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_waitqueue_head(&cli->cl_destroy_waitq);
395 atomic_set(&cli->cl_destroy_in_flight, 0);
396 #ifdef ENABLE_CHECKSUM
397 /* Turn on checksumming by default. */
398 cli->cl_checksum = 1;
400 * The supported checksum types will be worked out at connect time
401 * Set cl_chksum* to CRC32 for now to avoid returning screwed info
404 cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
406 atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
408 /* Set it to possible maximum size. It may be reduced by ocd_brw_size
409 * from OFD after connecting. */
410 cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
412 cli->cl_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
414 /* set cl_chunkbits default value to PAGE_SHIFT,
415 * it will be updated at OSC connection time. */
416 cli->cl_chunkbits = PAGE_SHIFT;
418 if (!strcmp(name, LUSTRE_MDC_NAME)) {
419 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
420 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
421 cli->cl_max_rpcs_in_flight = 2;
422 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
423 cli->cl_max_rpcs_in_flight = 3;
424 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
425 cli->cl_max_rpcs_in_flight = 4;
427 if (osc_on_mdt(obddev->obd_name))
428 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
430 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
433 spin_lock_init(&cli->cl_mod_rpcs_lock);
434 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
435 cli->cl_max_mod_rpcs_in_flight = 0;
436 cli->cl_mod_rpcs_in_flight = 0;
437 cli->cl_close_rpcs_in_flight = 0;
438 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
439 cli->cl_mod_tag_bitmap = NULL;
441 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
443 if (connect_op == MDS_CONNECT) {
444 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
445 OBD_ALLOC(cli->cl_mod_tag_bitmap,
446 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
447 if (cli->cl_mod_tag_bitmap == NULL)
448 GOTO(err, rc = -ENOMEM);
453 CERROR("ldlm_get_ref failed: %d\n", rc);
457 ptlrpc_init_client(rq_portal, rp_portal, name,
458 &obddev->obd_ldlm_client);
460 imp = class_new_import(obddev);
462 GOTO(err_ldlm, rc = -ENOENT);
463 imp->imp_client = &obddev->obd_ldlm_client;
464 imp->imp_connect_op = connect_op;
465 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
466 LUSTRE_CFG_BUFLEN(lcfg, 1));
467 class_import_put(imp);
469 if (lustre_cfg_buf(lcfg, 4)) {
470 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
472 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
474 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
475 obddev->obd_name, lustre_cfg_string(lcfg, 4),
477 GOTO(err_import, rc);
479 fake_conn.c_self = LNET_MKNID(refnet, 0);
480 imp->imp_connection = &fake_conn;
483 rc = client_import_add_conn(imp, &server_uuid, 1);
485 CERROR("can't add initial connection\n");
486 GOTO(err_import, rc);
488 imp->imp_connection = NULL;
490 cli->cl_import = imp;
491 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
492 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
494 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
495 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
496 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
497 name, obddev->obd_name,
498 cli->cl_target_uuid.uuid);
499 spin_lock(&imp->imp_lock);
500 imp->imp_deactive = 1;
501 spin_unlock(&imp->imp_lock);
505 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
506 LDLM_NAMESPACE_CLIENT,
507 LDLM_NAMESPACE_GREEDY,
509 if (obddev->obd_namespace == NULL) {
510 CERROR("Unable to create client namespace - %s\n",
512 GOTO(err_import, rc = -ENOMEM);
518 class_destroy_import(imp);
522 if (cli->cl_mod_tag_bitmap != NULL)
523 OBD_FREE(cli->cl_mod_tag_bitmap,
524 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
525 cli->cl_mod_tag_bitmap = NULL;
529 EXPORT_SYMBOL(client_obd_setup);
531 int client_obd_cleanup(struct obd_device *obddev)
533 struct client_obd *cli = &obddev->u.cli;
536 ldlm_namespace_free_post(obddev->obd_namespace);
537 obddev->obd_namespace = NULL;
539 obd_cleanup_client_import(obddev);
540 LASSERT(obddev->u.cli.cl_import == NULL);
544 if (cli->cl_mod_tag_bitmap != NULL)
545 OBD_FREE(cli->cl_mod_tag_bitmap,
546 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
547 cli->cl_mod_tag_bitmap = NULL;
551 EXPORT_SYMBOL(client_obd_cleanup);
553 /* ->o_connect() method for client side (OSC and MDC and MGC) */
554 int client_connect_import(const struct lu_env *env,
555 struct obd_export **exp,
556 struct obd_device *obd, struct obd_uuid *cluuid,
557 struct obd_connect_data *data, void *localdata)
559 struct client_obd *cli = &obd->u.cli;
560 struct obd_import *imp = cli->cl_import;
561 struct obd_connect_data *ocd;
562 struct lustre_handle conn = { 0 };
567 down_write(&cli->cl_sem);
568 if (cli->cl_conn_count > 0)
569 GOTO(out_sem, rc = -EALREADY);
571 rc = class_connect(&conn, obd, cluuid);
575 cli->cl_conn_count++;
576 *exp = class_conn2export(&conn);
578 LASSERT(obd->obd_namespace);
580 imp->imp_dlm_handle = conn;
581 rc = ptlrpc_init_import(imp);
585 ocd = &imp->imp_connect_data;
588 imp->imp_connect_flags_orig = data->ocd_connect_flags;
589 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
592 rc = ptlrpc_connect_import(imp);
594 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
597 LASSERT(*exp != NULL && (*exp)->exp_connection);
600 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
601 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
602 data->ocd_connect_flags, ocd->ocd_connect_flags);
603 data->ocd_connect_flags = ocd->ocd_connect_flags;
604 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
607 ptlrpc_pinger_add_import(imp);
613 cli->cl_conn_count--;
614 class_disconnect(*exp);
618 up_write(&cli->cl_sem);
622 EXPORT_SYMBOL(client_connect_import);
624 int client_disconnect_export(struct obd_export *exp)
626 struct obd_device *obd = class_exp2obd(exp);
627 struct client_obd *cli;
628 struct obd_import *imp;
633 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
634 exp, exp ? exp->exp_handle.h_cookie : -1);
639 imp = cli->cl_import;
641 down_write(&cli->cl_sem);
642 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
645 if (cli->cl_conn_count == 0) {
646 CERROR("disconnecting disconnected device (%s)\n",
648 GOTO(out_disconnect, rc = -EINVAL);
651 cli->cl_conn_count--;
652 if (cli->cl_conn_count != 0)
653 GOTO(out_disconnect, rc = 0);
655 /* Mark import deactivated now, so we don't try to reconnect if any
656 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
657 * fully deactivate the import, or that would drop all requests. */
658 spin_lock(&imp->imp_lock);
659 imp->imp_deactive = 1;
660 spin_unlock(&imp->imp_lock);
662 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
663 * delete it regardless. (It's safe to delete an import that was
665 (void)ptlrpc_pinger_del_import(imp);
667 if (obd->obd_namespace != NULL) {
668 /* obd_force == local only */
669 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
670 obd->obd_force ? LCF_LOCAL : 0, NULL);
671 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
674 /* There's no need to hold sem while disconnecting an import,
675 * and it may actually cause deadlock in GSS. */
676 up_write(&cli->cl_sem);
677 rc = ptlrpc_disconnect_import(imp, 0);
678 down_write(&cli->cl_sem);
680 ptlrpc_invalidate_import(imp);
685 /* Use server style - class_disconnect should be always called for
687 err = class_disconnect(exp);
691 up_write(&cli->cl_sem);
695 EXPORT_SYMBOL(client_disconnect_export);
697 #ifdef HAVE_SERVER_SUPPORT
698 int server_disconnect_export(struct obd_export *exp)
703 /* Disconnect early so that clients can't keep using export. */
704 rc = class_disconnect(exp);
705 /* Close import to avoid sending any requests. */
706 if (exp->exp_imp_reverse)
707 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
709 ldlm_bl_thread_wakeup();
711 /* complete all outstanding replies */
712 spin_lock(&exp->exp_lock);
713 while (!list_empty(&exp->exp_outstanding_replies)) {
714 struct ptlrpc_reply_state *rs =
715 list_entry(exp->exp_outstanding_replies.next,
716 struct ptlrpc_reply_state, rs_exp_list);
717 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
719 spin_lock(&svcpt->scp_rep_lock);
721 list_del_init(&rs->rs_exp_list);
723 spin_lock(&rs->rs_lock);
724 /* clear rs_convert_lock to make sure rs is handled and put */
725 rs->rs_convert_lock = 0;
726 ptlrpc_schedule_difficult_reply(rs);
727 spin_unlock(&rs->rs_lock);
729 spin_unlock(&svcpt->scp_rep_lock);
731 spin_unlock(&exp->exp_lock);
735 EXPORT_SYMBOL(server_disconnect_export);
737 /* --------------------------------------------------------------------------
738 * from old lib/target.c
739 * -------------------------------------------------------------------------- */
741 static int target_handle_reconnect(struct lustre_handle *conn,
742 struct obd_export *exp,
743 struct obd_uuid *cluuid)
745 struct obd_device *target;
746 struct lustre_handle *hdl;
753 hdl = &exp->exp_imp_reverse->imp_remote_handle;
754 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
755 conn->cookie = exp->exp_handle.h_cookie;
756 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
757 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
761 target = exp->exp_obd;
763 /* Might be a re-connect after a partition. */
764 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
765 LCONSOLE_WARN("%s: already connected client %s (at %s) "
766 "with handle %#llx. Rejecting client "
767 "with the same UUID trying to reconnect "
768 "with handle %#llx\n", target->obd_name,
769 obd_uuid2str(&exp->exp_client_uuid),
770 obd_export_nid2str(exp),
771 hdl->cookie, conn->cookie);
772 memset(conn, 0, sizeof *conn);
773 /* target_handle_connect() treats EALREADY and
774 * -EALREADY differently. -EALREADY is an error
775 * (same UUID, different handle). */
779 if (!target->obd_recovering) {
780 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
781 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
782 obd_export_nid2str(exp));
783 GOTO(out_already, rc);
786 now = ktime_get_seconds();
787 deadline = cfs_duration_sec(target->obd_recovery_timer.expires);
788 if (now < deadline) {
789 struct target_distribute_txn_data *tdtd;
794 timeout = deadline - now;
795 tdtd = class_exp2tgt(exp)->lut_tdtd;
796 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
797 buf = tdtd->tdtd_show_update_logs_retrievers(
798 tdtd->tdtd_show_retrievers_cbdata,
802 LCONSOLE_WARN("%s: Recovery already passed deadline "
803 "%lld:%.02lld. It is due to DNE recovery "
804 "failed/stuck on the %d MDT(s):%s. "
805 "Please wait until all MDTs recovered "
806 "or abort the recovery by force.\n",
807 target->obd_name, timeout / 60,
809 buf ? buf : "unknown (not enough RAM)");
811 LCONSOLE_WARN("%s: Recovery already passed deadline "
812 "%lld:%.02lld. If you do not want to wait "
813 "more, please abort the recovery by "
814 "force.\n", target->obd_name,
815 timeout / 60, timeout % 60);
820 timeout = now - deadline;
821 LCONSOLE_WARN("%s: Recovery already passed deadline"
822 " %lld:%.02lld, It is most likely due to DNE"
823 " recovery is failed or stuck, please wait a"
824 " few more minutes or abort the recovery.\n",
825 target->obd_name, timeout / 60, timeout % 60);
829 conn->cookie = exp->exp_handle.h_cookie;
830 /* target_handle_connect() treats EALREADY and
831 * -EALREADY differently. EALREADY means we are
832 * doing a valid reconnect from the same client. */
837 check_and_start_recovery_timer(struct obd_device *obd,
838 struct ptlrpc_request *req, int new_client);
841 * update flags for import during reconnect process
843 static int rev_import_flags_update(struct obd_import *revimp,
844 struct ptlrpc_request *req)
847 struct obd_connect_data *data;
849 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
851 if (data->ocd_connect_flags & OBD_CONNECT_AT)
852 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
854 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
856 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
858 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
860 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
861 revimp->imp_client->cli_name,
862 libcfs_id2str(req->rq_peer), rc);
870 * Allocate a new reverse import for an export.
872 * \retval -errno in case error hit
873 * \retval 0 if reverse import correctly init
875 int rev_import_init(struct obd_export *export)
877 struct obd_device *obd = export->exp_obd;
878 struct obd_import *revimp;
880 LASSERT(export->exp_imp_reverse == NULL);
882 revimp = class_new_import(obd);
886 revimp->imp_remote_handle.cookie = 0ULL;
887 revimp->imp_client = &obd->obd_ldlm_client;
888 revimp->imp_dlm_fake = 1;
890 /* it is safe to connect import in new state as no sends possible */
891 spin_lock(&export->exp_lock);
892 export->exp_imp_reverse = revimp;
893 spin_unlock(&export->exp_lock);
894 class_import_put(revimp);
898 EXPORT_SYMBOL(rev_import_init);
901 * Handle reconnect for an export.
903 * \param exp export to handle reconnect process
904 * \param req client reconnect request
906 * \retval -rc in case securitfy flavor can't be changed
907 * \retval 0 in case none problems
909 static int rev_import_reconnect(struct obd_export *exp,
910 struct ptlrpc_request *req)
912 struct obd_import *revimp = exp->exp_imp_reverse;
913 struct lustre_handle *lh;
916 /* avoid sending a request until import flags are changed */
917 ptlrpc_import_enter_resend(revimp);
919 if (revimp->imp_connection != NULL)
920 ptlrpc_connection_put(revimp->imp_connection);
923 * client from recovery don't have a handle so we need to take from
924 * request. it may produce situation when wrong client connected
925 * to recovery as we trust a client uuid
927 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
928 revimp->imp_remote_handle = *lh;
930 /* unknown versions will be caught in
931 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
932 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
934 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
936 rc = rev_import_flags_update(revimp, req);
938 /* it is safe to still be in RECOVERY phase as we are not able
939 * to setup correct security flavor so requests are not able to
940 * be delivered correctly */
944 /* resend all rpc's via new connection */
945 return ptlrpc_import_recovery_state_machine(revimp);
948 int target_handle_connect(struct ptlrpc_request *req)
950 struct obd_device *target = NULL;
951 struct obd_export *export = NULL;
952 /* connect handle - filled from target_handle_reconnect in
954 struct lustre_handle conn;
955 struct lustre_handle *tmp;
956 struct obd_uuid cluuid;
961 bool mds_conn = false, lw_client = false, initial_conn = false;
962 bool mds_mds_conn = false;
963 bool new_mds_mds_conn = false;
964 struct obd_connect_data *data, *tmpdata;
966 lnet_nid_t *client_nid = NULL;
969 OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
971 str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
973 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
974 GOTO(out, rc = -EINVAL);
977 target = class_dev_by_str(str);
979 deuuidify(str, NULL, &target_start, &target_len);
980 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
981 "from %s (no target). If you are running "
982 "an HA pair check that the target is "
983 "mounted on the other server.\n", str,
984 libcfs_nid2str(req->rq_peer.nid));
985 GOTO(out, rc = -ENODEV);
988 spin_lock(&target->obd_dev_lock);
990 target->obd_conn_inprogress++;
992 if (target->obd_stopping || !target->obd_set_up) {
993 spin_unlock(&target->obd_dev_lock);
995 deuuidify(str, NULL, &target_start, &target_len);
996 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
997 target_len, target_start,
998 libcfs_nid2str(req->rq_peer.nid),
999 (target->obd_stopping ?
1000 "stopping" : "not set up"));
1001 GOTO(out, rc = -ENODEV);
1004 if (target->obd_no_conn) {
1005 spin_unlock(&target->obd_dev_lock);
1007 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1008 "from %s\n", target->obd_name,
1009 libcfs_nid2str(req->rq_peer.nid));
1010 GOTO(out, rc = -EAGAIN);
1013 spin_unlock(&target->obd_dev_lock);
1015 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1017 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1018 GOTO(out, rc = -EINVAL);
1021 obd_str2uuid(&cluuid, str);
1023 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1025 GOTO(out, rc = -EPROTO);
1029 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1031 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1033 GOTO(out, rc = -EPROTO);
1035 rc = req_capsule_server_pack(&req->rq_pill);
1039 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1040 /* Don't allow clients to connect that are using old 1.8 format
1041 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1042 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1043 * FULL20 flag should be set on all connections since 2.0, but no
1044 * longer affects behaviour.
1046 * Later this check will be disabled and the flag can be retired
1047 * completely once interop with 3.0 is no longer needed.
1049 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1050 GOTO(out, rc = -EPROTO);
1052 /* Don't allow liblustre clients to connect.
1053 * - testing was disabled in v2_2_50_0-61-g6a75d65
1054 * - building was disabled in v2_5_58_0-28-g7277179
1055 * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1056 * - clients were refused connect for version difference > 0.0.1.32 */
1057 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1058 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1059 GOTO(out, rc = -EPROTO);
1063 /* Note: lw_client is needed in MDS-MDS failover during update log
1064 * processing, so we needs to allow lw_client to be connected at
1065 * anytime, instead of only the initial connection */
1066 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1068 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1069 initial_conn = true;
1070 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1071 mds_mds_conn = (data->ocd_connect_flags &
1072 OBD_CONNECT_MDS_MDS) != 0;
1074 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1075 * for Imperative Recovery connection from MGC to MGS.
1077 * Via check OBD_CONNECT_FID, we can distinguish whether
1078 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1081 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1082 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1083 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1084 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1085 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1086 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1088 /* We do not support the MDT-MDT interoperations with
1089 * different version MDT because of protocol changes. */
1090 if (unlikely(major != LUSTRE_MAJOR ||
1091 minor != LUSTRE_MINOR ||
1092 abs(patch - LUSTRE_PATCH) > 3)) {
1093 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1094 "connection from different version MDT "
1095 "(%d.%d.%d.%d) %s %s\n",
1096 target->obd_name, LUSTRE_MAJOR,
1097 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1098 major, minor, patch,
1099 OBD_OCD_VERSION_FIX(data->ocd_version),
1100 libcfs_nid2str(req->rq_peer.nid), str);
1102 GOTO(out, rc = -EPROTO);
1107 /* lctl gets a backstage, all-access pass. */
1108 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1109 goto dont_check_exports;
1111 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1115 /* We've found an export in the hash. */
1117 spin_lock(&export->exp_lock);
1119 if (export->exp_connecting) { /* bug 9635, et. al. */
1120 spin_unlock(&export->exp_lock);
1121 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1122 export->exp_obd->obd_name, export,
1123 libcfs_nid2str(req->rq_peer.nid));
1124 class_export_put(export);
1127 } else if ((mds_conn || (lw_client && initial_conn) ||
1128 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1129 export->exp_connection != NULL) {
1130 spin_unlock(&export->exp_lock);
1131 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1132 /* MDS or LWP reconnected after failover. */
1133 LCONSOLE_WARN("%s: Received %s connection from "
1134 "%s, removing former export from %s\n",
1135 target->obd_name, mds_conn ? "MDS" : "LWP",
1136 libcfs_nid2str(req->rq_peer.nid),
1137 libcfs_nid2str(export->exp_connection->c_peer.nid));
1139 /* New MDS connection from the same NID. */
1140 LCONSOLE_WARN("%s: Received new %s connection from "
1141 "%s, removing former export from same NID\n",
1142 target->obd_name, mds_conn ? "MDS" : "LWP",
1143 libcfs_nid2str(req->rq_peer.nid));
1146 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1147 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1148 /* Because exports between MDTs will always be
1149 * kept, let's do not fail such export if they
1150 * come from the same NID, otherwise it might
1151 * cause eviction between MDTs, which might
1152 * cause namespace inconsistency */
1153 spin_lock(&export->exp_lock);
1154 export->exp_connecting = 1;
1155 spin_unlock(&export->exp_lock);
1156 conn.cookie = export->exp_handle.h_cookie;
1159 class_fail_export(export);
1160 class_export_put(export);
1164 } else if (export->exp_connection != NULL && initial_conn &&
1165 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1166 spin_unlock(&export->exp_lock);
1167 /* In MDS failover we have static UUID but NID can change. */
1168 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1169 "existing nid %s is already connected\n",
1170 target->obd_name, cluuid.uuid,
1171 libcfs_nid2str(req->rq_peer.nid),
1173 export->exp_connection->c_peer.nid));
1175 class_export_put(export);
1178 export->exp_connecting = 1;
1179 spin_unlock(&export->exp_lock);
1180 LASSERT(export->exp_obd == target);
1182 rc = target_handle_reconnect(&conn, export, &cluuid);
1185 /* If we found an export, we already unlocked. */
1188 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1189 } else if (req->rq_export == NULL &&
1190 atomic_read(&export->exp_rpc_count) > 0) {
1191 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1192 "still busy with %d references\n",
1193 target->obd_name, cluuid.uuid,
1194 libcfs_nid2str(req->rq_peer.nid),
1195 atomic_read(&export->exp_refcount));
1196 GOTO(out, rc = -EBUSY);
1197 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1198 if (!strstr(cluuid.uuid, "mdt"))
1199 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1200 "known client %s (at %s) because it "
1201 "is indicating it is a new client",
1202 target->obd_name, cluuid.uuid,
1203 libcfs_nid2str(req->rq_peer.nid));
1204 GOTO(out, rc = -EALREADY);
1206 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1213 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1214 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1215 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1216 export, ktime_get_real_seconds(),
1217 export ? export->exp_last_request_time : 0);
1219 /* If this is the first time a client connects, reset the recovery
1220 * timer. Discard lightweight connections which might be local. */
1221 if (!lw_client && rc == 0 && target->obd_recovering)
1222 check_and_start_recovery_timer(target, req, export == NULL);
1224 /* We want to handle EALREADY but *not* -EALREADY from
1225 * target_handle_reconnect(), return reconnection state in a flag. */
1226 if (rc == EALREADY) {
1227 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1233 /* Tell the client if we support replayable requests. */
1234 if (target->obd_replayable)
1235 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1236 client_nid = &req->rq_peer.nid;
1238 if (export == NULL) {
1239 /* allow lightweight connections during recovery */
1240 /* allow "new" MDT to be connected during recovery, since we
1241 * need retrieve recovery update records from it */
1242 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1244 int c; /* connected */
1245 int i; /* in progress */
1247 int s; /* stale/evicted */
1249 c = atomic_read(&target->obd_connected_clients);
1250 i = atomic_read(&target->obd_lock_replay_clients);
1251 k = target->obd_max_recoverable_clients;
1252 s = target->obd_stale_clients;
1253 t = target->obd_recovery_timer.expires;
1254 t = cfs_duration_sec(target->obd_recovery_timer.expires);
1255 t -= ktime_get_seconds();
1256 LCONSOLE_WARN("%s: Denying connection for new client %s"
1257 "(at %s), waiting for %d known clients "
1258 "(%d recovered, %d in progress, and %d "
1259 "evicted) to recover in %lld:%.02lld\n",
1260 target->obd_name, cluuid.uuid,
1261 libcfs_nid2str(req->rq_peer.nid), k,
1262 c - i, i, s, t / 60, t % 60);
1266 rc = obd_connect(req->rq_svc_thread->t_env,
1267 &export, target, &cluuid, data,
1269 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1270 lustre_msg_add_op_flags(req->rq_repmsg,
1271 MSG_CONNECT_RECOVERING);
1273 conn.cookie = export->exp_handle.h_cookie;
1274 rc = rev_import_init(export);
1278 new_mds_mds_conn = true;
1281 rc = obd_reconnect(req->rq_svc_thread->t_env,
1282 export, target, &cluuid, data, client_nid);
1287 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1288 data->ocd_instance = target->u.obt.obt_instance;
1290 /* Return only the parts of obd_connect_data that we understand, so the
1291 * client knows that we don't understand the rest. */
1293 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1295 tmpdata = req_capsule_server_get(&req->rq_pill,
1297 /* Don't use struct assignment here, because the client reply
1298 * buffer may be smaller/larger than the local struct
1299 * obd_connect_data. */
1300 memcpy(tmpdata, data, min(tmpsize, size));
1303 /* If the client and the server are the same node, we will already
1304 * have an export that really points to the client's DLM export,
1305 * because we have a shared handles table.
1307 * XXX this will go away when shaver stops sending the "connect" handle
1308 * in the real "remote handle" field of the request --phik 24 Apr 2003
1310 ptlrpc_request_change_export(req, export);
1312 spin_lock(&export->exp_lock);
1313 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1314 spin_unlock(&export->exp_lock);
1315 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1316 "or equal conn_cnt: %d >= %d\n",
1317 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1318 export->exp_conn_cnt,
1319 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1321 GOTO(out, rc = -EALREADY);
1323 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1324 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1325 spin_unlock(&export->exp_lock);
1327 if (export->exp_connection != NULL) {
1328 /* Check to see if connection came from another NID. */
1329 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1330 !hlist_unhashed(&export->exp_nid_hash))
1331 cfs_hash_del(export->exp_obd->obd_nid_hash,
1332 &export->exp_connection->c_peer.nid,
1333 &export->exp_nid_hash);
1335 ptlrpc_connection_put(export->exp_connection);
1338 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1341 if (hlist_unhashed(&export->exp_nid_hash))
1342 cfs_hash_add(export->exp_obd->obd_nid_hash,
1343 &export->exp_connection->c_peer.nid,
1344 &export->exp_nid_hash);
1346 lustre_msg_set_handle(req->rq_repmsg, &conn);
1348 rc = rev_import_reconnect(export, req);
1352 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1354 __u64 transno = data->ocd_transno;
1356 spin_lock(&export->exp_lock);
1357 /* possible race with class_disconnect_stale_exports,
1358 * export may be already in the eviction process */
1359 if (export->exp_failed) {
1360 spin_unlock(&export->exp_lock);
1361 GOTO(out, rc = -ENODEV);
1363 export->exp_in_recovery = 1;
1364 export->exp_req_replay_needed = 1;
1365 export->exp_lock_replay_needed = 1;
1366 spin_unlock(&export->exp_lock);
1368 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1369 MSG_CONNECT_TRANSNO);
1370 if (has_transno && transno == 0)
1371 CWARN("Connect with zero transno!\n");
1373 if (has_transno && transno > 0 &&
1374 transno < target->obd_next_recovery_transno &&
1375 transno > target->obd_last_committed) {
1376 /* Another way is to use cmpxchg() to be lock-free. */
1377 spin_lock(&target->obd_recovery_task_lock);
1378 if (transno < target->obd_next_recovery_transno)
1379 target->obd_next_recovery_transno = transno;
1380 spin_unlock(&target->obd_recovery_task_lock);
1383 atomic_inc(&target->obd_req_replay_clients);
1384 atomic_inc(&target->obd_lock_replay_clients);
1385 /* Note: MDS-MDS connection is allowed to be connected during
1386 * recovery, no matter if the exports needs to be recoveried.
1387 * Because we need retrieve updates logs from all other MDTs.
1388 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1389 * also needs to be increased to match other recovery checking
1391 if (new_mds_mds_conn)
1392 target->obd_max_recoverable_clients++;
1393 if (atomic_inc_return(&target->obd_connected_clients) ==
1394 target->obd_max_recoverable_clients)
1395 wake_up(&target->obd_next_transno_waitq);
1398 /* Tell the client we're in recovery, when client is involved in it. */
1399 if (target->obd_recovering && !lw_client)
1400 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1404 spin_lock(&export->exp_lock);
1405 export->exp_connecting = 0;
1406 spin_unlock(&export->exp_lock);
1408 class_export_put(export);
1410 if (target != NULL) {
1411 spin_lock(&target->obd_dev_lock);
1412 target->obd_conn_inprogress--;
1413 spin_unlock(&target->obd_dev_lock);
1414 class_decref(target, "find", current);
1416 req->rq_status = rc;
1420 int target_handle_disconnect(struct ptlrpc_request *req)
1425 rc = req_capsule_server_pack(&req->rq_pill);
1429 /* Keep the rq_export around so we can send the reply. */
1430 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1435 void target_destroy_export(struct obd_export *exp)
1437 struct obd_import *imp = NULL;
1438 /* exports created from last_rcvd data, and "fake"
1439 exports created by lctl don't have an import */
1440 spin_lock(&exp->exp_lock);
1441 if (exp->exp_imp_reverse != NULL) {
1442 imp = exp->exp_imp_reverse;
1443 exp->exp_imp_reverse = NULL;
1445 spin_unlock(&exp->exp_lock);
1447 client_destroy_import(imp);
1449 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1450 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1451 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1452 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1454 EXPORT_SYMBOL(target_destroy_export);
1457 * Recovery functions
1459 static void target_request_copy_get(struct ptlrpc_request *req)
1461 class_export_rpc_inc(req->rq_export);
1462 LASSERT(list_empty(&req->rq_list));
1463 INIT_LIST_HEAD(&req->rq_replay_list);
1465 /* Increase refcount to keep request in queue. */
1466 atomic_inc(&req->rq_refcount);
1467 /* Let export know it has replays to be handled. */
1468 atomic_inc(&req->rq_export->exp_replay_count);
1471 static void target_request_copy_put(struct ptlrpc_request *req)
1473 LASSERT(list_empty(&req->rq_replay_list));
1474 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1476 atomic_dec(&req->rq_export->exp_replay_count);
1477 class_export_rpc_dec(req->rq_export);
1478 ptlrpc_server_drop_request(req);
1481 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1483 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1484 struct obd_export *exp = req->rq_export;
1485 struct ptlrpc_request *reqiter;
1486 struct ptlrpc_request *dup_req = NULL;
1491 spin_lock(&exp->exp_lock);
1492 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1494 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1502 /* We expect it with RESENT and REPLAY flags. */
1503 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1504 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1505 CERROR("invalid flags %x of resent replay\n",
1506 lustre_msg_get_flags(req->rq_reqmsg));
1508 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1511 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1513 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1514 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1518 list_add_tail(&req->rq_replay_list,
1519 &exp->exp_req_replay_queue);
1522 spin_unlock(&exp->exp_lock);
1526 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1528 LASSERT(!list_empty(&req->rq_replay_list));
1529 LASSERT(req->rq_export);
1531 spin_lock(&req->rq_export->exp_lock);
1532 list_del_init(&req->rq_replay_list);
1533 spin_unlock(&req->rq_export->exp_lock);
1536 static void target_finish_recovery(struct lu_target *lut)
1538 struct obd_device *obd = lut->lut_obd;
1541 /* Only log a recovery message when recovery has occurred. */
1542 if (obd->obd_recovery_start) {
1543 time64_t now = ktime_get_real_seconds();
1544 time64_t elapsed_time;
1546 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1547 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1548 "%d recovered and %d %s evicted.\n", obd->obd_name,
1549 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1550 obd->obd_max_recoverable_clients,
1551 atomic_read(&obd->obd_connected_clients),
1552 obd->obd_stale_clients,
1553 obd->obd_stale_clients == 1 ? "was" : "were");
1556 ldlm_reprocess_recovery_done(obd->obd_namespace);
1557 spin_lock(&obd->obd_recovery_task_lock);
1558 if (!list_empty(&obd->obd_req_replay_queue) ||
1559 !list_empty(&obd->obd_lock_replay_queue) ||
1560 !list_empty(&obd->obd_final_req_queue)) {
1561 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1563 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1564 list_empty(&obd->obd_lock_replay_queue) ? \
1566 list_empty(&obd->obd_final_req_queue) ? \
1568 spin_unlock(&obd->obd_recovery_task_lock);
1571 spin_unlock(&obd->obd_recovery_task_lock);
1573 obd->obd_recovery_end = ktime_get_real_seconds();
1575 /* When recovery finished, cleanup orphans on MDS and OST. */
1576 if (obd->obd_type && OBP(obd, postrecov)) {
1577 int rc = OBP(obd, postrecov)(obd);
1580 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1586 static void abort_req_replay_queue(struct obd_device *obd)
1588 struct ptlrpc_request *req, *n;
1589 struct list_head abort_list;
1591 INIT_LIST_HEAD(&abort_list);
1592 spin_lock(&obd->obd_recovery_task_lock);
1593 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1594 spin_unlock(&obd->obd_recovery_task_lock);
1595 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1596 DEBUG_REQ(D_WARNING, req, "aborted:");
1597 req->rq_status = -ENOTCONN;
1598 if (ptlrpc_error(req)) {
1599 DEBUG_REQ(D_ERROR, req,
1600 "failed abort_req_reply; skipping");
1602 target_exp_dequeue_req_replay(req);
1603 target_request_copy_put(req);
1607 static void abort_lock_replay_queue(struct obd_device *obd)
1609 struct ptlrpc_request *req, *n;
1610 struct list_head abort_list;
1612 INIT_LIST_HEAD(&abort_list);
1613 spin_lock(&obd->obd_recovery_task_lock);
1614 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1615 spin_unlock(&obd->obd_recovery_task_lock);
1616 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1617 DEBUG_REQ(D_ERROR, req, "aborted:");
1618 req->rq_status = -ENOTCONN;
1619 if (ptlrpc_error(req)) {
1620 DEBUG_REQ(D_ERROR, req,
1621 "failed abort_lock_reply; skipping");
1623 target_request_copy_put(req);
1627 /* Called from a cleanup function if the device is being cleaned up
1628 forcefully. The exports should all have been disconnected already,
1629 the only thing left to do is
1630 - clear the recovery flags
1632 - free queued requests and replies, but don't send replies
1633 Because the obd_stopping flag is set, no new requests should be received.
1636 void target_cleanup_recovery(struct obd_device *obd)
1638 struct ptlrpc_request *req, *n;
1639 struct list_head clean_list;
1641 INIT_LIST_HEAD(&clean_list);
1642 spin_lock(&obd->obd_dev_lock);
1643 if (!obd->obd_recovering) {
1644 spin_unlock(&obd->obd_dev_lock);
1648 obd->obd_recovering = obd->obd_abort_recovery = 0;
1649 spin_unlock(&obd->obd_dev_lock);
1651 spin_lock(&obd->obd_recovery_task_lock);
1652 target_cancel_recovery_timer(obd);
1653 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1654 spin_unlock(&obd->obd_recovery_task_lock);
1656 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1657 LASSERT(req->rq_reply_state == NULL);
1658 target_exp_dequeue_req_replay(req);
1659 target_request_copy_put(req);
1662 spin_lock(&obd->obd_recovery_task_lock);
1663 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1664 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1665 spin_unlock(&obd->obd_recovery_task_lock);
1667 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1668 LASSERT(req->rq_reply_state == NULL);
1669 target_request_copy_put(req);
1674 EXPORT_SYMBOL(target_cleanup_recovery);
1676 /* obd_recovery_task_lock should be held */
1677 void target_cancel_recovery_timer(struct obd_device *obd)
1679 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1680 del_timer(&obd->obd_recovery_timer);
1683 static void target_start_recovery_timer(struct obd_device *obd)
1685 if (obd->obd_recovery_start != 0)
1688 spin_lock(&obd->obd_dev_lock);
1689 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1690 spin_unlock(&obd->obd_dev_lock);
1694 LASSERT(obd->obd_recovery_timeout != 0);
1696 if (obd->obd_recovery_start != 0) {
1697 spin_unlock(&obd->obd_dev_lock);
1701 mod_timer(&obd->obd_recovery_timer,
1702 jiffies + cfs_time_seconds(obd->obd_recovery_timeout));
1703 obd->obd_recovery_start = ktime_get_real_seconds();
1704 spin_unlock(&obd->obd_dev_lock);
1706 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1708 obd->obd_recovery_timeout / 60,
1709 obd->obd_recovery_timeout % 60,
1710 obd->obd_max_recoverable_clients,
1711 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1712 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1716 * extend recovery window.
1718 * if @extend is true, extend recovery window to have @drt remaining at least;
1719 * otherwise, make sure the recovery timeout value is not less than @drt.
1721 static void extend_recovery_timer(struct obd_device *obd, time64_t drt,
1729 spin_lock(&obd->obd_dev_lock);
1730 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1731 spin_unlock(&obd->obd_dev_lock);
1734 LASSERT(obd->obd_recovery_start != 0);
1736 now = ktime_get_real_seconds();
1737 to = obd->obd_recovery_timeout;
1738 end = obd->obd_recovery_start + to;
1741 if (extend && (drt > left)) {
1743 } else if (!extend && (drt > to)) {
1747 if (to > obd->obd_recovery_time_hard) {
1748 to = obd->obd_recovery_time_hard;
1749 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1750 obd->obd_name, to, extend);
1753 if (obd->obd_recovery_timeout < to) {
1754 obd->obd_recovery_timeout = to;
1755 end = obd->obd_recovery_start + to;
1756 mod_timer(&obd->obd_recovery_timer,
1757 jiffies + cfs_time_seconds(end - now));
1759 spin_unlock(&obd->obd_dev_lock);
1761 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1762 obd->obd_name, (s64)(end - now));
1765 /* Reset the timer with each new client connection */
1767 * This timer is actually reconnect_timer, which is for making sure
1768 * the total recovery window is at least as big as my reconnect
1769 * attempt timing. So the initial recovery time_out will be set to
1770 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1771 * from client is bigger than this, then the recovery time_out will
1772 * be extended to make sure the client could be reconnected, in the
1773 * process, the timeout from the new client should be ignored.
1777 check_and_start_recovery_timer(struct obd_device *obd,
1778 struct ptlrpc_request *req,
1781 time64_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1782 struct obd_device_target *obt = &obd->u.obt;
1784 if (!new_client && service_time)
1785 /* Teach server about old server's estimates, as first guess
1786 * at how long new requests will take. */
1787 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1790 target_start_recovery_timer(obd);
1792 /* Convert the service time to RPC timeout,
1793 * and reuse service_time to limit stack usage.
1795 service_time = at_est2timeout(service_time);
1797 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1798 service_time < at_extra)
1799 service_time = at_extra;
1801 /* We expect other clients to timeout within service_time, then try
1802 * to reconnect, then try the failover server. The max delay between
1803 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1804 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1806 LASSERT(obt->obt_magic == OBT_MAGIC);
1807 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1808 if (service_time > obd->obd_recovery_timeout && !new_client)
1809 extend_recovery_timer(obd, service_time, false);
1812 /** Health checking routines */
1813 static inline int exp_connect_healthy(struct obd_export *exp)
1815 return (exp->exp_in_recovery);
1818 /** if export done req_replay or has replay in queue */
1819 static inline int exp_req_replay_healthy(struct obd_export *exp)
1821 return (!exp->exp_req_replay_needed ||
1822 atomic_read(&exp->exp_replay_count) > 0);
1826 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1828 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1829 exp_req_replay_healthy(exp);
1832 /** if export done lock_replay or has replay in queue */
1833 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1835 return (!exp->exp_lock_replay_needed ||
1836 atomic_read(&exp->exp_replay_count) > 0);
1839 static inline int exp_vbr_healthy(struct obd_export *exp)
1841 return (!exp->exp_vbr_failed);
1844 static inline int exp_finished(struct obd_export *exp)
1846 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1849 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1851 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1855 static int check_for_next_transno(struct lu_target *lut)
1857 struct ptlrpc_request *req = NULL;
1858 struct obd_device *obd = lut->lut_obd;
1859 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1860 int wake_up = 0, connected, completed, queue_len;
1861 __u64 req_transno = 0;
1862 __u64 update_transno = 0;
1863 __u64 next_transno = 0;
1866 spin_lock(&obd->obd_recovery_task_lock);
1867 if (!list_empty(&obd->obd_req_replay_queue)) {
1868 req = list_entry(obd->obd_req_replay_queue.next,
1869 struct ptlrpc_request, rq_list);
1870 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1874 update_transno = distribute_txn_get_next_transno(tdtd);
1876 connected = atomic_read(&obd->obd_connected_clients);
1877 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1878 queue_len = obd->obd_requests_queued_for_recovery;
1879 next_transno = obd->obd_next_recovery_transno;
1881 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1882 "req_transno: %llu, next_transno: %llu\n",
1883 obd->obd_max_recoverable_clients, connected, completed,
1884 queue_len, req_transno, next_transno);
1886 if (obd->obd_abort_recovery) {
1887 CDEBUG(D_HA, "waking for aborted recovery\n");
1889 } else if (obd->obd_recovery_expired) {
1890 CDEBUG(D_HA, "waking for expired recovery\n");
1892 } else if (tdtd != NULL && req != NULL &&
1893 is_req_replayed_by_update(req)) {
1894 LASSERTF(req_transno < next_transno, "req_transno %llu"
1895 "next_transno%llu\n", req_transno, next_transno);
1896 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1899 } else if (req_transno == next_transno ||
1900 (update_transno != 0 && update_transno <= next_transno)) {
1901 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1903 } else if (queue_len > 0 &&
1904 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1905 /** handle gaps occured due to lost reply or VBR */
1906 LASSERTF(req_transno >= next_transno,
1907 "req_transno: %llu, next_transno: %llu\n",
1908 req_transno, next_transno);
1910 "%s: waking for gap in transno, VBR is %s (skip: "
1911 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1912 ", next_update %lld last_committed: %lld)\n",
1913 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1914 next_transno, queue_len, completed, connected,
1915 req_transno, update_transno, obd->obd_last_committed);
1916 obd->obd_next_recovery_transno = req_transno;
1918 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1919 CDEBUG(D_HA, "waking for completed recovery\n");
1921 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1922 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1923 " by fail_lock, waking up (%lld)\n", next_transno);
1924 obd->obd_next_recovery_transno = req_transno;
1927 spin_unlock(&obd->obd_recovery_task_lock);
1931 static int check_for_next_lock(struct lu_target *lut)
1933 struct obd_device *obd = lut->lut_obd;
1936 spin_lock(&obd->obd_recovery_task_lock);
1937 if (!list_empty(&obd->obd_lock_replay_queue)) {
1938 CDEBUG(D_HA, "waking for next lock\n");
1940 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1941 CDEBUG(D_HA, "waking for completed lock replay\n");
1943 } else if (obd->obd_abort_recovery) {
1944 CDEBUG(D_HA, "waking for aborted recovery\n");
1946 } else if (obd->obd_recovery_expired) {
1947 CDEBUG(D_HA, "waking for expired recovery\n");
1950 spin_unlock(&obd->obd_recovery_task_lock);
1956 * wait for recovery events,
1957 * check its status with help of check_routine
1958 * evict dead clients via health_check
1960 static int target_recovery_overseer(struct lu_target *lut,
1961 int (*check_routine)(struct lu_target *),
1962 int (*health_check)(struct obd_export *))
1964 struct obd_device *obd = lut->lut_obd;
1965 struct target_distribute_txn_data *tdtd;
1967 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
1968 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
1969 __u64 next_update_transno = 0;
1971 /* Only abort the recovery if there are no update recovery
1972 * left in the queue */
1973 spin_lock(&obd->obd_recovery_task_lock);
1974 if (lut->lut_tdtd != NULL) {
1975 next_update_transno =
1976 distribute_txn_get_next_transno(lut->lut_tdtd);
1978 tdtd = lut->lut_tdtd;
1979 /* If next_update_transno == 0, it probably because
1980 * updatelog retrieve threads did not get any records
1981 * yet, let's wait those threads stopped */
1982 if (next_update_transno == 0) {
1983 struct l_wait_info lwi = { 0 };
1985 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1987 &tdtd->tdtd_recovery_threads_count) == 0,
1990 next_update_transno =
1991 distribute_txn_get_next_transno(
1996 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
1997 obd->obd_next_recovery_transno = next_update_transno;
1998 spin_unlock(&obd->obd_recovery_task_lock);
1999 /* Disconnect unfinished exports from clients, and
2000 * keep connection from MDT to make sure the update
2001 * recovery will still keep trying until some one
2002 * manually abort the recovery */
2003 class_disconnect_stale_exports(obd,
2004 exp_finished_or_from_mdt);
2005 /* Abort all of replay and replay lock req from
2007 abort_req_replay_queue(obd);
2008 abort_lock_replay_queue(obd);
2009 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2010 ")in the queue.\n", obd->obd_name,
2011 next_update_transno);
2013 obd->obd_abort_recovery = 1;
2014 spin_unlock(&obd->obd_recovery_task_lock);
2015 CWARN("%s recovery is aborted by hard timeout\n",
2020 while (wait_event_timeout(obd->obd_next_transno_waitq,
2022 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2023 /* wait indefinitely for event, but don't trigger watchdog */;
2025 if (obd->obd_abort_recovery) {
2026 CWARN("recovery is aborted, evict exports in recovery\n");
2027 if (lut->lut_tdtd != NULL) {
2028 struct l_wait_info lwi = { 0 };
2030 tdtd = lut->lut_tdtd;
2031 /* Let's wait all of the update log recovery thread
2033 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2034 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2036 /* Then abort the update recovery list */
2037 dtrq_list_destroy(lut->lut_tdtd);
2040 /** evict exports which didn't finish recovery yet */
2041 class_disconnect_stale_exports(obd, exp_finished);
2043 } else if (obd->obd_recovery_expired) {
2044 obd->obd_recovery_expired = 0;
2045 /** If some clients died being recovered, evict them */
2046 LCONSOLE_WARN("%s: recovery is timed out, "
2047 "evict stale exports\n", obd->obd_name);
2048 /** evict cexports with no replay in queue, they are stalled */
2049 class_disconnect_stale_exports(obd, health_check);
2051 /** continue with VBR */
2052 spin_lock(&obd->obd_dev_lock);
2053 obd->obd_version_recov = 1;
2054 spin_unlock(&obd->obd_dev_lock);
2056 * reset timer, recovery will proceed with versions now,
2057 * timeout is set just to handle reconnection delays
2059 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2060 /** Wait for recovery events again, after evicting bad clients */
2066 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2068 struct obd_device *obd = lut->lut_obd;
2069 struct ptlrpc_request *req = NULL;
2071 CDEBUG(D_HA, "Waiting for lock\n");
2072 if (target_recovery_overseer(lut, check_for_next_lock,
2073 exp_lock_replay_healthy))
2074 abort_lock_replay_queue(obd);
2076 spin_lock(&obd->obd_recovery_task_lock);
2077 if (!list_empty(&obd->obd_lock_replay_queue)) {
2078 req = list_entry(obd->obd_lock_replay_queue.next,
2079 struct ptlrpc_request, rq_list);
2080 list_del_init(&req->rq_list);
2081 spin_unlock(&obd->obd_recovery_task_lock);
2083 spin_unlock(&obd->obd_recovery_task_lock);
2084 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2085 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2086 /** evict exports failed VBR */
2087 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2092 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2094 struct ptlrpc_request *req = NULL;
2096 spin_lock(&obd->obd_recovery_task_lock);
2097 if (!list_empty(&obd->obd_final_req_queue)) {
2098 req = list_entry(obd->obd_final_req_queue.next,
2099 struct ptlrpc_request, rq_list);
2100 list_del_init(&req->rq_list);
2101 spin_unlock(&obd->obd_recovery_task_lock);
2102 if (req->rq_export->exp_in_recovery) {
2103 spin_lock(&req->rq_export->exp_lock);
2104 req->rq_export->exp_in_recovery = 0;
2105 spin_unlock(&req->rq_export->exp_lock);
2108 spin_unlock(&obd->obd_recovery_task_lock);
2113 static void handle_recovery_req(struct ptlrpc_thread *thread,
2114 struct ptlrpc_request *req,
2115 svc_handler_t handler)
2120 * export can be evicted during recovery, no need to handle replays for
2121 * it after that, discard such request silently
2123 if (req->rq_export->exp_disconnected)
2126 req->rq_session.lc_thread = thread;
2127 req->rq_svc_thread = thread;
2128 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2130 /* thread context */
2131 lu_context_enter(&thread->t_env->le_ctx);
2133 lu_context_exit(&thread->t_env->le_ctx);
2135 /* don't reset timer for final stage */
2136 if (!exp_finished(req->rq_export)) {
2137 time64_t to = obd_timeout;
2140 * Add request timeout to the recovery time so next request from
2141 * this client may come in recovery time
2144 struct ptlrpc_service_part *svcpt;
2146 svcpt = req->rq_rqbd->rqbd_svcpt;
2147 /* If the server sent early reply for this request,
2148 * the client will recalculate the timeout according to
2149 * current server estimate service time, so we will
2150 * use the maxium timeout here for waiting the client
2151 * sending the next req */
2152 to = max((int)at_est2timeout(
2153 at_get(&svcpt->scp_at_estimate)),
2154 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2155 /* Add 2 net_latency, one for balance rq_deadline
2156 * (see ptl_send_rpc), one for resend the req to server,
2157 * Note: client will pack net_latency in replay req
2158 * (see ptlrpc_replay_req) */
2159 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2161 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2166 /** Checking routines for recovery */
2167 static int check_for_recovery_ready(struct lu_target *lut)
2169 struct obd_device *obd = lut->lut_obd;
2170 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2172 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2173 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2174 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2175 obd->obd_recovery_expired);
2177 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2178 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2179 if (clnts + obd->obd_stale_clients <
2180 obd->obd_max_recoverable_clients)
2184 if (lut->lut_tdtd != NULL) {
2185 if (!lut->lut_tdtd->tdtd_replay_ready &&
2186 !obd->obd_abort_recovery) {
2187 /* Let's extend recovery timer, in case the recovery
2188 * timer expired, and some clients got evicted */
2189 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2191 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2192 obd->obd_name, obd->obd_recovery_timeout);
2201 REQUEST_RECOVERY = 1,
2202 UPDATE_RECOVERY = 2,
2205 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2209 if (!list_empty(&obd->obd_req_replay_queue)) {
2210 struct ptlrpc_request *req;
2212 req = list_entry(obd->obd_req_replay_queue.next,
2213 struct ptlrpc_request, rq_list);
2214 transno = lustre_msg_get_transno(req->rq_reqmsg);
2220 static __u64 get_next_transno(struct lu_target *lut, int *type)
2222 struct obd_device *obd = lut->lut_obd;
2223 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2225 __u64 update_transno;
2228 transno = get_next_replay_req_transno(obd);
2230 *type = REQUEST_RECOVERY;
2235 update_transno = distribute_txn_get_next_transno(tdtd);
2236 if (transno == 0 || (transno >= update_transno &&
2237 update_transno != 0)) {
2238 transno = update_transno;
2240 *type = UPDATE_RECOVERY;
2247 * drop duplicate replay request
2249 * Because the operation has been replayed by update recovery, the request
2250 * with the same transno will be dropped and also notify the client to send
2251 * next replay request.
2253 * \param[in] env execution environment
2254 * \param[in] obd failover obd device
2255 * \param[in] req request to be dropped
2257 static void drop_duplicate_replay_req(struct lu_env *env,
2258 struct obd_device *obd,
2259 struct ptlrpc_request *req)
2261 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2262 " update records are found.\n",
2263 lustre_msg_get_transno(req->rq_reqmsg),
2264 libcfs_nid2str(req->rq_peer.nid));
2266 /* Right now, only for MDS reint operation update replay and
2267 * normal request replay can have the same transno */
2268 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2269 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2270 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2271 if (likely(req->rq_export))
2272 target_committed_to_req(req);
2273 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2274 target_send_reply(req, req->rq_status, 0);
2276 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2277 libcfs_nid2str(req->rq_peer.nid));
2279 target_exp_dequeue_req_replay(req);
2280 target_request_copy_put(req);
2281 obd->obd_replayed_requests++;
2284 static void replay_request_or_update(struct lu_env *env,
2285 struct lu_target *lut,
2286 struct target_recovery_data *trd,
2287 struct ptlrpc_thread *thread)
2289 struct obd_device *obd = lut->lut_obd;
2290 struct ptlrpc_request *req = NULL;
2295 CDEBUG(D_HA, "Waiting for transno %lld\n",
2296 obd->obd_next_recovery_transno);
2298 /* Replay all of request and update by transno */
2300 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2302 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2304 /** It is needed to extend recovery window above
2305 * recovery_time_soft. Extending is possible only in the
2306 * end of recovery window (see more details in
2307 * handle_recovery_req()).
2309 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2311 if (target_recovery_overseer(lut, check_for_next_transno,
2312 exp_req_replay_healthy_or_from_mdt)) {
2313 abort_req_replay_queue(obd);
2314 abort_lock_replay_queue(obd);
2318 spin_lock(&obd->obd_recovery_task_lock);
2319 transno = get_next_transno(lut, &type);
2320 if (type == REQUEST_RECOVERY && transno != 0) {
2321 /* Drop replay request from client side, if the
2322 * replay has been executed by update with the
2324 req = list_entry(obd->obd_req_replay_queue.next,
2325 struct ptlrpc_request, rq_list);
2327 list_del_init(&req->rq_list);
2328 obd->obd_requests_queued_for_recovery--;
2329 spin_unlock(&obd->obd_recovery_task_lock);
2331 /* Let's check if the request has been redone by
2333 if (is_req_replayed_by_update(req)) {
2334 struct distribute_txn_replay_req *dtrq;
2336 dtrq = distribute_txn_lookup_finish_list(tdtd,
2338 LASSERT(dtrq != NULL);
2339 spin_lock(&tdtd->tdtd_replay_list_lock);
2340 list_del_init(&dtrq->dtrq_list);
2341 spin_unlock(&tdtd->tdtd_replay_list_lock);
2344 drop_duplicate_replay_req(env, obd, req);
2349 LASSERT(trd->trd_processing_task == current_pid());
2350 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2351 lustre_msg_get_transno(req->rq_reqmsg),
2352 libcfs_nid2str(req->rq_peer.nid));
2354 handle_recovery_req(thread, req,
2355 trd->trd_recovery_handler);
2357 * bz18031: increase next_recovery_transno before
2358 * target_request_copy_put() will drop exp_rpc reference
2360 spin_lock(&obd->obd_recovery_task_lock);
2361 obd->obd_next_recovery_transno++;
2362 spin_unlock(&obd->obd_recovery_task_lock);
2363 target_exp_dequeue_req_replay(req);
2364 target_request_copy_put(req);
2365 obd->obd_replayed_requests++;
2366 } else if (type == UPDATE_RECOVERY && transno != 0) {
2367 struct distribute_txn_replay_req *dtrq;
2370 spin_unlock(&obd->obd_recovery_task_lock);
2372 LASSERT(tdtd != NULL);
2373 dtrq = distribute_txn_get_next_req(tdtd);
2374 lu_context_enter(&thread->t_env->le_ctx);
2375 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2376 lu_context_exit(&thread->t_env->le_ctx);
2377 extend_recovery_timer(obd, obd_timeout, true);
2379 if (rc == 0 && dtrq->dtrq_xid != 0) {
2380 CDEBUG(D_HA, "Move x%llu t%llu"
2381 " to finish list\n", dtrq->dtrq_xid,
2382 dtrq->dtrq_master_transno);
2384 /* Add it to the replay finish list */
2385 spin_lock(&tdtd->tdtd_replay_list_lock);
2386 list_add(&dtrq->dtrq_list,
2387 &tdtd->tdtd_replay_finish_list);
2388 spin_unlock(&tdtd->tdtd_replay_list_lock);
2390 spin_lock(&obd->obd_recovery_task_lock);
2391 if (transno == obd->obd_next_recovery_transno)
2392 obd->obd_next_recovery_transno++;
2394 obd->obd_next_recovery_transno)
2395 obd->obd_next_recovery_transno =
2397 spin_unlock(&obd->obd_recovery_task_lock);
2402 spin_unlock(&obd->obd_recovery_task_lock);
2404 LASSERT(list_empty(&obd->obd_req_replay_queue));
2405 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2406 /** evict exports failed VBR */
2407 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2413 static int target_recovery_thread(void *arg)
2415 struct lu_target *lut = arg;
2416 struct obd_device *obd = lut->lut_obd;
2417 struct ptlrpc_request *req;
2418 struct target_recovery_data *trd = &obd->obd_recovery_data;
2419 unsigned long delta;
2421 struct ptlrpc_thread *thread = NULL;
2425 unshare_fs_struct();
2426 OBD_ALLOC_PTR(thread);
2432 OBD_FREE_PTR(thread);
2436 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2438 OBD_FREE_PTR(thread);
2443 thread->t_env = env;
2444 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2445 env->le_ctx.lc_thread = thread;
2446 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2447 thread->t_watchdog = NULL;
2449 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2451 trd->trd_processing_task = current_pid();
2453 spin_lock(&obd->obd_dev_lock);
2454 obd->obd_recovering = 1;
2455 spin_unlock(&obd->obd_dev_lock);
2456 complete(&trd->trd_starting);
2458 /* first of all, we have to know the first transno to replay */
2459 if (target_recovery_overseer(lut, check_for_recovery_ready,
2460 exp_connect_healthy)) {
2461 abort_req_replay_queue(obd);
2462 abort_lock_replay_queue(obd);
2463 if (lut->lut_tdtd != NULL)
2464 dtrq_list_destroy(lut->lut_tdtd);
2467 /* next stage: replay requests or update */
2469 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2470 atomic_read(&obd->obd_req_replay_clients),
2471 obd->obd_next_recovery_transno);
2472 replay_request_or_update(env, lut, trd, thread);
2475 * The second stage: replay locks
2477 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2478 atomic_read(&obd->obd_lock_replay_clients));
2479 while ((req = target_next_replay_lock(lut))) {
2480 LASSERT(trd->trd_processing_task == current_pid());
2481 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2482 libcfs_nid2str(req->rq_peer.nid));
2483 handle_recovery_req(thread, req,
2484 trd->trd_recovery_handler);
2485 target_request_copy_put(req);
2486 obd->obd_replayed_locks++;
2490 * The third stage: reply on final pings, at this moment all clients
2491 * must have request in final queue
2493 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2494 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2495 /** Update server last boot epoch */
2496 tgt_boot_epoch_update(lut);
2497 /* We drop recoverying flag to forward all new requests
2498 * to regular mds_handle() since now */
2499 spin_lock(&obd->obd_dev_lock);
2500 obd->obd_recovering = obd->obd_abort_recovery = 0;
2501 spin_unlock(&obd->obd_dev_lock);
2502 spin_lock(&obd->obd_recovery_task_lock);
2503 target_cancel_recovery_timer(obd);
2504 spin_unlock(&obd->obd_recovery_task_lock);
2505 while ((req = target_next_final_ping(obd))) {
2506 LASSERT(trd->trd_processing_task == current_pid());
2507 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2508 libcfs_nid2str(req->rq_peer.nid));
2509 handle_recovery_req(thread, req,
2510 trd->trd_recovery_handler);
2511 /* Because the waiting client can not send ping to server,
2512 * so we need refresh the last_request_time, to avoid the
2513 * export is being evicted */
2514 ptlrpc_update_export_timer(req->rq_export, 0);
2515 target_request_copy_put(req);
2518 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2519 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2520 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2521 if (delta > OBD_RECOVERY_TIME_SOFT) {
2522 CWARN("too long recovery - read logs\n");
2523 libcfs_debug_dumplog();
2526 target_finish_recovery(lut);
2528 lu_context_fini(&env->le_ctx);
2529 trd->trd_processing_task = 0;
2530 complete(&trd->trd_finishing);
2532 tgt_io_thread_done(thread);
2533 OBD_FREE_PTR(thread);
2538 static int target_start_recovery_thread(struct lu_target *lut,
2539 svc_handler_t handler)
2541 struct obd_device *obd = lut->lut_obd;
2543 struct target_recovery_data *trd = &obd->obd_recovery_data;
2546 memset(trd, 0, sizeof(*trd));
2547 init_completion(&trd->trd_starting);
2548 init_completion(&trd->trd_finishing);
2549 trd->trd_recovery_handler = handler;
2551 rc = server_name2index(obd->obd_name, &index, NULL);
2555 if (!IS_ERR(kthread_run(target_recovery_thread,
2556 lut, "tgt_recover_%d", index))) {
2557 wait_for_completion(&trd->trd_starting);
2558 LASSERT(obd->obd_recovering != 0);
2566 void target_stop_recovery_thread(struct obd_device *obd)
2568 if (obd->obd_recovery_data.trd_processing_task > 0) {
2569 struct target_recovery_data *trd = &obd->obd_recovery_data;
2570 /** recovery can be done but postrecovery is not yet */
2571 spin_lock(&obd->obd_dev_lock);
2572 if (obd->obd_recovering) {
2573 CERROR("%s: Aborting recovery\n", obd->obd_name);
2574 obd->obd_abort_recovery = 1;
2575 wake_up(&obd->obd_next_transno_waitq);
2577 spin_unlock(&obd->obd_dev_lock);
2578 wait_for_completion(&trd->trd_finishing);
2581 EXPORT_SYMBOL(target_stop_recovery_thread);
2583 void target_recovery_fini(struct obd_device *obd)
2585 class_disconnect_exports(obd);
2586 target_stop_recovery_thread(obd);
2587 target_cleanup_recovery(obd);
2589 EXPORT_SYMBOL(target_recovery_fini);
2591 static void target_recovery_expired(unsigned long castmeharder)
2593 struct obd_device *obd = (struct obd_device *)castmeharder;
2594 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2595 " after %llus (%d clients connected)\n",
2596 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2597 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2598 atomic_read(&obd->obd_connected_clients));
2600 obd->obd_recovery_expired = 1;
2601 wake_up(&obd->obd_next_transno_waitq);
2604 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2606 struct obd_device *obd = lut->lut_obd;
2608 if (lut->lut_bottom->dd_rdonly)
2611 if (obd->obd_max_recoverable_clients == 0) {
2612 /** Update server last boot epoch */
2613 tgt_boot_epoch_update(lut);
2617 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2618 "last_transno %llu\n", obd->obd_name,
2619 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2620 LASSERT(obd->obd_stopping == 0);
2621 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2622 obd->obd_recovery_start = 0;
2623 obd->obd_recovery_end = 0;
2625 setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2626 (unsigned long)obd);
2627 target_start_recovery_thread(lut, handler);
2629 EXPORT_SYMBOL(target_recovery_init);
2631 static int target_process_req_flags(struct obd_device *obd,
2632 struct ptlrpc_request *req)
2634 struct obd_export *exp = req->rq_export;
2635 LASSERT(exp != NULL);
2636 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2637 /* client declares he's ready to replay locks */
2638 spin_lock(&exp->exp_lock);
2639 if (exp->exp_req_replay_needed) {
2640 exp->exp_req_replay_needed = 0;
2641 spin_unlock(&exp->exp_lock);
2643 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2644 atomic_dec(&obd->obd_req_replay_clients);
2646 spin_unlock(&exp->exp_lock);
2649 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2650 /* client declares he's ready to complete recovery
2651 * so, we put the request on th final queue */
2652 spin_lock(&exp->exp_lock);
2653 if (exp->exp_lock_replay_needed) {
2654 exp->exp_lock_replay_needed = 0;
2655 spin_unlock(&exp->exp_lock);
2657 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2658 atomic_dec(&obd->obd_lock_replay_clients);
2660 spin_unlock(&exp->exp_lock);
2666 int target_queue_recovery_request(struct ptlrpc_request *req,
2667 struct obd_device *obd)
2669 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2670 struct ptlrpc_request *reqiter;
2674 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2675 /* Processing the queue right now, don't re-add. */
2679 target_process_req_flags(obd, req);
2681 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2682 /* client declares he's ready to complete recovery
2683 * so, we put the request on th final queue */
2684 target_request_copy_get(req);
2685 DEBUG_REQ(D_HA, req, "queue final req");
2686 wake_up(&obd->obd_next_transno_waitq);
2687 spin_lock(&obd->obd_recovery_task_lock);
2688 if (obd->obd_recovering) {
2689 struct ptlrpc_request *tmp;
2690 struct ptlrpc_request *duplicate = NULL;
2692 if (likely(!req->rq_export->exp_replay_done)) {
2693 req->rq_export->exp_replay_done = 1;
2694 list_add_tail(&req->rq_list,
2695 &obd->obd_final_req_queue);
2696 spin_unlock(&obd->obd_recovery_task_lock);
2700 /* XXX O(n), but only happens if final ping is
2701 * timed out, probably reorganize the list as
2702 * a hash list later */
2703 list_for_each_entry_safe(reqiter, tmp,
2704 &obd->obd_final_req_queue,
2706 if (reqiter->rq_export == req->rq_export) {
2707 list_del_init(&reqiter->rq_list);
2708 duplicate = reqiter;
2713 list_add_tail(&req->rq_list,
2714 &obd->obd_final_req_queue);
2715 req->rq_export->exp_replay_done = 1;
2716 spin_unlock(&obd->obd_recovery_task_lock);
2718 if (duplicate != NULL) {
2719 DEBUG_REQ(D_HA, duplicate,
2720 "put prev final req\n");
2721 target_request_copy_put(duplicate);
2725 spin_unlock(&obd->obd_recovery_task_lock);
2726 target_request_copy_put(req);
2727 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2730 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2731 /* client declares he's ready to replay locks */
2732 target_request_copy_get(req);
2733 DEBUG_REQ(D_HA, req, "queue lock replay req");
2734 wake_up(&obd->obd_next_transno_waitq);
2735 spin_lock(&obd->obd_recovery_task_lock);
2736 LASSERT(obd->obd_recovering);
2737 /* usually due to recovery abort */
2738 if (!req->rq_export->exp_in_recovery) {
2739 spin_unlock(&obd->obd_recovery_task_lock);
2740 target_request_copy_put(req);
2743 LASSERT(req->rq_export->exp_lock_replay_needed);
2744 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2745 spin_unlock(&obd->obd_recovery_task_lock);
2749 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2750 * (i.e. buflens etc are in my own byte order), but type-dependent
2751 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2754 INIT_LIST_HEAD(&req->rq_list);
2755 DEBUG_REQ(D_HA, req, "not queueing");
2759 /* If we're processing the queue, we want don't want to queue this
2762 * Also, if this request has a transno less than the one we're waiting
2763 * for, we should process it now. It could (and currently always will)
2764 * be an open request for a descriptor that was opened some time ago.
2766 * Also, a resent, replayed request that has already been
2767 * handled will pass through here and be processed immediately.
2769 CDEBUG(D_HA, "Next recovery transno: %llu"
2770 ", current: %llu, replaying\n",
2771 obd->obd_next_recovery_transno, transno);
2773 /* If the request has been replayed by update replay, then sends this
2774 * request to the recovery thread (replay_request_or_update()), where
2775 * it will be handled */
2776 spin_lock(&obd->obd_recovery_task_lock);
2777 if (transno < obd->obd_next_recovery_transno &&
2778 !is_req_replayed_by_update(req)) {
2779 /* Processing the queue right now, don't re-add. */
2780 LASSERT(list_empty(&req->rq_list));
2781 spin_unlock(&obd->obd_recovery_task_lock);
2784 spin_unlock(&obd->obd_recovery_task_lock);
2786 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2789 target_request_copy_get(req);
2790 if (!req->rq_export->exp_in_recovery) {
2791 target_request_copy_put(req);
2794 LASSERT(req->rq_export->exp_req_replay_needed);
2796 if (target_exp_enqueue_req_replay(req)) {
2797 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2798 target_request_copy_put(req);
2803 spin_lock(&obd->obd_recovery_task_lock);
2804 LASSERT(obd->obd_recovering);
2805 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2806 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2807 list_add_tail(&req->rq_list, &reqiter->rq_list);
2812 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2814 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2815 "has been claimed by another client");
2816 spin_unlock(&obd->obd_recovery_task_lock);
2817 target_exp_dequeue_req_replay(req);
2818 target_request_copy_put(req);
2824 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2826 obd->obd_requests_queued_for_recovery++;
2827 spin_unlock(&obd->obd_recovery_task_lock);
2828 wake_up(&obd->obd_next_transno_waitq);
2832 int target_handle_ping(struct ptlrpc_request *req)
2834 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2835 return req_capsule_server_pack(&req->rq_pill);
2838 void target_committed_to_req(struct ptlrpc_request *req)
2840 struct obd_export *exp = req->rq_export;
2842 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2843 lustre_msg_set_last_committed(req->rq_repmsg,
2844 exp->exp_last_committed);
2846 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2847 "%d)", exp->exp_obd->obd_no_transno,
2848 req->rq_repmsg == NULL);
2850 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2851 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2854 #endif /* HAVE_SERVER_SUPPORT */
2857 * Packs current SLV and Limit into \a req.
2859 int target_pack_pool_reply(struct ptlrpc_request *req)
2861 struct obd_device *obd;
2864 /* Check that we still have all structures alive as this may
2865 * be some late RPC at shutdown time. */
2866 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2867 !exp_connect_lru_resize(req->rq_export))) {
2868 lustre_msg_set_slv(req->rq_repmsg, 0);
2869 lustre_msg_set_limit(req->rq_repmsg, 0);
2873 /* OBD is alive here as export is alive, which we checked above. */
2874 obd = req->rq_export->exp_obd;
2876 read_lock(&obd->obd_pool_lock);
2877 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2878 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2879 read_unlock(&obd->obd_pool_lock);
2884 static int target_send_reply_msg(struct ptlrpc_request *req,
2885 int rc, int fail_id)
2887 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2888 DEBUG_REQ(D_ERROR, req, "dropping reply");
2891 /* We can have a null rq_reqmsg in the event of bad signature or
2892 * no context when unwrapping */
2893 if (req->rq_reqmsg &&
2894 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2895 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2899 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2900 req->rq_status = rc;
2901 return ptlrpc_send_error(req, 1);
2903 DEBUG_REQ(D_NET, req, "sending reply");
2906 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2909 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2911 struct ptlrpc_service_part *svcpt;
2913 struct ptlrpc_reply_state *rs;
2914 struct obd_export *exp;
2917 if (req->rq_no_reply) {
2922 svcpt = req->rq_rqbd->rqbd_svcpt;
2923 rs = req->rq_reply_state;
2924 if (rs == NULL || !rs->rs_difficult) {
2926 target_send_reply_msg (req, rc, fail_id);
2931 /* must be an export if locks saved */
2932 LASSERT(req->rq_export != NULL);
2933 /* req/reply consistent */
2934 LASSERT(rs->rs_svcpt == svcpt);
2937 LASSERT(!rs->rs_scheduled);
2938 LASSERT(!rs->rs_scheduled_ever);
2939 LASSERT(!rs->rs_handled);
2940 LASSERT(!rs->rs_on_net);
2941 LASSERT(rs->rs_export == NULL);
2942 LASSERT(list_empty(&rs->rs_obd_list));
2943 LASSERT(list_empty(&rs->rs_exp_list));
2945 exp = class_export_get(req->rq_export);
2947 /* disable reply scheduling while I'm setting up */
2948 rs->rs_scheduled = 1;
2950 rs->rs_xid = req->rq_xid;
2951 rs->rs_transno = req->rq_transno;
2952 rs->rs_export = exp;
2953 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2955 spin_lock(&exp->exp_uncommitted_replies_lock);
2956 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2957 rs->rs_transno, exp->exp_last_committed);
2958 if (rs->rs_transno > exp->exp_last_committed) {
2959 /* not committed already */
2960 list_add_tail(&rs->rs_obd_list,
2961 &exp->exp_uncommitted_replies);
2963 spin_unlock(&exp->exp_uncommitted_replies_lock);
2965 spin_lock(&exp->exp_lock);
2966 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2967 spin_unlock(&exp->exp_lock);
2969 netrc = target_send_reply_msg(req, rc, fail_id);
2971 spin_lock(&svcpt->scp_rep_lock);
2973 atomic_inc(&svcpt->scp_nreps_difficult);
2976 /* error sending: reply is off the net. Also we need +1
2977 * reply ref until ptlrpc_handle_rs() is done
2978 * with the reply state (if the send was successful, there
2979 * would have been +1 ref for the net, which
2980 * reply_out_callback leaves alone) */
2982 ptlrpc_rs_addref(rs);
2985 spin_lock(&rs->rs_lock);
2986 if (rs->rs_transno <= exp->exp_last_committed ||
2987 (!rs->rs_on_net && !rs->rs_no_ack) ||
2988 list_empty(&rs->rs_exp_list) || /* completed already */
2989 list_empty(&rs->rs_obd_list)) {
2990 CDEBUG(D_HA, "Schedule reply immediately\n");
2991 ptlrpc_dispatch_difficult_reply(rs);
2993 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2994 rs->rs_scheduled = 0; /* allow notifier to schedule */
2996 spin_unlock(&rs->rs_lock);
2997 spin_unlock(&svcpt->scp_rep_lock);
3001 enum ldlm_mode lck_compat_array[] = {
3002 [LCK_EX] = LCK_COMPAT_EX,
3003 [LCK_PW] = LCK_COMPAT_PW,
3004 [LCK_PR] = LCK_COMPAT_PR,
3005 [LCK_CW] = LCK_COMPAT_CW,
3006 [LCK_CR] = LCK_COMPAT_CR,
3007 [LCK_NL] = LCK_COMPAT_NL,
3008 [LCK_GROUP] = LCK_COMPAT_GROUP,
3009 [LCK_COS] = LCK_COMPAT_COS,
3013 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3014 * not escape to the user level.
3016 int ldlm_error2errno(enum ldlm_error error)
3022 case ELDLM_LOCK_MATCHED:
3025 case ELDLM_LOCK_CHANGED:
3028 case ELDLM_LOCK_ABORTED:
3031 case ELDLM_LOCK_REPLACED:
3034 case ELDLM_NO_LOCK_DATA:
3037 case ELDLM_NAMESPACE_EXISTS:
3040 case ELDLM_BAD_NAMESPACE:
3044 if (((int)error) < 0) { /* cast to signed type */
3045 result = error; /* as ldlm_error can be unsigned */
3047 CERROR("Invalid DLM result code: %d\n", error);
3053 EXPORT_SYMBOL(ldlm_error2errno);
3056 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3058 enum ldlm_error ldlm_errno2error(int err_no)
3067 error = ELDLM_LOCK_CHANGED;
3070 error = ELDLM_LOCK_ABORTED;
3073 error = ELDLM_LOCK_REPLACED;
3076 error = ELDLM_NO_LOCK_DATA;
3079 error = ELDLM_NAMESPACE_EXISTS;
3082 error = ELDLM_BAD_NAMESPACE;
3090 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3091 void ldlm_dump_export_locks(struct obd_export *exp)
3093 spin_lock(&exp->exp_locks_list_guard);
3094 if (!list_empty(&exp->exp_locks_list)) {
3095 struct ldlm_lock *lock;
3097 CERROR("dumping locks for export %p,"
3098 "ignore if the unmount doesn't hang\n", exp);
3099 list_for_each_entry(lock, &exp->exp_locks_list,
3101 LDLM_ERROR(lock, "lock:");
3103 spin_unlock(&exp->exp_locks_list_guard);
3107 #ifdef HAVE_SERVER_SUPPORT
3108 static int target_bulk_timeout(void *data)
3111 /* We don't fail the connection here, because having the export
3112 * killed makes the (vital) call to commitrw very sad.
3117 static inline const char *bulk2type(struct ptlrpc_request *req)
3119 if (req->rq_bulk_read)
3121 if (req->rq_bulk_write)
3126 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3127 struct l_wait_info *lwi)
3129 struct ptlrpc_request *req = desc->bd_req;
3130 time64_t start = ktime_get_real_seconds();
3136 /* If there is eviction in progress, wait for it to finish. */
3137 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3138 *lwi = LWI_INTR(NULL, NULL);
3139 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3140 !atomic_read(&exp->exp_obd->
3141 obd_evict_inprogress),
3145 /* Check if client was evicted or reconnected already. */
3146 if (exp->exp_failed ||
3147 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3150 if (req->rq_bulk_read)
3151 rc = sptlrpc_svc_wrap_bulk(req, desc);
3153 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3154 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3155 else /* old version, bulk matchbits is rq_xid */
3156 req->rq_mbits = req->rq_xid;
3159 rc = ptlrpc_start_bulk_transfer(desc);
3163 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3164 bulk2type(req), rc);
3168 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3169 ptlrpc_abort_bulk(desc);
3173 /* limit actual bulk transfer to bulk_timeout seconds */
3174 deadline = start + bulk_timeout;
3175 if (deadline > req->rq_deadline)
3176 deadline = req->rq_deadline;
3179 time64_t timeoutl = deadline - ktime_get_real_seconds();
3180 long timeout_jiffies = timeoutl <= 0 ?
3181 1 : cfs_time_seconds(timeoutl);
3182 time64_t rq_deadline;
3184 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3185 cfs_time_seconds(1),
3186 target_bulk_timeout, desc);
3187 rc = l_wait_event(desc->bd_waitq,
3188 !ptlrpc_server_bulk_active(desc) ||
3191 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3193 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3194 /* Wait again if we changed rq_deadline. */
3195 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3196 deadline = start + bulk_timeout;
3197 if (deadline > rq_deadline)
3198 deadline = rq_deadline;
3199 } while (rc == -ETIMEDOUT &&
3200 deadline > ktime_get_real_seconds());
3202 if (rc == -ETIMEDOUT) {
3203 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3204 bulk2type(req), deadline - start,
3205 ktime_get_real_seconds() - deadline);
3206 ptlrpc_abort_bulk(desc);
3207 } else if (exp->exp_failed) {
3208 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3211 ptlrpc_abort_bulk(desc);
3212 } else if (exp->exp_conn_cnt >
3213 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3214 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3216 /* We don't reply anyway. */
3218 ptlrpc_abort_bulk(desc);
3219 } else if (desc->bd_failure) {
3220 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3222 /* XXX should this be a different errno? */
3225 if (req->rq_bulk_write)
3226 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3227 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3228 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3229 bulk2type(req), desc->bd_nob_transferred,
3231 /* XXX should this be a different errno? */
3238 EXPORT_SYMBOL(target_bulk_io);
3240 #endif /* HAVE_SERVER_SUPPORT */