 LASSERT(list_empty(&rs->rs_obd_list));
3279 LASSERT(list_empty(&rs->rs_exp_list));
3281 exp = class_export_get(req->rq_export);
3283 /* disable reply scheduling while I'm setting up */
3284 rs->rs_scheduled = 1;
3286 rs->rs_unlinked = 0;
3287 rs->rs_xid = req->rq_xid;
3288 rs->rs_transno = req->rq_transno;
3289 rs->rs_export = exp;
3290 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3292 spin_lock(&exp->exp_uncommitted_replies_lock);
3293 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3294 rs->rs_transno, exp->exp_last_committed);
3295 if (rs->rs_transno > exp->exp_last_committed) {
3296 /* not committed already */
3297 list_add_tail(&rs->rs_obd_list,
3298 &exp->exp_uncommitted_replies);
3300 spin_unlock(&exp->exp_uncommitted_replies_lock);
3302 spin_lock(&exp->exp_lock);
3303 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3304 spin_unlock(&exp->exp_lock);
3306 netrc = target_send_reply_msg(req, rc, fail_id);
3308 spin_lock(&svcpt->scp_rep_lock);
3310 atomic_inc(&svcpt->scp_nreps_difficult);
3314 * error sending: reply is off the net. Also we need +1
3315 * reply ref until ptlrpc_handle_rs() is done
3316 * with the reply state (if the send was successful, there
3317 * would have been +1 ref for the net, which
3318 * reply_out_callback leaves alone)
3321 rs->rs_unlinked = 1;
3322 ptlrpc_rs_addref(rs);
3325 spin_lock(&rs->rs_lock);
3326 if (rs->rs_transno <= exp->exp_last_committed ||
3327 (rs->rs_unlinked && !rs->rs_no_ack) ||
3328 list_empty(&rs->rs_exp_list) || /* completed already */
3329 list_empty(&rs->rs_obd_list)) {
3330 CDEBUG(D_HA, "Schedule reply immediately\n");
3331 ptlrpc_dispatch_difficult_reply(rs);
3333 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3334 rs->rs_scheduled = 0; /* allow notifier to schedule */
3336 spin_unlock(&rs->rs_lock);
3337 spin_unlock(&svcpt->scp_rep_lock);
3341 enum ldlm_mode lck_compat_array[] = {
3342 [LCK_EX] = LCK_COMPAT_EX,
3343 [LCK_PW] = LCK_COMPAT_PW,
3344 [LCK_PR] = LCK_COMPAT_PR,
3345 [LCK_CW] = LCK_COMPAT_CW,
3346 [LCK_CR] = LCK_COMPAT_CR,
3347 [LCK_NL] = LCK_COMPAT_NL,
3348 [LCK_GROUP] = LCK_COMPAT_GROUP,
3349 [LCK_COS] = LCK_COMPAT_COS,
3353 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3354 * not escape to the user level.
3356 int ldlm_error2errno(enum ldlm_error error)
3362 case ELDLM_LOCK_MATCHED:
3365 case ELDLM_LOCK_CHANGED:
3368 case ELDLM_LOCK_ABORTED:
3371 case ELDLM_LOCK_REPLACED:
3374 case ELDLM_NO_LOCK_DATA:
3377 case ELDLM_NAMESPACE_EXISTS:
3380 case ELDLM_BAD_NAMESPACE:
3384 if (((int)error) < 0) { /* cast to signed type */
3385 result = error; /* as ldlm_error can be unsigned */
3387 CERROR("Invalid DLM result code: %d\n", error);
3393 EXPORT_SYMBOL(ldlm_error2errno);
3396 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3398 enum ldlm_error ldlm_errno2error(int err_no)
3407 error = ELDLM_LOCK_CHANGED;
3410 error = ELDLM_LOCK_ABORTED;
3413 error = ELDLM_LOCK_REPLACED;
3416 error = ELDLM_NO_LOCK_DATA;
3419 error = ELDLM_NAMESPACE_EXISTS;
3422 error = ELDLM_BAD_NAMESPACE;
3430 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3431 void ldlm_dump_export_locks(struct obd_export *exp)
3433 spin_lock(&exp->exp_locks_list_guard);
3434 if (!list_empty(&exp->exp_locks_list)) {
3435 struct ldlm_lock *lock;
3437 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3439 list_for_each_entry(lock, &exp->exp_locks_list,
3441 LDLM_ERROR(lock, "lock:");
3443 spin_unlock(&exp->exp_locks_list_guard);
3447 #ifdef HAVE_SERVER_SUPPORT
3448 static inline const char *bulk2type(struct ptlrpc_request *req)
3450 if (req->rq_bulk_read)
3452 if (req->rq_bulk_write)
3457 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3459 struct ptlrpc_request *req = desc->bd_req;
3460 time64_t start = ktime_get_seconds();
3466 /* If there is eviction in progress, wait for it to finish. */
3468 exp->exp_obd->obd_evict_inprogress_waitq,
3469 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3471 /* Check if client was evicted or reconnected already. */
3472 if (exp->exp_failed ||
3473 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3476 if (req->rq_bulk_read)
3477 rc = sptlrpc_svc_wrap_bulk(req, desc);
3479 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3480 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3481 else /* old version, bulk matchbits is rq_xid */
3482 req->rq_mbits = req->rq_xid;
3485 rc = ptlrpc_start_bulk_transfer(desc);
3489 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3490 bulk2type(req), rc);
3494 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3495 ptlrpc_abort_bulk(desc);
3499 /* limit actual bulk transfer to bulk_timeout seconds */
3500 deadline = start + bulk_timeout;
3501 if (deadline > req->rq_deadline)
3502 deadline = req->rq_deadline;
3505 time64_t timeoutl = deadline - ktime_get_seconds();
3506 time64_t rq_deadline;
3508 while (timeoutl >= 0 &&
3509 wait_event_idle_timeout(
3511 !ptlrpc_server_bulk_active(desc) ||
3514 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3515 timeoutl ? cfs_time_seconds(1) : 1) == 0)
3517 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3519 /* Wait again if we changed rq_deadline. */
3520 rq_deadline = READ_ONCE(req->rq_deadline);
3521 deadline = start + bulk_timeout;
3522 if (deadline > rq_deadline)
3523 deadline = rq_deadline;
3524 } while (rc == -ETIMEDOUT &&
3525 deadline > ktime_get_seconds());
3527 if (rc == -ETIMEDOUT) {
3528 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3529 bulk2type(req), deadline - start,
3530 ktime_get_real_seconds() - deadline);
3531 ptlrpc_abort_bulk(desc);
3532 } else if (exp->exp_failed) {
3533 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3536 ptlrpc_abort_bulk(desc);
3537 } else if (exp->exp_conn_cnt >
3538 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3539 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3541 /* We don't reply anyway. */
3543 ptlrpc_abort_bulk(desc);
3544 } else if (desc->bd_failure) {
3545 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3547 /* XXX should this be a different errno? */
3550 if (req->rq_bulk_write)
3551 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3552 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3553 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3554 bulk2type(req), desc->bd_nob_transferred,
3556 /* XXX should this be a different errno? */
3563 EXPORT_SYMBOL(target_bulk_io);
3565 #endif /* HAVE_SERVER_SUPPORT */