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, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
34 * This file deals with various client/target related logic including recovery.
36 * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
40 #define DEBUG_SUBSYSTEM S_LDLM
42 #include <linux/kthread.h>
43 #include <libcfs/libcfs.h>
45 #include <obd_class.h>
46 #include <lustre_dlm.h>
47 #include <lustre_net.h>
48 #include <lustre_sec.h>
49 #include "ldlm_internal.h"
51 /* @priority: If non-zero, move the selected connection to the list head.
52 * @create: If zero, only search in existing connections.
54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
55 int priority, int create)
57 struct ptlrpc_connection *ptlrpc_conn;
58 struct obd_import_conn *imp_conn = NULL, *item;
59 lnet_nid_t nid4refnet = LNET_NID_ANY;
63 if (!create && !priority) {
64 CDEBUG(D_HA, "Nothing to do\n");
68 if (imp->imp_connection &&
69 imp->imp_connection->c_remote_uuid.uuid[0] == 0)
70 /* nid4refnet is used to restrict network connections */
71 nid4refnet = imp->imp_connection->c_self;
72 ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
74 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
79 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
81 GOTO(out_put, rc = -ENOMEM);
84 spin_lock(&imp->imp_lock);
85 list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
86 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
88 list_del(&item->oic_item);
89 list_add(&item->oic_item,
91 item->oic_last_attempt = 0;
93 CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
94 imp, imp->imp_obd->obd_name, uuid->uuid,
95 (priority ? ", moved to head" : ""));
96 spin_unlock(&imp->imp_lock);
97 GOTO(out_free, rc = 0);
100 /* No existing import connection found for \a uuid. */
102 imp_conn->oic_conn = ptlrpc_conn;
103 imp_conn->oic_uuid = *uuid;
104 imp_conn->oic_last_attempt = 0;
106 list_add(&imp_conn->oic_item, &imp->imp_conn_list);
108 list_add_tail(&imp_conn->oic_item,
109 &imp->imp_conn_list);
110 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
111 imp, imp->imp_obd->obd_name, uuid->uuid,
112 (priority ? "head" : "tail"));
114 spin_unlock(&imp->imp_lock);
115 GOTO(out_free, rc = -ENOENT);
118 spin_unlock(&imp->imp_lock);
122 OBD_FREE(imp_conn, sizeof(*imp_conn));
124 ptlrpc_connection_put(ptlrpc_conn);
128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
130 return import_set_conn(imp, uuid, 1, 0);
133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
136 return import_set_conn(imp, uuid, priority, 1);
138 EXPORT_SYMBOL(client_import_add_conn);
140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
142 struct obd_import_conn *imp_conn;
143 struct obd_export *dlmexp;
147 spin_lock(&imp->imp_lock);
148 if (list_empty(&imp->imp_conn_list)) {
149 LASSERT(!imp->imp_connection);
153 list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
154 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
156 LASSERT(imp_conn->oic_conn);
158 if (imp_conn == imp->imp_conn_current) {
159 LASSERT(imp_conn->oic_conn == imp->imp_connection);
161 if (imp->imp_state != LUSTRE_IMP_CLOSED &&
162 imp->imp_state != LUSTRE_IMP_DISCON) {
163 CERROR("can't remove current connection\n");
164 GOTO(out, rc = -EBUSY);
167 ptlrpc_connection_put(imp->imp_connection);
168 imp->imp_connection = NULL;
170 dlmexp = class_conn2export(&imp->imp_dlm_handle);
171 if (dlmexp && dlmexp->exp_connection) {
172 LASSERT(dlmexp->exp_connection ==
174 ptlrpc_connection_put(dlmexp->exp_connection);
175 dlmexp->exp_connection = NULL;
179 class_export_put(dlmexp);
182 list_del(&imp_conn->oic_item);
183 ptlrpc_connection_put(imp_conn->oic_conn);
184 OBD_FREE(imp_conn, sizeof(*imp_conn));
185 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
186 imp, imp->imp_obd->obd_name, uuid->uuid);
191 spin_unlock(&imp->imp_lock);
193 CERROR("connection %s not found\n", uuid->uuid);
196 EXPORT_SYMBOL(client_import_del_conn);
199 * Find conn UUID by peer NID. \a peer is a server NID. This function is used
200 * to find a conn uuid of \a imp which can reach \a peer.
202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
203 struct obd_uuid *uuid)
205 struct obd_import_conn *conn;
209 spin_lock(&imp->imp_lock);
210 list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
211 /* Check if conn UUID does have this peer NID. */
212 if (class_check_uuid(&conn->oic_uuid, peer)) {
213 *uuid = conn->oic_uuid;
218 spin_unlock(&imp->imp_lock);
221 EXPORT_SYMBOL(client_import_find_conn);
223 void client_destroy_import(struct obd_import *imp)
225 /* Drop security policy instance after all RPCs have finished/aborted
226 * to let all busy contexts be released. */
227 class_import_get(imp);
228 class_destroy_import(imp);
229 sptlrpc_import_sec_put(imp);
230 class_import_put(imp);
232 EXPORT_SYMBOL(client_destroy_import);
235 * Check whether or not the OSC is on MDT.
238 * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
240 * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
243 static int osc_on_mdt(char *obdname)
247 ptr = strrchr(obdname, '-');
251 if (strncmp(ptr + 1, "MDT", 3) == 0)
257 /* Configure an RPC client OBD device.
262 * 3 - inactive-on-startup
263 * 4 - restrictive net
265 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
267 struct client_obd *cli = &obddev->u.cli;
268 struct obd_import *imp;
269 struct obd_uuid server_uuid;
270 int rq_portal, rp_portal, connect_op;
271 char *name = obddev->obd_type->typ_name;
272 enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
273 char *cli_name = lustre_cfg_buf(lcfg, 0);
274 struct ptlrpc_connection fake_conn = { .c_self = 0,
275 .c_remote_uuid.uuid[0] = 0 };
279 /* In a more perfect world, we would hang a ptlrpc_client off of
280 * obd_type and just use the values from there. */
281 if (!strcmp(name, LUSTRE_OSC_NAME)) {
282 rq_portal = OST_REQUEST_PORTAL;
283 rp_portal = OSC_REPLY_PORTAL;
284 connect_op = OST_CONNECT;
285 cli->cl_sp_me = LUSTRE_SP_CLI;
286 cli->cl_sp_to = LUSTRE_SP_OST;
287 ns_type = LDLM_NS_TYPE_OSC;
288 } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
289 !strcmp(name, LUSTRE_LWP_NAME)) {
290 rq_portal = MDS_REQUEST_PORTAL;
291 rp_portal = MDC_REPLY_PORTAL;
292 connect_op = MDS_CONNECT;
293 if (is_lwp_on_ost(cli_name))
294 cli->cl_sp_me = LUSTRE_SP_OST;
295 else if (is_lwp_on_mdt(cli_name))
296 cli->cl_sp_me = LUSTRE_SP_MDT;
298 cli->cl_sp_me = LUSTRE_SP_CLI;
299 cli->cl_sp_to = LUSTRE_SP_MDT;
300 ns_type = LDLM_NS_TYPE_MDC;
301 } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
302 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
303 /* OSP_on_MDT for other MDTs */
304 connect_op = MDS_CONNECT;
305 cli->cl_sp_to = LUSTRE_SP_MDT;
306 ns_type = LDLM_NS_TYPE_MDC;
307 rq_portal = OUT_PORTAL;
309 /* OSP on MDT for OST */
310 connect_op = OST_CONNECT;
311 cli->cl_sp_to = LUSTRE_SP_OST;
312 ns_type = LDLM_NS_TYPE_OSC;
313 rq_portal = OST_REQUEST_PORTAL;
315 rp_portal = OSC_REPLY_PORTAL;
316 cli->cl_sp_me = LUSTRE_SP_MDT;
317 } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
318 rq_portal = MGS_REQUEST_PORTAL;
319 rp_portal = MGC_REPLY_PORTAL;
320 connect_op = MGS_CONNECT;
321 cli->cl_sp_me = LUSTRE_SP_MGC;
322 cli->cl_sp_to = LUSTRE_SP_MGS;
323 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
324 ns_type = LDLM_NS_TYPE_MGC;
326 CERROR("unknown client OBD type \"%s\", can't setup\n",
331 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
332 CERROR("requires a TARGET UUID\n");
336 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
337 CERROR("client UUID must be less than 38 characters\n");
341 if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
342 CERROR("setup requires a SERVER UUID\n");
346 if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
347 CERROR("target UUID must be less than 38 characters\n");
351 init_rwsem(&cli->cl_sem);
352 mutex_init(&cli->cl_mgc_mutex);
354 init_rwsem(&cli->cl_seq_rwsem);
355 cli->cl_conn_count = 0;
356 memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
357 min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
358 sizeof(server_uuid)));
360 cli->cl_dirty_pages = 0;
361 cli->cl_avail_grant = 0;
362 /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
363 /* cl_dirty_max_pages may be changed at connect time in
364 * ptlrpc_connect_interpret(). */
365 client_adjust_max_dirty(cli);
366 INIT_LIST_HEAD(&cli->cl_cache_waiters);
367 INIT_LIST_HEAD(&cli->cl_loi_ready_list);
368 INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
369 INIT_LIST_HEAD(&cli->cl_loi_write_list);
370 INIT_LIST_HEAD(&cli->cl_loi_read_list);
371 spin_lock_init(&cli->cl_loi_list_lock);
372 atomic_set(&cli->cl_pending_w_pages, 0);
373 atomic_set(&cli->cl_pending_r_pages, 0);
374 cli->cl_r_in_flight = 0;
375 cli->cl_w_in_flight = 0;
377 spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
378 spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
379 spin_lock_init(&cli->cl_read_page_hist.oh_lock);
380 spin_lock_init(&cli->cl_write_page_hist.oh_lock);
381 spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
382 spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
385 INIT_LIST_HEAD(&cli->cl_lru_osc);
386 atomic_set(&cli->cl_lru_shrinkers, 0);
387 atomic_long_set(&cli->cl_lru_busy, 0);
388 atomic_long_set(&cli->cl_lru_in_list, 0);
389 INIT_LIST_HEAD(&cli->cl_lru_list);
390 spin_lock_init(&cli->cl_lru_list_lock);
391 atomic_long_set(&cli->cl_unstable_count, 0);
392 INIT_LIST_HEAD(&cli->cl_shrink_list);
394 INIT_LIST_HEAD(&cli->cl_flight_waiters);
395 cli->cl_rpcs_in_flight = 0;
397 init_waitqueue_head(&cli->cl_destroy_waitq);
398 atomic_set(&cli->cl_destroy_in_flight, 0);
399 #ifdef ENABLE_CHECKSUM
400 /* Turn on checksumming by default. */
401 cli->cl_checksum = 1;
403 * The supported checksum types will be worked out at connect time
404 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
407 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
409 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
411 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
412 * from OFD after connecting. */
413 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
415 /* set cl_chunkbits default value to PAGE_SHIFT,
416 * it will be updated at OSC connection time. */
417 cli->cl_chunkbits = PAGE_SHIFT;
419 if (!strcmp(name, LUSTRE_MDC_NAME)) {
420 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
421 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
422 cli->cl_max_rpcs_in_flight = 2;
423 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
424 cli->cl_max_rpcs_in_flight = 3;
425 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
426 cli->cl_max_rpcs_in_flight = 4;
428 if (osc_on_mdt(obddev->obd_name))
429 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
431 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
434 spin_lock_init(&cli->cl_mod_rpcs_lock);
435 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
436 cli->cl_max_mod_rpcs_in_flight = 0;
437 cli->cl_mod_rpcs_in_flight = 0;
438 cli->cl_close_rpcs_in_flight = 0;
439 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
440 cli->cl_mod_tag_bitmap = NULL;
442 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
444 if (connect_op == MDS_CONNECT) {
445 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
446 OBD_ALLOC(cli->cl_mod_tag_bitmap,
447 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
448 if (cli->cl_mod_tag_bitmap == NULL)
449 GOTO(err, rc = -ENOMEM);
454 CERROR("ldlm_get_ref failed: %d\n", rc);
458 ptlrpc_init_client(rq_portal, rp_portal, name,
459 &obddev->obd_ldlm_client);
461 imp = class_new_import(obddev);
463 GOTO(err_ldlm, rc = -ENOENT);
464 imp->imp_client = &obddev->obd_ldlm_client;
465 imp->imp_connect_op = connect_op;
466 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
467 LUSTRE_CFG_BUFLEN(lcfg, 1));
468 class_import_put(imp);
470 if (lustre_cfg_buf(lcfg, 4)) {
471 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
473 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
475 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
476 obddev->obd_name, lustre_cfg_string(lcfg, 4),
478 GOTO(err_import, rc);
480 fake_conn.c_self = LNET_MKNID(refnet, 0);
481 imp->imp_connection = &fake_conn;
484 rc = client_import_add_conn(imp, &server_uuid, 1);
486 CERROR("can't add initial connection\n");
487 GOTO(err_import, rc);
489 imp->imp_connection = NULL;
491 cli->cl_import = imp;
492 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
493 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
495 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
496 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
497 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
498 name, obddev->obd_name,
499 cli->cl_target_uuid.uuid);
500 spin_lock(&imp->imp_lock);
501 imp->imp_deactive = 1;
502 spin_unlock(&imp->imp_lock);
506 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
507 LDLM_NAMESPACE_CLIENT,
508 LDLM_NAMESPACE_GREEDY,
510 if (obddev->obd_namespace == NULL) {
511 CERROR("Unable to create client namespace - %s\n",
513 GOTO(err_import, rc = -ENOMEM);
519 class_destroy_import(imp);
523 if (cli->cl_mod_tag_bitmap != NULL)
524 OBD_FREE(cli->cl_mod_tag_bitmap,
525 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
526 cli->cl_mod_tag_bitmap = NULL;
530 EXPORT_SYMBOL(client_obd_setup);
532 int client_obd_cleanup(struct obd_device *obddev)
534 struct client_obd *cli = &obddev->u.cli;
537 ldlm_namespace_free_post(obddev->obd_namespace);
538 obddev->obd_namespace = NULL;
540 obd_cleanup_client_import(obddev);
541 LASSERT(obddev->u.cli.cl_import == NULL);
545 if (cli->cl_mod_tag_bitmap != NULL)
546 OBD_FREE(cli->cl_mod_tag_bitmap,
547 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
548 cli->cl_mod_tag_bitmap = NULL;
552 EXPORT_SYMBOL(client_obd_cleanup);
554 /* ->o_connect() method for client side (OSC and MDC and MGC) */
555 int client_connect_import(const struct lu_env *env,
556 struct obd_export **exp,
557 struct obd_device *obd, struct obd_uuid *cluuid,
558 struct obd_connect_data *data, void *localdata)
560 struct client_obd *cli = &obd->u.cli;
561 struct obd_import *imp = cli->cl_import;
562 struct obd_connect_data *ocd;
563 struct lustre_handle conn = { 0 };
568 down_write(&cli->cl_sem);
569 if (cli->cl_conn_count > 0)
570 GOTO(out_sem, rc = -EALREADY);
572 rc = class_connect(&conn, obd, cluuid);
576 cli->cl_conn_count++;
577 *exp = class_conn2export(&conn);
579 LASSERT(obd->obd_namespace);
581 imp->imp_dlm_handle = conn;
582 rc = ptlrpc_init_import(imp);
586 ocd = &imp->imp_connect_data;
589 imp->imp_connect_flags_orig = data->ocd_connect_flags;
590 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
593 rc = ptlrpc_connect_import(imp);
595 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
598 LASSERT(*exp != NULL && (*exp)->exp_connection);
601 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
602 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
603 data->ocd_connect_flags, ocd->ocd_connect_flags);
604 data->ocd_connect_flags = ocd->ocd_connect_flags;
605 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
608 ptlrpc_pinger_add_import(imp);
614 cli->cl_conn_count--;
615 class_disconnect(*exp);
619 up_write(&cli->cl_sem);
623 EXPORT_SYMBOL(client_connect_import);
625 int client_disconnect_export(struct obd_export *exp)
627 struct obd_device *obd = class_exp2obd(exp);
628 struct client_obd *cli;
629 struct obd_import *imp;
634 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
635 exp, exp ? exp->exp_handle.h_cookie : -1);
640 imp = cli->cl_import;
642 down_write(&cli->cl_sem);
643 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
646 if (cli->cl_conn_count == 0) {
647 CERROR("disconnecting disconnected device (%s)\n",
649 GOTO(out_disconnect, rc = -EINVAL);
652 cli->cl_conn_count--;
653 if (cli->cl_conn_count != 0)
654 GOTO(out_disconnect, rc = 0);
656 /* Mark import deactivated now, so we don't try to reconnect if any
657 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
658 * fully deactivate the import, or that would drop all requests. */
659 spin_lock(&imp->imp_lock);
660 imp->imp_deactive = 1;
661 spin_unlock(&imp->imp_lock);
663 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
664 * delete it regardless. (It's safe to delete an import that was
666 (void)ptlrpc_pinger_del_import(imp);
668 if (obd->obd_namespace != NULL) {
669 /* obd_force == local only */
670 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
671 obd->obd_force ? LCF_LOCAL : 0, NULL);
672 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
675 /* There's no need to hold sem while disconnecting an import,
676 * and it may actually cause deadlock in GSS. */
677 up_write(&cli->cl_sem);
678 rc = ptlrpc_disconnect_import(imp, 0);
679 down_write(&cli->cl_sem);
681 ptlrpc_invalidate_import(imp);
686 /* Use server style - class_disconnect should be always called for
688 err = class_disconnect(exp);
692 up_write(&cli->cl_sem);
696 EXPORT_SYMBOL(client_disconnect_export);
698 #ifdef HAVE_SERVER_SUPPORT
699 int server_disconnect_export(struct obd_export *exp)
704 /* Disconnect early so that clients can't keep using export. */
705 rc = class_disconnect(exp);
706 /* Close import to avoid sending any requests. */
707 if (exp->exp_imp_reverse)
708 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
710 ldlm_bl_thread_wakeup();
712 /* complete all outstanding replies */
713 spin_lock(&exp->exp_lock);
714 while (!list_empty(&exp->exp_outstanding_replies)) {
715 struct ptlrpc_reply_state *rs =
716 list_entry(exp->exp_outstanding_replies.next,
717 struct ptlrpc_reply_state, rs_exp_list);
718 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
720 spin_lock(&svcpt->scp_rep_lock);
722 list_del_init(&rs->rs_exp_list);
724 spin_lock(&rs->rs_lock);
725 /* clear rs_convert_lock to make sure rs is handled and put */
726 rs->rs_convert_lock = 0;
727 ptlrpc_schedule_difficult_reply(rs);
728 spin_unlock(&rs->rs_lock);
730 spin_unlock(&svcpt->scp_rep_lock);
732 spin_unlock(&exp->exp_lock);
736 EXPORT_SYMBOL(server_disconnect_export);
738 /* --------------------------------------------------------------------------
739 * from old lib/target.c
740 * -------------------------------------------------------------------------- */
742 static int target_handle_reconnect(struct lustre_handle *conn,
743 struct obd_export *exp,
744 struct obd_uuid *cluuid)
746 struct obd_device *target;
747 struct lustre_handle *hdl;
754 hdl = &exp->exp_imp_reverse->imp_remote_handle;
755 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
756 conn->cookie = exp->exp_handle.h_cookie;
757 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
758 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
762 target = exp->exp_obd;
764 /* Might be a re-connect after a partition. */
765 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
766 LCONSOLE_WARN("%s: already connected client %s (at %s) "
767 "with handle %#llx. Rejecting client "
768 "with the same UUID trying to reconnect "
769 "with handle %#llx\n", target->obd_name,
770 obd_uuid2str(&exp->exp_client_uuid),
771 obd_export_nid2str(exp),
772 hdl->cookie, conn->cookie);
773 memset(conn, 0, sizeof *conn);
774 /* target_handle_connect() treats EALREADY and
775 * -EALREADY differently. -EALREADY is an error
776 * (same UUID, different handle). */
780 if (!target->obd_recovering) {
781 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
782 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
783 obd_export_nid2str(exp));
784 GOTO(out_already, rc);
787 now = cfs_time_current();
788 deadline = target->obd_recovery_timer.expires;
789 if (cfs_time_before(now, deadline)) {
790 struct target_distribute_txn_data *tdtd =
791 class_exp2tgt(exp)->lut_tdtd;
796 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
797 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
798 buf = tdtd->tdtd_show_update_logs_retrievers(
799 tdtd->tdtd_show_retrievers_cbdata,
803 LCONSOLE_WARN("%s: Recovery already passed deadline "
804 "%d:%.02d. It is due to DNE recovery "
805 "failed/stuck on the %d MDT(s):%s. "
806 "Please wait until all MDTs recovered "
807 "or abort the recovery by force.\n",
808 target->obd_name, timeout / 60,
810 buf ? buf : "unknown (not enough RAM)");
812 LCONSOLE_WARN("%s: Recovery already passed deadline "
813 "%d:%.02d. If you do not want to wait "
814 "more, please abort the recovery by "
815 "force.\n", target->obd_name,
816 timeout / 60, timeout % 60);
821 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
822 LCONSOLE_WARN("%s: Recovery already passed deadline"
823 " %d:%.02d, It is most likely due to DNE"
824 " recovery is failed or stuck, please wait a"
825 " few more minutes or abort the recovery.\n",
826 target->obd_name, timeout / 60, timeout % 60);
830 conn->cookie = exp->exp_handle.h_cookie;
831 /* target_handle_connect() treats EALREADY and
832 * -EALREADY differently. EALREADY means we are
833 * doing a valid reconnect from the same client. */
838 check_and_start_recovery_timer(struct obd_device *obd,
839 struct ptlrpc_request *req, int new_client);
842 * update flags for import during reconnect process
844 static int rev_import_flags_update(struct obd_import *revimp,
845 struct ptlrpc_request *req)
848 struct obd_connect_data *data;
850 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
852 if (data->ocd_connect_flags & OBD_CONNECT_AT)
853 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
855 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
857 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
859 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
861 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
862 revimp->imp_client->cli_name,
863 libcfs_id2str(req->rq_peer), rc);
871 * Allocate a new reverse import for an export.
873 * \retval -errno in case error hit
874 * \retval 0 if reverse import correctly init
876 int rev_import_init(struct obd_export *export)
878 struct obd_device *obd = export->exp_obd;
879 struct obd_import *revimp;
881 LASSERT(export->exp_imp_reverse == NULL);
883 revimp = class_new_import(obd);
887 revimp->imp_remote_handle.cookie = 0ULL;
888 revimp->imp_client = &obd->obd_ldlm_client;
889 revimp->imp_dlm_fake = 1;
891 /* it is safe to connect import in new state as no sends possible */
892 spin_lock(&export->exp_lock);
893 export->exp_imp_reverse = revimp;
894 spin_unlock(&export->exp_lock);
895 class_import_put(revimp);
899 EXPORT_SYMBOL(rev_import_init);
902 * Handle reconnect for an export.
904 * \param exp export to handle reconnect process
905 * \param req client reconnect request
907 * \retval -rc in case securitfy flavor can't be changed
908 * \retval 0 in case none problems
910 static int rev_import_reconnect(struct obd_export *exp,
911 struct ptlrpc_request *req)
913 struct obd_import *revimp = exp->exp_imp_reverse;
914 struct lustre_handle *lh;
917 /* avoid sending a request until import flags are changed */
918 ptlrpc_import_enter_resend(revimp);
920 if (revimp->imp_connection != NULL)
921 ptlrpc_connection_put(revimp->imp_connection);
924 * client from recovery don't have a handle so we need to take from
925 * request. it may produce situation when wrong client connected
926 * to recovery as we trust a client uuid
928 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
929 revimp->imp_remote_handle = *lh;
931 /* unknown versions will be caught in
932 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
933 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
935 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
937 rc = rev_import_flags_update(revimp, req);
939 /* it is safe to still be in RECOVERY phase as we are not able
940 * to setup correct security flavor so requests are not able to
941 * be delivered correctly */
945 /* resend all rpc's via new connection */
946 return ptlrpc_import_recovery_state_machine(revimp);
949 int target_handle_connect(struct ptlrpc_request *req)
951 struct obd_device *target = NULL;
952 struct obd_export *export = NULL;
953 /* connect handle - filled from target_handle_reconnect in
955 struct lustre_handle conn;
956 struct lustre_handle *tmp;
957 struct obd_uuid tgtuuid;
958 struct obd_uuid cluuid;
963 bool mds_conn = false, lw_client = false, initial_conn = false;
964 bool mds_mds_conn = false;
965 bool new_mds_mds_conn = false;
966 bool target_referenced = false;
967 struct obd_connect_data *data, *tmpdata;
969 lnet_nid_t *client_nid = NULL;
972 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
974 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
976 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
977 GOTO(out, rc = -EINVAL);
980 obd_str2uuid(&tgtuuid, str);
981 target = class_uuid2obd(&tgtuuid);
983 target = class_name2obd(str);
986 deuuidify(str, NULL, &target_start, &target_len);
987 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
988 "from %s (no target). If you are running "
989 "an HA pair check that the target is "
990 "mounted on the other server.\n", str,
991 libcfs_nid2str(req->rq_peer.nid));
992 GOTO(out, rc = -ENODEV);
995 spin_lock(&target->obd_dev_lock);
996 if (target->obd_stopping || !target->obd_set_up) {
997 spin_unlock(&target->obd_dev_lock);
999 deuuidify(str, NULL, &target_start, &target_len);
1000 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1001 target_len, target_start,
1002 libcfs_nid2str(req->rq_peer.nid),
1003 (target->obd_stopping ?
1004 "stopping" : "not set up"));
1005 GOTO(out, rc = -ENODEV);
1008 if (target->obd_no_conn) {
1009 spin_unlock(&target->obd_dev_lock);
1011 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1012 "from %s\n", target->obd_name,
1013 libcfs_nid2str(req->rq_peer.nid));
1014 GOTO(out, rc = -EAGAIN);
1017 /* Make sure the target isn't cleaned up while we're here. Yes,
1018 * there's still a race between the above check and our incref here.
1019 * Really, class_uuid2obd should take the ref. */
1020 class_incref(target, __func__, current);
1021 target_referenced = true;
1023 target->obd_conn_inprogress++;
1024 spin_unlock(&target->obd_dev_lock);
1026 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1028 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1029 GOTO(out, rc = -EINVAL);
1032 obd_str2uuid(&cluuid, str);
1034 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1036 GOTO(out, rc = -EPROTO);
1040 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1042 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1044 GOTO(out, rc = -EPROTO);
1046 rc = req_capsule_server_pack(&req->rq_pill);
1050 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1051 /* Don't allow clients to connect that are using old 1.8 format
1052 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1053 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1054 * FULL20 flag should be set on all connections since 2.0, but no
1055 * longer affects behaviour.
1057 * Later this check will be disabled and the flag can be retired
1058 * completely once interop with 3.0 is no longer needed.
1060 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1061 GOTO(out, rc = -EPROTO);
1064 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1065 if (data->ocd_version < LUSTRE_VERSION_CODE -
1066 LUSTRE_VERSION_ALLOWED_OFFSET ||
1067 data->ocd_version > LUSTRE_VERSION_CODE +
1068 LUSTRE_VERSION_ALLOWED_OFFSET) {
1069 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1070 "libclient connection attempt",
1071 data->ocd_version < LUSTRE_VERSION_CODE ?
1073 OBD_OCD_VERSION_MAJOR(data->ocd_version),
1074 OBD_OCD_VERSION_MINOR(data->ocd_version),
1075 OBD_OCD_VERSION_PATCH(data->ocd_version),
1076 OBD_OCD_VERSION_FIX(data->ocd_version));
1077 data = req_capsule_server_sized_get(&req->rq_pill,
1079 offsetof(typeof(*data), ocd_version) +
1080 sizeof(data->ocd_version));
1082 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1083 data->ocd_version = LUSTRE_VERSION_CODE;
1085 GOTO(out, rc = -EPROTO);
1089 /* Note: lw_client is needed in MDS-MDS failover during update log
1090 * processing, so we needs to allow lw_client to be connected at
1091 * anytime, instead of only the initial connection */
1092 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1094 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1095 initial_conn = true;
1096 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1097 mds_mds_conn = (data->ocd_connect_flags &
1098 OBD_CONNECT_MDS_MDS) != 0;
1100 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1101 * for Imperative Recovery connection from MGC to MGS.
1103 * Via check OBD_CONNECT_FID, we can distinguish whether
1104 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1107 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1108 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1109 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1110 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1111 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1112 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1114 /* We do not support the MDT-MDT interoperations with
1115 * different version MDT because of protocol changes. */
1116 if (unlikely(major != LUSTRE_MAJOR ||
1117 minor != LUSTRE_MINOR ||
1118 abs(patch - LUSTRE_PATCH) > 3)) {
1119 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1120 "connection from different version MDT "
1121 "(%d.%d.%d.%d) %s %s\n",
1122 target->obd_name, LUSTRE_MAJOR,
1123 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1124 major, minor, patch,
1125 OBD_OCD_VERSION_FIX(data->ocd_version),
1126 libcfs_nid2str(req->rq_peer.nid), str);
1128 GOTO(out, rc = -EPROTO);
1133 /* lctl gets a backstage, all-access pass. */
1134 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1135 goto dont_check_exports;
1137 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1141 /* We've found an export in the hash. */
1143 spin_lock(&export->exp_lock);
1145 if (export->exp_connecting) { /* bug 9635, et. al. */
1146 spin_unlock(&export->exp_lock);
1147 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1148 export->exp_obd->obd_name, export,
1149 libcfs_nid2str(req->rq_peer.nid));
1150 class_export_put(export);
1153 } else if ((mds_conn || (lw_client && initial_conn) ||
1154 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1155 export->exp_connection != NULL) {
1156 spin_unlock(&export->exp_lock);
1157 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1158 /* MDS or LWP reconnected after failover. */
1159 LCONSOLE_WARN("%s: Received %s connection from "
1160 "%s, removing former export from %s\n",
1161 target->obd_name, mds_conn ? "MDS" : "LWP",
1162 libcfs_nid2str(req->rq_peer.nid),
1163 libcfs_nid2str(export->exp_connection->c_peer.nid));
1165 /* New MDS connection from the same NID. */
1166 LCONSOLE_WARN("%s: Received new %s connection from "
1167 "%s, removing former export from same NID\n",
1168 target->obd_name, mds_conn ? "MDS" : "LWP",
1169 libcfs_nid2str(req->rq_peer.nid));
1172 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1173 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1174 /* Because exports between MDTs will always be
1175 * kept, let's do not fail such export if they
1176 * come from the same NID, otherwise it might
1177 * cause eviction between MDTs, which might
1178 * cause namespace inconsistency */
1179 spin_lock(&export->exp_lock);
1180 export->exp_connecting = 1;
1181 spin_unlock(&export->exp_lock);
1182 conn.cookie = export->exp_handle.h_cookie;
1185 class_fail_export(export);
1186 class_export_put(export);
1190 } else if (export->exp_connection != NULL && initial_conn &&
1191 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1192 spin_unlock(&export->exp_lock);
1193 /* In MDS failover we have static UUID but NID can change. */
1194 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1195 "existing nid %s is already connected\n",
1196 target->obd_name, cluuid.uuid,
1197 libcfs_nid2str(req->rq_peer.nid),
1199 export->exp_connection->c_peer.nid));
1201 class_export_put(export);
1204 export->exp_connecting = 1;
1205 spin_unlock(&export->exp_lock);
1206 LASSERT(export->exp_obd == target);
1208 rc = target_handle_reconnect(&conn, export, &cluuid);
1211 /* If we found an export, we already unlocked. */
1214 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1215 } else if (req->rq_export == NULL &&
1216 atomic_read(&export->exp_rpc_count) > 0) {
1217 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1218 "still busy with %d references\n",
1219 target->obd_name, cluuid.uuid,
1220 libcfs_nid2str(req->rq_peer.nid),
1221 atomic_read(&export->exp_refcount));
1222 GOTO(out, rc = -EBUSY);
1223 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1224 if (!strstr(cluuid.uuid, "mdt"))
1225 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1226 "known client %s (at %s) because it "
1227 "is indicating it is a new client",
1228 target->obd_name, cluuid.uuid,
1229 libcfs_nid2str(req->rq_peer.nid));
1230 GOTO(out, rc = -EALREADY);
1232 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1239 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1240 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1241 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1242 export, (long)cfs_time_current_sec(),
1243 export ? (long)export->exp_last_request_time : 0);
1245 /* If this is the first time a client connects, reset the recovery
1246 * timer. Discard lightweight connections which might be local. */
1247 if (!lw_client && rc == 0 && target->obd_recovering)
1248 check_and_start_recovery_timer(target, req, export == NULL);
1250 /* We want to handle EALREADY but *not* -EALREADY from
1251 * target_handle_reconnect(), return reconnection state in a flag. */
1252 if (rc == EALREADY) {
1253 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1259 /* Tell the client if we support replayable requests. */
1260 if (target->obd_replayable)
1261 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1262 client_nid = &req->rq_peer.nid;
1264 if (export == NULL) {
1265 /* allow lightweight connections during recovery */
1266 /* allow "new" MDT to be connected during recovery, since we
1267 * need retrieve recovery update records from it */
1268 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1270 int c; /* connected */
1271 int i; /* in progress */
1273 int s; /* stale/evicted */
1275 c = atomic_read(&target->obd_connected_clients);
1276 i = atomic_read(&target->obd_lock_replay_clients);
1277 k = target->obd_max_recoverable_clients;
1278 s = target->obd_stale_clients;
1279 t = target->obd_recovery_timer.expires;
1280 t = cfs_time_sub(t, cfs_time_current());
1281 t = cfs_duration_sec(t);
1282 LCONSOLE_WARN("%s: Denying connection for new client %s"
1283 "(at %s), waiting for %d known clients "
1284 "(%d recovered, %d in progress, and %d "
1285 "evicted) to recover in %d:%.02d\n",
1286 target->obd_name, cluuid.uuid,
1287 libcfs_nid2str(req->rq_peer.nid), k,
1288 c - i, i, s, (int)t / 60,
1293 rc = obd_connect(req->rq_svc_thread->t_env,
1294 &export, target, &cluuid, data,
1296 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1297 lustre_msg_add_op_flags(req->rq_repmsg,
1298 MSG_CONNECT_RECOVERING);
1300 conn.cookie = export->exp_handle.h_cookie;
1301 rc = rev_import_init(export);
1305 new_mds_mds_conn = true;
1308 rc = obd_reconnect(req->rq_svc_thread->t_env,
1309 export, target, &cluuid, data, client_nid);
1314 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1315 data->ocd_instance = target->u.obt.obt_instance;
1317 /* Return only the parts of obd_connect_data that we understand, so the
1318 * client knows that we don't understand the rest. */
1320 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1322 tmpdata = req_capsule_server_get(&req->rq_pill,
1324 /* Don't use struct assignment here, because the client reply
1325 * buffer may be smaller/larger than the local struct
1326 * obd_connect_data. */
1327 memcpy(tmpdata, data, min(tmpsize, size));
1330 /* If the client and the server are the same node, we will already
1331 * have an export that really points to the client's DLM export,
1332 * because we have a shared handles table.
1334 * XXX this will go away when shaver stops sending the "connect" handle
1335 * in the real "remote handle" field of the request --phik 24 Apr 2003
1337 ptlrpc_request_change_export(req, export);
1339 spin_lock(&export->exp_lock);
1340 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1341 spin_unlock(&export->exp_lock);
1342 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1343 "or equal conn_cnt: %d >= %d\n",
1344 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1345 export->exp_conn_cnt,
1346 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1348 GOTO(out, rc = -EALREADY);
1350 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1351 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1353 /* Don't evict liblustre clients for not pinging. */
1354 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1355 export->exp_libclient = 1;
1356 spin_unlock(&export->exp_lock);
1358 spin_lock(&target->obd_dev_lock);
1359 list_del_init(&export->exp_obd_chain_timed);
1360 spin_unlock(&target->obd_dev_lock);
1362 spin_unlock(&export->exp_lock);
1365 if (export->exp_connection != NULL) {
1366 /* Check to see if connection came from another NID. */
1367 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1368 !hlist_unhashed(&export->exp_nid_hash))
1369 cfs_hash_del(export->exp_obd->obd_nid_hash,
1370 &export->exp_connection->c_peer.nid,
1371 &export->exp_nid_hash);
1373 ptlrpc_connection_put(export->exp_connection);
1376 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1379 if (hlist_unhashed(&export->exp_nid_hash))
1380 cfs_hash_add(export->exp_obd->obd_nid_hash,
1381 &export->exp_connection->c_peer.nid,
1382 &export->exp_nid_hash);
1384 lustre_msg_set_handle(req->rq_repmsg, &conn);
1386 rc = rev_import_reconnect(export, req);
1390 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1392 __u64 transno = data->ocd_transno;
1394 spin_lock(&export->exp_lock);
1395 /* possible race with class_disconnect_stale_exports,
1396 * export may be already in the eviction process */
1397 if (export->exp_failed) {
1398 spin_unlock(&export->exp_lock);
1399 GOTO(out, rc = -ENODEV);
1401 export->exp_in_recovery = 1;
1402 export->exp_req_replay_needed = 1;
1403 export->exp_lock_replay_needed = 1;
1404 spin_unlock(&export->exp_lock);
1406 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1407 MSG_CONNECT_TRANSNO);
1408 if (has_transno && transno == 0)
1409 CWARN("Connect with zero transno!\n");
1411 if (has_transno && transno > 0 &&
1412 transno < target->obd_next_recovery_transno &&
1413 transno > target->obd_last_committed) {
1414 /* Another way is to use cmpxchg() to be lock-free. */
1415 spin_lock(&target->obd_recovery_task_lock);
1416 if (transno < target->obd_next_recovery_transno)
1417 target->obd_next_recovery_transno = transno;
1418 spin_unlock(&target->obd_recovery_task_lock);
1421 atomic_inc(&target->obd_req_replay_clients);
1422 atomic_inc(&target->obd_lock_replay_clients);
1423 /* Note: MDS-MDS connection is allowed to be connected during
1424 * recovery, no matter if the exports needs to be recoveried.
1425 * Because we need retrieve updates logs from all other MDTs.
1426 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1427 * also needs to be increased to match other recovery checking
1429 if (new_mds_mds_conn)
1430 target->obd_max_recoverable_clients++;
1431 if (atomic_inc_return(&target->obd_connected_clients) ==
1432 target->obd_max_recoverable_clients)
1433 wake_up(&target->obd_next_transno_waitq);
1436 /* Tell the client we're in recovery, when client is involved in it. */
1437 if (target->obd_recovering && !lw_client)
1438 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1442 spin_lock(&export->exp_lock);
1443 export->exp_connecting = 0;
1444 spin_unlock(&export->exp_lock);
1446 class_export_put(export);
1448 if (target_referenced == true && target != NULL) {
1449 spin_lock(&target->obd_dev_lock);
1450 target->obd_conn_inprogress--;
1451 spin_unlock(&target->obd_dev_lock);
1453 class_decref(target, __func__, current);
1455 req->rq_status = rc;
1459 int target_handle_disconnect(struct ptlrpc_request *req)
1464 rc = req_capsule_server_pack(&req->rq_pill);
1468 /* Keep the rq_export around so we can send the reply. */
1469 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1474 void target_destroy_export(struct obd_export *exp)
1476 struct obd_import *imp = NULL;
1477 /* exports created from last_rcvd data, and "fake"
1478 exports created by lctl don't have an import */
1479 spin_lock(&exp->exp_lock);
1480 if (exp->exp_imp_reverse != NULL) {
1481 imp = exp->exp_imp_reverse;
1482 exp->exp_imp_reverse = NULL;
1484 spin_unlock(&exp->exp_lock);
1486 client_destroy_import(imp);
1488 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1489 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1490 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1491 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1493 EXPORT_SYMBOL(target_destroy_export);
1496 * Recovery functions
1498 static void target_request_copy_get(struct ptlrpc_request *req)
1500 class_export_rpc_inc(req->rq_export);
1501 LASSERT(list_empty(&req->rq_list));
1502 INIT_LIST_HEAD(&req->rq_replay_list);
1504 /* Increase refcount to keep request in queue. */
1505 atomic_inc(&req->rq_refcount);
1506 /* Let export know it has replays to be handled. */
1507 atomic_inc(&req->rq_export->exp_replay_count);
1510 static void target_request_copy_put(struct ptlrpc_request *req)
1512 LASSERT(list_empty(&req->rq_replay_list));
1513 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1515 atomic_dec(&req->rq_export->exp_replay_count);
1516 class_export_rpc_dec(req->rq_export);
1517 ptlrpc_server_drop_request(req);
1520 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1522 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1523 struct obd_export *exp = req->rq_export;
1524 struct ptlrpc_request *reqiter;
1525 struct ptlrpc_request *dup_req = NULL;
1530 spin_lock(&exp->exp_lock);
1531 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1533 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1541 /* We expect it with RESENT and REPLAY flags. */
1542 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1543 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1544 CERROR("invalid flags %x of resent replay\n",
1545 lustre_msg_get_flags(req->rq_reqmsg));
1547 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1550 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1552 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1553 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1557 list_add_tail(&req->rq_replay_list,
1558 &exp->exp_req_replay_queue);
1561 spin_unlock(&exp->exp_lock);
1565 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1567 LASSERT(!list_empty(&req->rq_replay_list));
1568 LASSERT(req->rq_export);
1570 spin_lock(&req->rq_export->exp_lock);
1571 list_del_init(&req->rq_replay_list);
1572 spin_unlock(&req->rq_export->exp_lock);
1575 static void target_finish_recovery(struct lu_target *lut)
1577 struct obd_device *obd = lut->lut_obd;
1580 /* Only log a recovery message when recovery has occurred. */
1581 if (obd->obd_recovery_start) {
1582 time64_t now = ktime_get_real_seconds();
1583 time64_t elapsed_time;
1585 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1586 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1587 "%d recovered and %d %s evicted.\n", obd->obd_name,
1588 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1589 obd->obd_max_recoverable_clients,
1590 atomic_read(&obd->obd_connected_clients),
1591 obd->obd_stale_clients,
1592 obd->obd_stale_clients == 1 ? "was" : "were");
1595 ldlm_reprocess_recovery_done(obd->obd_namespace);
1596 spin_lock(&obd->obd_recovery_task_lock);
1597 if (!list_empty(&obd->obd_req_replay_queue) ||
1598 !list_empty(&obd->obd_lock_replay_queue) ||
1599 !list_empty(&obd->obd_final_req_queue)) {
1600 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1602 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1603 list_empty(&obd->obd_lock_replay_queue) ? \
1605 list_empty(&obd->obd_final_req_queue) ? \
1607 spin_unlock(&obd->obd_recovery_task_lock);
1610 spin_unlock(&obd->obd_recovery_task_lock);
1612 obd->obd_recovery_end = ktime_get_real_seconds();
1614 /* When recovery finished, cleanup orphans on MDS and OST. */
1615 if (OBT(obd) && OBP(obd, postrecov)) {
1616 int rc = OBP(obd, postrecov)(obd);
1618 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1624 static void abort_req_replay_queue(struct obd_device *obd)
1626 struct ptlrpc_request *req, *n;
1627 struct list_head abort_list;
1629 INIT_LIST_HEAD(&abort_list);
1630 spin_lock(&obd->obd_recovery_task_lock);
1631 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1632 spin_unlock(&obd->obd_recovery_task_lock);
1633 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1634 DEBUG_REQ(D_WARNING, req, "aborted:");
1635 req->rq_status = -ENOTCONN;
1636 if (ptlrpc_error(req)) {
1637 DEBUG_REQ(D_ERROR, req,
1638 "failed abort_req_reply; skipping");
1640 target_exp_dequeue_req_replay(req);
1641 target_request_copy_put(req);
1645 static void abort_lock_replay_queue(struct obd_device *obd)
1647 struct ptlrpc_request *req, *n;
1648 struct list_head abort_list;
1650 INIT_LIST_HEAD(&abort_list);
1651 spin_lock(&obd->obd_recovery_task_lock);
1652 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1653 spin_unlock(&obd->obd_recovery_task_lock);
1654 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1655 DEBUG_REQ(D_ERROR, req, "aborted:");
1656 req->rq_status = -ENOTCONN;
1657 if (ptlrpc_error(req)) {
1658 DEBUG_REQ(D_ERROR, req,
1659 "failed abort_lock_reply; skipping");
1661 target_request_copy_put(req);
1665 /* Called from a cleanup function if the device is being cleaned up
1666 forcefully. The exports should all have been disconnected already,
1667 the only thing left to do is
1668 - clear the recovery flags
1670 - free queued requests and replies, but don't send replies
1671 Because the obd_stopping flag is set, no new requests should be received.
1674 void target_cleanup_recovery(struct obd_device *obd)
1676 struct ptlrpc_request *req, *n;
1677 struct list_head clean_list;
1679 INIT_LIST_HEAD(&clean_list);
1680 spin_lock(&obd->obd_dev_lock);
1681 if (!obd->obd_recovering) {
1682 spin_unlock(&obd->obd_dev_lock);
1686 obd->obd_recovering = obd->obd_abort_recovery = 0;
1687 spin_unlock(&obd->obd_dev_lock);
1689 spin_lock(&obd->obd_recovery_task_lock);
1690 target_cancel_recovery_timer(obd);
1691 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1692 spin_unlock(&obd->obd_recovery_task_lock);
1694 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1695 LASSERT(req->rq_reply_state == NULL);
1696 target_exp_dequeue_req_replay(req);
1697 target_request_copy_put(req);
1700 spin_lock(&obd->obd_recovery_task_lock);
1701 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1702 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1703 spin_unlock(&obd->obd_recovery_task_lock);
1705 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1706 LASSERT(req->rq_reply_state == NULL);
1707 target_request_copy_put(req);
1712 EXPORT_SYMBOL(target_cleanup_recovery);
1714 /* obd_recovery_task_lock should be held */
1715 void target_cancel_recovery_timer(struct obd_device *obd)
1717 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1718 del_timer(&obd->obd_recovery_timer);
1721 static void target_start_recovery_timer(struct obd_device *obd)
1723 if (obd->obd_recovery_start != 0)
1726 spin_lock(&obd->obd_dev_lock);
1727 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1728 spin_unlock(&obd->obd_dev_lock);
1732 LASSERT(obd->obd_recovery_timeout != 0);
1734 if (obd->obd_recovery_start != 0) {
1735 spin_unlock(&obd->obd_dev_lock);
1739 mod_timer(&obd->obd_recovery_timer,
1740 cfs_time_shift(obd->obd_recovery_timeout));
1741 obd->obd_recovery_start = ktime_get_real_seconds();
1742 spin_unlock(&obd->obd_dev_lock);
1744 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1746 obd->obd_recovery_timeout / 60,
1747 obd->obd_recovery_timeout % 60,
1748 obd->obd_max_recoverable_clients,
1749 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1750 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1754 * extend recovery window.
1756 * if @extend is true, extend recovery window to have @drt remaining at least;
1757 * otherwise, make sure the recovery timeout value is not less than @drt.
1759 static void extend_recovery_timer(struct obd_device *obd, int drt,
1767 spin_lock(&obd->obd_dev_lock);
1768 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1769 spin_unlock(&obd->obd_dev_lock);
1772 LASSERT(obd->obd_recovery_start != 0);
1774 now = ktime_get_real_seconds();
1775 to = obd->obd_recovery_timeout;
1776 end = obd->obd_recovery_start + to;
1779 if (extend && (drt > left)) {
1781 } else if (!extend && (drt > to)) {
1785 if (to > obd->obd_recovery_time_hard) {
1786 to = obd->obd_recovery_time_hard;
1787 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1788 obd->obd_name, to, extend);
1791 if (obd->obd_recovery_timeout < to) {
1792 obd->obd_recovery_timeout = to;
1793 end = obd->obd_recovery_start + to;
1794 mod_timer(&obd->obd_recovery_timer,
1795 cfs_time_shift(end - now));
1797 spin_unlock(&obd->obd_dev_lock);
1799 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1800 obd->obd_name, (s64)(end - now));
1803 /* Reset the timer with each new client connection */
1805 * This timer is actually reconnect_timer, which is for making sure
1806 * the total recovery window is at least as big as my reconnect
1807 * attempt timing. So the initial recovery time_out will be set to
1808 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1809 * from client is bigger than this, then the recovery time_out will
1810 * be extended to make sure the client could be reconnected, in the
1811 * process, the timeout from the new client should be ignored.
1815 check_and_start_recovery_timer(struct obd_device *obd,
1816 struct ptlrpc_request *req,
1819 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1820 struct obd_device_target *obt = &obd->u.obt;
1822 if (!new_client && service_time)
1823 /* Teach server about old server's estimates, as first guess
1824 * at how long new requests will take. */
1825 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1828 target_start_recovery_timer(obd);
1830 /* Convert the service time to RPC timeout,
1831 * and reuse service_time to limit stack usage. */
1832 service_time = at_est2timeout(service_time);
1834 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1835 service_time < at_extra)
1836 service_time = at_extra;
1838 /* We expect other clients to timeout within service_time, then try
1839 * to reconnect, then try the failover server. The max delay between
1840 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1841 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1843 LASSERT(obt->obt_magic == OBT_MAGIC);
1844 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1845 if (service_time > obd->obd_recovery_timeout && !new_client)
1846 extend_recovery_timer(obd, service_time, false);
1849 /** Health checking routines */
1850 static inline int exp_connect_healthy(struct obd_export *exp)
1852 return (exp->exp_in_recovery);
1855 /** if export done req_replay or has replay in queue */
1856 static inline int exp_req_replay_healthy(struct obd_export *exp)
1858 return (!exp->exp_req_replay_needed ||
1859 atomic_read(&exp->exp_replay_count) > 0);
1863 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1865 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1866 exp_req_replay_healthy(exp);
1869 /** if export done lock_replay or has replay in queue */
1870 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1872 return (!exp->exp_lock_replay_needed ||
1873 atomic_read(&exp->exp_replay_count) > 0);
1876 static inline int exp_vbr_healthy(struct obd_export *exp)
1878 return (!exp->exp_vbr_failed);
1881 static inline int exp_finished(struct obd_export *exp)
1883 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1886 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1888 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1892 static int check_for_next_transno(struct lu_target *lut)
1894 struct ptlrpc_request *req = NULL;
1895 struct obd_device *obd = lut->lut_obd;
1896 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1897 int wake_up = 0, connected, completed, queue_len;
1898 __u64 req_transno = 0;
1899 __u64 update_transno = 0;
1900 __u64 next_transno = 0;
1903 spin_lock(&obd->obd_recovery_task_lock);
1904 if (!list_empty(&obd->obd_req_replay_queue)) {
1905 req = list_entry(obd->obd_req_replay_queue.next,
1906 struct ptlrpc_request, rq_list);
1907 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1911 update_transno = distribute_txn_get_next_transno(tdtd);
1913 connected = atomic_read(&obd->obd_connected_clients);
1914 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1915 queue_len = obd->obd_requests_queued_for_recovery;
1916 next_transno = obd->obd_next_recovery_transno;
1918 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1919 "req_transno: %llu, next_transno: %llu\n",
1920 obd->obd_max_recoverable_clients, connected, completed,
1921 queue_len, req_transno, next_transno);
1923 if (obd->obd_abort_recovery) {
1924 CDEBUG(D_HA, "waking for aborted recovery\n");
1926 } else if (obd->obd_recovery_expired) {
1927 CDEBUG(D_HA, "waking for expired recovery\n");
1929 } else if (tdtd != NULL && req != NULL &&
1930 is_req_replayed_by_update(req)) {
1931 LASSERTF(req_transno < next_transno, "req_transno %llu"
1932 "next_transno%llu\n", req_transno, next_transno);
1933 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1936 } else if (req_transno == next_transno ||
1937 (update_transno != 0 && update_transno <= next_transno)) {
1938 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1940 } else if (queue_len > 0 &&
1941 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1942 /** handle gaps occured due to lost reply or VBR */
1943 LASSERTF(req_transno >= next_transno,
1944 "req_transno: %llu, next_transno: %llu\n",
1945 req_transno, next_transno);
1947 "%s: waking for gap in transno, VBR is %s (skip: "
1948 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1949 ", next_update %lld last_committed: %lld)\n",
1950 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1951 next_transno, queue_len, completed, connected,
1952 req_transno, update_transno, obd->obd_last_committed);
1953 obd->obd_next_recovery_transno = req_transno;
1955 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1956 CDEBUG(D_HA, "waking for completed recovery\n");
1958 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1959 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1960 " by fail_lock, waking up (%lld)\n", next_transno);
1961 obd->obd_next_recovery_transno = req_transno;
1964 spin_unlock(&obd->obd_recovery_task_lock);
1968 static int check_for_next_lock(struct lu_target *lut)
1970 struct obd_device *obd = lut->lut_obd;
1973 spin_lock(&obd->obd_recovery_task_lock);
1974 if (!list_empty(&obd->obd_lock_replay_queue)) {
1975 CDEBUG(D_HA, "waking for next lock\n");
1977 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1978 CDEBUG(D_HA, "waking for completed lock replay\n");
1980 } else if (obd->obd_abort_recovery) {
1981 CDEBUG(D_HA, "waking for aborted recovery\n");
1983 } else if (obd->obd_recovery_expired) {
1984 CDEBUG(D_HA, "waking for expired recovery\n");
1987 spin_unlock(&obd->obd_recovery_task_lock);
1993 * wait for recovery events,
1994 * check its status with help of check_routine
1995 * evict dead clients via health_check
1997 static int target_recovery_overseer(struct lu_target *lut,
1998 int (*check_routine)(struct lu_target *),
1999 int (*health_check)(struct obd_export *))
2001 struct obd_device *obd = lut->lut_obd;
2002 struct target_distribute_txn_data *tdtd;
2004 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
2005 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2006 __u64 next_update_transno = 0;
2008 /* Only abort the recovery if there are no update recovery
2009 * left in the queue */
2010 spin_lock(&obd->obd_recovery_task_lock);
2011 if (lut->lut_tdtd != NULL) {
2012 next_update_transno =
2013 distribute_txn_get_next_transno(lut->lut_tdtd);
2015 tdtd = lut->lut_tdtd;
2016 /* If next_update_transno == 0, it probably because
2017 * updatelog retrieve threads did not get any records
2018 * yet, let's wait those threads stopped */
2019 if (next_update_transno == 0) {
2020 struct l_wait_info lwi = { 0 };
2022 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2024 &tdtd->tdtd_recovery_threads_count) == 0,
2027 next_update_transno =
2028 distribute_txn_get_next_transno(
2033 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2034 obd->obd_next_recovery_transno = next_update_transno;
2035 spin_unlock(&obd->obd_recovery_task_lock);
2036 /* Disconnect unfinished exports from clients, and
2037 * keep connection from MDT to make sure the update
2038 * recovery will still keep trying until some one
2039 * manually abort the recovery */
2040 class_disconnect_stale_exports(obd,
2041 exp_finished_or_from_mdt);
2042 /* Abort all of replay and replay lock req from
2044 abort_req_replay_queue(obd);
2045 abort_lock_replay_queue(obd);
2046 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2047 ")in the queue.\n", obd->obd_name,
2048 next_update_transno);
2050 obd->obd_abort_recovery = 1;
2051 spin_unlock(&obd->obd_recovery_task_lock);
2052 CWARN("%s recovery is aborted by hard timeout\n",
2057 while (wait_event_timeout(obd->obd_next_transno_waitq,
2059 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2060 /* wait indefinitely for event, but don't trigger watchdog */;
2062 if (obd->obd_abort_recovery) {
2063 CWARN("recovery is aborted, evict exports in recovery\n");
2064 if (lut->lut_tdtd != NULL) {
2065 struct l_wait_info lwi = { 0 };
2067 tdtd = lut->lut_tdtd;
2068 /* Let's wait all of the update log recovery thread
2070 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2071 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2073 /* Then abort the update recovery list */
2074 dtrq_list_destroy(lut->lut_tdtd);
2077 /** evict exports which didn't finish recovery yet */
2078 class_disconnect_stale_exports(obd, exp_finished);
2080 } else if (obd->obd_recovery_expired) {
2081 obd->obd_recovery_expired = 0;
2082 /** If some clients died being recovered, evict them */
2083 LCONSOLE_WARN("%s: recovery is timed out, "
2084 "evict stale exports\n", obd->obd_name);
2085 /** evict cexports with no replay in queue, they are stalled */
2086 class_disconnect_stale_exports(obd, health_check);
2088 /** continue with VBR */
2089 spin_lock(&obd->obd_dev_lock);
2090 obd->obd_version_recov = 1;
2091 spin_unlock(&obd->obd_dev_lock);
2093 * reset timer, recovery will proceed with versions now,
2094 * timeout is set just to handle reconnection delays
2096 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2097 /** Wait for recovery events again, after evicting bad clients */
2103 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2105 struct obd_device *obd = lut->lut_obd;
2106 struct ptlrpc_request *req = NULL;
2108 CDEBUG(D_HA, "Waiting for lock\n");
2109 if (target_recovery_overseer(lut, check_for_next_lock,
2110 exp_lock_replay_healthy))
2111 abort_lock_replay_queue(obd);
2113 spin_lock(&obd->obd_recovery_task_lock);
2114 if (!list_empty(&obd->obd_lock_replay_queue)) {
2115 req = list_entry(obd->obd_lock_replay_queue.next,
2116 struct ptlrpc_request, rq_list);
2117 list_del_init(&req->rq_list);
2118 spin_unlock(&obd->obd_recovery_task_lock);
2120 spin_unlock(&obd->obd_recovery_task_lock);
2121 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2122 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2123 /** evict exports failed VBR */
2124 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2129 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2131 struct ptlrpc_request *req = NULL;
2133 spin_lock(&obd->obd_recovery_task_lock);
2134 if (!list_empty(&obd->obd_final_req_queue)) {
2135 req = list_entry(obd->obd_final_req_queue.next,
2136 struct ptlrpc_request, rq_list);
2137 list_del_init(&req->rq_list);
2138 spin_unlock(&obd->obd_recovery_task_lock);
2139 if (req->rq_export->exp_in_recovery) {
2140 spin_lock(&req->rq_export->exp_lock);
2141 req->rq_export->exp_in_recovery = 0;
2142 spin_unlock(&req->rq_export->exp_lock);
2145 spin_unlock(&obd->obd_recovery_task_lock);
2150 static void handle_recovery_req(struct ptlrpc_thread *thread,
2151 struct ptlrpc_request *req,
2152 svc_handler_t handler)
2157 * export can be evicted during recovery, no need to handle replays for
2158 * it after that, discard such request silently
2160 if (req->rq_export->exp_disconnected)
2163 req->rq_session.lc_thread = thread;
2164 req->rq_svc_thread = thread;
2165 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2167 /* thread context */
2168 lu_context_enter(&thread->t_env->le_ctx);
2170 lu_context_exit(&thread->t_env->le_ctx);
2172 /* don't reset timer for final stage */
2173 if (!exp_finished(req->rq_export)) {
2174 int to = obd_timeout;
2177 * Add request timeout to the recovery time so next request from
2178 * this client may come in recovery time
2181 struct ptlrpc_service_part *svcpt;
2183 svcpt = req->rq_rqbd->rqbd_svcpt;
2184 /* If the server sent early reply for this request,
2185 * the client will recalculate the timeout according to
2186 * current server estimate service time, so we will
2187 * use the maxium timeout here for waiting the client
2188 * sending the next req */
2189 to = max((int)at_est2timeout(
2190 at_get(&svcpt->scp_at_estimate)),
2191 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2192 /* Add 2 net_latency, one for balance rq_deadline
2193 * (see ptl_send_rpc), one for resend the req to server,
2194 * Note: client will pack net_latency in replay req
2195 * (see ptlrpc_replay_req) */
2196 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2198 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2203 /** Checking routines for recovery */
2204 static int check_for_recovery_ready(struct lu_target *lut)
2206 struct obd_device *obd = lut->lut_obd;
2207 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2209 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2210 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2211 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2212 obd->obd_recovery_expired);
2214 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2215 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2216 if (clnts + obd->obd_stale_clients <
2217 obd->obd_max_recoverable_clients)
2221 if (lut->lut_tdtd != NULL) {
2222 if (!lut->lut_tdtd->tdtd_replay_ready &&
2223 !obd->obd_abort_recovery) {
2224 /* Let's extend recovery timer, in case the recovery
2225 * timer expired, and some clients got evicted */
2226 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2228 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2229 obd->obd_name, obd->obd_recovery_timeout);
2238 REQUEST_RECOVERY = 1,
2239 UPDATE_RECOVERY = 2,
2242 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2246 if (!list_empty(&obd->obd_req_replay_queue)) {
2247 struct ptlrpc_request *req;
2249 req = list_entry(obd->obd_req_replay_queue.next,
2250 struct ptlrpc_request, rq_list);
2251 transno = lustre_msg_get_transno(req->rq_reqmsg);
2257 static __u64 get_next_transno(struct lu_target *lut, int *type)
2259 struct obd_device *obd = lut->lut_obd;
2260 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2262 __u64 update_transno;
2265 transno = get_next_replay_req_transno(obd);
2267 *type = REQUEST_RECOVERY;
2272 update_transno = distribute_txn_get_next_transno(tdtd);
2273 if (transno == 0 || (transno >= update_transno &&
2274 update_transno != 0)) {
2275 transno = update_transno;
2277 *type = UPDATE_RECOVERY;
2284 * drop duplicate replay request
2286 * Because the operation has been replayed by update recovery, the request
2287 * with the same transno will be dropped and also notify the client to send
2288 * next replay request.
2290 * \param[in] env execution environment
2291 * \param[in] obd failover obd device
2292 * \param[in] req request to be dropped
2294 static void drop_duplicate_replay_req(struct lu_env *env,
2295 struct obd_device *obd,
2296 struct ptlrpc_request *req)
2298 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2299 " update records are found.\n",
2300 lustre_msg_get_transno(req->rq_reqmsg),
2301 libcfs_nid2str(req->rq_peer.nid));
2303 /* Right now, only for MDS reint operation update replay and
2304 * normal request replay can have the same transno */
2305 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2306 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2307 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2308 if (likely(req->rq_export))
2309 target_committed_to_req(req);
2310 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2311 target_send_reply(req, req->rq_status, 0);
2313 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2314 libcfs_nid2str(req->rq_peer.nid));
2316 target_exp_dequeue_req_replay(req);
2317 target_request_copy_put(req);
2318 obd->obd_replayed_requests++;
2321 static void replay_request_or_update(struct lu_env *env,
2322 struct lu_target *lut,
2323 struct target_recovery_data *trd,
2324 struct ptlrpc_thread *thread)
2326 struct obd_device *obd = lut->lut_obd;
2327 struct ptlrpc_request *req = NULL;
2332 CDEBUG(D_HA, "Waiting for transno %lld\n",
2333 obd->obd_next_recovery_transno);
2335 /* Replay all of request and update by transno */
2337 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2339 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2341 /** It is needed to extend recovery window above
2342 * recovery_time_soft. Extending is possible only in the
2343 * end of recovery window (see more details in
2344 * handle_recovery_req()).
2346 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2348 if (target_recovery_overseer(lut, check_for_next_transno,
2349 exp_req_replay_healthy_or_from_mdt)) {
2350 abort_req_replay_queue(obd);
2351 abort_lock_replay_queue(obd);
2355 spin_lock(&obd->obd_recovery_task_lock);
2356 transno = get_next_transno(lut, &type);
2357 if (type == REQUEST_RECOVERY && transno != 0) {
2358 /* Drop replay request from client side, if the
2359 * replay has been executed by update with the
2361 req = list_entry(obd->obd_req_replay_queue.next,
2362 struct ptlrpc_request, rq_list);
2364 list_del_init(&req->rq_list);
2365 obd->obd_requests_queued_for_recovery--;
2366 spin_unlock(&obd->obd_recovery_task_lock);
2368 /* Let's check if the request has been redone by
2370 if (is_req_replayed_by_update(req)) {
2371 struct distribute_txn_replay_req *dtrq;
2373 dtrq = distribute_txn_lookup_finish_list(tdtd,
2375 LASSERT(dtrq != NULL);
2376 spin_lock(&tdtd->tdtd_replay_list_lock);
2377 list_del_init(&dtrq->dtrq_list);
2378 spin_unlock(&tdtd->tdtd_replay_list_lock);
2381 drop_duplicate_replay_req(env, obd, req);
2386 LASSERT(trd->trd_processing_task == current_pid());
2387 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2388 lustre_msg_get_transno(req->rq_reqmsg),
2389 libcfs_nid2str(req->rq_peer.nid));
2391 handle_recovery_req(thread, req,
2392 trd->trd_recovery_handler);
2394 * bz18031: increase next_recovery_transno before
2395 * target_request_copy_put() will drop exp_rpc reference
2397 spin_lock(&obd->obd_recovery_task_lock);
2398 obd->obd_next_recovery_transno++;
2399 spin_unlock(&obd->obd_recovery_task_lock);
2400 target_exp_dequeue_req_replay(req);
2401 target_request_copy_put(req);
2402 obd->obd_replayed_requests++;
2403 } else if (type == UPDATE_RECOVERY && transno != 0) {
2404 struct distribute_txn_replay_req *dtrq;
2407 spin_unlock(&obd->obd_recovery_task_lock);
2409 LASSERT(tdtd != NULL);
2410 dtrq = distribute_txn_get_next_req(tdtd);
2411 lu_context_enter(&thread->t_env->le_ctx);
2412 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2413 lu_context_exit(&thread->t_env->le_ctx);
2414 extend_recovery_timer(obd, obd_timeout, true);
2416 if (rc == 0 && dtrq->dtrq_xid != 0) {
2417 CDEBUG(D_HA, "Move x%llu t%llu"
2418 " to finish list\n", dtrq->dtrq_xid,
2419 dtrq->dtrq_master_transno);
2421 /* Add it to the replay finish list */
2422 spin_lock(&tdtd->tdtd_replay_list_lock);
2423 list_add(&dtrq->dtrq_list,
2424 &tdtd->tdtd_replay_finish_list);
2425 spin_unlock(&tdtd->tdtd_replay_list_lock);
2427 spin_lock(&obd->obd_recovery_task_lock);
2428 if (transno == obd->obd_next_recovery_transno)
2429 obd->obd_next_recovery_transno++;
2431 obd->obd_next_recovery_transno)
2432 obd->obd_next_recovery_transno =
2434 spin_unlock(&obd->obd_recovery_task_lock);
2439 spin_unlock(&obd->obd_recovery_task_lock);
2441 LASSERT(list_empty(&obd->obd_req_replay_queue));
2442 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2443 /** evict exports failed VBR */
2444 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2450 static int target_recovery_thread(void *arg)
2452 struct lu_target *lut = arg;
2453 struct obd_device *obd = lut->lut_obd;
2454 struct ptlrpc_request *req;
2455 struct target_recovery_data *trd = &obd->obd_recovery_data;
2456 unsigned long delta;
2458 struct ptlrpc_thread *thread = NULL;
2462 unshare_fs_struct();
2463 OBD_ALLOC_PTR(thread);
2469 OBD_FREE_PTR(thread);
2473 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2475 OBD_FREE_PTR(thread);
2480 thread->t_env = env;
2481 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2482 env->le_ctx.lc_thread = thread;
2483 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2484 thread->t_watchdog = NULL;
2486 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2488 trd->trd_processing_task = current_pid();
2490 spin_lock(&obd->obd_dev_lock);
2491 obd->obd_recovering = 1;
2492 spin_unlock(&obd->obd_dev_lock);
2493 complete(&trd->trd_starting);
2495 /* first of all, we have to know the first transno to replay */
2496 if (target_recovery_overseer(lut, check_for_recovery_ready,
2497 exp_connect_healthy)) {
2498 abort_req_replay_queue(obd);
2499 abort_lock_replay_queue(obd);
2500 if (lut->lut_tdtd != NULL)
2501 dtrq_list_destroy(lut->lut_tdtd);
2504 /* next stage: replay requests or update */
2506 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2507 atomic_read(&obd->obd_req_replay_clients),
2508 obd->obd_next_recovery_transno);
2509 replay_request_or_update(env, lut, trd, thread);
2512 * The second stage: replay locks
2514 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2515 atomic_read(&obd->obd_lock_replay_clients));
2516 while ((req = target_next_replay_lock(lut))) {
2517 LASSERT(trd->trd_processing_task == current_pid());
2518 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2519 libcfs_nid2str(req->rq_peer.nid));
2520 handle_recovery_req(thread, req,
2521 trd->trd_recovery_handler);
2522 target_request_copy_put(req);
2523 obd->obd_replayed_locks++;
2527 * The third stage: reply on final pings, at this moment all clients
2528 * must have request in final queue
2530 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2531 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2532 /** Update server last boot epoch */
2533 tgt_boot_epoch_update(lut);
2534 /* We drop recoverying flag to forward all new requests
2535 * to regular mds_handle() since now */
2536 spin_lock(&obd->obd_dev_lock);
2537 obd->obd_recovering = obd->obd_abort_recovery = 0;
2538 spin_unlock(&obd->obd_dev_lock);
2539 spin_lock(&obd->obd_recovery_task_lock);
2540 target_cancel_recovery_timer(obd);
2541 spin_unlock(&obd->obd_recovery_task_lock);
2542 while ((req = target_next_final_ping(obd))) {
2543 LASSERT(trd->trd_processing_task == current_pid());
2544 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2545 libcfs_nid2str(req->rq_peer.nid));
2546 handle_recovery_req(thread, req,
2547 trd->trd_recovery_handler);
2548 /* Because the waiting client can not send ping to server,
2549 * so we need refresh the last_request_time, to avoid the
2550 * export is being evicted */
2551 ptlrpc_update_export_timer(req->rq_export, 0);
2552 target_request_copy_put(req);
2555 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2556 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2557 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2558 if (delta > OBD_RECOVERY_TIME_SOFT) {
2559 CWARN("too long recovery - read logs\n");
2560 libcfs_debug_dumplog();
2563 target_finish_recovery(lut);
2565 lu_context_fini(&env->le_ctx);
2566 trd->trd_processing_task = 0;
2567 complete(&trd->trd_finishing);
2569 tgt_io_thread_done(thread);
2570 OBD_FREE_PTR(thread);
2575 static int target_start_recovery_thread(struct lu_target *lut,
2576 svc_handler_t handler)
2578 struct obd_device *obd = lut->lut_obd;
2580 struct target_recovery_data *trd = &obd->obd_recovery_data;
2583 memset(trd, 0, sizeof(*trd));
2584 init_completion(&trd->trd_starting);
2585 init_completion(&trd->trd_finishing);
2586 trd->trd_recovery_handler = handler;
2588 rc = server_name2index(obd->obd_name, &index, NULL);
2592 if (!IS_ERR(kthread_run(target_recovery_thread,
2593 lut, "tgt_recover_%d", index))) {
2594 wait_for_completion(&trd->trd_starting);
2595 LASSERT(obd->obd_recovering != 0);
2603 void target_stop_recovery_thread(struct obd_device *obd)
2605 if (obd->obd_recovery_data.trd_processing_task > 0) {
2606 struct target_recovery_data *trd = &obd->obd_recovery_data;
2607 /** recovery can be done but postrecovery is not yet */
2608 spin_lock(&obd->obd_dev_lock);
2609 if (obd->obd_recovering) {
2610 CERROR("%s: Aborting recovery\n", obd->obd_name);
2611 obd->obd_abort_recovery = 1;
2612 wake_up(&obd->obd_next_transno_waitq);
2614 spin_unlock(&obd->obd_dev_lock);
2615 wait_for_completion(&trd->trd_finishing);
2618 EXPORT_SYMBOL(target_stop_recovery_thread);
2620 void target_recovery_fini(struct obd_device *obd)
2622 class_disconnect_exports(obd);
2623 target_stop_recovery_thread(obd);
2624 target_cleanup_recovery(obd);
2626 EXPORT_SYMBOL(target_recovery_fini);
2628 static void target_recovery_expired(unsigned long castmeharder)
2630 struct obd_device *obd = (struct obd_device *)castmeharder;
2631 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2632 " after %llus (%d clients connected)\n",
2633 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2634 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2635 atomic_read(&obd->obd_connected_clients));
2637 obd->obd_recovery_expired = 1;
2638 wake_up(&obd->obd_next_transno_waitq);
2641 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2643 struct obd_device *obd = lut->lut_obd;
2645 if (lut->lut_bottom->dd_rdonly)
2648 if (obd->obd_max_recoverable_clients == 0) {
2649 /** Update server last boot epoch */
2650 tgt_boot_epoch_update(lut);
2654 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2655 "last_transno %llu\n", obd->obd_name,
2656 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2657 LASSERT(obd->obd_stopping == 0);
2658 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2659 obd->obd_recovery_start = 0;
2660 obd->obd_recovery_end = 0;
2662 setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2663 (unsigned long)obd);
2664 target_start_recovery_thread(lut, handler);
2666 EXPORT_SYMBOL(target_recovery_init);
2668 static int target_process_req_flags(struct obd_device *obd,
2669 struct ptlrpc_request *req)
2671 struct obd_export *exp = req->rq_export;
2672 LASSERT(exp != NULL);
2673 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2674 /* client declares he's ready to replay locks */
2675 spin_lock(&exp->exp_lock);
2676 if (exp->exp_req_replay_needed) {
2677 exp->exp_req_replay_needed = 0;
2678 spin_unlock(&exp->exp_lock);
2680 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2681 atomic_dec(&obd->obd_req_replay_clients);
2683 spin_unlock(&exp->exp_lock);
2686 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2687 /* client declares he's ready to complete recovery
2688 * so, we put the request on th final queue */
2689 spin_lock(&exp->exp_lock);
2690 if (exp->exp_lock_replay_needed) {
2691 exp->exp_lock_replay_needed = 0;
2692 spin_unlock(&exp->exp_lock);
2694 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2695 atomic_dec(&obd->obd_lock_replay_clients);
2697 spin_unlock(&exp->exp_lock);
2703 int target_queue_recovery_request(struct ptlrpc_request *req,
2704 struct obd_device *obd)
2706 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2707 struct ptlrpc_request *reqiter;
2711 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2712 /* Processing the queue right now, don't re-add. */
2716 target_process_req_flags(obd, req);
2718 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2719 /* client declares he's ready to complete recovery
2720 * so, we put the request on th final queue */
2721 target_request_copy_get(req);
2722 DEBUG_REQ(D_HA, req, "queue final req");
2723 wake_up(&obd->obd_next_transno_waitq);
2724 spin_lock(&obd->obd_recovery_task_lock);
2725 if (obd->obd_recovering) {
2726 struct ptlrpc_request *tmp;
2727 struct ptlrpc_request *duplicate = NULL;
2729 if (likely(!req->rq_export->exp_replay_done)) {
2730 req->rq_export->exp_replay_done = 1;
2731 list_add_tail(&req->rq_list,
2732 &obd->obd_final_req_queue);
2733 spin_unlock(&obd->obd_recovery_task_lock);
2737 /* XXX O(n), but only happens if final ping is
2738 * timed out, probably reorganize the list as
2739 * a hash list later */
2740 list_for_each_entry_safe(reqiter, tmp,
2741 &obd->obd_final_req_queue,
2743 if (reqiter->rq_export == req->rq_export) {
2744 list_del_init(&reqiter->rq_list);
2745 duplicate = reqiter;
2750 list_add_tail(&req->rq_list,
2751 &obd->obd_final_req_queue);
2752 req->rq_export->exp_replay_done = 1;
2753 spin_unlock(&obd->obd_recovery_task_lock);
2755 if (duplicate != NULL) {
2756 DEBUG_REQ(D_HA, duplicate,
2757 "put prev final req\n");
2758 target_request_copy_put(duplicate);
2762 spin_unlock(&obd->obd_recovery_task_lock);
2763 target_request_copy_put(req);
2764 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2767 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2768 /* client declares he's ready to replay locks */
2769 target_request_copy_get(req);
2770 DEBUG_REQ(D_HA, req, "queue lock replay req");
2771 wake_up(&obd->obd_next_transno_waitq);
2772 spin_lock(&obd->obd_recovery_task_lock);
2773 LASSERT(obd->obd_recovering);
2774 /* usually due to recovery abort */
2775 if (!req->rq_export->exp_in_recovery) {
2776 spin_unlock(&obd->obd_recovery_task_lock);
2777 target_request_copy_put(req);
2780 LASSERT(req->rq_export->exp_lock_replay_needed);
2781 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2782 spin_unlock(&obd->obd_recovery_task_lock);
2786 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2787 * (i.e. buflens etc are in my own byte order), but type-dependent
2788 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2791 INIT_LIST_HEAD(&req->rq_list);
2792 DEBUG_REQ(D_HA, req, "not queueing");
2796 /* If we're processing the queue, we want don't want to queue this
2799 * Also, if this request has a transno less than the one we're waiting
2800 * for, we should process it now. It could (and currently always will)
2801 * be an open request for a descriptor that was opened some time ago.
2803 * Also, a resent, replayed request that has already been
2804 * handled will pass through here and be processed immediately.
2806 CDEBUG(D_HA, "Next recovery transno: %llu"
2807 ", current: %llu, replaying\n",
2808 obd->obd_next_recovery_transno, transno);
2810 /* If the request has been replayed by update replay, then sends this
2811 * request to the recovery thread (replay_request_or_update()), where
2812 * it will be handled */
2813 spin_lock(&obd->obd_recovery_task_lock);
2814 if (transno < obd->obd_next_recovery_transno &&
2815 !is_req_replayed_by_update(req)) {
2816 /* Processing the queue right now, don't re-add. */
2817 LASSERT(list_empty(&req->rq_list));
2818 spin_unlock(&obd->obd_recovery_task_lock);
2821 spin_unlock(&obd->obd_recovery_task_lock);
2823 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2826 target_request_copy_get(req);
2827 if (!req->rq_export->exp_in_recovery) {
2828 target_request_copy_put(req);
2831 LASSERT(req->rq_export->exp_req_replay_needed);
2833 if (target_exp_enqueue_req_replay(req)) {
2834 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2835 target_request_copy_put(req);
2840 spin_lock(&obd->obd_recovery_task_lock);
2841 LASSERT(obd->obd_recovering);
2842 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2843 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2844 list_add_tail(&req->rq_list, &reqiter->rq_list);
2849 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2851 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2852 "has been claimed by another client");
2853 spin_unlock(&obd->obd_recovery_task_lock);
2854 target_exp_dequeue_req_replay(req);
2855 target_request_copy_put(req);
2861 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2863 obd->obd_requests_queued_for_recovery++;
2864 spin_unlock(&obd->obd_recovery_task_lock);
2865 wake_up(&obd->obd_next_transno_waitq);
2869 int target_handle_ping(struct ptlrpc_request *req)
2871 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2872 return req_capsule_server_pack(&req->rq_pill);
2875 void target_committed_to_req(struct ptlrpc_request *req)
2877 struct obd_export *exp = req->rq_export;
2879 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2880 lustre_msg_set_last_committed(req->rq_repmsg,
2881 exp->exp_last_committed);
2883 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2884 "%d)", exp->exp_obd->obd_no_transno,
2885 req->rq_repmsg == NULL);
2887 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2888 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2891 #endif /* HAVE_SERVER_SUPPORT */
2894 * Packs current SLV and Limit into \a req.
2896 int target_pack_pool_reply(struct ptlrpc_request *req)
2898 struct obd_device *obd;
2901 /* Check that we still have all structures alive as this may
2902 * be some late RPC at shutdown time. */
2903 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2904 !exp_connect_lru_resize(req->rq_export))) {
2905 lustre_msg_set_slv(req->rq_repmsg, 0);
2906 lustre_msg_set_limit(req->rq_repmsg, 0);
2910 /* OBD is alive here as export is alive, which we checked above. */
2911 obd = req->rq_export->exp_obd;
2913 read_lock(&obd->obd_pool_lock);
2914 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2915 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2916 read_unlock(&obd->obd_pool_lock);
2921 static int target_send_reply_msg(struct ptlrpc_request *req,
2922 int rc, int fail_id)
2924 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2925 DEBUG_REQ(D_ERROR, req, "dropping reply");
2928 /* We can have a null rq_reqmsg in the event of bad signature or
2929 * no context when unwrapping */
2930 if (req->rq_reqmsg &&
2931 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2932 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2936 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2937 req->rq_status = rc;
2938 return ptlrpc_send_error(req, 1);
2940 DEBUG_REQ(D_NET, req, "sending reply");
2943 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2946 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2948 struct ptlrpc_service_part *svcpt;
2950 struct ptlrpc_reply_state *rs;
2951 struct obd_export *exp;
2954 if (req->rq_no_reply) {
2959 svcpt = req->rq_rqbd->rqbd_svcpt;
2960 rs = req->rq_reply_state;
2961 if (rs == NULL || !rs->rs_difficult) {
2963 target_send_reply_msg (req, rc, fail_id);
2968 /* must be an export if locks saved */
2969 LASSERT(req->rq_export != NULL);
2970 /* req/reply consistent */
2971 LASSERT(rs->rs_svcpt == svcpt);
2974 LASSERT(!rs->rs_scheduled);
2975 LASSERT(!rs->rs_scheduled_ever);
2976 LASSERT(!rs->rs_handled);
2977 LASSERT(!rs->rs_on_net);
2978 LASSERT(rs->rs_export == NULL);
2979 LASSERT(list_empty(&rs->rs_obd_list));
2980 LASSERT(list_empty(&rs->rs_exp_list));
2982 exp = class_export_get(req->rq_export);
2984 /* disable reply scheduling while I'm setting up */
2985 rs->rs_scheduled = 1;
2987 rs->rs_xid = req->rq_xid;
2988 rs->rs_transno = req->rq_transno;
2989 rs->rs_export = exp;
2990 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2992 spin_lock(&exp->exp_uncommitted_replies_lock);
2993 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2994 rs->rs_transno, exp->exp_last_committed);
2995 if (rs->rs_transno > exp->exp_last_committed) {
2996 /* not committed already */
2997 list_add_tail(&rs->rs_obd_list,
2998 &exp->exp_uncommitted_replies);
3000 spin_unlock(&exp->exp_uncommitted_replies_lock);
3002 spin_lock(&exp->exp_lock);
3003 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3004 spin_unlock(&exp->exp_lock);
3006 netrc = target_send_reply_msg(req, rc, fail_id);
3008 spin_lock(&svcpt->scp_rep_lock);
3010 atomic_inc(&svcpt->scp_nreps_difficult);
3013 /* error sending: reply is off the net. Also we need +1
3014 * reply ref until ptlrpc_handle_rs() is done
3015 * with the reply state (if the send was successful, there
3016 * would have been +1 ref for the net, which
3017 * reply_out_callback leaves alone) */
3019 ptlrpc_rs_addref(rs);
3022 spin_lock(&rs->rs_lock);
3023 if (rs->rs_transno <= exp->exp_last_committed ||
3024 (!rs->rs_on_net && !rs->rs_no_ack) ||
3025 list_empty(&rs->rs_exp_list) || /* completed already */
3026 list_empty(&rs->rs_obd_list)) {
3027 CDEBUG(D_HA, "Schedule reply immediately\n");
3028 ptlrpc_dispatch_difficult_reply(rs);
3030 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3031 rs->rs_scheduled = 0; /* allow notifier to schedule */
3033 spin_unlock(&rs->rs_lock);
3034 spin_unlock(&svcpt->scp_rep_lock);
3038 enum ldlm_mode lck_compat_array[] = {
3039 [LCK_EX] = LCK_COMPAT_EX,
3040 [LCK_PW] = LCK_COMPAT_PW,
3041 [LCK_PR] = LCK_COMPAT_PR,
3042 [LCK_CW] = LCK_COMPAT_CW,
3043 [LCK_CR] = LCK_COMPAT_CR,
3044 [LCK_NL] = LCK_COMPAT_NL,
3045 [LCK_GROUP] = LCK_COMPAT_GROUP,
3046 [LCK_COS] = LCK_COMPAT_COS,
3050 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3051 * not escape to the user level.
3053 int ldlm_error2errno(enum ldlm_error error)
3059 case ELDLM_LOCK_MATCHED:
3062 case ELDLM_LOCK_CHANGED:
3065 case ELDLM_LOCK_ABORTED:
3068 case ELDLM_LOCK_REPLACED:
3071 case ELDLM_NO_LOCK_DATA:
3074 case ELDLM_NAMESPACE_EXISTS:
3077 case ELDLM_BAD_NAMESPACE:
3081 if (((int)error) < 0) { /* cast to signed type */
3082 result = error; /* as ldlm_error can be unsigned */
3084 CERROR("Invalid DLM result code: %d\n", error);
3090 EXPORT_SYMBOL(ldlm_error2errno);
3093 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3095 enum ldlm_error ldlm_errno2error(int err_no)
3104 error = ELDLM_LOCK_CHANGED;
3107 error = ELDLM_LOCK_ABORTED;
3110 error = ELDLM_LOCK_REPLACED;
3113 error = ELDLM_NO_LOCK_DATA;
3116 error = ELDLM_NAMESPACE_EXISTS;
3119 error = ELDLM_BAD_NAMESPACE;
3127 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3128 void ldlm_dump_export_locks(struct obd_export *exp)
3130 spin_lock(&exp->exp_locks_list_guard);
3131 if (!list_empty(&exp->exp_locks_list)) {
3132 struct ldlm_lock *lock;
3134 CERROR("dumping locks for export %p,"
3135 "ignore if the unmount doesn't hang\n", exp);
3136 list_for_each_entry(lock, &exp->exp_locks_list,
3138 LDLM_ERROR(lock, "lock:");
3140 spin_unlock(&exp->exp_locks_list_guard);
3144 #ifdef HAVE_SERVER_SUPPORT
3145 static int target_bulk_timeout(void *data)
3148 /* We don't fail the connection here, because having the export
3149 * killed makes the (vital) call to commitrw very sad.
3154 static inline const char *bulk2type(struct ptlrpc_request *req)
3156 if (req->rq_bulk_read)
3158 if (req->rq_bulk_write)
3163 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3164 struct l_wait_info *lwi)
3166 struct ptlrpc_request *req = desc->bd_req;
3167 time_t start = cfs_time_current_sec();
3173 /* If there is eviction in progress, wait for it to finish. */
3174 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3175 *lwi = LWI_INTR(NULL, NULL);
3176 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3177 !atomic_read(&exp->exp_obd->
3178 obd_evict_inprogress),
3182 /* Check if client was evicted or reconnected already. */
3183 if (exp->exp_failed ||
3184 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3187 if (req->rq_bulk_read)
3188 rc = sptlrpc_svc_wrap_bulk(req, desc);
3190 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3191 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3192 else /* old version, bulk matchbits is rq_xid */
3193 req->rq_mbits = req->rq_xid;
3196 rc = ptlrpc_start_bulk_transfer(desc);
3200 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3201 bulk2type(req), rc);
3205 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3206 ptlrpc_abort_bulk(desc);
3210 /* limit actual bulk transfer to bulk_timeout seconds */
3211 deadline = start + bulk_timeout;
3212 if (deadline > req->rq_deadline)
3213 deadline = req->rq_deadline;
3216 long timeoutl = deadline - cfs_time_current_sec();
3217 cfs_duration_t timeout = timeoutl <= 0 ?
3218 CFS_TICK : cfs_time_seconds(timeoutl);
3221 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3222 target_bulk_timeout, desc);
3223 rc = l_wait_event(desc->bd_waitq,
3224 !ptlrpc_server_bulk_active(desc) ||
3227 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3229 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3230 /* Wait again if we changed rq_deadline. */
3231 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3232 deadline = start + bulk_timeout;
3233 if (deadline > rq_deadline)
3234 deadline = rq_deadline;
3235 } while ((rc == -ETIMEDOUT) &&
3236 (deadline > cfs_time_current_sec()));
3238 if (rc == -ETIMEDOUT) {
3239 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3240 bulk2type(req), deadline - start,
3241 cfs_time_current_sec() - deadline);
3242 ptlrpc_abort_bulk(desc);
3243 } else if (exp->exp_failed) {
3244 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3247 ptlrpc_abort_bulk(desc);
3248 } else if (exp->exp_conn_cnt >
3249 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3250 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3252 /* We don't reply anyway. */
3254 ptlrpc_abort_bulk(desc);
3255 } else if (desc->bd_failure) {
3256 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3258 /* XXX should this be a different errno? */
3261 if (req->rq_bulk_write)
3262 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3263 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3264 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3265 bulk2type(req), desc->bd_nob_transferred,
3267 /* XXX should this be a different errno? */
3274 EXPORT_SYMBOL(target_bulk_io);
3276 #endif /* HAVE_SERVER_SUPPORT */