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 <uapi/linux/lustre/lustre_ioctl.h>
53 #include "ldlm_internal.h"
56 * @priority: If non-zero, move the selected connection to the list head.
57 * @create: If zero, only search in existing connections.
59 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
60 int priority, int create)
62 struct ptlrpc_connection *ptlrpc_conn;
63 struct obd_import_conn *imp_conn = NULL, *item;
64 u32 refnet = LNET_NET_ANY;
69 if (!create && !priority) {
70 CDEBUG(D_HA, "Nothing to do\n");
74 if (imp->imp_connection &&
75 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
76 /* refnet is used to restrict network connections */
77 refnet = LNET_NID_NET(&imp->imp_connection->c_self);
78 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, refnet);
80 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
85 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
87 GOTO(out_put, rc = -ENOMEM);
90 spin_lock(&imp->imp_lock);
91 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
92 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
94 list_move(&item->oic_item,
96 item->oic_last_attempt = 0;
98 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
99 imp, imp->imp_obd->obd_name, uuid->uuid,
100 (priority ? ", moved to head" : ""));
101 spin_unlock(&imp->imp_lock);
102 GOTO(out_free, rc = 0);
105 /* No existing import connection found for \a uuid. */
107 imp_conn->oic_conn = ptlrpc_conn;
108 imp_conn->oic_uuid = *uuid;
109 imp_conn->oic_last_attempt = 0;
111 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
113 list_add_tail(&imp_conn->oic_item,
114 &imp->imp_conn_list);
115 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
116 imp, imp->imp_obd->obd_name, uuid->uuid,
117 (priority ? "head" : "tail"));
119 spin_unlock(&imp->imp_lock);
120 GOTO(out_free, rc = -ENOENT);
123 spin_unlock(&imp->imp_lock);
127 OBD_FREE(imp_conn, sizeof(*imp_conn));
129 ptlrpc_connection_put(ptlrpc_conn);
133 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
135 return import_set_conn(imp, uuid, 1, 0);
138 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
141 return import_set_conn(imp, uuid, priority, 1);
143 EXPORT_SYMBOL(client_import_add_conn);
145 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
146 lnet_nid_t prim_nid, int priority)
148 struct ptlrpc_connection *ptlrpc_conn;
151 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, LNET_NIDNET(prim_nid));
153 const char *str_uuid = obd_uuid2str(uuid);
155 rc = class_add_uuid(str_uuid, prim_nid);
157 CERROR("%s: failed to add UUID '%s': rc = %d\n",
158 imp->imp_obd->obd_name, str_uuid, rc);
162 return import_set_conn(imp, uuid, priority, 1);
164 EXPORT_SYMBOL(client_import_dyn_add_conn);
166 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
167 int nid_count, struct obd_uuid *uuid)
169 struct obd_import_conn *conn;
173 if (nid_count <= 0 || !nids)
176 spin_lock(&imp->imp_lock);
177 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
178 if (class_check_uuid(&conn->oic_uuid, nids[0])) {
179 *uuid = conn->oic_uuid;
180 spin_unlock(&imp->imp_lock);
181 rc = class_add_nids_to_uuid(&conn->oic_uuid, nids,
186 spin_unlock(&imp->imp_lock);
189 EXPORT_SYMBOL(client_import_add_nids_to_conn);
191 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
193 struct obd_import_conn *imp_conn;
194 struct obd_export *dlmexp;
199 spin_lock(&imp->imp_lock);
200 if (list_empty(&imp->imp_conn_list)) {
201 LASSERT(!imp->imp_connection);
205 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
206 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
208 LASSERT(imp_conn->oic_conn);
210 if (imp_conn == imp->imp_conn_current) {
211 LASSERT(imp_conn->oic_conn == imp->imp_connection);
213 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
214 imp->imp_state != LUSTRE_IMP_DISCON) {
215 CERROR("can't remove current connection\n");
216 GOTO(out, rc = -EBUSY);
219 ptlrpc_connection_put(imp->imp_connection);
220 imp->imp_connection = NULL;
222 dlmexp = class_conn2export(&imp->imp_dlm_handle);
223 if (dlmexp && dlmexp->exp_connection) {
224 LASSERT(dlmexp->exp_connection ==
226 ptlrpc_connection_put(dlmexp->exp_connection);
227 dlmexp->exp_connection = NULL;
231 class_export_put(dlmexp);
234 list_del(&imp_conn->oic_item);
235 ptlrpc_connection_put(imp_conn->oic_conn);
236 OBD_FREE(imp_conn, sizeof(*imp_conn));
237 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
238 imp, imp->imp_obd->obd_name, uuid->uuid);
243 spin_unlock(&imp->imp_lock);
245 CERROR("connection %s not found\n", uuid->uuid);
248 EXPORT_SYMBOL(client_import_del_conn);
251 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
252 * to find a conn uuid of \a imp which can reach \a peer.
254 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
255 struct obd_uuid *uuid)
257 struct obd_import_conn *conn;
262 spin_lock(&imp->imp_lock);
263 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
264 /* Check if conn UUID does have this peer NID. */
265 if (class_check_uuid(&conn->oic_uuid, peer)) {
266 *uuid = conn->oic_uuid;
271 spin_unlock(&imp->imp_lock);
274 EXPORT_SYMBOL(client_import_find_conn);
276 void client_destroy_import(struct obd_import *imp)
279 * Drop security policy instance after all RPCs have finished/aborted
280 * to let all busy contexts be released.
282 class_import_get(imp);
283 class_destroy_import(imp);
284 sptlrpc_import_sec_put(imp);
285 class_import_put(imp);
287 EXPORT_SYMBOL(client_destroy_import);
290 * Check whether or not the OSC is on MDT.
293 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
295 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
298 static int osc_on_mdt(char *obdname)
302 ptr = strrchr(obdname, '-');
306 if (strncmp(ptr + 1, "MDT", 3) == 0)
313 * Configure an RPC client OBD device.
318 * 3 - inactive-on-startup
319 * 4 - restrictive net
321 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
323 struct client_obd *cli = &obd->u.cli;
324 struct obd_import *imp;
325 struct obd_uuid server_uuid;
326 int rq_portal, rp_portal, connect_op;
327 const char *name = obd->obd_type->typ_name;
328 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
329 char *cli_name = lustre_cfg_buf(lcfg, 0);
330 struct ptlrpc_connection fake_conn = { .c_self = {},
331 .c_remote_uuid.uuid[0] = 0 };
337 * In a more perfect world, we would hang a ptlrpc_client off of
338 * obd_type and just use the values from there.
340 if (!strcmp(name, LUSTRE_OSC_NAME)) {
341 rq_portal = OST_REQUEST_PORTAL;
342 rp_portal = OSC_REPLY_PORTAL;
343 connect_op = OST_CONNECT;
344 cli->cl_sp_me = LUSTRE_SP_CLI;
345 cli->cl_sp_to = LUSTRE_SP_OST;
346 ns_type = LDLM_NS_TYPE_OSC;
347 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
348 !strcmp(name, LUSTRE_LWP_NAME)) {
349 rq_portal = MDS_REQUEST_PORTAL;
350 rp_portal = MDC_REPLY_PORTAL;
351 connect_op = MDS_CONNECT;
352 if (is_lwp_on_ost(cli_name))
353 cli->cl_sp_me = LUSTRE_SP_OST;
354 else if (is_lwp_on_mdt(cli_name))
355 cli->cl_sp_me = LUSTRE_SP_MDT;
357 cli->cl_sp_me = LUSTRE_SP_CLI;
358 cli->cl_sp_to = LUSTRE_SP_MDT;
359 ns_type = LDLM_NS_TYPE_MDC;
360 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
361 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
362 /* OSP_on_MDT for other MDTs */
363 connect_op = MDS_CONNECT;
364 cli->cl_sp_to = LUSTRE_SP_MDT;
365 ns_type = LDLM_NS_TYPE_MDC;
366 rq_portal = OUT_PORTAL;
368 /* OSP on MDT for OST */
369 connect_op = OST_CONNECT;
370 cli->cl_sp_to = LUSTRE_SP_OST;
371 ns_type = LDLM_NS_TYPE_OSC;
372 rq_portal = OST_REQUEST_PORTAL;
374 rp_portal = OSC_REPLY_PORTAL;
375 cli->cl_sp_me = LUSTRE_SP_MDT;
376 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
377 rq_portal = MGS_REQUEST_PORTAL;
378 rp_portal = MGC_REPLY_PORTAL;
379 connect_op = MGS_CONNECT;
380 cli->cl_sp_me = LUSTRE_SP_MGC;
381 cli->cl_sp_to = LUSTRE_SP_MGS;
382 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
383 ns_type = LDLM_NS_TYPE_MGC;
385 CERROR("unknown client OBD type \"%s\", can't setup\n",
390 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
391 CERROR("requires a TARGET UUID\n");
395 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
396 CERROR("client UUID must be less than 38 characters\n");
400 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
401 CERROR("setup requires a SERVER UUID\n");
405 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
406 CERROR("target UUID must be less than 38 characters\n");
410 init_rwsem(&cli->cl_sem);
411 mutex_init(&cli->cl_mgc_mutex);
413 init_rwsem(&cli->cl_seq_rwsem);
414 cli->cl_conn_count = 0;
415 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
416 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
417 sizeof(server_uuid)));
419 cli->cl_dirty_pages = 0;
420 cli->cl_dirty_max_pages = 0;
421 cli->cl_avail_grant = 0;
422 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
424 * cl_dirty_max_pages may be changed at connect time in
425 * ptlrpc_connect_interpret().
427 client_adjust_max_dirty(cli);
428 init_waitqueue_head(&cli->cl_cache_waiters);
429 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
430 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
431 INIT_LIST_HEAD(&cli->cl_loi_write_list);
432 INIT_LIST_HEAD(&cli->cl_loi_read_list);
433 spin_lock_init(&cli->cl_loi_list_lock);
434 atomic_set(&cli->cl_pending_w_pages, 0);
435 atomic_set(&cli->cl_pending_r_pages, 0);
436 cli->cl_r_in_flight = 0;
437 cli->cl_w_in_flight = 0;
439 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
440 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
441 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
442 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
443 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
444 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
447 INIT_LIST_HEAD(&cli->cl_lru_osc);
448 atomic_set(&cli->cl_lru_shrinkers, 0);
449 atomic_long_set(&cli->cl_lru_busy, 0);
450 atomic_long_set(&cli->cl_lru_in_list, 0);
451 INIT_LIST_HEAD(&cli->cl_lru_list);
452 spin_lock_init(&cli->cl_lru_list_lock);
453 atomic_long_set(&cli->cl_unstable_count, 0);
454 INIT_LIST_HEAD(&cli->cl_shrink_list);
455 INIT_LIST_HEAD(&cli->cl_grant_chain);
457 INIT_LIST_HEAD(&cli->cl_flight_waiters);
458 cli->cl_rpcs_in_flight = 0;
460 init_waitqueue_head(&cli->cl_destroy_waitq);
461 atomic_set(&cli->cl_destroy_in_flight, 0);
464 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
465 cli->cl_preferred_cksum_type = 0;
466 #ifdef ENABLE_CHECKSUM
467 /* Turn on checksumming by default. */
468 cli->cl_checksum = 1;
470 * The supported checksum types will be worked out at connect time
471 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
474 cli->cl_cksum_type = cli->cl_supp_cksum_types;
476 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
479 * Set it to possible maximum size. It may be reduced by ocd_brw_size
480 * from OFD after connecting.
482 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
484 cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
487 * set cl_chunkbits default value to PAGE_SHIFT,
488 * it will be updated at OSC connection time.
490 cli->cl_chunkbits = PAGE_SHIFT;
492 if (!strcmp(name, LUSTRE_MDC_NAME)) {
493 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
494 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
495 cli->cl_max_rpcs_in_flight = 2;
496 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
497 cli->cl_max_rpcs_in_flight = 3;
498 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
499 cli->cl_max_rpcs_in_flight = 4;
501 if (osc_on_mdt(obd->obd_name))
502 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
504 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
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_rpcs_init = ktime_get_real();
513 cli->cl_mod_tag_bitmap = NULL;
515 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
517 if (connect_op == MDS_CONNECT) {
518 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
519 OBD_ALLOC(cli->cl_mod_tag_bitmap,
520 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
521 if (cli->cl_mod_tag_bitmap == NULL)
522 GOTO(err, rc = -ENOMEM);
527 CERROR("ldlm_get_ref failed: %d\n", rc);
531 ptlrpc_init_client(rq_portal, rp_portal, name,
532 &obd->obd_ldlm_client);
534 imp = class_new_import(obd);
536 GOTO(err_ldlm, rc = -ENOENT);
537 imp->imp_client = &obd->obd_ldlm_client;
538 imp->imp_connect_op = connect_op;
539 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
540 LUSTRE_CFG_BUFLEN(lcfg, 1));
541 class_import_put(imp);
543 if (lustre_cfg_buf(lcfg, 4)) {
544 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
546 if (refnet == LNET_NET_ANY) {
548 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
549 obd->obd_name, lustre_cfg_string(lcfg, 4),
551 GOTO(err_import, rc);
553 lnet_nid4_to_nid(LNET_MKNID(refnet, 0), &fake_conn.c_self);
554 imp->imp_connection = &fake_conn;
557 rc = client_import_add_conn(imp, &server_uuid, 1);
559 CERROR("can't add initial connection\n");
560 GOTO(err_import, rc);
562 imp->imp_connection = NULL;
564 cli->cl_import = imp;
565 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
566 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
568 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
569 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
570 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
572 cli->cl_target_uuid.uuid);
573 spin_lock(&imp->imp_lock);
574 imp->imp_deactive = 1;
575 spin_unlock(&imp->imp_lock);
579 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
580 LDLM_NAMESPACE_CLIENT,
581 LDLM_NAMESPACE_GREEDY,
583 if (IS_ERR(obd->obd_namespace)) {
584 rc = PTR_ERR(obd->obd_namespace);
585 CERROR("%s: unable to create client namespace: rc = %d\n",
587 obd->obd_namespace = NULL;
588 GOTO(err_import, rc);
594 class_destroy_import(imp);
598 if (cli->cl_mod_tag_bitmap != NULL)
599 OBD_FREE(cli->cl_mod_tag_bitmap,
600 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
601 cli->cl_mod_tag_bitmap = NULL;
605 EXPORT_SYMBOL(client_obd_setup);
607 int client_obd_cleanup(struct obd_device *obd)
609 struct client_obd *cli = &obd->u.cli;
613 ldlm_namespace_free_post(obd->obd_namespace);
614 obd->obd_namespace = NULL;
616 obd_cleanup_client_import(obd);
617 LASSERT(obd->u.cli.cl_import == NULL);
621 if (cli->cl_mod_tag_bitmap != NULL)
622 OBD_FREE(cli->cl_mod_tag_bitmap,
623 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
624 cli->cl_mod_tag_bitmap = NULL;
628 EXPORT_SYMBOL(client_obd_cleanup);
630 /* ->o_connect() method for client side (OSC and MDC and MGC) */
631 int client_connect_import(const struct lu_env *env,
632 struct obd_export **exp,
633 struct obd_device *obd, struct obd_uuid *cluuid,
634 struct obd_connect_data *data, void *localdata)
636 struct client_obd *cli = &obd->u.cli;
637 struct obd_import *imp = cli->cl_import;
638 struct obd_connect_data *ocd;
639 struct lustre_handle conn = { 0 };
645 down_write(&cli->cl_sem);
646 if (cli->cl_conn_count > 0)
647 GOTO(out_sem, rc = -EALREADY);
649 rc = class_connect(&conn, obd, cluuid);
653 cli->cl_conn_count++;
654 *exp = class_conn2export(&conn);
656 LASSERT(obd->obd_namespace);
658 spin_lock(&imp->imp_lock);
659 if (imp->imp_state == LUSTRE_IMP_CLOSED && imp->imp_deactive) {
660 /* need to reactivate import if trying to connect
661 * to a previously disconnected
663 imp->imp_deactive = 0;
664 imp->imp_invalid = 0;
666 spin_unlock(&imp->imp_lock);
668 imp->imp_dlm_handle = conn;
669 rc = ptlrpc_init_import(imp);
673 ocd = &imp->imp_connect_data;
676 imp->imp_connect_flags_orig = data->ocd_connect_flags;
677 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
680 rc = ptlrpc_connect_import(imp);
682 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
685 LASSERT(*exp != NULL && (*exp)->exp_connection);
688 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
689 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
690 data->ocd_connect_flags, ocd->ocd_connect_flags);
691 data->ocd_connect_flags = ocd->ocd_connect_flags;
692 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
695 ptlrpc_pinger_add_import(imp);
701 cli->cl_conn_count--;
702 class_disconnect(*exp);
706 up_write(&cli->cl_sem);
708 if (!rc && localdata && !cli->cl_cache) {
709 cli->cl_cache = (struct cl_client_cache *)localdata;
710 cl_cache_incref(cli->cl_cache);
711 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
713 /* add this osc into entity list */
714 LASSERT(list_empty(&cli->cl_lru_osc));
715 spin_lock(&cli->cl_cache->ccc_lru_lock);
716 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
717 spin_unlock(&cli->cl_cache->ccc_lru_lock);
722 EXPORT_SYMBOL(client_connect_import);
724 int client_disconnect_export(struct obd_export *exp)
726 struct obd_device *obd = class_exp2obd(exp);
727 struct client_obd *cli;
728 struct obd_import *imp;
734 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
735 exp, exp ? exp->exp_handle.h_cookie : -1);
740 imp = cli->cl_import;
742 down_write(&cli->cl_sem);
743 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
746 if (cli->cl_conn_count == 0) {
747 CERROR("disconnecting disconnected device (%s)\n",
749 GOTO(out_disconnect, rc = -EINVAL);
752 cli->cl_conn_count--;
753 if (cli->cl_conn_count != 0)
754 GOTO(out_disconnect, rc = 0);
757 * Mark import deactivated now, so we don't try to reconnect if any
758 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
759 * fully deactivate the import, or that would drop all requests.
761 spin_lock(&imp->imp_lock);
762 imp->imp_deactive = 1;
763 spin_unlock(&imp->imp_lock);
766 * Some non-replayable imports (MDS's OSCs) are pinged, so just
767 * delete it regardless. (It's safe to delete an import that was
770 (void)ptlrpc_pinger_del_import(imp);
772 if (obd->obd_namespace != NULL) {
773 /* obd_force == local only */
774 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
775 obd->obd_force ? LCF_LOCAL : 0, NULL);
776 ldlm_namespace_free_prior(obd->obd_namespace, imp,
781 * There's no need to hold sem while disconnecting an import,
782 * and it may actually cause deadlock in GSS.
784 up_write(&cli->cl_sem);
785 rc = ptlrpc_disconnect_import(imp, 0);
786 down_write(&cli->cl_sem);
788 ptlrpc_invalidate_import(imp);
794 * Use server style - class_disconnect should be always called for
797 err = class_disconnect(exp);
801 up_write(&cli->cl_sem);
805 EXPORT_SYMBOL(client_disconnect_export);
807 #ifdef HAVE_SERVER_SUPPORT
808 int server_disconnect_export(struct obd_export *exp)
814 /* Disconnect early so that clients can't keep using export. */
815 rc = class_disconnect(exp);
816 /* Close import to avoid sending any requests. */
817 if (exp->exp_imp_reverse)
818 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
820 ldlm_bl_thread_wakeup();
822 /* complete all outstanding replies */
823 spin_lock(&exp->exp_lock);
824 while (!list_empty(&exp->exp_outstanding_replies)) {
825 struct ptlrpc_reply_state *rs =
826 list_entry(exp->exp_outstanding_replies.next,
827 struct ptlrpc_reply_state, rs_exp_list);
828 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
830 spin_lock(&svcpt->scp_rep_lock);
832 list_del_init(&rs->rs_exp_list);
834 spin_lock(&rs->rs_lock);
835 /* clear rs_convert_lock to make sure rs is handled and put */
836 rs->rs_convert_lock = 0;
837 ptlrpc_schedule_difficult_reply(rs);
838 spin_unlock(&rs->rs_lock);
840 spin_unlock(&svcpt->scp_rep_lock);
842 spin_unlock(&exp->exp_lock);
846 EXPORT_SYMBOL(server_disconnect_export);
848 static inline int target_check_recovery_timer(struct obd_device *target)
853 if (!target->obd_recovering || target->obd_recovery_start == 0)
856 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
857 timeout = ktime_divns(remaining, NSEC_PER_SEC);
861 /* the recovery timer should expire, but it isn't triggered,
862 * it's better to abort the recovery of this target to speed up
863 * the recovery of the whole cluster. */
864 spin_lock(&target->obd_dev_lock);
865 if (target->obd_recovering) {
866 CERROR("%s: Aborting recovery\n", target->obd_name);
867 target->obd_abort_recovery = 1;
868 wake_up(&target->obd_next_transno_waitq);
870 spin_unlock(&target->obd_dev_lock);
875 * --------------------------------------------------------------------------
876 * from old lib/target.c
877 * --------------------------------------------------------------------------
879 static int target_handle_reconnect(struct lustre_handle *conn,
880 struct obd_export *exp,
881 struct obd_uuid *cluuid)
883 struct obd_device *target;
884 struct lustre_handle *hdl;
890 hdl = &exp->exp_imp_reverse->imp_remote_handle;
891 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
892 conn->cookie = exp->exp_handle.h_cookie;
894 "connect export for UUID '%s' at %p, cookie %#llx\n",
895 cluuid->uuid, exp, conn->cookie);
899 target = exp->exp_obd;
901 /* Might be a re-connect after a partition. */
902 if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
903 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",
905 obd_uuid2str(&exp->exp_client_uuid),
906 obd_export_nid2str(exp),
907 hdl->cookie, conn->cookie);
908 memset(conn, 0, sizeof(*conn));
910 * target_handle_connect() treats EALREADY and
911 * -EALREADY differently. -EALREADY is an error
912 * (same UUID, different handle).
917 if (!target->obd_recovering) {
918 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
919 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
920 obd_export_nid2str(exp));
921 GOTO(out_already, rc);
924 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
925 timeout = ktime_divns(remaining, NSEC_PER_SEC);
927 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
929 obd_uuid2str(&exp->exp_client_uuid),
930 obd_export_nid2str(exp),
931 atomic_read(&target->obd_max_recoverable_clients),
932 timeout / 60, timeout % 60);
934 struct target_distribute_txn_data *tdtd;
939 target_check_recovery_timer(target);
941 tdtd = class_exp2tgt(exp)->lut_tdtd;
942 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
943 buf = tdtd->tdtd_show_update_logs_retrievers(
944 tdtd->tdtd_show_retrievers_cbdata,
948 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",
950 obd_uuid2str(&exp->exp_client_uuid),
951 obd_export_nid2str(exp), count,
952 buf ? buf : "unknown (not enough RAM)",
953 (abs(timeout) + target->obd_recovery_timeout) / 60,
954 (abs(timeout) + target->obd_recovery_timeout) % 60,
957 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",
958 target->obd_name, abs(timeout) / 60,
959 abs(timeout) % 60, target->obd_name);
966 conn->cookie = exp->exp_handle.h_cookie;
968 * target_handle_connect() treats EALREADY and
969 * -EALREADY differently. EALREADY means we are
970 * doing a valid reconnect from the same client.
976 check_and_start_recovery_timer(struct obd_device *obd,
977 struct ptlrpc_request *req, int new_client);
980 * update flags for import during reconnect process
982 static int rev_import_flags_update(struct obd_import *revimp,
983 struct ptlrpc_request *req)
986 struct obd_connect_data *data;
988 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
990 if (data->ocd_connect_flags & OBD_CONNECT_AT)
991 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
993 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
995 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
997 revimp->imp_connect_data = *data;
998 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1000 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
1001 revimp->imp_client->cli_name,
1002 libcfs_idstr(&req->rq_peer), rc);
1010 * Allocate a new reverse import for an export.
1012 * \retval -errno in case error hit
1013 * \retval 0 if reverse import correctly init
1015 int rev_import_init(struct obd_export *export)
1017 struct obd_device *obd = export->exp_obd;
1018 struct obd_import *revimp;
1020 LASSERT(export->exp_imp_reverse == NULL);
1022 revimp = class_new_import(obd);
1026 revimp->imp_remote_handle.cookie = 0ULL;
1027 revimp->imp_client = &obd->obd_ldlm_client;
1028 revimp->imp_dlm_fake = 1;
1030 /* it is safe to connect import in new state as no sends possible */
1031 spin_lock(&export->exp_lock);
1032 export->exp_imp_reverse = revimp;
1033 spin_unlock(&export->exp_lock);
1034 class_import_put(revimp);
1038 EXPORT_SYMBOL(rev_import_init);
1041 * Handle reconnect for an export.
1043 * \param exp export to handle reconnect process
1044 * \param req client reconnect request
1046 * \retval -rc in case securitfy flavor can't be changed
1047 * \retval 0 in case none problems
1049 static int rev_import_reconnect(struct obd_export *exp,
1050 struct ptlrpc_request *req)
1052 struct obd_import *revimp = exp->exp_imp_reverse;
1053 struct lustre_handle *lh;
1056 /* avoid sending a request until import flags are changed */
1057 ptlrpc_import_enter_resend(revimp);
1059 ptlrpc_connection_put(revimp->imp_connection);
1062 * client from recovery don't have a handle so we need to take from
1063 * request. it may produce situation when wrong client connected
1064 * to recovery as we trust a client uuid
1066 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1067 revimp->imp_remote_handle = *lh;
1070 * unknown versions will be caught in
1071 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1073 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1075 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1077 rc = rev_import_flags_update(revimp, req);
1080 * it is safe to still be in RECOVERY phase as we are not able
1081 * to setup correct security flavor so requests are not able to
1082 * be delivered correctly
1087 /* resend all rpc's via new connection */
1088 return ptlrpc_import_recovery_state_machine(revimp);
1091 int target_handle_connect(struct ptlrpc_request *req)
1093 struct obd_device *target = NULL;
1094 struct obd_export *export = NULL;
1096 * connect handle - filled from target_handle_reconnect in
1099 struct lustre_handle conn;
1100 struct lustre_handle *tmp;
1101 struct obd_uuid cluuid;
1106 bool mds_conn = false, lw_client = false, initial_conn = false;
1107 bool mds_mds_conn = false;
1108 bool new_mds_mds_conn = false;
1109 struct obd_connect_data *data, *tmpdata;
1111 lnet_nid_t client_nid;
1112 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1113 int tmp_exp_old_falloc;
1115 struct ptlrpc_connection *pcon = NULL;
1119 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1121 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1123 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1124 GOTO(out, rc = -EINVAL);
1127 target = class_dev_by_str(str);
1129 deuuidify(str, NULL, &target_start, &target_len);
1130 LCONSOLE_ERROR_MSG(0x137,
1131 "%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",
1132 str, libcfs_nidstr(&req->rq_peer.nid));
1133 GOTO(out, rc = -ENODEV);
1136 spin_lock(&target->obd_dev_lock);
1138 target->obd_conn_inprogress++;
1140 if (target->obd_stopping || !target->obd_set_up) {
1141 spin_unlock(&target->obd_dev_lock);
1143 deuuidify(str, NULL, &target_start, &target_len);
1144 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1145 target_len, target_start,
1146 libcfs_nidstr(&req->rq_peer.nid),
1147 (target->obd_stopping ?
1148 "stopping" : "not set up"));
1149 GOTO(out, rc = -ENODEV);
1152 if (target->obd_no_conn) {
1153 spin_unlock(&target->obd_dev_lock);
1156 "%s: Temporarily refusing client connection from %s\n",
1157 target->obd_name, libcfs_nidstr(&req->rq_peer.nid));
1158 GOTO(out, rc = -EAGAIN);
1161 spin_unlock(&target->obd_dev_lock);
1163 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1165 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1166 GOTO(out, rc = -EINVAL);
1169 obd_str2uuid(&cluuid, str);
1171 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1173 GOTO(out, rc = -EPROTO);
1177 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1179 if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1180 GOTO(out, rc = -EPROTO);
1181 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1183 GOTO(out, rc = -EPROTO);
1185 rc = req_capsule_server_pack(&req->rq_pill);
1189 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1191 * Don't allow clients to connect that are using old 1.8 format
1192 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1193 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1194 * FULL20 flag should be set on all connections since 2.0, but no
1195 * longer affects behaviour.
1197 * Later this check will be disabled and the flag can be retired
1198 * completely once interop with 3.0 is no longer needed.
1200 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1201 GOTO(out, rc = -EPROTO);
1203 /* Old clients will have 'tmp_exp_old_falloc' as 1.
1204 * Newer clients (2.15) and beyond will have it set as 0
1206 tmp_exp_old_falloc =
1207 !!(data->ocd_connect_flags & OBD_CONNECT_OLD_FALLOC);
1209 CDEBUG(D_INFO, "%s: ocd_connect_flags: %#llx tmp_exp_old_falloc: %d\n",
1210 target->obd_name, data->ocd_connect_flags, tmp_exp_old_falloc);
1213 * Don't allow liblustre clients to connect.
1214 * - testing was disabled in v2_2_50_0-61-g6a75d65
1215 * - building was disabled in v2_5_58_0-28-g7277179
1216 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1217 * - clients were refused connect for version difference > 0.0.1.32
1219 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1220 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1221 GOTO(out, rc = -EPROTO);
1226 * Note: lw_client is needed in MDS-MDS failover during update log
1227 * processing, so we needs to allow lw_client to be connected at
1228 * anytime, instead of only the initial connection
1230 lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1232 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1233 initial_conn = true;
1234 mds_conn = OCD_HAS_FLAG(data, MDS);
1235 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1238 * OBD_CONNECT_MNE_SWAB is removed at 2.14
1239 * Checking OBD_CONNECT_FID can be removed in the future.
1241 * Via check OBD_CONNECT_FID, we can distinguish whether
1242 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1243 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1246 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1247 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1248 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1249 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1250 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1251 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1254 * We do not support the MDT-MDT interoperations with
1255 * different version MDT because of protocol changes.
1257 if (unlikely(major != LUSTRE_MAJOR ||
1258 minor != LUSTRE_MINOR ||
1259 abs(patch - LUSTRE_PATCH) > 3)) {
1260 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1261 target->obd_name, LUSTRE_MAJOR,
1262 LUSTRE_MINOR, LUSTRE_PATCH,
1263 LUSTRE_FIX, major, minor, patch,
1264 OBD_OCD_VERSION_FIX(data->ocd_version),
1265 libcfs_nidstr(&req->rq_peer.nid),
1267 GOTO(out, rc = -EPROTO);
1272 /* lctl gets a backstage, all-access pass. */
1273 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1274 goto dont_check_exports;
1276 export = obd_uuid_lookup(target, &cluuid);
1280 /* We've found an export in the hash. */
1282 spin_lock(&export->exp_lock);
1284 if (export->exp_connecting) { /* b=9635, et. al. */
1285 spin_unlock(&export->exp_lock);
1286 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1287 export->exp_obd->obd_name, export,
1288 libcfs_nidstr(&req->rq_peer.nid));
1289 class_export_put(export);
1292 } else if ((mds_conn || (lw_client && initial_conn) ||
1293 OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
1294 spin_unlock(&export->exp_lock);
1295 if (!nid_same(&req->rq_peer.nid,
1296 &export->exp_connection->c_peer.nid)) {
1297 /* MDS or LWP reconnected after failover. */
1298 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1300 lw_client ? "LWP" : "MDS",
1301 libcfs_nidstr(&req->rq_peer.nid),
1302 libcfs_nidstr(&export->exp_connection->c_peer.nid));
1304 /* New connection from the same NID. */
1305 LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1307 lw_client ? "LWP" : "MDS",
1308 libcfs_nidstr(&req->rq_peer.nid),
1309 OCD_HAS_FLAG(data, MDS_MDS) ?
1313 if (nid_same(&req->rq_peer.nid,
1314 &export->exp_connection->c_peer.nid) &&
1315 OCD_HAS_FLAG(data, MDS_MDS)) {
1317 * Because exports between MDTs will always be
1318 * kept, let's do not fail such export if they
1319 * come from the same NID, otherwise it might
1320 * cause eviction between MDTs, which might
1321 * cause namespace inconsistency
1323 spin_lock(&export->exp_lock);
1324 export->exp_connecting = 1;
1325 export->exp_conn_cnt = 0;
1326 spin_unlock(&export->exp_lock);
1327 conn.cookie = export->exp_handle.h_cookie;
1330 class_fail_export(export);
1331 class_export_put(export);
1335 } else if (export->exp_connection != NULL && initial_conn &&
1336 !nid_same(&req->rq_peer.nid,
1337 &export->exp_connection->c_peer.nid)) {
1338 spin_unlock(&export->exp_lock);
1339 /* In MDS failover we have static UUID but NID can change. */
1340 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1341 target->obd_name, cluuid.uuid,
1342 libcfs_nidstr(&req->rq_peer.nid),
1344 &export->exp_connection->c_peer.nid));
1346 class_export_put(export);
1348 } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1350 spin_unlock(&export->exp_lock);
1353 export->exp_connecting = 1;
1354 spin_unlock(&export->exp_lock);
1355 LASSERT(export->exp_obd == target);
1357 rc = target_handle_reconnect(&conn, export, &cluuid);
1360 /* If we found an export, we already unlocked. */
1363 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1364 } else if (req->rq_export == NULL &&
1365 atomic_read(&export->exp_rpc_count) > 0) {
1366 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1367 target->obd_name, cluuid.uuid,
1368 libcfs_nidstr(&req->rq_peer.nid),
1369 refcount_read(&export->exp_handle.h_ref));
1370 GOTO(out, rc = -EBUSY);
1371 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1373 if (!strstr(cluuid.uuid, "mdt"))
1374 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1375 target->obd_name, cluuid.uuid,
1376 libcfs_nidstr(&req->rq_peer.nid));
1377 GOTO(out, rc = -EALREADY);
1379 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1385 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1386 target->obd_name, cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1387 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1388 export, ktime_get_seconds(),
1389 export ? export->exp_last_request_time : 0);
1392 * If this is the first time a client connects, reset the recovery
1393 * timer. Discard lightweight connections which might be local.
1395 if (!lw_client && rc == 0 && target->obd_recovering)
1396 check_and_start_recovery_timer(target, req, export == NULL);
1399 * We want to handle EALREADY but *not* -EALREADY from
1400 * target_handle_reconnect(), return reconnection state in a flag.
1402 if (rc == EALREADY) {
1403 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1409 /* Tell the client if we support replayable requests. */
1410 if (target->obd_replayable)
1411 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1412 client_nid = lnet_nid_to_nid4(&req->rq_peer.nid);
1414 if (export == NULL) {
1415 /* allow lightweight connections during recovery */
1417 * allow "new" MDT to be connected during recovery, since we
1418 * need retrieve recovery update records from it
1420 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1421 struct hrtimer *timer = &target->obd_recovery_timer;
1430 connected = atomic_read(&target->obd_connected_clients);
1431 in_progress = atomic_read(&target->obd_lock_replay_clients);
1433 atomic_read(&target->obd_max_recoverable_clients);
1434 stale = target->obd_stale_clients;
1435 remaining = hrtimer_get_remaining(timer);
1436 left = ktime_divns(remaining, NSEC_PER_SEC);
1438 if (ktime_to_ns(remaining) > 0) {
1439 msg = "to recover in";
1442 msg = "already passed deadline";
1445 target_check_recovery_timer(target);
1448 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",
1449 target->obd_name, cluuid.uuid,
1450 libcfs_nidstr(&req->rq_peer.nid), known,
1451 connected - in_progress, in_progress,
1452 stale, msg, timeout / 60, timeout % 60);
1456 rc = obd_connect(req->rq_svc_thread->t_env,
1457 &export, target, &cluuid, data,
1459 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1460 lustre_msg_add_op_flags(req->rq_repmsg,
1461 MSG_CONNECT_RECOVERING);
1463 conn.cookie = export->exp_handle.h_cookie;
1464 rc = rev_import_init(export);
1468 new_mds_mds_conn = true;
1471 rc = obd_reconnect(req->rq_svc_thread->t_env,
1472 export, target, &cluuid, data, &client_nid);
1477 data->ocd_instance = obd2obt(target)->obt_instance;
1480 * Return only the parts of obd_connect_data that we understand, so the
1481 * client knows that we don't understand the rest.
1484 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1486 tmpdata = req_capsule_server_get(&req->rq_pill,
1489 * Don't use struct assignment here, because the client reply
1490 * buffer may be smaller/larger than the local struct
1493 memcpy(tmpdata, data, min(tmpsize, size));
1497 * If the client and the server are the same node, we will already
1498 * have an export that really points to the client's DLM export,
1499 * because we have a shared handles table.
1501 * XXX this will go away when shaver stops sending the "connect" handle
1502 * in the real "remote handle" field of the request --phik 24 Apr 2003
1504 ptlrpc_request_change_export(req, export);
1506 pcon = ptlrpc_connection_get(&req->rq_peer,
1507 &req->rq_self, &cluuid);
1509 GOTO(out, rc = -ENOTCONN);
1511 spin_lock(&export->exp_lock);
1513 if (export->exp_disconnected) {
1514 spin_unlock(&export->exp_lock);
1515 GOTO(out, rc = -ENODEV);
1517 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1518 spin_unlock(&export->exp_lock);
1520 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1521 cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1522 export->exp_conn_cnt,
1523 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1525 GOTO(out, rc = -EALREADY);
1527 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1528 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1530 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1531 /* make 'tmp_exp_old_falloc' persistent by saving it into
1532 * server side export object(obd_export)
1534 export->exp_old_falloc = tmp_exp_old_falloc;
1537 /* Check to see if connection came from another NID. */
1538 if (export->exp_connection != NULL &&
1539 !nid_same(&export->exp_connection->c_peer.nid,
1540 &req->rq_peer.nid)) {
1541 obd_nid_del(export->exp_obd, export);
1542 ptlrpc_connection_put(export->exp_connection);
1543 export->exp_connection = NULL;
1546 if (export->exp_connection == NULL) {
1547 export->exp_connection = pcon;
1550 obd_nid_add(export->exp_obd, export);
1552 spin_unlock(&export->exp_lock);
1554 lustre_msg_set_handle(req->rq_repmsg, &conn);
1556 rc = rev_import_reconnect(export, req);
1560 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1562 __u64 transno = data->ocd_transno;
1564 spin_lock(&export->exp_lock);
1566 * possible race with class_disconnect_stale_exports,
1567 * export may be already in the eviction process
1569 if (export->exp_failed) {
1570 spin_unlock(&export->exp_lock);
1571 GOTO(out, rc = -ENODEV);
1573 export->exp_in_recovery = 1;
1574 export->exp_req_replay_needed = 1;
1575 export->exp_lock_replay_needed = 1;
1576 spin_unlock(&export->exp_lock);
1578 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1579 MSG_CONNECT_TRANSNO);
1580 if (has_transno && transno == 0)
1581 CWARN("Connect with zero transno!\n");
1583 if (has_transno && transno > 0 &&
1584 transno < target->obd_next_recovery_transno &&
1585 transno > target->obd_last_committed) {
1586 /* Another way is to use cmpxchg() to be lock-free. */
1587 spin_lock(&target->obd_recovery_task_lock);
1588 if (transno < target->obd_next_recovery_transno)
1589 target->obd_next_recovery_transno = transno;
1590 spin_unlock(&target->obd_recovery_task_lock);
1593 atomic_inc(&target->obd_req_replay_clients);
1594 atomic_inc(&target->obd_lock_replay_clients);
1596 * Note: MDS-MDS connection is allowed to be connected during
1597 * recovery, no matter if the exports needs to be recoveried.
1598 * Because we need retrieve updates logs from all other MDTs.
1599 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1600 * also needs to be increased to match other recovery checking
1603 if (new_mds_mds_conn)
1604 atomic_inc(&target->obd_max_recoverable_clients);
1606 if (atomic_inc_return(&target->obd_connected_clients) ==
1607 atomic_read(&target->obd_max_recoverable_clients))
1608 wake_up(&target->obd_next_transno_waitq);
1611 /* Tell the client we're in recovery, when client is involved in it. */
1612 if (target->obd_recovering && !lw_client)
1613 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1617 spin_lock(&export->exp_lock);
1618 export->exp_connecting = 0;
1619 spin_unlock(&export->exp_lock);
1621 class_export_put(export);
1623 if (target != NULL) {
1624 spin_lock(&target->obd_dev_lock);
1625 target->obd_conn_inprogress--;
1626 spin_unlock(&target->obd_dev_lock);
1627 class_decref(target, "find", current);
1630 ptlrpc_connection_put(pcon);
1631 req->rq_status = rc;
1635 int target_handle_disconnect(struct ptlrpc_request *req)
1641 rc = req_capsule_server_pack(&req->rq_pill);
1645 /* In case of target disconnect, updating sec ctx immediately is
1646 * required in order to record latest sequence number used.
1647 * Sequence is normally updated on export destroy, but this event
1648 * can occur too late, ie after a new target connect request has
1650 * Maintaining correct sequence when client connection becomes idle
1651 * ensures that GSS does not erroneously consider requests as replays.
1653 rc = sptlrpc_export_update_ctx(req->rq_export);
1657 /* Keep the rq_export around so we can send the reply. */
1658 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1663 void target_destroy_export(struct obd_export *exp)
1665 struct obd_import *imp = NULL;
1667 * exports created from last_rcvd data, and "fake"
1668 * exports created by lctl don't have an import
1670 spin_lock(&exp->exp_lock);
1671 if (exp->exp_imp_reverse != NULL) {
1672 imp = exp->exp_imp_reverse;
1673 exp->exp_imp_reverse = NULL;
1675 spin_unlock(&exp->exp_lock);
1677 client_destroy_import(imp);
1679 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1680 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1681 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1682 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1684 EXPORT_SYMBOL(target_destroy_export);
1687 * Recovery functions
1689 static void target_request_copy_get(struct ptlrpc_request *req)
1691 class_export_rpc_inc(req->rq_export);
1692 LASSERT(list_empty(&req->rq_list));
1693 INIT_LIST_HEAD(&req->rq_replay_list);
1695 /* Increase refcount to keep request in queue. */
1696 atomic_inc(&req->rq_refcount);
1697 /* Let export know it has replays to be handled. */
1698 atomic_inc(&req->rq_export->exp_replay_count);
1701 static void target_request_copy_put(struct ptlrpc_request *req)
1703 LASSERT(list_empty(&req->rq_replay_list));
1704 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1706 atomic_dec(&req->rq_export->exp_replay_count);
1707 class_export_rpc_dec(req->rq_export);
1708 ptlrpc_server_drop_request(req);
1711 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1713 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1714 struct obd_export *exp = req->rq_export;
1715 struct ptlrpc_request *reqiter;
1716 struct ptlrpc_request *dup_req = NULL;
1721 spin_lock(&exp->exp_lock);
1722 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1724 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1732 /* We expect it with RESENT and REPLAY flags. */
1733 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1734 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1735 CERROR("invalid flags %x of resent replay\n",
1736 lustre_msg_get_flags(req->rq_reqmsg));
1738 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1741 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1743 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1744 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1748 list_add_tail(&req->rq_replay_list,
1749 &exp->exp_req_replay_queue);
1752 spin_unlock(&exp->exp_lock);
1756 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1758 LASSERT(!list_empty(&req->rq_replay_list));
1759 LASSERT(req->rq_export);
1761 spin_lock(&req->rq_export->exp_lock);
1762 list_del_init(&req->rq_replay_list);
1763 spin_unlock(&req->rq_export->exp_lock);
1766 static void target_finish_recovery(struct lu_target *lut)
1768 struct obd_device *obd = lut->lut_obd;
1772 /* Only log a recovery message when recovery has occurred. */
1773 if (obd->obd_recovery_start) {
1774 time64_t now = ktime_get_seconds();
1775 time64_t elapsed_time;
1777 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1779 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1780 obd->obd_name, elapsed_time / 60,
1782 atomic_read(&obd->obd_max_recoverable_clients),
1783 atomic_read(&obd->obd_connected_clients),
1784 obd->obd_stale_clients,
1785 obd->obd_stale_clients == 1 ? "was" : "were");
1786 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1787 libcfs_debug_dumplog();
1790 ldlm_reprocess_recovery_done(obd->obd_namespace);
1791 spin_lock(&obd->obd_recovery_task_lock);
1792 if (!list_empty(&obd->obd_req_replay_queue) ||
1793 !list_empty(&obd->obd_lock_replay_queue) ||
1794 !list_empty(&obd->obd_final_req_queue)) {
1795 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1797 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1798 list_empty(&obd->obd_lock_replay_queue) ? \
1800 list_empty(&obd->obd_final_req_queue) ? \
1802 spin_unlock(&obd->obd_recovery_task_lock);
1805 spin_unlock(&obd->obd_recovery_task_lock);
1807 obd->obd_recovery_end = ktime_get_seconds();
1809 /* When recovery finished, cleanup orphans on MDS and OST. */
1810 if (obd->obd_type && OBP(obd, postrecov)) {
1811 int rc = OBP(obd, postrecov)(obd);
1814 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1820 static void abort_req_replay_queue(struct obd_device *obd)
1822 struct ptlrpc_request *req, *n;
1823 LIST_HEAD(abort_list);
1825 spin_lock(&obd->obd_recovery_task_lock);
1826 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1827 spin_unlock(&obd->obd_recovery_task_lock);
1828 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1829 DEBUG_REQ(D_WARNING, req, "aborted:");
1830 req->rq_status = -ENOTCONN;
1831 if (ptlrpc_error(req)) {
1832 DEBUG_REQ(D_ERROR, req,
1833 "failed abort_req_reply; skipping");
1835 target_exp_dequeue_req_replay(req);
1836 target_request_copy_put(req);
1840 static void abort_lock_replay_queue(struct obd_device *obd)
1842 struct ptlrpc_request *req, *n;
1843 LIST_HEAD(abort_list);
1845 spin_lock(&obd->obd_recovery_task_lock);
1846 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1847 spin_unlock(&obd->obd_recovery_task_lock);
1848 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1849 DEBUG_REQ(D_ERROR, req, "aborted:");
1850 req->rq_status = -ENOTCONN;
1851 if (ptlrpc_error(req)) {
1852 DEBUG_REQ(D_ERROR, req,
1853 "failed abort_lock_reply; skipping");
1855 target_request_copy_put(req);
1860 * Called from a cleanup function if the device is being cleaned up
1861 * forcefully. The exports should all have been disconnected already,
1862 * the only thing left to do is
1863 * - clear the recovery flags
1864 * - cancel the timer
1865 * - free queued requests and replies, but don't send replies
1866 * Because the obd_stopping flag is set, no new requests should be received.
1868 void target_cleanup_recovery(struct obd_device *obd)
1870 struct ptlrpc_request *req, *n;
1871 LIST_HEAD(clean_list);
1873 spin_lock(&obd->obd_dev_lock);
1874 if (!obd->obd_recovering) {
1875 spin_unlock(&obd->obd_dev_lock);
1879 obd->obd_recovering = 0;
1880 obd->obd_abort_recovery = 0;
1881 obd->obd_abort_mdt_recovery = 0;
1882 spin_unlock(&obd->obd_dev_lock);
1884 spin_lock(&obd->obd_recovery_task_lock);
1885 target_cancel_recovery_timer(obd);
1886 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1887 spin_unlock(&obd->obd_recovery_task_lock);
1889 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1890 LASSERT(req->rq_reply_state == NULL);
1891 target_exp_dequeue_req_replay(req);
1892 target_request_copy_put(req);
1895 spin_lock(&obd->obd_recovery_task_lock);
1896 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1897 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1898 spin_unlock(&obd->obd_recovery_task_lock);
1900 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1901 LASSERT(req->rq_reply_state == NULL);
1902 target_request_copy_put(req);
1907 EXPORT_SYMBOL(target_cleanup_recovery);
1909 /* obd_recovery_task_lock should be held */
1910 void target_cancel_recovery_timer(struct obd_device *obd)
1912 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1913 hrtimer_cancel(&obd->obd_recovery_timer);
1916 static void target_start_recovery_timer(struct obd_device *obd)
1920 if (obd->obd_recovery_start != 0)
1923 spin_lock(&obd->obd_dev_lock);
1924 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1925 spin_unlock(&obd->obd_dev_lock);
1929 LASSERT(obd->obd_recovery_timeout != 0);
1931 if (obd->obd_recovery_start != 0) {
1932 spin_unlock(&obd->obd_dev_lock);
1936 obd->obd_recovery_start = ktime_get_seconds();
1937 delay = ktime_set(obd->obd_recovery_start +
1938 obd->obd_recovery_timeout, 0);
1939 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1940 spin_unlock(&obd->obd_dev_lock);
1942 LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1944 obd->obd_recovery_timeout / 60,
1945 obd->obd_recovery_timeout % 60,
1946 atomic_read(&obd->obd_max_recoverable_clients),
1947 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1949 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1954 * extend recovery window.
1956 * if @extend is true, extend recovery window to have @dr_timeout remaining
1957 * at least; otherwise, make sure the recovery timeout value is not less
1960 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1967 spin_lock(&obd->obd_dev_lock);
1968 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1969 obd->obd_stopping) {
1970 spin_unlock(&obd->obd_dev_lock);
1973 LASSERT(obd->obd_recovery_start != 0);
1975 left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1976 left = ktime_divns(left_ns, NSEC_PER_SEC);
1979 timeout = obd->obd_recovery_timeout;
1980 /* dr_timeout will happen after the hrtimer has expired.
1981 * Add the excess time to the soft recovery timeout without
1982 * exceeding the hard recovery timeout.
1984 if (dr_timeout > left) {
1985 timeout += dr_timeout - left;
1986 timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1990 timeout = clamp_t(timeout_t, dr_timeout,
1991 obd->obd_recovery_timeout,
1992 obd->obd_recovery_time_hard);
1995 if (timeout == obd->obd_recovery_time_hard)
1996 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1997 obd->obd_name, timeout, extend);
1999 if (obd->obd_recovery_timeout < timeout) {
2002 obd->obd_recovery_timeout = timeout;
2003 end = ktime_set(obd->obd_recovery_start + timeout, 0);
2004 now = ktime_set(ktime_get_seconds(), 0);
2005 left_ns = ktime_sub(end, now);
2006 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
2007 left = ktime_divns(left_ns, NSEC_PER_SEC);
2009 spin_unlock(&obd->obd_dev_lock);
2011 CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
2012 obd->obd_name, left);
2015 /* Reset the timer with each new client connection */
2017 * This timer is actually reconnect_timer, which is for making sure
2018 * the total recovery window is at least as big as my reconnect
2019 * attempt timing. So the initial recovery time_out will be set to
2020 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
2021 * from client is bigger than this, then the recovery time_out will
2022 * be extended to make sure the client could be reconnected, in the
2023 * process, the timeout from the new client should be ignored.
2026 check_and_start_recovery_timer(struct obd_device *obd,
2027 struct ptlrpc_request *req,
2030 timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
2031 struct obd_device_target *obt = obd2obt(obd);
2033 if (!new_client && service_timeout)
2035 * Teach server about old server's estimates, as first guess
2036 * at how long new requests will take.
2038 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2041 target_start_recovery_timer(obd);
2044 * Convert the service time to RPC timeout,
2045 * and reuse service_timeout to limit stack usage.
2047 service_timeout = at_est2timeout(service_timeout);
2049 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2050 service_timeout < at_extra)
2051 service_timeout = at_extra;
2054 * We expect other clients to timeout within service_timeout, then try
2055 * to reconnect, then try the failover server. The max delay between
2056 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2058 service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2060 LASSERT(obt->obt_magic == OBT_MAGIC);
2061 service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2062 if (service_timeout > obd->obd_recovery_timeout && !new_client)
2063 extend_recovery_timer(obd, service_timeout, false);
2066 /** Health checking routines */
2067 static inline int exp_connect_healthy(struct obd_export *exp)
2069 return exp->exp_in_recovery;
2072 /** if export done req_replay or has replay in queue */
2073 static inline int exp_req_replay_healthy(struct obd_export *exp)
2075 return (!exp->exp_req_replay_needed ||
2076 atomic_read(&exp->exp_replay_count) > 0);
2080 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2082 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2083 exp_req_replay_healthy(exp);
2086 /** if export done lock_replay or has replay in queue */
2087 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2089 return (!exp->exp_lock_replay_needed ||
2090 atomic_read(&exp->exp_replay_count) > 0);
2093 static inline int exp_vbr_healthy(struct obd_export *exp)
2095 return !exp->exp_vbr_failed;
2098 static inline int exp_finished(struct obd_export *exp)
2100 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2103 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2105 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2109 static int check_for_next_transno(struct lu_target *lut)
2111 struct ptlrpc_request *req = NULL;
2112 struct obd_device *obd = lut->lut_obd;
2113 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2114 int wake_up = 0, connected, completed, queue_len;
2115 __u64 req_transno = 0;
2116 __u64 update_transno = 0;
2117 __u64 next_transno = 0;
2121 spin_lock(&obd->obd_recovery_task_lock);
2122 if (!list_empty(&obd->obd_req_replay_queue)) {
2123 req = list_entry(obd->obd_req_replay_queue.next,
2124 struct ptlrpc_request, rq_list);
2125 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2128 if (!obd_mdt_recovery_abort(obd) && tdtd)
2129 update_transno = distribute_txn_get_next_transno(tdtd);
2131 connected = atomic_read(&obd->obd_connected_clients);
2132 completed = connected - atomic_read(&obd->obd_req_replay_clients);
2133 queue_len = obd->obd_requests_queued_for_recovery;
2134 next_transno = obd->obd_next_recovery_transno;
2137 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2138 atomic_read(&obd->obd_max_recoverable_clients),
2139 connected, completed,
2140 queue_len, req_transno, next_transno);
2142 if (obd_recovery_abort(obd)) {
2143 CDEBUG(D_HA, "waking for aborted recovery\n");
2145 } else if (obd->obd_recovery_expired) {
2146 CDEBUG(D_HA, "waking for expired recovery\n");
2148 } else if (!obd_mdt_recovery_abort(obd) && tdtd && req &&
2149 is_req_replayed_by_update(req)) {
2150 LASSERTF(req_transno < next_transno,
2151 "req_transno %llu next_transno%llu\n", req_transno,
2153 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2156 } else if (req_transno == next_transno ||
2157 (update_transno != 0 && update_transno <= next_transno)) {
2158 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2160 } else if (queue_len > 0 &&
2161 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2162 /** handle gaps occured due to lost reply or VBR */
2163 LASSERTF(req_transno >= next_transno,
2164 "req_transno: %llu, next_transno: %llu\n",
2165 req_transno, next_transno);
2167 "%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",
2168 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2169 next_transno, queue_len, completed, connected,
2170 req_transno, update_transno, obd->obd_last_committed);
2171 obd->obd_next_recovery_transno = req_transno;
2173 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2174 CDEBUG(D_HA, "waking for completed recovery\n");
2176 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2178 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2180 obd->obd_next_recovery_transno = req_transno;
2183 spin_unlock(&obd->obd_recovery_task_lock);
2187 static int check_for_next_lock(struct lu_target *lut)
2189 struct obd_device *obd = lut->lut_obd;
2192 spin_lock(&obd->obd_recovery_task_lock);
2193 if (!list_empty(&obd->obd_lock_replay_queue)) {
2194 CDEBUG(D_HA, "waking for next lock\n");
2196 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2197 CDEBUG(D_HA, "waking for completed lock replay\n");
2199 } else if (obd->obd_abort_recovery) {
2200 CDEBUG(D_HA, "waking for aborted recovery\n");
2202 } else if (obd->obd_recovery_expired) {
2203 CDEBUG(D_HA, "waking for expired recovery\n");
2206 spin_unlock(&obd->obd_recovery_task_lock);
2211 static int check_update_llog(struct lu_target *lut)
2213 struct obd_device *obd = lut->lut_obd;
2214 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2216 if (obd_mdt_recovery_abort(obd)) {
2217 CDEBUG(D_HA, "waking for aborted recovery\n");
2221 if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2222 CDEBUG(D_HA, "waking for completion of reading update log\n");
2230 * wait for recovery events,
2231 * check its status with help of check_routine
2232 * evict dead clients via health_check
2234 static int target_recovery_overseer(struct lu_target *lut,
2235 int (*check_routine)(struct lu_target *),
2236 int (*health_check)(struct obd_export *))
2238 struct obd_device *obd = lut->lut_obd;
2239 struct target_distribute_txn_data *tdtd;
2243 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2244 now = ktime_get_seconds();
2245 if (now - last > 600) {
2246 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2251 if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2252 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2253 __u64 next_update_transno = 0;
2256 * Only abort the recovery if there are no update recovery
2259 spin_lock(&obd->obd_recovery_task_lock);
2260 if (!obd_mdt_recovery_abort(obd) && lut->lut_tdtd) {
2261 next_update_transno =
2262 distribute_txn_get_next_transno(lut->lut_tdtd);
2264 tdtd = lut->lut_tdtd;
2266 * If next_update_transno == 0, it probably because
2267 * updatelog retrieve threads did not get any records
2268 * yet, let's wait those threads stopped
2270 if (next_update_transno == 0) {
2271 spin_unlock(&obd->obd_recovery_task_lock);
2273 while (wait_event_timeout(
2274 tdtd->tdtd_recovery_threads_waitq,
2275 check_update_llog(lut),
2276 cfs_time_seconds(60)) == 0);
2278 spin_lock(&obd->obd_recovery_task_lock);
2279 next_update_transno =
2280 distribute_txn_get_next_transno(tdtd);
2284 if (next_update_transno != 0 && !obd_recovery_abort(obd)) {
2285 obd->obd_next_recovery_transno = next_update_transno;
2286 spin_unlock(&obd->obd_recovery_task_lock);
2288 * Disconnect unfinished exports from clients, and
2289 * keep connection from MDT to make sure the update
2290 * recovery will still keep trying until some one
2291 * manually abort the recovery
2293 class_disconnect_stale_exports(obd,
2294 exp_finished_or_from_mdt);
2295 /* Abort all of replay & replay lock req from clients */
2296 abort_req_replay_queue(obd);
2297 abort_lock_replay_queue(obd);
2299 "%s: there are still update replay (%#llx)in the queue.\n",
2300 obd->obd_name, next_update_transno);
2302 obd->obd_abort_recovery = 1;
2303 spin_unlock(&obd->obd_recovery_task_lock);
2304 CWARN("%s recovery is aborted by hard timeout\n",
2309 while (wait_event_timeout(obd->obd_next_transno_waitq,
2311 cfs_time_seconds(60)) == 0)
2312 ; /* wait indefinitely for event, but don't trigger watchdog */
2314 if (obd_recovery_abort(obd)) {
2315 CWARN("recovery is aborted, evict exports in recovery\n");
2316 if (lut->lut_tdtd != NULL) {
2317 tdtd = lut->lut_tdtd;
2319 * Let's wait all of the update log recovery thread
2323 tdtd->tdtd_recovery_threads_waitq,
2324 atomic_read(&tdtd->tdtd_recovery_threads_count)
2326 /* Then abort the update recovery list */
2327 dtrq_list_destroy(lut->lut_tdtd);
2330 /** evict exports which didn't finish recovery yet */
2331 class_disconnect_stale_exports(obd, exp_finished);
2333 } else if (obd->obd_recovery_expired) {
2334 obd->obd_recovery_expired = 0;
2336 /** If some clients died being recovered, evict them */
2337 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2339 /** evict cexports with no replay in queue, they are stalled */
2340 class_disconnect_stale_exports(obd, health_check);
2342 /** continue with VBR */
2343 spin_lock(&obd->obd_dev_lock);
2344 obd->obd_version_recov = 1;
2345 spin_unlock(&obd->obd_dev_lock);
2347 * reset timer, recovery will proceed with versions now,
2348 * timeout is set just to handle reconnection delays
2350 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2352 * Wait for recovery events again, after evicting bad clients
2359 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2361 struct obd_device *obd = lut->lut_obd;
2362 struct ptlrpc_request *req = NULL;
2364 CDEBUG(D_HA, "Waiting for lock\n");
2365 if (target_recovery_overseer(lut, check_for_next_lock,
2366 exp_lock_replay_healthy))
2367 abort_lock_replay_queue(obd);
2369 spin_lock(&obd->obd_recovery_task_lock);
2370 if (!list_empty(&obd->obd_lock_replay_queue)) {
2371 req = list_entry(obd->obd_lock_replay_queue.next,
2372 struct ptlrpc_request, rq_list);
2373 list_del_init(&req->rq_list);
2374 spin_unlock(&obd->obd_recovery_task_lock);
2376 spin_unlock(&obd->obd_recovery_task_lock);
2377 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2378 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2379 /** evict exports failed VBR */
2380 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2385 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2387 struct ptlrpc_request *req = NULL;
2389 spin_lock(&obd->obd_recovery_task_lock);
2390 if (!list_empty(&obd->obd_final_req_queue)) {
2391 req = list_entry(obd->obd_final_req_queue.next,
2392 struct ptlrpc_request, rq_list);
2393 list_del_init(&req->rq_list);
2394 spin_unlock(&obd->obd_recovery_task_lock);
2395 if (req->rq_export->exp_in_recovery) {
2396 spin_lock(&req->rq_export->exp_lock);
2397 req->rq_export->exp_in_recovery = 0;
2398 spin_unlock(&req->rq_export->exp_lock);
2401 spin_unlock(&obd->obd_recovery_task_lock);
2406 static void handle_recovery_req(struct ptlrpc_thread *thread,
2407 struct ptlrpc_request *req,
2408 svc_handler_t handler)
2413 * export can be evicted during recovery, no need to handle replays for
2414 * it after that, discard such request silently
2416 if (req->rq_export->exp_disconnected)
2419 req->rq_session.lc_thread = thread;
2420 req->rq_svc_thread = thread;
2421 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2423 /* thread context */
2424 lu_context_enter(&thread->t_env->le_ctx);
2426 lu_context_exit(&thread->t_env->le_ctx);
2428 req->rq_svc_thread->t_env->le_ses = NULL;
2430 /* don't reset timer for final stage */
2431 if (!exp_finished(req->rq_export)) {
2432 timeout_t timeout = obd_timeout;
2435 * Add request @timeout to the recovery time so next request from
2436 * this client may come in recovery time
2439 struct ptlrpc_service_part *svcpt;
2440 timeout_t est_timeout;
2442 svcpt = req->rq_rqbd->rqbd_svcpt;
2444 * If the server sent early reply for this request,
2445 * the client will recalculate the timeout according to
2446 * current server estimate service time, so we will
2447 * use the maxium timeout here for waiting the client
2448 * sending the next req
2450 est_timeout = at_get(&svcpt->scp_at_estimate);
2451 timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2452 lustre_msg_get_timeout(req->rq_reqmsg));
2454 * Add 2 net_latency, one for balance rq_deadline
2455 * (see ptl_send_rpc), one for resend the req to server,
2456 * Note: client will pack net_latency in replay req
2457 * (see ptlrpc_replay_req)
2459 timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2461 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2467 /** Checking routines for recovery */
2468 static int check_for_recovery_ready(struct lu_target *lut)
2470 struct obd_device *obd = lut->lut_obd;
2471 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2474 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2475 clnts, obd->obd_stale_clients,
2476 atomic_read(&obd->obd_max_recoverable_clients),
2477 obd->obd_abort_recovery, obd->obd_recovery_expired);
2479 if (!obd_recovery_abort(obd) && !obd->obd_recovery_expired) {
2481 atomic_read(&obd->obd_max_recoverable_clients));
2482 if (clnts + obd->obd_stale_clients <
2483 atomic_read(&obd->obd_max_recoverable_clients))
2487 if (!obd_mdt_recovery_abort(obd) && lut->lut_tdtd &&
2488 !lut->lut_tdtd->tdtd_replay_ready) {
2489 /* Let's extend recovery timer, in case the recovery timer
2490 * expired, and some clients got evicted
2492 extend_recovery_timer(obd, obd->obd_recovery_timeout, true);
2494 "%s update recovery is not ready, extend recovery %d\n",
2495 obd->obd_name, obd->obd_recovery_timeout);
2503 REQUEST_RECOVERY = 1,
2504 UPDATE_RECOVERY = 2,
2507 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2511 if (!list_empty(&obd->obd_req_replay_queue)) {
2512 struct ptlrpc_request *req;
2514 req = list_entry(obd->obd_req_replay_queue.next,
2515 struct ptlrpc_request, rq_list);
2516 transno = lustre_msg_get_transno(req->rq_reqmsg);
2522 static __u64 get_next_transno(struct lu_target *lut, int *type)
2524 struct obd_device *obd = lut->lut_obd;
2525 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2527 __u64 update_transno;
2531 transno = get_next_replay_req_transno(obd);
2533 *type = REQUEST_RECOVERY;
2535 if (!tdtd || obd_mdt_recovery_abort(obd))
2538 update_transno = distribute_txn_get_next_transno(tdtd);
2539 if (transno == 0 || (transno >= update_transno &&
2540 update_transno != 0)) {
2541 transno = update_transno;
2543 *type = UPDATE_RECOVERY;
2550 * drop duplicate replay request
2552 * Because the operation has been replayed by update recovery, the request
2553 * with the same transno will be dropped and also notify the client to send
2554 * next replay request.
2556 * \param[in] env execution environment
2557 * \param[in] obd failover obd device
2558 * \param[in] req request to be dropped
2560 static void drop_duplicate_replay_req(struct lu_env *env,
2561 struct obd_device *obd,
2562 struct ptlrpc_request *req)
2564 DEBUG_REQ(D_HA, req,
2565 "remove t%lld from %s because duplicate update records found",
2566 lustre_msg_get_transno(req->rq_reqmsg),
2567 libcfs_nidstr(&req->rq_peer.nid));
2570 * Right now, only for MDS reint operation update replay and
2571 * normal request replay can have the same transno
2573 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2574 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2575 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2576 if (likely(req->rq_export))
2577 target_committed_to_req(req);
2578 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2579 target_send_reply(req, req->rq_status, 0);
2581 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2582 libcfs_nidstr(&req->rq_peer.nid));
2584 target_exp_dequeue_req_replay(req);
2585 target_request_copy_put(req);
2586 obd->obd_replayed_requests++;
2589 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2591 static void replay_request_or_update(struct lu_env *env,
2592 struct lu_target *lut,
2593 struct target_recovery_data *trd,
2594 struct ptlrpc_thread *thread)
2596 struct obd_device *obd = lut->lut_obd;
2597 struct ptlrpc_request *req = NULL;
2603 CDEBUG(D_HA, "Waiting for transno %lld\n",
2604 obd->obd_next_recovery_transno);
2606 /* Replay all of request and update by transno */
2608 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2610 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2613 * It is needed to extend recovery window above
2614 * recovery_time_soft. Extending is possible only in the
2615 * end of recovery window (see more details in
2616 * handle_recovery_req()).
2618 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2620 if (target_recovery_overseer(lut, check_for_next_transno,
2621 exp_req_replay_healthy_or_from_mdt)) {
2622 abort_req_replay_queue(obd);
2623 abort_lock_replay_queue(obd);
2627 spin_lock(&obd->obd_recovery_task_lock);
2628 transno = get_next_transno(lut, &type);
2629 if (type == REQUEST_RECOVERY && transno != 0) {
2631 * Drop replay request from client side, if the
2632 * replay has been executed by update with the
2635 req = list_entry(obd->obd_req_replay_queue.next,
2636 struct ptlrpc_request, rq_list);
2638 list_del_init(&req->rq_list);
2639 obd->obd_requests_queued_for_recovery--;
2640 spin_unlock(&obd->obd_recovery_task_lock);
2643 * Let's check if the request has been redone by
2646 if (is_req_replayed_by_update(req)) {
2647 struct distribute_txn_replay_req *dtrq;
2649 dtrq = distribute_txn_lookup_finish_list(tdtd,
2651 LASSERT(dtrq != NULL);
2652 spin_lock(&tdtd->tdtd_replay_list_lock);
2653 list_del_init(&dtrq->dtrq_list);
2654 spin_unlock(&tdtd->tdtd_replay_list_lock);
2657 drop_duplicate_replay_req(env, obd, req);
2662 LASSERT(trd->trd_processing_task == current->pid);
2663 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2665 lustre_msg_get_transno(req->rq_reqmsg),
2666 libcfs_nidstr(&req->rq_peer.nid));
2668 ptlrpc_watchdog_init(&thread->t_watchdog,
2670 handle_recovery_req(thread, req,
2671 trd->trd_recovery_handler);
2672 ptlrpc_watchdog_disable(&thread->t_watchdog);
2675 * bz18031: increase next_recovery_transno before
2676 * target_request_copy_put() will drop exp_rpc reference
2678 spin_lock(&obd->obd_recovery_task_lock);
2679 obd->obd_next_recovery_transno++;
2680 spin_unlock(&obd->obd_recovery_task_lock);
2681 target_exp_dequeue_req_replay(req);
2682 target_request_copy_put(req);
2683 obd->obd_replayed_requests++;
2684 } else if (type == UPDATE_RECOVERY && transno != 0) {
2685 struct distribute_txn_replay_req *dtrq;
2688 spin_unlock(&obd->obd_recovery_task_lock);
2690 LASSERT(tdtd != NULL);
2691 dtrq = distribute_txn_get_next_req(tdtd);
2692 lu_context_enter(&thread->t_env->le_ctx);
2693 ptlrpc_watchdog_init(&thread->t_watchdog,
2695 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2696 ptlrpc_watchdog_disable(&thread->t_watchdog);
2697 lu_context_exit(&thread->t_env->le_ctx);
2698 extend_recovery_timer(obd, obd_timeout, true);
2700 if (rc == 0 && dtrq->dtrq_xid != 0) {
2702 "Move x%llu t%llu to finish list\n",
2704 dtrq->dtrq_master_transno);
2706 /* Add it to the replay finish list */
2707 spin_lock(&tdtd->tdtd_replay_list_lock);
2708 list_add(&dtrq->dtrq_list,
2709 &tdtd->tdtd_replay_finish_list);
2710 spin_unlock(&tdtd->tdtd_replay_list_lock);
2712 spin_lock(&obd->obd_recovery_task_lock);
2713 if (transno == obd->obd_next_recovery_transno)
2714 obd->obd_next_recovery_transno++;
2716 obd->obd_next_recovery_transno)
2717 obd->obd_next_recovery_transno =
2719 spin_unlock(&obd->obd_recovery_task_lock);
2724 spin_unlock(&obd->obd_recovery_task_lock);
2726 LASSERT(list_empty(&obd->obd_req_replay_queue));
2727 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2728 /** evict exports failed VBR */
2729 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2735 static int target_recovery_thread(void *arg)
2737 struct lu_target *lut = arg;
2738 struct obd_device *obd = lut->lut_obd;
2739 struct ptlrpc_request *req;
2740 struct target_recovery_data *trd = &obd->obd_recovery_data;
2741 unsigned long delta;
2743 struct ptlrpc_thread *thread = NULL;
2747 unshare_fs_struct();
2748 OBD_ALLOC_PTR(thread);
2754 GOTO(out_thread, rc = -ENOMEM);
2755 rc = lu_env_add(env);
2759 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2761 GOTO(out_env_remove, rc);
2763 thread->t_env = env;
2764 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2765 thread->t_task = current;
2766 env->le_ctx.lc_thread = thread;
2767 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2769 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2771 trd->trd_processing_task = current->pid;
2773 spin_lock(&obd->obd_dev_lock);
2774 obd->obd_recovering = 1;
2775 spin_unlock(&obd->obd_dev_lock);
2776 complete(&trd->trd_starting);
2778 /* first of all, we have to know the first transno to replay */
2779 if (target_recovery_overseer(lut, check_for_recovery_ready,
2780 exp_connect_healthy)) {
2781 abort_req_replay_queue(obd);
2782 abort_lock_replay_queue(obd);
2783 if (lut->lut_tdtd != NULL)
2784 dtrq_list_destroy(lut->lut_tdtd);
2787 /* next stage: replay requests or update */
2789 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2790 atomic_read(&obd->obd_req_replay_clients),
2791 obd->obd_next_recovery_transno);
2792 replay_request_or_update(env, lut, trd, thread);
2795 * The second stage: replay locks
2797 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2798 atomic_read(&obd->obd_lock_replay_clients));
2799 while ((req = target_next_replay_lock(lut))) {
2800 LASSERT(trd->trd_processing_task == current->pid);
2801 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2802 libcfs_nidstr(&req->rq_peer.nid));
2803 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2804 req->rq_status = -ENODEV;
2805 target_request_copy_put(req);
2808 handle_recovery_req(thread, req,
2809 trd->trd_recovery_handler);
2810 target_request_copy_put(req);
2811 obd->obd_replayed_locks++;
2815 * The third stage: reply on final pings, at this moment all clients
2816 * must have request in final queue
2818 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2819 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2820 /** Update server last boot epoch */
2821 tgt_boot_epoch_update(lut);
2823 /* cancel update llogs upon recovery abort */
2824 if (obd->obd_abort_recovery || obd->obd_abort_mdt_recovery)
2825 OBP(obd, iocontrol)(OBD_IOC_LLOG_CANCEL, obd->obd_self_export,
2828 list_for_each_entry(req, &obd->obd_final_req_queue, rq_list) {
2830 * Because the waiting client can not send ping to server,
2831 * so we need refresh the last_request_time, to avoid the
2832 * export is being evicted
2834 ptlrpc_update_export_timer(req->rq_export, 0);
2838 * We drop recoverying flag to forward all new requests
2839 * to regular mds_handle() since now
2841 spin_lock(&obd->obd_dev_lock);
2842 obd->obd_recovering = 0;
2843 obd->obd_abort_recovery = 0;
2844 obd->obd_abort_mdt_recovery = 0;
2845 spin_unlock(&obd->obd_dev_lock);
2846 spin_lock(&obd->obd_recovery_task_lock);
2847 target_cancel_recovery_timer(obd);
2848 spin_unlock(&obd->obd_recovery_task_lock);
2849 while ((req = target_next_final_ping(obd))) {
2850 LASSERT(trd->trd_processing_task == current->pid);
2851 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2852 libcfs_nidstr(&req->rq_peer.nid));
2853 handle_recovery_req(thread, req,
2854 trd->trd_recovery_handler);
2855 target_request_copy_put(req);
2858 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2859 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2860 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2861 if (delta > OBD_RECOVERY_TIME_SOFT) {
2862 CWARN("too long recovery - read logs\n");
2863 libcfs_debug_dumplog();
2866 target_finish_recovery(lut);
2867 lu_context_fini(&env->le_ctx);
2868 trd->trd_processing_task = 0;
2869 complete_all(&trd->trd_finishing);
2870 tgt_io_thread_done(thread);
2876 OBD_FREE_PTR(thread);
2880 static int target_start_recovery_thread(struct lu_target *lut,
2881 svc_handler_t handler)
2883 struct obd_device *obd = lut->lut_obd;
2885 struct target_recovery_data *trd = &obd->obd_recovery_data;
2888 memset(trd, 0, sizeof(*trd));
2889 init_completion(&trd->trd_starting);
2890 init_completion(&trd->trd_finishing);
2891 trd->trd_recovery_handler = handler;
2893 rc = server_name2index(obd->obd_name, &index, NULL);
2897 if (!IS_ERR(kthread_run(target_recovery_thread,
2898 lut, "tgt_recover_%d", index))) {
2899 wait_for_completion(&trd->trd_starting);
2900 LASSERT(obd->obd_recovering != 0);
2908 void target_stop_recovery_thread(struct obd_device *obd)
2910 if (obd->obd_recovery_data.trd_processing_task > 0) {
2911 struct target_recovery_data *trd = &obd->obd_recovery_data;
2912 /** recovery can be done but postrecovery is not yet */
2913 spin_lock(&obd->obd_dev_lock);
2914 if (obd->obd_recovering) {
2915 CERROR("%s: Aborting recovery\n", obd->obd_name);
2916 obd->obd_abort_recovery = 1;
2917 wake_up(&obd->obd_next_transno_waitq);
2919 spin_unlock(&obd->obd_dev_lock);
2920 wait_for_completion(&trd->trd_finishing);
2923 EXPORT_SYMBOL(target_stop_recovery_thread);
2925 void target_recovery_fini(struct obd_device *obd)
2927 class_disconnect_exports(obd);
2928 target_stop_recovery_thread(obd);
2929 target_cleanup_recovery(obd);
2931 EXPORT_SYMBOL(target_recovery_fini);
2933 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2935 struct obd_device *obd = container_of(timer, struct obd_device,
2936 obd_recovery_timer);
2939 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2940 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2941 ktime_get_seconds() - obd->obd_recovery_start,
2942 atomic_read(&obd->obd_connected_clients));
2944 obd->obd_recovery_expired = 1;
2945 wake_up(&obd->obd_next_transno_waitq);
2946 return HRTIMER_NORESTART;
2949 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2951 struct obd_device *obd = lut->lut_obd;
2953 if (lut->lut_bottom->dd_rdonly)
2956 if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2957 /** Update server last boot epoch */
2958 tgt_boot_epoch_update(lut);
2962 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2963 "last_transno %llu\n", obd->obd_name,
2964 atomic_read(&obd->obd_max_recoverable_clients),
2965 obd->obd_last_committed);
2966 LASSERT(obd->obd_stopping == 0);
2967 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2968 obd->obd_recovery_start = 0;
2969 obd->obd_recovery_end = 0;
2971 hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2973 obd->obd_recovery_timer.function = &target_recovery_expired;
2974 target_start_recovery_thread(lut, handler);
2976 EXPORT_SYMBOL(target_recovery_init);
2978 static int target_process_req_flags(struct obd_device *obd,
2979 struct ptlrpc_request *req)
2981 struct obd_export *exp = req->rq_export;
2983 LASSERT(exp != NULL);
2984 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2985 /* client declares he's ready to replay locks */
2986 spin_lock(&exp->exp_lock);
2987 if (exp->exp_req_replay_needed) {
2988 exp->exp_req_replay_needed = 0;
2989 spin_unlock(&exp->exp_lock);
2991 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2992 atomic_dec(&obd->obd_req_replay_clients);
2994 spin_unlock(&exp->exp_lock);
2997 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2999 * client declares he's ready to complete recovery
3000 * so, we put the request on th final queue
3002 spin_lock(&exp->exp_lock);
3003 if (exp->exp_lock_replay_needed) {
3004 exp->exp_lock_replay_needed = 0;
3005 spin_unlock(&exp->exp_lock);
3007 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
3008 atomic_dec(&obd->obd_lock_replay_clients);
3010 spin_unlock(&exp->exp_lock);
3016 int target_queue_recovery_request(struct ptlrpc_request *req,
3017 struct obd_device *obd)
3019 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
3020 struct ptlrpc_request *reqiter;
3025 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
3026 /* Processing the queue right now, don't re-add. */
3030 target_process_req_flags(obd, req);
3032 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
3033 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
3034 if (cfs_fail_val == 1) {
3037 wake_up(&cfs_race_waitq);
3039 schedule_timeout_interruptible(
3040 cfs_time_seconds(1));
3045 * client declares he's ready to complete recovery
3046 * so, we put the request on th final queue
3048 target_request_copy_get(req);
3049 DEBUG_REQ(D_HA, req, "queue final req");
3050 wake_up(&obd->obd_next_transno_waitq);
3051 spin_lock(&obd->obd_recovery_task_lock);
3052 if (obd->obd_recovering) {
3053 struct ptlrpc_request *tmp;
3054 struct ptlrpc_request *duplicate = NULL;
3056 if (likely(!req->rq_export->exp_replay_done)) {
3057 req->rq_export->exp_replay_done = 1;
3058 list_add_tail(&req->rq_list,
3059 &obd->obd_final_req_queue);
3060 spin_unlock(&obd->obd_recovery_task_lock);
3065 * XXX O(n), but only happens if final ping is
3066 * timed out, probably reorganize the list as
3069 list_for_each_entry_safe(reqiter, tmp,
3070 &obd->obd_final_req_queue,
3072 if (reqiter->rq_export == req->rq_export) {
3073 list_del_init(&reqiter->rq_list);
3074 duplicate = reqiter;
3079 list_add_tail(&req->rq_list,
3080 &obd->obd_final_req_queue);
3081 req->rq_export->exp_replay_done = 1;
3082 spin_unlock(&obd->obd_recovery_task_lock);
3084 if (duplicate != NULL) {
3085 DEBUG_REQ(D_HA, duplicate,
3086 "put prev final req");
3087 target_request_copy_put(duplicate);
3091 spin_unlock(&obd->obd_recovery_task_lock);
3092 target_request_copy_put(req);
3093 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3096 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3097 /* client declares he's ready to replay locks */
3098 target_request_copy_get(req);
3099 DEBUG_REQ(D_HA, req, "queue lock replay req");
3100 wake_up(&obd->obd_next_transno_waitq);
3101 spin_lock(&obd->obd_recovery_task_lock);
3102 LASSERT(obd->obd_recovering);
3103 /* usually due to recovery abort */
3104 if (!req->rq_export->exp_in_recovery) {
3105 spin_unlock(&obd->obd_recovery_task_lock);
3106 target_request_copy_put(req);
3109 LASSERT(req->rq_export->exp_lock_replay_needed);
3110 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3111 spin_unlock(&obd->obd_recovery_task_lock);
3116 * CAVEAT EMPTOR: The incoming request message has been swabbed
3117 * (i.e. buflens etc are in my own byte order), but type-dependent
3118 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3122 INIT_LIST_HEAD(&req->rq_list);
3123 DEBUG_REQ(D_HA, req, "not queueing");
3128 * If we're processing the queue, we want don't want to queue this
3131 * Also, if this request has a transno less than the one we're waiting
3132 * for, we should process it now. It could (and currently always will)
3133 * be an open request for a descriptor that was opened some time ago.
3135 * Also, a resent, replayed request that has already been
3136 * handled will pass through here and be processed immediately.
3139 "Next recovery transno: %llu, current: %llu, replaying\n",
3140 obd->obd_next_recovery_transno, transno);
3143 * If the request has been replayed by update replay, then sends this
3144 * request to the recovery thread (replay_request_or_update()), where
3145 * it will be handled
3147 spin_lock(&obd->obd_recovery_task_lock);
3148 if (transno < obd->obd_next_recovery_transno &&
3149 !is_req_replayed_by_update(req)) {
3150 /* Processing the queue right now, don't re-add. */
3151 LASSERT(list_empty(&req->rq_list));
3152 spin_unlock(&obd->obd_recovery_task_lock);
3155 spin_unlock(&obd->obd_recovery_task_lock);
3157 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3160 target_request_copy_get(req);
3161 if (!req->rq_export->exp_in_recovery) {
3162 target_request_copy_put(req);
3165 LASSERT(req->rq_export->exp_req_replay_needed);
3167 if (target_exp_enqueue_req_replay(req)) {
3168 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3169 target_request_copy_put(req);
3174 spin_lock(&obd->obd_recovery_task_lock);
3175 LASSERT(obd->obd_recovering);
3176 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3177 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3178 list_add_tail(&req->rq_list, &reqiter->rq_list);
3183 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3185 DEBUG_REQ(D_ERROR, req,
3186 "dropping replay: transno has been claimed by another client");
3187 spin_unlock(&obd->obd_recovery_task_lock);
3188 target_exp_dequeue_req_replay(req);
3189 target_request_copy_put(req);
3195 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3197 obd->obd_requests_queued_for_recovery++;
3198 spin_unlock(&obd->obd_recovery_task_lock);
3199 wake_up(&obd->obd_next_transno_waitq);
3203 void target_committed_to_req(struct ptlrpc_request *req)
3205 struct obd_export *exp = req->rq_export;
3207 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3208 lustre_msg_set_last_committed(req->rq_repmsg,
3209 exp->exp_last_committed);
3211 DEBUG_REQ(D_IOCTL, req,
3212 "not sending last_committed update (%d/%d)",
3213 exp->exp_obd->obd_no_transno,
3214 req->rq_repmsg == NULL);
3216 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3217 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3220 #endif /* HAVE_SERVER_SUPPORT */
3223 * Packs current SLV and Limit into \a req.
3225 int target_pack_pool_reply(struct ptlrpc_request *req)
3227 struct obd_device *obd;
3232 * Check that we still have all structures alive as this may
3233 * be some late RPC at shutdown time.
3235 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3236 !exp_connect_lru_resize(req->rq_export))) {
3237 lustre_msg_set_slv(req->rq_repmsg, 0);
3238 lustre_msg_set_limit(req->rq_repmsg, 0);
3242 /* OBD is alive here as export is alive, which we checked above. */
3243 obd = req->rq_export->exp_obd;
3245 read_lock(&obd->obd_pool_lock);
3246 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3247 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3248 read_unlock(&obd->obd_pool_lock);
3253 static int target_send_reply_msg(struct ptlrpc_request *req,
3254 int rc, int fail_id)
3256 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3257 DEBUG_REQ(D_ERROR, req, "dropping reply");
3261 * We can have a null rq_reqmsg in the event of bad signature or
3262 * no context when unwrapping
3264 if (req->rq_reqmsg &&
3265 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3266 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3270 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3271 req->rq_status = rc;
3272 return ptlrpc_send_error(req, 1);
3274 DEBUG_REQ(D_NET, req, "sending reply");
3276 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3279 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3281 struct ptlrpc_service_part *svcpt;
3283 struct ptlrpc_reply_state *rs;
3284 struct obd_export *exp;
3288 if (req->rq_no_reply) {
3293 svcpt = req->rq_rqbd->rqbd_svcpt;
3294 rs = req->rq_reply_state;
3295 if (rs == NULL || !rs->rs_difficult) {
3297 target_send_reply_msg(req, rc, fail_id);
3302 /* must be an export if locks saved */
3303 LASSERT(req->rq_export != NULL);
3304 /* req/reply consistent */
3305 LASSERT(rs->rs_svcpt == svcpt);
3308 LASSERT(!rs->rs_scheduled);
3309 LASSERT(!rs->rs_scheduled_ever);
3310 LASSERT(!rs->rs_handled);
3311 LASSERT(!rs->rs_sent);
3312 LASSERT(!rs->rs_unlinked);
3313 LASSERT(rs->rs_export == NULL);
3314 LASSERT(list_empty(&rs->rs_obd_list));
3315 LASSERT(list_empty(&rs->rs_exp_list));
3317 exp = class_export_get(req->rq_export);
3319 /* disable reply scheduling while I'm setting up */
3320 rs->rs_scheduled = 1;
3322 rs->rs_unlinked = 0;
3323 rs->rs_xid = req->rq_xid;
3324 rs->rs_transno = req->rq_transno;
3325 rs->rs_export = exp;
3326 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3328 spin_lock(&exp->exp_uncommitted_replies_lock);
3329 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3330 rs->rs_transno, exp->exp_last_committed);
3331 if (rs->rs_transno > exp->exp_last_committed) {
3332 /* not committed already */
3333 list_add_tail(&rs->rs_obd_list,
3334 &exp->exp_uncommitted_replies);
3336 spin_unlock(&exp->exp_uncommitted_replies_lock);
3338 spin_lock(&exp->exp_lock);
3339 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3340 spin_unlock(&exp->exp_lock);
3342 netrc = target_send_reply_msg(req, rc, fail_id);
3344 spin_lock(&svcpt->scp_rep_lock);
3346 atomic_inc(&svcpt->scp_nreps_difficult);
3350 * error sending: reply is off the net. Also we need +1
3351 * reply ref until ptlrpc_handle_rs() is done
3352 * with the reply state (if the send was successful, there
3353 * would have been +1 ref for the net, which
3354 * reply_out_callback leaves alone)
3357 rs->rs_unlinked = 1;
3358 ptlrpc_rs_addref(rs);
3361 spin_lock(&rs->rs_lock);
3362 if (rs->rs_transno <= exp->exp_last_committed ||
3363 (rs->rs_unlinked && !rs->rs_no_ack) ||
3364 list_empty(&rs->rs_exp_list) || /* completed already */
3365 list_empty(&rs->rs_obd_list)) {
3366 CDEBUG(D_HA, "Schedule reply immediately\n");
3367 ptlrpc_dispatch_difficult_reply(rs);
3369 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3370 rs->rs_scheduled = 0; /* allow notifier to schedule */
3372 spin_unlock(&rs->rs_lock);
3373 spin_unlock(&svcpt->scp_rep_lock);
3377 enum ldlm_mode lck_compat_array[] = {
3378 [LCK_EX] = LCK_COMPAT_EX,
3379 [LCK_PW] = LCK_COMPAT_PW,
3380 [LCK_PR] = LCK_COMPAT_PR,
3381 [LCK_CW] = LCK_COMPAT_CW,
3382 [LCK_CR] = LCK_COMPAT_CR,
3383 [LCK_NL] = LCK_COMPAT_NL,
3384 [LCK_GROUP] = LCK_COMPAT_GROUP,
3385 [LCK_COS] = LCK_COMPAT_COS,
3389 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3390 * not escape to the user level.
3392 int ldlm_error2errno(enum ldlm_error error)
3398 case ELDLM_LOCK_MATCHED:
3401 case ELDLM_LOCK_CHANGED:
3404 case ELDLM_LOCK_ABORTED:
3407 case ELDLM_LOCK_REPLACED:
3410 case ELDLM_NO_LOCK_DATA:
3413 case ELDLM_NAMESPACE_EXISTS:
3416 case ELDLM_BAD_NAMESPACE:
3420 if (((int)error) < 0) { /* cast to signed type */
3421 result = error; /* as ldlm_error can be unsigned */
3423 CERROR("Invalid DLM result code: %d\n", error);
3429 EXPORT_SYMBOL(ldlm_error2errno);
3432 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3434 enum ldlm_error ldlm_errno2error(int err_no)
3443 error = ELDLM_LOCK_CHANGED;
3446 error = ELDLM_LOCK_ABORTED;
3449 error = ELDLM_LOCK_REPLACED;
3452 error = ELDLM_NO_LOCK_DATA;
3455 error = ELDLM_NAMESPACE_EXISTS;
3458 error = ELDLM_BAD_NAMESPACE;
3466 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3467 void ldlm_dump_export_locks(struct obd_export *exp)
3469 spin_lock(&exp->exp_locks_list_guard);
3470 if (!list_empty(&exp->exp_locks_list)) {
3471 struct ldlm_lock *lock;
3473 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3475 list_for_each_entry(lock, &exp->exp_locks_list,
3477 LDLM_ERROR(lock, "lock:");
3479 spin_unlock(&exp->exp_locks_list_guard);
3483 #ifdef HAVE_SERVER_SUPPORT
3484 static inline const char *bulk2type(struct ptlrpc_request *req)
3486 if (req->rq_bulk_read)
3488 if (req->rq_bulk_write)
3493 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3495 struct ptlrpc_request *req = desc->bd_req;
3496 time64_t start = ktime_get_seconds();
3502 /* If there is eviction in progress, wait for it to finish. */
3504 exp->exp_obd->obd_evict_inprogress_waitq,
3505 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3507 /* Check if client was evicted or reconnected already. */
3508 if (exp->exp_failed ||
3509 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3512 if (req->rq_bulk_read)
3513 rc = sptlrpc_svc_wrap_bulk(req, desc);
3515 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3516 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3517 else /* old version, bulk matchbits is rq_xid */
3518 req->rq_mbits = req->rq_xid;
3521 rc = ptlrpc_start_bulk_transfer(desc);
3525 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3526 bulk2type(req), rc);
3530 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3531 ptlrpc_abort_bulk(desc);
3535 /* limit actual bulk transfer to bulk_timeout seconds */
3536 deadline = start + bulk_timeout;
3537 if (deadline > req->rq_deadline)
3538 deadline = req->rq_deadline;
3541 time64_t timeoutl = deadline - ktime_get_seconds();
3542 time64_t rq_deadline;
3544 while (timeoutl >= 0 &&
3545 wait_event_idle_timeout(
3547 !ptlrpc_server_bulk_active(desc) ||
3550 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3551 timeoutl ? cfs_time_seconds(1) : 1) == 0)
3553 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3555 /* Wait again if we changed rq_deadline. */
3556 rq_deadline = READ_ONCE(req->rq_deadline);
3557 deadline = start + bulk_timeout;
3558 if (deadline > rq_deadline)
3559 deadline = rq_deadline;
3560 } while (rc == -ETIMEDOUT &&
3561 deadline > ktime_get_seconds());
3563 if (rc == -ETIMEDOUT) {
3564 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3565 bulk2type(req), deadline - start,
3566 ktime_get_real_seconds() - deadline);
3567 ptlrpc_abort_bulk(desc);
3568 } else if (exp->exp_failed) {
3569 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3572 ptlrpc_abort_bulk(desc);
3573 } else if (exp->exp_conn_cnt >
3574 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3575 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3577 /* We don't reply anyway. */
3579 ptlrpc_abort_bulk(desc);
3580 } else if (desc->bd_failure) {
3581 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3583 /* XXX should this be a different errno? */
3586 if (req->rq_bulk_write)
3587 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3588 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3589 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3590 bulk2type(req), desc->bd_nob_transferred,
3592 /* XXX should this be a different errno? */
3599 EXPORT_SYMBOL(target_bulk_io);
3601 #endif /* HAVE_SERVER_SUPPORT */