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 /* set cl_chunkbits default value to PAGE_SHIFT,
413 * it will be updated at OSC connection time. */
414 cli->cl_chunkbits = PAGE_SHIFT;
416 if (!strcmp(name, LUSTRE_MDC_NAME)) {
417 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
418 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
419 cli->cl_max_rpcs_in_flight = 2;
420 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
421 cli->cl_max_rpcs_in_flight = 3;
422 } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
423 cli->cl_max_rpcs_in_flight = 4;
425 if (osc_on_mdt(obddev->obd_name))
426 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
428 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
431 spin_lock_init(&cli->cl_mod_rpcs_lock);
432 spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
433 cli->cl_max_mod_rpcs_in_flight = 0;
434 cli->cl_mod_rpcs_in_flight = 0;
435 cli->cl_close_rpcs_in_flight = 0;
436 init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
437 cli->cl_mod_tag_bitmap = NULL;
439 INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
441 if (connect_op == MDS_CONNECT) {
442 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
443 OBD_ALLOC(cli->cl_mod_tag_bitmap,
444 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
445 if (cli->cl_mod_tag_bitmap == NULL)
446 GOTO(err, rc = -ENOMEM);
451 CERROR("ldlm_get_ref failed: %d\n", rc);
455 ptlrpc_init_client(rq_portal, rp_portal, name,
456 &obddev->obd_ldlm_client);
458 imp = class_new_import(obddev);
460 GOTO(err_ldlm, rc = -ENOENT);
461 imp->imp_client = &obddev->obd_ldlm_client;
462 imp->imp_connect_op = connect_op;
463 memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
464 LUSTRE_CFG_BUFLEN(lcfg, 1));
465 class_import_put(imp);
467 if (lustre_cfg_buf(lcfg, 4)) {
468 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
470 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
472 CERROR("%s: bad mount option 'network=%s': rc = %d\n",
473 obddev->obd_name, lustre_cfg_string(lcfg, 4),
475 GOTO(err_import, rc);
477 fake_conn.c_self = LNET_MKNID(refnet, 0);
478 imp->imp_connection = &fake_conn;
481 rc = client_import_add_conn(imp, &server_uuid, 1);
483 CERROR("can't add initial connection\n");
484 GOTO(err_import, rc);
486 imp->imp_connection = NULL;
488 cli->cl_import = imp;
489 /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
490 cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
492 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
493 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
494 CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
495 name, obddev->obd_name,
496 cli->cl_target_uuid.uuid);
497 spin_lock(&imp->imp_lock);
498 imp->imp_deactive = 1;
499 spin_unlock(&imp->imp_lock);
503 obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
504 LDLM_NAMESPACE_CLIENT,
505 LDLM_NAMESPACE_GREEDY,
507 if (obddev->obd_namespace == NULL) {
508 CERROR("Unable to create client namespace - %s\n",
510 GOTO(err_import, rc = -ENOMEM);
516 class_destroy_import(imp);
520 if (cli->cl_mod_tag_bitmap != NULL)
521 OBD_FREE(cli->cl_mod_tag_bitmap,
522 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
523 cli->cl_mod_tag_bitmap = NULL;
527 EXPORT_SYMBOL(client_obd_setup);
529 int client_obd_cleanup(struct obd_device *obddev)
531 struct client_obd *cli = &obddev->u.cli;
534 ldlm_namespace_free_post(obddev->obd_namespace);
535 obddev->obd_namespace = NULL;
537 obd_cleanup_client_import(obddev);
538 LASSERT(obddev->u.cli.cl_import == NULL);
542 if (cli->cl_mod_tag_bitmap != NULL)
543 OBD_FREE(cli->cl_mod_tag_bitmap,
544 BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
545 cli->cl_mod_tag_bitmap = NULL;
549 EXPORT_SYMBOL(client_obd_cleanup);
551 /* ->o_connect() method for client side (OSC and MDC and MGC) */
552 int client_connect_import(const struct lu_env *env,
553 struct obd_export **exp,
554 struct obd_device *obd, struct obd_uuid *cluuid,
555 struct obd_connect_data *data, void *localdata)
557 struct client_obd *cli = &obd->u.cli;
558 struct obd_import *imp = cli->cl_import;
559 struct obd_connect_data *ocd;
560 struct lustre_handle conn = { 0 };
565 down_write(&cli->cl_sem);
566 if (cli->cl_conn_count > 0)
567 GOTO(out_sem, rc = -EALREADY);
569 rc = class_connect(&conn, obd, cluuid);
573 cli->cl_conn_count++;
574 *exp = class_conn2export(&conn);
576 LASSERT(obd->obd_namespace);
578 imp->imp_dlm_handle = conn;
579 rc = ptlrpc_init_import(imp);
583 ocd = &imp->imp_connect_data;
586 imp->imp_connect_flags_orig = data->ocd_connect_flags;
587 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
590 rc = ptlrpc_connect_import(imp);
592 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
595 LASSERT(*exp != NULL && (*exp)->exp_connection);
598 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
599 ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
600 data->ocd_connect_flags, ocd->ocd_connect_flags);
601 data->ocd_connect_flags = ocd->ocd_connect_flags;
602 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
605 ptlrpc_pinger_add_import(imp);
611 cli->cl_conn_count--;
612 class_disconnect(*exp);
616 up_write(&cli->cl_sem);
620 EXPORT_SYMBOL(client_connect_import);
622 int client_disconnect_export(struct obd_export *exp)
624 struct obd_device *obd = class_exp2obd(exp);
625 struct client_obd *cli;
626 struct obd_import *imp;
631 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
632 exp, exp ? exp->exp_handle.h_cookie : -1);
637 imp = cli->cl_import;
639 down_write(&cli->cl_sem);
640 CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
643 if (cli->cl_conn_count == 0) {
644 CERROR("disconnecting disconnected device (%s)\n",
646 GOTO(out_disconnect, rc = -EINVAL);
649 cli->cl_conn_count--;
650 if (cli->cl_conn_count != 0)
651 GOTO(out_disconnect, rc = 0);
653 /* Mark import deactivated now, so we don't try to reconnect if any
654 * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't
655 * fully deactivate the import, or that would drop all requests. */
656 spin_lock(&imp->imp_lock);
657 imp->imp_deactive = 1;
658 spin_unlock(&imp->imp_lock);
660 /* Some non-replayable imports (MDS's OSCs) are pinged, so just
661 * delete it regardless. (It's safe to delete an import that was
663 (void)ptlrpc_pinger_del_import(imp);
665 if (obd->obd_namespace != NULL) {
666 /* obd_force == local only */
667 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
668 obd->obd_force ? LCF_LOCAL : 0, NULL);
669 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
672 /* There's no need to hold sem while disconnecting an import,
673 * and it may actually cause deadlock in GSS. */
674 up_write(&cli->cl_sem);
675 rc = ptlrpc_disconnect_import(imp, 0);
676 down_write(&cli->cl_sem);
678 ptlrpc_invalidate_import(imp);
683 /* Use server style - class_disconnect should be always called for
685 err = class_disconnect(exp);
689 up_write(&cli->cl_sem);
693 EXPORT_SYMBOL(client_disconnect_export);
695 #ifdef HAVE_SERVER_SUPPORT
696 int server_disconnect_export(struct obd_export *exp)
701 /* Disconnect early so that clients can't keep using export. */
702 rc = class_disconnect(exp);
703 /* Close import to avoid sending any requests. */
704 if (exp->exp_imp_reverse)
705 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
707 ldlm_bl_thread_wakeup();
709 /* complete all outstanding replies */
710 spin_lock(&exp->exp_lock);
711 while (!list_empty(&exp->exp_outstanding_replies)) {
712 struct ptlrpc_reply_state *rs =
713 list_entry(exp->exp_outstanding_replies.next,
714 struct ptlrpc_reply_state, rs_exp_list);
715 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
717 spin_lock(&svcpt->scp_rep_lock);
719 list_del_init(&rs->rs_exp_list);
721 spin_lock(&rs->rs_lock);
722 /* clear rs_convert_lock to make sure rs is handled and put */
723 rs->rs_convert_lock = 0;
724 ptlrpc_schedule_difficult_reply(rs);
725 spin_unlock(&rs->rs_lock);
727 spin_unlock(&svcpt->scp_rep_lock);
729 spin_unlock(&exp->exp_lock);
733 EXPORT_SYMBOL(server_disconnect_export);
735 /* --------------------------------------------------------------------------
736 * from old lib/target.c
737 * -------------------------------------------------------------------------- */
739 static int target_handle_reconnect(struct lustre_handle *conn,
740 struct obd_export *exp,
741 struct obd_uuid *cluuid)
743 struct obd_device *target;
744 struct lustre_handle *hdl;
751 hdl = &exp->exp_imp_reverse->imp_remote_handle;
752 if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
753 conn->cookie = exp->exp_handle.h_cookie;
754 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
755 " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
759 target = exp->exp_obd;
761 /* Might be a re-connect after a partition. */
762 if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
763 LCONSOLE_WARN("%s: already connected client %s (at %s) "
764 "with handle %#llx. Rejecting client "
765 "with the same UUID trying to reconnect "
766 "with handle %#llx\n", target->obd_name,
767 obd_uuid2str(&exp->exp_client_uuid),
768 obd_export_nid2str(exp),
769 hdl->cookie, conn->cookie);
770 memset(conn, 0, sizeof *conn);
771 /* target_handle_connect() treats EALREADY and
772 * -EALREADY differently. -EALREADY is an error
773 * (same UUID, different handle). */
777 if (!target->obd_recovering) {
778 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
779 target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
780 obd_export_nid2str(exp));
781 GOTO(out_already, rc);
784 now = cfs_time_current();
785 deadline = target->obd_recovery_timer.expires;
786 if (cfs_time_before(now, deadline)) {
787 struct target_distribute_txn_data *tdtd =
788 class_exp2tgt(exp)->lut_tdtd;
793 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
794 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
795 buf = tdtd->tdtd_show_update_logs_retrievers(
796 tdtd->tdtd_show_retrievers_cbdata,
800 LCONSOLE_WARN("%s: Recovery already passed deadline "
801 "%d:%.02d. It is due to DNE recovery "
802 "failed/stuck on the %d MDT(s):%s. "
803 "Please wait until all MDTs recovered "
804 "or abort the recovery by force.\n",
805 target->obd_name, timeout / 60,
807 buf ? buf : "unknown (not enough RAM)");
809 LCONSOLE_WARN("%s: Recovery already passed deadline "
810 "%d:%.02d. If you do not want to wait "
811 "more, please abort the recovery by "
812 "force.\n", target->obd_name,
813 timeout / 60, timeout % 60);
818 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
819 LCONSOLE_WARN("%s: Recovery already passed deadline"
820 " %d:%.02d, It is most likely due to DNE"
821 " recovery is failed or stuck, please wait a"
822 " few more minutes or abort the recovery.\n",
823 target->obd_name, timeout / 60, timeout % 60);
827 conn->cookie = exp->exp_handle.h_cookie;
828 /* target_handle_connect() treats EALREADY and
829 * -EALREADY differently. EALREADY means we are
830 * doing a valid reconnect from the same client. */
835 check_and_start_recovery_timer(struct obd_device *obd,
836 struct ptlrpc_request *req, int new_client);
839 * update flags for import during reconnect process
841 static int rev_import_flags_update(struct obd_import *revimp,
842 struct ptlrpc_request *req)
845 struct obd_connect_data *data;
847 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
849 if (data->ocd_connect_flags & OBD_CONNECT_AT)
850 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
852 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
854 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
856 rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
858 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
859 revimp->imp_client->cli_name,
860 libcfs_id2str(req->rq_peer), rc);
868 * Allocate a new reverse import for an export.
870 * \retval -errno in case error hit
871 * \retval 0 if reverse import correctly init
873 int rev_import_init(struct obd_export *export)
875 struct obd_device *obd = export->exp_obd;
876 struct obd_import *revimp;
878 LASSERT(export->exp_imp_reverse == NULL);
880 revimp = class_new_import(obd);
884 revimp->imp_remote_handle.cookie = 0ULL;
885 revimp->imp_client = &obd->obd_ldlm_client;
886 revimp->imp_dlm_fake = 1;
888 /* it is safe to connect import in new state as no sends possible */
889 spin_lock(&export->exp_lock);
890 export->exp_imp_reverse = revimp;
891 spin_unlock(&export->exp_lock);
892 class_import_put(revimp);
896 EXPORT_SYMBOL(rev_import_init);
899 * Handle reconnect for an export.
901 * \param exp export to handle reconnect process
902 * \param req client reconnect request
904 * \retval -rc in case securitfy flavor can't be changed
905 * \retval 0 in case none problems
907 static int rev_import_reconnect(struct obd_export *exp,
908 struct ptlrpc_request *req)
910 struct obd_import *revimp = exp->exp_imp_reverse;
911 struct lustre_handle *lh;
914 /* avoid sending a request until import flags are changed */
915 ptlrpc_import_enter_resend(revimp);
917 if (revimp->imp_connection != NULL)
918 ptlrpc_connection_put(revimp->imp_connection);
921 * client from recovery don't have a handle so we need to take from
922 * request. it may produce situation when wrong client connected
923 * to recovery as we trust a client uuid
925 lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
926 revimp->imp_remote_handle = *lh;
928 /* unknown versions will be caught in
929 * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
930 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
932 revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
934 rc = rev_import_flags_update(revimp, req);
936 /* it is safe to still be in RECOVERY phase as we are not able
937 * to setup correct security flavor so requests are not able to
938 * be delivered correctly */
942 /* resend all rpc's via new connection */
943 return ptlrpc_import_recovery_state_machine(revimp);
946 int target_handle_connect(struct ptlrpc_request *req)
948 struct obd_device *target = NULL;
949 struct obd_export *export = NULL;
950 /* connect handle - filled from target_handle_reconnect in
952 struct lustre_handle conn;
953 struct lustre_handle *tmp;
954 struct obd_uuid tgtuuid;
955 struct obd_uuid cluuid;
960 bool mds_conn = false, lw_client = false, initial_conn = false;
961 bool mds_mds_conn = false;
962 bool new_mds_mds_conn = false;
963 bool target_referenced = 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 obd_str2uuid(&tgtuuid, str);
978 target = class_uuid2obd(&tgtuuid);
980 target = class_name2obd(str);
983 deuuidify(str, NULL, &target_start, &target_len);
984 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
985 "from %s (no target). If you are running "
986 "an HA pair check that the target is "
987 "mounted on the other server.\n", str,
988 libcfs_nid2str(req->rq_peer.nid));
989 GOTO(out, rc = -ENODEV);
992 spin_lock(&target->obd_dev_lock);
993 if (target->obd_stopping || !target->obd_set_up) {
994 spin_unlock(&target->obd_dev_lock);
996 deuuidify(str, NULL, &target_start, &target_len);
997 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
998 target_len, target_start,
999 libcfs_nid2str(req->rq_peer.nid),
1000 (target->obd_stopping ?
1001 "stopping" : "not set up"));
1002 GOTO(out, rc = -ENODEV);
1005 if (target->obd_no_conn) {
1006 spin_unlock(&target->obd_dev_lock);
1008 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1009 "from %s\n", target->obd_name,
1010 libcfs_nid2str(req->rq_peer.nid));
1011 GOTO(out, rc = -EAGAIN);
1014 /* Make sure the target isn't cleaned up while we're here. Yes,
1015 * there's still a race between the above check and our incref here.
1016 * Really, class_uuid2obd should take the ref. */
1017 class_incref(target, __func__, current);
1018 target_referenced = true;
1020 target->obd_conn_inprogress++;
1021 spin_unlock(&target->obd_dev_lock);
1023 str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1025 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1026 GOTO(out, rc = -EINVAL);
1029 obd_str2uuid(&cluuid, str);
1031 tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1033 GOTO(out, rc = -EPROTO);
1037 size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1039 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1041 GOTO(out, rc = -EPROTO);
1043 rc = req_capsule_server_pack(&req->rq_pill);
1047 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1048 /* Don't allow clients to connect that are using old 1.8 format
1049 * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1050 * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The
1051 * FULL20 flag should be set on all connections since 2.0, but no
1052 * longer affects behaviour.
1054 * Later this check will be disabled and the flag can be retired
1055 * completely once interop with 3.0 is no longer needed.
1057 if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1058 GOTO(out, rc = -EPROTO);
1061 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1062 if (data->ocd_version < LUSTRE_VERSION_CODE -
1063 LUSTRE_VERSION_ALLOWED_OFFSET ||
1064 data->ocd_version > LUSTRE_VERSION_CODE +
1065 LUSTRE_VERSION_ALLOWED_OFFSET) {
1066 DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1067 "libclient connection attempt",
1068 data->ocd_version < LUSTRE_VERSION_CODE ?
1070 OBD_OCD_VERSION_MAJOR(data->ocd_version),
1071 OBD_OCD_VERSION_MINOR(data->ocd_version),
1072 OBD_OCD_VERSION_PATCH(data->ocd_version),
1073 OBD_OCD_VERSION_FIX(data->ocd_version));
1074 data = req_capsule_server_sized_get(&req->rq_pill,
1076 offsetof(typeof(*data), ocd_version) +
1077 sizeof(data->ocd_version));
1079 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1080 data->ocd_version = LUSTRE_VERSION_CODE;
1082 GOTO(out, rc = -EPROTO);
1086 /* Note: lw_client is needed in MDS-MDS failover during update log
1087 * processing, so we needs to allow lw_client to be connected at
1088 * anytime, instead of only the initial connection */
1089 lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1091 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1092 initial_conn = true;
1093 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1094 mds_mds_conn = (data->ocd_connect_flags &
1095 OBD_CONNECT_MDS_MDS) != 0;
1097 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1098 * for Imperative Recovery connection from MGC to MGS.
1100 * Via check OBD_CONNECT_FID, we can distinguish whether
1101 * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1104 (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1105 (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1106 (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1107 __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1108 __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1109 __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1111 /* We do not support the MDT-MDT interoperations with
1112 * different version MDT because of protocol changes. */
1113 if (unlikely(major != LUSTRE_MAJOR ||
1114 minor != LUSTRE_MINOR ||
1115 abs(patch - LUSTRE_PATCH) > 3)) {
1116 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1117 "connection from different version MDT "
1118 "(%d.%d.%d.%d) %s %s\n",
1119 target->obd_name, LUSTRE_MAJOR,
1120 LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1121 major, minor, patch,
1122 OBD_OCD_VERSION_FIX(data->ocd_version),
1123 libcfs_nid2str(req->rq_peer.nid), str);
1125 GOTO(out, rc = -EPROTO);
1130 /* lctl gets a backstage, all-access pass. */
1131 if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1132 goto dont_check_exports;
1134 export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1138 /* We've found an export in the hash. */
1140 spin_lock(&export->exp_lock);
1142 if (export->exp_connecting) { /* bug 9635, et. al. */
1143 spin_unlock(&export->exp_lock);
1144 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1145 export->exp_obd->obd_name, export,
1146 libcfs_nid2str(req->rq_peer.nid));
1147 class_export_put(export);
1150 } else if ((mds_conn || (lw_client && initial_conn) ||
1151 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1152 export->exp_connection != NULL) {
1153 spin_unlock(&export->exp_lock);
1154 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1155 /* MDS or LWP reconnected after failover. */
1156 LCONSOLE_WARN("%s: Received %s connection from "
1157 "%s, removing former export from %s\n",
1158 target->obd_name, mds_conn ? "MDS" : "LWP",
1159 libcfs_nid2str(req->rq_peer.nid),
1160 libcfs_nid2str(export->exp_connection->c_peer.nid));
1162 /* New MDS connection from the same NID. */
1163 LCONSOLE_WARN("%s: Received new %s connection from "
1164 "%s, removing former export from same NID\n",
1165 target->obd_name, mds_conn ? "MDS" : "LWP",
1166 libcfs_nid2str(req->rq_peer.nid));
1169 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1170 data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1171 /* Because exports between MDTs will always be
1172 * kept, let's do not fail such export if they
1173 * come from the same NID, otherwise it might
1174 * cause eviction between MDTs, which might
1175 * cause namespace inconsistency */
1176 spin_lock(&export->exp_lock);
1177 export->exp_connecting = 1;
1178 spin_unlock(&export->exp_lock);
1179 conn.cookie = export->exp_handle.h_cookie;
1182 class_fail_export(export);
1183 class_export_put(export);
1187 } else if (export->exp_connection != NULL && initial_conn &&
1188 req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1189 spin_unlock(&export->exp_lock);
1190 /* In MDS failover we have static UUID but NID can change. */
1191 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1192 "existing nid %s is already connected\n",
1193 target->obd_name, cluuid.uuid,
1194 libcfs_nid2str(req->rq_peer.nid),
1196 export->exp_connection->c_peer.nid));
1198 class_export_put(export);
1201 export->exp_connecting = 1;
1202 spin_unlock(&export->exp_lock);
1203 LASSERT(export->exp_obd == target);
1205 rc = target_handle_reconnect(&conn, export, &cluuid);
1208 /* If we found an export, we already unlocked. */
1211 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1212 } else if (req->rq_export == NULL &&
1213 atomic_read(&export->exp_rpc_count) > 0) {
1214 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1215 "still busy with %d references\n",
1216 target->obd_name, cluuid.uuid,
1217 libcfs_nid2str(req->rq_peer.nid),
1218 atomic_read(&export->exp_refcount));
1219 GOTO(out, rc = -EBUSY);
1220 } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1221 if (!strstr(cluuid.uuid, "mdt"))
1222 LCONSOLE_WARN("%s: Rejecting reconnect from the "
1223 "known client %s (at %s) because it "
1224 "is indicating it is a new client",
1225 target->obd_name, cluuid.uuid,
1226 libcfs_nid2str(req->rq_peer.nid));
1227 GOTO(out, rc = -EALREADY);
1229 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1236 CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1237 target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1238 target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1239 export, (long)cfs_time_current_sec(),
1240 export ? (long)export->exp_last_request_time : 0);
1242 /* If this is the first time a client connects, reset the recovery
1243 * timer. Discard lightweight connections which might be local. */
1244 if (!lw_client && rc == 0 && target->obd_recovering)
1245 check_and_start_recovery_timer(target, req, export == NULL);
1247 /* We want to handle EALREADY but *not* -EALREADY from
1248 * target_handle_reconnect(), return reconnection state in a flag. */
1249 if (rc == EALREADY) {
1250 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1256 /* Tell the client if we support replayable requests. */
1257 if (target->obd_replayable)
1258 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1259 client_nid = &req->rq_peer.nid;
1261 if (export == NULL) {
1262 /* allow lightweight connections during recovery */
1263 /* allow "new" MDT to be connected during recovery, since we
1264 * need retrieve recovery update records from it */
1265 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1267 int c; /* connected */
1268 int i; /* in progress */
1270 int s; /* stale/evicted */
1272 c = atomic_read(&target->obd_connected_clients);
1273 i = atomic_read(&target->obd_lock_replay_clients);
1274 k = target->obd_max_recoverable_clients;
1275 s = target->obd_stale_clients;
1276 t = target->obd_recovery_timer.expires;
1277 t = cfs_time_sub(t, cfs_time_current());
1278 t = cfs_duration_sec(t);
1279 LCONSOLE_WARN("%s: Denying connection for new client %s"
1280 "(at %s), waiting for %d known clients "
1281 "(%d recovered, %d in progress, and %d "
1282 "evicted) to recover in %d:%.02d\n",
1283 target->obd_name, cluuid.uuid,
1284 libcfs_nid2str(req->rq_peer.nid), k,
1285 c - i, i, s, (int)t / 60,
1290 rc = obd_connect(req->rq_svc_thread->t_env,
1291 &export, target, &cluuid, data,
1293 if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1294 lustre_msg_add_op_flags(req->rq_repmsg,
1295 MSG_CONNECT_RECOVERING);
1297 conn.cookie = export->exp_handle.h_cookie;
1298 rc = rev_import_init(export);
1302 new_mds_mds_conn = true;
1305 rc = obd_reconnect(req->rq_svc_thread->t_env,
1306 export, target, &cluuid, data, client_nid);
1311 LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1312 data->ocd_instance = target->u.obt.obt_instance;
1314 /* Return only the parts of obd_connect_data that we understand, so the
1315 * client knows that we don't understand the rest. */
1317 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1319 tmpdata = req_capsule_server_get(&req->rq_pill,
1321 /* Don't use struct assignment here, because the client reply
1322 * buffer may be smaller/larger than the local struct
1323 * obd_connect_data. */
1324 memcpy(tmpdata, data, min(tmpsize, size));
1327 /* If the client and the server are the same node, we will already
1328 * have an export that really points to the client's DLM export,
1329 * because we have a shared handles table.
1331 * XXX this will go away when shaver stops sending the "connect" handle
1332 * in the real "remote handle" field of the request --phik 24 Apr 2003
1334 ptlrpc_request_change_export(req, export);
1336 spin_lock(&export->exp_lock);
1337 if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1338 spin_unlock(&export->exp_lock);
1339 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1340 "or equal conn_cnt: %d >= %d\n",
1341 cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1342 export->exp_conn_cnt,
1343 lustre_msg_get_conn_cnt(req->rq_reqmsg));
1345 GOTO(out, rc = -EALREADY);
1347 LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1348 export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1350 /* Don't evict liblustre clients for not pinging. */
1351 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1352 export->exp_libclient = 1;
1353 spin_unlock(&export->exp_lock);
1355 spin_lock(&target->obd_dev_lock);
1356 list_del_init(&export->exp_obd_chain_timed);
1357 spin_unlock(&target->obd_dev_lock);
1359 spin_unlock(&export->exp_lock);
1362 if (export->exp_connection != NULL) {
1363 /* Check to see if connection came from another NID. */
1364 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1365 !hlist_unhashed(&export->exp_nid_hash))
1366 cfs_hash_del(export->exp_obd->obd_nid_hash,
1367 &export->exp_connection->c_peer.nid,
1368 &export->exp_nid_hash);
1370 ptlrpc_connection_put(export->exp_connection);
1373 export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1376 if (hlist_unhashed(&export->exp_nid_hash))
1377 cfs_hash_add(export->exp_obd->obd_nid_hash,
1378 &export->exp_connection->c_peer.nid,
1379 &export->exp_nid_hash);
1381 lustre_msg_set_handle(req->rq_repmsg, &conn);
1383 rc = rev_import_reconnect(export, req);
1387 if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1389 __u64 transno = data->ocd_transno;
1391 spin_lock(&export->exp_lock);
1392 /* possible race with class_disconnect_stale_exports,
1393 * export may be already in the eviction process */
1394 if (export->exp_failed) {
1395 spin_unlock(&export->exp_lock);
1396 GOTO(out, rc = -ENODEV);
1398 export->exp_in_recovery = 1;
1399 export->exp_req_replay_needed = 1;
1400 export->exp_lock_replay_needed = 1;
1401 spin_unlock(&export->exp_lock);
1403 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1404 MSG_CONNECT_TRANSNO);
1405 if (has_transno && transno == 0)
1406 CWARN("Connect with zero transno!\n");
1408 if (has_transno && transno > 0 &&
1409 transno < target->obd_next_recovery_transno &&
1410 transno > target->obd_last_committed) {
1411 /* Another way is to use cmpxchg() to be lock-free. */
1412 spin_lock(&target->obd_recovery_task_lock);
1413 if (transno < target->obd_next_recovery_transno)
1414 target->obd_next_recovery_transno = transno;
1415 spin_unlock(&target->obd_recovery_task_lock);
1418 atomic_inc(&target->obd_req_replay_clients);
1419 atomic_inc(&target->obd_lock_replay_clients);
1420 /* Note: MDS-MDS connection is allowed to be connected during
1421 * recovery, no matter if the exports needs to be recoveried.
1422 * Because we need retrieve updates logs from all other MDTs.
1423 * So if the MDS-MDS export is new, obd_max_recoverable_clients
1424 * also needs to be increased to match other recovery checking
1426 if (new_mds_mds_conn)
1427 target->obd_max_recoverable_clients++;
1428 if (atomic_inc_return(&target->obd_connected_clients) ==
1429 target->obd_max_recoverable_clients)
1430 wake_up(&target->obd_next_transno_waitq);
1433 /* Tell the client we're in recovery, when client is involved in it. */
1434 if (target->obd_recovering && !lw_client)
1435 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1439 spin_lock(&export->exp_lock);
1440 export->exp_connecting = 0;
1441 spin_unlock(&export->exp_lock);
1443 class_export_put(export);
1445 if (target_referenced == true && target != NULL) {
1446 spin_lock(&target->obd_dev_lock);
1447 target->obd_conn_inprogress--;
1448 spin_unlock(&target->obd_dev_lock);
1450 class_decref(target, __func__, current);
1452 req->rq_status = rc;
1456 int target_handle_disconnect(struct ptlrpc_request *req)
1461 rc = req_capsule_server_pack(&req->rq_pill);
1465 /* Keep the rq_export around so we can send the reply. */
1466 req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1471 void target_destroy_export(struct obd_export *exp)
1473 struct obd_import *imp = NULL;
1474 /* exports created from last_rcvd data, and "fake"
1475 exports created by lctl don't have an import */
1476 spin_lock(&exp->exp_lock);
1477 if (exp->exp_imp_reverse != NULL) {
1478 imp = exp->exp_imp_reverse;
1479 exp->exp_imp_reverse = NULL;
1481 spin_unlock(&exp->exp_lock);
1483 client_destroy_import(imp);
1485 LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1486 LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1487 LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1488 LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1490 EXPORT_SYMBOL(target_destroy_export);
1493 * Recovery functions
1495 static void target_request_copy_get(struct ptlrpc_request *req)
1497 class_export_rpc_inc(req->rq_export);
1498 LASSERT(list_empty(&req->rq_list));
1499 INIT_LIST_HEAD(&req->rq_replay_list);
1501 /* Increase refcount to keep request in queue. */
1502 atomic_inc(&req->rq_refcount);
1503 /* Let export know it has replays to be handled. */
1504 atomic_inc(&req->rq_export->exp_replay_count);
1507 static void target_request_copy_put(struct ptlrpc_request *req)
1509 LASSERT(list_empty(&req->rq_replay_list));
1510 LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1512 atomic_dec(&req->rq_export->exp_replay_count);
1513 class_export_rpc_dec(req->rq_export);
1514 ptlrpc_server_drop_request(req);
1517 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1519 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1520 struct obd_export *exp = req->rq_export;
1521 struct ptlrpc_request *reqiter;
1522 struct ptlrpc_request *dup_req = NULL;
1527 spin_lock(&exp->exp_lock);
1528 list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1530 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1538 /* We expect it with RESENT and REPLAY flags. */
1539 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1540 (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1541 CERROR("invalid flags %x of resent replay\n",
1542 lustre_msg_get_flags(req->rq_reqmsg));
1544 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1547 new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1549 lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1550 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1554 list_add_tail(&req->rq_replay_list,
1555 &exp->exp_req_replay_queue);
1558 spin_unlock(&exp->exp_lock);
1562 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1564 LASSERT(!list_empty(&req->rq_replay_list));
1565 LASSERT(req->rq_export);
1567 spin_lock(&req->rq_export->exp_lock);
1568 list_del_init(&req->rq_replay_list);
1569 spin_unlock(&req->rq_export->exp_lock);
1572 static void target_finish_recovery(struct lu_target *lut)
1574 struct obd_device *obd = lut->lut_obd;
1577 /* Only log a recovery message when recovery has occurred. */
1578 if (obd->obd_recovery_start) {
1579 time64_t now = ktime_get_real_seconds();
1580 time64_t elapsed_time;
1582 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1583 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1584 "%d recovered and %d %s evicted.\n", obd->obd_name,
1585 (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1586 obd->obd_max_recoverable_clients,
1587 atomic_read(&obd->obd_connected_clients),
1588 obd->obd_stale_clients,
1589 obd->obd_stale_clients == 1 ? "was" : "were");
1592 ldlm_reprocess_recovery_done(obd->obd_namespace);
1593 spin_lock(&obd->obd_recovery_task_lock);
1594 if (!list_empty(&obd->obd_req_replay_queue) ||
1595 !list_empty(&obd->obd_lock_replay_queue) ||
1596 !list_empty(&obd->obd_final_req_queue)) {
1597 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1599 list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1600 list_empty(&obd->obd_lock_replay_queue) ? \
1602 list_empty(&obd->obd_final_req_queue) ? \
1604 spin_unlock(&obd->obd_recovery_task_lock);
1607 spin_unlock(&obd->obd_recovery_task_lock);
1609 obd->obd_recovery_end = ktime_get_real_seconds();
1611 /* When recovery finished, cleanup orphans on MDS and OST. */
1612 if (OBT(obd) && OBP(obd, postrecov)) {
1613 int rc = OBP(obd, postrecov)(obd);
1615 LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1621 static void abort_req_replay_queue(struct obd_device *obd)
1623 struct ptlrpc_request *req, *n;
1624 struct list_head abort_list;
1626 INIT_LIST_HEAD(&abort_list);
1627 spin_lock(&obd->obd_recovery_task_lock);
1628 list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1629 spin_unlock(&obd->obd_recovery_task_lock);
1630 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1631 DEBUG_REQ(D_WARNING, req, "aborted:");
1632 req->rq_status = -ENOTCONN;
1633 if (ptlrpc_error(req)) {
1634 DEBUG_REQ(D_ERROR, req,
1635 "failed abort_req_reply; skipping");
1637 target_exp_dequeue_req_replay(req);
1638 target_request_copy_put(req);
1642 static void abort_lock_replay_queue(struct obd_device *obd)
1644 struct ptlrpc_request *req, *n;
1645 struct list_head abort_list;
1647 INIT_LIST_HEAD(&abort_list);
1648 spin_lock(&obd->obd_recovery_task_lock);
1649 list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1650 spin_unlock(&obd->obd_recovery_task_lock);
1651 list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1652 DEBUG_REQ(D_ERROR, req, "aborted:");
1653 req->rq_status = -ENOTCONN;
1654 if (ptlrpc_error(req)) {
1655 DEBUG_REQ(D_ERROR, req,
1656 "failed abort_lock_reply; skipping");
1658 target_request_copy_put(req);
1662 /* Called from a cleanup function if the device is being cleaned up
1663 forcefully. The exports should all have been disconnected already,
1664 the only thing left to do is
1665 - clear the recovery flags
1667 - free queued requests and replies, but don't send replies
1668 Because the obd_stopping flag is set, no new requests should be received.
1671 void target_cleanup_recovery(struct obd_device *obd)
1673 struct ptlrpc_request *req, *n;
1674 struct list_head clean_list;
1676 INIT_LIST_HEAD(&clean_list);
1677 spin_lock(&obd->obd_dev_lock);
1678 if (!obd->obd_recovering) {
1679 spin_unlock(&obd->obd_dev_lock);
1683 obd->obd_recovering = obd->obd_abort_recovery = 0;
1684 spin_unlock(&obd->obd_dev_lock);
1686 spin_lock(&obd->obd_recovery_task_lock);
1687 target_cancel_recovery_timer(obd);
1688 list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1689 spin_unlock(&obd->obd_recovery_task_lock);
1691 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1692 LASSERT(req->rq_reply_state == NULL);
1693 target_exp_dequeue_req_replay(req);
1694 target_request_copy_put(req);
1697 spin_lock(&obd->obd_recovery_task_lock);
1698 list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1699 list_splice_init(&obd->obd_final_req_queue, &clean_list);
1700 spin_unlock(&obd->obd_recovery_task_lock);
1702 list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1703 LASSERT(req->rq_reply_state == NULL);
1704 target_request_copy_put(req);
1709 EXPORT_SYMBOL(target_cleanup_recovery);
1711 /* obd_recovery_task_lock should be held */
1712 void target_cancel_recovery_timer(struct obd_device *obd)
1714 CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1715 del_timer(&obd->obd_recovery_timer);
1718 static void target_start_recovery_timer(struct obd_device *obd)
1720 if (obd->obd_recovery_start != 0)
1723 spin_lock(&obd->obd_dev_lock);
1724 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1725 spin_unlock(&obd->obd_dev_lock);
1729 LASSERT(obd->obd_recovery_timeout != 0);
1731 if (obd->obd_recovery_start != 0) {
1732 spin_unlock(&obd->obd_dev_lock);
1736 mod_timer(&obd->obd_recovery_timer,
1737 cfs_time_shift(obd->obd_recovery_timeout));
1738 obd->obd_recovery_start = ktime_get_real_seconds();
1739 spin_unlock(&obd->obd_dev_lock);
1741 LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1743 obd->obd_recovery_timeout / 60,
1744 obd->obd_recovery_timeout % 60,
1745 obd->obd_max_recoverable_clients,
1746 (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1747 (obd->obd_max_recoverable_clients == 1) ? "s": "");
1751 * extend recovery window.
1753 * if @extend is true, extend recovery window to have @drt remaining at least;
1754 * otherwise, make sure the recovery timeout value is not less than @drt.
1756 static void extend_recovery_timer(struct obd_device *obd, int drt,
1764 spin_lock(&obd->obd_dev_lock);
1765 if (!obd->obd_recovering || obd->obd_abort_recovery) {
1766 spin_unlock(&obd->obd_dev_lock);
1769 LASSERT(obd->obd_recovery_start != 0);
1771 now = ktime_get_real_seconds();
1772 to = obd->obd_recovery_timeout;
1773 end = obd->obd_recovery_start + to;
1776 if (extend && (drt > left)) {
1778 } else if (!extend && (drt > to)) {
1782 if (to > obd->obd_recovery_time_hard) {
1783 to = obd->obd_recovery_time_hard;
1784 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1785 obd->obd_name, to, extend);
1788 if (obd->obd_recovery_timeout < to) {
1789 obd->obd_recovery_timeout = to;
1790 end = obd->obd_recovery_start + to;
1791 mod_timer(&obd->obd_recovery_timer,
1792 cfs_time_shift(end - now));
1794 spin_unlock(&obd->obd_dev_lock);
1796 CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1797 obd->obd_name, (s64)(end - now));
1800 /* Reset the timer with each new client connection */
1802 * This timer is actually reconnect_timer, which is for making sure
1803 * the total recovery window is at least as big as my reconnect
1804 * attempt timing. So the initial recovery time_out will be set to
1805 * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1806 * from client is bigger than this, then the recovery time_out will
1807 * be extended to make sure the client could be reconnected, in the
1808 * process, the timeout from the new client should be ignored.
1812 check_and_start_recovery_timer(struct obd_device *obd,
1813 struct ptlrpc_request *req,
1816 int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1817 struct obd_device_target *obt = &obd->u.obt;
1819 if (!new_client && service_time)
1820 /* Teach server about old server's estimates, as first guess
1821 * at how long new requests will take. */
1822 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1825 target_start_recovery_timer(obd);
1827 /* Convert the service time to RPC timeout,
1828 * and reuse service_time to limit stack usage. */
1829 service_time = at_est2timeout(service_time);
1831 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1832 service_time < at_extra)
1833 service_time = at_extra;
1835 /* We expect other clients to timeout within service_time, then try
1836 * to reconnect, then try the failover server. The max delay between
1837 * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1838 service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1840 LASSERT(obt->obt_magic == OBT_MAGIC);
1841 service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1842 if (service_time > obd->obd_recovery_timeout && !new_client)
1843 extend_recovery_timer(obd, service_time, false);
1846 /** Health checking routines */
1847 static inline int exp_connect_healthy(struct obd_export *exp)
1849 return (exp->exp_in_recovery);
1852 /** if export done req_replay or has replay in queue */
1853 static inline int exp_req_replay_healthy(struct obd_export *exp)
1855 return (!exp->exp_req_replay_needed ||
1856 atomic_read(&exp->exp_replay_count) > 0);
1860 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1862 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1863 exp_req_replay_healthy(exp);
1866 /** if export done lock_replay or has replay in queue */
1867 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1869 return (!exp->exp_lock_replay_needed ||
1870 atomic_read(&exp->exp_replay_count) > 0);
1873 static inline int exp_vbr_healthy(struct obd_export *exp)
1875 return (!exp->exp_vbr_failed);
1878 static inline int exp_finished(struct obd_export *exp)
1880 return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1883 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1885 return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1889 static int check_for_next_transno(struct lu_target *lut)
1891 struct ptlrpc_request *req = NULL;
1892 struct obd_device *obd = lut->lut_obd;
1893 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1894 int wake_up = 0, connected, completed, queue_len;
1895 __u64 req_transno = 0;
1896 __u64 update_transno = 0;
1897 __u64 next_transno = 0;
1900 spin_lock(&obd->obd_recovery_task_lock);
1901 if (!list_empty(&obd->obd_req_replay_queue)) {
1902 req = list_entry(obd->obd_req_replay_queue.next,
1903 struct ptlrpc_request, rq_list);
1904 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1908 update_transno = distribute_txn_get_next_transno(tdtd);
1910 connected = atomic_read(&obd->obd_connected_clients);
1911 completed = connected - atomic_read(&obd->obd_req_replay_clients);
1912 queue_len = obd->obd_requests_queued_for_recovery;
1913 next_transno = obd->obd_next_recovery_transno;
1915 CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1916 "req_transno: %llu, next_transno: %llu\n",
1917 obd->obd_max_recoverable_clients, connected, completed,
1918 queue_len, req_transno, next_transno);
1920 if (obd->obd_abort_recovery) {
1921 CDEBUG(D_HA, "waking for aborted recovery\n");
1923 } else if (obd->obd_recovery_expired) {
1924 CDEBUG(D_HA, "waking for expired recovery\n");
1926 } else if (tdtd != NULL && req != NULL &&
1927 is_req_replayed_by_update(req)) {
1928 LASSERTF(req_transno < next_transno, "req_transno %llu"
1929 "next_transno%llu\n", req_transno, next_transno);
1930 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1933 } else if (req_transno == next_transno ||
1934 (update_transno != 0 && update_transno <= next_transno)) {
1935 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1937 } else if (queue_len > 0 &&
1938 queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1939 /** handle gaps occured due to lost reply or VBR */
1940 LASSERTF(req_transno >= next_transno,
1941 "req_transno: %llu, next_transno: %llu\n",
1942 req_transno, next_transno);
1944 "%s: waking for gap in transno, VBR is %s (skip: "
1945 "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1946 ", next_update %lld last_committed: %lld)\n",
1947 obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1948 next_transno, queue_len, completed, connected,
1949 req_transno, update_transno, obd->obd_last_committed);
1950 obd->obd_next_recovery_transno = req_transno;
1952 } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1953 CDEBUG(D_HA, "waking for completed recovery\n");
1955 } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1956 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1957 " by fail_lock, waking up (%lld)\n", next_transno);
1958 obd->obd_next_recovery_transno = req_transno;
1961 spin_unlock(&obd->obd_recovery_task_lock);
1965 static int check_for_next_lock(struct lu_target *lut)
1967 struct obd_device *obd = lut->lut_obd;
1970 spin_lock(&obd->obd_recovery_task_lock);
1971 if (!list_empty(&obd->obd_lock_replay_queue)) {
1972 CDEBUG(D_HA, "waking for next lock\n");
1974 } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1975 CDEBUG(D_HA, "waking for completed lock replay\n");
1977 } else if (obd->obd_abort_recovery) {
1978 CDEBUG(D_HA, "waking for aborted recovery\n");
1980 } else if (obd->obd_recovery_expired) {
1981 CDEBUG(D_HA, "waking for expired recovery\n");
1984 spin_unlock(&obd->obd_recovery_task_lock);
1990 * wait for recovery events,
1991 * check its status with help of check_routine
1992 * evict dead clients via health_check
1994 static int target_recovery_overseer(struct lu_target *lut,
1995 int (*check_routine)(struct lu_target *),
1996 int (*health_check)(struct obd_export *))
1998 struct obd_device *obd = lut->lut_obd;
1999 struct target_distribute_txn_data *tdtd;
2001 if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
2002 (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2003 __u64 next_update_transno = 0;
2005 /* Only abort the recovery if there are no update recovery
2006 * left in the queue */
2007 spin_lock(&obd->obd_recovery_task_lock);
2008 if (lut->lut_tdtd != NULL) {
2009 next_update_transno =
2010 distribute_txn_get_next_transno(lut->lut_tdtd);
2012 tdtd = lut->lut_tdtd;
2013 /* If next_update_transno == 0, it probably because
2014 * updatelog retrieve threads did not get any records
2015 * yet, let's wait those threads stopped */
2016 if (next_update_transno == 0) {
2017 struct l_wait_info lwi = { 0 };
2019 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2021 &tdtd->tdtd_recovery_threads_count) == 0,
2024 next_update_transno =
2025 distribute_txn_get_next_transno(
2030 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2031 obd->obd_next_recovery_transno = next_update_transno;
2032 spin_unlock(&obd->obd_recovery_task_lock);
2033 /* Disconnect unfinished exports from clients, and
2034 * keep connection from MDT to make sure the update
2035 * recovery will still keep trying until some one
2036 * manually abort the recovery */
2037 class_disconnect_stale_exports(obd,
2038 exp_finished_or_from_mdt);
2039 /* Abort all of replay and replay lock req from
2041 abort_req_replay_queue(obd);
2042 abort_lock_replay_queue(obd);
2043 CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2044 ")in the queue.\n", obd->obd_name,
2045 next_update_transno);
2047 obd->obd_abort_recovery = 1;
2048 spin_unlock(&obd->obd_recovery_task_lock);
2049 CWARN("%s recovery is aborted by hard timeout\n",
2054 while (wait_event_timeout(obd->obd_next_transno_waitq,
2056 msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2057 /* wait indefinitely for event, but don't trigger watchdog */;
2059 if (obd->obd_abort_recovery) {
2060 CWARN("recovery is aborted, evict exports in recovery\n");
2061 if (lut->lut_tdtd != NULL) {
2062 struct l_wait_info lwi = { 0 };
2064 tdtd = lut->lut_tdtd;
2065 /* Let's wait all of the update log recovery thread
2067 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2068 atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2070 /* Then abort the update recovery list */
2071 dtrq_list_destroy(lut->lut_tdtd);
2074 /** evict exports which didn't finish recovery yet */
2075 class_disconnect_stale_exports(obd, exp_finished);
2077 } else if (obd->obd_recovery_expired) {
2078 obd->obd_recovery_expired = 0;
2079 /** If some clients died being recovered, evict them */
2080 LCONSOLE_WARN("%s: recovery is timed out, "
2081 "evict stale exports\n", obd->obd_name);
2082 /** evict cexports with no replay in queue, they are stalled */
2083 class_disconnect_stale_exports(obd, health_check);
2085 /** continue with VBR */
2086 spin_lock(&obd->obd_dev_lock);
2087 obd->obd_version_recov = 1;
2088 spin_unlock(&obd->obd_dev_lock);
2090 * reset timer, recovery will proceed with versions now,
2091 * timeout is set just to handle reconnection delays
2093 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2094 /** Wait for recovery events again, after evicting bad clients */
2100 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2102 struct obd_device *obd = lut->lut_obd;
2103 struct ptlrpc_request *req = NULL;
2105 CDEBUG(D_HA, "Waiting for lock\n");
2106 if (target_recovery_overseer(lut, check_for_next_lock,
2107 exp_lock_replay_healthy))
2108 abort_lock_replay_queue(obd);
2110 spin_lock(&obd->obd_recovery_task_lock);
2111 if (!list_empty(&obd->obd_lock_replay_queue)) {
2112 req = list_entry(obd->obd_lock_replay_queue.next,
2113 struct ptlrpc_request, rq_list);
2114 list_del_init(&req->rq_list);
2115 spin_unlock(&obd->obd_recovery_task_lock);
2117 spin_unlock(&obd->obd_recovery_task_lock);
2118 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2119 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2120 /** evict exports failed VBR */
2121 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2126 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2128 struct ptlrpc_request *req = NULL;
2130 spin_lock(&obd->obd_recovery_task_lock);
2131 if (!list_empty(&obd->obd_final_req_queue)) {
2132 req = list_entry(obd->obd_final_req_queue.next,
2133 struct ptlrpc_request, rq_list);
2134 list_del_init(&req->rq_list);
2135 spin_unlock(&obd->obd_recovery_task_lock);
2136 if (req->rq_export->exp_in_recovery) {
2137 spin_lock(&req->rq_export->exp_lock);
2138 req->rq_export->exp_in_recovery = 0;
2139 spin_unlock(&req->rq_export->exp_lock);
2142 spin_unlock(&obd->obd_recovery_task_lock);
2147 static void handle_recovery_req(struct ptlrpc_thread *thread,
2148 struct ptlrpc_request *req,
2149 svc_handler_t handler)
2154 * export can be evicted during recovery, no need to handle replays for
2155 * it after that, discard such request silently
2157 if (req->rq_export->exp_disconnected)
2160 req->rq_session.lc_thread = thread;
2161 req->rq_svc_thread = thread;
2162 req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2164 /* thread context */
2165 lu_context_enter(&thread->t_env->le_ctx);
2167 lu_context_exit(&thread->t_env->le_ctx);
2169 /* don't reset timer for final stage */
2170 if (!exp_finished(req->rq_export)) {
2171 int to = obd_timeout;
2174 * Add request timeout to the recovery time so next request from
2175 * this client may come in recovery time
2178 struct ptlrpc_service_part *svcpt;
2180 svcpt = req->rq_rqbd->rqbd_svcpt;
2181 /* If the server sent early reply for this request,
2182 * the client will recalculate the timeout according to
2183 * current server estimate service time, so we will
2184 * use the maxium timeout here for waiting the client
2185 * sending the next req */
2186 to = max((int)at_est2timeout(
2187 at_get(&svcpt->scp_at_estimate)),
2188 (int)lustre_msg_get_timeout(req->rq_reqmsg));
2189 /* Add 2 net_latency, one for balance rq_deadline
2190 * (see ptl_send_rpc), one for resend the req to server,
2191 * Note: client will pack net_latency in replay req
2192 * (see ptlrpc_replay_req) */
2193 to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2195 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2200 /** Checking routines for recovery */
2201 static int check_for_recovery_ready(struct lu_target *lut)
2203 struct obd_device *obd = lut->lut_obd;
2204 unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2206 CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2207 " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2208 obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2209 obd->obd_recovery_expired);
2211 if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2212 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2213 if (clnts + obd->obd_stale_clients <
2214 obd->obd_max_recoverable_clients)
2218 if (lut->lut_tdtd != NULL) {
2219 if (!lut->lut_tdtd->tdtd_replay_ready &&
2220 !obd->obd_abort_recovery) {
2221 /* Let's extend recovery timer, in case the recovery
2222 * timer expired, and some clients got evicted */
2223 extend_recovery_timer(obd, obd->obd_recovery_timeout,
2225 CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2226 obd->obd_name, obd->obd_recovery_timeout);
2235 REQUEST_RECOVERY = 1,
2236 UPDATE_RECOVERY = 2,
2239 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2243 if (!list_empty(&obd->obd_req_replay_queue)) {
2244 struct ptlrpc_request *req;
2246 req = list_entry(obd->obd_req_replay_queue.next,
2247 struct ptlrpc_request, rq_list);
2248 transno = lustre_msg_get_transno(req->rq_reqmsg);
2254 static __u64 get_next_transno(struct lu_target *lut, int *type)
2256 struct obd_device *obd = lut->lut_obd;
2257 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2259 __u64 update_transno;
2262 transno = get_next_replay_req_transno(obd);
2264 *type = REQUEST_RECOVERY;
2269 update_transno = distribute_txn_get_next_transno(tdtd);
2270 if (transno == 0 || (transno >= update_transno &&
2271 update_transno != 0)) {
2272 transno = update_transno;
2274 *type = UPDATE_RECOVERY;
2281 * drop duplicate replay request
2283 * Because the operation has been replayed by update recovery, the request
2284 * with the same transno will be dropped and also notify the client to send
2285 * next replay request.
2287 * \param[in] env execution environment
2288 * \param[in] obd failover obd device
2289 * \param[in] req request to be dropped
2291 static void drop_duplicate_replay_req(struct lu_env *env,
2292 struct obd_device *obd,
2293 struct ptlrpc_request *req)
2295 DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2296 " update records are found.\n",
2297 lustre_msg_get_transno(req->rq_reqmsg),
2298 libcfs_nid2str(req->rq_peer.nid));
2300 /* Right now, only for MDS reint operation update replay and
2301 * normal request replay can have the same transno */
2302 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2303 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2304 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2305 if (likely(req->rq_export))
2306 target_committed_to_req(req);
2307 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2308 target_send_reply(req, req->rq_status, 0);
2310 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2311 libcfs_nid2str(req->rq_peer.nid));
2313 target_exp_dequeue_req_replay(req);
2314 target_request_copy_put(req);
2315 obd->obd_replayed_requests++;
2318 static void replay_request_or_update(struct lu_env *env,
2319 struct lu_target *lut,
2320 struct target_recovery_data *trd,
2321 struct ptlrpc_thread *thread)
2323 struct obd_device *obd = lut->lut_obd;
2324 struct ptlrpc_request *req = NULL;
2329 CDEBUG(D_HA, "Waiting for transno %lld\n",
2330 obd->obd_next_recovery_transno);
2332 /* Replay all of request and update by transno */
2334 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2336 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2338 /** It is needed to extend recovery window above
2339 * recovery_time_soft. Extending is possible only in the
2340 * end of recovery window (see more details in
2341 * handle_recovery_req()).
2343 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2345 if (target_recovery_overseer(lut, check_for_next_transno,
2346 exp_req_replay_healthy_or_from_mdt)) {
2347 abort_req_replay_queue(obd);
2348 abort_lock_replay_queue(obd);
2352 spin_lock(&obd->obd_recovery_task_lock);
2353 transno = get_next_transno(lut, &type);
2354 if (type == REQUEST_RECOVERY && transno != 0) {
2355 /* Drop replay request from client side, if the
2356 * replay has been executed by update with the
2358 req = list_entry(obd->obd_req_replay_queue.next,
2359 struct ptlrpc_request, rq_list);
2361 list_del_init(&req->rq_list);
2362 obd->obd_requests_queued_for_recovery--;
2363 spin_unlock(&obd->obd_recovery_task_lock);
2365 /* Let's check if the request has been redone by
2367 if (is_req_replayed_by_update(req)) {
2368 struct distribute_txn_replay_req *dtrq;
2370 dtrq = distribute_txn_lookup_finish_list(tdtd,
2372 LASSERT(dtrq != NULL);
2373 spin_lock(&tdtd->tdtd_replay_list_lock);
2374 list_del_init(&dtrq->dtrq_list);
2375 spin_unlock(&tdtd->tdtd_replay_list_lock);
2378 drop_duplicate_replay_req(env, obd, req);
2383 LASSERT(trd->trd_processing_task == current_pid());
2384 DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2385 lustre_msg_get_transno(req->rq_reqmsg),
2386 libcfs_nid2str(req->rq_peer.nid));
2388 handle_recovery_req(thread, req,
2389 trd->trd_recovery_handler);
2391 * bz18031: increase next_recovery_transno before
2392 * target_request_copy_put() will drop exp_rpc reference
2394 spin_lock(&obd->obd_recovery_task_lock);
2395 obd->obd_next_recovery_transno++;
2396 spin_unlock(&obd->obd_recovery_task_lock);
2397 target_exp_dequeue_req_replay(req);
2398 target_request_copy_put(req);
2399 obd->obd_replayed_requests++;
2400 } else if (type == UPDATE_RECOVERY && transno != 0) {
2401 struct distribute_txn_replay_req *dtrq;
2404 spin_unlock(&obd->obd_recovery_task_lock);
2406 LASSERT(tdtd != NULL);
2407 dtrq = distribute_txn_get_next_req(tdtd);
2408 lu_context_enter(&thread->t_env->le_ctx);
2409 rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2410 lu_context_exit(&thread->t_env->le_ctx);
2411 extend_recovery_timer(obd, obd_timeout, true);
2413 if (rc == 0 && dtrq->dtrq_xid != 0) {
2414 CDEBUG(D_HA, "Move x%llu t%llu"
2415 " to finish list\n", dtrq->dtrq_xid,
2416 dtrq->dtrq_master_transno);
2418 /* Add it to the replay finish list */
2419 spin_lock(&tdtd->tdtd_replay_list_lock);
2420 list_add(&dtrq->dtrq_list,
2421 &tdtd->tdtd_replay_finish_list);
2422 spin_unlock(&tdtd->tdtd_replay_list_lock);
2424 spin_lock(&obd->obd_recovery_task_lock);
2425 if (transno == obd->obd_next_recovery_transno)
2426 obd->obd_next_recovery_transno++;
2428 obd->obd_next_recovery_transno)
2429 obd->obd_next_recovery_transno =
2431 spin_unlock(&obd->obd_recovery_task_lock);
2436 spin_unlock(&obd->obd_recovery_task_lock);
2438 LASSERT(list_empty(&obd->obd_req_replay_queue));
2439 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2440 /** evict exports failed VBR */
2441 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2447 static int target_recovery_thread(void *arg)
2449 struct lu_target *lut = arg;
2450 struct obd_device *obd = lut->lut_obd;
2451 struct ptlrpc_request *req;
2452 struct target_recovery_data *trd = &obd->obd_recovery_data;
2453 unsigned long delta;
2455 struct ptlrpc_thread *thread = NULL;
2459 unshare_fs_struct();
2460 OBD_ALLOC_PTR(thread);
2466 OBD_FREE_PTR(thread);
2470 rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2472 OBD_FREE_PTR(thread);
2477 thread->t_env = env;
2478 thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2479 env->le_ctx.lc_thread = thread;
2480 tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2481 thread->t_watchdog = NULL;
2483 CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2485 trd->trd_processing_task = current_pid();
2487 spin_lock(&obd->obd_dev_lock);
2488 obd->obd_recovering = 1;
2489 spin_unlock(&obd->obd_dev_lock);
2490 complete(&trd->trd_starting);
2492 /* first of all, we have to know the first transno to replay */
2493 if (target_recovery_overseer(lut, check_for_recovery_ready,
2494 exp_connect_healthy)) {
2495 abort_req_replay_queue(obd);
2496 abort_lock_replay_queue(obd);
2497 if (lut->lut_tdtd != NULL)
2498 dtrq_list_destroy(lut->lut_tdtd);
2501 /* next stage: replay requests or update */
2503 CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2504 atomic_read(&obd->obd_req_replay_clients),
2505 obd->obd_next_recovery_transno);
2506 replay_request_or_update(env, lut, trd, thread);
2509 * The second stage: replay locks
2511 CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2512 atomic_read(&obd->obd_lock_replay_clients));
2513 while ((req = target_next_replay_lock(lut))) {
2514 LASSERT(trd->trd_processing_task == current_pid());
2515 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2516 libcfs_nid2str(req->rq_peer.nid));
2517 handle_recovery_req(thread, req,
2518 trd->trd_recovery_handler);
2519 target_request_copy_put(req);
2520 obd->obd_replayed_locks++;
2524 * The third stage: reply on final pings, at this moment all clients
2525 * must have request in final queue
2527 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2528 CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2529 /** Update server last boot epoch */
2530 tgt_boot_epoch_update(lut);
2531 /* We drop recoverying flag to forward all new requests
2532 * to regular mds_handle() since now */
2533 spin_lock(&obd->obd_dev_lock);
2534 obd->obd_recovering = obd->obd_abort_recovery = 0;
2535 spin_unlock(&obd->obd_dev_lock);
2536 spin_lock(&obd->obd_recovery_task_lock);
2537 target_cancel_recovery_timer(obd);
2538 spin_unlock(&obd->obd_recovery_task_lock);
2539 while ((req = target_next_final_ping(obd))) {
2540 LASSERT(trd->trd_processing_task == current_pid());
2541 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2542 libcfs_nid2str(req->rq_peer.nid));
2543 handle_recovery_req(thread, req,
2544 trd->trd_recovery_handler);
2545 /* Because the waiting client can not send ping to server,
2546 * so we need refresh the last_request_time, to avoid the
2547 * export is being evicted */
2548 ptlrpc_update_export_timer(req->rq_export, 0);
2549 target_request_copy_put(req);
2552 delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2553 CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2554 delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2555 if (delta > OBD_RECOVERY_TIME_SOFT) {
2556 CWARN("too long recovery - read logs\n");
2557 libcfs_debug_dumplog();
2560 target_finish_recovery(lut);
2562 lu_context_fini(&env->le_ctx);
2563 trd->trd_processing_task = 0;
2564 complete(&trd->trd_finishing);
2566 tgt_io_thread_done(thread);
2567 OBD_FREE_PTR(thread);
2572 static int target_start_recovery_thread(struct lu_target *lut,
2573 svc_handler_t handler)
2575 struct obd_device *obd = lut->lut_obd;
2577 struct target_recovery_data *trd = &obd->obd_recovery_data;
2580 memset(trd, 0, sizeof(*trd));
2581 init_completion(&trd->trd_starting);
2582 init_completion(&trd->trd_finishing);
2583 trd->trd_recovery_handler = handler;
2585 rc = server_name2index(obd->obd_name, &index, NULL);
2589 if (!IS_ERR(kthread_run(target_recovery_thread,
2590 lut, "tgt_recover_%d", index))) {
2591 wait_for_completion(&trd->trd_starting);
2592 LASSERT(obd->obd_recovering != 0);
2600 void target_stop_recovery_thread(struct obd_device *obd)
2602 if (obd->obd_recovery_data.trd_processing_task > 0) {
2603 struct target_recovery_data *trd = &obd->obd_recovery_data;
2604 /** recovery can be done but postrecovery is not yet */
2605 spin_lock(&obd->obd_dev_lock);
2606 if (obd->obd_recovering) {
2607 CERROR("%s: Aborting recovery\n", obd->obd_name);
2608 obd->obd_abort_recovery = 1;
2609 wake_up(&obd->obd_next_transno_waitq);
2611 spin_unlock(&obd->obd_dev_lock);
2612 wait_for_completion(&trd->trd_finishing);
2615 EXPORT_SYMBOL(target_stop_recovery_thread);
2617 void target_recovery_fini(struct obd_device *obd)
2619 class_disconnect_exports(obd);
2620 target_stop_recovery_thread(obd);
2621 target_cleanup_recovery(obd);
2623 EXPORT_SYMBOL(target_recovery_fini);
2625 static void target_recovery_expired(unsigned long castmeharder)
2627 struct obd_device *obd = (struct obd_device *)castmeharder;
2628 CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2629 " after %llus (%d clients connected)\n",
2630 obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2631 (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2632 atomic_read(&obd->obd_connected_clients));
2634 obd->obd_recovery_expired = 1;
2635 wake_up(&obd->obd_next_transno_waitq);
2638 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2640 struct obd_device *obd = lut->lut_obd;
2642 if (lut->lut_bottom->dd_rdonly)
2645 if (obd->obd_max_recoverable_clients == 0) {
2646 /** Update server last boot epoch */
2647 tgt_boot_epoch_update(lut);
2651 CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2652 "last_transno %llu\n", obd->obd_name,
2653 obd->obd_max_recoverable_clients, obd->obd_last_committed);
2654 LASSERT(obd->obd_stopping == 0);
2655 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2656 obd->obd_recovery_start = 0;
2657 obd->obd_recovery_end = 0;
2659 setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2660 (unsigned long)obd);
2661 target_start_recovery_thread(lut, handler);
2663 EXPORT_SYMBOL(target_recovery_init);
2665 static int target_process_req_flags(struct obd_device *obd,
2666 struct ptlrpc_request *req)
2668 struct obd_export *exp = req->rq_export;
2669 LASSERT(exp != NULL);
2670 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2671 /* client declares he's ready to replay locks */
2672 spin_lock(&exp->exp_lock);
2673 if (exp->exp_req_replay_needed) {
2674 exp->exp_req_replay_needed = 0;
2675 spin_unlock(&exp->exp_lock);
2677 LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2678 atomic_dec(&obd->obd_req_replay_clients);
2680 spin_unlock(&exp->exp_lock);
2683 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2684 /* client declares he's ready to complete recovery
2685 * so, we put the request on th final queue */
2686 spin_lock(&exp->exp_lock);
2687 if (exp->exp_lock_replay_needed) {
2688 exp->exp_lock_replay_needed = 0;
2689 spin_unlock(&exp->exp_lock);
2691 LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2692 atomic_dec(&obd->obd_lock_replay_clients);
2694 spin_unlock(&exp->exp_lock);
2700 int target_queue_recovery_request(struct ptlrpc_request *req,
2701 struct obd_device *obd)
2703 __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2704 struct ptlrpc_request *reqiter;
2708 if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2709 /* Processing the queue right now, don't re-add. */
2713 target_process_req_flags(obd, req);
2715 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2716 /* client declares he's ready to complete recovery
2717 * so, we put the request on th final queue */
2718 target_request_copy_get(req);
2719 DEBUG_REQ(D_HA, req, "queue final req");
2720 wake_up(&obd->obd_next_transno_waitq);
2721 spin_lock(&obd->obd_recovery_task_lock);
2722 if (obd->obd_recovering) {
2723 struct ptlrpc_request *tmp;
2724 struct ptlrpc_request *duplicate = NULL;
2726 if (likely(!req->rq_export->exp_replay_done)) {
2727 req->rq_export->exp_replay_done = 1;
2728 list_add_tail(&req->rq_list,
2729 &obd->obd_final_req_queue);
2730 spin_unlock(&obd->obd_recovery_task_lock);
2734 /* XXX O(n), but only happens if final ping is
2735 * timed out, probably reorganize the list as
2736 * a hash list later */
2737 list_for_each_entry_safe(reqiter, tmp,
2738 &obd->obd_final_req_queue,
2740 if (reqiter->rq_export == req->rq_export) {
2741 list_del_init(&reqiter->rq_list);
2742 duplicate = reqiter;
2747 list_add_tail(&req->rq_list,
2748 &obd->obd_final_req_queue);
2749 req->rq_export->exp_replay_done = 1;
2750 spin_unlock(&obd->obd_recovery_task_lock);
2752 if (duplicate != NULL) {
2753 DEBUG_REQ(D_HA, duplicate,
2754 "put prev final req\n");
2755 target_request_copy_put(duplicate);
2759 spin_unlock(&obd->obd_recovery_task_lock);
2760 target_request_copy_put(req);
2761 RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2764 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2765 /* client declares he's ready to replay locks */
2766 target_request_copy_get(req);
2767 DEBUG_REQ(D_HA, req, "queue lock replay req");
2768 wake_up(&obd->obd_next_transno_waitq);
2769 spin_lock(&obd->obd_recovery_task_lock);
2770 LASSERT(obd->obd_recovering);
2771 /* usually due to recovery abort */
2772 if (!req->rq_export->exp_in_recovery) {
2773 spin_unlock(&obd->obd_recovery_task_lock);
2774 target_request_copy_put(req);
2777 LASSERT(req->rq_export->exp_lock_replay_needed);
2778 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2779 spin_unlock(&obd->obd_recovery_task_lock);
2783 /* CAVEAT EMPTOR: The incoming request message has been swabbed
2784 * (i.e. buflens etc are in my own byte order), but type-dependent
2785 * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2788 INIT_LIST_HEAD(&req->rq_list);
2789 DEBUG_REQ(D_HA, req, "not queueing");
2793 /* If we're processing the queue, we want don't want to queue this
2796 * Also, if this request has a transno less than the one we're waiting
2797 * for, we should process it now. It could (and currently always will)
2798 * be an open request for a descriptor that was opened some time ago.
2800 * Also, a resent, replayed request that has already been
2801 * handled will pass through here and be processed immediately.
2803 CDEBUG(D_HA, "Next recovery transno: %llu"
2804 ", current: %llu, replaying\n",
2805 obd->obd_next_recovery_transno, transno);
2807 /* If the request has been replayed by update replay, then sends this
2808 * request to the recovery thread (replay_request_or_update()), where
2809 * it will be handled */
2810 spin_lock(&obd->obd_recovery_task_lock);
2811 if (transno < obd->obd_next_recovery_transno &&
2812 !is_req_replayed_by_update(req)) {
2813 /* Processing the queue right now, don't re-add. */
2814 LASSERT(list_empty(&req->rq_list));
2815 spin_unlock(&obd->obd_recovery_task_lock);
2818 spin_unlock(&obd->obd_recovery_task_lock);
2820 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2823 target_request_copy_get(req);
2824 if (!req->rq_export->exp_in_recovery) {
2825 target_request_copy_put(req);
2828 LASSERT(req->rq_export->exp_req_replay_needed);
2830 if (target_exp_enqueue_req_replay(req)) {
2831 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2832 target_request_copy_put(req);
2837 spin_lock(&obd->obd_recovery_task_lock);
2838 LASSERT(obd->obd_recovering);
2839 list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2840 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2841 list_add_tail(&req->rq_list, &reqiter->rq_list);
2846 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2848 DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2849 "has been claimed by another client");
2850 spin_unlock(&obd->obd_recovery_task_lock);
2851 target_exp_dequeue_req_replay(req);
2852 target_request_copy_put(req);
2858 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2860 obd->obd_requests_queued_for_recovery++;
2861 spin_unlock(&obd->obd_recovery_task_lock);
2862 wake_up(&obd->obd_next_transno_waitq);
2866 int target_handle_ping(struct ptlrpc_request *req)
2868 obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2869 return req_capsule_server_pack(&req->rq_pill);
2872 void target_committed_to_req(struct ptlrpc_request *req)
2874 struct obd_export *exp = req->rq_export;
2876 if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2877 lustre_msg_set_last_committed(req->rq_repmsg,
2878 exp->exp_last_committed);
2880 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2881 "%d)", exp->exp_obd->obd_no_transno,
2882 req->rq_repmsg == NULL);
2884 CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2885 exp->exp_last_committed, req->rq_transno, req->rq_xid);
2888 #endif /* HAVE_SERVER_SUPPORT */
2891 * Packs current SLV and Limit into \a req.
2893 int target_pack_pool_reply(struct ptlrpc_request *req)
2895 struct obd_device *obd;
2898 /* Check that we still have all structures alive as this may
2899 * be some late RPC at shutdown time. */
2900 if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2901 !exp_connect_lru_resize(req->rq_export))) {
2902 lustre_msg_set_slv(req->rq_repmsg, 0);
2903 lustre_msg_set_limit(req->rq_repmsg, 0);
2907 /* OBD is alive here as export is alive, which we checked above. */
2908 obd = req->rq_export->exp_obd;
2910 read_lock(&obd->obd_pool_lock);
2911 lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2912 lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2913 read_unlock(&obd->obd_pool_lock);
2918 static int target_send_reply_msg(struct ptlrpc_request *req,
2919 int rc, int fail_id)
2921 if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2922 DEBUG_REQ(D_ERROR, req, "dropping reply");
2925 /* We can have a null rq_reqmsg in the event of bad signature or
2926 * no context when unwrapping */
2927 if (req->rq_reqmsg &&
2928 unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2929 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2933 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2934 req->rq_status = rc;
2935 return ptlrpc_send_error(req, 1);
2937 DEBUG_REQ(D_NET, req, "sending reply");
2940 return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2943 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2945 struct ptlrpc_service_part *svcpt;
2947 struct ptlrpc_reply_state *rs;
2948 struct obd_export *exp;
2951 if (req->rq_no_reply) {
2956 svcpt = req->rq_rqbd->rqbd_svcpt;
2957 rs = req->rq_reply_state;
2958 if (rs == NULL || !rs->rs_difficult) {
2960 target_send_reply_msg (req, rc, fail_id);
2965 /* must be an export if locks saved */
2966 LASSERT(req->rq_export != NULL);
2967 /* req/reply consistent */
2968 LASSERT(rs->rs_svcpt == svcpt);
2971 LASSERT(!rs->rs_scheduled);
2972 LASSERT(!rs->rs_scheduled_ever);
2973 LASSERT(!rs->rs_handled);
2974 LASSERT(!rs->rs_on_net);
2975 LASSERT(rs->rs_export == NULL);
2976 LASSERT(list_empty(&rs->rs_obd_list));
2977 LASSERT(list_empty(&rs->rs_exp_list));
2979 exp = class_export_get(req->rq_export);
2981 /* disable reply scheduling while I'm setting up */
2982 rs->rs_scheduled = 1;
2984 rs->rs_xid = req->rq_xid;
2985 rs->rs_transno = req->rq_transno;
2986 rs->rs_export = exp;
2987 rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg);
2989 spin_lock(&exp->exp_uncommitted_replies_lock);
2990 CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2991 rs->rs_transno, exp->exp_last_committed);
2992 if (rs->rs_transno > exp->exp_last_committed) {
2993 /* not committed already */
2994 list_add_tail(&rs->rs_obd_list,
2995 &exp->exp_uncommitted_replies);
2997 spin_unlock(&exp->exp_uncommitted_replies_lock);
2999 spin_lock(&exp->exp_lock);
3000 list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3001 spin_unlock(&exp->exp_lock);
3003 netrc = target_send_reply_msg(req, rc, fail_id);
3005 spin_lock(&svcpt->scp_rep_lock);
3007 atomic_inc(&svcpt->scp_nreps_difficult);
3010 /* error sending: reply is off the net. Also we need +1
3011 * reply ref until ptlrpc_handle_rs() is done
3012 * with the reply state (if the send was successful, there
3013 * would have been +1 ref for the net, which
3014 * reply_out_callback leaves alone) */
3016 ptlrpc_rs_addref(rs);
3019 spin_lock(&rs->rs_lock);
3020 if (rs->rs_transno <= exp->exp_last_committed ||
3021 (!rs->rs_on_net && !rs->rs_no_ack) ||
3022 list_empty(&rs->rs_exp_list) || /* completed already */
3023 list_empty(&rs->rs_obd_list)) {
3024 CDEBUG(D_HA, "Schedule reply immediately\n");
3025 ptlrpc_dispatch_difficult_reply(rs);
3027 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3028 rs->rs_scheduled = 0; /* allow notifier to schedule */
3030 spin_unlock(&rs->rs_lock);
3031 spin_unlock(&svcpt->scp_rep_lock);
3035 enum ldlm_mode lck_compat_array[] = {
3036 [LCK_EX] = LCK_COMPAT_EX,
3037 [LCK_PW] = LCK_COMPAT_PW,
3038 [LCK_PR] = LCK_COMPAT_PR,
3039 [LCK_CW] = LCK_COMPAT_CW,
3040 [LCK_CR] = LCK_COMPAT_CR,
3041 [LCK_NL] = LCK_COMPAT_NL,
3042 [LCK_GROUP] = LCK_COMPAT_GROUP,
3043 [LCK_COS] = LCK_COMPAT_COS,
3047 * Rather arbitrary mapping from LDLM error codes to errno values. This should
3048 * not escape to the user level.
3050 int ldlm_error2errno(enum ldlm_error error)
3056 case ELDLM_LOCK_MATCHED:
3059 case ELDLM_LOCK_CHANGED:
3062 case ELDLM_LOCK_ABORTED:
3065 case ELDLM_LOCK_REPLACED:
3068 case ELDLM_NO_LOCK_DATA:
3071 case ELDLM_NAMESPACE_EXISTS:
3074 case ELDLM_BAD_NAMESPACE:
3078 if (((int)error) < 0) { /* cast to signed type */
3079 result = error; /* as ldlm_error can be unsigned */
3081 CERROR("Invalid DLM result code: %d\n", error);
3087 EXPORT_SYMBOL(ldlm_error2errno);
3090 * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3092 enum ldlm_error ldlm_errno2error(int err_no)
3101 error = ELDLM_LOCK_CHANGED;
3104 error = ELDLM_LOCK_ABORTED;
3107 error = ELDLM_LOCK_REPLACED;
3110 error = ELDLM_NO_LOCK_DATA;
3113 error = ELDLM_NAMESPACE_EXISTS;
3116 error = ELDLM_BAD_NAMESPACE;
3124 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3125 void ldlm_dump_export_locks(struct obd_export *exp)
3127 spin_lock(&exp->exp_locks_list_guard);
3128 if (!list_empty(&exp->exp_locks_list)) {
3129 struct ldlm_lock *lock;
3131 CERROR("dumping locks for export %p,"
3132 "ignore if the unmount doesn't hang\n", exp);
3133 list_for_each_entry(lock, &exp->exp_locks_list,
3135 LDLM_ERROR(lock, "lock:");
3137 spin_unlock(&exp->exp_locks_list_guard);
3141 #ifdef HAVE_SERVER_SUPPORT
3142 static int target_bulk_timeout(void *data)
3145 /* We don't fail the connection here, because having the export
3146 * killed makes the (vital) call to commitrw very sad.
3151 static inline const char *bulk2type(struct ptlrpc_request *req)
3153 if (req->rq_bulk_read)
3155 if (req->rq_bulk_write)
3160 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3161 struct l_wait_info *lwi)
3163 struct ptlrpc_request *req = desc->bd_req;
3164 time_t start = cfs_time_current_sec();
3170 /* If there is eviction in progress, wait for it to finish. */
3171 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3172 *lwi = LWI_INTR(NULL, NULL);
3173 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3174 !atomic_read(&exp->exp_obd->
3175 obd_evict_inprogress),
3179 /* Check if client was evicted or reconnected already. */
3180 if (exp->exp_failed ||
3181 exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3184 if (req->rq_bulk_read)
3185 rc = sptlrpc_svc_wrap_bulk(req, desc);
3187 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3188 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3189 else /* old version, bulk matchbits is rq_xid */
3190 req->rq_mbits = req->rq_xid;
3193 rc = ptlrpc_start_bulk_transfer(desc);
3197 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3198 bulk2type(req), rc);
3202 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3203 ptlrpc_abort_bulk(desc);
3207 /* limit actual bulk transfer to bulk_timeout seconds */
3208 deadline = start + bulk_timeout;
3209 if (deadline > req->rq_deadline)
3210 deadline = req->rq_deadline;
3213 long timeoutl = deadline - cfs_time_current_sec();
3214 cfs_duration_t timeout = timeoutl <= 0 ?
3215 CFS_TICK : cfs_time_seconds(timeoutl);
3218 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3219 target_bulk_timeout, desc);
3220 rc = l_wait_event(desc->bd_waitq,
3221 !ptlrpc_server_bulk_active(desc) ||
3224 lustre_msg_get_conn_cnt(req->rq_reqmsg),
3226 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3227 /* Wait again if we changed rq_deadline. */
3228 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3229 deadline = start + bulk_timeout;
3230 if (deadline > rq_deadline)
3231 deadline = rq_deadline;
3232 } while ((rc == -ETIMEDOUT) &&
3233 (deadline > cfs_time_current_sec()));
3235 if (rc == -ETIMEDOUT) {
3236 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3237 bulk2type(req), deadline - start,
3238 cfs_time_current_sec() - deadline);
3239 ptlrpc_abort_bulk(desc);
3240 } else if (exp->exp_failed) {
3241 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3244 ptlrpc_abort_bulk(desc);
3245 } else if (exp->exp_conn_cnt >
3246 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3247 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3249 /* We don't reply anyway. */
3251 ptlrpc_abort_bulk(desc);
3252 } else if (desc->bd_failure) {
3253 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3255 /* XXX should this be a different errno? */
3258 if (req->rq_bulk_write)
3259 rc = sptlrpc_svc_unwrap_bulk(req, desc);
3260 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3261 DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3262 bulk2type(req), desc->bd_nob_transferred,
3264 /* XXX should this be a different errno? */
3271 EXPORT_SYMBOL(target_bulk_io);
3273 #endif /* HAVE_SERVER_SUPPORT */