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 = imp->imp_conn_restricted_net;
69 if (!create && !priority) {
70 CDEBUG(D_HA, "Nothing to do\n");
74 /* refnet is used to restrict network connections */
75 if (refnet != LNET_NET_ANY)
76 CDEBUG(D_HA, "imp %s: restrict %s to %s net\n",
77 imp->imp_obd->obd_name, uuid->uuid,
78 libcfs_net2str(refnet));
80 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, refnet);
82 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
87 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
89 GOTO(out_put, rc = -ENOMEM);
92 spin_lock(&imp->imp_lock);
93 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
94 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
96 list_move(&item->oic_item,
98 item->oic_last_attempt = 0;
100 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
101 imp, imp->imp_obd->obd_name, uuid->uuid,
102 (priority ? ", moved to head" : ""));
103 spin_unlock(&imp->imp_lock);
104 GOTO(out_free, rc = 0);
107 /* No existing import connection found for \a uuid. */
109 imp_conn->oic_conn = ptlrpc_conn;
110 imp_conn->oic_uuid = *uuid;
111 imp_conn->oic_last_attempt = 0;
113 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
115 list_add_tail(&imp_conn->oic_item,
116 &imp->imp_conn_list);
117 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
118 imp, imp->imp_obd->obd_name, uuid->uuid,
119 (priority ? "head" : "tail"));
121 spin_unlock(&imp->imp_lock);
122 GOTO(out_free, rc = -ENOENT);
125 spin_unlock(&imp->imp_lock);
129 OBD_FREE(imp_conn, sizeof(*imp_conn));
131 ptlrpc_connection_put(ptlrpc_conn);
135 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
137 return import_set_conn(imp, uuid, 1, 0);
140 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
143 return import_set_conn(imp, uuid, priority, 1);
145 EXPORT_SYMBOL(client_import_add_conn);
147 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
148 struct lnet_nid *prim_nid, int priority)
150 struct ptlrpc_connection *ptlrpc_conn;
153 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, LNET_NID_NET(prim_nid));
155 const char *str_uuid = obd_uuid2str(uuid);
157 rc = class_add_uuid(str_uuid, prim_nid);
159 CERROR("%s: failed to add UUID '%s': rc = %d\n",
160 imp->imp_obd->obd_name, str_uuid, rc);
164 return import_set_conn(imp, uuid, priority, 1);
166 EXPORT_SYMBOL(client_import_dyn_add_conn);
168 int client_import_add_nids_to_conn(struct obd_import *imp,
169 struct lnet_nid *nidlist,
170 int nid_count, int nid_size,
171 struct obd_uuid *uuid)
173 struct obd_import_conn *conn;
177 if (nid_count <= 0 || !nidlist)
180 spin_lock(&imp->imp_lock);
181 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
182 if (class_check_uuid(&conn->oic_uuid, &nidlist[0])) {
183 *uuid = conn->oic_uuid;
184 spin_unlock(&imp->imp_lock);
185 rc = class_add_nids_to_uuid(&conn->oic_uuid, nidlist,
186 nid_count, nid_size);
190 spin_unlock(&imp->imp_lock);
193 EXPORT_SYMBOL(client_import_add_nids_to_conn);
195 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
197 struct obd_import_conn *imp_conn;
198 struct obd_export *dlmexp;
203 spin_lock(&imp->imp_lock);
204 if (list_empty(&imp->imp_conn_list)) {
205 LASSERT(!imp->imp_connection);
209 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
210 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
212 LASSERT(imp_conn->oic_conn);
214 if (imp_conn == imp->imp_conn_current) {
215 LASSERT(imp_conn->oic_conn == imp->imp_connection);
217 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
218 imp->imp_state != LUSTRE_IMP_DISCON) {
219 CERROR("can't remove current connection\n");
220 GOTO(out, rc = -EBUSY);
223 ptlrpc_connection_put(imp->imp_connection);
224 imp->imp_connection = NULL;
226 dlmexp = class_conn2export(&imp->imp_dlm_handle);
227 if (dlmexp && dlmexp->exp_connection) {
228 LASSERT(dlmexp->exp_connection ==
230 ptlrpc_connection_put(dlmexp->exp_connection);
231 dlmexp->exp_connection = NULL;
235 class_export_put(dlmexp);
238 list_del(&imp_conn->oic_item);
239 ptlrpc_connection_put(imp_conn->oic_conn);
240 OBD_FREE(imp_conn, sizeof(*imp_conn));
241 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
242 imp, imp->imp_obd->obd_name, uuid->uuid);
247 spin_unlock(&imp->imp_lock);
249 CERROR("connection %s not found\n", uuid->uuid);
252 EXPORT_SYMBOL(client_import_del_conn);
254 void client_destroy_import(struct obd_import *imp)
257 * Drop security policy instance after all RPCs have finished/aborted
258 * to let all busy contexts be released.
260 class_import_get(imp);
261 class_destroy_import(imp);
262 sptlrpc_import_sec_put(imp);
263 class_import_put(imp);
265 EXPORT_SYMBOL(client_destroy_import);
268 * Check whether or not the OSC is on MDT.
271 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
273 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
276 static int osc_on_mdt(char *obdname)
280 ptr = strrchr(obdname, '-');
284 if (strncmp(ptr + 1, "MDT", 3) == 0)
291 * Configure an RPC client OBD device.
296 * 3 - inactive-on-startup
297 * 4 - restrictive net
299 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
301 struct client_obd *cli = &obd->u.cli;
302 struct obd_import *imp;
303 struct obd_uuid server_uuid;
304 int rq_portal, rp_portal, connect_op;
305 const char *name = obd->obd_type->typ_name;
306 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
307 char *cli_name = lustre_cfg_buf(lcfg, 0);
313 * In a more perfect world, we would hang a ptlrpc_client off of
314 * obd_type and just use the values from there.
316 if (!strcmp(name, LUSTRE_OSC_NAME)) {
317 rq_portal = OST_REQUEST_PORTAL;
318 rp_portal = OSC_REPLY_PORTAL;
319 connect_op = OST_CONNECT;
320 cli->cl_sp_me = LUSTRE_SP_CLI;
321 cli->cl_sp_to = LUSTRE_SP_OST;
322 ns_type = LDLM_NS_TYPE_OSC;
323 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
324 !strcmp(name, LUSTRE_LWP_NAME)) {
325 rq_portal = MDS_REQUEST_PORTAL;
326 rp_portal = MDC_REPLY_PORTAL;
327 connect_op = MDS_CONNECT;
328 if (is_lwp_on_ost(cli_name))
329 cli->cl_sp_me = LUSTRE_SP_OST;
330 else if (is_lwp_on_mdt(cli_name))
331 cli->cl_sp_me = LUSTRE_SP_MDT;
333 cli->cl_sp_me = LUSTRE_SP_CLI;
334 cli->cl_sp_to = LUSTRE_SP_MDT;
335 ns_type = LDLM_NS_TYPE_MDC;
336 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
337 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
338 /* OSP_on_MDT for other MDTs */
339 connect_op = MDS_CONNECT;
340 cli->cl_sp_to = LUSTRE_SP_MDT;
341 ns_type = LDLM_NS_TYPE_MDC;
342 rq_portal = OUT_PORTAL;
344 /* OSP on MDT for OST */
345 connect_op = OST_CONNECT;
346 cli->cl_sp_to = LUSTRE_SP_OST;
347 ns_type = LDLM_NS_TYPE_OSC;
348 rq_portal = OST_REQUEST_PORTAL;
350 rp_portal = OSC_REPLY_PORTAL;
351 cli->cl_sp_me = LUSTRE_SP_MDT;
352 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
353 rq_portal = MGS_REQUEST_PORTAL;
354 rp_portal = MGC_REPLY_PORTAL;
355 connect_op = MGS_CONNECT;
356 cli->cl_sp_me = LUSTRE_SP_MGC;
357 cli->cl_sp_to = LUSTRE_SP_MGS;
358 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
359 ns_type = LDLM_NS_TYPE_MGC;
361 CERROR("unknown client OBD type \"%s\", can't setup\n",
366 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
367 CERROR("requires a TARGET UUID\n");
371 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
372 CERROR("client UUID must be less than 38 characters\n");
376 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
377 CERROR("setup requires a SERVER UUID\n");
381 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
382 CERROR("target UUID must be less than 38 characters\n");
386 init_rwsem(&cli->cl_sem);
387 mutex_init(&cli->cl_mgc_mutex);
389 init_rwsem(&cli->cl_seq_rwsem);
390 cli->cl_conn_count = 0;
391 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
392 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
393 sizeof(server_uuid)));
395 cli->cl_dirty_pages = 0;
396 cli->cl_dirty_max_pages = 0;
397 cli->cl_avail_grant = 0;
398 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
400 * cl_dirty_max_pages may be changed at connect time in
401 * ptlrpc_connect_interpret().
403 client_adjust_max_dirty(cli);
404 init_waitqueue_head(&cli->cl_cache_waiters);
405 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
406 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
407 INIT_LIST_HEAD(&cli->cl_loi_write_list);
408 INIT_LIST_HEAD(&cli->cl_loi_read_list);
409 spin_lock_init(&cli->cl_loi_list_lock);
410 atomic_set(&cli->cl_pending_w_pages, 0);
411 atomic_set(&cli->cl_pending_r_pages, 0);
412 cli->cl_r_in_flight = 0;
413 cli->cl_w_in_flight = 0;
415 cli->cl_stats_init = ktime_get_real();
416 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
417 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
418 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
419 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
420 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
421 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
422 spin_lock_init(&cli->cl_batch_rpc_hist.oh_lock);
425 INIT_LIST_HEAD(&cli->cl_lru_osc);
426 atomic_set(&cli->cl_lru_shrinkers, 0);
427 atomic_long_set(&cli->cl_lru_busy, 0);
428 atomic_long_set(&cli->cl_lru_in_list, 0);
429 INIT_LIST_HEAD(&cli->cl_lru_list);
430 spin_lock_init(&cli->cl_lru_list_lock);
431 atomic_long_set(&cli->cl_unstable_count, 0);
432 INIT_LIST_HEAD(&cli->cl_shrink_list);
433 INIT_LIST_HEAD(&cli->cl_grant_chain);
435 INIT_LIST_HEAD(&cli->cl_flight_waiters);
436 cli->cl_rpcs_in_flight = 0;
438 init_waitqueue_head(&cli->cl_destroy_waitq);
439 atomic_set(&cli->cl_destroy_in_flight, 0);
442 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
443 cli->cl_preferred_cksum_type = 0;
444 #ifdef CONFIG_ENABLE_CHECKSUM
445 /* Turn on checksumming by default. */
446 cli->cl_checksum = 1;
448 * The supported checksum types will be worked out at connect time
449 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
452 cli->cl_cksum_type = cli->cl_supp_cksum_types;
454 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
457 * Set it to possible maximum size. It may be reduced by ocd_brw_size
458 * from OFD after connecting.
460 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
462 cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
465 * set cl_chunkbits default value to PAGE_SHIFT,
466 * it will be updated at OSC connection time.
468 cli->cl_chunkbits = PAGE_SHIFT;
470 if (!strcmp(name, LUSTRE_MDC_NAME)) {
471 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
472 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
473 cli->cl_max_rpcs_in_flight = 2;
474 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
475 cli->cl_max_rpcs_in_flight = 3;
476 } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
477 cli->cl_max_rpcs_in_flight = 4;
479 if (osc_on_mdt(obd->obd_name))
480 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
482 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
485 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
486 cli->cl_max_mod_rpcs_in_flight = 0;
487 cli->cl_mod_rpcs_in_flight = 0;
488 cli->cl_close_rpcs_in_flight = 0;
489 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
490 cli->cl_mod_rpcs_init = ktime_get_real();
491 cli->cl_mod_tag_bitmap = NULL;
493 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
495 if (connect_op == MDS_CONNECT) {
496 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
497 OBD_ALLOC(cli->cl_mod_tag_bitmap,
498 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
499 if (cli->cl_mod_tag_bitmap == NULL)
500 GOTO(err, rc = -ENOMEM);
505 CERROR("ldlm_get_ref failed: %d\n", rc);
509 ptlrpc_init_client(rq_portal, rp_portal, name,
510 &obd->obd_ldlm_client);
512 imp = class_new_import(obd);
514 GOTO(err_ldlm, rc = -ENOENT);
515 imp->imp_client = &obd->obd_ldlm_client;
516 imp->imp_connect_op = connect_op;
517 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
518 LUSTRE_CFG_BUFLEN(lcfg, 1));
519 class_import_put(imp);
521 if (lustre_cfg_buf(lcfg, 4)) {
522 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
524 if (refnet == LNET_NET_ANY) {
526 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
527 obd->obd_name, lustre_cfg_string(lcfg, 4),
529 GOTO(err_import, rc);
531 imp->imp_conn_restricted_net = refnet;
533 imp->imp_conn_restricted_net = LNET_NET_ANY;
536 rc = client_import_add_conn(imp, &server_uuid, 1);
538 CERROR("can't add initial connection\n");
539 GOTO(err_import, rc);
541 imp->imp_connection = NULL;
543 cli->cl_import = imp;
544 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
545 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
547 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
548 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
549 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
551 cli->cl_target_uuid.uuid);
552 spin_lock(&imp->imp_lock);
553 imp->imp_deactive = 1;
554 spin_unlock(&imp->imp_lock);
558 obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
559 LDLM_NAMESPACE_CLIENT,
560 LDLM_NAMESPACE_GREEDY,
562 if (IS_ERR(obd->obd_namespace)) {
563 rc = PTR_ERR(obd->obd_namespace);
564 CERROR("%s: unable to create client namespace: rc = %d\n",
566 obd->obd_namespace = NULL;
567 GOTO(err_import, rc);
573 class_destroy_import(imp);
577 if (cli->cl_mod_tag_bitmap != NULL)
578 OBD_FREE(cli->cl_mod_tag_bitmap,
579 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
580 cli->cl_mod_tag_bitmap = NULL;
584 EXPORT_SYMBOL(client_obd_setup);
586 int client_obd_cleanup(struct obd_device *obd)
588 struct client_obd *cli = &obd->u.cli;
592 ldlm_namespace_free_post(obd->obd_namespace);
593 obd->obd_namespace = NULL;
595 obd_cleanup_client_import(obd);
596 LASSERT(obd->u.cli.cl_import == NULL);
600 if (cli->cl_mod_tag_bitmap != NULL)
601 OBD_FREE(cli->cl_mod_tag_bitmap,
602 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
603 cli->cl_mod_tag_bitmap = NULL;
607 EXPORT_SYMBOL(client_obd_cleanup);
609 /* ->o_connect() method for client side (OSC and MDC and MGC) */
610 int client_connect_import(const struct lu_env *env,
611 struct obd_export **exp,
612 struct obd_device *obd, struct obd_uuid *cluuid,
613 struct obd_connect_data *data, void *localdata)
615 struct client_obd *cli = &obd->u.cli;
616 struct obd_import *imp = cli->cl_import;
617 struct obd_connect_data *ocd;
618 struct lustre_handle conn = { 0 };
624 down_write(&cli->cl_sem);
625 if (cli->cl_conn_count > 0)
626 GOTO(out_sem, rc = -EALREADY);
628 rc = class_connect(&conn, obd, cluuid);
632 cli->cl_conn_count++;
633 *exp = class_conn2export(&conn);
635 LASSERT(obd->obd_namespace);
637 spin_lock(&imp->imp_lock);
638 if (imp->imp_state == LUSTRE_IMP_CLOSED && imp->imp_deactive) {
639 /* need to reactivate import if trying to connect
640 * to a previously disconnected
642 imp->imp_deactive = 0;
643 imp->imp_invalid = 0;
645 spin_unlock(&imp->imp_lock);
647 imp->imp_dlm_handle = conn;
648 rc = ptlrpc_init_import(imp);
652 ocd = &imp->imp_connect_data;
655 imp->imp_connect_flags_orig = data->ocd_connect_flags;
656 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
659 rc = ptlrpc_connect_import(imp);
661 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
664 LASSERT(*exp != NULL && (*exp)->exp_connection);
667 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
668 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
669 data->ocd_connect_flags, ocd->ocd_connect_flags);
670 data->ocd_connect_flags = ocd->ocd_connect_flags;
671 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
674 ptlrpc_pinger_add_import(imp);
680 cli->cl_conn_count--;
681 class_disconnect(*exp);
685 up_write(&cli->cl_sem);
687 if (!rc && localdata && !cli->cl_cache) {
688 cli->cl_cache = (struct cl_client_cache *)localdata;
689 cl_cache_incref(cli->cl_cache);
690 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
692 /* add this osc into entity list */
693 LASSERT(list_empty(&cli->cl_lru_osc));
694 spin_lock(&cli->cl_cache->ccc_lru_lock);
695 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
696 spin_unlock(&cli->cl_cache->ccc_lru_lock);
701 EXPORT_SYMBOL(client_connect_import);
703 int client_disconnect_export(struct obd_export *exp)
705 struct obd_device *obd = class_exp2obd(exp);
706 struct client_obd *cli;
707 struct obd_import *imp;
713 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
714 exp, exp ? exp->exp_handle.h_cookie : -1);
719 imp = cli->cl_import;
721 down_write(&cli->cl_sem);
722 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
725 if (cli->cl_conn_count == 0) {
726 CERROR("disconnecting disconnected device (%s)\n",
728 GOTO(out_disconnect, rc = -EINVAL);
731 cli->cl_conn_count--;
732 if (cli->cl_conn_count != 0)
733 GOTO(out_disconnect, rc = 0);
736 * Mark import deactivated now, so we don't try to reconnect if any
737 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
738 * fully deactivate the import, or that would drop all requests.
740 spin_lock(&imp->imp_lock);
741 imp->imp_deactive = 1;
742 spin_unlock(&imp->imp_lock);
745 * Some non-replayable imports (MDS's OSCs) are pinged, so just
746 * delete it regardless. (It's safe to delete an import that was
749 (void)ptlrpc_pinger_del_import(imp);
751 if (obd->obd_namespace != NULL) {
752 /* obd_force == local only */
753 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
754 obd->obd_force ? LCF_LOCAL : 0, NULL);
755 ldlm_namespace_free_prior(obd->obd_namespace, imp,
760 * There's no need to hold sem while disconnecting an import,
761 * and it may actually cause deadlock in GSS.
763 up_write(&cli->cl_sem);
764 rc = ptlrpc_disconnect_import(imp, 0);
765 down_write(&cli->cl_sem);
767 ptlrpc_invalidate_import(imp);
773 * Use server style - class_disconnect should be always called for
776 err = class_disconnect(exp);
780 up_write(&cli->cl_sem);
784 EXPORT_SYMBOL(client_disconnect_export);
786 #ifdef HAVE_SERVER_SUPPORT
787 int server_disconnect_export(struct obd_export *exp)
793 /* Disconnect early so that clients can't keep using export. */
794 rc = class_disconnect(exp);
795 /* Close import to avoid sending any requests. */
796 if (exp->exp_imp_reverse)
797 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
799 ldlm_bl_thread_wakeup();
801 /* complete all outstanding replies */
802 spin_lock(&exp->exp_lock);
803 while (!list_empty(&exp->exp_outstanding_replies)) {
804 struct ptlrpc_reply_state *rs =
805 list_first_entry(&exp->exp_outstanding_replies,
806 struct ptlrpc_reply_state,
808 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
810 spin_lock(&svcpt->scp_rep_lock);
812 list_del_init(&rs->rs_exp_list);
814 spin_lock(&rs->rs_lock);
815 ptlrpc_schedule_difficult_reply(rs);
816 spin_unlock(&rs->rs_lock);
818 spin_unlock(&svcpt->scp_rep_lock);
820 spin_unlock(&exp->exp_lock);
824 EXPORT_SYMBOL(server_disconnect_export);
826 static inline int target_check_recovery_timer(struct obd_device *target)
831 if (!target->obd_recovering || target->obd_recovery_start == 0)
834 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
835 timeout = ktime_divns(remaining, NSEC_PER_SEC);
839 /* the recovery timer should expire, but it isn't triggered,
840 * it's better to abort the recovery of this target to speed up
841 * the recovery of the whole cluster. */
842 spin_lock(&target->obd_dev_lock);
843 if (target->obd_recovering) {
844 CERROR("%s: Aborting recovery\n", target->obd_name);
845 target->obd_abort_recovery = 1;
846 wake_up(&target->obd_next_transno_waitq);
848 spin_unlock(&target->obd_dev_lock);
853 * --------------------------------------------------------------------------
854 * from old lib/target.c
855 * --------------------------------------------------------------------------
857 static int target_handle_reconnect(struct lustre_handle *conn,
858 struct obd_export *exp,
859 struct obd_uuid *cluuid)
861 struct obd_device *target;
862 struct lustre_handle *hdl;
868 hdl = &exp->exp_imp_reverse->imp_remote_handle;
869 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
870 conn->cookie = exp->exp_handle.h_cookie;
872 "connect export for UUID '%s' at %p, cookie %#llx\n",
873 cluuid->uuid, exp, conn->cookie);
877 target = exp->exp_obd;
879 /* Might be a re-connect after a partition. */
880 if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
881 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",
883 obd_uuid2str(&exp->exp_client_uuid),
884 obd_export_nid2str(exp),
885 hdl->cookie, conn->cookie);
886 memset(conn, 0, sizeof(*conn));
888 * target_handle_connect() treats EALREADY and
889 * -EALREADY differently. -EALREADY is an error
890 * (same UUID, different handle).
895 if (!target->obd_recovering) {
896 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
897 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
898 obd_export_nid2str(exp));
899 GOTO(out_already, rc);
902 remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
903 timeout = ktime_divns(remaining, NSEC_PER_SEC);
905 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
907 obd_uuid2str(&exp->exp_client_uuid),
908 obd_export_nid2str(exp),
909 atomic_read(&target->obd_max_recoverable_clients),
910 timeout / 60, timeout % 60);
912 struct target_distribute_txn_data *tdtd;
917 target_check_recovery_timer(target);
919 tdtd = class_exp2tgt(exp)->lut_tdtd;
920 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
921 buf = tdtd->tdtd_show_update_logs_retrievers(
922 tdtd->tdtd_show_retrievers_cbdata,
926 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",
928 obd_uuid2str(&exp->exp_client_uuid),
929 obd_export_nid2str(exp), count,
930 buf ? buf : "unknown (not enough RAM)",
931 (abs(timeout) + target->obd_recovery_timeout) / 60,
932 (abs(timeout) + target->obd_recovery_timeout) % 60,
935 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",
936 target->obd_name, abs(timeout) / 60,
937 abs(timeout) % 60, target->obd_name);
944 conn->cookie = exp->exp_handle.h_cookie;
946 * target_handle_connect() treats EALREADY and
947 * -EALREADY differently. EALREADY means we are
948 * doing a valid reconnect from the same client.
954 check_and_start_recovery_timer(struct obd_device *obd,
955 struct ptlrpc_request *req, int new_client);
958 * update flags for import during reconnect process
960 static int rev_import_flags_update(struct obd_import *revimp,
961 struct ptlrpc_request *req)
964 struct obd_connect_data *data;
966 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
968 if (data->ocd_connect_flags & OBD_CONNECT_AT)
969 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
971 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
973 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
975 revimp->imp_connect_data = *data;
976 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
978 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
979 revimp->imp_client->cli_name,
980 libcfs_idstr(&req->rq_peer), rc);
988 * Allocate a new reverse import for an export.
990 * \retval -errno in case error hit
991 * \retval 0 if reverse import correctly init
993 int rev_import_init(struct obd_export *export)
995 struct obd_device *obd = export->exp_obd;
996 struct obd_import *revimp;
998 LASSERT(export->exp_imp_reverse == NULL);
1000 revimp = class_new_import(obd);
1004 revimp->imp_remote_handle.cookie = 0ULL;
1005 revimp->imp_client = &obd->obd_ldlm_client;
1006 revimp->imp_dlm_fake = 1;
1008 /* it is safe to connect import in new state as no sends possible */
1009 spin_lock(&export->exp_lock);
1010 export->exp_imp_reverse = revimp;
1011 spin_unlock(&export->exp_lock);
1012 class_import_put(revimp);
1016 EXPORT_SYMBOL(rev_import_init);
1019 * Handle reconnect for an export.
1021 * \param exp export to handle reconnect process
1022 * \param req client reconnect request
1024 * \retval -rc in case securitfy flavor can't be changed
1025 * \retval 0 in case none problems
1027 static int rev_import_reconnect(struct obd_export *exp,
1028 struct ptlrpc_request *req)
1030 struct obd_import *revimp = exp->exp_imp_reverse;
1031 struct lustre_handle *lh;
1034 /* avoid sending a request until import flags are changed */
1035 ptlrpc_import_enter_resend(revimp);
1037 ptlrpc_connection_put(revimp->imp_connection);
1040 * client from recovery don't have a handle so we need to take from
1041 * request. it may produce situation when wrong client connected
1042 * to recovery as we trust a client uuid
1044 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1045 revimp->imp_remote_handle = *lh;
1048 * unknown versions will be caught in
1049 * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1051 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1053 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1055 rc = rev_import_flags_update(revimp, req);
1058 * it is safe to still be in RECOVERY phase as we are not able
1059 * to setup correct security flavor so requests are not able to
1060 * be delivered correctly
1065 /* resend all rpc's via new connection */
1066 return ptlrpc_import_recovery_state_machine(revimp);
1069 int target_handle_connect(struct ptlrpc_request *req)
1071 struct obd_device *target = NULL;
1072 struct obd_export *export = NULL;
1074 * connect handle - filled from target_handle_reconnect in
1077 struct lustre_handle conn;
1078 struct lustre_handle *tmp;
1079 struct obd_uuid cluuid;
1084 bool mds_conn = false, lw_client = false, initial_conn = false;
1085 bool mds_mds_conn = false;
1086 bool new_mds_mds_conn = false;
1087 struct obd_connect_data *data, *tmpdata;
1089 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1090 int tmp_exp_old_falloc;
1092 struct ptlrpc_connection *pcon = NULL;
1093 bool reconnected = false;
1097 CFS_RACE(OBD_FAIL_TGT_CONN_RACE);
1099 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1101 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1102 GOTO(out, rc = -EINVAL);
1105 target = class_str2obd(str);
1107 deuuidify(str, NULL, &target_start, &target_len);
1108 LCONSOLE_ERROR_MSG(0x137,
1109 "%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",
1110 str, libcfs_nidstr(&req->rq_peer.nid));
1111 GOTO(out, rc = -ENODEV);
1114 spin_lock(&target->obd_dev_lock);
1116 target->obd_conn_inprogress++;
1118 if (target->obd_stopping || !target->obd_set_up) {
1119 spin_unlock(&target->obd_dev_lock);
1121 deuuidify(str, NULL, &target_start, &target_len);
1122 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1123 target_len, target_start,
1124 libcfs_nidstr(&req->rq_peer.nid),
1125 (target->obd_stopping ?
1126 "stopping" : "not set up"));
1127 GOTO(out, rc = -ENODEV);
1130 if (target->obd_no_conn) {
1131 spin_unlock(&target->obd_dev_lock);
1134 "%s: Temporarily refusing client connection from %s\n",
1135 target->obd_name, libcfs_nidstr(&req->rq_peer.nid));
1136 GOTO(out, rc = -EAGAIN);
1139 spin_unlock(&target->obd_dev_lock);
1141 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1143 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1144 GOTO(out, rc = -EINVAL);
1147 obd_str2uuid(&cluuid, str);
1149 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1151 GOTO(out, rc = -EPROTO);
1155 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1157 if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1158 GOTO(out, rc = -EPROTO);
1159 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1161 GOTO(out, rc = -EPROTO);
1163 rc = req_capsule_server_pack(&req->rq_pill);
1167 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1169 * Don't allow clients to connect that are using old 1.8 format
1170 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1171 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1172 * FULL20 flag should be set on all connections since 2.0, but no
1173 * longer affects behaviour.
1175 * Later this check will be disabled and the flag can be retired
1176 * completely once interop with 3.0 is no longer needed.
1178 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1179 GOTO(out, rc = -EPROTO);
1181 /* Old clients will have 'tmp_exp_old_falloc' as 1.
1182 * Newer clients (2.15) and beyond will have it set as 0
1184 tmp_exp_old_falloc =
1185 !!(data->ocd_connect_flags & OBD_CONNECT_OLD_FALLOC);
1187 CDEBUG(D_INFO, "%s: ocd_connect_flags: %#llx tmp_exp_old_falloc: %d\n",
1188 target->obd_name, data->ocd_connect_flags, tmp_exp_old_falloc);
1191 * Don't allow liblustre clients to connect.
1192 * - testing was disabled in v2_2_50_0-61-g6a75d65
1193 * - building was disabled in v2_5_58_0-28-g7277179
1194 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1195 * - clients were refused connect for version difference > 0.0.1.32
1197 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1198 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1199 GOTO(out, rc = -EPROTO);
1204 * Note: lw_client is needed in MDS-MDS failover during update log
1205 * processing, so we needs to allow lw_client to be connected at
1206 * anytime, instead of only the initial connection
1208 lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1210 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1211 initial_conn = true;
1212 mds_conn = OCD_HAS_FLAG(data, MDS);
1213 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1216 * OBD_CONNECT_MNE_SWAB is removed at 2.14
1217 * Checking OBD_CONNECT_FID can be removed in the future.
1219 * Via check OBD_CONNECT_FID, we can distinguish whether
1220 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1221 * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1224 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1225 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1226 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1227 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1228 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1229 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1232 * We do not support the MDT-MDT interoperations with
1233 * different version MDT because of protocol changes.
1235 if (unlikely(major != LUSTRE_MAJOR ||
1236 minor != LUSTRE_MINOR ||
1237 abs(patch - LUSTRE_PATCH) > 3)) {
1238 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1239 target->obd_name, LUSTRE_MAJOR,
1240 LUSTRE_MINOR, LUSTRE_PATCH,
1241 LUSTRE_FIX, major, minor, patch,
1242 OBD_OCD_VERSION_FIX(data->ocd_version),
1243 libcfs_nidstr(&req->rq_peer.nid),
1245 GOTO(out, rc = -EPROTO);
1250 /* lctl gets a backstage, all-access pass. */
1251 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1252 goto dont_check_exports;
1254 export = obd_uuid_lookup(target, &cluuid);
1258 /* We've found an export in the hash. */
1260 spin_lock(&export->exp_lock);
1262 if (export->exp_connecting) { /* b=9635, et. al. */
1263 spin_unlock(&export->exp_lock);
1264 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1265 export->exp_obd->obd_name, export,
1266 libcfs_nidstr(&req->rq_peer.nid));
1267 class_export_put(export);
1270 } else if ((mds_conn || (lw_client && initial_conn) ||
1271 OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
1272 spin_unlock(&export->exp_lock);
1273 if (!nid_same(&req->rq_peer.nid,
1274 &export->exp_connection->c_peer.nid)) {
1275 /* MDS or LWP reconnected after failover. */
1276 LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1278 lw_client ? "LWP" : "MDS",
1279 libcfs_nidstr(&req->rq_peer.nid),
1280 libcfs_nidstr(&export->exp_connection->c_peer.nid));
1282 /* New connection from the same NID. */
1283 LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1285 lw_client ? "LWP" : "MDS",
1286 libcfs_nidstr(&req->rq_peer.nid),
1287 OCD_HAS_FLAG(data, MDS_MDS) ?
1291 if (nid_same(&req->rq_peer.nid,
1292 &export->exp_connection->c_peer.nid) &&
1293 OCD_HAS_FLAG(data, MDS_MDS)) {
1295 * Because exports between MDTs will always be
1296 * kept, let's do not fail such export if they
1297 * come from the same NID, otherwise it might
1298 * cause eviction between MDTs, which might
1299 * cause namespace inconsistency
1301 spin_lock(&export->exp_lock);
1302 export->exp_connecting = 1;
1303 export->exp_conn_cnt = 0;
1304 spin_unlock(&export->exp_lock);
1305 conn.cookie = export->exp_handle.h_cookie;
1308 class_fail_export(export);
1309 class_export_put(export);
1313 } else if (export->exp_connection != NULL && initial_conn &&
1314 !nid_same(&req->rq_peer.nid,
1315 &export->exp_connection->c_peer.nid)) {
1316 spin_unlock(&export->exp_lock);
1317 /* In MDS failover we have static UUID but NID can change. */
1318 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1319 target->obd_name, cluuid.uuid,
1320 libcfs_nidstr(&req->rq_peer.nid),
1322 &export->exp_connection->c_peer.nid));
1324 class_export_put(export);
1326 } else if (CFS_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1328 spin_unlock(&export->exp_lock);
1331 export->exp_connecting = 1;
1332 spin_unlock(&export->exp_lock);
1333 LASSERT(export->exp_obd == target);
1335 rc = target_handle_reconnect(&conn, export, &cluuid);
1338 /* If we found an export, we already unlocked. */
1341 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1342 } else if (req->rq_export == NULL &&
1343 atomic_read(&export->exp_rpc_count) > 0) {
1344 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1345 target->obd_name, cluuid.uuid,
1346 libcfs_nidstr(&req->rq_peer.nid),
1347 refcount_read(&export->exp_handle.h_ref));
1348 GOTO(out, rc = -EBUSY);
1349 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1351 if (!strstr(cluuid.uuid, "mdt"))
1352 LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1353 target->obd_name, cluuid.uuid,
1354 libcfs_nidstr(&req->rq_peer.nid));
1355 GOTO(out, rc = -EALREADY);
1357 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1363 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1364 target->obd_name, cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1365 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1366 export, ktime_get_seconds(),
1367 export ? export->exp_last_request_time : 0);
1370 * If this is the first time a client connects, reset the recovery
1371 * timer. Discard lightweight connections which might be local.
1373 if (!lw_client && rc == 0 && target->obd_recovering)
1374 check_and_start_recovery_timer(target, req, export == NULL);
1377 * We want to handle EALREADY but *not* -EALREADY from
1378 * target_handle_reconnect(), return reconnection state in a flag.
1380 if (rc == EALREADY) {
1381 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1387 /* Tell the client if we support replayable requests. */
1388 if (target->obd_replayable)
1389 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1391 if (export == NULL) {
1392 /* allow lightweight connections during recovery */
1394 * allow "new" MDT to be connected during recovery, since we
1395 * need retrieve recovery update records from it
1397 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1398 struct hrtimer *timer = &target->obd_recovery_timer;
1407 connected = atomic_read(&target->obd_connected_clients);
1408 in_progress = atomic_read(&target->obd_lock_replay_clients);
1410 atomic_read(&target->obd_max_recoverable_clients);
1411 stale = target->obd_stale_clients;
1412 remaining = hrtimer_get_remaining(timer);
1413 left = ktime_divns(remaining, NSEC_PER_SEC);
1415 if (ktime_to_ns(remaining) > 0) {
1416 msg = "to recover in";
1419 msg = "already passed deadline";
1422 target_check_recovery_timer(target);
1425 LCONSOLE_WARN("%s: Denying connection for new client %s (at %s), waiting for %d known clients (%d recovered, %d in progress, and %d evicted) %s %lld:%.02lld\n",
1426 target->obd_name, cluuid.uuid,
1427 libcfs_nidstr(&req->rq_peer.nid), known,
1428 connected - in_progress, in_progress,
1429 stale, msg, timeout / 60, timeout % 60);
1433 rc = obd_connect(req->rq_svc_thread->t_env,
1434 &export, target, &cluuid, data,
1436 if (mds_conn && CFS_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1437 lustre_msg_add_op_flags(req->rq_repmsg,
1438 MSG_CONNECT_RECOVERING);
1440 conn.cookie = export->exp_handle.h_cookie;
1441 rc = rev_import_init(export);
1445 new_mds_mds_conn = true;
1448 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_CONNECT_VS_EVICT)) {
1449 class_export_get(export);
1450 class_fail_export(export);
1451 class_export_put(export);
1453 rc = obd_reconnect(req->rq_svc_thread->t_env,
1454 export, target, &cluuid, data,
1462 data->ocd_instance = obd2obt(target)->obt_instance;
1465 * Return only the parts of obd_connect_data that we understand, so the
1466 * client knows that we don't understand the rest.
1469 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1471 tmpdata = req_capsule_server_get(&req->rq_pill,
1474 * Don't use struct assignment here, because the client reply
1475 * buffer may be smaller/larger than the local struct
1478 memcpy(tmpdata, data, min(tmpsize, size));
1482 * If the client and the server are the same node, we will already
1483 * have an export that really points to the client's DLM export,
1484 * because we have a shared handles table.
1486 * XXX this will go away when shaver stops sending the "connect" handle
1487 * in the real "remote handle" field of the request --phik 24 Apr 2003
1489 ptlrpc_request_change_export(req, export);
1491 pcon = ptlrpc_connection_get(&req->rq_peer,
1492 &req->rq_self, &cluuid);
1494 GOTO(out, rc = -ENOTCONN);
1496 spin_lock(&export->exp_lock);
1498 if (export->exp_disconnected) {
1499 spin_unlock(&export->exp_lock);
1502 * for each connect called disconnect
1503 * should be called to cleanup stuff
1505 class_export_get(export);
1506 obd_disconnect(export);
1509 GOTO(out, rc = -ENODEV);
1511 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1512 spin_unlock(&export->exp_lock);
1514 "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1515 cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1516 export->exp_conn_cnt,
1517 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1519 GOTO(out, rc = -EALREADY);
1521 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1522 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1524 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1525 /* make 'tmp_exp_old_falloc' persistent by saving it into
1526 * server side export object(obd_export)
1528 export->exp_old_falloc = tmp_exp_old_falloc;
1531 /* Check to see if connection came from another NID. */
1532 if (export->exp_connection != NULL &&
1533 !nid_same(&export->exp_connection->c_peer.nid,
1534 &req->rq_peer.nid)) {
1535 obd_nid_del(export->exp_obd, export);
1536 ptlrpc_connection_put(export->exp_connection);
1537 export->exp_connection = NULL;
1540 if (export->exp_connection == NULL) {
1541 export->exp_connection = pcon;
1544 obd_nid_add(export->exp_obd, export);
1546 spin_unlock(&export->exp_lock);
1548 lustre_msg_set_handle(req->rq_repmsg, &conn);
1550 rc = rev_import_reconnect(export, req);
1554 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1556 __u64 transno = data->ocd_transno;
1558 spin_lock(&export->exp_lock);
1560 * possible race with class_disconnect_stale_exports,
1561 * export may be already in the eviction process
1563 if (export->exp_failed) {
1564 spin_unlock(&export->exp_lock);
1565 GOTO(out, rc = -ENODEV);
1567 export->exp_in_recovery = 1;
1568 export->exp_req_replay_needed = 1;
1569 export->exp_lock_replay_needed = 1;
1570 spin_unlock(&export->exp_lock);
1572 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1573 MSG_CONNECT_TRANSNO);
1574 if (has_transno && transno == 0)
1575 CWARN("Connect with zero transno!\n");
1577 if (has_transno && transno > 0 &&
1578 transno < target->obd_next_recovery_transno &&
1579 transno > target->obd_last_committed) {
1580 /* Another way is to use cmpxchg() to be lock-free. */
1581 spin_lock(&target->obd_recovery_task_lock);
1582 if (transno < target->obd_next_recovery_transno)
1583 target->obd_next_recovery_transno = transno;
1584 spin_unlock(&target->obd_recovery_task_lock);
1587 atomic_inc(&target->obd_req_replay_clients);
1588 atomic_inc(&target->obd_lock_replay_clients);
1590 * Note: MDS-MDS connection is allowed to be connected during
1591 * recovery, no matter if the exports needs to be recoveried.
1592 * Because we need retrieve updates logs from all other MDTs.
1593 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1594 * also needs to be increased to match other recovery checking
1597 if (new_mds_mds_conn)
1598 atomic_inc(&target->obd_max_recoverable_clients);
1600 if (atomic_inc_return(&target->obd_connected_clients) ==
1601 atomic_read(&target->obd_max_recoverable_clients))
1602 wake_up(&target->obd_next_transno_waitq);
1605 /* Tell the client we're in recovery, when client is involved in it. */
1606 if (target->obd_recovering && !lw_client)
1607 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1611 spin_lock(&export->exp_lock);
1612 export->exp_connecting = 0;
1613 spin_unlock(&export->exp_lock);
1615 class_export_put(export);
1617 if (target != NULL) {
1618 spin_lock(&target->obd_dev_lock);
1619 target->obd_conn_inprogress--;
1620 spin_unlock(&target->obd_dev_lock);
1621 class_decref(target, "find", current);
1624 ptlrpc_connection_put(pcon);
1625 req->rq_status = rc;
1629 int target_handle_disconnect(struct ptlrpc_request *req)
1635 rc = req_capsule_server_pack(&req->rq_pill);
1639 /* In case of target disconnect, updating sec ctx immediately is
1640 * required in order to record latest sequence number used.
1641 * Sequence is normally updated on export destroy, but this event
1642 * can occur too late, ie after a new target connect request has
1644 * Maintaining correct sequence when client connection becomes idle
1645 * ensures that GSS does not erroneously consider requests as replays.
1647 rc = sptlrpc_export_update_ctx(req->rq_export);
1651 /* Keep the rq_export around so we can send the reply. */
1652 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1657 void target_destroy_export(struct obd_export *exp)
1659 struct obd_import *imp = NULL;
1661 * exports created from last_rcvd data, and "fake"
1662 * exports created by lctl don't have an import
1664 spin_lock(&exp->exp_lock);
1665 if (exp->exp_imp_reverse != NULL) {
1666 imp = exp->exp_imp_reverse;
1667 exp->exp_imp_reverse = NULL;
1669 spin_unlock(&exp->exp_lock);
1671 client_destroy_import(imp);
1673 LASSERT(atomic_read(&exp->exp_locks_count) == 0);
1674 LASSERT(atomic_read(&exp->exp_rpc_count) == 0);
1675 LASSERT(atomic_read(&exp->exp_cb_count) == 0);
1676 LASSERT(atomic_read(&exp->exp_replay_count) == 0);
1678 EXPORT_SYMBOL(target_destroy_export);
1681 * Recovery functions
1683 static void target_request_copy_get(struct ptlrpc_request *req)
1685 class_export_rpc_inc(req->rq_export);
1686 LASSERT(list_empty(&req->rq_list));
1687 INIT_LIST_HEAD(&req->rq_replay_list);
1689 /* Increase refcount to keep request in queue. */
1690 atomic_inc(&req->rq_refcount);
1691 /* Let export know it has replays to be handled. */
1692 atomic_inc(&req->rq_export->exp_replay_count);
1695 static void target_request_copy_put(struct ptlrpc_request *req)
1697 LASSERT(list_empty(&req->rq_replay_list));
1698 LASSERT(atomic_read(&(req)->rq_export->exp_replay_count) > 0);
1700 atomic_dec(&req->rq_export->exp_replay_count);
1701 class_export_rpc_dec(req->rq_export);
1702 ptlrpc_server_drop_request(req);
1705 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1707 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1708 struct obd_export *exp = req->rq_export;
1709 struct ptlrpc_request *reqiter;
1710 struct ptlrpc_request *dup_req = NULL;
1715 spin_lock(&exp->exp_lock);
1716 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1718 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1726 /* We expect it with RESENT and REPLAY flags. */
1727 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1728 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1729 CERROR("invalid flags %x of resent replay\n",
1730 lustre_msg_get_flags(req->rq_reqmsg));
1732 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1735 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1737 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1738 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1742 list_add_tail(&req->rq_replay_list,
1743 &exp->exp_req_replay_queue);
1746 spin_unlock(&exp->exp_lock);
1750 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1752 LASSERT(!list_empty(&req->rq_replay_list));
1753 LASSERT(req->rq_export);
1755 spin_lock(&req->rq_export->exp_lock);
1756 list_del_init(&req->rq_replay_list);
1757 spin_unlock(&req->rq_export->exp_lock);
1760 static void target_finish_recovery(struct lu_target *lut)
1762 struct obd_device *obd = lut->lut_obd;
1766 /* Only log a recovery message when recovery has occurred. */
1767 if (obd->obd_recovery_start) {
1768 time64_t now = ktime_get_seconds();
1769 time64_t elapsed_time;
1771 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1773 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1774 obd->obd_name, elapsed_time / 60,
1776 atomic_read(&obd->obd_max_recoverable_clients),
1777 atomic_read(&obd->obd_connected_clients),
1778 obd->obd_stale_clients,
1779 obd->obd_stale_clients == 1 ? "was" : "were");
1780 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1781 libcfs_debug_dumplog();
1784 ldlm_reprocess_recovery_done(obd->obd_namespace);
1785 spin_lock(&obd->obd_recovery_task_lock);
1786 if (!list_empty(&obd->obd_req_replay_queue) ||
1787 !list_empty(&obd->obd_lock_replay_queue) ||
1788 !list_empty(&obd->obd_final_req_queue)) {
1789 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1791 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1792 list_empty(&obd->obd_lock_replay_queue) ? \
1794 list_empty(&obd->obd_final_req_queue) ? \
1796 spin_unlock(&obd->obd_recovery_task_lock);
1799 spin_unlock(&obd->obd_recovery_task_lock);
1801 obd->obd_recovery_end = ktime_get_seconds();
1803 /* When recovery finished, cleanup orphans on MDS and OST. */
1804 if (obd->obd_type && obd->obd_type->typ_dt_ops->o_postrecov) {
1805 int rc = obd->obd_type->typ_dt_ops->o_postrecov(obd);
1808 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1814 static void abort_req_replay_queue(struct obd_device *obd)
1816 struct ptlrpc_request *req, *n;
1817 LIST_HEAD(abort_list);
1819 spin_lock(&obd->obd_recovery_task_lock);
1820 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1821 spin_unlock(&obd->obd_recovery_task_lock);
1822 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1823 DEBUG_REQ(D_WARNING, req, "aborted:");
1824 req->rq_status = -ENOTCONN;
1825 if (ptlrpc_error(req)) {
1826 DEBUG_REQ(D_ERROR, req,
1827 "failed abort_req_reply; skipping");
1829 target_exp_dequeue_req_replay(req);
1830 target_request_copy_put(req);
1834 static void abort_lock_replay_queue(struct obd_device *obd)
1836 struct ptlrpc_request *req, *n;
1837 LIST_HEAD(abort_list);
1839 spin_lock(&obd->obd_recovery_task_lock);
1840 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1841 spin_unlock(&obd->obd_recovery_task_lock);
1842 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1843 DEBUG_REQ(D_ERROR, req, "aborted:");
1844 req->rq_status = -ENOTCONN;
1845 if (ptlrpc_error(req)) {
1846 DEBUG_REQ(D_ERROR, req,
1847 "failed abort_lock_reply; skipping");
1849 target_request_copy_put(req);
1854 * Called from a cleanup function if the device is being cleaned up
1855 * forcefully. The exports should all have been disconnected already,
1856 * the only thing left to do is
1857 * - clear the recovery flags
1858 * - cancel the timer
1859 * - free queued requests and replies, but don't send replies
1860 * Because the obd_stopping flag is set, no new requests should be received.
1862 void target_cleanup_recovery(struct obd_device *obd)
1864 struct ptlrpc_request *req, *n;
1865 LIST_HEAD(clean_list);
1867 spin_lock(&obd->obd_dev_lock);
1868 if (!obd->obd_recovering) {
1869 spin_unlock(&obd->obd_dev_lock);
1873 obd->obd_recovering = 0;
1874 obd->obd_abort_recovery = 0;
1875 obd->obd_abort_mdt_recovery = 0;
1876 spin_unlock(&obd->obd_dev_lock);
1878 spin_lock(&obd->obd_recovery_task_lock);
1879 target_cancel_recovery_timer(obd);
1880 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1881 spin_unlock(&obd->obd_recovery_task_lock);
1883 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1884 LASSERT(req->rq_reply_state == NULL);
1885 target_exp_dequeue_req_replay(req);
1886 target_request_copy_put(req);
1889 spin_lock(&obd->obd_recovery_task_lock);
1890 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1891 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1892 spin_unlock(&obd->obd_recovery_task_lock);
1894 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1895 LASSERT(req->rq_reply_state == NULL);
1896 target_request_copy_put(req);
1901 EXPORT_SYMBOL(target_cleanup_recovery);
1903 /* obd_recovery_task_lock should be held */
1904 void target_cancel_recovery_timer(struct obd_device *obd)
1906 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1907 hrtimer_cancel(&obd->obd_recovery_timer);
1910 static void target_start_recovery_timer(struct obd_device *obd)
1914 if (obd->obd_recovery_start != 0)
1917 spin_lock(&obd->obd_dev_lock);
1918 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1919 spin_unlock(&obd->obd_dev_lock);
1923 LASSERT(obd->obd_recovery_timeout != 0);
1925 if (obd->obd_recovery_start != 0) {
1926 spin_unlock(&obd->obd_dev_lock);
1930 obd->obd_recovery_start = ktime_get_seconds();
1931 delay = ktime_set(obd->obd_recovery_start +
1932 obd->obd_recovery_timeout, 0);
1933 hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1934 spin_unlock(&obd->obd_dev_lock);
1936 LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1938 obd->obd_recovery_timeout / 60,
1939 obd->obd_recovery_timeout % 60,
1940 atomic_read(&obd->obd_max_recoverable_clients),
1941 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1943 (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1948 * extend recovery window.
1950 * if @extend is true, extend recovery window to have @dr_timeout remaining
1951 * at least; otherwise, make sure the recovery timeout value is not less
1954 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1961 spin_lock(&obd->obd_dev_lock);
1962 if (!obd->obd_recovering || obd->obd_abort_recovery ||
1963 obd->obd_stopping) {
1964 spin_unlock(&obd->obd_dev_lock);
1967 LASSERT(obd->obd_recovery_start != 0);
1969 left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1970 left = ktime_divns(left_ns, NSEC_PER_SEC);
1973 timeout = obd->obd_recovery_timeout;
1974 /* dr_timeout will happen after the hrtimer has expired.
1975 * Add the excess time to the soft recovery timeout without
1976 * exceeding the hard recovery timeout.
1978 if (dr_timeout > left) {
1979 timeout += dr_timeout - left;
1980 timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1984 timeout = clamp_t(timeout_t, dr_timeout,
1985 obd->obd_recovery_timeout,
1986 obd->obd_recovery_time_hard);
1989 if (timeout == obd->obd_recovery_time_hard)
1990 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1991 obd->obd_name, timeout, extend);
1993 if (obd->obd_recovery_timeout < timeout) {
1996 obd->obd_recovery_timeout = timeout;
1997 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1998 now = ktime_set(ktime_get_seconds(), 0);
1999 left_ns = ktime_sub(end, now);
2000 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
2001 left = ktime_divns(left_ns, NSEC_PER_SEC);
2003 spin_unlock(&obd->obd_dev_lock);
2005 CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
2006 obd->obd_name, left);
2009 /* Reset the timer with each new client connection */
2011 * This timer is actually reconnect_timer, which is for making sure
2012 * the total recovery window is at least as big as my reconnect
2013 * attempt timing. So the initial recovery time_out will be set to
2014 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
2015 * from client is bigger than this, then the recovery time_out will
2016 * be extended to make sure the client could be reconnected, in the
2017 * process, the timeout from the new client should be ignored.
2020 check_and_start_recovery_timer(struct obd_device *obd,
2021 struct ptlrpc_request *req,
2024 timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
2025 struct obd_device_target *obt = obd2obt(obd);
2027 if (!new_client && service_timeout)
2029 * Teach server about old server's estimates, as first guess
2030 * at how long new requests will take.
2032 obd_at_measure(obd, &req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2035 target_start_recovery_timer(obd);
2038 * Convert the service time to RPC timeout,
2039 * and reuse service_timeout to limit stack usage.
2041 service_timeout = at_est2timeout(service_timeout);
2043 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2044 service_timeout < at_extra)
2045 service_timeout = at_extra;
2048 * We expect other clients to timeout within service_timeout, then try
2049 * to reconnect, then try the failover server. The max delay between
2050 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2052 service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2054 LASSERT(obt->obt_magic == OBT_MAGIC);
2055 service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2056 if (service_timeout > obd->obd_recovery_timeout && !new_client)
2057 extend_recovery_timer(obd, service_timeout, false);
2060 /** Health checking routines */
2061 static inline int exp_connect_healthy(struct obd_export *exp)
2063 return exp->exp_in_recovery;
2066 /** if export done req_replay or has replay in queue */
2067 static inline int exp_req_replay_healthy(struct obd_export *exp)
2069 return (!exp->exp_req_replay_needed ||
2070 atomic_read(&exp->exp_replay_count) > 0);
2074 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2076 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2077 exp_req_replay_healthy(exp);
2080 /** if export done lock_replay or has replay in queue */
2081 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2083 return (!exp->exp_lock_replay_needed ||
2084 atomic_read(&exp->exp_replay_count) > 0);
2087 static inline int exp_vbr_healthy(struct obd_export *exp)
2089 return !exp->exp_vbr_failed;
2092 static inline int exp_finished(struct obd_export *exp)
2094 return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2097 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2099 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2103 static int check_for_next_transno(struct lu_target *lut)
2105 struct ptlrpc_request *req = NULL;
2106 struct obd_device *obd = lut->lut_obd;
2107 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2108 int wake_up = 0, connected, completed, queue_len;
2109 __u64 req_transno = 0;
2110 __u64 update_transno = 0;
2111 __u64 next_transno = 0;
2115 spin_lock(&obd->obd_recovery_task_lock);
2116 if (!list_empty(&obd->obd_req_replay_queue)) {
2117 req = list_first_entry(&obd->obd_req_replay_queue,
2118 struct ptlrpc_request, rq_list);
2119 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2122 if (!obd_mdt_recovery_abort(obd) && tdtd)
2123 update_transno = distribute_txn_get_next_transno(tdtd);
2125 connected = atomic_read(&obd->obd_connected_clients);
2126 completed = connected - atomic_read(&obd->obd_req_replay_clients);
2127 queue_len = obd->obd_requests_queued_for_recovery;
2128 next_transno = obd->obd_next_recovery_transno;
2131 "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2132 atomic_read(&obd->obd_max_recoverable_clients),
2133 connected, completed,
2134 queue_len, req_transno, next_transno);
2136 if (obd_recovery_abort(obd)) {
2137 CDEBUG(D_HA, "waking for aborted recovery\n");
2139 } else if (obd->obd_recovery_expired) {
2140 CDEBUG(D_HA, "waking for expired recovery\n");
2142 } else if (!obd_mdt_recovery_abort(obd) && tdtd && req &&
2143 is_req_replayed_by_update(req)) {
2144 LASSERTF(req_transno < next_transno,
2145 "req_transno %llu next_transno%llu\n", req_transno,
2147 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2150 } else if (req_transno == next_transno ||
2151 (update_transno != 0 && update_transno <= next_transno)) {
2152 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2154 } else if (queue_len > 0 &&
2155 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2156 /** handle gaps occured due to lost reply or VBR */
2157 LASSERTF(req_transno >= next_transno,
2158 "req_transno: %llu, next_transno: %llu\n",
2159 req_transno, next_transno);
2161 "%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",
2162 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2163 next_transno, queue_len, completed, connected,
2164 req_transno, update_transno, obd->obd_last_committed);
2165 obd->obd_next_recovery_transno = req_transno;
2167 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2168 CDEBUG(D_HA, "waking for completed recovery\n");
2170 } else if (CFS_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2172 "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2174 obd->obd_next_recovery_transno = req_transno;
2177 spin_unlock(&obd->obd_recovery_task_lock);
2181 static int check_for_next_lock(struct lu_target *lut)
2183 struct obd_device *obd = lut->lut_obd;
2186 spin_lock(&obd->obd_recovery_task_lock);
2187 if (!list_empty(&obd->obd_lock_replay_queue)) {
2188 CDEBUG(D_HA, "waking for next lock\n");
2190 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2191 CDEBUG(D_HA, "waking for completed lock replay\n");
2193 } else if (obd->obd_abort_recovery) {
2194 CDEBUG(D_HA, "waking for aborted recovery\n");
2196 } else if (obd->obd_recovery_expired) {
2197 CDEBUG(D_HA, "waking for expired recovery\n");
2200 spin_unlock(&obd->obd_recovery_task_lock);
2205 static int check_update_llog(struct lu_target *lut)
2207 struct obd_device *obd = lut->lut_obd;
2208 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2210 if (obd_mdt_recovery_abort(obd)) {
2211 CDEBUG(D_HA, "waking for aborted recovery\n");
2215 if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2216 CDEBUG(D_HA, "waking for completion of reading update log\n");
2224 * wait for recovery events,
2225 * check its status with help of check_routine
2226 * evict dead clients via health_check
2228 static int target_recovery_overseer(struct lu_target *lut,
2229 int (*check_routine)(struct lu_target *),
2230 int (*health_check)(struct obd_export *))
2232 struct obd_device *obd = lut->lut_obd;
2233 struct target_distribute_txn_data *tdtd;
2237 if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2238 now = ktime_get_seconds();
2239 if (now - last > 600) {
2240 LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2245 if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2246 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2247 __u64 next_update_transno = 0;
2250 * Only abort the recovery if there are no update recovery
2253 spin_lock(&obd->obd_recovery_task_lock);
2254 if (!obd_mdt_recovery_abort(obd) && lut->lut_tdtd) {
2255 next_update_transno =
2256 distribute_txn_get_next_transno(lut->lut_tdtd);
2258 tdtd = lut->lut_tdtd;
2260 * If next_update_transno == 0, it probably because
2261 * updatelog retrieve threads did not get any records
2262 * yet, let's wait those threads stopped
2264 if (next_update_transno == 0) {
2265 spin_unlock(&obd->obd_recovery_task_lock);
2267 while (wait_event_timeout(
2268 tdtd->tdtd_recovery_threads_waitq,
2269 check_update_llog(lut),
2270 cfs_time_seconds(60)) == 0);
2272 spin_lock(&obd->obd_recovery_task_lock);
2273 next_update_transno =
2274 distribute_txn_get_next_transno(tdtd);
2278 if (next_update_transno != 0 && !obd_recovery_abort(obd)) {
2279 obd->obd_next_recovery_transno = next_update_transno;
2280 spin_unlock(&obd->obd_recovery_task_lock);
2282 * Disconnect unfinished exports from clients, and
2283 * keep connection from MDT to make sure the update
2284 * recovery will still keep trying until some one
2285 * manually abort the recovery
2287 class_disconnect_stale_exports(obd,
2288 exp_finished_or_from_mdt);
2289 /* Abort all of replay & replay lock req from clients */
2290 abort_req_replay_queue(obd);
2291 abort_lock_replay_queue(obd);
2293 "%s: there are still update replay (%#llx)in the queue.\n",
2294 obd->obd_name, next_update_transno);
2296 obd->obd_abort_recovery = 1;
2297 spin_unlock(&obd->obd_recovery_task_lock);
2298 CWARN("%s recovery is aborted by hard timeout\n",
2303 while (wait_event_timeout(obd->obd_next_transno_waitq,
2305 cfs_time_seconds(60)) == 0)
2306 ; /* wait indefinitely for event, but don't trigger watchdog */
2308 if (obd_recovery_abort(obd)) {
2309 CWARN("recovery is aborted, evict exports in recovery\n");
2310 if (lut->lut_tdtd != NULL) {
2311 tdtd = lut->lut_tdtd;
2313 * Let's wait all of the update log recovery thread
2317 tdtd->tdtd_recovery_threads_waitq,
2318 atomic_read(&tdtd->tdtd_recovery_threads_count)
2320 /* Then abort the update recovery list */
2321 dtrq_list_destroy(lut->lut_tdtd);
2324 /** evict exports which didn't finish recovery yet */
2325 class_disconnect_stale_exports(obd, exp_finished);
2327 } else if (obd->obd_recovery_expired) {
2328 obd->obd_recovery_expired = 0;
2330 /** If some clients died being recovered, evict them */
2331 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2333 /** evict cexports with no replay in queue, they are stalled */
2334 class_disconnect_stale_exports(obd, health_check);
2336 /** continue with VBR */
2337 spin_lock(&obd->obd_dev_lock);
2338 obd->obd_version_recov = 1;
2339 spin_unlock(&obd->obd_dev_lock);
2341 * reset timer, recovery will proceed with versions now,
2342 * timeout is set just to handle reconnection delays
2344 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2346 * Wait for recovery events again, after evicting bad clients
2353 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2355 struct obd_device *obd = lut->lut_obd;
2356 struct ptlrpc_request *req = NULL;
2358 CDEBUG(D_HA, "Waiting for lock\n");
2359 if (target_recovery_overseer(lut, check_for_next_lock,
2360 exp_lock_replay_healthy))
2361 abort_lock_replay_queue(obd);
2363 spin_lock(&obd->obd_recovery_task_lock);
2364 if (!list_empty(&obd->obd_lock_replay_queue)) {
2365 req = list_first_entry(&obd->obd_lock_replay_queue,
2366 struct ptlrpc_request, rq_list);
2367 list_del_init(&req->rq_list);
2368 spin_unlock(&obd->obd_recovery_task_lock);
2370 spin_unlock(&obd->obd_recovery_task_lock);
2371 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2372 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2373 /** evict exports failed VBR */
2374 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2379 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2381 struct ptlrpc_request *req = NULL;
2383 spin_lock(&obd->obd_recovery_task_lock);
2384 if (!list_empty(&obd->obd_final_req_queue)) {
2385 req = list_first_entry(&obd->obd_final_req_queue,
2386 struct ptlrpc_request, rq_list);
2387 list_del_init(&req->rq_list);
2388 spin_unlock(&obd->obd_recovery_task_lock);
2389 if (req->rq_export->exp_in_recovery) {
2390 spin_lock(&req->rq_export->exp_lock);
2391 req->rq_export->exp_in_recovery = 0;
2392 spin_unlock(&req->rq_export->exp_lock);
2395 spin_unlock(&obd->obd_recovery_task_lock);
2400 static void handle_recovery_req(struct ptlrpc_thread *thread,
2401 struct ptlrpc_request *req,
2402 svc_handler_t handler)
2407 * export can be evicted during recovery, no need to handle replays for
2408 * it after that, discard such request silently
2410 if (req->rq_export->exp_disconnected)
2413 req->rq_session.lc_thread = thread;
2414 req->rq_svc_thread = thread;
2415 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2417 /* thread context */
2418 lu_context_enter(&thread->t_env->le_ctx);
2420 lu_context_exit(&thread->t_env->le_ctx);
2422 req->rq_svc_thread->t_env->le_ses = NULL;
2424 /* don't reset timer for final stage */
2425 if (!exp_finished(req->rq_export)) {
2426 timeout_t timeout = obd_timeout;
2427 struct obd_device *obd = req->rq_export->exp_obd;
2430 * Add request @timeout to the recovery time so next request from
2431 * this client may come in recovery time
2433 if (!obd_at_off(obd)) {
2434 struct ptlrpc_service_part *svcpt;
2435 timeout_t est_timeout;
2437 svcpt = req->rq_rqbd->rqbd_svcpt;
2439 * If the server sent early reply for this request,
2440 * the client will recalculate the timeout according to
2441 * current server estimate service time, so we will
2442 * use the maxium timeout here for waiting the client
2443 * sending the next req
2445 est_timeout = obd_at_get(obd, &svcpt->scp_at_estimate);
2446 timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2447 lustre_msg_get_timeout(req->rq_reqmsg));
2449 * Add 2 net_latency, one for balance rq_deadline
2450 * (see ptl_send_rpc), one for resend the req to server,
2451 * Note: client will pack net_latency in replay req
2452 * (see ptlrpc_replay_req)
2454 timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2456 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2462 /** Checking routines for recovery */
2463 static int check_for_recovery_ready(struct lu_target *lut)
2465 struct obd_device *obd = lut->lut_obd;
2466 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2469 "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2470 clnts, obd->obd_stale_clients,
2471 atomic_read(&obd->obd_max_recoverable_clients),
2472 obd->obd_abort_recovery, obd->obd_recovery_expired);
2474 if (!obd_recovery_abort(obd) && !obd->obd_recovery_expired) {
2476 atomic_read(&obd->obd_max_recoverable_clients));
2477 if (clnts + obd->obd_stale_clients <
2478 atomic_read(&obd->obd_max_recoverable_clients))
2482 if (!obd_mdt_recovery_abort(obd) && lut->lut_tdtd &&
2483 !lut->lut_tdtd->tdtd_replay_ready) {
2484 /* Let's extend recovery timer, in case the recovery timer
2485 * expired, and some clients got evicted
2487 extend_recovery_timer(obd, obd->obd_recovery_timeout, true);
2489 "%s update recovery is not ready, extend recovery %d\n",
2490 obd->obd_name, obd->obd_recovery_timeout);
2498 REQUEST_RECOVERY = 1,
2499 UPDATE_RECOVERY = 2,
2502 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2506 if (!list_empty(&obd->obd_req_replay_queue)) {
2507 struct ptlrpc_request *req;
2509 req = list_first_entry(&obd->obd_req_replay_queue,
2510 struct ptlrpc_request, rq_list);
2511 transno = lustre_msg_get_transno(req->rq_reqmsg);
2517 static __u64 get_next_transno(struct lu_target *lut, int *type)
2519 struct obd_device *obd = lut->lut_obd;
2520 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2522 __u64 update_transno;
2526 transno = get_next_replay_req_transno(obd);
2528 *type = REQUEST_RECOVERY;
2530 if (!tdtd || obd_mdt_recovery_abort(obd))
2533 update_transno = distribute_txn_get_next_transno(tdtd);
2534 if (transno == 0 || (transno >= update_transno &&
2535 update_transno != 0)) {
2536 transno = update_transno;
2538 *type = UPDATE_RECOVERY;
2545 * drop duplicate replay request
2547 * Because the operation has been replayed by update recovery, the request
2548 * with the same transno will be dropped and also notify the client to send
2549 * next replay request.
2551 * \param[in] env execution environment
2552 * \param[in] obd failover obd device
2553 * \param[in] req request to be dropped
2555 * \retval true duplicate replay update is dropped
2556 * \retval false duplicate replay update is not dropped
2558 static bool drop_duplicate_replay_req(struct lu_env *env,
2559 struct obd_device *obd,
2560 struct ptlrpc_request *req)
2562 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
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 (opc == 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);
2580 } else if (opc == MDS_CLOSE) {
2581 DEBUG_REQ(D_HA, req, "duplicate close replay from %s\n",
2582 libcfs_nidstr(&req->rq_peer.nid));
2585 DEBUG_REQ(D_ERROR, req, "wrong opc %d from %s\n", opc,
2586 libcfs_nidstr(&req->rq_peer.nid));
2588 target_exp_dequeue_req_replay(req);
2589 target_request_copy_put(req);
2590 obd->obd_replayed_requests++;
2595 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2597 static void replay_request_or_update(struct lu_env *env,
2598 struct lu_target *lut,
2599 struct target_recovery_data *trd,
2600 struct ptlrpc_thread *thread)
2602 struct obd_device *obd = lut->lut_obd;
2603 struct ptlrpc_request *req = NULL;
2609 CDEBUG(D_HA, "Waiting for transno %lld\n",
2610 obd->obd_next_recovery_transno);
2612 /* Replay all of request and update by transno */
2614 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2616 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2619 * It is needed to extend recovery window above
2620 * recovery_time_soft. Extending is possible only in the
2621 * end of recovery window (see more details in
2622 * handle_recovery_req()).
2624 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2626 if (target_recovery_overseer(lut, check_for_next_transno,
2627 exp_req_replay_healthy_or_from_mdt)) {
2628 abort_req_replay_queue(obd);
2629 abort_lock_replay_queue(obd);
2633 spin_lock(&obd->obd_recovery_task_lock);
2634 transno = get_next_transno(lut, &type);
2635 if (type == REQUEST_RECOVERY && transno != 0) {
2636 bool update = false;
2638 * Drop replay request from client side, if the
2639 * replay has been executed by update with the
2642 req = list_first_entry(&obd->obd_req_replay_queue,
2643 struct ptlrpc_request, rq_list);
2645 list_del_init(&req->rq_list);
2646 obd->obd_requests_queued_for_recovery--;
2647 spin_unlock(&obd->obd_recovery_task_lock);
2650 * Let's check if the request has been redone by
2653 if (is_req_replayed_by_update(req)) {
2654 struct distribute_txn_replay_req *dtrq;
2656 dtrq = distribute_txn_lookup_finish_list(tdtd,
2658 LASSERT(dtrq != NULL);
2659 spin_lock(&tdtd->tdtd_replay_list_lock);
2660 list_del_init(&dtrq->dtrq_list);
2661 spin_unlock(&tdtd->tdtd_replay_list_lock);
2664 if (drop_duplicate_replay_req(env, obd, req))
2666 /* not dropped yet */
2670 LASSERT(trd->trd_processing_task == current->pid);
2671 DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2673 lustre_msg_get_transno(req->rq_reqmsg),
2674 libcfs_nidstr(&req->rq_peer.nid));
2676 ptlrpc_watchdog_init(&thread->t_watchdog,
2678 handle_recovery_req(thread, req,
2679 trd->trd_recovery_handler);
2680 ptlrpc_watchdog_disable(&thread->t_watchdog);
2683 * bz18031: increase next_recovery_transno before
2684 * target_request_copy_put() will drop exp_rpc reference
2687 spin_lock(&obd->obd_recovery_task_lock);
2688 obd->obd_next_recovery_transno++;
2689 spin_unlock(&obd->obd_recovery_task_lock);
2691 target_exp_dequeue_req_replay(req);
2692 target_request_copy_put(req);
2693 obd->obd_replayed_requests++;
2694 } else if (type == UPDATE_RECOVERY && transno != 0) {
2695 struct distribute_txn_replay_req *dtrq;
2698 spin_unlock(&obd->obd_recovery_task_lock);
2700 LASSERT(tdtd != NULL);
2701 dtrq = distribute_txn_get_next_req(tdtd);
2702 lu_context_enter(&thread->t_env->le_ctx);
2703 ptlrpc_watchdog_init(&thread->t_watchdog,
2705 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2706 ptlrpc_watchdog_disable(&thread->t_watchdog);
2707 lu_context_exit(&thread->t_env->le_ctx);
2708 extend_recovery_timer(obd, obd_timeout, true);
2710 if (rc == 0 && dtrq->dtrq_xid != 0) {
2712 "Move x%llu t%llu to finish list\n",
2714 dtrq->dtrq_master_transno);
2716 /* Add it to the replay finish list */
2717 spin_lock(&tdtd->tdtd_replay_list_lock);
2718 list_add(&dtrq->dtrq_list,
2719 &tdtd->tdtd_replay_finish_list);
2720 spin_unlock(&tdtd->tdtd_replay_list_lock);
2722 spin_lock(&obd->obd_recovery_task_lock);
2723 if (transno == obd->obd_next_recovery_transno)
2724 obd->obd_next_recovery_transno++;
2726 obd->obd_next_recovery_transno)
2727 obd->obd_next_recovery_transno =
2729 spin_unlock(&obd->obd_recovery_task_lock);
2734 spin_unlock(&obd->obd_recovery_task_lock);
2736 LASSERT(list_empty(&obd->obd_req_replay_queue));
2737 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2738 /** evict exports failed VBR */
2739 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2745 static int target_recovery_thread(void *arg)
2747 struct lu_target *lut = arg;
2748 struct obd_device *obd = lut->lut_obd;
2749 struct ptlrpc_request *req;
2750 struct target_recovery_data *trd = &obd->obd_recovery_data;
2751 unsigned long delta;
2753 struct ptlrpc_thread *thread = NULL;
2757 unshare_fs_struct();
2758 OBD_ALLOC_PTR(thread);
2764 GOTO(out_thread, rc = -ENOMEM);
2765 rc = lu_env_add(env);
2769 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2771 GOTO(out_env_remove, rc);
2773 thread->t_env = env;
2774 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2775 thread->t_task = current;
2776 env->le_ctx.lc_thread = thread;
2777 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2779 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2781 trd->trd_processing_task = current->pid;
2783 spin_lock(&obd->obd_dev_lock);
2784 obd->obd_recovering = 1;
2785 spin_unlock(&obd->obd_dev_lock);
2786 complete(&trd->trd_starting);
2788 /* first of all, we have to know the first transno to replay */
2789 if (target_recovery_overseer(lut, check_for_recovery_ready,
2790 exp_connect_healthy)) {
2791 abort_req_replay_queue(obd);
2792 abort_lock_replay_queue(obd);
2793 if (lut->lut_tdtd != NULL)
2794 dtrq_list_destroy(lut->lut_tdtd);
2797 /* next stage: replay requests or update */
2799 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2800 atomic_read(&obd->obd_req_replay_clients),
2801 obd->obd_next_recovery_transno);
2802 replay_request_or_update(env, lut, trd, thread);
2805 * The second stage: replay locks
2807 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2808 atomic_read(&obd->obd_lock_replay_clients));
2809 while ((req = target_next_replay_lock(lut))) {
2810 LASSERT(trd->trd_processing_task == current->pid);
2811 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2812 libcfs_nidstr(&req->rq_peer.nid));
2813 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2814 req->rq_status = -ENODEV;
2815 target_request_copy_put(req);
2818 handle_recovery_req(thread, req,
2819 trd->trd_recovery_handler);
2820 target_request_copy_put(req);
2821 obd->obd_replayed_locks++;
2825 * The third stage: reply on final pings, at this moment all clients
2826 * must have request in final queue
2828 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2829 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2830 /** Update server last boot epoch */
2831 tgt_boot_epoch_update(lut);
2833 /* cancel update llogs upon recovery abort */
2834 if (obd->obd_abort_recovery || obd->obd_abort_mdt_recovery)
2835 obd->obd_type->typ_dt_ops->o_iocontrol(OBD_IOC_LLOG_CANCEL,
2836 obd->obd_self_export,
2839 list_for_each_entry(req, &obd->obd_final_req_queue, rq_list) {
2841 * Because the waiting client can not send ping to server,
2842 * so we need refresh the last_request_time, to avoid the
2843 * export is being evicted
2845 ptlrpc_update_export_timer(req->rq_export, 0);
2849 * We drop recoverying flag to forward all new requests
2850 * to regular mds_handle() since now
2852 spin_lock(&obd->obd_dev_lock);
2853 obd->obd_recovering = 0;
2854 obd->obd_abort_recovery = 0;
2855 obd->obd_abort_mdt_recovery = 0;
2856 spin_unlock(&obd->obd_dev_lock);
2857 spin_lock(&obd->obd_recovery_task_lock);
2858 target_cancel_recovery_timer(obd);
2859 spin_unlock(&obd->obd_recovery_task_lock);
2860 while ((req = target_next_final_ping(obd))) {
2861 LASSERT(trd->trd_processing_task == current->pid);
2862 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2863 libcfs_nidstr(&req->rq_peer.nid));
2864 handle_recovery_req(thread, req,
2865 trd->trd_recovery_handler);
2866 target_request_copy_put(req);
2869 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2870 CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2871 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2872 if (delta > OBD_RECOVERY_TIME_SOFT) {
2873 CWARN("too long recovery - read logs\n");
2874 libcfs_debug_dumplog();
2877 target_finish_recovery(lut);
2878 lu_context_fini(&env->le_ctx);
2879 trd->trd_processing_task = 0;
2880 complete_all(&trd->trd_finishing);
2881 tgt_io_thread_done(thread);
2887 OBD_FREE_PTR(thread);
2891 static int target_start_recovery_thread(struct lu_target *lut,
2892 svc_handler_t handler)
2894 struct obd_device *obd = lut->lut_obd;
2896 struct target_recovery_data *trd = &obd->obd_recovery_data;
2899 memset(trd, 0, sizeof(*trd));
2900 init_completion(&trd->trd_starting);
2901 init_completion(&trd->trd_finishing);
2902 trd->trd_recovery_handler = handler;
2904 rc = server_name2index(obd->obd_name, &index, NULL);
2908 if (!IS_ERR(kthread_run(target_recovery_thread,
2909 lut, "tgt_recover_%d", index))) {
2910 wait_for_completion(&trd->trd_starting);
2911 LASSERT(obd->obd_recovering != 0);
2919 void target_stop_recovery_thread(struct obd_device *obd)
2921 if (obd->obd_recovery_data.trd_processing_task > 0) {
2922 struct target_recovery_data *trd = &obd->obd_recovery_data;
2923 /** recovery can be done but postrecovery is not yet */
2924 spin_lock(&obd->obd_dev_lock);
2925 if (obd->obd_recovering) {
2926 CERROR("%s: Aborting recovery\n", obd->obd_name);
2927 obd->obd_abort_recovery = 1;
2928 wake_up(&obd->obd_next_transno_waitq);
2930 spin_unlock(&obd->obd_dev_lock);
2931 wait_for_completion(&trd->trd_finishing);
2934 EXPORT_SYMBOL(target_stop_recovery_thread);
2936 void target_recovery_fini(struct obd_device *obd)
2938 class_disconnect_exports(obd);
2939 target_stop_recovery_thread(obd);
2940 target_cleanup_recovery(obd);
2942 EXPORT_SYMBOL(target_recovery_fini);
2944 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2946 struct obd_device *obd = container_of(timer, struct obd_device,
2947 obd_recovery_timer);
2950 "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2951 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2952 ktime_get_seconds() - obd->obd_recovery_start,
2953 atomic_read(&obd->obd_connected_clients));
2955 obd->obd_recovery_expired = 1;
2956 wake_up(&obd->obd_next_transno_waitq);
2957 return HRTIMER_NORESTART;
2960 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2962 struct obd_device *obd = lut->lut_obd;
2964 if (lut->lut_bottom->dd_rdonly)
2967 if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2968 /** Update server last boot epoch */
2969 tgt_boot_epoch_update(lut);
2973 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2974 "last_transno %llu\n", obd->obd_name,
2975 atomic_read(&obd->obd_max_recoverable_clients),
2976 obd->obd_last_committed);
2977 LASSERT(obd->obd_stopping == 0);
2978 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2979 obd->obd_recovery_start = 0;
2980 obd->obd_recovery_end = 0;
2982 hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2984 obd->obd_recovery_timer.function = &target_recovery_expired;
2985 target_start_recovery_thread(lut, handler);
2987 EXPORT_SYMBOL(target_recovery_init);
2989 static int target_process_req_flags(struct obd_device *obd,
2990 struct ptlrpc_request *req)
2992 struct obd_export *exp = req->rq_export;
2994 LASSERT(exp != NULL);
2995 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2996 /* client declares he's ready to replay locks */
2997 spin_lock(&exp->exp_lock);
2998 if (exp->exp_req_replay_needed) {
2999 exp->exp_req_replay_needed = 0;
3000 spin_unlock(&exp->exp_lock);
3002 LASSERT(atomic_read(&(obd)->obd_req_replay_clients) >
3004 atomic_dec(&obd->obd_req_replay_clients);
3006 spin_unlock(&exp->exp_lock);
3009 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
3011 * client declares he's ready to complete recovery
3012 * so, we put the request on th final queue
3014 spin_lock(&exp->exp_lock);
3015 if (exp->exp_lock_replay_needed) {
3016 exp->exp_lock_replay_needed = 0;
3017 spin_unlock(&exp->exp_lock);
3019 LASSERT(atomic_read(&(obd)->obd_lock_replay_clients) >
3021 atomic_dec(&obd->obd_lock_replay_clients);
3023 spin_unlock(&exp->exp_lock);
3029 int target_queue_recovery_request(struct ptlrpc_request *req,
3030 struct obd_device *obd)
3032 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
3033 struct ptlrpc_request *reqiter;
3038 if (obd->obd_recovery_data.trd_processing_task == current->pid) {
3039 /* Processing the queue right now, don't re-add. */
3043 target_process_req_flags(obd, req);
3045 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
3046 if (unlikely(CFS_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
3047 if (cfs_fail_val == 1) {
3050 wake_up(&cfs_race_waitq);
3052 schedule_timeout_interruptible(
3053 cfs_time_seconds(1));
3058 * client declares he's ready to complete recovery
3059 * so, we put the request on th final queue
3061 target_request_copy_get(req);
3062 DEBUG_REQ(D_HA, req, "queue final req");
3063 wake_up(&obd->obd_next_transno_waitq);
3064 spin_lock(&obd->obd_recovery_task_lock);
3065 if (obd->obd_recovering) {
3066 struct ptlrpc_request *tmp;
3067 struct ptlrpc_request *duplicate = NULL;
3069 if (likely(!req->rq_export->exp_replay_done)) {
3070 req->rq_export->exp_replay_done = 1;
3071 list_add_tail(&req->rq_list,
3072 &obd->obd_final_req_queue);
3073 spin_unlock(&obd->obd_recovery_task_lock);
3078 * XXX O(n), but only happens if final ping is
3079 * timed out, probably reorganize the list as
3082 list_for_each_entry_safe(reqiter, tmp,
3083 &obd->obd_final_req_queue,
3085 if (reqiter->rq_export == req->rq_export) {
3086 list_del_init(&reqiter->rq_list);
3087 duplicate = reqiter;
3092 list_add_tail(&req->rq_list,
3093 &obd->obd_final_req_queue);
3094 req->rq_export->exp_replay_done = 1;
3095 spin_unlock(&obd->obd_recovery_task_lock);
3097 if (duplicate != NULL) {
3098 DEBUG_REQ(D_HA, duplicate,
3099 "put prev final req");
3100 target_request_copy_put(duplicate);
3104 spin_unlock(&obd->obd_recovery_task_lock);
3105 target_request_copy_put(req);
3106 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3109 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3110 /* client declares he's ready to replay locks */
3111 target_request_copy_get(req);
3112 DEBUG_REQ(D_HA, req, "queue lock replay req");
3113 wake_up(&obd->obd_next_transno_waitq);
3114 spin_lock(&obd->obd_recovery_task_lock);
3115 LASSERT(obd->obd_recovering);
3116 /* usually due to recovery abort */
3117 if (!req->rq_export->exp_in_recovery) {
3118 spin_unlock(&obd->obd_recovery_task_lock);
3119 target_request_copy_put(req);
3122 LASSERT(req->rq_export->exp_lock_replay_needed);
3123 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3124 spin_unlock(&obd->obd_recovery_task_lock);
3129 * CAVEAT EMPTOR: The incoming request message has been swabbed
3130 * (i.e. buflens etc are in my own byte order), but type-dependent
3131 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3135 INIT_LIST_HEAD(&req->rq_list);
3136 DEBUG_REQ(D_HA, req, "not queueing");
3141 * If we're processing the queue, we want don't want to queue this
3144 * Also, if this request has a transno less than the one we're waiting
3145 * for, we should process it now. It could (and currently always will)
3146 * be an open request for a descriptor that was opened some time ago.
3148 * Also, a resent, replayed request that has already been
3149 * handled will pass through here and be processed immediately.
3152 "Next recovery transno: %llu, current: %llu, replaying\n",
3153 obd->obd_next_recovery_transno, transno);
3156 * If the request has been replayed by update replay, then sends this
3157 * request to the recovery thread (replay_request_or_update()), where
3158 * it will be handled
3160 spin_lock(&obd->obd_recovery_task_lock);
3161 if (transno < obd->obd_next_recovery_transno &&
3162 !is_req_replayed_by_update(req)) {
3163 /* Processing the queue right now, don't re-add. */
3164 LASSERT(list_empty(&req->rq_list));
3165 spin_unlock(&obd->obd_recovery_task_lock);
3168 spin_unlock(&obd->obd_recovery_task_lock);
3170 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3173 target_request_copy_get(req);
3174 if (!req->rq_export->exp_in_recovery) {
3175 target_request_copy_put(req);
3178 LASSERT(req->rq_export->exp_req_replay_needed);
3180 if (target_exp_enqueue_req_replay(req)) {
3181 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3182 target_request_copy_put(req);
3187 spin_lock(&obd->obd_recovery_task_lock);
3188 LASSERT(obd->obd_recovering);
3189 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3190 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3191 list_add_tail(&req->rq_list, &reqiter->rq_list);
3196 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3198 DEBUG_REQ(D_ERROR, req,
3199 "dropping replay: transno has been claimed by another client");
3200 spin_unlock(&obd->obd_recovery_task_lock);
3201 target_exp_dequeue_req_replay(req);
3202 target_request_copy_put(req);
3208 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3210 obd->obd_requests_queued_for_recovery++;
3211 spin_unlock(&obd->obd_recovery_task_lock);
3212 wake_up(&obd->obd_next_transno_waitq);
3216 void target_committed_to_req(struct ptlrpc_request *req)
3218 struct obd_export *exp = req->rq_export;
3220 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3221 lustre_msg_set_last_committed(req->rq_repmsg,
3222 exp->exp_last_committed);
3224 DEBUG_REQ(D_IOCTL, req,
3225 "not sending last_committed update (%d/%d)",
3226 exp->exp_obd->obd_no_transno,
3227 req->rq_repmsg == NULL);
3229 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3230 exp->exp_last_committed, req->rq_transno, req->rq_xid);
3233 #endif /* HAVE_SERVER_SUPPORT */
3236 * Packs current SLV and Limit into \a req.
3238 int target_pack_pool_reply(struct ptlrpc_request *req)
3240 struct obd_device *obd;
3245 * Check that we still have all structures alive as this may
3246 * be some late RPC at shutdown time.
3248 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3249 !exp_connect_lru_resize(req->rq_export))) {
3250 lustre_msg_set_slv(req->rq_repmsg, 0);
3251 lustre_msg_set_limit(req->rq_repmsg, 0);
3255 /* OBD is alive here as export is alive, which we checked above. */
3256 obd = req->rq_export->exp_obd;
3258 read_lock(&obd->obd_pool_lock);
3259 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3260 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3261 read_unlock(&obd->obd_pool_lock);
3266 static int target_send_reply_msg(struct ptlrpc_request *req,
3267 int rc, int fail_id)
3269 if (CFS_FAIL_CHECK_ORSET(fail_id & ~CFS_FAIL_ONCE, CFS_FAIL_ONCE)) {
3270 DEBUG_REQ(D_ERROR, req, "dropping reply");
3274 * We can have a null rq_reqmsg in the event of bad signature or
3275 * no context when unwrapping
3277 if (req->rq_reqmsg &&
3278 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3279 CFS_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3283 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3284 req->rq_status = rc;
3285 return ptlrpc_send_error(req, 1);
3287 DEBUG_REQ(D_NET, req, "sending reply");
3289 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3292 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3294 struct ptlrpc_service_part *svcpt;
3296 struct ptlrpc_reply_state *rs;
3297 struct obd_export *exp;
3301 if (req->rq_no_reply) {
3306 svcpt = req->rq_rqbd->rqbd_svcpt;
3307 rs = req->rq_reply_state;
3308 if (rs == NULL || !rs->rs_difficult) {
3310 target_send_reply_msg(req, rc, fail_id);
3315 /* must be an export if locks saved */
3316 LASSERT(req->rq_export != NULL);
3317 /* req/reply consistent */
3318 LASSERT(rs->rs_svcpt == svcpt);
3321 LASSERT(!rs->rs_scheduled);
3322 LASSERT(!rs->rs_scheduled_ever);
3323 LASSERT(!rs->rs_handled);
3324 LASSERT(!rs->rs_sent);
3325 LASSERT(!rs->rs_unlinked);
3326 LASSERT(rs->rs_export == NULL);
3327 LASSERT(list_empty(&rs->rs_obd_list));
3328 LASSERT(list_empty(&rs->rs_exp_list));
3330 exp = class_export_get(req->rq_export);
3332 /* disable reply scheduling while I'm setting up */
3333 rs->rs_scheduled = 1;
3335 rs->rs_unlinked = 0;
3336 rs->rs_xid = req->rq_xid;
3337 rs->rs_transno = req->rq_transno;
3338 rs->rs_export = exp;
3339 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
3341 spin_lock(&exp->exp_uncommitted_replies_lock);
3342 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3343 rs->rs_transno, exp->exp_last_committed);
3344 if (rs->rs_transno > exp->exp_last_committed) {
3345 /* not committed already */
3346 list_add_tail(&rs->rs_obd_list,
3347 &exp->exp_uncommitted_replies);
3349 spin_unlock(&exp->exp_uncommitted_replies_lock);
3351 spin_lock(&exp->exp_lock);
3352 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3353 spin_unlock(&exp->exp_lock);
3355 netrc = target_send_reply_msg(req, rc, fail_id);
3357 spin_lock(&svcpt->scp_rep_lock);
3359 atomic_inc(&svcpt->scp_nreps_difficult);
3363 * error sending: reply is off the net. Also we need +1
3364 * reply ref until ptlrpc_handle_rs() is done
3365 * with the reply state (if the send was successful, there
3366 * would have been +1 ref for the net, which
3367 * reply_out_callback leaves alone)
3370 rs->rs_unlinked = 1;
3371 ptlrpc_rs_addref(rs);
3374 spin_lock(&rs->rs_lock);
3375 if (rs->rs_transno <= exp->exp_last_committed ||
3376 (rs->rs_unlinked && !rs->rs_no_ack) ||
3377 list_empty(&rs->rs_exp_list) || /* completed already */
3378 list_empty(&rs->rs_obd_list)) {
3379 CDEBUG(D_HA, "Schedule reply immediately\n");
3380 ptlrpc_dispatch_difficult_reply(rs);
3382 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3383 rs->rs_scheduled = 0; /* allow notifier to schedule */
3385 spin_unlock(&rs->rs_lock);
3386 spin_unlock(&svcpt->scp_rep_lock);
3390 enum ldlm_mode lck_compat_array[] = {
3391 [LCK_EX] = LCK_COMPAT_EX,
3392 [LCK_PW] = LCK_COMPAT_PW,
3393 [LCK_PR] = LCK_COMPAT_PR,
3394 [LCK_CW] = LCK_COMPAT_CW,
3395 [LCK_CR] = LCK_COMPAT_CR,
3396 [LCK_NL] = LCK_COMPAT_NL,
3397 [LCK_GROUP] = LCK_COMPAT_GROUP,
3398 [LCK_COS] = LCK_COMPAT_COS,
3399 [LCK_TXN] = LCK_COMPAT_TXN,
3403 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3404 * not escape to the user level.
3406 int ldlm_error2errno(enum ldlm_error error)
3412 case ELDLM_LOCK_MATCHED:
3415 case ELDLM_LOCK_CHANGED:
3418 case ELDLM_LOCK_ABORTED:
3421 case ELDLM_LOCK_REPLACED:
3424 case ELDLM_NO_LOCK_DATA:
3427 case ELDLM_NAMESPACE_EXISTS:
3430 case ELDLM_BAD_NAMESPACE:
3434 if (((int)error) < 0) { /* cast to signed type */
3435 result = error; /* as ldlm_error can be unsigned */
3437 CERROR("Invalid DLM result code: %d\n", error);
3443 EXPORT_SYMBOL(ldlm_error2errno);
3446 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3448 enum ldlm_error ldlm_errno2error(int err_no)
3457 error = ELDLM_LOCK_CHANGED;
3460 error = ELDLM_LOCK_ABORTED;
3463 error = ELDLM_LOCK_REPLACED;
3466 error = ELDLM_NO_LOCK_DATA;
3469 error = ELDLM_NAMESPACE_EXISTS;
3472 error = ELDLM_BAD_NAMESPACE;
3480 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3481 void ldlm_dump_export_locks(struct obd_export *exp)
3483 spin_lock(&exp->exp_locks_list_guard);
3484 if (!list_empty(&exp->exp_locks_list)) {
3485 struct ldlm_lock *lock;
3487 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3489 list_for_each_entry(lock, &exp->exp_locks_list,
3491 LDLM_ERROR(lock, "lock:");
3493 spin_unlock(&exp->exp_locks_list_guard);
3497 #ifdef HAVE_SERVER_SUPPORT
3498 static inline const char *bulk2type(struct ptlrpc_request *req)
3500 if (req->rq_bulk_read)
3502 if (req->rq_bulk_write)
3507 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3509 struct ptlrpc_request *req = desc->bd_req;
3510 time64_t start = ktime_get_seconds();
3516 /* Check if client was evicted or reconnected already. */
3517 if (exp->exp_failed ||
3518 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3521 if (req->rq_bulk_read)
3522 rc = sptlrpc_svc_wrap_bulk(req, desc);
3524 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3525 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3526 else /* old version, bulk matchbits is rq_xid */
3527 req->rq_mbits = req->rq_xid;
3530 rc = ptlrpc_start_bulk_transfer(desc);
3534 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3535 bulk2type(req), rc);
3539 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3540 ptlrpc_abort_bulk(desc);
3544 /* limit actual bulk transfer to bulk_timeout seconds */
3545 deadline = start + bulk_timeout;
3546 if (deadline > req->rq_deadline)
3547 deadline = req->rq_deadline;
3550 time64_t timeoutl = deadline - ktime_get_seconds();
3551 time64_t rq_deadline;
3553 while (timeoutl >= 0 &&
3554 wait_event_idle_timeout(
3556 !ptlrpc_server_bulk_active(desc) ||
3559 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3560 timeoutl ? cfs_time_seconds(1) : 1) == 0)
3562 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3564 /* Wait again if we changed rq_deadline. */
3565 rq_deadline = READ_ONCE(req->rq_deadline);
3566 deadline = start + bulk_timeout;
3567 if (deadline > rq_deadline)
3568 deadline = rq_deadline;
3569 } while (rc == -ETIMEDOUT &&
3570 deadline > ktime_get_seconds());
3572 if (rc == -ETIMEDOUT) {
3573 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3574 bulk2type(req), deadline - start,
3575 ktime_get_real_seconds() - deadline);
3576 ptlrpc_abort_bulk(desc);
3577 } else if (exp->exp_failed) {
3578 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3581 ptlrpc_abort_bulk(desc);
3582 } else if (exp->exp_conn_cnt >
3583 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3584 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3586 /* We don't reply anyway. */
3588 ptlrpc_abort_bulk(desc);
3589 } else if (desc->bd_failure) {
3590 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3592 /* XXX should this be a different errno? */
3595 if (req->rq_bulk_write)
3596 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3597 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3598 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3599 bulk2type(req), desc->bd_nob_transferred,
3601 /* XXX should this be a different errno? */
3608 EXPORT_SYMBOL(target_bulk_io);
3610 #endif /* HAVE_SERVER_SUPPORT */